Author’s Note
I have been interested in the problem considered in this manuscript since the completion of [7]. It seemed clear that the transfer operator strategy in [7] was the right way to approach the problem, but also that there were significant modelspecific challenges in implementing it in the context of sparse random matrices. For various reasons, my attention became focused on other things, but I always intended to return to this question.
Shortly after the release of GPT-5.6 Sol, I became curious about whether the model could complete the details of the argument given an appropriate high-level sketch of the proof, and indeed it could (after exchanging many messages). I had hoped to produce a polished, human-written version of this manuscript, but given the current pace of mathematical progress and recent improvements in LLM writing abilities, it does not feel appropriate to delay the dissemination of these results for the rather long time that would take. The current document, while not perfect, should be accessible to those with the relevant background. Further, given the implications these results have for certain predictions in the physics literature, as described in the introductory material below, I thought it was best to release them now.
With the exception of this note and small tweaks to the introduction to add some recent references, the entire document is machine-written. I have taken care to supervise this writing to ensure that the proofs meet a minimum standard of readability and that previous literature is appropriately cited. I welcome suggestions concerning mathematical corrections, omitted citations, and attribution of ideas.
P.L. was partially supported by NSF grant DMS-2450004.
Introduction
Background
Disordered quantum systems often exhibit one of two contrasting behaviors. In a localized phase, eigenfunctions are concentrated in bounded regions of space and the system is an insulator. In a delocalized phase, eigenfunctions are spread out and transport is possible. Anderson proposed in [26] that sufficiently strong disorder localizes all eigenfunctions of a tight-binding Hamiltonian. At weaker disorder, localized and delocalized states may coexist at different energies, separated by mobility edges , a notion introduced by Mott [103]; see also [65] and the review [71]. Localization is by now well understood mathematically in many regimes, for instance through the fractional moment method of Aizenman and Molchanov [8, 9]. We refer to [13] for an account. Delocalization is much harder to prove, and tree graphs are the main setting in which both phases, and the transition between them, are accessible.
Trees. The Anderson model on the Bethe lattice was studied by Abou-Chacra, Anderson, and Thouless in [2], who wrote self-consistent equations for the local Green function that are exact on the tree. They predicted that the localized phase corresponds to a real solution of these equations, and they located its stability limit by linearizing the equations around this solution. The case of energies near the band edges was treated in [1]. Kunz and Souillard obtained exact results on the transition between extended and localized states on the Bethe lattice [93], and Miller and Derrida studied localization on trees through a weak-disorder expansion [99]. This linear stability criterion has since been reformulated in several ways: through the free energy of a directed polymer on the tree [40, 62, 89, 102], through supersymmetric field theory [66, 100, 136], and through the largest eigenvalue of an integral operator acting on real Green functions [43, 109, 127]. In these formulations the criterion is evaluated at the symmetric point of a symmetry of the linearized kernel, which goes back to [2]; see [109] for a discussion. A pedagogical account of the cavity approach, including its analysis at large connectivity, is given in [111]. Biroli, Semerjian, and Tarzia computed the mobility edge and the density of states of the Anderson model on Bethe lattices with the cavity method in [43]. More recently, the delocalized phase on trees and random regular graphs has been studied in detail, in particular the question of its ergodicity [39, 40, 44, 45, 59, 89, 130, 131]. The critical properties of the transition and the localized phase near it were studied in [114, 128].
On the mathematical side, Klein proved that the Anderson model on the Bethe lattice has absolutely continuous spectrum at weak disorder [87, 88]; other proofs and extensions were given in [14, 63, 72, 80, 84]. Aizenman and Warzel developed a resonance mechanism for delocalization in [11, 12]. They proved that the spectrum is absolutely continuous wherever a certain free-energy function, denoted φ ( 1 ; E ) \varphi(1; E) φ ( 1 ; E ) in [12] and describing the decay of fractional moments of the Green function along paths, exceeds the logarithm of the inverse branching number. As a consequence, at weak disorder the absolutely continuous spectrum extends well beyond the band of the hopping operator, into a Lifshitz-tail regime [95] of very small density of states (see [30] for Lifshitz tails on the Bethe lattice). For bounded potentials there is no mobility edge of the usual kind at weak disorder [10]; see also the survey [135]. Bapst analyzed this free energy at large connectivity in [29] and confirmed the prediction of [2] for the critical disorder. Building on these works, Aggarwal and Lopatto proved in [4] that mobility edges exist for the Anderson model on the Bethe lattice with unbounded potential at large degree, near the locations predicted in [2]. Liu and Lopatto transferred this result to random regular graphs in [96].
For random trees, much less is known. The spectral measure of the unweighted adjacency operator of the Poisson–Galton–Watson tree, the local limit of the Erdős–Rényi graph, has a dense set of atoms [36, 119], and its continuous part was studied in [53, 57, 120]. Absolutely continuous spectrum for multi-type Galton–Watson trees close to trees of finite cone type was established by Keller [83]. Arras and Bordenave proved in [28] that, conditionally on survival, the adjacency operator of a Poisson–Galton–Watson tree with large mean degree has a nontrivial absolutely continuous spectral component, which carries most of the spectral mass.
Sparse random matrices. The spectra of sparse random matrices, with a bounded mean number of nonzero entries per row, were first studied in physics in [115, 116]. Fyodorov and Mirlin showed that such matrices undergo a localization transition, and derived equations for a mobility edge and for the inverse participation ratio in [73, 74, 101]; numerical studies include [69, 70]. Biroli and Monasson proposed in [38] that localized states in the tails of the spectra of random graphs are created by rare local defects, such as vertices of abnormally large degree; see also [121], and [132] for the analogous picture for non-Hermitian sparse matrices. In the limit of high connectivity, Silva and Metz solved the resolvent equations of sparse random graphs analytically and found that all eigenvectors are extended in this limit [122]. The cavity method was developed for sparse matrices in [91, 117, 126], and mobility edges in the tails of the spectra of sparse matrices were computed numerically in [91, 92, 98, 125]. Tarzia studied the tails of the spectrum of Erdős–Rényi graphs in the critical regime in [129]. Most relevant for us, Cugliandolo, Schehr, Tarzia, and Venturelli studied in [58] the giant component of Erdős–Rényi graphs with real Gaussian edge weights, which is the real symmetric analogue of the model considered here. They found a delocalized but weakly multifractal phase at small p p p , separated by a mobility edge from localized states in the tails. The multifractality of extended states in weighted random networks was also studied by Baron and Rogers in [31].
The mathematical literature on sparse Erdős–Rényi graphs is extensive when the mean degree grows with the number of vertices; see for example [33, 34, 67, 76]. In the critical regime where the mean degree is of order log n \log n log n , Alt, Ducatez, and Knowles proved in a series of works [18–22] that eigenvectors are delocalized in the middle of the spectrum and localized near high-degree vertices at the spectral edges, with a mobility edge between the two regimes. See also the recent work [78]. For bounded mean degree, the global eigenvalue distribution of weighted random graphs was studied in [50, 85], and the extreme eigenvalues and their eigenvectors in [55, 79]. For the model (1.1) itself, Afanasiev showed in [3], using supersymmetry, that for p > 2 p > 2 p > 2 the second correlation function of characteristic polynomials has a GUE-type limit inside an energy interval and factorizes outside it.
Heavy-tailed matrices. A mobility edge was predicted by Cizeau and Bouchaud for Lévy matrices, whose entries are in the domain of attraction of an α \alpha α -stable law with α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) [56]. Biroli, Bouchaud, and Potters showed that the largest eigenvalues of heavy-tailed matrices are governed by their largest entries [42]. Tarquini, Biroli, and Tarzia computed its location in [127], through the stability of the real solution of the cavity equations, and Biroli and Tarzia used cavity equations to determine the phase diagram of a Lévy version of the Rosenzweig–Porter ensemble [41]. The eigenvalues and eigenvectors of Lévy matrices were studied rigorously in [5, 6, 32, 48, 49]. Aggarwal, Bordenave, and Lopatto proved in [7] that Lévy matrices have a mobility edge, in the sense of the imaginary part of the resolvent of their local limit and of eigenvector participation ratios, at the location predicted in [127], and that it is unique for α \alpha α close to 0 or 1.
Recent advances. Several recent preprints produced by an internal OpenAI large language model announce results on longstanding problems in localization and spectral universality. For the Anderson model with independent uniform site potentials, these include purely absolutely continuous spectrum on a nonempty energy interval at sufficiently weak disorder on Z d \mathbb{Z}^{d} Z d for every fixed d ≥ 3 d \ge3 d ≥ 3 [104], and pure-point spectrum throughout the spectrum at every positive disorder strength on Z 2 \mathbb{Z}^{2} Z 2 [107]. For random regular graphs of every fixed degree d ≥ 3 d \ge3 d ≥ 3 , further preprints announce fixed-energy GOE bulk universality for the adjacency matrix [106] and for the Anderson model with sufficiently weak independent uniform diagonal disorder [105]. The fixed-energy GOE statistics result for the adjacency matrix was obtained independently and simultaneously in [77].
This paper. In this article we establish that the sparse Hermitian matrix with complex Gaussian weights on G ( n , p / n ) G(n,p/n) G ( n , p / n ) has exactly one positive mobility edge in its local limit, for every fixed sufficiently large p p p , and we determine its location to leading order. Its location, of order p \sqrt{p} p , is far in the tail of the spectrum, which contradicts the large-p p p behavior predicted in the physics literature; we discuss this in § 2.5 . We also derive consequences for the giant component of the finite graph.
Model and main results
The model
Fix p > 1 p>1 p > 1 . For each integer n > p n>p n > p , let G n G_n G n be an Erdős–Rényi random graph on the vertex set { 1 , … , n } \{1,\ldots,n\} { 1 , … , n } in which each edge is present independently with probability p / n p/n p / n . Conditional on G n G_n G n , let ( w i j ) 1 ≤ i < j ≤ n (w_{ij})_{1\leq i<j\leq n} ( w ij ) 1 ≤ i < j ≤ n be independent standard complex Gaussian random variables, so that each w i j w_{ij} w ij has density π − 1 e − ∣ z ∣ 2 \pi^{-1}e^{-|z|^2} π − 1 e − ∣ z ∣ 2 with respect to Lebesgue measure on C \mathbb{C} C . We consider the n × n n\times n n × n Hermitian matrix H n H_n H n defined by
H n ( i , j ) = 1 { i j ∈ E ( G n ) } p w i j , H n ( j , i ) = H n ( i , j ) ‾ , H n ( i , i ) = 0 , (1) H_n(i,j)=\frac{\mathbf{1}_{\{ij\in E(G_n)\}}}{\sqrt{p}}w_{ij},\qquad H_n(j,i)=\overline{H_n(i,j)},\qquad H_n(i,i)=0, \tag*{(1)} H n ( i , j ) = p 1 { ij ∈ E ( G n )} w ij , H n ( j , i ) = H n ( i , j ) , H n ( i , i ) = 0 , (1)
for 1 ≤ i < j ≤ n 1\leq i<j\leq n 1 ≤ i < j ≤ n . Throughout the paper, p p p is fixed while n n n tends to infinity, and statements about large p p p concern the resulting limiting objects.
Since p > 1 p>1 p > 1 , the graph G n G_n G n has with high probability a unique giant component, which contains a positive fraction of the vertices, together with linearly many finite components. We write g c \mathrm{gc} gc for the vertex set of the giant component, that is, of the largest connected component of G n G_n G n (if there are several, the one containing the smallest vertex; see Definition 3.1 ), N n = ∣ g c ∣ N_n=|\mathrm{gc}| N n = ∣ gc ∣ for its size, and H n g c H_n^{\mathrm{gc}} H n gc for the restriction of H n H_n H n to g c \mathrm{gc} gc . Since the finite components produce a positive density of localized eigenvectors for trivial reasons, the natural object of study is the giant component.
The local weak limit of H n H_n H n is described by a random tree. Let T \mathcal{T} T be a Poisson–Galton–Watson tree with mean offspring p p p and root o o o , and attach to each edge e e e an independent weight h e = w e / p h_e=w_e/\sqrt{p} h e = w e / p , where the w e w_e w e are standard complex Gaussian variables; we set h v u = h u v ‾ h_{vu}=\overline{h_{uv}} h v u = h uv . Let H \mathcal{H} H be the weighted adjacency operator of T \mathcal{T} T ,
( H f ) ( v ) = ∑ u ∼ v h v u f ( u ) . (\mathcal{H}f)(v)=\sum_{u\sim v}h_{vu}f(u). ( H f ) ( v ) = u ∼ v ∑ h v u f ( u ) .
Although the degrees in T \mathcal{T} T are unbounded, H \mathcal{H} H is almost surely essentially self-adjoint on finitely supported vectors (Theorem 3.10 ), and we use the same notation for its closure. A uniformly chosen vertex of G n G_n G n converges locally to the root of ( T , ( h e ) ) (\mathcal{T},(h_e)) ( T , ( h e )) , and a uniformly chosen vertex of the giant component converges locally to the root of the same tree conditioned on the event S S S that it is infinite. We write P \mathbb{P} P for the law of the weighted tree, ϱ p = P ( S ) \varrho_p=\mathbb{P}(S) ϱ p = P ( S ) for its survival probability, and P ∞ = P ( ⋅ ∣ S ) \mathbb{P}^{\infty}=\mathbb{P}(\mathord{\cdot}\mid S) P ∞ = P ( ⋅ ∣ S ) for the survival-conditioned law.
For an interval I I I , we say that H \mathcal{H} H has purely absolutely continuous spectrum (respectively pure point spectrum ) in I I I if, for every vector f f f , the restriction to I I I of the spectral measure of f f f is absolutely continuous with respect to Lebesgue measure (respectively a countable sum of point masses). Intervals of the first kind are regions of delocalization, and intervals of the second kind are regions of localization. An energy separating a region of one kind from a region of the other kind is a mobility edge .
Main results
Set
μ ∗ = E ∣ w ∣ = π 2 , w ∼ N C ( 0 , 1 ) . (2) \mu_* = \mathbb{E}|w| = \frac{\sqrt{\pi}}{2}, \qquad w \sim\mathcal{N}_{\mathbb{C}}(0,1). \tag*{(2)} μ ∗ = E ∣ w ∣ = 2 π , w ∼ N C ( 0 , 1 ) . (2)
Theorem 1.1. There exists p 0 < ∞ p_0 < \infty p 0 < ∞ such that for every p ≥ p 0 p \ge p_0 p ≥ p 0 there exists E c ( p ) > 0 E_c(p) > 0 E c ( p ) > 0 with the following properties. Almost surely with respect to P ∞ \mathbb{P}^\infty P ∞ , the operator H \mathcal{H} H has purely absolutely continuous spectrum in
( − E c ( p ) , 0 ) ∪ ( 0 , E c ( p ) ) \left(-E_c(p),0\right) \cup\left(0,E_c(p)\right) ( − E c ( p ) , 0 ) ∪ ( 0 , E c ( p ) ) and pure point spectrum in
( − ∞ , − E c ( p ) ) ∪ ( E c ( p ) , ∞ ) . \left(-\infty,-E_c(p)\right) \cup\left(E_c(p),\infty\right). ( − ∞ , − E c ( p ) ) ∪ ( E c ( p ) , ∞ ) . Moreover,
lim p → ∞ E c ( p ) p = μ ∗ = π 2 . (3) \lim_{p \to\infty} \frac{E_c(p)}{\sqrt{p}} = \mu_* = \frac{\sqrt{\pi}}{2}. \tag*{(3)} p → ∞ lim p E c ( p ) = μ ∗ = 2 π . (3) Finally, P ∞ \mathbb{P}^\infty P ∞ -almost surely, H \mathcal{H} H has no eigenvector with finite support and nonzero eigenvalue, and it has infinitely many linearly independent eigenvectors with finite support and eigenvalue zero.
Thus, for every fixed large p p p , the positive half-line contains exactly one mobility edge. The theorem makes no statement about the spectral type at the two energies ± E c ( p ) \pm E_c(p) ± E c ( p ) . The energy zero is exceptional: it is an eigenvalue of infinite multiplicity, created by pairs of leaves attached to a common vertex (localized states of this kind in diluted models go back to Kirkpatrick and Eggarter [86]), and Theorem 1.1 does not describe the spectral measure at zero.
The mobility edge lies far in the tail of the spectrum. As p → ∞ p \to\infty p → ∞ , the limiting empirical spectral measure of H n H_n H n converges to the semicircle law on [ − 2 , 2 ] [-2,2] [ − 2 , 2 ] , as a computation with the moment method shows, while E c ( p ) E_c(p) E c ( p ) is of order p \sqrt{p} p . Eigenvalues near E c ( p ) E_c(p) E c ( p ) therefore come from rare local configurations with unusually large edge weights. The constant μ ∗ \mu_* μ ∗ arises from a first-moment computation, which we explain in Section 2.5: at the energy E = ξ p E = \xi\sqrt{p} E = ξ p , a typical path of length L L L in the tree contributes a product of about L L L factors ∣ w e ∣ / ( ξ p ) |w_e|/(\xi p) ∣ w e ∣/ ( ξ p ) to the resolvent, and there are about p L p^L p L such paths. Their total contribution therefore grows or decays like ( μ ∗ / ξ ) L (\mu_*/\xi)^L ( μ ∗ / ξ ) L .
Our second result identifies E c ( p ) E_c(p) E c ( p ) as the unique zero of a function defined through the cavity equation, which is a recursive distributional equation for the diagonal resolvent entry at the root. Let Γ = ⟨ δ o , ( H − z ) − 1 δ o ⟩ \Gamma= \langle\delta_o,(\mathcal{H}-z)^{-1}\delta_o\rangle Γ = ⟨ δ o , ( H − z ) − 1 δ o ⟩ for z z z in the upper half-plane. Its law satisfies the cavity equation
Γ = d − ( z + 1 p ∑ j = 1 N γ j Γ j ) − 1 , N ∼ Pois ( p ) , γ j ∼ Exp ( 1 ) , \Gamma\overset{\mathrm{d}}{=} -\left(z+\frac{1}{p}\sum_{j=1}^{N}\gamma_j\Gamma_j\right)^{-1},\qquad N\sim\operatorname{Pois}(p),\qquad\gamma_j\sim\operatorname{Exp}(1), Γ = d − ( z + p 1 j = 1 ∑ N γ j Γ j ) − 1 , N ∼ Pois ( p ) , γ j ∼ Exp ( 1 ) ,
where the variables on the right are independent and the Γ j \Gamma_j Γ j are distributed as Γ \Gamma Γ . For E > 2 E>2 E > 2 , let g ( E ) = ( E − E 2 − 4 ) / 2 g(E)=(E-\sqrt{E^2-4})/2 g ( E ) = ( E − E 2 − 4 ) /2 and θ = g ( E ) 2 \theta=g(E)^2 θ = g ( E ) 2 . For a real energy E E E , the variable B = − g ( E ) / Γ B=-g(E)/\Gamma B = − g ( E ) /Γ formally satisfies
B = d 1 + θ − θ p ∑ j = 1 N γ j B j . (4) B\overset{\mathrm{d}}{=}1+\theta-\frac{\theta}{p}\sum_{j=1}^{N}\frac{\gamma_j}{B_j}. \tag*{(4)} B = d 1 + θ − p θ j = 1 ∑ N B j γ j . (4)
For every real solution of this equation we define in § 6.2 a function φ ( s ) \varphi(s) φ ( s ) , s ∈ ( 0 , 2 ) s\in(0,2) s ∈ ( 0 , 2 ) , as the logarithm of the Perron–Frobenius eigenvalue λ ( s ) \lambda(s) λ ( s ) of a positive transfer operator on the real projective line. We call it the free energy , as in [4, 12]. For s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) , φ ( s ) \varphi(s) φ ( s ) is the exponential growth rate of the expected generation sums of the s s s -th powers of the formal path products. Its limiting value φ ( 1 ) = lim s ↑ 1 φ ( s ) \varphi(1)=\lim_{s\uparrow1}\varphi(s) φ ( 1 ) = lim s ↑ 1 φ ( s ) decides the spectral type.
Theorem 1.2. Fix ζ 0 ∈ ( 0 , μ ∗ ) \zeta_0\in(0,\mu_*) ζ 0 ∈ ( 0 , μ ∗ ) . There exists p 1 ( ζ 0 ) ≥ p 0 p_1(\zeta_0)\ge p_0 p 1 ( ζ 0 ) ≥ p 0 such that for every p ≥ p 1 ( ζ 0 ) p\ge p_1(\zeta_0) p ≥ p 1 ( ζ 0 ) the following holds.
(i) For every E ≥ ζ 0 p E\ge\zeta_0\sqrt{p} E ≥ ζ 0 p , the iterates of the map defined by the right-hand side of (4) , started from the point mass at one, converge weakly to a solution ν p , E \nu_{p,E} ν p , E of (1.4).
(ii) Let φ E ( 1 ) \varphi_E(1) φ E ( 1 ) be the free energy of ν p , E \nu_{p,E} ν p , E at s = 1 s=1 s = 1 . The function E ↦ φ E ( 1 ) E\mapsto\varphi_E(1) E ↦ φ E ( 1 ) is continuous on [ ζ 0 p , ∞ ) [\zeta_0\sqrt{p},\infty) [ ζ 0 p , ∞ ) . It is positive on [ ζ 0 p , E c ( p ) ) [\zeta_0\sqrt{p},E_c(p)) [ ζ 0 p , E c ( p )) , vanishes at E c ( p ) E_c(p) E c ( p ) , and is negative on ( E c ( p ) , ∞ ) (E_c(p),\infty) ( E c ( p ) , ∞ ) , where E c ( p ) E_c(p) E c ( p ) is the energy of Theorem 1.1. It is real-analytic in a neighborhood of E c ( p ) E_c(p) E c ( p ) , and its derivative at E c ( p ) E_c(p) E c ( p ) is negative.
(iii) For every compact set Ξ ⊂ [ ζ 0 , ∞ ) \Xi\subset[\zeta_0,\infty) Ξ ⊂ [ ζ 0 , ∞ ) ,
lim p → ∞ sup ξ ∈ Ξ ∣ exp φ ξ p ( 1 ) − μ ∗ ζ ∣ = 0. \lim_{p\to\infty}\sup_{\xi\in\Xi}\left|\exp\varphi_{\xi\sqrt{p}}(1)-\frac{\mu_*}{\zeta}\right|=0. p → ∞ lim ξ ∈ Ξ sup exp φ ξ p ( 1 ) − ζ μ ∗ = 0.
Theorem 1.2 is an instance of the general principle, going back to Abou-Chacra, Anderson, and Thouless [2], that the mobility edge on a tree is located by the linear stability of the real solutions of the cavity equation. Its content is that, at large p p p , the free energy at s = 1 s=1 s = 1 of the relevant solution changes sign exactly once, and transversally. The asymptotic statement (iii) gives (3) .
The giant component
We now state the consequences for the finite graph. Since the spectral type of a finite matrix is trivial, we describe localization through the distribution of spectral weight over vertices in small energy windows. For an interval I ⊂ R I \subset\mathbb{R} I ⊂ R and a vertex v ∈ g c v \in\mathrm{gc} v ∈ gc , let
μ n , v ( I ) = ⟨ δ v , 1 I ( H n g c ) δ v ⟩ , μ ˉ n ( I ) = 1 N n Tr 1 I ( H n g c ) . \mu_{n,v}(I)=\langle\delta_v,\mathbf{1}_I(H_n^{\mathrm{gc}})\delta_v\rangle,\qquad\bar{\mu}_n(I)=\frac{1}{N_n}\operatorname{Tr}\mathbf{1}_I(H_n^{\mathrm{gc}}). μ n , v ( I ) = ⟨ δ v , 1 I ( H n gc ) δ v ⟩ , μ ˉ n ( I ) = N n 1 Tr 1 I ( H n gc ) .
so that μ ˉ n ( I ) \bar{\mu}_n(I) μ ˉ n ( I ) is the average of μ n , v ( I ) \mu_{n,v}(I) μ n , v ( I ) over v ∈ g c v \in\mathrm{gc} v ∈ gc . When μ ˉ n ( I ) > 0 \bar{\mu}_n(I)>0 μ ˉ n ( I ) > 0 , let P I , n ( v ) = μ n , v ( I ) / μ ˉ n ( I ) P_{I,n}(v)=\mu_{n,v}(I)/\bar{\mu}_n(I) P I , n ( v ) = μ n , v ( I ) / μ ˉ n ( I ) be the normalized local mass, and for R > 0 R>0 R > 0 define the deficit
D I , n ( R ) = 1 − 1 N n ∑ v ∈ g c min { P I , n ( v ) , R } . D_{I,n}(R)=1-\frac{1}{N_n}\sum_{v\in\mathrm{gc}}\min\{P_{I,n}(v),R\}. D I , n ( R ) = 1 − N n 1 v ∈ gc ∑ min { P I , n ( v ) , R } .
Set D I , n ( R ) = 1 D_{I,n}(R)=1 D I , n ( R ) = 1 when μ ˉ n ( I ) = 0 \bar{\mu}_n(I)=0 μ ˉ n ( I ) = 0 . When μ ˉ n ( I ) > 0 \bar{\mu}_n(I)>0 μ ˉ n ( I ) > 0 , the deficit is the excess normalized mass above the truncation level R R R , averaged over the vertices.
Definition 1.3. Let I h ( E ) = ( E − h , E + h ] I_h(E)=(E-h,E+h] I h ( E ) = ( E − h , E + h ] . The giant component is uniformly integrable delocalized at E E E if, for every δ > 0 \delta>0 δ > 0 ,
lim R → ∞ lim sup h ↓ 0 lim sup n → ∞ P ( D I h ( E ) , n ( R ) > δ ) = 0. \lim_{R\to\infty}\limsup_{h\downarrow0}\limsup_{n\to\infty}\mathbb{P}\bigl(D_{I_h(E),n}(R)>\delta\bigr)=0. R → ∞ lim h ↓ 0 lim sup n → ∞ lim sup P ( D I h ( E ) , n ( R ) > δ ) = 0. It is spectrally condensed at E E E if, for every R < ∞ R<\infty R < ∞ and every δ > 0 \delta>0 δ > 0 ,
lim h ↓ 0 lim inf n → ∞ P ( D I h ( E ) , n ( R ) > 1 − δ ) = 1. \lim_{h\downarrow0}\liminf_{n\to\infty}\mathbb{P}\bigl(D_{I_h(E),n}(R)>1-\delta\bigr)=1. h ↓ 0 lim n → ∞ lim inf P ( D I h ( E ) , n ( R ) > 1 − δ ) = 1.
In words, the giant component is uniformly integrable delocalized at E E E if the normalized local masses in small windows around E E E are uniformly integrable over the vertices, and it is spectrally condensed at E E E if almost all of the normalized mass is carried by a vanishing fraction of the vertices. In both definitions the limit n → ∞ n\to\infty n → ∞ is taken before the window width h h h tends to zero. These notions concern averaged spectral projections, and they do not describe individual eigenvectors.
Theorem 1.4. The following statements hold.
(i) For every p > 1 p>1 p > 1 and every compact set J ⊂ R ∖ { 0 } J\subset\mathbb{R}\setminus\{0\} J ⊂ R ∖ { 0 } , there exist constants 0 < c p , J ≤ C p , J < ∞ 0<c_{p,J}\leq C_{p,J}<\infty 0 < c p , J ≤ C p , J < ∞ such that, for every fixed interval I ⊂ J I\subset J I ⊂ J of positive length,
lim n → ∞ P ( c p , J N n ∣ I ∣ ≤ Tr 1 I ( H n g c ) ≤ C p , J N n ∣ I ∣ ) = 1. \lim_{n\to\infty}\mathbb{P}\left(c_{p,J}N_n|I|\leq\operatorname{Tr}\mathbf{1}_I(H_n^{\mathrm{gc}})\leq C_{p,J}N_n|I|\right)=1. n → ∞ lim P ( c p , J N n ∣ I ∣ ≤ Tr 1 I ( H n gc ) ≤ C p , J N n ∣ I ∣ ) = 1. (ii) Let p ≥ p 0 p\geq p_0 p ≥ p 0 , where p 0 p_0 p 0 is as in Theorem 1.1 . The giant component is uniformly integrable delocalized at Lebesgue-almost every E E E with 0 < ∣ E ∣ < E c ( p ) 0<|E|<E_c(p) 0 < ∣ E ∣ < E c ( p ) , and spectrally condensed at Lebesgue-almost every E E E with ∣ E ∣ > E c ( p ) |E|>E_c(p) ∣ E ∣ > E c ( p ) .
Discussion
Comparison with the physics literature. Theorem 1.1 shows that, for large p p p and apart from the eigenvalue zero, the localized states of the limiting operator are confined to energies ∣ E ∣ > E c ( p ) ≈ μ ∗ p |E| > E_c(p) \approx\mu_* \sqrt{p} ∣ E ∣ > E c ( p ) ≈ μ ∗ p , far beyond the semicircle bulk [ − 2 , 2 ] [-2,2] [ − 2 , 2 ] . This disagrees with the large-p p p behavior described in the physics literature on this model. For the real symmetric version of (1) restricted to the giant component, Cugliandolo, Schehr, Tarzia, and Venturelli state that the mobility edge tends to the edge 2 2 2 of the semicircle law as p → ∞ p \to\infty p → ∞ [58, Section V]. This statement is an extrapolation: their numerical study covers p ∈ [ 1.6 , 5 ] p \in[1.6,5] p ∈ [ 1.6 , 5 ] , and the mobility edge in their phase diagram is drawn only for p ≲ 2.5 p \lesssim2.5 p ≲ 2.5 [58, Section V and Fig. 1]. Similarly, Baron and Rogers describe the eigenvalues of weighted Erdős–Rényi graphs outside the continuous part of the spectrum, which ends at 2 + O ( p − 1 ) 2 + O(p^{-1}) 2 + O ( p − 1 ) , as produced by well-connected vertices or heavily weighted edges, with eigenvectors exponentially localized around these defects [31, Section VIII]. This is the picture of Biroli and Monasson [38].
Theorem 1.1 shows that this picture is not correct at large p p p on the limiting tree. At energies 2 < ∣ E ∣ < E c ( p ) 2 < |E| < E_c(p) 2 < ∣ E ∣ < E c ( p ) , rare local configurations with large edge weights do produce spectral weight, but they are not isolated from each other: their number grows exponentially with the distance, and resonances between distant configurations produce absolutely continuous spectrum. This is the mechanism discovered by Aizenman and Warzel for the Anderson model on regular trees, where extended states appear through disorder-enabled resonances well beyond the band of the hopping operator [11, 12]. Since the spectral mass of these energies tends to zero as p → ∞ p \to\infty p → ∞ , this delocalization is difficult to observe in numerical studies of moderate system size. The available numerical values are nonetheless consistent with a mobility edge that grows with p p p : for real Gaussian weights of variance 1 / p 1/p 1/ p on the whole graph, Kühn found the localization threshold at approximately 2.295 2.295 2.295 for p = 2 p = 2 p = 2 and 2.581 2.581 2.581 for p = 4 p = 4 p = 4 [91]. The first-moment heuristic described after Theorem 1.1 also applies to real Gaussian weights, and predicts a mobility edge near ∣ E ∣ E ∣ w ∣ p = 2 p / π |E| \mathbb{E}|w| \sqrt{p} = \sqrt{2p/\pi} ∣ E ∣ E ∣ w ∣ p = 2 p / π in that case; we do not prove this. At energies of order p \sqrt{p} p , the relevant real solution of the cavity equation is concentrated near − 1 / E -1/E − 1/ E , which explains why the leading-order location of the mobility edge is given by Anderson’s upper limit approximation, as reformulated in [2]; see also the forward approximation of [110]. This is in contrast with the critical disorder of the Anderson model on the Bethe lattice at large connectivity, where the upper limit approximation misses the correct constant because of correlations along paths [109].
Relation to previous rigorous results. Our proof follows the general strategy of [4, 7, 12]: the spectral type at an energy E E E is decided by the sign of an exponential growth rate of fractional moments of the Green function along paths, which is computed through the Perron–Frobenius eigenvalue of a transfer operator. Aizenman and Warzel normalize this rate per vertex, so that the threshold for their free energy φ ( 1 ; E ) \varphi(1; E) φ ( 1 ; E ) is minus the logarithm of the branching number; as in [4, 7], we sum over generations, so that the threshold for our free energy φ ( 1 ) \varphi(1) φ ( 1 ) is zero. There are four main differences.
First, the model has no random potential. In [4, 12], the random potential at a vertex is the rank-one parameter in the Simon–Wolff criterion [124] and in spectral averaging. Here the randomness is entirely in the hopping amplitudes. We use instead that the tree is bipartite, so that H 2 \mathcal{H}^{2} H 2 is a direct sum of two operators of the form A ∗ A A^{*}A A ∗ A , and that A ∗ A A^{*}A A ∗ A is a sum of rank-one projections onto the normalized rows of A A A , weighted by the squared row norms. For complex Gaussian weights these squared norms have Gamma densities and are independent of the row directions, so they can play the role of the random potential.
Second, our localization argument does not use a continuity argument in the energy. In [4, 7], the transfer operator describes the fractional moments only if the boundary values of the resolvent are already known to be real, and localization is propagated from very large energies by a continuity argument in the energy. We show instead that any real solution of the cavity equation with negative free energy at s = 1 s = 1 s = 1 produces, on an enlarged probability space, a square-summable vector F F F with ( H − E ) F = δ o (\mathcal{H} - E)F = \delta_{o} ( H − E ) F = δ o . This gives the input of the Simon–Wolff criterion directly, at every energy where the free energy at s = 1 s = 1 s = 1 is negative, without identifying the actual boundary values of the resolvent.
Third, the delocalization criterion applies to arbitrary real boundary laws, not only to the solution of the cavity equation that we construct. This allows us to treat all small positive energies, where we do not identify the real solutions of the cavity equation, by proving that every real solution there has positive free energy at s = 1 s = 1 s = 1 .
Fourth, we prove that the mobility edge is unique on the whole positive half-line for every fixed large p p p . In [4], the mobility edges are shown to be isolated through a strict monotonicity estimate for the Perron–Frobenius eigenvalue near each predicted location. In [7], uniqueness is derived from an explicit formula for the critical eigenvalue, and is proved for α \alpha α near 0 and 1. For the present model no explicit formula is available, and the relevant cavity law has rare cancellation events whose contribution to the critical eigenvalue must be controlled together with their energy derivatives. We obtain uniqueness from an exact positive renewal decomposition of the square of the critical operator, which reduces the monotonicity of the Perron–Frobenius eigenvalue to that of a 2 × 2 2 \times2 2 × 2 matrix, and from estimates on the energy derivatives of the cavity law.
Concerning finite graphs, Theorem 1.4 concerns averaged local spectral masses in small energy windows, in the spirit of the localization and delocalization criteria for heavy-tailed matrices in [7, 48] and their analogues for random regular graphs in [96]. Related transfers of delocalization from the absolutely continuous spectrum of a limiting tree to finite graphs are the quantum ergodicity theorems for regular graphs [23, 24] and for general graphs converging to trees [25], and the quantitative delocalization results of [ref-46 , ref-54 ]. We do not prove statements about individual eigenvectors of H n H_n H n .
The complex Gaussian assumption. The assumption that the weights are complex Gaussian is used in two ways. First, the squared moduli Y e = ∣ w e ∣ 2 Y_e = |w_e|^2 Y e = ∣ w e ∣ 2 of the weights are exponential random variables, independent of the phases, and the cavity variables depend on the weights only through these moduli. Exponential laws make many estimates explicit, in particular negative moment bounds for affine combinations of independent squared moduli, which we use throughout. Second, as explained above, the squared row norms of the bipartite square have Gamma densities independent of the row directions, which provides the rank-one randomness needed for the Simon–Wolff criterion and for spectral averaging. For real Gaussian weights the squared row norms are still independent of the directions, and we expect our methods to extend, with additional work and with μ ∗ \mu_* μ ∗ replaced by 2 / π \sqrt{2/\pi} 2/ π , but we have not checked the details.
Open questions. We mention several questions that our results leave open. The first concerns small p p p : does the limiting operator have a unique mobility edge for every p > 1 p > 1 p > 1 , or only for p p p large? For small p p p , the analysis of [58] suggests that the delocalized phase has unusual properties. The second concerns the spectral type at the energies ± E c ( p ) \pm E_c(p) ± E c ( p ) and at zero. The third concerns finite graphs: it would be interesting to prove localization and delocalization of individual eigenvectors of H n g c H_n^{\mathrm{gc}} H n gc in the two phases, and to determine their local eigenvalue statistics, in analogy with [6, 21]. Finally, it would be natural to study the multifractal properties of the delocalized eigenvectors predicted in [31, 58].
Notation and conventions
We write H = { z ∈ C : Im z > 0 } \mathbb{H} = \{z \in\mathbb{C} : \operatorname{Im} z > 0\} H = { z ∈ C : Im z > 0 } for the upper half-plane, and z = E + i η z = E + i\eta z = E + i η with E ∈ R E \in\mathbb{R} E ∈ R and η > 0 \eta> 0 η > 0 . For a self-adjoint operator K K K and a vector f f f , the spectral measure of f f f is the measure μ \mu μ with ⟨ f , ( K − z ) − 1 f ⟩ = ∫ ( λ − z ) − 1 μ ( d λ ) \langle f,(K-z)^{-1}f\rangle= \int(\lambda-z)^{-1}\mu(\mathrm{d}\lambda) ⟨ f , ( K − z ) − 1 f ⟩ = ∫ ( λ − z ) − 1 μ ( d λ ) . The letters c c c and C C C denote positive constants whose dependence on other parameters is indicated in each statement; within a proof they may change from line to line. We write Pois ( λ ) \operatorname{Pois}(\lambda) Pois ( λ ) and Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) for the Poisson law with mean λ \lambda λ and the exponential law with mean one, and = d \overset{d}{=} = d for equality in distribution. For a tree vertex v v v other than the root, v − v^- v − is its parent and ∣ v ∣ |v| ∣ v ∣ its generation, and T v T_v T v is the subtree of descendants of v v v , rooted at v v v . For an edge e e e , we call Y e = ∣ w e ∣ 2 = p ∣ h e ∣ 2 Y_e = |w_e|^2 = p|h_e|^2 Y e = ∣ w e ∣ 2 = p ∣ h e ∣ 2 the squared modulus of the edge weight. The squared moduli have the Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) law, and are independent of the phases w e / ∣ w e ∣ w_e/|w_e| w e /∣ w e ∣ , which are uniform on the unit circle. Notation that is used only in one section is introduced in that section.
Organization
Section 2 outlines the proofs of our main results. Part I develops the basic theory for a fixed degree p p p : the operator on the tree (Section 3), spine decompositions and estimates for fresh squared moduli (Section 4), real solutions of the cavity equation (Section 5), and the transfer operator and the free energy (§ 6.2 ). Part II shows that the sign of the free energy at s = 1 s = 1 s = 1 determines the spectral type at a fixed degree: Section 7 collects the rank-one perturbation arguments, § 8.2 proves that negative free energy implies localization, and Sections 9 and 10 prove that positive free energy is incompatible with real boundary values. Part III studies the free energy at large degree: we construct the principal solution of the cavity equation in Section 11, prove its analyticity and derivative bounds in Section 12, determine the leading asymptotics of the free energy at s = 1 s = 1 s = 1 in Section 13, prove the uniqueness of its zero in Section 14, and treat small energies in Section 15. Finally, Part IV contains the proofs of Theorems 1.1 and 1.2 in Section 16 and of Theorem 1.4 in Section 17.
Outline of the proof
In this section we explain the main steps of the proofs of Theorems 1.1 and 1.2 , and indicate where each step is carried out. The discussion is informal, but the objects and identities that appear are the ones used in the proofs. We only consider positive energies, since H \mathcal{H} H is unitarily equivalent to − H -\mathcal{H} − H (Proposition 3.21 ). Throughout, z = E + i η z=E+i\eta z = E + i η with E > 0 E>0 E > 0 and η > 0 \eta>0 η > 0 .
Cavity variables and the two alternatives
Let T \mathcal{T} T be the rooted Poisson–Galton–Watson tree with mean degree p p p , with edge weights h e = w e / p h_e=w_e/\sqrt{p} h e = w e / p , and let H \mathcal{H} H be its weighted adjacency operator. Although the degrees of T \mathcal{T} T are unbounded, H \mathcal{H} H is essentially self-adjoint on finitely supported vectors (Theorem 3.10 ), and its resolvent G ( z ) = ( H − z ) − 1 G(z)=(\mathcal{H}-z)^{-1} G ( z ) = ( H − z ) − 1 is canonically defined. For a vertex v ≠ o v\ne o v = o , let Γ v ( z ) \Gamma_v(z) Γ v ( z ) denote the diagonal resolvent entry at v v v of the operator restricted to the subtree of descendants of v v v . The Schur complement formula gives the cavity recursion
Γ v ( z ) = − ( z + ∑ u : u − = v ∣ h v u ∣ 2 Γ u ( z ) ) − 1 (5) \Gamma_v(z)=-\left(z+\sum_{u:u^-=v}|h_{vu}|^2\Gamma_u(z)\right)^{-1} \tag*{(5)} Γ v ( z ) = − ( z + u : u − = v ∑ ∣ h v u ∣ 2 Γ u ( z ) ) − 1 (5)
and, since the subtrees of distinct children are independent copies of T \mathcal{T} T , the root variable Γ o = G o o \Gamma_o=G_{oo} Γ o = G oo satisfies
Γ o = d − ( z + 1 p ∑ j = 1 N Y j Γ j ) − 1 , N ∼ Pois ( p ) , Y j ∼ Exp ( 1 ) , (6) \Gamma_o\mathrel{\overset{\mathrm d}{=}}-\left(z+\frac{1}{p}\sum_{j=1}^{N}Y_j\Gamma_j\right)^{-1},\qquad N\sim\operatorname{Pois}(p),\qquad Y_j\sim\operatorname{Exp}(1), \tag*{(6)} Γ o = d − ( z + p 1 j = 1 ∑ N Y j Γ j ) − 1 , N ∼ Pois ( p ) , Y j ∼ Exp ( 1 ) , (6)
with all variables on the right independent and Γ j = d Γ o \Gamma_j\mathrel{\overset{\mathrm d}{=}}\Gamma_o Γ j = d Γ o . Off-diagonal resolvent entries factor along paths: if o = v 0 , v 1 , … , v L = v o=v_0,v_1,\ldots,v_L=v o = v 0 , v 1 , … , v L = v is the path from the root to v v v , then
G o v ( z ) = ( − 1 ) L Γ o ( z ) ∏ j = 1 L h v j − 1 v j Γ v j ( z ) . (7) G_{ov}(z)=(-1)^L\Gamma_o(z)\prod_{j=1}^{L}h_{v_{j-1}v_j}\Gamma_{v_j}(z). \tag*{(7)} G o v ( z ) = ( − 1 ) L Γ o ( z ) j = 1 ∏ L h v j − 1 v j Γ v j ( z ) . (7)
These facts are proved in Section 3.
At Lebesgue-almost every energy E E E , the boundary values Γ v ( E + i 0 ) \Gamma_v(E+i0) Γ v ( E + i 0 ) exist, and a zero–one law holds: under the survival-conditioned law P ∞ \mathbb{P}^{\infty} P ∞ , either $\operatorname{Im}\Gamma_o(E+i0)>0 almost surely, or Im Γ o ( E + i 0 ) = 0 \operatorname{Im}\Gamma_o(E+i0)=0 Im Γ o ( E + i 0 ) = 0 almost surely (Proposition 5.11 ). In the first case, the full diagonal boundary values satisfy Im G v v ( E + i 0 ) > 0 \operatorname{Im}G_{vv}(E+i0)>0 Im G v v ( E + i 0 ) > 0 at every vertex of the infinite tree. If this alternative holds for Lebesgue-almost every energy in an interval I I I , a spectral averaging argument shows that the spectral measures are absolutely continuous in I I I (Proposition 7.11 ). In the second case, which we call the no-ac hypothesis following [12], all cavity variables at E + i 0 E+i0 E + i 0 are finite and real, and the inverse cavity variables D v = − 1 / Γ v ( E + i 0 ) D_v=-1/\Gamma_v(E+i0) D v = − 1/ Γ v ( E + i 0 ) , which we call denominators, solve the real cavity equation
D ≡ E − 1 p ∑ j = 1 N Y j D j . (8) D \mathrel{\mathop{\equiv}} E-\frac{1}{p}\sum_{j=1}^{N}\frac{Y_j}{D_j}. \tag*{(8)} D ≡ E − p 1 j = 1 ∑ N D j Y j . (8)
A law on R ∖ { 0 } \mathbb{R}\setminus\{0\} R ∖ { 0 } satisfying (2.4) is called a real solution at energy E E E . The proof decides between the two alternatives by attaching to every real solution a number, the value φ ( 1 ) \varphi(1) φ ( 1 ) at s = 1 s=1 s = 1 of its free energy, and showing that its sign determines the spectral type.
The free energy
By (7) , the size of G o v G_{ov} G o v is governed by products of the variables ∣ h v j − 1 v j Γ v j ∣ |h_{v_j-1v_j}\Gamma_{v_j}| ∣ h v j − 1 v j Γ v j ∣ along paths. Under the no-ac hypothesis these products can be generated by a transfer operator. Fix a scaling g > 0 g>0 g > 0 and set B = g D B=gD B = g D and β = g 2 γ / p \beta=g^2\gamma/p β = g 2 γ / p , so that (8) becomes a recursion for B B B in which a child with state x = 1 / B x=1/B x = 1/ B is mapped to the parent state ( B ′ − β x ) − 1 (B'-\beta x)^{-1} ( B ′ − β x ) − 1 , where B ′ B' B ′ collects the contributions of the other children. Each step of this recursion is the projective action of the matrix
M ( B , β ) = β − 1 / 2 ( 0 1 − β B ) , det M ( B , β ) = 1. (9) M(B,\beta)=\beta^{-1/2}\begin{pmatrix}0&1\\-\beta&B\end{pmatrix},\qquad\det M(B,\beta)=1. \tag*{(9)} M ( B , β ) = β − 1/2 ( 0 − β 1 B ) , det M ( B , β ) = 1. (9)
The Mecke formula for Poisson processes turns sums over vertices of generation L L L into expectations over a single path, at the cost of a factor p p p per generation (Proposition 4.8). For s ∈ ( 0 , 2 ) s\in(0,2) s ∈ ( 0 , 2 ) this leads to the positive transfer operator
( Q s F ) ( x ) = p E [ ∥ M ( B , β ) v x ∥ − s F ( f B , β ( x ) ) ] , v x = ( x , 1 ) 1 + x 2 , (Q_sF)(x)=p\,\mathbb{E}\left[\|M(B,\beta)v_x\|^{-s}F(f_{B,\beta}(x))\right],\qquad v_x=\frac{(x,1)}{\sqrt{1+x^2}}, ( Q s F ) ( x ) = p E [ ∥ M ( B , β ) v x ∥ − s F ( f B , β ( x )) ] , v x = 1 + x 2 ( x , 1 ) ,
acting on continuous functions on the real projective line R P 1 \mathbb{R}\mathbb{P}^1 R P 1 , where f B , β f_{B,\beta} f B , β is the projective action of M ( B , β ) M(B,\beta) M ( B , β ) and B B B is distributed according to the real solution. We show in § 6.2 that Q s Q_s Q s is compact and strongly positive, so that Q s Q_s Q s has a simple dominant Perron–Frobenius eigenvalue λ ( s ) \lambda(s) λ ( s ) . The free energy is φ ( s ) = log λ ( s ) \varphi(s)=\log\lambda(s) φ ( s ) = log λ ( s ) , and its value φ ( 1 ) \varphi(1) φ ( 1 ) at s = 1 s=1 s = 1 decides the spectral type. We have
φ ( s ) = log p + lim L → ∞ 1 L log E ∫ R P 1 ∥ M L ⋯ M 1 v ∥ − s m ( d v ) , \varphi(s)=\log p+\lim_{L\to\infty}\frac{1}{L}\log\mathbb{E}\int_{\mathbb{R}\mathbb{P}^1}\|M_L\cdots M_1v\|^{-s}m(dv), φ ( s ) = log p + L → ∞ lim L 1 log E ∫ R P 1 ∥ M L ⋯ M 1 v ∥ − s m ( d v ) ,
where M 1 , M 2 , … M_1, M_2, \ldots M 1 , M 2 , … are independent copies of M ( B , β ) M(B,\beta) M ( B , β ) and m m m is the uniform measure on R P 1 \mathbb{R}\mathbb{P}^1 R P 1 . For s < 1 s<1 s < 1 , the quantity φ ( s ) \varphi(s) φ ( s ) is also the exponential growth rate of the expected generation sums of s s s -th powers of the path products in (7) .
The free energy is convex, real-analytic, and independent of the scaling. Its key property is the symmetry
φ ( s ) = φ ( 2 − s ) , (10) \varphi(s)=\varphi(2-s), \tag*{(10)} φ ( s ) = φ ( 2 − s ) , (10)
which follows from the change of variables v ↦ M v / ∥ M v ∥ v\mapsto Mv/\lVert Mv\rVert v ↦ M v / ∥ M v ∥ on R P 1 \mathbb{R}\mathbb{P}^1 R P 1 , whose Jacobian is ∥ M v ∥ − 2 \lVert Mv\rVert^{-2} ∥ M v ∥ − 2 , and from the fact that a 2 × 2 2\times2 2 × 2 matrix of determinant one has the same singular values as its inverse. Then φ ′ ( 1 ) = 0 \varphi'(1)=0 φ ′ ( 1 ) = 0 , the point s = 1 s=1 s = 1 is the minimum of φ \varphi φ , and φ \varphi φ decreases from φ ( 0 + ) = log p \varphi(0+)=\log p φ ( 0 + ) = log p to φ ( 1 ) \varphi(1) φ ( 1 ) on ( 0 , 1 ] (0,1] ( 0 , 1 ] . The value φ ( 1 ) \varphi(1) φ ( 1 ) is the analogue, for the present model, of the criterion of Abou-Chacra, Anderson, and Thouless [2], of the free-energy criterion of Aizenman and Warzel [12], and of the quantity λ ( E , α ) \lambda(E,\alpha) λ ( E , α ) in the work of Aggarwal, Bordenave, and Lopatto on Lévy matrices [7]. A symmetry of this kind for the linearized cavity kernel goes back to [2]; see [109]. As in [4, 7], and unlike [12], we sum over generations rather than normalizing per vertex, and the threshold is zero.
Negative free energy implies localization
Suppose that some real solution π \pi π at energy E E E has φ π ( 1 ) < 0 \varphi_\pi(1)<0 φ π ( 1 ) < 0 . We do not need to know that π \pi π is the law of the actual boundary values. By Kolmogorov’s extension theorem, we may enlarge the probability space so that every vertex v v v carries a real variable D v D_v D v with law π \pi π , the variables satisfy the recursion (8) pathwise on the tree, and the tree itself keeps its law (Lemma 5.13). Set Γ v = − 1 / D v \Gamma_v=-1/D_v Γ v = − 1/ D v and define the formal inverse column
F ( o ) = Γ o , F ( v ) = Γ o ∏ j = 1 L ( − h v j v j − 1 Γ v j ) . F(o)=\Gamma_o,\qquad F(v)=\Gamma_o\prod_{j=1}^{L}\left(-h_{v_jv_{j-1}}\Gamma_{v_j}\right). F ( o ) = Γ o , F ( v ) = Γ o j = 1 ∏ L ( − h v j v j − 1 Γ v j ) .
The recursion gives ( H − E ) F = δ o (\mathcal{H}-E)F=\delta_o ( H − E ) F = δ o pointwise. By the many-to-one formula, E ∑ ∣ v ∣ = L ∣ F ( v ) ∣ s \mathbb{E}\sum_{|v|=L}|F(v)|^s E ∑ ∣ v ∣ = L ∣ F ( v ) ∣ s grows at the exponential rate φ π ( s ) \varphi_\pi(s) φ π ( s ) , and since φ π ( 1 ) < 0 \varphi_\pi(1)<0 φ π ( 1 ) < 0 we may choose s < 1 s<1 s < 1 with φ π ( s ) < 0 \varphi_\pi(s)<0 φ π ( s ) < 0 . Then F ∈ ℓ s ⊂ ℓ 2 F\in\ell^s\subset\ell^2 F ∈ ℓ s ⊂ ℓ 2 almost surely. In particular, F F F lies in the domain of H \mathcal{H} H and δ o = ( H − E ) F \delta_o=(\mathcal{H}-E)F δ o = ( H − E ) F lies in the range of H − E \mathcal{H}-E H − E . By the spectral theorem, ∫ μ o ( d λ ) / ( λ − E ) 2 < ∞ \int\mu_o(\mathrm{d}\lambda)/(\lambda-E)^2<\infty ∫ μ o ( d λ ) / ( λ − E ) 2 < ∞ , where μ o \mu_o μ o is the spectral measure at the root. The same holds at every vertex and on every component of the tree with one vertex removed (Proposition 8.3).
These reciprocal-square integrals are the input of the Simon–Wolff criterion [124]. Since the model has no random potential, we need another source of rank-one randomness. It is supplied by the complex Gaussian weights. The tree is bipartite, so H 2 = A ∗ A ⊕ A A ∗ \mathcal{H}^2=A^*A\oplus AA^* H 2 = A ∗ A ⊕ A A ∗ , and
A ∗ A = ∑ y U y ∣ v y ⟩ ⟨ v y ∣ . A^*A=\sum_y U_y\lvert v_y\rangle\langle v_y\rvert. A ∗ A = y ∑ U y ∣ v y ⟩ ⟨ v y ∣ .
where y y y runs over the odd generations, v y v_y v y is the normalized vector of edge weights adjacent to y y y , and U y U_y U y is its squared norm. Conditional on the tree, the squared row norm U y U_y U y has a Gamma density and is independent of the direction v y v_y v y and of all other rows (Proposition 3.23). Thus each U y U_y U y is the coupling constant of a rank-one perturbation with an absolutely continuous law, and the Simon–Wolff argument applies in the square blocks. This yields pure point spectrum on every open set of energies at which a real solution with negative free energy at s = 1 s=1 s = 1 exists (Theorem 8.4 ).
Positive free energy implies delocalization
Suppose now that the no-ac hypothesis holds at a regular energy E E E , and let π \pi π be the law of the boundary denominators. We show that φ π ( 1 ) ≤ 0 \varphi_\pi(1)\leq0 φ π ( 1 ) ≤ 0 (Theorem 10.1 ). Together with the spectral averaging argument mentioned above, this proves pure absolute continuity on any interval where, for almost every regular energy, all possible real boundary laws have positive free energy at s = 1 s=1 s = 1 (Corollary 10.2). The argument follows the resonance strategy of Aizenman and Warzel [12]: we look for vertices at large depth n n n that are strongly coupled to the root, and show that their contributions force Im Γ o \operatorname{Im}\Gamma_o Im Γ o to be large. Variants of this strategy were used in [4, 7].
Suppose, for a contradiction, that 0 < φ π ( 1 ) 0<\varphi_\pi(1) 0 < φ π ( 1 ) . We always have φ π ( 1 ) < log p \varphi_\pi(1)<\log p φ π ( 1 ) < log p (Proposition 6.15 ). By convexity and symmetry, there exists s ∈ ( 1 / 2 , 1 ) s\in(1/2,1) s ∈ ( 1/2 , 1 ) such that
γ = − φ π ′ ( s ) > 0 , F = φ π ( s ) − ( 1 − s ) γ > 0. \gamma=-\varphi_\pi'(s)>0,\qquad F=\varphi_\pi(s)-(1-s)\gamma>0. γ = − φ π ′ ( s ) > 0 , F = φ π ( s ) − ( 1 − s ) γ > 0.
Consider a path of length n n n from the root, and the event that the product of the transfer matrices along it is of size about e γ n e^{\gamma n} e γ n . After an exponential tilt of the path law, this is a moderate deviation event for a sum of bounded increments, with probability about e − ( log p − φ π ( s ) − s γ ) n e^{-(\log p-\varphi_\pi(s)-s\gamma)n} e − ( l o g p − φ π ( s ) − s γ ) n . Since there are about p n p^n p n paths, the expected number of such paths is about e ( φ π ( s ) + s γ ) n e^{(\varphi_\pi(s)+s\gamma)n} e ( φ π ( s ) + s γ ) n . Along such a path, an exact continuant identity shows that ∣ G o x / G x x ∣ ≳ e − γ n \lvert G_{ox}/G_{xx}\rvert\gtrsim e^{-\gamma n} ∣ G o x / G xx ∣ ≳ e − γ n , where x x x is the endpoint. At x x x we attach a pendant leaf whose squared edge weight t t t has an exponential density. Since the diagonal entry G x x G_{xx} G xx is a Möbius function of t t t with a real pole, the event ∣ G x x ∣ ≥ T \lvert G_{xx}\rvert\geq T ∣ G xx ∣ ≥ T has probability of order T − 1 T^{-1} T − 1 . With T = e ( γ + δ ) n T=e^{(\gamma+\delta)n} T = e ( γ + δ ) n , we obtain an expected number
E N n ≥ e ( F − δ ) n − o ( n ) \mathbb{E}N_n\geq e^{(F-\delta)n-o(n)} E N n ≥ e ( F − δ ) n − o ( n )
of resonant endpoints x x x , each satisfying ∣ G o x ∣ ≥ e δ n \lvert G_{ox}\rvert\geq e^{\delta n} ∣ G o x ∣ ≥ e δ n . A surviving child of x x x , whose imaginary part is at least a fixed quantile ζ α \zeta_\alpha ζ α of the distribution of Im Γ \operatorname{Im}\Gamma Im Γ , then gives Im Γ o ≥ e 2 δ n − o ( n ) ζ α \operatorname{Im}\Gamma_o\geq e^{2\delta n-o(n)}\zeta_\alpha Im Γ o ≥ e 2 δ n − o ( n ) ζ α by the Ward identity.
The main difficulty is the second moment. For two resonant endpoints x x x and y y y whose paths split at depth j j j , the two pendant tunings are nearly independent unless the Green function between the two branches is large. This interaction is controlled. by the off-diagonal entry of the effective 2 × 2 2 \times2 2 × 2 Feshbach matrix at x x x and y y y , which is the bilinear form of a product of transfer matrices along the path from x x x to y y y . To average this product we reserve, near the split and near both endpoints, three consecutive factors whose squared moduli are not constrained by any of the events retained in the second moment. Three fresh factors suffice to smooth an arbitrary fixed bilinear form (Proposition 6.18). The interaction contribution is then bounded by the norm moments of the two arms, which decay at the rate φ π ( s ) − log p \varphi_{\pi}(s) - \log p φ π ( s ) − log p per step. The resulting exponent is an affine function of j j j , and the choice of s s s makes it negative at j = 0 j = 0 j = 0 and at j = n j = n j = n . Thus E [ N n ( N n − 1 ) ] ≤ e o ( n ) ( E N n ) 2 \mathbb{E}[N_n(N_n - 1)] \leq e^{o(n)}(\mathbb{E}N_n)^2 E [ N n ( N n − 1 )] ≤ e o ( n ) ( E N n ) 2 , and the Paley–Zygmund inequality gives P ( N n > 0 ) ≥ e − o ( n ) \mathbb{P}(N_n > 0) \geq e^{-o(n)} P ( N n > 0 ) ≥ e − o ( n ) .
To conclude, we need to raise this subexponentially small probability to a fixed level. We use the Poisson structure of the tree: the root has many children, and each of them independently carries a copy of the rare event. An amplification argument, which uses only a fixed number of amplifying children per generation and a reserve of other children that keeps the cavity variables in a fixed compact region, raises the probability to a fixed number q ∗ > α q_* > \alpha q ∗ > α after o ( n ) o(n) o ( n ) generations, at the cost of a factor e − o ( n ) e^{-o(n)} e − o ( n ) in the lower bound for Im Γ o \operatorname{Im}\Gamma_o Im Γ o . We obtain P ∞ ( Im Γ o ≥ e 2 δ n − o ( n ) ζ α ) ≥ q ∗ > α \mathbb{P}^{\infty}(\operatorname{Im}\Gamma_o \geq e^{2\delta n-o(n)}\zeta_\alpha) \geq q_* > \alpha P ∞ ( Im Γ o ≥ e 2 δ n − o ( n ) ζ α ) ≥ q ∗ > α , which contradicts the definition of the quantile ζ α \zeta_\alpha ζ α once n n n is large. All estimates are first made for the real recursive tree process at a fixed depth and then transferred to a small positive imaginary part η \eta η , chosen after n n n .
The free energy at large degree
By the previous two steps, Theorem 1.1 reduces to determining the sign of the free energy at s = 1 s = 1 s = 1 of the relevant real solutions. This is where the degree p p p is taken large.
The principal solution. For E > 2 E > 2 E > 2 it is convenient to use the scaling g ( E ) = ( E − E 2 − 4 ) / 2 g(E) = (E - \sqrt{E^2 - 4})/2 g ( E ) = ( E − E 2 − 4 ) /2 , in which the real cavity equation becomes
B = d 1 + θ − θ p ∑ j = 1 N γ j B j , θ = g ( E ) 2 . B \overset{\mathrm d}= 1 + \theta- \frac{\theta}{p}\sum_{j=1}^{N}\frac{\gamma_j}{B_j}, \qquad\theta= g(E)^2. B = d 1 + θ − p θ j = 1 ∑ N B j γ j , θ = g ( E ) 2 .
At energies E = ξ p E = \xi\sqrt{p} E = ξ p with ξ \xi ξ of order one, we have θ ≍ 1 / p \theta\asymp1/p θ ≍ 1/ p , so the sum is an average of about p p p terms of size θ / p \theta/p θ / p , and one expects B ≈ 1 B \approx1 B ≈ 1 . This is correct except on rare events where some B j B_j B j is close to zero. In Section 11 we construct, for every compact window of scaled energies ξ \xi ξ and all large p p p , a solution ν p , E \nu_{p,E} ν p , E whose law is concentrated within p − 1 / 4 p^{-1/4} p − 1/4 of one, has density at most p A e − a p p^A e^{-ap} p A e − a p near zero, and has a tail decaying like p A e − a p / L p^A e^{-ap}/L p A e − a p / L . We denote by S \mathcal{S} S the class of laws with these three properties. The solution ν p , E \nu_{p,E} ν p , E , which we call the principal solution , is the unique fixed point of the cavity map in S \mathcal{S} S , on which the map is contracted by a metric adapted to these three regions. The same construction covers all energies E ≥ M 0 p E \ge M_0\sqrt{p} E ≥ M 0 p , where the cavity map is a contraction in the same class. We then show that at large p p p and energies E = ξ p E = \xi\sqrt{p} E = ξ p in a fixed window, every real solution of the cavity equation is the principal solution (Proposition 11.27). The proof is a one-big-jump bootstrap: it starts from the bound p + 1 \sqrt{p+1} p + 1 on the density near zero, which every real solution satisfies, and improves it until the solution lies in the class S \mathcal{S} S . In particular, under the no-ac hypothesis the boundary law can only be the principal solution.
The location of the mobility edge. The leading behavior of the free energy at s = 1 s = 1 s = 1 of the principal solution is easy to guess. Since B ≈ 1 B \approx1 B ≈ 1 , we have ∣ Γ v ∣ ≈ g ( E ) ≈ E − 1 |\Gamma_v| \approx g(E) \approx E^{-1} ∣ Γ v ∣ ≈ g ( E ) ≈ E − 1 along a typical path. Each step of (7) contributes a factor ∣ h e ∣ / E = ∣ w e ∣ / ( ξ p ) |h_e|/E = |w_e|/(\xi p) ∣ h e ∣/ E = ∣ w e ∣/ ( ξ p ) , and there are about p p p choices of the next vertex. The expected sum of the products over the p L p^L p L paths of length L L L is therefore about
( p ⋅ E ∣ w ∣ ξ p ) L = ( μ ∗ ξ ) L , μ ∗ = E ∣ w ∣ = π 2 . \left(p \cdot\frac{\mathbb{E}|w|}{\xi p}\right)^L = \left(\frac{\mu_*}{\xi}\right)^L,\qquad\mu_* = \mathbb{E}|w| = \frac{\sqrt{\pi}}{2}. ( p ⋅ ξ p E ∣ w ∣ ) L = ( ξ μ ∗ ) L , μ ∗ = E ∣ w ∣ = 2 π .
This suggests exp ϕ ξ p ( 1 ) ≈ μ ∗ / ξ \exp\phi_{\xi\sqrt{p}}(1) \approx\mu_*/\xi exp ϕ ξ p ( 1 ) ≈ μ ∗ / ξ , and in Section 13 we prove that this is correct uniformly on compact sets of ξ \xi ξ (Theorem 13.1). The lower bound follows from the behavior of the transfer operator on states near one. The upper bound is harder, because the rare cancellations B ≈ 0 B \approx0 B ≈ 0 produce large states, which may feed back into the core region of states near one. We decompose the state space into six regions and show that the only order-one transitions between the five other regions form an acyclic graph, so that every return to the core region passes through a transition of small weight. A finite-dimensional diagonal balancing then bounds the growth rate.
Uniqueness of the crossing. The leading asymptotics show that the free energy at s = 1 s = 1 s = 1 changes sign near μ ∗ p \mu_*\sqrt{p} μ ∗ p , but by themselves they do not exclude several sign changes in a window of width o ( p ) o(\sqrt{p}) o ( p ) . In Section 12 we show that the principal law is a real-analytic function of the energy, with polynomially bounded derivatives in suitable norms, and that its first two derivatives are small when tested against smooth functions supported near one. In Section 14 we work with the L 2 L^2 L 2 version of the critical transfer operator, on a fixed space on which the energy enters only through the law of B B B and through exponential densities. We then decompose its square exactly as
A ξ 2 = G ξ + h c ⊗ ℓ c + h t ⊗ ℓ t , A_\xi^2 = G_\xi+ h_c \otimes\ell_c + h_t \otimes\ell_t, A ξ 2 = G ξ + h c ⊗ ℓ c + h t ⊗ ℓ t ,
where A ξ A_\xi A ξ is the critical operator on this space, the first rank-one term is a positive operator describing the region where B B B is close to 1, the second is its image under an involution J J J that exchanges this region with the region where B B B is close to 0, and G ξ G_\xi G ξ is a positive operator with small spectral radius. The Perron–Frobenius eigenvalue of A ξ 2 A_\xi^2 A ξ 2 is then determined by the Perron–Frobenius eigenvalue of a 2 × 2 2 \times2 2 × 2 positive renewal matrix. Its diagonal entries are equal, by the symmetry under J J J , and decrease strictly in ξ \xi ξ . Its off-diagonal entries are positive, and one of them is exponentially small. A positivity argument, based on Glaeser’s inequality, controls the derivative of this Perron–Frobenius eigenvalue without dividing by the possibly tiny gap between the two components. This shows that, for scaled energies in a fixed neighborhood of μ ∗ \mu_* μ ∗ , the free energy at s = 1 s = 1 s = 1 has exactly one zero, at which it crosses zero transversally and downward. Outside this neighborhood, its sign is given by the leading asymptotics and, at large energies, by the large-energy estimate. Together, these facts give exactly one zero on the half-line, and the crossing is transverse (Theorem 14.1 and Corollary 14.2).
Small energies. For energies below a small multiple of p \sqrt{p} p , the real cavity equation need not have a principal solution, and we do not try to identify the real boundary law. Instead we show in Section 15 that every real solution at an energy $0 < E \leq \xi_{\mathrm{in}}\sqrt{p}$ has free energy at s = 1 s = 1 s = 1 at least a fixed $\delta > 0 (Proposition 15.2). The proof bounds the critical eigenvalue from below by p E ∥ M ∥ − 1 pE\|M\|^{-1} pE ∥ M ∥ − 1 for two scalings and shows, using only the recursion, that one of the two lower bounds exceeds one. Since the delocalization criterion applies to arbitrary real boundary laws, this gives absolute continuity at all small positive energies for the same fixed p p p .
Assembly and the finite graph
The proof of Theorem 1.1 in Section 16 combines these results. On ( 0 , ξ i n p ] (0, \xi_{\mathrm{in}}\sqrt{p}] ( 0 , ξ in p ] , every possible real boundary law has positive free energy at s = 1 s = 1 s = 1 , and the delocalization criterion shows that the spectrum is absolutely continuous. At larger energies of order p \sqrt{p} p , a real boundary law must be the principal solution, whose free energy at s = 1 s = 1 s = 1 is positive below E c ( p ) E_c(p) E c ( p ) and negative above it. Above E c ( p ) E_c(p) E c ( p ) the principal solution is a real solution with negative free energy at s = 1 s = 1 s = 1 , which gives pure point spectrum, including at all large energies. The two absolutely continuous regions overlap, which leaves E c ( p ) E_c(p) E c ( p ) as the only transition.
Finally, in Section 17 we transfer the conclusions to the giant component of the finite graph. A uniformly chosen vertex of the giant component converges locally to the survival-conditioned tree, jointly for finitely many vertices, and the local spectral measures converge accordingly. The density of states of the limiting operator is bounded above and below on compact sets away from zero. On the absolutely continuous side, the normalized local spectral masses are uniformly integrable over vertices, by Lebesgue differentiation and Scheffé’s lemma; on the pure point side, they concentrate on a vanishing fraction of vertices. This proves Theorem 1.4 .
I The limiting tree and the cavity equation
The weighted tree and its operator
In this section, we define the finite model and the marked Poisson–Galton–Watson tree, construct the self-adjoint operator of the tree and of the forests obtained from it by finitely many deletions, and collect the algebraic identities for its resolvent that are used throughout the paper. Section 3.1 defines the models. Section 3.2 defines finite removals and shows that marks which are not inspected by a selection remain independent standard Gaussians. Section 3.3 proves essential self-adjointness of all these operators on finitely supported vectors, identifies their domains, and proves convergence of Dirichlet truncations. Section 3.4 proves the Schur complement formula (the cavity recursion), the path product formula, and the Ward identity. Section 3.5 records a gauge transformation and the parity symmetry. Section 3.6 develops the bipartite factorization of the square of the operator and its decomposition into rank-one terms indexed by rows, which are the basis of the rank-one arguments in Section 7. Finally, Section 3.7 shows that on the survival event the operator has no finitely supported eigenvector with nonzero eigenvalue, while finitely supported eigenvectors with eigenvalue zero occur infinitely often.
Conventions. For a countable set S S S , we write ℓ 2 ( S ) \ell^{2}(S) ℓ 2 ( S ) for the Hilbert space of square-summable functions f : S → C f:S\to\mathbb{C} f : S → C , with the inner product
⟨ f , g ⟩ = ∑ s ∈ S f ( s ) ‾ g ( s ) , \langle f,g\rangle=\sum_{s\in S}\overline{f(s)}g(s), ⟨ f , g ⟩ = s ∈ S ∑ f ( s ) g ( s ) ,
which is antilinear in the first argument. We write δ s \delta_s δ s for the coordinate vector at s ∈ S s\in S s ∈ S and c c ( S ) c_c(S) c c ( S ) for the space of finitely supported functions on S S S . For S ′ ⊂ S S'\subset S S ′ ⊂ S , we identify ℓ 2 ( S ′ ) \ell^{2}(S') ℓ 2 ( S ′ ) with the subspace of ℓ 2 ( S ) \ell^{2}(S) ℓ 2 ( S ) of functions vanishing off S ′ S' S ′ , and we write P S ′ P_{S'} P S ′ for the orthogonal projection onto it. For a self-adjoint operator T T T and a vector f f f , we write μ f T \mu_f^T μ f T for the spectral measure of f f f , defined by ⟨ f , ϕ ( T ) f ⟩ = ∫ ϕ d μ f T \langle f,\phi(T)f\rangle=\int\phi\,d\mu_f^T ⟨ f , ϕ ( T ) f ⟩ = ∫ ϕ d μ f T for bounded Borel functions ϕ \phi ϕ . For a Borel set I ⊂ R I\subset\mathbb{R} I ⊂ R , we say that T T T has pure point spectrum in I I I if Ran 1 I ( T ) \operatorname{Ran}1_I(T) Ran 1 I ( T ) is contained in the closed span of the eigenvectors of T T T , and that T T T is purely absolutely continuous in I I I if the restriction of μ f T \mu_f^T μ f T to I I I is absolutely continuous with respect to Lebesgue measure for every vector f f f . Since the closed span of the eigenvectors consists of the vectors whose spectral measure is purely atomic [113, Section VII.2], the operator T T T has pure point spectrum in I I I if and only if the restriction of μ f T \mu_f^T μ f T to I I I is purely atomic for every f f f . Spectral parameters are written z = E + i η z=E+i\eta z = E + i η with E , η ∈ R E,\eta\in\mathbb{R} E , η ∈ R , and H = { z ∈ C : Im z > 0 } \mathbb{H}=\{z\in\mathbb{C}:\operatorname{Im}z>0\} H = { z ∈ C : Im z > 0 } .
The finite model and the marked tree
A standard complex Gaussian variable is a random variable w w w with density π − 1 e − ∣ ζ ∣ 2 \pi^{-1}e^{-|\zeta|^2} π − 1 e − ∣ ζ ∣ 2 with respect to Lebesgue measure on C \mathbb{C} C . Writing ζ = r e i ϑ \zeta= re^{\mathrm{i}\vartheta} ζ = r e i ϑ , Lebesgue measure on C \mathbb{C} C is r d r d ϑ r\,\mathrm{d}r\,\mathrm{d}\vartheta r d r d ϑ , and in the variables t = r 2 t = r^2 t = r 2 and ϑ ∈ [ 0 , 2 π ) \vartheta\in[0,2\pi) ϑ ∈ [ 0 , 2 π ) the law of w w w is e − t d t ⊗ d ϑ / ( 2 π ) e^{-t}\,\mathrm{d}t \otimes\mathrm{d}\vartheta/(2\pi) e − t d t ⊗ d ϑ / ( 2 π ) . This implies that Y = ∣ w ∣ 2 Y = |w|^2 Y = ∣ w ∣ 2 has the law Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) , the phase w / ∣ w ∣ w/|w| w /∣ w ∣ is uniform on the unit circle, and the two are independent. Also w ‾ \overline{w} w and − w -w − w are standard complex Gaussian variables.
Definition 3.1. Let p > 0 p > 0 p > 0 and let n ≥ 1 n \ge1 n ≥ 1 be an integer with n > p n > p n > p . Let G n G_n G n be the Erdős–Rényi graph on [ n ] = { 1 , … , n } [n] = \{1,\ldots,n\} [ n ] = { 1 , … , n } , in which each pair { i , j } \{i,j\} { i , j } is an edge independently with probability p / n p/n p / n . Let ( w i j ) 1 ≤ i < j ≤ n (w_{ij})_{1\le i<j\le n} ( w ij ) 1 ≤ i < j ≤ n be independent standard complex Gaussian variables, independent of G n G_n G n . Define the Hermitian n × n n \times n n × n matrix H n H_n H n by
H n ( i , j ) = 1 { i j ∈ E ( G n ) } p w i j , H n ( j , i ) = H n ( i , j ) ‾ , H n ( i , i ) = 0 , 1 ≤ i < j ≤ n . H_n(i,j) = \frac{\mathbf{1}_{\{ij\in E(G_n)\}}}{\sqrt{p}}w_{ij}, \qquad H_n(j,i) = \overline{H_n(i,j)}, \qquad H_n(i,i) = 0,\qquad1\le i<j\le n. H n ( i , j ) = p 1 { ij ∈ E ( G n )} w ij , H n ( j , i ) = H n ( i , j ) , H n ( i , i ) = 0 , 1 ≤ i < j ≤ n . The giant component g c ⊂ [ n ] \mathrm{gc} \subset[n] gc ⊂ [ n ] is the vertex set of a connected component of G n G_n G n of maximal size; if there are several, we take the one containing the smallest vertex label. Thus g c \mathrm{gc} gc is a deterministic function of G n G_n G n . We set N n = ∣ g c ∣ N_n = |\mathrm{gc}| N n = ∣ gc ∣ and write H n g c = ( H n ( i , j ) ) i , j ∈ g c H_n^{\mathrm{gc}} = (H_n(i,j))_{i,j\in\mathrm{gc}} H n gc = ( H n ( i , j ) ) i , j ∈ gc for the restriction of H n H_n H n to ℓ 2 ( g c ) \ell^2(\mathrm{gc}) ℓ 2 ( gc ) .
Throughout the paper, the limit n → ∞ n \to\infty n → ∞ is taken with p p p fixed. The local limit of G n G_n G n seen from a uniformly chosen vertex is the Poisson–Galton–Watson tree, and seen from a uniformly chosen vertex of g c \mathrm{gc} gc it is that tree conditioned on survival; this is made precise in Proposition 17.13. We now define the tree and its marks on an explicit probability space.
Let U = ⋃ k ≥ 0 N k \mathbb{U} = \bigcup_{k\ge0}\mathbb{N}^k U = ⋃ k ≥ 0 N k be the set of finite words over N = { 1 , 2 , … } \mathbb{N} = \{1,2,\ldots\} N = { 1 , 2 , … } , with the empty word o o o , and let U ∗ = U ∖ { o } \mathbb{U}_* = \mathbb{U}\setminus\{o\} U ∗ = U ∖ { o } . For u ∈ U u \in\mathbb{U} u ∈ U and j ∈ N j \in\mathbb{N} j ∈ N , we write u j uj u j for the concatenation, ∣ u ∣ |u| ∣ u ∣ for the length of u u u , and u − u^{-} u − for the word obtained from u ≠ o u \ne o u = o by deleting its last letter.
Definition 3.2. Let p > 0 p > 0 p > 0 . Let ( Ω , F , P ) (\Omega,\mathcal{F},\mathbb{P}) ( Ω , F , P ) be a probability space carrying three independent families: random variables ( N u ) u ∈ U (N_u)_{u\in\mathbb{U}} ( N u ) u ∈ U , independent with the law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) ; random variables ( w u ) u ∈ U ∗ (w_u)_{u\in\mathbb{U}_*} ( w u ) u ∈ U ∗ , independent standard complex Gaussians; and an auxiliary random element a a a of a standard Borel space.
(i) The genealogy is the rooted tree T \mathcal{T} T with root o o o , vertex set
V = { j 1 j 2 ⋯ j k ∈ U : k ≥ 0 , j i ≤ N j 1 ⋯ j i − 1 for 1 ≤ i ≤ k } , V = \left\{j_1j_2\cdots j_k \in\mathbb{U}: k \ge0,\ j_i \le N_{j_1\cdots j_{i-1}}\ \text{for }1 \le i \le k\right\}, V = { j 1 j 2 ⋯ j k ∈ U : k ≥ 0 , j i ≤ N j 1 ⋯ j i − 1 for 1 ≤ i ≤ k } , and edges { u − , u } \{u^{-},u\} { u − , u } for u ∈ V ∖ { o } u \in V\setminus\{o\} u ∈ V ∖ { o } . It is a PGW( p ) (p) ( p ) tree: the children of v ∈ V v \in V v ∈ V are v 1 , … , v N v v1,\ldots,vN_v v 1 , … , v N v . The generation of v v v is ∣ v ∣ |v| ∣ v ∣ , the vertex v − v^{-} v − is the parent of v ≠ o v \ne o v = o , and we write u ∼ v u \sim v u ∼ v when u u u and v v v are adjacent. The degree of v v v is deg ( v ) = N v + 1 { v ≠ o } \deg(v) = N_v + \mathbf{1}_{\{v\ne o\}} deg ( v ) = N v + 1 { v = o } . We write N = N o N = N_o N = N o for the number of children of the root.
(ii) The edge { u − , u } \{u^{-},u\} { u − , u } is identified with the word u u u , and w u w_u w u is its mark ; we also write w e w_e w e for the mark of an edge e e e . The edge weights are
h u − u = w u p , h u u − = h u − u ‾ , u ∈ V ∖ { o } , h_{u^{-}u}=\frac{w_u}{\sqrt{p}}, \qquad h_{uu^{-}}=\overline{h_{u^{-}u}}, \qquad u\in V\setminus\{o\}, h u − u = p w u , h u u − = h u − u , u ∈ V ∖ { o } ,
and h u v = 0 h_{uv}=0 h uv = 0 when u u u and v v v are not adjacent. The squared modulus of the mark of the edge u u u is Y u = ∣ w u ∣ 2 = p ∣ h u − u ∣ 2 Y_u=|w_u|^2=p|h_{u^{-}u}|^2 Y u = ∣ w u ∣ 2 = p ∣ h u − u ∣ 2 , which has the law Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) and is independent of the phase w u / ∣ w u ∣ w_u/|w_u| w u /∣ w u ∣ . The pair ( T , ( h u v ) u , v ∈ V ) (\mathcal{T},(h_{uv})_{u,v\in V}) ( T , ( h uv ) u , v ∈ V ) is the marked tree .
(iii) For v ∈ V v\in V v ∈ V , the forward subtree T v \mathcal{T}_v T v is the subtree of T \mathcal{T} T induced on V ( T v ) = { v u : u ∈ U } ∩ V V(\mathcal{T}_v)=\{vu:u\in\mathbb{U}\}\cap V V ( T v ) = { v u : u ∈ U } ∩ V , rooted at v v v . In particular T o = T \mathcal{T}_o=\mathcal{T} T o = T .
(iv) The survival event is S = { ∣ V ∣ = ∞ } S=\{|V|=\infty\} S = { ∣ V ∣ = ∞ } , and ϱ p = P ( S ) \varrho_p=\mathbb{P}(S) ϱ p = P ( S ) . When p > 1 p>1 p > 1 , we have ϱ p ∈ ( 0 , 1 ) \varrho_p\in(0,1) ϱ p ∈ ( 0 , 1 ) , and ϱ p \varrho_p ϱ p is the unique solution in ( 0 , 1 ] (0,1] ( 0 , 1 ] of 1 − ϱ = e − p ϱ 1-\varrho=e^{-p\varrho} 1 − ϱ = e − pϱ . For p > 1 p>1 p > 1 we set P ∞ = P ( ⋅ ∣ S ) \mathbb{P}^{\infty}=\mathbb{P}(\mathord{\cdot}\mid S) P ∞ = P ( ⋅ ∣ S ) and write E ∞ \mathbb{E}^{\infty} E ∞ for the corresponding expectation.
We write G 0 = σ ( ( N u ) u ∈ U , α ) \mathcal{G}_0=\sigma((N_u)_{u\in\mathbb{U}},\alpha) G 0 = σ (( N u ) u ∈ U , α ) for the σ \sigma σ -algebra generated by the genealogy variables and the auxiliary randomness.
The marks w u w_u w u with u ∉ V u\notin V u ∈ / V do not enter the marked tree; including them makes the edge marks of T \mathcal{T} T a subfamily of an independent family indexed by the fixed set U ∗ \mathbb{U}_* U ∗ . Conditionally on G 0 \mathcal{G}_0 G 0 , the marks of the edges of T \mathcal{T} T are independent standard complex Gaussians, and the same holds under P ∞ \mathbb{P}^{\infty} P ∞ , since S ∈ G 0 S\in\mathcal{G}_0 S ∈ G 0 . The value of ϱ p \varrho_p ϱ p is the classical extinction computation for Galton–Watson trees [133, Chapter 3]. The auxiliary element α \alpha α is used in later sections for selections by independent ranks; in Section 4, it is realized as a family ( ζ v ) v ∈ U (\zeta_v)_{v\in\mathbb{U}} ( ζ v ) v ∈ U of independent auxiliary marks indexed by the words.
The families indexed by disjoint subsets of U \mathbb{U} U are independent. We record the consequence that is used to derive distributional identities. For j ∈ N j\in\mathbb{N} j ∈ N , let θ j \theta_j θ j be the shift which maps the families above to ( ( N j u ) u ∈ U , ( w j u ) u ∈ U ∗ ) ((N_{ju})_{u\in\mathbb{U}},(w_{ju})_{u\in\mathbb{U}_*}) (( N j u ) u ∈ U , ( w j u ) u ∈ U ∗ ) . The genealogy and marks built from θ j \theta_j θ j by Definition 3.2 form a marked tree θ j T \theta_j\mathcal{T} θ j T , and if j ≤ N o j\le N_o j ≤ N o , then θ j T \theta_j\mathcal{T} θ j T is the forward subtree T j \mathcal{T}_j T j relabelled by j u ↦ u ju\mapsto u j u ↦ u . The shifted families use the variables indexed by words beginning with j j j , other than w j w_j w j , and we conclude that
θ 1 T , θ 2 T , … are independent copies of T , independent of ( N o , ( w j ) j ≥ 1 ) . (11) \theta_1\mathcal{T},\theta_2\mathcal{T},\ldots\text{ are independent copies of }\mathcal{T},\text{ independent of }(N_o,(w_j)_{j\geq1}). \tag*{(11)} θ 1 T , θ 2 T , … are independent copies of T , independent of ( N o , ( w j ) j ≥ 1 ) . (11)
Finite removals
Many arguments in this paper delete a vertex, an edge, or a few selected leaves from the tree and study the resulting operator. We now define these operations and the class of selections under which the retained marks keep their law.
A finite removal of T \mathcal{T} T is given by a pair ( F , D ) (F,D) ( F , D ) , called its removed set , where F ⊂ V F \subset V F ⊂ V is a finite set of vertices and D D D is a finite set of edges of T \mathcal{T} T ; we also write ( F , D ) (F,D) ( F , D ) for the finite removal. The deleted edges are the edges in D D D and the edges with an endpoint in F F F ; all other edges of T \mathcal{T} T are retained . The reduced forest T F , D \mathcal{T}^{F,D} T F , D is the graph with vertex set V ∖ F V \setminus F V ∖ F whose edges are the retained edges. A component of ( F , D ) (F,D) ( F , D ) is the vertex set of a connected component of T F , D \mathcal{T}^{F,D} T F , D .
An inspected set is a random finite set Q ⊂ U ∗ Q \subset\mathbb{U}_* Q ⊂ U ∗ such that { Q = q } ∈ G 0 q \{Q=q\} \in\mathcal{G}_0^q { Q = q } ∈ G 0 q for every finite q ⊂ U ∗ q \subset\mathbb{U}_* q ⊂ U ∗ , where G 0 q = G 0 ∨ σ ( w u : u ∈ q ) \mathcal{G}_0^q=\mathcal{G}_0\vee\sigma(w_u:u\in q) G 0 q = G 0 ∨ σ ( w u : u ∈ q ) . Its σ \sigma σ -algebra is
G Q = { A ∈ F : A ∩ { Q = q } ∈ G 0 q for every finite q ⊂ U ∗ } . \mathcal{G}_Q=\{A\in\mathcal{F}:A\cap\{Q=q\}\in\mathcal{G}_0^q\text{ for every finite }q\subset\mathbb{U}_*\}. G Q = { A ∈ F : A ∩ { Q = q } ∈ G 0 q for every finite q ⊂ U ∗ } .
A random finite removal ( F , D ) (F,D) ( F , D ) is selected without inspecting the retained marks if there exists an inspected set Q Q Q such that { ( F , D ) = ( f , d ) } ∈ G Q \{(F,D)=(f,d)\}\in\mathcal{G}_Q {( F , D ) = ( f , d )} ∈ G Q for all finite sets f , d f,d f , d , and such that almost surely no edge with label in Q Q Q is retained. If this holds with Q = ∅ Q=\varnothing Q = ∅ , that is, if ( F , D ) (F,D) ( F , D ) is a function of the genealogy and the auxiliary randomness, we say that ( F , D ) (F,D) ( F , D ) is selected from the genealogy .
An inspected set is the set of labels of the marks that a selection procedure looks at; the procedure may decide which marks to look at using what it has seen so far, and G Q \mathcal{G}_Q G Q is the information available at the end. The finite removals used later are the following.
(a) Deleting a vertex y y y : ( F , D ) = ( { y } , ∅ ) (F,D)= (\{y\},\varnothing) ( F , D ) = ({ y } , ∅ ) . Its components are the forward subtrees V ( T y j ) V(\mathcal{T}_{y_j}) V ( T y j ) , 1 ≤ j ≤ N y 1\le j\le N_y 1 ≤ j ≤ N y , and, if y ≠ o y\ne o y = o , the set V ∖ V ( T y ) V\setminus V(\mathcal{T}_y) V ∖ V ( T y ) .
(b) Deleting an edge { v − , v } \{v^-,v\} { v − , v } : ( F , D ) = ( ∅ , { v } ) (F,D)=(\varnothing,\{v\}) ( F , D ) = ( ∅ , { v }) . Its components are V ( T v ) V(\mathcal{T}_v) V ( T v ) and V ∖ V ( T v ) V\setminus V(\mathcal{T}_v) V ∖ V ( T v ) . In particular, every forward subtree is a component of a finite removal.
(c) Deleting selected pendant leaves : F F F is a finite set of leaf children (children ℓ \ell ℓ with N ℓ = 0 N_\ell=0 N ℓ = 0 ) chosen as a function of the genealogy and of auxiliary ranks carried by α \alpha α , and D = ∅ D=\varnothing D = ∅ . This finite removal is selected from the genealogy.
(d) Dirichlet balls : deleting the finitely many edges between generations R R R and R + 1 R+1 R + 1 leaves the ball { v ∈ V : ∣ v ∣ ≤ R } \{v\in V:|v|\le R\} { v ∈ V : ∣ v ∣ ≤ R } as the component of o o o .
Lemma 3.4. Let Q Q Q be an inspected set. For every A ∈ G Q A\in\mathcal{G}_Q A ∈ G Q , every finite q ⊂ U ∗ q\subset\mathbb{U}_* q ⊂ U ∗ , and every bounded measurable function Ψ \Psi Ψ of the family ( w u ) u ∈ U ∗ ∖ q (w_u)_{u\in\mathbb{U}_*\setminus q} ( w u ) u ∈ U ∗ ∖ q , we have
E [ 1 A ∩ { Q = q } Ψ ( ( w u ) u ∉ q ) ] = P ( A ∩ { Q = q } ) E [ Ψ ( ( w u ) u ∉ q ) ] . (12) \mathbb{E}\left[1_{A\cap\{Q=q\}}\Psi\left((w_u)_{u\notin q}\right)\right]
=\mathbb{P}(A\cap\{Q=q\})\mathbb{E}\left[\Psi\left((w_u)_{u\notin q}\right)\right]. \tag*{(12)} E [ 1 A ∩ { Q = q } Ψ ( ( w u ) u ∈ / q ) ] = P ( A ∩ { Q = q }) E [ Ψ ( ( w u ) u ∈ / q ) ] . (12) The same identity holds with P , E \mathbb{P},\mathbb{E} P , E replaced by P ∞ , E ∞ \mathbb{P}^{\infty},\mathbb{E}^{\infty} P ∞ , E ∞ when p > 1 p>1 p > 1 . If ( F , D ) (F,D) ( F , D ) is selected without inspecting the retained marks, then conditionally on G Q \mathcal{G}_Q G Q the marks of the retained edges are independent standard complex Gaussian variables, under P \mathbb{P} P and under P ∞ \mathbb{P}^{\infty} P ∞ .
Proof. Fix A A A and q q q . By the definition of G Q \mathcal{G}_Q G Q , the event A ∩ { Q = q } A\cap\{Q=q\} A ∩ { Q = q } belongs to G 0 q = σ ( ( N u ) u ∈ U , α , ( w u ) u ∈ q ) \mathcal{G}_0^q=\sigma((N_u)_{u\in\mathbb{U}},\alpha,(w_u)_{u\in q}) G 0 q = σ (( N u ) u ∈ U , α , ( w u ) u ∈ q ) . Since the three families in Definition 3.2 are independent and
the marks are independent of each other, G 0 q \mathcal{G}_0^q G 0 q is independent of ( w u ) u ∉ q (w_u)_{u\notin q} ( w u ) u ∈ / q , and (3.2) follows. Since S ∈ G 0 S \in\mathcal{G}_0 S ∈ G 0 , the event A ∩ S A \cap S A ∩ S belongs to G Q \mathcal{G}_Q G Q ; applying (3.2) to it and dividing by ϱ p \varrho_p ϱ p gives the identity under P ∞ \mathbb{P}^{\infty} P ∞ , since E [ Ψ ] = E ∞ [ Ψ ] \mathbb{E}[\Psi] = \mathbb{E}^{\infty}[\Psi] E [ Ψ ] = E ∞ [ Ψ ] by the independence of S ∈ G 0 S \in\mathcal{G}_0 S ∈ G 0 and the marks.
For the last assertion, let ( w u ′ ) u ∈ U ∗ (w'_u)_{u\in\mathcal{U}_*} ( w u ′ ) u ∈ U ∗ be an independent copy of the marks, defined on a product extension of the probability space, and set w ~ u = w u \widetilde{w}_u = w_u w u = w u for u ∉ Q u \notin Q u ∈ / Q and w ~ u = w u ′ \widetilde{w}_u = w'_u w u = w u ′ for u ∈ Q u \in Q u ∈ Q . For fixed q q q , the family ( ( w u ) u ∉ q , ( w u ′ ) u ∈ q ) \big((w_u)_{u\notin q},(w'_u)_{u\in q}\big) ( ( w u ) u ∈ / q , ( w u ′ ) u ∈ q ) has the law of ( w u ) u ≠ o (w_u)_{u\ne o} ( w u ) u = o and is independent of G 0 q \mathcal{G}_0^q G 0 q , by the argument above and since w ′ w' w ′ is independent of everything else. For A ∈ G Q A \in\mathcal{G}_Q A ∈ G Q and every bounded measurable function Ψ \Psi Ψ of the whole family, we then have
E [ 1 A Ψ ( w ~ ) ] = ∑ q E [ 1 A ∩ { Q = q } Ψ ( ( w u ) u ∉ q , ( w u ′ ) u ∈ q ) ] = ∑ q P ( A ∩ { Q = q } ) E [ Ψ ( w ) ] = P ( A ) E [ Ψ ( w ) ] . \begin{aligned}
\mathbb{E}\left[1_A\Psi(\widetilde{w})\right] = \sum_q \mathbb{E}\left[1_{A\cap\{Q=q\}}\Psi\left((w_u)_{u\notin q},(w'_u)_{u\in q}\right)\right] \\
&= \sum_q \mathbb{P}(A\cap\{Q=q\})\mathbb{E}[\Psi(w)] = \mathbb{P}(A)\mathbb{E}[\Psi(w)].
\end{aligned} E [ 1 A Ψ ( w ) ] = q ∑ E [ 1 A ∩ { Q = q } Ψ ( ( w u ) u ∈ / q , ( w u ′ ) u ∈ q ) ] = q ∑ P ( A ∩ { Q = q }) E [ Ψ ( w )] = P ( A ) E [ Ψ ( w )] .
where the sums run over the countably many finite sets q q q . This shows that w ~ \widetilde{w} w is independent of G Q \mathcal{G}_Q G Q and is a family of independent standard complex Gaussians. The set of retained edges is a function of T \mathcal{T} T and ( F , D ) (F,D) ( F , D ) , and in particular it is G Q \mathcal{G}_Q G Q -measurable. Since it is almost surely disjoint from Q Q Q , the retained marks coincide with the corresponding entries of w ~ \widetilde{w} w . The conditional law, given G Q \mathcal{G}_Q G Q , of the entries of an independent i.i.d. family indexed by a G Q \mathcal{G}_Q G Q -measurable random set is the product law over that set [81, Theorem 6.4]. This proves the claim under P \mathbb{P} P , and the same argument applies under P ∞ \mathbb{P}^{\infty} P ∞ .
For a finite removal selected from the genealogy, conditionally on G 0 \mathcal{G}_0 G 0 all marks, deleted or retained, are independent standard complex Gaussians. For the pendant leaves in (c), this means that the squared moduli of the marks of the deleted leaf edges are independent Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) variables, independent of the genealogy and of all retained marks.
We next record some elementary facts about the shape of the components. For a component C C C of a finite removal and adjacent vertices u , w ∈ C u,w \in C u , w ∈ C , the branch C u → w C_{u\to w} C u → w is the component containing u u u of the graph obtained from T F , D \mathcal{T}^{F,D} T F , D by also deleting the edge { u , w } \{u,w\} { u , w } .
Lemma 3.5. Let ( F , D ) (F,D) ( F , D ) be a finite removal of T \mathcal{T} T .
(i) The reduced forest T F , D \mathcal{T}^{F,D} T F , D has finitely many components. Every component C C C is the vertex set of an induced subtree of T \mathcal{T} T : every edge of T \mathcal{T} T with both endpoints in C C C is retained.
(ii) Let C C C be a component and let u , w ∈ C u,w \in C u , w ∈ C be adjacent. Then C u → w C_{u\to w} C u → w is a component of the finite removal ( F , D ∪ { { u , w } } ) (F,D \cup\{\{u,w\}\}) ( F , D ∪ {{ u , w }}) , and it consists of the vertices x ∈ C x \in C x ∈ C such that the path in T \mathcal{T} T from x x x to w w w contains u u u . The neighbors of u u u in C C C lie in distinct components of ( F ∪ { u } , D ) (F \cup\{u\},D) ( F ∪ { u } , D ) , and the component containing a neighbor x x x is C x → u C_{x\to u} C x → u .
(iii) Let C C C be as in (ii), let u , w ∈ C u,w \in C u , w ∈ C be adjacent, and let C ′ C' C ′ be a component of a finite removal ( F ′ , D ′ ) (F',D') ( F ′ , D ′ ) with F ⊂ F ′ F \subset F' F ⊂ F ′ and D ⊂ D ′ D \subset D' D ⊂ D ′ . If C u → w ⊂ C ′ C_{u\to w} \subset C' C u → w ⊂ C ′ and w ∈ C ′ w \in C' w ∈ C ′ , then C u → w ′ = C u → w C'_{u\to w} = C_{u\to w} C u → w ′ = C u → w .
Proof. Every edge of a tree disconnects it, and every edge of a forest disconnects the component that contains it. By induction on the number of deleted edges, deleting them from T \mathcal{T} T produces a forest with exactly one more component than deleted edges. The vertices of F F F are isolated in this forest, and its other components are the components of ( F , D ) (F,D) ( F , D ) ; this proves the first claim of (i). If an edge { u , v } \{u,v\} { u , v } of T \mathcal{T} T with u , v ∈ C u,v \in C u , v ∈ C were deleted, then the path from u u u to v v v in the connected graph T F , D \mathcal{T}^{F,D} T F , D restricted to C C C would avoid this edge, and together with the edge it would form a cycle in T \mathcal{T} T . This proves (i).
For (ii), the first assertion holds because adding { u , w } \{u,w\} { u , w } to D D D changes only the component C C C , which splits into two. Since C C C is a subtree, paths in C C C are paths in T \mathcal{T} T . A vertex x ∈ C x \in C x ∈ C is connected to u u u in C C C without using the edge { u , w } \{u,w\} { u , w } exactly when the path from x x x to u u u does not contain w w w , and this holds exactly when the path from x x x to w w w is the path from x x x to u u u followed by the edge { u , w } \{u,w\} { u , w } . This proves the description of C u → w C_{u\to w} C u → w . If x x x and x ′ x' x ′ are distinct neighbors of u u u in C C C , then they lie in distinct components after deleting u u u , since the unique path between them is x , u , x ′ x,u,x' x , u , x ′ . The component of x x x after deleting u u u consists of the vertices of C C C connected to x x x by a path avoiding u u u , and this is C x → u C_{x\to u} C x → u by the description just proved.
For (iii), we have C ′ ⊂ C C' \subset C C ′ ⊂ C , because every edge retained by ( F ′ , D ′ ) (F',D') ( F ′ , D ′ ) is retained by ( F , D ) (F,D) ( F , D ) . Since a path in C ′ C' C ′ which avoids the edge { u , w } \{u,w\} { u , w } is such a path in C C C , we have C u → w ′ ⊂ C u → w C'_{u\to w} \subset C_{u\to w} C u → w ′ ⊂ C u → w . Conversely, C u → w C_{u\to w} C u → w induces a connected subtree of T \mathcal{T} T by (i) and (ii), it is contained in C ′ C' C ′ , and C ′ C' C ′ is an induced subtree by (i). Then every vertex of C u → w C_{u\to w} C u → w is joined to u u u by a path in C ′ C' C ′ avoiding { u , w } \{u,w\} { u , w } , which gives C u → w ⊂ C u → w ′ C_{u\to w} \subset C'_{u\to w} C u → w ⊂ C u → w ′ .
Essential self-adjointness
The weighted adjacency operator of T \mathcal{T} T need not be bounded. In this section, we show that it is essentially self-adjoint on finitely supported vectors, simultaneously for all components of all finite removals. The argument follows [51, Appendix A]: by [51, Lemma A.3], an exhaustion of a weighted tree by finite connected sets whose boundary weights are uniformly bounded implies essential self-adjointness, and such an exhaustion of the Poisson weighted infinite tree is constructed in [51, Proposition A.2] by a subcritical exploration; see also [52] for the unweighted case. We include the proof for our model for completeness.
Hermitian networks
A locally finite Hermitian network is a countable set V \mathcal{V} V with weights k : V × V → C k:\mathcal{V}\times\mathcal{V}\to\mathbb{C} k : V × V → C such that k v u = k u v k_{vu}=k_{uv} k v u = k uv and k v w = 0 k_{vw}=0 k v w = 0 for all u , v u,v u , v , and such that { u : k v u ≠ 0 } \{u:k_{vu}\ne0\} { u : k v u = 0 } is finite for every v v v . Its formal operator acts on all functions f : V → C f:\mathcal{V}\to\mathbb{C} f : V → C by
( T f o r m f ) ( v ) = ∑ u ∈ V k v u f ( u ) , (T_{\mathrm{form}}f)(v)=\sum_{u\in\mathcal{V}} k_{vu}f(u), ( T form f ) ( v ) = u ∈ V ∑ k v u f ( u ) ,
which is a finite sum. The minimal operator T 0 T_0 T 0 is the restriction of T f o r m T_{\mathrm{form}} T form to c c ( V ) c_c(\mathcal{V}) c c ( V ) , viewed as an operator in ℓ 2 ( V ) \ell^2(\mathcal{V}) ℓ 2 ( V ) , and the maximal domain is { f ∈ ℓ 2 ( V ) : T f o r m f ∈ ℓ 2 ( V ) } \{f \in\ell^2(\mathcal{V}) : T_{\mathrm{form}}f \in\ell^2(\mathcal{V})\} { f ∈ ℓ 2 ( V ) : T form f ∈ ℓ 2 ( V )} . For a finite set S ⊂ V S \subset\mathcal{V} S ⊂ V , the boundary operator B S B_S B S is defined by
( B S f ) ( v ) = 1 S ( v ) ∑ u ∉ S k v u f ( u ) , f ∈ ℓ 2 ( V ) . (B_Sf)(v)=\mathbf{1}_S(v)\sum_{u\notin S}k_{vu}f(u),\qquad f\in\ell^2(\mathcal{V}). ( B S f ) ( v ) = 1 S ( v ) u ∈ / S ∑ k v u f ( u ) , f ∈ ℓ 2 ( V ) .
Since only the finitely many u ∉ S u\notin S u ∈ / S with a neighbor in S S S enter this sum, B S B_S B S is a bounded operator of finite rank.
On the marked tree, ( V , ( h u v ) u , v ∈ V ) (\mathcal{V},(h_{uv})_{u,v\in\mathcal{V}}) ( V , ( h uv ) u , v ∈ V ) is a locally finite Hermitian network. If C C C is a component of a finite removal, then ( C , ( h u v ) u , v ∈ C ) (C,(h_{uv})_{u,v\in C}) ( C , ( h uv ) u , v ∈ C ) is a locally finite Hermitian network as well, and by Lemma 3.5(i) its weights are those of the retained edges inside C C C . We write H C , f o r m \mathcal{H}_{C,\mathrm{form}} H C , form and H C , 0 \mathcal{H}_{C,0} H C , 0 for its formal and minimal operators, and H f o r m \mathcal{H}_{\mathrm{form}} H form and H 0 \mathcal{H}_0 H 0 when C = V C=\mathcal{V} C = V .
Lemma 3.6. Let ( V , k ) (\mathcal{V},k) ( V , k ) be a locally finite Hermitian network. Then T 0 T_0 T 0 is symmetric, and T 0 ∗ T_0^* T 0 ∗ is the restriction of T f o r m T_{\mathrm{form}} T form to the maximal domain. If T 0 T_0 T 0 is essentially self-adjoint, then its closure T T T satisfies
Dom ( T ) = { f ∈ ℓ 2 ( V ) : T f o r m f ∈ ℓ 2 ( V ) } , T f = T f o r m f for f ∈ Dom ( T ) . \operatorname{Dom}(T)=\{f\in\ell^2(\mathcal{V}):T_{\mathrm{form}}f\in\ell^2(\mathcal{V})\},\qquad Tf=T_{\mathrm{form}}f\quad\text{for }f\in\operatorname{Dom}(T). Dom ( T ) = { f ∈ ℓ 2 ( V ) : T form f ∈ ℓ 2 ( V )} , T f = T form f for f ∈ Dom ( T ) .
Proof. Let g ∈ c c ( V ) g\in c_c(\mathcal{V}) g ∈ c c ( V ) and f ∈ ℓ 2 ( V ) f\in\ell^2(\mathcal{V}) f ∈ ℓ 2 ( V ) . Only the finitely many pairs ( u , v ) (u,v) ( u , v ) with u ∈ supp g u\in\operatorname{supp}g u ∈ supp g and k u v ≠ 0 k_{uv}\ne0 k uv = 0 contribute to the following sums. Exchanging these finite sums and using k v u ‾ = k u v \overline{k_{vu}}=k_{uv} k v u = k uv , we have
⟨ T 0 g , f ⟩ = ∑ v ∑ u k v u ‾ g ( u ) f ( v ) ‾ = ∑ u g ( u ) ∑ v k u v f ( v ) ‾ = ⟨ g , T f o r m f ⟩ . (13) \langle T_0g,f\rangle=\sum_v\sum_u\overline{k_{vu}}g(u)\overline{f(v)}=\sum_u g(u)\sum_v k_{uv}\overline{f(v)}=\langle g,T_{\mathrm{form}}f\rangle. \tag*{(13)} ⟨ T 0 g , f ⟩ = v ∑ u ∑ k v u g ( u ) f ( v ) = u ∑ g ( u ) v ∑ k uv f ( v ) = ⟨ g , T form f ⟩ . (13)
Taking f ∈ c c ( V ) f\in c_c(\mathcal{V}) f ∈ c c ( V ) , we see that T 0 T_0 T 0 is symmetric. If f f f is in the maximal domain, then (13) implies that f ∈ Dom ( T 0 ∗ ) f\in\operatorname{Dom}(T_0^*) f ∈ Dom ( T 0 ∗ ) and T 0 ∗ f = T f o r m f T_0^*f=T_{\mathrm{form}}f T 0 ∗ f = T form f . Conversely, if f ∈ Dom ( T 0 ∗ ) f\in\operatorname{Dom}(T_0^*) f ∈ Dom ( T 0 ∗ ) , then applying (13) with g = δ v g=\delta_v g = δ v , we have ( T 0 ∗ f ) ( v ) = ( T f o r m f ) ( v ) (T_0^*f)(v)=(T_{\mathrm{form}}f)(v) ( T 0 ∗ f ) ( v ) = ( T form f ) ( v ) for every v v v , that is, T f o r m f = T 0 ∗ f ∈ ℓ 2 T_{\mathrm{form}}f=T_0^*f\in\ell^2 T form f = T 0 ∗ f ∈ ℓ 2 . If T 0 T_0 T 0 is essentially self-adjoint, then its closure T T T is self-adjoint and T = T ∗ = T 0 ∗ T=T^*=T_0^* T = T ∗ = T 0 ∗ , since an operator and its closure have the same adjoint. □ \square □
Lemma 3.7. Let ( V , k ) (\mathcal{V},k) ( V , k ) be a locally finite Hermitian network. Suppose that there exist finite sets S 1 ⊂ S 2 ⊂ ⋯ S_1\subset S_2\subset\cdots S 1 ⊂ S 2 ⊂ ⋯ with ⋃ n S n = V \bigcup_n S_n=\mathcal{V} ⋃ n S n = V and M = sup n ∥ B S n ∥ < ∞ M=\sup_n\lVert B_{S_n}\rVert<\infty M = sup n ∥ B S n ∥ < ∞ . Then T 0 T_0 T 0 is essentially self-adjoint on c c ( V ) c_c(\mathcal{V}) c c ( V ) .
Proof. By [113, Theorem VIII.3 and its corollary], the symmetric operator T 0 T_0 T 0 is essentially self-adjoint if ker ( T 0 ∗ − z ) = { 0 } \ker(T_0^*-z)=\{0\} ker ( T 0 ∗ − z ) = { 0 } for z = i z=i z = i and z = − i z=-i z = − i . Fix z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R and ψ ∈ ker ( T 0 ∗ − z ) \psi\in\ker(T_0^*-z) ψ ∈ ker ( T 0 ∗ − z ) . By Lemma 3.6, we have T f o r m ψ = z ψ T_{\mathrm{form}}\psi=z\psi T form ψ = z ψ pointwise. Let S ⊂ V S\subset\mathcal{V} S ⊂ V be finite. Multiplying the equation at v v v by ψ ( v ) ‾ \overline{\psi(v)} ψ ( v ) and summing over v ∈ S v\in S v ∈ S , we obtain
z ∥ P S ψ ∥ 2 = ∑ v ∈ S ∑ u ∈ S ψ ( v ) ‾ k v u ψ ( u ) + ∑ v ∈ S ∑ u ∉ S ψ ( v ) ‾ k v u ψ ( u ) = ∑ u , v ∈ S ψ ( v ) ‾ k v u ψ ( u ) + ⟨ P S ψ , B S ψ ⟩ . z\lVert P_S\psi\rVert^2=\sum_{v\in S}\sum_{u\in S}\overline{\psi(v)}k_{vu}\psi(u)+\sum_{v\in S}\sum_{u\notin S}\overline{\psi(v)}k_{vu}\psi(u)=\sum_{u,v\in S}\overline{\psi(v)}k_{vu}\psi(u)+\langle P_S\psi,B_S\psi\rangle. z ∥ P S ψ ∥ 2 = v ∈ S ∑ u ∈ S ∑ ψ ( v ) k v u ψ ( u ) + v ∈ S ∑ u ∈ / S ∑ ψ ( v ) k v u ψ ( u ) = u , v ∈ S ∑ ψ ( v ) k v u ψ ( u ) + ⟨ P S ψ , B S ψ ⟩ .
The first sum on the right is real, because the matrix ( k v u ) u , v ∈ S (k_{vu})_{u,v\in S} ( k v u ) u , v ∈ S is Hermitian. Taking imaginary parts, and using that B S ψ = B S P V ∖ S ψ B_S\psi=B_SP_{\mathcal{V}\setminus S}\psi B S ψ = B S P V ∖ S ψ , we find
∣ Im z ∣ ∥ P S ψ ∥ 2 ≤ ∣ ⟨ P S ψ , B S P V ∖ S ψ ⟩ ∣ ≤ ∥ B S ∥ ∥ P S ψ ∥ ∥ P V ∖ S ψ ∥ . (14) |\operatorname{Im}z|\,\lVert P_S\psi\rVert^2\leq\left|\langle P_S\psi,B_SP_{\mathcal{V}\setminus S}\psi\rangle\right|\leq\lVert B_S\rVert\,\lVert P_S\psi\rVert\,\lVert P_{\mathcal{V}\setminus S}\psi\rVert. \tag*{(14)} ∣ Im z ∣ ∥ P S ψ ∥ 2 ≤ ⟨ P S ψ , B S P V ∖ S ψ ⟩ ≤ ∥ B S ∥ ∥ P S ψ ∥ ∥ P V ∖ S ψ ∥ . (14)
We apply (14) with S = S n S=S_n S = S n . As n → ∞ n\to\infty n → ∞ , we have ∥ P S n ψ ∥ → ∥ ψ ∥ \lVert P_{S_n}\psi\rVert\to\lVert\psi\rVert ∥ P S n ψ ∥ → ∥ ψ ∥ and ∥ P V ∖ S n ψ ∥ → 0 \lVert P_{\mathcal{V}\setminus S_n}\psi\rVert\to0 ∥ P V ∖ S n ψ ∥ → 0 , since S n ↑ V S_n\uparrow\mathcal{V} S n ↑ V and ψ ∈ ℓ 2 ( V ) \psi\in\ell^2(\mathcal{V}) ψ ∈ ℓ 2 ( V ) . The right side of (14) is at most M ∥ ψ ∥ ∥ P V ∖ S n ψ ∥ M\lVert\psi\rVert\lVert P_{\mathcal{V}\setminus S_n}\psi\rVert M ∥ ψ ∥ ∥ P V ∖ S n ψ ∥ , which tends to zero. We conclude that ∣ Im z ∣ ∥ ψ ∥ 2 = 0 |\operatorname{Im}z|\lVert\psi\rVert^2=0 ∣ Im z ∣ ∥ ψ ∥ 2 = 0 and ψ = 0 \psi=0 ψ = 0 .
A bounded-boundary exhaustion of the tree
For u ∈ U u\in\mathbb{U} u ∈ U , set
R u 2 = 1 p ∑ j = 1 N u ∣ w u j ∣ 2 . R_u^2=\frac{1}{p}\sum_{j=1}^{N_u}|w_{uj}|^2. R u 2 = p 1 j = 1 ∑ N u ∣ w u j ∣ 2 .
For u ∈ V u\in\mathcal{V} u ∈ V , this is ∑ j ≤ N u ∣ h u , u j ∣ 2 \sum_{j\leq N_u}|h_{u,u_j}|^2 ∑ j ≤ N u ∣ h u , u j ∣ 2 , the sum of the squared weights of the edges from u u u to its children. Since R o < ∞ R_o<\infty R o < ∞ almost surely and N o 1 { R o > M } ≤ N o N_o\mathbf{1}_{\{R_o>M\}}\leq N_o N o 1 { R o > M } ≤ N o , which is integrable, by dominated convergence we have E [ N o 1 { R o > M } ] → 0 \mathbb{E}[N_o\mathbf{1}_{\{R_o>M\}}]\to0 E [ N o 1 { R o > M } ] → 0 as M → ∞ M\to\infty M → ∞ . We fix a deterministic M = M ( p ) < ∞ M=M(p)<\infty M = M ( p ) < ∞ such that
m M = E [ N o 1 { R o > M } ] < 1. (15) m_M=\mathbb{E}\left[N_o\mathbf{1}_{\{R_o>M\}}\right]<1. \tag*{(15)} m M = E [ N o 1 { R o > M } ] < 1. (15)
We call u ∈ U u\in\mathbb{U} u ∈ U good if R u ≤ M R_u\leq M R u ≤ M and bad otherwise. For v ∈ V v\in\mathcal{V} v ∈ V , let λ ( v ) \lambda(v) λ ( v ) be the number of good vertices among the strict ancestors of v v v , and set
S n = { v ∈ V : λ ( v ) ≤ n } , n ≥ 0. S_n=\{v\in\mathcal{V}:\lambda(v)\leq n\},\qquad n\geq0. S n = { v ∈ V : λ ( v ) ≤ n } , n ≥ 0.
Lemma 3.8. Let p > 0 p>0 p > 0 . There exists an event Ω s a \Omega_{\mathrm{sa}} Ω sa with P ( Ω s a ) = 1 \mathbb{P}(\Omega_{\mathrm{sa}})=1 P ( Ω sa ) = 1 on which the following holds. Every S n S_n S n is finite, o ∈ S n ⊂ S n + 1 o\in S_n\subset S_{n+1} o ∈ S n ⊂ S n + 1 , every ancestor of a vertex of S n S_n S n belongs to S n S_n S n , and ⋃ n S n = V \bigcup_n S_n=\mathcal{V} ⋃ n S n = V . Further, for every n n n , every vertex outside S n S_n S n has at most one neighbor in S n S_n S n , and
∑ u ∉ S n , u ∼ v ∣ h v u ∣ 2 ≤ M 2 for every v ∈ S n . (16) \sum_{u\notin S_n,\ u\sim v}|h_{vu}|^2\leq M^2\qquad\text{for every }v\in S_n. \tag*{(16)} u ∈ / S n , u ∼ v ∑ ∣ h v u ∣ 2 ≤ M 2 for every v ∈ S n . (16) In particular, the boundary operators of the network ( V , h ) (\mathcal{V},h) ( V , h ) satisfy ∥ B S n ∥ ≤ M \lVert B_{S_n}\rVert\leq M ∥ B S n ∥ ≤ M .
Proof. The proof shows that the vertices reachable from a given vertex through bad vertices form a subcritical Galton–Watson tree, and that the sets S n S_n S n are finite unions of such trees.
Step 1. The bad clusters are subcritical Galton–Watson trees. For u ∈ U u\in\mathbb{U} u ∈ U , let χ u = N u 1 { R u > M } \chi_u=N_u\mathbf{1}_{\{R_u>M\}} χ u = N u 1 { R u > M } . This is a function of N u N_u N u and of ( w u j ) j ∈ N (w_{uj})_{j\in\mathbb{N}} ( w u j ) j ∈ N . Since for distinct u u u the variables N u N_u N u are distinct and the families ( w u j ) j ∈ N (w_{uj})_{j\in\mathbb{N}} ( w u j ) j ∈ N are indexed by disjoint subsets of U \mathbb{U} U , the family ( χ u ) u ∈ U (\chi_u)_{u\in\mathbb{U}} ( χ u ) u ∈ U is independent and identically distributed, with mean m M < 1 m_M<1 m M < 1 . For x ∈ U x\in\mathbb{U} x ∈ U , define the bad cluster
K x = { x j 1 ⋯ j k : k ≥ 0 , j i ≤ χ x j 1 ⋯ j i − 1 for 1 ≤ i ≤ k } . K_x=\{xj_1\cdots j_k:k\geq0,\ j_i\leq\chi_{xj_1\cdots j_{i-1}}\text{ for }1\leq i\leq k\}. K x = { x j 1 ⋯ j k : k ≥ 0 , j i ≤ χ x j 1 ⋯ j i − 1 for 1 ≤ i ≤ k } .
Then K x K_x K x is a Galton–Watson tree whose offspring numbers ( χ x u ) u ∈ U (\chi_{xu})_{u\in\mathbb{U}} ( χ xu ) u ∈ U are independent with mean m M m_M m M . Let Z k x Z_k^x Z k x be the number of words of length ∣ x ∣ + k |x|+k ∣ x ∣ + k in K x K_x K x . The event { u ∈ K x } \{u\in K_x\} { u ∈ K x } is determined by the variables χ \chi χ at the strict ancestors of u u u , which are independent of χ u \chi_u χ u . Summing over the countably many words u u u of length ∣ x ∣ + k |x|+k ∣ x ∣ + k , we obtain
E Z k + 1 x = ∑ ∣ u ∣ = ∣ x ∣ + k E [ 1 { u ∈ K x } χ u ] = m M E Z k x \mathbb{E}Z_{k+1}^x=\sum_{|u|=|x|+k}\mathbb{E}\left[1_{\{u\in K_x\}}\chi_u\right]=m_M\mathbb{E}Z_k^x E Z k + 1 x = ∣ u ∣ = ∣ x ∣ + k ∑ E [ 1 { u ∈ K x } χ u ] = m M E Z k x
so E Z k x = m M k \mathbb{E}Z_k^x=m_M^k E Z k x = m M k and E ∣ K x ∣ = ( 1 − m M ) − 1 < ∞ \mathbb{E}|K_x|=(1-m_M)^{-1}<\infty E ∣ K x ∣ = ( 1 − m M ) − 1 < ∞ . Let Ω s a \Omega_{\mathrm{sa}} Ω sa be the event that K x K_x K x is finite for every x ∈ U x\in\mathbb{U} x ∈ U ; as a countable intersection of events of probability one, it satisfies P ( Ω s a ) = 1 \mathbb{P}(\Omega_{\mathrm{sa}})=1 P ( Ω sa ) = 1 .
Since j ≤ χ u j\leq\chi_u j ≤ χ u holds exactly when u u u is bad and j ≤ N u j\leq N_u j ≤ N u , we have, for x ∈ V x\in V x ∈ V ,
K x = { y ∈ V ( T x ) : every vertex u ≠ y on the path from x to y is bad } . (17) K_x=\{y\in V(T_x):\text{ every vertex }u\ne y\text{ on the path from }x\text{ to }y\text{ is bad}\}. \tag*{(17)} K x = { y ∈ V ( T x ) : every vertex u = y on the path from x to y is bad } . (17)
Step 2. The sets S n S_n S n are finite and exhaust V V V . For v ∈ V v\in V v ∈ V and j ≤ N v j\leq N_v j ≤ N v , we have λ ( v j ) = λ ( v ) + 1 { v g o o d } \lambda(vj)=\lambda(v)+1_{\{v\ \mathrm{good}\}} λ ( v j ) = λ ( v ) + 1 { v good } . This implies that λ \lambda λ is nondecreasing along every path away from the root, and S n S_n S n contains every ancestor of its vertices, and o ∈ S n o\in S_n o ∈ S n since λ ( o ) = 0 \lambda(o)=0 λ ( o ) = 0 . Also λ ( v ) ≤ ∣ v ∣ \lambda(v)\leq|v| λ ( v ) ≤ ∣ v ∣ , which gives ⋃ n S n = V \bigcup_n S_n=V ⋃ n S n = V . We show by induction that S n S_n S n is finite on Ω s a \Omega_{\mathrm{sa}} Ω sa . By (17) , we have S 0 = K o S_0=K_o S 0 = K o . Let n ≥ 0 n\geq0 n ≥ 0 and let v ∈ V v\in V v ∈ V satisfy λ ( v ) = n + 1 \lambda(v)=n+1 λ ( v ) = n + 1 . Let g g g be the strict ancestor of v v v of largest generation which is good, and let t t t be the child of g g g on the path from g g g to v v v (possibly t = v t=v t = v ). Since the good strict ancestors of v v v are g g g and the good strict ancestors of g g g , we have λ ( g ) = n \lambda(g)=n λ ( g ) = n . All vertices on the path from t t t to v v v , other than v v v , are bad, and (17) implies that v ∈ K t v\in K_t v ∈ K t . We conclude that
{ v ∈ V : λ ( v ) = n + 1 } ⊂ ⋃ g ∈ S n ⋃ j = 1 N g K g j . \{v\in V:\lambda(v)=n+1\}\subset\bigcup_{g\in S_n}\bigcup_{j=1}^{N_g}K_{gj}. { v ∈ V : λ ( v ) = n + 1 } ⊂ g ∈ S n ⋃ j = 1 ⋃ N g K g j .
which is a finite union of finite sets on Ω s a \Omega_{\mathrm{sa}} Ω sa when S n S_n S n is finite.
Step 3. The boundary of S n S_n S n . Let v ∈ S n v\in S_n v ∈ S n and let u ∼ v u\sim v u ∼ v with u ∉ S n u\notin S_n u ∈ / S n . Since the parent of v v v belongs to S n S_n S n , the vertex u u u is a child of v v v , and λ ( u ) = λ ( v ) + 1 { v g o o d } > n \lambda(u)=\lambda(v)+1_{\{v\ \mathrm{good}\}}>n λ ( u ) = λ ( v ) + 1 { v good } > n forces v v v to be good. It follows that the left side of (16) vanishes if v v v is bad, and is at most R v 2 ≤ M 2 R_v^2\leq M^2 R v 2 ≤ M 2 if v v v is good. If u ∉ S n u\notin S_n u ∈ / S n , then the children of u u u satisfy λ ≥ λ ( u ) > n \lambda\geq\lambda(u)>n λ ≥ λ ( u ) > n , and the only possible neighbor of u u u in S n S_n S n is its parent.
The boundary operator B = B S n B=B_{S_n} B = B S n has matrix entries B v u = h v u B_{vu}=h_{vu} B v u = h v u for v ∈ S n v\in S_n v ∈ S n and u ∉ S n u\notin S_n u ∈ / S n , and all its other entries vanish. Since each column of B B B has at most one nonzero entry, the matrix B B ∗ BB^* B B ∗ , which acts on ℓ 2 ( S n ) \ell^2(S_n) ℓ 2 ( S n ) and has entries ∑ u ∉ S n h v u h v u ‾ \sum_{u\notin S_n}h_{vu}\overline{h_{vu}} ∑ u ∈ / S n h v u h v u , is diagonal. Its diagonal entries are the left side of (16) , which gives ∥ B ∥ 2 = ∥ B B ∗ ∥ ≤ M 2 \|B\|^2=\|BB^*\|\leq M^2 ∥ B ∥ 2 = ∥ B B ∗ ∥ ≤ M 2 . □
Dirichlet truncations
Lemma 3.9. Let T T T and T 1 , T 2 , … T_1,T_2,\ldots T 1 , T 2 , … be self-adjoint operators on a Hilbert space h \mathfrak{h} h . Let D D D be a core for T T T with D ⊂ Dom ( T n ) D\subset\operatorname{Dom}(T_n) D ⊂ Dom ( T n ) for every n n n , and suppose that T n φ → T φ T_n\varphi\to T\varphi T n φ → T φ for every φ ∈ D \varphi\in D φ ∈ D . Then
( T n − z ) − 1 f → ( T − z ) − 1 f for every f ∈ h and z ∈ C ∖ R . Further, for all f , g ∈ h , we have ⟨ f , ( T n − z ) − 1 g ⟩ → ⟨ f , ( T − z ) − 1 g ⟩ uniformly for z in compact subsets of C ∖ R . (T_n-z)^{-1}f \to(T-z)^{-1}f \text{ for every } f\in\mathfrak{h}\text{ and }z\in\mathbb{C}\setminus\mathbb{R}. \text{ Further, for all }f,g\in\mathfrak{h},\text{ we have}\\
\langle f,(T_n-z)^{-1}g\rangle\to\langle f,(T-z)^{-1}g\rangle\text{ uniformly for }z\text{ in compact subsets of }\mathbb{C}\setminus\mathbb{R}. ( T n − z ) − 1 f → ( T − z ) − 1 f for every f ∈ h and z ∈ C ∖ R . Further, for all f , g ∈ h , we have ⟨ f , ( T n − z ) − 1 g ⟩ → ⟨ f , ( T − z ) − 1 g ⟩ uniformly for z in compact subsets of C ∖ R .
Proof. The first assertion is a variant of [113, Theorem VIII.25(a)], where D \mathcal{D} D is assumed to be a common core for T T T and all T n T_n T n ; the following argument does not use this assumption. Fix z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R . For φ ∈ D \varphi\in\mathcal{D} φ ∈ D and f = ( T − z ) φ f=(T-z)\varphi f = ( T − z ) φ , we have
( T n − z ) − 1 f − ( T − z ) − 1 f = ( T n − z ) − 1 ( ( T − z ) φ − ( T n − z ) φ ) = ( T n − z ) − 1 ( T − T n ) φ , (T_n-z)^{-1}f-(T-z)^{-1}f=(T_n-z)^{-1}\bigl((T-z)\varphi-(T_n-z)\varphi\bigr)=(T_n-z)^{-1}(T-T_n)\varphi, ( T n − z ) − 1 f − ( T − z ) − 1 f = ( T n − z ) − 1 ( ( T − z ) φ − ( T n − z ) φ ) = ( T n − z ) − 1 ( T − T n ) φ ,
whose norm is at most ∣ Im z ∣ − 1 ∥ ( T − T n ) φ ∥ → 0 |\operatorname{Im}z|^{-1}\lVert(T-T_n)\varphi\rVert\to0 ∣ Im z ∣ − 1 ∥( T − T n ) φ ∥ → 0 . The set ( T − z ) D (T-z)\mathcal{D} ( T − z ) D is dense in h \mathfrak{h} h : every f ∈ h f\in\mathfrak{h} f ∈ h equals ( T − z ) ψ (T-z)\psi ( T − z ) ψ for some ψ ∈ Dom ( T ) \psi\in\operatorname{Dom}(T) ψ ∈ Dom ( T ) , and if φ k ∈ D \varphi_k\in\mathcal{D} φ k ∈ D converge to ψ \psi ψ in the graph norm of T T T , then ( T − z ) φ k → f (T-z)\varphi_k\to f ( T − z ) φ k → f . Since all resolvents have norm at most ∣ Im z ∣ − 1 |\operatorname{Im}z|^{-1} ∣ Im z ∣ − 1 , the convergence extends from this dense set to all f f f .
For the second assertion, let K ⊂ C ∖ R K\subset\mathbb{C}\setminus\mathbb{R} K ⊂ C ∖ R be compact and d = dist ( K , R ) > 0 d=\operatorname{dist}(K,\mathbb{R})>0 d = dist ( K , R ) > 0 . For every self-adjoint S S S and z , z ′ ∈ K z,z'\in K z , z ′ ∈ K , the resolvent identity ( S − z ) − 1 − ( S − z ′ ) − 1 = ( z − z ′ ) ( S − z ) − 1 ( S − z ′ ) − 1 (S-z)^{-1}-(S-z')^{-1}=(z-z')(S-z)^{-1}(S-z')^{-1} ( S − z ) − 1 − ( S − z ′ ) − 1 = ( z − z ′ ) ( S − z ) − 1 ( S − z ′ ) − 1 gives
∣ ⟨ f , ( S − z ) − 1 g ⟩ − ⟨ f , ( S − z ′ ) − 1 g ⟩ ∣ ≤ ∣ z − z ′ ∣ ∥ f ∥ ∥ g ∥ d 2 . \left|\langle f,(S-z)^{-1}g\rangle-\langle f,(S-z')^{-1}g\rangle\right|\leq\frac{|z-z'|\lVert f\rVert\lVert g\rVert}{d^2}. ⟨ f , ( S − z ) − 1 g ⟩ − ⟨ f , ( S − z ′ ) − 1 g ⟩ ≤ d 2 ∣ z − z ′ ∣ ∥ f ∥ ∥ g ∥ .
Then the functions z ↦ ⟨ f , ( T n − z ) − 1 g ⟩ z\mapsto\langle f,(T_n-z)^{-1}g\rangle z ↦ ⟨ f , ( T n − z ) − 1 g ⟩ and z ↦ ⟨ f , ( T − z ) − 1 g ⟩ z\mapsto\langle f,(T-z)^{-1}g\rangle z ↦ ⟨ f , ( T − z ) − 1 g ⟩ are Lipschitz on K K K with a common constant. Given ε > 0 \varepsilon>0 ε > 0 , we cover K K K by finitely many discs of radius ε d 2 / ( 3 ∥ f ∥ ∥ g ∥ + 1 ) \varepsilon d^2/(3\lVert f\rVert\lVert g\rVert+1) ε d 2 / ( 3 ∥ f ∥ ∥ g ∥ + 1 ) centered in K K K , and choose n 0 n_0 n 0 such that the pointwise convergence holds within ε / 3 \varepsilon/3 ε /3 at the finitely many centers for n ≥ n 0 n\geq n_0 n ≥ n 0 . Then the difference is at most ε \varepsilon ε on K K K for n ≥ n 0 n\geq n_0 n ≥ n 0 .
For a component C C C of a finite removal and a finite set Λ ⊂ C \Lambda\subset C Λ ⊂ C , the Dirichlet truncation H C [ Λ ] = P Λ H C , 0 P Λ \mathcal{H}_C^{[\Lambda]}=P_\Lambda\mathcal{H}_{C,0}P_\Lambda H C [ Λ ] = P Λ H C , 0 P Λ is the bounded self-adjoint operator on ℓ 2 ( C ) \ell^2(C) ℓ 2 ( C ) which acts on ℓ 2 ( Λ ) \ell^2(\Lambda) ℓ 2 ( Λ ) by the finite matrix ( h u v ) u , v ∈ Λ (h_{uv})_{u,v\in\Lambda} ( h uv ) u , v ∈ Λ and vanishes on ℓ 2 ( C ∖ Λ ) \ell^2(C\setminus\Lambda) ℓ 2 ( C ∖ Λ ) . When C = V C=V C = V , we write H [ Λ ] \mathcal{H}^{[\Lambda]} H [ Λ ] .
The self-adjoint operators
For a finite removal ( F , D ) (F,D) ( F , D ) , define the bounded self-adjoint operator of finite rank
W F , D = ∑ { u , v } d e l e t e d ( h u v ∣ δ u ⟩ ⟨ δ v ∣ + h v u ∣ δ v ⟩ ⟨ δ u ∣ ) W_{F,D}=\sum_{\{u,v\}\ \mathrm{deleted}}\left(h_{uv}\lvert\delta_u\rangle\langle\delta_v\rvert+h_{vu}\lvert\delta_v\rangle\langle\delta_u\rvert\right) W F , D = { u , v } deleted ∑ ( h uv ∣ δ u ⟩ ⟨ δ v ∣ + h v u ∣ δ v ⟩ ⟨ δ u ∣ )
on ℓ 2 ( V ) \ell^2(V) ℓ 2 ( V ) , where the sum runs over the deleted edges of ( F , D ) (F,D) ( F , D ) , and ∣ f ⟩ ⟨ g ∣ \lvert f\rangle\langle g\rvert ∣ f ⟩ ⟨ g ∣ denotes the operator φ ↦ ⟨ g , φ ⟩ f \varphi\mapsto\langle g,\varphi\rangle f φ ↦ ⟨ g , φ ⟩ f .
Theorem 3.10. Let p > 0 p>0 p > 0 , and let Ω s a \Omega_{\mathrm{sa}} Ω sa be the event of Lemma 3.8; we have P ( Ω s a ) = 1 \mathbb{P}(\Omega_{\mathrm{sa}})=1 P ( Ω sa ) = 1 , and P ∞ ( Ω s a ) = 1 \mathbb{P}^{\infty}(\Omega_{\mathrm{sa}})=1 P ∞ ( Ω sa ) = 1 when p > 1 p>1 p > 1 . On Ω s a \Omega_{\mathrm{sa}} Ω sa , the following statements hold simultaneously for every finite removal ( F , D ) (F,D) ( F , D ) of T T T and every component C C C of ( F , D ) (F,D) ( F , D ) .
(i) The operator H C , 0 \mathcal{H}_{C,0} H C , 0 is essentially self-adjoint on c c ( C ) c_c(\mathbb{C}) c c ( C ) . We write H C \mathcal{H}_C H C for its closure, H = H V \mathcal{H}=\mathcal{H}_V H = H V for the operator of the full tree, and H v = H V ( T v ) \mathcal{H}_v=\mathcal{H}_{V(T_v)} H v = H V ( T v ) for the operator of the forward subtree at v ∈ V v\in V v ∈ V .
(ii) The domain of H C \mathcal{H}_C H C is the maximal domain:
Dom ( H C ) = { f ∈ ℓ 2 ( C ) : H C , f o r m f ∈ ℓ 2 ( C ) } , \operatorname{Dom}(\mathcal{H}_C)=\{f\in\ell^2(C):\mathcal{H}_{C,\mathrm{form}}f\in\ell^2(C)\}, Dom ( H C ) = { f ∈ ℓ 2 ( C ) : H C , form f ∈ ℓ 2 ( C )} ,
and H C f = H C , f o r m f \mathcal{H}_Cf=\mathcal{H}_{C,\mathrm{form}}f H C f = H C , form f for f ∈ Dom ( H C ) f\in\operatorname{Dom}(\mathcal{H}_C) f ∈ Dom ( H C ) .
(iii) Let Λ 1 ⊂ Λ 2 ⊂ ⋯ \Lambda_1\subset\Lambda_2\subset\cdots Λ 1 ⊂ Λ 2 ⊂ ⋯ be finite sets with ⋃ n Λ n = C \bigcup_n\Lambda_n=C ⋃ n Λ n = C . Then ( H C [ Λ n ] − z ) − 1 f → ( H C − z ) − 1 f (\mathcal{H}_C^{[\Lambda_n]}-z)^{-1}f\to(\mathcal{H}_C-z)^{-1}f ( H C [ Λ n ] − z ) − 1 f → ( H C − z ) − 1 f for every f ∈ ℓ 2 ( C ) f\in\ell^2(C) f ∈ ℓ 2 ( C ) and z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R , and for all u , v ∈ C u,v\in C u , v ∈ C ,
⟨ δ u , ( H C [ Λ n ] − z ) − 1 δ v ⟩ → ⟨ δ u , ( H C − z ) − 1 δ v ⟩ \langle\delta_u,(\mathcal{H}_C^{[\Lambda_n]}-z)^{-1}\delta_v\rangle\to\langle\delta_u,(\mathcal{H}_C-z)^{-1}\delta_v\rangle ⟨ δ u , ( H C [ Λ n ] − z ) − 1 δ v ⟩ → ⟨ δ u , ( H C − z ) − 1 δ v ⟩
uniformly for z z z in compact subsets of C ∖ R \mathbb{C}\setminus\mathbb{R} C ∖ R . This applies to the balls Λ n = { v ∈ C : dist C ( v , v 0 ) ≤ n } \Lambda_n=\{v\in C:\operatorname{dist}_C(v,v_0)\le n\} Λ n = { v ∈ C : dist C ( v , v 0 ) ≤ n } around any v 0 ∈ C v_0\in C v 0 ∈ C , and for C = V C=V C = V to the balls { v ∈ V : ∣ v ∣ ≤ n } \{v\in V:|v|\le n\} { v ∈ V : ∣ v ∣ ≤ n } .
(iv) Let C 1 , … , C m C_1,\ldots,C_m C 1 , … , C m be the components of ( F , D ) (F,D) ( F , D ) . Then
H − W F , D = H C 1 ⊕ ⋯ ⊕ H C m ⊕ 0 F \mathcal{H}-W_{F,D}=\mathcal{H}_{C_1}\oplus\cdots\oplus\mathcal{H}_{C_m}\oplus0_F H − W F , D = H C 1 ⊕ ⋯ ⊕ H C m ⊕ 0 F
with respect to ℓ 2 ( V ) = ℓ 2 ( C 1 ) ⊕ ⋯ ⊕ ℓ 2 ( C m ) ⊕ ℓ 2 ( F ) \ell^2(V)=\ell^2(C_1)\oplus\cdots\oplus\ell^2(C_m)\oplus\ell^2(F) ℓ 2 ( V ) = ℓ 2 ( C 1 ) ⊕ ⋯ ⊕ ℓ 2 ( C m ) ⊕ ℓ 2 ( F ) , where 0 F 0_F 0 F is the zero operator on ℓ 2 ( F ) \ell^2(F) ℓ 2 ( F ) . In particular Dom ( H ) = Dom ( H C 1 ) ⊕ ⋯ ⊕ Dom ( H C m ) ⊕ ℓ 2 ( F ) \operatorname{Dom}(\mathcal{H})=\operatorname{Dom}(\mathcal{H}_{C_1})\oplus\cdots\oplus\operatorname{Dom}(\mathcal{H}_{C_m})\oplus\ell^2(F) Dom ( H ) = Dom ( H C 1 ) ⊕ ⋯ ⊕ Dom ( H C m ) ⊕ ℓ 2 ( F ) .
We write H F , D = H C 1 ⊕ ⋯ ⊕ H C m \mathcal{H}^{F,D}=\mathcal{H}_{C_1}\oplus\cdots\oplus\mathcal{H}_{C_m} H F , D = H C 1 ⊕ ⋯ ⊕ H C m for the reduced operator on ℓ 2 ( V ∖ F ) \ell^2(V\setminus F) ℓ 2 ( V ∖ F ) , and H ( y ) = H { y } , ∅ \mathcal{H}^{(y)}=\mathcal{H}_{\{y\},\varnothing} H ( y ) = H { y } , ∅ for the operator with the vertex y y y deleted.
Proof. We fix a realization in Ω s a \Omega_{\mathrm{sa}} Ω sa ; all statements below are deterministic consequences of the conclusions of Lemma 3.8 , which do not depend on ( F , D ) (F,D) ( F , D ) .
Step 1. The full tree. By Lemma 3.8 , the network ( V , h ) (V,h) ( V , h ) satisfies the hypotheses of Lemma 3.7 , which implies that H 0 \mathcal{H}_0 H 0 is essentially self-adjoint. Lemma 3.6 identifies the domain of its closure H \mathcal{H} H with the maximal domain.
Step 2. Finite removals. Let ( F , D ) (F,D) ( F , D ) be a finite removal and W = W F , D W=W_{F,D} W = W F , D . Since W W W is bounded and self-adjoint, the Kato–Rellich theorem [112, Theorem X.12] (with relative bound zero) shows that H − W \mathcal{H}-W H − W is self-adjoint on Dom ( H ) \operatorname{Dom}(\mathcal{H}) Dom ( H ) and essentially self-adjoint on every core of H \mathcal{H} H , in particular on c c ( V ) c_c(V) c c ( V ) . On c c ( V ) c_c(V) c c ( V ) , the operator H 0 − W \mathcal{H}_0-W H 0 − W is the weighted adjacency operator of the graph on V V V with the retained edges. By Lemma 3.5 (i), this graph has the components C 1 , … , C m C_1,\ldots,C_m C 1 , … , C m and the isolated vertices of F F F . Hence
( H 0 − W ) ∣ c c ( V ) = H C 1 , 0 ⊕ ⋯ ⊕ H C m , 0 ⊕ 0 (18) (\mathcal{H}_0-W)|_{c_c(V)}=\mathcal{H}_{C_1,0}\oplus\cdots\oplus\mathcal{H}_{C_m,0}\oplus0 \tag*{(18)} ( H 0 − W ) ∣ c c ( V ) = H C 1 , 0 ⊕ ⋯ ⊕ H C m , 0 ⊕ 0 (18)
with respect to c c ( V ) = c c ( C 1 ) ⊕ ⋯ ⊕ c c ( C m ) ⊕ c c ( F ) c_c(V)=c_c(C_1)\oplus\cdots\oplus c_c(C_m)\oplus c_c(F) c c ( V ) = c c ( C 1 ) ⊕ ⋯ ⊕ c c ( C m ) ⊕ c c ( F ) .
The adjoint of a finite orthogonal sum of densely defined operators is the orthogonal sum of their adjoints. Indeed, a vector ( φ j ) j (\varphi_j)_j ( φ j ) j is in the domain of the adjoint of ⨁ j T j \bigoplus_jT_j ⨁ j T j exactly when ( f j ) j ↦ ∑ j ⟨ φ j , T j f j ⟩ (f_j)_j \mapsto\sum_j \langle\varphi_j,T_jf_j\rangle ( f j ) j ↦ ∑ j ⟨ φ j , T j f j ⟩ is bounded, and taking f f f supported in a single summand shows that this holds exactly when φ j ∈ Dom ( T j ∗ ) \varphi_j \in\operatorname{Dom}(T_j^*) φ j ∈ Dom ( T j ∗ ) for every j j j . For z = ± i z=\pm i z = ± i , it follows that the kernel of T ∗ − z T^*-z T ∗ − z , where T T T is the operator in (18) , is the orthogonal sum of the kernels of H C j , 0 ∗ − z \mathcal{H}_{C_j,0}^*-z H C j , 0 ∗ − z and of the kernel of − z -z − z on ℓ 2 ( F ) \ell^2(F) ℓ 2 ( F ) , which is trivial. Since T T T is essentially self-adjoint, the kernel of T ∗ − z T^*-z T ∗ − z is trivial, and so are the kernels of H C j , 0 ∗ − z \mathcal{H}_{C_j,0}^*-z H C j , 0 ∗ − z . Then [113, Theorem VIII.3 and its corollary] shows that every H C j , 0 \mathcal{H}_{C_j,0} H C j , 0 is essentially self-adjoint. The closure of a finite orthogonal sum is the orthogonal sum of the closures, since the graph of the sum is the orthogonal sum of the graphs. Taking closures in (18) proves (iv), and (i) holds for every component. Statement (ii) follows from Lemma 3.6 applied to the network ( C , h ) (C,h) ( C , h ) .
Step 3. Dirichlet truncations. Let Λ n \Lambda_n Λ n be as in (iii) and f ∈ c c ( C ) f \in c_c(C) f ∈ c c ( C ) . The vector H C , 0 f \mathcal{H}_{C,0}f H C , 0 f is supported on the union of supp f \operatorname{supp} f supp f and its neighbors in C C C , which is finite. When Λ n \Lambda_n Λ n contains this union, we have H C [ Λ n ] f = P Λ n H C , 0 f = H C , 0 f \mathcal{H}_C^{[\Lambda_n]}f=P_{\Lambda_n}\mathcal{H}_{C,0}f=\mathcal{H}_{C,0}f H C [ Λ n ] f = P Λ n H C , 0 f = H C , 0 f . Since c c ( C ) c_c(C) c c ( C ) is a core for H C \mathcal{H}_C H C , Lemma 3.9 implies (iii). Balls in the locally finite graph C C C are finite and exhaust C C C .
Remark 3.11. Statement (ii) is used in the following form. Let C C C be as in Theorem 3.10, let E ∈ R E \in\mathbb{R} E ∈ R , and let g ∈ ℓ 2 ( C ) g \in\ell^2(C) g ∈ ℓ 2 ( C ) . If F ∈ ℓ 2 ( C ) F \in\ell^2(C) F ∈ ℓ 2 ( C ) satisfies ( H C , f o r m F ) ( v ) − E F ( v ) = g ( v ) (\mathcal{H}_{C,\mathrm{form}}F)(v)-EF(v)=g(v) ( H C , form F ) ( v ) − E F ( v ) = g ( v ) for every v ∈ C v \in C v ∈ C , then H C , f o r m F = E F + g ∈ ℓ 2 ( C ) \mathcal{H}_{C,\mathrm{form}}F=EF+g \in\ell^2(C) H C , form F = E F + g ∈ ℓ 2 ( C ) , and (ii) implies that F ∈ Dom ( H C ) F \in\operatorname{Dom}(\mathcal{H}_C) F ∈ Dom ( H C ) and ( H C − E ) F = g (\mathcal{H}_C-E)F=g ( H C − E ) F = g . In particular, a vector F ∈ ℓ 2 ( V ) F \in\ell^2(V) F ∈ ℓ 2 ( V ) satisfying ( H f o r m − E ) F = δ o (\mathcal{H}_{\mathrm{form}}-E)F=\delta_o ( H form − E ) F = δ o pointwise lies in Dom ( H ) \operatorname{Dom}(\mathcal{H}) Dom ( H ) and satisfies ( H − E ) F = δ o (\mathcal{H}-E)F=\delta_o ( H − E ) F = δ o ; in particular, δ o ∈ Ran ( H − E ) \delta_o \in\operatorname{Ran}(\mathcal{H}-E) δ o ∈ Ran ( H − E ) .
Remark 3.12. The exhaustion of Lemma 3.8 also applies directly to a component C C C of a finite removal: the sets S n ∩ C S_n \cap C S n ∩ C exhaust C C C , each vertex of C ∖ S n C \setminus S_n C ∖ S n has at most one neighbor in S n ∩ C S_n \cap C S n ∩ C , and the boundary sums in C C C are bounded by the left side of (16) . Then Lemma 3.7 yields (i) without the perturbation argument. Statement (iv), which identifies the domains, is used repeatedly below.
The following consequence of (iv) decouples a component along finitely many further deleted edges.
Lemma 3.13. On Ω s a \Omega_{\mathrm{sa}} Ω sa , let C C C be a component of a finite removal ( F , D ) (F,D) ( F , D ) , and let D ′ D' D ′ be a finite set of edges of T \mathcal{T} T with both endpoints in C C C . Let C 1 ′ , … , C k ′ C'_1,\ldots,C'_k C 1 ′ , … , C k ′ be the components of ( F , D ∪ D ′ ) (F,D \cup D') ( F , D ∪ D ′ ) contained in C C C . Then
H C = ( H C 1 ′ ⊕ ⋯ ⊕ H C k ′ ) + W D ′ , Dom ( H C ) = Dom ( H C 1 ′ ) ⊕ ⋯ ⊕ Dom ( H C k ′ ) , \mathcal{H}_C=(\mathcal{H}_{C'_1}\oplus\cdots\oplus\mathcal{H}_{C'_k})+W_{D'},\qquad\operatorname{Dom}(\mathcal{H}_C)=\operatorname{Dom}(\mathcal{H}_{C'_1})\oplus\cdots\oplus\operatorname{Dom}(\mathcal{H}_{C'_k}), H C = ( H C 1 ′ ⊕ ⋯ ⊕ H C k ′ ) + W D ′ , Dom ( H C ) = Dom ( H C 1 ′ ) ⊕ ⋯ ⊕ Dom ( H C k ′ ) , where W D ′ = ∑ { u , v } ∈ D ′ ( h u v ∣ δ u ⟩ ⟨ δ v ∣ + h v u ∣ δ v ⟩ ⟨ δ u ∣ ) W_{D'}=\sum_{\{u,v\}\in D'}(h_{uv}\lvert\delta_u\rangle\langle\delta_v\rvert+h_{vu}\lvert\delta_v\rangle\langle\delta_u\rvert) W D ′ = ∑ { u , v } ∈ D ′ ( h uv ∣ δ u ⟩ ⟨ δ v ∣ + h v u ∣ δ v ⟩ ⟨ δ u ∣) acts on ℓ 2 ( C ) \ell^2(C) ℓ 2 ( C ) . In particular, if y ∈ C y \in C y ∈ C has the neighbors x 1 , … , x d x_1,\ldots,x_d x 1 , … , x d in C C C , then with respect to ℓ 2 ( C ) = C δ y ⊕ ℓ 2 ( C x 1 → y ) ⊕ ⋯ ⊕ ℓ 2 ( C x d → y ) \ell^2(C)=\mathbb{C}\delta_y\oplus\ell^2(C_{x_1\to y})\oplus\cdots\oplus\ell^2(C_{x_d\to y}) ℓ 2 ( C ) = C δ y ⊕ ℓ 2 ( C x 1 → y ) ⊕ ⋯ ⊕ ℓ 2 ( C x d → y ) ,
H C = ( 0 ⊕ H C x 1 → y ⊕ ⋯ ⊕ H C x d → y ) + ∑ i = 1 d ( h y x i ∣ δ y ⟩ ⟨ δ x i ∣ + h x i y ∣ δ x i ⟩ ⟨ δ y ∣ ) . (19) \mathcal{H}_C=\left(0\oplus\mathcal{H}_{C_{x_1\to y}}\oplus\cdots\oplus\mathcal{H}_{C_{x_d\to y}}\right)+\sum_{i=1}^{d}\left(h_{yx_i}\lvert\delta_y\rangle\langle\delta_{x_i}\rvert+h_{x_i y}\lvert\delta_{x_i}\rangle\langle\delta_y\rvert\right). \tag*{(19)} H C = ( 0 ⊕ H C x 1 → y ⊕ ⋯ ⊕ H C x d → y ) + i = 1 ∑ d ( h y x i ∣ δ y ⟩ ⟨ δ x i ∣ + h x i y ∣ δ x i ⟩ ⟨ δ y ∣ ) . (19)
Proof. Since the edges of D ′ D' D ′ are retained by ( F , D ) (F,D) ( F , D ) , we have W F , D ∪ D ′ = W F , D + W D ′ W_{F,D\cup D'}=W_{F,D}+W_{D'} W F , D ∪ D ′ = W F , D + W D ′ . The components of ( F , D ∪ D ′ ) (F,D\cup D') ( F , D ∪ D ′ ) are the components of ( F , D ) (F,D) ( F , D ) other than C C C , together with C 1 ′ , … , C k ′ C'_1,\ldots,C'_k C 1 ′ , … , C k ′ . Applying Theorem 3.10(iv) to both finite removals, we find that H − W F , D − W D ′ \mathcal{H}-W_{F,D}-W_{D'} H − W F , D − W D ′ equals the orthogonal sum of the operators H C ′ ′ \mathcal{H}_{C''} H C ′′ over the components C ′ ′ ≠ C C''\ne C C ′′ = C of ( F , D ) (F,D) ( F , D ) , of H C − W D ′ \mathcal{H}_C-W_{D'} H C − W D ′ , and of 0 F 0_F 0 F , and that it also equals the same sum with H C − W D ′ \mathcal{H}_C-W_{D'} H C − W D ′ replaced by H C 1 ′ ⊕ ⋯ ⊕ H C k ′ \mathcal{H}_{C'_1}\oplus\cdots\oplus\mathcal{H}_{C'_k} H C 1 ′ ⊕ ⋯ ⊕ H C k ′ . Both decompositions are orthogonal with respect to the same splitting of ℓ 2 ( V ) \ell^2(V) ℓ 2 ( V ) , and their parts acting in ℓ 2 ( C ) \ell^2(C) ℓ 2 ( C ) agree, including their domains. Since W D ′ W_{D'} W D ′ is bounded, H C − W D ′ \mathcal{H}_C-W_{D'} H C − W D ′ has the domain Dom ( H C ) \operatorname{Dom}(\mathcal{H}_C) Dom ( H C ) .
For (3.9), take D ′ D' D ′ to be the set of edges of C C C at y y y . By Lemma 3.5(ii), the components of ( F , D ∪ D ′ ) (F,D\cup D') ( F , D ∪ D ′ ) inside C C C are { y } \{y\} { y } and the branches C x i → y C_{x_i\to y} C x i → y , and the operator of the isolated vertex y y y is zero.
Resolvents and the cavity recursion
Definitions
We begin with the elementary properties of resolvent entries of a self-adjoint operator.
Lemma 3.14. Let T T T be a self-adjoint operator on a Hilbert space h \mathfrak{h} h , let f , g ∈ h f,g\in\mathfrak{h} f , g ∈ h , and let z = E + i η ∈ C ∖ R z=E+i\eta\in\mathbb{C}\setminus\mathbb{R} z = E + i η ∈ C ∖ R .
(i) The function z ↦ ⟨ f , ( T − z ) − 1 g ⟩ z\mapsto\langle f,(T-z)^{-1}g\rangle z ↦ ⟨ f , ( T − z ) − 1 g ⟩ is analytic on C ∖ R \mathbb{C}\setminus\mathbb{R} C ∖ R , bounded by ∥ f ∥ ∥ g ∥ / ∣ η ∣ \lVert f\rVert\lVert g\rVert/|\eta| ∥ f ∥ ∥ g ∥ /∣ η ∣ , and satisfies ⟨ f , ( T − z ) − 1 g ⟩ = ⟨ g , ( T − z ‾ ) − 1 f ⟩ ‾ \langle f,(T-z)^{-1}g\rangle=\overline{\langle g,(T-\overline{z})^{-1}f\rangle} ⟨ f , ( T − z ) − 1 g ⟩ = ⟨ g , ( T − z ) − 1 f ⟩ .
(ii) The function m ( z ) = ⟨ f , ( T − z ) − 1 f ⟩ m(z)=\langle f,(T-z)^{-1}f\rangle m ( z ) = ⟨ f , ( T − z ) − 1 f ⟩ satisfies
m ( z ) = ∫ R μ f T ( d λ ) λ − z , Im m ( z ) = η ∥ ( T − z ) − 1 f ∥ 2 = η ∫ R μ f T ( d λ ) ∣ λ − z ∣ 2 , m(z)=\int_{\mathbb{R}}\frac{\mu_f^T(\mathrm{d}\lambda)}{\lambda-z},\qquad\operatorname{Im}m(z)=\eta\lVert(T-z)^{-1}f\rVert^2=\eta\int_{\mathbb{R}}\frac{\mu_f^T(\mathrm{d}\lambda)}{|\lambda-z|^2}, m ( z ) = ∫ R λ − z μ f T ( d λ ) , Im m ( z ) = η ∥( T − z ) − 1 f ∥ 2 = η ∫ R ∣ λ − z ∣ 2 μ f T ( d λ ) ,
and m ( z ‾ ) = m ( z ) ‾ m(\overline{z})=\overline{m(z)} m ( z ) = m ( z ) . In particular, if f ≠ 0 f\ne0 f = 0 and η > 0 \eta>0 η > 0 , then Im m ( z ) > 0 \operatorname{Im}m(z)>0 Im m ( z ) > 0 .
Proof. The resolvent is an analytic operator-valued function on C ∖ R \mathbb{C}\setminus\mathbb{R} C ∖ R with norm at most ∣ η ∣ − 1 |\eta|^{-1} ∣ η ∣ − 1 , and ( ( T − z ) − 1 ) ∗ = ( T − z ‾ ) − 1 \bigl((T-z)^{-1}\bigr)^*=(T-\overline{z})^{-1} ( ( T − z ) − 1 ) ∗ = ( T − z ) − 1 ; this gives (i). The integral formula is the spectral theorem. By the resolvent identity, ( T − z ) − 1 − ( T − z ‾ ) − 1 = ( z − z ‾ ) ( T − z ‾ ) − 1 ( T − z ) − 1 (T-z)^{-1}-(T-\overline{z})^{-1}=(z-\overline{z})(T-\overline{z})^{-1}(T-z)^{-1} ( T − z ) − 1 − ( T − z ) − 1 = ( z − z ) ( T − z ) − 1 ( T − z ) − 1 , and we have
2 i Im m ( z ) = m ( z ) − m ( z ) ‾ = ⟨ f , ( ( T − z ) − 1 − ( T − z ‾ ) − 1 ) f ⟩ = 2 i η ⟨ ( T − z ) − 1 f , ( T − z ) − 1 f ⟩ . 2i\operatorname{Im}m(z)=m(z)-\overline{m(z)}=\langle f,\bigl((T-z)^{-1}-(T-\overline{z})^{-1}\bigr)f\rangle
=2i\eta\langle(T-z)^{-1}f,(T-z)^{-1}f\rangle. 2 i Im m ( z ) = m ( z ) − m ( z ) = ⟨ f , ( ( T − z ) − 1 − ( T − z ) − 1 ) f ⟩ = 2 i η ⟨( T − z ) − 1 f , ( T − z ) − 1 f ⟩ .
where we used (i) in the form m ( z ) ‾ = m ( z ‾ ) \overline{m(z)}=m(\overline{z}) m ( z ) = m ( z ) . Since the resolvent is injective, Im m ( z ) > 0 \operatorname{Im}m(z)>0 Im m ( z ) > 0 when f ≠ 0 f\ne0 f = 0 and η > 0 \eta>0 η > 0 .
Definition 3.15. On Ω s a \Omega_{\mathrm{sa}} Ω sa , let ( F , D ) (F,D) ( F , D ) be a finite removal, let C C C be one of its components, and let z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R .
(i) The resolvent entries are
G u v ( z ) = ⟨ δ u , ( H − z ) − 1 δ v ⟩ , G u v C ( z ) = ⟨ δ u , ( H C − z ) − 1 δ v ⟩ , G_{uv}(z)=\langle\delta_u,(\mathcal{H}-z)^{-1}\delta_v\rangle,\qquad G^{C}_{uv}(z)=\langle\delta_u,(\mathcal{H}_{C}-z)^{-1}\delta_v\rangle, G uv ( z ) = ⟨ δ u , ( H − z ) − 1 δ v ⟩ , G uv C ( z ) = ⟨ δ u , ( H C − z ) − 1 δ v ⟩ ,
G u v F , D ( z ) = ⟨ δ u , ( H F , D − z ) − 1 δ v ⟩ , G^{F,D}_{uv}(z)=\langle\delta_u,(\mathcal{H}^{F,D}-z)^{-1}\delta_v\rangle, G uv F , D ( z ) = ⟨ δ u , ( H F , D − z ) − 1 δ v ⟩ ,
for u , v ∈ V u,v\in V u , v ∈ V , u , v ∈ C u,v\in C u , v ∈ C , and u , v ∈ V ∖ F u,v\in V\setminus F u , v ∈ V ∖ F , respectively. We write G ( y ) G^{(y)} G ( y ) for G { y } , ∅ G^{\{y\},\varnothing} G { y } , ∅ . Since H F , D \mathcal{H}^{F,D} H F , D is the orthogonal sum of the operators of its components, G u v F , D = G u v C G^{F,D}_{uv}=G^{C}_{uv} G uv F , D = G uv C if u , v ∈ C u,v\in C u , v ∈ C , and G u v F , D = 0 G^{F,D}_{uv}=0 G uv F , D = 0 if u u u and v v v lie in different components.
(ii) For adjacent u , w ∈ C u,w\in C u , w ∈ C , the cavity resolvent is the diagonal entry at u u u of the operator of the branch C u → w C_{u\to w} C u → w :
Γ u → w C ( z ) = G u u C u → w ( z ) . \Gamma^{C}_{u\to w}(z)=G^{C_{u\to w}}_{uu}(z). Γ u → w C ( z ) = G uu C u → w ( z ) .
For C = V C=V C = V we write Γ u → w \Gamma_{u\to w} Γ u → w . The forward cavity at v ∈ V ∖ { o } v\in V\setminus\{o\} v ∈ V ∖ { o } is Γ v = Γ v → v − \Gamma_v=\Gamma_{v\to v^-} Γ v = Γ v → v − ; since V v → v − = V ( T v ) V_{v\to v^-}=V(\mathcal{T}_v) V v → v − = V ( T v ) , it equals ⟨ δ v , ( H v − z ) − 1 δ v ⟩ \langle\delta_v,(\mathcal{H}_v-z)^{-1}\delta_v\rangle ⟨ δ v , ( H v − z ) − 1 δ v ⟩ . At the root we set Γ o = G o o \Gamma_o=G_{oo} Γ o = G oo , the diagonal entry at the root of the operator of its forward subtree T o = T \mathcal{T}_o=\mathcal{T} T o = T .
(iii) The spectral measure of v ∈ V v\in V v ∈ V is μ v = μ δ v H \mu_v=\mu^{\mathcal{H}}_{\delta_v} μ v = μ δ v H , and G v v ( z ) = ∫ μ v ( d λ ) / ( λ − z ) G_{vv}(z)=\int\mu_v(\mathrm{d}\lambda)/(\lambda-z) G v v ( z ) = ∫ μ v ( d λ ) / ( λ − z ) .
By Lemma 3.14, all these entries are analytic in z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R and bounded by 1 / ∣ Im z ∣ 1/|\operatorname{Im}z| 1/∣ Im z ∣ , and the diagonal entries and cavity resolvents map H \mathbb{H} H to H \mathbb{H} H . They are also measurable functions of the marked tree. Indeed, for a fixed finite removal and fixed words u , v u,v u , v , Theorem 3.10 (iii) with Λ n = { x ∈ C : ∣ x ∣ ≤ n } \Lambda_n=\{x\in C:|x|\le n\} Λ n = { x ∈ C : ∣ x ∣ ≤ n } expresses G u v C ( z ) G^{C}_{uv}(z) G uv C ( z ) as the limit of entries of the inverses of finite matrices, which are measurable functions of ( N x ) ∣ x ∣ < n (N_x)_{|x|<n} ( N x ) ∣ x ∣ < n and ( w x ) ∣ x ∣ ≤ n (w_x)_{|x|\le n} ( w x ) ∣ x ∣ ≤ n . For a random finite removal, we decompose according to its countably many values. In particular, let Φ z \Phi_z Φ z be the measurable function of the families ( ( N u ) u ∈ U , ( w u ) u ≠ o ) ((N_u)_{u\in U},(w_u)_{u\ne o}) (( N u ) u ∈ U , ( w u ) u = o ) defined as the limit of ⟨ δ o , ( H { ∣ v ∣ ≤ n } − z ) − 1 δ o ⟩ \langle\delta_o,(\mathcal{H}^{\{|v|\le n\}}-z)^{-1}\delta_o\rangle ⟨ δ o , ( H { ∣ v ∣ ≤ n } − z ) − 1 δ o ⟩ when this limit exists, and as zero otherwise. We write Φ ( T ) \Phi(\mathcal{T}) Φ ( T ) for the function z ↦ Φ z ( T ) z\mapsto\Phi_z(\mathcal{T}) z ↦ Φ z ( T ) on C ∖ R \mathbb{C}\setminus\mathbb{R} C ∖ R . On Ω s a \Omega_{\mathrm{sa}} Ω sa we have Γ o ( z ) = Φ z ( T ) \Gamma_o(z)=\Phi_z(\mathcal{T}) Γ o ( z ) = Φ z ( T ) , and for j ≤ N o j\le N_o j ≤ N o the relabelling j u ↦ u ju\mapsto u j u ↦ u identifies the truncations of H j \mathcal{H}_j H j to balls around j j j with the truncations of the operator of θ j T \theta_j\mathcal{T} θ j T to balls around o o o . We conclude that
Γ j ( z ) = Φ z ( θ j T ) , 1 ≤ j ≤ N o . (20) \Gamma_j(z)=\Phi_z(\theta_j\mathcal{T}),\qquad1\le j\le N_o. \tag*{(20)} Γ j ( z ) = Φ z ( θ j T ) , 1 ≤ j ≤ N o . (20)
Proposition 3.16. Let p > 0 p>0 p > 0 . On Ω s a \Omega_{\mathrm{sa}} Ω sa , the following statements hold for every z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R ; statements (i) and (ii) hold for every finite removal ( F , D ) (F,D) ( F , D ) and every component C C C of ( F , D ) (F,D) ( F , D ) .
(i) For every u ∈ C u\in C u ∈ C ,
G u u C ( z ) = − ( z + ∑ x ∈ C , x ∼ u ∣ h u x ∣ 2 Γ x → u C ( z ) ) − 1 . (21) G^{C}_{uu}(z)=-\left(z+\sum_{x\in C,\,x\sim u}|h_{ux}|^2\Gamma^{C}_{x\to u}(z)\right)^{-1}. \tag*{(21)} G uu C ( z ) = − z + x ∈ C , x ∼ u ∑ ∣ h ux ∣ 2 Γ x → u C ( z ) − 1 . (21)
(ii) For all adjacent u , w ∈ C u,w\in C u , w ∈ C ,
Γ u → w C ( z ) = − ( z + ∑ x ∈ C , x ∼ u , x ≠ w ∣ h u x ∣ 2 Γ x → u C ( z ) ) − 1 . (22) \Gamma^{C}_{u\to w}(z)=-\left(z+\sum_{x\in C,\,x\sim u,\,x\ne w}|h_{ux}|^2\Gamma^{C}_{x\to u}(z)\right)^{-1}. \tag*{(22)} Γ u → w C ( z ) = − z + x ∈ C , x ∼ u , x = w ∑ ∣ h ux ∣ 2 Γ x → u C ( z ) − 1 . (22)
(iii) Let N = N o N=N_o N = N o , Y j = ∣ w j ∣ 2 Y_j=|w_j|^2 Y j = ∣ w j ∣ 2 , and Γ j ( z ) = Φ z ( θ j T ) \Gamma_j(z)=\Phi_z(\theta_j\mathcal{T}) Γ j ( z ) = Φ z ( θ j T ) for j ≥ 1 j\geq1 j ≥ 1 . Then
Γ o ( z ) = − ( z + 1 p ∑ j = 1 N Y j Γ j ( z ) ) − 1 , (23) \Gamma_o(z)=-\left(z+\frac{1}{p}\sum_{j=1}^{N}Y_j\Gamma_j(z)\right)^{-1}, \tag*{(23)} Γ o ( z ) = − ( z + p 1 j = 1 ∑ N Y j Γ j ( z ) ) − 1 , (23)
where N N N has the law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) , the Y j Y_j Y j are independent with the law Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) , the random analytic functions Γ j \Gamma_j Γ j on C ∖ R \mathbb{C}\setminus\mathbb{R} C ∖ R are independent copies of Γ o \Gamma_o Γ o , and the three families N N N , ( Y j ) j ≥ 1 (Y_j)_{j\geq1} ( Y j ) j ≥ 1 , ( Γ j ) j ≥ 1 (\Gamma_j)_{j\geq1} ( Γ j ) j ≥ 1 are independent. In particular, Γ o ( z ) = d − ( z + p − 1 ∑ j = 1 N Y j Γ j ( z ) ) − 1 \Gamma_o(z)\mathrel{\stackrel{\mathrm{d}}{=}}-\left(z+p^{-1}\sum_{j=1}^{N}Y_j\Gamma_j(z)\right)^{-1} Γ o ( z ) = d − ( z + p − 1 ∑ j = 1 N Y j Γ j ( z ) ) − 1 . For j ≤ N j\leq N j ≤ N , the function Γ j \Gamma_j Γ j is the forward cavity at the child j j j of the root, and Γ o ( z ) \Gamma_o(z) Γ o ( z ) is the limit of the root entries of the resolvents of the Dirichlet truncations to the balls { v ∈ V : ∣ v ∣ ≤ n } \{v\in V:|v|\leq n\} { v ∈ V : ∣ v ∣ ≤ n } .
In (i) and (ii), the imaginary part of the expression in parentheses has the sign of Im z \operatorname{Im}z Im z and absolute value at least ∣ Im z ∣ |\operatorname{Im}z| ∣ Im z ∣ . For the full tree, these formulas read
Γ v ( z ) = − ( z + ∑ j = 1 N v ∣ h v , v j ∣ 2 Γ v j ( z ) ) − 1 , v ∈ V , (24) \Gamma_v(z)=-\left(z+\sum_{j=1}^{N_v}|h_{v,v_j}|^2\Gamma_{v_j}(z)\right)^{-1},\qquad v\in V, \tag*{(24)} Γ v ( z ) = − ( z + j = 1 ∑ N v ∣ h v , v j ∣ 2 Γ v j ( z ) ) − 1 , v ∈ V , (24)
G v v ( z ) = − ( z + ∣ h v v − ∣ 2 Γ v − → v ( z ) + ∑ j = 1 N v ∣ h v , v j ∣ 2 Γ v j ( z ) ) − 1 , v ∈ V ∖ { o } , (25) G_{vv}(z)=-\left(z+|h_{vv_-}|^2\Gamma_{v_-\to v}(z)+\sum_{j=1}^{N_v}|h_{v,v_j}|^2\Gamma_{v_j}(z)\right)^{-1},\qquad v\in V\setminus\{o\}, \tag*{(25)} G v v ( z ) = − ( z + ∣ h v v − ∣ 2 Γ v − → v ( z ) + j = 1 ∑ N v ∣ h v , v j ∣ 2 Γ v j ( z ) ) − 1 , v ∈ V ∖ { o } , (25)
Γ u → v ( z ) = − ( z + 1 { u ≠ o } ∣ h u u − ∣ 2 Γ u − → u ( z ) + ∑ j ≤ N u u j ≠ v ∣ h u , u j ∣ 2 Γ u j ( z ) ) − 1 , (26) \Gamma_{u\to v}(z)=-\left(z+\mathbf{1}_{\{u\ne o\}}|h_{uu_-}|^2\Gamma_{u_-\to u}(z)+\sum_{\substack{j\leq N_u\\u_j\ne v}}|h_{u,u_j}|^2\Gamma_{u_j}(z)\right)^{-1}, \tag*{(26)} Γ u → v ( z ) = − z + 1 { u = o } ∣ h u u − ∣ 2 Γ u − → u ( z ) + j ≤ N u u j = v ∑ ∣ h u , u j ∣ 2 Γ u j ( z ) − 1 , (26)
where in (3.16), u ∈ V u\in V u ∈ V and v v v is a child of u u u .
The cavity Γ u → v \Gamma_{u\to v} Γ u → v in (3.16) is a rootward cavity: it is the diagonal entry at u u u of the operator on V ∖ V ( T v ) V\setminus V(\mathcal{T}_v) V ∖ V ( T v ) . Formula (3.16) expresses it through the rootward cavity one generation closer to the root and through forward cavities; along a path from the root, the rootward cavities can be computed recursively from the root downwards. Note that (3.14) with v = o v=o v = o is the formula for Γ o = G o o \Gamma_o=G_{oo} Γ o = G oo .
Proof. Fix C C C , u ∈ C u\in C u ∈ C and z z z , and let x 1 , … , x d x_1,\ldots,x_d x 1 , … , x d be the neighbors of u u u in C C C . By Lemma 3.13, with respect to ℓ 2 ( C ) = C δ u ⊕ k \ell^2(C)=\mathbb{C}\delta_u\oplus\mathfrak{k} ℓ 2 ( C ) = C δ u ⊕ k , where k = ⨁ i ℓ 2 ( C x i → u ) \mathfrak{k}=\bigoplus_i\ell^2(C_{x_i\to u}) k = ⨁ i ℓ 2 ( C x i → u ) , we have Dom ( H C ) = C δ u ⊕ Dom ( D ) \operatorname{Dom}(H_C)=\mathbb{C}\delta_u\oplus\operatorname{Dom}(D) Dom ( H C ) = C δ u ⊕ Dom ( D ) with D = ⨁ i H C x i → u D=\bigoplus_i H_{C_{x_i\to u}} D = ⨁ i H C x i → u , and
H C ( α δ u + g ) = ⟨ a , g ⟩ δ u + α a + D g , α ∈ C , g ∈ Dom ( D ) , H_C(\alpha\delta_u+g)=\langle a,g\rangle\delta_u+\alpha a+Dg,\qquad\alpha\in\mathbb{C},\ g\in\operatorname{Dom}(D), H C ( α δ u + g ) = ⟨ a , g ⟩ δ u + α a + D g , α ∈ C , g ∈ Dom ( D ) ,
where a = ∑ i h x i u δ x i ∈ k a=\sum_i h_{x_i u}\delta_{x_i}\in\mathfrak{k} a = ∑ i h x i u δ x i ∈ k . Here we used that ⟨ a , g ⟩ = ∑ i h x i u ‾ g ( x i ) = ∑ i h u x i g ( x i ) \langle a,g\rangle=\sum_i\overline{h_{x_i u}}g(x_i)=\sum_i h_{u x_i}g(x_i) ⟨ a , g ⟩ = ∑ i h x i u g ( x i ) = ∑ i h u x i g ( x i ) . The operator D D D is self-adjoint on k \mathfrak{k} k , and since it is block diagonal,
⟨ a , ( D − z ) − 1 a ⟩ = ∑ i = 1 d ∣ h u x i ∣ 2 ⟨ δ x i , ( H C x i → u − z ) − 1 δ x i ⟩ = ∑ i = 1 d ∣ h u x i ∣ 2 Γ x i → u C ( z ) . \left\langle a,(D-z)^{-1}a\right\rangle=\sum_{i=1}^{d}|h_{u x_i}|^2\left\langle\delta_{x_i},(H_{C_{x_i\to u}}-z)^{-1}\delta_{x_i}\right\rangle=\sum_{i=1}^{d}|h_{u x_i}|^2\Gamma^{C}_{x_i\to u}(z). ⟨ a , ( D − z ) − 1 a ⟩ = i = 1 ∑ d ∣ h u x i ∣ 2 ⟨ δ x i , ( H C x i → u − z ) − 1 δ x i ⟩ = i = 1 ∑ d ∣ h u x i ∣ 2 Γ x i → u C ( z ) .
By Lemma 3.14(ii), the imaginary part of this quantity has the sign of Im z \operatorname{Im} z Im z (or vanishes), and s = − z − ⟨ a , ( D − z ) − 1 a ⟩ s=-z-\langle a,(\mathcal{D}-z)^{-1}a\rangle s = − z − ⟨ a , ( D − z ) − 1 a ⟩ satisfies ∣ Im s ∣ ≥ ∣ Im z ∣ > 0 |\operatorname{Im}s|\ge|\operatorname{Im}z|>0 ∣ Im s ∣ ≥ ∣ Im z ∣ > 0 . Set α = s − 1 \alpha=s^{-1} α = s − 1 and g = − α ( D − z ) − 1 a ∈ Dom ( D ) g=-\alpha(\mathcal{D}-z)^{-1}a\in\operatorname{Dom}(\mathcal{D}) g = − α ( D − z ) − 1 a ∈ Dom ( D ) . Then f = α δ u + g ∈ Dom ( H C ) f=\alpha\delta_u+g\in\operatorname{Dom}(\mathcal{H}_C) f = α δ u + g ∈ Dom ( H C ) , and
( H C − z ) f = ( ⟨ a , g ⟩ − z α ) δ u + α a + ( D − z ) g = α ( − ⟨ a , ( D − z ) − 1 a ⟩ − z ) δ u = δ u . (\mathcal{H}_C-z)f=\bigl(\langle a,g\rangle-z\alpha\bigr)\delta_u+\alpha a+(\mathcal{D}-z)g=\alpha\bigl(-\langle a,(\mathcal{D}-z)^{-1}a\rangle-z\bigr)\delta_u=\delta_u. ( H C − z ) f = ( ⟨ a , g ⟩ − z α ) δ u + α a + ( D − z ) g = α ( − ⟨ a , ( D − z ) − 1 a ⟩ − z ) δ u = δ u .
This shows that f = ( H C − z ) − 1 δ u f=(\mathcal{H}_C-z)^{-1}\delta_u f = ( H C − z ) − 1 δ u , and G u u C ( z ) = ⟨ δ u , f ⟩ = α = s − 1 G^C_{uu}(z)=\langle\delta_u,f\rangle=\alpha=s^{-1} G uu C ( z ) = ⟨ δ u , f ⟩ = α = s − 1 , which is (21) . Reading off the other coordinates of f f f , we also obtain, for every i i i and every v ∈ C x i → u v\in C_{x_i\to u} v ∈ C x i → u ,
G v u C ( z ) = − G u u C ( z ) h x i u G v x i C x i → u ( z ) . (27) G^C_{vu}(z)=-G^C_{uu}(z)h_{x_i u}G^{C_{x_i\to u}}_{v x_i}(z). \tag*{(27)} G v u C ( z ) = − G uu C ( z ) h x i u G v x i C x i → u ( z ) . (27)
For (ii), let C ′ = C u → w C'=C_{u\to w} C ′ = C u → w , which is a component of the finite removal ( F , D ∪ { { u , w } } ) (F,D\cup\{\{u,w\}\}) ( F , D ∪ {{ u , w }}) by Lemma 3.5(ii). The neighbors of u u u in C ′ C' C ′ are its neighbors in C C C other than w w w . For such a neighbor x x x , every vertex of C x → u C_{x\to u} C x → u is joined to w w w by a path through x x x and u u u . This gives C x → u ⊂ C ′ C_{x\to u}\subset C' C x → u ⊂ C ′ , and Lemma 3.5(iii) implies that C x → u ′ = C x → u C'_{x\to u}=C_{x\to u} C x → u ′ = C x → u . Applying (21) to C ′ C' C ′ at u u u , and using G u u C ′ = Γ u → w C G^{C'}_{uu}=\Gamma^C_{u\to w} G uu C ′ = Γ u → w C and Γ x → u C ′ = Γ x → u C \Gamma^{C'}_{x\to u}=\Gamma^C_{x\to u} Γ x → u C ′ = Γ x → u C , we obtain (22) .
For the full tree, the branch V v j → v V_{v_j\to v} V v j → v is V ( T v j ) V(\mathcal{T}_{v_j}) V ( T v j ) and Γ v j → v = Γ v j \Gamma_{v_j\to v}=\Gamma_{v_j} Γ v j → v = Γ v j , and the neighbors of v v v are its children and, if v ≠ o v\ne o v = o , its parent. With these identifications, (24) and (25) are (22) with ( u , w ) = ( v , v − ) (u,w)=(v,v^-) ( u , w ) = ( v , v − ) (or (21) at u = o u=o u = o ) and (21) at u = v u=v u = v . Formula (26) is (22) with ( u , w ) = ( u , v ) (u,w)=(u,v) ( u , w ) = ( u , v ) , since the neighbors of u u u other than v v v are its parent (if u ≠ o u\ne o u = o ) and its other children.
For (iii), on Ω s a \Omega_{\mathrm{sa}} Ω sa , (24) at v = o v=o v = o , the identity ∣ h o j ∣ 2 = Y j / p |h_{o j}|^2=Y_j/p ∣ h o j ∣ 2 = Y j / p , and (20) give (23) ; the functions Γ j \Gamma_j Γ j with j > N j>N j > N do not enter it. By (11) , the marked trees θ j T \theta_j\mathcal{T} θ j T are independent copies of T \mathcal{T} T , independent of N o N_o N o and ( w j ) j ≥ 1 (w_j)_{j\ge1} ( w j ) j ≥ 1 . The functions Γ j = Φ ( θ j T ) \Gamma_j=\Phi(\theta_j\mathcal{T}) Γ j = Φ ( θ j T ) are independent copies of Φ ( T ) \Phi(\mathcal{T}) Φ ( T ) , which equals Γ o \Gamma_o Γ o almost surely, and they are independent of N N N and ( Y j ) j ≥ 1 (Y_j)_{j\ge1} ( Y j ) j ≥ 1 . The variables N o N_o N o and ( w j ) j ≥ 1 (w_j)_{j\ge1} ( w j ) j ≥ 1 are independent by construction, and Y j = ∣ w j ∣ 2 Y_j=|w_j|^2 Y j = ∣ w j ∣ 2 has the law Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) . The last statement is Theorem 3.10(iii).
Remark 3.17. Let x ∈ V x\in V x ∈ V and let o = v 0 , v 1 , … , v k = x o=v_0,v_1,\ldots,v_k=x o = v 0 , v 1 , … , v k = x be the path from the root to x x x . When T \mathcal{T} T is rerooted at x x x , the parent of a vertex v ≠ x v\ne x v = x is the next vertex on the path from v v v to x x x . If v v v is not on the path, this is v − v^- v − , since the path from v v v to x x x first climbs to the last common ancestor of v v v and x x x . The forward cavities of the rerooted tree are Γ v \Gamma_v Γ v for v v v off the path, and Γ v j → v j + 1 \Gamma_{v_j\to v_{j+1}} Γ v j → v j + 1 for 0 ≤ j < k 0\le j<k 0 ≤ j < k ; the latter are rootward cavities, and (22) computes them recursively from v 0 v_0 v 0 to v k − 1 v_{k-1} v k − 1 in terms of the forward cavities of the subtrees hanging off the path. The same holds for every component of a finite removal, rerooted at any of its vertices.
Proposition 3.18. On Ω s a \Omega_{\mathrm{sa}} Ω sa , let C C C be a component of a finite removal, let u = v 0 , v 1 , … , v L = v u=v_0,v_1,\ldots,v_L=v u = v 0 , v 1 , … , v L = v be a path in C C C with L ≥ 1 L\ge1 L ≥ 1 , and let z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R . Then
G u v C ( z ) = ( − 1 ) L G u u C ( z ) ∏ j = 1 L h v j − 1 v j Γ v j → v j − 1 C ( z ) . (28) G^C_{uv}(z)=(-1)^L G^C_{uu}(z)\prod_{j=1}^{L}h_{v_{j-1}v_j}\Gamma^C_{v_j\to v_{j-1}}(z). \tag*{(28)} G uv C ( z ) = ( − 1 ) L G uu C ( z ) j = 1 ∏ L h v j − 1 v j Γ v j → v j − 1 C ( z ) . (28)
G v u C ( z ) = ( − 1 ) L G u u C ( z ) ∏ j = 1 L h v j v j − 1 Γ v j → v j − 1 C ( z ) . (29) G^C_{vu}(z)=(-1)^L G^C_{uu}(z)\prod_{j=1}^{L}h_{v_jv_{j-1}}\Gamma^C_{v_j\to v_{j-1}}(z). \tag*{(29)} G v u C ( z ) = ( − 1 ) L G uu C ( z ) j = 1 ∏ L h v j v j − 1 Γ v j → v j − 1 C ( z ) . (29)
If u u u and v v v lie in different components of a finite removal ( F , D ) (F,D) ( F , D ) , then G u v F , D ( z ) = 0 G^{F,D}_{uv}(z)=0 G uv F , D ( z ) = 0 . Every finite connected set Λ ⊂ V \Lambda\subset V Λ ⊂ V is a component of the finite removal that deletes the edges between Λ \Lambda Λ and V ∖ Λ V\setminus\Lambda V ∖ Λ ; in particular, these identities also hold for the Dirichlet truncation H [ Λ ] \mathcal{H}^{[\Lambda]} H [ Λ ] , for every finite connected Λ \Lambda Λ .
Proof. We prove (3.19) by induction on L L L , simultaneously for all components of all finite removals. For L = 1 L=1 L = 1 , it is (3.17) with x i = v = v 1 x_i=v=v_1 x i = v = v 1 , since G v 1 v 1 C v 1 → u = Γ v 1 → u C G^{C_{v_1\to u}}_{v_1v_1}=\Gamma^C_{v_1\to u} G v 1 v 1 C v 1 → u = Γ v 1 → u C . Let L ≥ 2 L\ge2 L ≥ 2 and let C ′ = C v 1 → u C'=C_{v_1\to u} C ′ = C v 1 → u , which is a component of a finite removal. The vertices v 1 , … , v L v_1,\ldots,v_L v 1 , … , v L belong to C ′ C' C ′ , because their paths to u u u pass through v 1 v_1 v 1 . By (3.17) and the induction hypothesis applied in C ′ C' C ′ to the path v 1 , … , v L v_1,\ldots,v_L v 1 , … , v L , we have
G v u C = − G u u C h v 1 u G v v 1 C ′ = − G u u C h v 1 u ( − 1 ) L − 1 G v 1 v 1 C ′ ∏ j = 2 L h v j v j − 1 Γ v j → v j − 1 C ′ . G^C_{vu}=-G^C_{uu}h_{v_1u}G^{C'}_{vv_1}=-G^C_{uu}h_{v_1u}(-1)^{L-1}G^{C'}_{v_1v_1}\prod_{j=2}^{L}h_{v_jv_{j-1}}\Gamma^{C'}_{v_j\to v_{j-1}}. G v u C = − G uu C h v 1 u G v v 1 C ′ = − G uu C h v 1 u ( − 1 ) L − 1 G v 1 v 1 C ′ j = 2 ∏ L h v j v j − 1 Γ v j → v j − 1 C ′ .
We have G v 1 v 1 C ′ = Γ v 1 → u C G^{C'}_{v_1v_1}=\Gamma^C_{v_1\to u} G v 1 v 1 C ′ = Γ v 1 → u C . For j ≥ 2 j\ge2 j ≥ 2 , the path from a vertex of C v j → v j − 1 C_{v_j\to v_{j-1}} C v j → v j − 1 to u u u passes through v j , v j − 1 , … , v 1 v_j,v_{j-1},\ldots,v_1 v j , v j − 1 , … , v 1 . Then C v j → v j − 1 ⊂ C ′ C_{v_j\to v_{j-1}}\subset C' C v j → v j − 1 ⊂ C ′ , and Lemma 3.5(iii) implies that Γ v j → v j − 1 C ′ = Γ v j → v j − 1 C \Gamma^{C'}_{v_j\to v_{j-1}}=\Gamma^C_{v_j\to v_{j-1}} Γ v j → v j − 1 C ′ = Γ v j → v j − 1 C . This proves (3.19).
By Lemma 3.14(i), we have G u v C ( z ) = G v u C ( z ‾ ) ‾ G^C_{uv}(z)=\overline{G^C_{vu}(\overline{z})} G uv C ( z ) = G v u C ( z ) . Applying (3.19) at z ‾ \overline{z} z , and using G u u C ( z ‾ ) ‾ = G u u C ( z ) \overline{G^C_{uu}(\overline{z})}=G^C_{uu}(z) G uu C ( z ) = G uu C ( z ) , Γ v j → v j − 1 C ( z ‾ ) ‾ = Γ v j → v j − 1 C ( z ) \overline{\Gamma^C_{v_j\to v_{j-1}}(\overline{z})}=\Gamma^C_{v_j\to v_{j-1}}(z) Γ v j → v j − 1 C ( z ) = Γ v j → v j − 1 C ( z ) and h v j v j − 1 = h v j − 1 v j h_{v_jv_{j-1}}=h_{v_{j-1}v_j} h v j v j − 1 = h v j − 1 v j , we obtain (3.18). The statement for different components holds because H F , D \mathcal{H}^{F,D} H F , D is an orthogonal sum over components. □
The Ward identity
Lemma 3.19. On Ω s a \Omega_{\mathrm{sa}} Ω sa , let C C C be a component of a finite removal and let z = E + i η ∈ C ∖ R z=E+i\eta\in\mathbb{C}\setminus\mathbb{R} z = E + i η ∈ C ∖ R .
(i) For every u ∈ C u\in C u ∈ C ,
∑ v ∈ C ∣ G u v C ( z ) ∣ 2 = ∑ v ∈ C ∣ G v u C ( z ) ∣ 2 = Im G u u C ( z ) η . \sum_{v\in C}\left|G^C_{uv}(z)\right|^2=\sum_{v\in C}\left|G^C_{vu}(z)\right|^2=\frac{\operatorname{Im}G^C_{uu}(z)}{\eta}. v ∈ C ∑ G uv C ( z ) 2 = v ∈ C ∑ G v u C ( z ) 2 = η Im G uu C ( z ) . In particular, if η > 0 \eta>0 η > 0 , then Im G u u C ( z ) > 0 \operatorname{Im}G^C_{uu}(z)>0 Im G uu C ( z ) > 0 and Im Γ u → w C ( z ) > 0 \operatorname{Im}\Gamma^C_{u\to w}(z)>0 Im Γ u → w C ( z ) > 0 for all u u u and all neighbors w w w of u u u in C C C .
(ii) For all adjacent u , w ∈ C u,w\in C u , w ∈ C ,
Im Γ u → w C ( z ) = ∣ Γ u → w C ( z ) ∣ 2 ( η + ∑ x ∈ C , x ∼ u x ≠ w ∣ h u x ∣ 2 Im Γ x → u C ( z ) ) , \operatorname{Im}\Gamma^C_{u\to w}(z)=\left|\Gamma^C_{u\to w}(z)\right|^2\left(\eta+\sum_{\substack{x\in C,\ x\sim u\\x\ne w}}\left|h_{ux}\right|^2\operatorname{Im}\Gamma^C_{x\to u}(z)\right), Im Γ u → w C ( z ) = Γ u → w C ( z ) 2 η + x ∈ C , x ∼ u x = w ∑ ∣ h ux ∣ 2 Im Γ x → u C ( z ) , Im G u u C ( z ) = ∣ G u u C ( z ) ∣ 2 ( η + ∑ x ∈ C , x ∼ u ∣ h u x ∣ 2 Im Γ x → u C ( z ) ) . \operatorname{Im}G^C_{uu}(z)=\left|G^C_{uu}(z)\right|^2\left(\eta+\sum_{x\in C,\ x\sim u}\left|h_{ux}\right|^2\operatorname{Im}\Gamma^C_{x\to u}(z)\right). Im G uu C ( z ) = G uu C ( z ) 2 η + x ∈ C , x ∼ u ∑ ∣ h ux ∣ 2 Im Γ x → u C ( z ) .
(iii) Let u , x ∈ C u, x \in C u , x ∈ C , and let y y y be a neighbor of x x x in C C C which does not lie on the path from u u u to x x x . If η > 0 \eta> 0 η > 0 , then
Im G u u C ( z ) ≥ ∣ G u x C ( z ) ∣ 2 ∣ h x y ∣ 2 Im Γ y → x C ( z ) . \operatorname{Im} G^{C}_{uu}(z) \ge\lvert G^{C}_{ux}(z)\rvert^{2} \lvert h_{xy}\rvert^{2} \operatorname{Im} \Gamma^{C}_{y\to x}(z). Im G uu C ( z ) ≥ ∣ G ux C ( z ) ∣ 2 ∣ h x y ∣ 2 Im Γ y → x C ( z ) .
In particular, Im G o o ( z ) ≥ ∣ G o x ( z ) ∣ 2 ∣ h x y ∣ 2 Im Γ y ( z ) \operatorname{Im} G_{oo}(z) \ge\lvert G_{ox}(z)\rvert^{2} \lvert h_{xy}\rvert^{2} \operatorname{Im} \Gamma_y(z) Im G oo ( z ) ≥ ∣ G o x ( z ) ∣ 2 ∣ h x y ∣ 2 Im Γ y ( z ) for every x ∈ V x \in V x ∈ V and every child y y y of x x x .
Proof. For (i), we have ∑ v ∣ G v u C ( z ) ∣ 2 = ∥ ( H C − z ) − 1 δ u ∥ 2 \sum_v \lvert G^{C}_{vu}(z)\rvert^2 = \lVert(H_C-z)^{-1}\delta_u\rVert^2 ∑ v ∣ G v u C ( z ) ∣ 2 = ∥( H C − z ) − 1 δ u ∥ 2 , which equals Im G u u C ( z ) / η \operatorname{Im} G^{C}_{uu}(z)/\eta Im G uu C ( z ) / η by Lemma 3.14(ii). By Lemma 3.14(i), we have ∣ G u v C ( z ) ∣ = ∣ G v u C ( z ˉ ) ∣ \lvert G^{C}_{uv}(z)\rvert= \lvert G^{C}_{vu}(\bar{z})\rvert ∣ G uv C ( z )∣ = ∣ G v u C ( z ˉ )∣ , and the first sum equals Im G u u C ( z ˉ ) / ( − η ) = Im G u u C ( z ) / η \operatorname{Im} G^{C}_{uu}(\bar{z})/(-\eta) = \operatorname{Im} G^{C}_{uu}(z)/\eta Im G uu C ( z ˉ ) / ( − η ) = Im G uu C ( z ) / η . The positivity statements follow from Lemma 3.14(ii), since Γ u → w C \Gamma^{C}_{u\to w} Γ u → w C is a diagonal entry of the operator of C u → w C_{u\to w} C u → w .
For (ii), write (3.12) as Γ = − 1 / s \Gamma=-1/s Γ = − 1/ s . Then Im Γ = Im s / ∣ s ∣ 2 = ∣ Γ ∣ 2 Im s \operatorname{Im}\Gamma=\operatorname{Im}s/\lvert s\rvert^2=\lvert\Gamma\rvert^2\operatorname{Im}s Im Γ = Im s / ∣ s ∣ 2 = ∣ Γ ∣ 2 Im s , and Im s \operatorname{Im}s Im s is the expression in parentheses. The same computation applies to (3.11).
For (iii), let B = C y → x B=C_{y\to x} B = C y → x . Let u = v 0 , … , v k = x u=v_0,\ldots,v_k=x u = v 0 , … , v k = x be the path from u u u to x x x (with k = 0 k=0 k = 0 if u = x u=x u = x ). No vertex of this path lies in B B B : otherwise, by Lemma 3.5 (ii), its path to x x x , which is part of the path from u u u to x x x , would contain y y y . Let v ∈ B v\in B v ∈ B and let y = v k + 1 , … , v L = v y=v_{k+1},\ldots,v_L=v y = v k + 1 , … , v L = v be the path from y y y to v v v , which lies in B B B . Then v 0 , … , v L v_0,\ldots,v_L v 0 , … , v L is the path from u u u to v v v . Applying (28) in C C C to the paths from u u u to x x x and from u u u to v v v , and in B B B to the path from y y y to v v v , and using G y y B = Γ y → x C G^{B}_{yy}=\Gamma^{C}_{y\to x} G y y B = Γ y → x C and Γ v j → v j − 1 B = Γ v j → v j − 1 C \Gamma^{B}_{v_j\to v_{j-1}}=\Gamma^{C}_{v_j\to v_{j-1}} Γ v j → v j − 1 B = Γ v j → v j − 1 C for j ≥ k + 2 j\ge k+2 j ≥ k + 2 (by Lemma 3.5 (iii), as in the proof of Proposition 3.18 ), we obtain
G u v C ( z ) = − G u x C ( z ) h x y G y v B ( z ) , v ∈ B . G^{C}_{uv}(z)=-G^{C}_{ux}(z)h_{xy}G^{B}_{yv}(z), \qquad v\in B. G uv C ( z ) = − G ux C ( z ) h x y G y v B ( z ) , v ∈ B .
Using (i) in C C C and in B B B , we find
Im G u u C ( z ) η ≥ ∑ v ∈ B ∣ G u v C ( z ) ∣ 2 = ∣ G u x C ( z ) ∣ 2 ∣ h x y ∣ 2 ∑ v ∈ B ∣ G y v B ( z ) ∣ 2 = ∣ G u x C ( z ) ∣ 2 ∣ h x y ∣ 2 Im Γ y → x C ( z ) η . \frac{\operatorname{Im}G^{C}_{uu}(z)}{\eta}\ge\sum_{v\in B}\lvert G^{C}_{uv}(z)\rvert^2=\lvert G^{C}_{ux}(z)\rvert^2\lvert h_{xy}\rvert^2\sum_{v\in B}\lvert G^{B}_{yv}(z)\rvert^2=\lvert G^{C}_{ux}(z)\rvert^2\lvert h_{xy}\rvert^2\frac{\operatorname{Im}\Gamma^{C}_{y\to x}(z)}{\eta}. η Im G uu C ( z ) ≥ v ∈ B ∑ ∣ G uv C ( z ) ∣ 2 = ∣ G ux C ( z ) ∣ 2 ∣ h x y ∣ 2 v ∈ B ∑ ∣ G y v B ( z ) ∣ 2 = ∣ G ux C ( z ) ∣ 2 ∣ h x y ∣ 2 η Im Γ y → x C ( z ) .
For the last statement, a child y y y of x x x is never on the path from o o o to x x x , and Γ y → x = Γ y \Gamma_{y\to x}=\Gamma_y Γ y → x = Γ y .
Gauge and parity symmetries
The phases of the edge weights can be removed by a diagonal unitary transformation, because a tree has no cycles. For ζ ∈ C \zeta\in\mathbb{C} ζ ∈ C , let σ ( ζ ) = ζ / ∣ ζ ∣ \sigma(\zeta)=\zeta/\lvert\zeta\rvert σ ( ζ ) = ζ / ∣ ζ ∣ if ζ ≠ 0 \zeta\ne0 ζ = 0 and σ ( 0 ) = 1 \sigma(0)=1 σ ( 0 ) = 1 .
Lemma 3.20. On Ω s a \Omega_{\mathrm{sa}} Ω sa , let C C C be a component of a finite removal and fix r ∈ C r\in C r ∈ C . For v ∈ C v\in C v ∈ C , let r = v 0 , v 1 , … , v k = v r=v_0,v_1,\ldots,v_k=v r = v 0 , v 1 , … , v k = v be the path from r r r to v v v and set φ v = ∏ j = 1 k σ ( h v j − 1 v j ) \varphi_v=\prod_{j=1}^{k}\sigma(h_{v_{j-1}v_j}) φ v = ∏ j = 1 k σ ( h v j − 1 v j ) , and let U U U be the unitary operator ( U f ) ( v ) = φ v f ( v ) (Uf)(v)=\varphi_v f(v) ( U f ) ( v ) = φ v f ( v ) on ℓ 2 ( C ) \ell^2(C) ℓ 2 ( C ) . Then the weighted adjacency operator of C C C with the nonnegative weights ∣ h u v ∣ \lvert h_{uv}\rvert ∣ h uv ∣ is essentially self-adjoint on c c ( C ) c_c(C) c c ( C ) , and its closure H C a b s H_C^{\mathrm{abs}} H C abs satisfies U H C U ∗ = H C a b s UH_CU^*=H_C^{\mathrm{abs}} U H C U ∗ = H C abs . Further, for all u , v ∈ C u,v\in C u , v ∈ C and z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R , writing G a b s G^{\mathrm{abs}} G abs for the resolvent entries of H C a b s H_C^{\mathrm{abs}} H C abs ,
G u v C ( z ) = φ u ‾ φ v G u v a b s ( z ) , G u v a b s ( z ) = G v u a b s ( z ) , ∣ G u v C ( z ) ∣ = ∣ G v u C ( z ) ∣ . G^{C}_{uv}(z)=\overline{\varphi_u}\varphi_vG^{\mathrm{abs}}_{uv}(z), \qquad G^{\mathrm{abs}}_{uv}(z)=G^{\mathrm{abs}}_{vu}(z), \qquad\lvert G^{C}_{uv}(z)\rvert=\lvert G^{C}_{vu}(z)\rvert. G uv C ( z ) = φ u φ v G uv abs ( z ) , G uv abs ( z ) = G v u abs ( z ) , ∣ G uv C ( z )∣ = ∣ G v u C ( z )∣ . The same identities hold for the entries of G F , D G^{F,D} G F , D , which vanish between different components.
Proof. Let u , v ∈ C u,v \in C u , v ∈ C be adjacent, and suppose that u u u lies on the path from r r r to v v v . Then φ v = φ u σ ( h u v ) \varphi_v=\varphi_u\sigma(h_{uv}) φ v = φ u σ ( h uv ) , and
φ u ‾ h u v φ v ‾ = h u v σ ( h u v ) ‾ = ∣ h u v ∣ , φ v ‾ h v u φ u ‾ = σ ( h u v ) h u v ‾ = ∣ h u v ∣ . \overline{\varphi_u}h_{uv}\overline{\varphi_v}=h_{uv}\overline{\sigma(h_{uv})}=|h_{uv}|,\qquad\overline{\varphi_v}h_{vu}\overline{\varphi_u}=\sigma(h_{uv})\overline{h_{uv}}=|h_{uv}|. φ u h uv φ v = h uv σ ( h uv ) = ∣ h uv ∣ , φ v h v u φ u = σ ( h uv ) h uv = ∣ h uv ∣.
where we used ∣ φ u ∣ = 1 |\varphi_u|=1 ∣ φ u ∣ = 1 and h v u = h u v ‾ h_{vu}=\overline{h_{uv}} h v u = h uv . Every pair of adjacent vertices is of this form after possibly exchanging u u u and v v v . Since U U U is diagonal, it maps c c ( C ) c_c(C) c c ( C ) onto itself and U H C , 0 U ∗ UH_{C,0}U^* U H C , 0 U ∗ is the weighted adjacency operator with the weights ∣ h u v ∣ |h_{uv}| ∣ h uv ∣ on c c ( C ) c_c(C) c c ( C ) . Since conjugation by a unitary operator commutes with taking closures and adjoints, this operator is essentially self-adjoint with closure U H C U ∗ UH_CU^* U H C U ∗ . Then U ( H C − z ) − 1 U ∗ = ( H C a b s − z ) − 1 U(H_C-z)^{-1}U^*=(H_C^{\mathrm{abs}}-z)^{-1} U ( H C − z ) − 1 U ∗ = ( H C abs − z ) − 1 , and taking the ( u , v ) (u,v) ( u , v ) entry gives the first identity.
For the second identity, let K K K be complex conjugation, ( K f ) ( v ) = f ( v ) ‾ (Kf)(v)=\overline{f(v)} ( K f ) ( v ) = f ( v ) . It is antiunitary, maps c c ( C ) c_c(C) c c ( C ) onto itself, and commutes with the minimal operator with real weights. Taking closures, we have K H C a b s K = H C a b s KH_C^{\mathrm{abs}}K=H_C^{\mathrm{abs}} K H C abs K = H C abs . If f = ( H C a b s − z ) − 1 K g f=(H_C^{\mathrm{abs}}-z)^{-1}Kg f = ( H C abs − z ) − 1 K g , then K f ∈ Dom ( H C a b s ) Kf\in\operatorname{Dom}(H_C^{\mathrm{abs}}) K f ∈ Dom ( H C abs ) and ( H C a b s − z ‾ ) K f = K ( H C a b s − z ) f = g (H_C^{\mathrm{abs}}-\overline{z})Kf=K(H_C^{\mathrm{abs}}-z)f=g ( H C abs − z ) K f = K ( H C abs − z ) f = g . This proves that K ( H C a b s − z ) − 1 K = ( H C a b s − z ‾ ) − 1 K(H_C^{\mathrm{abs}}-z)^{-1}K=(H_C^{\mathrm{abs}}-\overline{z})^{-1} K ( H C abs − z ) − 1 K = ( H C abs − z ) − 1 . Using Lemma 3.14(i) and ⟨ δ u , K f ⟩ = ⟨ δ u , f ⟩ ‾ \langle\delta_u,Kf\rangle=\overline{\langle\delta_u,f\rangle} ⟨ δ u , K f ⟩ = ⟨ δ u , f ⟩ , we have
G v u a b s ( z ) = ⟨ δ u , ( H C a b s − z ‾ ) − 1 δ v ⟩ = ⟨ δ u , K ( H C a b s − z ) − 1 K δ v ⟩ = G u v a b s ( z ) ‾ . G^{\mathrm{abs}}_{vu}(z)=\langle\delta_u,(H_C^{\mathrm{abs}}-\overline{z})^{-1}\delta_v\rangle=\langle\delta_u,K(H_C^{\mathrm{abs}}-z)^{-1}K\delta_v\rangle=\overline{G^{\mathrm{abs}}_{uv}(z)}. G v u abs ( z ) = ⟨ δ u , ( H C abs − z ) − 1 δ v ⟩ = ⟨ δ u , K ( H C abs − z ) − 1 K δ v ⟩ = G uv abs ( z ) .
The third identity follows from the first two, since ∣ φ u ∣ = ∣ φ v ∣ = 1 |\varphi_u|=|\varphi_v|=1 ∣ φ u ∣ = ∣ φ v ∣ = 1 . The statement for G F , D G^{F,D} G F , D follows by applying the lemma to each component.
Recall that V 0 V_0 V 0 and V 1 V_1 V 1 are the sets of vertices of even and odd generation. Every edge of T \mathcal{T} T joins V 0 V_0 V 0 and V 1 V_1 V 1 .
Proposition 3.21.
(i) On Ω s a \Omega_{\mathrm{sa}} Ω sa , let C C C be a component of a finite removal, and let J J J be the unitary operator ( J f ) ( v ) = ( − 1 ) ∣ v ∣ f ( v ) (Jf)(v)=(-1)^{|v|}f(v) ( J f ) ( v ) = ( − 1 ) ∣ v ∣ f ( v ) on ℓ 2 ( C ) \ell^2(C) ℓ 2 ( C ) . Then J H C J = − H C JH_CJ=-H_C J H C J = − H C . Further, J 1 B ( H C ) J = 1 − B ( H C ) J\mathbf{1}_B(H_C)J=\mathbf{1}_{-B}(H_C) J 1 B ( H C ) J = 1 − B ( H C ) for every Borel set B ⊂ R B\subset\mathbb{R} B ⊂ R ; the spectral measure of every δ v \delta_v δ v , v ∈ C v\in C v ∈ C , is invariant under λ ↦ − λ \lambda\mapsto-\lambda λ ↦ − λ ;
G u v C ( − z ) = − ( − 1 ) ∣ u ∣ + ∣ v ∣ G u v C ( z ) , u , v ∈ C , z ∈ C ∖ R ; G^C_{uv}(-z)=-(-1)^{|u|+|v|}G^C_{uv}(z),\qquad u,v\in C,\ z\in\mathbb{C}\setminus\mathbb{R}; G uv C ( − z ) = − ( − 1 ) ∣ u ∣ + ∣ v ∣ G uv C ( z ) , u , v ∈ C , z ∈ C ∖ R ; and for every Borel set I ⊂ R I\subset\mathbb{R} I ⊂ R , the operator H C H_C H C has pure point spectrum in I I I if and only if it has pure point spectrum in − I -I − I , and it is purely absolutely continuous in I I I if and only if it is purely absolutely continuous in − I -I − I . In particular, this holds for H H H under P \mathbb{P} P and under P ∞ \mathbb{P}^{\infty} P ∞ .
(ii) For the finite model of Definition 3.1, the pairs ( G n , H n ) (G_n,H_n) ( G n , H n ) and ( G n , − H n ) (G_n,-H_n) ( G n , − H n ) have the same law. In particular H n = d − H n H_n\overset{\mathrm{d}}{=}-H_n H n = d − H n and H n g c = d − H n g c H_n^{\mathrm{gc}}\overset{\mathrm{d}}{=}-H_n^{\mathrm{gc}} H n gc = d − H n gc .
Proof. For (i), the operator J J J is a self-adjoint unitary which maps c c ( C ) c_c(C) c c ( C ) onto itself. For f ∈ c c ( C ) f\in c_c(C) f ∈ c c ( C ) and v ∈ C v\in C v ∈ C , we have
( J H C , 0 J f ) ( v ) = ( − 1 ) ∣ v ∣ ∑ u ∼ v h v u ( − 1 ) ∣ u ∣ f ( u ) = − ( H C , 0 f ) ( v ) , (JH_{C,0}Jf)(v)=(-1)^{|v|}\sum_{u\sim v}h_{vu}(-1)^{|u|}f(u)=-(H_{C,0}f)(v), ( J H C , 0 J f ) ( v ) = ( − 1 ) ∣ v ∣ u ∼ v ∑ h v u ( − 1 ) ∣ u ∣ f ( u ) = − ( H C , 0 f ) ( v ) ,
since adjacent vertices have generations differing by one. Taking closures, we find J H C J = − H C J\mathcal{H}_{C}J=-\mathcal{H}_{C} J H C J = − H C . Since the functional calculus is invariant under unitary conjugation, we have J ϕ ( H C ) J = ϕ ( − H C ) J\phi(\mathcal{H}_{C})J=\phi(-\mathcal{H}_{C}) J ϕ ( H C ) J = ϕ ( − H C ) for every bounded Borel function ϕ \phi ϕ . With ϕ = 1 B \phi=\mathbf{1}_{B} ϕ = 1 B , this gives J 1 B ( H C ) J = 1 − B ( H C ) J\mathbf{1}_{B}(\mathcal{H}_{C})J=\mathbf{1}_{-B}(\mathcal{H}_{C}) J 1 B ( H C ) J = 1 − B ( H C ) , and with ϕ ( λ ) = ( λ − z ) − 1 \phi(\lambda)=(\lambda-z)^{-1} ϕ ( λ ) = ( λ − z ) − 1 it gives J ( H C − z ) − 1 J = − ( H C + z ) − 1 J(\mathcal{H}_{C}-z)^{-1}J=-(\mathcal{H}_{C}+z)^{-1} J ( H C − z ) − 1 J = − ( H C + z ) − 1 , whose ( u , v ) (u,v) ( u , v ) entry is the displayed identity. For every f ∈ ℓ 2 ( C ) f\in\ell^{2}(\mathbb{C}) f ∈ ℓ 2 ( C ) and Borel B B B , we find
μ f H C ( B ) = ⟨ f , J 1 B ( H C ) J f ⟩ = μ f H C ( − B ) . \mu_{f}^{\mathcal{H}_{C}}(B)=\langle f,J\mathbf{1}_{B}(\mathcal{H}_{C})Jf\rangle=\mu_{f}^{\mathcal{H}_{C}}(-B). μ f H C ( B ) = ⟨ f , J 1 B ( H C ) J f ⟩ = μ f H C ( − B ) .
Since J δ v = ± δ v J\delta_{v}=\pm\delta_{v} J δ v = ± δ v , the spectral measure of δ v \delta_{v} δ v is reflection invariant. If H C \mathcal{H}_{C} H C has pure point spectrum in I I I , then for every f f f the restriction of μ f H C \mu_{f}^{\mathcal{H}_{C}} μ f H C to I I I is purely atomic, and by the last display the restriction of μ f H C \mu_{f}^{\mathcal{H}_{C}} μ f H C to − I -I − I is purely atomic, and H C \mathcal{H}_{C} H C has pure point spectrum in − I -I − I . The same argument applies to absolute continuity, since reflection preserves Lebesgue null sets.
For (ii), since the variables − w i j -w_{ij} − w ij are independent standard complex Gaussians, independent of G n G_{n} G n , the pairs ( G n , ( w i j ) ) (G_{n},(w_{ij})) ( G n , ( w ij )) and ( G n , ( − w i j ) ) (G_{n},(-w_{ij})) ( G n , ( − w ij )) have the same law. The matrix H n H_{n} H n is a function of ( G n , ( w i j ) ) (G_{n},(w_{ij})) ( G n , ( w ij )) , and replacing ( w i j ) (w_{ij}) ( w ij ) by ( − w i j ) (-w_{ij}) ( − w ij ) replaces H n H_{n} H n by − H n -H_{n} − H n . The set g c \mathrm{gc} gc is a function of G n G_{n} G n .
The bipartite factorization
Since every edge joins V 0 V_{0} V 0 and V 1 V_{1} V 1 , the operator H \mathcal{H} H maps functions on V 0 V_{0} V 0 to functions on V 1 V_{1} V 1 and conversely, and its square splits into two nonnegative blocks. In this section, we make this precise for the unbounded operator, and we write each block as a sum of rank-one terms, one for each vertex of the opposite parity, with independent Gamma-distributed squared norms. We begin with an abstract lemma.
Lemma 3.22. Let h 0 , h 1 \mathfrak{h}_{0},\mathfrak{h}_{1} h 0 , h 1 be Hilbert spaces and let A A A be a closed, densely defined operator from Dom ( A ) ⊂ h 0 \operatorname{Dom}(A)\subset\mathfrak{h}_{0} Dom ( A ) ⊂ h 0 to h 1 \mathfrak{h}_{1} h 1 . Define the operator T A T_{A} T A on h 0 ⊕ h 1 \mathfrak{h}_{0}\oplus\mathfrak{h}_{1} h 0 ⊕ h 1 by
Dom ( T A ) = Dom ( A ) ⊕ Dom ( A ∗ ) , T A ( f , g ) = ( A ∗ g , A f ) . \operatorname{Dom}(T_{A})=\operatorname{Dom}(A)\oplus\operatorname{Dom}(A^{*}),\qquad T_{A}(f,g)=(A^{*}g,Af). Dom ( T A ) = Dom ( A ) ⊕ Dom ( A ∗ ) , T A ( f , g ) = ( A ∗ g , A f ) . (i) The operator T A T_{A} T A is self-adjoint, Dom ( T A 2 ) = Dom ( A ∗ A ) ⊕ Dom ( A A ∗ ) \operatorname{Dom}(T_{A}^{2})=\operatorname{Dom}(A^{*}A)\oplus\operatorname{Dom}(AA^{*}) Dom ( T A 2 ) = Dom ( A ∗ A ) ⊕ Dom ( A A ∗ ) , and T A 2 = A ∗ A ⊕ A A ∗ T_{A}^{2}=A^{*}A\oplus AA^{*} T A 2 = A ∗ A ⊕ A A ∗ . The operators A ∗ A A^{*}A A ∗ A and A A ∗ AA^{*} A A ∗ are nonnegative and self-adjoint. Further, Dom ( ( A ∗ A ) 1 / 2 ) = Dom ( A ) \operatorname{Dom}((A^{*}A)^{1/2})=\operatorname{Dom}(A) Dom (( A ∗ A ) 1/2 ) = Dom ( A ) and ∥ ( A ∗ A ) 1 / 2 f ∥ = ∥ A f ∥ \|(A^{*}A)^{1/2}f\|=\|Af\| ∥ ( A ∗ A ) 1/2 f ∥ = ∥ A f ∥ for f ∈ Dom ( A ) f\in\operatorname{Dom}(A) f ∈ Dom ( A ) .
(ii) Let z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R . Then z 2 ∉ [ 0 , ∞ ) z^{2}\notin[0,\infty) z 2 ∈ / [ 0 , ∞ ) , and with Q 0 = ( A ∗ A − z 2 ) − 1 Q_{0}=(A^{*}A-z^{2})^{-1} Q 0 = ( A ∗ A − z 2 ) − 1 and Q 1 = ( A A ∗ − z 2 ) − 1 Q_{1}=(AA^{*}-z^{2})^{-1} Q 1 = ( A A ∗ − z 2 ) − 1 ,
( T A − z ) − 1 ( f , g ) = ( z Q 0 f + A ∗ Q 1 g , A Q 0 f + z Q 1 g ) , f ∈ h 0 , g ∈ h 1 . (T_{A}-z)^{-1}(f,g)=(zQ_{0}f+A^{*}Q_{1}g,AQ_{0}f+zQ_{1}g),\qquad f\in\mathfrak{h}_{0},\ g\in\mathfrak{h}_{1}. ( T A − z ) − 1 ( f , g ) = ( z Q 0 f + A ∗ Q 1 g , A Q 0 f + z Q 1 g ) , f ∈ h 0 , g ∈ h 1 . In particular, P 0 ( T A − z ) − 1 P 0 = z Q 0 P_{0}(T_{A}-z)^{-1}P_{0}=zQ_{0} P 0 ( T A − z ) − 1 P 0 = z Q 0 and P 1 ( T A − z ) − 1 P 1 = z Q 1 P_{1}(T_{A}-z)^{-1}P_{1}=zQ_{1} P 1 ( T A − z ) − 1 P 1 = z Q 1 , where P 0 , P 1 P_{0},P_{1} P 0 , P 1 are the projections onto h 0 , h 1 \mathfrak{h}_{0},\mathfrak{h}_{1} h 0 , h 1 .
(iii) Let z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R and g ∈ Dom ( A ∗ ) g\in\operatorname{Dom}(A^{*}) g ∈ Dom ( A ∗ ) . Then Q 0 A ∗ g = A ∗ Q 1 g Q_{0}A^{*}g=A^{*}Q_{1}g Q 0 A ∗ g = A ∗ Q 1 g , and
⟨ A ∗ g , ( A ∗ A − z 2 ) − 1 A ∗ g ⟩ = ∥ g ∥ 2 + z ⟨ g , ( T A − z ) − 1 g ⟩ . \langle A^{*}g,(A^{*}A-z^{2})^{-1}A^{*}g\rangle=\|g\|^{2}+z\langle g,(T_{A}-z)^{-1}g\rangle. ⟨ A ∗ g , ( A ∗ A − z 2 ) − 1 A ∗ g ⟩ = ∥ g ∥ 2 + z ⟨ g , ( T A − z ) − 1 g ⟩ .
(iv) Let B B B be a densely defined operator from Dom ( B ) ⊂ h 1 \operatorname{Dom}(B) \subset\mathfrak{h}_1 Dom ( B ) ⊂ h 1 to h 0 \mathfrak{h}_0 h 0 . If the operator ( f , g ) ↦ ( B g , A f ) (f,g) \mapsto(Bg,Af) ( f , g ) ↦ ( B g , A f ) with domain Dom ( A ) ⊕ Dom ( B ) \operatorname{Dom}(A) \oplus\operatorname{Dom}(B) Dom ( A ) ⊕ Dom ( B ) is self-adjoint, then B = A ∗ B=A^* B = A ∗ , including domains.
Proof. We begin with (iv) and a computation of adjoints. Let T ′ T' T ′ be the operator in (iv). A pair ( φ , ψ ) (\varphi,\psi) ( φ , ψ ) is in Dom ( T ′ ∗ ) \operatorname{Dom}(T'^*) Dom ( T ′ ∗ ) exactly when the map ( f , g ) ↦ ⟨ ( φ , ψ ) , T ′ ( f , g ) ⟩ = ⟨ φ , B g ⟩ + ⟨ ψ , A f ⟩ (f,g) \mapsto\langle(\varphi,\psi),T'(f,g)\rangle= \langle\varphi,Bg\rangle+ \langle\psi,Af\rangle ( f , g ) ↦ ⟨( φ , ψ ) , T ′ ( f , g )⟩ = ⟨ φ , B g ⟩ + ⟨ ψ , A f ⟩ is bounded on Dom ( A ) ⊕ Dom ( B ) \operatorname{Dom}(A) \oplus\operatorname{Dom}(B) Dom ( A ) ⊕ Dom ( B ) . Taking g = 0 g=0 g = 0 and f = 0 f=0 f = 0 separately, we see that this holds exactly when ψ ∈ Dom ( A ∗ ) \psi\in\operatorname{Dom}(A^*) ψ ∈ Dom ( A ∗ ) and φ ∈ Dom ( B ∗ ) \varphi\in\operatorname{Dom}(B^*) φ ∈ Dom ( B ∗ ) , and then T ′ ∗ ( φ , ψ ) = ( A ∗ ψ , B ∗ φ ) T'^*(\varphi,\psi)=(A^*\psi,B^*\varphi) T ′ ∗ ( φ , ψ ) = ( A ∗ ψ , B ∗ φ ) . If T ′ = T ′ ∗ T'=T'^* T ′ = T ′ ∗ , then comparing domains, we find Dom ( B ) = Dom ( A ∗ ) \operatorname{Dom}(B)=\operatorname{Dom}(A^*) Dom ( B ) = Dom ( A ∗ ) , and comparing first components, we find B ψ = A ∗ ψ B\psi=A^*\psi B ψ = A ∗ ψ for ψ ∈ Dom ( B ) \psi\in\operatorname{Dom}(B) ψ ∈ Dom ( B ) . This proves (iv).
For (i), the same computation with B = A ∗ B=A^* B = A ∗ gives Dom ( T A ∗ ) = Dom ( A ∗ ∗ ) ⊕ Dom ( A ∗ ) \operatorname{Dom}(T_A^*)=\operatorname{Dom}(A^{**}) \oplus\operatorname{Dom}(A^*) Dom ( T A ∗ ) = Dom ( A ∗∗ ) ⊕ Dom ( A ∗ ) and T A ∗ ( φ , ψ ) = ( A ∗ ψ , A ∗ ∗ φ ) T_A^*(\varphi,\psi)=(A^*\psi,A^{**}\varphi) T A ∗ ( φ , ψ ) = ( A ∗ ψ , A ∗∗ φ ) . Since A A A is closed, we have A ∗ ∗ = A A^{**}=A A ∗∗ = A and T A ∗ = T A T_A^*=T_A T A ∗ = T A . A pair ( f , g ) ∈ Dom ( T A ) (f,g) \in\operatorname{Dom}(T_A) ( f , g ) ∈ Dom ( T A ) belongs to Dom ( T A 2 ) \operatorname{Dom}(T_A^2) Dom ( T A 2 ) exactly when ( A ∗ g , A f ) ∈ Dom ( A ) ⊕ Dom ( A ∗ ) (A^*g,Af) \in\operatorname{Dom}(A) \oplus\operatorname{Dom}(A^*) ( A ∗ g , A f ) ∈ Dom ( A ) ⊕ Dom ( A ∗ ) , that is, when f ∈ Dom ( A ∗ A ) f \in\operatorname{Dom}(A^*A) f ∈ Dom ( A ∗ A ) and g ∈ Dom ( A A ∗ ) g \in\operatorname{Dom}(AA^*) g ∈ Dom ( A A ∗ ) , and then T A 2 ( f , g ) = ( A ∗ A f , A A ∗ g ) T_A^2(f,g)=(A^*Af,AA^*g) T A 2 ( f , g ) = ( A ∗ A f , A A ∗ g ) . The operator T A 2 T_A^2 T A 2 is self-adjoint by the functional calculus. As shown in the proof of Theorem 3.10, the adjoint of an orthogonal sum is the orthogonal sum of the adjoints. Hence A ∗ A A^*A A ∗ A and A A ∗ AA^* A A ∗ are self-adjoint; they are nonnegative because ⟨ f , A ∗ A f ⟩ = ∥ A f ∥ 2 \langle f,A^*Af\rangle=\|Af\|^2 ⟨ f , A ∗ A f ⟩ = ∥ A f ∥ 2 . By the functional calculus, ∣ T A ∣ = ( T A 2 ) 1 / 2 = ( A ∗ A ) 1 / 2 ⊕ ( A A ∗ ) 1 / 2 |T_A|=(T_A^2)^{1/2}=(A^*A)^{1/2}\oplus(AA^*)^{1/2} ∣ T A ∣ = ( T A 2 ) 1/2 = ( A ∗ A ) 1/2 ⊕ ( A A ∗ ) 1/2 , and Dom ( ∣ T A ∣ ) = Dom ( T A ) \operatorname{Dom}(|T_A|)=\operatorname{Dom}(T_A) Dom ( ∣ T A ∣ ) = Dom ( T A ) with ∥ ∣ T A ∣ φ ∥ = ∥ T A φ ∥ \||T_A|\varphi\|=\|T_A\varphi\| ∥∣ T A ∣ φ ∥ = ∥ T A φ ∥ , since both sides equal ( ∫ λ 2 d μ φ T A ) 1 / 2 \left(\int\lambda^2\,\mathrm{d}\mu_\varphi^{T_A}\right)^{1/2} ( ∫ λ 2 d μ φ T A ) 1/2 . Applying this to φ = ( f , 0 ) \varphi=(f,0) φ = ( f , 0 ) gives the last assertion of (i).
For (ii), since z 2 ∈ [ 0 , ∞ ) z^2 \in[0,\infty) z 2 ∈ [ 0 , ∞ ) only if z z z is real, the operators Q 0 Q_0 Q 0 and Q 1 Q_1 Q 1 are bounded. Let f ∈ h 0 f \in\mathfrak{h}_0 f ∈ h 0 and φ = Q 0 f ∈ Dom ( A ∗ A ) \varphi=Q_0f \in\operatorname{Dom}(A^*A) φ = Q 0 f ∈ Dom ( A ∗ A ) . Then φ ∈ Dom ( A ) \varphi\in\operatorname{Dom}(A) φ ∈ Dom ( A ) and A φ ∈ Dom ( A ∗ ) A\varphi\in\operatorname{Dom}(A^*) A φ ∈ Dom ( A ∗ ) , which gives ( z φ , A φ ) ∈ Dom ( T A ) (z\varphi,A\varphi) \in\operatorname{Dom}(T_A) ( z φ , A φ ) ∈ Dom ( T A ) and
( T A − z ) ( z φ , A φ ) = ( A ∗ A φ − z 2 φ , z A φ − z A φ ) = ( f , 0 ) . (T_A-z)(z\varphi,A\varphi)=(A^*A\varphi-z^2\varphi,zA\varphi-zA\varphi)=(f,0). ( T A − z ) ( z φ , A φ ) = ( A ∗ A φ − z 2 φ , z A φ − z A φ ) = ( f , 0 ) .
Similarly, for g ∈ h 1 g \in\mathfrak{h}_1 g ∈ h 1 and ψ = Q 1 g ∈ Dom ( A A ∗ ) \psi=Q_1g \in\operatorname{Dom}(AA^*) ψ = Q 1 g ∈ Dom ( A A ∗ ) , we have ( A ∗ ψ , z ψ ) ∈ Dom ( T A ) (A^*\psi,z\psi) \in\operatorname{Dom}(T_A) ( A ∗ ψ , z ψ ) ∈ Dom ( T A ) and ( T A − z ) ( A ∗ ψ , z ψ ) = ( 0 , g ) (T_A-z)(A^*\psi,z\psi)=(0,g) ( T A − z ) ( A ∗ ψ , z ψ ) = ( 0 , g ) . Adding these gives (ii).
For (iii), let ψ = Q 1 g \psi=Q_1g ψ = Q 1 g . Then A ∗ ψ ∈ Dom ( A ) A^*\psi\in\operatorname{Dom}(A) A ∗ ψ ∈ Dom ( A ) , and A A ∗ ψ = g + z 2 ψ AA^*\psi=g+z^2\psi A A ∗ ψ = g + z 2 ψ belongs to Dom ( A ∗ ) \operatorname{Dom}(A^*) Dom ( A ∗ ) , because g g g and ψ \psi ψ do. We conclude that A ∗ ψ ∈ Dom ( A ∗ A ) A^*\psi\in\operatorname{Dom}(A^*A) A ∗ ψ ∈ Dom ( A ∗ A ) and
( A ∗ A − z 2 ) A ∗ ψ = A ∗ ( A A ∗ ψ ) − z 2 A ∗ ψ = A ∗ ( g + z 2 ψ ) − z 2 A ∗ ψ = A ∗ g , (A^*A-z^2)A^*\psi=A^*(AA^*\psi)-z^2A^*\psi=A^*(g+z^2\psi)-z^2A^*\psi=A^*g, ( A ∗ A − z 2 ) A ∗ ψ = A ∗ ( A A ∗ ψ ) − z 2 A ∗ ψ = A ∗ ( g + z 2 ψ ) − z 2 A ∗ ψ = A ∗ g ,
so Q 0 A ∗ g = A ∗ ψ Q_0A^*g=A^*\psi Q 0 A ∗ g = A ∗ ψ . Since A ∗ ψ ∈ Dom ( A ) A^*\psi\in\operatorname{Dom}(A) A ∗ ψ ∈ Dom ( A ) and g ∈ Dom ( A ∗ ) g \in\operatorname{Dom}(A^*) g ∈ Dom ( A ∗ ) , we obtain
⟨ A ∗ g , Q 0 A ∗ g ⟩ = ⟨ A ∗ g , A ∗ ψ ⟩ = ⟨ g , A A ∗ ψ ⟩ = ∥ g ∥ 2 + z 2 ⟨ g , Q 1 g ⟩ = ∥ g ∥ 2 + z ⟨ g , ( T A − z ) − 1 g ⟩ , \begin{aligned}
\langle A^*g,Q_0A^*g\rangle=\langle A^*g,A^*\psi\rangle=\langle g,AA^*\psi\rangle=\|g\|^2+z^2\langle g,Q_1g\rangle\\
&=\|g\|^2+z\langle g,(T_A-z)^{-1}g\rangle,
\end{aligned} ⟨ A ∗ g , Q 0 A ∗ g ⟩ = ⟨ A ∗ g , A ∗ ψ ⟩ = ⟨ g , A A ∗ ψ ⟩ = ∥ g ∥ 2 + z 2 ⟨ g , Q 1 g ⟩ = ∥ g ∥ 2 + z ⟨ g , ( T A − z ) − 1 g ⟩ ,
where in the last step we used (ii).
We now apply this to the tree. For a component C C C of a finite removal, set C 0 = C ∩ V 0 C_0=C\cap V_0 C 0 = C ∩ V 0 and C 1 = C ∩ V 1 C_1=C\cap V_1 C 1 = C ∩ V 1 , identify ℓ 2 ( C ) = ℓ 2 ( C 0 ) ⊕ ℓ 2 ( C 1 ) \ell^2(C)=\ell^2(C_0)\oplus\ell^2(C_1) ℓ 2 ( C ) = ℓ 2 ( C 0 ) ⊕ ℓ 2 ( C 1 ) , and write P 0 P_0 P 0 and P 1 P_1 P 1 for the corresponding projections. For y ∈ C 1 y \in C_1 y ∈ C 1 , let d y d_y d y be the number of neighbors of y y y in C C C . For y ∈ C 1 y \in C_1 y ∈ C 1 , the row vector , the squared row norm and, when U y > 0 U_y > 0 U y > 0 , the row direction are
a y = ∑ x ∈ C , x ∼ y h x y δ x ∈ ℓ 2 ( C 0 ) , U y = ∥ a y ∥ 2 = ∑ x ∈ C , x ∼ y ∣ h x y ∣ 2 , v y = a y U y . (30) a_y = \sum_{x\in C,\,x\sim y} h_{xy}\,\delta_x \in\ell^2(C_0), \qquad U_y = \lVert a_y\rVert^2 = \sum_{x\in C,\,x\sim y} \lvert h_{xy}\rvert^2, \qquad v_y = \frac{a_y}{\sqrt{U_y}}. \tag*{(30)} a y = x ∈ C , x ∼ y ∑ h x y δ x ∈ ℓ 2 ( C 0 ) , U y = ∥ a y ∥ 2 = x ∈ C , x ∼ y ∑ ∣ h x y ∣ 2 , v y = U y a y . (30)
These depend on C C C ; when C = V C = V C = V we have d y = deg ( y ) = N y + 1 d_y = \deg(y) = N_y + 1 d y = deg ( y ) = N y + 1 . Almost surely all edge weights are nonzero, and then U y > 0 U_y > 0 U y > 0 exactly when d y ≥ 1 d_y \ge1 d y ≥ 1 . For y ∈ C 1 y \in C_1 y ∈ C 1 , let H C ( y ) = ⨁ x ∈ C , x ∼ y H C x → y \mathcal{H}_C^{(y)} = \bigoplus_{x\in C,\,x\sim y} \mathcal{H}_{C_{x\to y}} H C ( y ) = ⨁ x ∈ C , x ∼ y H C x → y be the operator on ℓ 2 ( C ∖ { y } ) \ell^2(C\setminus\{y\}) ℓ 2 ( C ∖ { y }) with the vertex y y y deleted; by Lemma 3.5(ii), the branches C x → y C_{x\to y} C x → y are the components of C ∖ { y } C\setminus\{y\} C ∖ { y } . For C = V C = V C = V , this is H ( y ) \mathcal{H}^{(y)} H ( y ) .
Proposition 3.23. On Ω s a \Omega_{\mathrm{sa}} Ω sa , let C C C be a component of a finite removal ( F , D ) (F,D) ( F , D ) ; for instance C = V C = V C = V . Write H C \mathcal{H}_C H C and G C G^C G C for its operator and resolvent.
(i) There exists a unique closed, densely defined operator A A A from Dom ( A ) ⊂ ℓ 2 ( C 0 ) \operatorname{Dom}(A) \subset\ell^2(C_0) Dom ( A ) ⊂ ℓ 2 ( C 0 ) to ℓ 2 ( C 1 ) \ell^2(C_1) ℓ 2 ( C 1 ) such that H C = T A \mathcal{H}_C = T_A H C = T A , that is,
H C = ( 0 A ∗ A 0 ) , Dom ( H C ) = Dom ( A ) ⊕ Dom ( A ∗ ) . \mathcal{H}_C = \begin{pmatrix} 0 & A^* \\ A & 0 \end{pmatrix}, \qquad\operatorname{Dom}(\mathcal{H}_C) = \operatorname{Dom}(A) \oplus\operatorname{Dom}(A^*). H C = ( 0 A A ∗ 0 ) , Dom ( H C ) = Dom ( A ) ⊕ Dom ( A ∗ ) . Its domain is Dom ( A ) = { f ∈ ℓ 2 ( C 0 ) : H C , f o r m f ∈ ℓ 2 ( C 1 ) } \operatorname{Dom}(A) = \{f \in\ell^2(C_0) : \mathcal{H}_{C,\mathrm{form}}f \in\ell^2(C_1)\} Dom ( A ) = { f ∈ ℓ 2 ( C 0 ) : H C , form f ∈ ℓ 2 ( C 1 )} . The space c c ( C 0 ) c_c(C_0) c c ( C 0 ) is a core for A A A , and c c ( C 1 ) c_c(C_1) c c ( C 1 ) is a core for A ∗ A^* A ∗ . Further, H C 2 = A ∗ A ⊕ A A ∗ \mathcal{H}_C^2 = A^*A \oplus AA^* H C 2 = A ∗ A ⊕ A A ∗ , with Dom ( H C 2 ) = Dom ( A ∗ A ) ⊕ Dom ( A A ∗ ) \operatorname{Dom}(\mathcal{H}_C^2) = \operatorname{Dom}(A^*A) \oplus\operatorname{Dom}(AA^*) Dom ( H C 2 ) = Dom ( A ∗ A ) ⊕ Dom ( A A ∗ ) .
(ii) For y ∈ C 1 y \in C_1 y ∈ C 1 , we have δ y ∈ Dom ( A ∗ ) \delta_y \in\operatorname{Dom}(A^*) δ y ∈ Dom ( A ∗ ) and A ∗ δ y = a y A^*\delta_y = a_y A ∗ δ y = a y . For f ∈ Dom ( A ) f \in\operatorname{Dom}(A) f ∈ Dom ( A ) and y ∈ C 1 y \in C_1 y ∈ C 1 , we have ( A f ) ( y ) = ⟨ a y , f ⟩ (Af)(y) = \langle a_y,f\rangle ( A f ) ( y ) = ⟨ a y , f ⟩ . Further,
Dom ( A ) = { f ∈ ℓ 2 ( C 0 ) : ∑ y ∈ C 1 : U y > 0 U y ∣ ⟨ v y , f ⟩ ∣ 2 < ∞ } . \operatorname{Dom}(A) = \left\{f \in\ell^2(C_0) : \sum_{y\in C_1:U_y>0} U_y\lvert\langle v_y,f\rangle\rvert^2 < \infty\right\}. Dom ( A ) = ⎩ ⎨ ⎧ f ∈ ℓ 2 ( C 0 ) : y ∈ C 1 : U y > 0 ∑ U y ∣⟨ v y , f ⟩ ∣ 2 < ∞ ⎭ ⎬ ⎫ . and ∥ A f ∥ 2 = ∑ y ∈ C 1 : U y > 0 U y ∣ ⟨ v y , f ⟩ ∣ 2 \lVert Af\rVert^2 = \sum_{y\in C_1:U_y>0} U_y\lvert\langle v_y,f\rangle\rvert^2 ∥ A f ∥ 2 = ∑ y ∈ C 1 : U y > 0 U y ∣⟨ v y , f ⟩ ∣ 2 for f ∈ Dom ( A ) f \in\operatorname{Dom}(A) f ∈ Dom ( A ) , so the closed quadratic form f ↦ ∥ ( A ∗ A ) 1 / 2 f ∥ 2 f \mapsto\lVert(A^*A)^{1/2}f\rVert^2 f ↦ ∥( A ∗ A ) 1/2 f ∥ 2 of A ∗ A A^*A A ∗ A , with form domain Dom ( A ) \operatorname{Dom}(A) Dom ( A ) , is the sum of the rank-one forms f ↦ U y ∣ ⟨ v y , f ⟩ ∣ 2 f \mapsto U_y\lvert\langle v_y,f\rangle\rvert^2 f ↦ U y ∣⟨ v y , f ⟩ ∣ 2 . The closed linear span of { v y : y ∈ C 1 , U y > 0 } \{v_y : y \in C_1,\ U_y > 0\} { v y : y ∈ C 1 , U y > 0 } is Ran A ∗ = ( ker A ) ⊥ \operatorname{Ran} A^* = (\ker A)^\perp Ran A ∗ = ( ker A ) ⊥ .
(iii) Suppose that ( F , D ) (F,D) ( F , D ) is selected without inspecting the retained marks, with inspected set Q Q Q , and that C C C is determined by G Q \mathcal{G}_Q G Q (for instance, the component containing a fixed word, on the event that this word is in V ∖ F V\setminus F V ∖ F ). Conditionally on G Q \mathcal{G}_Q G Q , the pairs ( U y , v y ) (U_y,v_y) ( U y , v y ) , for y ∈ C 1 y \in C_1 y ∈ C 1 with d y ≥ 1 d_y \ge1 d y ≥ 1 , are independent; the direction v y v_y v y is uniformly distributed on the unit sphere of the d y d_y d y -dimensional space span { δ x : x ∈ C , x ∼ y } \operatorname{span}\{\delta_x : x \in C,\ x\sim y\} span { δ x : x ∈ C , x ∼ y } and is independent of U y U_y U y ; and U y U_y U y has the density
p d Γ ( d ) u d − 1 e − p u 1 ( 0 , ∞ ) ( u ) , d = d y . \frac{p^d}{\Gamma(d)}u^{d-1}e^{-pu}\,1_{(0,\infty)}(u), \qquad d=d_y. Γ ( d ) p d u d − 1 e − p u 1 ( 0 , ∞ ) ( u ) , d = d y . The same holds under P ∞ \mathbb{P}^{\infty} P ∞ . For C = V C = V C = V , the hypothesis holds with Q = ∅ Q = \varnothing Q = ∅ and G Q = G 0 \mathcal{G}_Q = \mathcal{G}_0 G Q = G 0 .
(iv) Let y ∈ C 1 y \in\mathcal{C}_1 y ∈ C 1 with U y > 0 U_y > 0 U y > 0 . The operator A ( y ) = A − ∣ δ y ⟩ ⟨ a y ∣ A^{(y)} = A - |\delta_y\rangle\langle a_y| A ( y ) = A − ∣ δ y ⟩ ⟨ a y ∣ with domain Dom ( A ) \operatorname{Dom}(A) Dom ( A ) is closed, and A ~ ( y ) = 0 ⊕ H C ( y ) \widetilde{A}^{(y)} = 0 \oplus\mathcal{H}_{\mathcal{C}}^{(y)} A ( y ) = 0 ⊕ H C ( y ) with respect to ℓ 2 ( C ) = C δ y ⊕ ℓ 2 ( C ∖ { y } ) \ell^2(\mathcal{C}) = \mathbb{C}\delta_y \oplus\ell^2(\mathcal{C} \setminus\{y\}) ℓ 2 ( C ) = C δ y ⊕ ℓ 2 ( C ∖ { y }) . The square block
K 0 ( y ) = ( A ( y ) ) ∗ A ( y ) = A ∗ A − U y ∣ v y ⟩ ⟨ v y ∣ , Dom ( K 0 ( y ) ) = Dom ( A ∗ A ) , K_0^{(y)} = (A^{(y)})^* A^{(y)} = A^*A - U_y|v_y\rangle\langle v_y|,\qquad\operatorname{Dom}(K_0^{(y)}) = \operatorname{Dom}(A^*A), K 0 ( y ) = ( A ( y ) ) ∗ A ( y ) = A ∗ A − U y ∣ v y ⟩ ⟨ v y ∣ , Dom ( K 0 ( y ) ) = Dom ( A ∗ A ) ,
is nonnegative and self-adjoint, and for z ∈ C ∖ R z \in\mathbb{C} \setminus\mathbb{R} z ∈ C ∖ R ,
P 0 ( H C ( y ) − z ) − 1 P 0 = z ( K 0 ( y ) − z 2 ) − 1 , (31) P_0\left(\mathcal{H}_{\mathcal{C}}^{(y)} - z\right)^{-1}P_0 = z\left(K_0^{(y)} - z^2\right)^{-1}, \tag*{(31)} P 0 ( H C ( y ) − z ) − 1 P 0 = z ( K 0 ( y ) − z 2 ) − 1 , (31)
and
z ⟨ v y , ( K 0 ( y ) − z 2 ) − 1 v y ⟩ = ∑ x ∈ C , x ∼ y ∣ v y ( x ) ∣ 2 Γ C x → y ( z ) . (32) z\langle v_y,(K_0^{(y)} - z^2)^{-1}v_y\rangle= \sum_{x\in\mathcal{C},\,x\sim y}|v_y(x)|^2\Gamma_{\mathcal{C}}^{x\to y}(z). \tag*{(32)} z ⟨ v y , ( K 0 ( y ) − z 2 ) − 1 v y ⟩ = x ∈ C , x ∼ y ∑ ∣ v y ( x ) ∣ 2 Γ C x → y ( z ) . (32)
Under the hypotheses of (iii), let y y y be determined by G Q \mathcal{G}_Q G Q , let R y R_y R y be the set of retained edges which are not incident to y y y , and let G y \mathcal{G}^y G y be the σ \sigma σ -algebra generated by G Q \mathcal{G}_Q G Q , by v y v_y v y , and by the variables w e 1 { e ∈ R y } w_e\mathbf{1}_{\{e\in R_y\}} w e 1 { e ∈ R y } , e ∈ U ∗ e \in\mathcal{U}_* e ∈ U ∗ . Then v y v_y v y , the operator H C ( y ) \mathcal{H}_{\mathcal{C}}^{(y)} H C ( y ) and the operator K 0 ( y ) K_0^{(y)} K 0 ( y ) are determined by G y \mathcal{G}^y G y , and conditionally on G y \mathcal{G}^y G y the squared row norm U y U_y U y has the density in (iii) with d = d y d = d_y d = d y , under P \mathbb{P} P and under P ∞ \mathbb{P}^{\infty} P ∞ .
(v) For z ∈ C ∖ R z \in\mathbb{C} \setminus\mathbb{R} z ∈ C ∖ R , we have P 0 ( H C − z ) − 1 P 0 = z ( A ∗ A − z 2 ) − 1 P_0(\mathcal{H}_{\mathcal{C}} - z)^{-1}P_0 = z(A^*A-z^2)^{-1} P 0 ( H C − z ) − 1 P 0 = z ( A ∗ A − z 2 ) − 1 and P 1 ( H C − z ) − 1 P 1 = z ( A A ∗ − z 2 ) − 1 P_1(\mathcal{H}_{\mathcal{C}} - z)^{-1}P_1 = z(AA^*-z^2)^{-1} P 1 ( H C − z ) − 1 P 1 = z ( A A ∗ − z 2 ) − 1 , and for every y ∈ C 1 y \in\mathcal{C}_1 y ∈ C 1 with U y > 0 U_y > 0 U y > 0 ,
U y ⟨ v y , ( A ∗ A − z 2 ) − 1 v y ⟩ = 1 + z G C y ( z ) . (33) U_y\langle v_y,(A^*A-z^2)^{-1}v_y\rangle= 1+zG_{\mathcal{C}}^y(z). \tag*{(33)} U y ⟨ v y , ( A ∗ A − z 2 ) − 1 v y ⟩ = 1 + z G C y ( z ) . (33)
(vi) For every Borel set B ⊂ ( 0 , ∞ ) B \subset(0,\infty) B ⊂ ( 0 , ∞ ) , with B 2 = { b 2 : b ∈ B } B^2 = \{b^2:b \in B\} B 2 = { b 2 : b ∈ B } ,
1 B ( H C ) = 1 2 ( I + sgn ( H C ) ) 1 B 2 ( H C 2 ) , 1 − B ( H C ) = 1 2 ( I − sgn ( H C ) ) 1 B 2 ( H C 2 ) , \mathbf{1}_B(\mathcal{H}_{\mathcal{C}}) = \frac{1}{2}(I+\operatorname{sgn}(\mathcal{H}_{\mathcal{C}}))\mathbf{1}_{B^2}(\mathcal{H}_{\mathcal{C}}^2),\qquad\mathbf{1}_{-B}(\mathcal{H}_{\mathcal{C}}) = \frac{1}{2}(I-\operatorname{sgn}(\mathcal{H}_{\mathcal{C}}))\mathbf{1}_{B^2}(\mathcal{H}_{\mathcal{C}}^2), 1 B ( H C ) = 2 1 ( I + sgn ( H C )) 1 B 2 ( H C 2 ) , 1 − B ( H C ) = 2 1 ( I − sgn ( H C )) 1 B 2 ( H C 2 ) ,
and 1 B 2 ( H C 2 ) = 1 B 2 ( A ∗ A ) ⊕ 1 B 2 ( A A ∗ ) \mathbf{1}_{B^2}(\mathcal{H}_{\mathcal{C}}^2) = \mathbf{1}_{B^2}(A^*A) \oplus\mathbf{1}_{B^2}(AA^*) 1 B 2 ( H C 2 ) = 1 B 2 ( A ∗ A ) ⊕ 1 B 2 ( A A ∗ ) . Further, H C \mathcal{H}_{\mathcal{C}} H C has pure point spectrum in B ∪ ( − B ) B \cup(-B) B ∪ ( − B ) if and only if both A ∗ A A^*A A ∗ A and A A ∗ AA^* A A ∗ have pure point spectrum in B 2 B^2 B 2 , and the same holds with “pure point” replaced by “purely absolutely continuous”.
(vii) All statements hold with the roles of C 0 \mathcal{C}_0 C 0 and C 1 \mathcal{C}_1 C 1 exchanged and A A A replaced by A ∗ A^* A ∗ ; the rows are then indexed by x ∈ C 0 x \in\mathcal{C}_0 x ∈ C 0 , with a x = A δ x a_x = A\delta_x a x = A δ x .
Proof. The proof proceeds part by part.
Proof of (i). By Proposition 3.21(i), we have J H C J = − H C J\mathcal{H}_{\mathcal{C}}J = -\mathcal{H}_{\mathcal{C}} J H C J = − H C . It follows that J J J maps Dom ( H C ) \operatorname{Dom}(\mathcal{H}_{\mathcal{C}}) Dom ( H C ) onto itself, and so do P 0 = ( I + J ) / 2 P_0 = (I+J)/2 P 0 = ( I + J ) /2 and P 1 = ( I − J ) / 2 P_1 = (I-J)/2 P 1 = ( I − J ) /2 , which gives Dom ( H C ) = D 0 ⊕ D 1 \operatorname{Dom}(\mathcal{H}_{\mathcal{C}}) = \mathcal{D}_0 \oplus\mathcal{D}_1 Dom ( H C ) = D 0 ⊕ D 1 with D i = Dom ( H C ) ∩ ℓ 2 ( C i ) \mathcal{D}_i = \operatorname{Dom}(\mathcal{H}_{\mathcal{C}}) \cap\ell^2(\mathcal{C}_i) D i = Dom ( H C ) ∩ ℓ 2 ( C i ) . For f ∈ D 0 f \in\mathcal{D}_0 f ∈ D 0 , the vector H C f = H C , f o r m f \mathcal{H}_{\mathcal{C}}f = \mathcal{H}_{\mathcal{C},\mathrm{form}}f H C f = H C , form f is supported on C 1 \mathcal{C}_1 C 1 , because every edge joins the two parity classes; similarly, H C \mathcal{H}_{\mathcal{C}} H C maps D 1 \mathcal{D}_1 D 1 into ℓ 2 ( C 0 ) \ell^2(\mathcal{C}_0) ℓ 2 ( C 0 ) . Define A A A on Dom ( A ) = D 0 \operatorname{Dom}(A) = \mathcal{D}_0 Dom ( A ) = D 0 and A ~ \widetilde{A} A on D 1 \mathcal{D}_1 D 1 as the restrictions of H C \mathcal{H}_{\mathcal{C}} H C . Then H C ( f , g ) = ( A ~ g , A f ) \mathcal{H}_{\mathcal{C}}(f,g) = (\widetilde{A}g,Af) H C ( f , g ) = ( A g , A f ) on Dom ( A ) ⊕ Dom ( A ~ ) \operatorname{Dom}(A) \oplus\operatorname{Dom}(\widetilde{A}) Dom ( A ) ⊕ Dom ( A ) . Both operators are densely defined, since they are defined on the finitely supported vectors, and A A A is closed: if f n → f f_n \to f f n → f and A f n → g Af_n \to g A f n → g , then H C f n → g \mathcal{H}_{\mathcal{C}}f_n \to g H C f n → g , and since H C \mathcal{H}_{\mathcal{C}} H C is closed, f ∈ Dom ( H C ) ∩ ℓ 2 ( C 0 ) f \in\operatorname{Dom}(\mathcal{H}_{\mathcal{C}}) \cap\ell^2(\mathcal{C}_0) f ∈ Dom ( H C ) ∩ ℓ 2 ( C 0 ) and A f = g A f = g A f = g . By Lemma 3.22(iv), we have A ~ = A ∗ \widetilde{A}=A^* A = A ∗ , that is, H C = T A \mathcal{H}_C=T_A H C = T A . If H C = T A ′ \mathcal{H}_C=T_{A'} H C = T A ′ for another closed operator A ′ A' A ′ , then Dom ( A ′ ) = { f : ( f , 0 ) ∈ Dom ( H C ) } = Dom ( A ) \operatorname{Dom}(A')=\{f:(f,0)\in\operatorname{Dom}(\mathcal{H}_C)\}=\operatorname{Dom}(A) Dom ( A ′ ) = { f : ( f , 0 ) ∈ Dom ( H C )} = Dom ( A ) and A ′ f = P 1 H C ( f , 0 ) = A f A'f=P_1\mathcal{H}_C(f,0)=Af A ′ f = P 1 H C ( f , 0 ) = A f , which proves uniqueness. The description of Dom ( A ) \operatorname{Dom}(A) Dom ( A ) follows from Theorem 3.10(ii).
The graph norm of H C \mathcal{H}_C H C splits as ∥ ( f , g ) ∥ 2 + ∥ H C ( f , g ) ∥ 2 = ∥ f ∥ 2 + ∥ A f ∥ 2 + ∥ g ∥ 2 + ∥ A ∗ g ∥ 2 \|(f,g)\|^2+\|\mathcal{H}_C(f,g)\|^2=\|f\|^2+\|Af\|^2+\|g\|^2+\|A^*g\|^2 ∥ ( f , g ) ∥ 2 + ∥ H C ( f , g ) ∥ 2 = ∥ f ∥ 2 + ∥ A f ∥ 2 + ∥ g ∥ 2 + ∥ A ∗ g ∥ 2 . Let f ∈ Dom ( A ) f\in\operatorname{Dom}(A) f ∈ Dom ( A ) . Since c c ( C ) c_c(C) c c ( C ) is a core for H C \mathcal{H}_C H C , there exist φ n ∈ c c ( C ) \varphi_n\in c_c(C) φ n ∈ c c ( C ) converging to ( f , 0 ) (f,0) ( f , 0 ) in the graph norm of H C \mathcal{H}_C H C . Then P 0 φ n ∈ c c ( C 0 ) P_0\varphi_n\in c_c(C_0) P 0 φ n ∈ c c ( C 0 ) converges to f f f in the graph norm of A A A , by the splitting. This shows that c c ( C 0 ) c_c(C_0) c c ( C 0 ) is a core for A A A , and the same argument applied to ( 0 , g ) (0,g) ( 0 , g ) proves that c c ( C 1 ) c_c(C_1) c c ( C 1 ) is a core for A ∗ A^* A ∗ . The statement about H C 2 \mathcal{H}_C^2 H C 2 is Lemma 3.22(i).
Proof of (ii). Since δ y ∈ c c ( C 1 ) ⊂ Dom ( A ∗ ) \delta_y\in c_c(C_1)\subset\operatorname{Dom}(A^*) δ y ∈ c c ( C 1 ) ⊂ Dom ( A ∗ ) , we have A ∗ δ y = H C , f o r m δ y = ∑ x ∼ y h x y δ x = a y A^*\delta_y=\mathcal{H}_{C,\mathrm{form}}\delta_y=\sum_{x\sim y}h_{xy}\delta_x=a_y A ∗ δ y = H C , form δ y = ∑ x ∼ y h x y δ x = a y . For f ∈ Dom ( A ) f\in\operatorname{Dom}(A) f ∈ Dom ( A ) , by Theorem 3.10(ii),
( A f ) ( y ) = ( H C , f o r m f ) ( y ) = ∑ x ∼ y h y x f ( x ) = ∑ x ∼ y h x y ‾ f ( x ) = ⟨ a y , f ⟩ . (Af)(y)=(\mathcal{H}_{C,\mathrm{form}}f)(y)=\sum_{x\sim y}h_{yx}f(x)=\sum_{x\sim y}\overline{h_{xy}}f(x)=\langle a_y,f\rangle. ( A f ) ( y ) = ( H C , form f ) ( y ) = x ∼ y ∑ h y x f ( x ) = x ∼ y ∑ h x y f ( x ) = ⟨ a y , f ⟩ .
The same computation applies to every f ∈ ℓ 2 ( C 0 ) f\in\ell^2(C_0) f ∈ ℓ 2 ( C 0 ) , with H C , f o r m \mathcal{H}_{C,\mathrm{form}} H C , form in place of A A A . Together with Theorem 3.10(ii), this implies that f ∈ Dom ( A ) f\in\operatorname{Dom}(A) f ∈ Dom ( A ) exactly when ∑ y ∣ ⟨ a y , f ⟩ ∣ 2 < ∞ \sum_y|\langle a_y,f\rangle|^2<\infty ∑ y ∣ ⟨ a y , f ⟩ ∣ 2 < ∞ , and then ∥ A f ∥ 2 = ∑ y ∣ ⟨ a y , f ⟩ ∣ 2 \|Af\|^2=\sum_y|\langle a_y,f\rangle|^2 ∥ A f ∥ 2 = ∑ y ∣ ⟨ a y , f ⟩ ∣ 2 . The terms with U y = 0 U_y=0 U y = 0 vanish, and the others equal U y ∣ ⟨ v y , f ⟩ ∣ 2 U_y|\langle v_y,f\rangle|^2 U y ∣ ⟨ v y , f ⟩ ∣ 2 . By Lemma 3.22(i), the form domain of A ∗ A A^*A A ∗ A is Dom ( A ) \operatorname{Dom}(A) Dom ( A ) and ∥ ( A ∗ A ) 1 / 2 f ∥ 2 = ∥ A f ∥ 2 \|(A^*A)^{1/2}f\|^2=\|Af\|^2 ∥ ( A ∗ A ) 1/2 f ∥ 2 = ∥ A f ∥ 2 , which gives the form identity.
For the span, f ⊥ Ran A ∗ f\perp\operatorname{Ran}A^* f ⊥ Ran A ∗ means that ⟨ A ∗ g , f ⟩ = 0 \langle A^*g,f\rangle=0 ⟨ A ∗ g , f ⟩ = 0 for all g ∈ Dom ( A ∗ ) g\in\operatorname{Dom}(A^*) g ∈ Dom ( A ∗ ) , which holds exactly when f ∈ Dom ( A ∗ ∗ ) = Dom ( A ) f\in\operatorname{Dom}(A^{**})=\operatorname{Dom}(A) f ∈ Dom ( A ∗∗ ) = Dom ( A ) and A f = 0 Af=0 A f = 0 . We conclude that ( Ran A ∗ ) ⊥ = ker A (\operatorname{Ran}A^*)^\perp=\ker A ( Ran A ∗ ) ⊥ = ker A and Ran A ∗ = ( ker A ) ⊥ \operatorname{Ran}A^*=(\ker A)^\perp Ran A ∗ = ( ker A ) ⊥ . Since c c ( C 1 ) c_c(C_1) c c ( C 1 ) is a core for A ∗ A^* A ∗ , the set A ∗ c c ( C 1 ) = span { a y : y ∈ C 1 } A^*c_c(C_1)=\operatorname{span}\{a_y:y\in C_1\} A ∗ c c ( C 1 ) = span { a y : y ∈ C 1 } is dense in Ran A ∗ \operatorname{Ran}A^* Ran A ∗ . The nonzero vectors among the a y a_y a y are multiples of the v y v_y v y with U y > 0 U_y>0 U y > 0 .
Proof of (iii). By Lemma 3.4, conditionally on G Q \mathcal{G}_Q G Q , the marks of the retained edges are independent standard complex Gaussians. Fix y ∈ C 1 y\in C_1 y ∈ C 1 with d y = d ≥ 1 d_y=d\ge1 d y = d ≥ 1 . The coordinates of p a y \sqrt{p}\,a_y p a y are h x y p h_{xy}\sqrt{p} h x y p for the d d d neighbors x x x of y y y in C C C ; each of them is the mark of a retained edge or its complex conjugate. Since complex conjugation preserves the standard complex Gaussian law, the vector ζ = p a y ∈ C d \zeta=\sqrt{p}\,a_y\in\mathbb{C}^d ζ = p a y ∈ C d has, conditionally on G Q \mathcal{G}_Q G Q , the density π − d e − ∣ ζ ∣ 2 \pi^{-d}e^{-|\zeta|^2} π − d e − ∣ ζ ∣ 2 on C d ≅ R 2 d \mathbb{C}^d\cong\mathbb{R}^{2d} C d ≅ R 2 d . In polar coordinates ζ = r ϑ \zeta=r\vartheta ζ = r ϑ , with r > 0 r>0 r > 0 and ϑ \vartheta ϑ in the unit sphere S 2 d − 1 S^{2d-1} S 2 d − 1 , Lebesgue measure is r 2 d − 1 d r σ ( d ϑ ) r^{2d-1}\,dr\,\sigma(d\vartheta) r 2 d − 1 d r σ ( d ϑ ) , where σ \sigma σ is the surface measure, of total mass 2 π d / Γ ( d ) 2\pi^d/\Gamma(d) 2 π d /Γ ( d ) . Then ( r , ϑ ) (r,\vartheta) ( r , ϑ ) has the product density π − d e − r 2 r 2 d − 1 d r σ ( d ϑ ) \pi^{-d}e^{-r^2}r^{2d-1}\,dr\,\sigma(d\vartheta) π − d e − r 2 r 2 d − 1 d r σ ( d ϑ ) ; in particular, ϑ \vartheta ϑ is uniform on the sphere and independent of r r r , and r r r has the density 2 Γ ( d ) − 1 r 2 d − 1 e − r 2 2\Gamma(d)^{-1}r^{2d-1}e^{-r^2} 2Γ ( d ) − 1 r 2 d − 1 e − r 2 . The change of variables u = r 2 / p u=r^2/p u = r 2 / p shows that U y = ∥ a y ∥ 2 = r 2 / p U_y=\|a_y\|^2=r^2/p U y = ∥ a y ∥ 2 = r 2 / p has the density in (iii), and v y = ζ / ∣ ζ ∣ = ϑ v_y=\zeta/|\zeta|=\vartheta v y = ζ /∣ ζ ∣ = ϑ . Since each edge has exactly one endpoint in V 1 V_1 V 1 , the vectors a y a_y a y for distinct y ∈ C 1 y\in C_1 y ∈ C 1 are built from disjoint sets of retained marks, and they are independent conditionally on G Q \mathcal{G}_Q G Q . The statement under P ∞ \mathbb{P}^{\infty} P ∞ follows in the same way from the P ∞ \mathbb{P}^{\infty} P ∞ version of Lemma 3.4.
Proof of (iv). By (3.9), we have H C = ( 0 ⊕ H C ( y ) ) + W y \mathcal{H}_{C} = (0 \oplus\mathcal{H}_{C}^{(y)}) + W_y H C = ( 0 ⊕ H C ( y ) ) + W y , where
W y = ∑ x ∈ C , x ∼ y ( h y x ∣ δ y ⟩ ⟨ δ x ∣ + h x y ∣ δ x ⟩ ⟨ δ y ∣ ) = ∣ δ y ⟩ ⟨ a y ∣ + ∣ a y ⟩ ⟨ δ y ∣ . W_y = \sum_{\substack{x \in C, x \sim y}} \left(h_{yx} |\delta_y\rangle\langle\delta_x| + h_{xy} |\delta_x\rangle\langle\delta_y|\right) = |\delta_y\rangle\langle a_y| + |a_y\rangle\langle\delta_y|. W y = x ∈ C , x ∼ y ∑ ( h y x ∣ δ y ⟩ ⟨ δ x ∣ + h x y ∣ δ x ⟩ ⟨ δ y ∣ ) = ∣ δ y ⟩ ⟨ a y ∣ + ∣ a y ⟩ ⟨ δ y ∣.
and the domains agree. With respect to ℓ 2 ( C 0 ) ⊕ ℓ 2 ( C 1 ) \ell^2(C_0) \oplus\ell^2(C_1) ℓ 2 ( C 0 ) ⊕ ℓ 2 ( C 1 ) , the operator W y W_y W y maps ( f , g ) (f,g) ( f , g ) to ( g ( y ) a y , ⟨ a y , f ⟩ δ y ) (g(y)a_y,\langle a_y,f\rangle\delta_y) ( g ( y ) a y , ⟨ a y , f ⟩ δ y ) . Since ∣ δ y ⟩ ⟨ a y ∣ |\delta_y\rangle\langle a_y| ∣ δ y ⟩ ⟨ a y ∣ is bounded, the operator A ( y ) A^{(y)} A ( y ) is closed with domain Dom ( A ) \operatorname{Dom}(A) Dom ( A ) , and its adjoint is ( A ( y ) ) ∗ = A ∗ − ∣ a y ⟩ ⟨ δ y ∣ (A^{(y)})^* = A^* - |a_y\rangle\langle\delta_y| ( A ( y ) ) ∗ = A ∗ − ∣ a y ⟩ ⟨ δ y ∣ with domain Dom ( A ∗ ) \operatorname{Dom}(A^*) Dom ( A ∗ ) ; here we used that the adjoint of the sum of a densely defined operator and a bounded everywhere defined operator is the sum of the adjoints. We conclude that
0 ⊕ H C ( y ) = H C − W y = ( 0 A ∗ − ∣ a y ⟩ ⟨ δ y ∣ A − ∣ δ y ⟩ ⟨ a y ∣ 0 ) = T A ( y ) . 0 \oplus\mathcal{H}_{C}^{(y)} = \mathcal{H}_{C} - W_y = \begin{pmatrix} 0 & A^* - |a_y\rangle\langle\delta_y| \\ A - |\delta_y\rangle\langle a_y| & 0 \end{pmatrix} = T_{A^{(y)}}. 0 ⊕ H C ( y ) = H C − W y = ( 0 A − ∣ δ y ⟩ ⟨ a y ∣ A ∗ − ∣ a y ⟩ ⟨ δ y ∣ 0 ) = T A ( y ) .
By Lemma 3.22(i), K 0 ( y ) K_0^{(y)} K 0 ( y ) is nonnegative and self-adjoint. A vector f ∈ Dom ( A ) f \in\operatorname{Dom}(A) f ∈ Dom ( A ) belongs to Dom ( K 0 ( y ) ) \operatorname{Dom}(K_0^{(y)}) Dom ( K 0 ( y ) ) exactly when A ( y ) f = A f − ⟨ a y , f ⟩ δ y A^{(y)}f = Af - \langle a_y,f\rangle\delta_y A ( y ) f = A f − ⟨ a y , f ⟩ δ y belongs to Dom ( A ∗ ) \operatorname{Dom}(A^*) Dom ( A ∗ ) , and since δ y ∈ Dom ( A ∗ ) \delta_y \in\operatorname{Dom}(A^*) δ y ∈ Dom ( A ∗ ) , this holds exactly when A f ∈ Dom ( A ∗ ) Af \in\operatorname{Dom}(A^*) A f ∈ Dom ( A ∗ ) . This proves that Dom ( K 0 ( y ) ) = Dom ( A ∗ A ) \operatorname{Dom}(K_0^{(y)}) = \operatorname{Dom}(A^*A) Dom ( K 0 ( y ) ) = Dom ( A ∗ A ) , and for such f f f , using ⟨ δ y , A f ⟩ = ⟨ a y , f ⟩ \langle\delta_y,Af\rangle= \langle a_y,f\rangle ⟨ δ y , A f ⟩ = ⟨ a y , f ⟩ ,
K 0 ( y ) f = ( A ∗ − ∣ a y ⟩ ⟨ δ y ∣ ) ( A f − ⟨ a y , f ⟩ δ y ) = A ∗ A f − ⟨ a y , f ⟩ a y = A ∗ A f − U y ⟨ v y , f ⟩ v y . K_0^{(y)}f = (A^* - |a_y\rangle\langle\delta_y|)(Af - \langle a_y,f\rangle\delta_y) = A^*Af - \langle a_y,f\rangle a_y = A^*Af - U_y\langle v_y,f\rangle v_y. K 0 ( y ) f = ( A ∗ − ∣ a y ⟩ ⟨ δ y ∣ ) ( A f − ⟨ a y , f ⟩ δ y ) = A ∗ A f − ⟨ a y , f ⟩ a y = A ∗ A f − U y ⟨ v y , f ⟩ v y .
By Lemma 3.22(ii) applied to A ( y ) A^{(y)} A ( y ) , we have P 0 ( T A ( y ) − z ) − 1 P 0 = z ( K 0 ( y ) − z 2 ) − 1 P_0(T_{A^{(y)}} - z)^{-1}P_0 = z(K_0^{(y)} - z^2)^{-1} P 0 ( T A ( y ) − z ) − 1 P 0 = z ( K 0 ( y ) − z 2 ) − 1 . Since ( T A ( y ) − z ) − 1 = ( − z ) − 1 ⊕ ( H C ( y ) − z ) − 1 (T_{A^{(y)}} - z)^{-1} = (-z)^{-1} \oplus(\mathcal{H}_{C}^{(y)} - z)^{-1} ( T A ( y ) − z ) − 1 = ( − z ) − 1 ⊕ ( H C ( y ) − z ) − 1 and ℓ 2 ( C 0 ) ⊂ ℓ 2 ( C ∖ { y } ) \ell^2(C_0) \subset\ell^2(C \setminus\{y\}) ℓ 2 ( C 0 ) ⊂ ℓ 2 ( C ∖ { y }) , this is (3.21). For (3.22), the neighbors of y y y in C C C lie in distinct branches C x → y C_{x\to y} C x → y , which are distinct components of C ∖ { y } C \setminus\{y\} C ∖ { y } . We obtain
⟨ v y , ( H C ( y ) − z ) − 1 v y ⟩ = ∑ x ∼ y ∣ v y ( x ) ∣ 2 ⟨ δ x , ( H C x → y − z ) − 1 δ x ⟩ = ∑ x ∼ y ∣ v y ( x ) ∣ 2 Γ C x → y ( z ) . \langle v_y,(\mathcal{H}_{C}^{(y)} - z)^{-1}v_y\rangle= \sum_{x\sim y}|v_y(x)|^2\langle\delta_x,(\mathcal{H}_{C_{x\to y}} - z)^{-1}\delta_x\rangle= \sum_{x\sim y}|v_y(x)|^2\Gamma_{C_{x\to y}}(z). ⟨ v y , ( H C ( y ) − z ) − 1 v y ⟩ = x ∼ y ∑ ∣ v y ( x ) ∣ 2 ⟨ δ x , ( H C x → y − z ) − 1 δ x ⟩ = x ∼ y ∑ ∣ v y ( x ) ∣ 2 Γ C x → y ( z ) .
For the last assertion, since the components C x → y C_{x\to y} C x → y are determined by G Q \mathcal{G}_Q G Q and the weights of their edges are the marks in R y R_y R y , the operator H C ( y ) \mathcal{H}_{C}^{(y)} H C ( y ) is determined by G y \mathcal{G}^y G y ; so is K 0 ( y ) K_0^{(y)} K 0 ( y ) , which by Lemma 3.22(i) is the restriction of ( 0 ⊕ H C ( y ) ) 2 (0 \oplus\mathcal{H}_{C}^{(y)})^2 ( 0 ⊕ H C ( y ) ) 2 to ℓ 2 ( C 0 ) \ell^2(C_0) ℓ 2 ( C 0 ) . Conditionally on G Q \mathcal{G}_Q G Q , the marks of the d y d_y d y retained edges at y y y are independent of the marks in R y R_y R y , and by the proof of (iii) the squared row norm U y U_y U y is independent of the direction v y v_y v y , with the stated density. Therefore, conditionally on G y \mathcal{G}^y G y , the squared row norm U y U_y U y has the stated density. Since the conditional statements given G Q \mathcal{G}_Q G Q hold under P ∞ \mathbb{P}^{\infty} P ∞ as well, by Lemma 3.4, the same argument applies under P ∞ \mathbb{P}^{\infty} P ∞ .
Proof of (v). The first two identities are Lemma 3.22(ii). For the third, we apply Lemma 3.22(iii) with g = δ y g = \delta_y g = δ y , using A ∗ δ y = a y = U y v y A^*\delta_y = a_y = \sqrt{U_y}v_y A ∗ δ y = a y = U y v y , ∥ δ y ∥ = 1 \|\delta_y\| = 1 ∥ δ y ∥ = 1 and ⟨ δ y , ( H C − z ) − 1 δ y ⟩ = G y y C ( z ) \langle\delta_y,(\mathcal{H}_{C} - z)^{-1}\delta_y\rangle= G^C_{yy}(z) ⟨ δ y , ( H C − z ) − 1 δ y ⟩ = G y y C ( z ) . Proof of (vi). Since the square map s ( λ ) = λ 2 s(\lambda)=\lambda^2 s ( λ ) = λ 2 is a homeomorphism of ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , the set B 2 B^2 B 2 is Borel. For every λ ∈ R \lambda\in\mathbb{R} λ ∈ R , with sgn ( 0 ) = 0 \operatorname{sgn}(0)=0 sgn ( 0 ) = 0 , we have
1 B ( λ ) = 1 B 2 ( λ 2 ) 1 2 ( 1 + sgn λ ) , 1 − B ( λ ) = 1 B 2 ( λ 2 ) 1 2 ( 1 − sgn λ ) . \mathbb{1}_B(\lambda)=\mathbb{1}_{B^2}(\lambda^2)\frac{1}{2}(1+\operatorname{sgn}\lambda),\qquad\mathbb{1}_{-B}(\lambda)=\mathbb{1}_{B^2}(\lambda^2)\frac{1}{2}(1-\operatorname{sgn}\lambda). 1 B ( λ ) = 1 B 2 ( λ 2 ) 2 1 ( 1 + sgn λ ) , 1 − B ( λ ) = 1 B 2 ( λ 2 ) 2 1 ( 1 − sgn λ ) .
Indeed, for λ > 0 \lambda>0 λ > 0 we have λ 2 ∈ B 2 \lambda^2\in B^2 λ 2 ∈ B 2 exactly when λ ∈ B \lambda\in B λ ∈ B ; for λ < 0 \lambda<0 λ < 0 the first right side vanishes, and λ 2 ∈ B 2 \lambda^2\in B^2 λ 2 ∈ B 2 exactly when − λ ∈ B -\lambda\in B − λ ∈ B ; and 0 ∉ B 2 0\notin B^2 0 ∈ / B 2 . The functional calculus is multiplicative and satisfies ϕ ( H C 2 ) = ( ϕ ∘ s ) ( H C ) \phi(\mathcal{H}_C^2)=(\phi\circ s)(\mathcal{H}_C) ϕ ( H C 2 ) = ( ϕ ∘ s ) ( H C ) , which gives the two identities. Since H C 2 = A ∗ A ⊕ A A ∗ \mathcal{H}_C^2=A^*A\oplus AA^* H C 2 = A ∗ A ⊕ A A ∗ is an orthogonal sum of self-adjoint operators, its functional calculus acts blockwise.
For the spectral types, let φ = ( f 0 , f 1 ) ∈ ℓ 2 ( C ) \varphi=(f_0,f_1)\in\ell^2(C) φ = ( f 0 , f 1 ) ∈ ℓ 2 ( C ) . The same composition rule shows that the spectral measure of φ \varphi φ for H C 2 \mathcal{H}_C^2 H C 2 is the image of μ φ H C \mu_\varphi^{\mathcal{H}_C} μ φ H C under s s s , and the block structure shows that it equals μ f 0 A ∗ A + μ f 1 A A ∗ \mu_{f_0}^{A^*A}+\mu_{f_1}^{AA^*} μ f 0 A ∗ A + μ f 1 A A ∗ . The restriction of μ φ H C 2 \mu_\varphi^{\mathcal{H}_C^2} μ φ H C 2 to B 2 B^2 B 2 is the image under s s s of the restriction of μ φ H C \mu_\varphi^{\mathcal{H}_C} μ φ H C to B ∪ ( − B ) B\cup(-B) B ∪ ( − B ) . The map s s s sends countable sets to countable sets, and the preimage under s s s of a countable subset of B 2 B^2 B 2 is countable. Further, s s s and its local inverses ± ⋅ \pm\sqrt{\cdot} ± ⋅ are locally Lipschitz on R ∖ { 0 } \mathbb{R}\setminus\{0\} R ∖ { 0 } and ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , and therefore map Lebesgue null sets to Lebesgue null sets. We conclude that the restriction of μ φ H C \mu_\varphi^{\mathcal{H}_C} μ φ H C to B ∪ ( − B ) B\cup(-B) B ∪ ( − B ) is purely atomic (respectively absolutely continuous) if and only if the restriction of μ φ H C 2 \mu_\varphi^{\mathcal{H}_C^2} μ φ H C 2 to B 2 B^2 B 2 is. Taking φ = ( f 0 , 0 ) \varphi=(f_0,0) φ = ( f 0 , 0 ) and φ = ( 0 , f 1 ) \varphi=(0,f_1) φ = ( 0 , f 1 ) , and using that a sum of two purely atomic (absolutely continuous) measures is purely atomic (absolutely continuous), we obtain the last assertion.
Proof of (vii). With respect to ℓ 2 ( C 1 ) ⊕ ℓ 2 ( C 0 ) \ell^2(C_1)\oplus\ell^2(C_0) ℓ 2 ( C 1 ) ⊕ ℓ 2 ( C 0 ) , we have H C ( g , f ) = ( A f , A ∗ g ) = T A ∗ ( g , f ) \mathcal{H}_C(g,f)=(Af,A^*g)=T_{A^*}(g,f) H C ( g , f ) = ( A f , A ∗ g ) = T A ∗ ( g , f ) , since A ∗ ∗ = A A^{**}=A A ∗∗ = A . All the arguments above apply to A ∗ A^* A ∗ in place of A A A .
Compactly supported eigenvectors
Let C C C be a component of a finite removal. A finitely supported vector ψ ∈ c c ( C ) \psi\in c_c(C) ψ ∈ c c ( C ) lies in the domain of H C \mathcal{H}_C H C and satisfies H C ψ = H C , f o r m ψ \mathcal{H}_C\psi=\mathcal{H}_{C,\mathrm{form}}\psi H C ψ = H C , form ψ . In particular, ψ \psi ψ is an eigenvector of H C \mathcal{H}_C H C with eigenvalue E E E exactly when ( H C , f o r m ψ ) ( v ) = E ψ ( v ) (\mathcal{H}_{C,\mathrm{form}}\psi)(v)=E\psi(v) ( H C , form ψ ) ( v ) = E ψ ( v ) for every v ∈ C v\in C v ∈ C . In this section, we show that such eigenvectors do not exist for E ≠ 0 E\ne0 E = 0 on infinite components. For E = 0 E=0 E = 0 , they occur infinitely often on the surviving tree. For a finite set T T T of vertices, we write M T = ( h u v ) u , v ∈ T M_T=(h_{uv})_{u,v\in T} M T = ( h uv ) u , v ∈ T for the corresponding finite Hermitian matrix.
Proposition 3.24.
(i) Let T ′ T' T ′ be a deterministic countable, infinite, connected, locally finite tree with vertex set V ′ V' V ′ . Let ( h e ) e (h_e)_e ( h e ) e be independent complex random variables indexed by the edges of T ′ T' T ′ , each with a law that is absolutely continuous with respect to Lebesgue measure on C \mathbb{C} C , and set h v u = h u v ‾ h_{vu}=\overline{h_{uv}} h v u = h uv , with an arbitrary fixed orientation of each edge (for example, independent nondegenerate complex Gaussian variables). Then almost surely there exist no E ∈ R ∖ { 0 } E\in\mathbb{R}\setminus\{0\} E ∈ R ∖ { 0 } and no nonzero finitely supported ψ : V ′ → C \psi:V'\to\mathbb{C} ψ : V ′ → C with ∑ u ∼ v h v u ψ ( u ) = E ψ ( v ) \sum_{u\sim v}h_{vu}\psi(u)=E\psi(v) ∑ u ∼ v h v u ψ ( u ) = E ψ ( v ) for every v ∈ V ′ v\in V' v ∈ V ′ .
(ii) Let p > 0 p > 0 p > 0 . There exists an event Ω c m \Omega_{\mathrm{cm}} Ω cm with P ( Ω c m ) = 1 \mathbb{P}(\Omega_{\mathrm{cm}}) = 1 P ( Ω cm ) = 1 such that on Ω s a ∩ Ω c m \Omega_{\mathrm{sa}} \cap\Omega_{\mathrm{cm}} Ω sa ∩ Ω cm , for every finite removal of T \mathcal{T} T and every infinite component C C C of it, the operator H C \mathcal{H}_C H C has no finitely supported eigenvector with a nonzero eigenvalue. In particular, if p > 1 p > 1 p > 1 , then P ∞ \mathbb{P}^{\infty} P ∞ -almost surely H \mathcal{H} H has no finitely supported eigenvector with a nonzero eigenvalue.
The restriction to infinite components cannot be removed: on the extinction event S c S^{c} S c , which has probability 1 − ϱ p > 0 1-\varrho_p > 0 1 − ϱ p > 0 , the operator H \mathcal{H} H is a finite Hermitian matrix, and it has nonzero eigenvalues as soon as T \mathcal{T} T has an edge. The conclusion for H \mathcal{H} H holds P \mathbb{P} P -almost surely on S S S , but not P \mathbb{P} P -almost surely.
Proof. The proof shows that a finitely supported eigenvector produces two disjoint finite subtrees whose matrices share a nonzero eigenvalue, and that this has probability zero for each fixed pair.
Step 1. Two disjoint subtrees with a common eigenvalue. We work on the event that all weights are nonzero, which has probability one. Suppose that E ≠ 0 E \ne0 E = 0 and ψ ≠ 0 \psi\ne0 ψ = 0 are as in (i), and let S = supp ψ S=\operatorname{supp}\psi S = supp ψ , a finite nonempty set. Since T ′ \mathcal{T}' T ′ is infinite and connected, there exists an edge { a , b } \{a,b\} { a , b } with a ∉ S a \notin S a ∈ / S and b ∈ S b \in S b ∈ S . The equation at a a a reads 0 = E ψ ( a ) = ∑ u ∼ a h a u ψ ( u ) 0=E\psi(a)=\sum_{u\sim a}h_{au}\psi(u) 0 = E ψ ( a ) = ∑ u ∼ a h a u ψ ( u ) . Since the term h a b ψ ( b ) h_{ab}\psi(b) h ab ψ ( b ) is nonzero, there exists a second neighbor b ′ ≠ b b'\ne b b ′ = b of a a a with ψ ( b ′ ) ≠ 0 \psi(b')\ne0 ψ ( b ′ ) = 0 . Let T 1 T_1 T 1 and T 2 T_2 T 2 be the vertex sets of the connected components of the subgraph of T ′ \mathcal{T}' T ′ induced on S S S which contain b b b and b ′ b' b ′ , respectively. Both sets are connected and do not contain a a a , so in the graph obtained from T ′ \mathcal{T}' T ′ by removing a a a , the set T 1 T_1 T 1 lies in the component of b b b and T 2 T_2 T 2 lies in the component of b ′ b' b ′ . These components are different, because the only path from b b b to b ′ b' b ′ in T ′ \mathcal{T}' T ′ is b , a , b ′ b,a,b' b , a , b ′ . In particular, T 1 ∩ T 2 = ∅ T_1\cap T_2=\varnothing T 1 ∩ T 2 = ∅ . For v ∈ T i v\in T_i v ∈ T i , every neighbor u u u of v v v with ψ ( u ) ≠ 0 \psi(u)\ne0 ψ ( u ) = 0 lies in S S S , and therefore in T i T_i T i . It follows that
( M T i ψ ∣ T i ) ( v ) = ∑ u ∈ T i h v u ψ ( u ) = ∑ u ∼ v h v u ψ ( u ) = E ψ ( v ) , v ∈ T i , (M_{T_i}\psi|_{T_i})(v)=\sum_{u\in T_i}h_{vu}\psi(u)=\sum_{u\sim v}h_{vu}\psi(u)=E\psi(v),\qquad v\in T_i, ( M T i ψ ∣ T i ) ( v ) = u ∈ T i ∑ h v u ψ ( u ) = u ∼ v ∑ h v u ψ ( u ) = E ψ ( v ) , v ∈ T i ,
and ψ ∣ T i ≠ 0 \psi|_{T_i}\ne0 ψ ∣ T i = 0 . We conclude that E E E is a common nonzero eigenvalue of M T 1 M_{T_1} M T 1 and M T 2 M_{T_2} M T 2 , where T 1 T_1 T 1 and T 2 T_2 T 2 are disjoint finite sets which induce connected subgraphs.
Step 2. A fixed pair has no common nonzero eigenvalue. Fix disjoint finite sets T 1 , T 2 ⊂ V ′ T_1,T_2\subset V' T 1 , T 2 ⊂ V ′ which induce connected subgraphs. We show that, almost surely, M T 1 M_{T_1} M T 1 and M T 2 M_{T_2} M T 2 have no common nonzero eigenvalue. (The event is Borel, because the eigenvalues of a Hermitian matrix, listed in increasing order, depend continuously on its entries.) If T 2 T_2 T 2 is a single vertex, then M T 2 = 0 M_{T_2}=0 M T 2 = 0 . Otherwise, let e 1 , … , e k e_1,\ldots,e_k e 1 , … , e k , with k ≥ 1 k\ge1 k ≥ 1 , be the edges with both endpoints in T 2 T_2 T 2 , and let ζ = ( h e 1 , … , h e k ) ∈ C k \zeta=(h_{e_1},\ldots,h_{e_k})\in\mathbb{C}^k ζ = ( h e 1 , … , h e k ) ∈ C k . Since T 1 ∩ T 2 = ∅ T_1\cap T_2=\varnothing T 1 ∩ T 2 = ∅ , the vector ζ \zeta ζ is independent of the weights X 1 X_1 X 1 of the edges inside T 1 T_1 T 1 , and it has a density ρ \rho ρ on C k ≅ R 2 k \mathbb{C}^k\cong\mathbb{R}^{2k} C k ≅ R 2 k , the product of the densities of its coordinates. We write ζ = r ϑ \zeta=r\vartheta ζ = r ϑ with r = ∣ ζ ∣ r=|\zeta| r = ∣ ζ ∣ and ϑ ∈ S 2 k − 1 \vartheta\in S^{2k-1} ϑ ∈ S 2 k − 1 . Then M T 2 = r M ( ϑ ) M_{T_2}=rM(\vartheta) M T 2 = r M ( ϑ ) , where M ( ϑ ) M(\vartheta) M ( ϑ ) is the Hermitian matrix built in the same way from ϑ \vartheta ϑ , and the nonzero eigenvalues of M T 2 M_{T_2} M T 2 are the numbers r λ r\lambda r λ with λ \lambda λ a nonzero eigenvalue of M ( ϑ ) M(\vartheta) M ( ϑ ) . Let Σ ( ϑ , X 1 ) \Sigma(\vartheta, X_1) Σ ( ϑ , X 1 ) be the finite set of quotients κ / λ \kappa/\lambda κ / λ , with κ \kappa κ a nonzero eigenvalue of M T 1 M_{T_1} M T 1 and λ \lambda λ a nonzero eigenvalue of M ( ϑ ) M(\vartheta) M ( ϑ ) . A common nonzero eigenvalue exists only if r ∈ Σ ( ϑ , X 1 ) r \in\Sigma(\vartheta, X_1) r ∈ Σ ( ϑ , X 1 ) . In polar coordinates, the law of ( r , ϑ ) (r,\vartheta) ( r , ϑ ) has the density ρ ( r ϑ ) r 2 k − 1 \rho(r\vartheta)r^{2k-1} ρ ( r ϑ ) r 2 k − 1 with respect to d r σ ( d ϑ ) dr\,\sigma(d\vartheta) d r σ ( d ϑ ) . By Fubini’s theorem and the independence of ζ \zeta ζ and X 1 X_1 X 1 , we have
P ( r ∈ Σ ( ϑ , X 1 ) ) = E [ ∫ S 2 k − 1 ∫ 0 ∞ 1 { r ∈ Σ ( ϑ , X 1 ) } ρ ( r ϑ ) r 2 k − 1 d r σ ( d ϑ ) ] = 0 , \mathbb{P}(r \in\Sigma(\vartheta,X_1))=\mathbb{E}\left[\int_{\mathbb{S}^{2k-1}}\int_0^\infty\mathbf{1}_{\{r\in\Sigma(\vartheta,X_1)\}}\rho(r\vartheta)r^{2k-1}\,dr\,\sigma(d\vartheta)\right]=0, P ( r ∈ Σ ( ϑ , X 1 )) = E [ ∫ S 2 k − 1 ∫ 0 ∞ 1 { r ∈ Σ ( ϑ , X 1 )} ρ ( r ϑ ) r 2 k − 1 d r σ ( d ϑ ) ] = 0 ,
where the expectation is over X 1 X_1 X 1 . The inner integral vanishes because Σ ( ϑ , X 1 ) \Sigma(\vartheta, X_1) Σ ( ϑ , X 1 ) is a finite set.
Step 3. Conclusion. Since there are countably many pairs of finite subsets of V ′ V' V ′ , almost surely no pair as in Step 2 has a common nonzero eigenvalue. By Step 1, this proves (i).
For (ii), for disjoint finite sets T 1 , T 2 ⊂ U T_1,T_2\subset U T 1 , T 2 ⊂ U , let A ( T 1 , T 2 ) A(T_1,T_2) A ( T 1 , T 2 ) be the event that T 1 , T 2 ⊂ V T_1,T_2\subset V T 1 , T 2 ⊂ V , that both induce connected subgraphs of T \mathcal{T} T , and that M T 1 M_{T_1} M T 1 and M T 2 M_{T_2} M T 2 have a common nonzero eigenvalue. The first two conditions are G 0 \mathcal{G}_0 G 0 -measurable. Conditionally on G 0 \mathcal{G}_0 G 0 , the marks are independent standard complex Gaussians. In particular, the weights of the edges inside T 1 T_1 T 1 and inside T 2 T_2 T 2 are independent and have densities, and Step 2 implies that P ( A ( T 1 , T 2 ) ∣ G 0 ) = 0 \mathbb{P}(A(T_1,T_2)\mid\mathcal{G}_0)=0 P ( A ( T 1 , T 2 ) ∣ G 0 ) = 0 . Let Ω c m \Omega_{\mathrm{cm}} Ω cm be the event that all marks are nonzero and that no event A ( T 1 , T 2 ) A(T_1,T_2) A ( T 1 , T 2 ) occurs; then P ( Ω c m ) = 1 \mathbb{P}(\Omega_{\mathrm{cm}})=1 P ( Ω cm ) = 1 . On Ω s a ∩ Ω c m \Omega_{\mathrm{sa}}\cap\Omega_{\mathrm{cm}} Ω sa ∩ Ω cm , let C C C be an infinite component of a finite removal, and suppose that H C \mathcal{H}_C H C has a finitely supported eigenvector with eigenvalue E ≠ 0 E\ne0 E = 0 . The network ( C , h ) (C,h) ( C , h ) is an infinite, connected, locally finite tree with nonzero weights. Step 1 applies to it and produces disjoint finite sets T 1 , T 2 ⊂ C T_1,T_2\subset C T 1 , T 2 ⊂ C which induce connected subgraphs of C C C and whose matrices share the eigenvalue E E E . By Lemma 3.5(i), C C C is an induced subtree of T \mathcal{T} T , and T 1 T_1 T 1 and T 2 T_2 T 2 therefore induce connected subgraphs of T \mathcal{T} T with the same matrices, and A ( T 1 , T 2 ) A(T_1,T_2) A ( T 1 , T 2 ) occurs. This contradiction proves (ii). The last assertion is the case of the empty finite removal on S S S , where C = V C=V C = V is infinite.
At energy zero, the situation is different. Call a vertex v ∈ V v\in V v ∈ V a twin-leaf vertex if N v = 2 N_v=2 N v = 2 and N v 1 = N v 2 = 0 N_{v1}=N_{v2}=0 N v 1 = N v 2 = 0 , that is, if v v v has exactly two children and both are leaves.
Proposition 3.25.
(i) On Ω s a \Omega_{\mathrm{sa}} Ω sa , let C C C be a component of a finite removal, let v ∈ C v\in C v ∈ C , and let ℓ 1 ≠ ℓ 2 \ell_1\ne\ell_2 ℓ 1 = ℓ 2 be neighbors of v v v in C C C whose only neighbor in C C C is v v v . Then
ψ = h v ℓ 2 δ ℓ 1 − h v ℓ 1 δ ℓ 2 \psi=h_{v\ell_2}\delta_{\ell_1}-h_{v\ell_1}\delta_{\ell_2} ψ = h v ℓ 2 δ ℓ 1 − h v ℓ 1 δ ℓ 2 satisfies H C ψ = 0 \mathcal{H}_C\psi=0 H C ψ = 0 , and ψ ≠ 0 \psi\ne0 ψ = 0 if the weights are nonzero.
(ii) Let p > 1 p>1 p > 1 . Then P ∞ \mathbb{P}^{\infty} P ∞ -almost surely there are infinitely many twin-leaf vertices. Consequently, P ∞ \mathbb{P}^{\infty} P ∞ -almost surely, the vectors ψ v = h v , v 2 δ v 1 − h v , v 1 δ v 2 \psi_v=h_{v,v2}\delta_{v1}-h_{v,v1}\delta_{v2} ψ v = h v , v 2 δ v 1 − h v , v 1 δ v 2 , for v v v a twin-leaf vertex, form an infinite family of nonzero, pairwise orthogonal, finitely supported eigenvectors of H \mathcal{H} H with eigenvalue $0. ∗ I n p a r t i c u l a r , ∗ . *In particular,* . ∗ I n p a r t i c u l a r , ∗ 0 is an eigenvalue of H \mathcal{H} H of infinite multiplicity.
Proof. For (i), since the vector ψ \psi ψ is finitely supported, it suffices to check the pointwise equation. At v v v , we have ( H C , f o r m ψ ) ( v ) = h v ℓ 1 h v ℓ 2 − h v ℓ 2 h v ℓ 1 = 0 (\mathcal{H}_{C,\mathrm{form}}\psi)(v)=h_{v\ell_1}h_{v\ell_2}-h_{v\ell_2}h_{v\ell_1}=0 ( H C , form ψ ) ( v ) = h v ℓ 1 h v ℓ 2 − h v ℓ 2 h v ℓ 1 = 0 . At ℓ 1 \ell_1 ℓ 1 and ℓ 2 \ell_2 ℓ 2 , the only neighbor is v v v , and ψ ( v ) = 0 \psi(v)=0 ψ ( v ) = 0 . Every other vertex u u u is not adjacent to ℓ 1 \ell_1 ℓ 1 or ℓ 2 \ell_2 ℓ 2 , since their only neighbor is v v v , and ( H C , f o r m ψ ) ( u ) = 0 (\mathcal{H}_{C,\mathrm{form}}\psi)(u)=0 ( H C , form ψ ) ( u ) = 0 .
For (ii), let V n = { v ∈ V : ∣ v ∣ = n } V_n=\{v\in V:|v|=n\} V n = { v ∈ V : ∣ v ∣ = n } . We first recall that the generation sizes ∣ V n ∣ |V_n| ∣ V n ∣ tend to infinity almost surely on S \mathcal{S} S (see [64, Section 4.3] for background on Galton–Watson processes); for completeness, we give the argument. Let N n = σ ( N u : ∣ u ∣ < n ) \mathcal{N}_n=\sigma(N_u:|u|<n) N n = σ ( N u : ∣ u ∣ < n ) . Then ∣ V n ∣ |V_n| ∣ V n ∣ is N n \mathcal{N}_n N n -measurable and, conditionally on N n \mathcal{N}_n N n , the variable ∣ V n + 1 ∣ |V_{n+1}| ∣ V n + 1 ∣ is a sum of ∣ V n ∣ |V_n| ∣ V n ∣ independent Pois ( p ) \operatorname{Pois}(p) Pois ( p ) variables; hence P ( ∣ V n + 1 ∣ = 0 ∣ N n ) = e − p ∣ V n ∣ \mathbb{P}(|V_{n+1}|=0\mid\mathcal{N}_n)=e^{-p|V_n|} P ( ∣ V n + 1 ∣ = 0 ∣ N n ) = e − p ∣ V n ∣ . Fix K ≥ 1 K\ge1 K ≥ 1 , and let τ 1 < τ 2 < ⋯ \tau_1<\tau_2<\cdots τ 1 < τ 2 < ⋯ be the successive times n n n with 1 ≤ ∣ V n ∣ ≤ K 1\le|V_n|\le K 1 ≤ ∣ V n ∣ ≤ K , which are stopping times for ( N n ) (\mathcal{N}_n) ( N n ) . On { τ j < ∞ } \{\tau_j<\infty\} { τ j < ∞ } , we have P ( ∣ V τ j + 1 ∣ = 0 ∣ N τ j ) ≥ e − p K \mathbb{P}(|V_{\tau_j+1}|=0\mid\mathcal{N}_{\tau_j})\ge e^{-pK} P ( ∣ V τ j + 1 ∣ = 0 ∣ N τ j ) ≥ e − p K , as one checks by decomposing according to the value of τ j \tau_j τ j . If ∣ V τ j + 1 ∣ = 0 |V_{\tau_j+1}|=0 ∣ V τ j + 1 ∣ = 0 , then ∣ V n ∣ = 0 |V_n|=0 ∣ V n ∣ = 0 for all n > τ j n>\tau_j n > τ j and τ j + 1 = ∞ \tau_{j+1}=\infty τ j + 1 = ∞ . Therefore P ( τ j + 1 < ∞ ) ≤ ( 1 − e − p K ) P ( τ j < ∞ ) \mathbb{P}(\tau_{j+1}<\infty)\le(1-e^{-pK})\mathbb{P}(\tau_j<\infty) P ( τ j + 1 < ∞ ) ≤ ( 1 − e − p K ) P ( τ j < ∞ ) , and P ( τ j < ∞ ) → 0 \mathbb{P}(\tau_j<\infty)\to0 P ( τ j < ∞ ) → 0 as j → ∞ j\to\infty j → ∞ , and almost surely 1 ≤ ∣ V n ∣ ≤ K 1\le|V_n|\le K 1 ≤ ∣ V n ∣ ≤ K for only finitely many n n n . On S \mathcal{S} S , we have ∣ V n ∣ ≥ 1 |V_n|\ge1 ∣ V n ∣ ≥ 1 for all n n n , and it follows that ∣ V n ∣ > K |V_n|>K ∣ V n ∣ > K for all large n n n . Since K K K is arbitrary, ∣ V n ∣ → ∞ |V_n|\to\infty ∣ V n ∣ → ∞ almost surely on S \mathcal{S} S .
For a generation-n n n vertex v v v , the event that v v v is a twin-leaf vertex depends on N v , N v 1 , N v 2 N_v,N_{v1},N_{v2} N v , N v 1 , N v 2 , which are independent of N n \mathcal{N}_n N n , and these variables are distinct for distinct v v v . Conditionally on N n \mathcal{N}_n N n , these events are independent with probability
a p = P ( N o = 2 ) P ( N o = 0 ) 2 = p 2 e − p 2 e − 2 p = p 2 e − 3 p 2 > 0 a_p=\mathbb{P}(N_o=2)\mathbb{P}(N_o=0)^2=\frac{p^2e^{-p}}{2}e^{-2p}=\frac{p^2e^{-3p}}{2}>0 a p = P ( N o = 2 ) P ( N o = 0 ) 2 = 2 p 2 e − p e − 2 p = 2 p 2 e − 3 p > 0
each. Since S ⊂ { ∣ V n ∣ ≥ 1 } \mathcal{S}\subset\{|V_n|\ge1\} S ⊂ { ∣ V n ∣ ≥ 1 } , we obtain
P ( S ∩ { no twin-leaf vertex in generation n } ) ≤ E [ 1 { ∣ V n ∣ ≥ 1 } ( 1 − a p ) ∣ V n ∣ ] . \mathbb{P}\bigl(\mathcal{S}\cap\{\text{no twin-leaf vertex in generation }n\}\bigr)\le\mathbb{E}\left[1_{\{|V_n|\ge1\}}(1-a_p)^{|V_n|}\right]. P ( S ∩ { no twin-leaf vertex in generation n } ) ≤ E [ 1 { ∣ V n ∣ ≥ 1 } ( 1 − a p ) ∣ V n ∣ ] .
As n → ∞ n\to\infty n → ∞ , the integrand tends to zero almost surely: on S \mathcal{S} S because ∣ V n ∣ → ∞ |V_n|\to\infty ∣ V n ∣ → ∞ , and on S c \mathcal{S}^c S c because ∣ V n ∣ = 0 |V_n|=0 ∣ V n ∣ = 0 for all large n n n . It is bounded by one, so the right side tends to zero by dominated convergence. For every m m m , the event that S \mathcal{S} S occurs and no generation n ≥ m n\ge m n ≥ m contains a twin-leaf vertex is contained in each of the events on the left with n ≥ m n\ge m n ≥ m , and it has probability zero. Taking the union over m m m , we find that P \mathbb{P} P -almost surely on S \mathcal{S} S there are infinitely many twin-leaf vertices.
The two children of a twin-leaf vertex are leaves of T \mathcal{T} T , whose only neighbor is their parent. Then (i) applies with C = V C=V C = V and shows that H ψ v = 0 \mathcal{H}\psi_v=0 H ψ v = 0 ; on the event that all marks are nonzero, ψ v ≠ 0 \psi_v\ne0 ψ v = 0 . Since the supports { v 1 , v 2 } \{v_1,v_2\} { v 1 , v 2 } of the vectors ψ v \psi_v ψ v are pairwise disjoint for distinct v v v , these vectors are pairwise orthogonal, and in particular linearly independent.
Spine decompositions and fresh squared moduli
This section collects the probabilistic tools that we use to compute expectations of sums over the vertices of the tree. They are of three kinds.
The first are spine decompositions. If F ( v ) F(v) F ( v ) is a nonnegative functional of the marked tree and of a vertex v v v of generation L L L , then the expectation of ∑ ∣ v ∣ = L F ( v ) \sum_{|v|=L} F(v) ∑ ∣ v ∣ = L F ( v ) equals p L p^L p L times the expectation of F F F evaluated at the endpoint of a spine: a path of length L L L from the root whose vertices carry independent Poisson families of further children, and whose endpoint carries an independent copy of the whole tree. This identity is the many-to-one formula. We prove it, its analogue for ordered pairs of vertices, which is the many-to-two formula, and their versions under the survival-conditioned law P ∞ \mathbb{P}^{\infty} P ∞ . The second kind concerns endpoints: the law of an endpoint functional under P ∞ \mathbb{P}^{\infty} P ∞ , the children of an endpoint classified by the fate of their forward subtrees, and the growth of the set of vertices with infinite forward subtree. The third is an anti-concentration estimate for affine combinations of independent exponential variables, together with a definition of fresh squared moduli, which says when this estimate may be applied conditionally on a σ \sigma σ -field.
In Section 4.1, we realize the marked tree on the space of finite words and record its branching property. Section 4.2 recalls the Mecke formula and splits the children of the root according to the survival of their forward subtrees. Section 4.3 defines the spine laws and proves the many-to-one and many-to-two formulas. Section 4.4 compares endpoint laws under P \mathbb{P} P and P ∞ \mathbb{P}^{\infty} P ∞ . Section 4.5 defines fresh squared moduli and proves the affine negative-moment lemma. Section 4.6 treats the children of a single vertex and the growth of the backbone.
Throughout this section, p > 0 p>0 p > 0 is fixed. Whenever the survival event enters, we assume that p > 1 p>1 p > 1 , so that ϱ p > 0 \varrho_p>0 ϱ p > 0 .
The Ulam–Harris construction
Let U 0 = { o } \mathbb{U}_0=\{o\} U 0 = { o } , where o o o denotes the empty word, let U n = { 1 , 2 , … } n \mathbb{U}_n=\{1,2,\ldots\}^n U n = { 1 , 2 , … } n for n ≥ 1 n\ge1 n ≥ 1 , and set U = ⋃ n ≥ 0 U n \mathbb{U}=\bigcup_{n\ge0}\mathbb{U}_n U = ⋃ n ≥ 0 U n and U ∗ = U ∖ { o } \mathbb{U}_*=\mathbb{U}\setminus\{o\} U ∗ = U ∖ { o } . For a word u = i 1 ⋯ i n u=i_1\cdots i_n u = i 1 ⋯ i n and a word v v v , we write ∣ u ∣ = n |u|=n ∣ u ∣ = n for the length of u u u and u v uv uv for the concatenation of u u u and v v v . For $0\le l\le n$, we let u l = i 1 ⋯ i l u_l=i_1\cdots i_l u l = i 1 ⋯ i l denote the prefix of u u u of length l l l , so that u 0 = o u_0=o u 0 = o and u n = u u_n=u u n = u , and for u ≠ o u\ne o u = o we write u − = u n − 1 u^-=u_{n-1} u − = u n − 1 . For u , v ∈ U u,v\in\mathbb{U} u , v ∈ U , we let u ∧ v u\wedge v u ∧ v denote the longest common prefix of $u and v v v .
Let γ \gamma γ denote the standard complex Gaussian law, which has density π − 1 e − ∣ w ∣ 2 \pi^{-1}e^{-|w|^{2}} π − 1 e − ∣ w ∣ 2 with respect to Lebesgue measure on C \mathbb{C} C . If w w w has law γ \gamma γ , then ∣ w ∣ 2 |w|^{2} ∣ w ∣ 2 has law Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) , and w / ∣ w ∣ w/|w| w /∣ w ∣ is uniform on the unit circle and independent of ∣ w ∣ 2 |w|^{2} ∣ w ∣ 2 . Let Z = [ 0 , 1 ] { 1 , 2 , … } \mathsf{Z}=[0,1]^{\{1,2,\ldots\}} Z = [ 0 , 1 ] { 1 , 2 , … } with its product σ \sigma σ -field, and let κ \kappa κ be the product of the Lebesgue measures on the factors.
Definition 4.1. The configuration space is
Ω = { 0 , 1 , 2 , … } U × Z U × C U ∗ , \Omega=\{0,1,2,\ldots\}^{\mathbb{U}}\times\mathsf{Z}^{\mathbb{U}}\times\mathbb{C}^{\mathbb{U}_{*}}, Ω = { 0 , 1 , 2 , … } U × Z U × C U ∗ , with the product σ \sigma σ -field F \mathcal{F} F . Its elements are written
ω = ( ( N v ) v ∈ U , ( ζ v ) v ∈ U , ( w v ) v ∈ U ∗ ) . \omega=((N_v)_{v\in\mathbb{U}},(\zeta_v)_{v\in\mathbb{U}},(w_v)_{v\in\mathbb{U}_{*}}). ω = (( N v ) v ∈ U , ( ζ v ) v ∈ U , ( w v ) v ∈ U ∗ ) .
We let P \mathbb{P} P be the probability measure on ( Ω , F ) (\Omega,\mathcal{F}) ( Ω , F ) under which all coordinates are independent, each N v N_v N v has law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) , each ζ v \zeta_v ζ v has law κ \kappa κ , and each w v w_v w v has law γ \gamma γ . The rank of v ∈ U v\in\mathbb{U} v ∈ U is the first coordinate ζ v ( 1 ) \zeta_v(1) ζ v ( 1 ) of ζ v = ( ζ v ( 1 ) , ζ v ( 2 ) , … ) \zeta_v=(\zeta_v(1),\zeta_v(2),\ldots) ζ v = ( ζ v ( 1 ) , ζ v ( 2 ) , … ) ; the variables ζ v \zeta_v ζ v are called auxiliary marks.
(a) The vertex set is
V ( ω ) = { i 1 ⋯ i n ∈ U : n ≥ 0 , i l ≤ N i 1 ⋯ i l − 1 for 1 ≤ l ≤ n } , V(\omega)=\{i_1\cdots i_n\in\mathbb{U}:n\geq0,\ i_l\leq N_{i_1\cdots i_{l-1}}\text{ for }1\leq l\leq n\}, V ( ω ) = { i 1 ⋯ i n ∈ U : n ≥ 0 , i l ≤ N i 1 ⋯ i l − 1 for 1 ≤ l ≤ n } ,
and T = T ( ω ) \mathcal{T}=\mathcal{T}(\omega) T = T ( ω ) is the tree with vertex set V ( ω ) V(\omega) V ( ω ) , root o o o , and edges { v − , v } \{v^{-},v\} { v − , v } for v ∈ V ( ω ) ∖ { o } v\in V(\omega)\setminus\{o\} v ∈ V ( ω ) ∖ { o } . The children of v ∈ V ( ω ) v\in V(\omega) v ∈ V ( ω ) are v 1 , … , v N v v1,\ldots,vN_v v 1 , … , v N v .
(b) For v ∈ V ( ω ) ∖ { o } v\in V(\omega)\setminus\{o\} v ∈ V ( ω ) ∖ { o } , the edge { v − , v } \{v^{-},v\} { v − , v } carries the mark w v − v = w v w_{v^{-}v}=w_v w v − v = w v . We set h v − v = p − 1 / 2 w v h_{v^{-}v}=p^{-1/2}w_v h v − v = p − 1/2 w v , h v v − = h v − v ‾ h_{vv^{-}}=\overline{h_{v^{-}v}} h v v − = h v − v and Y v = Y v − v = ∣ w v ∣ 2 Y_v=Y_{v^{-}v}=|w_v|^2 Y v = Y v − v = ∣ w v ∣ 2 .
(c) For u ∈ U u\in\mathbb{U} u ∈ U , the shift θ u : Ω → Ω \theta_u:\Omega\to\Omega θ u : Ω → Ω is defined by
θ u ω = ( ( N u v ) v ∈ U , ( ζ u v ) v ∈ U , ( w u v ) v ∈ U ∗ ) . \theta_u\omega=((N_{uv})_{v\in\mathbb{U}},(\zeta_{uv})_{v\in\mathbb{U}},(w_{uv})_{v\in\mathbb{U}_{*}}). θ u ω = (( N uv ) v ∈ U , ( ζ uv ) v ∈ U , ( w uv ) v ∈ U ∗ ) .
(d) The survival event is S = { ω ∈ Ω : ∣ V ( ω ) ∣ = ∞ } \mathcal{S}=\{\omega\in\Omega:|V(\omega)|=\infty\} S = { ω ∈ Ω : ∣ V ( ω ) ∣ = ∞ } , and ϱ p = P ( S ) \varrho_p=\mathbb{P}(\mathcal{S}) ϱ p = P ( S ) . For p > 1 p>1 p > 1 , we set P ∞ = P ( ⋅ ∣ S ) \mathbb{P}^{\infty}=\mathbb{P}(\,\cdot\mid\mathcal{S}) P ∞ = P ( ⋅ ∣ S ) and write E ∞ \mathbb{E}^{\infty} E ∞ for the corresponding expectation.
(e) For n ≥ 0 n\geq0 n ≥ 0 , we set V n = V ( ω ) ∩ U n \mathcal{V}_n=V(\omega)\cap\mathbb{U}_n V n = V ( ω ) ∩ U n , V n ∞ = { v ∈ V n : θ v ω ∈ S } \mathcal{V}^{\infty}_n=\{v\in\mathcal{V}_n:\theta_v\omega\in\mathcal{S}\} V n ∞ = { v ∈ V n : θ v ω ∈ S } , and Z n = ∣ V n ∞ ∣ Z_n=|\mathcal{V}^{\infty}_n| Z n = ∣ V n ∞ ∣ . The backbone is the set ⋃ n ≥ 0 V n ∞ \bigcup_{n\geq0}\mathcal{V}^{\infty}_n ⋃ n ≥ 0 V n ∞ of vertices whose forward subtree is infinite.
Under P \mathbb{P} P , the offspring numbers N v N_v N v are independent with law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) , and T \mathcal{T} T is a Poisson–Galton–Watson tree with mean offspring p p p . The marks w v w_v w v , v ∈ U ∗ v\in\mathbb{U}_{*} v ∈ U ∗ , are independent of ( N v ) v ∈ U (N_v)_{v\in\mathbb{U}} ( N v ) v ∈ U . This implies that, conditionally on T \mathcal{T} T , the marks of its edges are independent with law γ \gamma γ ; in particular, the squared moduli Y v = ∣ w v ∣ 2 Y_v=|w_v|^2 Y v = ∣ w v ∣ 2 are independent Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) variables, and they are independent of the phases w v / ∣ w v ∣ w_v/|w_v| w v /∣ w v ∣ , which are uniform on the unit circle. We conclude that T \mathcal{T} T , together with the weights h e h_e h e , is a realization of the marked tree of Definition 3.2, with the auxiliary element a = ( ζ v ) v ∈ U a=(\zeta_v)_{v\in\mathbb{U}} a = ( ζ v ) v ∈ U of the standard Borel space Z U \mathbb{Z}^{\mathbb{U}} Z U , which is independent of the marked tree. In particular, the σ \sigma σ -algebra G 0 \mathcal{G}_0 G 0 of Definition 3.2 is generated by the offspring numbers N v N_v N v and the auxiliary marks ζ v \zeta_v ζ v , v ∈ U v \in\mathbb{U} v ∈ U . From now on we work with this realization: P \mathbb{P} P and E \mathbb{E} E refer to the law of ω \omega ω , and S \mathcal{S} S , ϱ p \varrho_p ϱ p , P ∞ \mathbb{P}^{\infty} P ∞ are the objects of Definition 3.2. The event S \mathcal{S} S is measurable, since every vertex has finitely many children and therefore, by König’s lemma, S = ⋂ n ≥ 0 { V n ≠ ∅ } \mathcal{S} = \bigcap_{n\geq0}\{\mathcal{V}_n \ne\varnothing\} S = ⋂ n ≥ 0 { V n = ∅ } . Recall that 0 < ϱ p < 1 0 < \varrho_p < 1 0 < ϱ p < 1 for p > 1 p > 1 p > 1 .
The coordinates of ω \omega ω at labels outside V ( ω ) V(\omega) V ( ω ) do not affect the marked tree. We keep them because they make the shifts measure preserving and the spine laws of Section 4.3 simple to write down. Note that each shift θ u \theta_u θ u preserves P \mathbb{P} P , since it moves the coordinates at the labels u v uv uv to the labels v v v , and all coordinates of the same type have the same law.
If u ∈ V ( ω ) u \in V(\omega) u ∈ V ( ω ) , then V ( θ u ω ) = { v ∈ U : u v ∈ V ( ω ) } V(\theta_u\omega) = \{v \in\mathbb{U} : uv \in V(\omega)\} V ( θ u ω ) = { v ∈ U : uv ∈ V ( ω )} , and the map v ↦ u v v \mapsto uv v ↦ uv is an isomorphism from T ( θ u ω ) \mathcal{T}(\theta_u\omega) T ( θ u ω ) onto the forward subtree T u \mathcal{T}_u T u . The edge { v − , v } \{v^{-}, v\} { v − , v } of T ( θ u ω ) \mathcal{T}(\theta_u\omega) T ( θ u ω ) carries the mark w u v w_{uv} w uv , which is also the mark of the corresponding edge { ( u v ) − , u v } \{(uv)^{-}, uv\} {( uv ) − , uv } of T u \mathcal{T}_u T u . We use this correspondence throughout: if a quantity f ( ω ) f(\omega) f ( ω ) is defined in terms of the marked tree T ( ω ) \mathcal{T}(\omega) T ( ω ) and its root o o o , then the same quantity defined in terms of the marked forward subtree T u \mathcal{T}_u T u and its root u u u equals f ( θ u ω ) f(\theta_u\omega) f ( θ u ω ) . For example, { T u is infinite } = { θ u ω ∈ S } \{\mathcal{T}_u \text{ is infinite}\} = \{\theta_u\omega\in\mathcal{S}\} { T u is infinite } = { θ u ω ∈ S } , so that Z n Z_n Z n is the number of vertices of generation n n n whose forward subtree is infinite. Similarly, the forward cavity variable of Definition 3.15 satisfies Γ u ( z ) = G o o ( z ) ∘ θ u \Gamma_u(z) = G_{oo}(z) \circ\theta_u Γ u ( z ) = G oo ( z ) ∘ θ u for every u ∈ V ( ω ) ∖ { o } u \in V(\omega) \setminus\{o\} u ∈ V ( ω ) ∖ { o } and z ∈ C ∖ R z \in\mathbb{C} \setminus\mathbb{R} z ∈ C ∖ R . Here, and in the rest of this section, a function such as G o o ( z ) G_{oo}(z) G oo ( z ) which is defined only P \mathbb{P} P -almost surely is extended by 0 0 0 to the exceptional null set; since every shift preserves P \mathbb{P} P , the identity Γ u ( z ) = G o o ( z ) ∘ θ u \Gamma_u(z) = G_{oo}(z) \circ\theta_u Γ u ( z ) = G oo ( z ) ∘ θ u then holds for all u u u simultaneously, P \mathbb{P} P -almost surely.
Lemma 4.2. Let p > 0 p > 0 p > 0 , let M = C × Ω M = \mathbb{C} \times\Omega M = C × Ω with the product σ \sigma σ -field, and let m = γ ⊗ P m = \gamma\otimes\mathbb{P} m = γ ⊗ P .
(i) For every n ≥ 0 n \geq0 n ≥ 0 , the configurations θ v ω \theta_v\omega θ v ω , v ∈ U n v \in\mathbb{U}_n v ∈ U n , are independent, each has law P \mathbb{P} P , and they are independent of the σ \sigma σ -field
F n = σ ( N u , ζ u : u ∈ U , ∣ u ∣ < n ; w u : u ∈ U , 1 ≤ ∣ u ∣ ≤ n ) . \mathcal{F}_n = \sigma(N_u,\zeta_u : u \in\mathbb{U},\ |u| < n; \ w_u : u \in\mathbb{U},\ 1 \leq|u| \leq n). F n = σ ( N u , ζ u : u ∈ U , ∣ u ∣ < n ; w u : u ∈ U , 1 ≤ ∣ u ∣ ≤ n ) . The set V n \mathcal{V}_n V n is F n \mathcal{F}_n F n -measurable.
(ii) The random elements N o N_o N o , ζ o \zeta_o ζ o , and ( w k , θ k ω ) (w_k,\theta_k\omega) ( w k , θ k ω ) , k = 1 , 2 , … k = 1,2,\ldots k = 1 , 2 , … , are independent, and each ( w k , θ k ω ) (w_k,\theta_k\omega) ( w k , θ k ω ) has law m m m . The map ω ↦ ( N o , ζ o , ( ( w k , θ k ω ) k ≥ 1 ) \omega\mapsto(N_o,\zeta_o,((w_k,\theta_k\omega)_{k\geq1}) ω ↦ ( N o , ζ o , (( w k , θ k ω ) k ≥ 1 ) is a bijection from Ω \Omega Ω onto { 0 , 1 , 2 , … } × Z × M { 1 , 2 , … } \{0,1,2,\ldots\} \times\mathbb{Z} \times M^{\{1,2,\ldots\}} { 0 , 1 , 2 , … } × Z × M { 1 , 2 , … } , and both this map and its inverse are measurable.
(iii) The child process Π o = ∑ k = 1 N o δ ( w k , θ k ω ) \Pi_o = \sum_{k=1}^{N_o} \delta_{(w_k,\theta_k\omega)} Π o = ∑ k = 1 N o δ ( w k , θ k ω ) is a Poisson process on M M M with intensity measure p m p\,m p m , and it is independent of ζ o \zeta_o ζ o . For u ∈ U u \in\mathbb{U} u ∈ U , the process Π u = Π o ∘ θ u = ∑ k = 1 N u δ ( w u k , θ u k ω ) \Pi_u = \Pi_o \circ\theta_u = \sum_{k=1}^{N_u} \delta_{(w_{uk},\theta_{uk}\omega)} Π u = Π o ∘ θ u = ∑ k = 1 N u δ ( w u k , θ u k ω ) lists the edge marks and the forward configurations of the children of u u u .
Proof. The coordinates of θ v ω \theta_v\omega θ v ω are the coordinates N v u , ζ v u N_{vu},\zeta_{vu} N v u , ζ v u with u ∈ U u \in\mathbb{U} u ∈ U and w v u w_{vu} w v u with u ∈ U ∗ u \in\mathbb{U}_* u ∈ U ∗ . For distinct v , v ′ ∈ U n v,v' \in\mathbb{U}_n v , v ′ ∈ U n , no label of the form v u vu v u is of the form v ′ u ′ v'u' v ′ u ′ . Further, every label v u vu v u with u ∈ U u \in\mathbb{U} u ∈ U has length at least n n n , and every label v u vu v u with u ∈ U ∗ u \in\mathbb{U}_* u ∈ U ∗ has length larger than n n n . The configurations θ v ω \theta_v\omega θ v ω , v ∈ U n v \in U_n v ∈ U n , and the generators of F n \mathcal{F}_n F n are therefore functions of pairwise disjoint sets of coordinates. Under the product measure P \mathbb{P} P , such functions are independent. Recall that θ v \theta_v θ v preserves P \mathbb{P} P , so that each θ v ω \theta_v\omega θ v ω has law P \mathbb{P} P . Finally, whether v ∈ U n v \in U_n v ∈ U n belongs to V ( ω ) V(\omega) V ( ω ) depends only on N v 0 , … , N v n − 1 N_{v0},\ldots,N_{v n-1} N v 0 , … , N v n − 1 , which are F n \mathcal{F}_n F n -measurable. This proves (i).
For (ii), we note that N o N_o N o , ζ o \zeta_o ζ o , and the pairs ( w k , θ k ω ) (w_k,\theta_k\omega) ( w k , θ k ω ) , k ≥ 1 k \ge1 k ≥ 1 , are again functions of pairwise disjoint sets of coordinates. They are independent, and ( w k , θ k ω ) (w_k,\theta_k\omega) ( w k , θ k ω ) has law γ ⊗ P \gamma\otimes\mathbb{P} γ ⊗ P by (i). Every coordinate of ω \omega ω appears exactly once in ( N o , ζ o , ( ( w k , θ k ω ) k ≥ 1 ) (N_o,\zeta_o,((w_k,\theta_k\omega)_{k\ge1}) ( N o , ζ o , (( w k , θ k ω ) k ≥ 1 ) : the coordinates at a label k u ku k u appear in θ k ω \theta_k\omega θ k ω at the label u u u , except for w k w_k w k , which appears separately. This gives the bijection; it is measurable in both directions because both directions act coordinatewise.
By (ii), conditionally on N o = n N_o=n N o = n , the points ( w k , θ k ω ) (w_k,\theta_k\omega) ( w k , θ k ω ) , 1 ≤ k ≤ n 1 \le k \le n 1 ≤ k ≤ n , are independent with law m \mathfrak{m} m , and they are independent of ζ o \zeta_o ζ o . A mixed binomial process whose number of points has law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) is a Poisson process with intensity measure p m p\mathfrak{m} p m [94, Proposition 3.5], which proves the first claim in (iii). The second follows from the definition of θ u \theta_u θ u . □
Remark 4.3. The auxiliary marks do not enter the marked tree. They serve two purposes. The rank ζ v ( 1 ) \zeta_v(1) ζ v ( 1 ) is used to select children by a rule which does not depend on the labels, such as “the leaf child with the smallest rank”. The remaining coordinates of ζ v \zeta_v ζ v may be used to attach further independent randomness to v v v , for instance a variable with a prescribed law on R \mathbb{R} R , obtained by applying the corresponding quantile function to the second coordinate of ζ v \zeta_v ζ v . All statements below concern arbitrary measurable functionals of ω \omega ω and apply to such constructions without change.
The Mecke formula and survival
Let ( X , X ) (X,\mathcal{X}) ( X , X ) be a measurable space. We let N ( X ) \mathcal{N}(X) N ( X ) denote the set of finite counting measures
π = ∑ i = 1 n δ x i , \pi=\sum_{i=1}^{n}\delta_{x_i}, π = i = 1 ∑ n δ x i ,
where n ≥ 0 n \ge0 n ≥ 0 and x 1 , … , x n ∈ X x_1,\ldots,x_n \in X x 1 , … , x n ∈ X , with the σ \sigma σ -field generated by the maps π ↦ π ( B ) \pi\mapsto\pi(B) π ↦ π ( B ) , B ∈ X B \in\mathcal{X} B ∈ X . For such π \pi π and k ≥ 1 k \ge1 k ≥ 1 , we let
π ( k ) = ∑ ( i 1 , … , i k ) δ ( x i 1 , … , x i k ) \pi^{(k)}=\sum_{(i_1,\ldots,i_k)}\delta_{(x_{i_1},\ldots,x_{i_k})} π ( k ) = ( i 1 , … , i k ) ∑ δ ( x i 1 , … , x i k )
where the sum runs over all ordered k k k -tuples of pairwise distinct indices in { 1 , … , n } \{1,\ldots,n\} { 1 , … , n } ; this measure on X k X^k X k does not depend on the enumeration of the atoms of π \pi π . Let μ \mu μ be a probability measure on X X X . A Poisson process with intensity measure p μ p\mu p μ is a random element Π \Pi Π of N ( X ) \mathcal{N}(X) N ( X ) such that, for all pairwise disjoint B 1 , … , B r ∈ X B_1,\ldots,B_r \in\mathcal{X} B 1 , … , B r ∈ X , the variables Π ( B 1 ) , … , Π ( B r ) \Pi(B_1),\ldots,\Pi(B_r) Π ( B 1 ) , … , Π ( B r ) are independent and Π ( B i ) \Pi(B_i) Π ( B i ) has law Pois ( p μ ( B i ) ) \operatorname{Pois}(p\mu(B_i)) Pois ( p μ ( B i )) . Its law is determined by p μ p\mu p μ , and it coincides with the law of ∑ i = 1 N δ X i \sum_{i=1}^{N}\delta_{X_i} ∑ i = 1 N δ X i , where N N N has law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) and X 1 , X 2 , … X_1,X_2,\ldots X 1 , X 2 , … are independent with law μ \mu μ and independent of N N N ; see [94, Chapter 3].
Lemma 4.4. Let ( X , X ) (X,\mathcal{X}) ( X , X ) be a measurable space, let μ \mu μ be a probability measure on X X X , let p > 0 p>0 p > 0 , and let Π \Pi Π be a Poisson process on X X X with intensity measure p μ p\mu p μ .
(i) For every measurable F : X × N ( X ) → [ 0 , ∞ ] F:X\times\mathbb{N}(X)\to[0,\infty] F : X × N ( X ) → [ 0 , ∞ ] ,
E ∫ F ( x , Π − δ x ) Π ( d x ) = p ∫ E F ( x , Π ) μ ( d x ) . (34) \mathbb{E}\int F(x,\Pi-\delta_x)\,\Pi(\mathrm{d}x)=p\int\mathbb{E}F(x,\Pi)\,\mu(\mathrm{d}x). \tag*{(34)} E ∫ F ( x , Π − δ x ) Π ( d x ) = p ∫ E F ( x , Π ) μ ( d x ) . (34) (ii) For every k ≥ 1 k\geq1 k ≥ 1 and every measurable F : X k × N ( X ) → [ 0 , ∞ ] F:X^k\times\mathbb{N}(X)\to[0,\infty] F : X k × N ( X ) → [ 0 , ∞ ] ,
E ∫ F ( x 1 , … , x k , Π − ∑ i = 1 k δ x i ) Π ( k ) ( d x 1 ⋯ d x k ) \mathbb{E}\int F\left(x_1,\ldots,x_k,\Pi-\sum_{i=1}^{k}\delta_{x_i}\right)\,\Pi^{(k)}(\mathrm{d}x_1\cdots\mathrm{d}x_k) E ∫ F ( x 1 , … , x k , Π − i = 1 ∑ k δ x i ) Π ( k ) ( d x 1 ⋯ d x k ) = p k ∫ E F ( x 1 , … , x k , Π ) μ ⊗ k ( d x 1 ⋯ d x k ) . (35) =p^k\int\mathbb{E}F(x_1,\ldots,x_k,\Pi)\,\mu^{\otimes k}(\mathrm{d}x_1\cdots\mathrm{d}x_k). \tag*{(35)} = p k ∫ E F ( x 1 , … , x k , Π ) μ ⊗ k ( d x 1 ⋯ d x k ) . (35)
Identity (34) is the Mecke equation [97]; see [94, Theorem 4.1], and (35) is its multivariate form [94, Theorem 4.4]. In words, (34) states that the measure B ↦ E ∫ 1 B ( x , Π − δ x ) Π ( d x ) B\mapsto\mathbb{E}\int\mathbf{1}_B(x,\Pi-\delta_x)\,\Pi(\mathrm{d}x) B ↦ E ∫ 1 B ( x , Π − δ x ) Π ( d x ) on X × N ( X ) X\times\mathbb{N}(X) X × N ( X ) equals p μ ⊗ Law ( Π ) p\mu\otimes\operatorname{Law}(\Pi) p μ ⊗ Law ( Π ) . Summing a functional over the points of Π \Pi Π amounts to drawing the selected point from μ \mu μ , with total weight p p p , while the remaining points form a Poisson process with the original intensity measure which is independent of the selected point (Slivnyak’s theorem). We include the short proof for finite intensity measures.
Proof of Lemma 4.4. Since both sides of (35) depend only on the law of Π \Pi Π , we may assume that Π = ∑ i = 1 N δ X i \Pi=\sum_{i=1}^{N}\delta_{X_i} Π = ∑ i = 1 N δ X i , where N N N has law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) and X 1 , X 2 , … X_1,X_2,\ldots X 1 , X 2 , … are independent with law μ \mu μ and independent of N N N . Then the left side of (35) equals
E ∑ ( i 1 , … , i k ) F ( X i 1 , … , X i k , ∑ i ≤ N i ∉ { i 1 , … , i k } δ X i ) , \mathbb{E}\sum_{(i_1,\ldots,i_k)}F\left(X_{i_1},\ldots,X_{i_k},\sum_{\substack{i\leq N\\i\notin\{i_1,\ldots,i_k\}}}\delta_{X_i}\right), E ( i 1 , … , i k ) ∑ F X i 1 , … , X i k , i ≤ N i ∈ / { i 1 , … , i k } ∑ δ X i ,
where the sum runs over ordered k k k -tuples of pairwise distinct indices in { 1 , … , N } \{1,\ldots,N\} { 1 , … , N } . Conditionally on N = n ≥ k N=n\geq k N = n ≥ k , there are n ! / ( n − k ) ! n!/(n-k)! n ! / ( n − k )! such tuples, and by exchangeability of ( X 1 , … , X n ) (X_1,\ldots,X_n) ( X 1 , … , X n ) , each term has the same conditional expectation as F ( X 1 , … , X k , ∑ i = k + 1 n δ X i ) F(X_1,\ldots,X_k,\sum_{i=k+1}^{n}\delta_{X_i}) F ( X 1 , … , X k , ∑ i = k + 1 n δ X i ) . Since P ( N = n ) n ! / ( n − k ) ! = p k P ( N = n − k ) \mathbb{P}(N=n)n!/(n-k)!=p^k\mathbb{P}(N=n-k) P ( N = n ) n ! / ( n − k )! = p k P ( N = n − k ) , we obtain
p k ∑ n ≥ k P ( N = n − k ) E F ( X 1 , … , X k , ∑ i = k + 1 n δ X i ) = p k E F ( X 1 , … , X k , ∑ i = k + 1 k + N ′ δ X i ) . p^k\sum_{n\geq k}\mathbb{P}(N=n-k)\mathbb{E}F\left(X_1,\ldots,X_k,\sum_{i=k+1}^{n}\delta_{X_i}\right)
=p^k\mathbb{E}F\left(X_1,\ldots,X_k,\sum_{i=k+1}^{k+N'}\delta_{X_i}\right). p k n ≥ k ∑ P ( N = n − k ) E F ( X 1 , … , X k , i = k + 1 ∑ n δ X i ) = p k E F X 1 , … , X k , i = k + 1 ∑ k + N ′ δ X i .
where N ′ N' N ′ is a Pois ( p ) \operatorname{Pois}(p) Pois ( p ) variable independent of ( X i ) i ≥ 1 (X_i)_{i\geq1} ( X i ) i ≥ 1 . The process ∑ i = k + 1 k + N ′ δ X i \sum_{i=k+1}^{k+N'}\delta_{X_i} ∑ i = k + 1 k + N ′ δ X i is a Poisson process with intensity p μ p\mu p μ which is independent of ( X 1 , … , X k ) (X_1,\ldots,X_k) ( X 1 , … , X k ) . Then (35) follows from Tonelli’s theorem. The case k = 1 k=1 k = 1 is (34) . □
We now split the children of the root according to the survival of their forward subtrees. Recall the space M = C × Ω \mathrm{M}=\mathbb{C}\times\Omega M = C × Ω , the law m = γ ⊗ P m=\gamma\otimes\mathbb{P} m = γ ⊗ P , and the child process Π o \Pi_o Π o from Lemma 4.2.
Lemma 4.5. Let p > 1 p>1 p > 1 . Set M ∞ = C × S M^\infty=\mathbb{C}\times S M ∞ = C × S and M f = C × ( Ω ∖ S ) M^{\mathrm f}=\mathbb{C}\times(\Omega\setminus S) M f = C × ( Ω ∖ S ) , let m ∞ = γ ⊗ P ∞ m^\infty=\gamma\otimes\mathbb{P}^\infty m ∞ = γ ⊗ P ∞ and m f = γ ⊗ P ( ⋅ ∣ Ω ∖ S ) m^{\mathrm f}=\gamma\otimes\mathbb{P}(\,\cdot\mid\Omega\setminus S) m f = γ ⊗ P ( ⋅ ∣ Ω ∖ S ) , and let Π ∞ \Pi^\infty Π ∞ and Π f \Pi^{\mathrm f} Π f be the restrictions of Π o \Pi_o Π o to M ∞ M^\infty M ∞ and M f M^{\mathrm f} M f . Let N ∞ = Π ∞ ( M ) N^\infty=\Pi^\infty(M) N ∞ = Π ∞ ( M ) and N f = Π f ( M ) N^{\mathrm f}=\Pi^{\mathrm f}(M) N f = Π f ( M ) be the numbers of children of the root with infinite and with finite forward subtree.
(i) Under P \mathbb{P} P , the random elements Π ∞ \Pi^\infty Π ∞ , Π f \Pi^{\mathrm f} Π f , and ζ o \zeta_o ζ o are independent, Π ∞ \Pi^\infty Π ∞ is a Poisson process with intensity measure p ϱ p m ∞ p\varrho_p m^\infty p ϱ p m ∞ , and Π f \Pi^{\mathrm f} Π f is a Poisson process with intensity measure p ( 1 − ϱ p ) m f p(1-\varrho_p)m^{\mathrm f} p ( 1 − ϱ p ) m f . In particular, N ∞ N^\infty N ∞ and N f N^{\mathrm f} N f are independent with laws Pois ( p ϱ p ) \operatorname{Pois}(p\varrho_p) Pois ( p ϱ p ) and Pois ( p ( 1 − ϱ p ) ) \operatorname{Pois}(p(1-\varrho_p)) Pois ( p ( 1 − ϱ p )) . Moreover, S = { N ∞ ≥ 1 } S=\{N^\infty\geq1\} S = { N ∞ ≥ 1 } and ϱ p = 1 − e − p ϱ p \varrho_p=1-e^{-p\varrho_p} ϱ p = 1 − e − p ϱ p .
(ii) Under P ∞ \mathbb{P}^\infty P ∞ , the random elements Π ∞ \Pi^\infty Π ∞ , Π f \Pi^{\mathrm f} Π f , and ζ o \zeta_o ζ o are independent; Π f \Pi^{\mathrm f} Π f and ζ o \zeta_o ζ o have the same laws as under P \mathbb{P} P , and Π ∞ \Pi^\infty Π ∞ has the law of a Poisson process with intensity measure p ϱ p m ∞ p\varrho_p m^\infty p ϱ p m ∞ conditioned to have at least one point.
(iii) For every measurable Z : Ω → [ 0 , ∞ ) Z:\Omega\to[0,\infty) Z : Ω → [ 0 , ∞ ) ,
E ∞ Z = ϱ p − 1 E [ Z 1 S ] ≤ ϱ p − 1 E Z . (36) \mathbb{E}^\infty Z=\varrho_p^{-1}\mathbb{E}[Z1_S]\leq\varrho_p^{-1}\mathbb{E}Z. \tag*{(36)} E ∞ Z = ϱ p − 1 E [ Z 1 S ] ≤ ϱ p − 1 E Z . (36) (iv) For every k ≥ 1 k\geq1 k ≥ 1 and every measurable F : M k × N ( M ) → [ 0 , ∞ ] F:M^k\times\mathbb{N}(M)\to[0,\infty] F : M k × N ( M ) → [ 0 , ∞ ] ,
E ∞ ∫ F ( m 1 , … , m k , Π o − ∑ i = 1 k δ m i ) ( Π ∞ ) ( k ) ( d m 1 ⋯ d m k ) = p k ϱ p k − 1 ∫ E F ( m 1 , … , m k , Π o ) ( m ∞ ) ⊗ k ( d m 1 ⋯ d m k ) . (37) \mathbb{E}^\infty\int F\left(m_1,\ldots,m_k,\Pi_o-\sum_{i=1}^{k}\delta_{m_i}\right)(\Pi^\infty)^{(k)}(\mathrm{d}m_1\cdots\mathrm{d}m_k)
=
p^k\varrho_p^{k-1}\int\mathbb{E}F(m_1,\ldots,m_k,\Pi_o)(m^\infty)^{\otimes k}(\mathrm{d}m_1\cdots\mathrm{d}m_k). \tag*{(37)} E ∞ ∫ F ( m 1 , … , m k , Π o − i = 1 ∑ k δ m i ) ( Π ∞ ) ( k ) ( d m 1 ⋯ d m k ) = p k ϱ p k − 1 ∫ E F ( m 1 , … , m k , Π o ) ( m ∞ ) ⊗ k ( d m 1 ⋯ d m k ) . (37) The left side of (4.4) is the sum, over ordered k k k -tuples of distinct children of the root with infinite forward subtrees, of F F F evaluated at their marks and at the process of all remaining children. On the right side, the selected marks are independent with law m ∞ m^\infty m ∞ , so that their edge marks have law γ \gamma γ and their forward configurations have law P ∞ \mathbb{P}^\infty P ∞ , while the residual process is the unconditioned child process Π o \Pi_o Π o under P \mathbb{P} P , independent of the selected marks. For k = 1 k=1 k = 1 the factor is p p p .
Proof. We begin with (i). Let u 1 , u 2 : M → [ 0 , ∞ ) u_1,u_2:M\to[0,\infty) u 1 , u 2 : M → [ 0 , ∞ ) be measurable and set u = u 1 1 M ∞ + u 2 1 M f u=u_1 1_{M^\infty}+u_2 1_{M^{\mathrm f}} u = u 1 1 M ∞ + u 2 1 M f . Using the representation of Π o \Pi_o Π o from Lemma 4.2, we have
E exp ( − Π ∞ ( u 1 ) − Π f ( u 2 ) ) = E ( ( ∫ e − u d m ) N o ) = exp ( − p ∫ ( 1 − e − u ) d m ) \mathbb{E}\exp\bigl(-\Pi^\infty(u_1)-\Pi^{\mathrm f}(u_2)\bigr)
=\mathbb{E}\left(\left(\int e^{-u}\,\mathrm{d}m\right)^{N_o}\right)
=\exp\left(-p\int(1-e^{-u})\,\mathrm{d}m\right) E exp ( − Π ∞ ( u 1 ) − Π f ( u 2 ) ) = E ( ( ∫ e − u d m ) N o ) = exp ( − p ∫ ( 1 − e − u ) d m ) = exp ( − p ∫ M ∞ ( 1 − e − u 1 ) d m ) exp ( − p ∫ M f ( 1 − e − u 2 ) d m ) . =\exp\left(-p\int_{M^\infty}(1-e^{-u_1})\,\mathrm{d}m\right)
\exp\left(-p\int_{M^{\mathrm f}}(1-e^{-u_2})\,\mathrm{d}m\right). = exp ( − p ∫ M ∞ ( 1 − e − u 1 ) d m ) exp ( − p ∫ M f ( 1 − e − u 2 ) d m ) . The right side is the joint Laplace functional of two independent Poisson processes with intensity measures p m ( ⋅ ∩ M ∞ ) = p ϱ p m ∞ p\,m(\,\cdot\cap M^\infty)=p\varrho_p m^\infty p m ( ⋅ ∩ M ∞ ) = p ϱ p m ∞ and p m ( ⋅ ∩ M f ) = p ( 1 − ϱ p ) m f p\,m(\,\cdot\cap M^{\mathrm f})=p(1-\varrho_p)m^{\mathrm f} p m ( ⋅ ∩ M f ) = p ( 1 − ϱ p ) m f . Since the joint Laplace functional determines the joint law of ( Π ∞ , Π f ) (\Pi^\infty,\Pi^{\mathrm f}) ( Π ∞ , Π f ) (see [94, Chapter 2]; alternatively, this is the restriction theorem [94, Theorem 5.2]), the first claim follows; ζ o \zeta_o ζ o is independent of Π o \Pi_o Π o by Lemma 4.2. Since ∣ V ( ω ) ∣ = 1 + ∑ k = 1 N o ∣ V ( θ k ω ) ∣ |V(\omega)|=1+\sum_{k=1}^{N_o}|V(\theta_k\omega)| ∣ V ( ω ) ∣ = 1 + ∑ k = 1 N o ∣ V ( θ k ω ) ∣ and N o < ∞ N_o<\infty N o < ∞ , the tree is infinite if
and only if the forward subtree of some child of the root is infinite, that is, S = { N ∞ ≥ 1 } \mathcal{S}=\{N^\infty\ge1\} S = { N ∞ ≥ 1 } . Then ϱ p = P ( N ∞ ≥ 1 ) = 1 − e − p ϱ p \varrho_p=\mathbb{P}(N^\infty\ge1)=1-e^{-p\varrho_p} ϱ p = P ( N ∞ ≥ 1 ) = 1 − e − p ϱ p .
For (ii), let A A A , B B B , C C C be measurable sets. Since S = { Π ∞ ≠ 0 } \mathcal{S}=\{\Pi^\infty\ne0\} S = { Π ∞ = 0 } , part (i) implies that
P ∞ ( Π ∞ ∈ A , Π f ∈ B , ζ o ∈ C ) = ϱ p − 1 P ( Π ∞ ∈ A , Π ∞ ≠ 0 ) P ( Π f ∈ B ) P ( ζ o ∈ C ) , \mathbb{P}^{\infty}(\Pi^\infty\in A,\Pi^{\mathrm{f}}\in B,\zeta_o\in C)=\varrho_p^{-1}\mathbb{P}(\Pi^\infty\in A,\Pi^\infty\ne0)\mathbb{P}(\Pi^{\mathrm{f}}\in B)\mathbb{P}(\zeta_o\in C), P ∞ ( Π ∞ ∈ A , Π f ∈ B , ζ o ∈ C ) = ϱ p − 1 P ( Π ∞ ∈ A , Π ∞ = 0 ) P ( Π f ∈ B ) P ( ζ o ∈ C ) ,
and ϱ p = P ( Π ∞ ≠ 0 ) \varrho_p=\mathbb{P}(\Pi^\infty\ne0) ϱ p = P ( Π ∞ = 0 ) . This is (ii). Part (iii) is the definition of P ∞ \mathbb{P}^{\infty} P ∞ together with Z 1 1 S ≤ Z Z_1\mathbb{1}_{\mathcal{S}}\le Z Z 1 1 S ≤ Z .
For (iv), we note that the integrand on the left side vanishes unless N ∞ ≥ k ≥ 1 N^\infty\ge k\ge1 N ∞ ≥ k ≥ 1 , that is, unless ω ∈ S \omega\in\mathcal{S} ω ∈ S . By (iii), the left side equals ϱ p − 1 \varrho_p^{-1} ϱ p − 1 times the corresponding expectation under P \mathbb{P} P . The measure ( Π ∞ ) ⊗ k (\Pi^\infty)^{\otimes k} ( Π ∞ ) ⊗ k is the restriction of Π o ⊗ k \Pi_o^{\otimes k} Π o ⊗ k to ( M ∞ ) k (\mathcal{M}^\infty)^k ( M ∞ ) k . Applying (4.2) to Π o \Pi_o Π o and to the function F ( m 1 , … , m k , π ) ∏ i = 1 k 1 M ∞ ( m i ) F(m_1,\ldots,m_k,\pi)\prod_{i=1}^k\mathbb{1}_{\mathcal{M}^\infty}(m_i) F ( m 1 , … , m k , π ) ∏ i = 1 k 1 M ∞ ( m i ) , we find that the left side of (4.4) equals
ϱ p − 1 p k ∫ E F ( m 1 , … , m k , Π o ) ∏ i = 1 k 1 M ∞ ( m i ) m ⊗ k ( d m 1 ⋯ d m k ) . \varrho_p^{-1}p^k\int\mathbb{E}F(m_1,\ldots,m_k,\Pi_o)\prod_{i=1}^k\mathbb{1}_{\mathcal{M}^\infty}(m_i)\,m^{\otimes k}(dm_1\cdots dm_k). ϱ p − 1 p k ∫ E F ( m 1 , … , m k , Π o ) i = 1 ∏ k 1 M ∞ ( m i ) m ⊗ k ( d m 1 ⋯ d m k ) .
Since 1 M ∞ m = ϱ p m ∞ \mathbb{1}_{\mathcal{M}^\infty}m=\varrho_p m^\infty 1 M ∞ m = ϱ p m ∞ , this is the right side of (4.4).
Spine decompositions
We now define the laws of the spine trees and prove the many-to-one and many-to-two formulas. They are instances of the Palm (Mecke) formula for the Poisson process; see Remark 4.9. The idea is the classical size-biasing of the offspring numbers along a path. In a sum over the vertices v v v of generation L L L , the vertex v v v exists only if each of its ancestors has enough children; weighting by this event replaces the law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) of each ancestor’s offspring number by the law of 1 + Pois ( p ) 1+\operatorname{Pois}(p) 1 + Pois ( p ) , at the cost of a factor p p p per generation, and makes the position of the next ancestor among its siblings uniform.
Definition 4.6. Let L ≥ 0 L\ge0 L ≥ 0 .
(a) One-spine law. Let ω ′ \omega' ω ′ be a random configuration with law P \mathbb{P} P , let N 0 ′ , … , N L − 1 ′ N'_0,\ldots,N'_{L-1} N 0 ′ , … , N L − 1 ′ be random variables with law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) , and let I 1 , … , I L I_1,\ldots,I_L I 1 , … , I L be random variables such that ω ′ , N 0 ′ , … , N L − 1 ′ \omega',N'_0,\ldots,N'_{L-1} ω ′ , N 0 ′ , … , N L − 1 ′ are independent and, conditionally on them, I 1 , … , I L I_1,\ldots,I_L I 1 , … , I L are independent and I l + 1 I_{l+1} I l + 1 is uniformly distributed on { 1 , … , N l ′ + 1 } \{1,\ldots,N'_l+1\} { 1 , … , N l ′ + 1 } . Set v = I 1 I 2 ⋯ I L ∈ U L v=I_1I_2\cdots I_L\in\mathbb{U}_L v = I 1 I 2 ⋯ I L ∈ U L , so that v l = I 1 ⋯ I l v_l=I_1\cdots I_l v l = I 1 ⋯ I l , and let ω \omega ω be the configuration obtained from ω ′ \omega' ω ′ by setting N v l = N l ′ + 1 N_{v_l}=N'_l+1 N v l = N l ′ + 1 for 0 ≤ l < L 0\le l<L 0 ≤ l < L , all other coordinates of ω \omega ω being equal to those of ω ′ \omega' ω ′ . The one-spine law Q L Q_L Q L is the law of ( ω , v ) (\omega,v) ( ω , v ) on Ω × U L \Omega\times\mathbb{U}_L Ω × U L .
(b) Two-spine law. Let 0 ≤ j < L 0\le j<L 0 ≤ j < L . Let ω ′ \omega' ω ′ be a random configuration with law P \mathbb{P} P , and let N l ′ N'_l N l ′ (0 ≤ l ≤ j 0\le l\le j 0 ≤ l ≤ j ) and N l ( 1 ) , N l ( 2 ) N_l^{(1)},N_l^{(2)} N l ( 1 ) , N l ( 2 ) (j < l < L j<l<L j < l < L ) be random variables with law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) , all of them independent. Conditionally on these, let the following random variables be independent: for 0 ≤ l < j 0\le l<j 0 ≤ l < j , a variable I l + 1 I_{l+1} I l + 1 uniform on { 1 , … , N l ′ + 1 } \{1,\ldots,N'_l+1\} { 1 , … , N l ′ + 1 } ; a pair ( J 1 , J 2 ) (J_1,J_2) ( J 1 , J 2 )
uniform on the set of ordered pairs of distinct elements of { 1 , … , N j ′ + 2 } \{1,\ldots,N'_j+2\} { 1 , … , N j ′ + 2 } ; and for j < l < L j<l<L j < l < L and i ∈ { 1 , 2 } i\in\{1,2\} i ∈ { 1 , 2 } , a variable I l + 1 ( i ) I^{(i)}_{l+1} I l + 1 ( i ) uniform on { 1 , … , N l ( i ) + 1 } \{1,\ldots,N^{(i)}_l+1\} { 1 , … , N l ( i ) + 1 } . Set
u = I 1 ⋯ I j J 1 I j + 2 ( 1 ) ⋯ I L ( 1 ) , v = I 1 ⋯ I j J 2 I j + 2 ( 2 ) ⋯ I L ( 2 ) , \mathbf{u}=I_1\cdots I_jJ_1I^{(1)}_{j+2}\cdots I^{(1)}_L,\qquad\mathbf{v}=I_1\cdots I_jJ_2I^{(2)}_{j+2}\cdots I^{(2)}_L, u = I 1 ⋯ I j J 1 I j + 2 ( 1 ) ⋯ I L ( 1 ) , v = I 1 ⋯ I j J 2 I j + 2 ( 2 ) ⋯ I L ( 2 ) ,
and let ω \omega ω be obtained from ω ′ \omega' ω ′ by setting N v l = N l ′ + 1 N_{v_l}=N'_l+1 N v l = N l ′ + 1 for 0 ≤ l < j 0\le l<j 0 ≤ l < j , N v j = N j ′ + 2 N_{v_j}=N'_j+2 N v j = N j ′ + 2 , and N u l = N l ( 1 ) + 1 N_{u_l}=N^{(1)}_l+1 N u l = N l ( 1 ) + 1 , N v l = N l ( 2 ) + 1 N_{v_l}=N^{(2)}_l+1 N v l = N l ( 2 ) + 1 for j < l < L j<l<L j < l < L , all other coordinates being equal to those of ω ′ \omega' ω ′ . The two-spine law Q L , j Q_{L,j} Q L , j is the law of ( ω , u , v ) (\omega,\mathbf{u},\mathbf{v}) ( ω , u , v ) on Ω × U L × U L \Omega\times U_L\times U_L Ω × U L × U L .
(c) Backbone laws. Let p > 1 p>1 p > 1 . We set
Q L ∞ = Q L ( ⋅ ∣ θ v ω ∈ S ) , Q L , j ∞ = Q L , j ( ⋅ ∣ θ u ω ∈ S , θ v ω ∈ S ) . Q^\infty_L=Q_L(\,\cdot\mid\theta_{\mathbf{v}}\omega\in\mathcal{S}),\qquad Q^\infty_{L,j}=Q_{L,j}(\,\cdot\mid\theta_{\mathbf{u}}\omega\in\mathcal{S},\ \theta_{\mathbf{v}}\omega\in\mathcal{S}). Q L ∞ = Q L ( ⋅ ∣ θ v ω ∈ S ) , Q L , j ∞ = Q L , j ( ⋅ ∣ θ u ω ∈ S , θ v ω ∈ S ) .
We write E Q \mathbb{E}_Q E Q for the expectation with respect to one of these laws Q Q Q , and we use ( ω , v ) (\omega,\mathbf{v}) ( ω , v ) , respectively ( ω , u , v ) (\omega,\mathbf{u},\mathbf{v}) ( ω , u , v ) , for the canonical random elements; the labels I l I_l I l , J i J_i J i , I l ( i ) I^{(i)}_l I l ( i ) are the letters of u \mathbf{u} u and v \mathbf{v} v .
Note that under Q L Q_L Q L the vertex v \mathbf{v} v belongs to V ( ω ) V(\omega) V ( ω ) , because I l + 1 ≤ N v l I_{l+1}\le N_{v_l} I l + 1 ≤ N v l for every l < L l<L l < L . We call the path o = v 0 , v 1 , … , v L = v o=v_0,v_1,\ldots,v_L=\mathbf{v} o = v 0 , v 1 , … , v L = v the spine, the edges { v l − 1 , v l } \{v_{l-1},v_l\} { v l − 1 , v l } the spine edges, v \mathbf{v} v the endpoint, and T ( ω ) \mathcal{T}(\omega) T ( ω ) under Q L Q_L Q L the spine tree of length L L L . For l < L l<L l < L , the children of v l v_l v l other than v l + 1 v_{l+1} v l + 1 are the side children of v l v_l v l ; there are N l ′ N'_l N l ′ of them. Under Q L , j Q_{L,j} Q L , j , both u \mathbf{u} u and v \mathbf{v} v belong to V ( ω ) V(\omega) V ( ω ) , and their last common ancestor u ∧ v = v j = u j \mathbf{u}\wedge\mathbf{v}=v_j=u_j u ∧ v = v j = u j , the fork, has generation j j j , because J 1 ≠ J 2 J_1\ne J_2 J 1 = J 2 . The fork has N j ′ N'_j N j ′ side children besides u j + 1 u_{j+1} u j + 1 and v j + 1 v_{j+1} v j + 1 . The well-definedness of the backbone laws follows from the next lemma, which shows that the endpoint configurations have law P \mathbb{P} P , so that θ v ω ∈ S \theta_{\mathbf{v}}\omega\in\mathcal{S} θ v ω ∈ S has probability ϱ p > 0 \varrho_p>0 ϱ p > 0 .
Lemma 4.7. Let L ≥ 0 L\ge0 L ≥ 0 , and let M \mathrm{M} M and m = γ ⊗ P \mathfrak{m}=\gamma\otimes\mathbb{P} m = γ ⊗ P be as in Lemma 4.2. Under Q L Q_L Q L , for 0 ≤ l < L 0\le l<L 0 ≤ l < L , let S l S_l S l be the finite sequence of the pairs ( w v l k , θ v l k ω ) (w_{v_lk},\theta_{v_lk}\omega) ( w v l k , θ v l k ω ) , where k k k runs in increasing order through { 1 , … , N v l } ∖ { I l + 1 } \{1,\ldots,N_{v_l}\}\setminus\{I_{l+1}\} { 1 , … , N v l } ∖ { I l + 1 } , and let Σ l ∈ N ( M ) \Sigma_l\in\mathbb{N}(\mathrm{M}) Σ l ∈ N ( M ) be the counting measure whose atoms are the entries of S l S_l S l .
(i) The random elements
w v 1 , … , w v L , ζ v 0 , … , ζ v L − 1 , S 0 , … , S L − 1 , θ v ω w_{v_1},\ldots,w_{v_L},\qquad\zeta_{v_0},\ldots,\zeta_{v_{L-1}},\qquad S_0,\ldots,S_{L-1},\qquad\theta_{\mathbf{v}}\omega w v 1 , … , w v L , ζ v 0 , … , ζ v L − 1 , S 0 , … , S L − 1 , θ v ω are independent. Each w v l w_{v_l} w v l has law γ \gamma γ , each ζ v l \zeta_{v_l} ζ v l has law κ \kappa κ , and θ v ω \theta_{\mathbf{v}}\omega θ v ω has law P \mathbb{P} P . Each S l S_l S l has length N l ′ N'_l N l ′ , with law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) , and conditionally on its length its entries are independent with law m \mathfrak{m} m . In particular, Σ 0 , … , Σ L − 1 \Sigma_0,\ldots,\Sigma_{L-1} Σ 0 , … , Σ L − 1 are independent Poisson processes on M \mathrm{M} M with intensity measure p m p\mathfrak{m} p m .
(ii) Conditionally on the random elements in (i), the labels I 1 , … , I L I_1,\ldots,I_L I 1 , … , I L are independent, and I l + 1 I_{l+1} I l + 1 is uniform on { 1 , … , ∣ S l ∣ + 1 } \{1,\ldots,|S_l|+1\} { 1 , … , ∣ S l ∣ + 1 } , where ∣ S l ∣ |S_l| ∣ S l ∣ is the length of S l S_l S l .
(iii) Let p > 1 p>1 p > 1 . Under Q L ∞ Q^\infty_L Q L ∞ , assertions (i) and (ii) hold with the single change that θ v ω \theta_{\mathbf{v}}\omega θ v ω has law P ∞ \mathbb{P}^\infty P ∞ .
(iv) Let 0 ≤ j < L 0 \le j < L 0 ≤ j < L . Under Q L , j Q_{L,j} Q L , j , define sequences S S S as above at the vertices v l v_l v l with l < j l < j l < j (omitting the child v l + 1 v_{l+1} v l + 1 ), at the fork v j v_j v j (omitting the two children u j + 1 u_{j+1} u j + 1 and v j + 1 v_{j+1} v j + 1 ), and at the vertices u l u_l u l and v l v_l v l with j < l < L j < l < L j < l < L (omitting u l + 1 u_{l+1} u l + 1 , respectively v l + 1 v_{l+1} v l + 1 ). Then the 2 L − j 2L-j 2 L − j spine edge marks w v l w_{v_l} w v l (1 ≤ l ≤ L 1 \le l \le L 1 ≤ l ≤ L ) and w u l w_{u_l} w u l (j < l ≤ L j < l \le L j < l ≤ L ), the auxiliary marks of the 2 L − j − 1 2L-j-1 2 L − j − 1 spine vertices other than u u u and v v v , the 2 L − j − 1 2L-j-1 2 L − j − 1 sequences S S S , and the two endpoint configurations θ u ω \theta_u\omega θ u ω and θ v ω \theta_v\omega θ v ω are independent, with the laws described in (i). Under Q ∞ , j ∞ Q_{\infty,j}^{\infty} Q ∞ , j ∞ (p > 1 p > 1 p > 1 ), the same holds, except that the two endpoint configurations have law P ∞ \mathbb{P}^{\infty} P ∞ .
The marked spine tree is assembled from the random elements in (i) and the labels: the spine edges carry the marks w v l w_{v_l} w v l , the side children of v l v_l v l carry the entries of S l S_l S l (edge mark and forward configuration), and the endpoint carries the forward configuration θ v ω \theta_v\omega θ v ω . In particular, under Q L Q_L Q L the squared moduli Y v 1 , … , Y v L Y_{v_1}, \ldots, Y_{v_L} Y v 1 , … , Y v L of the spine marks are independent Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) variables, every side child has a forward subtree which is an independent Poisson–Galton–Watson tree with independent marks, and the endpoint has an independent forward subtree with the law of the whole tree.
Proof. Fix n = ( n 0 , … , n L − 1 ) ∈ { 0 , 1 , … } L n=(n_0,\ldots,n_{L-1}) \in\{0,1,\ldots\}^{L} n = ( n 0 , … , n L − 1 ) ∈ { 0 , 1 , … } L and i = ( i 1 , … , i L ) i=(i_1,\ldots,i_L) i = ( i 1 , … , i L ) with 1 ≤ i l + 1 ≤ n l + 1 1 \le i_{l+1} \le n_l+1 1 ≤ i l + 1 ≤ n l + 1 , and let v = i 1 ⋯ i L v=i_1\cdots i_L v = i 1 ⋯ i L . On the event E n , i = { N ′ = n , I = i } E_{n,i}=\{N'=n,I=i\} E n , i = { N ′ = n , I = i } , which has probability ∏ l < L P ( N = n l ) ( n l + 1 ) − 1 \prod_{l<L}\mathbb{P}(N=n_l)(n_l+1)^{-1} ∏ l < L P ( N = n l ) ( n l + 1 ) − 1 for a Pois ( p ) \operatorname{Pois}(p) Pois ( p ) variable N N N , we have v = v v=v v = v , and ω \omega ω is obtained from ω ′ \omega' ω ′ by setting N v l = n l + 1 N_{v_l}=n_l+1 N v l = n l + 1 for l < L l<L l < L . Since the pair ( N ′ , I ) (N',I) ( N ′ , I ) is independent of ω ′ \omega' ω ′ , the configuration ω ′ \omega' ω ′ still has law P \mathbb{P} P conditionally on E n , i E_{n,i} E n , i .
We note that the modified coordinates N v 0 , … , N v L − 1 N_{v_0},\ldots,N_{v_{L-1}} N v 0 , … , N v L − 1 are not coordinates of θ v l k ω \theta_{v_l^k}\omega θ v l k ω for k ≠ i l + 1 k \ne i_{l+1} k = i l + 1 , because no v l ′ v_l' v l ′ is a descendant of v l k v_l^k v l k ; they are not coordinates of θ v ω \theta_v\omega θ v ω either. We conclude that, on E n , i E_{n,i} E n , i , the random elements in (i) are equal to
w v l ′ ( 1 ≤ l ≤ L ) , ζ v l ′ ( 0 ≤ l < L ) , θ v ω ′ , w'_{v_l}\ (1 \le l \le L), \qquad\zeta'_{v_l}\ (0 \le l < L), \qquad\theta_v\omega', w v l ′ ( 1 ≤ l ≤ L ) , ζ v l ′ ( 0 ≤ l < L ) , θ v ω ′ ,
and, for 0 ≤ l < L 0 \le l < L 0 ≤ l < L , the sequence of pairs ( w v l k ′ , θ v l k ω ′ ) (w'_{v_l^k},\theta_{v_l^k}\omega') ( w v l k ′ , θ v l k ω ′ ) with k k k running in increasing order through { 1 , … , n l + 1 } ∖ { i l + 1 } \{1,\ldots,n_l+1\}\setminus\{i_{l+1}\} { 1 , … , n l + 1 } ∖ { i l + 1 } . These are functions of pairwise disjoint sets of coordinates of ω ′ \omega' ω ′ . Indeed, the forward configurations of distinct side children and of v v v use the coordinates at pairwise disjoint sets of labels, none of which is one of v 0 , … , v L − 1 v_0,\ldots,v_{L-1} v 0 , … , v L − 1 , and they do not use the edge marks w v l w_{v_l} w v l and w v l k w_{v_l^k} w v l k , which appear separately. Since the coordinates of ω ′ \omega' ω ′ are independent, these random elements are, conditionally on E n , i E_{n,i} E n , i , independent with laws γ \gamma γ , κ \kappa κ , m ⊗ n l \mathfrak{m}^{\otimes n_l} m ⊗ n l , and P \mathbb{P} P , respectively. The conditional joint law depends on ( n , i ) (n,i) ( n , i ) only through the lengths n l n_l n l . Summing over i i i and then over n n n , and using that N 0 ′ , … , N L − 1 ′ N'_0,\ldots,N'_{L-1} N 0 ′ , … , N L − 1 ′ are independent with law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) , we find (i), except for its last sentence. Further, the conditional probability of { I = i } \{I=i\} { I = i } given N ′ = n N'=n N ′ = n and the random elements in (i) equals ∏ l < L ( n l + 1 ) − 1 \prod_{l<L}(n_l+1)^{-1} ∏ l < L ( n l + 1 ) − 1 , and this is (ii), since n l = ∣ S l ∣ n_l=|S_l| n l = ∣ S l ∣ . The process Σ l \Sigma_l Σ l is a mixed binomial process with a Pois ( p ) \operatorname{Pois}(p) Pois ( p ) number of independent m \mathfrak{m} m -distributed points, hence a Poisson process with intensity measure p m p\mathfrak{m} p m [94, Proposition 3.5], and the processes Σ l \Sigma_l Σ l are independent because the sequences S l S_l S l are.
For (iii), let g g g be a nonnegative measurable function of the random elements in (i) other than θ v ω \theta_{v\omega} θ v ω , together with the labels, and let h : Ω → [ 0 , ∞ ] h:\Omega\to[0,\infty] h : Ω → [ 0 , ∞ ] be measurable. By (i), (ii), and Q L ( θ v ω ∈ S ) = P ( S ) = ϱ p \mathbb{Q}_L(\theta_{v\omega}\in S)=\mathbb{P}(S)=\varrho_p Q L ( θ v ω ∈ S ) = P ( S ) = ϱ p , we have
E Q L [ g h ( θ v ω ) ] = ϱ p − 1 E Q L [ g h ( θ v ω ) 1 S ( θ v ω ) ] = E Q L [ g ] E ∞ h , \mathbb{E}_{\mathbb{Q}_L}\left[g h(\theta_{v\omega})\right]=\varrho_p^{-1}\mathbb{E}_{\mathbb{Q}_L}\left[g h(\theta_{v\omega})\mathbf{1}_S(\theta_{v\omega})\right]=\mathbb{E}_{\mathbb{Q}_L}[g]\mathbb{E}^{\infty}h, E Q L [ g h ( θ v ω ) ] = ϱ p − 1 E Q L [ g h ( θ v ω ) 1 S ( θ v ω ) ] = E Q L [ g ] E ∞ h ,
which is (iii). The proof of (iv) is the same. The only change is at the fork: on the corresponding event E E E , we have N v j = n j + 2 N_{v_j}=n_j+2 N v j = n j + 2 , the two omitted children are v j J 1 v_{jJ_1} v j J 1 and v j J 2 v_{jJ_2} v j J 2 , and ( J 1 , J 2 ) (J_1,J_2) ( J 1 , J 2 ) takes each of the ( n j + 2 ) ( n j + 1 ) (n_j+2)(n_j+1) ( n j + 2 ) ( n j + 1 ) possible values with the same conditional probability. The backbone version follows as in (iii), since the two endpoint configurations are independent with law P \mathbb{P} P under Q L , j \mathbb{Q}_{L,j} Q L , j .
Proposition 4.8. Let p > 0 p>0 p > 0 and L ≥ 0 L\geq0 L ≥ 0 . All functions below are measurable; a function on Ω × U \Omega\times\mathcal{U} Ω × U is measurable if it is measurable in ω \omega ω for each fixed word.
(i) For every F : Ω × U → [ 0 , ∞ ] F:\Omega\times\mathcal{U}\to[0,\infty] F : Ω × U → [ 0 , ∞ ] ,
E ∑ v ∈ V L F ( ω , v ) = p L E Q L F ( ω , v ) . (38) \mathbb{E}\sum_{v\in V_L}F(\omega,v)=p^L\mathbb{E}_{\mathbb{Q}_L}F(\omega,v). \tag*{(38)} E v ∈ V L ∑ F ( ω , v ) = p L E Q L F ( ω , v ) . (38) (ii) Let 0 ≤ j < L 0\leq j<L 0 ≤ j < L . For every F : Ω × U × U → [ 0 , ∞ ] F:\Omega\times\mathcal{U}\times\mathcal{U}\to[0,\infty] F : Ω × U × U → [ 0 , ∞ ] ,
E ∑ u , v ∈ V L ∣ u ∧ v ∣ = j F ( ω , u , v ) = p 2 L − j E Q L , j F ( ω , u , v ) . (39) \mathbb{E}\sum_{\substack{u,v\in V_L\\|u\wedge v|=j}}F(\omega,u,v)=p^{2L-j}\mathbb{E}_{\mathbb{Q}_{L,j}}F(\omega,u,v). \tag*{(39)} E u , v ∈ V L ∣ u ∧ v ∣ = j ∑ F ( ω , u , v ) = p 2 L − j E Q L , j F ( ω , u , v ) . (39) (iii) Let p > 1 p>1 p > 1 . For every F F F as in (i),
E ∞ ∑ v ∈ V L F ( ω , v ) = ϱ p − 1 p L E Q L [ F ( ω , v ) 1 S ( ω ) ] ≤ ϱ p − 1 p L E Q L F ( ω , v ) , (40) \mathbb{E}^{\infty}\sum_{v\in V_L}F(\omega,v)=\varrho_p^{-1}p^L\mathbb{E}_{\mathbb{Q}_L}\left[F(\omega,v)\mathbf{1}_S(\omega)\right]\leq\varrho_p^{-1}p^L\mathbb{E}_{\mathbb{Q}_L}F(\omega,v), \tag*{(40)} E ∞ v ∈ V L ∑ F ( ω , v ) = ϱ p − 1 p L E Q L [ F ( ω , v ) 1 S ( ω ) ] ≤ ϱ p − 1 p L E Q L F ( ω , v ) , (40) E ∞ ∑ v ∈ V L ∞ F ( ω , v ) = p L E Q L ∞ F ( ω , v ) ≤ ϱ p − 1 p L E Q L F ( ω , v ) . (41) \mathbb{E}^{\infty}\sum_{v\in V_L^\infty}F(\omega,v)=p^L\mathbb{E}_{\mathbb{Q}_L^\infty}F(\omega,v)\leq\varrho_p^{-1}p^L\mathbb{E}_{\mathbb{Q}_L}F(\omega,v). \tag*{(41)} E ∞ v ∈ V L ∞ ∑ F ( ω , v ) = p L E Q L ∞ F ( ω , v ) ≤ ϱ p − 1 p L E Q L F ( ω , v ) . (41) (iv) Let p > 1 p>1 p > 1 and 0 ≤ j < L 0\leq j<L 0 ≤ j < L . For every F F F as in (ii),
E ∞ ∑ u , v ∈ V L ∣ u ∧ v ∣ = j F ( ω , u , v ) = ϱ p − 1 p 2 L − j E Q L , j [ F ( ω , u , v ) 1 S ( ω ) ] \mathbb{E}^{\infty}\sum_{\substack{u,v\in V_L\\|u\wedge v|=j}}F(\omega,u,v)=\varrho_p^{-1}p^{2L-j}\mathbb{E}_{\mathbb{Q}_{L,j}}\left[F(\omega,u,v)\mathbf{1}_S(\omega)\right] E ∞ u , v ∈ V L ∣ u ∧ v ∣ = j ∑ F ( ω , u , v ) = ϱ p − 1 p 2 L − j E Q L , j [ F ( ω , u , v ) 1 S ( ω ) ] ≤ ϱ p − 1 p 2 L − j E Q L , j F ( ω , u , v ) , (42) \leq\varrho_p^{-1}p^{2L-j}\mathbb{E}_{\mathbb{Q}_{L,j}}F(\omega,u,v), \tag*{(42)} ≤ ϱ p − 1 p 2 L − j E Q L , j F ( ω , u , v ) , (42) E ∞ ∑ u , v ∈ V L ∞ ∣ u ∧ v ∣ = j F ( ω , u , v ) = ϱ p p 2 L − j E Q L , j ∞ F ( ω , u , v ) \mathbb{E}^{\infty}\sum_{\substack{u,v\in V_L^\infty\\|u\wedge v|=j}}F(\omega,u,v)=\varrho_p p^{2L-j}\mathbb{E}_{\mathbb{Q}_{L,j}^\infty}F(\omega,u,v) E ∞ u , v ∈ V L ∞ ∣ u ∧ v ∣ = j ∑ F ( ω , u , v ) = ϱ p p 2 L − j E Q L , j ∞ F ( ω , u , v ) ≤ ϱ p − 1 p 2 L − j E Q L , j F ( ω , u , v ) . (43) \leq\varrho_p^{-1}p^{2L-j}\mathbb{E}_{\mathbb{Q}_{L,j}}F(\omega,u,v). \tag*{(43)} ≤ ϱ p − 1 p 2 L − j E Q L , j F ( ω , u , v ) . (43)
The sum in (39) runs over ordered pairs of distinct vertices of generation L L L whose last common ancestor has generation j j j ; summing over 0 ≤ j < L 0\leq j<L 0 ≤ j < L gives all ordered pairs of distinct vertices of generation L L L , and the diagonal pairs are covered by (38) . Under the survival-conditioned law, the backbone sums in (41) and (43) have exact representations by the backbone spine laws, in which only the endpoint configurations are conditioned on survival, while the side processes keep their unconditioned Poisson laws (Lemma 4.7). For the full sums, (40) and (42) are exact as well, but the indicator 1 S ( ω ) \mathbf{1}_{S}(\omega) 1 S ( ω ) depends on the side subtrees; in applications we use only the resulting upper bounds.
Proof. We prove the four parts in order.
Step 1. One spine. Fix v = i 1 ⋯ i L ∈ U L v=i_1\cdots i_L\in\mathbb{U}_L v = i 1 ⋯ i L ∈ U L . Then v ∈ V ( ω ) v\in V(\omega) v ∈ V ( ω ) if and only if N v l ≥ i l + 1 N_{v_l}\ge i_{l+1} N v l ≥ i l + 1 for 0 ≤ l < L 0\le l<L 0 ≤ l < L . For n = ( n 0 , … , n L − 1 ) ∈ { 0 , 1 , … } L n=(n_0,\ldots,n_{L-1})\in\{0,1,\ldots\}^{L} n = ( n 0 , … , n L − 1 ) ∈ { 0 , 1 , … } L , let ω v , n \omega^{v,n} ω v , n denote the configuration obtained from ω \omega ω by setting N v l = n l N_{v_l}=n_l N v l = n l for 0 ≤ l < L 0\le l<L 0 ≤ l < L . Under P \mathbb{P} P , the coordinates N v 0 , … , N v L − 1 N_{v_0},\ldots,N_{v_{L-1}} N v 0 , … , N v L − 1 are independent of each other and of the remaining coordinates, and ω v , n \omega^{v,n} ω v , n does not depend on them. Let N N N denote a Pois ( p ) \operatorname{Pois}(p) Pois ( p ) variable. By Tonelli’s theorem, we have
E [ F ( ω , v ) 1 { v ∈ V ( ω ) } ] = ∑ n : n l ≥ i l + 1 ∀ l ∏ l = 0 L − 1 P ( N = n l ) E F ( ω v , n , v ) . \mathbb{E}\left[F(\omega,v)\mathbf{1}_{\{v\in V(\omega)\}}\right]
=
\sum_{\substack{n:\ n_l\ge i_{l+1}\\ \forall l}}
\prod_{l=0}^{L-1}\mathbb{P}(N=n_l)\,\mathbb{E}F(\omega^{v,n},v). E [ F ( ω , v ) 1 { v ∈ V ( ω )} ] = n : n l ≥ i l + 1 ∀ l ∑ l = 0 ∏ L − 1 P ( N = n l ) E F ( ω v , n , v ) .
Since P ( N = n ) = p n − 1 P ( N = n − 1 ) \mathbb{P}(N=n)=p\,n^{-1}\mathbb{P}(N=n-1) P ( N = n ) = p n − 1 P ( N = n − 1 ) for n ≥ 1 n\ge1 n ≥ 1 , the substitution n l = n l ′ + 1 n_l=n_l'+1 n l = n l ′ + 1 yields
E [ F ( ω , v ) 1 { v ∈ V ( ω ) } ] = p L ∑ n ′ ∈ { 0 , 1 , … } L ∏ l = 0 L − 1 P ( N = n l ′ ) 1 { i l + 1 ≤ n l ′ + 1 } n l ′ + 1 E F ( ω v , n ′ + 1 , v ) . \mathbb{E}\left[F(\omega,v)\mathbf{1}_{\{v\in V(\omega)\}}\right]
=
p^L\sum_{n'\in\{0,1,\ldots\}^{L}}
\prod_{l=0}^{L-1}
\frac{\mathbb{P}(N=n_l')\mathbf{1}_{\{i_{l+1}\le n_l'+1\}}}{n_l'+1}
\mathbb{E}F(\omega^{v,n'+1},v). E [ F ( ω , v ) 1 { v ∈ V ( ω )} ] = p L n ′ ∈ { 0 , 1 , … } L ∑ l = 0 ∏ L − 1 n l ′ + 1 P ( N = n l ′ ) 1 { i l + 1 ≤ n l ′ + 1 } E F ( ω v , n ′ + 1 , v ) .
where n ′ + 1 = ( n 0 ′ + 1 , … , n L − 1 ′ + 1 ) n'+1=(n_0'+1,\ldots,n_{L-1}'+1) n ′ + 1 = ( n 0 ′ + 1 , … , n L − 1 ′ + 1 ) . In the construction of Q L \mathbb{Q}_L Q L , the event { N ′ = n ′ , I = ( i 1 , … , i L ) } \{N'=n', I=(i_1,\ldots,i_L)\} { N ′ = n ′ , I = ( i 1 , … , i L )} has probability equal to the product in the last display, and on this event ( ω , v ) = ( ( ω ′ ) v , n ′ + 1 , v ) (\omega,v)=((\omega')^{v,n'+1},v) ( ω , v ) = (( ω ′ ) v , n ′ + 1 , v ) , where ω ′ \omega' ω ′ has law P \mathbb{P} P and is independent of ( N ′ , I ) (N',I) ( N ′ , I ) . Summing the last display over v ∈ U L v\in\mathbb{U}_L v ∈ U L , that is, over ( i 1 , … , i L ) (i_1,\ldots,i_L) ( i 1 , … , i L ) , we obtain
E ∑ v ∈ V L F ( ω , v ) = ∑ v ∈ U L E [ F ( ω , v ) 1 { v ∈ V ( ω ) } ] = p L E Q L F ( ω , v ) , \mathbb{E}\sum_{v\in\mathbb{V}_L}F(\omega,v)
=
\sum_{v\in\mathbb{U}_L}\mathbb{E}\left[F(\omega,v)\mathbf{1}_{\{v\in V(\omega)\}}\right]
=
p^L\mathbb{E}_{\mathbb{Q}_L}F(\omega,v), E v ∈ V L ∑ F ( ω , v ) = v ∈ U L ∑ E [ F ( ω , v ) 1 { v ∈ V ( ω )} ] = p L E Q L F ( ω , v ) ,
which is (38) . In the first equality we used Tonelli’s theorem to exchange the sum over the countable set U L \mathbb{U}_L U L with the expectation.
Step 2. Two spines. Fix u , v ∈ U L u,v\in\mathbb{U}_L u , v ∈ U L with ∣ u ∧ v ∣ = j |u\wedge v|=j ∣ u ∧ v ∣ = j , and write a = u ∧ v a=u\wedge v a = u ∧ v , so that u j + 1 ≠ v j + 1 u_{j+1}\ne v_{j+1} u j + 1 = v j + 1 . Then u , v ∈ V ( ω ) u,v\in V(\omega) u , v ∈ V ( ω ) if and only if the following constraints hold: N a l N_{a_l} N a l is at least the ( l + 1 ) (l+1) ( l + 1 ) -th letter of a a a for 0 ≤ l < j 0\le l<j 0 ≤ l < j ; N a N_a N a is at least the last letters of both u j + 1 u_{j+1} u j + 1 and v j + 1 v_{j+1} v j + 1 ; and for j < l < L j<l<L j < l < L , N u l N_{u_l} N u l and N v l N_{v_l} N v l are at least the last letters of u l + 1 u_{l+1} u l + 1 and v l + 1 v_{l+1} v l + 1 , respectively. These constraints involve the offspring numbers at 2 L − j − 1 2L-j-1 2 L − j − 1 distinct labels. We expand in these offspring numbers as in Step 1. At the j + 2 ( L − j − 1 ) j+2(L-j-1) j + 2 ( L − j − 1 ) labels other than the fork, we use P ( N = n ) = p n − 1 P ( N = n − 1 ) \mathbb{P}(N=n)=p\,n^{-1}\mathbb{P}(N=n-1) P ( N = n ) = p n − 1 P ( N = n − 1 ) . At the fork, the two letters are distinct, which forces N a ≥ 2 N_a\ge2 N a ≥ 2 , and we use P ( N = n ) = p 2 ( n ( n − 1 ) ) − 1 P ( N = n − 2 ) \mathbb{P}(N=n)=p^2(n(n-1))^{-1}\mathbb{P}(N=n-2) P ( N = n ) = p 2 ( n ( n − 1 ) ) − 1 P ( N = n − 2 ) for n ≥ 2 n\ge2 n ≥ 2 . For a fixed value n n n of N a N_a N a , there are n ( n − 1 ) n(n-1) n ( n − 1 ) ordered pairs of distinct letters in { 1 , … , n } \{1,\ldots,n\} { 1 , … , n } , which matches the uniform law of ( J 1 , J 2 ) (J_1,J_2) ( J 1 , J 2 ) in Definition 4.6. Summing over ( u , v ) (u,v) ( u , v ) as in Step 1 and collecting the powers of p p p , namely p j ⋅ p 2 ⋅ p 2 ( L − j − 1 ) = p 2 L − j p^j \cdot p^2 \cdot p^{2(L-j-1)} = p^{2L-j} p j ⋅ p 2 ⋅ p 2 ( L − j − 1 ) = p 2 L − j , we arrive at (39) .
Step 3. Conditioned sums. Let p > 1 p>1 p > 1 . Applying (36) and then (38) to F ( ω , v ) 1 S ( ω ) F(\omega,v)\mathbf{1}_{\mathcal{S}}(\omega) F ( ω , v ) 1 S ( ω ) , we have the identity in (40) ; the inequality follows from 1 S ≤ 1 \mathbf{1}_{\mathcal{S}}\leq1 1 S ≤ 1 . If v ∈ V L ∞ v\in\mathcal{V}^{\infty}_L v ∈ V L ∞ , then the forward subtree of v v v is infinite, and in particular ω ∈ S \omega\in\mathcal{S} ω ∈ S . By (36) and (38) , we have
E p ∞ ∑ v ∈ V L ∞ F ( ω , v ) = ϱ p − 1 E ∑ v ∈ V L F ( ω , v ) 1 S ( θ v ω ) = ϱ p − 1 p L E Q L [ F ( ω , v ) 1 S ( θ v ω ) ] . \mathbb{E}^{\infty}_{p}\sum_{v\in\mathcal{V}^{\infty}_L}F(\omega,v)=\varrho_p^{-1}\mathbb{E}\sum_{v\in\mathcal{V}_L}F(\omega,v)\mathbf{1}_{\mathcal{S}}(\theta_v\omega)=\varrho_p^{-1}p^L\mathbb{E}_{Q_L}\left[F(\omega,v)\mathbf{1}_{\mathcal{S}}(\theta_v\omega)\right]. E p ∞ v ∈ V L ∞ ∑ F ( ω , v ) = ϱ p − 1 E v ∈ V L ∑ F ( ω , v ) 1 S ( θ v ω ) = ϱ p − 1 p L E Q L [ F ( ω , v ) 1 S ( θ v ω ) ] .
By Lemma 4.7, we have Q L ( θ v ω ∈ S ) = ϱ p Q_L(\theta_v\omega\in\mathcal{S})=\varrho_p Q L ( θ v ω ∈ S ) = ϱ p . Then the right side equals p L E Q L ∞ F ( ω , v ) p^L\mathbb{E}_{Q^\infty_L}F(\omega,v) p L E Q L ∞ F ( ω , v ) , and it is at most ϱ p − 1 p L E Q L F ( ω , v ) \varrho_p^{-1}p^L\mathbb{E}_{Q_L}F(\omega,v) ϱ p − 1 p L E Q L F ( ω , v ) . This proves (41) .
Step 4. Conditioned pairs. The identity and inequality in (42) follow from (39) exactly as in Step 3. For (43) , we argue as in Step 3 and obtain
E p ∞ ∑ u , v ∈ V L ∞ ∣ u ∧ v ∣ = j F ( ω , u , v ) = ϱ p − 1 p 2 L − j E Q L , j [ F ( ω , u , v ) 1 S ( θ u ω ) 1 S ( θ v ω ) ] . \mathbb{E}^{\infty}_{p}\sum_{\substack{u,v\in\mathcal{V}^{\infty}_L\\ |u\wedge v|=j}}F(\omega,u,v)=\varrho_p^{-1}p^{2L-j}\mathbb{E}_{Q_{L,j}}\left[F(\omega,u,v)\mathbf{1}_{\mathcal{S}}(\theta_u\omega)\mathbf{1}_{\mathcal{S}}(\theta_v\omega)\right]. E p ∞ u , v ∈ V L ∞ ∣ u ∧ v ∣ = j ∑ F ( ω , u , v ) = ϱ p − 1 p 2 L − j E Q L , j [ F ( ω , u , v ) 1 S ( θ u ω ) 1 S ( θ v ω ) ] .
By Lemma 4.7(iv), the two endpoint configurations are independent with law P \mathbb{P} P under Q L , j Q_{L,j} Q L , j , and the probability that both survive is ϱ p 2 \varrho_p^2 ϱ p 2 . Then the right side equals ϱ p p 2 L − j E Q L , j ∞ F ( ω , u , v ) \varrho_p p^{2L-j}\mathbb{E}_{Q^\infty_{L,j}}F(\omega,u,v) ϱ p p 2 L − j E Q L , j ∞ F ( ω , u , v ) , and it is at most ϱ p − 1 p 2 L − j E Q L , j F ( ω , u , v ) \varrho_p^{-1}p^{2L-j}\mathbb{E}_{Q_{L,j}}F(\omega,u,v) ϱ p − 1 p 2 L − j E Q L , j F ( ω , u , v ) . This completes the proof.
Remark 4.9. The identity n P ( N = n ) = p P ( N = n − 1 ) n\mathbb{P}(N=n)=p\mathbb{P}(N=n-1) n P ( N = n ) = p P ( N = n − 1 ) , used at each ancestor in Step 1, is the Mecke formula (34) for the number of points of a Poisson process. If F ( ω , v ) F(\omega,v) F ( ω , v ) depends on ω \omega ω and v v v only through the marked tree and the auxiliary marks, and is invariant under relabelings of siblings, then (38) can also be obtained by applying (34) to the child process at each of the L L L ancestors, by induction on L L L ; the labels I l I_l I l then play no role, and only Lemma 4.7(i) is needed to evaluate the right side. We have stated the formulas for labelled configurations so that they apply without a separate invariance check, for example to functionals that select children by their labels.
The spine laws are absolutely continuous with respect to P \mathbb{P} P . In particular, every almost sure property of the tree holds along spines.
Corollary 4.10. Let p > 0 p>0 p > 0 , L ≥ 0 L\geq0 L ≥ 0 , and 0 ≤ j < L 0\leq j<L 0 ≤ j < L . The law of ω \omega ω under Q L Q_L Q L has density p − L ∣ V L ∣ p^{-L}|\mathcal{V}_L| p − L ∣ V L ∣ with respect to P \mathbb{P} P , and the law of ω \omega ω under Q L , j Q_{L,j} Q L , j has density p − ( 2 L − j ) ∣ { ( u , v ) ∈ V L 2 : ∣ u ∧ v ∣ = j } ∣ p^{-(2L-j)}\left|\{(u,v)\in\mathcal{V}_L^2:|u\wedge v|=j\}\right| p − ( 2 L − j ) {( u , v ) ∈ V L 2 : ∣ u ∧ v ∣ = j } with respect to P \mathbb{P} P . If p > 1 p>1 p > 1 , the laws of ω \omega ω under Q L ∞ Q^\infty_L Q L ∞ and Q L , j ∞ Q^\infty_{L,j} Q L , j ∞ are absolutely continuous with respect to P ∞ \mathbb{P}^\infty P ∞ . Further, if A ∈ F A\in\mathcal{F} A ∈ F satisfies P ( A ) = 1 \mathbb{P}(A)=1 P ( A ) = 1 , then almost surely under each of the laws P \mathbb{P} P , P ∞ \mathbb{P}^\infty P ∞ , Q L Q_L Q L , Q L , j Q_{L,j} Q L , j , Q L ∞ Q^\infty_L Q L ∞ , and Q L , j ∞ Q^\infty_{L,j} Q L , j ∞ , we have θ u ω ∈ A \theta_u\omega\in A θ u ω ∈ A for all u ∈ U u\in\mathbb{U} u ∈ U simultaneously.
Proof. Applying (38) to F ( ω , v ) = 1 B ( ω ) F(\omega,v)=\mathbf{1}_B(\omega) F ( ω , v ) = 1 B ( ω ) , we find that Q L ( ω ∈ B ) = p − L E [ ∣ V L ∣ 1 B ] Q_L(\omega\in B)=p^{-L}\mathbb{E}[|\mathcal{V}_L|\mathbf{1}_B] Q L ( ω ∈ B ) = p − L E [ ∣ V L ∣ 1 B ] for B ∈ F B\in\mathcal{F} B ∈ F , which is the first claim. The second follows in the same way from (39) . For p > 1 p>1 p > 1 , (4.8) implies that Q L ∞ ( ω ∈ B ) = p − L E ∞ [ ∣ V L ∞ ∣ 1 B ] Q^\infty_L(\omega\in B)=p^{-L}\mathbb{E}^\infty[|V^\infty_L|\mathbf{1}_B] Q L ∞ ( ω ∈ B ) = p − L E ∞ [ ∣ V L ∞ ∣ 1 B ] , and (4.10) gives the analogous formula for Q L , j ∞ Q^\infty_{L,j} Q L , j ∞ . Finally, P ( θ u − 1 A ) = P ( A ) = 1 \mathbb{P}(\theta_u^{-1}A)=\mathbb{P}(A)=1 P ( θ u − 1 A ) = P ( A ) = 1 for every u u u , because θ u \theta_u θ u preserves P \mathbb{P} P . The countable intersection ⋂ u ∈ U θ u − 1 A \bigcap_{u\in\mathbb{U}}\theta_u^{-1}A ⋂ u ∈ U θ u − 1 A therefore has P \mathbb{P} P -probability one, and by absolute continuity it has probability one under the other laws.
For instance, the conclusions of Theorem 3.10 , Proposition 3.16, and Proposition 3.18 hold almost surely under all spine laws.
We next describe recursively defined variables along a spine. This is the form in which the many-to-one formula enters the computation of path moments.
Corollary 4.11. Let p > 0 p>0 p > 0 and L ≥ 1 L\geq1 L ≥ 1 . Let ( X , X ) (X,\mathcal{X}) ( X , X ) be a measurable space, let f 0 , … , f L : Ω → X f_0,\ldots,f_L:\Omega\to X f 0 , … , f L : Ω → X be measurable, and let Ψ 0 , … , Ψ L − 1 : Z × N ( C × X ) → X \Psi_0,\ldots,\Psi_{L-1}:\mathbb{Z}\times\mathcal{N}(\mathcal{C}\times X)\to X Ψ 0 , … , Ψ L − 1 : Z × N ( C × X ) → X be measurable maps such that, for P \mathbb{P} P -almost every ω \omega ω and every 0 ≤ l < L 0\leq l<L 0 ≤ l < L ,
f l ( ω ) = Ψ l ( ζ ω , ∑ k = 1 N ∘ δ ( w k , f l + 1 ( θ k ω ) ) ) . (44) f_l(\omega)=\Psi_l\left(\zeta_\omega,\sum_{k=1}^{N_\circ}\delta_{(w_k,f_{l+1}(\theta_k\omega))}\right). \tag*{(44)} f l ( ω ) = Ψ l ( ζ ω , k = 1 ∑ N ∘ δ ( w k , f l + 1 ( θ k ω )) ) . (44) Under Q L Q_L Q L , set X l = f l ( θ v l ω ) X_l=f_l(\theta_{v_l}\omega) X l = f l ( θ v l ω ) for 0 ≤ l ≤ L 0\leq l\leq L 0 ≤ l ≤ L , and
Σ l f = ∑ k ∈ { 1 , … , N v l } ∖ { I l + 1 } δ ( w v l k , f l + 1 ( θ v l k ω ) ) , 0 ≤ l < L . \Sigma_l^f=\sum_{k\in\{1,\ldots,N_{v_l}\}\setminus\{I_{l+1}\}}\delta_{(w_{v_lk},f_{l+1}(\theta_{v_lk}\omega))},\qquad0\leq l<L. Σ l f = k ∈ { 1 , … , N v l } ∖ { I l + 1 } ∑ δ ( w v l k , f l + 1 ( θ v l k ω )) , 0 ≤ l < L . (i) The random elements X L X_L X L , w v 1 , … , w v L w_{v_1},\ldots,w_{v_L} w v 1 , … , w v L , ζ v 0 , … , ζ v L − 1 \zeta_{v_0},\ldots,\zeta_{v_{L-1}} ζ v 0 , … , ζ v L − 1 , and Σ 0 f , … , Σ L − 1 f \Sigma_0^f,\ldots,\Sigma_{L-1}^f Σ 0 f , … , Σ L − 1 f are independent. The variable X L X_L X L has law P ∘ f L − 1 \mathbb{P}\circ f_L^{-1} P ∘ f L − 1 , and Σ l f \Sigma_l^f Σ l f is a Poisson process on C × X \mathcal{C}\times X C × X with intensity measure p γ ⊗ ( P ∘ f l + 1 − 1 ) p\gamma\otimes(\mathbb{P}\circ f_{l+1}^{-1}) p γ ⊗ ( P ∘ f l + 1 − 1 ) .
(ii) Almost surely, X l = Ψ l ( ζ v l , Σ l f + δ ( w v l + 1 , X l + 1 ) ) X_l=\Psi_l(\zeta_{v_l},\Sigma_l^f+\delta_{(w_{v_{l+1}},X_{l+1})}) X l = Ψ l ( ζ v l , Σ l f + δ ( w v l + 1 , X l + 1 ) ) for 0 ≤ l < L 0\leq l<L 0 ≤ l < L .
(iii) Let p > 1 p>1 p > 1 . Under Q L ∞ Q^\infty_L Q L ∞ , (i) and (ii) hold with the single change that X L X_L X L has law P ∞ ∘ f L − 1 \mathbb{P}^\infty\circ f_L^{-1} P ∞ ∘ f L − 1 . Further, for every measurable Φ : X L + 1 × C L → [ 0 , ∞ ] \Phi:X^{L+1}\times\mathcal{C}^L\to[0,\infty] Φ : X L + 1 × C L → [ 0 , ∞ ] ,
E ∑ v ∈ V L Φ ( ( f l ( θ v l ω ) ) l = 0 L , ( w v l ) l = 1 L ) = p L E Q L Φ ( ( X l ) l = 0 L , ( w v l ) l = 1 L ) , (45) \mathbb{E}\sum_{v\in V_L}\Phi\left(\left(f_l(\theta_{v_l}\omega)\right)_{l=0}^{L},(w_{v_l})_{l=1}^{L}\right)
=p^L\mathbb{E}_{Q_L}\Phi\left((X_l)_{l=0}^{L},(w_{v_l})_{l=1}^{L}\right), \tag*{(45)} E v ∈ V L ∑ Φ ( ( f l ( θ v l ω ) ) l = 0 L , ( w v l ) l = 1 L ) = p L E Q L Φ ( ( X l ) l = 0 L , ( w v l ) l = 1 L ) , (45) and the right side is the expectation of a function of the independent random elements in (i), computed through the backward recursion in (ii). The same holds under P ∞ \mathbb{P}^\infty P ∞ , with the sum over V L ∞ V^\infty_L V L ∞ and Q L ∞ Q^\infty_L Q L ∞ in place of Q L Q_L Q L .
Proof. The process Σ l f \Sigma_l^f Σ l f is the image of the process Σ l \Sigma_l Σ l of Lemma 4.7 under the measurable map ( w , ω ′ ′ ) ↦ ( w , f l + 1 ( ω ′ ′ ) ) (w,\omega'')\mapsto(w,f_{l+1}(\omega'')) ( w , ω ′′ ) ↦ ( w , f l + 1 ( ω ′′ )) , and X L = f L ( θ v ω ) X_L=f_L(\theta_v\omega) X L = f L ( θ v ω ) . Then (i) follows from Lemma 4.7 (i) and the fact that the image of a Poisson process under a measurable map is a Poisson process with the image intensity measure [94, Theorem 5.1]; for mixed binomial processes this is immediate, because the images of independent points are independent. For (ii), let A A A be the set of ω \omega ω for which (44) holds for every l l l . By Corollary 4.10 , almost surely under Q L \mathbb{Q}_L Q L , we have θ v l ω ∈ A \theta_{v_l}\omega\in A θ v l ω ∈ A for every l l l . The configuration θ v l ω \theta_{v_l}\omega θ v l ω has root offspring number N v l N_{v_l} N v l , root auxiliary mark ζ v l \zeta_{v_l} ζ v l , and child points ( w v l k , θ v l k ω ) (w_{v_lk},\theta_{v_lk}\omega) ( w v l k , θ v l k ω ) , 1 ≤ k ≤ N v l 1 \le k \le N_{v_l} 1 ≤ k ≤ N v l . The child with k = I l + 1 k=I_{l+1} k = I l + 1 is v l + 1 v_{l+1} v l + 1 , and its point contributes δ ( w v l + 1 , X l + 1 ) \delta(w_{v_{l+1}},X_{l+1}) δ ( w v l + 1 , X l + 1 ) in (44) ; the other children contribute Ξ l f \Xi_l^f Ξ l f . This proves (ii). Part (iii) follows in the same way from Lemma 4.7 (iii). Finally, (45) is (38) applied to F ( ω , v ) = Φ ( ( f l ( θ v l ω ) ) l , ( w v l ) l ) F(\omega,v)=\Phi((f_l(\theta_{v_l}\omega))_l,(w_{v_l})_l) F ( ω , v ) = Φ (( f l ( θ v l ω ) ) l , ( w v l ) l ) , and its conditioned version is (41) .
The same argument applies to the two-spine laws. Under Q L , j \mathbb{Q}_{L,j} Q L , j , the variables along the two arms are computed from the two endpoint variables f L ( θ u ω ) f_L(\theta_u\omega) f L ( θ u ω ) and f L ( θ v ω ) f_L(\theta_v\omega) f L ( θ v ω ) , which are independent with law P ∘ f L − 1 \mathbb{P}\circ f_L^{-1} P ∘ f L − 1 , by the recursion of Corollary 4.11 (ii). At the fork, Ψ j \Psi_j Ψ j receives the side process of the fork together with the two points ( w u j + 1 , f j + 1 ( θ u j + 1 ω ) ) (w_{u_{j+1}},f_{j+1}(\theta_{u_{j+1}}\omega)) ( w u j + 1 , f j + 1 ( θ u j + 1 ω )) and ( w v j + 1 , f j + 1 ( θ v j + 1 ω ) ) (w_{v_{j+1}},f_{j+1}(\theta_{v_{j+1}}\omega)) ( w v j + 1 , f j + 1 ( θ v j + 1 ω )) , and along the common part the recursion is that of (ii). All spine edge marks, auxiliary marks, and side processes are independent, by Lemma 4.7 (iv).
Remark 4.12. As an example, fix z ∈ C ∖ R z \in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R and s > 0 s>0 s > 0 , let f l = G ∞ ( z ) f_l=G_{\infty}(z) f l = G ∞ ( z ) for every l l l , and let Ψ l ( ζ , π ) = − ( z + p − 1 ∫ ∣ w ∣ 2 x π ( d w d x ) ) − 1 \Psi_l(\zeta,\pi)=-\left(z+p^{-1}\int|w|^2x\,\pi(dw\,dx)\right)^{-1} Ψ l ( ζ , π ) = − ( z + p − 1 ∫ ∣ w ∣ 2 x π ( d w d x ) ) − 1 . By Proposition 3.16 at the root, (44) holds P \mathbb{P} P -almost surely, and X 0 = G ∞ ( z ) X_0=G_{\infty}(z) X 0 = G ∞ ( z ) and X l = Γ v l ( z ) X_l=\Gamma_{v_l}(z) X l = Γ v l ( z ) for 1 ≤ l ≤ L 1\le l\le L 1 ≤ l ≤ L . By Proposition 3.18 , for v ∈ V L v\in\mathcal{V}_L v ∈ V L we have
∣ G o v ( z ) ∣ s = ∣ G o o ( z ) ∣ s ∏ l = 1 L p − s / 2 γ v l s / 2 ∣ Γ v l ( z ) ∣ s . |G_{ov}(z)|^s=|G_{oo}(z)|^s\prod_{l=1}^{L}p^{-s/2}\gamma_{v_l}^{s/2}|\Gamma_{v_l}(z)|^s. ∣ G o v ( z ) ∣ s = ∣ G oo ( z ) ∣ s l = 1 ∏ L p − s /2 γ v l s /2 ∣ Γ v l ( z ) ∣ s .
By (45) , we have
E ∑ v ∈ V L ∣ G o v ( z ) ∣ s = p L ( 1 − s / 2 ) E Q L [ ∣ X 0 ∣ s ∏ l = 1 L γ v l s / 2 ∣ X l ∣ s ] . \mathbb{E}\sum_{v\in\mathcal{V}_L}|G_{ov}(z)|^s=p^{L(1-s/2)}\mathbb{E}_{\mathbb{Q}_L}\left[|X_0|^s\prod_{l=1}^{L}\gamma_{v_l}^{s/2}|X_l|^s\right]. E v ∈ V L ∑ ∣ G o v ( z ) ∣ s = p L ( 1 − s /2 ) E Q L [ ∣ X 0 ∣ s l = 1 ∏ L γ v l s /2 ∣ X l ∣ s ] .
On the right, X L X_L X L has the law of G ∞ ( z ) G_{\infty}(z) G ∞ ( z ) , the squared moduli γ v l \gamma_{v_l} γ v l of the spine marks are independent Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) variables, and X l = − ( z + p − 1 γ v l + 1 X l + 1 + Ξ l ) − 1 X_l=-\left(z+p^{-1}\gamma_{v_{l+1}}X_{l+1}+\Xi_l\right)^{-1} X l = − ( z + p − 1 γ v l + 1 X l + 1 + Ξ l ) − 1 for l < L l<L l < L , where Ξ l = p − 1 ∫ ∣ w ∣ 2 x Σ l f ( d w d x ) \Xi_l=p^{-1}\int|w|^2x\,\Sigma_l^f(dw\,dx) Ξ l = p − 1 ∫ ∣ w ∣ 2 x Σ l f ( d w d x ) are independent copies of the side sum p − 1 ∑ k = 1 N γ k Γ k p^{-1}\sum_{k=1}^{N}\gamma_k\Gamma_k p − 1 ∑ k = 1 N γ k Γ k in the cavity equation, independent of the squared moduli of the spine marks and of X L X_L X L .
Remark 4.13. Corollary 4.11 also applies to recursions which are not driven by the tree alone. Let E > 0 E>0 E > 0 and let π \pi π be a real solution at energy E E E in the sense of Definition 5.1 . Let q π q_\pi q π be a measurable map from [ 0 , 1 ] [0,1] [ 0 , 1 ] to R ∖ { 0 } \mathbb{R}\setminus\{0\} R ∖ { 0 } which pushes Lebesgue measure forward to π \pi π , and define f ( 0 ) ( ω ) = q π ( ζ o ( 2 ) ) f^{(0)}(\omega)=q_\pi(\zeta_o(2)) f ( 0 ) ( ω ) = q π ( ζ o ( 2 )) , where ζ o ( 2 ) \zeta_o(2) ζ o ( 2 ) is the second coordinate of ζ o \zeta_o ζ o , and, for m ≥ 1 m\ge1 m ≥ 1 ,
f ( m ) ( ω ) = E − p − 1 ∑ k = 1 N o γ k f ( m − 1 ) ( θ k ω ) f^{(m)}(\omega)=E-p^{-1}\sum_{k=1}^{N_o}\frac{\gamma_k}{f^{(m-1)}(\theta_k\omega)} f ( m ) ( ω ) = E − p − 1 k = 1 ∑ N o f ( m − 1 ) ( θ k ω ) γ k
(we set f ( m ) ( ω ) = E f^{(m)}(\omega)=E f ( m ) ( ω ) = E if some denominator vanishes, and we use the same convention for Ψ l \Psi_l Ψ l below). By induction on m m m , f ( m ) f^{(m)} f ( m ) has law π \pi π under P \mathbb{P} P for every m ≥ 0 m\ge0 m ≥ 0 . Indeed, by Lemma 4.2(ii), the pairs ( Y k , f ( m − 1 ) ( θ k ω ) ) (Y_k, f^{(m-1)}(\theta_k\omega)) ( Y k , f ( m − 1 ) ( θ k ω )) are independent with law Exp ( 1 ) ⊗ π \operatorname{Exp}(1) \otimes\pi Exp ( 1 ) ⊗ π and independent of N o N_o N o . Then f ( m ) f^{(m)} f ( m ) has the law of E − p − 1 ∑ k ≤ N Y k / D k E-p^{-1}\sum_{k\le N}Y_k/D_k E − p − 1 ∑ k ≤ N Y k / D k with independent D k ∼ π D_k\sim\pi D k ∼ π , which is π \pi π ; in particular the denominators vanish with probability zero. Fix m ≥ L m\ge L m ≥ L and apply Corollary 4.11 with f l = f ( m − l ) f_l=f^{(m-l)} f l = f ( m − l ) and Ψ l ( ζ , π ′ ) = E − p − 1 ∫ ∣ w ∣ 2 x − 1 π ′ ( d w d x ) \Psi_l(\zeta,\pi')=E-p^{-1}\int|w|^{2}x^{-1}\pi'(dw\,dx) Ψ l ( ζ , π ′ ) = E − p − 1 ∫ ∣ w ∣ 2 x − 1 π ′ ( d w d x ) . In this case, X L X_L X L has law π \pi π and the side processes are Poisson with intensity measure p γ ⊗ π p\gamma\otimes\pi p γ ⊗ π , for every m ≥ L m\ge L m ≥ L . The family D v = f ( m − ∣ v ∣ ) ( θ v ω ) D_v=f^{(m-|v|)}(\theta_v\omega) D v = f ( m − ∣ v ∣ ) ( θ v ω ) , ∣ v ∣ ≤ m |v|\le m ∣ v ∣ ≤ m , satisfies D v = E − p − 1 ∑ k Y v k / D v k D_v=E-p^{-1}\sum_k Y_{v k}/D_{v k} D v = E − p − 1 ∑ k Y v k / D v k for ∣ v ∣ < m |v|<m ∣ v ∣ < m , and its marks at generation m m m are independent with law π \pi π and independent of ( N u ) ∣ u ∣ < m (N_u)_{|u|<m} ( N u ) ∣ u ∣ < m and ( w u ) 1 ≤ ∣ u ∣ ≤ m (w_u)_{1\le|u|\le m} ( w u ) 1 ≤ ∣ u ∣ ≤ m . Every family ( D v ) ∣ v ∣ ≤ m (D_v)_{|v|\le m} ( D v ) ∣ v ∣ ≤ m with these two properties has the same joint law with ( N u ) ∣ u ∣ < m (N_u)_{|u|<m} ( N u ) ∣ u ∣ < m and ( w u ) 1 ≤ ∣ u ∣ ≤ m (w_u)_{1\le|u|\le m} ( w u ) 1 ≤ ∣ u ∣ ≤ m , because the marks at generation m m m determine the others through the recursion. This applies in particular to a recursive tree process as in Lemma 5.13, restricted to generations at most m m m , provided its marks at generation m m m are independent with law π \pi π and independent of ( N u ) ∣ u ∣ < m (N_u)_{|u|<m} ( N u ) ∣ u ∣ < m and ( w u ) 1 ≤ ∣ u ∣ ≤ m (w_u)_{1\le|u|\le m} ( w u ) 1 ≤ ∣ u ∣ ≤ m ; this is part of the definition of a recursive tree process [17]. The many-to-one and many-to-two formulas of this section, and Corollary 4.11 in particular, therefore apply to functionals of ( N u ) ∣ u ∣ < m (N_u)_{|u|<m} ( N u ) ∣ u ∣ < m , ( w u ) 1 ≤ ∣ u ∣ ≤ m (w_u)_{1\le|u|\le m} ( w u ) 1 ≤ ∣ u ∣ ≤ m , and ( D v ) ∣ v ∣ ≤ m (D_v)_{|v|\le m} ( D v ) ∣ v ∣ ≤ m for such a process. The offspring numbers at generation m m m must not enter these functionals, since a mark D v D_v D v of a recursive tree process with ∣ v ∣ = m |v|=m ∣ v ∣ = m depends on N v N_v N v .
Conditioned endpoint laws
The following proposition compares laws of endpoint functionals under P \mathbb{P} P and P ∞ \mathbb{P}^{\infty} P ∞ , and it transfers limits from P \mathbb{P} P to P ∞ \mathbb{P}^{\infty} P ∞ and to the backbone spine laws. Here an endpoint functional is any measurable function of the root configuration ω \omega ω , or of an endpoint configuration θ v ω \theta_v\omega θ v ω under a spine law; it may involve the forest obtained from the forward subtree by removing finitely many vertices and edges, or a finite decoration, since these are again measurable functions of the configuration.
Proposition 4.14. Let p > 1 p>1 p > 1 , let ( X , X ) (X,\mathcal{X}) ( X , X ) be a measurable space, and let X : Ω → X X:\Omega\to X X : Ω → X be measurable.
(i) As measures on X X X , P ∞ ∘ X − 1 ≤ ϱ p − 1 P ∘ X − 1 \mathbb{P}^{\infty}\circ X^{-1}\le\varrho_p^{-1}\mathbb{P}\circ X^{-1} P ∞ ∘ X − 1 ≤ ϱ p − 1 P ∘ X − 1 . Equivalently, E ∞ φ ( X ) ≤ ϱ p − 1 E φ ( X ) \mathbb{E}^{\infty}\varphi(X)\le\varrho_p^{-1}\mathbb{E}\varphi(X) E ∞ φ ( X ) ≤ ϱ p − 1 E φ ( X ) for every measurable φ : X → [ 0 , ∞ ] \varphi:X\to[0,\infty] φ : X → [ 0 , ∞ ] . In particular, if C ∈ X C\in\mathcal{X} C ∈ X and P ( X ∉ C ) ≤ δ \mathbb{P}(X\notin C)\le\delta P ( X ∈ / C ) ≤ δ , then P ∞ ( X ∈ C ) ≥ 1 − ϱ p − 1 δ \mathbb{P}^{\infty}(X\in C)\ge1-\varrho_p^{-1}\delta P ∞ ( X ∈ C ) ≥ 1 − ϱ p − 1 δ .
(ii) Let L > 0 L>0 L > 0 and 0 ≤ j < L 0\le j<L 0 ≤ j < L . Under Q L ∞ \mathbb{Q}^{\infty}_L Q L ∞ , the endpoint variable X ( θ v ω ) X(\theta_v\omega) X ( θ v ω ) has law P ∞ ∘ X − 1 \mathbb{P}^{\infty}\circ X^{-1} P ∞ ∘ X − 1 , and it is independent of the spine edge marks, the auxiliary marks of the spine vertices, and the side sequences. Under Q L , j ∞ \mathbb{Q}^{\infty}_{L,j} Q L , j ∞ , the pair ( X ( θ u ω ) , X ( θ v ω ) ) (X(\theta_u\omega),X(\theta_v\omega)) ( X ( θ u ω ) , X ( θ v ω )) has law ( P ∞ ∘ X − 1 ) ⊗ 2 (\mathbb{P}^{\infty}\circ X^{-1})^{\otimes2} ( P ∞ ∘ X − 1 ) ⊗ 2 , and is independent of the remaining random elements of Lemma 4.7(iv). In particular, these laws are bounded by ϱ p − 1 P ∘ X − 1 \varrho_p^{-1}\mathbb{P}\circ X^{-1} ϱ p − 1 P ∘ X − 1 and ϱ p − 2 ( P ∘ X − 1 ) ⊗ 2 \varrho_p^{-2}(\mathbb{P}\circ X^{-1})^{\otimes2} ϱ p − 2 ( P ∘ X − 1 ) ⊗ 2 , respectively.
(iii) For 0 < η ≤ 1 0<\eta\le1 0 < η ≤ 1 , let Z η Z_\eta Z η and Z 0 Z_0 Z 0 be real random variables on Ω \Omega Ω such that Z η → Z 0 Z_\eta\to Z_0 Z η → Z 0 P \mathbb{P} P -almost surely as η ↓ 0 \eta\downarrow0 η ↓ 0 , and such that the family { Z η : 0 < η ≤ 1 } \{Z_\eta:0<\eta\le1\} { Z η : 0 < η ≤ 1 } is uniformly integrable under P \mathbb{P} P .
Then Z 0 Z_0 Z 0 is P \mathbb{P} P -integrable, Z η → Z 0 Z_\eta\to Z_0 Z η → Z 0 in L 1 ( P ∞ ) L^1(\mathbb{P}^\infty) L 1 ( P ∞ ) , and in particular E ∞ Z η → E ∞ Z 0 \mathbb{E}^\infty Z_\eta\to\mathbb{E}^\infty Z_0 E ∞ Z η → E ∞ Z 0 . Further, E Q L ∞ Z η ( θ v ω ) → E Q L ∞ Z 0 ( θ v ω ) \mathbb{E}_{\mathbb{Q}^\infty_L}Z_\eta(\theta_v\omega) \to\mathbb{E}_{\mathbb{Q}^\infty_L}Z_0(\theta_v\omega) E Q L ∞ Z η ( θ v ω ) → E Q L ∞ Z 0 ( θ v ω ) for every L ≥ 0 L \ge0 L ≥ 0 .
(iv) For 0 ≤ η ≤ 1 0 \le\eta\le1 0 ≤ η ≤ 1 , let X η : Ω → R k X_\eta:\Omega\to\mathbb{R}^k X η : Ω → R k be measurable with X η → X 0 X_\eta\to X_0 X η → X 0 P \mathbb{P} P -almost surely as η ↓ 0 \eta\downarrow0 η ↓ 0 , and let φ : R k → R \varphi:\mathbb{R}^k \to\mathbb{R} φ : R k → R be a bounded Borel function which is continuous at every point outside a Borel set D D D with P ( X 0 ∈ D ) = 0 \mathbb{P}(X_0 \in D) = 0 P ( X 0 ∈ D ) = 0 . Then E ∞ φ ( X η ) → E ∞ φ ( X 0 ) \mathbb{E}^\infty\varphi(X_\eta) \to\mathbb{E}^\infty\varphi(X_0) E ∞ φ ( X η ) → E ∞ φ ( X 0 ) as η ↓ 0 \eta\downarrow0 η ↓ 0 .
Proof. For B ∈ X B \in\mathcal{X} B ∈ X , we have by (36) that P ∞ ( X ∈ B ) = ϱ p − 1 P ( X ∈ B , S ) ≤ ϱ p − 1 P ( X ∈ B ) \mathbb{P}^\infty(X \in B) = \varrho_p^{-1}\mathbb{P}(X \in B, S) \le\varrho_p^{-1}\mathbb{P}(X \in B) P ∞ ( X ∈ B ) = ϱ p − 1 P ( X ∈ B , S ) ≤ ϱ p − 1 P ( X ∈ B ) , which is (i); the last assertion of (i) follows by taking B = X ∖ C B = \mathcal{X} \setminus C B = X ∖ C . Part (ii) follows from Lemma 4.7 (iii) and (iv), together with (i).
For (iii), it suffices to consider an arbitrary sequence η n ↓ 0 \eta_n \downarrow0 η n ↓ 0 . Since P ∞ ≤ ϱ p − 1 P \mathbb{P}^\infty\le\varrho_p^{-1}\mathbb{P} P ∞ ≤ ϱ p − 1 P , the convergence Z η n → Z 0 Z_{\eta_n} \to Z_0 Z η n → Z 0 holds P ∞ \mathbb{P}^\infty P ∞ -almost surely. For K > 0 K > 0 K > 0 , we have
sup n E ∞ [ ∣ Z η n ∣ 1 { ∣ Z η n ∣ > K } ] ≤ ϱ p − 1 sup n E [ ∣ Z η n ∣ 1 { ∣ Z η n ∣ > K } ] , \sup_n \mathbb{E}^\infty\left[|Z_{\eta_n}|\mathbf{1}_{\{|Z_{\eta_n}|>K\}}\right] \le\varrho_p^{-1}\sup_n \mathbb{E}\left[|Z_{\eta_n}|\mathbf{1}_{\{|Z_{\eta_n}|>K\}}\right], n sup E ∞ [ ∣ Z η n ∣ 1 { ∣ Z η n ∣ > K } ] ≤ ϱ p − 1 n sup E [ ∣ Z η n ∣ 1 { ∣ Z η n ∣ > K } ] ,
and the right side tends to zero as K → ∞ K \to\infty K → ∞ . This shows that the sequence is uniformly integrable under P ∞ \mathbb{P}^\infty P ∞ as well. By Fatou’s lemma, Z 0 Z_0 Z 0 is integrable under P \mathbb{P} P . Then Vitali’s convergence theorem (see [64, Section 4.6]), applied under P ∞ \mathbb{P}^\infty P ∞ , implies that Z η n → Z 0 Z_{\eta_n} \to Z_0 Z η n → Z 0 in L 1 ( P ∞ ) L^1(\mathbb{P}^\infty) L 1 ( P ∞ ) . The last statement follows from (ii), since E Q L ∞ Z ( θ v ω ) = E ∞ Z \mathbb{E}_{\mathbb{Q}^\infty_L}Z(\theta_v\omega) = \mathbb{E}^\infty Z E Q L ∞ Z ( θ v ω ) = E ∞ Z for every integrable Z Z Z .
For (iv), φ ( X η ) → φ ( X 0 ) \varphi(X_\eta) \to\varphi(X_0) φ ( X η ) → φ ( X 0 ) on the event { X η → X 0 } ∩ { X 0 ∉ D } \{X_\eta\to X_0\} \cap\{X_0 \notin D\} { X η → X 0 } ∩ { X 0 ∈ / D } , which has P \mathbb{P} P -probability one and therefore P ∞ \mathbb{P}^\infty P ∞ -probability one. The bounded convergence theorem concludes the proof.
Remark 4.15. The domination in Proposition 4.14 (i) cannot be reversed with a constant independent of X X X . For instance, P ( N o = 0 ) = e − p \mathbb{P}(N_o = 0) = e^{-p} P ( N o = 0 ) = e − p , while P ∞ ( N o = 0 ) = 0 \mathbb{P}^\infty(N_o = 0) = 0 P ∞ ( N o = 0 ) = 0 , because a root without children has a finite tree. If a functional X X X equals a fixed value x 0 x_0 x 0 on { N o = 0 } \{N_o = 0\} { N o = 0 } , as the real boundary denominator does (it equals E E E there), then the part of the atom of P ∘ X − 1 \mathbb{P} \circ X^{-1} P ∘ X − 1 at x 0 x_0 x 0 coming from the event { N o = 0 } \{N_o = 0\} { N o = 0 } is absent from P ∞ ∘ X − 1 \mathbb{P}^\infty\circ X^{-1} P ∞ ∘ X − 1 . In particular, conditioned endpoint laws are in general not equal to the corresponding unconditioned laws, and only upper bounds transfer from P \mathbb{P} P to P ∞ \mathbb{P}^\infty P ∞ through (i).
Fresh squared moduli
Many estimates in this paper bound negative moments of quantities of the form c + ∑ j a j Y j c + \sum_j a_jY_j c + ∑ j a j Y j , where the Y j Y_j Y j are squared moduli of edge marks and c c c and the a j a_j a j depend on the rest of the tree. The following definition makes precise when such an estimate may be applied conditionally.
Definition 4.16. Let ( Ω ^ , A , P ) (\widehat{\Omega},\mathcal{A},\mathbb{P}) ( Ω , A , P ) be a probability space and let G ⊂ A \mathcal{G} \subset\mathcal{A} G ⊂ A be a σ \sigma σ -field. Let M M M be a G \mathcal{G} G -measurable random variable with values in { 0 , 1 , 2 , … } \{0,1,2,\ldots\} { 0 , 1 , 2 , … } , and let Y 1 , Y 2 , … Y_1,Y_2,\ldots Y 1 , Y 2 , … be random variables with values in [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . We say that Y 1 , … , Y M Y_1,\ldots,Y_M Y 1 , … , Y M are fresh squared moduli given G \mathcal{G} G , or fresh given G \mathcal{G} G , if for every m ≥ 0 m\ge0 m ≥ 0 and every bounded Borel function φ : [ 0 , ∞ ) m → R \varphi:[0,\infty)^m\to\mathbb{R} φ : [ 0 , ∞ ) m → R ,
E [ φ ( Y 1 , … , Y m ) 1 { M = m } ∣ G ] = 1 { M = m } ∫ [ 0 , ∞ ) m φ ( y ) e − ( y 1 + ⋯ + y m ) d y (46) \mathbb{E}\left[\varphi(Y_1,\ldots,Y_m)\mathbf{1}\{M=m\}\mid\mathcal{G}\right]=\mathbf{1}\{M=m\}\int_{[0,\infty)^m}\varphi(y)e^{-(y_1+\cdots+y_m)}\,dy \tag*{(46)} E [ φ ( Y 1 , … , Y m ) 1 { M = m } ∣ G ] = 1 { M = m } ∫ [ 0 , ∞ ) m φ ( y ) e − ( y 1 + ⋯ + y m ) d y (46)
almost surely. Random variables that are G \mathcal{G} G -measurable are called revealed.
In words, on the event { M = m } \{M=m\} { M = m } , the variables Y 1 , … , Y m Y_1,\ldots,Y_m Y 1 , … , Y m are independent Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) variables conditionally on G \mathcal{G} G . The number M M M of fresh variables is revealed, while their values are not; the variables Y j Y_j Y j with j > M j>M j > M play no role. In the applications, Ω ^ \widehat{\Omega} Ω is Ω \Omega Ω with one of the laws P \mathbb{P} P , P ∞ \mathbb{P}^{\infty} P ∞ , Q L \mathbb{Q}_L Q L , Q L ∞ \mathbb{Q}^{\infty}_L Q L ∞ , Q L , j \mathbb{Q}_{L,j} Q L , j , Q L , j ∞ \mathbb{Q}^{\infty}_{L,j} Q L , j ∞ , and the Y j Y_j Y j are the squared moduli of the marks of a random finite set of edges, listed in a revealed order. We record three elementary consequences of the definition.
(a) If Y 1 , … , Y M Y_1,\ldots,Y_M Y 1 , … , Y M are fresh given G \mathcal{G} G , and G ′ ⊂ G \mathcal{G}'\subset\mathcal{G} G ′ ⊂ G is a σ \sigma σ -field with respect to which M M M is measurable, then Y 1 , … , Y M Y_1,\ldots,Y_M Y 1 , … , Y M are fresh given G ′ \mathcal{G}' G ′ . This follows from (46) by the tower property of conditional expectation.
(b) Revealing some of the fresh variables leaves the others fresh. Precisely, let M ′ ≤ M M'\le M M ′ ≤ M be G \mathcal{G} G -measurable, and let G ′ \mathcal{G}' G ′ be generated by G \mathcal{G} G and the variables Y j 1 { M ′ < j ≤ M } Y_j\mathbf{1}\{M'<j\le M\} Y j 1 { M ′ < j ≤ M } , j ≥ 1 j\ge1 j ≥ 1 . Then Y 1 , … , Y M ′ Y_1,\ldots,Y_{M'} Y 1 , … , Y M ′ are fresh given G ′ \mathcal{G}' G ′ , because on { M = m , M ′ = m ′ } \{M=m,M'=m'\} { M = m , M ′ = m ′ } the variables Y 1 , … , Y m ′ Y_1,\ldots,Y_{m'} Y 1 , … , Y m ′ are independent of Y m ′ + 1 , … , Y m Y_{m'+1},\ldots,Y_m Y m ′ + 1 , … , Y m conditionally on G \mathcal{G} G . In a sum c + ∑ j ≤ M a j Y j c+\sum_{j\le M}a_jY_j c + ∑ j ≤ M a j Y j , the revealed terms a j Y j a_jY_j a j Y j , M ′ < j ≤ M M'<j\le M M ′ < j ≤ M , are then absorbed into the constant, and the required lower bound must be checked for the new constant c + ∑ M ′ < j ≤ M a j Y j c+\sum_{M'<j\le M}a_jY_j c + ∑ M ′ < j ≤ M a j Y j .
(c) Revealed variables may be treated as constants. Let Y 1 , … , Y M Y_1,\ldots,Y_M Y 1 , … , Y M be fresh given G \mathcal{G} G , let k ≥ 1 k\ge1 k ≥ 1 , and let X X X be a G \mathcal{G} G -measurable random variable with values in C k \mathbb{C}^k C k . Then for every m ≥ 0 m\ge0 m ≥ 0 and every Borel function Φ : C k × [ 0 , ∞ ) m → [ 0 , ∞ ] \Phi:\mathbb{C}^k\times[0,\infty)^m\to[0,\infty] Φ : C k × [ 0 , ∞ ) m → [ 0 , ∞ ] ,
E [ Φ ( X , Y 1 , … , Y m ) 1 { M = m } ∣ G ] = 1 { M = m } ∫ [ 0 , ∞ ) m Φ ( X , y ) e − ( y 1 + ⋯ + y m ) d y (47) \mathbb{E}\left[\Phi(X,Y_1,\ldots,Y_m)\mathbf{1}\{M=m\}\mid\mathcal{G}\right]=\mathbf{1}\{M=m\}\int_{[0,\infty)^m}\Phi(X,y)e^{-(y_1+\cdots+y_m)}\,dy \tag*{(47)} E [ Φ ( X , Y 1 , … , Y m ) 1 { M = m } ∣ G ] = 1 { M = m } ∫ [ 0 , ∞ ) m Φ ( X , y ) e − ( y 1 + ⋯ + y m ) d y (47)
almost surely. Indeed, if Φ = 1 D 1 × D 2 \Phi=\mathbf{1}_{D_1\times D_2} Φ = 1 D 1 × D 2 with Borel sets D 1 ⊂ C k D_1\subset\mathbb{C}^k D 1 ⊂ C k and D 2 ⊂ [ 0 , ∞ ) m D_2\subset[0,\infty)^m D 2 ⊂ [ 0 , ∞ ) m , then 1 D 1 ( X ) \mathbf{1}_{D_1}(X) 1 D 1 ( X ) is G \mathcal{G} G -measurable and may be taken out of the conditional expectation, and (47) reduces to (46) with φ = 1 D 2 \varphi=\mathbf{1}_{D_2} φ = 1 D 2 . The Borel sets D ⊂ C k × [ 0 , ∞ ) m D\subset\mathbb{C}^k\times[0,\infty)^m D ⊂ C k × [ 0 , ∞ ) m for which (47) holds with Φ = 1 D \Phi=\mathbf{1}_D Φ = 1 D form a Dynkin system, by linearity and the monotone convergence theorem for conditional expectations, and this system contains the π \pi π -system of measurable rectangles. By the π \pi π -λ \lambda λ theorem [81, Chapter 1], it contains all Borel sets. Approximating Φ \Phi Φ from below by simple functions and using monotone convergence once more, we obtain (47) . We refer to (47) as the substitution identity.
The main source of fresh variables is the following lemma: edge marks of edges that are selected without looking at their marks are independent standard complex Gaussian variables given everything else. It complements Lemma 3.4: there, a selection may inspect some marks, which are then removed, and the retained marks are fresh; here, a selection may inspect every mark outside the selected set. For a finite set a ⊂ U ∗ a \subset\mathbb{U}_* a ⊂ U ∗ , let
G a = σ ( N v , ζ v : v ∈ U ; w v : v ∈ U ∗ ∖ a ) , \mathcal{G}_a = \sigma\left(N_v,\zeta_v : v \in\mathbb{U};\ w_v : v \in\mathbb{U}_* \setminus a\right), G a = σ ( N v , ζ v : v ∈ U ; w v : v ∈ U ∗ ∖ a ) ,
and for x = ( x v ) v ∈ a ∈ C a x = (x_v)_{v\in a} \in\mathbb{C}^a x = ( x v ) v ∈ a ∈ C a , let ω a ← x \omega^{a\leftarrow x} ω a ← x denote the configuration obtained from ω \omega ω by replacing w v w_v w v by x v x_v x v for every v ∈ a v \in a v ∈ a .
Lemma 4.17. Let p > 0 p > 0 p > 0 , and let A A A be a random finite subset of U ∗ \mathbb{U}_* U ∗ such that { A = a } ∈ G a \{A = a\} \in\mathcal{G}_a { A = a } ∈ G a for every finite set a ⊂ U ∗ a \subset\mathbb{U}_* a ⊂ U ∗ .
(i) For every measurable F : Ω → [ 0 , ∞ ] F : \Omega\to[0,\infty] F : Ω → [ 0 , ∞ ] ,
E F ( ω ) = E ∫ C A F ( ω A ← x ) γ ⊗ A ( d x ) . (48) \mathbb{E}F(\omega) = \mathbb{E}\int_{\mathbb{C}^A} F\left(\omega^{A\leftarrow x}\right)\gamma^{\otimes A}(\mathrm{d}x). \tag*{(48)} E F ( ω ) = E ∫ C A F ( ω A ← x ) γ ⊗ A ( d x ) . (48)
If p > 1 p > 1 p > 1 , the same identity holds with E ∞ \mathbb{E}^{\infty} E ∞ in place of E \mathbb{E} E on both sides.
(ii) Let L ≥ 0 L \ge0 L ≥ 0 , 0 ≤ j < L 0 \le j < L 0 ≤ j < L , and let A : Ω × U L → { finite subsets of U ∗ } A : \Omega\times\mathbb{U}_L \to\{\text{finite subsets of } \mathbb{U}_*\} A : Ω × U L → { finite subsets of U ∗ } be such that A ( ⋅ , v ) A(\mathord\cdot,v) A ( ⋅ , v ) satisfies the hypothesis above for every v ∈ U L v \in\mathbb{U}_L v ∈ U L . Then for every measurable F : Ω × U L → [ 0 , ∞ ] F : \Omega\times\mathbb{U}_L \to[0,\infty] F : Ω × U L → [ 0 , ∞ ] , with A = A ( ω , v ) A = A(\omega,v) A = A ( ω , v ) ,
E Q L F ( ω , v ) = E Q L ∫ C A F ( ω A ← x , v ) γ ⊗ A ( d x ) , \mathbb{E}_{\mathbb{Q}_L}F(\omega,v) = \mathbb{E}_{\mathbb{Q}_L}\int_{\mathbb{C}^A} F\left(\omega^{A\leftarrow x},v\right)\gamma^{\otimes A}(\mathrm{d}x), E Q L F ( ω , v ) = E Q L ∫ C A F ( ω A ← x , v ) γ ⊗ A ( d x ) ,
and the same identity holds under Q L ∞ \mathbb{Q}^{\infty}_L Q L ∞ if p > 1 p > 1 p > 1 . The analogous identities hold under Q L , j \mathbb{Q}_{L,j} Q L , j and Q L , j ∞ \mathbb{Q}^{\infty}_{L,j} Q L , j ∞ for A = A ( ω , u , v ) A = A(\omega,u,v) A = A ( ω , u , v ) such that A ( ⋅ , u , v ) A(\mathord\cdot,u,v) A ( ⋅ , u , v ) satisfies the hypothesis above for all u , v ∈ U L u,v \in\mathbb{U}_L u , v ∈ U L .
(iii) In the setting of (i) or (ii), list the elements of A A A as v ( 1 ) , … , v ( M ) v^{(1)},\ldots,v^{(M)} v ( 1 ) , … , v ( M ) , where M = ∣ A ∣ M = |A| M = ∣ A ∣ , in a fixed total order of U \mathbb{U} U . Let H A \mathcal{H}_A H A be the σ \sigma σ -field generated by A A A , by the configuration ω A ← 1 \omega^{A\leftarrow1} ω A ← 1 in which all selected marks are replaced by 1 1 1 , by the phases w v / ∣ w v ∣ w_v/|w_v| w v /∣ w v ∣ , v ∈ A v \in A v ∈ A (set to 1 1 1 if w v = 0 w_v = 0 w v = 0 ), and, in the setting of (ii), by the endpoint v v v , respectively ( u , v ) (u,v) ( u , v ) . Then γ v ( 1 ) , … , γ v ( M ) \gamma_{v^{(1)}},\ldots,\gamma_{v^{(M)}} γ v ( 1 ) , … , γ v ( M ) are fresh squared moduli given H A \mathcal{H}_A H A .
Proof. For (i), fix a finite set a ⊂ U ∗ a \subset\mathbb{U}_* a ⊂ U ∗ . The indicator 1 { A = a } 1\{A = a\} 1 { A = a } and the configuration ω a ← x \omega^{a\leftarrow x} ω a ← x , for fixed x x x , are functions of the coordinates of ω \omega ω other than ( w v ) v ∈ a (w_v)_{v\in a} ( w v ) v ∈ a , and under P \mathbb{P} P the coordinates ( w v ) v ∈ a (w_v)_{v\in a} ( w v ) v ∈ a are independent of the other coordinates and have law γ ⊗ a \gamma^{\otimes a} γ ⊗ a . Since ω = ω a ← ( w v ) v ∈ a \omega= \omega^{a\leftarrow(w_v)_{v\in a}} ω = ω a ← ( w v ) v ∈ a , we have by Fubini’s theorem that
E [ 1 { A = a } F ( ω ) ] = E [ 1 { A = a } ∫ C a F ( ω a ← x ) γ ⊗ a ( d x ) ] . \mathbb{E}\left[1\{A = a\}F(\omega)\right] = \mathbb{E}\left[1\{A = a\}\int_{\mathbb{C}^a} F\left(\omega^{a\leftarrow x}\right)\gamma^{\otimes a}(\mathrm{d}x)\right]. E [ 1 { A = a } F ( ω ) ] = E [ 1 { A = a } ∫ C a F ( ω a ← x ) γ ⊗ a ( d x ) ] .
Summing over the countably many finite sets a a a gives (4.15). Under P ∞ \mathbb{P}^{\infty} P ∞ , we apply (4.15) to F 1 S F1_{\mathcal{S}} F 1 S and use that 1 S ( ω A ← x ) = 1 S ( ω ) 1_{\mathcal{S}}(\omega^{A\leftarrow x}) = 1_{\mathcal{S}}(\omega) 1 S ( ω A ← x ) = 1 S ( ω ) , since S \mathcal{S} S depends only on the offspring numbers.
For (ii), we use (4.5) to write E Q L F ( ω , v ) = p − L ∑ v ∈ U L E [ 1 { v ∈ V ( ω ) } F ( ω , v ) ] \mathbb{E}_{\mathbb{Q}_L}F(\omega,v) = p^{-L}\sum_{v\in\mathbb{U}_L}\mathbb{E}[1\{v \in V(\omega)\}F(\omega,v)] E Q L F ( ω , v ) = p − L ∑ v ∈ U L E [ 1 { v ∈ V ( ω )} F ( ω , v )] . For each fixed v v v , we apply (i) with the random set A ( ⋅ , v ) A(\mathord\cdot,v) A ( ⋅ , v ) to the function $\omega \mapsto 1{v \in V ( ω ) F ( ω , v ) V(\omega)\mathcal{F}(\omega,v) V ( ω ) F ( ω , v ) , and we use that V ( ω a ← x ) = V ( ω ) V(\omega^{a\leftarrow x})=V(\omega) V ( ω a ← x ) = V ( ω ) . Then we apply (38) backwards to the function ( ω , v ) ↦ ∫ F ( ω A ( ω , v ) ← x , v ) γ ⊗ A ( ω , v ) ( d x ) (\omega,v)\mapsto\int F(\omega^{A(\omega,v)\leftarrow x},v)\,\gamma^{\otimes A(\omega,v)}(\mathrm{d}x) ( ω , v ) ↦ ∫ F ( ω A ( ω , v ) ← x , v ) γ ⊗ A ( ω , v ) ( d x ) . This proves the identity under Q L \mathbb{Q}_{L} Q L . Under Q ∞ L \mathbb{Q}_{\infty}^{L} Q ∞ L , we argue in the same way with (41) and the second part of (i), using that 1 S ( θ v ω ) \mathbf{1}_{S}(\theta_{v}\omega) 1 S ( θ v ω ) depends only on the offspring numbers. The two-spine identities follow in the same way from (39) and (43) .
For (iii), we first note that changing the selected marks does not change A A A : if ω ∈ { A = a } \omega\in\{A=a\} ω ∈ { A = a } , then ω a ← x ∈ { A = a } \omega^{a\leftarrow x}\in\{A=a\} ω a ← x ∈ { A = a } for every x ∈ C a x\in\mathbb{C}^{a} x ∈ C a , since { A = a } ∈ G a \{A=a\}\in\mathcal{G}_{a} { A = a } ∈ G a . Fix m ≥ 0 m\geq0 m ≥ 0 , a bounded Borel function φ ≥ 0 \varphi\geq0 φ ≥ 0 on [ 0 , ∞ ) m [0,\infty)^{m} [ 0 , ∞ ) m , and a bounded measurable function g ≥ 0 g\geq0 g ≥ 0 of the generators of H A \mathcal{H}_{A} H A . We write x v = y v 1 / 2 e i ϕ v x_{v}=y_{v}^{1/2}e^{i\phi_{v}} x v = y v 1/2 e i ϕ v ; under γ \gamma γ , y v y_{v} y v has law Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) , e i ϕ v e^{i\phi_{v}} e i ϕ v is uniform on the unit circle, and the two are independent. Applying (i) to the function F = g ⋅ 1 { M = m } φ ( Y v ( 1 ) , … , Y v ( m ) ) F=g\cdot\mathbf{1}_{\{M=m\}}\varphi(Y_{v_{(1)}},\ldots,Y_{v_{(m)}}) F = g ⋅ 1 { M = m } φ ( Y v ( 1 ) , … , Y v ( m ) ) , we obtain
E [ g 1 { M = m } φ ( Y v ( 1 ) , … , Y v ( m ) ) ] = E [ 1 { M = m } ∫ g ( ω A ← x ) γ ⊗ A ( d x ) ] ∫ φ ( y ) e − ( y 1 + ⋯ + y m ) d y . \mathbb{E}\left[g\mathbf{1}_{\{M=m\}}\varphi(Y_{v_{(1)}},\ldots,Y_{v_{(m)}})\right]
=
\mathbb{E}\left[\mathbf{1}_{\{M=m\}}\int g(\omega^{A\leftarrow x})\,\gamma^{\otimes A}(\mathrm{d}x)\right]\int\varphi(y)e^{-(y_{1}+\cdots+y_{m})}\,\mathrm{d}y. E [ g 1 { M = m } φ ( Y v ( 1 ) , … , Y v ( m ) ) ] = E [ 1 { M = m } ∫ g ( ω A ← x ) γ ⊗ A ( d x ) ] ∫ φ ( y ) e − ( y 1 + ⋯ + y m ) d y .
Here we used that g ( ω A ← x ) g(\omega^{A\leftarrow x}) g ( ω A ← x ) depends on x x x only through the phases of the x v x_{v} x v , while φ \varphi φ depends only on their moduli. Applying (i) once more, to the function g 1 { M = m } g\mathbf{1}_{\{M=m\}} g 1 { M = m } , shows that the first factor on the right equals E [ g 1 { M = m } ] \mathbb{E}[g\mathbf{1}_{\{M=m\}}] E [ g 1 { M = m } ] . This is (46) for G = H A \mathcal{G}=\mathcal{H}_{A} G = H A . The same argument applies in the setting of (ii).
For example, take A = { 1 , … , N o } A=\{1,\ldots,N_{o}\} A = { 1 , … , N o } , the set of children of the root. We have { A = a } ∈ σ ( N o ) ⊂ G a \{A=a\}\in\sigma(N_{o})\subset\mathcal{G}_{a} { A = a } ∈ σ ( N o ) ⊂ G a , and H A \mathcal{H}_{A} H A contains N o N_{o} N o , ζ o \zeta_{o} ζ o , all forward configurations θ k ω \theta_{k}\omega θ k ω , k ≥ 1 k\geq1 k ≥ 1 , and the phases of the w k w_{k} w k . By Lemma 4.17(iii) and consequence (a) after Definition 4.16, the squared moduli Y 1 , … , Y N o Y_{1},\ldots,Y_{N_{o}} Y 1 , … , Y N o of the root edge marks are fresh given the σ \sigma σ -field generated by N o N_{o} N o , ζ o \zeta_{o} ζ o , the forward configurations of the children, and the phases. This is the situation of the cavity equation. Similarly, under Q L \mathbb{Q}_{L} Q L , the choice A ( ω , v ) = { v 1 , … , v L } A(\omega,v)=\{v_{1},\ldots,v_{L}\} A ( ω , v ) = { v 1 , … , v L } in Lemma 4.17(ii) shows that the squared moduli Y v 1 , … , Y v L Y_{v_{1}},\ldots,Y_{v_{L}} Y v 1 , … , Y v L of the spine marks are fresh given the σ \sigma σ -field generated by the endpoint v v v , the phases of the spine edge marks, and all coordinates of ω \omega ω other than the spine edge marks.
We now state the affine negative-moment lemma.
Lemma 4.18. Let 0 < r < 1 0<r<1 0 < r < 1 and d > 0 d>0 d > 0 , and set C r , d = ( 1 − r ) − 1 ( 2 / d ) r C_{r,d}=(1-r)^{-1}(2/d)^{r} C r , d = ( 1 − r ) − 1 ( 2/ d ) r . Let m ≥ 0 m\geq0 m ≥ 0 , let Y 1 , … , Y m Y_{1},\ldots,Y_{m} Y 1 , … , Y m be independent random variables with law Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) , let a 1 , … , a m ∈ C a_{1},\ldots,a_{m}\in\mathbb{C} a 1 , … , a m ∈ C and c ∈ C c\in\mathbb{C} c ∈ C , and set S = c + ∑ j = 1 m a j Y j S=c+\sum_{j=1}^{m}a_{j}Y_{j} S = c + ∑ j = 1 m a j Y j .
(i) If c ≠ 0 c\neq0 c = 0 , then for every ε > 0 \varepsilon>0 ε > 0 ,
P ( ∣ S ∣ ≤ ε ) ≤ min { 1 , 2 ε m + 1 ∣ c ∣ } . (49) \mathbb{P}(|S|\leq\varepsilon)\leq\min\left\{1,\frac{2\varepsilon\sqrt{m+1}}{|c|}\right\}. \tag*{(49)} P ( ∣ S ∣ ≤ ε ) ≤ min { 1 , ∣ c ∣ 2 ε m + 1 } . (49) (ii) If c ≠ 0 c\neq0 c = 0 , then
E ∣ S ∣ − r ≤ 1 1 − r ( 2 m + 1 ∣ c ∣ ) r . (50) \mathbb{E}|S|^{-r}\leq\frac{1}{1-r}\left(\frac{2\sqrt{m+1}}{|c|}\right)^{r}. \tag*{(50)} E ∣ S ∣ − r ≤ 1 − r 1 ( ∣ c ∣ 2 m + 1 ) r . (50)
In particular, if ∣ c ∣ ≥ d |c| \ge d ∣ c ∣ ≥ d , then E ∣ S ∣ − r ≤ C r , d ( m + 1 ) r / 2 \mathbb{E}|S|^{-r} \le C_{r,d}(m+1)^{r/2} E ∣ S ∣ − r ≤ C r , d ( m + 1 ) r /2 .
(iii) Suppose that c , a 1 , … , a m c,a_1,\ldots,a_m c , a 1 , … , a m are real. If a 1 = ⋯ = a m = 0 a_1=\cdots=a_m=0 a 1 = ⋯ = a m = 0 , then the law of S S S is δ c \delta_c δ c . Otherwise, the law of S S S is absolutely continuous, and it has a density ρ \rho ρ such that
ρ ( x ) ≤ m 2 π ∣ x − c ∣ for every x ∈ R ∖ { c } . (51) \rho(x) \le\frac{\sqrt{m}}{\sqrt{2\pi}|x-c|} \qquad\text{for every } x\in\mathbb{R}\setminus\{c\}. \tag*{(51)} ρ ( x ) ≤ 2 π ∣ x − c ∣ m for every x ∈ R ∖ { c } . (51)
In either case, the absolutely continuous part of the law of S S S has a density which is at most m + 1 / ( 2 d ) \sqrt{m+1}/(2d) m + 1 / ( 2 d ) at every x x x with ∣ x − c ∣ ≥ d |x-c|\ge d ∣ x − c ∣ ≥ d .
(iv) Let G \mathcal{G} G , M M M , and Y 1 , Y 2 , … Y_1,Y_2,\ldots Y 1 , Y 2 , … be as in Definition 4.16, with Y 1 , … , Y M Y_1,\ldots,Y_M Y 1 , … , Y M fresh given G \mathcal{G} G . Let c , a 1 , a 2 , … c,a_1,a_2,\ldots c , a 1 , a 2 , … be G \mathcal{G} G -measurable complex random variables, and set S = c + ∑ j = 1 M a j Y j S=c+\sum_{j=1}^{M}a_jY_j S = c + ∑ j = 1 M a j Y j . Then, for every ε > 0 \varepsilon>0 ε > 0 , almost surely on the event { c ≠ 0 } \{c\ne0\} { c = 0 } ,
P ( ∣ S ∣ ≤ ε ∣ G ) ≤ min { 1 , 2 ε M + 1 ∣ c ∣ } , E [ ∣ S ∣ − r ∣ G ] ≤ 1 1 − r ( 2 M + 1 ∣ c ∣ ) r . \mathbb{P}(|S|\le\varepsilon\mid\mathcal{G})\le\min\left\{1,\frac{2\varepsilon\sqrt{M+1}}{|c|}\right\},\qquad
\mathbb{E}\left[|S|^{-r}\mid\mathcal{G}\right]\le\frac{1}{1-r}\left(\frac{2\sqrt{M+1}}{|c|}\right)^r. P ( ∣ S ∣ ≤ ε ∣ G ) ≤ min { 1 , ∣ c ∣ 2 ε M + 1 } , E [ ∣ S ∣ − r ∣ G ] ≤ 1 − r 1 ( ∣ c ∣ 2 M + 1 ) r .
If ∣ c ∣ ≥ d |c|\ge d ∣ c ∣ ≥ d almost surely, then E ∣ S ∣ − r ≤ C r , d E ( M + 1 ) r / 2 ≤ C r , d ( 1 + E M ) r / 2 \mathbb{E}|S|^{-r}\le C_{r,d}\mathbb{E}(M+1)^{r/2}\le C_{r,d}(1+\mathbb{E}M)^{r/2} E ∣ S ∣ − r ≤ C r , d E ( M + 1 ) r /2 ≤ C r , d ( 1 + E M ) r /2 . If c c c and all a j a_j a j are real, then for every Borel set B ⊂ R B\subset\mathbb{R} B ⊂ R , almost surely,
P ( S ∈ B , ∣ S − c ∣ ≥ d ∣ G ) ≤ M + 1 2 d ∣ B ∣ , \mathbb{P}(S\in B,\ |S-c|\ge d\mid\mathcal{G})\le\frac{\sqrt{M+1}}{2d}|B|, P ( S ∈ B , ∣ S − c ∣ ≥ d ∣ G ) ≤ 2 d M + 1 ∣ B ∣ ,
where ∣ B ∣ |B| ∣ B ∣ denotes the Lebesgue measure of B B B .
The bounds in (i)--(iii) depend on the coefficients a j a_j a j only through their number. This is what makes the lemma applicable when the coefficients are cavity variables of subtrees, about which nothing is known except that they are revealed.
Proof. We first prove the density bound (iii) for real coefficients using the Gamma--Dirichlet decomposition of ( Y 1 , … , Y m ) (Y_1,\ldots,Y_m) ( Y 1 , … , Y m ) . Then we deduce the small-ball bound (i) for complex coefficients by rotating c c c to the positive real axis and projecting onto the real axis, and we integrate it to obtain (ii). Finally, we derive the conditional version (iv) by treating the revealed variables as constants.
Step 1. The Gamma--Dirichlet decomposition. Let m ≥ 1 m\ge1 m ≥ 1 , set T = Y 1 + ⋯ + Y m T=Y_1+\cdots+Y_m T = Y 1 + ⋯ + Y m , and set U j = Y j / T U_j=Y_j/T U j = Y j / T for 1 ≤ j ≤ m 1\le j\le m 1 ≤ j ≤ m . Then T T T has the Gamma density
f m ( t ) = t m − 1 e − t ( m − 1 ) ! , t > 0 , f_m(t)=\frac{t^{m-1}e^{-t}}{(m-1)!},\qquad t>0, f m ( t ) = ( m − 1 )! t m − 1 e − t , t > 0 ,
and the vector ( U 1 , … , U m ) (U_1,\ldots,U_m) ( U 1 , … , U m ) is independent of T T T (it is uniformly distributed on the simplex). Indeed, the joint density of ( Y 1 , … , Y m ) (Y_1,\ldots,Y_m) ( Y 1 , … , Y m ) is e − ( y 1 + ⋯ + y m ) = e − t e^{-(y_1+\cdots+y_m)}=e^{-t} e − ( y 1 + ⋯ + y m ) = e − t , and the change of variables ( y 1 , … , y m ) ↦ ( t , u 1 , … , u m − 1 ) (y_1,\ldots,y_m)\mapsto(t,u_1,\ldots,u_{m-1}) ( y 1 , … , y m ) ↦ ( t , u 1 , … , u m − 1 ) has Jacobian t m − 1 t^{m-1} t m − 1 ; it follows that the joint density of ( T , U 1 , … , U m − 1 ) (T,U_1,\ldots,U_{m-1}) ( T , U 1 , … , U m − 1 ) is a product of a function of t t t and a function of ( u 1 , … , u m − 1 ) (u_1,\ldots,u_{m-1}) ( u 1 , … , u m − 1 ) . We also need the bound
σ m = sup t > 0 t f m ( t ) = m m e − m ( m − 1 ) ! = m m + 1 e − m m ! ≤ m 2 π . (52) \sigma_m=\sup_{t>0}t f_m(t)=\frac{m^m e^{-m}}{(m-1)!}=\frac{m^{m+1}e^{-m}}{m!}\le\frac{\sqrt{m}}{\sqrt{2\pi}}. \tag*{(52)} σ m = t > 0 sup t f m ( t ) = ( m − 1 )! m m e − m = m ! m m + 1 e − m ≤ 2 π m . (52)
The supremum is attained at t = m t=m t = m , since t ↦ t m e − t t\mapsto t^m e^{-t} t ↦ t m e − t is maximal there, and in the last inequality we used Stirling’s formula in the form m ! ≥ 2 π m ( m / e ) m m!\ge\sqrt{2\pi m}(m/e)^m m ! ≥ 2 π m ( m / e ) m , valid for every m ≥ 1 m\ge1 m ≥ 1 .
Step 2. Real coefficients. Let c , a 1 , … , a m c,a_1,\ldots,a_m c , a 1 , … , a m be real. If all a j a_j a j vanish (including the case m = 0 m=0 m = 0 ), then S = c S=c S = c . Otherwise m ≥ 1 m\ge1 m ≥ 1 , and we may assume that a 1 ≠ 0 a_1\ne0 a 1 = 0 . Then S = a 1 Y 1 + ( c + ∑ j ≥ 2 a j Y j ) S=a_1Y_1+(c+\sum_{j\ge2}a_jY_j) S = a 1 Y 1 + ( c + ∑ j ≥ 2 a j Y j ) is the sum of an absolutely continuous variable and an independent variable, and its law is therefore absolutely continuous. With the notation of Step 1, set A = ∑ j = 1 m a j U j A=\sum_{j=1}^{m}a_jU_j A = ∑ j = 1 m a j U j , so that S = c + A T S=c+AT S = c + A T , where A A A is independent of T T T . Since { A = 0 } ⊂ { S = c } \{A=0\}\subset\{S=c\} { A = 0 } ⊂ { S = c } and P ( S = c ) = 0 \mathbb{P}(S=c)=0 P ( S = c ) = 0 , we have P ( A = 0 ) = 0 \mathbb{P}(A=0)=0 P ( A = 0 ) = 0 . For a Borel set B ⊂ R B\subset\mathbb{R} B ⊂ R , Tonelli’s theorem and the change of variables x = c + A t x=c+At x = c + A t give
P ( S ∈ B ) = E [ ∫ 0 ∞ 1 B ( c + A t ) f m ( t ) d t ] = ∫ B ρ ( x ) d x , \mathbb{P}(S\in B)=\mathbb{E}\left[\int_0^\infty\mathbf{1}_B(c+At)f_m(t)\,dt\right]=\int_B\rho(x)\,dx, P ( S ∈ B ) = E [ ∫ 0 ∞ 1 B ( c + A t ) f m ( t ) d t ] = ∫ B ρ ( x ) d x ,
where
ρ ( x ) = E [ 1 { x − c A > 0 } 1 ∣ A ∣ f m ( x − c A ) ] , \rho(x)=\mathbb{E}\left[\mathbf{1}_{\left\{\frac{x-c}{A}>0\right\}}\frac{1}{|A|}f_m\left(\frac{x-c}{A}\right)\right], ρ ( x ) = E [ 1 { A x − c > 0 } ∣ A ∣ 1 f m ( A x − c ) ] ,
and where the integrand in the definition of ρ \rho ρ is set to zero on the null event { A = 0 } \{A=0\} { A = 0 } . The function ρ \rho ρ is Borel by Tonelli’s theorem. Let x ≠ c x\ne c x = c , and suppose that t = ( x − c ) / A > 0 t=(x-c)/A>0 t = ( x − c ) / A > 0 . Then ∣ A ∣ = ∣ x − c ∣ / t |A|=|x-c|/t ∣ A ∣ = ∣ x − c ∣/ t , so that
1 ∣ A ∣ f m ( x − c A ) = t f m ( t ) ∣ x − c ∣ ≤ σ m ∣ x − c ∣ . \frac{1}{|A|}f_m\left(\frac{x-c}{A}\right)=\frac{t f_m(t)}{|x-c|}\le\frac{\sigma_m}{|x-c|}. ∣ A ∣ 1 f m ( A x − c ) = ∣ x − c ∣ t f m ( t ) ≤ ∣ x − c ∣ σ m .
Taking expectations and using (52) , we arrive at (51) . The last assertion of (iii) follows because m / 2 π ≤ m + 1 / 2 \sqrt{m}/\sqrt{2\pi}\le\sqrt{m+1}/2 m / 2 π ≤ m + 1 /2 , and because the absolutely continuous part of δ c \delta_c δ c is zero.
Step 3. The small-ball bound. Let c ≠ 0 c\ne0 c = 0 , let θ \theta θ be an argument of c c c , and set
Q = Re ( e − i θ S ) = ∣ c ∣ + ∑ j = 1 m b j Y j , b j = Re ( e − i θ a j ) ∈ R . Q=\operatorname{Re}(e^{-i\theta}S)=|c|+\sum_{j=1}^{m}b_jY_j,\qquad b_j=\operatorname{Re}(e^{-i\theta}a_j)\in\mathbb{R}. Q = Re ( e − i θ S ) = ∣ c ∣ + j = 1 ∑ m b j Y j , b j = Re ( e − i θ a j ) ∈ R .
Then ∣ S ∣ = ∣ e − i θ S ∣ ≥ ∣ Q ∣ |S|=|e^{-i\theta}S|\ge|Q| ∣ S ∣ = ∣ e − i θ S ∣ ≥ ∣ Q ∣ , which implies that P ( ∣ S ∣ ≤ ε ) ≤ P ( ∣ Q ∣ ≤ ε ) \mathbb{P}(|S|\le\varepsilon)\le\mathbb{P}(|Q|\le\varepsilon) P ( ∣ S ∣ ≤ ε ) ≤ P ( ∣ Q ∣ ≤ ε ) . If ε > ∣ c ∣ / 2 \varepsilon>|c|/2 ε > ∣ c ∣/2 , then the right side of (49) equals 1 1 1 and there is nothing to prove. Let 0 < ε ≤ ∣ c ∣ / 2 0<\varepsilon\le|c|/2 0 < ε ≤ ∣ c ∣/2 . If all b j b_j b j vanish, then Q = ∣ c ∣ > ε Q=|c|>\varepsilon Q = ∣ c ∣ > ε , and P ( ∣ Q ∣ ≤ ε ) = 0 \mathbb{P}(|Q|\le\varepsilon)=0 P ( ∣ Q ∣ ≤ ε ) = 0 . Otherwise, by Step 2 applied to Q Q Q , whose constant is ∣ c ∣ |c| ∣ c ∣ , the law of Q Q Q has a density bounded by σ m / ∣ x − ∣ c ∣ ∣ \sigma_m/|x-|c|| σ m /∣ x − ∣ c ∣∣ . Every x ∈ [ − ε , ε ] x\in[-\varepsilon,\varepsilon] x ∈ [ − ε , ε ] satisfies ∣ x − ∣ c ∣ ∣ ≥ ∣ c ∣ / 2 |x-|c||\ge|c|/2 ∣ x − ∣ c ∣∣ ≥ ∣ c ∣/2 , so
P ( ∣ Q ∣ ≤ ε ) ≤ 2 ε ⋅ 2 σ m ∣ c ∣ ≤ 4 2 π ⋅ ε m ∣ c ∣ ≤ 2 ε m + 1 ∣ c ∣ . \mathbb{P}(|Q|\le\varepsilon)\le2\varepsilon\cdot\frac{2\sigma_m}{|c|}\le\frac{4}{\sqrt{2\pi}}\cdot\frac{\varepsilon\sqrt{m}}{|c|}\le\frac{2\varepsilon\sqrt{m+1}}{|c|}. P ( ∣ Q ∣ ≤ ε ) ≤ 2 ε ⋅ ∣ c ∣ 2 σ m ≤ 2 π 4 ⋅ ∣ c ∣ ε m ≤ ∣ c ∣ 2 ε m + 1 .
This proves (i).
Step 4. The negative moment. By (i), P ( S = 0 ) = 0 \mathbb{P}(S=0)=0 P ( S = 0 ) = 0 . For t > 0 t>0 t > 0 , we have t − r = ∫ t ∞ r ε − r − 1 d ε t^{-r}=\int_t^\infty r\varepsilon^{-r-1}\,d\varepsilon t − r = ∫ t ∞ r ε − r − 1 d ε , and Tonelli’s theorem implies that
E ∣ S ∣ − r = ∫ 0 ∞ r ε − r − 1 P ( ∣ S ∣ ≤ ε ) d ε . \mathbb{E}|S|^{-r}=\int_0^\infty r\varepsilon^{-r-1}\mathbb{P}(|S|\leq\varepsilon)\,d\varepsilon. E ∣ S ∣ − r = ∫ 0 ∞ r ε − r − 1 P ( ∣ S ∣ ≤ ε ) d ε .
Set K = 2 m + 1 / ∣ c ∣ K=2\sqrt{m+1}/|c| K = 2 m + 1 /∣ c ∣ . By (i), the right side is at most
∫ 0 1 / K r K ε − r d ε + ∫ 1 / K ∞ r ε − r − 1 d ε = r 1 − r K r + K r = K r 1 − r , (53) \int_0^{1/K}rK\varepsilon^{-r}\,d\varepsilon+\int_{1/K}^{\infty}r\varepsilon^{-r-1}\,d\varepsilon=\frac{r}{1-r}K^r+K^r=\frac{K^r}{1-r}, \tag*{(53)} ∫ 0 1/ K r K ε − r d ε + ∫ 1/ K ∞ r ε − r − 1 d ε = 1 − r r K r + K r = 1 − r K r , (53)
which is (4.17). If ∣ c ∣ ≥ d |c|\geq d ∣ c ∣ ≥ d , then K r ≤ ( 2 / d ) r ( m + 1 ) r / 2 K^r\leq(2/d)^r(m+1)^{r/2} K r ≤ ( 2/ d ) r ( m + 1 ) r /2 , which gives the second assertion of (ii).
Step 5. The conditional version. Fix m ≥ 0 m\geq0 m ≥ 0 . By consequence (c) after Definition 4.16, applied with X = ( c , a 1 , … , a m ) X=(c,a_1,\ldots,a_m) X = ( c , a 1 , … , a m ) , the substitution identity (4.14) holds for every Borel function Φ : C 1 + m × [ 0 , ∞ ) m → [ 0 , ∞ ] \Phi:\mathbb{C}^{1+m}\times[0,\infty)^m\to[0,\infty] Φ : C 1 + m × [ 0 , ∞ ) m → [ 0 , ∞ ] .
We now apply (4.14) with Φ ( c , a , y ) = 1 { ∣ c + ∑ j a j y j ∣ ≤ ε } 1 { c ≠ 0 } \Phi(c,a,y)=1\{|c+\sum_j a_jy_j|\leq\varepsilon\}1\{c\neq0\} Φ ( c , a , y ) = 1 { ∣ c + ∑ j a j y j ∣ ≤ ε } 1 { c = 0 } . By (i), applied for fixed values of c ≠ 0 c\neq0 c = 0 and a 1 , … , a m a_1,\ldots,a_m a 1 , … , a m , the right side of (4.14) is at most 1 { M = m } 1 { c ≠ 0 } min { 1 , 2 ε m + 1 / ∣ c ∣ } 1\{M=m\}1\{c\neq0\}\min\{1,2\varepsilon\sqrt{m+1}/|c|\} 1 { M = m } 1 { c = 0 } min { 1 , 2 ε m + 1 /∣ c ∣ } . Summing over m m m gives the small-ball bound in (iv). The bound on E [ ∣ S ∣ − r ∣ G ] \mathbb{E}[|S|^{-r}\mid\mathcal{G}] E [ ∣ S ∣ − r ∣ G ] follows in the same way from (ii), with Φ ( c , a , y ) = ∣ c + ∑ j a j y j ∣ − r 1 { c ≠ 0 } \Phi(c,a,y)=|c+\sum_j a_jy_j|^{-r}1\{c\neq0\} Φ ( c , a , y ) = ∣ c + ∑ j a j y j ∣ − r 1 { c = 0 } . If ∣ c ∣ ≥ d |c|\geq d ∣ c ∣ ≥ d almost surely, then taking expectations, we have E ∣ S ∣ − r ≤ C r , d E ( M + 1 ) r / 2 \mathbb{E}|S|^{-r}\leq C_{r,d}\mathbb{E}(M+1)^{r/2} E ∣ S ∣ − r ≤ C r , d E ( M + 1 ) r /2 , and Jensen’s inequality for the concave function t ↦ t r / 2 t\mapsto t^{r/2} t ↦ t r /2 gives E ( M + 1 ) r / 2 ≤ ( 1 + E M ) r / 2 \mathbb{E}(M+1)^{r/2}\leq(1+\mathbb{E}M)^{r/2} E ( M + 1 ) r /2 ≤ ( 1 + E M ) r /2 . Finally, if c c c and the a j a_j a j are real, we apply (4.14) with Φ ( c , a , y ) = 1 { c + ∑ j a j y j ∈ B , ∣ ∑ j a j y j ∣ ≥ d } \Phi(c,a,y)=1\{c+\sum_j a_jy_j\in B,\ |\sum_j a_jy_j|\geq d\} Φ ( c , a , y ) = 1 { c + ∑ j a j y j ∈ B , ∣ ∑ j a j y j ∣ ≥ d } . For fixed real c , a 1 , … , a m c,a_1,\ldots,a_m c , a 1 , … , a m , the integral of Φ \Phi Φ vanishes if all a j a_j a j vanish, and otherwise it is at most ∫ B ∩ { ∣ x − c ∣ ≥ d } ρ ( x ) d x ≤ m + 1 ∣ B ∣ / ( 2 d ) \int_{B\cap\{|x-c|\geq d\}}\rho(x)\,dx\leq\sqrt{m+1}|B|/(2d) ∫ B ∩ { ∣ x − c ∣ ≥ d } ρ ( x ) d x ≤ m + 1 ∣ B ∣/ ( 2 d ) by (iii). Summing over m m m completes the proof. □
Remark 4.19. Freshness is a property of the variables together with the σ \sigma σ -field, and it is lost as soon as the σ \sigma σ -field carries information about their values. We give an example. Let Y 1 Y_1 Y 1 have law Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) , let 0 < δ ≤ 1 / 2 0<\delta\leq1/2 0 < δ ≤ 1/2 , and let G = σ ( A δ ) \mathcal{G}=\sigma(A_\delta) G = σ ( A δ ) , where A δ = { ∣ Y 1 − 1 ∣ ≤ δ } A_\delta=\{|Y_1-1|\leq\delta\} A δ = { ∣ Y 1 − 1∣ ≤ δ } is a truncation event for Y 1 Y_1 Y 1 . The variable Y 1 Y_1 Y 1 is not fresh given G \mathcal{G} G : on A δ A_\delta A δ , its conditional law is supported in [ 1 − δ , 1 + δ ] [1-\delta,1+\delta] [ 1 − δ , 1 + δ ] . The conclusion of Lemma 4.18(iv) fails as well. Let M = 1 M=1 M = 1 , c = 1 c=1 c = 1 , and a 1 = − 1 a_1=-1 a 1 = − 1 , which are constants and hence G \mathcal{G} G -measurable, so that S = 1 − Y 1 S=1-Y_1 S = 1 − Y 1 . On A δ A_\delta A δ , we have
E [ ∣ S ∣ − r ∣ G ] = ∫ 1 − δ 1 + δ ∣ 1 − y ∣ − r e − y d y ∫ 1 − δ 1 + δ e − y d y ≥ e − 1 − δ ⋅ 2 δ 1 − r ( 1 − r ) − 1 2 δ e − 1 + δ = e − 2 δ δ − r 1 − r . \mathbb{E}[|S|^{-r}\mid\mathcal{G}]=\frac{\int_{1-\delta}^{1+\delta}|1-y|^{-r}e^{-y}\,dy}{\int_{1-\delta}^{1+\delta}e^{-y}\,dy}\geq\frac{e^{-1-\delta}\cdot2\delta^{1-r}(1-r)^{-1}}{2\delta e^{-1+\delta}}=\frac{e^{-2\delta}\delta^{-r}}{1-r}. E [ ∣ S ∣ − r ∣ G ] = ∫ 1 − δ 1 + δ e − y d y ∫ 1 − δ 1 + δ ∣1 − y ∣ − r e − y d y ≥ 2 δ e − 1 + δ e − 1 − δ ⋅ 2 δ 1 − r ( 1 − r ) − 1 = 1 − r e − 2 δ δ − r .
which tends to infinity as δ → 0 \delta\to0 δ → 0 , while the bound ( 1 − r ) − 1 2 3 r / 2 (1-r)^{-1}2^{3r/2} ( 1 − r ) − 1 2 3 r /2 of Lemma 4.18(iv) does not depend on δ \delta δ . The same failure occurs if G \mathcal{G} G contains σ ( Y 1 ) \sigma(Y_1) σ ( Y 1 ) , that is, if Y 1 Y_1 Y 1 is conditioned on its own value. In this case Y 1 Y_1 Y 1 is revealed, the term a 1 Y 1 a_1Y_1 a 1 Y 1 must be absorbed into the constant as in (b) after Definition 4.16, and the new constant 1 − Y 1 1-Y_1 1 − Y 1 is not bounded away from zero. In particular, whenever an event defined in terms of a squared modulus is included in the conditioning σ \sigma σ -field, that squared modulus can no longer be treated as fresh.
As a first application, we bound fractional moments of the diagonal resolvent at the root, uniformly in the imaginary part of the spectral parameter.
Corollary 4.20. Let p > 0 p>0 p > 0 , 0 < r < 1 0<r<1 0 < r < 1 , and z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R . Then
E ∣ G o o ( z ) ∣ r ≤ 1 1 − r ( 2 ∣ z ∣ ) r ( p + 1 ) r / 2 . \mathbb{E}|G_{oo}(z)|^r \le\frac{1}{1-r}\left(\frac{2}{|z|}\right)^r(p+1)^{r/2}. E ∣ G oo ( z ) ∣ r ≤ 1 − r 1 ( ∣ z ∣ 2 ) r ( p + 1 ) r /2 . If p > 1 p>1 p > 1 , then E ∞ ∣ G o o ( z ) ∣ r \mathbb{E}^{\infty}|G_{oo}(z)|^r E ∞ ∣ G oo ( z ) ∣ r is at most ϱ p − 1 \varrho_p^{-1} ϱ p − 1 times the same bound.
Proof. By Proposition 3.16 at the root and the identity Γ k ( z ) = G o o ( z ) ∘ θ k \Gamma_k(z)=G_{oo}(z)\circ\theta_k Γ k ( z ) = G oo ( z ) ∘ θ k , we have, P \mathbb{P} P -almost surely,
G o o ( z ) = − ( z + 1 p ∑ k = 1 N o γ k G o o ( z ) ∘ θ k ) − 1 . G_{oo}(z)=-\left(z+\frac{1}{p}\sum_{k=1}^{N_o}\gamma_kG_{oo}(z)\circ\theta_k\right)^{-1}. G oo ( z ) = − ( z + p 1 k = 1 ∑ N o γ k G oo ( z ) ∘ θ k ) − 1 . Let G \mathcal{G} G be the σ \sigma σ -field generated by N o N_o N o , ζ o \zeta_o ζ o , the configurations θ k ω \theta_k\omega θ k ω for k ≥ 1 k\geq1 k ≥ 1 , and the phases w k / ∣ w k ∣ w_k/|w_k| w k /∣ w k ∣ for k ≥ 1 k\geq1 k ≥ 1 . By the example following the proof of Lemma 4.17, together with consequence (a) after Definition 4.16, the squared moduli γ 1 , … , γ N o \gamma_1,\ldots,\gamma_{N_o} γ 1 , … , γ N o are fresh given G \mathcal{G} G . The coefficients a k = p − 1 G o o ( z ) ∘ θ k a_k=p^{-1}G_{oo}(z)\circ\theta_k a k = p − 1 G oo ( z ) ∘ θ k are G \mathcal{G} G -measurable, and the constant c = z c=z c = z is deterministic and nonzero. Then Lemma 4.18(iv) implies that E [ ∣ G o o ( z ) ∣ r ∣ G ] ≤ ( 1 − r ) − 1 ( 2 N o + 1 / ∣ z ∣ ) r \mathbb{E}[|G_{oo}(z)|^r\mid\mathcal{G}]\leq(1-r)^{-1}(2\sqrt{N_o}+1/|z|)^r E [ ∣ G oo ( z ) ∣ r ∣ G ] ≤ ( 1 − r ) − 1 ( 2 N o + 1/∣ z ∣ ) r . Taking expectations and using Jensen’s inequality, E ( N o + 1 ) r / 2 ≤ ( p + 1 ) r / 2 \mathbb{E}(N_o+1)^{r/2}\leq(p+1)^{r/2} E ( N o + 1 ) r /2 ≤ ( p + 1 ) r /2 , we find the first claim. The second follows from (4.3).
Since ∣ z ∣ ≥ ∣ Re z ∣ |z|\geq|\operatorname{Re}z| ∣ z ∣ ≥ ∣ Re z ∣ , the bound is uniform over all z z z with real part bounded away from zero, which is the uniformity required at energies E ≠ 0 E\ne0 E = 0 .
Selections at a vertex and the backbone
We now classify the children of a vertex by the shape of their forward subtrees. For a vertex x ∈ V ( ω ) x\in V(\omega) x ∈ V ( ω ) , a child x k x_k x k of x x x is a leaf child if N x k = 0 N_{x_k}=0 N x k = 0 , so that its forward subtree consists of x k x_k x k alone; it is a surviving child if θ x k ω ∈ S \theta_{x_k}\omega\in S θ x k ω ∈ S ; and it is an other child if neither holds. We write N x l e a f N_x^{\mathrm{leaf}} N x leaf , N x s u r v N_x^{\mathrm{surv}} N x surv , and N x o t h N_x^{\mathrm{oth}} N x oth for the numbers of these children. For a vertex x x x with N x l e a f ≥ 1 N_x^{\mathrm{leaf}}\geq1 N x leaf ≥ 1 , we let ℓ ( x ) \ell(x) ℓ ( x ) denote the leaf child x k x_k x k of x x x for which the pair ( ζ x k ( 1 ) , k ) (\zeta_{x_k}(1),k) ( ζ x k ( 1 ) , k ) is lexicographically smallest; thus ℓ ( x ) \ell(x) ℓ ( x ) is the leaf child with the smallest rank, ties (which have probability zero) being broken by the label. Recall M = C × Ω M=\mathcal{C}\times\Omega M = C × Ω and m = γ ⊗ P m=\gamma\otimes\mathbb{P} m = γ ⊗ P , and set
m l e a f = γ ⊗ P ( ⋅ ∣ N o = 0 ) , m ∞ = γ ⊗ P ∞ , m o t h = γ ⊗ P ( ⋅ ∣ Ω ∖ ( S ∪ { N o = 0 } ) ) . m^{\mathrm{leaf}}=\gamma\otimes\mathbb{P}(\,\cdot\mid N_o=0),\qquad
m^\infty=\gamma\otimes\mathbb{P}^\infty,\qquad
m^{\mathrm{oth}}=\gamma\otimes\mathbb{P}(\,\cdot\mid\Omega\setminus(S\cup\{N_o=0\})). m leaf = γ ⊗ P ( ⋅ ∣ N o = 0 ) , m ∞ = γ ⊗ P ∞ , m oth = γ ⊗ P ( ⋅ ∣ Ω ∖ ( S ∪ { N o = 0 })) .
The last conditioning event has probability 1 − ϱ p − e − p ≥ P ( N o = 1 , N 1 = 0 ) = p e − 2 p > 0 1-\varrho_p-e^{-p}\geq\mathbb{P}(N_o=1,N_1=0)=pe^{-2p}>0 1 − ϱ p − e − p ≥ P ( N o = 1 , N 1 = 0 ) = p e − 2 p > 0 . Finally, let
G l e a f = σ ( N v , ζ v : v ∈ U ; w v 1 { N v ≥ 1 } : v ∈ U ∗ ) \mathcal{G}^{\mathrm{leaf}}=\sigma\left(N_{v},\zeta_v:v\in\mathbb{U};\ w_v\mathbf{1}\{N_v\geq1\}:v\in\mathbb{U}_*\right) G leaf = σ ( N v , ζ v : v ∈ U ; w v 1 { N v ≥ 1 } : v ∈ U ∗ )
be the σ \sigma σ -field generated by the configuration in which the marks of all edges leading to leaves have been erased.
Lemma 4.21. Let p > 1 p>1 p > 1 .
(i) Under P \mathbb{P} P , the processes Π l e a f \Pi^{\mathrm{leaf}} Π leaf , Π s u r v \Pi^{\mathrm{surv}} Π surv , and Π o t h \Pi^{\mathrm{oth}} Π oth formed by the points ( w k , θ k ω ) (w_k,\theta_k\omega) ( w k , θ k ω ) of the leaf children, the surviving children, and the other children of the root are independent Poisson processes with intensity measures p e − p m l e a f pe^{-p}m^{\mathrm{leaf}} p e − p m leaf , p ϱ p m ∞ p\varrho_pm^\infty p ϱ p m ∞ , and p ( 1 − ϱ p − e − p ) m o t h p(1-\varrho_p-e^{-p})m^{\mathrm{oth}} p ( 1 − ϱ p − e − p ) m oth , and they are independent of ζ o \zeta_o ζ o . In particular, N o l e a f N_o^{\mathrm{leaf}} N o leaf , N o s u r v N_o^{\mathrm{surv}} N o surv , and N o o t h N_o^{\mathrm{oth}} N o oth are independent Poisson variables with means p e − p pe^{-p} p e − p , p ϱ p p\varrho_p p ϱ p , and p ( 1 − ϱ p − e − p ) p(1-\varrho_p-e^{-p}) p ( 1 − ϱ p − e − p ) . Under P ∞ \mathbb{P}^\infty P ∞ , the processes Π l e a f \Pi^{\mathrm{leaf}} Π leaf , Π o t h \Pi^{\mathrm{oth}} Π oth , and the mark ζ o \zeta_o ζ o have the same laws as under P \mathbb{P} P and are independent of Π s u r v \Pi^{\mathrm{surv}} Π surv , which has the law of a Poisson process with intensity measure p ϱ p m ∞ p\varrho_pm^\infty p ϱ p m ∞ conditioned to have at least one point. In particular, for every measurable set B ⊂ M ∞ = C × S B\subset M^\infty=\mathbb{C}\times S B ⊂ M ∞ = C × S ,
P ∞ ( N o l e a f ≥ 1 , Π s u r v ( B ) ≥ 1 ) = ϱ p − 1 ( 1 − e − p e − p ) ( 1 − e − p m ( B ) ) . (54) \mathbb{P}^\infty\left(N_o^{\mathrm{leaf}}\geq1,\Pi^{\mathrm{surv}}(B)\geq1\right)=\varrho_p^{-1}\left(1-e^{-pe^{-p}}\right)\left(1-e^{-pm(B)}\right). \tag*{(54)} P ∞ ( N o leaf ≥ 1 , Π surv ( B ) ≥ 1 ) = ϱ p − 1 ( 1 − e − p e − p ) ( 1 − e − p m ( B ) ) . (54) (ii) Let R = { N o l e a f = 1 , N o s u r v = 1 , N o o t h = 0 } \mathcal{R}=\{N_o^{\mathrm{leaf}}=1,N_o^{\mathrm{surv}}=1,N_o^{\mathrm{oth}}=0\} R = { N o leaf = 1 , N o surv = 1 , N o oth = 0 } be the event that the root has exactly one leaf child, exactly one surviving child, and no other child, and on R \mathcal{R} R let ℓ , y ∈ { 1 , 2 } \ell,y\in\{1,2\} ℓ , y ∈ { 1 , 2 } be the labels of the leaf child and of the surviving child. Then R ⊂ S \mathcal{R}\subset S R ⊂ S , P ( R ) = p 2 ϱ p e − 2 p \mathbb{P}(\mathcal{R})=p^2\varrho_pe^{-2p} P ( R ) = p 2 ϱ p e − 2 p , and P ∞ ( R ) = p 2 e − 2 p \mathbb{P}^\infty(\mathcal{R})=p^2e^{-2p} P ∞ ( R ) = p 2 e − 2 p . Moreover, for all measurable F 1 , F 2 : M → [ 0 , ∞ ] F_1,F_2:M\to[0,\infty] F 1 , F 2 : M → [ 0 , ∞ ] and F 3 : Z → [ 0 , ∞ ] F_3:\mathbb{Z}\to[0,\infty] F 3 : Z → [ 0 , ∞ ] ,
E ∞ [ 1 R F 1 ( w y , θ y ω ) F 2 ( w ℓ , θ ℓ ω ) F 3 ( ζ o ) ] = p 2 e − 2 p ∫ F 1 d m ∞ ∫ F 2 d m l e a f ∫ F 3 d κ . (55) \mathbb{E}^\infty\left[\mathbf{1}_{\mathcal{R}}F_1(w_y,\theta_y\omega)F_2(w_\ell,\theta_\ell\omega)F_3(\zeta_o)\right]=p^2e^{-2p}\int F_1\,dm^\infty\int F_2\,dm^{\mathrm{leaf}}\int F_3\,d\kappa. \tag*{(55)} E ∞ [ 1 R F 1 ( w y , θ y ω ) F 2 ( w ℓ , θ ℓ ω ) F 3 ( ζ o ) ] = p 2 e − 2 p ∫ F 1 d m ∞ ∫ F 2 d m leaf ∫ F 3 d κ . (55) In particular, conditionally on R \mathcal{R} R under P ∞ \mathbb{P}^\infty P ∞ , the squared moduli Y ℓ Y_\ell Y ℓ and Y y Y_y Y y are independent Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) variables, the forward configuration θ y ω \theta_y\omega θ y ω of the surviving child has law P ∞ \mathbb{P}^\infty P ∞ , and these three random elements are independent of each other and of the phases w ℓ / ∣ w ℓ ∣ w_\ell/|w_\ell| w ℓ /∣ w ℓ ∣ , w y / ∣ w y ∣ w_y/|w_y| w y /∣ w y ∣ .
(iii) The root has exactly one child with probability P ( N o = 1 ) = P ∞ ( N o = 1 ) = p e − p \mathbb{P}(N_o=1)=\mathbb{P}^\infty(N_o=1)=pe^{-p} P ( N o = 1 ) = P ∞ ( N o = 1 ) = p e − p . Conditionally on { N o = 1 } \{N_o=1\} { N o = 1 } , the point ( w 1 , θ 1 ω ) (w_1,\theta_1\omega) ( w 1 , θ 1 ω ) has law m m m under P \mathbb{P} P and law m ∞ m^\infty m ∞ under P ∞ \mathbb{P}^\infty P ∞ . Along a spine, let L ≥ 1 L\geq1 L ≥ 1 , let J ⊂ { 0 , … , L − 1 } J\subset\{0,\ldots,L-1\} J ⊂ { 0 , … , L − 1 } , and let C J = { N v l = 1 for all l ∈ J } C_J=\{N_{v_l}=1\text{ for all }l\in J\} C J = { N v l = 1 for all l ∈ J } be the event that the spine vertices v l v_l v l , l ∈ J l\in J l ∈ J , have no side children. Then Q L ( C J ) = Q L ∞ ( C J ) = e − p ∣ J ∣ Q_L(C_J)=Q_L^\infty(C_J)=e^{-p|J|} Q L ( C J ) = Q L ∞ ( C J ) = e − p ∣ J ∣ ; conditionally on C J C_J C J , the conclusions of Lemma 4.7 hold with S l S_l S l empty for l ∈ J l\in J l ∈ J ; and for every measurable F : Ω × U → [ 0 , ∞ ] F:\Omega\times\mathbb{U}\to[0,\infty] F : Ω × U → [ 0 , ∞ ] ,
E ∑ v ∈ V L F ( ω , v ) ∏ l ∈ J 1 { N v l = 1 } = p L e − p ∣ J ∣ E Q L [ F ( ω , v ) ∣ C J ] . \mathbb{E}\sum_{v\in\mathcal{V}_L}F(\omega,v)\prod_{l\in J}\mathbf{1}\{N_{v_l}=1\}=p^Le^{-p|J|}\mathbb{E}_{Q_L}[F(\omega,v)\mid C_J]. E v ∈ V L ∑ F ( ω , v ) l ∈ J ∏ 1 { N v l = 1 } = p L e − p ∣ J ∣ E Q L [ F ( ω , v ) ∣ C J ] . The same identity holds with E ∞ \mathbb{E}^\infty E ∞ , V L ∞ \mathcal{V}_L^\infty V L ∞ , and Q L ∞ Q_L^\infty Q L ∞ in place of E \mathbb{E} E , V L \mathcal{V}_L V L , and Q L Q_L Q L .
(iv) Let X X X be a random finite subset of U \mathbb{U} U which is G l e a f \mathcal{G}^{\mathrm{leaf}} G leaf -measurable, and let A = { ℓ ( x ) : x ∈ X ∩ V ( ω ) , N x l e a f ≥ 1 } A=\{\ell(x):x\in X\cap V(\omega),\ N_x^{\mathrm{leaf}}\geq1\} A = { ℓ ( x ) : x ∈ X ∩ V ( ω ) , N x leaf ≥ 1 } be the set of selected leaf children of the vertices in X X X . Then
{ A = a } ∈ G a \{A=a\}\in\mathcal{G}_a { A = a } ∈ G a for every finite a ⊂ U ∗ a\subset\mathbb{U}_* a ⊂ U ∗ , and the conclusions of Lemma 4.17 hold for A A A under P \mathbb{P} P and P ∞ \mathbb{P}^{\infty} P ∞ . In particular, the selected squared moduli ( Y v ) v ∈ A (Y_v)_{v\in A} ( Y v ) v ∈ A are fresh given H A \mathcal{H}_A H A : conditionally on H A \mathcal{H}_A H A , they are independent Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) variables. The same holds under the spine laws if X = X ( ω , v ) X=X(\omega,v) X = X ( ω , v ) , respectively X ( ω , u , v ) X(\omega,u,v) X ( ω , u , v ) , is G l e a f \mathcal{G}_{\mathrm{leaf}} G leaf -measurable for each fixed value of the endpoints; for example, X = { v } X=\{v\} X = { v } or X = { u , v } X=\{u,v\} X = { u , v } .
(v) Under Q L \mathbb{Q}_L Q L , respectively Q L , j ∞ \mathbb{Q}^{\infty}_{L,j} Q L , j ∞ , the endpoint configuration θ v ω \theta_v\omega θ v ω has law P \mathbb{P} P , respectively P ∞ \mathbb{P}^{\infty} P ∞ , and is independent of the spine edge marks, the auxiliary marks of the spine vertices, and the side sequences. In particular, (i)–(iii) hold for the children of the endpoint v v v , with θ v ω \theta_v\omega θ v ω in place of ω \omega ω ; for example, Q L , j ∞ ( R ∘ θ v ) = p 2 e − 2 p \mathbb{Q}^{\infty}_{L,j}(\mathcal{R}\circ\theta_v)=p^2e^{-2p} Q L , j ∞ ( R ∘ θ v ) = p 2 e − 2 p , and (4.21) holds conditionally on the spine data. Under Q L , j ∞ \mathbb{Q}^{\infty}_{L,j} Q L , j ∞ , the events R ∘ θ u \mathcal{R}\circ\theta_u R ∘ θ u and R ∘ θ v \mathcal{R}\circ\theta_v R ∘ θ v are independent, each has probability p 2 e − 2 p p^2e^{-2p} p 2 e − 2 p , and (4.21) holds jointly at the two endpoints, with independent factors.
Proof. For (i), the sets M l e a f = C × { N o = 0 } M^{\mathrm{leaf}}=\mathcal{C}\times\{N_o=0\} M leaf = C × { N o = 0 } , M ∞ = C × S M^\infty=\mathcal{C}\times S M ∞ = C × S , and M o t h = C × ( Ω ∖ ( S ∪ { N o = 0 } ) ) M^{\mathrm{oth}}=\mathcal{C}\times(\Omega\setminus(S\cup\{N_o=0\})) M oth = C × ( Ω ∖ ( S ∪ { N o = 0 })) partition M M M , and their m m m -masses are e − p e^{-p} e − p , ϱ p \varrho_p ϱ p , and 1 − ϱ p − e − p 1-\varrho_p-e^{-p} 1 − ϱ p − e − p . For measurable u 1 , u 2 , u 3 : M → [ 0 , ∞ ) u_1,u_2,u_3:M\to[0,\infty) u 1 , u 2 , u 3 : M → [ 0 , ∞ ) , the computation in the proof of Lemma 4.5 (i) gives
E exp ( − Π l e a f ( u 1 ) − Π s u r v ( u 2 ) − Π o t h ( u 3 ) ) = ∏ i = 1 3 exp ( − p ∫ M i ( 1 − e − u i ) d m ) , \mathbb{E}\exp\left(-\Pi^{\mathrm{leaf}}(u_1)-\Pi^{\mathrm{surv}}(u_2)-\Pi^{\mathrm{oth}}(u_3)\right)=\prod_{i=1}^{3}\exp\left(-p\int_{M_i}(1-e^{-u_i})\,\mathrm{d}m\right), E exp ( − Π leaf ( u 1 ) − Π surv ( u 2 ) − Π oth ( u 3 ) ) = i = 1 ∏ 3 exp ( − p ∫ M i ( 1 − e − u i ) d m ) ,
where ( M 1 , M 2 , M 3 ) = ( M l e a f , M ∞ , M o t h ) (M_1,M_2,M_3)=(M^{\mathrm{leaf}},M^\infty,M^{\mathrm{oth}}) ( M 1 , M 2 , M 3 ) = ( M leaf , M ∞ , M oth ) . As there, this shows that the three processes are independent Poisson processes with intensity measures p m ( ⋅ ∩ M i ) p\,m(\mathord{\cdot}\cap M_i) p m ( ⋅ ∩ M i ) , which are the stated ones; they are independent of ζ o \zeta_o ζ o by Lemma 4.2 . The statement under P ∞ \mathbb{P}^{\infty} P ∞ follows as in Lemma 4.5 (ii), since S = { N o s u r v ≥ 1 } S=\{N_o^{\mathrm{surv}}\geq1\} S = { N o surv ≥ 1 } . For (4.20), we note that { Π s u r v ( B ) ≥ 1 } ⊂ S \{\Pi^{\mathrm{surv}}(B)\geq1\}\subset S { Π surv ( B ) ≥ 1 } ⊂ S , that Π s u r v ( B ) = Π o ( B ) \Pi^{\mathrm{surv}}(B)=\Pi_o(B) Π surv ( B ) = Π o ( B ) has law Pois ( p m ( B ) ) \operatorname{Pois}(p\,m(B)) Pois ( p m ( B )) , and that it is independent of N o l e a f N_o^{\mathrm{leaf}} N o leaf , which has law Pois ( p e − p ) \operatorname{Pois}(pe^{-p}) Pois ( p e − p ) ; dividing by ϱ p \varrho_p ϱ p gives (4.20).
For (ii), write M k = ( w k , θ k ω ) M_k=(w_k,\theta_k\omega) M k = ( w k , θ k ω ) . On R \mathcal{R} R , we have N o = 2 N_o=2 N o = 2 , and exactly one of M 1 , M 2 M_1,M_2 M 1 , M 2 lies in M ∞ M^\infty M ∞ while the other lies in M l e a f M^{\mathrm{leaf}} M leaf . By Lemma 4.2 (ii), N o N_o N o , ζ o \zeta_o ζ o , M 1 M_1 M 1 , M 2 M_2 M 2 are independent, and M 1 , M 2 M_1,M_2 M 1 , M 2 have law m m m . Using also P ( N o = 2 ) = p 2 e − p / 2 \mathbb{P}(N_o=2)=p^2e^{-p}/2 P ( N o = 2 ) = p 2 e − p /2 , we obtain
E [ 1 R F 1 ( M y ) F 2 ( M ℓ ) F 3 ( ζ o ) ] = P ( N o = 2 ) ⋅ 2 ∫ M ∞ F 1 d m ∫ M l e a f F 2 d m ∫ F 3 d κ = p 2 e − p ⋅ ϱ p ∫ F 1 d m ∞ ⋅ e − p ∫ F 2 d m l e a f ∫ F 3 d κ , \begin{aligned}
\mathbb{E}\left[\mathbf{1}_{\mathcal{R}}F_1(M_y)F_2(M_\ell)F_3(\zeta_o)\right]=\mathbb{P}(N_o=2)\cdot2\int_{M^\infty}F_1\,\mathrm{d}m\int_{M^{\mathrm{leaf}}}F_2\,\mathrm{d}m\int F_3\,\mathrm{d}\kappa\\
&=p^2e^{-p}\cdot\varrho_p\int F_1\,\mathrm{d}m^\infty\cdot e^{-p}\int F_2\,\mathrm{d}m^{\mathrm{leaf}}\int F_3\,\mathrm{d}\kappa,
\end{aligned} E [ 1 R F 1 ( M y ) F 2 ( M ℓ ) F 3 ( ζ o ) ] = P ( N o = 2 ) ⋅ 2 ∫ M ∞ F 1 d m ∫ M leaf F 2 d m ∫ F 3 d κ = p 2 e − p ⋅ ϱ p ∫ F 1 d m ∞ ⋅ e − p ∫ F 2 d m leaf ∫ F 3 d κ ,
where the factor 2 accounts for the two possible values of ( y , ℓ ) (y,\ell) ( y , ℓ ) . The surviving child has an infinite forward subtree, which gives R ⊂ S \mathcal{R}\subset S R ⊂ S . Dividing by ϱ p \varrho_p ϱ p gives (4.21), and taking F 1 = F 2 = F 3 = 1 F_1=F_2=F_3=1 F 1 = F 2 = F 3 = 1 gives the two probabilities. For the last assertion, we take F 1 ( w , ω ′ ′ ) = g 1 ( ∣ w ∣ 2 ) g 2 ( w / ∣ w ∣ ) g 3 ( ω ′ ′ ) F_1(w,\omega'')=g_1(|w|^2)g_2(w/|w|)g_3(\omega'') F 1 ( w , ω ′′ ) = g 1 ( ∣ w ∣ 2 ) g 2 ( w /∣ w ∣ ) g 3 ( ω ′′ ) and F 2 ( w , ω ′ ′ ) = g 4 ( ∣ w ∣ 2 ) g 5 ( w / ∣ w ∣ ) F_2(w,\omega'')=g_4(|w|^2)g_5(w/|w|) F 2 ( w , ω ′′ ) = g 4 ( ∣ w ∣ 2 ) g 5 ( w /∣ w ∣ ) , and use that under γ \gamma γ the modulus squared and the phase are independent with laws Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) and uniform.
For (iii), P ( N o = 1 ) = p e − p \mathbb{P}(N_o=1)=pe^{-p} P ( N o = 1 ) = p e − p , and { N o = 1 } ∩ S = { N o = 1 , θ 1 ω ∈ S } \{N_o=1\}\cap S=\{N_o=1,\theta_1\omega\in S\} { N o = 1 } ∩ S = { N o = 1 , θ 1 ω ∈ S } has probability p e − p ϱ p pe^{-p}\varrho_p p e − p ϱ p by Lemma 4.2 (ii); dividing by ϱ p \varrho_p ϱ p gives P ∞ ( N o = 1 ) = p e − p \mathbb{P}^{\infty}(N_o=1)=pe^{-p} P ∞ ( N o = 1 ) = p e − p . The conditional laws follow in the same way. Along a spine, we have N v l = N l ′ + 1 N_{v_l}=N'_l+1 N v l = N l ′ + 1 and therefore C J = { ∣ S l ∣ = 0 for l ∈ J } C_J=\{|S_l|=0\text{ for }l\in J\} C J = { ∣ S l ∣ = 0 for l ∈ J } , which has Q L Q_L Q L -probability e − p ∣ J ∣ e^{-p|J|} e − p ∣ J ∣ by Lemma 4.7 (i). Since C J C_J C J is an event in terms of the sequences S l S_l S l , conditioning on it preserves the independence in Lemma 4.7 and only changes the laws of S l S_l S l , l ∈ J l\in J l ∈ J , to the point mass at the empty sequence. Under Q ∞ Q^\infty Q ∞ the event C J C_J C J has the same probability, because it is independent of θ v ω \theta_v\omega θ v ω . The displayed identity is (38) applied to F ( ω , v ) ∏ l ∈ J 1 { N v l = 1 } F(\omega,v)\prod_{l\in J}1\{N_{v_l}=1\} F ( ω , v ) ∏ l ∈ J 1 { N v l = 1 } , and its conditioned version is (4.8).
For (iv), fix a finite set a ⊂ U ∗ a\subset\mathbb{U}_* a ⊂ U ∗ , and let ω ~ = ω a ← 0 \widetilde{\omega}=\omega^{a\leftarrow0} ω = ω a ← 0 be the configuration in which the marks w v w_v w v , v ∈ a v\in a v ∈ a , are replaced by 0. Every element of A A A is a leaf, and therefore { A = a } ⊂ { N v = 0 for all v ∈ a } \{A=a\}\subset\{N_v=0\text{ for all }v\in a\} { A = a } ⊂ { N v = 0 for all v ∈ a } . On the latter event, the generators w v 1 { N v ≥ 1 } w_v1\{N_v\geq1\} w v 1 { N v ≥ 1 } of G l e a f \mathcal{G}^{\mathrm{leaf}} G leaf take the same values at ω \omega ω and at ω ~ \widetilde{\omega} ω , and so do all offspring numbers and auxiliary marks. Since X X X is G l e a f \mathcal{G}^{\mathrm{leaf}} G leaf -measurable and the selection x ↦ ℓ ( x ) x\mapsto\ell(x) x ↦ ℓ ( x ) depends only on the offspring numbers and the ranks, we have 1 { A = a } ( ω ) = 1 { A = a } ( ω ~ ) 1\{A=a\}(\omega)=1\{A=a\}(\widetilde{\omega}) 1 { A = a } ( ω ) = 1 { A = a } ( ω ) on this event. Off this event, both sides vanish, because ω \omega ω and ω ~ \widetilde{\omega} ω have the same offspring numbers. Since the right side does not depend on ( w v ) v ∈ a (w_v)_{v\in a} ( w v ) v ∈ a , we conclude that { A = a } ∈ G a \{A=a\}\in\mathcal{G}_a { A = a } ∈ G a . The remaining assertions are Lemma 4.17; the independence of the selected leaf marks at distinct vertices of X X X is part of (4.15), which identifies their joint conditional law as γ ⊗ A \gamma^{\otimes A} γ ⊗ A . Under the spine laws, we argue in the same way for each fixed value of the endpoints and apply Lemma 4.17(ii).
Part (v) follows from Lemma 4.7 (i), (iii), and (iv).
We emphasize that the structural events in Lemma 4.21 are defined in terms of offspring numbers and survival, and that the selections use only offspring numbers and ranks; neither inspects an edge mark. This is why the selected squared moduli remain fresh; compare Remark 4.19, where the conditioning event is defined in terms of the squared modulus itself.
We conclude with the growth of the backbone. By Lemma 4.5 (ii), the probability under P ∞ \mathbb{P}^{\infty} P ∞ that the root has exactly one surviving child is
q = p ϱ p e − p ϱ p 1 − e − p ϱ p = p e − p ϱ p , q=\frac{p\varrho_p e^{-p\varrho_p}}{1-e^{-p\varrho_p}}=pe^{-p\varrho_p}, q = 1 − e − p ϱ p p ϱ p e − p ϱ p = p e − p ϱ p ,
where we used ϱ p = 1 − e − p ϱ p \varrho_p=1-e^{-p\varrho_p} ϱ p = 1 − e − p ϱ p . With λ = p ϱ p > 0 \lambda=p\varrho_p>0 λ = p ϱ p > 0 , we have q = λ / ( e λ − 1 ) < 1 q=\lambda/(e^\lambda-1)<1 q = λ / ( e λ − 1 ) < 1 , since e λ − 1 > λ e^\lambda-1>\lambda e λ − 1 > λ .
Lemma 4.22. Let p > 1 p>1 p > 1 and q = p e − p ϱ p q=pe^{-p\varrho_p} q = p e − p ϱ p .
(i) On S S S , we have Z 0 = 1 Z_0=1 Z 0 = 1 and Z n + 1 ≥ Z n Z_{n+1}\geq Z_n Z n + 1 ≥ Z n for every $n\geq 0.
(ii) For all n ≥ 0 n\geq0 n ≥ 0 , m ≥ 1 m\geq1 m ≥ 1 , and K ≥ 0 K\geq0 K ≥ 0 , P ∞ ( Z n = m , Z n + K = m ) = q m K P ∞ ( Z n = m ) \mathbb{P}^{\infty}(Z_n=m,Z_{n+K}=m)=q^{mK}\mathbb{P}^{\infty}(Z_n=m) P ∞ ( Z n = m , Z n + K = m ) = q m K P ∞ ( Z n = m ) .
(iii) Z n → ∞ Z_n\to\infty Z n → ∞ as n → ∞ n\to\infty n → ∞ , P ∞ \mathbb{P}^{\infty} P ∞ -almost surely.
Proof. The proof proceeds in four steps. After the elementary monotonicity, we compute the probability that the backbone of a surviving tree consists of a single line for $K generations, using the Mecke formula. Then we use the branching property at generation n n n to obtain (ii), from which (iii) follows.
Step 1. Monotonicity. On S \mathcal{S} S , the root has an infinite forward subtree, and Z 0 = 1 Z_0=1 Z 0 = 1 . If v ∈ V n ∞ v\in\mathcal{V}^{\infty}_n v ∈ V n ∞ , then the forward subtree of v v v is infinite and v v v has finitely many children, so at least one child of v v v has an infinite forward subtree and belongs to V n + 1 ∞ \mathcal{V}^{\infty}_{n+1} V n + 1 ∞ . Since distinct vertices have distinct children, we have Z n + 1 ≥ Z n Z_{n+1}\ge Z_n Z n + 1 ≥ Z n . This proves (i). Further, for all n , K ≥ 0 n,K\ge0 n , K ≥ 0 and ω ∈ Ω \omega\in\Omega ω ∈ Ω , we have
Z n + K ( ω ) = ∑ v ∈ V n ∞ Z K ( θ v ω ) . (56) Z_{n+K}(\omega)=\sum_{v\in\mathcal{V}^{\infty}_n}Z_K(\theta_v\omega). \tag*{(56)} Z n + K ( ω ) = v ∈ V n ∞ ∑ Z K ( θ v ω ) . (56)
Indeed, for v ∈ V n v\in\mathcal{V}_n v ∈ V n , the vertices of V n + K ∞ \mathcal{V}^{\infty}_{n+K} V n + K ∞ that descend from v v v are exactly the vertices v u vu v u with u ∈ V K ∞ ( θ v ω ) u\in\mathcal{V}^{\infty}_K(\theta_v\omega) u ∈ V K ∞ ( θ v ω ) , and there are none unless the forward subtree of v v v is infinite, that is, unless v ∈ V n ∞ v\in\mathcal{V}^{\infty}_n v ∈ V n ∞ .
Step 2. A single line. For K ≥ 0 K\ge0 K ≥ 0 , let A K = S ∩ { Z K = 1 } \mathcal{A}_K=\mathcal{S}\cap\{Z_K=1\} A K = S ∩ { Z K = 1 } . We claim that P ∞ ( A K ) = q K \mathbb{P}^{\infty}(\mathcal{A}_K)=q^K P ∞ ( A K ) = q K . For K = 0 K=0 K = 0 this holds since A 0 = S \mathcal{A}_0=\mathcal{S} A 0 = S . Let K ≥ 1 K\ge1 K ≥ 1 . By Step 1, we have Z 1 = 1 Z_1=1 Z 1 = 1 on A K \mathcal{A}_K A K , that is, the root has exactly one child k k k with θ k ω ∈ S \theta_k\omega\in\mathcal{S} θ k ω ∈ S , and (4.22) with n = 1 n=1 n = 1 gives Z K − 1 ( θ k ω ) = 1 Z_{K-1}(\theta_k\omega)=1 Z K − 1 ( θ k ω ) = 1 . Conversely, if the root has exactly one surviving child k k k and θ k ω ∈ A K − 1 \theta_k\omega\in\mathcal{A}_{K-1} θ k ω ∈ A K − 1 , then ω ∈ A K \omega\in\mathcal{A}_K ω ∈ A K . In terms of the child process, this reads
1 A K ( ω ) = ∑ k = 1 N 0 1 A K − 1 ( θ k ω ) ∏ k ′ ≤ N 0 k ′ ≠ k 1 Ω ∖ S ( θ k ′ ω ) = ∫ F ( m , Π 0 − δ m ) Π 0 ( d m ) , \mathbf{1}_{\mathcal{A}_K}(\omega)=\sum_{k=1}^{N_0}\mathbf{1}_{\mathcal{A}_{K-1}}(\theta_k\omega)\prod_{\substack{k'\le N_0\\ k'\ne k}}\mathbf{1}_{\Omega\setminus\mathcal{S}}(\theta_{k'}\omega)=\int F(m,\Pi_0-\delta_m)\,\Pi_0(dm), 1 A K ( ω ) = k = 1 ∑ N 0 1 A K − 1 ( θ k ω ) k ′ ≤ N 0 k ′ = k ∏ 1 Ω ∖ S ( θ k ′ ω ) = ∫ F ( m , Π 0 − δ m ) Π 0 ( d m ) ,
where F ( ( w , ω ′ ′ ) , π ) = 1 A K − 1 ( ω ′ ′ ) 1 { π ( M ∞ ) = 0 } F((w,\omega''),\pi)=\mathbf{1}_{\mathcal{A}_{K-1}}(\omega'')\mathbf{1}_{\{\pi(M^\infty)=0\}} F (( w , ω ′′ ) , π ) = 1 A K − 1 ( ω ′′ ) 1 { π ( M ∞ ) = 0 } and M ∞ = C × S M^\infty=\mathcal{C}\times\mathcal{S} M ∞ = C × S . In the first equality we used A K − 1 ⊂ S \mathcal{A}_{K-1}\subset\mathcal{S} A K − 1 ⊂ S . By Lemma 4.2(iii) and (4.1), we obtain
P ( A K ) = p P ( A K − 1 ) P ( Π 0 ( M ∞ ) = 0 ) = p e − p ϱ p P ( A K − 1 ) . \mathbb{P}(\mathcal{A}_K)=p\mathbb{P}(\mathcal{A}_{K-1})\mathbb{P}(\Pi_0(M^\infty)=0)=p e^{-p\varrho_p}\mathbb{P}(\mathcal{A}_{K-1}). P ( A K ) = p P ( A K − 1 ) P ( Π 0 ( M ∞ ) = 0 ) = p e − p ϱ p P ( A K − 1 ) .
Dividing by ϱ p \varrho_p ϱ p , we find P ∞ ( A K ) = q P ∞ ( A K − 1 ) \mathbb{P}^{\infty}(\mathcal{A}_K)=q\mathbb{P}^{\infty}(\mathcal{A}_{K-1}) P ∞ ( A K ) = q P ∞ ( A K − 1 ) , and the claim follows by induction.
Step 3. Proof of (ii). Let n ≥ 0 n\ge0 n ≥ 0 , m ≥ 1 m\ge1 m ≥ 1 , and K ≥ 0 K\ge0 K ≥ 0 . On { Z n = m } \{Z_n=m\} { Z n = m } , each term in (4.22) is at least 1 by Step 1, and Z n + K = m Z_{n+K}=m Z n + K = m holds if and only if θ v ω ∈ A K \theta_v\omega\in\mathcal{A}_K θ v ω ∈ A K for every v ∈ V n ∞ v\in\mathcal{V}^{\infty}_n v ∈ V n ∞ . We obtain
1 { Z n = m , Z n + K = m } = ∑ a ⊂ U n ∣ a ∣ = m 1 { a ⊂ V n } ∏ v ∈ a 1 A K ( θ v ω ) ∏ v ∈ V n ∖ a 1 Ω ∖ S ( θ v ω ) , \mathbf{1}_{\{Z_n=m,\ Z_{n+K}=m\}}=\sum_{\substack{a\subset\mathcal{U}_n\\ |a|=m}}\mathbf{1}_{\{a\subset\mathcal{V}_n\}}\prod_{v\in a}\mathbf{1}_{\mathcal{A}_K}(\theta_v\omega)\prod_{v\in\mathcal{V}_n\setminus a}\mathbf{1}_{\Omega\setminus\mathcal{S}}(\theta_v\omega), 1 { Z n = m , Z n + K = m } = a ⊂ U n ∣ a ∣ = m ∑ 1 { a ⊂ V n } v ∈ a ∏ 1 A K ( θ v ω ) v ∈ V n ∖ a ∏ 1 Ω ∖ S ( θ v ω ) ,
since only the term with a = V n ∞ a=\mathcal{V}^{\infty}_n a = V n ∞ can be nonzero. By Lemma 4.2(i), the configurations θ v ω \theta_v\omega θ v ω , v ∈ U n v\in\mathcal{U}_n v ∈ U n , are independent with law P \mathbb{P} P and independent of F n \mathcal{F}_n F n , and the finite set V n \mathcal{V}_n V n is F n \mathcal{F}_n F n -measurable. Conditioning on F n \mathcal{F}_n F n , we find that
P ( Z n = m , Z n + K = m ∣ F n ) = ∑ a ⊂ U n ∣ a ∣ = m 1 { a ⊂ V n } P ( A K ) m ( 1 − ϱ p ) ∣ V n ∣ − m . \mathbb{P}(Z_n=m,\ Z_{n+K}=m\mid\mathcal{F}_n)=\sum_{\substack{a\subset\mathcal{U}_n\\ |a|=m}}\mathbf{1}_{\{a\subset\mathcal{V}_n\}}\mathbb{P}(\mathcal{A}_K)^m(1-\varrho_p)^{|\mathcal{V}_n|-m}. P ( Z n = m , Z n + K = m ∣ F n ) = a ⊂ U n ∣ a ∣ = m ∑ 1 { a ⊂ V n } P ( A K ) m ( 1 − ϱ p ) ∣ V n ∣ − m .
The same computation with S S S in place of A K A_K A K gives the conditional probability of { Z n = m } \{Z_n=m\} { Z n = m } , with q p m q_p^m q p m in place of P ( A K ) m \mathbb{P}(A_K)^m P ( A K ) m . By Step 2, we have P ( A K ) = q p q K \mathbb{P}(A_K)=q_pq^K P ( A K ) = q p q K , and the first conditional probability is therefore q m K q^{mK} q m K times the second. Taking expectations, and using that both events are contained in S S S because m ≥ 1 m\geq1 m ≥ 1 , we deduce (ii) after division by q p q_p q p .
Step 4. Proof of (iii). By (i), the limit Z ∞ = lim n Z n ∈ { 1 , 2 , … } ∪ { ∞ } Z_\infty=\lim_{n}Z_n\in\{1,2,\ldots\}\cup\{\infty\} Z ∞ = lim n Z n ∈ { 1 , 2 , … } ∪ { ∞ } exists on S S S . For m ≥ 1 m\geq1 m ≥ 1 , we have
{ Z ∞ = m } ⊂ ⋃ n ≥ 0 ⋂ K ≥ 0 { Z n = m , Z n + K = m } . \{Z_\infty=m\}\subset\bigcup_{n\geq0}\bigcap_{K\geq0}\{Z_n=m,Z_{n+K}=m\}. { Z ∞ = m } ⊂ n ≥ 0 ⋃ K ≥ 0 ⋂ { Z n = m , Z n + K = m } .
By (ii), each event in the union has P ∞ \mathbb{P}^{\infty} P ∞ -probability at most inf K q m K = 0 \inf_K q^{mK}=0 inf K q m K = 0 , since q < 1 q<1 q < 1 . We conclude that P ∞ ( Z ∞ < ∞ ) = 0 \mathbb{P}^{\infty}(Z_\infty<\infty)=0 P ∞ ( Z ∞ < ∞ ) = 0 , which is (iii).
Real solutions of the cavity equation
At an energy E > 0 E > 0 E > 0 where the imaginary parts of the cavity variables vanish at the real axis, the cavity equation of Proposition 3.16 becomes an equation for a probability measure on the real line. This section studies that equation and its relation with the tree.
In § 5.2.1 , we define real solutions and the normalizations used in the rest of the paper. In § 5.2.2 , we prove properties shared by all real solutions: the structure of their atoms, a density bound near zero, fractional moments of both signs, full support, a fractional cancellation estimate, and a bound on the probability that the denominator is small on the scale p − 1 / 2 p^{-1/2} p − 1/2 . These properties hold for every p > 1 p > 1 p > 1 , every E > 0 E > 0 E > 0 and every real solution. The atom structure, the moments, the full support and the fractional cancellation estimate are the hypotheses under which the free energy is defined and identified in § 6.2 .
The rest of the section connects real solutions with the tree. In § 5.2.3 , we define a deterministic set E p \mathcal{E}_p E p of regular energies, of full Lebesgue measure, at which all boundary values used in this paper exist, and we prove a zero–one law for the imaginary part of the root cavity variable. If this imaginary part vanishes under P ∞ \mathbb{P}^{\infty} P ∞ , then the denominators at the real axis are finite, real and nonzero, and the law of the root denominator is a real solution, which we call the real boundary law. In § 5.5.1 , we construct, for an arbitrary real solution, a recursive tree process with that solution as its marginal. § 5.5.2 records fractional moment bounds at z = E + i η z = E + i\eta z = E + i η with η > 0 \eta> 0 η > 0 that are uniform in η \eta η . Finally, § 5.5.3 proves that when the boundary is real, every cavity variable in a fixed neighborhood of the root, of a spine, or of two spines converges as the imaginary part η \eta η tends to zero, and identifies the limit with the real recursive tree process.
Throughout this section, p > 1 p > 1 p > 1 is fixed. Dependence of constants on parameters is indicated explicitly.
Real solutions and their normalizations
By Proposition 3.16, the root cavity variable Γ o ( z ) \Gamma_o(z) Γ o ( z ) satisfies the cavity equation
Γ o ( z ) = d − ( z + 1 p ∑ j = 1 N γ j Γ j ( z ) ) − 1 , (57) \Gamma_o(z) \overset{\mathrm{d}}{=} -\left(z+\frac{1}{p}\sum_{j=1}^{N}\gamma_j\Gamma_j(z)\right)^{-1}, \tag*{(57)} Γ o ( z ) = d − ( z + p 1 j = 1 ∑ N γ j Γ j ( z ) ) − 1 , (57)
where N ∼ Pois ( p ) N \sim\operatorname{Pois}(p) N ∼ Pois ( p ) , the Y j Y_j Y j are Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) variables, the Γ j ( z ) \Gamma_j(z) Γ j ( z ) are copies of Γ o ( z ) \Gamma_o(z) Γ o ( z ) , and all these variables are independent. If the variables are real at z = E z=E z = E , then the inverse cavity variable D = − 1 / Γ D=-1/\Gamma D = − 1/Γ , which we call the denominator, satisfies the equation in the following definition.
Definition 5.1. Let E > 0 E>0 E > 0 . A real solution at energy E E E is a Borel probability measure π \pi π on R ∖ { 0 } \mathbb{R}\setminus\{0\} R ∖ { 0 } with the following property. Let N N N , ( Y j ) j ≥ 1 (Y_j)_{j\ge1} ( Y j ) j ≥ 1 and ( D j ) j ≥ 1 (D_j)_{j\ge1} ( D j ) j ≥ 1 be independent, with N ∼ Pois ( p ) N\sim\operatorname{Pois}(p) N ∼ Pois ( p ) , Y j ∼ Exp ( 1 ) Y_j\sim\operatorname{Exp}(1) Y j ∼ Exp ( 1 ) and D j ∼ π D_j\sim\pi D j ∼ π for every j j j . Then the random variable
E − 1 p ∑ j = 1 N Y j D j (58) E-\frac{1}{p}\sum_{j=1}^{N}\frac{Y_j}{D_j} \tag*{(58)} E − p 1 j = 1 ∑ N D j Y j (58) has law π \pi π . We abbreviate this property by
D = d E − 1 p ∑ j = 1 N Y j D j , D \mathrel{\overset{\mathrm{d}}{=}} E-\frac{1}{p}\sum_{j=1}^{N}\frac{Y_j}{D_j}, D = d E − p 1 j = 1 ∑ N D j Y j , where D ∼ π D\sim\pi D ∼ π and the variables on the right are as above, and we call D D D the denominator .
Since D j ≠ 0 D_j\ne0 D j = 0 almost surely and N N N is finite, the sum in (5.1) is a real number almost surely, and the definition makes sense. Since π \pi π is carried by R ∖ { 0 } \mathbb{R}\setminus\{0\} R ∖ { 0 } , the definition includes the requirement that the variable (5.1) is nonzero almost surely. We only consider positive energies. By Proposition 3.21 , spectral statements at negative energies follow from those at positive energies, and if π \pi π is a real solution at energy E E E , then the law of − D -D − D satisfies the same equation with E E E replaced by − E -E − E .
The following normalizations are used throughout the paper.
Definition 5.2. Let E > 0 E>0 E > 0 , let π \pi π be a real solution at energy E E E , and let D ∼ π D\sim\pi D ∼ π . Let Y ∼ Exp ( 1 ) Y\sim\operatorname{Exp}(1) Y ∼ Exp ( 1 ) be independent of D D D .
For a scaling g > 0 g>0 g > 0 , set B = g D B=gD B = g D and β = g 2 Y / p \beta=g^2Y/p β = g 2 Y / p . The transfer matrix is
M ( B , β ) = β − 1 / 2 ( 0 1 − β B ) . M(B,\beta)=\beta^{-1/2}
\begin{pmatrix}
0 & 1\\
-\beta& B
\end{pmatrix}. M ( B , β ) = β − 1/2 ( 0 − β 1 B ) . For E > 2 E>2 E > 2 , set
g ( E ) = E − E 2 − 4 2 , θ = g ( E ) 2 , K = p θ , g(E)=\frac{E-\sqrt{E^2-4}}{2},\qquad\theta=g(E)^2,\qquad K=\frac{p}{\theta}, g ( E ) = 2 E − E 2 − 4 , θ = g ( E ) 2 , K = θ p , and, with the scaling g = g ( E ) g=g(E) g = g ( E ) , set B = g ( E ) D B=g(E)D B = g ( E ) D and X = 1 / B X=1/B X = 1/ B . We call this the B B B -normalization .
The scaled energy is ξ = E / p \xi=E/\sqrt{p} ξ = E / p . For ξ > 2 / p \xi>2/\sqrt{p} ξ > 2/ p , set
a p ( ξ ) = 1 p g ( ξ p ) . a_p(\xi)=\frac{1}{\sqrt{p}\,g(\xi\sqrt{p})}. a p ( ξ ) = p g ( ξ p ) 1 .
(iv) The scaling g = 1 / E g=1/E g = 1/ E gives B ~ = D / E \widetilde{B}=D/E B = D / E , in which the atom of the law sits at 1 (see Proposition 5.5(i)). We use the parameter τ = 1 / ( p E 2 ) \tau=1/(pE^2) τ = 1/ ( p E 2 ) .
We record the identities behind these normalizations.
Lemma 5.3. Let E > 0 E>0 E > 0 , let π \pi π be a real solution at energy E E E , and let N N N , ( Y j ) (Y_j) ( Y j ) , ( D j ) (D_j) ( D j ) be as in Definition 5.1.
(i) For every g > 0 g>0 g > 0 , the matrix M ( B , β ) M(B,\beta) M ( B , β ) has determinant one, and with B j = g D j B_j=gD_j B j = g D j and β j = g 2 Y j / p \beta_j=g^2Y_j/p β j = g 2 Y j / p we have
B = d g E − ∑ j = 1 N β j B j . B \mathrel{\mathop{=}^{\mathrm d}} gE-\sum_{j=1}^{N}\frac{\beta_j}{B_j}. B = d g E − j = 1 ∑ N B j β j . (ii) Let E > 2 E>2 E > 2 . Then E g ( E ) = 1 + θ Eg(E)=1+\theta E g ( E ) = 1 + θ , β = Y / K \beta=Y/K β = Y / K , X = − Γ / g ( E ) X=-\Gamma/g(E) X = − Γ/ g ( E ) when D = − 1 / Γ D=-1/\Gamma D = − 1/Γ , and
B = d 1 + θ − θ p ∑ j = 1 N Y j B j . B \mathrel{\mathop{=}^{\mathrm d}} 1+\theta-\frac{\theta}{p}\sum_{j=1}^{N}\frac{Y_j}{B_j}. B = d 1 + θ − p θ j = 1 ∑ N B j Y j . Further, E − 1 < g ( E ) ≤ 2 E − 1 E^{-1}<g(E)\leq2E^{-1} E − 1 < g ( E ) ≤ 2 E − 1 ; in particular g ( E ) < 1 g(E)<1 g ( E ) < 1 and E − 2 < θ ≤ 4 E − 2 E^{-2}<\theta\leq4E^{-2} E − 2 < θ ≤ 4 E − 2 .
(iii) Let E = ξ p > 2 E=\xi\sqrt{p}>2 E = ξ p > 2 . Then
a p ( ξ ) = ξ + ξ 2 − 4 / p 2 , K = p 2 a p ( ξ ) 2 , ξ 2 ≤ a p ( ξ ) < ξ . a_p(\xi)=\frac{\xi+\sqrt{\xi^2-4/p}}{2},\qquad K=p^2a_p(\xi)^2,\qquad\frac{\xi}{2}\leq a_p(\xi)<\xi. a p ( ξ ) = 2 ξ + ξ 2 − 4/ p , K = p 2 a p ( ξ ) 2 , 2 ξ ≤ a p ( ξ ) < ξ . (iv) With B ~ j = D j / E \widetilde{B}_j=D_j/E B j = D j / E , we have
B ~ = d 1 − τ ∑ j = 1 N Y j B ~ j . \widetilde{B}\mathrel{\mathop{=}^{\mathrm d}}1-\tau\sum_{j=1}^{N}\frac{Y_j}{\widetilde{B}_j}. B = d 1 − τ j = 1 ∑ N B j Y j . If E > 2 E>2 E > 2 , then B ~ = B / ( 1 + θ ) \widetilde{B}=B/(1+\theta) B = B / ( 1 + θ ) and τ = θ / ( p ( 1 + θ ) 2 ) \tau=\theta/(p(1+\theta)^2) τ = θ / ( p ( 1 + θ ) 2 ) .
Proof. The determinant of M ( B , β ) M(B,\beta) M ( B , β ) is β − 1 ( 0 ⋅ B + β ) = 1 \beta^{-1}(0\cdot B+\beta)=1 β − 1 ( 0 ⋅ B + β ) = 1 . Multiplying (5.1) by g g g , we obtain
g D = d g E − g p ∑ j = 1 N Y j D j = g E − ∑ j = 1 N g 2 Y j / p g D j , gD\mathrel{\mathop{=}^{\mathrm d}}gE-\frac{g}{p}\sum_{j=1}^{N}\frac{Y_j}{D_j}
=gE-\sum_{j=1}^{N}\frac{g^2Y_j/p}{gD_j}, g D = d g E − p g j = 1 ∑ N D j Y j = g E − j = 1 ∑ N g D j g 2 Y j / p ,
which is (i).
For (ii), the numbers ( E ± E 2 − 4 ) / 2 (E\pm\sqrt{E^2-4})/2 ( E ± E 2 − 4 ) /2 are the roots of t 2 − E t + 1 = 0 t^2-Et+1=0 t 2 − E t + 1 = 0 . In particular, g ( E ) 2 − E g ( E ) + 1 = 0 g(E)^2-Eg(E)+1=0 g ( E ) 2 − E g ( E ) + 1 = 0 , that is, E g ( E ) = 1 + θ Eg(E)=1+\theta E g ( E ) = 1 + θ , and since the product of the two roots is one, we have
1 g ( E ) = E + E 2 − 4 2 . (59) \frac{1}{g(E)}=\frac{E+\sqrt{E^2-4}}{2}. \tag*{(59)} g ( E ) 1 = 2 E + E 2 − 4 . (59)
Inserting g E = 1 + θ gE=1+\theta g E = 1 + θ and g 2 / p = θ / p g^2/p=\theta/p g 2 / p = θ / p into (i), we obtain the equation for B B B . The identities β = g ( E ) 2 Y / p = Y / K \beta=g(E)^2Y/p=Y/K β = g ( E ) 2 Y / p = Y / K and X = 1 / ( g ( E ) D ) = − Γ / g ( E ) X=1/(g(E)D)=-\Gamma/g(E) X = 1/ ( g ( E ) D ) = − Γ/ g ( E ) follow from the definitions. Since E ≤ E + E 2 − 4 < 2 E E \le E + \sqrt{E^2 - 4} < 2E E ≤ E + E 2 − 4 < 2 E , (59) implies that E − 1 < g ( E ) ≤ 2 E − 1 E^{-1} < g(E) \le2E^{-1} E − 1 < g ( E ) ≤ 2 E − 1 , and the remaining bounds follow because E > 2 E > 2 E > 2 .
For (iii), using (59) , we find
a p ( ζ ) = E + E 2 − 4 2 p = ζ p + ζ 2 p − 4 2 p = ζ + ζ 2 − 4 / p 2 , a_p(\zeta) = \frac{E+\sqrt{E^2-4}}{2\sqrt{p}} = \frac{\zeta\sqrt{p}+\sqrt{\zeta^2p-4}}{2\sqrt{p}} = \frac{\zeta+\sqrt{\zeta^2-4/p}}{2}, a p ( ζ ) = 2 p E + E 2 − 4 = 2 p ζ p + ζ 2 p − 4 = 2 ζ + ζ 2 − 4/ p ,
and K = p / g ( E ) 2 = p ⋅ p a p ( ζ ) 2 K = p/g(E)^2 = p\cdot p a_p(\zeta)^2 K = p / g ( E ) 2 = p ⋅ p a p ( ζ ) 2 . The bounds on a p ( ζ ) a_p(\zeta) a p ( ζ ) follow from 0 ≤ ζ 2 − 4 / p < ζ 0 \le\sqrt{\zeta^2-4/p} < \zeta 0 ≤ ζ 2 − 4/ p < ζ .
For (iv), we divide (5.1) by E E E and use Y j / ( p E D j ) = τ Y j / B ~ j Y_j/(pED_j)=\tau Y_j/\widetilde{B}_j Y j / ( pE D j ) = τ Y j / B j . If E > 2 E>2 E > 2 , then B / ( 1 + θ ) = g ( E ) D / ( g ( E ) E ) = B ~ B/(1+\theta)=g(E)D/(g(E)E)=\widetilde{B} B / ( 1 + θ ) = g ( E ) D / ( g ( E ) E ) = B , and θ / ( p ( 1 + θ ) 2 ) = g ( E ) 2 / ( p g ( E ) 2 E 2 ) = τ \theta/(p(1+\theta)^2)=g(E)^2/(pg(E)^2E^2)=\tau θ / ( p ( 1 + θ ) 2 ) = g ( E ) 2 / ( p g ( E ) 2 E 2 ) = τ .
The following table summarizes the normalizations. The last column is the atom produced by the event { N = 0 } \{N=0\} { N = 0 } , on which the denominator equals the constant term of its equation; by Proposition 5.5 , this is the only atom of a real solution.
Table 1.
A priori properties of real solutions
We use the density bound of Lemma 4.18 (iii) for affine combinations of independent exponential variables, which is proved there by the Gamma–Dirichlet decomposition, and the following bound on translated Gamma integrals.
Lemma 5.4. Let q ∈ ( 0 , 1 ) q\in(0,1) q ∈ ( 0 , 1 ) , and let T q T_q T q be a random variable with density
ω q ( t ) = t q / 2 e − t Γ ( 1 + q / 2 ) , t > 0. \omega_q(t)=\frac{t^{q/2}e^{-t}}{\Gamma(1+q/2)},\qquad t>0. ω q ( t ) = Γ ( 1 + q /2 ) t q /2 e − t , t > 0. Then for every u ∈ R u\in\mathbb{R} u ∈ R ,
E ∣ T q − u ∣ − q ≤ 70 1 − q ( 1 + ∣ u ∣ ) − q . \mathbb{E}|T_q-u|^{-q}\le\frac{70}{1-q}(1+|u|)^{-q}. E ∣ T q − u ∣ − q ≤ 1 − q 70 ( 1 + ∣ u ∣ ) − q . Proof. Let I ( u ) = E ∣ T q − u ∣ − q I(u)=\mathbb{E}|T_q-u|^{-q} I ( u ) = E ∣ T q − u ∣ − q . Since t q / 2 e − t / 2 ≤ ( q / e ) q / 2 ≤ 1 t^{q/2}e^{-t/2}\le(q/e)^{q/2}\le1 t q /2 e − t /2 ≤ ( q / e ) q /2 ≤ 1 for t > 0 t>0 t > 0 and Γ ( 1 + q / 2 ) ≥ 1 / 2 \Gamma(1+q/2)\ge1/2 Γ ( 1 + q /2 ) ≥ 1/2 , we have
ω q ( t ) ≤ 2 e − t / 2 ≤ 2 , t > 0. (60) \omega_q(t)\le2e^{-t/2}\le2,\qquad t>0. \tag*{(60)} ω q ( t ) ≤ 2 e − t /2 ≤ 2 , t > 0. (60) We consider three cases.
Let ∣ u ∣ ≤ 2 |u| \le2 ∣ u ∣ ≤ 2 . Splitting the integral according to whether ∣ t − u ∣ ≤ 1 |t-u| \le1 ∣ t − u ∣ ≤ 1 and using (5.3) on the first part, we have
I ( u ) ≤ 2 ∫ ∣ t − u ∣ ≤ 1 ∣ t − u ∣ − q d t + ∫ 0 ∞ ω q ( t ) d t = 4 1 − q + 1 ≤ 5 1 − q . I(u) \le2\int_{|t-u|\le1}|t-u|^{-q}\,\mathrm{d}t+\int_{0}^{\infty}\omega_q(t)\,\mathrm{d}t=\frac{4}{1-q}+1\le\frac{5}{1-q}. I ( u ) ≤ 2 ∫ ∣ t − u ∣ ≤ 1 ∣ t − u ∣ − q d t + ∫ 0 ∞ ω q ( t ) d t = 1 − q 4 + 1 ≤ 1 − q 5 .
Since ( 1 + ∣ u ∣ ) − q ≥ 3 − q ≥ 1 / 3 (1+|u|)^{-q}\ge3^{-q}\ge1/3 ( 1 + ∣ u ∣ ) − q ≥ 3 − q ≥ 1/3 , this gives I ( u ) ≤ 15 ( 1 − q ) − 1 ( 1 + ∣ u ∣ ) − q I(u)\le15(1-q)^{-1}(1+|u|)^{-q} I ( u ) ≤ 15 ( 1 − q ) − 1 ( 1 + ∣ u ∣ ) − q .
Let u < − 2 u<-2 u < − 2 . Since ∣ t − u ∣ = t + ∣ u ∣ ≥ ∣ u ∣ |t-u|=t+|u|\ge|u| ∣ t − u ∣ = t + ∣ u ∣ ≥ ∣ u ∣ for t > 0 t>0 t > 0 , we have I ( u ) ≤ ∣ u ∣ − q ≤ 2 ( 1 + ∣ u ∣ ) − q I(u)\le|u|^{-q}\le2(1+|u|)^{-q} I ( u ) ≤ ∣ u ∣ − q ≤ 2 ( 1 + ∣ u ∣ ) − q , where we used 1 + ∣ u ∣ ≤ 2 ∣ u ∣ 1+|u|\le2|u| 1 + ∣ u ∣ ≤ 2∣ u ∣ .
Let u > 2 u>2 u > 2 . We split ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) into ( 0 , u / 2 ] (0,u/2] ( 0 , u /2 ] , the set of t > u / 2 t>u/2 t > u /2 with ∣ t − u ∣ ≤ 1 |t-u|\le1 ∣ t − u ∣ ≤ 1 , and the set of t > u / 2 t>u/2 t > u /2 with ∣ t − u ∣ > 1 |t-u|>1 ∣ t − u ∣ > 1 . On the first set, we have ∣ t − u ∣ ≥ u / 2 |t-u|\ge u/2 ∣ t − u ∣ ≥ u /2 , and its contribution is at most ( u / 2 ) − q ≤ 2 u − q (u/2)^{-q}\le2u^{-q} ( u /2 ) − q ≤ 2 u − q . On the second set, we have ω q ( t ) ≤ 2 e − u / 4 \omega_q(t)\le2e^{-u/4} ω q ( t ) ≤ 2 e − u /4 by (5.3), and its contribution is at most 4 e − u / 4 / ( 1 − q ) 4e^{-u/4}/(1-q) 4 e − u /4 / ( 1 − q ) . The contribution of the third set is at most
∫ u / 2 ∞ 2 e − t / 2 d t = 4 e − u / 4 . \int_{u/2}^{\infty}2e^{-t/2}\,\mathrm{d}t=4e^{-u/4}. ∫ u /2 ∞ 2 e − t /2 d t = 4 e − u /4 .
Since the maximum of u q e − u / 4 u^qe^{-u/4} u q e − u /4 over u > 0 u>0 u > 0 is ( 4 q / e ) q ≤ 4 (4q/e)^q\le4 ( 4 q / e ) q ≤ 4 , we have e − u / 4 ≤ 4 u − q e^{-u/4}\le4u^{-q} e − u /4 ≤ 4 u − q . Combining these bounds, we obtain
I ( u ) ≤ ( 2 + 16 1 − q + 16 ) u − q ≤ 34 1 − q u − q ≤ 68 1 − q ( 1 + u ) − q . I(u)\le\left(2+\frac{16}{1-q}+16\right)u^{-q}\le\frac{34}{1-q}u^{-q}\le\frac{68}{1-q}(1+u)^{-q}. I ( u ) ≤ ( 2 + 1 − q 16 + 16 ) u − q ≤ 1 − q 34 u − q ≤ 1 − q 68 ( 1 + u ) − q .
where in the last step we used 1 + u ≤ 2 u 1+u\le2u 1 + u ≤ 2 u . The three cases prove the lemma.
Proposition 5.5. Let E > 0 E>0 E > 0 , let π \pi π be a real solution at energy E E E , and let D ∼ π D\sim\pi D ∼ π .
(i) We have π = e − p δ E + π a c \pi=e^{-p}\delta_E+\pi_{\mathrm{ac}} π = e − p δ E + π ac , where π a c \pi_{\mathrm{ac}} π ac is absolutely continuous with respect to Lebesgue measure. In particular, π \pi π has an atom of mass e − p e^{-p} e − p at E E E and no other atom.
(ii) The density f f f of π a c \pi_{\mathrm{ac}} π ac can be chosen so that
f ( x ) ≤ p 2 π 1 ∣ x − E ∣ for all x ≠ E . f(x)\le\sqrt{\frac{p}{2\pi}}\frac{1}{|x-E|}\qquad\text{for all }x\ne E. f ( x ) ≤ 2 π p ∣ x − E ∣ 1 for all x = E . In particular, f ( x ) ≤ p / E f(x)\le\sqrt{p}/E f ( x ) ≤ p / E for ∣ x ∣ ≤ E / 2 |x|\le E/2 ∣ x ∣ ≤ E /2 , and P ( ∣ D ∣ ≤ t ) ≤ 2 p t / E \mathbb{P}(|D|\le t)\le2\sqrt{p}\,t/E P ( ∣ D ∣ ≤ t ) ≤ 2 p t / E for 0 ≤ t ≤ E / 2 0\le t\le E/2 0 ≤ t ≤ E /2 .
(iii) For every q ∈ ( 0 , 1 ) q\in(0,1) q ∈ ( 0 , 1 ) ,
E ∣ D ∣ − q ≤ 2 ( 1 + p 1 − q ) E − q , E ∣ D ∣ q ≤ E q + p 1 − q Γ ( 1 + q ) E ∣ D ∣ − q . \mathbb{E}|D|^{-q}\le2\left(1+\frac{\sqrt{p}}{1-q}\right)E^{-q},\qquad\mathbb{E}|D|^q\le E^q+p^{1-q}\Gamma(1+q)\mathbb{E}|D|^{-q}. E ∣ D ∣ − q ≤ 2 ( 1 + 1 − q p ) E − q , E ∣ D ∣ q ≤ E q + p 1 − q Γ ( 1 + q ) E ∣ D ∣ − q . (iv) The support of π \pi π is R \mathbb{R} R .
(v) Let q ∈ ( 0 , 1 ) q\in(0,1) q ∈ ( 0 , 1 ) and g > 0 g>0 g > 0 , set B = g D B=gD B = g D , and let T q T_q T q be as in Lemma 5.4 and independent of $D. Then for every a ≥ 0 a\ge0 a ≥ 0 and σ ∈ { − 1 , 1 } \sigma\in\{-1,1\} σ ∈ { − 1 , 1 } ,
E ∣ B − σ a T q ∣ − q ≤ 70 1 − q E ( a + ∣ B ∣ ) − q ≤ C min { 1 , a − q } , \mathbb{E}|B-\sigma aT_q|^{-q}\le\frac{70}{1-q}\mathbb{E}(a+|B|)^{-q}\le C\min\{1,a^{-q}\}, E ∣ B − σ a T q ∣ − q ≤ 1 − q 70 E ( a + ∣ B ∣ ) − q ≤ C min { 1 , a − q } , where C = 70 ( 1 − q ) − 1 max { 1 , g − q E ∣ D ∣ − q } C=70(1-q)^{-1}\max\{1,g^{-q}\mathbb{E}|D|^{-q}\} C = 70 ( 1 − q ) − 1 max { 1 , g − q E ∣ D ∣ − q } depends only on p p p , E E E , g g g and q q q , and min { 1 , a − q } = 1 \min\{1,a^{-q}\}=1 min { 1 , a − q } = 1 for a = 0 a=0 a = 0 .
By (i), (iii) and (iv), for every g > 0 g>0 g > 0 , the law of B = g D B=gD B = g D has no atom at zero, has full support, and satisfies E ∣ B ∣ q < ∞ \mathbb{E}|B|^q<\infty E ∣ B ∣ q < ∞ for every q ∈ ( 0 , 1 ) q\in(0,1) q ∈ ( 0 , 1 ) . These are the hypotheses under which the free energy is defined in Section 6.
Proof. Throughout the proof, N N N , ( Y j ) (Y_j) ( Y j ) and ( D j ) (D_j) ( D j ) are as in Definition 5.1, and D D D has the law of
E − R , R = 1 p ∑ j = 1 N Y j D j . (61) E-R,\qquad R=\frac{1}{p}\sum_{j=1}^{N}\frac{Y_j}{D_j}. \tag*{(61)} E − R , R = p 1 j = 1 ∑ N D j Y j . (61)
Step 1. Atoms and density. On the event { N = 0 } \{N=0\} { N = 0 } , which has probability e − p e^{-p} e − p , the variable (5.4) equals E E E . Fix n ≥ 1 n\geq1 n ≥ 1 and condition on N = n N=n N = n and on D 1 , … , D n D_1,\ldots,D_n D 1 , … , D n . Then E − R = E + ∑ j ≤ n a j Y j E-R=E+\sum_{j\leq n}a_jY_j E − R = E + ∑ j ≤ n a j Y j with real nonzero coefficients a j = − 1 / ( p D j ) a_j=-1/(pD_j) a j = − 1/ ( p D j ) , and Y 1 , … , Y n Y_1,\ldots,Y_n Y 1 , … , Y n are independent Exp( 1 ) (1) ( 1 ) variables independent of the conditioning. By Lemma 4.18(iii), the conditional law of E − R E-R E − R has a density ρ n ( x ) = ρ n ( x ; D 1 , … , D n ) \rho_n(x)=\rho_n(x;D_1,\ldots,D_n) ρ n ( x ) = ρ n ( x ; D 1 , … , D n ) with
ρ n ( x ) ≤ n 2 π 1 ∣ x − E ∣ , x ≠ E . (62) \rho_n(x)\leq\sqrt{\frac{n}{2\pi}}\frac{1}{|x-E|},\qquad x\ne E. \tag*{(62)} ρ n ( x ) ≤ 2 π n ∣ x − E ∣ 1 , x = E . (62)
and the formula for ρ n \rho_n ρ n in the proof of that lemma is jointly measurable in ( x , D 1 , … , D n ) (x,D_1,\ldots,D_n) ( x , D 1 , … , D n ) . Using Tonelli’s theorem, we find that for every Borel set A ⊂ R A\subset\mathbb{R} A ⊂ R ,
π ( A ) = e − p 1 A ( E ) + ∑ n ≥ 1 P ( N = n ) E ∫ A ρ n ( x ; D 1 , … , D n ) d x = e − p 1 A ( E ) + ∫ A f ( x ) d x , \begin{aligned}
\pi(A)=e^{-p}\mathbf{1}_A(E)+\sum_{n\geq1}\mathbb{P}(N=n)\mathbb{E}\int_A\rho_n(x;D_1,\ldots,D_n)\,dx \\
&=e^{-p}\mathbf{1}_A(E)+\int_A f(x)\,dx,
\end{aligned} π ( A ) = e − p 1 A ( E ) + n ≥ 1 ∑ P ( N = n ) E ∫ A ρ n ( x ; D 1 , … , D n ) d x = e − p 1 A ( E ) + ∫ A f ( x ) d x ,
with
f ( x ) = ∑ n ≥ 1 P ( N = n ) E ρ n ( x ; D 1 , … , D n ) . f(x)=\sum_{n\geq1}\mathbb{P}(N=n)\mathbb{E}\rho_n(x;D_1,\ldots,D_n). f ( x ) = n ≥ 1 ∑ P ( N = n ) E ρ n ( x ; D 1 , … , D n ) .
This proves (i). For x ≠ E x\ne E x = E , (5.5) implies that
f ( x ) ≤ E N 2 π ∣ x − E ∣ ≤ p 2 π 1 ∣ x − E ∣ , f(x)\leq\frac{\mathbb{E}\sqrt{N}}{\sqrt{2\pi}|x-E|}\leq\sqrt{\frac{p}{2\pi}}\frac{1}{|x-E|}, f ( x ) ≤ 2 π ∣ x − E ∣ E N ≤ 2 π p ∣ x − E ∣ 1 ,
where we used Jensen’s inequality E N ≤ E N = p \mathbb{E}\sqrt{N}\leq\sqrt{\mathbb{E}N}=\sqrt{p} E N ≤ E N = p . If ∣ x ∣ ≤ E / 2 |x|\leq E/2 ∣ x ∣ ≤ E /2 , then ∣ x − E ∣ ≥ E / 2 |x-E|\geq E/2 ∣ x − E ∣ ≥ E /2 and f ( x ) ≤ 2 p / ( 2 π ) / E ≤ p / E f(x)\leq2\sqrt{p/(2\pi)}/E\leq\sqrt{p}/E f ( x ) ≤ 2 p / ( 2 π ) / E ≤ p / E . Since the interval [ − t , t ] [-t,t] [ − t , t ] does not contain the atom E E E for 0 ≤ t ≤ E / 2 0\leq t\leq E/2 0 ≤ t ≤ E /2 , we have P ( ∣ D ∣ ≤ t ) = ∫ − t t f ≤ 2 p t / E \mathbb{P}(|D|\leq t)=\int_{-t}^{t}f\leq2\sqrt{p}\,t/E P ( ∣ D ∣ ≤ t ) = ∫ − t t f ≤ 2 p t / E . This proves (ii).
Step 2. Moments. By (ii),
E ∣ D ∣ − q ≤ ( E 2 ) − q + ∫ − E / 2 E / 2 ∣ x ∣ − q f ( x ) d x ≤ ( 2 E ) q + p E 2 ( E / 2 ) 1 − q 1 − q ≤ 2 E − q + 2 p 1 − q E − q . \mathbb{E}|D|^{-q}\leq\left(\frac{E}{2}\right)^{-q}+\int_{-E/2}^{E/2}|x|^{-q}f(x)\,dx\leq\left(\frac{2}{E}\right)^q+\frac{\sqrt{p}}{E}\frac{2(E/2)^{1-q}}{1-q}\leq2E^{-q}+\frac{2\sqrt{p}}{1-q}E^{-q}. E ∣ D ∣ − q ≤ ( 2 E ) − q + ∫ − E /2 E /2 ∣ x ∣ − q f ( x ) d x ≤ ( E 2 ) q + E p 1 − q 2 ( E /2 ) 1 − q ≤ 2 E − q + 1 − q 2 p E − q .
In the first inequality, we bounded ∣ D ∣ − q |D|^{-q} ∣ D ∣ − q by ( E / 2 ) − q (E/2)^{-q} ( E /2 ) − q on the event { ∣ D ∣ > E / 2 } \{|D| > E/2\} { ∣ D ∣ > E /2 } . For the positive moment, we use (61) and the inequality ∣ a + b ∣ q ≤ ∣ a ∣ q + ∣ b ∣ q |a+b|^q \le|a|^q + |b|^q ∣ a + b ∣ q ≤ ∣ a ∣ q + ∣ b ∣ q , valid for q ∈ ( 0 , 1 ] q \in(0,1] q ∈ ( 0 , 1 ] , to obtain
∣ D ∣ q ≤ E q + p − q ∑ j = 1 N Y j q ∣ D j ∣ − q . |D|^q \le E^q + p^{-q}\sum_{j=1}^{N}Y_j^q|D_j|^{-q}. ∣ D ∣ q ≤ E q + p − q j = 1 ∑ N Y j q ∣ D j ∣ − q .
Taking expectations and using the independence of N N N , ( Y j ) (Y_j) ( Y j ) and ( D j ) (D_j) ( D j ) , we find
E ∣ D ∣ q ≤ E q + p − q E N E Y 1 q E ∣ D ∣ − q = E q + p 1 − q Γ ( 1 + q ) E ∣ D ∣ − q , \mathbb{E}|D|^q \le E^q+p^{-q}\mathbb{E}N\,\mathbb{E}Y_1^q\,\mathbb{E}|D|^{-q}=E^q+p^{1-q}\Gamma(1+q)\,\mathbb{E}|D|^{-q}, E ∣ D ∣ q ≤ E q + p − q E N E Y 1 q E ∣ D ∣ − q = E q + p 1 − q Γ ( 1 + q ) E ∣ D ∣ − q ,
which proves (iii).
Step 3. Full support. It suffices to show that every open interval I I I has π ( I ) > 0 \pi(I)>0 π ( I ) > 0 . By (i), the event { N = 1 , D 1 = E } \{N=1,D_1=E\} { N = 1 , D 1 = E } has probability p e − p π ( { E } ) = p e − 2 p > 0 pe^{-p}\pi(\{E\})=pe^{-2p}>0 p e − p π ({ E }) = p e − 2 p > 0 . On this event, the variable (61) equals E − Y 1 / ( p E ) E-Y_1/(pE) E − Y 1 / ( pE ) , which has the density p E e − p E ( E − x ) pEe^{-pE(E-x)} pE e − pE ( E − x ) on ( − ∞ , E ) (-\infty,E) ( − ∞ , E ) . This density is strictly positive on ( − ∞ , E ) (-\infty,E) ( − ∞ , E ) , and we conclude that π ( I ) > 0 \pi(I)>0 π ( I ) > 0 for every open interval I I I that meets ( − ∞ , E ) (-\infty,E) ( − ∞ , E ) . Since E > 0 E>0 E > 0 , this applies to I = ( − ∞ , 0 ) I=(-\infty,0) I = ( − ∞ , 0 ) , and the event { N = 1 , D 1 < 0 } \{N=1,D_1<0\} { N = 1 , D 1 < 0 } has probability p e − p π ( ( − ∞ , 0 ) ) > 0 pe^{-p}\pi((-\infty,0))>0 p e − p π (( − ∞ , 0 )) > 0 . On this event, the variable (61) equals E + Y 1 / ( p ∣ D 1 ∣ ) E+Y_1/(p|D_1|) E + Y 1 / ( p ∣ D 1 ∣ ) , and conditionally on D 1 D_1 D 1 it has the strictly positive density p ∣ D 1 ∣ e − p ∣ D 1 ∣ ( x − E ) p|D_1|e^{-p|D_1|(x-E)} p ∣ D 1 ∣ e − p ∣ D 1 ∣ ( x − E ) on ( E , ∞ ) (E,\infty) ( E , ∞ ) . As before, we conclude that π ( I ) > 0 \pi(I)>0 π ( I ) > 0 for every open interval I I I that meets ( E , ∞ ) (E,\infty) ( E , ∞ ) . Every open interval meets ( − ∞ , E ) (-\infty,E) ( − ∞ , E ) or ( E , ∞ ) (E,\infty) ( E , ∞ ) , which proves (iv).
Step 4. Fractional cancellation. Fix b ∈ R ∖ { 0 } b\in\mathbb{R}\setminus\{0\} b ∈ R ∖ { 0 } . For a > 0 a>0 a > 0 , we have ∣ b − σ a T q ∣ = a ∣ T q − σ b / a ∣ |b-\sigma aT_q|=a|T_q-\sigma b/a| ∣ b − σ a T q ∣ = a ∣ T q − σ b / a ∣ since ∣ σ ∣ = 1 |\sigma|=1 ∣ σ ∣ = 1 , and Lemma 5.4 implies that
E ∣ b − σ a T q ∣ − q = a − q E ∣ T q − σ b a ∣ − q ≤ 70 1 − q a − q ( 1 + ∣ b ∣ a ) − q = 70 1 − q ( a + ∣ b ∣ ) − q . \mathbb{E}|b-\sigma aT_q|^{-q}=a^{-q}\mathbb{E}\left|T_q-\frac{\sigma b}{a}\right|^{-q}\le\frac{70}{1-q}a^{-q}\left(1+\frac{|b|}{a}\right)^{-q}=\frac{70}{1-q}(a+|b|)^{-q}. E ∣ b − σ a T q ∣ − q = a − q E T q − a σ b − q ≤ 1 − q 70 a − q ( 1 + a ∣ b ∣ ) − q = 1 − q 70 ( a + ∣ b ∣ ) − q .
For a = 0 a=0 a = 0 , the left side equals ∣ b ∣ − q = ( a + ∣ b ∣ ) − q |b|^{-q}=(a+|b|)^{-q} ∣ b ∣ − q = ( a + ∣ b ∣ ) − q . Since T q T_q T q is independent of B B B and B ≠ 0 B\ne0 B = 0 almost surely, we may condition on B B B and obtain the first inequality in (v). For the second, we have E ( a + ∣ B ∣ ) − q ≤ min { a − q , E ∣ B ∣ − q } \mathbb{E}(a+|B|)^{-q}\le\min\{a^{-q},\mathbb{E}|B|^{-q}\} E ( a + ∣ B ∣ ) − q ≤ min { a − q , E ∣ B ∣ − q } , and E ∣ B ∣ − q = g − q E ∣ D ∣ − q \mathbb{E}|B|^{-q}=g^{-q}\mathbb{E}|D|^{-q} E ∣ B ∣ − q = g − q E ∣ D ∣ − q is finite by (iii). Finally, min { a − q , M } ≤ max { 1 , M } min { 1 , a − q } \min\{a^{-q},M\}\le\max\{1,M\}\min\{1,a^{-q}\} min { a − q , M } ≤ max { 1 , M } min { 1 , a − q } for M ≥ 0 M\ge0 M ≥ 0 , as one sees by treating a ≥ 1 a\ge1 a ≥ 1 and a < 1 a<1 a < 1 separately. This proves (v).
The next lemma bounds the probability that the denominator is small on the scale p − 1 / 2 p^{-1/2} p − 1/2 , uniformly in the energy and in the real solution.
Lemma 5.6. Let E > 0 E>0 E > 0 , let π \pi π be a real solution at energy E E E , let D ∼ π D\sim\pi D ∼ π , and for a > 0 a>0 a > 0 set q a = P ( ∣ D ∣ ≤ a / p ) q_a=\mathbb{P}(|D|\le a/\sqrt{p}) q a = P ( ∣ D ∣ ≤ a / p ) . Then
q a ≤ e − p q a + 2 a 2 . q_a\le e^{-pq_a}+2a^2. q a ≤ e − p q a + 2 a 2 . In particular, let ε ∈ ( 0 , 1 ) \varepsilon\in(0,1) ε ∈ ( 0 , 1 ) , and set a ( ε ) = ε / 2 a(\varepsilon)=\sqrt{\varepsilon}/2 a ( ε ) = ε /2 and p ( ε ) = ε − 1 log ( 2 / ε ) p(\varepsilon)=\varepsilon^{-1}\log(2/\varepsilon) p ( ε ) = ε − 1 log ( 2/ ε ) . Then q a < ε q_a<\varepsilon q a < ε for every p > max { 1 , p ( ε ) } p>\max\{1,p(\varepsilon)\} p > max { 1 , p ( ε )} , every E > 0 E>0 E > 0 , every real solution at energy E E E , and every a ∈ ( 0 , a ( ε ) ] a\in(0,a(\varepsilon)] a ∈ ( 0 , a ( ε )] .
Proof. Let N N N , ( Y j ) (Y_j) ( Y j ) and ( D j ) (D_j) ( D j ) be as in Definition 5.1, and recall that D D D has the law of (61) . We call an index j ≤ N j \le N j ≤ N small if ∣ D j ∣ ≤ a / p |D_j| \le a/\sqrt{p} ∣ D j ∣ ≤ a / p , and we let K K K be the number of small indices. Since the D j D_j D j are independent of N N N , the variable K K K has, conditionally on N N N , the binomial law with parameters N N N and q a q_a q a . Its generating function is E s K = E ( 1 − q a + q a s ) N = e − p q a ( 1 − s ) \mathbb{E}s^K = \mathbb{E}(1-q_a+q_as)^N = e^{-pq_a(1-s)} E s K = E ( 1 − q a + q a s ) N = e − p q a ( 1 − s ) , s ∈ [ 0 , 1 ] s \in[0,1] s ∈ [ 0 , 1 ] , which is that of Pois ( p q a ) \operatorname{Pois}(pq_a) Pois ( p q a ) ; in particular P ( K = 0 ) = e − p q a \mathbb{P}(K=0)=e^{-pq_a} P ( K = 0 ) = e − p q a .
On the event { K ≥ 1 } \{K \ge1\} { K ≥ 1 } , let j ∗ j_* j ∗ be the smallest small index. Then { K ≥ 1 } \{K \ge1\} { K ≥ 1 } and j ∗ j_* j ∗ are measurable with respect to σ ( N , ( D j ) j ≥ 1 ) \sigma(N,(D_j)_{j\ge1}) σ ( N , ( D j ) j ≥ 1 ) . Let G \mathcal{G} G be the σ \sigma σ -algebra generated by N N N , ( D j ) j ≥ 1 (D_j)_{j\ge1} ( D j ) j ≥ 1 and the variables Y j 1 { j ≠ j ∗ } Y_j1_{\{j\ne j_*\}} Y j 1 { j = j ∗ } , j ≥ 1 j \ge1 j ≥ 1 . For each k ≥ 1 k \ge1 k ≥ 1 , the variable Y k Y_k Y k is independent of N N N , ( D j ) j ≥ 1 (D_j)_{j\ge1} ( D j ) j ≥ 1 and ( Y j ) j ≠ k (Y_j)_{j\ne k} ( Y j ) j = k , and the event { j ∗ = k } \{j_*=k\} { j ∗ = k } is determined by N N N and ( D j ) (D_j) ( D j ) . Then on { K ≥ 1 } \{K \ge1\} { K ≥ 1 } , conditionally on G \mathcal{G} G , the variable Y j ∗ Y_{j_*} Y j ∗ has the Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) law. Further,
E − R = c − Y j ∗ p D j ∗ , c = E − 1 p ∑ j ≤ N , j ≠ j ∗ Y j D j . E-R=c-\frac{Y_{j_*}}{pD_{j_*}}, \qquad c=E-\frac{1}{p}\sum_{\substack{j\le N,\,j\ne j_*}}\frac{Y_j}{D_j}. E − R = c − p D j ∗ Y j ∗ , c = E − p 1 j ≤ N , j = j ∗ ∑ D j Y j .
where c c c is G \mathcal{G} G -measurable. For a real constant c c c and κ ≠ 0 \kappa\ne0 κ = 0 , the variable c + κ Y j ∗ c+\kappa Y_{j_*} c + κ Y j ∗ has a density bounded by ∣ κ ∣ − 1 |\kappa|^{-1} ∣ κ ∣ − 1 . Here ∣ κ ∣ − 1 = p ∣ D j ∗ ∣ ≤ a p |\kappa|^{-1}=p|D_{j_*}|\le a\sqrt{p} ∣ κ ∣ − 1 = p ∣ D j ∗ ∣ ≤ a p , and on { K ≥ 1 } \{K\ge1\} { K ≥ 1 } we have
P ( ∣ E − R ∣ ≤ a p ∣ G ) ≤ 2 a p ⋅ a p = 2 a 2 . \mathbb{P}\left(|E-R|\le\frac{a}{\sqrt{p}}\mathrel{\big|}\mathcal{G}\right)\le\frac{2a}{\sqrt{p}}\cdot a\sqrt{p}=2a^2. P ( ∣ E − R ∣ ≤ p a G ) ≤ p 2 a ⋅ a p = 2 a 2 .
Since E − R E-R E − R has the law of D D D , we conclude that
q a ≤ P ( K = 0 ) + E [ P ( ∣ E − R ∣ ≤ a p ∣ G ) 1 { K ≥ 1 } ] ≤ e − p q a + 2 a 2 . q_a\le\mathbb{P}(K=0)+\mathbb{E}\left[\mathbb{P}\left(|E-R|\le\frac{a}{\sqrt{p}}\mathrel{\big|}\mathcal{G}\right)1_{\{K\ge1\}}\right]\le e^{-pq_a}+2a^2. q a ≤ P ( K = 0 ) + E [ P ( ∣ E − R ∣ ≤ p a G ) 1 { K ≥ 1 } ] ≤ e − p q a + 2 a 2 .
For the second assertion, let a ≤ a ( ε ) a\le a(\varepsilon) a ≤ a ( ε ) and p > p ( ε ) p>p(\varepsilon) p > p ( ε ) ; then 2 a 2 ≤ ε / 2 2a^2\le\varepsilon/2 2 a 2 ≤ ε /2 and e − p ε < ε / 2 e^{-p\varepsilon}<\varepsilon/2 e − pε < ε /2 . If q a ≥ ε q_a\ge\varepsilon q a ≥ ε , then the first assertion gives q a ≤ e − p ε + ε / 2 < ε q_a\le e^{-p\varepsilon}+\varepsilon/2<\varepsilon q a ≤ e − pε + ε /2 < ε , which is impossible. This proves the second assertion. □ \square □
Regular energies and boundary values
In this section, we define the regular energies and study the boundary values of resolvent entries at a regular energy. The main result is Proposition 5.11.
Notation
We use the realization of the marked tree on the configuration space Ω \Omega Ω of Definition 4.1. Thus ω = ( ( N v ) v ∈ U , ( ζ v ) v ∈ U , ( w v ) v ∈ U ∗ ) \omega=((N_v)_{v\in\mathbb{U}},(\zeta_v)_{v\in\mathbb{U}},(w_v)_{v\in\mathbb{U}_*}) ω = (( N v ) v ∈ U , ( ζ v ) v ∈ U , ( w v ) v ∈ U ∗ ) has law P \mathbb{P} P , the children of v ∈ V = V ( ω ) v\in V=V(\omega) v ∈ V = V ( ω ) are v 1 , … , v N v v1,\ldots,vN_v v 1 , … , v N v , the edge { v − , v } \{v^-,v\} { v − , v } carries the mark w v w_v w v , and Y v = ∣ w v ∣ 2 = p ∣ h v − v ∣ 2 Y_v=|w_v|^2=p|h_{v^-v}|^2 Y v = ∣ w v ∣ 2 = p ∣ h v − v ∣ 2 . We write θ u \theta_u θ u for the shifts, U ≤ m = ⋃ n ≤ m U n \mathbb{U}_{\le m}=\bigcup_{n\le m}\mathbb{U}_n U ≤ m = ⋃ n ≤ m U n , u U = { u v : v ∈ U } u\mathbb{U}=\{uv:v\in\mathbb{U}\} u U = { uv : v ∈ U } , and V n = V ∩ U n V_n=V\cap\mathbb{U}_n V n = V ∩ U n for the set of vertices of generation n n n . For u ∈ U u\in\mathbb{U} u ∈ U , we set S u = { θ u ω ∈ S } S_u=\{\theta_u\omega\in S\} S u = { θ u ω ∈ S } , and for u ∈ V u\in V u ∈ V we set ς u = 1 S u \varsigma_u=1_{S_u} ς u = 1 S u , so that ς u = 1 \varsigma_u=1 ς u = 1 if and only if the forward subtree T u T_u T u is infinite. We recall from Lemma 4.2(i) that for every n ≥ 0 n \ge0 n ≥ 0 , the configurations θ u ω \theta_u\omega θ u ω , u ∈ U n u \in\mathbb{U}_n u ∈ U n , are independent with law P \mathbb{P} P and independent of the σ \sigma σ -field
F n = σ ( N u , ζ u : ∣ u ∣ < n ; w u : 1 ≤ ∣ u ∣ ≤ n ) , \mathcal{F}_n = \sigma\left(N_u,\zeta_u : |u| < n; \ w_u : 1 \le|u| \le n\right), F n = σ ( N u , ζ u : ∣ u ∣ < n ; w u : 1 ≤ ∣ u ∣ ≤ n ) ,
which determines V n \mathcal{V}_n V n . We also recall from the discussion after Definition 4.1 that, P \mathbb{P} P -almost surely, Γ u ( z ) = G o o ( z ) ∘ θ u \Gamma_u(z) = G_{oo}(z) \circ\theta_u Γ u ( z ) = G oo ( z ) ∘ θ u for every u ∈ V ∖ { o } u \in\mathcal{V} \setminus\{o\} u ∈ V ∖ { o } and z ∈ C ∖ R z \in\mathbb{C} \setminus\mathbb{R} z ∈ C ∖ R .
A removed set is a pair Δ = ( Δ V , Δ E ) \Delta= (\Delta_{\mathsf{V}},\Delta_{\mathsf{E}}) Δ = ( Δ V , Δ E ) of finite sets Δ V ⊂ U \Delta_{\mathsf{V}} \subset\mathbb{U} Δ V ⊂ U and Δ E ⊂ U ∗ \Delta_{\mathsf{E}} \subset\mathbb{U}_* Δ E ⊂ U ∗ , where a word u ∈ Δ E u \in\Delta_{\mathsf{E}} u ∈ Δ E stands for the edge { u − , u } \{u^-,u\} { u − , u } . It determines the finite removal of Definition 3.3 in which the vertices of Δ V ∩ V \Delta_{\mathsf{V}} \cap\mathcal{V} Δ V ∩ V and the edges { u − , u } \{u^-,u\} { u − , u } with u ∈ Δ E ∩ V u \in\Delta_{\mathsf{E}} \cap\mathcal{V} u ∈ Δ E ∩ V are removed, and every finite removal arises in this way. Its depth d ( Δ ) d(\Delta) d ( Δ ) is the maximal length of a word in Δ V ∪ Δ E \Delta_{\mathsf{V}} \cup\Delta_{\mathsf{E}} Δ V ∪ Δ E , and d ( Δ ) = 0 d(\Delta)=0 d ( Δ ) = 0 if Δ \Delta Δ is empty. We write T Δ \mathcal{T}^{\Delta} T Δ for the reduced forest, H Δ \mathcal{H}^{\Delta} H Δ for the reduced operator, and G x y Δ ( z ) G_{xy}^{\Delta}(z) G x y Δ ( z ) , G x y C ( z ) G_{xy}^{\mathsf{C}}(z) G x y C ( z ) , C x → w C_{x\to w} C x → w and Γ x → w C ( z ) \Gamma_{x\to w}^{\mathsf{C}}(z) Γ x → w C ( z ) for the resolvent entries, components, branches and cavity resolvents of Definition 3.15 and Lemma 3.5. These objects are defined for all removed sets simultaneously on the almost sure event Ω s a \Omega_{\mathrm{sa}} Ω sa of Theorem 3.10. If Δ \Delta Δ is empty, then C = V \mathsf{C}=\mathcal{V} C = V , and we recover G x y G_{xy} G x y , Γ x → w \Gamma_{x\to w} Γ x → w , the forward cavities Γ v = Γ v → v − \Gamma_v=\Gamma_{v\to v^-} Γ v = Γ v → v − for v ≠ o v\ne o v = o , and Γ o = G o o \Gamma_o=G_{oo} Γ o = G oo . Since there are countably many removed sets, a statement that holds almost surely for each fixed removed set holds almost surely for all removed sets simultaneously; this covers finite removals selected at random, as in Definition 3.3.
Let K \mathcal{K} K be the countable set of quadruples κ = ( Δ , x , y , k ) \kappa=(\Delta,x,y,k) κ = ( Δ , x , y , k ) , where Δ \Delta Δ is a removed set, x , y ∈ U x,y\in\mathbb{U} x , y ∈ U and k ∈ { 0 , 1 , 2 , 3 } k\in\{0,1,2,3\} k ∈ { 0 , 1 , 2 , 3 } . For κ ∈ K \kappa\in\mathcal{K} κ ∈ K and z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R , we set ϕ κ = δ x + i k δ y \phi_\kappa=\delta_x+i^k\delta_y ϕ κ = δ x + i k δ y and
F κ ( z ) = ⟨ ϕ κ , ( H Δ − z ) − 1 ϕ κ ⟩ if x , y ∈ V ∖ Δ V , F_\kappa(z)=\left\langle\phi_\kappa,(\mathcal{H}^{\Delta}-z)^{-1}\phi_\kappa\right\rangle\quad\text{if } x,y\in\mathcal{V}\setminus\Delta_{\mathsf{V}}, F κ ( z ) = ⟨ ϕ κ , ( H Δ − z ) − 1 ϕ κ ⟩ if x , y ∈ V ∖ Δ V ,
and F κ ( z ) = 0 F_\kappa(z)=0 F κ ( z ) = 0 otherwise. By the polarization identity, we have
G x y Δ ( z ) = 1 4 ∑ k = 0 3 i − k F ( Δ , x , y , k ) ( z ) , (63) G_{xy}^{\Delta}(z)=\frac{1}{4}\sum_{k=0}^{3}i^{-k}F_{(\Delta,x,y,k)}(z), \tag*{(63)} G x y Δ ( z ) = 4 1 k = 0 ∑ 3 i − k F ( Δ , x , y , k ) ( z ) , (63)
so every resolvent entry of every reduced operator is a fixed linear combination of four functions F κ F_\kappa F κ . For instance, Γ v = G v v Δ \Gamma_v=G_{vv}^{\Delta} Γ v = G v v Δ for the removed set with Δ V = ∅ \Delta_{\mathsf{V}}=\varnothing Δ V = ∅ and Δ E = { v } \Delta_{\mathsf{E}}=\{v\} Δ E = { v } , and Γ x → w C = G x x Δ ′ \Gamma_{x\to w}^{\mathsf{C}}=G_{xx}^{\Delta'} Γ x → w C = G xx Δ ′ , where Δ ′ \Delta' Δ ′ is obtained from the removed set defining C \mathsf{C} C by adding the edge { x , w } \{x,w\} { x , w } .
If x , y ∈ V ∖ Δ V x,y\in\mathcal{V}\setminus\Delta_{\mathsf{V}} x , y ∈ V ∖ Δ V , then Lemma 3.14(ii) gives F κ ( z ) = ∫ μ κ ( d λ ) / ( λ − z ) F_\kappa(z)=\int\mu_\kappa(\mathrm{d}\lambda)/(\lambda-z) F κ ( z ) = ∫ μ κ ( d λ ) / ( λ − z ) , where μ κ \mu_\kappa μ κ is the spectral measure of ϕ κ \phi_\kappa ϕ κ for H Δ \mathcal{H}^{\Delta} H Δ , a positive measure of total mass ∥ ϕ κ ∥ 2 ≤ 4 \|\phi_\kappa\|^2\le4 ∥ ϕ κ ∥ 2 ≤ 4 . Thus F κ F_\kappa F κ is the Borel transform of a finite measure. For such a function, the limit lim η ↓ 0 F κ ( E + i η ) \lim_{\eta\downarrow0}F_\kappa(E+i\eta) lim η ↓ 0 F κ ( E + i η ) exists and is finite for Lebesgue-almost every E ∈ R E\in\mathbb{R} E ∈ R ; see, for example, [123] or [13].
Regular energies
Definition 5.7. The set E p \mathcal{E}_{p} E p of regular energies is the set of E ∈ R ∖ { 0 } E\in\mathbb{R}\setminus\{0\} E ∈ R ∖ { 0 } such that, P \mathbb{P} P -almost surely, the following two conditions hold.
(a) For every κ ∈ K \kappa\in\mathcal{K} κ ∈ K , the limit F κ ( E + i 0 ) = lim η ↓ 0 F κ ( E + i η ) F_{\kappa}(E+i0)=\lim_{\eta\downarrow0}F_{\kappa}(E+i\eta) F κ ( E + i 0 ) = lim η ↓ 0 F κ ( E + i η ) exists in C \mathbb{C} C .
(b) For every finite, nonempty, connected set Λ ⊂ V \Lambda\subset V Λ ⊂ V , the energy E E E is not an eigenvalue of the matrix H Λ = ( ⟨ δ x , H δ y ⟩ ) x , y ∈ Λ \mathcal{H}_{\Lambda}=(\langle\delta_x,\mathcal{H}\delta_y\rangle)_{x,y\in\Lambda} H Λ = (⟨ δ x , H δ y ⟩ ) x , y ∈ Λ .
For E ∈ E p E\in\mathcal{E}_p E ∈ E p , we write G x y Δ ( E + i 0 ) G^{\Delta}_{xy}(E+i0) G x y Δ ( E + i 0 ) , Γ x → w C ( E + i 0 ) \Gamma^{C}_{x\to w}(E+i0) Γ x → w C ( E + i 0 ) , Γ v ( E + i 0 ) \Gamma_v(E+i0) Γ v ( E + i 0 ) , and so on, for the corresponding limits, which exist and are finite almost surely by (a) and (5.6).
By Lemma 5.8 below, E p \mathcal{E}_p E p is a Borel set whose complement is Lebesgue-null. Since P ∞ ≤ ϱ p − 1 P \mathbb{P}^{\infty}\leq\varrho_p^{-1}\mathbb{P} P ∞ ≤ ϱ p − 1 P , every P \mathbb{P} P -almost sure statement is also a P ∞ \mathbb{P}^{\infty} P ∞ -almost sure statement. Condition (b) states that every finite subtree of T \mathcal{T} T , with its edge weights, has E E E outside its spectrum. Every finite component of a reduced forest is such a subtree, and so is every finite forward subtree T v \mathcal{T}_v T v together with all its sub-bushes. For such a finite component C C C and x , y ∈ C x,y\in C x , y ∈ C , the function G x y C ( z ) G^C_{xy}(z) G x y C ( z ) is a rational function of z z z with poles in the spectrum of H C \mathcal{H}_C H C . At E ∈ E p E\in\mathcal{E}_p E ∈ E p it is analytic near E E E , and its boundary value G x y C ( E + i 0 ) = G x y C ( E ) G^C_{xy}(E+i0)=G^C_{xy}(E) G x y C ( E + i 0 ) = G x y C ( E ) is the corresponding entry of ( H C − E ) − 1 (\mathcal{H}_C-E)^{-1} ( H C − E ) − 1 . In particular, the diagonal entries G x x C ( E + i 0 ) G^C_{xx}(E+i0) G xx C ( E + i 0 ) are real.
Lemma 5.8. The set E p \mathcal{E}_p E p is a Borel set, its complement in R \mathbb{R} R has Lebesgue measure zero, and E p = − E p \mathcal{E}_p=-\mathcal{E}_p E p = − E p .
Proof. Step 1. Joint measurability. Fix κ ∈ K \kappa\in\mathcal{K} κ ∈ K . By Theorem 3.10(iii) and the discussion after Definition 3.15, on Ω s a \Omega_{\mathrm{sa}} Ω sa the function F κ ( z ) F_{\kappa}(z) F κ ( z ) is the limit, uniformly for z z z in compact subsets of C ∖ R \mathbb{C}\setminus\mathbb{R} C ∖ R , of the corresponding quantities for the restrictions of H Δ \mathcal{H}^{\Delta} H Δ to finite balls. These are rational functions of z z z whose coefficients are measurable functions of ω \omega ω . Then ( ω , E , η ) ↦ F κ ( E + i η ) (\omega,E,\eta)\mapsto F_{\kappa}(E+i\eta) ( ω , E , η ) ↦ F κ ( E + i η ) is jointly measurable on Ω × R × ( 0 , ∞ ) \Omega\times\mathbb{R}\times(0,\infty) Ω × R × ( 0 , ∞ ) .
Since η ↦ F κ ( E + i η ) \eta\mapsto F_{\kappa}(E+i\eta) η ↦ F κ ( E + i η ) is continuous, the limit as η ↓ 0 \eta\downarrow0 η ↓ 0 exists in C \mathbb{C} C if and only if for every k ∈ N k\in\mathbb{N} k ∈ N there exists n ∈ N n\in\mathbb{N} n ∈ N such that ∣ F κ ( E + i η ) − F κ ( E + i η ′ ) ∣ ≤ 1 / k \lvert F_{\kappa}(E+i\eta)-F_{\kappa}(E+i\eta')\rvert\leq1/k ∣ F κ ( E + i η ) − F κ ( E + i η ′ )∣ ≤ 1/ k for all rational η , η ′ ∈ ( 0 , 1 / n ) \eta,\eta'\in(0,1/n) η , η ′ ∈ ( 0 , 1/ n ) . By this criterion, the set B a \mathcal{B}_a B a of pairs ( ω , E ) (\omega,E) ( ω , E ) for which (a) fails is measurable. The set B b \mathcal{B}_b B b of pairs for which (b) fails is the union over finite sets Λ ⊂ U \Lambda\subset\mathcal{U} Λ ⊂ U of the sets { Λ ⊂ V , Λ connected , det ( H Λ − E ) = 0 } \{\Lambda\subset V,\ \Lambda\text{ connected},\ \det(\mathcal{H}_{\Lambda}-E)=0\} { Λ ⊂ V , Λ connected , det ( H Λ − E ) = 0 } , and it is measurable as well. Let B = B a ∪ B b ∪ ( ( Ω ∖ Ω s a ) × R ) \mathcal{B}=\mathcal{B}_a\cup\mathcal{B}_b\cup((\Omega\setminus\Omega_{\mathrm{sa}})\times\mathbb{R}) B = B a ∪ B b ∪ (( Ω ∖ Ω sa ) × R ) .
Step 2. Full measure. Fix ω ∈ Ω s a \omega\in\Omega_{\mathrm{sa}} ω ∈ Ω sa . For each κ \kappa κ , the function F κ F_{\kappa} F κ is either zero or the Borel transform of a finite measure; in both cases its boundary value exists in C \mathbb{C} C for Lebesgue-almost every E E E . Since K \mathcal{K} K is countable, the section { E : ( ω , E ) ∈ B a } \{E:(\omega,E)\in\mathcal{B}_a\} { E : ( ω , E ) ∈ B a } is Lebesgue-null. The section { E : ( ω , E ) ∈ B b } \{E:(\omega,E)\in\mathcal{B}_b\} { E : ( ω , E ) ∈ B b } is contained in the union of the spectra of the countably many finite matrices H Λ \mathcal{H}_{\Lambda} H Λ , which is a countable set. By Tonelli’s theorem,
∫ R P { ω : ( ω , E ) ∈ B } d E = E ∣ { E : ( ω , E ) ∈ B } ∣ = 0 , \int_{\mathbb{R}}\mathbb{P}\{\omega:(\omega,E)\in\mathcal{B}\}\,\mathrm{d}E
=\mathbb{E}\lvert\{E:(\omega,E)\in\mathcal{B}\}\rvert=0, ∫ R P { ω : ( ω , E ) ∈ B } d E = E ∣{ E : ( ω , E ) ∈ B }∣ = 0 ,
where ∣ ⋅ ∣ \lvert\cdot\rvert ∣ ⋅ ∣ denotes Lebesgue measure. The function E ↦ P { ω : ( ω , E ) ∈ B } E\mapsto\mathbb{P}\{\omega:(\omega,E)\in\mathcal{B}\} E ↦ P { ω : ( ω , E ) ∈ B } is Borel by Tonelli’s theorem, and E p = { E ≠ 0 : P { ω : ( ω , E ) ∈ B } = 0 } \mathcal{E}_p=\{E\neq0:\mathbb{P}\{\omega:(\omega,E)\in\mathcal{B}\}=0\} E p = { E = 0 : P { ω : ( ω , E ) ∈ B } = 0 } . We conclude that E p \mathcal{E}_p E p is a Borel set of full measure.
Step 3. Symmetry. Let U U U be the diagonal unitary U δ v = ( − 1 ) ∣ v ∣ δ v U\delta_v=(-1)^{|v|}\delta_v U δ v = ( − 1 ) ∣ v ∣ δ v on ℓ 2 ( V ∖ Δ V ) \ell^2(V\setminus\Delta V) ℓ 2 ( V ∖ Δ V ) . Since the edges of T Δ T^\Delta T Δ join consecutive generations, U H Δ U = − H Δ UH^\Delta U=-H^\Delta U H Δ U = − H Δ on finitely supported vectors, and hence on the closure. For κ = ( Δ , x , y , k ) \kappa=(\Delta,x,y,k) κ = ( Δ , x , y , k ) with x , y ∈ V ∖ Δ V x,y\in V\setminus\Delta V x , y ∈ V ∖ Δ V , we have U ϕ κ = ( − 1 ) ∣ x ∣ ϕ κ ′ U\phi_\kappa=(-1)^{|x|}\phi_{\kappa'} U ϕ κ = ( − 1 ) ∣ x ∣ ϕ κ ′ with κ ′ = ( Δ , x , y , k ′ ) \kappa'=(\Delta,x,y,k') κ ′ = ( Δ , x , y , k ′ ) , where k ′ = k k'=k k ′ = k if ∣ x ∣ + ∣ y ∣ |x|+|y| ∣ x ∣ + ∣ y ∣ is even and k ′ ≡ k + 2 ( m o d 4 ) k'\equiv k+2\pmod4 k ′ ≡ k + 2 ( mod 4 ) otherwise. We have, for z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R ,
F κ ( z ) = ⟨ U ϕ κ , ( U H Δ U − z ) − 1 U ϕ κ ⟩ = − ⟨ ϕ κ ′ , ( H Δ + z ) − 1 ϕ κ ′ ⟩ . F_\kappa(z)=\langle U\phi_\kappa,(UH^\Delta U-z)^{-1}U\phi_\kappa\rangle=-\langle\phi_{\kappa'},(H^\Delta+z)^{-1}\phi_{\kappa'}\rangle. F κ ( z ) = ⟨ U ϕ κ , ( U H Δ U − z ) − 1 U ϕ κ ⟩ = − ⟨ ϕ κ ′ , ( H Δ + z ) − 1 ϕ κ ′ ⟩ .
Taking z = − E + i η z=-E+i\eta z = − E + i η and using Lemma 3.14(i), we obtain
F κ ( − E + i η ) = − F κ ′ ( E + i η ) ‾ . F_\kappa(-E+i\eta)=-\overline{F_{\kappa'}(E+i\eta)}. F κ ( − E + i η ) = − F κ ′ ( E + i η ) .
This shows that condition (a) holds at − E -E − E if and only if it holds at E E E . Similarly, U H Λ U = − H Λ UH_\Lambda U=-H_\Lambda U H Λ U = − H Λ for every finite connected Λ \Lambda Λ , which shows that the spectrum of H Λ H_\Lambda H Λ is symmetric and that condition (b) holds at − E -E − E if and only if it holds at E E E . Hence E p = − E p \mathcal{E}_p=-\mathcal{E}_p E p = − E p . □
Boundary values of the Schur recursion
We next pass the Schur formulas of Proposition 3.16 to the real axis. For a component C C C of T Δ T^\Delta T Δ and x ∈ C x\in C x ∈ C , we write y ∼ C x y\sim_C x y ∼ C x if y y y is a neighbor of x x x in C C C .
Lemma 5.9. Let E ∈ E p E\in\mathcal{E}_p E ∈ E p . Almost surely, all edge weights of T T T are nonzero, and the following holds for every removed set Δ \Delta Δ , every component C C C of T Δ T^\Delta T Δ , every x ∈ C x\in C x ∈ C , and every neighbor w w w of x x x in C C C . The numbers
D x C = E + ∑ y ∼ C x ∣ h x y ∣ 2 Γ y → x C ( E + i 0 ) , D x → w C = E + ∑ y ∼ C x y ≠ w ∣ h x y ∣ 2 Γ y → x C ( E + i 0 ) D_x^C=E+\sum_{y\sim_C x}|h_{xy}|^2\Gamma^C_{y\to x}(E+i0),\qquad D_{x\to w}^C=E+\sum_{\substack{y\sim_C x\\y\ne w}}|h_{xy}|^2\Gamma^C_{y\to x}(E+i0) D x C = E + y ∼ C x ∑ ∣ h x y ∣ 2 Γ y → x C ( E + i 0 ) , D x → w C = E + y ∼ C x y = w ∑ ∣ h x y ∣ 2 Γ y → x C ( E + i 0 ) are finite and nonzero, G x x C ( E + i 0 ) = − 1 / D x C G_{xx}^C(E+i0)=-1/D_x^C G xx C ( E + i 0 ) = − 1/ D x C , and Γ x → w C ( E + i 0 ) = − 1 / D x → w C \Gamma^C_{x\to w}(E+i0)=-1/D_{x\to w}^C Γ x → w C ( E + i 0 ) = − 1/ D x → w C . In particular, G x x C ( E + i 0 ) G_{xx}^C(E+i0) G xx C ( E + i 0 ) and Γ x → w C ( E + i 0 ) \Gamma^C_{x\to w}(E+i0) Γ x → w C ( E + i 0 ) are nonzero, and
Im G x x C ( E + i 0 ) = ∣ G x x C ( E + i 0 ) ∣ 2 ∑ y ∼ C x ∣ h x y ∣ 2 Im Γ y → x C ( E + i 0 ) . (64) \operatorname{Im}G_{xx}^C(E+i0)=|G_{xx}^C(E+i0)|^2\sum_{y\sim_C x}|h_{xy}|^2\operatorname{Im}\Gamma^C_{y\to x}(E+i0). \tag*{(64)} Im G xx C ( E + i 0 ) = ∣ G xx C ( E + i 0 ) ∣ 2 y ∼ C x ∑ ∣ h x y ∣ 2 Im Γ y → x C ( E + i 0 ) . (64) Im Γ x → w C ( E + i 0 ) = ∣ Γ x → w C ( E + i 0 ) ∣ 2 ∑ y ∼ C x y ≠ w ∣ h x y ∣ 2 Im Γ y → x C ( E + i 0 ) . (65) \operatorname{Im}\Gamma^C_{x\to w}(E+i0)=|\Gamma^C_{x\to w}(E+i0)|^2\sum_{\substack{y\sim_C x\\y\ne w}}|h_{xy}|^2\operatorname{Im}\Gamma^C_{y\to x}(E+i0). \tag*{(65)} Im Γ x → w C ( E + i 0 ) = ∣ Γ x → w C ( E + i 0 ) ∣ 2 y ∼ C x y = w ∑ ∣ h x y ∣ 2 Im Γ y → x C ( E + i 0 ) . (65)
Proof. Each mark w v w_v w v is nonzero almost surely, and since there are countably many marks, all edge weights are nonzero almost surely. By Proposition 3.16(i), almost surely, for all z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R ,
G x x C ( z ) = − 1 D x C ( z ) , D x C ( z ) = z + ∑ y ∼ C x ∣ h x y ∣ 2 Γ y → x C ( z ) . G_{xx}^C(z)=-\frac{1}{D_x^C(z)},\qquad D_x^C(z)=z+\sum_{y\sim_C x}|h_{xy}|^2\Gamma^C_{y\to x}(z). G xx C ( z ) = − D x C ( z ) 1 , D x C ( z ) = z + y ∼ C x ∑ ∣ h x y ∣ 2 Γ y → x C ( z ) .
The sum has finitely many terms, and each Γ y → x C ( E + i η ) \Gamma^C_{y\to x}(E+i\eta) Γ y → x C ( E + i η ) converges to a finite limit by condition (a) of Definition 5.7. Then D x C ( E + i η ) → D x C D_x^C(E+i\eta)\to D_x^C D x C ( E + i η ) → D x C , which is finite. If D x C = 0 D_x^C=0 D x C = 0 , then ∣ G x x C ( E + i η ) ∣ = 1 / ∣ D x C ( E + i η ) ∣ → ∞ |G^{\mathcal{C}}_{xx}(E+i\eta)|=1/|D^{\mathcal{C}}_x(E+i\eta)|\to\infty ∣ G xx C ( E + i η ) ∣ = 1/∣ D x C ( E + i η ) ∣ → ∞ , which contradicts the existence of a finite limit of G x x C ( E + i η ) G^{\mathcal{C}}_{xx}(E+i\eta) G xx C ( E + i η ) . We conclude that D x C ≠ 0 D^{\mathcal{C}}_x\ne0 D x C = 0 and G x x C ( E + i 0 ) = − 1 / D x C G^{\mathcal{C}}_{xx}(E+i0)=-1/D^{\mathcal{C}}_x G xx C ( E + i 0 ) = − 1/ D x C , which is nonzero because D x C D^{\mathcal{C}}_x D x C is finite. For Im z > 0 \operatorname{Im}z>0 Im z > 0 , we have Im ( − 1 / D ) = Im D / ∣ D ∣ 2 \operatorname{Im}(-1/D)=\operatorname{Im}D/|D|^2 Im ( − 1/ D ) = Im D /∣ D ∣ 2 , and
Im G x x C ( z ) = ∣ G x x C ( z ) ∣ 2 ( η + ∑ y ∼ C x ∣ h x y ∣ 2 Im Γ y → x C ( z ) ) . \operatorname{Im}G^{\mathcal{C}}_{xx}(z)=|G^{\mathcal{C}}_{xx}(z)|^2\left(\eta+\sum_{y\sim_{\mathcal{C}}x}|h_{xy}|^2\operatorname{Im}\Gamma^{\mathcal{C}}_{y\to x}(z)\right). Im G xx C ( z ) = ∣ G xx C ( z ) ∣ 2 ( η + y ∼ C x ∑ ∣ h x y ∣ 2 Im Γ y → x C ( z ) ) .
Letting η ↓ 0 \eta\downarrow0 η ↓ 0 gives (64) . The assertions about Γ x → w C \Gamma^{\mathcal{C}}_{x\to w} Γ x → w C follow in the same way from Proposition 3.16(ii). Since K \mathcal{K} K is countable, all these statements hold simultaneously on one almost sure event.
All imaginary parts in (64) and (65) are nonnegative, since they are limits of imaginary parts of Borel transforms of positive measures. The next lemma shows that a positive imaginary part at the real axis can only be created by an infinite forward subtree.
Lemma 5.10. Let E ∈ E p E\in\mathcal{E}_p E ∈ E p . Almost surely, the following holds for every removed set Δ \Delta Δ , every component C C C of T Δ \mathcal{T}^{\Delta} T Δ , every x ∈ C x\in C x ∈ C , and every neighbor w w w of x x x in C C C .
(i) If Im G x x C ( E + i 0 ) > 0 \operatorname{Im}G^{\mathcal{C}}_{xx}(E+i0)>0 Im G xx C ( E + i 0 ) > 0 or Im Γ x → w C ( E + i 0 ) > 0 \operatorname{Im}\Gamma^{\mathcal{C}}_{x\to w}(E+i0)>0 Im Γ x → w C ( E + i 0 ) > 0 , then there exists y ∈ V y\in V y ∈ V with Im Γ y ( E + i 0 ) > 0 \operatorname{Im}\Gamma_y(E+i0)>0 Im Γ y ( E + i 0 ) > 0 .
(ii) On the event that Im Γ y ( E + i 0 ) > 0 \operatorname{Im}\Gamma_y(E+i0)>0 Im Γ y ( E + i 0 ) > 0 for every y ∈ V y\in V y ∈ V such that T y \mathcal{T}_y T y is infinite, we have Im G x x C ( E + i 0 ) > 0 \operatorname{Im}G^{\mathcal{C}}_{xx}(E+i0)>0 Im G xx C ( E + i 0 ) > 0 if and only if C C C is infinite, and Im Γ x → w C ( E + i 0 ) > 0 \operatorname{Im}\Gamma^{\mathcal{C}}_{x\to w}(E+i0)>0 Im Γ x → w C ( E + i 0 ) > 0 if and only if C x → w C_{x\to w} C x → w is infinite.
Proof. We work on the intersection of the almost sure events of Definition 5.7 and Lemma 5.9, and we use two facts about paths in the tree. First, a non-backtracking path x 0 , x 1 , x 2 , … x_0,x_1,x_2,\ldots x 0 , x 1 , x 2 , … , that is, a path with x k + 1 ∼ x k x_{k+1}\sim x_k x k + 1 ∼ x k and x k + 1 ≠ x k − 1 x_{k+1}\ne x_{k-1} x k + 1 = x k − 1 , visits distinct vertices. If it is infinite, then there exists k 0 ≤ ∣ x 0 ∣ k_0\le|x_0| k 0 ≤ ∣ x 0 ∣ such that x k + 1 x_{k+1} x k + 1 is a child of x k x_k x k for all k ≥ k 0 k\ge k_0 k ≥ k 0 . Indeed, a step to the parent cannot follow a step to a child without returning to the previous vertex; the path takes at most ∣ x 0 ∣ |x_0| ∣ x 0 ∣ steps toward the root, and then it moves away from the root forever. Second, if ∣ y ∣ > d ( Δ ) |y|>d(\Delta) ∣ y ∣ > d ( Δ ) , then the forward subtree T y \mathcal{T}_y T y contains no vertex of Δ V \Delta_V Δ V and no edge of Δ E \Delta_E Δ E , since all its vertices and all lower endpoints of its edges have length at least ∣ y ∣ |y| ∣ y ∣ . If moreover y ∈ C y\in C y ∈ C and y − ∈ C y^-\in C y − ∈ C , then by Lemma 3.5(ii) the branch C y → y − C_{y\to y^-} C y → y − is the vertex set of T y \mathcal{T}_y T y , and Γ y → y − C = Γ y \Gamma^{\mathcal{C}}_{y\to y^-}=\Gamma_y Γ y → y − C = Γ y .
Proof of (i). Suppose that Im Γ x → w C ( E + i 0 ) > 0 \operatorname{Im}\Gamma^{\mathcal{C}}_{x\to w}(E+i0)>0 Im Γ x → w C ( E + i 0 ) > 0 , and set x 0 = x x_0=x x 0 = x and x − 1 = w x_{-1}=w x − 1 = w . By (65) , there exists a neighbor x 1 ≠ x − 1 x_1\ne x_{-1} x 1 = x − 1 of x 0 x_0 x 0 in C C C with Im Γ x 1 → x 0 C ( E + i 0 ) > 0 \operatorname{Im}\Gamma^{\mathcal{C}}_{x_1\to x_0}(E+i0)>0 Im Γ x 1 → x 0 C ( E + i 0 ) > 0 . Applying (65) again to the oriented edge x 1 → x 0 x_1\to x_0 x 1 → x 0 , and repeating, we obtain an infinite non-backtracking path ( x k ) k ≥ 0 (x_k)_{k\ge0} ( x k ) k ≥ 0 in C C C with Im Γ x k → x k − 1 C ( E + i 0 ) > 0 \operatorname{Im}\Gamma^{\mathcal{C}}_{x_k\to x_{k-1}}(E+i0)>0 Im Γ x k → x k − 1 C ( E + i 0 ) > 0 for all k ≥ 0 k\ge0 k ≥ 0 . Let k > k 0 k>k_0 k > k 0 satisfy ∣ x k ∣ > d ( Δ ) |x_k|>d(\Delta) ∣ x k ∣ > d ( Δ ) . Then x k x_k x k is a child of x k − 1 x_{k-1} x k − 1 , and by the second fact, Γ x k → x k − 1 C = Γ x k \Gamma^{\mathcal{C}}_{x_k\to x_{k-1}}=\Gamma_{x_k} Γ x k → x k − 1 C = Γ x k . The vertex y = x k y=x_k y = x k satisfies Im Γ y ( E + i 0 ) > 0 \operatorname{Im}\Gamma_y(E+i0)>0 Im Γ y ( E + i 0 ) > 0 . If instead Im G x x C ( E + i 0 ) > 0 \operatorname{Im}G^{\mathcal{C}}_{xx}(E+i0)>0 Im G xx C ( E + i 0 ) > 0 , then (64) provides a neighbor w ′ w' w ′ of x x x in C C C with Im Γ w ′ → x C ( E + i 0 ) > 0 \operatorname{Im}\Gamma^{\mathcal{C}}_{w'\to x}(E+i0)>0 Im Γ w ′ → x C ( E + i 0 ) > 0 , and we apply the previous argument to the oriented edge w ′ → x w'\to x w ′ → x . Proof of (ii). If C x → w C_{x\to w} C x → w is finite, then condition (b) of Definition 5.7 and the discussion after it show that Γ x → w C ( E + i 0 ) \Gamma^C_{x\to w}(E+i0) Γ x → w C ( E + i 0 ) is real. Suppose that C x → w C_{x\to w} C x → w is infinite. It induces an infinite, connected, locally finite subtree, and König’s lemma provides an infinite non-backtracking path x = x 0 , x 1 , x 2 , … x=x_0,x_1,x_2,\ldots x = x 0 , x 1 , x 2 , … in C x → w C_{x\to w} C x → w ; its first step is not w w w , since w ∉ C x → w w\notin C_{x\to w} w ∈ / C x → w . Let k > k 0 k>k_0 k > k 0 satisfy ∣ x k ∣ > max { d ( Δ ) , ∣ x ∣ , ∣ w ∣ } |x_k|>\max\{d(\Delta),|x|,|w|\} ∣ x k ∣ > max { d ( Δ ) , ∣ x ∣ , ∣ w ∣ } . Then x k x_k x k is a child of x k − 1 x_{k-1} x k − 1 , and by the second fact, Γ x k → x k − 1 C = Γ x k \Gamma^C_{x_k\to x_{k-1}}=\Gamma_{x_k} Γ x k → x k − 1 C = Γ x k . The forward subtree T x k T_{x_k} T x k is infinite, since it contains x k , x k + 1 , … x_k,x_{k+1},\ldots x k , x k + 1 , … , and by assumption Im Γ x k ( E + i 0 ) > 0 \operatorname{Im}\Gamma_{x_k}(E+i0)>0 Im Γ x k ( E + i 0 ) > 0 . For 0 ≤ i ≤ k − 1 0\leq i\leq k-1 0 ≤ i ≤ k − 1 , we keep only the term y = x i + 1 y=x_{i+1} y = x i + 1 in (65) for the oriented edge x i → x i − 1 x_i\to x_{i-1} x i → x i − 1 , where x − 1 = w x_{-1}=w x − 1 = w , and obtain
Im Γ x i → x i − 1 C ( E + i 0 ) ≥ ∣ Γ x i → x i − 1 C ( E + i 0 ) ∣ 2 ∣ h x i x i + 1 ∣ 2 Im Γ x i + 1 → x i C ( E + i 0 ) . \operatorname{Im}\Gamma^C_{x_i\to x_{i-1}}(E+i0)\geq\left|\Gamma^C_{x_i\to x_{i-1}}(E+i0)\right|^2\left|h_{x_ix_{i+1}}\right|^2\operatorname{Im}\Gamma^C_{x_{i+1}\to x_i}(E+i0). Im Γ x i → x i − 1 C ( E + i 0 ) ≥ Γ x i → x i − 1 C ( E + i 0 ) 2 ∣ h x i x i + 1 ∣ 2 Im Γ x i + 1 → x i C ( E + i 0 ) .
The first two factors on the right are positive by Lemma 5.9. By induction from i = k − 1 i=k-1 i = k − 1 down to i = 0 i=0 i = 0 , we obtain Im Γ x → w C ( E + i 0 ) > 0 \operatorname{Im}\Gamma^C_{x\to w}(E+i0)>0 Im Γ x → w C ( E + i 0 ) > 0 .
For the diagonal entries, if C C C is finite, then G x x C ( E + i 0 ) G^C_{xx}(E+i0) G xx C ( E + i 0 ) is real. If C C C is infinite, then one of the finitely many branches C y → x C_{y\to x} C y → x , y ∼ C x y\sim_C x y ∼ C x , is infinite, since by Lemma 3.5(ii) these branches together with { x } \{x\} { x } make up C C C . For this y y y , the previous paragraph gives Im Γ y → x C ( E + i 0 ) > 0 \operatorname{Im}\Gamma^C_{y\to x}(E+i0)>0 Im Γ y → x C ( E + i 0 ) > 0 , and (64) gives Im G x x C ( E + i 0 ) > 0 \operatorname{Im}G^C_{xx}(E+i0)>0 Im G xx C ( E + i 0 ) > 0 .
The zero–one law and the real boundary law
Recall from Definition 3.2 that, for p > 1 p>1 p > 1 , the survival probability ϱ p \varrho_p ϱ p is the unique solution in ( 0 , 1 ) (0,1) ( 0 , 1 ) of 1 − ϱ = e − p ϱ 1-\varrho=e^{-p\varrho} 1 − ϱ = e − pϱ .
Proposition 5.11. Let E ∈ E p E\in\mathcal{E}_p E ∈ E p with E > 0 E>0 E > 0 , and set u = P ( Im Γ o ( E + i 0 ) > 0 ) u=\mathbb{P}(\operatorname{Im}\Gamma_o(E+i0)>0) u = P ( Im Γ o ( E + i 0 ) > 0 ) .
(i) We have u = 1 − e − p u u=1-e^{-pu} u = 1 − e − p u and u ∈ { 0 , ϱ p } u\in\{0,\varrho_p\} u ∈ { 0 , ϱ p } , and
P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 − u ϱ p ∈ { 0 , 1 } . \mathbb{P}^{\infty}(\operatorname{Im}\Gamma_o(E+i0)=0)=1-\frac{u}{\varrho_p}\in\{0,1\}. P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 − ϱ p u ∈ { 0 , 1 } . If u = ϱ p u=\varrho_p u = ϱ p , then almost surely (under P \mathbb{P} P , and hence under P ∞ \mathbb{P}^{\infty} P ∞ ), the following holds for every vertex v ∈ V v\in V v ∈ V , every removed set Δ \Delta Δ , every component C C C of T Δ T^\Delta T Δ and every x ∈ C : x\in C: x ∈ C : \operatorname{Im}\Gamma_v(E+i0)>0$ if and only if T v T_v T v is infinite, and Im G x x C ( E + i 0 ) > 0 \operatorname{Im}G^C_{xx}(E+i0)>0 Im G xx C ( E + i 0 ) > 0 if and only if C C C is infinite. In particular, P ∞ \mathbb{P}^{\infty} P ∞ -almost surely, Im G v v ( E + i 0 ) > 0 \operatorname{Im}G_{vv}(E+i0)>0 Im G v v ( E + i 0 ) > 0 for every v ∈ V v\in V v ∈ V .
(ii) Suppose that P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 \mathbb{P}^{\infty}(\operatorname{Im}\Gamma_o(E+i0)=0)=1 P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 . Then u = 0 u=0 u = 0 , and almost surely (under P \mathbb{P} P , and hence under P ∞ \mathbb{P}^{\infty} P ∞ ) the following hold. For every v ∈ V v\in V v ∈ V , the boundary value Γ v ( E + i 0 ) \Gamma_v(E+i0) Γ v ( E + i 0 ) is finite, real and nonzero, and the denominators D v = − 1 / Γ v ( E + i 0 ) D_v=-1/\Gamma_v(E+i0) D v = − 1/ Γ v ( E + i 0 ) satisfy
D v = E − 1 p ∑ k = 1 N v γ v k D v k . (66) D_v=E-\frac{1}{p}\sum_{k=1}^{N_v}\frac{\gamma_{vk}}{D_{vk}}. \tag*{(66)} D v = E − p 1 k = 1 ∑ N v D v k γ v k . (66) For every removed set Δ \Delta Δ , every component C C C of T Δ T^\Delta T Δ , every x ∈ C x\in C x ∈ C and every neighbor w w w of x x x in C C C , the boundary values G x x C ( E + i 0 ) G^C_{xx}(E+i0) G xx C ( E + i 0 ) and Γ x → w C ( E + i 0 ) \Gamma^C_{x\to w}(E+i0) Γ x → w C ( E + i 0 ) are real and nonzero. Finally, the law of D o D_o D o under P \mathbb{P} P is a real solution at energy $E.
Proof. We work on the almost sure events of Definition 5.7 and Lemmas 5.9 and 5.10, and on the event where Γ v ( z ) = G ∞ ( z ) ∘ θ v \Gamma_v(z)=G_{\infty}(z)\circ\theta_v Γ v ( z ) = G ∞ ( z ) ∘ θ v for all v ∈ V ∖ { o } v\in V\setminus\{o\} v ∈ V ∖ { o } . Let A \mathcal{A} A be the event that lim k → ∞ G ∞ ( E + i / k ) \lim_{k\to\infty}G_{\infty}(E+\mathrm{i}/k) lim k → ∞ G ∞ ( E + i / k ) exists and has positive imaginary part. For v ∈ V v\in V v ∈ V , the event A v = { Im Γ v ( E + i 0 ) > 0 } A_v=\{\operatorname{Im}\Gamma_v(E+\mathrm{i}0)>0\} A v = { Im Γ v ( E + i 0 ) > 0 } then coincides with θ v − 1 A \theta_v^{-1}\mathcal{A} θ v − 1 A up to a null set.
Step 1. Finite trees. On the event S c \mathcal{S}^{\mathrm{c}} S c , the tree T \mathcal{T} T is finite, and condition (b) of Definition 5.7 with Λ = V \Lambda=V Λ = V shows that Γ o ( E + i 0 ) = G o o ( E ) \Gamma_o(E+\mathrm{i}0)=G_{oo}(E) Γ o ( E + i 0 ) = G oo ( E ) is real. In other words, A o ⊂ S A_o\subset\mathcal{S} A o ⊂ S up to a null set. Since θ v \theta_v θ v preserves P \mathbb{P} P , we also have θ v − 1 A ⊂ S v \theta_v^{-1}\mathcal{A}\subset\mathcal{S}_v θ v − 1 A ⊂ S v up to a null set, for every v ∈ U v\in\mathcal{U} v ∈ U .
Step 2. The fixed point equation. By (65) , applied with C = V C=V C = V and the oriented edge v → v − v\to v^{-} v → v − , and by (64) at the root, we have
Im Γ v ( E + i 0 ) = ∣ Γ v ( E + i 0 ) ∣ 2 ∑ k = 1 N v ∣ h v , v k ∣ 2 Im Γ v k ( E + i 0 ) . \operatorname{Im}\Gamma_v(E+\mathrm{i}0)=\lvert\Gamma_v(E+\mathrm{i}0)\rvert^2\sum_{k=1}^{N_v}\lvert h_{v,v_k}\rvert^2\operatorname{Im}\Gamma_{v_k}(E+\mathrm{i}0). Im Γ v ( E + i 0 ) = ∣ Γ v ( E + i 0 ) ∣ 2 k = 1 ∑ N v ∣ h v , v k ∣ 2 Im Γ v k ( E + i 0 ) .
Since Γ v ( E + i 0 ) ≠ 0 \Gamma_v(E+\mathrm{i}0)\ne0 Γ v ( E + i 0 ) = 0 and the edge weights are nonzero, we obtain A v = ⋃ k ≤ N v A v k A_v=\bigcup_{k\le N_v}A_{v_k} A v = ⋃ k ≤ N v A v k almost surely. At the root, A o = ⋃ k ≤ N o θ k − 1 A A_o=\bigcup_{k\le N_o}\theta_k^{-1}\mathcal{A} A o = ⋃ k ≤ N o θ k − 1 A up to a null set. By Lemma 4.2(ii), the configurations θ k ω \theta_k\omega θ k ω , k ≥ 1 k\ge1 k ≥ 1 , are independent with law P \mathbb{P} P and independent of N o N_o N o , and the events θ k − 1 A \theta_k^{-1}\mathcal{A} θ k − 1 A are independent of each other and of N o N_o N o , with probability u u u each. This gives
1 − u = P ( A o c ) = E ( 1 − u ) N o = e − p u . 1-u=\mathbb{P}(A_o^{\mathrm{c}})=\mathbb{E}(1-u)^{N_o}=e^{-pu}. 1 − u = P ( A o c ) = E ( 1 − u ) N o = e − p u .
The function u ↦ 1 − e − p u − u u\mapsto1-e^{-pu}-u u ↦ 1 − e − p u − u is strictly concave on [ 0 , 1 ] [0,1] [ 0 , 1 ] , vanishes at 0 0 0 , has derivative p − 1 > 0 p-1>0 p − 1 > 0 at 0 0 0 , and equals − e − p < 0 -e^{-p}<0 − e − p < 0 at 1 1 1 ; it has exactly two zeros in [ 0 , 1 ] [0,1] [ 0 , 1 ] , namely 0 0 0 and ϱ p \varrho_p ϱ p . By Step 1, we have P ∞ ( A o ) = P ( A o ∩ S ) / ϱ p = u / ϱ p \mathbb{P}^{\infty}(A_o)=\mathbb{P}(A_o\cap\mathcal{S})/\varrho_p=u/\varrho_p P ∞ ( A o ) = P ( A o ∩ S ) / ϱ p = u / ϱ p , which is 0 0 0 or 1 1 1 .
Step 3. Propagation when u = ϱ p u=\varrho_p u = ϱ p . Suppose u = ϱ p u=\varrho_p u = ϱ p . By Step 1, P ( A △ S ) = P ( S ) − P ( A ) = 0 \mathbb{P}(\mathcal{A}\mathbin{\triangle}\mathcal{S})=\mathbb{P}(\mathcal{S})-\mathbb{P}(\mathcal{A})=0 P ( A △ S ) = P ( S ) − P ( A ) = 0 . Since θ v \theta_v θ v preserves P \mathbb{P} P , the event θ v − 1 A △ S v \theta_v^{-1}\mathcal{A}\mathbin{\triangle}\mathcal{S}_v θ v − 1 A △ S v is null for every v ∈ U v\in\mathcal{U} v ∈ U . Taking the union over the countably many v v v , we find that almost surely A v = S v A_v=\mathcal{S}_v A v = S v for every v ∈ V v\in V v ∈ V . This is the first assertion about Γ v \Gamma_v Γ v , and on this event Lemma 5.10(ii) gives the assertion about G x x C G^C_{xx} G xx C . Under P ∞ \mathbb{P}^{\infty} P ∞ , the tree is infinite, and taking Δ \Delta Δ empty and C = V C=V C = V , we find that Im G v v ( E + i 0 ) > 0 \operatorname{Im}G_{vv}(E+\mathrm{i}0)>0 Im G v v ( E + i 0 ) > 0 for all v ∈ V v\in V v ∈ V . This proves (i).
Step 4. The case P ∞ ( A o ) = 0 \mathbb{P}^{\infty}(A_o)=0 P ∞ ( A o ) = 0 . Suppose P ∞ ( A o ) = 0 \mathbb{P}^{\infty}(A_o)=0 P ∞ ( A o ) = 0 . Decomposing according to survival and extinction, and using that P ( A o ∩ S c ) = 0 \mathbb{P}(A_o\cap\mathcal{S}^{\mathrm{c}})=0 P ( A o ∩ S c ) = 0 by Step 1, we have
u = P ( A o ) = ϱ p P ∞ ( A o ) + P ( A o ∩ S c ) = 0. u=\mathbb{P}(A_o)=\varrho_p\mathbb{P}^{\infty}(A_o)+\mathbb{P}(A_o\cap\mathcal{S}^{\mathrm{c}})=0. u = P ( A o ) = ϱ p P ∞ ( A o ) + P ( A o ∩ S c ) = 0.
Since θ v \theta_v θ v preserves P \mathbb{P} P , we have P ( θ v − 1 A ) = u = 0 \mathbb{P}(\theta_v^{-1}\mathcal{A})=u=0 P ( θ v − 1 A ) = u = 0 for every v ∈ U v\in\mathcal{U} v ∈ U , and almost surely Γ v ( E + i 0 ) \Gamma_v(E+\mathrm{i}0) Γ v ( E + i 0 ) is real for every v ∈ V v\in V v ∈ V . It is finite by condition (a) and nonzero by Lemma 5.9. By Lemma 5.10(i), the boundary values G x x C ( E + i 0 ) G^C_{xx}(E+\mathrm{i}0) G xx C ( E + i 0 ) and Γ x → w C ( E + i 0 ) \Gamma^C_{x\to w}(E+\mathrm{i}0) Γ x → w C ( E + i 0 ) are real for every removed set, component and vertex, and they are nonzero by Lemma 5.9. Further, Lemma 5.9 with C = V C=V C = V , applied to the oriented edge v → v ˉ v\to\bar v v → v ˉ if v ≠ o v\ne o v = o and to the diagonal entry G o o = Γ o G_{oo}=\Gamma_o G oo = Γ o if v = o v=o v = o , gives
D v = − 1 Γ v ( E + i 0 ) = E + ∑ k = 1 N v ∣ h v , v k ∣ 2 Γ v k ( E + i 0 ) = E − 1 p ∑ k = 1 N v Y v k D v k . (67) D_v=-\frac{1}{\Gamma_v(E+\mathrm{i}0)}=E+\sum_{k=1}^{N_v}|h_{v,v_k}|^2\Gamma_{v_k}(E+\mathrm{i}0)=E-\frac{1}{p}\sum_{k=1}^{N_v}\frac{Y_{v_k}}{D_{v_k}}. \tag*{(67)} D v = − Γ v ( E + i 0 ) 1 = E + k = 1 ∑ N v ∣ h v , v k ∣ 2 Γ v k ( E + i 0 ) = E − p 1 k = 1 ∑ N v D v k Y v k . (67)
which is (5.9).
Step 5. The law of D o D_o D o . Let D : Ω → R \mathcal{D}:\Omega\to\mathbb{R} D : Ω → R be the measurable map defined by D = − 1 / lim k → ∞ G o o ( E + i / k ) \mathcal{D}=-1/\lim_{k\to\infty}G_{oo}(E+\mathrm{i}/k) D = − 1/ lim k → ∞ G oo ( E + i / k ) if this limit exists and is real and nonzero, and D = E \mathcal{D}=E D = E otherwise. By Step 4 and the identity Γ v = G o o ∘ θ v \Gamma_v=G_{oo}\circ\theta_v Γ v = G oo ∘ θ v , almost surely D v = D ( θ v ω ) D_v=\mathcal{D}(\theta_v\omega) D v = D ( θ v ω ) for every v ∈ V v\in V v ∈ V . By (5.9) at the root,
D o = E − 1 p ∑ k = 1 N o Y k D ( θ k ω ) . D_o=E-\frac{1}{p}\sum_{k=1}^{N_o}\frac{Y_k}{\mathcal{D}(\theta_k\omega)}. D o = E − p 1 k = 1 ∑ N o D ( θ k ω ) Y k .
By Lemma 4.2(ii), the variables N o N_o N o , ( Y k ) k ≥ 1 (Y_k)_{k\ge1} ( Y k ) k ≥ 1 and ( θ k ω ) k ≥ 1 (\theta_k\omega)_{k\ge1} ( θ k ω ) k ≥ 1 are independent, and the θ k ω \theta_k\omega θ k ω have law P \mathbb{P} P . It follows that the variables D ( θ k ω ) \mathcal{D}(\theta_k\omega) D ( θ k ω ) , k ≥ 1 k\ge1 k ≥ 1 , are independent with the law of D o D_o D o , and independent of N o N_o N o and of the squared moduli of the root edge marks. Since D o ∈ R ∖ { 0 } D_o\in\mathbb{R}\setminus\{0\} D o ∈ R ∖ { 0 } almost surely, the law of D o D_o D o is a real solution at energy E E E .
The nonvanishing of the boundary denominators in Proposition 5.11(ii) is a consequence of the finiteness of the boundary values, which is part of the definition of E p \mathcal{E}_p E p . We also record the structure of the law of D o D_o D o given the children. Under the no-ac hypothesis, by Step 5 of the proof of Proposition 5.11, D o D_o D o is either the constant E E E (if N o = 0 N_o=0 N o = 0 ) or an affine function of Y 1 , … , Y N o Y_1,\ldots,Y_{N_o} Y 1 , … , Y N o with the nonzero real coefficients − 1 / ( p D ( θ k ω ) ) -1/(p\mathcal{D}(\theta_k\omega)) − 1/ ( p D ( θ k ω )) (if N o ≥ 1 N_o\ge1 N o ≥ 1 ). These coefficients are measurable with respect to the σ \sigma σ -field generated by N o N_o N o and ( θ k ω ) k ≥ 1 (\theta_k\omega)_{k\ge1} ( θ k ω ) k ≥ 1 , and the squared moduli Y 1 , … , Y N o Y_1,\ldots,Y_{N_o} Y 1 , … , Y N o are fresh given this σ \sigma σ -field, by the example after Lemma 4.17 and consequence (a) after Definition 4.16. By Lemma 4.18(iii) and (4.14), the conditional law of D o D_o D o on the event { N o ≥ 1 } \{N_o\ge1\} { N o ≥ 1 } has no atoms. The same argument is used in Section 5.6 for the real recursive tree process.
Definition 5.12. Let E ∈ E p E\in\mathcal{E}_p E ∈ E p with E > 0 E>0 E > 0 . Following [12], we say that the no-ac hypothesis holds at E E E if P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 \mathbb{P}^{\infty}(\operatorname{Im}\Gamma_o(E+\mathrm{i}0)=0)=1 P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 . In this case, with D o = − 1 / Γ o ( E + i 0 ) D_o=-1/\Gamma_o(E+\mathrm{i}0) D o = − 1/ Γ o ( E + i 0 ) , the real boundary law at E E E is π E = Law P ( D o ) \pi_E=\operatorname{Law}_{\mathbb{P}}(D_o) π E = Law P ( D o ) , and the conditioned laws are
π E ∞ = Law P ∞ ( D o ) , π E f = Law P ( D o ∣ S c ) . \pi_E^\infty=\operatorname{Law}_{\mathbb{P}^{\infty}}(D_o),\qquad\pi_E^{\mathrm{f}}=\operatorname{Law}_{\mathbb{P}}(D_o\mid\mathcal{S}^{\mathrm{c}}). π E ∞ = Law P ∞ ( D o ) , π E f = Law P ( D o ∣ S c ) .
In the B B B -normalization (E > 2 E>2 E > 2 ), the real boundary law is the law of g ( E ) D o g(E)D_o g ( E ) D o under P \mathbb{P} P .
By Proposition 5.11(ii), π E \pi_E π E is a real solution at energy E E E , and all the conclusions of Proposition 5.5 apply to it. Further,
π E = ϱ p π E ∞ + ( 1 − ϱ p ) π E f , π E ∞ ≤ ϱ p − 1 π E as measures, \pi_E=\varrho_p\pi_E^\infty+(1-\varrho_p)\pi_E^{\mathrm{f}},\qquad\pi_E^\infty\le\varrho_p^{-1}\pi_E\quad\text{as measures,} π E = ϱ p π E ∞ + ( 1 − ϱ p ) π E f , π E ∞ ≤ ϱ p − 1 π E as measures,
and π E \pi^{\mathrm{E}} π E has no atoms. To see the last claim, note that on S \mathcal{S} S the root has at least one child, and by the previous paragraph the law of D o D_o D o restricted to the event { N o ≥ 1 } \{N_o \ge1\} { N o ≥ 1 } has no atoms. Then P ∞ ( D o = c ) ≤ ϱ p − 1 P ( D o = c , N o ≥ 1 ) = 0 \mathbb{P}^{\infty}(D_o=c) \le\varrho_p^{-1}\mathbb{P}(D_o=c,N_o\ge1)=0 P ∞ ( D o = c ) ≤ ϱ p − 1 P ( D o = c , N o ≥ 1 ) = 0 for every c ∈ R c\in\mathbb{R} c ∈ R . In Remark 5.15 below, we note that the boundary denominators form a recursive tree process with marginal π E \pi_{\mathrm{E}} π E which is a function of the tree.
Recursive tree processes
For an arbitrary real solution, a family of denominators satisfying the recursion need not be a function of the tree. The following lemma constructs such a family on an extended probability space. It is used in § 8.2 , where the real solution need not be the real boundary law.
Lemma 5.13. Let E > 0 E>0 E > 0 and let π \pi π be a real solution at energy E E E . There exist a probability space and, on it, a random configuration ω \omega ω with law P \mathbb{P} P (Definition 4.1) and real random variables ( D v ) v ∈ U (D_v)_{v\in\mathbb{U}} ( D v ) v ∈ U with the following properties.
(i) For every v ∈ U v\in\mathbb{U} v ∈ U , we have D v ∼ π D_v\sim\pi D v ∼ π , and almost surely D v ∈ R ∖ { 0 } D_v\in\mathbb{R}\setminus\{0\} D v ∈ R ∖ { 0 } and
D v = E − 1 p ∑ k = 1 N v γ v k D v k . D_v=E-\frac{1}{p}\sum_{k=1}^{N_v}\frac{\gamma_{v_k}}{D_{v_k}}. D v = E − p 1 k = 1 ∑ N v D v k γ v k . (ii) For every n ≥ 0 n\ge0 n ≥ 0 , the variables D v D_v D v , v ∈ U n v\in\mathbb{U}_n v ∈ U n , are independent with law π \pi π , and they are independent of the σ \sigma σ -field F n \mathcal{F}_n F n of Section 5.3.
(iii) Let v 1 , … , v k ∈ U v_1,\ldots,v_k\in\mathbb{U} v 1 , … , v k ∈ U be distinct words none of which is a prefix of another, and let F ( v 1 , … , v k ) \mathcal{F}(v_1,\ldots,v_k) F ( v 1 , … , v k ) be the σ \sigma σ -field generated by the coordinates N u , ζ u N_u,\zeta_u N u , ζ u with u ∉ ⋃ i v i U u\notin\bigcup_i v_i\mathbb{U} u ∈ / ⋃ i v i U and w u w_u w u with u ∉ ⋃ i v i U ∗ u\notin\bigcup_i v_i\mathbb{U}_* u ∈ / ⋃ i v i U ∗ , where v U ∗ = v U ∖ { v } v\mathbb{U}_*=v\mathbb{U}\setminus\{v\} v U ∗ = v U ∖ { v } . Then the pairs ( θ v i ω , ( D v i u ) u ∈ U ) (\theta_{v_i}\omega,(D_{v_i u})_{u\in\mathbb{U}}) ( θ v i ω , ( D v i u ) u ∈ U ) , $1\le i\le k$, are independent of each other and of F ( v 1 , … , v k ) \mathcal{F}(v_1,\ldots,v_k) F ( v 1 , … , v k ) , and each has the law of ( ω , ( D u ) u ∈ U ) (\omega,(D_u)_{u\in\mathbb{U}}) ( ω , ( D u ) u ∈ U ) . In particular, for every $n\ge 0 this holds for the pairs indexed by v ∈ U n v\in\mathbb{U}_n v ∈ U n , with F n \mathcal{F}_n F n in place of F ( v 1 , … , v k ) \mathcal{F}(v_1,\ldots,v_k) F ( v 1 , … , v k ) .
We call ( ω , ( D v ) v ∈ U ) (\omega,(D_v)_{v\in\mathbb{U}}) ( ω , ( D v ) v ∈ U ) a recursive tree process with marginal π \pi π , in the sense of [17]. By (i), the restriction of ( D v ) (D_v) ( D v ) to V V V is a family of real denominators on the tree that satisfies the real recursion at every vertex.
Remark 5.14. We comment on the measurability structure of the process.
(i) For ∣ v ∣ = n |v|=n ∣ v ∣ = n , the variable D v D_v D v is not independent of θ v ω \theta_v\omega θ v ω ; for instance, D v = E D_v=E D v = E on the event { N v = 0 } \{N_v=0\} { N v = 0 } . Property (ii) asserts independence from F n \mathcal{F}_n F n only, and F n \mathcal{F}_n F n contains the offspring numbers N u N_u N u with ∣ u ∣ < n |u|<n ∣ u ∣ < n , but not N v N_v N v with ∣ v ∣ = n |v|=n ∣ v ∣ = n . By (i), the family ( D v ) v ∈ U ≤ n (D_v)_{v\in\mathbb{U}_{\le n}} ( D v ) v ∈ U ≤ n is determined by ( D v ) v ∈ U n (D_v)_{v\in\mathbb{U}_n} ( D v ) v ∈ U n and F n \mathcal{F}_n F n . Accordingly, for a measurable functional F ≥ 0 F\ge0 F ≥ 0 of F n \mathcal{F}_n F n , of ( D v ) v ∈ U ≤ n (D_v)_{v\in\mathbb{U}_{\le n}} ( D v ) v ∈ U ≤ n and of a word v ∈ U L v\in\mathbb{U}_L v ∈ U L with L ≤ n L\le n L ≤ n , property (ii) allows one to integrate out ( D v ) v ∈ U n (D_v)_{v\in\mathbb{U}_n} ( D v ) v ∈ U n as independent π \pi π -distributed variables; the result is a measurable functional of ω \omega ω and v v v , to which the many-to-one and many-to-two formulas of Proposition 4.8 apply. This is the setting of Remark 4.13.
(ii) The variables D v D_v D v need not be measurable with respect to ω \omega ω ; that is, the process need not be endogenous in the sense of [17]. Under the no-ac hypothesis at a regular energy, the boundary denominators of Proposition 5.11 (ii) form an endogenous recursive tree process with marginal π E \pi_E π E (Remark 5.15).
Proof of Lemma 5.13. For u ∈ U u \in\mathbb{U} u ∈ U , let ζ u = ( N u , ζ u , ( w u k ) k ≥ 1 ) \zeta_u = (N_u,\zeta_u,(w_{uk})_{k\geq1}) ζ u = ( N u , ζ u , ( w u k ) k ≥ 1 ) be the innovation at u u u [17], with values in the Polish space Ξ = N 0 × [ 0 , 1 ] N × C N \Xi= \mathbb{N}_0 \times[0,1]^{\mathbb{N}} \times\mathbb{C}^{\mathbb{N}} Ξ = N 0 × [ 0 , 1 ] N × C N , and let ν ζ \nu_\zeta ν ζ be its law under P \mathbb{P} P . Every coordinate of ω \omega ω appears in exactly one innovation, and ω \omega ω and ( ζ u ) u ∈ U (\zeta_u)_{u\in\mathbb{U}} ( ζ u ) u ∈ U determine each other, and under P \mathbb{P} P the innovations are independent with law ν ζ \nu_\zeta ν ζ . In terms of innovations, F n = σ ( ζ u : ∣ u ∣ < n ) \mathcal{F}_n = \sigma(\zeta_u : |u| < n) F n = σ ( ζ u : ∣ u ∣ < n ) , and θ v ω \theta_v\omega θ v ω corresponds to ( ζ v u ) u ∈ U (\zeta_{vu})_{u\in\mathbb{U}} ( ζ v u ) u ∈ U . For an innovation ζ = ( n , ζ , ( w k ) k ≥ 1 ) \zeta= (n,\zeta,(w_k)_{k\geq1}) ζ = ( n , ζ , ( w k ) k ≥ 1 ) and a sequence ( d k ) k ≥ 1 (d_k)_{k\geq1} ( d k ) k ≥ 1 in R ∖ { 0 } \mathbb{R} \setminus\{0\} R ∖ { 0 } , we set
Ψ ( ζ , ( d k ) k ≥ 1 ) = E − 1 p ∑ k = 1 n ∣ w k ∣ 2 d k . \Psi(\zeta,(d_k)_{k\geq1}) = E - \frac{1}{p}\sum_{k=1}^{n}\frac{|w_k|^2}{d_k}. Ψ ( ζ , ( d k ) k ≥ 1 ) = E − p 1 k = 1 ∑ n d k ∣ w k ∣ 2 .
Step 1. Finite-depth processes. Fix m ≥ 0 m \geq0 m ≥ 0 . On some probability space, let ( ζ u ) u ∈ U (\zeta_u)_{u\in\mathbb{U}} ( ζ u ) u ∈ U and ( Θ v ) v ∈ U m (\Theta_v)_{v\in\mathbb{U}_m} ( Θ v ) v ∈ U m be independent, with ζ u ∼ ν ζ \zeta_u \sim\nu_\zeta ζ u ∼ ν ζ and Θ v ∼ π \Theta_v \sim\pi Θ v ∼ π . We define D v ( m ) = Θ v D_v^{(m)} = \Theta_v D v ( m ) = Θ v for ∣ v ∣ = m |v| = m ∣ v ∣ = m , and for ∣ v ∣ < m |v| < m ∣ v ∣ < m , by backward induction on ∣ v ∣ |v| ∣ v ∣ ,
D v ( m ) = Ψ ( ζ v , ( D v k ( m ) ) k ≥ 1 ) . D_v^{(m)} = \Psi\left(\zeta_v,(D_{vk}^{(m)})_{k\geq1}\right). D v ( m ) = Ψ ( ζ v , ( D v k ( m ) ) k ≥ 1 ) .
We claim that for 0 ≤ l ≤ m 0 \leq l \leq m 0 ≤ l ≤ m , the family ( ( ζ u ) ∣ u ∣ < l , ( D v ( m ) ) ∣ v ∣ = l ) ((\zeta_u)_{|u|<l},(D_v^{(m)})_{|v|=l}) (( ζ u ) ∣ u ∣ < l , ( D v ( m ) ) ∣ v ∣ = l ) has independent coordinates, with ζ u ∼ ν ζ \zeta_u \sim\nu_\zeta ζ u ∼ ν ζ and D v ( m ) ∼ π D_v^{(m)} \sim\pi D v ( m ) ∼ π ; in particular D v ( m ) ≠ 0 D_v^{(m)} \ne0 D v ( m ) = 0 almost surely, and the recursion is well defined. For l = m l=m l = m this holds by construction. Suppose that it holds for l + 1 ≤ m l+1 \leq m l + 1 ≤ m . Then the pairs ( ζ v , ( D v k ( m ) ) k ≥ 1 ) (\zeta_v,(D_{vk}^{(m)})_{k\geq1}) ( ζ v , ( D v k ( m ) ) k ≥ 1 ) , ∣ v ∣ = l |v|=l ∣ v ∣ = l , are independent of each other and of ( ζ u ) ∣ u ∣ < l (\zeta_u)_{|u|<l} ( ζ u ) ∣ u ∣ < l , and in each pair the variables ζ v , D v 1 ( m ) , D v 2 ( m ) , … \zeta_v,D_{v1}^{(m)},D_{v2}^{(m)},\ldots ζ v , D v 1 ( m ) , D v 2 ( m ) , … are independent with laws ν ζ , π , π , … \nu_\zeta,\pi,\pi,\ldots ν ζ , π , π , … . Since π \pi π is a real solution, D v ( m ) = Ψ ( ζ v , ( D v k ( m ) ) k ) D_v^{(m)}=\Psi(\zeta_v,(D_{vk}^{(m)})_k) D v ( m ) = Ψ ( ζ v , ( D v k ( m ) ) k ) has law π \pi π . This proves the claim for l l l .
Step 2. Consistency. For l ≤ m l \leq m l ≤ m , the family ( ( ζ u ) ∣ u ∣ < l , ( D v ( m ) ) ∣ v ∣ ≤ l ) ((\zeta_u)_{|u|<l},(D_v^{(m)})_{|v|\leq l}) (( ζ u ) ∣ u ∣ < l , ( D v ( m ) ) ∣ v ∣ ≤ l ) is the image of ( ( ζ u ) ∣ u ∣ < l , ( D v ( m ) ) ∣ v ∣ = l ) ((\zeta_u)_{|u|<l},(D_v^{(m)})_{|v|=l}) (( ζ u ) ∣ u ∣ < l , ( D v ( m ) ) ∣ v ∣ = l ) under a map that does not depend on m m m , namely the recursion from generation l l l to the root. By Step 1, the law of the latter family does not depend on m ≥ l m \geq l m ≥ l . It follows that the law μ l \mu_l μ l of ( ( ζ u ) ∣ u ∣ < l , ( D v ( l ) ) ∣ v ∣ ≤ l ) ((\zeta_u)_{|u|<l},(D_v^{(l)})_{|v|\leq l}) (( ζ u ) ∣ u ∣ < l , ( D v ( l ) ) ∣ v ∣ ≤ l ) on Ω l = Ξ { ∣ u ∣ < l } × R { ∣ v ∣ ≤ l } \Omega_l = \Xi^{\{|u|<l\}} \times\mathbb{R}^{\{|v|\leq l\}} Ω l = Ξ { ∣ u ∣ < l } × R { ∣ v ∣ ≤ l } is the image of μ l + 1 \mu_{l+1} μ l + 1 under the coordinate projection Ω l + 1 → Ω l \Omega_{l+1} \to\Omega_l Ω l + 1 → Ω l . The spaces Ω l \Omega_l Ω l are countable products of Polish spaces, and the Kolmogorov extension theorem [81] provides a probability measure μ \mu μ on Ξ U × R U \Xi^{\mathbb{U}} \times\mathbb{R}^{\mathbb{U}} Ξ U × R U whose image under the projection onto every Ω l \Omega_l Ω l is μ l \mu_l μ l . We let ( ζ u ) (\zeta_u) ( ζ u ) and ( D v ) (D_v) ( D v ) be the coordinates under μ \mu μ , and ω \omega ω the configuration corresponding to ( ζ u ) (\zeta_u) ( ζ u ) .
Step 3. Properties (i) and (ii). For every l l l , the coordinates ( ξ u ) ∣ u ∣ < l (\xi_u)_{|u|<l} ( ξ u ) ∣ u ∣ < l are independent with law ν ξ \nu_\xi ν ξ under μ l \mu_l μ l , which shows that ω \omega ω has law P \mathbb{P} P . By Step 1 with m = l m=l m = l , the variables ( D v ) ∣ v ∣ = l (D_v)_{|v|=l} ( D v ) ∣ v ∣ = l are independent with law π \pi π and independent of ( ξ u ) ∣ u ∣ < l (\xi_u)_{|u|<l} ( ξ u ) ∣ u ∣ < l , which proves (ii) and the first part of (i); in particular D v ≠ 0 D_v \ne0 D v = 0 almost surely. The event { D v = Ψ ( ξ v , ( D v k ) k ≥ 1 ) } \{D_v=\Psi(\xi_v,(D_{vk})_{k\geq1})\} { D v = Ψ ( ξ v , ( D v k ) k ≥ 1 )} depends only on coordinates in Ω ∣ v ∣ + 1 \Omega_{|v|+1} Ω ∣ v ∣ + 1 , and it has probability one under μ ∣ v ∣ + 1 \mu_{|v|+1} μ ∣ v ∣ + 1 by construction. This proves (i).
Step 4. Property (iii). In terms of innovations, θ v i ω \theta_{v_i}\omega θ v i ω corresponds to ( ξ v i u ) u ∈ U (\xi_{v_i u})_{u\in\mathbb{U}} ( ξ v i u ) u ∈ U , and F ( v 1 , … , v k ) \mathcal{F}(v_1,\ldots,v_k) F ( v 1 , … , v k ) is generated by the innovations ξ u \xi_u ξ u with u ∉ ⋃ i v i U u\notin\bigcup_i v_i\mathbb{U} u ∈ / ⋃ i v i U ; indeed, the mark w v i w_{v_i} w v i of the edge into v i v_i v i belongs to the innovation at the parent of v i v_i v i . Let m ≥ max i ∣ v i ∣ m\geq\max_i|v_i| m ≥ max i ∣ v i ∣ . Under μ m \mu_m μ m , the family ( ξ v i u , D v i u ) (\xi_{v_i u},D_{v_i u}) ( ξ v i u , D v i u ) , ∣ v i u ∣ ≤ m |v_i u|\leq m ∣ v i u ∣ ≤ m (with ξ v i u \xi_{v_i u} ξ v i u only for ∣ v i u ∣ < m |v_i u|<m ∣ v i u ∣ < m ), is the image of ( ξ v i u ) ∣ v i u ∣ < m (\xi_{v_i u})_{|v_i u|<m} ( ξ v i u ) ∣ v i u ∣ < m and ( D v i u ) ∣ v i u ∣ = m (D_{v_i u})_{|v_i u|=m} ( D v i u ) ∣ v i u ∣ = m under the recursion. By Step 1, these inputs for i = 1 , … , k i=1,\ldots,k i = 1 , … , k and the family ( ξ u ) u ∉ ⋃ i v i U , ∣ u ∣ < m (\xi_u)_{u\notin\bigcup_i v_i\mathbb{U},\,|u|<m} ( ξ u ) u ∈ / ⋃ i v i U , ∣ u ∣ < m are independent, since they consist of distinct coordinates of a family with independent coordinates. This shows that the restrictions of the pairs in (iii) to words of length at most m m m are independent of each other and of the restriction of F ( v 1 , … , v k ) \mathcal{F}(v_1,\ldots,v_k) F ( v 1 , … , v k ) to innovations at words of length less than m m m . Further, the inputs for v i v_i v i , read in the coordinates v i u ↦ u v_i u\mapsto u v i u ↦ u , have the same law as the inputs of the process of depth m − ∣ v i ∣ m-|v_i| m − ∣ v i ∣ rooted at o o o , and the recursion is the same map. It follows that the restriction of ( ξ v i u , D v i u ) u ∈ U (\xi_{v_i u},D_{v_i u})_{u\in\mathbb{U}} ( ξ v i u , D v i u ) u ∈ U to ∣ v i u ∣ ≤ m |v_i u|\leq m ∣ v i u ∣ ≤ m has the law of the restriction of ( ξ u , D u ) u ∈ U (\xi_u,D_u)_{u\in\mathbb{U}} ( ξ u , D u ) u ∈ U to ∣ u ∣ ≤ m − ∣ v i ∣ |u|\leq m-|v_i| ∣ u ∣ ≤ m − ∣ v i ∣ . Since m m m is arbitrary, and laws on countable products are determined by their finite-dimensional marginals, (iii) follows. For { v 1 , … , v k } = U n \{v_1,\ldots,v_k\}=\mathbb{U}_n { v 1 , … , v k } = U n (which is infinite, but the argument is the same), the σ \sigma σ -field F ( v 1 , … , v k ) \mathcal{F}(v_1,\ldots,v_k) F ( v 1 , … , v k ) is F n \mathcal{F}_n F n .
Remark 5.15. Let E ∈ E p E\in\mathcal{E}_p E ∈ E p with E > 0 E>0 E > 0 , suppose that the no-ac hypothesis holds at E E E , let D \mathcal{D} D be the map of Step 5 of the proof of Proposition 5.11, and set D v = D ( θ v ω ) D_v=\mathcal{D}(\theta_v\omega) D v = D ( θ v ω ) for v ∈ U v\in\mathbb{U} v ∈ U , where ω \omega ω has law P \mathbb{P} P . Then ( ω , ( D v ) v ∈ U ) (\omega,(D_v)_{v\in\mathbb{U}}) ( ω , ( D v ) v ∈ U ) satisfies the conclusions of Lemma 5.13 with π = π E \pi=\pi_E π = π E , every D v D_v D v is a measurable function of ω \omega ω , and almost surely D v = − 1 / Γ v ( E + i 0 ) D_v=-1/\Gamma_v(E+i0) D v = − 1/ Γ v ( E + i 0 ) for every v ∈ V v\in V v ∈ V . In the terminology of [17], this is an endogenous recursive tree process with marginal π E \pi_E π E . In particular, when the real solution in Lemma 5.13 is the real boundary law, the process may be taken to be this one.
To verify the claim, note that D \mathcal{D} D takes values in R ∖ { 0 } \mathbb{R}\setminus\{0\} R ∖ { 0 } by its definition, and that the last assertion is Step 5 of the proof of Proposition 5.11. Let A A A be the event that D ( ω ) = E − p − 1 ∑ k ≤ N v γ k / D ( θ k ω ) \mathcal{D}(\omega)=E-p^{-1}\sum_{k\leq N_v}\gamma_k/\mathcal{D}(\theta_k\omega) D ( ω ) = E − p − 1 ∑ k ≤ N v γ k / D ( θ k ω ) ; it has probability one by Step 5 of that proof. Since θ v k = θ k ∘ θ v \theta_{v k}=\theta_k\circ\theta_v θ v k = θ k ∘ θ v , and since the offspring number and the root edge marks of θ v ω \theta_v\omega θ v ω are N v N_v N v and w v k w_{v k} w v k , k ≥ 1 k\geq1 k ≥ 1 , the recursion of Lemma 5.13(i) at the word v v v is the event θ v − 1 A \theta_v^{-1}A θ v − 1 A . This event has probability one because θ v \theta_v θ v preserves P \mathbb{P} P , and D v D_v D v has the law of D o D_o D o , which is π E \pi_E π E . This proves (i). For (ii) and (iii), we note that the pair ( θ v i ω , ( D v i u ) u ∈ U ) (\theta_{v_i}\omega,(D_{v_i u})_{u\in\mathbb{U}}) ( θ v i ω , ( D v i u ) u ∈ U ) is a measurable function of θ v i ω \theta_{v_i}\omega θ v i ω , since D v i u = D ( θ u ( θ v i ω ) ) D_{v_i u}=\mathcal{D}(\theta_u(\theta_{v_i}\omega)) D v i u = D ( θ u ( θ v i ω )) . The configurations θ v 1 ω , … , θ v k ω \theta_{v_1}\omega,\ldots,\theta_{v_k}\omega θ v 1 ω , … , θ v k ω and the generators of F ( v 1 , … , v k ) \mathcal{F}(v_1,\ldots,v_k) F ( v 1 , … , v k ) are functions of pairwise disjoint sets of the independent coordinates of ω \omega ω , and each θ v i ω \theta_{v_i}\omega θ v i ω has law P \mathbb{P} P ; this is the argument of the proof of Lemma 4.2(i). This gives (iii), and (ii) follows in the same way from Lemma 4.2 (i).
Fractional moments at positive imaginary part
The following lemma integrates one step of the cavity recursion, at a single vertex. It extends Lemma 4.18 to products of several negative powers sharing the same fresh variables, with additional weights given by powers of these variables. We write C r , d = ( 1 − r ) − 1 ( 2 / d ) r C_{r,d}=(1-r)^{-1}(2/d)^r C r , d = ( 1 − r ) − 1 ( 2/ d ) r for the constant of Lemma 4.18 .
Lemma 5.16. Let r ∈ ( 0 , 1 ) r\in(0,1) r ∈ ( 0 , 1 ) , d > 0 d>0 d > 0 , b ≥ 0 b\geq0 b ≥ 0 and k ∈ N 0 k\in\mathbb{N}_0 k ∈ N 0 , and set c b = max { 1 , ( 2 b / e ) b } c_b=\max\{1,(2b/e)^b\} c b = max { 1 , ( 2 b / e ) b } . Let n ≥ 0 n\geq0 n ≥ 0 , let Y 1 , … , Y n Y_1,\ldots,Y_n Y 1 , … , Y n be independent Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) variables, let c ∈ C c\in\mathbb{C} c ∈ C with ∣ c ∣ ≥ d |c|\geq d ∣ c ∣ ≥ d and γ 1 , … , γ n ∈ C \gamma_1,\ldots,\gamma_n\in\mathbb{C} γ 1 , … , γ n ∈ C , let J 1 , … , J l ⊂ { 1 , … , n } J_1,\ldots,J_l\subset\{1,\ldots,n\} J 1 , … , J l ⊂ { 1 , … , n } , let a 1 , … , a l ≥ 0 a_1,\ldots,a_l\geq0 a 1 , … , a l ≥ 0 with a 1 + ⋯ + a l ≤ r a_1+\cdots+a_l\leq r a 1 + ⋯ + a l ≤ r , and let K ⊂ { 1 , … , n } K\subset\{1,\ldots,n\} K ⊂ { 1 , … , n } with ∣ K ∣ ≤ k |K|\leq k ∣ K ∣ ≤ k and b j ∈ [ 0 , b ] b_j\in[0,b] b j ∈ [ 0 , b ] for j ∈ K j\in K j ∈ K . Then
E [ ∏ i = 1 l ∣ c + ∑ j ∈ J i γ j Y j ∣ − a i ∏ j ∈ K Y j b j ] ≤ ( 2 c b ) k ( 1 + C r , d ( n + 1 ) r / 2 ) . \mathbb{E}\left[\prod_{i=1}^{l}\left|c+\sum_{j\in J_i}\gamma_jY_j\right|^{-a_i}\prod_{j\in K}Y_j^{b_j}\right]\leq(2c_b)^k\left(1+C_{r,d}(n+1)^{r/2}\right). E i = 1 ∏ l c + j ∈ J i ∑ γ j Y j − a i j ∈ K ∏ Y j b j ≤ ( 2 c b ) k ( 1 + C r , d ( n + 1 ) r /2 ) . The same bound holds for the conditional expectation given a σ \sigma σ -field G \mathcal{G} G if Y 1 , … , Y n Y_1,\ldots,Y_n Y 1 , … , Y n are fresh given G \mathcal{G} G (Definition 4.16), and n n n , c c c , the γ j \gamma_j γ j , the J i J_i J i , the a i a_i a i , K K K and the b j b_j b j are G \mathcal{G} G -measurable, with ∣ c ∣ ≥ d |c|\geq d ∣ c ∣ ≥ d almost surely.
Proof. For j ∈ K j\in K j ∈ K and y > 0 y>0 y > 0 , we have y b j e − y ≤ c b e − y / 2 y^{b_j}e^{-y}\leq c_be^{-y/2} y b j e − y ≤ c b e − y /2 , since the maximum of y b j e − y / 2 y^{b_j}e^{-y/2} y b j e − y /2 is ( 2 b j / e ) b j ≤ c b (2b_j/e)^{b_j}\leq c_b ( 2 b j / e ) b j ≤ c b . For every measurable Φ ≥ 0 \Phi\geq0 Φ ≥ 0 , this gives
E [ Y j b j Φ ( Y j ) ] ≤ c b ∫ 0 ∞ e − y / 2 Φ ( y ) d y = 2 c b E Φ ( 2 Y j ) . \mathbb{E}\left[Y_j^{b_j}\Phi(Y_j)\right]\leq c_b\int_0^\infty e^{-y/2}\Phi(y)\,\mathrm{d}y=2c_b\mathbb{E}\Phi(2Y_j). E [ Y j b j Φ ( Y j ) ] ≤ c b ∫ 0 ∞ e − y /2 Φ ( y ) d y = 2 c b E Φ ( 2 Y j ) .
Applying this to each j ∈ K j\in K j ∈ K in turn, using the independence of the Y j Y_j Y j and Tonelli’s theorem, we find that the left side of the claimed bound is at most ( 2 c b ) ∣ K ∣ (2c_b)^{|K|} ( 2 c b ) ∣ K ∣ times E ∏ i ∣ X i ∣ − a i \mathbb{E}\prod_i|X_i|^{-a_i} E ∏ i ∣ X i ∣ − a i , where X i = c + ∑ j ∈ J i γ j ′ Y j X_i=c+\sum_{j\in J_i}\gamma'_jY_j X i = c + ∑ j ∈ J i γ j ′ Y j with γ j ′ = 2 γ j \gamma'_j=2\gamma_j γ j ′ = 2 γ j for j ∈ K j\in K j ∈ K and γ j ′ = γ j \gamma'_j=\gamma_j γ j ′ = γ j otherwise. Let A = a 1 + ⋯ + a l A=a_1+\cdots+a_l A = a 1 + ⋯ + a l . If A = 0 A=0 A = 0 , there is nothing more to prove. Otherwise, Hölder’s inequality with the exponents A / a i A/a_i A / a i gives
E ∏ i = 1 l ∣ X i ∣ − a i ≤ ∏ i = 1 l ( E ∣ X i ∣ − A ) a i / A ≤ max i E ∣ X i ∣ − A . \mathbb{E}\prod_{i=1}^{l}|X_i|^{-a_i}\leq\prod_{i=1}^{l}\left(\mathbb{E}|X_i|^{-A}\right)^{a_i/A}\leq\max_i\mathbb{E}|X_i|^{-A}. E i = 1 ∏ l ∣ X i ∣ − a i ≤ i = 1 ∏ l ( E ∣ X i ∣ − A ) a i / A ≤ i max E ∣ X i ∣ − A .
Since ∣ x ∣ − A ≤ 1 + ∣ x ∣ − r |x|^{-A}\leq1+|x|^{-r} ∣ x ∣ − A ≤ 1 + ∣ x ∣ − r for 0 < A ≤ r 0<A\leq r 0 < A ≤ r , Lemma 4.18 (ii) with m = ∣ J i ∣ ≤ n m=|J_i|\leq n m = ∣ J i ∣ ≤ n gives E ∣ X i ∣ − A ≤ 1 + C r , d ( n + 1 ) r / 2 \mathbb{E}|X_i|^{-A}\leq1+C_{r,d}(n+1)^{r/2} E ∣ X i ∣ − A ≤ 1 + C r , d ( n + 1 ) r /2 . For the conditional version, we use the substitution identity (4.14) (consequence (c) after Definition 4.16), with X X X a finite-dimensional vector encoding the G \mathcal{G} G -measurable parameters c c c , γ j \gamma_j γ j , J i J_i J i , a i a_i a i , K K K and b j b_j b j . On each event { n = n 0 } \{n=n_0\} { n = n 0 } these are finitely many numbers and finite sets of indices, so such an encoding exists. The identity reduces the claim on this event to the unconditional bound for fixed values of the revealed variables. ∎
For v ∈ V v \in V v ∈ V and z ∈ C ∖ R z \in\mathbb{C} \setminus\mathbb{R} z ∈ C ∖ R , we write D v ( z ) = − 1 / Γ v ( z ) D_v(z)=-1/\Gamma_v(z) D v ( z ) = − 1/ Γ v ( z ) for the denominator at z z z . By Proposition 3.16,
D v ( z ) = z + 1 p ∑ k = 1 N v γ v k Γ v k ( z ) . (68) D_v(z)=z+\frac{1}{p}\sum_{k=1}^{N_v}\gamma_{vk}\Gamma_{vk}(z). \tag*{(68)} D v ( z ) = z + p 1 k = 1 ∑ N v γ v k Γ v k ( z ) . (68)
Lemma 5.17. Let r ∈ ( 0 , 1 ) r \in(0,1) r ∈ ( 0 , 1 ) and d > 0 d>0 d > 0 . Then the following hold for every z ∈ C ∖ R z \in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R with ∣ z ∣ ≥ d |z|\ge d ∣ z ∣ ≥ d .
(i) E ∣ Γ o ( z ) ∣ r ≤ C r , d ( p + 1 ) r / 2 \mathbb{E}|\Gamma_o(z)|^r \le C_{r,d}(p+1)^{r/2} E ∣ Γ o ( z ) ∣ r ≤ C r , d ( p + 1 ) r /2 and E ∞ ∣ Γ o ( z ) ∣ r ≤ ϱ − 1 C r , d ( p + 1 ) r / 2 \mathbb{E}^{\infty}|\Gamma_o(z)|^r \le\varrho^{-1}C_{r,d}(p+1)^{r/2} E ∞ ∣ Γ o ( z ) ∣ r ≤ ϱ − 1 C r , d ( p + 1 ) r /2 .
(ii) Let J J J be a random set of children of the root that is a function of the offspring numbers and the auxiliary marks, and let
Γ o J ( z ) = − ( z + 1 p ∑ k ∈ J γ k Γ k ( z ) ) − 1 . \Gamma_o^J(z)=-\left(z+\frac{1}{p}\sum_{k\in J}\gamma_k\Gamma_k(z)\right)^{-1}. Γ o J ( z ) = − ( z + p 1 k ∈ J ∑ γ k Γ k ( z ) ) − 1 . be the diagonal entry at the root of the resolvent after the edges from the root to the children outside J J J are deleted. Then E ∣ Γ o J ( z ) ∣ r ≤ C r , d ( p + 1 ) r / 2 \mathbb{E}|\Gamma_o^J(z)|^r\le C_{r,d}(p+1)^{r/2} E ∣ Γ o J ( z ) ∣ r ≤ C r , d ( p + 1 ) r /2 . If Q Q Q is a probability measure on Ω \Omega Ω with Q ≤ K P Q\le K\mathbb{P} Q ≤ K P for a constant K K K , then all bounds of (i) and (ii) hold under Q Q Q with an extra factor K K K .
(iii) For q ∈ ( 0 , 1 ) q\in(0,1) q ∈ ( 0 , 1 ) and g > 0 g>0 g > 0 ,
E ∣ D o ( z ) ∣ q ≤ ∣ z ∣ q + p 1 − q Γ ( 1 + q ) C q , d ( p + 1 ) q / 2 , E ∣ g D o ( z ) ∣ − q ≤ g − q C q , d ( p + 1 ) q / 2 . \mathbb{E}|D_o(z)|^q\le|z|^q+p^{1-q}\Gamma(1+q)C_{q,d}(p+1)^{q/2},\qquad\mathbb{E}|gD_o(z)|^{-q}\le g^{-q}C_{q,d}(p+1)^{q/2}. E ∣ D o ( z ) ∣ q ≤ ∣ z ∣ q + p 1 − q Γ ( 1 + q ) C q , d ( p + 1 ) q /2 , E ∣ g D o ( z ) ∣ − q ≤ g − q C q , d ( p + 1 ) q /2 . (iv) Let E > 2 E>2 E > 2 and X η = − Γ o ( E + i η ) / g ( E ) X_\eta=-\Gamma_o(E+i\eta)/g(E) X η = − Γ o ( E + i η ) / g ( E ) . Then sup η > 0 E ∣ X η ∣ r ≤ C r , 1 ( p + 1 ) r / 2 \sup_{\eta>0}\mathbb{E}|X_\eta|^r\le C_{r,1}(p+1)^{r/2} sup η > 0 E ∣ X η ∣ r ≤ C r , 1 ( p + 1 ) r /2 .
In particular, for every compact interval I ⊂ ( 0 , ∞ ) I\subset(0,\infty) I ⊂ ( 0 , ∞ ) and r ∈ ( 0 , 1 ) r\in(0,1) r ∈ ( 0 , 1 ) , there exists C = C ( p , I , r ) C=C(p,I,r) C = C ( p , I , r ) such that
sup E ∈ I , 0 < η ≤ 1 ( E ∣ Γ o ( E + i η ) ∣ r + E ∞ ∣ Γ o ( E + i η ) ∣ r + E ∣ D o ( E + i η ) ∣ r + E ∣ D o ( E + i η ) ∣ − r ) ≤ C . \sup_{\substack{E\in I,\,0<\eta\le1}}\left(\mathbb{E}|\Gamma_o(E+i\eta)|^r+\mathbb{E}^{\infty}|\Gamma_o(E+i\eta)|^r+\mathbb{E}|D_o(E+i\eta)|^r+\mathbb{E}|D_o(E+i\eta)|^{-r}\right)\le C. E ∈ I , 0 < η ≤ 1 sup ( E ∣ Γ o ( E + i η ) ∣ r + E ∞ ∣ Γ o ( E + i η ) ∣ r + E ∣ D o ( E + i η ) ∣ r + E ∣ D o ( E + i η ) ∣ − r ) ≤ C .
Proof. Part (i) is Corollary 4.20, since Γ o = G o o \Gamma_o=G_{oo} Γ o = G oo and ∣ z ∣ ≥ d |z|\ge d ∣ z ∣ ≥ d . For (ii), let G \mathcal{G} G be the σ \sigma σ -field generated by N o N_o N o , the auxiliary marks, the configurations θ k ω \theta_k\omega θ k ω , k ≥ 1 k\ge1 k ≥ 1 , and the phases of the root edge marks. As shown after Lemma 4.17, the squared moduli Y 1 , … , Y N o Y_1,\ldots,Y_{N_o} Y 1 , … , Y N o are fresh given G \mathcal{G} G , and the set J J J and the coefficients 1 { k ∈ J } Γ k ( z ) / p \mathbf{1}_{\{k\in J\}}\Gamma_k(z)/p 1 { k ∈ J } Γ k ( z ) / p are G \mathcal{G} G -measurable. Lemma 4.18(iv), with constant z z z and these coefficients, gives E [ ∣ Γ o J ( z ) ∣ r ∣ G ] ≤ C r , d ( N o + 1 ) r / 2 \mathbb{E}[|\Gamma_o^J(z)|^r\mid\mathcal{G}]\le C_{r,d}(N_o+1)^{r/2} E [ ∣ Γ o J ( z ) ∣ r ∣ G ] ≤ C r , d ( N o + 1 ) r /2 , and taking expectations and using Jensen’s inequality, we find the bound. For Q ≤ K P Q\le K\mathbb{P} Q ≤ K P , we have E Q Z ≤ K E Z \mathbb{E}_Q Z\le K\mathbb{E}Z E Q Z ≤ K E Z for every Z ≥ 0 Z\ge0 Z ≥ 0 .
For (iii), we use (5.10) and ∣ a + b ∣ q ≤ ∣ a ∣ q + ∣ b ∣ q |a+b|^q\le|a|^q+|b|^q ∣ a + b ∣ q ≤ ∣ a ∣ q + ∣ b ∣ q to obtain
E ∣ D o ( z ) ∣ q ≤ ∣ z ∣ q + p − q E ∑ k = 1 N o γ k q ∣ Γ k ( z ) ∣ q = ∣ z ∣ q + p 1 − q Γ ( 1 + q ) E ∣ Γ o ( z ) ∣ q , \mathbb{E}|D_o(z)|^q\le|z|^q+p^{-q}\mathbb{E}\sum_{k=1}^{N_o}\gamma_k^q|\Gamma_k(z)|^q=|z|^q+p^{1-q}\Gamma(1+q)\mathbb{E}|\Gamma_o(z)|^q, E ∣ D o ( z ) ∣ q ≤ ∣ z ∣ q + p − q E k = 1 ∑ N o γ k q ∣ Γ k ( z ) ∣ q = ∣ z ∣ q + p 1 − q Γ ( 1 + q ) E ∣ Γ o ( z ) ∣ q ,
where we used that N o N_o N o , the Y k Y_k Y k and the Γ k ( z ) = G ∞ ( z ) ∘ θ k \Gamma_k(z)=G_\infty(z)\circ\theta_k Γ k ( z ) = G ∞ ( z ) ∘ θ k are independent and that the Γ k ( z ) \Gamma_k(z) Γ k ( z ) have the law of Γ o ( z ) \Gamma_o(z) Γ o ( z ) , by Lemma 4.2(ii). Combined with (i), this gives the first bound in (iii). The second bound is (i) with r = q r=q r = q , since ∣ g D o ( z ) ∣ − q = g − q ∣ Γ o ( z ) ∣ q |gD_o(z)|^{-q}=g^{-q}|\Gamma_o(z)|^q ∣ g D o ( z ) ∣ − q = g − q ∣ Γ o ( z ) ∣ q .
For (iv), let z = E + i η z=E+i\eta z = E + i η and X k = − Γ k ( z ) / g ( E ) X_k=-\Gamma_k(z)/g(E) X k = − Γ k ( z ) / g ( E ) ; then X η = 1 / ( g ( E ) D o ( z ) ) X_\eta=1/(g(E)D_o(z)) X η = 1/ ( g ( E ) D o ( z )) . We multiply (5.10) at the root by g ( E ) g(E) g ( E ) and use g ( E ) E = 1 + θ g(E)E=1+\theta g ( E ) E = 1 + θ from Lemma 5.3 and g ( E ) Γ k ( z ) = − θ X k g(E)\Gamma_k(z)=-\theta X_k g ( E ) Γ k ( z ) = − θ X k . We obtain
X η = ( 1 + θ + i g ( E ) η − θ p ∑ k = 1 N o Y k X k ) − 1 . X_\eta=\left(1+\theta+ig(E)\eta-\frac{\theta}{p}\sum_{k=1}^{N_o}Y_kX_k\right)^{-1}. X η = ( 1 + θ + i g ( E ) η − p θ k = 1 ∑ N o Y k X k ) − 1 .
The constant term has modulus at least 1 + θ ≥ 1 1+\theta\geq1 1 + θ ≥ 1 , and the argument of (ii) with d = 1 d=1 d = 1 applies.
The final assertion follows from (i) and (iii), with d = min I d=\min I d = min I and ∣ z ∣ ≤ max I + 1 |z|\leq\max I+1 ∣ z ∣ ≤ max I + 1 .
The bounds of Lemma 5.17(ii) also apply when structural decorations are prescribed at the root. For example, under the law conditioned on the event that the root has exactly one child, which has P \mathbb P P -probability p e − p pe^{-p} p e − p , all bounds hold with the factor K = e p / p K=e^p/p K = e p / p . If a vertex is required to have no children other than a selected one, then its denominator without the selected child is the deterministic number z z z .
Fixed-depth convergence
In Sections 9 and 10, the resonance argument compares, at a fixed depth, expectations of functionals of the cavity variables at E + i η E+i\eta E + i η , η > 0 \eta>0 η > 0 , with the corresponding expectations for real denominators, and then lets η \eta η tend to zero with the depth fixed. This section provides the convergence statement for this step. The functionals are sums over one or two vertices of a fixed generation, which the many-to-one and many-to-two formulas of Proposition 4.8 express through the spine laws. We first record the properties of the spine laws that we use, then define the depth-m m m data and the real recursive tree process, and finally prove the convergence.
Spine laws
We use the spine laws of Definition 4.6. These are the one-spine laws Q L Q_L Q L and Q L ∞ Q_L^\infty Q L ∞ , with canonical random elements ( ω , v ) (\omega,v) ( ω , v ) , and the two-spine laws Q L , j Q_{L,j} Q L , j and Q L , j ∞ Q_{L,j}^\infty Q L , j ∞ , with canonical random elements ( ω , u , v ) (\omega,u,v) ( ω , u , v ) ; here L ≥ 0 L\geq0 L ≥ 0 and 0 ≤ j < L 0\leq j<L 0 ≤ j < L . We call L L L the length of the spine law. For L = 0 L=0 L = 0 , the law Q 0 Q_0 Q 0 is the law of ( ω , o ) (\omega,o) ( ω , o ) with ω ∼ P \omega\sim\mathbb P ω ∼ P , and Q 0 ∞ Q_0^\infty Q 0 ∞ is the law of ( ω , o ) (\omega,o) ( ω , o ) with ω ∼ P ∞ \omega\sim\mathbb P^\infty ω ∼ P ∞ ; in this way P \mathbb P P and P ∞ \mathbb P^\infty P ∞ are spine laws of length zero. The spine vertices are the prefixes v l v_l v l , 0 ≤ l ≤ L 0\leq l\leq L 0 ≤ l ≤ L , of v v v , together with the prefixes u l u_l u l , j < l ≤ L j<l\leq L j < l ≤ L , of u u u for a two-spine law. The endpoints are v v v , and u u u for a two-spine law; the endpoint law is P \mathbb P P for Q L Q_L Q L and Q L , j Q_{L,j} Q L , j , and P ∞ \mathbb P^\infty P ∞ for Q L ∞ Q_L^\infty Q L ∞ and Q L , j ∞ Q_{L,j}^\infty Q L , j ∞ . For a spine vertex x x x which is not an endpoint, the spine children of x x x are its children that are spine vertices (two at the fork v j v_j v j of a two-spine law, one otherwise), and its other children are its side children .
Lemma 5.18. Let Q \mathsf{Q} Q be a spine law of length L L L .
(i) If A ∈ F A \in\mathcal{F} A ∈ F satisfies P ( A ) = 1 \mathbb{P}(A) = 1 P ( A ) = 1 , then Q \mathsf{Q} Q -almost surely, θ u ω ∈ A \theta_u\omega\in A θ u ω ∈ A for every u ∈ U u \in\mathbb{U} u ∈ U .
(ii) Let A \mathcal{A} A be a random finite subset of U ∗ \mathbb{U}_* U ∗ which is a function of the offspring numbers, the auxiliary marks, and the distinguished vertices. Then, under Q \mathsf{Q} Q , the squared moduli Y u = ∣ w u ∣ 2 Y_u = |w_u|^2 Y u = ∣ w u ∣ 2 , u ∈ A u \in\mathcal{A} u ∈ A , listed in a fixed order of U \mathbb{U} U , are fresh given the σ \sigma σ -field generated by the offspring numbers, the auxiliary marks, the distinguished vertices, the phases w u / ∣ w u ∣ w_u/|w_u| w u /∣ w u ∣ of all marks, and the marks w u w_u w u with u ∉ A u \notin\mathcal{A} u ∈ / A .
(iii) Let L ≥ 1 L \ge1 L ≥ 1 . If Q \mathsf{Q} Q is Q L \mathsf{Q}_L Q L or Q L ∞ \mathsf{Q}^{\infty}_L Q L ∞ , write v = I 1 v ′ v = I_1v' v = I 1 v ′ . Then N o − 1 N_o - 1 N o − 1 has law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) , and the pair ( θ I 1 ω , v ′ ) (\theta_{I_1}\omega,v') ( θ I 1 ω , v ′ ) is independent of N o N_o N o and has law Q L − 1 \mathsf{Q}_{L-1} Q L − 1 , respectively Q L − 1 ∞ \mathsf{Q}^{\infty}_{L-1} Q L − 1 ∞ . If Q \mathsf{Q} Q is Q L , j \mathsf{Q}_{L,j} Q L , j or Q L , j ∞ \mathsf{Q}^{\infty}_{L,j} Q L , j ∞ with j ≥ 1 j \ge1 j ≥ 1 , the same holds for the triple ( θ I 1 ω , u ′ , v ′ ) (\theta_{I_1}\omega,u',v') ( θ I 1 ω , u ′ , v ′ ) with u = I 1 u ′ u = I_1u' u = I 1 u ′ and v = I 1 v ′ v = I_1v' v = I 1 v ′ , whose law is Q L − 1 , j − 1 \mathsf{Q}_{L-1,j-1} Q L − 1 , j − 1 respectively Q L − 1 , j − 1 ∞ \mathsf{Q}^{\infty}_{L-1,j-1} Q L − 1 , j − 1 ∞ . If j = 0 j = 0 j = 0 , write u = J 1 u ′ u = J_1u' u = J 1 u ′ and v = J 2 v ′ v = J_2v' v = J 2 v ′ . Then N o − 2 N_o - 2 N o − 2 has law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) , and the pairs ( θ J 1 ω , u ′ ) (\theta_{J_1}\omega,u') ( θ J 1 ω , u ′ ) and ( θ J 2 ω , v ′ ) (\theta_{J_2}\omega,v') ( θ J 2 ω , v ′ ) are independent of each other and of N o N_o N o , each with law Q L − 1 \mathsf{Q}_{L-1} Q L − 1 , respectively Q L − 1 ∞ \mathsf{Q}^{\infty}_{L-1} Q L − 1 ∞ .
Proof. Part (i) is Corollary 4.10 . Part (ii) follows from Lemma 4.17(ii) and (iii) (and from Lemma 4.17(i) and (iii) for L = 0 L = 0 L = 0 ). Indeed, for fixed values of the distinguished vertices, the event { A = a } \{ \mathcal{A} = a \} { A = a } is determined by the offspring numbers and the auxiliary marks; it belongs to the σ \sigma σ -field G a \mathcal{G}_a G a of that lemma, and the σ \sigma σ -field H A \mathcal{H}_{\mathcal{A}} H A of Lemma 4.17(iii) is generated by the variables listed in (ii).
For (iii), we use the construction of Q L \mathsf{Q}_L Q L in Definition 4.6 , with ω ′ \omega' ω ′ , N 0 ′ , … , N L − 1 ′ N'_0,\ldots,N'_{L-1} N 0 ′ , … , N L − 1 ′ and I 1 , … , I L I_1,\ldots,I_L I 1 , … , I L . Set ω ′ ′ = θ I 1 ω ′ \omega'' = \theta_{I_1}\omega' ω ′′ = θ I 1 ω ′ . Since ω ′ \omega' ω ′ has law P \mathbb{P} P and is independent of ( N ′ , I ) (N',I) ( N ′ , I ) , and since every shift preserves P \mathbb{P} P , the configuration ω ′ ′ \omega'' ω ′′ has law P \mathbb{P} P conditionally on ( N ′ , I ) (N',I) ( N ′ , I ) ; that is, ω ′ ′ \omega'' ω ′′ has law P \mathbb{P} P and is independent of ( N ′ , I ) (N',I) ( N ′ , I ) . The label v l = I 1 ⋯ I l v_l = I_1\cdots I_l v l = I 1 ⋯ I l of ω \omega ω corresponds to the label I 2 ⋯ I l I_2\cdots I_l I 2 ⋯ I l of θ I 1 ω \theta_{I_1}\omega θ I 1 ω , and θ I 1 ω \theta_{I_1}\omega θ I 1 ω is obtained from ω ′ ′ \omega'' ω ′′ by setting N I 2 ⋯ I l = N l ′ + 1 N_{I_2\cdots I_l} = N'_{l} + 1 N I 2 ⋯ I l = N l ′ + 1 for 1 ≤ l < L 1 \le l < L 1 ≤ l < L , all other coordinates being unchanged. This is the construction of Q L − 1 \mathsf{Q}_{L-1} Q L − 1 from ω ′ ′ \omega'' ω ′′ , ( N 1 ′ , … , N L − 1 ′ ) (N'_1,\ldots,N'_{L-1}) ( N 1 ′ , … , N L − 1 ′ ) and ( I 2 , … , I L ) (I_2,\ldots,I_L) ( I 2 , … , I L ) . These inputs have the joint law required in Definition 4.6 , and they are independent of N 0 ′ = N o − 1 N'_0 = N_o - 1 N 0 ′ = N o − 1 , because the N l ′ N'_l N l ′ are independent and I l + 1 I_{l+1} I l + 1 depends only on N l ′ N'_l N l ′ and on independent randomness. This proves the claim for Q L \mathsf{Q}_L Q L . Under Q L ∞ \mathsf{Q}^{\infty}_L Q L ∞ , we condition on { θ v ω ∈ S } = { θ v ′ ( θ I 1 ω ) ∈ S } \{\theta_v\omega\in\mathcal{S}\} = \{\theta_{v'}(\theta_{I_1}\omega) \in\mathcal{S}\} { θ v ω ∈ S } = { θ v ′ ( θ I 1 ω ) ∈ S } , which is an event of the pair ( θ I 1 ω , v ′ ) (\theta_{I_1}\omega,v') ( θ I 1 ω , v ′ ) ; since this pair is independent of N o N_o N o , conditioning on the event keeps the independence and replaces Q L − 1 \mathsf{Q}_{L-1} Q L − 1 by Q L − 1 ∞ \mathsf{Q}^{\infty}_{L-1} Q L − 1 ∞ . The two-spine laws with j ≥ 1 j \ge1 j ≥ 1 are treated in the same way. For j = 0 j = 0 j = 0 , the fork is the root, and N o = N 0 ′ + 2 N_o = N'_0 + 2 N o = N 0 ′ + 2 . The configurations θ J 1 ω ′ \theta_{J_1}\omega' θ J 1 ω ′ and θ J 2 ω ′ \theta_{J_2}\omega' θ J 2 ω ′ are independent with law P \mathbb{P} P , and independent of ( N ′ , J , I ) (N',J,I) ( N ′ , J , I ) , because J 1 ≠ J 2 J_1 \ne J_2 J 1 = J 2 and the two configurations use disjoint sets of coordinates of ω ′ \omega' ω ′ . Each of ( θ J 1 ω , u ′ ) (\theta_{J_1}\omega,u') ( θ J 1 ω , u ′ ) and ( θ J 2 ω , v ′ ) (\theta_{J_2}\omega,v') ( θ J 2 ω , v ′ ) is obtained by the construction of Q L − 1 \mathsf{Q}_{L-1} Q L − 1 from one of these configurations and from the variables N l ( i ) N_l^{(i)} N l ( i ) and I l ( i ) I_l^{(i)} I l ( i ) of the corresponding arm, and the two sets of inputs are independent of each other and of N 0 ′ N'_0 N 0 ′ . Under Q L , 0 ∞ Q^\infty_{L,0} Q L , 0 ∞ we condition on the intersection of an event of the first pair and an event of the second pair, which preserves the independence.
Depth-m m m data and side denominators
For m ≥ 0 m \ge0 m ≥ 0 , let K m \mathcal{K}_m K m be the set of κ = ( Δ , x , y , k ) ∈ K \kappa=(\Delta,x,y,k)\in\mathcal{K} κ = ( Δ , x , y , k ) ∈ K such that x x x , y y y and the words in Δ V ∪ Δ E \Delta_V\cup\Delta_E Δ V ∪ Δ E have length at most m m m , and let Z = [ 0 , 1 ] N Z=[0,1]^{\mathbb{N}} Z = [ 0 , 1 ] N be the space of auxiliary marks.
Definition 5.19. Let m ≥ 0 m \ge0 m ≥ 0 . The space of depth-m m m data is the Polish space
X m = N 0 U < m × Z U < m × C U ≤ m ∖ { o } × { 0 , 1 } U ≤ m × C K m , \mathcal{X}_m=\mathbb{N}_0^{\mathbb{U}_{<m}}\times Z^{\mathbb{U}_{<m}}\times\mathbb{C}^{\mathbb{U}_{\le m}\setminus\{o\}}\times\{0,1\}^{\mathbb{U}_{\le m}}\times\mathbb{C}^{\mathcal{K}_m}, X m = N 0 U < m × Z U < m × C U ≤ m ∖ { o } × { 0 , 1 } U ≤ m × C K m ,
with the product topology, where U < m = U ≤ m − 1 \mathbb{U}_{<m}=\mathbb{U}_{\le m-1} U < m = U ≤ m − 1 . For z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R , the variables of the first m m m generations, which we call the depth-m m m data , form the element of X m \mathcal{X}_m X m
L m ( z ) = ( ( N v ) ∣ v ∣ < m , ( ζ v ) ∣ v ∣ < m , ( w v ) 1 ≤ ∣ v ∣ ≤ m , ( ς v ) ∣ v ∣ ≤ m , ( F κ ( z ) ) κ ∈ K m ) , \mathcal{L}_m(z)=\left((N_v)_{|v|<m},(\zeta_v)_{|v|<m},(w_v)_{1\le|v|\le m},(\varsigma_v)_{|v|\le m},(F_\kappa(z))_{\kappa\in\mathcal{K}_m}\right), L m ( z ) = ( ( N v ) ∣ v ∣ < m , ( ζ v ) ∣ v ∣ < m , ( w v ) 1 ≤ ∣ v ∣ ≤ m , ( ς v ) ∣ v ∣ ≤ m , ( F κ ( z ) ) κ ∈ K m ) ,
where ς v = 0 \varsigma_v=0 ς v = 0 if v ∉ V v\notin V v ∈ / V . If E ∈ E p E\in\mathcal{E}_p E ∈ E p , the depth-m m m data L m ( E ) \mathcal{L}_m(E) L m ( E ) is defined in the same way, with F κ ( z ) F_\kappa(z) F κ ( z ) replaced by F κ ( E + i 0 ) F_\kappa(E+\mathrm{i}0) F κ ( E + i 0 ) .
The first three components determine the marked ball of radius m m m around the root and the auxiliary marks at the vertices of generation less than m m m ; they generate the σ \sigma σ -field F m \mathcal{F}_m F m . By (63) , the last component contains every resolvent entry G x y Δ ( z ) G^\Delta_{xy}(z) G x y Δ ( z ) with x x x , y y y and Δ \Delta Δ inside the ball, in particular the forward cavities Γ x ( z ) \Gamma_x(z) Γ x ( z ) for ∣ x ∣ ≤ m |x|\le m ∣ x ∣ ≤ m , the cavity resolvents Γ x → w C ( z ) \Gamma^C_{x\to w}(z) Γ x → w C ( z ) , and the path entries G o x Δ ( z ) G^\Delta_{ox}(z) G o x Δ ( z ) . Functionals of the depth-m m m data therefore include the quantities that the resonance argument uses at a fixed depth: the marks and the genealogy near one or two spines, side denominators, selections by auxiliary ranks, survival indicators, and resolvent entries of the forests obtained by removing finitely many vertices and edges near the spines.
For x ∈ V x\in V x ∈ V , a set J J J of children of x x x , and z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R , we let
Γ x J ( z ) = − ( z + 1 p ∑ y ∈ J Γ y Γ y ( z ) ) − 1 (69) \Gamma^J_x(z)=-\left(z+\frac{1}{p}\sum_{y\in J}\Gamma_y\Gamma_y(z)\right)^{-1} \tag*{(69)} Γ x J ( z ) = − z + p 1 y ∈ J ∑ Γ y Γ y ( z ) − 1 (69)
be the diagonal entry at x x x of the resolvent of the forward subtree T x \mathcal{T}_x T x after the edges from x x x to its children outside J J J are deleted; the identity is Proposition 3.16(ii) for this finite removal. If ∣ x ∣ < m |x|<m ∣ x ∣ < m , then Γ x J ( z ) \Gamma^J_x(z) Γ x J ( z ) is one of the resolvent entries contained in L m ( z ) \mathcal{L}_m(z) L m ( z ) , selected by a function of x x x , J J J and N x N_x N x . For a spine vertex x x x which is not an endpoint, the side denominator at x x x is
D x s i d e ( z ) = z + 1 p ∑ y s i d e c h i l d o f x Γ y Γ y ( z ) = − 1 Γ x J ( z ) , J = { side children of x } , D^{\mathrm{side}}_x(z)=z+\frac{1}{p}\sum_{y\ \mathrm{side\ child\ of}\ x}\Gamma_y\Gamma_y(z)=-\frac{1}{\Gamma^J_x(z)},\qquad J=\{\text{side children of }x\}, D x side ( z ) = z + p 1 y side child of x ∑ Γ y Γ y ( z ) = − Γ x J ( z ) 1 , J = { side children of x } ,
so that, by (5.10),
D x ( z ) = D x s i d e ( z ) + 1 p ∑ y s p i n e c h i l d o f x γ y Γ y ( z ) . (70) D_x(z)=D_x^{\mathrm{side}}(z)+\frac{1}{p}\sum_{y\ \mathrm{spine\ child\ of}\ x}\gamma_y\Gamma_y(z). \tag*{(70)} D x ( z ) = D x side ( z ) + p 1 y spine child of x ∑ γ y Γ y ( z ) . (70)
For a spine law of length L L L and m ≥ L + 1 m\ge L+1 m ≥ L + 1 , the side denominators are functions of the depth-m m m data and of the distinguished vertices.
The real recursive tree process
Definition 5.20. Let E ∈ E p E\in\mathcal{E}_p E ∈ E p with E > 0 E>0 E > 0 , suppose that the no-ac hypothesis holds at E E E , let L ≥ 0 L\ge0 L ≥ 0 , and let π ′ ∈ { π E , π E ∞ } \pi'\in\{\pi_E,\pi_E^\infty\} π ′ ∈ { π E , π E ∞ } .
(i) The one-spine real recursive tree process of length L L L with endpoint law π ′ \pi' π ′ consists of independent random variables D l s i d e D_l^{\mathrm{side}} D l side , 0 ≤ l < L 0\le l<L 0 ≤ l < L , with law π E \pi_E π E , Y l Y_l Y l , 1 ≤ l ≤ L 1\le l\le L 1 ≤ l ≤ L , with law Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) , and D L D_L D L with law π ′ \pi' π ′ , together with the variables D l D_l D l , 0 ≤ l < L 0\le l<L 0 ≤ l < L , defined by backward induction by
D l = D l s i d e − Y l + 1 p D l + 1 . (71) D_l=D_l^{\mathrm{side}}-\frac{Y_{l+1}}{pD_{l+1}}. \tag*{(71)} D l = D l side − p D l + 1 Y l + 1 . (71) (ii) Let 0 ≤ j < L 0\le j<L 0 ≤ j < L . The two-spine real recursive tree process with fork at generation j j j and endpoint law π ′ \pi' π ′ consists of independent random variables D l s i d e D_l^{\mathrm{side}} D l side (0 ≤ l ≤ j 0\le l\le j 0 ≤ l ≤ j ) and D l s i d e , a D_l^{\mathrm{side},a} D l side , a (j < l < L j<l<L j < l < L , a ∈ { 1 , 2 } a\in\{1,2\} a ∈ { 1 , 2 } ) with law π E \pi_E π E , Y l Y_l Y l (1 ≤ l ≤ j 1\le l\le j 1 ≤ l ≤ j ) and Y l a Y_l^a Y l a (j < l ≤ L j<l\le L j < l ≤ L , a ∈ { 1 , 2 } a\in\{1,2\} a ∈ { 1 , 2 } ) with law Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) , and D L 1 , D L 2 D_L^1,D_L^2 D L 1 , D L 2 with law π ′ \pi' π ′ , together with the variables defined by backward induction by
D l a = D l s i d e , a − Y l + 1 a p D l + 1 a ( j < l < L , a ∈ { 1 , 2 } ) , D_l^a=D_l^{\mathrm{side},a}-\frac{Y_{l+1}^a}{pD_{l+1}^a}\qquad(j<l<L,\ a\in\{1,2\}), D l a = D l side , a − p D l + 1 a Y l + 1 a ( j < l < L , a ∈ { 1 , 2 }) , D j = D j s i d e − ∑ a = 1 2 Y j + 1 a p D j + 1 a , D l = D l s i d e − Y l + 1 p D l + 1 ( l < j ) . D_j=D_j^{\mathrm{side}}-\sum_{a=1}^{2}\frac{Y_{j+1}^a}{pD_{j+1}^a},\qquad D_l=D_l^{\mathrm{side}}-\frac{Y_{l+1}}{pD_{l+1}}\qquad(l<j). D j = D j side − a = 1 ∑ 2 p D j + 1 a Y j + 1 a , D l = D l side − p D l + 1 Y l + 1 ( l < j ) . The recursions are well defined almost surely. Indeed, by induction from the endpoints, each denominator is almost surely nonzero: the endpoint laws are carried by R ∖ { 0 } \mathbb{R}\setminus\{0\} R ∖ { 0 } , and each other denominator is, conditionally on the side denominator and on the denominators of the next generation, an affine function of one or two independent Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) variables with nonzero coefficients; such a variable has no atoms. In words, the one-spine real recursive tree process is the sequence of real denominators along a path. The side environments of the path carry independent denominators with the real boundary law, and its endpoint carries the real boundary law or the conditioned law.
Lemma 5.21. Let E ∈ E p E\in\mathcal{E}_p E ∈ E p with E > 0 E>0 E > 0 , suppose that the no-ac hypothesis holds at E E E , and let Q Q Q be a spine law of length L L L .
(i) Q-almost surely, all conclusions of Definition 5.7, Lemma 5.9, Lemma 5.10 and Proposition 5.11(ii) hold. In particular, for every v ∈ V v\in V v ∈ V , we have $D_v(E+i\eta)\to D_v=\mathcal{D}(\theta_v\omega)\in
R ∖ { 0 } \mathbb{R} \setminus\{0\} R ∖ { 0 } , and for every spine vertex x x x which is not an endpoint, D x s i d e ( E + i η ) → D x s i d e ( E ) ∈ R ∖ { 0 } D_x^{\mathrm{side}}(E+i\eta) \to D_x^{\mathrm{side}}(E) \in\mathbb{R} \setminus\{0\} D x side ( E + i η ) → D x side ( E ) ∈ R ∖ { 0 } , where
D x s i d e ( E ) = E − 1 p ∑ y s i d e c h i l d o f x Y y D y . D_x^{\mathrm{side}}(E)=E-\frac{1}{p}\sum_{y\ \mathrm{side\ child\ of}\ x}\frac{Y_y}{D_y}. D x side ( E ) = E − p 1 y side child of x ∑ D y Y y .
(ii) Under Q L \mathbb{Q}_L Q L , respectively Q L ∞ \mathbb{Q}_L^\infty Q L ∞ , the variables D v l s i d e ( E ) D_{v_l}^{\mathrm{side}}(E) D v l side ( E ) , 0 ≤ l < L 0\le l<L 0 ≤ l < L , Y v l Y_{v_l} Y v l , 1 ≤ l ≤ L 1\le l\le L 1 ≤ l ≤ L , and D v l D_{v_l} D v l have the joint law of the one-spine real recursive tree process of length L L L with endpoint law π E \pi_E π E , respectively π E ∞ \pi_E^\infty π E ∞ , and D v l D_{v_l} D v l is the corresponding variable D l D_l D l of this process. The analogous statement holds for the two-spine laws and the two-spine real recursive tree process.
(iii) Let m ≥ L m\ge L m ≥ L . Under Q \mathbb{Q} Q , conditionally on F m \mathcal{F}_m F m , on the distinguished vertices, and on ( ς v ) v ∈ V m (\varsigma_v)_{v\in\mathcal{V}_m} ( ς v ) v ∈ V m , the variables D v D_v D v , v ∈ V m v\in\mathcal{V}_m v ∈ V m , are independent, with law π E ∞ \pi_E^\infty π E ∞ if ς v = 1 \varsigma_v=1 ς v = 1 and π E \pi_E π E if ς v = 0 \varsigma_v=0 ς v = 0 . The denominators D v D_v D v with ∣ v ∣ < m |v|<m ∣ v ∣ < m are determined by these and by F m \mathcal{F}_m F m through (5.9).
Proof. The conclusions in (i) hold P \mathbb{P} P -almost surely, by Definition 5.7, Lemmas 5.9 and 5.10 , and Proposition 5.11 (ii) together with Step 5 of its proof, and by Lemma 5.18 (i) they hold Q \mathbb{Q} Q -almost surely. For a spine vertex x x x which is not an endpoint, the side denominator equals − 1 / Γ x J ( z ) -1/\Gamma_x^J(z) − 1/ Γ x J ( z ) , where J J J is the set of side children of x x x , and Γ x J ( z ) \Gamma_x^J(z) Γ x J ( z ) is a diagonal entry of the resolvent of a reduced operator. By Lemma 5.9 , its boundary value is finite and nonzero, and equals − 1 / ( E + ∑ y ∈ J ∣ h x y ∣ 2 Γ y ( E + i 0 ) ) -1/(E+\sum_{y\in J}|h_{xy}|^2\Gamma_y(E+i0)) − 1/ ( E + ∑ y ∈ J ∣ h x y ∣ 2 Γ y ( E + i 0 )) . Then D x s i d e ( E + i η ) D_x^{\mathrm{side}}(E+i\eta) D x side ( E + i η ) converges to the finite nonzero number D x s i d e ( E ) D_x^{\mathrm{side}}(E) D x side ( E ) . This proves (i).
For (ii), we consider Q L \mathbb{Q}_L Q L and use the notation of Lemma 4.7 . Let Ψ \Psi Ψ be the map on finite counting measures on C × Ω \mathbb{C}\times\Omega C × Ω defined by Ψ ( ∑ i δ ( w i , ω i ) ) = E − p − 1 ∑ i ∣ w i ∣ 2 / D ( ω i ) \Psi(\sum_i\delta_{(w_i,\omega_i)})=E-p^{-1}\sum_i|w_i|^2/\mathcal{D}(\omega_i) Ψ ( ∑ i δ ( w i , ω i ) ) = E − p − 1 ∑ i ∣ w i ∣ 2 / D ( ω i ) . By (i), almost surely D v l s i d e ( E ) = Ψ ( Σ l ) D_{v_l}^{\mathrm{side}}(E)=\Psi(\Sigma_l) D v l side ( E ) = Ψ ( Σ l ) for 0 ≤ l < L 0\le l<L 0 ≤ l < L , where Σ l \Sigma_l Σ l is the side process at v l v_l v l . By Lemma 4.7 (i), the processes Σ l \Sigma_l Σ l are independent Poisson processes with intensity measure p γ ⊗ P p\gamma\otimes\mathbb{P} p γ ⊗ P , and they are independent of the spine edge marks w v l w_{v_l} w v l and of the endpoint configuration θ v l ω \theta_{v_l}\omega θ v l ω , which has law P \mathbb{P} P . On the other hand, under P \mathbb{P} P , Proposition 5.11 (ii) and Step 5 of its proof give D o = Ψ ( Π o ) D_o=\Psi(\Pi_o) D o = Ψ ( Π o ) almost surely, where Π o \Pi_o Π o is the child process of the root, a Poisson process with intensity measure p γ ⊗ P p\gamma\otimes\mathbb{P} p γ ⊗ P by Lemma 4.2 (iii). Hence each D v l s i d e ( E ) D_{v_l}^{\mathrm{side}}(E) D v l side ( E ) has law π E \pi_E π E . Further, Y v l = ∣ w v l ∣ 2 Y_{v_l}=|w_{v_l}|^2 Y v l = ∣ w v l ∣ 2 has law Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) , and D v l = D ( θ v l ω ) D_{v_l}=\mathcal{D}(\theta_{v_l}\omega) D v l = D ( θ v l ω ) has law π E \pi_E π E . These variables are independent, since they are functions of the independent random elements of Lemma 4.7 (i). Finally, (5.9) at v l v_l v l , split as in (70) , gives (71) . Under Q L ∞ \mathbb{Q}_L^\infty Q L ∞ , Lemma 4.7 (iii) gives the same conclusions, except that θ v l ω \theta_{v_l}\omega θ v l ω has law P ∞ \mathbb{P}^\infty P ∞ and D v l D_{v_l} D v l has law π E ∞ \pi_E^\infty π E ∞ . The two-spine laws are treated in the same way, using Lemma 4.7 (iv).
For (iii), we first consider Q = Q L \mathbb{Q}=\mathbb{Q}_L Q = Q L . Let g ≥ 0 g\ge0 g ≥ 0 be a measurable function of the generators of F m \mathcal{F}_m F m and of a word in U L \mathcal{U}_L U L , and let h ≥ 0 h\ge0 h ≥ 0 be a measurable function of ( θ u ω ) u ∈ U m (\theta_u\omega)_{u\in\mathcal{U}_m} ( θ u ω ) u ∈ U m . By (38) and Lemma 4.2 (i), since V L \mathcal{V}_L V L is F m \mathcal{F}_m F m -measurable for L ≤ m L\le m L ≤ m , we have
E Q L [ g ( ω , v ) h ] = p − L E ∑ v ∈ V L g ( ω , v ) h = p − L E ∑ v ∈ V L g ( ω , v ) ⋅ E h = E Q L [ g ( ω , v ) ] E h . \mathbb{E}_{\mathbb{Q}_L}[g(\omega,\mathbf{v})h]
=p^{-L}\mathbb{E}\sum_{v\in\mathcal{V}_L}g(\omega,v)h
=p^{-L}\mathbb{E}\sum_{v\in\mathcal{V}_L}g(\omega,v)\cdot\mathbb{E}h
=\mathbb{E}_{\mathbb{Q}_L}[g(\omega,\mathbf{v})]\mathbb{E}h. E Q L [ g ( ω , v ) h ] = p − L E v ∈ V L ∑ g ( ω , v ) h = p − L E v ∈ V L ∑ g ( ω , v ) ⋅ E h = E Q L [ g ( ω , v )] E h .
Under Q L \mathbb{Q}_L Q L , the configurations θ u ω \theta_u\omega θ u ω , u ∈ U m u\in\mathcal{U}_m u ∈ U m , are therefore independent with law P \mathbb{P} P , and independent of F m \mathcal{F}_m F m and v \mathbf{v} v . Conditioning further on the survival indicators ς u = 1 S ( θ u ω ) \varsigma_u=1_{\mathsf{S}(\theta_u\omega)} ς u = 1 S ( θ u ω ) , u ∈ V m u \in\mathcal{V}_m u ∈ V m , we find that these configurations are conditionally independent with law P ( ⋅ ∣ S ) \mathbb{P}(\,\cdot\mid S) P ( ⋅ ∣ S ) or P ( ⋅ ∣ S c ) \mathbb{P}(\,\cdot\mid S^{c}) P ( ⋅ ∣ S c ) . Since D u = D ( θ u ω ) D_u=\mathcal{D}(\theta_u\omega) D u = D ( θ u ω ) , this proves the claim under Q L \mathbb{Q}_{L} Q L . Under Q L ∞ \mathbb{Q}_{L}^{\infty} Q L ∞ , the law is Q L \mathbb{Q}_{L} Q L conditioned on { θ v ω ∈ S } \{\theta_v\omega\in S\} { θ v ω ∈ S } , and this event occurs if and only if ς u = 1 \varsigma_u=1 ς u = 1 for some u ∈ V m u\in\mathcal{V}_m u ∈ V m with prefix v v v ; it is therefore measurable with respect to the conditioning σ \sigma σ -field, and conditioning on it does not change the conditional laws. The two-spine laws are treated in the same way, using (4.6). The last sentence of (iii) is (5.9).
The next lemma helps to verify that the boundaries of the events used in Section 10, which are defined by finitely many inequalities on denominators and squared moduli, are null sets for the limiting law.
Lemma 5.22. In the setting of Lemma 5.21, the following hold for every c ∈ R c\in\mathbb{R} c ∈ R . For every word v ∈ U v\in\mathcal{U} v ∈ U , we have Q ( v ∈ V , N v ≥ 1 , D v = c ) = 0 \mathbb{Q}(v\in V,N_v\geq1,D_v=c)=0 Q ( v ∈ V , N v ≥ 1 , D v = c ) = 0 ; in particular Q ( D x = c ) = 0 \mathbb{Q}(D_x=c)=0 Q ( D x = c ) = 0 for every spine vertex x x x which is not an endpoint. For every such x x x , we have Q ( D x s i d e ( E ) = c ) = 0 \mathbb{Q}(D_x^{\mathrm{side}}(E)=c)=0 Q ( D x side ( E ) = c ) = 0 if c ≠ E c\ne E c = E . For every endpoint t t t , we have Q ( D t = c ) = 0 \mathbb{Q}(D_t=c)=0 Q ( D t = c ) = 0 if c ≠ E c\ne E c = E or if the endpoint law is P ∞ \mathbb{P}^{\infty} P ∞ . Finally, Q ( Y u = c ) = 0 \mathbb{Q}(Y_u=c)=0 Q ( Y u = c ) = 0 for every u ∈ U ∗ u\in\mathcal{U}_* u ∈ U ∗ .
Proof. Fix v ∈ U v\in\mathcal{U} v ∈ U and let A = { v k : 1 ≤ k ≤ N v } \mathcal{A}=\{vk:1\leq k\leq N_v\} A = { v k : 1 ≤ k ≤ N v } if v ∈ V v\in V v ∈ V , and A = ∅ \mathcal{A}=\varnothing A = ∅ otherwise; this is a function of the offspring numbers. By Lemma 5.18(ii), the squared moduli Y v k Y_{vk} Y v k , k ≤ N v k\leq N_v k ≤ N v , are fresh given a σ \sigma σ -field with respect to which the denominators D v k = D ( θ v k ω ) D_{vk}=\mathcal{D}(\theta_{vk}\omega) D v k = D ( θ v k ω ) are measurable, since these depend only on coordinates of θ v k ω \theta_{vk}\omega θ v k ω . On { v ∈ V , N v ≥ 1 } \{v\in V,N_v\geq1\} { v ∈ V , N v ≥ 1 } , (5.9) expresses D v D_v D v as an affine function of these fresh variables with nonzero coefficients − 1 / ( p D v k ) -1/(pD_{vk}) − 1/ ( p D v k ) . By Lemma 4.18(iii) and (4.14), its conditional law has no atoms. A spine vertex which is not an endpoint has at least one child. The claims about D x s i d e ( E ) D_x^{\mathrm{side}}(E) D x side ( E ) and D t D_t D t follow from Lemma 5.21(ii), Proposition 5.5(i) applied to π E \pi_E π E , and the absence of atoms of π E ∞ \pi_E^\infty π E ∞ . The squared moduli have continuous laws.
In particular, let Φ \Phi Φ be a function of finitely many of the variables in Lemma 5.22, evaluated at E + i η E+i\eta E + i η through the real parts of the corresponding coordinates of the depth-m m m data, and suppose that Φ \Phi Φ is continuous outside a finite union of level sets of these variables, at levels c c c for which Lemma 5.22 gives probability zero. Since the denominators are continuous functions of the nonvanishing cavity coordinates, the discontinuity set of Φ \Phi Φ is then Q \mathbb{Q} Q -null for the limiting depth-m m m data.
Fractional products
For a word v ∈ U L v\in\mathcal{U}_L v ∈ U L , the spine vertices of v v v are its prefixes v 0 , … , v L v_0,\ldots,v_L v 0 , … , v L , and the spine child of v l v_l v l is v l + 1 v_{l+1} v l + 1 for l < L l<L l < L . For a pair ( u , v ) ∈ U L 2 (u,v)\in\mathcal{U}_L^2 ( u , v ) ∈ U L 2 with ∣ u ∧ v ∣ = j < L |u\wedge v|=j<L ∣ u ∧ v ∣ = j < L , the spine vertices are the prefixes of u u u and of v v v , and the spine children of a spine vertex are those of its children that are spine vertices. This agrees with Section 5.6 when v = u v=u v = u or ( u , v ) = ( u , v ) (u,v)=(u,v) ( u , v ) = ( u , v ) .
Definition 5.23. Let L ≥ 0 L \ge0 L ≥ 0 , m ≥ L + 1 m \ge L + 1 m ≥ L + 1 , s ∈ ( 0 , 1 ) s \in(0,1) s ∈ ( 0 , 1 ) , b ≥ 0 b \ge0 b ≥ 0 and k , l 0 ∈ N k,l_0 \in\mathbb{N} k , l 0 ∈ N , and let v \mathbf{v} v denote a word v ∈ U L v \in\mathbb{U}_L v ∈ U L or a pair ( u , v ) ∈ U L 2 (u,v) \in\mathbb{U}_L^2 ( u , v ) ∈ U L 2 with ∣ u ∧ v ∣ = j |u \wedge v| = j ∣ u ∧ v ∣ = j for a fixed j < L j < L j < L . A fractional product of depth L L L with cavity exponents at most s s s and exponents of the squared moduli at most b b b is a function on Ω × { v } × ( C ∖ R ) \Omega\times\{\mathbf{v}\} \times(\mathbb{C} \setminus\mathbb{R}) Ω × { v } × ( C ∖ R ) of the form
W ( ω , v ; z ) = ∏ x ( ∏ i = 1 l 0 ∣ Γ x J x , i ( z ) ∣ a x , i ∏ y ∈ K x γ y b x , y ) . W(\omega,\mathbf{v};z)=\prod_{x}\left(\prod_{i=1}^{l_0}|\Gamma_x^{J_{x,i}}(z)|^{a_{x,i}}\prod_{y\in K_x}\gamma_y^{b_{x,y}}\right). W ( ω , v ; z ) = x ∏ i = 1 ∏ l 0 ∣ Γ x J x , i ( z ) ∣ a x , i y ∈ K x ∏ γ y b x , y .
where the first product runs over the spine vertices x x x of v \mathbf{v} v in V V V (and W = 0 W=0 W = 0 if some spine vertex is not in V V V ), and where, for each spine vertex x x x :
the sets J x , 1 , … , J x , l 0 J_{x,1},\ldots,J_{x,l_0} J x , 1 , … , J x , l 0 and K x K_x K x of children of x x x , with ∣ K x ∣ ≤ k |K_x| \le k ∣ K x ∣ ≤ k , and the numbers b x , y ∈ [ 0 , b ] b_{x,y} \in[0,b] b x , y ∈ [ 0 , b ] are given by measurable rules depending only on the position of x x x among the spine vertices, on the offspring numbers and auxiliary marks of θ x ω \theta_x\omega θ x ω at words of length less than m − ∣ x ∣ m-|x| m − ∣ x ∣ , and on the labels of the spine children of x x x ;
Γ x J \Gamma_x^J Γ x J is defined in (69) ;
the exponents a x , i ≥ 0 a_{x,i} \ge0 a x , i ≥ 0 , which we call the cavity exponents, are deterministic, depend only on the position of x x x among the spine vertices, and satisfy ∑ i a x , i ≤ s \sum_i a_{x,i} \le s ∑ i a x , i ≤ s .
For E ∈ E p E \in\mathcal{E}_p E ∈ E p , we define W ( ω , v ; E ) W(\omega,\mathbf{v};E) W ( ω , v ; E ) by the same formula with Γ x J ( E + i 0 ) \Gamma_x^J(E+i0) Γ x J ( E + i 0 ) in place of Γ x J ( z ) \Gamma_x^J(z) Γ x J ( z ) .
For example, by Proposition 3.18 , we have
∣ G o v ( z ) ∣ s = p − s L / 2 ∣ Γ o ( z ) ∣ s ∏ l = 1 L γ v l s / 2 ∣ Γ v l ( z ) ∣ s |G_{ov}(z)|^s=p^{-sL/2}|\Gamma_o(z)|^s\prod_{l=1}^{L}\gamma_{v_l}^{s/2}|\Gamma_{v_l}(z)|^s ∣ G o v ( z ) ∣ s = p − s L /2 ∣ Γ o ( z ) ∣ s l = 1 ∏ L γ v l s /2 ∣ Γ v l ( z ) ∣ s
for v ∈ V L v \in\mathcal{V}_L v ∈ V L , and this is p − s L / 2 p^{-sL/2} p − s L /2 times a fractional product with cavity exponents s s s and exponents s / 2 s/2 s /2 of the squared moduli, in which J x , 1 J_{x,1} J x , 1 is the set of all children of x x x , K v l = { v l + 1 } K_{v_l}=\{v_{l+1}\} K v l = { v l + 1 } for 0 ≤ l < L 0 \le l < L 0 ≤ l < L , and K v L = ∅ K_{v_L}=\varnothing K v L = ∅ . A product ∣ G o u ( z ) ∣ s 1 ∣ G o v ( z ) ∣ s 2 |G_{ou}(z)|^{s_1}|G_{ov}(z)|^{s_2} ∣ G o u ( z ) ∣ s 1 ∣ G o v ( z ) ∣ s 2 along two spines is, up to a constant, a fractional product with cavity exponent s 1 + s 2 s_1+s_2 s 1 + s 2 on the common segment of the two spines; the definition requires s 1 + s 2 < 1 s_1+s_2<1 s 1 + s 2 < 1 there.
Lemma 5.24. Let Q Q Q be a spine law of length L L L , let W W W be a fractional product of depth L L L of the corresponding type, with cavity exponents at most r ∈ ( 0 , 1 ) r \in(0,1) r ∈ ( 0 , 1 ) , exponents of the squared moduli at most b b b , and ∣ K x ∣ ≤ k |K_x| \le k ∣ K x ∣ ≤ k , and let d > 0 d>0 d > 0 . Then for every z ∈ C ∖ R z \in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R with ∣ z ∣ ≥ d |z| \ge d ∣ z ∣ ≥ d ,
E Q W ( ω , v ; z ) ≤ ϱ p − 2 ( ( 2 c b ) k ( 1 + C r , d ( p + 3 ) r / 2 ) ) n , \mathbb{E}_QW(\omega,\mathbf{v};z)\le\varrho_p^{-2}\left((2c_b)^k\left(1+C_{r,d}(p+3)^{r/2}\right)\right)^n, E Q W ( ω , v ; z ) ≤ ϱ p − 2 ( ( 2 c b ) k ( 1 + C r , d ( p + 3 ) r /2 ) ) n ,
where v = v \mathbf{v}=v v = v or ( u , v ) (u,v) ( u , v ) , n n n is the number of spine vertices, and c b c_b c b and C r , d C_{r,d} C r , d are as in Lemma 5.16 .
Proof. We argue by induction on L L L , integrating the factor of the root first. Let R o ( z ) R_o(z) R o ( z ) be the factor of W W W indexed by x = o x=o x = o , and let W ′ ( z ) W'(z) W ′ ( z ) be the product of the other factors; then W = R o W ′ W=R_oW' W = R o W ′ . Let A A A be the set of children of the root, and let G \mathcal{G} G be the σ \sigma σ -field of Lemma 5.18(ii) for this set. Then the squared moduli Y 1 , … , Y N o Y_1,\ldots,Y_{N_o} Y 1 , … , Y N o of the root edge marks are fresh given G \mathcal{G} G . The forward cavities Γ k ( z ) = G o o ( z ) ∘ θ k \Gamma_k(z)=G_{oo}(z)\circ\theta_k Γ k ( z ) = G oo ( z ) ∘ θ k , the sets J o , i J_{o,i} J o , i and K o K_o K o , the numbers b o , y b_{o,y} b o , y , and the factor W ′ ( z ) W'(z) W ′ ( z ) are G \mathcal{G} G -measurable, since they depend only on the offspring numbers, the auxiliary marks, the distinguished vertices, and the configurations θ k ω \theta_k\omega θ k ω , which do not contain the marks w k w_k w k . The conditional version of Lemma 5.16, with n = N o n=N_o n = N o , c = z c=z c = z and γ k = Γ k ( z ) / p \gamma_k=\Gamma_k(z)/p γ k = Γ k ( z ) / p , gives
E Q [ R o ( z ) ∣ G ] ≤ ( 2 c b ) k ( 1 + C r , d ( N o + 1 ) r / 2 ) . \mathbb{E}_{\mathbb{Q}}\left[R_o(z)\mid\mathcal{G}\right]\leq(2c_b)^k\left(1+C_{r,d}(N_o+1)^{r/2}\right). E Q [ R o ( z ) ∣ G ] ≤ ( 2 c b ) k ( 1 + C r , d ( N o + 1 ) r /2 ) .
Taking expectations, we find
E Q W ( z ) ≤ ( 2 c b ) k E Q [ ( 1 + C r , d ( N o + 1 ) r / 2 ) W ′ ( z ) ] . (72) \mathbb{E}_{\mathbb{Q}}W(z)\leq(2c_b)^k\mathbb{E}_{\mathbb{Q}}\left[\left(1+C_{r,d}(N_o+1)^{r/2}\right)W'(z)\right]. \tag*{(72)} E Q W ( z ) ≤ ( 2 c b ) k E Q [ ( 1 + C r , d ( N o + 1 ) r /2 ) W ′ ( z ) ] . (72)
If L = 0 L=0 L = 0 , then W ′ = 1 W'=1 W ′ = 1 , the law of N o N_o N o is Pois ( p ) \operatorname{Pois}(p) Pois ( p ) under P \mathbb{P} P , and E ∞ Z ≤ ϱ p − 1 E Z \mathbb{E}^{\infty}Z\leq\varrho_p^{-1}\mathbb{E}Z E ∞ Z ≤ ϱ p − 1 E Z for Z ≥ 0 Z\geq0 Z ≥ 0 ; Jensen’s inequality and (72) give the claim. Let L ≥ 1 L\geq1 L ≥ 1 , and consider a one-spine law. By Lemma 5.18(iii), N o − 1 N_o-1 N o − 1 has law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) and is independent of the pair ( θ I 1 ω , v ′ ) (\theta_{I_1}\omega,v') ( θ I 1 ω , v ′ ) , which has law Q L − 1 \mathbb{Q}_{L-1} Q L − 1 or Q L − 1 ∞ \mathbb{Q}^{\infty}_{L-1} Q L − 1 ∞ . The factor W ′ W' W ′ is a function of this pair, and in the coordinates of θ I 1 ω \theta_{I_1}\omega θ I 1 ω it is a fractional product of depth L − 1 L-1 L − 1 with the same parameters, by Definition 5.23(i) and (iii) and since m − 1 ≥ L m-1\geq L m − 1 ≥ L . Using this independence in (72) and Jensen’s inequality E ( N o + 1 ) r / 2 ≤ ( p + 2 ) r / 2 \mathbb{E}(N_o+1)^{r/2}\leq(p+2)^{r/2} E ( N o + 1 ) r /2 ≤ ( p + 2 ) r /2 , we find
E Q W ( z ) ≤ ( 2 c b ) k ( 1 + C r , d ( p + 2 ) r / 2 ) E Q ′ W ′ ( z ) , \mathbb{E}_{\mathbb{Q}}W(z)\leq(2c_b)^k\left(1+C_{r,d}(p+2)^{r/2}\right)\mathbb{E}_{\mathbb{Q}'}W'(z), E Q W ( z ) ≤ ( 2 c b ) k ( 1 + C r , d ( p + 2 ) r /2 ) E Q ′ W ′ ( z ) ,
where Q ′ \mathbb{Q}' Q ′ is the law of the pair, and the induction hypothesis applies to the last factor. For two-spine laws with j ≥ 1 j\geq1 j ≥ 1 , the same argument applies with the triple of Lemma 5.18(iii). For j = 0 j=0 j = 0 , the factor W ′ W' W ′ is the product of two fractional products of depth L − 1 L-1 L − 1 , one for each of the two independent pairs of Lemma 5.18(iii), and N o − 2 N_o-2 N o − 2 has law Pois ( p ) \operatorname{Pois}(p) Pois ( p ) ; the expectation factorizes, and E ( N o + 1 ) r / 2 ≤ ( p + 3 ) r / 2 \mathbb{E}(N_o+1)^{r/2}\leq(p+3)^{r/2} E ( N o + 1 ) r /2 ≤ ( p + 3 ) r /2 . In each case, the factor ϱ p − 1 \varrho_p^{-1} ϱ p − 1 appears once for each endpoint with endpoint law P ∞ \mathbb{P}^{\infty} P ∞ , and there are at most two endpoints. □
Convergence at fixed depth
Proposition 5.25. Let E ∈ E p E\in\mathcal{E}_p E ∈ E p with E > 0 E>0 E > 0 , and suppose that the no-ac hypothesis holds at E E E , that is, P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 \mathbb{P}^{\infty}(\operatorname{Im}\Gamma_o(E+\mathrm{i}0)=0)=1 P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 . Let Q \mathbb{Q} Q be a spine law of length L L L , write v = v v=v v = v or v = ( u , v ) v=(u,v) v = ( u , v ) for its distinguished vertices, and let m ≥ L + 1 m\geq L+1 m ≥ L + 1 .
(i) Q \mathbb{Q} Q -almost surely, L m ( E + i η ) → L m ( E ) \mathcal{L}_m(E+\mathrm{i}\eta)\to\mathcal{L}_m(E) L m ( E + i η ) → L m ( E ) in X m \mathcal{X}_m X m as η ↓ 0 \eta\downarrow0 η ↓ 0 . The side denominators and the endpoint denominators converge almost surely, and their limits, together with the squared moduli of the spine edge marks, have the joint law of the real recursive tree process of Lemma 5.21(ii); the denominators at the spine vertices converge to the corresponding denominators of this process. In particular, the joint law of these variables at E + i η E+\mathrm{i}\eta E + i η converges weakly to that of the real recursive tree process. (ii) Let Φ \Phi Φ be a bounded measurable function of L m \mathcal{L}_m L m and v \mathbf{v} v , and suppose that Q \mathbb{Q} Q -almost surely Φ ( ⋅ , v ) \Phi(\cdot,\mathbf{v}) Φ ( ⋅ , v ) is continuous at L m ( E ) \mathcal{L}_m(E) L m ( E ) . Then
lim η ↓ 0 E Q Φ ( L m ( E + i η ) , v ) = E Q Φ ( L m ( E ) , v ) . \lim_{\eta\downarrow0}\mathbb{E}_{\mathbb{Q}}\Phi(\mathcal{L}_m(E+i\eta),\mathbf{v})=\mathbb{E}_{\mathbb{Q}}\Phi(\mathcal{L}_m(E),\mathbf{v}). η ↓ 0 lim E Q Φ ( L m ( E + i η ) , v ) = E Q Φ ( L m ( E ) , v ) .
(iii) Let 0 < s < r < 1 0<s<r<1 0 < s < r < 1 , and let W W W be a fractional product of depth L L L of the type of Q \mathbb{Q} Q with cavity exponents at most s s s . Then sup 0 < η ≤ 1 E Q W ( ω , v ; E + i η ) r / s < ∞ \sup_{0<\eta\leq1}\mathbb{E}_{\mathbb{Q}}W(\omega,\mathbf{v};E+i\eta)^{r/s}<\infty sup 0 < η ≤ 1 E Q W ( ω , v ; E + i η ) r / s < ∞ . In particular, the family { W ( ω , v ; E + i η ) } 0 < η ≤ 1 \{W(\omega,\mathbf{v};E+i\eta)\}_{0<\eta\leq1} { W ( ω , v ; E + i η ) } 0 < η ≤ 1 is uniformly integrable under Q \mathbb{Q} Q , and for every Φ \Phi Φ as in (ii),
lim η ↓ 0 E Q [ W ( ω , v ; E + i η ) Φ ( L m ( E + i η ) , v ) ] = E Q [ W ( ω , v ; E ) Φ ( L m ( E ) , v ) ] . \lim_{\eta\downarrow0}\mathbb{E}_{\mathbb{Q}}\left[W(\omega,\mathbf{v};E+i\eta)\Phi(\mathcal{L}_m(E+i\eta),\mathbf{v})\right]=\mathbb{E}_{\mathbb{Q}}\left[W(\omega,\mathbf{v};E)\Phi(\mathcal{L}_m(E),\mathbf{v})\right]. η ↓ 0 lim E Q [ W ( ω , v ; E + i η ) Φ ( L m ( E + i η ) , v ) ] = E Q [ W ( ω , v ; E ) Φ ( L m ( E ) , v ) ] .
(iv) Let W W W and Φ \Phi Φ be as in (iii) for Q = Q L \mathbb{Q}=\mathbb{Q}_L Q = Q L , and set F η ( ω , v ) = W ( ω , v ; E + i η ) Φ ( L m ( E + i η ) , v ) F_\eta(\omega,\mathbf{v})=W(\omega,\mathbf{v};E+i\eta)\Phi(\mathcal{L}_m(E+i\eta),\mathbf{v}) F η ( ω , v ) = W ( ω , v ; E + i η ) Φ ( L m ( E + i η ) , v ) for 0 ≤ η ≤ 1 0\leq\eta\leq1 0 ≤ η ≤ 1 , with the convention that η = 0 \eta=0 η = 0 refers to the boundary values. Then
lim η ↓ 0 E ∑ v ∈ V L F η ( ω , v ) = E ∑ v ∈ V L F 0 ( ω , v ) , lim η ↓ 0 E ∞ ∑ v ∈ V L ∞ F η ( ω , v ) = E ∞ ∑ v ∈ V L ∞ F 0 ( ω , v ) . \lim_{\eta\downarrow0}\mathbb{E}\sum_{\mathbf{v}\in V_L}F_\eta(\omega,\mathbf{v})=\mathbb{E}\sum_{\mathbf{v}\in V_L}F_0(\omega,\mathbf{v}),\qquad
\lim_{\eta\downarrow0}\mathbb{E}^{\infty}\sum_{\mathbf{v}\in V_L^\infty}F_\eta(\omega,\mathbf{v})=\mathbb{E}^{\infty}\sum_{\mathbf{v}\in V_L^\infty}F_0(\omega,\mathbf{v}). η ↓ 0 lim E v ∈ V L ∑ F η ( ω , v ) = E v ∈ V L ∑ F 0 ( ω , v ) , η ↓ 0 lim E ∞ v ∈ V L ∞ ∑ F η ( ω , v ) = E ∞ v ∈ V L ∞ ∑ F 0 ( ω , v ) .
The same holds for the backbone sums E ∞ ∑ v ∈ V L ∞ F η ( ω , v ) \mathbb{E}^{\infty}\sum_{\mathbf{v}\in V_L^\infty}F_\eta(\omega,\mathbf{v}) E ∞ ∑ v ∈ V L ∞ F η ( ω , v ) when the hypotheses of (iii) hold for Q L ∞ \mathbb{Q}_L^\infty Q L ∞ , and for the sums over pairs in Proposition 4.8(ii) and (iv) when they hold for Q L , j \mathbb{Q}_{L,j} Q L , j , respectively Q L , j ∞ \mathbb{Q}_{L,j}^{\infty} Q L , j ∞ . Each limit equals the right side of the corresponding identity in Proposition 4.8 with F = F 0 F=F_0 F = F 0 .
(v) Let finitely many triples ( Q k , W k , Φ k ) (\mathbb{Q}_k,W_k,\Phi_k) ( Q k , W k , Φ k ) , 1 ≤ k ≤ K 1\leq k\leq K 1 ≤ k ≤ K , be given at a common depth m m m , where each W k W_k W k is either the constant 1 1 1 or a fractional product with cavity exponents at most some s k < 1 s_k<1 s k < 1 , and each Φ k \Phi_k Φ k satisfies the hypothesis of (ii) for Q k \mathbb{Q}_k Q k . For every ε > 0 \varepsilon>0 ε > 0 there exists η 0 > 0 \eta_0>0 η 0 > 0 such that, for all 0 < η ≤ η 0 0<\eta\leq\eta_0 0 < η ≤ η 0 and all k k k ,
∣ E Q k [ W k ( ω , v ; E + i η ) Φ k ( L m ( E + i η ) , v ) ] − E Q k [ W k ( ω , v ; E ) Φ k ( L m ( E ) , v ) ] ∣ ≤ ε . \left|\mathbb{E}_{\mathbb{Q}_k}\left[W_k(\omega,\mathbf{v};E+i\eta)\Phi_k(\mathcal{L}_m(E+i\eta),\mathbf{v})\right]-\mathbb{E}_{\mathbb{Q}_k}\left[W_k(\omega,\mathbf{v};E)\Phi_k(\mathcal{L}_m(E),\mathbf{v})\right]\right|\leq\varepsilon. ∣ E Q k [ W k ( ω , v ; E + i η ) Φ k ( L m ( E + i η ) , v ) ] − E Q k [ W k ( ω , v ; E ) Φ k ( L m ( E ) , v ) ] ∣ ≤ ε .
The same holds for finitely many sums as in (iv).
Proof. Step 1. Almost sure convergence. By Definition 5.7(a) and Lemma 5.18(i), Q \mathbb{Q} Q -almost surely F κ ( E + i η ) → F κ ( E + i 0 ) F_\kappa(E+i\eta)\to F_\kappa(E+i0) F κ ( E + i η ) → F κ ( E + i 0 ) for every κ ∈ K m \kappa\in\mathcal{K}_m κ ∈ K m . The other components of L m ( z ) \mathcal{L}_m(z) L m ( z ) do not depend on z z z . Since convergence in the product topology is coordinatewise convergence, L m ( E + i η ) → L m ( E ) \mathcal{L}_m(E+i\eta)\to\mathcal{L}_m(E) L m ( E + i η ) → L m ( E ) almost surely. The convergence of the side denominators and the law of the limit are Lemma 5.21(i) and (ii), and almost sure convergence implies weak convergence of the joint law. This proves (i).
Step 2. Bounded functionals. By Step 1 and the continuity assumption, Φ ( L m ( E + i η ) , v ) → Φ ( L m ( E ) , v ) \Phi(\mathcal{L}_m(E+i\eta),\mathbf{v})\to\Phi(\mathcal{L}_m(E),\mathbf{v}) Φ ( L m ( E + i η ) , v ) → Φ ( L m ( E ) , v ) almost surely. Since Φ \Phi Φ is bounded, the dominated convergence theorem implies (ii).
Step 3. Fractional products. Let 0 < η ≤ 1 0<\eta\leq1 0 < η ≤ 1 and z = E + i η z=E+i\eta z = E + i η ; note that ∣ z ∣ ≥ E |z|\geq E ∣ z ∣ ≥ E . Raising W W W to the power r / s r/s r / s multiplies the exponents a x , i a_{x,i} a x , i and b x , y b_{x,y} b x , y by r / s r/s r / s , which shows that W r / s W^{r/s} W r / s is a fractional product with cavity exponents at most r r r and exponents of the squared moduli at most b r / s br/s b r / s . By Lemma 5.24 with d = E d=E d = E , we have a bound on E Q W ( ω , v ; z ) r / s \mathbb{E}_{\mathbb{Q}}W(\omega,\mathbf{v};z)^{r/s} E Q W ( ω , v ; z ) r / s that does not depend on η \eta η . Since r / s > 1 r/s>1 r / s > 1 , the family { W ( ω , v ; E + i η ) } 0 < η ≤ 1 \{W(\omega,\mathbf{v};E+i\eta)\}_{0<\eta\leq1} { W ( ω , v ; E + i η ) } 0 < η ≤ 1 is uniformly integrable. Since each Γ x j ( z ) \Gamma_x^j(z) Γ x j ( z ) appearing in W W W has ∣ x ∣ ≤ L < m |x|\leq L<m ∣ x ∣ ≤ L < m , it is one of the resolvent entries contained in L m ( z ) \mathcal{L}_m(z) L m ( z ) , selected by a function of quantities which do not depend on η \eta η . By Step 1, it converges almost surely to Γ x j ( E + i 0 ) \Gamma_x^j(E+i0) Γ x j ( E + i 0 ) , which implies that W ( ω , v ; E + i η ) → W ( ω , v ; E ) W(\omega,\mathbf{v};E+i\eta)\to W(\omega,\mathbf{v};E) W ( ω , v ; E + i η ) → W ( ω , v ; E ) almost surely. The products W Φ W\Phi W Φ converge almost surely by Step 2, and they are uniformly integrable, because their absolute values are bounded by sup ∣ Φ ∣ \sup|\Phi| sup ∣Φ∣ times W W W . By Vitali’s convergence theorem [64, Section 4.6], we conclude (iii).
Step 4. Sums over vertices. We write Φ = Φ + − Φ − \Phi=\Phi^+-\Phi^- Φ = Φ + − Φ − with Φ ± ≥ 0 \Phi^\pm\geq0 Φ ± ≥ 0 , and we apply the identities of Proposition 4.8 to W Φ + W\Phi^+ W Φ + and W Φ − W\Phi^- W Φ − separately, at E + i η E+i\eta E + i η and at the boundary. By (4.5), we have E ∑ v ∈ V L F η ( ω , v ) = p L E Q L F η ( ω , v ) \mathbb{E}\sum_{\mathbf{v}\in V_L}F_\eta(\omega,\mathbf{v})=p^L\mathbb{E}_{\mathbb{Q}_L}F_\eta(\omega,\mathbf{v}) E ∑ v ∈ V L F η ( ω , v ) = p L E Q L F η ( ω , v ) , and both sides are finite by (iii). By (iii) applied to Φ ± \Phi^\pm Φ ± , the right side converges to p L E Q L F 0 ( ω , v ) = E ∑ v ∈ V L F 0 ( ω , v ) p^L\mathbb{E}_{\mathbb{Q}_L}F_0(\omega,\mathbf{v})=\mathbb{E}\sum_{\mathbf{v}\in V_L}F_0(\omega,\mathbf{v}) p L E Q L F 0 ( ω , v ) = E ∑ v ∈ V L F 0 ( ω , v ) . Under P ∞ \mathbb{P}^\infty P ∞ , we have E ∞ ∑ v ∈ V L F η ( ω , v ) = ϱ ρ − 1 p L E Q L [ F η ( ω , v ) ζ o ] \mathbb{E}^\infty\sum_{\mathbf{v}\in V_L}F_\eta(\omega,\mathbf{v})=\varrho_{\rho}^{-1}p^L\mathbb{E}_{\mathbb{Q}_L}[F_\eta(\omega,\mathbf{v})\zeta_o] E ∞ ∑ v ∈ V L F η ( ω , v ) = ϱ ρ − 1 p L E Q L [ F η ( ω , v ) ζ o ] by (4.7), and Φ ζ o \Phi\zeta_o Φ ζ o satisfies the hypothesis of (ii) because ζ o \zeta_o ζ o is a coordinate of the depth-m m m data taking the values 0 and 1. The backbone sums and the sums over pairs are treated in the same way, using (4.8), (4.6), (4.9) and (4.10). This proves (iv).
Step 5. Uniformity over finite families. For each k k k , (ii) or (iii) implies that the difference in (v) tends to zero as η ↓ 0 \eta\downarrow0 η ↓ 0 ; in particular, it is at most ε \varepsilon ε for η ≤ η k \eta\leq\eta_k η ≤ η k , for some η k > 0 \eta_k>0 η k > 0 . We take η 0 = min k η k \eta_0=\min_k\eta_k η 0 = min k η k , and we treat finitely many sums as in (iv) in the same way. □ \square □
Remark 5.26. We comment on the use of Proposition 5.25.
(i) The order of choices is: the energy E E E and the depth m m m are fixed first, then the finite family of spine laws and functionals, then the tolerance ε \varepsilon ε , and finally the imaginary part η 0 \eta_0 η 0 . The proposition does not interchange the limits m → ∞ m\to\infty m → ∞ and η ↓ 0 \eta\downarrow0 η ↓ 0 .
(ii) In the backbone version, by Lemma 5.21(ii), the side denominators of the limiting real recursive tree process have the unconditioned real boundary law π E \pi_E π E , and only the endpoint denominators have the conditioned law π E ∞ \pi_E^\infty π E ∞ . Functionals that look beyond the endpoints are described by Lemma 5.21(iii).
(iii) Conditioning on a structural event of positive probability, such as a selection of decorations at an endpoint as in Lemma 4.21, amounts to multiplying Φ \Phi Φ by the indicator of an event depending only on the genealogy, the survival indicators and the auxiliary marks in the ball, and dividing by its probability. If the event depends on the auxiliary ranks only through their relative order, then its indicator is continuous at every point of X m \mathcal{X}_m X m at which the ranks in the ball are distinct, since the genealogy and survival coordinates are discrete; this holds almost surely. Indicators of events defined by finitely many inequalities on the real parts of side denominators, spine denominators and endpoint denominators, and on squared moduli, satisfy the hypothesis of (ii) when the levels are those allowed by Lemma 5.22.
The transfer operator and the free energy
In this section, we study the positive operator that governs the growth of fractional moments of products of transfer matrices, and we define the free energy. We work with a general law of the rescaled denominator. Throughout this section, we fix p > 1 p > 1 p > 1 and K > 0 K > 0 K > 0 , and we let ν \nu ν be a Borel probability measure on R \mathbb{R} R such that
ν ( { 0 } ) = 0 , supp ν = R , ∫ R ∣ b ∣ q ν ( d b ) < ∞ for every q ∈ ( 0 , 1 ) . (73) \nu(\{0\}) = 0,\qquad\operatorname{supp}\nu= \mathbb{R},\qquad\int_{\mathbb{R}} |b|^q \,\nu(\mathrm{d}b) < \infty\quad\text{for every }q \in(0,1). \tag*{(73)} ν ({ 0 }) = 0 , supp ν = R , ∫ R ∣ b ∣ q ν ( d b ) < ∞ for every q ∈ ( 0 , 1 ) . (73)
We let B B B be a random variable with law ν \nu ν , and we let β \beta β be a random variable independent of B B B with the exponential law of rate K K K , that is, with density K e − K t K e^{-Kt} K e − K t on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . If π \pi π is a real solution at energy E > 0 E > 0 E > 0 (Definition 5.1 ), D D D has law π \pi π , and g > 0 g > 0 g > 0 , then the law of B = g D B = gD B = g D satisfies (73) by Proposition 5.5 , and β = g 2 Y / p \beta= g^2Y/p β = g 2 Y / p with Y ∼ Exp ( 1 ) Y \sim\operatorname{Exp}(1) Y ∼ Exp ( 1 ) has the exponential law of rate K = p / g 2 K = p/g^2 K = p / g 2 . This is the situation in which the results of this section are applied.
We now describe the contents of this section. In § 6.2.1 , we define the transfer matrices, their action on the projective line, and the transfer operator Q s Q_s Q s , and we record some elementary facts about angular integrals of matrices. In § 6.3 , we prove Theorem 6.11 , which states that Q s Q_s Q s has a simple dominant eigenvalue λ ( s ) \lambda(s) λ ( s ) with a strictly positive eigenfunction, that λ ( s ) \lambda(s) λ ( s ) is real-analytic in s s s , and that λ ( s ) \lambda(s) λ ( s ) determines the growth of angular moments of products of independent transfer matrices. In § 6.6.1 , we define the free energy φ ( s ) = log λ ( s ) \varphi(s) = \log\lambda(s) φ ( s ) = log λ ( s ) and establish its basic properties: independence of the scaling, convexity, the symmetry φ ( s ) = φ ( 2 − s ) \varphi(s) = \varphi(2-s) φ ( s ) = φ ( 2 − s ) , monotonicity, and the strict inequality φ ( 1 ) < log p \varphi(1) < \log p φ ( 1 ) < log p . In § 7.0.1 , we prove inverse moment bounds for products of three independent transfer matrices, and we use them to compare the free energy with moments of norms of products. Finally, in § 7.2.1 , we identify the free energy with the exponential growth rate of the path moments Φ L ( s ) \Phi_L(s) Φ L ( s ) , which are the quantities controlled in § 8.2 .
Transfer matrices and projective coordinates
We identify the real projective line R P 1 \mathbb{RP}^1 RP 1 with R ∪ { ∞ } \mathbb{R} \cup\{\infty\} R ∪ { ∞ } , where x ∈ R x \in\mathbb{R} x ∈ R corresponds to the line through ( x , 1 ) (x,1) ( x , 1 ) and ∞ \infty ∞ corresponds to the line through ( 1 , 0 ) (1,0) ( 1 , 0 ) ; this is a homeomorphism when R ∪ { ∞ } \mathbb{R} \cup\{\infty\} R ∪ { ∞ } carries the topology of the one-point compactification. We write C ( R P 1 ) C(\mathbb{RP}^1) C ( RP 1 ) for the Banach space of continuous functions F : R P 1 → C F : \mathbb{RP}^1 \to\mathbb{C} F : RP 1 → C with the norm ∥ F ∥ ∞ = \lVert F\rVert_\infty= ∥ F ∥ ∞ = max x ∈ R P 1 ∣ F ( x ) ∣ \max_{x\in\mathbb{R}\mathbb{P}^{1}} |F(x)| max x ∈ R P 1 ∣ F ( x ) ∣ . We say that a real-valued function F F F is nonnegative, and write F ≥ 0 F \ge0 F ≥ 0 , if F ( x ) ≥ 0 F(x) \ge0 F ( x ) ≥ 0 for all x x x . Vectors in C 2 \mathbb{C}^{2} C 2 carry the Hermitian norm ∥ ⋅ ∥ \|\cdot\| ∥ ⋅ ∥ , and matrices carry the corresponding operator norm.
Definition 6.1. (i) For b ∈ R b \in\mathbb{R} b ∈ R and β > 0 \beta> 0 β > 0 , define
M ( b , β ) = β − 1 / 2 ( 0 1 − β b ) . (74) M(b,\beta)=\beta^{-1/2}\begin{pmatrix}0 & 1 \\ -\beta& b\end{pmatrix}. \tag*{(74)} M ( b , β ) = β − 1/2 ( 0 − β 1 b ) . (74) We have det M ( b , β ) = 1 \det M(b,\beta)=1 det M ( b , β ) = 1 .
(ii) For x ∈ R x \in\mathbb{R} x ∈ R , set v x = ( x , 1 ) / 1 + x 2 v_x=(x,1)/\sqrt{1+x^2} v x = ( x , 1 ) / 1 + x 2 , and set v ∞ = ( 1 , 0 ) v_\infty=(1,0) v ∞ = ( 1 , 0 ) . For an invertible real 2 × 2 2 \times2 2 × 2 matrix A A A and x ∈ R P 1 x \in\mathbb{R}\mathbb{P}^{1} x ∈ R P 1 , let A ⋅ x ∈ R P 1 A\cdot x \in\mathbb{R}\mathbb{P}^{1} A ⋅ x ∈ R P 1 denote the point corresponding to the line through A v x Av_x A v x . The number ∥ A v x ∥ \|Av_x\| ∥ A v x ∥ does not depend on the choice of sign of the unit vector representing x x x .
(iii) For b ∈ R b \in\mathbb{R} b ∈ R and β > 0 \beta> 0 β > 0 , define the projective map f b , β ( x ) = M ( b , β ) ⋅ x f_{b,\beta}(x)=M(b,\beta)\cdot x f b , β ( x ) = M ( b , β ) ⋅ x . In coordinates, we have
f b , β ( x ) = ( b − β x ) − 1 ( x ∈ R , x ≠ b / β ) , f b , β ( b / β ) = ∞ , f b , β ( ∞ ) = 0. f_{b,\beta}(x)=(b-\beta x)^{-1}\quad(x\in\mathbb{R},\ x\ne b/\beta),\qquad f_{b,\beta}(b/\beta)=\infty,\qquad f_{b,\beta}(\infty)=0. f b , β ( x ) = ( b − β x ) − 1 ( x ∈ R , x = b / β ) , f b , β ( b / β ) = ∞ , f b , β ( ∞ ) = 0. (iv) Let m m m be the Borel probability measure on R P 1 \mathbb{R}\mathbb{P}^{1} R P 1 given by m ( d x ) = π − 1 ( 1 + x 2 ) − 1 d x m(dx)=\pi^{-1}(1+x^2)^{-1}dx m ( d x ) = π − 1 ( 1 + x 2 ) − 1 d x on R \mathbb{R} R and m ( { ∞ } ) = 0 m(\{\infty\})=0 m ({ ∞ }) = 0 , where π = 3.14 … \pi=3.14\ldots π = 3.14 … (elsewhere, π \pi π denotes a real solution); it is the image of the uniform probability measure on the unit circle under the map sending a unit vector to the line it spans. For an invertible real 2 × 2 2 \times2 2 × 2 matrix A A A and s ∈ [ 0 , 2 ] s\in[0,2] s ∈ [ 0 , 2 ] , define
J s ( A ) = ∫ R P 1 ∥ A v x ∥ − s m ( d x ) . J_s(A)=\int_{\mathbb{R}\mathbb{P}^{1}}\|Av_x\|^{-s}m(dx). J s ( A ) = ∫ R P 1 ∥ A v x ∥ − s m ( d x ) . (v) For s ∈ ( 0 , 2 ) s\in(0,2) s ∈ ( 0 , 2 ) , define e s ( x ) = ∣ x ∣ s e_s(x)=|x|^s e s ( x ) = ∣ x ∣ s for x ∈ R x\in\mathbb{R} x ∈ R , and define h s : R P 1 → [ 0 , 1 ] h_s:\mathbb{R}\mathbb{P}^{1}\to[0,1] h s : R P 1 → [ 0 , 1 ] by
h s ( x ) = ∣ x ∣ s ( 1 + x 2 ) s / 2 ( x ∈ R ) , h s ( ∞ ) = 1. h_s(x)=\frac{|x|^s}{(1+x^2)^{s/2}}\quad(x\in\mathbb{R}),\qquad h_s(\infty)=1. h s ( x ) = ( 1 + x 2 ) s /2 ∣ x ∣ s ( x ∈ R ) , h s ( ∞ ) = 1. The function h s h_s h s is continuous on R P 1 \mathbb{R}\mathbb{P}^{1} R P 1 , vanishes only at 0 0 0 , and e s = h s ( 1 + x 2 ) s / 2 e_s=h_s(1+x^2)^{s/2} e s = h s ( 1 + x 2 ) s /2 on R \mathbb{R} R .
(vi) For s ∈ ( 0 , 2 ) s\in(0,2) s ∈ ( 0 , 2 ) and a Borel function F : R → [ 0 , ∞ ] F:\mathbb{R}\to[0,\infty] F : R → [ 0 , ∞ ] , define the transfer operator in the x x x -coordinate
( A s F ) ( x ) = p E [ β s / 2 ∣ x ∣ s F ( f B , β ( x ) ) ] , x ∈ R . (\mathcal{A}_sF)(x)=p\,\mathbb{E}\left[\beta^{s/2}|x|^sF\left(f_{B,\beta}(x)\right)\right],\qquad x\in\mathbb{R}. ( A s F ) ( x ) = p E [ β s /2 ∣ x ∣ s F ( f B , β ( x ) ) ] , x ∈ R . This is well defined, because f B , β ( x ) ∈ R f_{B,\beta}(x)\in\mathbb{R} f B , β ( x ) ∈ R almost surely for every x ∈ R x\in\mathbb{R} x ∈ R : for x = 0 x=0 x = 0 we have f B , β ( 0 ) = 1 / B f_{B,\beta}(0)=1/B f B , β ( 0 ) = 1/ B and P ( B = 0 ) = 0 \mathbb{P}(B=0)=0 P ( B = 0 ) = 0 , and for x ≠ 0 x\ne0 x = 0 we have f B , β ( x ) = ∞ f_{B,\beta}(x)=\infty f B , β ( x ) = ∞ only if β = B / x \beta=B/x β = B / x , an event of probability zero since β \beta β has a density and is independent of B B B .
(vii) For s ∈ ( 0 , 2 ) s \in(0,2) s ∈ ( 0 , 2 ) , a bounded Borel function F : R P 1 → C F : \mathbb{RP}^{1} \to\mathbb{C} F : RP 1 → C , and x ∈ R P 1 x \in\mathbb{RP}^{1} x ∈ RP 1 , define the transfer operator
( Q s F ) ( x ) = p E [ ∥ M ( B , β ) v x ∥ − s F ( f B , β ( x ) ) ] . (Q_sF)(x) = p\,\mathbb{E}\left[\lVert M(B,\beta)v_x\rVert^{-s}F(f_{B,\beta}(x))\right]. ( Q s F ) ( x ) = p E [ ∥ M ( B , β ) v x ∥ − s F ( f B , β ( x )) ] .
The expectation converges absolutely by Lemma 6.5 below.
For x ∈ R x \in\mathbb{R} x ∈ R , we have M ( b , β ) ( x , 1 ) T = β − 1 / 2 ( 1 , b − β x ) T M(b,\beta)(x,1)^{\mathsf{T}} = \beta^{-1/2}(1,b-\beta x)^{\mathsf{T}} M ( b , β ) ( x , 1 ) T = β − 1/2 ( 1 , b − β x ) T and M ( b , β ) ( 1 , 0 ) T = β − 1 / 2 ( 0 , − β ) T M(b,\beta)(1,0)^{\mathsf{T}} = \beta^{-1/2}(0,-\beta)^{\mathsf{T}} M ( b , β ) ( 1 , 0 ) T = β − 1/2 ( 0 , − β ) T . Taking norms, we find
∥ M ( b , β ) v x ∥ − s = β s / 2 ( 1 + x 2 ) s / 2 ( 1 + ( b − β x ) 2 ) s / 2 ( x ∈ R ) , ∥ M ( b , β ) v ∞ ∥ − s = β − s / 2 . (75) \lVert M(b,\beta)v_x\rVert^{-s} = \frac{\beta^{s/2}(1+x^2)^{s/2}}{\left(1+(b-\beta x)^2\right)^{s/2}} \quad(x \in\mathbb{R}), \qquad\lVert M(b,\beta)v_\infty\rVert^{-s} = \beta^{-s/2}. \tag*{(75)} ∥ M ( b , β ) v x ∥ − s = ( 1 + ( b − β x ) 2 ) s /2 β s /2 ( 1 + x 2 ) s /2 ( x ∈ R ) , ∥ M ( b , β ) v ∞ ∥ − s = β − s /2 . (75)
For fixed ( b , β ) (b,\beta) ( b , β ) , the functions x ↦ ∥ M ( b , β ) v x ∥ − s x \mapsto\lVert M(b,\beta)v_x\rVert^{-s} x ↦ ∥ M ( b , β ) v x ∥ − s and x ↦ f b , β ( x ) x \mapsto f_{b,\beta}(x) x ↦ f b , β ( x ) are continuous on R P 1 \mathbb{RP}^{1} RP 1 .
The next lemma collects two elementary facts about the projective action. Part (i) is the identity that makes the norm multipliers in Q s Q_s Q s telescope along products, and part (ii) computes the projective Jacobian.
Lemma 6.2. Let A A A and A ′ A' A ′ be real 2 × 2 2 \times2 2 × 2 matrices with determinant one.
(i) For every x ∈ R P 1 x \in\mathbb{RP}^{1} x ∈ RP 1 , we have ( A ′ A ) ⋅ x = A ′ ⋅ ( A ⋅ x ) (A'A)\cdot x = A'\cdot(A\cdot x) ( A ′ A ) ⋅ x = A ′ ⋅ ( A ⋅ x ) and
∥ A ′ A v x ∥ = ∥ A ′ v A ⋅ x ∥ ∥ A v x ∥ . (76) \lVert A'Av_x\rVert= \lVert A'v_{A\cdot x}\rVert\lVert Av_x\rVert. \tag*{(76)} ∥ A ′ A v x ∥ = ∥ A ′ v A ⋅ x ∥ ∥ A v x ∥ . (76) (ii) For every Borel function F : R P 1 → [ 0 , ∞ ] F : \mathbb{RP}^{1} \to[0,\infty] F : RP 1 → [ 0 , ∞ ] , we have
∫ R P 1 F ( A ⋅ x ) ∥ A v x ∥ − 2 m ( d x ) = ∫ R P 1 F ( y ) m ( d y ) . (77) \int_{\mathbb{RP}^{1}} F(A\cdot x)\lVert Av_x\rVert^{-2}m(\mathrm{d}x) = \int_{\mathbb{RP}^{1}} F(y)m(\mathrm{d}y). \tag*{(77)} ∫ RP 1 F ( A ⋅ x ) ∥ A v x ∥ − 2 m ( d x ) = ∫ RP 1 F ( y ) m ( d y ) . (77)
Proof. For (i), since the vector A v x Av_x A v x spans the line A ⋅ x A\cdot x A ⋅ x , we have A v x = ± ∥ A v x ∥ v A ⋅ x Av_x = \pm\lVert Av_x\rVert v_{A\cdot x} A v x = ± ∥ A v x ∥ v A ⋅ x . Applying A ′ A' A ′ , we obtain A ′ A v x = ± ∥ A v x ∥ A ′ v A ⋅ x A'Av_x = \pm\lVert Av_x\rVert A'v_{A\cdot x} A ′ A v x = ± ∥ A v x ∥ A ′ v A ⋅ x . This vector spans the line A ′ ⋅ ( A ⋅ x ) A'\cdot(A\cdot x) A ′ ⋅ ( A ⋅ x ) , and taking norms gives (6.4).
For (ii), we parametrize R P 1 \mathbb{RP}^{1} RP 1 by angles. For θ ∈ R \theta\in\mathbb{R} θ ∈ R , let u θ = ( cos θ , sin θ ) u_\theta= (\cos\theta,\sin\theta) u θ = ( cos θ , sin θ ) . The map θ ↦ [ u θ ] \theta\mapsto[u_\theta] θ ↦ [ u θ ] from [ 0 , π ) [0,\pi) [ 0 , π ) to R P 1 \mathbb{RP}^{1} RP 1 is a bijection, and the image of π − 1 d θ \pi^{-1}\mathrm{d}\theta π − 1 d θ under it is m m m ; indeed, u θ u_\theta u θ spans the line of x = cot θ x=\cot\theta x = cot θ , and d x = − ( 1 + x 2 ) d θ \mathrm{d}x=-(1+x^2)\mathrm{d}\theta d x = − ( 1 + x 2 ) d θ . The map u ↦ A u / ∥ A u ∥ u \mapsto Au/\lVert Au\rVert u ↦ A u / ∥ A u ∥ is a homeomorphism of the unit circle, with inverse u ↦ A − 1 u / ∥ A − 1 u ∥ u \mapsto A^{-1}u/\lVert A^{-1}u\rVert u ↦ A − 1 u / ∥ A − 1 u ∥ . Let Θ : R → R \Theta:\mathbb{R}\to\mathbb{R} Θ : R → R be a continuous lift, so that A u θ = ∥ A u θ ∥ u Θ ( θ ) Au_\theta=\lVert Au_\theta\rVert u_{\Theta(\theta)} A u θ = ∥ A u θ ∥ u Θ ( θ ) . Writing w ( θ ) = A u θ w(\theta)=Au_\theta w ( θ ) = A u θ , we have
Θ ′ ( θ ) = det ( w ( θ ) , w ′ ( θ ) ) ∥ w ( θ ) ∥ 2 = det ( A ) det ( u θ , u θ ′ ) ∥ A u θ ∥ 2 = ∥ A u θ ∥ − 2 > 0 , \Theta'(\theta) = \frac{\det\left(w(\theta),w'(\theta)\right)}{\lVert w(\theta)\rVert^2} = \frac{\det(A)\det\left(u_\theta,u'_\theta\right)}{\lVert Au_\theta\rVert^2} = \lVert Au_\theta\rVert^{-2} > 0, Θ ′ ( θ ) = ∥ w ( θ ) ∥ 2 det ( w ( θ ) , w ′ ( θ ) ) = ∥ A u θ ∥ 2 det ( A ) det ( u θ , u θ ′ ) = ∥ A u θ ∥ − 2 > 0 ,
where we used that w ′ ( θ ) = A u θ ′ w'(\theta)=Au'_\theta w ′ ( θ ) = A u θ ′ and det ( u θ , u θ ′ ) = 1 \det(u_\theta,u'_\theta)=1 det ( u θ , u θ ′ ) = 1 . Since a homeomorphism of the circle has degree ± 1 \pm1 ± 1 and Θ \Theta Θ is increasing, we have Θ ( θ + 2 π ) = Θ ( θ ) + 2 π \Theta(\theta+2\pi)=\Theta(\theta)+2\pi Θ ( θ + 2 π ) = Θ ( θ ) + 2 π . Further, since A u θ + π = − A u θ Au_{\theta+\pi}=-Au_\theta A u θ + π = − A u θ , the continuous function θ ↦ Θ ( θ + π ) − Θ ( θ ) \theta\mapsto\Theta(\theta+\pi)-\Theta(\theta) θ ↦ Θ ( θ + π ) − Θ ( θ ) takes values in π + 2 π Z \pi+2\pi\mathbb{Z} π + 2 π Z , and it is therefore constant; since twice this constant equals Θ ( θ + 2 π ) − Θ ( θ ) = 2 π \Theta(\theta+2\pi)-\Theta(\theta)=2\pi Θ ( θ + 2 π ) − Θ ( θ ) = 2 π , it equals π \pi π . We conclude that Θ \Theta Θ is an increasing C 1 C^1 C 1 bijection of R \mathbb{R} R with Θ ( θ + π ) = Θ ( θ ) + π \Theta(\theta+\pi)=\Theta(\theta)+\pi Θ ( θ + π ) = Θ ( θ ) + π , and it induces the projective action of A A A on R / π Z ≃ R P 1 \mathbb{R}/\pi\mathbb{Z}\simeq\mathbb{R}\mathrm{P}^1 R / π Z ≃ R P 1 . For G ( θ ) = F ( [ u θ ] ) G(\theta)=F([u_\theta]) G ( θ ) = F ([ u θ ]) , which is π \pi π -periodic, the change of variables ϑ = Θ ( θ ) \vartheta=\Theta(\theta) ϑ = Θ ( θ ) gives
∫ 0 π G ( Θ ( θ ) ) ∥ A u θ ∥ − 2 d θ = ∫ Θ ( 0 ) Θ ( 0 ) + π G ( ϑ ) d ϑ = ∫ 0 π G ( ϑ ) d ϑ . \int_0^\pi G(\Theta(\theta))\lVert Au_\theta\rVert^{-2}\,\mathrm{d}\theta=\int_{\Theta(0)}^{\Theta(0)+\pi}G(\vartheta)\,\mathrm{d}\vartheta=\int_0^\pi G(\vartheta)\,\mathrm{d}\vartheta. ∫ 0 π G ( Θ ( θ )) ∥ A u θ ∥ − 2 d θ = ∫ Θ ( 0 ) Θ ( 0 ) + π G ( ϑ ) d ϑ = ∫ 0 π G ( ϑ ) d ϑ .
Dividing by π \pi π gives (77) .
Lemma 6.3. Let A A A be a real 2 × 2 2\times2 2 × 2 matrix with determinant one. Then J 0 ( A ) = J 2 ( A ) = 1 J_0(A)=J_2(A)=1 J 0 ( A ) = J 2 ( A ) = 1 , and for every s ∈ [ 0 , 2 ] s\in[0,2] s ∈ [ 0 , 2 ] , we have
J s ( A ) = J 2 − s ( A − 1 ) = J 2 − s ( A ) . (78) J_s(A)=J_{2-s}(A^{-1})=J_{2-s}(A). \tag*{(78)} J s ( A ) = J 2 − s ( A − 1 ) = J 2 − s ( A ) . (78) Further, s ↦ log J s ( A ) s\mapsto\log J_s(A) s ↦ log J s ( A ) is convex on [ 0 , 2 ] [0,2] [ 0 , 2 ] , and J s ( A ) ≤ 1 J_s(A)\leq1 J s ( A ) ≤ 1 for s ∈ [ 0 , 2 ] s\in[0,2] s ∈ [ 0 , 2 ] .
Proof. We have J 0 ( A ) = m ( R P 1 ) = 1 J_0(A)=m(\mathbb{R}\mathrm{P}^1)=1 J 0 ( A ) = m ( R P 1 ) = 1 , and J 2 ( A ) = 1 J_2(A)=1 J 2 ( A ) = 1 by (77) with F = 1 F=1 F = 1 . Fix s ∈ [ 0 , 2 ] s\in[0,2] s ∈ [ 0 , 2 ] and x ∈ R P 1 x\in\mathbb{R}\mathrm{P}^1 x ∈ R P 1 , and set y = A ⋅ x y=A\cdot x y = A ⋅ x . By (76) applied to A − 1 A = I A^{-1}A=I A − 1 A = I , we have 1 = ∥ A − 1 v y ∥ ∥ A v x ∥ 1=\lVert A^{-1}v_y\rVert\lVert Av_x\rVert 1 = ∥ A − 1 v y ∥ ∥ A v x ∥ , and we can write
∥ A v x ∥ − s = ∥ A v x ∥ − 2 ∥ A v x ∥ 2 − s = ∥ A v x ∥ − 2 ∥ A − 1 v A ⋅ x ∥ − ( 2 − s ) . \lVert Av_x\rVert^{-s}=\lVert Av_x\rVert^{-2}\lVert Av_x\rVert^{2-s}=\lVert Av_x\rVert^{-2}\lVert A^{-1}v_{A\cdot x}\rVert^{-(2-s)}. ∥ A v x ∥ − s = ∥ A v x ∥ − 2 ∥ A v x ∥ 2 − s = ∥ A v x ∥ − 2 ∥ A − 1 v A ⋅ x ∥ − ( 2 − s ) .
Integrating against m m m and using (77) with F ( y ) = ∥ A − 1 v y ∥ − ( 2 − s ) F(y)=\lVert A^{-1}v_y\rVert^{-(2-s)} F ( y ) = ∥ A − 1 v y ∥ − ( 2 − s ) , we find J s ( A ) = J 2 − s ( A − 1 ) J_s(A)=J_{2-s}(A^{-1}) J s ( A ) = J 2 − s ( A − 1 ) .
To prove that J t ( A − 1 ) = J t ( A ) J_t(A^{-1})=J_t(A) J t ( A − 1 ) = J t ( A ) for t ∈ [ 0 , 2 ] t\in[0,2] t ∈ [ 0 , 2 ] , we use the singular value decomposition. There exist orthogonal matrices U , V U,V U , V and σ ≥ 1 \sigma\geq1 σ ≥ 1 such that A = U diag ( σ , σ − 1 ) V T A=U\operatorname{diag}(\sigma,\sigma^{-1})V^{\mathsf{T}} A = U diag ( σ , σ − 1 ) V T ; the two singular values are σ \sigma σ and σ − 1 \sigma^{-1} σ − 1 because det A = 1 \det A=1 det A = 1 . Then A − 1 = V diag ( σ − 1 , σ ) U T A^{-1}=V\operatorname{diag}(\sigma^{-1},\sigma)U^{\mathsf{T}} A − 1 = V diag ( σ − 1 , σ ) U T . The quantity J t J_t J t is invariant under multiplication by orthogonal matrices on either side: on the left because ∥ U A v ∥ = ∥ A v ∥ \lVert UAv\rVert=\lVert Av\rVert ∥ U A v ∥ = ∥ A v ∥ , and on the right because ∥ A V v x ∥ = ∥ A v V ⋅ x ∥ \lVert AVv_x\rVert=\lVert Av_{V\cdot x}\rVert ∥ A V v x ∥ = ∥ A v V ⋅ x ∥ for orthogonal V V V and the projective action of an orthogonal matrix preserves m m m (it is induced by a rotation or a reflection of the unit circle). It follows that J t ( A ) = J t ( D σ ) J_t(A)=J_t(D_\sigma) J t ( A ) = J t ( D σ ) and J t ( A − 1 ) = J t ( D σ − 1 ) J_t(A^{-1})=J_t(D_\sigma^{-1}) J t ( A − 1 ) = J t ( D σ − 1 ) , where D σ = diag ( σ , σ − 1 ) D_\sigma=\operatorname{diag}(\sigma,\sigma^{-1}) D σ = diag ( σ , σ − 1 ) . Let R R R be the rotation by π / 2 \pi/2 π /2 . Then D σ − 1 = R D σ R − 1 D_\sigma^{-1}=RD_\sigma R^{-1} D σ − 1 = R D σ R − 1 , and the same invariance implies J t ( D σ − 1 ) = J t ( D σ ) J_t(D_\sigma^{-1})=J_t(D_\sigma) J t ( D σ − 1 ) = J t ( D σ ) . This proves (78) .
For s 0 , s 1 ∈ [ 0 , 2 ] s_0,s_1\in[0,2] s 0 , s 1 ∈ [ 0 , 2 ] and τ ∈ ( 0 , 1 ) \tau\in(0,1) τ ∈ ( 0 , 1 ) , Hölder’s inequality with exponents 1 / ( 1 − τ ) 1/(1-\tau) 1/ ( 1 − τ ) and 1 / τ 1/\tau 1/ τ gives
J ( 1 − τ ) s 0 + τ s 1 ( A ) = ∫ ( ∥ A v x ∥ − s 0 ) 1 − τ ( ∥ A v x ∥ − s 1 ) τ m ( d x ) ≤ J s 0 ( A ) 1 − τ J s 1 ( A ) τ . J_{(1-\tau)s_0+\tau s_1}(A)=\int\bigl(\lVert Av_x\rVert^{-s_0}\bigr)^{1-\tau}\bigl(\lVert Av_x\rVert^{-s_1}\bigr)^\tau\,m(\mathrm{d}x)\leq J_{s_0}(A)^{1-\tau}J_{s_1}(A)^\tau. J ( 1 − τ ) s 0 + τ s 1 ( A ) = ∫ ( ∥ A v x ∥ − s 0 ) 1 − τ ( ∥ A v x ∥ − s 1 ) τ m ( d x ) ≤ J s 0 ( A ) 1 − τ J s 1 ( A ) τ .
which is the convexity of log J s ( A ) \log J_s(A) log J s ( A ) . Taking s 0 = 0 s_0=0 s 0 = 0 and s 1 = 2 s_1=2 s 1 = 2 , we find J s ( A ) ≤ J 0 ( A ) 1 − s / 2 J 2 ( A ) s / 2 = 1 J_s(A)\leq J_0(A)^{1-s/2}J_2(A)^{s/2}=1 J s ( A ) ≤ J 0 ( A ) 1 − s /2 J 2 ( A ) s /2 = 1 .
For exponents below one, the angular integral J s ( A ) J_s(A) J s ( A ) is comparable to a power of the norm of A A A .
Lemma 6.4. For every s ∈ ( 0 , 1 ) s \in(0,1) s ∈ ( 0 , 1 ) , there exists C s < ∞ C_s < \infty C s < ∞ such that for every real 2 × 2 2 \times2 2 × 2 matrix A A A with determinant one, setting t = ∥ A ∥ t = \lVert A\rVert t = ∥ A ∥ , we have
t − s ≤ J s ( A ) ≤ C s t − s . t^{-s} \le J_s(A) \le C_s t^{-s}. t − s ≤ J s ( A ) ≤ C s t − s . One can take C s = π − 1 ∫ 0 π ∣ cos θ ∣ − s d θ C_s = \pi^{-1} \int_0^\pi\lvert\cos\theta\rvert^{-s}\,\mathrm{d}\theta C s = π − 1 ∫ 0 π ∣ cos θ ∣ − s d θ .
Proof. The lower bound follows from ∥ A v x ∥ ≤ t \lVert Av_x\rVert\le t ∥ A v x ∥ ≤ t for every x x x . For the upper bound, write A = U diag ( t , t − 1 ) V T A = U\operatorname{diag}(t,t^{-1})V^{\mathsf T} A = U diag ( t , t − 1 ) V T with U , V U,V U , V orthogonal. As in the proof of Lemma 6.3, we have J s ( A ) = J s ( diag ( t , t − 1 ) ) J_s(A) = J_s(\operatorname{diag}(t,t^{-1})) J s ( A ) = J s ( diag ( t , t − 1 )) . In the angular parametrization u θ = ( cos θ , sin θ ) u_\theta= (\cos\theta,\sin\theta) u θ = ( cos θ , sin θ ) , we have ∥ diag ( t , t − 1 ) u θ ∥ ≥ t ∣ cos θ ∣ \lVert\operatorname{diag}(t,t^{-1})u_\theta\rVert\ge t\lvert\cos\theta\rvert ∥ diag ( t , t − 1 ) u θ ∥ ≥ t ∣ cos θ ∣ , and therefore
J s ( A ) ≤ t − s π ∫ 0 π ∣ cos θ ∣ − s d θ = C s t − s . J_s(A) \le\frac{t^{-s}}{\pi}\int_0^\pi\lvert\cos\theta\rvert^{-s}\,\mathrm{d}\theta= C_s t^{-s}. J s ( A ) ≤ π t − s ∫ 0 π ∣ cos θ ∣ − s d θ = C s t − s .
The constant C s C_s C s is finite because s < 1 s < 1 s < 1 .
The Perron–Frobenius eigenvalue of the transfer operator
Boundedness and holomorphy
We begin with a bound on the expected norm multiplier of one factor M ( b , β ) M(b,\beta) M ( b , β ) , uniform over the input direction.
Lemma 6.5. For s ∈ ( 0 , 2 ) s \in(0,2) s ∈ ( 0 , 2 ) and K > 0 K > 0 K > 0 , set
C s , K = 2 s / 2 K − s / 2 Γ ( 1 + s / 2 ) + 2 s K s / 2 Γ ( 1 − s / 2 ) . C_{s,K}=2^{s/2}K^{-s/2}\Gamma(1+s/2)+2^sK^{s/2}\Gamma(1-s/2). C s , K = 2 s /2 K − s /2 Γ ( 1 + s /2 ) + 2 s K s /2 Γ ( 1 − s /2 ) . Then for every b ∈ R b \in\mathbb{R} b ∈ R , we have
sup x ∈ R P 1 ∫ 0 ∞ K e − K t ∥ M ( b , t ) v x ∥ − s d t ≤ C s , K ( 1 + ∣ b ∣ s / 2 ) . (79) \sup_{x\in\mathbb{R}\mathbb{P}^{1}}\int_0^\infty Ke^{-Kt}\lVert M(b,t)v_x\rVert^{-s}\,\mathrm{d}t \le C_{s,K}(1+\lvert b\rvert^{s/2}). \tag*{(79)} x ∈ R P 1 sup ∫ 0 ∞ K e − K t ∥ M ( b , t ) v x ∥ − s d t ≤ C s , K ( 1 + ∣ b ∣ s /2 ) . (79) The constant C s , K C_{s,K} C s , K is continuous in ( s , K ) ∈ ( 0 , 2 ) × ( 0 , ∞ ) (s,K) \in(0,2)\times(0,\infty) ( s , K ) ∈ ( 0 , 2 ) × ( 0 , ∞ ) .
Proof. Let r = s / 2 ∈ ( 0 , 1 ) r=s/2 \in(0,1) r = s /2 ∈ ( 0 , 1 ) , and let I ( x ) I(x) I ( x ) denote the integral on the left side of (79) . By (6.3), we have I ( ∞ ) = K ∫ 0 ∞ t − r e − K t d t = K r Γ ( 1 − r ) I(\infty)=K\int_0^\infty t^{-r}e^{-Kt}\,\mathrm{d}t=K^r\Gamma(1-r) I ( ∞ ) = K ∫ 0 ∞ t − r e − K t d t = K r Γ ( 1 − r ) . For x ∈ R x\in\mathbb{R} x ∈ R with ∣ x ∣ ≤ 1 \lvert x\rvert\le1 ∣ x ∣ ≤ 1 , we discard the denominator in (6.3) and obtain I ( x ) ≤ 2 r ∫ 0 ∞ K t r e − K t d t = 2 r K − r Γ ( 1 + r ) I(x)\le2^r\int_0^\infty Kt^r e^{-Kt}\,\mathrm{d}t=2^rK^{-r}\Gamma(1+r) I ( x ) ≤ 2 r ∫ 0 ∞ K t r e − K t d t = 2 r K − r Γ ( 1 + r ) . Now let ∣ x ∣ = a ≥ 1 \lvert x\rvert=a\ge1 ∣ x ∣ = a ≥ 1 and σ = sgn x \sigma=\operatorname{sgn}x σ = sgn x . Substituting u = a t u=at u = a t in the integral defining I ( x ) I(x) I ( x ) and using (6.3), we have
I ( x ) = K ( 1 + a 2 ) r a 1 + r ∫ 0 ∞ e − K u / a u r ( 1 + ( b − σ u ) 2 ) r d u . (80) I(x)=K\frac{(1+a^2)^r}{a^{1+r}}\int_0^\infty e^{-Ku/a}\frac{u^r}{(1+(b-\sigma u)^2)^r}\,\mathrm{d}u. \tag*{(80)} I ( x ) = K a 1 + r ( 1 + a 2 ) r ∫ 0 ∞ e − K u / a ( 1 + ( b − σ u ) 2 ) r u r d u . (80)
Since u = ∣ σ u ∣ ≤ ∣ b ∣ + ∣ b − σ u ∣ u = |\sigma u| \le|b| + |b-\sigma u| u = ∣ σ u ∣ ≤ ∣ b ∣ + ∣ b − σ u ∣ and r < 1 r < 1 r < 1 , we have u r ≤ ∣ b ∣ r + ∣ b − σ u ∣ r u^r \le|b|^r + |b-\sigma u|^r u r ≤ ∣ b ∣ r + ∣ b − σ u ∣ r . Using ∣ b − σ u ∣ r ≤ ( 1 + ( b − σ u ) 2 ) r / 2 |b-\sigma u|^r \le(1+(b-\sigma u)^2)^{r/2} ∣ b − σ u ∣ r ≤ ( 1 + ( b − σ u ) 2 ) r /2 and ( 1 + ( b − σ u ) 2 ) − r ≤ ( 1 + ( b − σ u ) 2 ) − r / 2 (1+(b-\sigma u)^2)^{-r} \le(1+(b-\sigma u)^2)^{-r/2} ( 1 + ( b − σ u ) 2 ) − r ≤ ( 1 + ( b − σ u ) 2 ) − r /2 , we have
u r ( 1 + ( b − σ u ) 2 ) r ≤ ( 1 + ∣ b ∣ r ) H ( b − σ u ) , H ( v ) = ( 1 + v 2 ) − r / 2 . \frac{u^r}{(1+(b-\sigma u)^2)^r} \le(1+|b|^r)H(b-\sigma u), \qquad H(v)=(1+v^2)^{-r/2}. ( 1 + ( b − σ u ) 2 ) r u r ≤ ( 1 + ∣ b ∣ r ) H ( b − σ u ) , H ( v ) = ( 1 + v 2 ) − r /2 .
The function H H H is even and nonincreasing on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . Hence ∫ J ′ H ≤ ∫ − ℓ / 2 ℓ / 2 H \int_{J'}H\le\int_{-\ell/2}^{\ell/2}H ∫ J ′ H ≤ ∫ − ℓ /2 ℓ /2 H for every interval J ′ ⊂ R J'\subset\mathbb{R} J ′ ⊂ R of length ℓ \ell ℓ . (Indeed, if J J J denotes the centered interval of length ℓ \ell ℓ , then J ′ ∖ J J'\setminus J J ′ ∖ J and J ∖ J ′ J\setminus J' J ∖ J ′ have the same length, and H ≤ H ( ℓ / 2 ) H\le H(\ell/2) H ≤ H ( ℓ /2 ) on J ′ ∖ J J'\setminus J J ′ ∖ J while H ≥ H ( ℓ / 2 ) H\ge H(\ell/2) H ≥ H ( ℓ /2 ) on J ∖ J ′ J\setminus J' J ∖ J ′ .) We write e − K u / a = ∫ u ∞ ( K / a ) e − K t / a d t e^{-Ku/a}=\int_u^\infty(K/a)e^{-Kt/a}\,dt e − K u / a = ∫ u ∞ ( K / a ) e − K t / a d t and use Tonelli's theorem. Since { b − σ u : 0 ≤ u ≤ t } \{b-\sigma u:0\le u\le t\} { b − σ u : 0 ≤ u ≤ t } is an interval of length t t t , this gives
∫ 0 ∞ e − K u / a H ( b − σ u ) d u = ∫ 0 ∞ K a e − K t / a ∫ 0 t H ( b − σ u ) d u d t ≤ ∫ 0 ∞ K a e − K t / a 2 ∫ 0 t / 2 H ( v ) d v d t = 2 ∫ 0 ∞ H ( v ) e − 2 K v / a d v ≤ 2 ∫ 0 ∞ v − r e − 2 K v / a d v = 2 Γ ( 1 − r ) ( a 2 K ) 1 − r . \begin{aligned}
\int_0^\infty e^{-Ku/a}H(b-\sigma u)\,du
&=\int_0^\infty\frac{K}{a}e^{-Kt/a}\int_0^t H(b-\sigma u)\,du\,dt\\
&\le\int_0^\infty\frac{K}{a}e^{-Kt/a}2\int_0^{t/2}H(v)\,dv\,dt\\
&=2\int_0^\infty H(v)e^{-2Kv/a}\,dv\\
&\le2\int_0^\infty v^{-r}e^{-2Kv/a}\,dv
=2\Gamma(1-r)\left(\frac{a}{2K}\right)^{1-r}.
\end{aligned} ∫ 0 ∞ e − K u / a H ( b − σ u ) d u = ∫ 0 ∞ a K e − K t / a ∫ 0 t H ( b − σ u ) d u d t ≤ ∫ 0 ∞ a K e − K t / a 2 ∫ 0 t /2 H ( v ) d v d t = 2 ∫ 0 ∞ H ( v ) e − 2 K v / a d v ≤ 2 ∫ 0 ∞ v − r e − 2 K v / a d v = 2Γ ( 1 − r ) ( 2 K a ) 1 − r .
Inserting this bound into (6.8) and using ( 1 + a 2 ) r ≤ 2 r a 2 r (1+a^2)^r\le2^r a^{2r} ( 1 + a 2 ) r ≤ 2 r a 2 r , we find
I ( x ) ≤ K 2 r a r − 1 ⋅ 2 Γ ( 1 − r ) ( a 2 K ) 1 − r ( 1 + ∣ b ∣ r ) = 4 r K r Γ ( 1 − r ) ( 1 + ∣ b ∣ r ) . I(x)\le K2^r a^{r-1}\cdot2\Gamma(1-r)\left(\frac{a}{2K}\right)^{1-r}(1+|b|^r)=4^rK^r\Gamma(1-r)(1+|b|^r). I ( x ) ≤ K 2 r a r − 1 ⋅ 2Γ ( 1 − r ) ( 2 K a ) 1 − r ( 1 + ∣ b ∣ r ) = 4 r K r Γ ( 1 − r ) ( 1 + ∣ b ∣ r ) .
Combining the three cases gives (6.7). The continuity of C s , K C_{s,K} C s , K is evident from its formula.
To prove analyticity in s s s , we extend Q s Q_s Q s to complex exponents. Let Σ = { z ∈ C : 0 < Re z < 2 } \Sigma=\{z\in\mathbb{C}:0<\operatorname{Re}z<2\} Σ = { z ∈ C : 0 < Re z < 2 } . For z ∈ Σ z\in\Sigma z ∈ Σ , x ∈ R P 1 x\in\mathbb{RP}^1 x ∈ RP 1 , b ∈ R b\in\mathbb{R} b ∈ R , and β > 0 \beta>0 β > 0 , set
R z ( x ; b , β ) = exp ( − z log ∥ M ( b , β ) v x ∥ ) , R_z(x;b,\beta)=\exp\left(-z\log\|M(b,\beta)v_x\|\right), R z ( x ; b , β ) = exp ( − z log ∥ M ( b , β ) v x ∥ ) ,
so that ∣ R z ∣ = R Re z |R_z|=R_{\operatorname{Re}z} ∣ R z ∣ = R Re z and R s ( x ; b , β ) = ∥ M ( b , β ) v x ∥ − s R_s(x;b,\beta)=\|M(b,\beta)v_x\|^{-s} R s ( x ; b , β ) = ∥ M ( b , β ) v x ∥ − s for real s s s . For a bounded Borel function F : R P 1 → C F:\mathbb{RP}^1\to\mathbb{C} F : RP 1 → C , define
( Q z F ) ( x ) = p E [ R z ( x ; B , β ) F ( f B , β ( x ) ) ] , x ∈ R P 1 . (Q_zF)(x)=p\mathbb{E}\left[R_z(x;B,\beta)F\left(f_{B,\beta}(x)\right)\right],\qquad x\in\mathbb{RP}^1. ( Q z F ) ( x ) = p E [ R z ( x ; B , β ) F ( f B , β ( x ) ) ] , x ∈ RP 1 .
For real z = s ∈ ( 0 , 2 ) z=s\in(0,2) z = s ∈ ( 0 , 2 ) , this is the operator Q s Q_s Q s of Definition 6.1.
Lemma 6.6. Suppose that ν \nu ν satisfies (6.1).
(i) For every z ∈ Σ z\in\Sigma z ∈ Σ , the operator Q z Q_z Q z maps C ( R P 1 ) C(\mathbb{RP}^1) C ( RP 1 ) into C ( R P 1 ) C(\mathbb{RP}^1) C ( RP 1 ) , and
∥ Q z ∥ ≤ p C Re z , K ( 1 + E ∣ B ∣ Re z / 2 ) . \|Q_z\|\le pC_{\operatorname{Re}z,K}\left(1+\mathbb{E}|B|^{\operatorname{Re}z/2}\right). ∥ Q z ∥ ≤ p C Re z , K ( 1 + E ∣ B ∣ Re z /2 ) . The right side is bounded on every compact subset of $\Sigma.
(ii) For real s ∈ ( 0 , 2 ) s \in(0,2) s ∈ ( 0 , 2 ) , the operator Q s Q_s Q s is positive, that is, Q s F ≥ 0 Q_sF \ge0 Q s F ≥ 0 for F ≥ 0 F \ge0 F ≥ 0 , and ∥ Q s ∥ = ∥ Q s 1 ∥ ∞ \lVert Q_s\rVert= \lVert Q_s1\rVert_\infty ∥ Q s ∥ = ∥ Q s 1 ∥ ∞ . Further, Q s 1 ( x ) > 0 Q_s1(x) > 0 Q s 1 ( x ) > 0 for every x ∈ R P 1 x \in\mathbb{R}\mathbb{P}^1 x ∈ R P 1 .
(iii) The map z ↦ Q z z \mapsto Q_z z ↦ Q z from Σ \Sigma Σ to the bounded operators on C ( R P 1 ) C(\mathbb{R}\mathbb{P}^1) C ( R P 1 ) is holomorphic in the operator norm.
Proof. For σ ∈ ( 0 , 2 ) \sigma\in(0,2) σ ∈ ( 0 , 2 ) , conditioning on B B B and applying Lemma 6.5, we have
sup x ∈ R P 1 E [ R σ ( x ; B , β ) ] ≤ C σ , K ( 1 + E ∣ B ∣ σ / 2 ) < ∞ (81) \sup_{x \in\mathbb{R}\mathbb{P}^1} \mathbb{E}\left[R_\sigma(x;B,\beta)\right] \le C_{\sigma,K}\left(1+\mathbb{E}|B|^{\sigma/2}\right)<\infty \tag*{(81)} x ∈ R P 1 sup E [ R σ ( x ; B , β ) ] ≤ C σ , K ( 1 + E ∣ B ∣ σ /2 ) < ∞ (81)
where the finiteness follows from (6.1) because σ / 2 < 1 \sigma/2<1 σ /2 < 1 . Since E ∣ B ∣ σ / 2 \mathbb{E}|B|^{\sigma/2} E ∣ B ∣ σ /2 is a finite continuous function of σ ∈ ( 0 , 2 ) \sigma\in(0,2) σ ∈ ( 0 , 2 ) (by dominated convergence, with dominating function 1 + ∣ B ∣ σ + / 2 1+|B|^{\sigma_+/2} 1 + ∣ B ∣ σ + /2 on σ ≤ σ + < 2 \sigma\le\sigma_+<2 σ ≤ σ + < 2 ), the right side of (6.9) is bounded on compact subsets of ( 0 , 2 ) (0,2) ( 0 , 2 ) .
We now prove (i). Fix z ∈ Σ z \in\Sigma z ∈ Σ and F ∈ C ( R P 1 ) F \in C(\mathbb{R}\mathbb{P}^1) F ∈ C ( R P 1 ) , and write σ = Re z \sigma=\operatorname{Re}z σ = Re z . Since ∣ R z ∣ = R σ |R_z|=R_\sigma ∣ R z ∣ = R σ , (6.9) implies that ∣ Q z F ( x ) ∣ ≤ p C σ , K ( 1 + E ∣ B ∣ σ / 2 ) ∥ F ∥ ∞ |Q_zF(x)| \le pC_{\sigma,K}(1+\mathbb{E}|B|^{\sigma/2})\lVert F\rVert_\infty ∣ Q z F ( x ) ∣ ≤ p C σ , K ( 1 + E ∣ B ∣ σ /2 ) ∥ F ∥ ∞ for every x x x . It remains to show that Q z F Q_zF Q z F is continuous. Let x n → x x_n \to x x n → x in R P 1 \mathbb{R}\mathbb{P}^1 R P 1 . For each fixed ( b , β ) (b,\beta) ( b , β ) , the maps y ↦ R z ( y ; b , β ) y \mapsto R_z(y;b,\beta) y ↦ R z ( y ; b , β ) and y ↦ f b , β ( y ) y \mapsto f_{b,\beta}(y) y ↦ f b , β ( y ) are continuous on R P 1 \mathbb{R}\mathbb{P}^1 R P 1 by (6.3). Then the random variables Z n = R z ( x n ; B , β ) F ( f B , β ( x n ) ) Z_n=R_z(x_n;B,\beta)F(f_{B,\beta}(x_n)) Z n = R z ( x n ; B , β ) F ( f B , β ( x n )) converge almost surely to Z = R z ( x ; B , β ) F ( f B , β ( x ) ) Z=R_z(x;B,\beta)F(f_{B,\beta}(x)) Z = R z ( x ; B , β ) F ( f B , β ( x )) . Fix t ∈ ( σ , 2 ) t \in(\sigma,2) t ∈ ( σ , 2 ) . Then ∣ Z n ∣ t / σ ≤ ∥ F ∥ ∞ t / σ R t ( x n ; B , β ) |Z_n|^{t/\sigma}\le\lVert F\rVert_\infty^{t/\sigma}R_t(x_n;B,\beta) ∣ Z n ∣ t / σ ≤ ∥ F ∥ ∞ t / σ R t ( x n ; B , β ) , and (6.9) with σ \sigma σ replaced by t t t shows that sup n E ∣ Z n ∣ t / σ < ∞ \sup_n\mathbb{E}|Z_n|^{t/\sigma}<\infty sup n E ∣ Z n ∣ t / σ < ∞ . Since t / σ > 1 t/\sigma>1 t / σ > 1 , the sequence ( Z n ) (Z_n) ( Z n ) is uniformly integrable, and by Vitali’s convergence theorem, we have E Z n → E Z \mathbb{E}Z_n\to\mathbb{E}Z E Z n → E Z . This proves (i).
For (ii), positivity is immediate from the definition, and for real F F F we have ∣ Q s F ∣ ≤ Q s ∣ F ∣ ≤ ∥ F ∥ ∞ Q s 1 |Q_sF|\le Q_s|F|\le\lVert F\rVert_\infty Q_s1 ∣ Q s F ∣ ≤ Q s ∣ F ∣ ≤ ∥ F ∥ ∞ Q s 1 , which gives ∥ Q s ∥ = ∥ Q s 1 ∥ ∞ \lVert Q_s\rVert=\lVert Q_s1\rVert_\infty ∥ Q s ∥ = ∥ Q s 1 ∥ ∞ ; the same bound holds for complex F F F since ∣ Q s F ∣ ≤ Q s ∣ F ∣ |Q_sF|\le Q_s|F| ∣ Q s F ∣ ≤ Q s ∣ F ∣ . Since the integrand R s ( x ; B , β ) R_s(x;B,\beta) R s ( x ; B , β ) in Q s 1 ( x ) Q_s1(x) Q s 1 ( x ) is strictly positive, we have Q s 1 ( x ) > 0 Q_s1(x)>0 Q s 1 ( x ) > 0 .
For (iii), fix z ∈ Σ z \in\Sigma z ∈ Σ , let σ = Re z \sigma=\operatorname{Re}z σ = Re z , and choose ε > 0 \varepsilon>0 ε > 0 with [ σ − 2 ε , σ + 2 ε ] ⊂ ( 0 , 2 ) [\sigma-2\varepsilon,\sigma+2\varepsilon]\subset(0,2) [ σ − 2 ε , σ + 2 ε ] ⊂ ( 0 , 2 ) . We write ℓ = log ∥ M ( B , β ) v x ∥ \ell=\log\lVert M(B,\beta)v_x\rVert ℓ = log ∥ M ( B , β ) v x ∥ , and we define the operator
( D z F ) ( x ) = − p E [ ℓ R z ( x ; B , β ) F ( f B , β ( x ) ) ] . (D_zF)(x)=-p\mathbb{E}\left[\ell R_z(x;B,\beta)F\left(f_{B,\beta}(x)\right)\right]. ( D z F ) ( x ) = − p E [ ℓ R z ( x ; B , β ) F ( f B , β ( x ) ) ] .
For every w ∈ C w \in\mathbb{C} w ∈ C , we have ∣ e w − 1 − w ∣ ≤ ∣ w ∣ 2 e ∣ w ∣ / 2 |e^w-1-w|\le|w|^2e^{|w|}/2 ∣ e w − 1 − w ∣ ≤ ∣ w ∣ 2 e ∣ w ∣ /2 . Applying this with w = − h ℓ w=-h\ell w = − h ℓ for h ∈ C h \in\mathbb{C} h ∈ C with ∣ h ∣ ≤ ε / 2 |h|\le\varepsilon/2 ∣ h ∣ ≤ ε /2 , we obtain
∣ R z + h − R z + h ℓ R z ∣ = ∣ R z ∣ ∣ e − h ℓ − 1 + h ℓ ∣ ≤ ∣ h ∣ 2 2 ℓ 2 e ε ∣ ℓ ∣ / 2 R σ ≤ 8 ∣ h ∣ 2 ε 2 e ε ∣ ℓ ∣ R σ ≤ 8 ∣ h ∣ 2 ε 2 ( R σ − ε + R σ + ε ) . \begin{aligned}
\left|R_{z+h}-R_z+h\ell R_z\right|=|R_z|\left|e^{-h\ell}-1+h\ell\right|\le\frac{|h|^2}{2}\ell^2e^{\varepsilon|\ell|/2}R_\sigma\\
&\le\frac{8|h|^2}{\varepsilon^2}e^{\varepsilon|\ell|}R_\sigma\le\frac{8|h|^2}{\varepsilon^2}\left(R_{\sigma-\varepsilon}+R_{\sigma+\varepsilon}\right).
\end{aligned} ∣ R z + h − R z + h ℓ R z ∣ = ∣ R z ∣ e − h ℓ − 1 + h ℓ ≤ 2 ∣ h ∣ 2 ℓ 2 e ε ∣ ℓ ∣/2 R σ ≤ ε 2 8∣ h ∣ 2 e ε ∣ ℓ ∣ R σ ≤ ε 2 8∣ h ∣ 2 ( R σ − ε + R σ + ε ) .
In the second inequality, we used ℓ 2 e − ε ∣ ℓ ∣ / 2 ≤ 16 ε − 2 e − 2 ≤ 16 ε − 2 \ell^2e^{-\varepsilon|\ell|/2}\le16\varepsilon^{-2}e^{-2}\le16\varepsilon^{-2} ℓ 2 e − ε ∣ ℓ ∣/2 ≤ 16 ε − 2 e − 2 ≤ 16 ε − 2 , and in the third, we used e ε ∣ ℓ ∣ e − σ ℓ ≤ e − ( σ − ε ) ℓ + e − ( σ + ε ) ℓ e^{\varepsilon|\ell|}e^{-\sigma\ell}\le e^{-(\sigma-\varepsilon)\ell}+e^{-(\sigma+\varepsilon)\ell} e ε ∣ ℓ ∣ e − σ ℓ ≤ e − ( σ − ε ) ℓ + e − ( σ + ε ) ℓ . By (6.9), we conclude that
∥ Q z + h − Q z − h D z ∥ ≤ 8 p ∣ h ∣ 2 ε 2 sup x ∈ R P 1 E [ R σ − ε ( x ; B , β ) + R σ + ε ( x ; B , β ) ] ≤ C ∣ h ∣ 2 , \lVert Q_{z+h}-Q_z-hD_z\rVert\le\frac{8p|h|^2}{\varepsilon^2}\sup_{x\in\mathbb{R}\mathbb{P}^1}\mathbb{E}\left[R_{\sigma-\varepsilon}(x;B,\beta)+R_{\sigma+\varepsilon}(x;B,\beta)\right]\le C|h|^2, ∥ Q z + h − Q z − h D z ∥ ≤ ε 2 8 p ∣ h ∣ 2 x ∈ R P 1 sup E [ R σ − ε ( x ; B , β ) + R σ + ε ( x ; B , β ) ] ≤ C ∣ h ∣ 2 ,
with C C C independent of h h h . Since ∣ ℓ ∣ R σ ≤ ε − 1 ( R σ − ε + R σ + ε ) \lvert\ell\rvert R_{\sigma} \le\varepsilon^{-1}(R_{\sigma-\varepsilon}+R_{\sigma+\varepsilon}) ∣ ℓ ∣ R σ ≤ ε − 1 ( R σ − ε + R σ + ε ) , the same reasoning as in (i) shows that D z \mathcal{D}_z D z is a bounded operator on C ( R P 1 ) C(\mathbb{R}\mathbb{P}^{1}) C ( R P 1 ) . We conclude that z ↦ Q z z \mapsto Q_z z ↦ Q z is complex differentiable in operator norm at every point of Σ \Sigma Σ , with derivative D z \mathcal{D}_z D z .
Kernel measures and total variation
For s ∈ ( 0 , 2 ) s \in(0,2) s ∈ ( 0 , 2 ) and x ∈ R P 1 x \in\mathbb{R}\mathbb{P}^{1} x ∈ R P 1 , we let κ s ( x , ⋅ ) \kappa_s(x,\cdot) κ s ( x , ⋅ ) be the finite Borel measure on R P 1 \mathbb{R}\mathbb{P}^{1} R P 1 defined by
κ s ( x , A ) = p E [ ∥ M ( B , β ) v x ∥ − s 1 A ( f B , β ( x ) ) ] , \kappa_s(x,A)=p\mathbb{E}\left[\lVert M(B,\beta)v_x\rVert^{-s}\mathbf{1}_A\left(f_{B,\beta}(x)\right)\right], κ s ( x , A ) = p E [ ∥ M ( B , β ) v x ∥ − s 1 A ( f B , β ( x ) ) ] ,
so that Q s F ( x ) = ∫ F ( y ) κ s ( x , d y ) Q_sF(x)=\int F(y)\,\kappa_s(x,dy) Q s F ( x ) = ∫ F ( y ) κ s ( x , d y ) for bounded Borel F F F . We call κ s ( x , ⋅ ) \kappa_s(x,\cdot) κ s ( x , ⋅ ) the kernel measure of Q s Q_s Q s at x x x . By Tonelli’s theorem, x ↦ κ s ( x , A ) x\mapsto\kappa_s(x,A) x ↦ κ s ( x , A ) is Borel for each Borel set A A A ; that is, κ s \kappa_s κ s is a finite kernel. We define its iterates by κ s ( 1 ) = κ s \kappa_s^{(1)}=\kappa_s κ s ( 1 ) = κ s and κ s ( n + 1 ) ( x , A ) = ∫ κ s ( x , d y ) κ s ( n ) ( y , A ) \kappa_s^{(n+1)}(x,A)=\int\kappa_s(x,dy)\,\kappa_s^{(n)}(y,A) κ s ( n + 1 ) ( x , A ) = ∫ κ s ( x , d y ) κ s ( n ) ( y , A ) . By induction and Tonelli’s theorem, we have Q s n F ( x ) = ∫ F d κ s ( n ) ( x , ⋅ ) Q_s^nF(x)=\int F\,d\kappa_s^{(n)}(x,\cdot) Q s n F ( x ) = ∫ F d κ s ( n ) ( x , ⋅ ) for bounded Borel F F F , the Chapman–Kolmogorov relations κ s ( m + n ) ( x , ⋅ ) = ∫ κ s ( m ) ( x , d y ) κ s ( n ) ( y , ⋅ ) \kappa_s^{(m+n)}(x,\cdot)=\int\kappa_s^{(m)}(x,dy)\,\kappa_s^{(n)}(y,\cdot) κ s ( m + n ) ( x , ⋅ ) = ∫ κ s ( m ) ( x , d y ) κ s ( n ) ( y , ⋅ ) hold, and
κ s ( n ) ( x , R P 1 ) = Q s n 1 ( x ) ≤ ∥ Q s ∥ n . (82) \kappa_s^{(n)}(x,\mathbb{R}\mathbb{P}^{1})=Q_s^n1(x)\leq\lVert Q_s\rVert^n. \tag*{(82)} κ s ( n ) ( x , R P 1 ) = Q s n 1 ( x ) ≤ ∥ Q s ∥ n . (82)
Directly from the definition, we also have
κ s ( n + 1 ) ( x , ⋅ ) = p E [ ∥ M ( B , β ) v x ∥ − s κ s ( n ) ( f B , β ( x ) , ⋅ ) ] . (83) \kappa_s^{(n+1)}(x,\cdot)=p\mathbb{E}\left[\lVert M(B,\beta)v_x\rVert^{-s}\kappa_s^{(n)}\left(f_{B,\beta}(x),\cdot\right)\right]. \tag*{(83)} κ s ( n + 1 ) ( x , ⋅ ) = p E [ ∥ M ( B , β ) v x ∥ − s κ s ( n ) ( f B , β ( x ) , ⋅ ) ] . (83)
For a finite signed Borel measure μ \mu μ on R P 1 \mathbb{R}\mathbb{P}^{1} R P 1 , we write ∥ μ ∥ T V = ∣ μ ∣ ( R P 1 ) \lVert\mu\rVert_{\mathrm{TV}}=\lvert\mu\rvert(\mathbb{R}\mathbb{P}^{1}) ∥ μ ∥ TV = ∣ μ ∣ ( R P 1 ) for its total variation norm. We say that the n n n -step kernel measure is continuous in total variation at x 0 ∈ R P 1 x_0\in\mathbb{R}\mathbb{P}^{1} x 0 ∈ R P 1 if ∥ κ s ( n ) ( x , ⋅ ) − κ s ( n ) ( x 0 , ⋅ ) ∥ T V → 0 \lVert\kappa_s^{(n)}(x,\cdot)-\kappa_s^{(n)}(x_0,\cdot)\rVert_{\mathrm{TV}}\to0 ∥ κ s ( n ) ( x , ⋅ ) − κ s ( n ) ( x 0 , ⋅ ) ∥ TV → 0 as x → x 0 x\to x_0 x → x 0 .
The kernel measures at the two special points are explicit. Since f b , β ( ∞ ) = 0 f_{b,\beta}(\infty)=0 f b , β ( ∞ ) = 0 and f b , β ( 0 ) = 1 / b f_{b,\beta}(0)=1/b f b , β ( 0 ) = 1/ b , (6.3) yields
κ s ( ∞ , ⋅ ) = p E [ β − s / 2 ] δ 0 , κ s ( 0 , A ) = p E [ β s / 2 ] E [ ( 1 + B 2 ) − s / 2 1 A ( 1 / B ) ] . (84) \kappa_s(\infty,\cdot)=p\mathbb{E}\left[\beta^{-s/2}\right]\delta_0,\qquad
\kappa_s(0,A)=p\mathbb{E}\left[\beta^{s/2}\right]\mathbb{E}\left[(1+B^2)^{-s/2}\mathbf{1}_A(1/B)\right]. \tag*{(84)} κ s ( ∞ , ⋅ ) = p E [ β − s /2 ] δ 0 , κ s ( 0 , A ) = p E [ β s /2 ] E [ ( 1 + B 2 ) − s /2 1 A ( 1/ B ) ] . (84)
At a finite nonzero input, the kernel measure has a density. To describe it, let ι : R → R P 1 \iota:\mathbb{R}\to\mathbb{R}\mathbb{P}^{1} ι : R → R P 1 be the map ι ( w ) = 1 / w \iota(w)=1/w ι ( w ) = 1/ w , with ι ( 0 ) = ∞ \iota(0)=\infty ι ( 0 ) = ∞ ; it is a Borel bijection onto R P 1 ∖ { 0 } \mathbb{R}\mathbb{P}^{1}\setminus\{0\} R P 1 ∖ { 0 } with Borel inverse. Set φ ( t ) = t s / 2 e − K t 1 { t > 0 } \varphi(t)=t^{s/2}e^{-Kt}\mathbf{1}_{\{t>0\}} φ ( t ) = t s /2 e − K t 1 { t > 0 } , which is a bounded continuous function on R \mathbb{R} R because s > 0 s>0 s > 0 . For x ∈ R ∖ { 0 } x\in\mathbb{R}\setminus\{0\} x ∈ R ∖ { 0 } and w ∈ R w\in\mathbb{R} w ∈ R , define
k x ( w ) = p K ∣ x ∣ ( 1 + x 2 ) s / 2 ( 1 + w 2 ) − s / 2 ∫ φ ( b − w x ) ν ( d b ) . (85) k_x(w)=\frac{pK}{\lvert x\rvert}(1+x^2)^{s/2}(1+w^2)^{-s/2}\int\varphi\left(\frac{b-w}{x}\right)\nu(db). \tag*{(85)} k x ( w ) = ∣ x ∣ p K ( 1 + x 2 ) s /2 ( 1 + w 2 ) − s /2 ∫ φ ( x b − w ) ν ( d b ) . (85)
Lemma 6.7. Suppose that ν \nu ν satisfies (6.1), and let s ∈ ( 0 , 2 ) s\in(0,2) s ∈ ( 0 , 2 ) .
(i) For x ∈ R ∖ { 0 } x\in\mathbb{R}\setminus\{0\} x ∈ R ∖ { 0 } , the kernel measure κ s ( x , ⋅ ) \kappa_s(x,\cdot) κ s ( x , ⋅ ) is the image of the measure k x ( w ) d w k_x(w)\,dw k x ( w ) d w under ι \iota ι . Further, k x ( w ) > 0 k_x(w)>0 k x ( w ) > 0 for every w ∈ R w\in\mathbb{R} w ∈ R .
(ii) The one-step kernel measure is continuous in total variation at every x ∈ R ∖ { 0 } x \in\mathbb{R} \setminus\{0\} x ∈ R ∖ { 0 } .
(iii) The two-step kernel measure is continuous in total variation at every x ∈ R x \in\mathbb{R} x ∈ R .
(iv) The three-step kernel measure is continuous in total variation at every x ∈ R P 1 x \in\mathbb{R}\mathbb{P}^{1} x ∈ R P 1 .
Further, Q s 3 Q_s^3 Q s 3 is a compact operator on C ( R P 1 ) C(\mathbb{R}\mathbb{P}^{1}) C ( R P 1 ) .
Proof. We begin with (i). Fix x ∈ R ∖ { 0 } x \in\mathbb{R} \setminus\{0\} x ∈ R ∖ { 0 } and a Borel set A ⊂ R P 1 A \subset\mathbb{R}\mathbb{P}^{1} A ⊂ R P 1 . For fixed b b b , we substitute w = b − β x w=b-\beta x w = b − β x in the integral over β \beta β , so that β = ( b − w ) / x \beta=(b-w)/x β = ( b − w ) / x , d β = d w / ∣ x ∣ \mathrm{d}\beta=\mathrm{d}w/|x| d β = d w /∣ x ∣ , and the constraint β > 0 \beta>0 β > 0 becomes ( b − w ) / x > 0 (b-w)/x>0 ( b − w ) / x > 0 . By (6.3), the integrand ∥ M ( b , β ) v x ∥ − s \lVert M(b,\beta)v_x\rVert^{-s} ∥ M ( b , β ) v x ∥ − s equals β s / 2 ( 1 + x 2 ) s / 2 ( 1 + w 2 ) − s / 2 \beta^{s/2}(1+x^2)^{s/2}(1+w^2)^{-s/2} β s /2 ( 1 + x 2 ) s /2 ( 1 + w 2 ) − s /2 , and f b , β ( x ) = ι ( w ) f_{b,\beta}(x)=\iota(w) f b , β ( x ) = ι ( w ) . Using Tonelli’s theorem to integrate over b b b last, we obtain
κ s ( x , A ) = ∫ R k x ( w ) 1 A ( ι ( w ) ) d w . \kappa_s(x,A)=\int_{\mathbb{R}}k_x(w)\mathbf{1}_A(\iota(w))\,\mathrm{d}w. κ s ( x , A ) = ∫ R k x ( w ) 1 A ( ι ( w )) d w .
which is the first claim. For the second claim, note that ϕ ( t ) > 0 \phi(t)>0 ϕ ( t ) > 0 for t > 0 t>0 t > 0 . If x > 0 x>0 x > 0 , the integral in (6.13) is positive because ν ( ( w , ∞ ) ) > 0 \nu((w,\infty))>0 ν (( w , ∞ )) > 0 , and if x < 0 x<0 x < 0 , it is positive because ν ( ( − ∞ , w ) ) > 0 \nu((-\infty,w))>0 ν (( − ∞ , w )) > 0 ; both follow from supp ν = R \operatorname{supp}\nu=\mathbb{R} supp ν = R .
For (ii), let x n → x x_n \to x x n → x in R ∖ { 0 } \mathbb{R} \setminus\{0\} R ∖ { 0 } . Since ϕ \phi ϕ is bounded and continuous, dominated convergence gives ∫ ϕ ( ( b − w ) / x n ) ν ( d b ) → ∫ ϕ ( ( b − w ) / x ) ν ( d b ) \int\phi((b-w)/x_n)\nu(\mathrm{d}b)\to\int\phi((b-w)/x)\nu(\mathrm{d}b) ∫ ϕ (( b − w ) / x n ) ν ( d b ) → ∫ ϕ (( b − w ) / x ) ν ( d b ) for every w w w . The prefactors in (6.13) are continuous in x x x , so k x n ( w ) → k x ( w ) k_{x_n}(w)\to k_x(w) k x n ( w ) → k x ( w ) for every w w w . Further, ∫ k x n = Q s 1 ( x n ) → Q s 1 ( x ) = ∫ k x \int k_{x_n}=Q_s1(x_n)\to Q_s1(x)=\int k_x ∫ k x n = Q s 1 ( x n ) → Q s 1 ( x ) = ∫ k x by Lemma 6.6(i). We now use Scheffé’s lemma: since ( k x − k x n ) + ≤ k x (k_x-k_{x_n})^+\leq k_x ( k x − k x n ) + ≤ k x and ( k x − k x n ) + → 0 (k_x-k_{x_n})^+\to0 ( k x − k x n ) + → 0 pointwise, dominated convergence gives ∫ ( k x − k x n ) + → 0 \int(k_x-k_{x_n})^+\to0 ∫ ( k x − k x n ) + → 0 , and then
∫ ∣ k x − k x n ∣ = 2 ∫ ( k x − k x n ) + − ∫ ( k x − k x n ) → 0. \int|k_x-k_{x_n}|=2\int(k_x-k_{x_n})^+-\int(k_x-k_{x_n})\to0. ∫ ∣ k x − k x n ∣ = 2 ∫ ( k x − k x n ) + − ∫ ( k x − k x n ) → 0.
Since ι \iota ι is a Borel bijection with Borel inverse, the total variation norm of the difference of the image measures equals ∫ ∣ k x − k x n ∣ \int|k_x-k_{x_n}| ∫ ∣ k x − k x n ∣ . This proves (ii).
The remaining parts rest on the following claim.
Claim. Let n ≥ 1 n\geq1 n ≥ 1 and x 0 ∈ R P 1 x_0\in\mathbb{R}\mathbb{P}^{1} x 0 ∈ R P 1 , and suppose that for ν ⊗ Exp ( K ) \nu\otimes\operatorname{Exp}(K) ν ⊗ Exp ( K ) -almost every ( b , β ) (b,\beta) ( b , β ) , the n n n -step kernel measure is continuous in total variation at f b , β ( x 0 ) f_{b,\beta}(x_0) f b , β ( x 0 ) . Then the ( n + 1 ) (n+1) ( n + 1 ) -step kernel measure is continuous in total variation at x 0 x_0 x 0 .
To prove the claim, let x k → x 0 x_k\to x_0 x k → x 0 , and write R ( k ) = ∥ M ( B , β ) v x k ∥ − s R^{(k)}=\lVert M(B,\beta)v_{x_k}\rVert^{-s} R ( k ) = ∥ M ( B , β ) v x k ∥ − s , R = ∥ M ( B , β ) v x 0 ∥ − s R=\lVert M(B,\beta)v_{x_0}\rVert^{-s} R = ∥ M ( B , β ) v x 0 ∥ − s , y k = f B , β ( x k ) y_k=f_{B,\beta}(x_k) y k = f B , β ( x k ) , and y = f B , β ( x 0 ) y=f_{B,\beta}(x_0) y = f B , β ( x 0 ) . By (6.11) and (6.10), we have
∥ κ s ( n + 1 ) ( x k , ⋅ ) − κ s ( n + 1 ) ( x 0 , ⋅ ) ∥ T V ≤ p ∥ Q s ∥ E ∣ R ( k ) − R ∣ + p E [ R ∥ κ s ( n ) ( y k , ⋅ ) − κ s ( n ) ( y , ⋅ ) ∥ T V ] . \left\lVert\kappa_s^{(n+1)}(x_k,\mathord\cdot)-\kappa_s^{(n+1)}(x_0,\mathord\cdot)\right\rVert_{\mathrm{TV}}
\leq p\lVert Q_s\rVert\,\mathbb{E}|R^{(k)}-R|+p\,\mathbb{E}\left[R\left\lVert\kappa_s^{(n)}(y_k,\mathord\cdot)-\kappa_s^{(n)}(y,\mathord\cdot)\right\rVert_{\mathrm{TV}}\right]. κ s ( n + 1 ) ( x k , ⋅ ) − κ s ( n + 1 ) ( x 0 , ⋅ ) TV ≤ p ∥ Q s ∥ E ∣ R ( k ) − R ∣ + p E [ R κ s ( n ) ( y k , ⋅ ) − κ s ( n ) ( y , ⋅ ) TV ] .
(The total variation distance is a Borel function of ( y k , y ) (y_k,y) ( y k , y ) , because it equals the supremum of ∣ Q s n F ( y k ) − Q s n F ( y ) ∣ |Q_s^nF(y_k)-Q_s^nF(y)| ∣ Q s n F ( y k ) − Q s n F ( y ) ∣ over a countable dense subset of the unit ball of C ( R P 1 ; R ) C(\mathbb{R}\mathbb{P}^{1};\mathbb{R}) C ( R P 1 ; R ) , and each Q s n F Q_s^nF Q s n F is continuous.) By continuity of the projective action, R ( k ) → R R^{(k)}\to R R ( k ) → R and y k → y y_k\to y y k → y almost surely. The family ( R ( k ) ) (R^{(k)}) ( R ( k ) ) is uniformly integrable by the argument in the proof of Lemma 6.6(i), and the first term tends to zero by Vitali’s convergence theorem. In second term, the integrand tends to zero almost surely by hypothesis, and it is bounded by 2 ∥ Q s ∥ n R 2\lVert Q_s\rVert^n R 2 ∥ Q s ∥ n R , which is integrable. By dominated convergence, the second term tends to zero. This proves the claim.
We now prove (iii). Let x 0 ∈ R ∖ { 0 } x_0 \in\mathbb{R} \setminus\{0\} x 0 ∈ R ∖ { 0 } . Almost surely, f B , β ( x 0 ) f_{B,\beta}(x_0) f B , β ( x 0 ) is finite (as noted in Definition 6.1(vi)) and nonzero (since f b , β ( x 0 ) = 0 f_{b,\beta}(x_0) = 0 f b , β ( x 0 ) = 0 only if x 0 = ∞ x_0 = \infty x 0 = ∞ ). By (ii) and the claim with n = 1 n = 1 n = 1 , the two-step kernel measure is continuous at x 0 x_0 x 0 . For x 0 = 0 x_0 = 0 x 0 = 0 , we have f B , β ( 0 ) = 1 / B f_{B,\beta}(0) = 1/B f B , β ( 0 ) = 1/ B , which lies in R ∖ { 0 } \mathbb{R} \setminus\{0\} R ∖ { 0 } almost surely because P ( B = 0 ) = 0 \mathbb{P}(B = 0) = 0 P ( B = 0 ) = 0 . Again (ii) and the claim imply continuity at 0 0 0 .
For (iv), let x 0 ∈ R x_0 \in\mathbb{R} x 0 ∈ R . By (iii), the two-step kernel measure is continuous at every point of R \mathbb{R} R , and f B , β ( x 0 ) ∈ R f_{B,\beta}(x_0) \in\mathbb{R} f B , β ( x 0 ) ∈ R almost surely. The claim with n = 2 n = 2 n = 2 implies continuity of the three-step kernel measure at x 0 x_0 x 0 . For x 0 = ∞ x_0 = \infty x 0 = ∞ , we have f b , β ( ∞ ) = 0 f_{b,\beta}(\infty) = 0 f b , β ( ∞ ) = 0 for all ( b , β ) (b,\beta) ( b , β ) , and the two-step kernel measure is continuous at 0 0 0 by (iii). The claim with n = 2 n = 2 n = 2 implies continuity at ∞ \infty ∞ .
Finally, we deduce compactness. By (iv) and the compactness of R P 1 \mathbb{RP}^1 RP 1 , the map x ↦ κ s ( 3 ) ( x , ⋅ ) x \mapsto\kappa_s^{(3)}(x,\mathord\cdot) x ↦ κ s ( 3 ) ( x , ⋅ ) is uniformly continuous in total variation. For F ∈ C ( R P 1 ) F \in C(\mathbb{RP}^1) F ∈ C ( RP 1 ) with ∥ F ∥ ∞ ≤ 1 \lVert F\rVert_\infty\leq1 ∥ F ∥ ∞ ≤ 1 and x , x ′ ∈ R P 1 x,x' \in\mathbb{RP}^1 x , x ′ ∈ RP 1 , we have
∣ Q s 3 F ( x ) − Q s 3 F ( x ′ ) ∣ ≤ ∥ κ s ( 3 ) ( x , ⋅ ) − κ s ( 3 ) ( x ′ , ⋅ ) ∥ T V . \left|Q_s^3F(x)-Q_s^3F(x')\right| \leq\left\lVert\kappa_s^{(3)}(x,\mathord\cdot)-\kappa_s^{(3)}(x',\mathord\cdot)\right\rVert_{\mathrm{TV}}. Q s 3 F ( x ) − Q s 3 F ( x ′ ) ≤ κ s ( 3 ) ( x , ⋅ ) − κ s ( 3 ) ( x ′ , ⋅ ) TV .
The image of the unit ball under Q s 3 Q_s^3 Q s 3 is therefore equicontinuous, and it is bounded by ∥ Q s ∥ 3 \lVert Q_s\rVert^3 ∥ Q s ∥ 3 . By the Arzelà–Ascoli theorem, it is relatively compact in C ( R P 1 ) C(\mathbb{RP}^1) C ( RP 1 ) .
Remark 6.8. The argument above does not apply to Q s Q_s Q s or Q s 2 Q_s^2 Q s 2 . By (6.12), the kernel measure at ∞ \infty ∞ is a multiple of the point mass δ 0 \delta_0 δ 0 , while by Lemma 6.7(i) the kernel measures at finite nonzero inputs are absolutely continuous with respect to m m m ; since such measures form a closed set in total variation, the one-step kernel measure is not continuous in total variation at ∞ \infty ∞ . For the second power, (6.12) gives κ s ( 2 ) ( ∞ , ⋅ ) = p E [ β − s / 2 ] κ s ( 0 , ⋅ ) \kappa_s^{(2)}(\infty,\mathord\cdot)=p\mathbb{E}[\beta^{-s/2}]\kappa_s(0,\mathord\cdot) κ s ( 2 ) ( ∞ , ⋅ ) = p E [ β − s /2 ] κ s ( 0 , ⋅ ) , and κ s ( 0 , ⋅ ) \kappa_s(0,\mathord\cdot) κ s ( 0 , ⋅ ) has an atom at 1 / b 0 1/b_0 1/ b 0 whenever ν ( { b 0 } ) > 0 \nu(\{b_0\})>0 ν ({ b 0 }) > 0 . A real solution always has such an atom: by Proposition 5.5, the law of D D D has an atom of mass e − p e^{-p} e − p at E E E , and the law of B = g D B=gD B = g D has an atom at g E gE g E . On the other hand, κ s ( 2 ) ( x , ⋅ ) \kappa_s^{(2)}(x,\mathord\cdot) κ s ( 2 ) ( x , ⋅ ) is absolutely continuous for x ∈ R ∖ { 0 } x \in\mathbb{R} \setminus\{0\} x ∈ R ∖ { 0 } , since it is a mixture of the absolutely continuous kernel measures κ s ( y , ⋅ ) \kappa_s(y,\mathord\cdot) κ s ( y , ⋅ ) , y ∈ R ∖ { 0 } y \in\mathbb{R} \setminus\{0\} y ∈ R ∖ { 0 } . It follows that the two-step kernel measure is not continuous in total variation at ∞ \infty ∞ in this case. The third step from ∞ \infty ∞ passes through 0 0 0 and then through the almost surely finite nonzero point 1 / B 1/B 1/ B , where the kernel measures are absolutely continuous; this is why the third power is the first one to which the argument applies.
Lemma 6.9. Suppose that ν \nu ν satisfies (6.1), and let s ∈ ( 0 , 2 ) s \in(0,2) s ∈ ( 0 , 2 ) . For every n ≥ 2 n \geq2 n ≥ 2 and x ∈ R P 1 x \in\mathbb{RP}^1 x ∈ RP 1 , the measure κ s ( n ) ( x , ⋅ ) \kappa_s^{(n)}(x,\mathord\cdot) κ s ( n ) ( x , ⋅ ) has full support, that is, κ s ( n ) ( x , O ) > 0 \kappa_s^{(n)}(x,O)>0 κ s ( n ) ( x , O ) > 0 for every nonempty open set O ⊂ R P 1 O \subset\mathbb{RP}^1 O ⊂ RP 1 . In particular, for every n ≥ 2 n \geq2 n ≥ 2 and every real-valued F ∈ C ( R P 1 ) F \in C(\mathbb{RP}^1) F ∈ C ( RP 1 ) with F ≥ 0 F \geq0 F ≥ 0 and F ≢ 0 F \not\equiv0 F ≡ 0 , we have min x ∈ R P 1 Q s 2 F ( x ) > 0 \min_{x\in\mathbb{RP}^1}Q_s^2F(x)>0 min x ∈ RP 1 Q s 2 F ( x ) > 0 .
Proof. Let O ⊂ R P 1 O \subset\mathbb{R}\mathbb{P}^{1} O ⊂ R P 1 be open and nonempty. We first show that κ s ( y , O ) > 0 \kappa_s(y,O)>0 κ s ( y , O ) > 0 for every y ∈ R y\in\mathbb{R} y ∈ R . For y ∈ R ∖ { 0 } y\in\mathbb{R}\setminus\{0\} y ∈ R ∖ { 0 } , this follows from Lemma 6.7(i), since ι − 1 ( O ∖ { 0 } ) \iota^{-1}(O\setminus\{0\}) ι − 1 ( O ∖ { 0 }) is a nonempty open subset of R \mathbb{R} R and k y > 0 k_y>0 k y > 0 everywhere. For y = 0 y=0 y = 0 , the set { b ∈ R ∖ { 0 } : 1 / b ∈ O } \{b\in\mathbb{R}\setminus\{0\}:1/b\in O\} { b ∈ R ∖ { 0 } : 1/ b ∈ O } is open and nonempty (because O ∖ { 0 , ∞ } O\setminus\{0,\infty\} O ∖ { 0 , ∞ } is open and nonempty), and it has positive ν \nu ν -measure since supp ν = R \operatorname{supp}\nu=\mathbb{R} supp ν = R . By (6.12), we conclude that κ s ( 0 , O ) > 0 \kappa_s(0,O)>0 κ s ( 0 , O ) > 0 .
Next, let x ∈ R P 1 x\in\mathbb{R}\mathbb{P}^{1} x ∈ R P 1 . If x ∈ R x\in\mathbb{R} x ∈ R , then κ s ( x , ⋅ ) \kappa_s(x,\cdot) κ s ( x , ⋅ ) has positive total mass and gives no mass to { ∞ } \{\infty\} { ∞ } (by Lemma 6.7(i) for x ≠ 0 x\ne0 x = 0 , and by P ( B = 0 ) = 0 \mathbb{P}(B=0)=0 P ( B = 0 ) = 0 for x = 0 x=0 x = 0 ), which implies that κ s ( 2 ) ( x , O ) = ∫ κ s ( x , d y ) κ s ( y , O ) > 0 \kappa_s^{(2)}(x,O)=\int\kappa_s(x,\mathrm{d}y)\kappa_s(y,O)>0 κ s ( 2 ) ( x , O ) = ∫ κ s ( x , d y ) κ s ( y , O ) > 0 . If x = ∞ x=\infty x = ∞ , then κ s ( 2 ) ( ∞ , O ) = p E [ β − s / 2 ] κ s ( 0 , O ) > 0 \kappa_s^{(2)}(\infty,O)=p\mathbb{E}[\beta^{-s/2}]\kappa_s(0,O)>0 κ s ( 2 ) ( ∞ , O ) = p E [ β − s /2 ] κ s ( 0 , O ) > 0 by (6.12). This proves the claim for n = 2 n=2 n = 2 . For n ≥ 3 n\ge3 n ≥ 3 , we have κ s ( n ) ( x , O ) = ∫ κ s ( n − 2 ) ( x , d y ) κ s ( 2 ) ( y , O ) > 0 \kappa_s^{(n)}(x,O)=\int\kappa_s^{(n-2)}(x,\mathrm{d}y)\kappa_s^{(2)}(y,O)>0 κ s ( n ) ( x , O ) = ∫ κ s ( n − 2 ) ( x , d y ) κ s ( 2 ) ( y , O ) > 0 , because κ s ( n − 2 ) ( x , R P 1 ) = Q s n − 2 1 ( x ) > 0 \kappa_s^{(n-2)}(x,\mathbb{R}\mathbb{P}^{1})=Q_s^{n-2}1(x)>0 κ s ( n − 2 ) ( x , R P 1 ) = Q s n − 2 1 ( x ) > 0 by Lemma 6.6(ii) and induction.
Finally, if F ≥ 0 F\ge0 F ≥ 0 is continuous and not identically zero, then { F > 0 } \{F>0\} { F > 0 } is open and nonempty, and we have Q s n F ( x ) = ∫ F d κ s ( n ) ( x , ⋅ ) > 0 Q_s^nF(x)=\int F\,\mathrm{d}\kappa_s^{(n)}(x,\cdot)>0 Q s n F ( x ) = ∫ F d κ s ( n ) ( x , ⋅ ) > 0 for every x x x . Since Q s n F Q_s^nF Q s n F is continuous and R P 1 \mathbb{R}\mathbb{P}^{1} R P 1 is compact, its minimum is positive.
The Perron–Frobenius eigenvalue
We use the following form of the stability of an isolated simple eigenvalue under small perturbations, both for analyticity in s s s and for continuity in the law of B B B . For a bounded operator T T T on a complex Banach space, we write σ ( T ) \sigma(T) σ ( T ) for its spectrum.
Lemma 6.10. Let X X X be a complex Banach space and let T 0 T_0 T 0 be a bounded operator on X X X . Suppose that μ 0 ∈ σ ( T 0 ) \mu_0\in\sigma(T_0) μ 0 ∈ σ ( T 0 ) is an isolated point of σ ( T 0 ) \sigma(T_0) σ ( T 0 ) whose Riesz projection has rank one, and that there exists r 0 < ∣ μ 0 ∣ r_0<|\mu_0| r 0 < ∣ μ 0 ∣ such that ∣ w ∣ ≤ r 0 |w|\le r_0 ∣ w ∣ ≤ r 0 for every w ∈ σ ( T 0 ) ∖ { μ 0 } w\in\sigma(T_0)\setminus\{\mu_0\} w ∈ σ ( T 0 ) ∖ { μ 0 } . Fix δ ∈ ( 0 , ( ∣ μ 0 ∣ − r 0 ) / 3 ) \delta\in(0,(|\mu_0|-r_0)/3) δ ∈ ( 0 , ( ∣ μ 0 ∣ − r 0 ) /3 ) , let Γ \Gamma Γ be the circle { ∣ w − μ 0 ∣ = δ } \{|w-\mu_0|=\delta\} { ∣ w − μ 0 ∣ = δ } , and let ψ ∈ X \psi\in X ψ ∈ X and ℓ ∈ X ∗ \ell\in X^* ℓ ∈ X ∗ satisfy ℓ ( Π 0 ψ ) ≠ 0 \ell(\Pi_0\psi)\ne0 ℓ ( Π 0 ψ ) = 0 , where Π 0 \Pi_0 Π 0 is the Riesz projection of T 0 T_0 T 0 at μ 0 \mu_0 μ 0 . Then there exists ε > 0 \varepsilon>0 ε > 0 such that every bounded operator T T T with ∥ T − T 0 ∥ < ε \|T-T_0\|<\varepsilon ∥ T − T 0 ∥ < ε has the following properties.
(i) We have σ ( T ) ⊂ { ∣ w ∣ < r 0 + δ } ∪ { ∣ w − μ 0 ∣ < δ } \sigma(T)\subset\{|w|<r_0+\delta\}\cup\{|w-\mu_0|<\delta\} σ ( T ) ⊂ { ∣ w ∣ < r 0 + δ } ∪ { ∣ w − μ 0 ∣ < δ } , and σ ( T ) ∩ { ∣ w − μ 0 ∣ < δ } \sigma(T)\cap\{|w-\mu_0|<\delta\} σ ( T ) ∩ { ∣ w − μ 0 ∣ < δ } consists of a single point μ ( T ) \mu(T) μ ( T ) , which is an eigenvalue of T T T whose Riesz projection Π T = ( 2 π i ) − 1 ∮ Γ ( w − T ) − 1 d w \Pi_T=(2\pi i)^{-1}\oint_{\Gamma}(w-T)^{-1}\,\mathrm{d}w Π T = ( 2 π i ) − 1 ∮ Γ ( w − T ) − 1 d w has rank one. In particular, μ ( T ) \mu(T) μ ( T ) is the unique point of σ ( T ) \sigma(T) σ ( T ) of maximal modulus.
(ii) We have ℓ ( Π T ψ ) ≠ 0 \ell(\Pi_T\psi)\ne0 ℓ ( Π T ψ ) = 0 and μ ( T ) = ℓ ( T Π T ψ ) / ℓ ( Π T ψ ) \mu(T)=\ell(T\Pi_T\psi)/\ell(\Pi_T\psi) μ ( T ) = ℓ ( T Π T ψ ) / ℓ ( Π T ψ ) . The maps T ↦ Π T T\mapsto\Pi_T T ↦ Π T and T ↦ μ ( T ) T\mapsto\mu(T) T ↦ μ ( T ) are continuous on { ∥ T − T 0 ∥ < ε } \{\|T-T_0\|<\varepsilon\} { ∥ T − T 0 ∥ < ε } .
(iii) If U ⊂ C U\subset\mathbb{C} U ⊂ C is open and z ↦ T ( z ) z\mapsto T(z) z ↦ T ( z ) is a holomorphic map from U U U into the bounded operators on X X X (in operator norm) with ∥ T ( z ) − T 0 ∥ < ε \|T(z)-T_0\|<\varepsilon ∥ T ( z ) − T 0 ∥ < ε for all z ∈ U z\in U z ∈ U , then z ↦ μ ( T ( z ) ) z\mapsto\mu(T(z)) z ↦ μ ( T ( z )) is holomorphic on U U U .
Proof. Let Ω = { w ∈ C : ∣ w ∣ ≥ r 0 + δ , ∣ w − μ 0 ∣ ≥ δ } \Omega=\{w\in\mathbb{C}:|w|\ge r_0+\delta,\ |w-\mu_0|\ge\delta\} Ω = { w ∈ C : ∣ w ∣ ≥ r 0 + δ , ∣ w − μ 0 ∣ ≥ δ } . Then Ω \Omega Ω is closed and contained in the resolvent set of T 0 T_0 T 0 . Since the map w ↦ ∥ ( w − T 0 ) − 1 ∥ w\mapsto\|(w-T_0)^{-1}\| w ↦ ∥ ( w − T 0 ) − 1 ∥ is continuous on the resolvent set and bounded by ( ∣ w ∣ − ∥ T 0 ∥ ) − 1 (|w|-\|T_0\|)^{-1} ( ∣ w ∣ − ∥ T 0 ∥ ) − 1 for ∣ w ∣ > ∥ T 0 ∥ |w|>\|T_0\| ∣ w ∣ > ∥ T 0 ∥ , we have M = sup w ∈ Ω ∥ ( w − T 0 ) − 1 ∥ < ∞ M=\sup_{w\in\Omega}\|(w-T_0)^{-1}\|<\infty M = sup w ∈ Ω ∥ ( w − T 0 ) − 1 ∥ < ∞ . Suppose that ∥ T − T 0 ∥ < 1 / ( 2 M ) \lVert T-T_{0}\rVert< 1/(2M) ∥ T − T 0 ∥ < 1/ ( 2 M ) . For w ∈ Ω w \in\Omega w ∈ Ω , we write w − T = ( w − T 0 ) ( I − ( w − T 0 ) − 1 ( T − T 0 ) ) w-T=(w-T_{0})\left(I-(w-T_{0})^{-1}(T-T_{0})\right) w − T = ( w − T 0 ) ( I − ( w − T 0 ) − 1 ( T − T 0 ) ) ; the second factor is invertible by the Neumann series, with inverse of norm at most 2. Then Ω \Omega Ω lies in the resolvent set of T T T , ∥ ( w − T ) − 1 ∥ ≤ 2 M \lVert(w-T)^{-1}\rVert\le2M ∥( w − T ) − 1 ∥ ≤ 2 M , and the resolvent identity gives
∥ ( w − T ) − 1 − ( w − T 0 ) − 1 ∥ = ∥ ( w − T ) − 1 ( T − T 0 ) ( w − T 0 ) − 1 ∥ ≤ 2 M 2 ∥ T − T 0 ∥ \lVert(w-T)^{-1}-(w-T_{0})^{-1}\rVert=\lVert(w-T)^{-1}(T-T_{0})(w-T_{0})^{-1}\rVert\le2M^{2}\lVert T-T_{0}\rVert ∥( w − T ) − 1 − ( w − T 0 ) − 1 ∥ = ∥( w − T ) − 1 ( T − T 0 ) ( w − T 0 ) − 1 ∥ ≤ 2 M 2 ∥ T − T 0 ∥
for w ∈ Ω w \in\Omega w ∈ Ω . This proves the first inclusion in (i). Since ∣ μ 0 ∣ − δ > r 0 + 2 δ |\mu_{0}|-\delta>r_{0}+2\delta ∣ μ 0 ∣ − δ > r 0 + 2 δ , the circle Γ \Gamma Γ lies in Ω \Omega Ω , and integrating the last bound over Γ \Gamma Γ gives ∥ Π T − Π 0 ∥ ≤ 2 δ M 2 ∥ T − T 0 ∥ \lVert\Pi_{T}-\Pi_{0}\rVert\le2\delta M^{2}\lVert T-T_{0}\rVert ∥ Π T − Π 0 ∥ ≤ 2 δ M 2 ∥ T − T 0 ∥ . We now choose ε ≤ 1 / ( 2 M ) \varepsilon\le1/(2M) ε ≤ 1/ ( 2 M ) so small that 2 δ M 2 ε < 1 2\delta M^{2}\varepsilon<1 2 δ M 2 ε < 1 and ∣ ℓ ( Π T ψ ) − ℓ ( Π 0 ψ ) ∣ < ∣ ℓ ( Π 0 ψ ) ∣ |\ell(\Pi_{T}\psi)-\ell(\Pi_{0}\psi)|<|\ell(\Pi_{0}\psi)| ∣ ℓ ( Π T ψ ) − ℓ ( Π 0 ψ ) ∣ < ∣ ℓ ( Π 0 ψ ) ∣ whenever ∥ T − T 0 ∥ < ε \lVert T-T_{0}\rVert<\varepsilon ∥ T − T 0 ∥ < ε . Since ∥ Π T − Π 0 ∥ < 1 \lVert\Pi_{T}-\Pi_{0}\rVert<1 ∥ Π T − Π 0 ∥ < 1 , the projection Π T \Pi_{T} Π T has rank one. Indeed, if f f f lies in the range of Π T \Pi_{T} Π T and Π 0 f = 0 \Pi_{0}f=0 Π 0 f = 0 , then ∥ f ∥ = ∥ ( Π T − Π 0 ) f ∥ < ∥ f ∥ \lVert f\rVert=\lVert(\Pi_{T}-\Pi_{0})f\rVert<\lVert f\rVert ∥ f ∥ = ∥( Π T − Π 0 ) f ∥ < ∥ f ∥ unless f = 0 f=0 f = 0 , so Π 0 \Pi_{0} Π 0 is injective on the range of Π T \Pi_{T} Π T ; in the same way, Π T \Pi_{T} Π T is injective on the range of Π 0 \Pi_{0} Π 0 , which is one-dimensional (see also [82, Section I.4.6]). By the properties of Riesz projections [82, Section III.6.4], the range of Π T \Pi_{T} Π T is invariant under T T T , and the spectrum of the restriction of T T T to it equals σ ( T ) ∩ { ∣ w − μ 0 ∣ < δ } \sigma(T)\cap\{|w-\mu_{0}|<\delta\} σ ( T ) ∩ { ∣ w − μ 0 ∣ < δ } . Since the range is one-dimensional, T T T acts on it as multiplication by a scalar μ ( T ) \mu(T) μ ( T ) , and this scalar is the only point of σ ( T ) \sigma(T) σ ( T ) in the disc. Every other point of σ ( T ) \sigma(T) σ ( T ) has modulus less than r 0 + δ < ∣ μ 0 ∣ − δ < ∣ μ ( T ) ∣ r_{0}+\delta<|\mu_{0}|-\delta<|\mu(T)| r 0 + δ < ∣ μ 0 ∣ − δ < ∣ μ ( T ) ∣ . This proves (i).
For (ii), we have T Π T = μ ( T ) Π T T\Pi_{T}=\mu(T)\Pi_{T} T Π T = μ ( T ) Π T , since T T T acts on the range of Π T \Pi_{T} Π T as multiplication by μ ( T ) \mu(T) μ ( T ) . Applying ℓ \ell ℓ to T Π T ψ = μ ( T ) Π T ψ T\Pi_{T}\psi=\mu(T)\Pi_{T}\psi T Π T ψ = μ ( T ) Π T ψ gives the formula for μ ( T ) \mu(T) μ ( T ) , and its denominator is nonzero by the choice of ε \varepsilon ε . For T T T and T 1 T_{1} T 1 in the ball { ∥ T − T 0 ∥ < ε } \{\lVert T-T_{0}\rVert<\varepsilon\} {∥ T − T 0 ∥ < ε } and w ∈ Γ w\in\Gamma w ∈ Γ , both resolvents are bounded by 2 M 2M 2 M , and the resolvent identity gives ∥ ( w − T ) − 1 − ( w − T 1 ) − 1 ∥ ≤ 4 M 2 ∥ T − T 1 ∥ \lVert(w-T)^{-1}-(w-T_{1})^{-1}\rVert\le4M^{2}\lVert T-T_{1}\rVert ∥( w − T ) − 1 − ( w − T 1 ) − 1 ∥ ≤ 4 M 2 ∥ T − T 1 ∥ . Integrating over Γ \Gamma Γ , we obtain ∥ Π T − Π T 1 ∥ ≤ 4 δ M 2 ∥ T − T 1 ∥ \lVert\Pi_{T}-\Pi_{T_{1}}\rVert\le4\delta M^{2}\lVert T-T_{1}\rVert ∥ Π T − Π T 1 ∥ ≤ 4 δ M 2 ∥ T − T 1 ∥ . This proves continuity of Π T \Pi_{T} Π T , and continuity of μ ( T ) \mu(T) μ ( T ) follows from the formula.
For (iii), fix w ∈ Γ w\in\Gamma w ∈ Γ . The map z ↦ ( w − T ( z ) ) − 1 z\mapsto(w-T(z))^{-1} z ↦ ( w − T ( z ) ) − 1 is holomorphic on U U U , being the composition of a holomorphic map with inversion, which is holomorphic on the invertible operators; its derivative ( w − T ( z ) ) − 1 T ′ ( z ) ( w − T ( z ) ) − 1 (w-T(z))^{-1}T'(z)(w-T(z))^{-1} ( w − T ( z ) ) − 1 T ′ ( z ) ( w − T ( z ) ) − 1 is jointly continuous in ( w , z ) ∈ Γ × U (w,z)\in\Gamma\times U ( w , z ) ∈ Γ × U . It follows that z ↦ Π T ( z ) z\mapsto\Pi_{T(z)} z ↦ Π T ( z ) is holomorphic in operator norm (see also [82, Section VII.1.3]), and z ↦ μ ( T ( z ) ) = ℓ ( T ( z ) Π T ( z ) ψ ) / ℓ ( Π T ( z ) ψ ) z\mapsto\mu(T(z))=\ell(T(z)\Pi_{T(z)}\psi)/\ell(\Pi_{T(z)}\psi) z ↦ μ ( T ( z )) = ℓ ( T ( z ) Π T ( z ) ψ ) / ℓ ( Π T ( z ) ψ ) is a quotient of holomorphic functions with nonvanishing denominator.
We now state and prove the main result of this section.
Theorem 6.11. Suppose that ν \nu ν satisfies (6.1), and let s ∈ ( 0 , 2 ) s\in(0,2) s ∈ ( 0 , 2 ) . Let λ ( s ) \lambda(s) λ ( s ) be the spectral radius of Q s Q_{s} Q s on C ( R P 1 ) C(\mathbb{RP}^{1}) C ( RP 1 ) .
(i) The operator Q s Q_{s} Q s is bounded and positive on C ( R P 1 ) C(\mathbb{RP}^{1}) C ( RP 1 ) . For every n ≥ 2 n\ge2 n ≥ 2 , the operator Q s n Q_{s}^{n} Q s n is strongly positive: min x ∈ R P 1 Q s n F ( x ) > 0 \min_{x\in\mathbb{RP}^{1}}Q_{s}^{n}F(x)>0 min x ∈ RP 1 Q s n F ( x ) > 0 for every real-valued F ∈ C ( R P 1 ) F\in C(\mathbb{RP}^{1}) F ∈ C ( RP 1 ) with F ≥ 0 F\ge0 F ≥ 0 and F ≠ 0 F\ne0 F = 0 . Further, Q s 3 Q_{s}^{3} Q s 3 is compact.
(ii) We have λ ( s ) > 0 \lambda(s)>0 λ ( s ) > 0 , and λ ( s ) \lambda(s) λ ( s ) is an eigenvalue of Q s Q_s Q s whose Riesz projection has rank one. There exists a real-valued ψ s ∈ C ( R P 1 ) \psi_s\in C(\mathbb{RP}^1) ψ s ∈ C ( RP 1 ) with min x ψ s ( x ) > 0 \min_x\psi_s(x)>0 min x ψ s ( x ) > 0 and Q s ψ s = λ ( s ) ψ s Q_s\psi_s=\lambda(s)\psi_s Q s ψ s = λ ( s ) ψ s , and there exists r ( s ) < λ ( s ) r(s)<\lambda(s) r ( s ) < λ ( s ) such that ∣ z ∣ ≤ r ( s ) |z|\le r(s) ∣ z ∣ ≤ r ( s ) for every z ∈ σ ( Q s ) ∖ { λ ( s ) } z\in\sigma(Q_s)\setminus\{\lambda(s)\} z ∈ σ ( Q s ) ∖ { λ ( s )} .
(iii) There exists a finite positive Borel measure ℓ s \ell_s ℓ s on R P 1 \mathbb{RP}^1 RP 1 with full support such that ∫ ψ s d ℓ s = 1 \int\psi_s\,\mathrm{d}\ell_s=1 ∫ ψ s d ℓ s = 1 , ∫ Q s F d ℓ s = λ ( s ) ∫ F d ℓ s \int Q_sF\,\mathrm{d}\ell_s=\lambda(s)\int F\,\mathrm{d}\ell_s ∫ Q s F d ℓ s = λ ( s ) ∫ F d ℓ s for F ∈ C ( R P 1 ) F\in C(\mathbb{RP}^1) F ∈ C ( RP 1 ) , and
lim L → ∞ sup ∥ F ∥ ∞ ≤ 1 ∥ λ ( s ) − L Q s L F − ( ∫ F d ℓ s ) ψ s ∥ ∞ = 0. \lim_{L\to\infty}\sup_{\lVert F\rVert_\infty\le1}\left\lVert\lambda(s)^{-L}Q_s^L F-\left(\int F\,\mathrm{d}\ell_s\right)\psi_s\right\rVert_\infty=0. L → ∞ lim ∥ F ∥ ∞ ≤ 1 sup λ ( s ) − L Q s L F − ( ∫ F d ℓ s ) ψ s ∞ = 0.
(iv) The function s ↦ λ ( s ) s\mapsto\lambda(s) s ↦ λ ( s ) is real-analytic on ( 0 , 2 ) (0,2) ( 0 , 2 ) .
(v) Let M 1 , M 2 , … M_1,M_2,\ldots M 1 , M 2 , … be independent copies of M ( B , β ) M(B,\beta) M ( B , β ) . For every L ≥ 1 L\ge1 L ≥ 1 , every bounded Borel function F F F , and every x ∈ R P 1 x\in\mathbb{RP}^1 x ∈ RP 1 , we have
Q s L F ( x ) = p L E [ ∥ M L ⋯ M 1 v x ∥ − s F ( ( M L ⋯ M 1 ) ⋅ x ) ] . (86) Q_s^L F(x)=p^L\mathbb{E}\left[\lVert M_L\cdots M_1v_x\rVert^{-s}F((M_L\cdots M_1)\cdot x)\right]. \tag*{(86)} Q s L F ( x ) = p L E [ ∥ M L ⋯ M 1 v x ∥ − s F (( M L ⋯ M 1 ) ⋅ x ) ] . (86)
Further, there exists C ≥ 1 C\ge1 C ≥ 1 , depending on s s s , p p p , K K K , and ν \nu ν , such that for every L ≥ 1 L\ge1 L ≥ 1 ,
C − 1 λ ( s ) L ≤ p L E J s ( M L ⋯ M 1 ) ≤ C λ ( s ) L . (87) C^{-1}\lambda(s)^L\le p^L\mathbb{E}J_s(M_L\cdots M_1)\le C\lambda(s)^L. \tag*{(87)} C − 1 λ ( s ) L ≤ p L E J s ( M L ⋯ M 1 ) ≤ C λ ( s ) L . (87)
In particular, the limit below exists and
log λ ( s ) = log p + lim L → ∞ 1 L log E J s ( M L ⋯ M 1 ) . \log\lambda(s)=\log p+\lim_{L\to\infty}\frac{1}{L}\log\mathbb{E}J_s(M_L\cdots M_1). log λ ( s ) = log p + L → ∞ lim L 1 log E J s ( M L ⋯ M 1 ) .
Proof. The proof proceeds as follows. Part (i) collects Lemmas 6.6, 6.7, and 6.9. We apply the Krein–Rutman theorem to the compact strongly positive operator T = Q s 3 T=Q_s^3 T = Q s 3 , and we transfer the conclusions to Q s Q_s Q s using the Riesz projection of T T T , which commutes with Q s Q_s Q s . Analyticity then follows from Lemma 6.10 and the holomorphy of z ↦ Q z z\mapsto Q_z z ↦ Q z , and (6.15) follows from two-sided bounds on ψ s \psi_s ψ s .
Step 1. Part (i). Boundedness and positivity are Lemma 6.6(i) and (ii), strong positivity is Lemma 6.9, and compactness of Q s 3 Q_s^3 Q s 3 is Lemma 6.7.
Step 2. The Perron–Frobenius eigenvector of T = Q s 3 T=Q_s^3 T = Q s 3 . Let K = { F ∈ C ( R P 1 ; R ) : F ≥ 0 } \mathcal{K}=\{F\in C(\mathbb{RP}^1;\mathbb{R}):F\ge0\} K = { F ∈ C ( RP 1 ; R ) : F ≥ 0 } . This is a closed cone in the real Banach space C ( R P 1 ; R ) C(\mathbb{RP}^1;\mathbb{R}) C ( RP 1 ; R ) , and its interior { F : min F > 0 } \{F:\min F>0\} { F : min F > 0 } is nonempty. By Step 1, the operator T T T is compact, and it maps K ∖ { 0 } \mathcal{K}\setminus\{0\} K ∖ { 0 } into the interior of K \mathcal{K} K . By the Krein–Rutman theorem [90], in the form for compact strongly positive operators [60, Theorem 19.3], the spectral radius r T r_T r T of T T T is positive, and there exists ψ \psi ψ in the interior of K \mathcal{K} K with T ψ = r T ψ T\psi=r_T\psi T ψ = r T ψ . Since ∣ T n F ∣ ≤ T n ∣ F ∣ |T^nF|\le T^n|F| ∣ T n F ∣ ≤ T n ∣ F ∣ for complex F F F , we have ∥ T n ∥ = ∥ T n 1 ∥ ∞ \lVert T^n\rVert=\lVert T^n1\rVert_\infty ∥ T n ∥ = ∥ T n 1 ∥ ∞ both on C ( R P 1 ; R ) C(\mathbb{RP}^1;\mathbb{R}) C ( RP 1 ; R ) and on C ( R P 1 ) C(\mathbb{RP}^1) C ( RP 1 ) , and by Gelfand’s formula, r T r_T r T is also the spectral radius of T T T on the complex space C ( R P 1 ) C(\mathbb{RP}^1) C ( RP 1 ) . The cited theorem also asserts that r T r_T r T is simple and is the only eigenvalue of modulus r T r_T r T . Since we work on the complex space C ( R P 1 ) C(\mathbb{RP}^1) C ( RP 1 ) and use algebraic simplicity below, we verify these two assertions directly.
First, let w ∈ C ( R P 1 ) w \in C(\mathbb{R}\mathbb{P}^{1}) w ∈ C ( R P 1 ) be nonzero with T w = μ w Tw=\mu w T w = μ w and ∣ μ ∣ = r T |\mu|=r_{T} ∣ μ ∣ = r T . Since ∣ T w ∣ ≤ T ∣ w ∣ |Tw| \leq T|w| ∣ T w ∣ ≤ T ∣ w ∣ pointwise, we have r T ∣ w ∣ ≤ T ∣ w ∣ r_{T}|w| \leq T|w| r T ∣ w ∣ ≤ T ∣ w ∣ . The function ψ / ∣ w ∣ \psi/|w| ψ /∣ w ∣ on { w ≠ 0 } \{w \neq0\} { w = 0 } tends to infinity at the boundary of this set, because min ψ > 0 \min\psi> 0 min ψ > 0 , and it attains its infimum t ∗ ∈ ( 0 , ∞ ) t_{*} \in(0,\infty) t ∗ ∈ ( 0 , ∞ ) at some x 0 x_{0} x 0 . The function χ = ψ − t ∗ ∣ w ∣ \chi=\psi-t_{*}|w| χ = ψ − t ∗ ∣ w ∣ is nonnegative and vanishes at x 0 x_{0} x 0 , and
T χ = r T ψ − t ∗ T ∣ w ∣ ≤ r T ψ − t ∗ r T ∣ w ∣ = r T χ . T\chi=r_{T}\psi-t_{*}T|w| \leq r_{T}\psi-t_{*}r_{T}|w|=r_{T}\chi. T χ = r T ψ − t ∗ T ∣ w ∣ ≤ r T ψ − t ∗ r T ∣ w ∣ = r T χ .
If χ ≠ 0 \chi\neq0 χ = 0 , then T χ ( x 0 ) > 0 T\chi(x_{0})>0 T χ ( x 0 ) > 0 by strong positivity, contradicting T χ ( x 0 ) ≤ r T χ ( x 0 ) = 0 T\chi(x_{0}) \leq r_{T}\chi(x_{0})=0 T χ ( x 0 ) ≤ r T χ ( x 0 ) = 0 . We conclude that ∣ w ∣ = ψ / t ∗ |w|=\psi/t_{*} ∣ w ∣ = ψ / t ∗ , and then T ∣ w ∣ = r T ∣ w ∣ = ∣ T w ∣ T|w|=r_{T}|w|=|Tw| T ∣ w ∣ = r T ∣ w ∣ = ∣ T w ∣ . Fix x ∈ R P 1 x \in\mathbb{R}\mathbb{P}^{1} x ∈ R P 1 . Writing T ∣ w ∣ ( x ) = ∫ ∣ w ∣ d κ s ( 3 ) ( x , ⋅ ) T|w|(x)=\int|w|\,\mathrm{d}\kappa_{s}^{(3)}(x,\cdot) T ∣ w ∣ ( x ) = ∫ ∣ w ∣ d κ s ( 3 ) ( x , ⋅ ) and ∣ T w ( x ) ∣ = ∣ ∫ w d κ s ( 3 ) ( x , ⋅ ) ∣ |Tw(x)|=\left|\int w\,\mathrm{d}\kappa_{s}^{(3)}(x,\cdot)\right| ∣ T w ( x ) ∣ = ∫ w d κ s ( 3 ) ( x , ⋅ ) , equality in the triangle inequality for integrals shows that w = c ∣ w ∣ w=c|w| w = c ∣ w ∣ holds κ s ( 3 ) ( x , ⋅ ) \kappa_{s}^{(3)}(x,\cdot) κ s ( 3 ) ( x , ⋅ ) -almost everywhere for some constant c c c with ∣ c ∣ = 1 |c|=1 ∣ c ∣ = 1 . Since both sides are continuous and κ s ( 3 ) ( x , ⋅ ) \kappa_{s}^{(3)}(x,\cdot) κ s ( 3 ) ( x , ⋅ ) has full support by Lemma 6.9, we have w = c ∣ w ∣ = ( c / t ∗ ) ψ w=c|w|=(c/t_{*})\psi w = c ∣ w ∣ = ( c / t ∗ ) ψ everywhere. We conclude that every eigenvector for an eigenvalue of modulus r T r_{T} r T is a multiple of ψ \psi ψ , and that r T r_{T} r T is the only eigenvalue of modulus r T r_{T} r T .
Second, we show that r T r_{T} r T is algebraically simple. Since T T T is compact, r T r_{T} r T is an isolated point of σ ( T ) \sigma(T) σ ( T ) and its Riesz projection Π \Pi Π has finite rank, equal to the dimension of the generalized eigenspace ⋃ k ker ( T − r T ) k \bigcup_{k}\ker(T-r_{T})^{k} ⋃ k ker ( T − r T ) k [82, Theorem III.6.26]. If this dimension were larger than one, then, since ker ( T − r T ) \ker(T-r_{T}) ker ( T − r T ) is spanned by ψ \psi ψ , there would exist w w w with ( T − r T ) w = ψ (T-r_{T})w=\psi ( T − r T ) w = ψ , and taking real parts we could assume that w w w is real. By induction,
T L w = r T L w + L r T L − 1 ψ for L ≥ 1. T^{L}w=r_{T}^{L}w+Lr_{T}^{L-1}\psi\quad\text{for } L\geq1. T L w = r T L w + L r T L − 1 ψ for L ≥ 1.
Let a = ∥ w ∥ ∞ / min ψ a=\|w\|_{\infty}/\min\psi a = ∥ w ∥ ∞ / min ψ , so that ∣ w ∣ ≤ a ψ |w|\leq a\psi ∣ w ∣ ≤ a ψ . Positivity gives ∣ T L w ∣ ≤ T L ∣ w ∣ ≤ a T L ψ = a r T L ψ |T^{L}w|\leq T^{L}|w|\leq aT^{L}\psi=ar_{T}^{L}\psi ∣ T L w ∣ ≤ T L ∣ w ∣ ≤ a T L ψ = a r T L ψ , and therefore
L r T L − 1 ψ ≤ ∣ T L w ∣ + r T L ∣ w ∣ ≤ 2 a r T L ψ . Lr_{T}^{L-1}\psi\leq|T^{L}w|+r_{T}^{L}|w|\leq2ar_{T}^{L}\psi. L r T L − 1 ψ ≤ ∣ T L w ∣ + r T L ∣ w ∣ ≤ 2 a r T L ψ .
This gives L ≤ 2 a r T L\leq2ar_{T} L ≤ 2 a r T for every L L L , which is impossible. This contradiction shows that Π \Pi Π has rank one, and Π F = ℓ ′ ( F ) ψ \Pi F=\ell^{\prime}(F)\psi Π F = ℓ ′ ( F ) ψ for some ℓ ′ ∈ C ( R P 1 ) ∗ \ell^{\prime}\in C(\mathbb{R}\mathbb{P}^{1})^{*} ℓ ′ ∈ C ( R P 1 ) ∗ with ℓ ′ ( ψ ) = 1 \ell^{\prime}(\psi)=1 ℓ ′ ( ψ ) = 1 .
Finally, by [82, Theorem III.6.26], the nonzero points of σ ( T ) \sigma(T) σ ( T ) are eigenvalues and can accumulate only at 0. By the first assertion above, every point of σ ( T ) ∖ { r T } \sigma(T)\setminus\{r_{T}\} σ ( T ) ∖ { r T } has modulus less than r T r_{T} r T , and only finitely many of these points have modulus at least r T / 2 r_{T}/2 r T /2 . Then there exists r ′ < r T r^{\prime}<r_{T} r ′ < r T with ∣ z ∣ ≤ r ′ |z|\leq r^{\prime} ∣ z ∣ ≤ r ′ for every z ∈ σ ( T ) ∖ { r T } z\in\sigma(T)\setminus\{r_{T}\} z ∈ σ ( T ) ∖ { r T } .
Step 3. Transfer to Q s Q_{s} Q s ; part (ii). Since Q s Q_{s} Q s commutes with T T T , it commutes with the resolvent of T T T and with Π \Pi Π . Then T ( Q s ψ ) = Q s T ψ = r T Q s ψ T(Q_{s}\psi)=Q_{s}T\psi=r_{T}Q_{s}\psi T ( Q s ψ ) = Q s T ψ = r T Q s ψ , and since ker ( T − r T ) \ker(T-r_{T}) ker ( T − r T ) is spanned by ψ \psi ψ , we have Q s ψ = λ ψ Q_{s}\psi=\lambda\psi Q s ψ = λ ψ for some scalar λ \lambda λ . Since Q s ψ ≥ ( min ψ ) Q s 1 Q_{s}\psi\geq(\min\psi)Q_{s}1 Q s ψ ≥ ( min ψ ) Q s 1 and Q s 1 > 0 Q_{s}1>0 Q s 1 > 0 , we have λ > 0 \lambda>0 λ > 0 , and λ 3 = r T \lambda^{3}=r_{T} λ 3 = r T . Let X 1 = ker Π X_{1}=\ker\Pi X 1 = ker Π . The decomposition C ( R P 1 ) = Ran Π ⊕ X 1 C(\mathbb{R}\mathbb{P}^{1})=\operatorname{Ran}\Pi\oplus X_{1} C ( R P 1 ) = Ran Π ⊕ X 1 is invariant under Q s Q_{s} Q s , and Q s Q_{s} Q s acts on Ran Π = span { ψ } \operatorname{Ran}\Pi=\operatorname{span}\{\psi\} Ran Π = span { ψ } as multiplication by λ \lambda λ . By the properties of Riesz projections [82, Section III.6.4], we have σ ( T ∣ X 1 ) = σ ( T ) ∖ { r T } ⊂ { ∣ z ∣ ≤ r ′ } \sigma(T|_{X_{1}})=\sigma(T)\setminus\{r_{T}\}\subset\{|z|\leq r^{\prime}\} σ ( T ∣ X 1 ) = σ ( T ) ∖ { r T } ⊂ { ∣ z ∣ ≤ r ′ } . By the spectral mapping theorem for polynomials, σ ( T ∣ X 1 ) = { z 3 : z ∈ σ ( Q s ∣ X 1 ) } \sigma(T|_{X_{1}})=\{z^{3}:z\in\sigma(Q_{s}|_{X_{1}})\} σ ( T ∣ X 1 ) = { z 3 : z ∈ σ ( Q s ∣ X 1 )} , and we obtain σ ( Q s ∣ X 1 ) ⊂ { ∣ z ∣ ≤ ( r ′ ) 1 / 3 } \sigma(Q_{s}|_{X_{1}})\subset\{|z|\leq(r^{\prime})^{1/3}\} σ ( Q s ∣ X 1 ) ⊂ { ∣ z ∣ ≤ ( r ′ ) 1/3 } . We set r ( s ) = ( r ′ ) 1 / 3 < λ r(s)=(r^{\prime})^{1/3}<\lambda r ( s ) = ( r ′ ) 1/3 < λ . Then σ ( Q s ) = { λ } ∪ σ ( Q s ∣ X 1 ) \sigma(Q_{s})=\{\lambda\}\cup\sigma(Q_{s}|_{X_{1}}) σ ( Q s ) = { λ } ∪ σ ( Q s ∣ X 1 ) , the point λ \lambda λ is isolated in σ ( Q s ) \sigma(Q_{s}) σ ( Q s ) , and its Riesz projection is the projection onto Ran Π \operatorname{Ran}\Pi Ran Π along X 1 X_1 X 1 , which is Π \Pi Π ; this has rank one. The spectral radius of Q s Q_s Q s is therefore λ \lambda λ , that is, λ = λ ( s ) \lambda=\lambda(s) λ = λ ( s ) . Setting ψ s = ψ \psi_s=\psi ψ s = ψ proves (ii).
Step 4. Part (iii). Let N = Q s ( I − Π ) N=Q_s(I-\Pi) N = Q s ( I − Π ) . Then Q s = λ ( s ) Π + N Q_s=\lambda(s)\Pi+N Q s = λ ( s ) Π + N and Π N = N Π = 0 \Pi N=N\Pi=0 Π N = N Π = 0 , which give Q s L = λ ( s ) L Π + N L Q_s^L=\lambda(s)^L\Pi+N^L Q s L = λ ( s ) L Π + N L for L ≥ 1 L\geq1 L ≥ 1 . The spectral radius of N N N equals that of Q s ∣ X 1 Q_s|_{X_1} Q s ∣ X 1 , which is at most r ( s ) r(s) r ( s ) . By Gelfand’s formula, we have ∥ N L ∥ 1 / L → r ( N ) ≤ r ( s ) < λ ( s ) \|N^L\|^{1/L}\to r(N)\leq r(s)<\lambda(s) ∥ N L ∥ 1/ L → r ( N ) ≤ r ( s ) < λ ( s ) . This gives ∥ λ ( s ) − L Q s L − Π ∥ = ∥ ( N / λ ( s ) ) L ∥ → 0 \|\lambda(s)^{-L}Q_s^L-\Pi\|=\|(N/\lambda(s))^L\|\to0 ∥ λ ( s ) − L Q s L − Π∥ = ∥ ( N / λ ( s ) ) L ∥ → 0 . In particular, Π \Pi Π is positive, being a norm limit of positive operators. For real F ≥ 0 F\geq0 F ≥ 0 , we have ℓ ′ ( F ) ψ = Π F ≥ 0 \ell'(F)\psi=\Pi F\geq0 ℓ ′ ( F ) ψ = Π F ≥ 0 . Then ℓ ′ \ell' ℓ ′ is a positive linear functional, and by the Riesz–Markov theorem it is given by integration against a finite positive Borel measure ℓ s \ell_s ℓ s . We have ∫ ψ s d ℓ s = ℓ ′ ( ψ ) = 1 \int\psi_s\,d\ell_s=\ell'(\psi)=1 ∫ ψ s d ℓ s = ℓ ′ ( ψ ) = 1 , and Π Q s = λ ( s ) Π \Pi Q_s=\lambda(s)\Pi Π Q s = λ ( s ) Π gives ∫ Q s F d ℓ s = λ ( s ) ∫ F d ℓ s \int Q_sF\,d\ell_s=\lambda(s)\int F\,d\ell_s ∫ Q s F d ℓ s = λ ( s ) ∫ F d ℓ s . For full support, let F ≥ 0 F\geq0 F ≥ 0 be real, continuous, and nonzero. By Step 1, there exists c > 0 c>0 c > 0 with Q s 3 F ≥ c ψ s Q_s^3F\geq c\psi_s Q s 3 F ≥ c ψ s , and positivity of Π \Pi Π gives
λ ( s ) 3 ( ∫ F d ℓ s ) ψ s = Π Q s 3 F ≥ c Π ψ s = c ψ s . \lambda(s)^3\left(\int F\,d\ell_s\right)\psi_s=\Pi Q_s^3F\geq c\Pi\psi_s=c\psi_s. λ ( s ) 3 ( ∫ F d ℓ s ) ψ s = Π Q s 3 F ≥ c Π ψ s = c ψ s .
Since ψ s > 0 \psi_s>0 ψ s > 0 , we have ∫ F d ℓ s > 0 \int F\,d\ell_s>0 ∫ F d ℓ s > 0 . Applying this to a nonzero continuous F ≥ 0 F\geq0 F ≥ 0 supported in a given nonempty open set shows that ℓ s \ell_s ℓ s has full support.
Step 5. Part (iv). Fix s 0 ∈ ( 0 , 2 ) s_0\in(0,2) s 0 ∈ ( 0 , 2 ) . By Step 3 and Step 4, the operator T 0 = Q s 0 T_0=Q_{s_0} T 0 = Q s 0 satisfies the hypotheses of Lemma 6.10 with μ 0 = λ ( s 0 ) \mu_0=\lambda(s_0) μ 0 = λ ( s 0 ) , r 0 = r ( s 0 ) r_0=r(s_0) r 0 = r ( s 0 ) , ψ = ψ s 0 \psi=\psi_{s_0} ψ = ψ s 0 , and ℓ = ℓ s 0 \ell=\ell_{s_0} ℓ = ℓ s 0 ; here ℓ ( Π 0 ψ ) = ∫ ψ s 0 d ℓ s 0 = 1 \ell(\Pi_0\psi)=\int\psi_{s_0}\,d\ell_{s_0}=1 ℓ ( Π 0 ψ ) = ∫ ψ s 0 d ℓ s 0 = 1 , since the Riesz projection Π 0 \Pi_0 Π 0 of Q s 0 Q_{s_0} Q s 0 at λ ( s 0 ) \lambda(s_0) λ ( s 0 ) is F ↦ ( ∫ F d ℓ s 0 ) ψ s 0 F\mapsto\left(\int F\,d\ell_{s_0}\right)\psi_{s_0} F ↦ ( ∫ F d ℓ s 0 ) ψ s 0 by Steps 3 and 4. Let ε > 0 \varepsilon>0 ε > 0 be as in that lemma. By Lemma 6.6 (iii), z ↦ Q z z\mapsto Q_z z ↦ Q z is holomorphic, in particular continuous, and there exists an open disc U ⊂ Σ U\subset\Sigma U ⊂ Σ centered at s 0 s_0 s 0 with ∥ Q z − Q s 0 ∥ < ε \|Q_z-Q_{s_0}\|<\varepsilon ∥ Q z − Q s 0 ∥ < ε for z ∈ U z\in U z ∈ U . By Lemma 6.10 (iii), z ↦ μ ( Q z ) z\mapsto\mu(Q_z) z ↦ μ ( Q z ) is holomorphic on U U U . For real s ∈ U s\in U s ∈ U , the spectral radius λ ( s ) \lambda(s) λ ( s ) belongs to σ ( Q s ) \sigma(Q_s) σ ( Q s ) by (ii) and has maximal modulus in σ ( Q s ) \sigma(Q_s) σ ( Q s ) , and Lemma 6.10 (i) implies that λ ( s ) = μ ( Q s ) \lambda(s)=\mu(Q_s) λ ( s ) = μ ( Q s ) . Then λ \lambda λ coincides on U ∩ R U\cap\mathbb{R} U ∩ R with a holomorphic function, which proves that λ \lambda λ is real-analytic near s 0 s_0 s 0 .
Step 6. Part (v). We prove (86) by induction on L L L ; the case L = 1 L=1 L = 1 is the definition of Q s Q_s Q s . Suppose (86) holds for L L L , and let M 1 , … , M L + 1 M_1,\ldots,M_{L+1} M 1 , … , M L + 1 be independent copies of M ( B , β ) M(B,\beta) M ( B , β ) . Then Q s L + 1 F ( x ) = p E [ ∥ M 1 v x ∥ − s ( Q s L F ) ( M 1 ⋅ x ) ] Q_s^{L+1}F(x)=p\,\mathbb{E}\left[\|M_1v_x\|^{-s}(Q_s^LF)(M_1\cdot x)\right] Q s L + 1 F ( x ) = p E [ ∥ M 1 v x ∥ − s ( Q s L F ) ( M 1 ⋅ x ) ] . We apply the induction hypothesis to Q s L F Q_s^LF Q s L F with the matrices M 2 , … , M L + 1 M_2,\ldots,M_{L+1} M 2 , … , M L + 1 , which are independent of M 1 M_1 M 1 , and use Fubini’s theorem. Fubini’s theorem applies because, by Tonelli’s theorem and the same computation with ∣ F ∣ |F| ∣ F ∣ in place of F F F , the absolute value of the integrand has expectation at most p − L − 1 Q s L + 1 ∣ F ∣ ( x ) ≤ p − L − 1 ∥ Q s ∥ L + 1 ∥ F ∥ ∞ p^{-L-1}Q_s^{L+1}|F|(x)\leq p^{-L-1}\|Q_s\|^{L+1}\|F\|_\infty p − L − 1 Q s L + 1 ∣ F ∣ ( x ) ≤ p − L − 1 ∥ Q s ∥ L + 1 ∥ F ∥ ∞ . We obtain
Q s L + 1 F ( x ) = p L + 1 E [ ∥ M 1 v x ∥ − s ∥ M L + 1 ⋯ M 2 v M 1 ⋅ x ∥ − s F ( ( M L + 1 ⋯ M 2 ) ⋅ ( M 1 ⋅ x ) ) ] . Q_s^{L+1}F(x)=p^{L+1}\mathbb{E}\left[\|M_1v_x\|^{-s}\|M_{L+1}\cdots M_2v_{M_1\cdot x}\|^{-s}F((M_{L+1}\cdots M_2)\cdot(M_1\cdot x))\right]. Q s L + 1 F ( x ) = p L + 1 E [ ∥ M 1 v x ∥ − s ∥ M L + 1 ⋯ M 2 v M 1 ⋅ x ∥ − s F (( M L + 1 ⋯ M 2 ) ⋅ ( M 1 ⋅ x )) ] .
By Lemma 6.2 (i), we have ∥ M L + 1 ⋯ M 2 v M 1 ⋅ x ∥ ∥ M 1 v x ∥ = ∥ M L + 1 ⋯ M 1 v x ∥ \|M_{L+1}\cdots M_2v_{M_1\cdot x}\|\|M_1v_x\|=\|M_{L+1}\cdots M_1v_x\| ∥ M L + 1 ⋯ M 2 v M 1 ⋅ x ∥∥ M 1 v x ∥ = ∥ M L + 1 ⋯ M 1 v x ∥ and ( M L + 1 ⋯ M 2 ) ⋅ ( M 1 ⋅ x ) = ( M L + 1 ⋯ M 1 ) ⋅ x (M_{L+1}\cdots M_2)\cdot(M_1\cdot x)=(M_{L+1}\cdots M_1)\cdot x ( M L + 1 ⋯ M 2 ) ⋅ ( M 1 ⋅ x ) = ( M L + 1 ⋯ M 1 ) ⋅ x . This completes the induction.
Taking F = 1 F=1 F = 1 in (86) , integrating against m m m , and using Tonelli’s theorem, we obtain ∫ Q s L 1 d m = λ L E J s ( M L ⋯ M 1 ) \int Q_s^L \mathbf{1}\,dm=\lambda^L\mathbb{E}J_s(M_L\cdots M_1) ∫ Q s L 1 d m = λ L E J s ( M L ⋯ M 1 ) . Let a = min ψ s > 0 a=\min\psi_s>0 a = min ψ s > 0 and A = max ψ s A=\max\psi_s A = max ψ s . Since Q s L Q_s^L Q s L is positive and a ≤ ψ s ≤ A a\leq\psi_s\leq A a ≤ ψ s ≤ A , we have
a A λ ( s ) L ≤ 1 A Q s L ψ s ≤ Q s L 1 ≤ 1 a Q s L ψ s ≤ A a λ ( s ) L \frac{a}{A}\lambda(s)^L\leq\frac{1}{A}Q_s^L\psi_s\leq Q_s^L\mathbf{1}\leq\frac{1}{a}Q_s^L\psi_s\leq\frac{A}{a}\lambda(s)^L A a λ ( s ) L ≤ A 1 Q s L ψ s ≤ Q s L 1 ≤ a 1 Q s L ψ s ≤ a A λ ( s ) L
pointwise on R P 1 \mathbb{RP}^1 RP 1 . Integrating against the probability measure m m m gives (6.15) with C = A / a C=A/a C = A / a . Taking logarithms, dividing by L L L , and letting L → ∞ L\to\infty L → ∞ gives the last statement.
Continuity in the law of the denominator
We now show that the Perron–Frobenius eigenvalue depends continuously on the law of B B B and on the rate of β \beta β , under weak convergence of the laws and a uniform moment bound. Since atoms of ν \nu ν produce atoms in the kernel measures of Q s Q_s Q s (see Remark 6.8), the operator Q s Q_s Q s itself is not norm-continuous in this sense in general, and we work with the third power.
Proposition 6.12. Let I ⊂ R I\subset\mathbb{R} I ⊂ R be an interval. For each E ∈ I E\in I E ∈ I , let ν E \nu_E ν E be a Borel probability measure on R \mathbb{R} R satisfying (73) , and let K ( E ) > 0 K(E)>0 K ( E ) > 0 . Suppose that E ↦ K ( E ) E\mapsto K(E) E ↦ K ( E ) is continuous, that E ↦ ν E E\mapsto\nu_E E ↦ ν E is continuous for the topology of weak convergence, and that there exists q 0 ∈ ( 0 , 1 ) q_0\in(0,1) q 0 ∈ ( 0 , 1 ) such that
sup E ∈ I ′ ∫ ∣ b ∣ q 0 ν E ( d b ) < ∞ for every compact I ′ ⊂ I . \sup_{E\in I'}\int|b|^{q_0}\,\nu_E(db)<\infty\qquad\text{for every compact }I'\subset I. E ∈ I ′ sup ∫ ∣ b ∣ q 0 ν E ( d b ) < ∞ for every compact I ′ ⊂ I . For E ∈ I E\in I E ∈ I and s ∈ ( 0 , 2 ) s\in(0,2) s ∈ ( 0 , 2 ) , let Q E , s Q_{E,s} Q E , s and λ E ( s ) \lambda_E(s) λ E ( s ) be the transfer operator and its Perron–Frobenius eigenvalue from Theorem 6.11 , defined with the law ν E \nu_E ν E and the rate K ( E ) K(E) K ( E ) . Then for every s ∈ ( 0 , 2 q 0 ) s\in(0,2q_0) s ∈ ( 0 , 2 q 0 ) , the map E ↦ Q E , s 3 E\mapsto Q_{E,s}^3 E ↦ Q E , s 3 is continuous from I I I to the bounded operators on C ( R P 1 ) C(\mathbb{RP}^1) C ( RP 1 ) in operator norm, and E ↦ λ E ( s ) E\mapsto\lambda_E(s) E ↦ λ E ( s ) is continuous on I I I . In particular, if the moment bound holds for some q 0 > 1 / 2 q_0>1/2 q 0 > 1/2 , then E ↦ λ E ( 1 ) E\mapsto\lambda_E(1) E ↦ λ E ( 1 ) is continuous on I I I , and if it holds for every q 0 ∈ ( 0 , 1 ) q_0\in(0,1) q 0 ∈ ( 0 , 1 ) , then E ↦ λ E ( s ) E\mapsto\lambda_E(s) E ↦ λ E ( s ) is continuous on I I I for every s ∈ ( 0 , 2 ) s\in(0,2) s ∈ ( 0 , 2 ) .
Proof. The proof repeats the total variation argument of Lemma 6.7, with the law and the input varying together, on a probability space on which the denominators converge almost surely. Fix s ∈ ( 0 , 2 q 0 ) s\in(0,2q_0) s ∈ ( 0 , 2 q 0 ) , E 0 ∈ I E_0\in I E 0 ∈ I , and a sequence E n → E 0 E_n\to E_0 E n → E 0 in I I I . The set I ′ = { E 0 } ∪ { E n : n ≥ 1 } I'=\{E_0\}\cup\{E_n:n\geq1\} I ′ = { E 0 } ∪ { E n : n ≥ 1 } is a compact subset of I I I . We write K n = K ( E n ) K_n=K(E_n) K n = K ( E n ) for n ≥ 0 n\geq0 n ≥ 0 , and let k − = inf n K n > 0 k_-=\inf_n K_n>0 k − = inf n K n > 0 and k + = sup n K n < ∞ k_+=\sup_n K_n<\infty k + = sup n K n < ∞ . Fix t ∈ ( s , 2 q 0 ) t\in(s,2q_0) t ∈ ( s , 2 q 0 ) . In the following, n n n ranges over { 0 , 1 , 2 , … } \{0,1,2,\ldots\} { 0 , 1 , 2 , … } , and the index 0 0 0 refers to E 0 E_0 E 0 .
Step 1. Coupling and uniform bounds. By the Skorokhod representation theorem [37, Theorem 6.7], there exist random variables B n B_n B n with law ν E n \nu_{E_n} ν E n , defined on one probability space, such that B n → B 0 B_n\to B_0 B n → B 0 almost surely. Enlarging the space if necessary, let Y Y Y be an Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) random variable independent of ( B n ) n ≥ 0 (B_n)_{n\geq0} ( B n ) n ≥ 0 , and set β n = Y / K n \beta_n=Y/K_n β n = Y / K n . Then ( B n , β n ) (B_n,\beta_n) ( B n , β n ) has law ν E n ⊗ Exp ( K n ) \nu_{E_n} \otimes\operatorname{Exp}(K_n) ν E n ⊗ Exp ( K n ) for each n n n , and β n → β 0 \beta_n \to\beta_0 β n → β 0 almost surely. For σ ∈ { s , t } \sigma\in\{s,t\} σ ∈ { s , t } , by Lemma 6.5, we have
m σ = sup n ≥ 0 sup x ∈ R P 1 E [ R σ ( x ; B n , β n ) ] ≤ sup n ≥ 0 C σ , K n ( 2 + E ∣ B n ∣ q 0 ) < ∞ , (88) m_\sigma= \sup_{n \ge0}\sup_{x \in\mathbb{R}\mathbb{P}^{1}} \mathbb{E}\left[R_\sigma(x;B_n,\beta_n)\right] \le\sup_{n \ge0} C_{\sigma,K_n}\left(2+\mathbb{E}|B_n|^{q_0}\right)<\infty, \tag*{(88)} m σ = n ≥ 0 sup x ∈ R P 1 sup E [ R σ ( x ; B n , β n ) ] ≤ n ≥ 0 sup C σ , K n ( 2 + E ∣ B n ∣ q 0 ) < ∞ , (88)
where we used ∣ b ∣ σ / 2 ≤ 1 + ∣ b ∣ q 0 |b|^{\sigma/2} \le1+|b|^{q_0} ∣ b ∣ σ /2 ≤ 1 + ∣ b ∣ q 0 (since σ / 2 < q 0 \sigma/2<q_0 σ /2 < q 0 ), the continuity of C σ , K C_{\sigma,K} C σ , K in K K K on [ k − , k + ] [k_-,k_+] [ k − , k + ] , and the moment bound on I ′ I' I ′ . In particular, sup n ∥ Q E n , s ∥ ≤ p m s \sup_n\|Q_{E_n,s}\|\le p m_s sup n ∥ Q E n , s ∥ ≤ p m s . Further, since R s t / s = R t R_s^{t/s}=R_t R s t / s = R t and t / s > 1 t/s>1 t / s > 1 , every family of random variables of the form R s ( x n ; B n , β n ) R_s(x_n;B_n,\beta_n) R s ( x n ; B n , β n ) with x n ∈ R P 1 x_n\in\mathbb{R}\mathbb{P}^{1} x n ∈ R P 1 is bounded in L t / s L^{t/s} L t / s , and in particular uniformly integrable.
Let κ n ( k ) ( x , ⋅ ) \kappa_n^{(k)}(x,\cdot) κ n ( k ) ( x , ⋅ ) denote the k k k -step kernel measure of Q E n , s Q_{E_n,s} Q E n , s at x x x . By (6.10), we have κ n ( k ) ( x , R P 1 ) ≤ ( p m s ) k \kappa_n^{(k)}(x,\mathbb{R}\mathbb{P}^{1})\le(p m_s)^k κ n ( k ) ( x , R P 1 ) ≤ ( p m s ) k for all n , k , x n,k,x n , k , x . We say that the k k k -step kernel measures are jointly continuous at y ∈ R P 1 y\in\mathbb{R}\mathbb{P}^{1} y ∈ R P 1 if ∥ κ n ( k ) ( y n , ⋅ ) − κ 0 ( k ) ( y , ⋅ ) ∥ T V → 0 \|\kappa_n^{(k)}(y_n,\cdot)-\kappa_0^{(k)}(y,\cdot)\|_{\mathrm{TV}}\to0 ∥ κ n ( k ) ( y n , ⋅ ) − κ 0 ( k ) ( y , ⋅ ) ∥ TV → 0 for every sequence y n → y y_n\to y y n → y in R P 1 \mathbb{R}\mathbb{P}^{1} R P 1 . The maps ( b , β , x ) ↦ ∥ M ( b , β ) v x ∥ − s (b,\beta,x)\mapsto\|M(b,\beta)v_x\|^{-s} ( b , β , x ) ↦ ∥ M ( b , β ) v x ∥ − s and ( b , β , x ) ↦ f b , β ( x ) (b,\beta,x)\mapsto f_{b,\beta}(x) ( b , β , x ) ↦ f b , β ( x ) are continuous on R × ( 0 , ∞ ) × R P 1 \mathbb{R}\times(0,\infty)\times\mathbb{R}\mathbb{P}^{1} R × ( 0 , ∞ ) × R P 1 , because M ( b , β ) M(b,\beta) M ( b , β ) depends continuously on ( b , β ) (b,\beta) ( b , β ) and x ↦ v x x\mapsto v_x x ↦ v x is continuous up to sign.
Step 2. The one-step kernel measures are jointly continuous at every y ∈ R ∖ { 0 } y\in\mathbb{R}\setminus\{0\} y ∈ R ∖ { 0 } . For K > 0 K>0 K > 0 , let ϕ K ( u ) = u s / 2 e − K u 1 { u > 0 } \phi_K(u)=u^{s/2}e^{-Ku}\mathbf{1}_{\{u>0\}} ϕ K ( u ) = u s /2 e − K u 1 { u > 0 } . The map ( K , u ) ↦ ϕ K ( u ) (K,u)\mapsto\phi_K(u) ( K , u ) ↦ ϕ K ( u ) is continuous on ( 0 , ∞ ) × R (0,\infty)\times\mathbb{R} ( 0 , ∞ ) × R , and ϕ K ≤ sup u > 0 u s / 2 e − k − u < ∞ \phi_K\le\sup_{u>0}u^{s/2}e^{-k_-u}<\infty ϕ K ≤ sup u > 0 u s /2 e − k − u < ∞ for K ≥ k − K\ge k_- K ≥ k − . Let y n → y ∈ R ∖ { 0 } y_n\to y\in\mathbb{R}\setminus\{0\} y n → y ∈ R ∖ { 0 } ; then y n ∈ R ∖ { 0 } y_n\in\mathbb{R}\setminus\{0\} y n ∈ R ∖ { 0 } for large n n n . By Lemma 6.7(i), applied to ν E n \nu_{E_n} ν E n and K n K_n K n , the kernel measure κ n ( 1 ) ( y n , ⋅ ) \kappa_n^{(1)}(y_n,\cdot) κ n ( 1 ) ( y n , ⋅ ) is the image under ι \iota ι of the measure with density
k n ( w ) = p K n ∣ y n ∣ ( 1 + y n 2 ) s / 2 ( 1 + w 2 ) − s / 2 E [ ϕ K n ( B n − w y n ) ] . k_n(w)=\frac{pK_n}{|y_n|}(1+y_n^2)^{s/2}(1+w^2)^{-s/2}\mathbb{E}\left[\phi_{K_n}\left(\frac{B_n-w}{y_n}\right)\right]. k n ( w ) = ∣ y n ∣ p K n ( 1 + y n 2 ) s /2 ( 1 + w 2 ) − s /2 E [ ϕ K n ( y n B n − w ) ] .
For each w w w , the random variables ϕ K n ( ( B n − w ) / y n ) \phi_{K_n}((B_n-w)/y_n) ϕ K n (( B n − w ) / y n ) converge almost surely to ϕ K 0 ( ( B 0 − w ) / y ) \phi_{K_0}((B_0-w)/y) ϕ K 0 (( B 0 − w ) / y ) and are uniformly bounded. By dominated convergence, k n ( w ) k_n(w) k n ( w ) converges to the density k ∞ ( w ) k_\infty(w) k ∞ ( w ) of κ 0 ( 1 ) ( y , ⋅ ) \kappa_0^{(1)}(y,\cdot) κ 0 ( 1 ) ( y , ⋅ ) . Further, ∫ k n = p E R s ( y n ; B n , β n ) → p E R s ( y ; B 0 , β 0 ) = ∫ k ∞ \int k_n=p\mathbb{E}R_s(y_n;B_n,\beta_n)\to p\mathbb{E}R_s(y;B_0,\beta_0)=\int k_\infty ∫ k n = p E R s ( y n ; B n , β n ) → p E R s ( y ; B 0 , β 0 ) = ∫ k ∞ , by almost sure convergence and uniform integrability (Step 1). Scheffé’s lemma, in the form proved in Lemma 6.7(ii), gives ∫ ∣ k n − k ∞ ∣ → 0 \int|k_n-k_\infty|\to0 ∫ ∣ k n − k ∞ ∣ → 0 , which is the claim.
Step 3. Propagation. We claim that if k ≥ 1 k\ge1 k ≥ 1 , x ∈ R P 1 x\in\mathbb{R}\mathbb{P}^{1} x ∈ R P 1 , and the k k k -step kernel measures are jointly continuous at f B 0 , β 0 ( x ) f_{B_0,\beta_0}(x) f B 0 , β 0 ( x ) almost surely, then the ( k + 1 ) (k+1) ( k + 1 ) -step kernel measures are jointly continuous at x x x . Let x n → x x_n\to x x n → x , and write R ( n ) = R s ( x n ; B n , β n ) R^{(n)}=R_s(x_n;B_n,\beta_n) R ( n ) = R s ( x n ; B n , β n ) and y n = f B n , β n ( x n ) y_n=f_{B_n,\beta_n}(x_n) y n = f B n , β n ( x n ) for n ≥ 1 n\ge1 n ≥ 1 , and R = R s ( x ; B 0 , β 0 ) R=R_s(x;B_0,\beta_0) R = R s ( x ; B 0 , β 0 ) and y = f B 0 , β 0 ( x ) y=f_{B_0,\beta_0}(x) y = f B 0 , β 0 ( x ) . Since (6.11) is an identity between expectations over the law of ( B n , β n ) (B_n,\beta_n) ( B n , β n ) , it holds on the coupled space, and we obtain
∥ κ n ( k + 1 ) ( x n , ⋅ ) − κ 0 ( k + 1 ) ( x , ⋅ ) ∥ T V ≤ p ( p m s ) k E ∣ R ( n ) − R ∣ + p E [ R ∥ κ n ( k ) ( y n , ⋅ ) − κ 0 ( k ) ( y , ⋅ ) ∥ T V ] . \left\|\kappa_n^{(k+1)}(x_n,\cdot)-\kappa_0^{(k+1)}(x,\cdot)\right\|_{\mathrm{TV}}
\le p(pm_s)^k\mathbb{E}\left|R^{(n)}-R\right|+p\mathbb{E}\left[R\left\|\kappa_n^{(k)}(y_n,\cdot)-\kappa_0^{(k)}(y,\cdot)\right\|_{\mathrm{TV}}\right]. κ n ( k + 1 ) ( x n , ⋅ ) − κ 0 ( k + 1 ) ( x , ⋅ ) TV ≤ p ( p m s ) k E R ( n ) − R + p E [ R κ n ( k ) ( y n , ⋅ ) − κ 0 ( k ) ( y , ⋅ ) TV ] .
By Step 1, R ( n ) → R R^{(n)}\to R R ( n ) → R and y n → y y_n\to y y n → y almost surely. The first term tends to zero by uniform integrability. The integrand in the second term is measurable, by the countable supremum used in the proof of Lemma 6.7. For almost every outcome, the hypothesis applies to the sequence y n → y y_n \to y y n → y , and this integrand tends to zero; it is bounded by 2 ( p m s ) k R 2(p m_s)^k R 2 ( p m s ) k R , which is integrable. Dominated convergence completes the proof of the claim.
Step 4. The three-step kernel measures are jointly continuous at every x ∈ R P 1 x \in\mathbb{R}\mathbb{P}^1 x ∈ R P 1 . As in the proof of Lemma 6.7, for x ∈ R x \in\mathbb{R} x ∈ R the point f B 0 , β 0 ( x ) f_{B_0,\beta_0}(x) f B 0 , β 0 ( x ) lies in R ∖ { 0 } \mathbb{R} \setminus\{0\} R ∖ { 0 } almost surely. By Step 2 and Step 3 with k = 1 k=1 k = 1 , the two-step kernel measures are jointly continuous at every x ∈ R x \in\mathbb{R} x ∈ R . For x ∈ R x \in\mathbb{R} x ∈ R , the point f B 0 , β 0 ( x ) f_{B_0,\beta_0}(x) f B 0 , β 0 ( x ) lies in R \mathbb{R} R almost surely, and f B 0 , β 0 ( ∞ ) = 0 f_{B_0,\beta_0}(\infty)=0 f B 0 , β 0 ( ∞ ) = 0 . Step 3 with k = 2 k=2 k = 2 implies the claim.
Step 5. Norm continuity of the third power. We show that sup x ∈ R P 1 ∥ κ n ( 3 ) ( x , ⋅ ) − κ 0 ( 3 ) ( x , ⋅ ) ∥ T V → 0 \sup_{x\in\mathbb{R}\mathbb{P}^1}\lVert\kappa_n^{(3)}(x,\cdot)-\kappa_0^{(3)}(x,\cdot)\rVert_{\mathrm{TV}}\to0 sup x ∈ R P 1 ∥ κ n ( 3 ) ( x , ⋅ ) − κ 0 ( 3 ) ( x , ⋅ ) ∥ TV → 0 . If not, then there exist η > 0 \eta>0 η > 0 , a subsequence (which we do not relabel), and points x n x_n x n with ∥ κ n ( 3 ) ( x n , ⋅ ) − κ 0 ( 3 ) ( x n , ⋅ ) ∥ T V ≥ η \lVert\kappa_n^{(3)}(x_n,\cdot)-\kappa_0^{(3)}(x_n,\cdot)\rVert_{\mathrm{TV}}\geq\eta ∥ κ n ( 3 ) ( x n , ⋅ ) − κ 0 ( 3 ) ( x n , ⋅ ) ∥ TV ≥ η . By compactness of R P 1 \mathbb{R}\mathbb{P}^1 R P 1 , we may pass to a further subsequence with x n → x x_n\to x x n → x . Steps 1–4 apply to every sequence converging to E 0 E_0 E 0 , in particular to this subsequence, and they give ∥ κ n ( 3 ) ( x n , ⋅ ) − κ 0 ( 3 ) ( x , ⋅ ) ∥ T V → 0 \lVert\kappa_n^{(3)}(x_n,\cdot)-\kappa_0^{(3)}(x,\cdot)\rVert_{\mathrm{TV}}\to0 ∥ κ n ( 3 ) ( x n , ⋅ ) − κ 0 ( 3 ) ( x , ⋅ ) ∥ TV → 0 . By Lemma 6.7(iv) for ν E 0 \nu_{E_0} ν E 0 , we also have ∥ κ 0 ( 3 ) ( x n , ⋅ ) − κ 0 ( 3 ) ( x , ⋅ ) ∥ T V → 0 \lVert\kappa_0^{(3)}(x_n,\cdot)-\kappa_0^{(3)}(x,\cdot)\rVert_{\mathrm{TV}}\to0 ∥ κ 0 ( 3 ) ( x n , ⋅ ) − κ 0 ( 3 ) ( x , ⋅ ) ∥ TV → 0 . This contradicts the choice of x n x_n x n . Since ∣ Q E n , s 3 F ( x ) − Q E 0 , s 3 F ( x ) ∣ ≤ ∥ F ∥ ∞ ∥ κ n ( 3 ) ( x , ⋅ ) − κ 0 ( 3 ) ( x , ⋅ ) ∥ T V \lvert Q_{E_n,s}^3F(x)-Q_{E_0,s}^3F(x)\rvert\leq\lVert F\rVert_\infty\lVert\kappa_n^{(3)}(x,\cdot)-\kappa_0^{(3)}(x,\cdot)\rVert_{\mathrm{TV}} ∣ Q E n , s 3 F ( x ) − Q E 0 , s 3 F ( x )∣ ≤ ∥ F ∥ ∞ ∥ κ n ( 3 ) ( x , ⋅ ) − κ 0 ( 3 ) ( x , ⋅ ) ∥ TV , we conclude that ∥ Q E n , s 3 − Q E 0 , s 3 ∥ → 0 \lVert Q_{E_n,s}^3-Q_{E_0,s}^3\rVert\to0 ∥ Q E n , s 3 − Q E 0 , s 3 ∥ → 0 .
Step 6. Continuity of the eigenvalue. Let T n = Q E n , s 3 T_n=Q_{E_n,s}^3 T n = Q E n , s 3 . By Step 2 of the proof of Theorem 6.11, applied to ν E 0 \nu_{E_0} ν E 0 and K ( E 0 ) K(E_0) K ( E 0 ) , the point μ 0 = λ E 0 ( s ) 3 \mu_0=\lambda_{E_0}(s)^3 μ 0 = λ E 0 ( s ) 3 is isolated in σ ( T 0 ) \sigma(T_0) σ ( T 0 ) , its Riesz projection Π \Pi Π has rank one, there exists r ′ < μ 0 r'<\mu_0 r ′ < μ 0 with ∣ z ∣ ≤ r ′ \lvert z\rvert\leq r' ∣ z ∣ ≤ r ′ for every z ∈ σ ( T 0 ) ∖ { μ 0 } z\in\sigma(T_0)\setminus\{\mu_0\} z ∈ σ ( T 0 ) ∖ { μ 0 } , and Π ψ = ψ \Pi\psi=\psi Π ψ = ψ for ψ = ψ E 0 , s \psi=\psi_{E_0,s} ψ = ψ E 0 , s . Let ℓ ∈ C ( R P 1 ) ∗ \ell\in C(\mathbb{R}\mathbb{P}^1)^* ℓ ∈ C ( R P 1 ) ∗ be the functional with Π F = ℓ ( F ) ψ \Pi F=\ell(F)\psi Π F = ℓ ( F ) ψ , so that ℓ ( Π ψ ) = 1 \ell(\Pi\psi)=1 ℓ ( Π ψ ) = 1 . Fix δ ∈ ( 0 , ( μ 0 − r ′ ) / 3 ) \delta\in(0,(\mu_0-r')/3) δ ∈ ( 0 , ( μ 0 − r ′ ) /3 ) and let ε \varepsilon ε be as in Lemma 6.10. By Step 5, ∥ T n − T 0 ∥ < ε \lVert T_n-T_0\rVert<\varepsilon ∥ T n − T 0 ∥ < ε for all large n n n . For such n n n , Lemma 6.10(i) shows that σ ( T n ) \sigma(T_n) σ ( T n ) has a unique point of maximal modulus, and that this point lies within distance δ \delta δ of μ 0 \mu_0 μ 0 . By Theorem 6.11(ii) for ν E n \nu_{E_n} ν E n and the spectral mapping theorem, the spectral radius of T n T_n T n is λ E n ( s ) 3 \lambda_{E_n}(s)^3 λ E n ( s ) 3 and belongs to σ ( T n ) \sigma(T_n) σ ( T n ) ; it is therefore this point, and ∣ λ E n ( s ) 3 − λ E 0 ( s ) 3 ∣ < δ \lvert\lambda_{E_n}(s)^3-\lambda_{E_0}(s)^3\rvert<\delta ∣ λ E n ( s ) 3 − λ E 0 ( s ) 3 ∣ < δ for all large n n n . Since δ \delta δ can be taken arbitrarily small, λ E n ( s ) → λ E 0 ( s ) \lambda_{E_n}(s)\to\lambda_{E_0}(s) λ E n ( s ) → λ E 0 ( s ) . This completes the proof.
The free energy
Independence of the scaling and the definition of the free energy
The transfer matrices of a real solution depend on the choice of scaling g > 0 g>0 g > 0 . The next lemma shows that different scalings give similar transfer operators.
Lemma 6.13 (Change of scaling). Let E > 0 E>0 E > 0 and let π \pi π be a real solution at energy E E E . For g > 0 g>0 g > 0 , let ν g \nu_g ν g be the law of g D gD g D , where D D D has law π \pi π , and let K g = p / g 2 K_g=p/g^2 K g = p / g 2 .
For every g > 0 g>0 g > 0 , the law ν g \nu_g ν g satisfies (6.1), and ∫ ∣ b ∣ − q ν g ( d b ) < ∞ \int\lvert b\rvert^{-q}\,\nu_g(db)<\infty ∫ ∣ b ∣ − q ν g ( d b ) < ∞ for every q ∈ ( 0 , 1 ) q\in(0,1) q ∈ ( 0 , 1 ) .
(ii) Let s ∈ ( 0 , 2 ) s \in(0,2) s ∈ ( 0 , 2 ) , and for g > 0 g > 0 g > 0 let Q s ( g ) Q_s^{(g)} Q s ( g ) be the transfer operator defined with ν g \nu_g ν g and K g K_g K g . For g , g ′ > 0 g,g' > 0 g , g ′ > 0 , set S = diag ( c − 1 / 2 , c 1 / 2 ) S = \operatorname{diag}(c^{-1/2},c^{1/2}) S = diag ( c − 1/2 , c 1/2 ) with c = g ′ / g c = g'/g c = g ′ / g , and define ( Θ F ) ( y ) = ∥ S v y ∥ − s F ( S ⋅ y ) (\Theta F)(y) = \lVert S v_y\rVert^{-s}F(S \mathbin{\cdot} y) ( Θ F ) ( y ) = ∥ S v y ∥ − s F ( S ⋅ y ) . Then Θ \Theta Θ is a bounded invertible operator on C ( R P 1 ) C(\mathbb{R}\mathbb{P}^{1}) C ( R P 1 ) , both Θ \Theta Θ and Θ − 1 \Theta^{-1} Θ − 1 are positive, and
Q s ( g ′ ) = Θ − 1 Q s ( g ) Θ . Q_s^{(g')} = \Theta^{-1} Q_s^{(g)} \Theta. Q s ( g ′ ) = Θ − 1 Q s ( g ) Θ.
In particular, the spectral radius of Q s ( g ) Q_s^{(g)} Q s ( g ) does not depend on g g g .
Proof. For (i), Proposition 5.5 states that π \pi π has full support, that its only atom is at E ≠ 0 E \ne0 E = 0 , and that E ∣ D ∣ ± q < ∞ \mathbb{E}|D|^{\pm q} < \infty E ∣ D ∣ ± q < ∞ for q ∈ ( 0 , 1 ) q \in(0,1) q ∈ ( 0 , 1 ) ; these properties are preserved by the scaling D ↦ g D D \mapsto gD D ↦ g D .
For (ii), let Y ∼ Exp ( 1 ) Y \sim\operatorname{Exp}(1) Y ∼ Exp ( 1 ) be independent of D D D and set α = Y / p \alpha= Y/p α = Y / p . Then ( g D , g 2 α ) (gD,g^2\alpha) ( g D , g 2 α ) has law ν g ⊗ Exp ( K g ) \nu_g \otimes\operatorname{Exp}(K_g) ν g ⊗ Exp ( K g ) , and Q s ( g ) Q_s^{(g)} Q s ( g ) is defined by the random matrix M = M ( g D , g 2 α ) M=M(gD,g^2\alpha) M = M ( g D , g 2 α ) , and similarly Q s ( g ′ ) Q_s^{(g')} Q s ( g ′ ) is defined by M ~ = M ( g ′ D , g ′ 2 α ) \widetilde{M}=M(g'D,g'^2\alpha) M = M ( g ′ D , g ′2 α ) . A direct computation from (74) shows that for every d ∈ R d \in\mathbb{R} d ∈ R and a , γ > 0 a,\gamma> 0 a , γ > 0 , we have
M ( γ d , γ 2 a ) = S γ M ( d , a ) S γ − 1 , S γ = diag ( γ − 1 / 2 , γ 1 / 2 ) . (89) M(\gamma d,\gamma^2 a)=S_\gamma M(d,a)S_\gamma^{-1},\qquad S_\gamma=\operatorname{diag}(\gamma^{-1/2},\gamma^{1/2}). \tag*{(89)} M ( γ d , γ 2 a ) = S γ M ( d , a ) S γ − 1 , S γ = diag ( γ − 1/2 , γ 1/2 ) . (89)
Indeed, the ( 1 , 2 ) (1,2) ( 1 , 2 ) , ( 2 , 1 ) (2,1) ( 2 , 1 ) , and ( 2 , 2 ) (2,2) ( 2 , 2 ) entries of both sides are γ − 1 a − 1 / 2 \gamma^{-1}a^{-1/2} γ − 1 a − 1/2 , − γ a 1 / 2 -\gamma a^{1/2} − γ a 1/2 , and a − 1 / 2 d a^{-1/2}d a − 1/2 d , respectively, and the ( 1 , 1 ) (1,1) ( 1 , 1 ) entries vanish. Since S g ′ = S c S g S_{g'}=S_cS_g S g ′ = S c S g , (89) gives M ~ = S M S − 1 \widetilde{M}=SMS^{-1} M = S M S − 1 with S = S c S=S_c S = S c .
The map y ↦ S ⋅ y y \mapsto S \mathbin{\cdot} y y ↦ S ⋅ y is a homeomorphism of R P 1 \mathbb{R}\mathbb{P}^{1} R P 1 and y ↦ ∥ S v y ∥ − s y \mapsto\lVert S v_y\rVert^{-s} y ↦ ∥ S v y ∥ − s is continuous and positive. Then Θ \Theta Θ is bounded and positive, with positive inverse ( Θ − 1 G ) ( x ) = ∥ S v S − 1 ⋅ x ∥ s G ( S − 1 ⋅ x ) (\Theta^{-1}G)(x)=\lVert S v_{S^{-1}\mathbin{\cdot}x}\rVert^{s}G(S^{-1}\mathbin{\cdot}x) ( Θ − 1 G ) ( x ) = ∥ S v S − 1 ⋅ x ∥ s G ( S − 1 ⋅ x ) . Fix y ∈ R P 1 y \in\mathbb{R}\mathbb{P}^{1} y ∈ R P 1 and a realization of M M M . By Lemma 6.2 (i), we have M ~ ⋅ ( S ⋅ y ) = S ⋅ ( M ⋅ y ) \widetilde{M}\mathbin{\cdot}(S\mathbin{\cdot}y)=S\mathbin{\cdot}(M\mathbin{\cdot}y) M ⋅ ( S ⋅ y ) = S ⋅ ( M ⋅ y ) and
∥ M ~ v S ⋅ y ∥ = ∥ S M v y ∥ ∥ S − 1 v S ⋅ y ∥ = ∥ S v M ⋅ y ∥ ∥ M v y ∥ ∥ S v y ∥ − 1 , \lVert\widetilde{M}v_{S\mathbin{\cdot}y}\rVert=\lVert SMv_y\rVert\lVert S^{-1}v_{S\mathbin{\cdot}y}\rVert=\lVert S v_{M\mathbin{\cdot}y}\rVert\lVert Mv_y\rVert\lVert S v_y\rVert^{-1}, ∥ M v S ⋅ y ∥ = ∥ S M v y ∥ ∥ S − 1 v S ⋅ y ∥ = ∥ S v M ⋅ y ∥ ∥ M v y ∥ ∥ S v y ∥ − 1 ,
where in the last step we used ∥ S − 1 v S ⋅ y ∥ ∥ S v y ∥ = ∥ S − 1 S v y ∥ = 1 \lVert S^{-1}v_{S\mathbin{\cdot}y}\rVert\lVert S v_y\rVert=\lVert S^{-1}S v_y\rVert=1 ∥ S − 1 v S ⋅ y ∥ ∥ S v y ∥ = ∥ S − 1 S v y ∥ = 1 . For F ∈ C ( R P 1 ) F \in C(\mathbb{R}\mathbb{P}^{1}) F ∈ C ( R P 1 ) , we obtain
( Q s ( g ′ ) F ) ( S ⋅ y ) = ∥ S v y ∥ s E [ ∥ M v y ∥ − s ∥ S v M ⋅ y ∥ − s F ( S ⋅ ( M ⋅ y ) ) ] = ∥ S v y ∥ s ( Q s ( g ) Θ F ) ( y ) . (Q_s^{(g')}F)(S\mathbin{\cdot}y)=\lVert S v_y\rVert^s\mathbb{E}\left[\lVert Mv_y\rVert^{-s}\lVert S v_{M\mathbin{\cdot}y}\rVert^{-s}F(S\mathbin{\cdot}(M\mathbin{\cdot}y))\right]=\lVert S v_y\rVert^s(Q_s^{(g)}\Theta F)(y). ( Q s ( g ′ ) F ) ( S ⋅ y ) = ∥ S v y ∥ s E [ ∥ M v y ∥ − s ∥ S v M ⋅ y ∥ − s F ( S ⋅ ( M ⋅ y )) ] = ∥ S v y ∥ s ( Q s ( g ) Θ F ) ( y ) .
Multiplying by ∥ S v y ∥ − s \lVert S v_y\rVert^{-s} ∥ S v y ∥ − s , we find Θ Q s ( g ′ ) F = Q s ( g ) Θ F \Theta Q_s^{(g')}F=Q_s^{(g)}\Theta F Θ Q s ( g ′ ) F = Q s ( g ) Θ F , which is the claimed identity. Since similar operators have the same spectrum, this implies the last statement.
Definition 6.14. Suppose that ν \nu ν satisfies (73) , and let λ ( s ) \lambda(s) λ ( s ) be the Perron–Frobenius eigenvalue of Q s Q_s Q s from Theorem 6.11 .
(i) The free energy is φ ( s ) = log λ ( s ) \varphi(s)=\log\lambda(s) φ ( s ) = log λ ( s ) for s ∈ ( 0 , 2 ) s\in(0,2) s ∈ ( 0 , 2 ) . The name follows [4, 12], where the same quantity, up to normalization, is called the free energy.
(ii) Let M 1 , M 2 , … M_1,M_2,\ldots M 1 , M 2 , … be independent copies of M ( B , β ) M(B,\beta) M ( B , β ) . For s ∈ ( 0 , 2 ) s\in(0,2) s ∈ ( 0 , 2 ) and L ≥ 0 L\geq0 L ≥ 0 , we set f L ( s ) = E ∥ M L ⋯ M 1 ∥ − s f_L(s)=\mathbb{E}\lVert M_L\cdots M_1\rVert^{-s} f L ( s ) = E ∥ M L ⋯ M 1 ∥ − s , with f 0 ( s ) = 1 f_0(s)=1 f 0 ( s ) = 1 .
(iii) Let π \pi π be a real solution at energy E > 0 E > 0 E > 0 . The free energy of π \pi π is φ π ( s ) = log λ ( s ) \varphi_\pi(s) = \log\lambda(s) φ π ( s ) = log λ ( s ) , where λ ( s ) \lambda(s) λ ( s ) is computed with the law ν g \nu_g ν g and the rate K g = p / g 2 K_g = p/g^2 K g = p / g 2 of Lemma 6.13 for some g > 0 g > 0 g > 0 ; by Lemma 6.13(ii), it does not depend on g g g . When the real solution is determined by the energy, we also write φ E ( s ) \varphi_E(s) φ E ( s ) . We define f L ( s ) f_L(s) f L ( s ) for π \pi π as in (ii), with a fixed scaling.
Properties of the free energy
Proposition 6.15. Suppose that ν \nu ν satisfies (6.1), and let φ \varphi φ be its free energy.
(i) The function φ \varphi φ is real-analytic and convex on ( 0 , 2 ) (0,2) ( 0 , 2 ) , and φ ( s ) = φ ( 2 − s ) \varphi(s) = \varphi(2-s) φ ( s ) = φ ( 2 − s ) for s ∈ ( 0 , 2 ) s \in(0,2) s ∈ ( 0 , 2 ) .
(ii) We have φ ′ ( 1 ) = 0 \varphi'(1) = 0 φ ′ ( 1 ) = 0 , and φ \varphi φ is nonincreasing on ( 0 , 1 ] (0,1] ( 0 , 1 ] and nondecreasing on [ 1 , 2 ) [1,2) [ 1 , 2 ) .
(iii) We have 0 ≤ E log ∥ M ( B , β ) ∥ < ∞ 0 \leq\mathbb{E}\log\lVert M(B,\beta)\rVert< \infty 0 ≤ E log ∥ M ( B , β )∥ < ∞ , and for every s ∈ ( 0 , 2 ) s \in(0,2) s ∈ ( 0 , 2 ) ,
log p − s E log ∥ M ( B , β ) ∥ ≤ φ ( s ) ≤ log p . \log p - s \mathbb{E}\log\lVert M(B,\beta)\rVert\leq\varphi(s) \leq\log p. log p − s E log ∥ M ( B , β )∥ ≤ φ ( s ) ≤ log p . In particular, lim s ↓ 0 φ ( s ) = log p \lim_{s\downarrow0}\varphi(s) = \log p lim s ↓ 0 φ ( s ) = log p .
(iv) We have φ ( 1 ) < log p \varphi(1) < \log p φ ( 1 ) < log p .
(v) We have φ ′ ( s ) < 0 \varphi'(s) < 0 φ ′ ( s ) < 0 for every s ∈ ( 0 , 1 ) s \in(0,1) s ∈ ( 0 , 1 ) . In particular, φ \varphi φ is strictly decreasing on ( 0 , 1 ] (0,1] ( 0 , 1 ] , and φ ( s ) < log p \varphi(s) < \log p φ ( s ) < log p for every s ∈ ( 0 , 2 ) s \in(0,2) s ∈ ( 0 , 2 ) .
If π \pi π is a real solution at energy E > 0 E > 0 E > 0 , then φ π \varphi_\pi φ π is independent of the scaling, and (i)–(v) hold for φ π \varphi_\pi φ π .
Proof. Throughout the proof, M 1 , M 2 , … M_1, M_2, \ldots M 1 , M 2 , … are independent copies of M = M ( B , β ) M = M(B,\beta) M = M ( B , β ) , and we abbreviate W L = M L ⋯ M 1 W_L = M_L \cdots M_1 W L = M L ⋯ M 1 and j L ( s ) = E J s ( W L ) j_L(s) = \mathbb{E}J_s(W_L) j L ( s ) = E J s ( W L ) . By Theorem 6.11 (v), we have
φ ( s ) = log p + lim L → ∞ 1 L log j L ( s ) , s ∈ ( 0 , 2 ) . (90) \varphi(s) = \log p + \lim_{L\to\infty} \frac{1}{L}\log j_L(s), \qquad s \in(0,2). \tag*{(90)} φ ( s ) = log p + L → ∞ lim L 1 log j L ( s ) , s ∈ ( 0 , 2 ) . (90)
Proof of (i). Real-analyticity is Theorem 6.11 (iv). By Lemma 6.3 , we have J s ( W L ) = J 2 − s ( W L ) J_s(W_L) = J_{2-s}(W_L) J s ( W L ) = J 2 − s ( W L ) for every realization. Then j L ( s ) = j L ( 2 − s ) j_L(s) = j_L(2-s) j L ( s ) = j L ( 2 − s ) , and (6.18) implies that φ ( s ) = φ ( 2 − s ) \varphi(s) = \varphi(2-s) φ ( s ) = φ ( 2 − s ) . For convexity, fix L L L , s 0 , s 1 ∈ ( 0 , 2 ) s_0,s_1 \in(0,2) s 0 , s 1 ∈ ( 0 , 2 ) , and τ ∈ ( 0 , 1 ) \tau\in(0,1) τ ∈ ( 0 , 1 ) . Applying Hölder’s inequality with exponents 1 / ( 1 − τ ) 1/(1-\tau) 1/ ( 1 − τ ) and 1 / τ 1/\tau 1/ τ to the integral over m ⊗ P m \otimes\mathbb{P} m ⊗ P of ∥ W L v x ∥ − ( 1 − τ ) s 0 ∥ W L v x ∥ − τ s 1 \lVert W_L v_x\rVert^{-(1-\tau)s_0}\lVert W_L v_x\rVert^{-\tau s_1} ∥ W L v x ∥ − ( 1 − τ ) s 0 ∥ W L v x ∥ − τ s 1 , we obtain j L ( ( 1 − τ ) s 0 + τ s 1 ) ≤ j L ( s 0 ) 1 − τ j L ( s 1 ) τ j_L((1-\tau)s_0+\tau s_1) \leq j_L(s_0)^{1-\tau}j_L(s_1)^\tau j L (( 1 − τ ) s 0 + τ s 1 ) ≤ j L ( s 0 ) 1 − τ j L ( s 1 ) τ . Then s ↦ L − 1 log j L ( s ) s \mapsto L^{-1}\log j_L(s) s ↦ L − 1 log j L ( s ) is convex, and so is its pointwise limit φ ( s ) − log p \varphi(s)-\log p φ ( s ) − log p .
Proof of (ii). Differentiating φ ( s ) = φ ( 2 − s ) \varphi(s) = \varphi(2-s) φ ( s ) = φ ( 2 − s ) gives φ ′ ( s ) = − φ ′ ( 2 − s ) \varphi'(s) = -\varphi'(2-s) φ ′ ( s ) = − φ ′ ( 2 − s ) , and taking s = 1 s=1 s = 1 gives φ ′ ( 1 ) = 0 \varphi'(1)=0 φ ′ ( 1 ) = 0 . Since φ \varphi φ is convex and differentiable, φ ′ \varphi' φ ′ is nondecreasing. Then φ ′ ≤ 0 \varphi' \leq0 φ ′ ≤ 0 on ( 0 , 1 ] (0,1] ( 0 , 1 ] and φ ′ ≥ 0 \varphi' \geq0 φ ′ ≥ 0 on [ 1 , 2 ) [1,2) [ 1 , 2 ) .
Proof of (iii). Since det M = 1 \det M = 1 det M = 1 , the product of the two singular values of M M M is one; in particular, ∥ M ∥ ≥ 1 \lVert M\rVert\geq1 ∥ M ∥ ≥ 1 and log ∥ M ∥ ≥ 0 \log\lVert M\rVert\geq0 log ∥ M ∥ ≥ 0 . Bounding the operator norm by the Hilbert–Schmidt norm, we have ∥ M ( b , β ) ∥ 2 ≤ β − 1 ( 1 + β 2 + b 2 ) \lVert M(b,\beta)\rVert^2 \leq\beta^{-1}(1+\beta^2+b^2) ∥ M ( b , β ) ∥ 2 ≤ β − 1 ( 1 + β 2 + b 2 ) , and therefore
log ∥ M ( b , β ) ∥ ≤ 1 2 ∣ log β ∣ + 1 2 log ( 1 + β 2 ) + 1 2 log ( 1 + b 2 ) . \log\lVert M(b,\beta)\rVert\leq\frac{1}{2}\lvert\log\beta\rvert+ \frac{1}{2}\log(1+\beta^2) + \frac{1}{2}\log(1+b^2). log ∥ M ( b , β )∥ ≤ 2 1 ∣ log β ∣ + 2 1 log ( 1 + β 2 ) + 2 1 log ( 1 + b 2 ) .
The first two terms have finite expectation since β \beta β is exponential. For the third, fix q ∈ ( 0 , 1 ) q \in(0,1) q ∈ ( 0 , 1 ) ; since log y ≤ ( 2 / q ) y q / 2 \log y \le(2/q)y^{q/2} log y ≤ ( 2/ q ) y q /2 for y ≥ 1 y \ge1 y ≥ 1 , we have log ( 1 + B 2 ) ≤ ( 2 / q ) ( 1 + B 2 ) q / 2 ≤ ( 2 / q ) ( 1 + ∣ B ∣ q ) \log(1+B^2) \le(2/q)(1+B^2)^{q/2} \le(2/q)(1+|B|^q) log ( 1 + B 2 ) ≤ ( 2/ q ) ( 1 + B 2 ) q /2 ≤ ( 2/ q ) ( 1 + ∣ B ∣ q ) , which is integrable by (6.1). This proves E log ∥ M ∥ < ∞ \mathbb{E}\log\lVert M\rVert< \infty E log ∥ M ∥ < ∞ .
By Lemma 6.3, we have J s ( W L ) ≤ 1 J_s(W_L) \le1 J s ( W L ) ≤ 1 . Then j L ( s ) ≤ 1 j_L(s) \le1 j L ( s ) ≤ 1 , which implies φ ( s ) ≤ log p \varphi(s) \le\log p φ ( s ) ≤ log p by (6.18). For the lower bound, by Jensen’s inequality for the probability measure m m m , we have
J s ( A ) = ∫ exp ( − s log ∥ A v x ∥ ) m ( d x ) ≥ exp ( − s ∫ log ∥ A v x ∥ m ( d x ) ) ≥ ∥ A ∥ − s . J_s(A)=\int\exp\left(-s\log\lVert Av_x\rVert\right)m(\mathrm{d}x)\ge\exp\left(-s\int\log\lVert Av_x\rVert m(\mathrm{d}x)\right)\ge\lVert A\rVert^{-s}. J s ( A ) = ∫ exp ( − s log ∥ A v x ∥ ) m ( d x ) ≥ exp ( − s ∫ log ∥ A v x ∥ m ( d x ) ) ≥ ∥ A ∥ − s .
for every A A A with determinant one. Using Jensen’s inequality again, and then log ∥ W L ∥ ≤ ∑ j = 1 L log ∥ M j ∥ \log\lVert W_L\rVert\le\sum_{j=1}^{L}\log\lVert M_j\rVert log ∥ W L ∥ ≤ ∑ j = 1 L log ∥ M j ∥ , we obtain
j L ( s ) ≥ E exp ( − s log ∥ W L ∥ ) ≥ exp ( − s E log ∥ W L ∥ ) ≥ exp ( − s L E log ∥ M ∥ ) . j_L(s)\ge\mathbb{E}\exp\left(-s\log\lVert W_L\rVert\right)\ge\exp\left(-s\mathbb{E}\log\lVert W_L\rVert\right)\ge\exp\left(-sL\mathbb{E}\log\lVert M\rVert\right). j L ( s ) ≥ E exp ( − s log ∥ W L ∥ ) ≥ exp ( − s E log ∥ W L ∥ ) ≥ exp ( − s L E log ∥ M ∥ ) .
Inserting this into (6.18) gives the lower bound. Letting s ↓ 0 s\downarrow0 s ↓ 0 gives lim s ↓ 0 φ ( s ) = log p \lim_{s\downarrow0}\varphi(s)=\log p lim s ↓ 0 φ ( s ) = log p .
Proof of (iv). We work in the real Hilbert space L 2 ( m ) L^2(m) L 2 ( m ) of real-valued functions, with inner product ⟨ ⋅ , ⋅ ⟩ \langle\cdot,\cdot\rangle ⟨ ⋅ , ⋅ ⟩ and norm ∥ ⋅ ∥ 2 \lVert\cdot\rVert_2 ∥ ⋅ ∥ 2 . For a real 2 × 2 2\times2 2 × 2 matrix A A A with determinant one, define ( U A F ) ( x ) = ∥ A v x ∥ − 1 F ( A ⋅ x ) (U_AF)(x)=\lVert Av_x\rVert^{-1}F(A\mathbin{\cdot}x) ( U A F ) ( x ) = ∥ A v x ∥ − 1 F ( A ⋅ x ) . By (6.5) applied to ∣ F ∣ 2 \lvert F\rvert^2 ∣ F ∣ 2 , U A U_A U A is an isometry of L 2 ( m ) L^2(m) L 2 ( m ) . By (6.4), we have U A U A ′ = U A ′ A U_AU_{A'}=U_{A'A} U A U A ′ = U A ′ A ; indeed,
( U A U A ′ F ) ( x ) = ∥ A v x ∥ − 1 ∥ A ′ v A ⋅ x ∥ − 1 F ( A ′ ⋅ ( A ⋅ x ) ) = ∥ A ′ A v x ∥ − 1 F ( ( A ′ A ) ⋅ x ) . (U_AU_{A'}F)(x)=\lVert Av_x\rVert^{-1}\lVert A'v_{A\cdot x}\rVert^{-1}F(A'\mathbin{\cdot}(A\mathbin{\cdot}x))=\lVert A'Av_x\rVert^{-1}F((A'A)\mathbin{\cdot}x). ( U A U A ′ F ) ( x ) = ∥ A v x ∥ − 1 ∥ A ′ v A ⋅ x ∥ − 1 F ( A ′ ⋅ ( A ⋅ x )) = ∥ A ′ A v x ∥ − 1 F (( A ′ A ) ⋅ x ) .
Taking A ′ = A − 1 A'=A^{-1} A ′ = A − 1 and then A = A ′ − 1 A=A'^{-1} A = A ′ − 1 , we have U A − 1 U A = U A U A − 1 = U I = I U_{A^{-1}}U_A=U_AU_{A^{-1}}=U_I=I U A − 1 U A = U A U A − 1 = U I = I , and U A U_A U A is unitary with U A − 1 = U A − 1 U_A^{-1}=U_{A^{-1}} U A − 1 = U A − 1 .
Let ψ = ψ 1 \psi=\psi_1 ψ = ψ 1 be the eigenfunction from Theorem 6.11(ii) with s = 1 s=1 s = 1 ; it is continuous, real-valued, and bounded below by a positive constant. By definition, Q 1 ψ ( x ) = p E [ ( U M ψ ) ( x ) ] Q_1\psi(x)=p\mathbb{E}[(U_M\psi)(x)] Q 1 ψ ( x ) = p E [( U M ψ ) ( x )] . By Tonelli’s theorem (all integrands are nonnegative), the Cauchy–Schwarz inequality, and unitarity, we have
λ ( 1 ) ∥ ψ ∥ 2 2 = ⟨ ψ , Q 1 ψ ⟩ = p E ⟨ ψ , U M ψ ⟩ ≤ p E [ ∥ ψ ∥ 2 ∥ U M ψ ∥ 2 ] = p ∥ ψ ∥ 2 2 . \lambda(1)\lVert\psi\rVert_2^2=\langle\psi,Q_1\psi\rangle=p\mathbb{E}\langle\psi,U_M\psi\rangle\le p\mathbb{E}\left[\lVert\psi\rVert_2\lVert U_M\psi\rVert_2\right]=p\lVert\psi\rVert_2^2. λ ( 1 ) ∥ ψ ∥ 2 2 = ⟨ ψ , Q 1 ψ ⟩ = p E ⟨ ψ , U M ψ ⟩ ≤ p E [ ∥ ψ ∥ 2 ∥ U M ψ ∥ 2 ] = p ∥ ψ ∥ 2 2 .
We have shown that λ ( 1 ) ≤ p \lambda(1)\le p λ ( 1 ) ≤ p . Suppose, for the sake of contradiction, that λ ( 1 ) = p \lambda(1)=p λ ( 1 ) = p . Then the display implies E [ ∥ ψ ∥ 2 2 − ⟨ ψ , U M ψ ⟩ ] = 0 \mathbb{E}[\lVert\psi\rVert_2^2-\langle\psi,U_M\psi\rangle]=0 E [∥ ψ ∥ 2 2 − ⟨ ψ , U M ψ ⟩] = 0 . Since ∥ U M ψ ∥ 2 = ∥ ψ ∥ 2 \lVert U_M\psi\rVert_2=\lVert\psi\rVert_2 ∥ U M ψ ∥ 2 = ∥ ψ ∥ 2 , we have ∥ ψ ∥ 2 2 − ⟨ ψ , U M ψ ⟩ = ∥ U M ψ − ψ ∥ 2 2 / 2 ≥ 0 \lVert\psi\rVert_2^2-\langle\psi,U_M\psi\rangle=\lVert U_M\psi-\psi\rVert_2^2/2\ge0 ∥ ψ ∥ 2 2 − ⟨ ψ , U M ψ ⟩ = ∥ U M ψ − ψ ∥ 2 2 /2 ≥ 0 , and therefore E ∥ U M ψ − ψ ∥ 2 2 = 0 \mathbb{E}\lVert U_M\psi-\psi\rVert_2^2=0 E ∥ U M ψ − ψ ∥ 2 2 = 0 . The function ( b , β ) ↦ ∥ U M ( b , β ) ψ − ψ ∥ 2 2 (b,\beta)\mapsto\lVert U_{M(b,\beta)}\psi-\psi\rVert_2^2 ( b , β ) ↦ ∥ U M ( b , β ) ψ − ψ ∥ 2 2 is continuous on R × ( 0 , ∞ ) \mathbb{R}\times(0,\infty) R × ( 0 , ∞ ) , by dominated convergence: the integrand is continuous in ( b , β ) (b,\beta) ( b , β ) for each x x x , and ∥ M ( b , β ) v x ∥ − 1 ≤ ∥ M ( b , β ) − 1 ∥ \lVert M(b,\beta)v_x\rVert^{-1}\le\lVert M(b,\beta)^{-1}\rVert ∥ M ( b , β ) v x ∥ − 1 ≤ ∥ M ( b , β ) − 1 ∥ is bounded uniformly in x x x and in ( b , β ) (b,\beta) ( b , β ) in compact sets. This function vanishes ν ⊗ Exp ( K ) \nu\otimes\operatorname{Exp}(K) ν ⊗ Exp ( K ) -almost everywhere, and this measure charges every nonempty open subset of R × ( 0 , ∞ ) \mathbb{R}\times(0,\infty) R × ( 0 , ∞ ) because supp ν = R \operatorname{supp}\nu=\mathbb{R} supp ν = R . Since the function is continuous, it vanishes identically. In particular, U M ( 0 , 1 ) ψ = ψ U_{M(0,1)}\psi=\psi U M ( 0 , 1 ) ψ = ψ and U M ( 0 , 4 ) ψ = ψ U_{M(0,4)}\psi=\psi U M ( 0 , 4 ) ψ = ψ . A direct computation from (6.2) gives
M ( 0 , 4 ) − 1 M ( 0 , 1 ) = ( 0 − 1 / 2 2 0 ) ( 0 1 − 1 0 ) = ( 1 / 2 0 0 2 ) = Δ . M(0,4)^{-1}M(0,1)=\begin{pmatrix}0 & -1/2 \\ 2 & 0\end{pmatrix}\begin{pmatrix}0 & 1 \\ -1 & 0\end{pmatrix}=\begin{pmatrix}1/2 & 0 \\ 0 & 2\end{pmatrix}=\Delta. M ( 0 , 4 ) − 1 M ( 0 , 1 ) = ( 0 2 − 1/2 0 ) ( 0 − 1 1 0 ) = ( 1/2 0 0 2 ) = Δ.
Then U Δ = U M ( 0 , 1 ) U M ( 0 , 4 ) − 1 = U M ( 0 , 1 ) U M ( 0 , 4 ) − 1 U_{\Delta}=U_{M(0,1)}U_{M(0,4)^{-1}}=U_{M(0,1)}U_{M(0,4)}^{-1} U Δ = U M ( 0 , 1 ) U M ( 0 , 4 ) − 1 = U M ( 0 , 1 ) U M ( 0 , 4 ) − 1 and U Δ ψ = ψ U_{\Delta}\psi=\psi U Δ ψ = ψ in L 2 ( m ) L^{2}(m) L 2 ( m ) . Since both sides are continuous and m m m has full support, we have ψ ( x ) = ∥ Δ v x ∥ − 1 ψ ( Δ ⋅ x ) \psi(x)=\lVert\Delta v_x\rVert^{-1}\psi(\Delta\mathbin{\cdot}x) ψ ( x ) = ∥ Δ v x ∥ − 1 ψ ( Δ ⋅ x ) for every x x x . Let μ \mu μ be the finite measure ψ 2 m \psi^{2}m ψ 2 m . For every Borel F ≥ 0 F\ge0 F ≥ 0 , (6.5) with A = Δ A=\Delta A = Δ gives
∫ F ( Δ ⋅ x ) μ ( d x ) = ∫ F ( Δ ⋅ x ) ψ ( Δ ⋅ x ) 2 ∥ Δ v x ∥ − 2 m ( d x ) = ∫ F ( y ) μ ( d y ) . \int F(\Delta\mathbin{\cdot}x)\,\mu(\mathrm{d}x)=\int F(\Delta\mathbin{\cdot}x)\psi(\Delta\mathbin{\cdot}x)^{2}\lVert\Delta v_x\rVert^{-2}\,m(\mathrm{d}x)=\int F(y)\,\mu(\mathrm{d}y). ∫ F ( Δ ⋅ x ) μ ( d x ) = ∫ F ( Δ ⋅ x ) ψ ( Δ ⋅ x ) 2 ∥ Δ v x ∥ − 2 m ( d x ) = ∫ F ( y ) μ ( d y ) .
so μ \mu μ is invariant under the map x ↦ Δ ⋅ x x\mapsto\Delta\mathbin{\cdot}x x ↦ Δ ⋅ x , which is x ↦ x / 4 x\mapsto x/4 x ↦ x /4 in coordinates. For k ∈ Z k\in\mathbb{Z} k ∈ Z , let I k = [ 4 k , 4 k + 1 ) I_k=[4^k,4^{k+1}) I k = [ 4 k , 4 k + 1 ) . The preimage of I k I_k I k under x ↦ x / 4 x\mapsto x/4 x ↦ x /4 is I k + 1 I_{k+1} I k + 1 , and invariance gives μ ( I k + 1 ) = μ ( I k ) \mu(I_{k+1})=\mu(I_k) μ ( I k + 1 ) = μ ( I k ) for every k k k . The intervals I k I_k I k are disjoint with union ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , and μ ( ( 0 , ∞ ) ) ≤ μ ( R P 1 ) < ∞ \mu((0,\infty))\le\mu(\mathbb{R}\mathbb{P}^{1})<\infty μ (( 0 , ∞ )) ≤ μ ( R P 1 ) < ∞ , which forces μ ( I k ) = 0 \mu(I_k)=0 μ ( I k ) = 0 for every k k k and μ ( ( 0 , ∞ ) ) = 0 \mu((0,\infty))=0 μ (( 0 , ∞ )) = 0 . This contradicts μ ( ( 0 , ∞ ) ) = ∫ ( 0 , ∞ ) ψ 2 d m > 0 \mu((0,\infty))=\int_{(0,\infty)}\psi^{2}\,\mathrm{d}m>0 μ (( 0 , ∞ )) = ∫ ( 0 , ∞ ) ψ 2 d m > 0 , which holds because ψ > 0 \psi>0 ψ > 0 and m ( ( 0 , ∞ ) ) > 0 m((0,\infty))>0 m (( 0 , ∞ )) > 0 . We conclude that λ ( 1 ) < p \lambda(1)<p λ ( 1 ) < p , that is, φ ( 1 ) < log p \varphi(1)<\log p φ ( 1 ) < log p .
Proof of (v). By (ii), φ ′ ≤ 0 \varphi'\le0 φ ′ ≤ 0 on ( 0 , 1 ) (0,1) ( 0 , 1 ) . Suppose that φ ′ ( s 0 ) = 0 \varphi'(s_0)=0 φ ′ ( s 0 ) = 0 for some s 0 ∈ ( 0 , 1 ) s_0\in(0,1) s 0 ∈ ( 0 , 1 ) . Since φ ′ \varphi' φ ′ is nondecreasing and φ ′ ( 1 ) = 0 \varphi'(1)=0 φ ′ ( 1 ) = 0 , we have φ ′ = 0 \varphi'=0 φ ′ = 0 on [ s 0 , 1 ] [s_0,1] [ s 0 , 1 ] , and φ \varphi φ is constant on [ s 0 , 1 ] [s_0,1] [ s 0 , 1 ] . Since φ \varphi φ is real-analytic on the interval ( 0 , 2 ) (0,2) ( 0 , 2 ) , it is constant on ( 0 , 2 ) (0,2) ( 0 , 2 ) . Then (iii) implies φ ( 1 ) = lim s ↓ 0 φ ( s ) = log p \varphi(1)=\lim_{s\downarrow0}\varphi(s)=\log p φ ( 1 ) = lim s ↓ 0 φ ( s ) = log p , contradicting (iv). We conclude that φ ′ < 0 \varphi'<0 φ ′ < 0 on ( 0 , 1 ) (0,1) ( 0 , 1 ) , and φ \varphi φ is strictly decreasing on ( 0 , 1 ] (0,1] ( 0 , 1 ] . For s ∈ ( 0 , 1 ] s\in(0,1] s ∈ ( 0 , 1 ] , this implies φ ( s ) < φ ( s / 2 ) ≤ log p \varphi(s)<\varphi(s/2)\le\log p φ ( s ) < φ ( s /2 ) ≤ log p , and for s ∈ [ 1 , 2 ) s\in[1,2) s ∈ [ 1 , 2 ) we use φ ( s ) = φ ( 2 − s ) \varphi(s)=\varphi(2-s) φ ( s ) = φ ( 2 − s ) .
Finally, if π \pi π is a real solution, then ν g \nu_g ν g satisfies (6.1) for every g > 0 g>0 g > 0 by Lemma 6.13(i), and φ π \varphi_\pi φ π does not depend on g g g by Lemma 6.13(ii). Applying (i)–(v) to ν g \nu_g ν g , we obtain the last statement.
The symmetry φ ( s ) = φ ( 2 − s ) \varphi(s)=\varphi(2-s) φ ( s ) = φ ( 2 − s ) is inherited from the identity J s ( A ) = J 2 − s ( A ) J_s(A)=J_{2-s}(A) J s ( A ) = J 2 − s ( A ) of Lemma 6.3. A symmetry of this kind for the linearized cavity equation goes back to Abou-Chacra, Anderson, and Thouless [2]; see [109] for a discussion. The proof of Proposition 6.15(iv) uses only (6.1), and in particular it does not use atoms of ν \nu ν . The next lemma selects an exponent below one at which the free energy has a useful tangent line.
Lemma 6.16. Suppose that ν \nu ν satisfies (6.1) and that φ ( 1 ) > 0 \varphi(1)>0 φ ( 1 ) > 0 . Then there exists s ∈ ( 1 / 2 , 1 ) s\in(1/2,1) s ∈ ( 1/2 , 1 ) such that
γ = − φ ′ ( s ) > 0 , F = φ ( s ) − ( 1 − s ) γ > 0 , φ ( s ) < log p . \gamma=-\varphi'(s)>0,\qquad F=\varphi(s)-(1-s)\gamma>0,\qquad\varphi(s)<\log p. γ = − φ ′ ( s ) > 0 , F = φ ( s ) − ( 1 − s ) γ > 0 , φ ( s ) < log p . In particular, this holds for the free energy φ π \varphi_\pi φ π of a real solution π \pi π with 0 < φ π ( 1 ) 0<\varphi_\pi(1) 0 < φ π ( 1 ) .
Proof. By Proposition 6.15(v), we have φ ′ ( s ) < 0 \varphi'(s)<0 φ ′ ( s ) < 0 and φ ( s ) < log p \varphi(s)<\log p φ ( s ) < log p for every s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) . Since φ \varphi φ is continuously differentiable on ( 0 , 2 ) (0,2) ( 0 , 2 ) and φ ′ ( 1 ) = 0 \varphi'(1)=0 φ ′ ( 1 ) = 0 , we have
lim s ↑ 1 ( φ ( s ) − ( 1 − s ) ( − φ ′ ( s ) ) ) = lim s ↑ 1 ( φ ( s ) + ( 1 − s ) φ ′ ( s ) ) = φ ( 1 ) > 0. \lim_{s\uparrow1}\left(\varphi(s)-(1-s)(-\varphi'(s))\right)=\lim_{s\uparrow1}\left(\varphi(s)+(1-s)\varphi'(s)\right)=\varphi(1)>0. s ↑ 1 lim ( φ ( s ) − ( 1 − s ) ( − φ ′ ( s )) ) = s ↑ 1 lim ( φ ( s ) + ( 1 − s ) φ ′ ( s ) ) = φ ( 1 ) > 0.
Choosing s ∈ ( 1 / 2 , 1 ) s\in(1/2,1) s ∈ ( 1/2 , 1 ) sufficiently close to one gives all three inequalities. □
Products of three independent factors
In this section, we show that the product of three independent transfer matrices has bounded inverse fractional moments of its bilinear forms, uniformly in the vectors. Here and below, a factor is a random matrix M ( B , β ) M(B,\beta) M ( B , β ) in which β \beta β has an exponential law and is independent of B B B , as in Section 6.2. The smoothing comes from the exponential variables β i \beta_i β i alone; the variables B i B_i B i enter only through their moments. We then use this bound to compare the free energy with moments of norms of products.
Inverse moments of multiaffine polynomials
A multiaffine polynomial in n n n variables is a polynomial of the form P ( t ) = ∑ α ⊂ { 1 , … , n } c α ∏ i ∈ α t i \mathfrak{P}(t)=\sum_{\alpha\subset\{1,\ldots,n\}}c_\alpha\prod_{i\in\alpha}t_i P ( t ) = ∑ α ⊂ { 1 , … , n } c α ∏ i ∈ α t i with coefficients c α ∈ C c_\alpha\in\mathbb{C} c α ∈ C , that is, a polynomial of degree at most one in each variable. For s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) and K > 0 K>0 K > 0 , let ω s , K \omega_{s,K} ω s , K be the measure on R \mathbb{R} R given by ω s , K ( d t ) = K t s / 2 e − K t 1 { t > 0 } d t \omega_{s,K}(\mathrm{d}t)=Kt^{s/2}e^{-Kt}\mathbf{1}_{\{t>0\}}\mathrm{d}t ω s , K ( d t ) = K t s /2 e − K t 1 { t > 0 } d t . Its total mass is ω s , K ( R ) = K − s / 2 Γ ( 1 + s / 2 ) \omega_{s,K}(\mathbb{R})=K^{-s/2}\Gamma(1+s/2) ω s , K ( R ) = K − s /2 Γ ( 1 + s /2 ) .
Lemma 6.17. Let s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) , K > 0 K>0 K > 0 , and n ≥ 1 n\ge1 n ≥ 1 . There exists a constant C s , K , n < ∞ C_{s,K,n}<\infty C s , K , n < ∞ such that for every nonzero multiaffine polynomial P \mathfrak{P} P in n n n variables with coefficients c α c_\alpha c α ,
∫ ( 0 , ∞ ) n ∣ P ( t ) ∣ − s ω s , K ( d t 1 ) ⋯ ω s , K ( d t n ) ≤ C s , K , n ( max α ∣ c α ∣ ) − s . \int_{(0,\infty)^n}|\mathfrak{P}(t)|^{-s}\omega_{s,K}(\mathrm{d}t_1)\cdots\omega_{s,K}(\mathrm{d}t_n)\le C_{s,K,n}\left(\max_\alpha|c_\alpha|\right)^{-s}. ∫ ( 0 , ∞ ) n ∣ P ( t ) ∣ − s ω s , K ( d t 1 ) ⋯ ω s , K ( d t n ) ≤ C s , K , n ( α max ∣ c α ∣ ) − s .
One can take C s , K , n = C s , K , 1 n C_{s,K,n}=C_{s,K,1}^{n} C s , K , n = C s , K , 1 n , and C s , K , 1 C_{s,K,1} C s , K , 1 is bounded for ( s , K ) (s,K) ( s , K ) in compact subsets of ( 0 , 1 ) × ( 0 , ∞ ) (0,1)\times(0,\infty) ( 0 , 1 ) × ( 0 , ∞ ) .
Proof. We argue by induction on n n n .
Step 1. A translated singularity. Let
A s , K = sup x 0 ∈ R ∫ 0 ∞ ∣ t − x 0 ∣ − s ω s , K ( d t ) . A_{s,K}=\sup_{x_0\in\mathbb{R}}\int_0^\infty|t-x_0|^{-s}\omega_{s,K}(\mathrm{d}t). A s , K = x 0 ∈ R sup ∫ 0 ∞ ∣ t − x 0 ∣ − s ω s , K ( d t ) .
Let w ∗ = sup t > 0 K t s / 2 e − K t < ∞ w_*=\sup_{t>0}Kt^{s/2}e^{-Kt}<\infty w ∗ = sup t > 0 K t s /2 e − K t < ∞ . On { ∣ t − x 0 ∣ < 1 } \{|t-x_0|<1\} { ∣ t − x 0 ∣ < 1 } , we bound the density of ω s , K \omega_{s,K} ω s , K by w ∗ w_* w ∗ , and ∫ ∣ t − x 0 ∣ < 1 ∣ t − x 0 ∣ − s d t = 2 / ( 1 − s ) \int_{|t-x_0|<1}|t-x_0|^{-s}\mathrm{d}t=2/(1-s) ∫ ∣ t − x 0 ∣ < 1 ∣ t − x 0 ∣ − s d t = 2/ ( 1 − s ) . On { ∣ t − x 0 ∣ ≥ 1 } \{|t-x_0|\ge1\} { ∣ t − x 0 ∣ ≥ 1 } , we bound ∣ t − x 0 ∣ − s |t-x_0|^{-s} ∣ t − x 0 ∣ − s by one. This gives A s , K ≤ 2 w ∗ / ( 1 − s ) + K − s / 2 Γ ( 1 + s / 2 ) < ∞ A_{s,K}\le2w_* /(1-s)+K^{-s/2}\Gamma(1+s/2)<\infty A s , K ≤ 2 w ∗ / ( 1 − s ) + K − s /2 Γ ( 1 + s /2 ) < ∞ .
Step 2. The case n = 1 n=1 n = 1 . Let P ( t ) = a t + b \mathfrak{P}(t)=at+b P ( t ) = a t + b with ( a , b ) ∈ C 2 ∖ { 0 } (a,b)\in\mathbb{C}^2\setminus\{0\} ( a , b ) ∈ C 2 ∖ { 0 } . Replacing P \mathfrak{P} P by P / max ( ∣ a ∣ , ∣ b ∣ ) \mathfrak{P}/\max(|a|,|b|) P / max ( ∣ a ∣ , ∣ b ∣ ) multiplies both sides of the claimed inequality by max ( ∣ a ∣ , ∣ b ∣ ) s \max(|a|,|b|)^s max ( ∣ a ∣ , ∣ b ∣ ) s , so we may assume that max ( ∣ a ∣ , ∣ b ∣ ) = 1 \max(|a|,|b|)=1 max ( ∣ a ∣ , ∣ b ∣ ) = 1 . Suppose first that ∣ a ∣ ≥ 1 / 2 |a| \ge1/2 ∣ a ∣ ≥ 1/2 . Then ∣ a t + b ∣ = ∣ a ∣ ∣ t − z ∣ |at+b|=|a||t-z| ∣ a t + b ∣ = ∣ a ∣∣ t − z ∣ with z = − b / a z=-b/a z = − b / a , and ∣ t − z ∣ ≥ ∣ t − Re z ∣ |t-z| \ge|t-\operatorname{Re} z| ∣ t − z ∣ ≥ ∣ t − Re z ∣ for real t t t , and integrating gives ∫ ∣ P ∣ − s d ω s , K ≤ ∣ a ∣ − s A s , K ≤ 2 s A s , K \int|\mathcal{P}|^{-s}\,\mathrm{d}\omega_{s,K} \le|a|^{-s}A_{s,K} \le2^{s}A_{s,K} ∫ ∣ P ∣ − s d ω s , K ≤ ∣ a ∣ − s A s , K ≤ 2 s A s , K .
Suppose next that ∣ a ∣ < 1 / 2 |a|<1/2 ∣ a ∣ < 1/2 , so that ∣ b ∣ = 1 |b|=1 ∣ b ∣ = 1 . Let R = 1 / ( 2 ∣ a ∣ ) R=1/(2|a|) R = 1/ ( 2∣ a ∣ ) , with R = ∞ R=\infty R = ∞ if a = 0 a=0 a = 0 . For 0 < t ≤ R 0<t\le R 0 < t ≤ R , we have ∣ a t + b ∣ ≥ 1 − ∣ a ∣ t ≥ 1 / 2 |at+b|\ge1-|a|t\ge1/2 ∣ a t + b ∣ ≥ 1 − ∣ a ∣ t ≥ 1/2 , and this range contributes at most 2 s ω s , K ( R ) 2^{s}\omega_{s,K}(R) 2 s ω s , K ( R ) . If a ≠ 0 a\ne0 a = 0 , then for t > R t>R t > R , we write ∣ a t + b ∣ − s = ∣ a ∣ − s ∣ t − z ∣ − s ≤ ∣ a ∣ − s ∣ t − Re z ∣ − s |at+b|^{-s}=|a|^{-s}|t-z|^{-s}\le|a|^{-s}|t-\operatorname{Re}z|^{-s} ∣ a t + b ∣ − s = ∣ a ∣ − s ∣ t − z ∣ − s ≤ ∣ a ∣ − s ∣ t − Re z ∣ − s and e − K t ≤ e − K R / 2 e − K t / 2 e^{-Kt}\le e^{-KR/2}e^{-Kt/2} e − K t ≤ e − K R /2 e − K t /2 . Since K t s / 2 e − K t / 2 d t = 2 ω s , K / 2 ( d t ) Kt^{s/2}e^{-Kt/2}\,\mathrm{d}t=2\omega_{s,K/2}(\mathrm{d}t) K t s /2 e − K t /2 d t = 2 ω s , K /2 ( d t ) , the range t > R t>R t > R contributes at most
∣ a ∣ − s e − K R / 2 ⋅ 2 A s , K / 2 = 2 A s , K / 2 ∣ a ∣ − s e − K / ( 4 ∣ a ∣ ) ≤ 2 A s , K / 2 ( 4 s e K ) s . |a|^{-s}e^{-KR/2}\cdot2A_{s,K/2}=2A_{s,K/2}|a|^{-s}e^{-K/(4|a|)}\le2A_{s,K/2}\left(\frac{4s}{eK}\right)^{s}. ∣ a ∣ − s e − K R /2 ⋅ 2 A s , K /2 = 2 A s , K /2 ∣ a ∣ − s e − K / ( 4∣ a ∣ ) ≤ 2 A s , K /2 ( eK 4 s ) s .
where we used sup y > 0 y − s e − K / ( 4 y ) = ( 4 s / ( e K ) ) s \sup_{y>0}y^{-s}e^{-K/(4y)}=(4s/(eK))^{s} sup y > 0 y − s e − K / ( 4 y ) = ( 4 s / ( eK ) ) s . We conclude that the lemma holds for n = 1 n=1 n = 1 with
C s , K , 1 = max { 2 s A s , K , 2 s K − s / 2 Γ ( 1 + s / 2 ) + 2 A s , K / 2 ( 4 s e K ) s } . C_{s,K,1}=\max\left\{2^{s}A_{s,K},\,2^{s}K^{-s/2}\Gamma(1+s/2)+2A_{s,K/2}\left(\frac{4s}{eK}\right)^{s}\right\}. C s , K , 1 = max { 2 s A s , K , 2 s K − s /2 Γ ( 1 + s /2 ) + 2 A s , K /2 ( eK 4 s ) s } .
The bounds on A s , K A_{s,K} A s , K in Step 1 are bounded for ( s , K ) (s,K) ( s , K ) in compact subsets of ( 0 , 1 ) × ( 0 , ∞ ) (0,1)\times(0,\infty) ( 0 , 1 ) × ( 0 , ∞ ) , and hence so is C s , K , 1 C_{s,K,1} C s , K , 1 .
Step 3. The induction step. Let n ≥ 2 n\ge2 n ≥ 2 and suppose that the lemma holds for n − 1 n-1 n − 1 variables with the constant C s , K , 1 n − 1 C_{s,K,1}^{n-1} C s , K , 1 n − 1 . We write t = ( t ′ , t n ) t=(t',t_n) t = ( t ′ , t n ) with t ′ ∈ R n − 1 t'\in\mathbb{R}^{n-1} t ′ ∈ R n − 1 . Every multiaffine polynomial P \mathcal{P} P in n n n variables can be written uniquely as P ( t ) = t n P 1 ( t ′ ) + P 0 ( t ′ ) \mathcal{P}(t)=t_n\mathcal{P}_1(t')+\mathcal{P}_0(t') P ( t ) = t n P 1 ( t ′ ) + P 0 ( t ′ ) , where P 1 \mathcal{P}_1 P 1 and P 0 \mathcal{P}_0 P 0 are multiaffine in n − 1 n-1 n − 1 variables; the coefficients of P 1 \mathcal{P}_1 P 1 are the c α c_\alpha c α with n ∈ α n\in\alpha n ∈ α , and those of P 0 \mathcal{P}_0 P 0 are the c α c_\alpha c α with n ∉ α n\notin\alpha n ∈ / α . Let Q ∈ { P 1 , P 0 } \mathcal{Q}\in\{\mathcal{P}_1,\mathcal{P}_0\} Q ∈ { P 1 , P 0 } be a polynomial that has a coefficient of maximal modulus among all coefficients of P \mathcal{P} P , so that Q ≠ 0 \mathcal{Q}\ne0 Q = 0 and max α ∣ c α ( Q ) ∣ = max α ∣ c α ( P ) ∣ \max_\alpha|c_\alpha(\mathcal{Q})|=\max_\alpha|c_\alpha(\mathcal{P})| max α ∣ c α ( Q ) ∣ = max α ∣ c α ( P ) ∣ . The zero set of the nonzero polynomial Q \mathcal{Q} Q in R n − 1 \mathbb{R}^{n-1} R n − 1 has Lebesgue measure zero (its real or imaginary part is a nonzero real polynomial); in particular, it is null for ω s , K ⊗ ( n − 1 ) \omega_{s,K}^{\otimes(n-1)} ω s , K ⊗ ( n − 1 ) . For t ′ t' t ′ with Q ( t ′ ) ≠ 0 \mathcal{Q}(t')\ne0 Q ( t ′ ) = 0 , we have ( P 1 ( t ′ ) , P 0 ( t ′ ) ) ≠ ( 0 , 0 ) (\mathcal{P}_1(t'),\mathcal{P}_0(t'))\ne(0,0) ( P 1 ( t ′ ) , P 0 ( t ′ )) = ( 0 , 0 ) , and the case n = 1 n=1 n = 1 applied to t n ↦ t n P 1 ( t ′ ) + P 0 ( t ′ ) t_n\mapsto t_n\mathcal{P}_1(t')+\mathcal{P}_0(t') t n ↦ t n P 1 ( t ′ ) + P 0 ( t ′ ) gives
∫ ∣ P ( t ′ , t n ) ∣ − s ω s , K ( d t n ) ≤ C s , K , 1 max ( ∣ P 1 ( t ′ ) ∣ , ∣ P 0 ( t ′ ) ∣ ) − s ≤ C s , K , 1 ∣ Q ( t ′ ) ∣ − s . \int|\mathcal{P}(t',t_n)|^{-s}\,\omega_{s,K}(\mathrm{d}t_n)\le C_{s,K,1}\max\left(|\mathcal{P}_1(t')|,|\mathcal{P}_0(t')|\right)^{-s}\le C_{s,K,1}|\mathcal{Q}(t')|^{-s}. ∫ ∣ P ( t ′ , t n ) ∣ − s ω s , K ( d t n ) ≤ C s , K , 1 max ( ∣ P 1 ( t ′ ) ∣ , ∣ P 0 ( t ′ ) ∣ ) − s ≤ C s , K , 1 ∣ Q ( t ′ ) ∣ − s .
Integrating over t ′ t' t ′ , and using Tonelli’s theorem and the induction hypothesis for Q \mathcal{Q} Q , we obtain
∫ ∣ P ∣ − s d ω s , K ⊗ n ≤ C s , K , 1 ∫ ∣ Q ∣ − s d ω s , K ⊗ ( n − 1 ) ≤ C s , K , 1 n ( max α ∣ c α ( P ) ∣ ) − s . \int|\mathcal{P}|^{-s}\,\mathrm{d}\omega_{s,K}^{\otimes n}\le C_{s,K,1}\int|\mathcal{Q}|^{-s}\,\mathrm{d}\omega_{s,K}^{\otimes(n-1)}\le C_{s,K,1}^{n}\left(\max_\alpha|c_\alpha(\mathcal{P})|\right)^{-s}. ∫ ∣ P ∣ − s d ω s , K ⊗ n ≤ C s , K , 1 ∫ ∣ Q ∣ − s d ω s , K ⊗ ( n − 1 ) ≤ C s , K , 1 n ( α max ∣ c α ( P ) ∣ ) − s .
This completes the induction.
Three independent factors
For b ∈ C b\in\mathbb{C} b ∈ C and β > 0 \beta>0 β > 0 , we define M ( b , β ) M(b,\beta) M ( b , β ) by the formula (6.2); it is a complex 2 × 2 2\times2 2 × 2 matrix with determinant one. For vectors u , v ∈ C 2 u,v\in\mathbb{C}^{2} u , v ∈ C 2 and a matrix G G G , we write u T G v = ∑ i , j u i G i j v j u^{\mathsf{T}}Gv=\sum_{i,j}u_iG_{ij}v_j u T G v = ∑ i , j u i G ij v j for the bilinear (not sesquilinear) pairing.
Proposition 6.18. Let s ∈ ( 0 , 1 ) s \in(0,1) s ∈ ( 0 , 1 ) , K > 0 K > 0 K > 0 , and a + , a − ≥ 0 a_{+},a_{-} \ge0 a + , a − ≥ 0 . Let B 1 , B 2 , B 3 B_{1},B_{2},B_{3} B 1 , B 2 , B 3 be independent real random variables such that
E ∣ B i ∣ s ≤ a + , E ∣ B i ∣ − s ≤ a − , i ∈ { 1 , 2 , 3 } , \mathbb{E}|B_i|^{s} \le a_{+}, \qquad\mathbb{E}|B_i|^{-s} \le a_{-}, \qquad i \in\{1,2,3\}, E ∣ B i ∣ s ≤ a + , E ∣ B i ∣ − s ≤ a − , i ∈ { 1 , 2 , 3 } , and let β 1 , β 2 , β 3 \beta_{1},\beta_{2},\beta_{3} β 1 , β 2 , β 3 be independent random variables with the exponential law of rate K K K , independent of ( B 1 , B 2 , B 3 ) (B_{1},B_{2},B_{3}) ( B 1 , B 2 , B 3 ) . Set M i = M ( B i , β i ) M_i=M(B_i,\beta_i) M i = M ( B i , β i ) and G = M 3 M 2 M 1 G=M_{3}M_{2}M_{1} G = M 3 M 2 M 1 , and let
C ∗ = 2 s C s , K , 3 ( 1 + a + ) ( 1 + 2 a − ) , C_{*}=2^{s}C_{s,K,3}(1+a_{+})(1+2a_{-}), C ∗ = 2 s C s , K , 3 ( 1 + a + ) ( 1 + 2 a − ) , where C s , K , 3 C_{s,K,3} C s , K , 3 is the constant of Lemma 6.17. Then the following hold.
(i) For all nonzero u , v ∈ R 2 u,v \in\mathbb{R}^{2} u , v ∈ R 2 , we have E ∣ u T G v ∣ − s ≤ C ∗ ∥ u ∥ − s ∥ v ∥ − s \mathbb{E}|u^{\mathsf{T}}Gv|^{-s} \le C_{*}\|u\|^{-s}\|v\|^{-s} E ∣ u T G v ∣ − s ≤ C ∗ ∥ u ∥ − s ∥ v ∥ − s .
(ii) For all nonzero real 2 × 2 2 \times2 2 × 2 matrices A A A and C C C , we have E ∥ C G A ∥ − s ≤ C ∗ ∥ C ∥ − s ∥ A ∥ − s \mathbb{E}\|CGA\|^{-s} \le C_{*}\|C\|^{-s}\|A\|^{-s} E ∥ C G A ∥ − s ≤ C ∗ ∥ C ∥ − s ∥ A ∥ − s .
The laws of B 1 , B 2 , B 3 B_{1},B_{2},B_{3} B 1 , B 2 , B 3 need not coincide, and each may be any law with the stated moment bounds, including a law with atoms or a point mass at a nonzero constant. The same bounds hold for complex factors, in the sense of Proposition 6.19.
Proof. Both sides of (i) are homogeneous of degree − s -s − s in u u u and in v v v , so we may assume that ∥ u ∥ = ∥ v ∥ = 1 \|u\|=\|v\|=1 ∥ u ∥ = ∥ v ∥ = 1 . We write u = ( u 1 , u 2 ) u=(u_{1},u_{2}) u = ( u 1 , u 2 ) and v = ( v 1 , v 2 ) v=(v_{1},v_{2}) v = ( v 1 , v 2 ) , and we let N i = β i 1 / 2 M i = N_i=\beta_i^{1/2}M_i= N i = β i 1/2 M i =
( 0 1 − β i B i ) . \begin{pmatrix}
0 & 1\\
-\beta_i & B_i
\end{pmatrix}. ( 0 − β i 1 B i ) .
Multiplying out, we have
N 2 N 1 = ( − β 1 B 1 − B 2 β 1 B 1 B 2 − β 2 ) , N_{2}N_{1}=
\begin{pmatrix}
-\beta_{1} & B_{1}\\
-B_{2}\beta_{1} & B_{1}B_{2}-\beta_{2}
\end{pmatrix}, N 2 N 1 = ( − β 1 − B 2 β 1 B 1 B 1 B 2 − β 2 ) ,
N 3 N 2 N 1 = ( − B 2 β 1 B 1 B 2 − β 2 β 1 β 3 − B 2 B 3 β 1 B 1 B 2 B 3 − B 1 β 3 − B 3 β 2 ) . N_{3}N_{2}N_{1}=
\begin{pmatrix}
-B_{2}\beta_{1} & B_{1}B_{2}-\beta_{2}\\
\beta_{1}\beta_{3}-B_{2}B_{3}\beta_{1} & B_{1}B_{2}B_{3}-B_{1}\beta_{3}-B_{3}\beta_{2}
\end{pmatrix}. N 3 N 2 N 1 = ( − B 2 β 1 β 1 β 3 − B 2 B 3 β 1 B 1 B 2 − β 2 B 1 B 2 B 3 − B 1 β 3 − B 3 β 2 ) .
We see that P = u T N 3 N 2 N 1 v \mathcal{P}=u^{\mathsf{T}}N_{3}N_{2}N_{1}v P = u T N 3 N 2 N 1 v is a multiaffine polynomial in ( β 1 , β 2 , β 3 ) (\beta_{1},\beta_{2},\beta_{3}) ( β 1 , β 2 , β 3 ) , with coefficients depending on B 1 , B 2 , B 3 , u , v B_{1},B_{2},B_{3},u,v B 1 , B 2 , B 3 , u , v . Set
a = u 1 + B 3 u 2 , b = u 2 . a=u_{1}+B_{3}u_{2}, \qquad b=u_{2}. a = u 1 + B 3 u 2 , b = u 2 .
Collecting monomials, we find that the coefficients of β 1 β 3 \beta_{1}\beta_{3} β 1 β 3 , β 1 \beta_{1} β 1 , β 2 \beta_{2} β 2 , β 3 \beta_{3} β 3 , and 1 1 1 are
c 13 = b v 1 , c 1 = − B 2 a v 1 , c 2 = − a v 2 , c 3 = − B 1 b v 2 , c ∅ = B 1 B 2 a v 2 , c_{13}=bv_{1}, \qquad c_{1}=-B_{2}av_{1}, \qquad c_{2}=-av_{2}, \qquad c_{3}=-B_{1}bv_{2}, \qquad c_{\varnothing}=B_{1}B_{2}av_{2}, c 13 = b v 1 , c 1 = − B 2 a v 1 , c 2 = − a v 2 , c 3 = − B 1 b v 2 , c ∅ = B 1 B 2 a v 2 ,
and the coefficients of β 1 β 2 \beta_{1}\beta_{2} β 1 β 2 , β 2 β 3 \beta_{2}\beta_{3} β 2 β 3 , and β 1 β 2 β 3 \beta_{1}\beta_{2}\beta_{3} β 1 β 2 β 3 vanish. Let d = min ( 1 , ∣ B 1 ∣ , ∣ B 2 ∣ ) d=\min(1,|B_{1}|,|B_{2}|) d = min ( 1 , ∣ B 1 ∣ , ∣ B 2 ∣ ) . The four coefficients c 13 , c 1 , c 2 , c 3 c_{13},c_{1},c_{2},c_{3} c 13 , c 1 , c 2 , c 3 are the products b v 1 , a v 1 , a v 2 , b v 2 bv_{1},av_{1},av_{2},bv_{2} b v 1 , a v 1 , a v 2 , b v 2 , multiplied by 1 , − B 2 , − 1 , − B 1 1,-B_{2},-1,-B_{1} 1 , − B 2 , − 1 , − B 1 , respectively. Since d ≤ 1 d \le1 d ≤ 1 , d ≤ ∣ B 1 ∣ d \le|B_{1}| d ≤ ∣ B 1 ∣ , and d ≤ ∣ B 2 ∣ d \le|B_{2}| d ≤ ∣ B 2 ∣ , we find
max α ∣ c α ∣ ≥ d max ( ∣ b v 1 ∣ , ∣ a v 1 ∣ , ∣ a v 2 ∣ , ∣ b v 2 ∣ ) ≥ d 2 ( ∣ b v 1 ∣ 2 + ∣ a v 1 ∣ 2 + ∣ a v 2 ∣ 2 + ∣ b v 2 ∣ 2 ) 1 / 2 = d 2 ( ∣ a ∣ 2 + ∣ b ∣ 2 ) 1 / 2 , \max_{\alpha}|c_{\alpha}| \ge d\max\left(|bv_{1}|,|av_{1}|,|av_{2}|,|bv_{2}|\right)
\ge\frac{d}{2}\left(|bv_{1}|^{2}+|av_{1}|^{2}+|av_{2}|^{2}+|bv_{2}|^{2}\right)^{1/2}
=\frac{d}{2}\left(|a|^{2}+|b|^{2}\right)^{1/2}, α max ∣ c α ∣ ≥ d max ( ∣ b v 1 ∣ , ∣ a v 1 ∣ , ∣ a v 2 ∣ , ∣ b v 2 ∣ ) ≥ 2 d ( ∣ b v 1 ∣ 2 + ∣ a v 1 ∣ 2 + ∣ a v 2 ∣ 2 + ∣ b v 2 ∣ 2 ) 1/2 = 2 d ( ∣ a ∣ 2 + ∣ b ∣ 2 ) 1/2 ,
where in the second inequality we bounded the maximum of four numbers below by half of their Euclidean norm, and in the last step we used ∣ v 1 ∣ 2 + ∣ v 2 ∣ 2 = 1 |v_1|^2+|v_2|^2=1 ∣ v 1 ∣ 2 + ∣ v 2 ∣ 2 = 1 . Further, ( a , b ) T = S u (a,b)^{\mathsf T}=Su ( a , b ) T = S u with S = ( 1 B 3 0 1 ) S=\begin{pmatrix}1&B_3\\0&1\end{pmatrix} S = ( 1 0 B 3 1 ) , and S − 1 = ( 1 − B 3 0 1 ) S^{-1}=\begin{pmatrix}1&-B_3\\0&1\end{pmatrix} S − 1 = ( 1 0 − B 3 1 ) has norm at most 1 + ∣ B 3 ∣ 1+|B_3| 1 + ∣ B 3 ∣ . Since 1 = ∥ u ∥ ≤ ∥ S − 1 ∥ ∥ S u ∥ ≤ ( 1 + ∣ B 3 ∣ ) ∥ S u ∥ 1=\|u\|\leq\|S^{-1}\|\|Su\|\leq(1+|B_3|)\|Su\| 1 = ∥ u ∥ ≤ ∥ S − 1 ∥∥ S u ∥ ≤ ( 1 + ∣ B 3 ∣ ) ∥ S u ∥ , we have ( ∣ a ∣ 2 + ∣ b ∣ 2 ) 1 / 2 ≥ ( 1 + ∣ B 3 ∣ ) − 1 (|a|^2+|b|^2)^{1/2}\geq(1+|B_3|)^{-1} ( ∣ a ∣ 2 + ∣ b ∣ 2 ) 1/2 ≥ ( 1 + ∣ B 3 ∣ ) − 1 . We conclude that
max α ∣ c α ∣ ≥ d 2 ( 1 + ∣ B 3 ∣ ) . (91) \max_{\alpha}|c_\alpha|\geq\frac{d}{2(1+|B_3|)}. \tag*{(91)} α max ∣ c α ∣ ≥ 2 ( 1 + ∣ B 3 ∣ ) d . (91)
Since u T G v = ( β 1 β 2 β 3 ) − 1 / 2 P u^{\mathsf T}Gv=(\beta_1\beta_2\beta_3)^{-1/2}\mathcal{P} u T G v = ( β 1 β 2 β 3 ) − 1/2 P , we have ∣ u T G v ∣ − s = ( β 1 β 2 β 3 ) s / 2 ∣ P ∣ − s |u^{\mathsf T}Gv|^{-s}=(\beta_1\beta_2\beta_3)^{s/2}|\mathcal{P}|^{-s} ∣ u T G v ∣ − s = ( β 1 β 2 β 3 ) s /2 ∣ P ∣ − s . The variables β 1 , β 2 , β 3 \beta_1,\beta_2,\beta_3 β 1 , β 2 , β 3 are independent of ( B 1 , B 2 , B 3 ) (B_1,B_2,B_3) ( B 1 , B 2 , B 3 ) and have joint density ∏ i K e − K β i \prod_i Ke^{-K\beta_i} ∏ i K e − K β i . By Fubini’s theorem, the conditional expectation given ( B 1 , B 2 , B 3 ) (B_1,B_2,B_3) ( B 1 , B 2 , B 3 ) is
E [ ∣ u T G v ∣ − s ∣ B 1 , B 2 , B 3 ] = ∫ ( 0 , ∞ ) 3 ∣ P ( β ) ∣ − s ω s , K ( d β 1 ) ω s , K ( d β 2 ) ω s , K ( d β 3 ) \mathbb{E}\left[|u^{\mathsf T}Gv|^{-s}\mid B_1,B_2,B_3\right]=\int_{(0,\infty)^3}|\mathcal{P}(\beta)|^{-s}\,\omega_{s,K}(\mathrm{d}\beta_1)\,\omega_{s,K}(\mathrm{d}\beta_2)\,\omega_{s,K}(\mathrm{d}\beta_3) E [ ∣ u T G v ∣ − s ∣ B 1 , B 2 , B 3 ] = ∫ ( 0 , ∞ ) 3 ∣ P ( β ) ∣ − s ω s , K ( d β 1 ) ω s , K ( d β 2 ) ω s , K ( d β 3 )
≤ 2 s C s , K , 3 ( 1 + ∣ B 3 ∣ ) s d − s , \leq2^s C_{s,K,3}(1+|B_3|)^s d^{-s}, ≤ 2 s C s , K , 3 ( 1 + ∣ B 3 ∣ ) s d − s ,
where we used Lemma 6.17 and (6.19). We have ( 1 + ∣ B 3 ∣ ) s ≤ 1 + ∣ B 3 ∣ s (1+|B_3|)^s\leq1+|B_3|^s ( 1 + ∣ B 3 ∣ ) s ≤ 1 + ∣ B 3 ∣ s and d − s ≤ 1 + ∣ B 1 ∣ − s + ∣ B 2 ∣ − s d^{-s}\leq1+|B_1|^{-s}+|B_2|^{-s} d − s ≤ 1 + ∣ B 1 ∣ − s + ∣ B 2 ∣ − s . Taking expectations and using the independence of B 3 B_3 B 3 and ( B 1 , B 2 ) (B_1,B_2) ( B 1 , B 2 ) , we obtain
E ∣ u T G v ∣ − s ≤ 2 s C s , K , 3 ( 1 + E ∣ B 3 ∣ s ) ( 1 + E ∣ B 1 ∣ − s + E ∣ B 2 ∣ − s ) ≤ C ∗ . \mathbb{E}|u^{\mathsf T}Gv|^{-s}\leq2^sC_{s,K,3}(1+\mathbb{E}|B_3|^s)(1+\mathbb{E}|B_1|^{-s}+\mathbb{E}|B_2|^{-s})\leq C_*. E ∣ u T G v ∣ − s ≤ 2 s C s , K , 3 ( 1 + E ∣ B 3 ∣ s ) ( 1 + E ∣ B 1 ∣ − s + E ∣ B 2 ∣ − s ) ≤ C ∗ .
This proves (i).
For (ii), choose a unit vector v 0 v_0 v 0 with ∥ A v 0 ∥ = ∥ A ∥ \|Av_0\|=\|A\| ∥ A v 0 ∥ = ∥ A ∥ and a unit vector w 0 w_0 w 0 with ∥ C T w 0 ∥ = ∥ C T ∥ = ∥ C ∥ \|C^{\mathsf T}w_0\|=\|C^{\mathsf T}\|=\|C\| ∥ C T w 0 ∥ = ∥ C T ∥ = ∥ C ∥ , and set v = A v 0 / ∥ A ∥ v=Av_0/\|A\| v = A v 0 /∥ A ∥ and u = C T w 0 / ∥ C ∥ u=C^{\mathsf T}w_0/\|C\| u = C T w 0 /∥ C ∥ . These are unit vectors depending only on A A A and C C C , and
∥ C G A ∥ ≥ ∣ w 0 T C G A v 0 ∣ = ∣ ( C T w 0 ) T G ( A v 0 ) ∣ = ∥ C ∥ ∥ A ∥ ∣ u T G v ∣ . \|CGA\|\geq|w_0^{\mathsf T}CGAv_0|=|(C^{\mathsf T}w_0)^{\mathsf T}G(Av_0)|=\|C\|\|A\||u^{\mathsf T}Gv|. ∥ C G A ∥ ≥ ∣ w 0 T C G A v 0 ∣ = ∣ ( C T w 0 ) T G ( A v 0 ) ∣ = ∥ C ∥∥ A ∥∣ u T G v ∣.
Raising to the power − s -s − s , taking expectations, and applying (i), we obtain (ii). □
The same argument applies to complex denominators, which arise when the spectral parameter has positive imaginary part.
Proposition 6.19. Let s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) , K > 0 K>0 K > 0 , and a + , a − ≥ 0 a_+,a_-\geq0 a + , a − ≥ 0 . Let B 1 , B 2 , B 3 B_1,B_2,B_3 B 1 , B 2 , B 3 be independent complex random variables with E ∣ B i ∣ s ≤ a + \mathbb{E}|B_i|^s\leq a_+ E ∣ B i ∣ s ≤ a + and E ∣ B i ∣ − s ≤ a − \mathbb{E}|B_i|^{-s}\leq a_- E ∣ B i ∣ − s ≤ a − for i ∈ { 1 , 2 , 3 } i\in\{1,2,3\} i ∈ { 1 , 2 , 3 } , let β 1 , β 2 , β 3 \beta_1,\beta_2,\beta_3 β 1 , β 2 , β 3 be as in Proposition 6.18, and set M i = M ( B i , β i ) M_i=M(B_i,\beta_i) M i = M ( B i , β i ) and G = M 3 M 2 M 1 G=M_3M_2M_1 G = M 3 M 2 M 1 . Let C ∗ C_* C ∗ be the constant of Proposition 6.18. Then the following hold.
(i) For all nonzero u , v ∈ C 2 u,v\in\mathbb{C}^2 u , v ∈ C 2 , we have E ∣ u T G v ∣ − s ≤ C ∗ ∥ u ∥ − s ∥ v ∥ − s \mathbb{E}|u^{\mathsf T}Gv|^{-s}\leq C_*\|u\|^{-s}\|v\|^{-s} E ∣ u T G v ∣ − s ≤ C ∗ ∥ u ∥ − s ∥ v ∥ − s .
(ii) For all nonzero complex 2 × 2 2\times2 2 × 2 matrices A A A and C C C , we have E ∥ C G A ∥ − s ≤ C ∗ ∥ C ∥ − s ∥ A ∥ − s \mathbb{E}\|CGA\|^{-s}\leq C_*\|C\|^{-s}\|A\|^{-s} E ∥ C G A ∥ − s ≤ C ∗ ∥ C ∥ − s ∥ A ∥ − s .
In particular, the bounds hold with the same constant C ∗ C_* C ∗ for every member of a family of triples ( B 1 , B 2 , B 3 ) (B_1,B_2,B_3) ( B 1 , B 2 , B 3 ) of independent complex random variables satisfying the moment bounds with the same a + a_+ a + and a − a_- a − .
Proof. For (i), the proof of Proposition 6.18(i) applies with the following changes. The expansion of u T N 3 N 2 N 1 v u^{\mathsf T}N_3N_2N_1v u T N 3 N 2 N 1 v is a polynomial identity and is unchanged, and P \mathcal{P} P is a multiaffine polynomial in ( β 1 , β 2 , β 3 ) (\beta_1,\beta_2,\beta_3) ( β 1 , β 2 , β 3 ) with the complex coefficients displayed there. The identity
∣ b v 1 ∣ 2 + ∣ a v 1 ∣ 2 + ∣ a v 2 ∣ 2 + ∣ b v 2 ∣ 2 = ( ∣ a ∣ 2 + ∣ b ∣ 2 ) ( ∣ v 1 ∣ 2 + ∣ v 2 ∣ 2 ) |bv_1|^2+|av_1|^2+|av_2|^2+|bv_2|^2=(|a|^2+|b|^2)(|v_1|^2+|v_2|^2) ∣ b v 1 ∣ 2 + ∣ a v 1 ∣ 2 + ∣ a v 2 ∣ 2 + ∣ b v 2 ∣ 2 = ( ∣ a ∣ 2 + ∣ b ∣ 2 ) ( ∣ v 1 ∣ 2 + ∣ v 2 ∣ 2 )
holds for complex numbers, the bound ∥ S − 1 ∥ ≤ 1 + ∣ B 3 ∣ \|S^{-1}\|\leq1+|B_3| ∥ S − 1 ∥ ≤ 1 + ∣ B 3 ∣ holds for the Hermitian operator norm, and Lemma 6.17 allows complex coefficients. Then (6.19) and the rest of the argument hold verbatim.
For (ii), choose unit vectors v 0 , w 0 ∈ C 2 v_0,w_0\in\mathbb{C}^2 v 0 , w 0 ∈ C 2 with ∥ A v 0 ∥ = ∥ A ∥ \|Av_0\|=\|A\| ∥ A v 0 ∥ = ∥ A ∥ and ∥ C ∗ w 0 ∥ = ∥ C ∥ \|C^*w_0\|=\|C\| ∥ C ∗ w 0 ∥ = ∥ C ∥ , where C ∗ C^* C ∗ is the conjugate transpose. Set v = A v 0 / ∥ A ∥ v=Av_0/\|A\| v = A v 0 /∥ A ∥ , and let u = C ∗ w 0 / ∥ C ∥ u=C^*w_0/\|C\| u = C ∗ w 0 /∥ C ∥ , the entrywise complex conjugate of C ∗ w 0 C^*w_0 C ∗ w 0 divided by ∥ C ∥ \|C\| ∥ C ∥ . Then ( C ∗ w 0 ) ∗ = ∥ C ∥ u T (C^*w_0)^*=\|C\|u^{\mathsf T} ( C ∗ w 0 ) ∗ = ∥ C ∥ u T , and
∥ C G A ∥ ≥ ∣ w 0 ∗ C G A v 0 ∣ = ∣ ( C ∗ w 0 ) ∗ G ( A v 0 ) ∣ = ∥ C ∥ ∥ A ∥ ∣ u T G v ∣ . \|CGA\|\geq|w_0^*CGA v_0|=|(C^*w_0)^*G(Av_0)|=\|C\|\|A\||u^{\mathsf T}Gv|. ∥ C G A ∥ ≥ ∣ w 0 ∗ C G A v 0 ∣ = ∣ ( C ∗ w 0 ) ∗ G ( A v 0 ) ∣ = ∥ C ∥∥ A ∥∣ u T G v ∣.
Part (i) completes the proof. The last statement holds because C ∗ C_* C ∗ depends only on s s s , K K K , a + a_+ a + , and a − a_- a − .
Remark 6.20. Proposition 6.19 applies to the denominators of the tree at the spectral parameter E + i η E+i\eta E + i η with η > 0 \eta>0 η > 0 . Let E > 0 E>0 E > 0 , g > 0 g>0 g > 0 , and 0 < η ≤ 1 0<\eta\leq1 0 < η ≤ 1 , and let B η = g D η B_\eta=gD_\eta B η = g D η with D η = − 1 / Γ o ( E + i η ) D_\eta=-1/\Gamma_o(E+i\eta) D η = − 1/ Γ o ( E + i η ) . By Lemma 5.17, both sup 0 < η ≤ 1 E ∣ B η ∣ − s = g − s sup 0 < η ≤ 1 E ∣ Γ o ( E + i η ) ∣ s \sup_{0<\eta\leq1}\mathbb{E}|B_\eta|^{-s}=g^{-s}\sup_{0<\eta\leq1}\mathbb{E}|\Gamma_o(E+i\eta)|^s sup 0 < η ≤ 1 E ∣ B η ∣ − s = g − s sup 0 < η ≤ 1 E ∣ Γ o ( E + i η ) ∣ s and sup 0 < η ≤ 1 E ∣ B η ∣ s = g s sup 0 < η ≤ 1 E ∣ D η ∣ s \sup_{0<\eta\leq1}\mathbb{E}|B_\eta|^s=g^s\sup_{0<\eta\leq1}\mathbb{E}|D_\eta|^s sup 0 < η ≤ 1 E ∣ B η ∣ s = g s sup 0 < η ≤ 1 E ∣ D η ∣ s are finite for s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) . Then the constant C ∗ C_* C ∗ can be chosen uniformly in η ∈ ( 0 , 1 ] \eta\in(0,1] η ∈ ( 0 , 1 ] .
Corollary 6.21. Let s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) , K > 0 K>0 K > 0 , a + , a − ≥ 0 a_+,a_-\geq0 a + , a − ≥ 0 , and let C ∗ C_* C ∗ be as in Proposition 6.18. Let G G G and G ′ G' G ′ be independent random matrices, each a product of three factors as in Proposition 6.19 (with possibly different laws of the B i B_i B i , all satisfying the moment bounds with a + a_+ a + and a − a_- a − ).
(i) For all nonzero u , v ∈ C 2 u,v\in\mathbb{C}^2 u , v ∈ C 2 and every nonzero complex 2 × 2 2\times2 2 × 2 matrix A A A , we have
E ∣ u T G A G ′ v ∣ − s ≤ C ∗ 2 ∥ u ∥ − s ∥ A ∥ − s ∥ v ∥ − s . \mathbb{E}|u^{\mathsf T}GAG'v|^{-s}\leq C_*^2\|u\|^{-s}\|A\|^{-s}\|v\|^{-s}. E ∣ u T G A G ′ v ∣ − s ≤ C ∗ 2 ∥ u ∥ − s ∥ A ∥ − s ∥ v ∥ − s . (ii) For all nonzero complex 2 × 2 2\times2 2 × 2 matrices A A A , C C C , and F F F , we have
E ∥ C G F G ′ A ∥ − s ≤ C ∗ 2 ∥ C ∥ − s ∥ F ∥ − s ∥ A ∥ − s . \mathbb{E}\|CGFG'A\|^{-s}\leq C_*^2\|C\|^{-s}\|F\|^{-s}\|A\|^{-s}. E ∥ C GF G ′ A ∥ − s ≤ C ∗ 2 ∥ C ∥ − s ∥ F ∥ − s ∥ A ∥ − s . If ∥ F ∥ ≥ 1 \|F\|\geq1 ∥ F ∥ ≥ 1 , which holds in particular if ∣ det F ∣ ≥ 1 |\det F|\geq1 ∣ det F ∣ ≥ 1 , the right side is at most C ∗ 2 ∥ C ∥ − s ∥ A ∥ − s C_*^2\|C\|^{-s}\|A\|^{-s} C ∗ 2 ∥ C ∥ − s ∥ A ∥ − s .
(iii) Let G \mathcal{G} G be a σ \sigma σ -algebra independent of ( G , G ′ ) (G,G') ( G , G ′ ) , and let u , v , A , C , F u,v,A,C,F u , v , A , C , F be G \mathcal{G} G -measurable random vectors and matrices that are almost surely nonzero. Then the bounds in (i) and (ii) hold almost surely with E \mathbb{E} E replaced by E [ ⋅ ∣ G ] \mathbb{E}[\cdot\mid\mathcal{G}] E [ ⋅ ∣ G ] .
Proof. For a nonzero complex matrix A A A , choose a unit vector w 0 w_0 w 0 with ∥ A ∗ w 0 ∥ = ∥ A ∥ \|A^*w_0\|=\|A\| ∥ A ∗ w 0 ∥ = ∥ A ∥ , and let u A = A ∗ w 0 / ∥ A ∥ u_A=A^*w_0/\|A\| u A = A ∗ w 0 /∥ A ∥ . As in the proof of Proposition 6.19(ii), we have
∥ A x ∥ ≥ ∣ w 0 ∗ A x ∣ = ∥ A ∥ ∣ u A T x ∣ for every x ∈ C 2 . (92) \|Ax\|\geq|w_0^*Ax|=\|A\||u_A^{\mathsf T}x|\qquad\text{for every }x\in\mathbb{C}^2. \tag*{(92)} ∥ A x ∥ ≥ ∣ w 0 ∗ A x ∣ = ∥ A ∥∣ u A T x ∣ for every x ∈ C 2 . (92)
For (i), we condition on G ′ G' G ′ ; since G G G is independent of G ′ G' G ′ , Proposition 6.19(i) applied to the vectors u u u and A G ′ v AG'v A G ′ v gives E [ ∣ u T G A G ′ v ∣ − s ∣ G ′ ] ≤ C ∗ ∥ u ∥ − s ∥ A G ′ v ∥ − s \mathbb{E}[|u^{\mathsf{T}}GAG'v|^{-s}\mid G'] \le C_* \|u\|^{-s}\|AG'v\|^{-s} E [ ∣ u T G A G ′ v ∣ − s ∣ G ′ ] ≤ C ∗ ∥ u ∥ − s ∥ A G ′ v ∥ − s (both sides are infinite if A G ′ v = 0 AG'v=0 A G ′ v = 0 ). By (6.20), we have ∥ A G ′ v ∥ − s ≤ ∥ A ∥ − s ∣ u T G ′ v ∣ − s \|AG'v\|^{-s}\le\|A\|^{-s}|u^{\mathsf{T}}G'v|^{-s} ∥ A G ′ v ∥ − s ≤ ∥ A ∥ − s ∣ u T G ′ v ∣ − s , and Proposition 6.19(i) implies E ∣ u T G ′ v ∣ − s ≤ C ∗ ∥ v ∥ − s \mathbb{E}|u^{\mathsf{T}}G'v|^{-s}\le C_*\|v\|^{-s} E ∣ u T G ′ v ∣ − s ≤ C ∗ ∥ v ∥ − s . Combining these bounds proves (i).
For (ii), we condition on G ′ G' G ′ and apply Proposition 6.19(ii) to G G G with the matrices C C C and F G ′ A FG'A F G ′ A , which gives E [ ∥ C G F G ′ A ∥ − s ∣ G ′ ] ≤ C ∗ ∥ C ∥ − s ∥ F G ′ A ∥ − s \mathbb{E}[\|CGFG'A\|^{-s}\mid G']\le C_*\|C\|^{-s}\|FG'A\|^{-s} E [ ∥ C GF G ′ A ∥ − s ∣ G ′ ] ≤ C ∗ ∥ C ∥ − s ∥ F G ′ A ∥ − s . Applying Proposition 6.19(ii) to G ′ G' G ′ with the matrices F F F and A A A gives E ∥ F G ′ A ∥ − s ≤ C ∗ ∥ F ∥ − s ∥ A ∥ − s \mathbb{E}\|FG'A\|^{-s}\le C_*\|F\|^{-s}\|A\|^{-s} E ∥ F G ′ A ∥ − s ≤ C ∗ ∥ F ∥ − s ∥ A ∥ − s . If ∣ det F ∣ ≥ 1 |\det F|\ge1 ∣ det F ∣ ≥ 1 , then ∥ F ∥ 2 ≥ ∣ det F ∣ ≥ 1 \|F\|^2\ge|\det F|\ge1 ∥ F ∥ 2 ≥ ∣ det F ∣ ≥ 1 , since ∣ det F ∣ |\det F| ∣ det F ∣ is the product of the two singular values of F F F .
For (iii), since ( G , G ′ ) (G,G') ( G , G ′ ) is independent of G \mathcal{G} G , the conditional expectation given G \mathcal{G} G of a nonnegative measurable function of ( G , G ′ ) (G,G') ( G , G ′ ) and of G \mathcal{G} G -measurable variables is obtained by integrating over ( G , G ′ ) (G,G') ( G , G ′ ) with the G \mathcal{G} G -measurable variables held fixed [81, Chapter 6]. Applying (i) and (ii) for fixed values of u , v , A , C , F u,v,A,C,F u , v , A , C , F gives (iii). □
Moments of norms of products
Corollary 6.22. Let s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) . Suppose that ν \nu ν satisfies (6.1) and ∫ ∣ b ∣ − s ν ( d b ) < ∞ \int|b|^{-s}\nu(\mathrm{d}b)<\infty ∫ ∣ b ∣ − s ν ( d b ) < ∞ , and let f L ( s ) = E ∥ M L ⋯ M 1 ∥ − s f_L(s)=\mathbb{E}\|M_L\cdots M_1\|^{-s} f L ( s ) = E ∥ M L ⋯ M 1 ∥ − s be as in Definition 6.14. Then the limit ϱ s = lim L → ∞ f L ( s ) 1 / L \varrho_s=\lim_{L\to\infty}f_L(s)^{1/L} ϱ s = lim L → ∞ f L ( s ) 1/ L exists, and ϱ s = λ ( s ) / p = e φ ( s ) − log p ∈ ( 0 , 1 ) \varrho_s=\lambda(s)/p=e^{\varphi(s)-\log p}\in(0,1) ϱ s = λ ( s ) / p = e φ ( s ) − l o g p ∈ ( 0 , 1 ) . Further, let C ∗ C_* C ∗ be the constant of Proposition 6.18 with a + = E ∣ B ∣ s a_+=\mathbb{E}|B|^s a + = E ∣ B ∣ s and a − = E ∣ B ∣ − s a_-=\mathbb{E}|B|^{-s} a − = E ∣ B ∣ − s . Then for every L ≥ 1 L\ge1 L ≥ 1 ,
C ∗ − 1 ϱ s L + 3 ≤ f L ( s ) ≤ ϱ s L . C_*^{-1}\varrho_s^{L+3}\le f_L(s)\le\varrho_s^L. C ∗ − 1 ϱ s L + 3 ≤ f L ( s ) ≤ ϱ s L .
Proof. We abbreviate f L = f L ( s ) f_L=f_L(s) f L = f L ( s ) and W L = M L ⋯ M 1 W_L=M_L\cdots M_1 W L = M L ⋯ M 1 . Since det M j = 1 \det M_j=1 det M j = 1 , every product has norm at least one, and 0 < f L ≤ 1 0<f_L\le1 0 < f L ≤ 1 . For L , N ≥ 0 L,N\ge0 L , N ≥ 0 , we have ∥ W L + N ∥ ≤ ∥ M L + N ⋯ M L + 1 ∥ ∥ W L ∥ \|W_{L+N}\|\le\|M_{L+N}\cdots M_{L+1}\|\|W_L\| ∥ W L + N ∥ ≤ ∥ M L + N ⋯ M L + 1 ∥∥ W L ∥ , and the two factors on the right are independent, the first having the law of W N W_N W N . This implies that f L + N ≥ f N f L f_{L+N}\ge f_Nf_L f L + N ≥ f N f L , that is, the sequence ( log f L ) (\log f_L) ( log f L ) is superadditive. By Fekete’s lemma, we have lim L → ∞ L − 1 log f L = sup L ≥ 1 L − 1 log f L ∈ [ log f 1 , 0 ] \lim_{L\to\infty}L^{-1}\log f_L=\sup_{L\ge1}L^{-1}\log f_L\in[\log f_1,0] lim L → ∞ L − 1 log f L = sup L ≥ 1 L − 1 log f L ∈ [ log f 1 , 0 ] . This proves that ϱ s \varrho_s ϱ s exists, that ϱ s > 0 \varrho_s>0 ϱ s > 0 , and that f L ≤ ϱ s L f_L\le\varrho_s^L f L ≤ ϱ s L .
By Lemma 6.4, we have f L ≤ E J s ( W L ) ≤ C s f L f_L\le\mathbb{E}J_s(W_L)\le C_sf_L f L ≤ E J s ( W L ) ≤ C s f L , with C s C_s C s independent of L L L . Then Theorem 6.11(v) implies that log ϱ s = log λ ( s ) − log p = φ ( s ) − log p \log\varrho_s=\log\lambda(s)-\log p=\varphi(s)-\log p log ϱ s = log λ ( s ) − log p = φ ( s ) − log p , and φ ( s ) < log p \varphi(s)<\log p φ ( s ) < log p by Proposition 6.15(v).
For the lower bound, let L , N ≥ 1 L,N\ge1 L , N ≥ 1 , and write W L + N + 3 = C ′ G A W_{L+N+3}=C'GA W L + N + 3 = C ′ G A with A = W L A=W_L A = W L , G = M L + 3 M L + 2 M L + 1 G=M_{L+3}M_{L+2}M_{L+1} G = M L + 3 M L + 2 M L + 1 , and C ′ = M L + N + 3 ⋯ M L + 4 C'=M_{L+N+3}\cdots M_{L+4} C ′ = M L + N + 3 ⋯ M L + 4 . The matrix G G G is a product of three independent factors as in Proposition 6.18, with a + = E ∣ B ∣ s a_+=\mathbb{E}|B|^s a + = E ∣ B ∣ s and a − = E ∣ B ∣ − s a_-=\mathbb{E}|B|^{-s} a − = E ∣ B ∣ − s , and it is independent of ( A , C ′ ) (A,C') ( A , C ′ ) . Since G G G is independent of ( A , C ′ ) (A,C') ( A , C ′ ) , we may apply Proposition 6.18(ii) conditionally on ( A , C ′ ) (A,C') ( A , C ′ ) , integrating over G G G with A A A and C ′ C' C ′ held fixed (as in the proof of Corollary 6.21(iii)), and we obtain E [ ∥ C ′ G A ∥ − s ∣ A , C ′ ] ≤ C ∗ ∥ C ′ ∥ − s ∥ A ∥ − s \mathbb{E}[\|C'GA\|^{-s}\mid A,C']\le C_*\|C'\|^{-s}\|A\|^{-s} E [ ∥ C ′ G A ∥ − s ∣ A , C ′ ] ≤ C ∗ ∥ C ′ ∥ − s ∥ A ∥ − s . Taking expectations and using the independence of A A A and C ′ C' C ′ , where C ′ C' C ′ has the law of W N W_N W N , we obtain
f L + N + 3 ≤ C ∗ f N f L . (93) f_{L+N+3}\le C_*f_Nf_L. \tag*{(93)} f L + N + 3 ≤ C ∗ f N f L . (93)
By induction on k ≥ 1 k \ge1 k ≥ 1 , this gives f k ( L + 3 ) − 3 ≤ C ∗ k − 1 f L k f_{k(L+3)-3} \le C_*^{k-1}f_L^k f k ( L + 3 ) − 3 ≤ C ∗ k − 1 f L k . Since n − 1 log f n → log ϱ s n^{-1}\log f_n \to\log\varrho_s n − 1 log f n → log ϱ s , taking logarithms and dividing by k k k gives
k ( L + 3 ) − 3 k ⋅ log f k ( L + 3 ) − 3 k ( L + 3 ) − 3 ≤ k − 1 k log C ∗ + log f L . \frac{k(L+3)-3}{k}\cdot\frac{\log f_{k(L+3)-3}}{k(L+3)-3}\le\frac{k-1}{k}\log C_*+\log f_L. k k ( L + 3 ) − 3 ⋅ k ( L + 3 ) − 3 log f k ( L + 3 ) − 3 ≤ k k − 1 log C ∗ + log f L .
Letting k → ∞ k\to\infty k → ∞ , we find ( L + 3 ) log ϱ s ≤ log C ∗ + log f L (L+3)\log\varrho_s\le\log C_*+\log f_L ( L + 3 ) log ϱ s ≤ log C ∗ + log f L , which is the lower bound.
Path moments
In this section, we identify the free energy with the growth rate of the path moments Φ L ( s ) \Phi_L(s) Φ L ( s ) , which are iterates of the transfer operator A s A_s A s in the x x x -coordinate. We begin by writing the iterates of A s A_s A s along a chain of independent factors.
Lemma 6.23. Suppose that ν \nu ν satisfies (6.1), and let s ∈ ( 0 , 2 ) s\in(0,2) s ∈ ( 0 , 2 ) and L ≥ 1 L\ge1 L ≥ 1 .
(i) Let B 0 ′ , … , B L − 1 ′ B'_0,\ldots,B'_{L-1} B 0 ′ , … , B L − 1 ′ have law ν \nu ν and β 1 , … , β L \beta_1,\ldots,\beta_L β 1 , … , β L have the exponential law of rate K K K , all independent. For x ∈ R x\in\mathbb{R} x ∈ R , set X L = x X_L=x X L = x and X j − 1 = f B j − 1 ′ , β j ( X j ) X_{j-1}=f_{B'_{j-1},\beta_j}(X_j) X j − 1 = f B j − 1 ′ , β j ( X j ) for j = L , … , 1 j=L,\ldots,1 j = L , … , 1 . Then almost surely X j ∈ R X_j\in\mathbb{R} X j ∈ R for every j j j , and for every Borel F : R → [ 0 , ∞ ] F:\mathbb{R}\to[0,\infty] F : R → [ 0 , ∞ ] ,
( A s L F ) ( x ) = p L E [ ∏ j = 1 L β j s / 2 ∣ X j ∣ s F ( X 0 ) ] . (94) (A_s^L F)(x)=p^L\mathbb{E}\left[\prod_{j=1}^{L}\beta_j^{s/2}|X_j|^sF(X_0)\right]. \tag*{(94)} ( A s L F ) ( x ) = p L E [ j = 1 ∏ L β j s /2 ∣ X j ∣ s F ( X 0 ) ] . (94) (ii) Let π \pi π be a real solution at energy E > 0 E>0 E > 0 , and for g > 0 g>0 g > 0 let A s ( g ) A_s^{(g)} A s ( g ) be the transfer operator in the x x x -coordinate defined with ν g \nu_g ν g and K g = p / g 2 K_g=p/g^2 K g = p / g 2 , and let π g x \pi_g^x π g x be the law of 1 / ( g D ) 1/(gD) 1/ ( g D ) , where D D D has law π \pi π . Let D L , D 0 ′ , … , D L − 1 ′ D_L,D'_0,\ldots,D'_{L-1} D L , D 0 ′ , … , D L − 1 ′ have law π \pi π and Y 1 , … , Y L Y_1,\ldots,Y_L Y 1 , … , Y L have law Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) , all independent, and set D j − 1 = D j − 1 ′ − Y j / ( p D j ) D_{j-1}=D'_{j-1}-Y_j/(pD_j) D j − 1 = D j − 1 ′ − Y j / ( p D j ) for j = L , … , 1 j=L,\ldots,1 j = L , … , 1 . Then almost surely D j ∈ R ∖ { 0 } D_j\in\mathbb{R}\setminus\{0\} D j ∈ R ∖ { 0 } for every j j j , and for every g > 0 g>0 g > 0 ,
g s ∫ A s ( g ) L e s d π g x = p L E [ ∣ D 0 ∣ − s ∏ j = 1 L ( Y j p ) s / 2 ∣ D j ∣ − s ] . (95) g^s\int A_s^{(g)L}e_s\,d\pi_g^x
=p^L\mathbb{E}\left[|D_0|^{-s}\prod_{j=1}^{L}\left(\frac{Y_j}{p}\right)^{s/2}|D_j|^{-s}\right]. \tag*{(95)} g s ∫ A s ( g ) L e s d π g x = p L E [ ∣ D 0 ∣ − s j = 1 ∏ L ( p Y j ) s /2 ∣ D j ∣ − s ] . (95) In particular, the left side does not depend on g g g .
Proof. For (i), we first show that X j ∈ R X_j\in\mathbb{R} X j ∈ R almost surely. Suppose that X j ∈ R X_j\in\mathbb{R} X j ∈ R almost surely. The variables ( B j − 1 ′ , β j ) (B'_{j-1},\beta_j) ( B j − 1 ′ , β j ) are independent of X j X_j X j , which is a function of B j ′ , … , B L − 1 ′ , β j + 1 , … , β L B'_j,\ldots,B'_{L-1},\beta_{j+1},\ldots,\beta_L B j ′ , … , B L − 1 ′ , β j + 1 , … , β L . By Definition 6.1(vi), for every fixed y ∈ R y\in\mathbb{R} y ∈ R we have f B j − 1 ′ , β j ( y ) ∈ R f_{B'_{j-1},\beta_j}(y)\in\mathbb{R} f B j − 1 ′ , β j ( y ) ∈ R almost surely, and Fubini’s theorem implies X j − 1 ∈ R X_{j-1}\in\mathbb{R} X j − 1 ∈ R almost surely. Formula (6.21) for L = 1 L=1 L = 1 is the definition of A s A_s A s . Suppose it holds for L L L . Then
( A s L + 1 F ) ( x ) = p E [ β L + 1 s / 2 ∣ x ∣ s ( A s L F ) ( f B L ′ , β L + 1 ( x ) ) ] . (A_s^{L+1}F)(x)=p\mathbb{E}\left[\beta_{L+1}^{s/2}|x|^s(A_s^LF)(f_{B'_L,\beta_{L+1}}(x))\right]. ( A s L + 1 F ) ( x ) = p E [ β L + 1 s /2 ∣ x ∣ s ( A s L F ) ( f B L ′ , β L + 1 ( x )) ] .
Inserting (6.21) for A s L F A^{L}_{s}F A s L F , evaluated at the point f B L ′ , β L + 1 ( x ) f_{B'_{L},\beta_{L+1}}(x) f B L ′ , β L + 1 ( x ) with the independent variables B 0 ′ , … , B L − 1 ′ , β 1 , … , β L B'_0,\ldots,B'_{L-1},\beta_1,\ldots,\beta_L B 0 ′ , … , B L − 1 ′ , β 1 , … , β L , and using Tonelli’s theorem, we obtain (6.21) for L + 1 L+1 L + 1 , with X L + 1 = x X_{L+1}=x X L + 1 = x and X L = f B L ′ , β L + 1 ( x ) X_L=f_{B'_{L},\beta_{L+1}}(x) X L = f B L ′ , β L + 1 ( x ) .
For (ii), the variables D j D_j D j are almost surely finite and nonzero by the same argument: given D j ∈ R ∖ { 0 } D_j\in\mathbb{R}\setminus\{0\} D j ∈ R ∖ { 0 } , the variable D j − 1 D_{j-1} D j − 1 is finite, and it vanishes only if D j − 1 ′ = Y j / ( p D j ) D'_{j-1}=Y_j/(pD_j) D j − 1 ′ = Y j / ( p D j ) , which has probability zero because Y j Y_j Y j has a density and is independent of ( D j − 1 ′ , D j ) (D'_{j-1},D_j) ( D j − 1 ′ , D j ) . Set B j ′ = g D j ′ B'_j=gD'_j B j ′ = g D j ′ , β j = g 2 Y j / p \beta_j=g^2Y_j/p β j = g 2 Y j / p , and X j = 1 / ( g D j ) X_j=1/(gD_j) X j = 1/ ( g D j ) . Then B j ′ B'_j B j ′ has law ν g \nu_g ν g , β j \beta_j β j has the exponential law of rate K g K_g K g , X L X_L X L has law π g x \pi_g^x π g x , all are independent, and
f B j − 1 ′ , β j ( X j ) = ( g D j − 1 ′ − g 2 Y j p ⋅ 1 g D j ) − 1 = 1 g D j − 1 = X j − 1 . f_{B'_{j-1},\beta_j}(X_j)=\left(gD'_{j-1}-\frac{g^2Y_j}{p}\cdot\frac{1}{gD_j}\right)^{-1}=\frac{1}{gD_{j-1}}=X_{j-1}. f B j − 1 ′ , β j ( X j ) = ( g D j − 1 ′ − p g 2 Y j ⋅ g D j 1 ) − 1 = g D j − 1 1 = X j − 1 .
By (i) with F = e s F=e_s F = e s , applied conditionally on X L X_L X L and integrated against π g x \pi_g^x π g x using Tonelli’s theorem, we obtain
g s ∫ A s ( g ) L e s d π g x = g s p L E [ ∣ X 0 ∣ s ∏ j = 1 L β j s / 2 ∣ X j ∣ s ] = g s p L E [ g − s ∣ D 0 ∣ − s ∏ j = 1 L g s ( Y j p ) s / 2 g − s ∣ D j ∣ − s ] , \begin{aligned}
g^s\int A_s^{(g)L}e_s\,\mathrm{d}\pi_g^x=g^sp^L\mathbb{E}\left[|X_0|^s\prod_{j=1}^{L}\beta_j^{s/2}|X_j|^s\right] \\
&=g^sp^L\mathbb{E}\left[g^{-s}|D_0|^{-s}\prod_{j=1}^{L}g^s\left(\frac{Y_j}{p}\right)^{s/2}g^{-s}|D_j|^{-s}\right],
\end{aligned} g s ∫ A s ( g ) L e s d π g x = g s p L E [ ∣ X 0 ∣ s j = 1 ∏ L β j s /2 ∣ X j ∣ s ] = g s p L E [ g − s ∣ D 0 ∣ − s j = 1 ∏ L g s ( p Y j ) s /2 g − s ∣ D j ∣ − s ] ,
which is (6.22).
Definition 6.24. Let π \pi π be a real solution at energy E > 0 E>0 E > 0 , let s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) , and let L ≥ 0 L\geq0 L ≥ 0 . The path moment is
Φ L ( s ) = Φ L ( s ; E ) = g s ∫ A s ( g ) L e s d π g x ∈ [ 0 , ∞ ] , \Phi_L(s)=\Phi_L(s;E)=g^s\int A_s^{(g)L}e_s\,\mathrm{d}\pi_g^x\in[0,\infty], Φ L ( s ) = Φ L ( s ; E ) = g s ∫ A s ( g ) L e s d π g x ∈ [ 0 , ∞ ] , where g > 0 g>0 g > 0 , and A s ( g ) A_s^{(g)} A s ( g ) and π g x \pi_g^x π g x are as in Lemma 6.23(ii); for L = 0 L=0 L = 0 , we set A s ( g ) 0 e s = e s A_s^{(g)0}e_s=e_s A s ( g ) 0 e s = e s . By Lemma 6.23(ii), Φ L ( s ) \Phi_L(s) Φ L ( s ) does not depend on g g g and is given by the right side of (6.22).
Remark 6.25. We explain the name. Let π \pi π be a real solution at E E E , and let ( D v ) (D_v) ( D v ) be the variables of the recursive tree process of Lemma 5.13, which have law π \pi π and satisfy D v = E − p − 1 ∑ u Y u / D u D_v=E-p^{-1}\sum_uY_u/D_u D v = E − p − 1 ∑ u Y u / D u , the sum running over the children u u u of v v v . (When the no-ac hypothesis of Proposition 5.11 holds and π \pi π is the real boundary law, the boundary denominators − 1 / Γ v ( E + i 0 ) -1/\Gamma_v(E+i0) − 1/ Γ v ( E + i 0 ) satisfy the same recursion.) For a vertex v v v at generation L L L with path o = v 0 , … , v L = v o=v_0,\ldots,v_L=v o = v 0 , … , v L = v , the formula of Proposition 3.18 with Γ v j = − 1 / D v j \Gamma_{v_j}=-1/D_{v_j} Γ v j = − 1/ D v j and ∣ h v j − 1 v j ∣ 2 = Y v j / p |h_{v_{j-1}v_j}|^2=Y_{v_j}/p ∣ h v j − 1 v j ∣ 2 = Y v j / p gives the formal column entry
∣ Γ o ∏ j = 1 L h v j − 1 v j Γ v j ∣ s = ∣ D o ∣ − s ∏ j = 1 L ( Y v j p ) s / 2 ∣ D v j ∣ − s . \left|\Gamma_o\prod_{j=1}^{L}h_{v_{j-1}v_j}\Gamma_{v_j}\right|^s=|D_o|^{-s}\prod_{j=1}^{L}\left(\frac{Y_{v_j}}{p}\right)^{s/2}|D_{v_j}|^{-s}. Γ o j = 1 ∏ L h v j − 1 v j Γ v j s = ∣ D o ∣ − s j = 1 ∏ L ( p Y v j ) s /2 ∣ D v j ∣ − s .
Summing over v v v and applying the many-to-one formula (Proposition 4.8), the sum becomes p L p^L p L times an expectation along a single path, on which D v j − 1 = D j − 1 ′ − Y v j / ( p D v j ) D_{v_{j-1}}=D'_{j-1}-Y_{v_j}/(pD_{v_j}) D v j − 1 = D j − 1 ′ − Y v j / ( p D v j ) , where D j − 1 ′ D'_{j-1} D j − 1 ′ is the contribution of the other children of v j − 1 v_{j-1} v j − 1 ; since these children form an independent Poisson family, the cavity equation shows that D j − 1 ′ D'_{j-1} D j − 1 ′ has law π \pi π . Comparing with (6.22), the expected generation-L L L sum is Φ L ( s ; E ) \Phi_L(s; E) Φ L ( s ; E ) . This identity is stated and proved in Proposition 8.3.
The operators A s A_s A s and Q s Q_s Q s are related by the weight function h s h_s h s of Definition 6.1(v). For the identification, we use the following conjugacy. Here, for a Borel function G : R P 1 → [ 0 , ∞ ] G : \mathbb{R}\mathbb{P}^1 \to[0,\infty] G : R P 1 → [ 0 , ∞ ] , we define Q s L G ( x ) = ∫ G d κ s ( L ) ( x , ⋅ ) ∈ [ 0 , ∞ ] Q_s^L G(x) = \int G \,\mathrm{d}\kappa_s^{(L)}(x,\cdot) \in[0,\infty] Q s L G ( x ) = ∫ G d κ s ( L ) ( x , ⋅ ) ∈ [ 0 , ∞ ] ; for bounded G G G , this agrees with the operator iterate.
Lemma 6.26. Suppose that ν \nu ν satisfies (6.1), and let s ∈ ( 0 , 2 ) s \in(0,2) s ∈ ( 0 , 2 ) . For x ∈ R ∖ { 0 } x \in\mathbb{R} \setminus\{0\} x ∈ R ∖ { 0 } , b ∈ R b \in\mathbb{R} b ∈ R , and β > 0 \beta> 0 β > 0 , we have
β s / 2 ∣ x ∣ s h s ( f b , β ( x ) ) = h s ( x ) ∥ M ( b , β ) v x ∥ − s . (96) \beta^{s/2}|x|^s h_s(f_{b,\beta}(x)) = h_s(x)\lVert M(b,\beta)v_x\rVert^{-s}. \tag*{(96)} β s /2 ∣ x ∣ s h s ( f b , β ( x )) = h s ( x ) ∥ M ( b , β ) v x ∥ − s . (96) Further, for every Borel G : R P 1 → [ 0 , ∞ ] G : \mathbb{R}\mathbb{P}^1 \to[0,\infty] G : R P 1 → [ 0 , ∞ ] , every L ≥ 1 L \ge1 L ≥ 1 , and every x ∈ R ∖ { 0 } x \in\mathbb{R} \setminus\{0\} x ∈ R ∖ { 0 } ,
A s L ( h s G ) ( x ) = h s ( x ) Q s L G ( x ) . (97) A_s^L(h_sG)(x) = h_s(x)Q_s^L G(x). \tag*{(97)} A s L ( h s G ) ( x ) = h s ( x ) Q s L G ( x ) . (97)
Proof. Let w = b − β x w=b-\beta x w = b − β x . If w ≠ 0 w\ne0 w = 0 , then f b , β ( x ) = 1 / w f_{b,\beta}(x)=1/w f b , β ( x ) = 1/ w and h s ( 1 / w ) = ( 1 + w 2 ) − s / 2 h_s(1/w)=(1+w^2)^{-s/2} h s ( 1/ w ) = ( 1 + w 2 ) − s /2 , and (6.3) gives
h s ( x ) ∥ M ( b , β ) v x ∥ − s = ∣ x ∣ s ( 1 + x 2 ) s / 2 ⋅ β s / 2 ( 1 + x 2 ) s / 2 ( 1 + w 2 ) s / 2 = β s / 2 ∣ x ∣ s h s ( 1 / w ) . h_s(x)\lVert M(b,\beta)v_x\rVert^{-s}
= \frac{|x|^s}{(1+x^2)^{s/2}}\cdot\frac{\beta^{s/2}(1+x^2)^{s/2}}{(1+w^2)^{s/2}}
= \beta^{s/2}|x|^s h_s(1/w). h s ( x ) ∥ M ( b , β ) v x ∥ − s = ( 1 + x 2 ) s /2 ∣ x ∣ s ⋅ ( 1 + w 2 ) s /2 β s /2 ( 1 + x 2 ) s /2 = β s /2 ∣ x ∣ s h s ( 1/ w ) .
If w = 0 w=0 w = 0 , then f b , β ( x ) = ∞ f_{b,\beta}(x)=\infty f b , β ( x ) = ∞ , h s ( ∞ ) = 1 h_s(\infty)=1 h s ( ∞ ) = 1 , and both sides of (6.23) equal β s / 2 ∣ x ∣ s \beta^{s/2}|x|^s β s /2 ∣ x ∣ s . This proves (6.23).
We prove (6.24) by induction on L L L . For x ∈ R ∖ { 0 } x\in\mathbb{R}\setminus\{0\} x ∈ R ∖ { 0 } , the point f B , β ( x ) f_{B,\beta}(x) f B , β ( x ) lies in R ∖ { 0 } \mathbb{R}\setminus\{0\} R ∖ { 0 } almost surely (it equals ∞ \infty ∞ only on a null event, and it equals 0 0 0 only if x = ∞ x=\infty x = ∞ ). For L = 1 L=1 L = 1 , by (6.23), we have
A s ( h s G ) ( x ) = p E [ β s / 2 ∣ x ∣ s h s ( f B , β ( x ) ) G ( f B , β ( x ) ) ] = h s ( x ) p E [ ∥ M ( B , β ) v x ∥ − s G ( f B , β ( x ) ) ] = h s ( x ) Q s G ( x ) . \begin{aligned}
A_s(h_sG)(x)
= p\mathbb{E}\left[\beta^{s/2}|x|^s h_s(f_{B,\beta}(x))G(f_{B,\beta}(x))\right] \\
&= h_s(x)p\mathbb{E}\left[\lVert M(B,\beta)v_x\rVert^{-s}G(f_{B,\beta}(x))\right]
= h_s(x)Q_sG(x).
\end{aligned} A s ( h s G ) ( x ) = p E [ β s /2 ∣ x ∣ s h s ( f B , β ( x )) G ( f B , β ( x )) ] = h s ( x ) p E [ ∥ M ( B , β ) v x ∥ − s G ( f B , β ( x )) ] = h s ( x ) Q s G ( x ) .
Suppose that (6.24) holds for L L L . Since f B , β ( x ) ∈ R ∖ { 0 } f_{B,\beta}(x)\in\mathbb{R}\setminus\{0\} f B , β ( x ) ∈ R ∖ { 0 } almost surely, we may apply the induction hypothesis at this point, and (6.23) gives
A s L + 1 ( h s G ) ( x ) = p E [ β s / 2 ∣ x ∣ s h s ( f B , β ( x ) ) ( Q s L G ) ( f B , β ( x ) ) ] = h s ( x ) Q s L + 1 G ( x ) . A_s^{L+1}(h_sG)(x)
= p\mathbb{E}\left[\beta^{s/2}|x|^s h_s(f_{B,\beta}(x))(Q_s^L G)(f_{B,\beta}(x))\right]
= h_s(x)Q_s^{L+1}G(x). A s L + 1 ( h s G ) ( x ) = p E [ β s /2 ∣ x ∣ s h s ( f B , β ( x )) ( Q s L G ) ( f B , β ( x )) ] = h s ( x ) Q s L + 1 G ( x ) .
In the last step, we used (6.11) to write Q s L + 1 G ( x ) = p E [ ∥ M ( B , β ) v x ∥ − s ( Q s L G ) ( f B , β ( x ) ) ] Q_s^{L+1}G(x)=p\mathbb{E}\left[\lVert M(B,\beta)v_x\rVert^{-s}(Q_s^L G)(f_{B,\beta}(x))\right] Q s L + 1 G ( x ) = p E [ ∥ M ( B , β ) v x ∥ − s ( Q s L G ) ( f B , β ( x )) ] .
The identification of the free energy with the path moments requires a fractional moment bound for the denominator perturbed by a tilted squared modulus. For q ∈ ( 0 , 1 ) q \in(0,1) q ∈ ( 0 , 1 ) , let T q T_q T q be a random variable with density t q / 2 e − t / Γ ( 1 + q / 2 ) t^{q/2}e^{-t}/\Gamma(1+q/2) t q /2 e − t /Γ ( 1 + q /2 ) on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , independent of B B B . We consider the bound
E ∣ B − σ a T q ∣ − q ≤ C q min { 1 , a − q } for all σ ∈ { − 1 , 1 } and a ≥ 0. (98) \mathbb{E}|B-\sigma aT_q|^{-q} \le C_q\min\{1,a^{-q}\}\qquad\text{for all }\sigma\in\{-1,1\}\text{ and }a\ge0. \tag*{(98)} E ∣ B − σ a T q ∣ − q ≤ C q min { 1 , a − q } for all σ ∈ { − 1 , 1 } and a ≥ 0. (98)
By Proposition 5.5, if B = g D B=gD B = g D with D D D distributed according to a real solution at energy E > 0 E>0 E > 0 , then (6.25) holds for every q ∈ ( 0 , 1 ) q\in(0,1) q ∈ ( 0 , 1 ) , with a constant C q C_q C q depending on p , E , g p,E,g p , E , g , and q q q .
Proposition 6.27. Let s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) . Suppose that ν \nu ν satisfies (6.1), and that there exist q ∈ ( s , 1 ) q\in(s,1) q ∈ ( s , 1 ) and C q < ∞ C_q<\infty C q < ∞ such that (6.25) holds. Let ν x \nu^x ν x be the law of 1 / B 1/B 1/ B .
(i) There exists C ≥ 1 C\ge1 C ≥ 1 , depending on s , p , K , ν , q s,p,K,\nu,q s , p , K , ν , q , and C q C_q C q , such that for every L ≥ 1 L\ge1 L ≥ 1 ,
C − 1 λ ( s ) L ≤ ∫ A s L e s d ν x ≤ C λ ( s ) L . C^{-1}\lambda(s)^L\le\int\mathcal{A}_s^L e_s\,\mathrm{d}\nu^x\le C\lambda(s)^L. C − 1 λ ( s ) L ≤ ∫ A s L e s d ν x ≤ C λ ( s ) L . In particular, ∫ A s L e s d ν x \int\mathcal{A}_s^L e_s\,\mathrm{d}\nu^x ∫ A s L e s d ν x is finite for every L L L , and lim L → ∞ L − 1 log ∫ A s L e s d ν x = φ ( s ) \lim_{L\to\infty}L^{-1}\log\int\mathcal{A}_s^L e_s\,\mathrm{d}\nu^x=\varphi(s) lim L → ∞ L − 1 log ∫ A s L e s d ν x = φ ( s ) .
(ii) We have E ∣ B ∣ − s < ∞ \mathbb{E}|B|^{-s}<\infty E ∣ B ∣ − s < ∞ , and the norm moments satisfy
C ∗ − 1 e ( L + 3 ) ( φ ( s ) − log p ) ≤ f L ( s ) ≤ e L ( φ ( s ) − log p ) , L ≥ 1 , C_*^{-1}e^{(L+3)(\varphi(s)-\log p)}\le f_L(s)\le e^{L(\varphi(s)-\log p)},\qquad L\ge1, C ∗ − 1 e ( L + 3 ) ( φ ( s ) − l o g p ) ≤ f L ( s ) ≤ e L ( φ ( s ) − l o g p ) , L ≥ 1 , where C ∗ C_* C ∗ is as in Corollary 6.22. In particular, lim L → ∞ L − 1 log f L ( s ) = φ ( s ) − log p \lim_{L\to\infty}L^{-1}\log f_L(s)=\varphi(s)-\log p lim L → ∞ L − 1 log f L ( s ) = φ ( s ) − log p .
(iii) Let π \pi π be a real solution at energy E > 0 E>0 E > 0 . Then the hypotheses above hold for ν g \nu_g ν g and K g = p / g 2 K_g=p/g^2 K g = p / g 2 , for every g > 0 g>0 g > 0 and every s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) . In particular, there exists C ≥ 1 C\ge1 C ≥ 1 , depending on s , p , E s,p,E s , p , E , and π \pi π , such that
C − 1 e L φ π ( s ) ≤ Φ L ( s ; E ) ≤ C e L φ π ( s ) , L ≥ 1 , C^{-1}e^{L\varphi_\pi(s)}\le\Phi_L(s;E)\le Ce^{L\varphi_\pi(s)},\qquad L\ge1, C − 1 e L φ π ( s ) ≤ Φ L ( s ; E ) ≤ C e L φ π ( s ) , L ≥ 1 , and therefore
lim L → ∞ 1 L log Φ L ( s ; E ) = φ π ( s ) . \lim_{L\to\infty}\frac{1}{L}\log\Phi_L(s;E)=\varphi_\pi(s). L → ∞ lim L 1 log Φ L ( s ; E ) = φ π ( s ) . Further, φ π ( 1 ) = lim s ↑ 1 lim L → ∞ L − 1 log Φ L ( s ; E ) \varphi_\pi(1)=\lim_{s\uparrow1}\lim_{L\to\infty}L^{-1}\log\Phi_L(s;E) φ π ( 1 ) = lim s ↑ 1 lim L → ∞ L − 1 log Φ L ( s ; E ) , and the bounds in (ii) hold for the norm moments of π \pi π with every scaling.
Proof. Since the weight function h s h_s h s vanishes at 0, the initial function e s = ∣ x ∣ s e_s=|x|^s e s = ∣ x ∣ s of the path moment corresponds under (6.24) to the function u s ( y ) = ( 1 + y 2 ) s / 2 u_s(y)=(1+y^2)^{s/2} u s ( y ) = ( 1 + y 2 ) s /2 (with u s ( ∞ ) = ∞ u_s(\infty)=\infty u s ( ∞ ) = ∞ ), which is unbounded on R P 1 \mathbb{R}P^1 R P 1 . We show that one application of Q s Q_s Q s turns u s u_s u s into a bounded continuous positive function, using (6.25); the uniform integrability needed for continuity comes from (6.25) at the larger exponent q q q . After this, the bounds follow from Theorem 6.11.
Step 1. A tilted moment bound. For r ∈ ( 0 , 1 ) r \in(0,1) r ∈ ( 0 , 1 ) , we have y r / 2 e − y d y = Γ ( 1 + r / 2 ) P ( T r ∈ d y ) y^{r/2}e^{-y}\,dy=\Gamma(1+r/2)\mathbb{P}(T_r\in dy) y r /2 e − y d y = Γ ( 1 + r /2 ) P ( T r ∈ d y ) . Writing β = Y / K \beta=Y/K β = Y / K with Y ∼ Exp ( 1 ) Y\sim\operatorname{Exp}(1) Y ∼ Exp ( 1 ) independent of B B B , we therefore have, for x ∈ R x\in\mathbb{R} x ∈ R ,
E [ β r / 2 ∣ B − β x ∣ − r ] = K − r / 2 E [ Y r / 2 ∣ B − x K Y ∣ − r ] = K − r / 2 Γ ( 1 + r / 2 ) E [ ∣ B − σ ∣ x ∣ K T r ∣ − r ] . (99) \mathbb{E}\left[\beta^{r/2}\lvert B-\beta x\rvert^{-r}\right]=K^{-r/2}\mathbb{E}\left[Y^{r/2}\left\lvert B-\frac{x}{K}Y\right\rvert^{-r}\right]=K^{-r/2}\Gamma(1+r/2)\mathbb{E}\left[\left\lvert B-\sigma\frac{\lvert x\rvert}{K}T_r\right\rvert^{-r}\right]. \tag*{(99)} E [ β r /2 ∣ B − β x ∣ − r ] = K − r /2 E [ Y r /2 B − K x Y − r ] = K − r /2 Γ ( 1 + r /2 ) E [ B − σ K ∣ x ∣ T r − r ] . (99)
where σ = sgn x \sigma=\operatorname{sgn}x σ = sgn x (and σ = 1 \sigma=1 σ = 1 if x = 0 x=0 x = 0 ). Applying (99) with r = q r=q r = q and (6.25), we obtain
( 1 + x 2 ) q / 2 E [ β q / 2 ∣ B − β x ∣ − q ] ≤ K − q / 2 Γ ( 1 + q / 2 ) C q ( 1 + x 2 ) q / 2 min { 1 , K q ∣ x ∣ − q } . (1+x^2)^{q/2}\mathbb{E}\left[\beta^{q/2}\lvert B-\beta x\rvert^{-q}\right]\le K^{-q/2}\Gamma(1+q/2)C_q(1+x^2)^{q/2}\min\{1,K^q\lvert x\rvert^{-q}\}. ( 1 + x 2 ) q /2 E [ β q /2 ∣ B − β x ∣ − q ] ≤ K − q /2 Γ ( 1 + q /2 ) C q ( 1 + x 2 ) q /2 min { 1 , K q ∣ x ∣ − q } .
The right side is at most ( 1 + K 2 ) q / 2 (1+K^2)^{q/2} ( 1 + K 2 ) q /2 times the prefactor, uniformly in x ∈ R x\in\mathbb{R} x ∈ R : for ∣ x ∣ ≤ K \lvert x\rvert\le K ∣ x ∣ ≤ K this is immediate, and for ∣ x ∣ ≥ K \lvert x\rvert\ge K ∣ x ∣ ≥ K we have ( 1 + x 2 ) q / 2 K q ∣ x ∣ − q = K q ( 1 + x − 2 ) q / 2 ≤ ( K 2 + 1 ) q / 2 (1+x^2)^{q/2}K^q\lvert x\rvert^{-q}=K^q(1+x^{-2})^{q/2}\le(K^2+1)^{q/2} ( 1 + x 2 ) q /2 K q ∣ x ∣ − q = K q ( 1 + x − 2 ) q /2 ≤ ( K 2 + 1 ) q /2 . Then there exists C 1 < ∞ C_1<\infty C 1 < ∞ , depending on q q q , K K K , and C q C_q C q , with
sup x ∈ R ( 1 + x 2 ) q / 2 E [ β q / 2 ∣ B − β x ∣ − q ] ≤ C 1 . (100) \sup_{x\in\mathbb{R}}(1+x^2)^{q/2}\mathbb{E}\left[\beta^{q/2}\lvert B-\beta x\rvert^{-q}\right]\le C_1. \tag*{(100)} x ∈ R sup ( 1 + x 2 ) q /2 E [ β q /2 ∣ B − β x ∣ − q ] ≤ C 1 . (100)
Step 2. The function Q s u s Q_su_s Q s u s is continuous and positive on R P 1 \mathbb{R}\mathbb{P}^1 R P 1 . Let x ∈ R x\in\mathbb{R} x ∈ R . Almost surely w = B − β x ≠ 0 w=B-\beta x\ne0 w = B − β x = 0 , and then f B , β ( x ) = 1 / w f_{B,\beta}(x)=1/w f B , β ( x ) = 1/ w and u s ( 1 / w ) = ( 1 + w − 2 ) s / 2 = ( 1 + w 2 ) s / 2 ∣ w ∣ − s u_s(1/w)=(1+w^{-2})^{s/2}=(1+w^2)^{s/2}\lvert w\rvert^{-s} u s ( 1/ w ) = ( 1 + w − 2 ) s /2 = ( 1 + w 2 ) s /2 ∣ w ∣ − s . By (6.3), we have
Q s u s ( x ) = p E Z x , Z x = β s / 2 ( 1 + x 2 ) s / 2 ∣ B − β x ∣ − s . Q_su_s(x)=p\,\mathbb{E}Z_x,\qquad Z_x=\beta^{s/2}(1+x^2)^{s/2}\lvert B-\beta x\rvert^{-s}. Q s u s ( x ) = p E Z x , Z x = β s /2 ( 1 + x 2 ) s /2 ∣ B − β x ∣ − s .
At ∞ \infty ∞ , we have f B , β ( ∞ ) = 0 f_{B,\beta}(\infty)=0 f B , β ( ∞ ) = 0 and u s ( 0 ) = 1 u_s(0)=1 u s ( 0 ) = 1 , which give Q s u s ( ∞ ) = p E β − s / 2 = p K s / 2 Γ ( 1 − s / 2 ) Q_su_s(\infty)=p\,\mathbb{E}\beta^{-s/2}=pK^{s/2}\Gamma(1-s/2) Q s u s ( ∞ ) = p E β − s /2 = p K s /2 Γ ( 1 − s /2 ) . Since Z x q / s = β q / 2 ( 1 + x 2 ) q / 2 ∣ B − β x ∣ − q Z_x^{q/s}=\beta^{q/2}(1+x^2)^{q/2}\lvert B-\beta x\rvert^{-q} Z x q / s = β q /2 ( 1 + x 2 ) q /2 ∣ B − β x ∣ − q and q / s > 1 q/s>1 q / s > 1 , (6.27) implies that the family { Z x : x ∈ R } \{Z_x:x\in\mathbb{R}\} { Z x : x ∈ R } is bounded in L q / s L^{q/s} L q / s ; in particular, it is uniformly integrable and bounded in L 1 L^1 L 1 . In particular, sup x ∈ R Q s u s ( x ) < ∞ \sup_{x\in\mathbb{R}}Q_su_s(x)<\infty sup x ∈ R Q s u s ( x ) < ∞ .
Let x n → x 0 ∈ R x_n\to x_0\in\mathbb{R} x n → x 0 ∈ R . The event { B ≠ β x 0 } \{B\ne\beta x_0\} { B = β x 0 } has probability one: for x 0 = 0 x_0=0 x 0 = 0 because P ( B = 0 ) = 0 \mathbb{P}(B=0)=0 P ( B = 0 ) = 0 , and for x 0 ≠ 0 x_0\ne0 x 0 = 0 because β \beta β has a density and is independent of B B B . On this event, Z x n → Z x 0 Z_{x_n}\to Z_{x_0} Z x n → Z x 0 , and by Vitali’s convergence theorem, we have Q s u s ( x n ) → Q s u s ( x 0 ) Q_su_s(x_n)\to Q_su_s(x_0) Q s u s ( x n ) → Q s u s ( x 0 ) . Now let x n ∈ R x_n\in\mathbb{R} x n ∈ R with ∣ x n ∣ → ∞ \lvert x_n\rvert\to\infty ∣ x n ∣ → ∞ . Since ∣ B − β x n ∣ ( 1 + x n 2 ) − 1 / 2 → β \lvert B-\beta x_n\rvert(1+x_n^2)^{-1/2}\to\beta ∣ B − β x n ∣ ( 1 + x n 2 ) − 1/2 → β almost surely, we have Z x n → β s / 2 β − s = β − s / 2 Z_{x_n}\to\beta^{s/2}\beta^{-s}=\beta^{-s/2} Z x n → β s /2 β − s = β − s /2 almost surely, and Vitali’s theorem gives Q s u s ( x n ) → p E β − s / 2 = Q s u s ( ∞ ) Q_su_s(x_n)\to p\,\mathbb{E}\beta^{-s/2}=Q_su_s(\infty) Q s u s ( x n ) → p E β − s /2 = Q s u s ( ∞ ) . This proves that Q s u s ∈ C ( R P 1 ) Q_su_s\in C(\mathbb{R}\mathbb{P}^1) Q s u s ∈ C ( R P 1 ) . Since Z x > 0 Z_x>0 Z x > 0 almost surely for every x ∈ R x\in\mathbb{R} x ∈ R and E β − s / 2 > 0 \mathbb{E}\beta^{-s/2}>0 E β − s /2 > 0 , the function Q s u s Q_su_s Q s u s is strictly positive, and by compactness c 1 = min Q s u s > 0 c_1=\min Q_su_s>0 c 1 = min Q s u s > 0 . Let c 2 = max Q s u s c_2=\max Q_su_s c 2 = max Q s u s .
Step 3. Proof of (i). Since e s = h s u s e_s=h_su_s e s = h s u s on R \mathbb{R} R , Lemma 6.26 with G = u s G=u_s G = u s implies A s L e s ( x ) = h s ( x ) Q s L u s ( x ) A_s^Le_s(x)=h_s(x)Q_s^Lu_s(x) A s L e s ( x ) = h s ( x ) Q s L u s ( x ) for x ∈ R ∖ { 0 } x\in\mathbb{R}\setminus\{0\} x ∈ R ∖ { 0 } . By the Chapman–Kolmogorov relations and Tonelli’s theorem, Q s L u s = Q s L − 1 ( Q s u s ) Q_s^Lu_s=Q_s^{L-1}(Q_su_s) Q s L u s = Q s L − 1 ( Q s u s ) , where on the right Q s L − 1 Q_s^{L-1} Q s L − 1 acts on the continuous function Q s u s Q_su_s Q s u s . The law ν x \nu^x ν x is concentrated on R ∖ { 0 } \mathbb{R}\setminus\{0\} R ∖ { 0 } , because B B B is finite and P ( B = 0 ) = 0 \mathbb{P}(B=0)=0 P ( B = 0 ) = 0 . We conclude that
∫ A s L e s d ν x = ∫ h s ( x ) Q s L − 1 ( Q s u s ) ( x ) ν x ( d x ) , L ≥ 1. (101) \int A_s^Le_s\,d\nu^x=\int h_s(x)Q_s^{L-1}(Q_su_s)(x)\,\nu^x(dx),\qquad L\ge1. \tag*{(101)} ∫ A s L e s d ν x = ∫ h s ( x ) Q s L − 1 ( Q s u s ) ( x ) ν x ( d x ) , L ≥ 1. (101)
Let ψ s \psi_s ψ s be the eigenfunction of Theorem 6.11 (ii), and let a = min ψ s a=\min\psi_s a = min ψ s and A = max ψ s A=\max\psi_s A = max ψ s . As in Step 6 of the proof of Theorem 6.11 , positivity of Q s Q_s Q s and c 1 ≤ Q s u s ≤ c 2 c_1\le Q_su_s\le c_2 c 1 ≤ Q s u s ≤ c 2 give
c 1 a A λ ( s ) L − 1 ≤ Q s L − 1 ( Q s u s ) ≤ c 2 A a λ ( s ) L − 1 \frac{c_1a}{A}\lambda(s)^{L-1}\le Q_s^{L-1}(Q_su_s)\le\frac{c_2A}{a}\lambda(s)^{L-1} A c 1 a λ ( s ) L − 1 ≤ Q s L − 1 ( Q s u s ) ≤ a c 2 A λ ( s ) L − 1
pointwise on R P 1 \mathbb{R}\mathbb{P}^{1} R P 1 . Further, since 0 < h s ≤ 1 0<h_s\le1 0 < h s ≤ 1 on R ∖ { 0 } \mathbb{R}\setminus\{0\} R ∖ { 0 } , we have 0 < ∫ h s d ν x ≤ 1 0<\int h_s\,d\nu^x\le1 0 < ∫ h s d ν x ≤ 1 . Inserting these bounds into (6.28) proves (i) with C = max { c 2 A / ( a λ ( s ) ) , A λ ( s ) / ( c 1 a ∫ h s d ν x ) } C=\max\{c_2A/(a\lambda(s)), A\lambda(s)/(c_1a\int h_s\,d\nu^x)\} C = max { c 2 A / ( aλ ( s )) , A λ ( s ) / ( c 1 a ∫ h s d ν x )} , enlarged to be at least one. The limit statement follows by taking logarithms.
Step 4. Proof of (ii). Taking a = 0 a=0 a = 0 in (6.25), we obtain E ∣ B ∣ − q ≤ C q \mathbb{E}|B|^{-q}\le C_q E ∣ B ∣ − q ≤ C q and E ∣ B ∣ − s ≤ 1 + C q < ∞ \mathbb{E}|B|^{-s}\le1+C_q<\infty E ∣ B ∣ − s ≤ 1 + C q < ∞ . Then Corollary 6.22 applies, and we have the bounds with ϱ s = e φ ( s ) − log p \varrho_s=e^{\varphi(s)-\log p} ϱ s = e φ ( s ) − l o g p .
Step 5. Proof of (iii). By Lemma 6.13(i) and Proposition 5.5, the law ν g \nu_g ν g satisfies (6.1) and (6.25) for every q ∈ ( 0 , 1 ) q\in(0,1) q ∈ ( 0 , 1 ) , and the hypotheses hold for every s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) , with any q ∈ ( s , 1 ) q\in(s,1) q ∈ ( s , 1 ) . By Definition 6.24, Φ L ( s ; E ) = g s ∫ A s ( g ) L e s d π g x \Phi_L(s;E)=g^s\int A_s^{(g)}\mathcal{L}_{e_s}\,d\pi_g^x Φ L ( s ; E ) = g s ∫ A s ( g ) L e s d π g x and π g x \pi_g^x π g x is the law of 1 / B 1/B 1/ B for B = g D B=gD B = g D . Part (i) and λ ( s ) = e φ π ( s ) \lambda(s)=e^{\varphi_\pi(s)} λ ( s ) = e φ π ( s ) imply the bounds on Φ L ( s ; E ) \Phi_L(s;E) Φ L ( s ; E ) , after multiplying C C C by max ( g δ , g − δ ) \max(g^\delta,g^{-\delta}) max ( g δ , g − δ ) . Since φ π \varphi_\pi φ π is continuous at 1 1 1 by Proposition 6.15(i), we have φ π ( 1 ) = lim s ↑ 1 φ π ( s ) \varphi_\pi(1)=\lim_{s\uparrow1}\varphi_\pi(s) φ π ( 1 ) = lim s ↑ 1 φ π ( s ) , which implies the formula for φ π ( 1 ) \varphi_\pi(1) φ π ( 1 ) . The last statement is (ii) applied with each scaling.
Spectral type from the sign of the free energy
Rank-one perturbations of the squared row norms
In this section, we prove two criteria for the spectral type of H \mathcal{H} H on a set of energies of the form I ∪ ( − I ) I \cup(-I) I ∪ ( − I ) , where I ⊂ ( 0 , ∞ ) I \subset(0,\infty) I ⊂ ( 0 , ∞ ) . Both criteria rest on the bipartite structure recorded in Proposition 3.23. The square block A ∗ A A^*A A ∗ A is a sum of the rank-one terms U y ∣ v y ⟩ ⟨ v y ∣ U_y|v_y\rangle\langle v_y| U y ∣ v y ⟩ ⟨ v y ∣ over the rows y ∈ V 1 y \in V_1 y ∈ V 1 , and, conditional on all other variables, the squared row norm U y U_y U y has a Lebesgue density. Removing the term of one row leaves an operator K 0 ( y ) K_0^{(y)} K 0 ( y ) that does not depend on U y U_y U y , and the spectral measure of v y v_y v y for A ∗ A A^*A A ∗ A is the member with coupling u = U y u = U_y u = U y of the rank-one family K 0 ( y ) + u ∣ v y ⟩ ⟨ v y ∣ K_0^{(y)} + u|v_y\rangle\langle v_y| K 0 ( y ) + u ∣ v y ⟩ ⟨ v y ∣ . For Lebesgue-almost every coupling, the spectral type of such a family is determined by the boundary values of the Borel transform of the spectral measure of v y v_y v y for K 0 ( y ) K_0^{(y)} K 0 ( y ) . This is the mechanism behind the Simon–Wolff criterion [124] and spectral averaging; see [123] and [13] for expositions.
The first criterion, Proposition 7.9, deduces pure point spectrum from the finiteness of reciprocal-square integrals of spectral measures of the vertex-deleted operators; it is used in § 8.2 . The second, Proposition 7.11 , deduces purely absolutely continuous spectrum from positivity, with positive probability, of the imaginary part of the boundary value of the root cavity resolvent; it is used in Section 10.
§ 8.1.1 collects facts about Borel transforms and rank-one perturbations, the main one being Lemma 7.3 . § 8.1.2 records the measurability of resolvent entries of the tree operators. § 8.1.3 relates the row-deleted square blocks to the vertex-deleted operators. § 8.1.4 and § 8.1.5 prove Proposition 7.9 and Proposition 7.11 , respectively.
Throughout this section, p > 1 p > 1 p > 1 is fixed. For a self-adjoint operator K K K on a separable Hilbert space h \mathfrak{h} h and a vector φ ∈ h \varphi\in\mathfrak{h} φ ∈ h , we write μ φ K \mu_\varphi^K μ φ K for the spectral measure of φ \varphi φ , so that μ φ K ( B ) = ⟨ φ , 1 B ( K ) φ ⟩ \mu_\varphi^K(B) = \langle\varphi,1_B(K)\varphi\rangle μ φ K ( B ) = ⟨ φ , 1 B ( K ) φ ⟩ for Borel sets B ⊂ R B \subset\mathbb{R} B ⊂ R . We let h p p ( K ) \mathfrak{h}_{\mathrm{pp}}(K) h pp ( K ) and h a c ( K ) \mathfrak{h}_{\mathrm{ac}}(K) h ac ( K ) denote the closed subspaces of vectors whose spectral measures are pure point and absolutely continuous, respectively. For a Borel set B ⊂ R B \subset\mathbb{R} B ⊂ R , we say that K K K has pure point spectrum in B B B if Ran 1 B ( K ) ⊂ h p p ( K ) \operatorname{Ran}1_B(K) \subset\mathfrak{h}_{\mathrm{pp}}(K) Ran 1 B ( K ) ⊂ h pp ( K ) , and that K K K is purely absolutely continuous in B B B if Ran 1 B ( K ) ⊂ h a c ( K ) \operatorname{Ran}1_B(K) \subset\mathfrak{h}_{\mathrm{ac}}(K) Ran 1 B ( K ) ⊂ h ac ( K ) . Since μ 1 B ( K ) φ K \mu_{1_B(K)\varphi}^K μ 1 B ( K ) φ K is the restriction of μ φ K \mu_\varphi^K μ φ K to B B B , the first property holds if and only if the restriction of μ φ K \mu_\varphi^K μ φ K to B B B is pure point for every φ ∈ h \varphi\in\mathfrak{h} φ ∈ h , and similarly for the second. We write ∣ B ∣ |B| ∣ B ∣ for the Lebesgue measure of B B B , and B 2 = { t 2 : t ∈ B } B^2 = \{t^2 : t \in B\} B 2 = { t 2 : t ∈ B } .
Borel transforms and rank-one perturbations
For a finite positive Borel measure μ \mu μ on R \mathbb{R} R , its Borel transform is
F μ ( z ) = ∫ R μ ( d t ) t − z , z ∈ H . F_\mu(z)=\int_{\mathbb{R}}\frac{\mu(\mathrm{d}t)}{t-z},\qquad z\in\mathbb{H}. F μ ( z ) = ∫ R t − z μ ( d t ) , z ∈ H .
It is holomorphic on H \mathbb{H} H , and for λ ∈ R \lambda\in\mathbb{R} λ ∈ R and η > 0 \eta>0 η > 0 we have
Im F μ ( λ + i η ) = ∫ R η μ ( d t ) ( t − λ ) 2 + η 2 . (102) \operatorname{Im}F_\mu(\lambda+\mathrm{i}\eta)=\int_{\mathbb{R}}\frac{\eta\mu(\mathrm{d}t)}{(t-\lambda)^2+\eta^2}. \tag*{(102)} Im F μ ( λ + i η ) = ∫ R ( t − λ ) 2 + η 2 η μ ( d t ) . (102)
If μ \mu μ is not the zero measure, then Im F μ > 0 \operatorname{Im}F_\mu>0 Im F μ > 0 on H \mathbb{H} H . We say that the vertical boundary value F μ ( λ + i 0 ) F_\mu(\lambda+\mathrm{i}0) F μ ( λ + i 0 ) exists if lim η ↓ 0 F μ ( λ + i η ) \lim_{\eta\downarrow0}F_\mu(\lambda+\mathrm{i}\eta) lim η ↓ 0 F μ ( λ + i η ) exists in C \mathbb{C} C ; in that case F μ ( λ + i 0 ) F_\mu(\lambda+\mathrm{i}0) F μ ( λ + i 0 ) denotes the limit. We also define the reciprocal-square integral
R μ ( λ ) = ∫ R μ ( d t ) ( t − λ ) 2 ∈ [ 0 , ∞ ] . (103) R_\mu(\lambda)=\int_{\mathbb{R}}\frac{\mu(\mathrm{d}t)}{(t-\lambda)^2}\in[0,\infty]. \tag*{(103)} R μ ( λ ) = ∫ R ( t − λ ) 2 μ ( d t ) ∈ [ 0 , ∞ ] . (103)
with the convention that the integrand equals + ∞ +\infty + ∞ at t = λ t=\lambda t = λ , so that R μ ( λ ) = ∞ R_\mu(\lambda)=\infty R μ ( λ ) = ∞ whenever μ ( { λ } ) > 0 \mu(\{\lambda\})>0 μ ({ λ }) > 0 . By (102) , the function η ↦ η − 1 Im F μ ( λ + i η ) \eta\mapsto\eta^{-1}\operatorname{Im}F_\mu(\lambda+\mathrm{i}\eta) η ↦ η − 1 Im F μ ( λ + i η ) is the integral of ( ( t − λ ) 2 + η 2 ) − 1 ((t-\lambda)^2+\eta^2)^{-1} (( t − λ ) 2 + η 2 ) − 1 against μ \mu μ , which increases as η ↓ 0 \eta\downarrow0 η ↓ 0 . By the monotone convergence theorem, we have
R μ ( λ ) = lim η ↓ 0 Im F μ ( λ + i η ) η = sup n ∈ N n Im F μ ( λ + i / n ) . (104) R_\mu(\lambda)=\lim_{\eta\downarrow0}\frac{\operatorname{Im}F_\mu(\lambda+\mathrm{i}\eta)}{\eta}=\sup_{n\in\mathbb{N}}n\operatorname{Im}F_\mu(\lambda+\mathrm{i}/n). \tag*{(104)} R μ ( λ ) = η ↓ 0 lim η Im F μ ( λ + i η ) = n ∈ N sup n Im F μ ( λ + i / n ) . (104)
Lemma 7.1. Let μ \mu μ be a finite positive Borel measure on R \mathbb{R} R , and let μ = μ a c + μ s i n g \mu=\mu_{\mathrm{ac}}+\mu_{\mathrm{sing}} μ = μ ac + μ sing be its decomposition into parts that are absolutely continuous and singular with respect to Lebesgue measure.
(i) For Lebesgue-almost every λ ∈ R \lambda\in\mathbb{R} λ ∈ R , the vertical boundary value F μ ( λ + i 0 ) F_\mu(\lambda+\mathrm{i}0) F μ ( λ + i 0 ) exists.
(ii) The measure μ s i n g \mu_{\mathrm{sing}} μ sing is carried by the set of λ \lambda λ such that lim η ↓ 0 Im F μ ( λ + i η ) = ∞ \lim_{\eta\downarrow0}\operatorname{Im}F_\mu(\lambda+\mathrm{i}\eta)=\infty lim η ↓ 0 Im F μ ( λ + i η ) = ∞ . In particular, μ s i n g ( R ∖ N μ ) = 0 \mu_{\mathrm{sing}}(\mathbb{R}\setminus N_\mu)=0 μ sing ( R ∖ N μ ) = 0 , where N μ N_\mu N μ is the set of λ \lambda λ at which F μ ( λ + i 0 ) F_\mu(\lambda+\mathrm{i}0) F μ ( λ + i 0 ) does not exist.
(iii) We have μ ( { λ ∈ R : lim η ↓ 0 Im F μ ( λ + i η ) = 0 } ) = 0 \mu(\{\lambda\in\mathbb{R}:\lim_{\eta\downarrow0}\operatorname{Im}F_\mu(\lambda+\mathrm{i}\eta)=0\})=0 μ ({ λ ∈ R : lim η ↓ 0 Im F μ ( λ + i η ) = 0 }) = 0 .
(iv) If R μ ( λ ) < ∞ R_\mu(\lambda)<\infty R μ ( λ ) < ∞ , then μ ( { λ } ) = 0 \mu(\{\lambda\})=0 μ ({ λ }) = 0 , the boundary value F μ ( λ + i 0 ) F_\mu(\lambda+\mathrm{i}0) F μ ( λ + i 0 ) exists, and it is real and equal to ∫ R ( t − λ ) − 1 μ ( d t ) \int_{\mathbb{R}}(t-\lambda)^{-1}\mu(\mathrm{d}t) ∫ R ( t − λ ) − 1 μ ( d t ) .
Proof. Part (i) is the classical theorem on the existence of finite boundary values of Borel transforms of finite measures; see [123] and [13]. (One proof applies Fatou’s theorem on boundary values of bounded holomorphic functions to ( F μ − i ) / ( F μ + i ) (F_\mu-\mathrm{i})/(F_\mu+\mathrm{i}) ( F μ − i ) / ( F μ + i ) ; see [118, Chapter 11].)
For (ii), we first note that if ∣ t − λ ∣ < η |t-\lambda|<\eta ∣ t − λ ∣ < η , then ( t − λ ) 2 + η 2 ≤ 2 η 2 (t-\lambda)^2+\eta^2\leq2\eta^2 ( t − λ ) 2 + η 2 ≤ 2 η 2 . Restricting the integral in (102) to such t t t , we obtain the Poisson-kernel lower bound
Im F μ ( λ + i η ) ≥ μ ( ( λ − η , λ + η ) ) 2 η . (105) \operatorname{Im}F_\mu(\lambda+\mathrm{i}\eta)\geq\frac{\mu((\lambda-\eta,\lambda+\eta))}{2\eta}. \tag*{(105)} Im F μ ( λ + i η ) ≥ 2 η μ (( λ − η , λ + η )) . (105)
By the differentiation theorem for positive measures that are singular with respect to Lebesgue measure [118, Chapter 7], for μ s i n g \mu_{\mathrm{sing}} μ sing -almost every λ \lambda λ we have
lim η ↓ 0 μ s i n g ( ( λ − η , λ + η ) ) 2 η = ∞ . \lim_{\eta\downarrow0}\frac{\mu_{\mathrm{sing}}((\lambda-\eta,\lambda+\eta))}{2\eta}=\infty. η ↓ 0 lim 2 η μ sing (( λ − η , λ + η )) = ∞.
Since μ ≥ μ s i n g \mu\geq\mu_{\mathrm{sing}} μ ≥ μ sing , (7.4) implies that Im F μ ( λ + i η ) → ∞ \operatorname{Im}F_\mu(\lambda+i\eta)\to\infty Im F μ ( λ + i η ) → ∞ for μ s i n g \mu_{\mathrm{sing}} μ sing -almost every λ \lambda λ . At such λ \lambda λ the boundary value F μ ( λ + i 0 ) F_\mu(\lambda+i0) F μ ( λ + i 0 ) does not exist, which proves (ii).
For (iii), let A A A be the set in (iii); it is a Borel set, because the limit of the continuous function η ↦ Im F μ ( λ + i η ) \eta\mapsto\operatorname{Im}F_\mu(\lambda+i\eta) η ↦ Im F μ ( λ + i η ) may be taken along rational η \eta η . By (ii), we have μ s i n g ( A ) = 0 \mu_{\mathrm{sing}}(A)=0 μ sing ( A ) = 0 . For λ ∈ A \lambda\in A λ ∈ A , the bound (7.4) implies that lim η ↓ 0 μ ( ( λ − η , λ + η ) ) / ( 2 η ) = 0 \lim_{\eta\downarrow0}\mu((\lambda-\eta,\lambda+\eta))/(2\eta)=0 lim η ↓ 0 μ (( λ − η , λ + η )) / ( 2 η ) = 0 . By the differentiation theorem for measures [118, Chapter 7], at Lebesgue-almost every λ \lambda λ this limit exists and equals the density f = d μ a c / d λ f=d\mu_{\mathrm{ac}}/d\lambda f = d μ ac / d λ ; the singular part contributes zero Lebesgue-almost everywhere. We conclude that f = 0 f=0 f = 0 almost everywhere on A A A , and μ a c ( A ) = ∫ A f ( λ ) d λ = 0 \mu_{\mathrm{ac}}(A)=\int_A f(\lambda)\,\mathrm{d}\lambda=0 μ ac ( A ) = ∫ A f ( λ ) d λ = 0 .
For (iv), the convention in (7.2) shows that μ ( { λ } ) = 0 \mu(\{\lambda\})=0 μ ({ λ }) = 0 . For t ≠ λ t\neq\lambda t = λ and η > 0 \eta>0 η > 0 , we have ∣ t − λ − i η ∣ − 1 ≤ ∣ t − λ ∣ − 1 |t-\lambda-i\eta|^{-1}\leq|t-\lambda|^{-1} ∣ t − λ − i η ∣ − 1 ≤ ∣ t − λ ∣ − 1 , and by the Cauchy–Schwarz inequality,
∫ R μ ( d t ) ∣ t − λ ∣ ≤ μ ( R ) 1 / 2 R μ ( λ ) 1 / 2 < ∞ . \int_{\mathbb{R}}\frac{\mu(\mathrm{d}t)}{|t-\lambda|}\leq\mu(\mathbb{R})^{1/2}R_\mu(\lambda)^{1/2}<\infty. ∫ R ∣ t − λ ∣ μ ( d t ) ≤ μ ( R ) 1/2 R μ ( λ ) 1/2 < ∞.
By the dominated convergence theorem, we have F μ ( λ + i η ) → ∫ R ( t − λ ) − 1 μ ( d t ) ∈ R F_\mu(\lambda+i\eta)\to\int_{\mathbb{R}}(t-\lambda)^{-1}\mu(\mathrm{d}t)\in\mathbb{R} F μ ( λ + i η ) → ∫ R ( t − λ ) − 1 μ ( d t ) ∈ R as η ↓ 0 \eta\downarrow0 η ↓ 0 .
The next lemma compares vertical boundary values at E 2 E^2 E 2 with limits along the curve η ↦ ( E + i η ) 2 \eta\mapsto(E+i\eta)^2 η ↦ ( E + i η ) 2 , which is the curve that appears when the resolvent of H \mathcal{H} H at E + i η E+i\eta E + i η is expressed through resolvents of H 2 \mathcal{H}^2 H 2 .
Lemma 7.2. Let μ \mu μ be a finite positive Borel measure on R \mathbb{R} R and let E > 0 E>0 E > 0 . Then lim η ↓ 0 F μ ( ( E + i η ) 2 ) \lim_{\eta\downarrow0}F_\mu((E+i\eta)^2) lim η ↓ 0 F μ (( E + i η ) 2 ) exists in C \mathbb{C} C if and only if F μ ( E 2 + i 0 ) F_\mu(E^2+i0) F μ ( E 2 + i 0 ) exists, and in that case the two limits coincide.
Proof. For η > 0 \eta>0 η > 0 , set w η = ( E + i η ) 2 = E 2 − η 2 + 2 i E η w_\eta=(E+i\eta)^2=E^2-\eta^2+2iE\eta w η = ( E + i η ) 2 = E 2 − η 2 + 2 i E η and w η ′ = E 2 + 2 i E η w'_\eta=E^2+2iE\eta w η ′ = E 2 + 2 i E η . Both lie in H \mathbb{H} H , they have the same imaginary part 2 E η 2E\eta 2 E η , and ∣ w η − w η ′ ∣ = η 2 |w_\eta-w'_\eta|=\eta^2 ∣ w η − w η ′ ∣ = η 2 . For w ∈ H w\in\mathbb{H} w ∈ H , we have ∫ R ∣ t − w ∣ − 2 μ ( d t ) = Im F μ ( w ) / Im w \int_{\mathbb{R}}|t-w|^{-2}\mu(\mathrm{d}t)=\operatorname{Im}F_\mu(w)/\operatorname{Im}w ∫ R ∣ t − w ∣ − 2 μ ( d t ) = Im F μ ( w ) / Im w by (7.1). Using the Cauchy–Schwarz inequality, we find
∣ F μ ( w η ) − F μ ( w η ′ ) ∣ = ∣ ∫ R ( w η − w η ′ ) μ ( d t ) ( t − w η ) ( t − w η ′ ) ∣ ≤ η 2 ( ∫ R μ ( d t ) ∣ t − w η ∣ 2 ) 1 / 2 ( ∫ R μ ( d t ) ∣ t − w η ′ ∣ 2 ) 1 / 2 = η 2 E ( Im F μ ( w η ) Im F μ ( w η ′ ) ) 1 / 2 . (106) \begin{aligned}
|F_\mu(w_\eta)-F_\mu(w'_\eta)|
&=\left|\int_{\mathbb{R}}\frac{(w_\eta-w'_\eta)\mu(\mathrm{d}t)}{(t-w_\eta)(t-w'_\eta)}\right| \\
&\leq\eta^2\left(\int_{\mathbb{R}}\frac{\mu(\mathrm{d}t)}{|t-w_\eta|^2}\right)^{1/2}
\left(\int_{\mathbb{R}}\frac{\mu(\mathrm{d}t)}{|t-w'_\eta|^2}\right)^{1/2} \\
&=\frac{\eta}{2E}\bigl(\operatorname{Im}F_\mu(w_\eta)\operatorname{Im}F_\mu(w'_\eta)\bigr)^{1/2}. \tag*{(106)}
\end{aligned} ∣ F μ ( w η ) − F μ ( w η ′ ) ∣ = ∫ R ( t − w η ) ( t − w η ′ ) ( w η − w η ′ ) μ ( d t ) ≤ η 2 ( ∫ R ∣ t − w η ∣ 2 μ ( d t ) ) 1/2 ( ∫ R ∣ t − w η ′ ∣ 2 μ ( d t ) ) 1/2 = 2 E η ( Im F μ ( w η ) Im F μ ( w η ′ ) ) 1/2 . (106)
Suppose that F μ ( E 2 + i 0 ) F_\mu(E^2+i0) F μ ( E 2 + i 0 ) exists. Since w η ′ = E 2 + i ( 2 E η ) w'_\eta=E^2+i(2E\eta) w η ′ = E 2 + i ( 2 E η ) , we have F μ ( w η ′ ) → F μ ( E 2 + i 0 ) F_\mu(w'_\eta)\to F_\mu(E^2+i0) F μ ( w η ′ ) → F μ ( E 2 + i 0 ) , and there exist M < ∞ M<\infty M < ∞ and η 0 ∈ ( 0 , E ] \eta_0\in(0,E] η 0 ∈ ( 0 , E ] such that Im F μ ( w η ′ ) ≤ M \operatorname{Im}F_\mu(w'_\eta)\leq M Im F μ ( w η ′ ) ≤ M for η ≤ η 0 \eta\leq\eta_0 η ≤ η 0 . For such η \eta η , write a = Im F μ ( w η ) a=\operatorname{Im}F_{\mu}(w_{\eta}) a = Im F μ ( w η ) . By (7.5) and the inequality a M ≤ ( a + M ) / 2 \sqrt{aM}\le(a+M)/2 a M ≤ ( a + M ) /2 , we have a ≤ M + ( a + M ) / 4 a\le M+(a+M)/4 a ≤ M + ( a + M ) /4 , which implies that a ≤ 2 M a\le2M a ≤ 2 M . Inserting this into (7.5), we obtain ∣ F μ ( w η ) − F μ ( w η ′ ) ∣ ≤ 2 M η / ( 2 E ) → 0 \lvert F_{\mu}(w_{\eta})-F_{\mu}(w'_{\eta})\rvert\le\sqrt{2}M\eta/(2E)\to0 ∣ F μ ( w η ) − F μ ( w η ′ )∣ ≤ 2 M η / ( 2 E ) → 0 , and therefore F μ ( w η ) → F μ ( E 2 + i 0 ) F_{\mu}(w_{\eta})\to F_{\mu}(E^{2}+i0) F μ ( w η ) → F μ ( E 2 + i 0 ) . The converse implication follows by exchanging the roles of w η w_{\eta} w η and w η ′ w'_{\eta} w η ′ in the same argument.
We now turn to rank-one perturbations. Let L L L be a self-adjoint operator on a separable Hilbert space h \mathfrak h h , and let v ∈ h v\in\mathfrak h v ∈ h with ∥ v ∥ = 1 \lVert v\rVert=1 ∥ v ∥ = 1 . For u ∈ R u\in\mathbb{R} u ∈ R , set
L u = L + u ∣ v ⟩ ⟨ v ∣ , Dom ( L u ) = Dom ( L ) , μ u = μ v L u , F u = F μ u . (107) L_u=L+u\lvert v\rangle\langle v\rvert,\qquad\operatorname{Dom}(L_u)=\operatorname{Dom}(L),\qquad\mu_u=\mu_v^{L_u},\qquad F_u=F_{\mu_u}. \tag*{(107)} L u = L + u ∣ v ⟩ ⟨ v ∣ , Dom ( L u ) = Dom ( L ) , μ u = μ v L u , F u = F μ u . (107)
Each L u L_u L u is self-adjoint, since it is a bounded symmetric perturbation of L L L , and each μ u \mu_u μ u is a probability measure. Let N ⊂ R N\subset\mathbb{R} N ⊂ R be the set of λ \lambda λ at which F 0 ( λ + i 0 ) F_0(\lambda+i0) F 0 ( λ + i 0 ) does not exist, and for u ∈ R u\in\mathbb{R} u ∈ R set
Z u = { λ ∈ R ∖ N : F 0 ( λ + i 0 ) ∈ R , 1 + u F 0 ( λ + i 0 ) = 0 } . (108) Z_u=\{\lambda\in\mathbb{R}\setminus N:F_0(\lambda+i0)\in\mathbb{R},\ 1+uF_0(\lambda+i0)=0\}. \tag*{(108)} Z u = { λ ∈ R ∖ N : F 0 ( λ + i 0 ) ∈ R , 1 + u F 0 ( λ + i 0 ) = 0 } . (108)
Note that Z 0 = ∅ Z_0=\varnothing Z 0 = ∅ .
Lemma 7.3. In the setting of (7.6) and (7.7), the following statements hold.
(i) For every u ∈ R u\in\mathbb{R} u ∈ R and z ∈ H z\in\mathbb{H} z ∈ H , we have 1 + u F 0 ( z ) ≠ 0 1+uF_0(z)\ne0 1 + u F 0 ( z ) = 0 and
F u ( z ) = F 0 ( z ) 1 + u F 0 ( z ) . (109) F_u(z)=\frac{F_0(z)}{1+uF_0(z)}. \tag*{(109)} F u ( z ) = 1 + u F 0 ( z ) F 0 ( z ) . (109) (ii) The set N N N is a Lebesgue-null Borel set, and for every u ∈ R u\in\mathbb{R} u ∈ R the singular part of μ u \mu_u μ u is carried by N ∪ Z u N\cup Z_u N ∪ Z u :
( μ u ) s i n g ( R ∖ ( N ∪ Z u ) ) = 0. (\mu_u)_{\mathrm{sing}}\bigl(\mathbb{R}\setminus(N\cup Z_u)\bigr)=0. ( μ u ) sing ( R ∖ ( N ∪ Z u ) ) = 0. (iii) For every Borel set B ⊂ R B\subset\mathbb{R} B ⊂ R , the map u ↦ μ u ( B ) u\mapsto\mu_u(B) u ↦ μ u ( B ) is Borel and
∫ R μ u ( B ) d u = ∣ B ∣ . (110) \int_{\mathbb{R}}\mu_u(B)\,\mathrm{d}u=\lvert B\rvert. \tag*{(110)} ∫ R μ u ( B ) d u = ∣ B ∣ . (110) (iv) Let ( Ω ′ , A ) (\Omega',\mathcal{A}) ( Ω ′ , A ) be a measurable space, and suppose that for each ω ∈ Ω ′ \omega\in\Omega' ω ∈ Ω ′ we are given L = L ( ω ) L=L(\omega) L = L ( ω ) and v = v ( ω ) v=v(\omega) v = v ( ω ) as above, possibly on a Hilbert space depending on ω \omega ω , such that ω ↦ F 0 ( ω , z ) \omega\mapsto F_0(\omega,z) ω ↦ F 0 ( ω , z ) is A \mathcal{A} A -measurable for every z ∈ H z\in\mathbb{H} z ∈ H . Write N ( ω ) N(\omega) N ( ω ) , Z u ( ω ) Z_u(\omega) Z u ( ω ) , μ u ( ω ) \mu_u(\omega) μ u ( ω ) for the corresponding objects. Then the sets
{ ( ω , λ ) : λ ∈ N ( ω ) } ∈ A ⊗ B ( R ) \{(\omega,\lambda):\lambda\in N(\omega)\}\in\mathcal{A}\otimes\mathcal{B}(\mathbb{R}) {( ω , λ ) : λ ∈ N ( ω )} ∈ A ⊗ B ( R ) , { ( ω , u , λ ) : λ ∈ Z u ( ω ) } ∈ A ⊗ B ( R ) ⊗ B ( R ) \{(\omega,u,\lambda):\lambda\in Z_u(\omega)\}\in\mathcal{A}\otimes\mathcal{B}(\mathbb{R})\otimes\mathcal{B}(\mathbb{R}) {( ω , u , λ ) : λ ∈ Z u ( ω )} ∈ A ⊗ B ( R ) ⊗ B ( R )
are measurable, the function ( ω , λ ) ↦ R μ 0 ( ω ) ( λ ) (\omega,\lambda)\mapsto R_{\mu_0(\omega)}(\lambda) ( ω , λ ) ↦ R μ 0 ( ω ) ( λ ) is A ⊗ B ( R ) \mathcal{A}\otimes\mathcal{B}(\mathbb{R}) A ⊗ B ( R ) -measurable, and for every T ∈ A ⊗ B ( R ) T\in\mathcal{A}\otimes\mathcal{B}(\mathbb{R}) T ∈ A ⊗ B ( R ) the map ( ω , u ) ↦ μ u ( ω ) ( T ω ) (\omega,u)\mapsto\mu_u(\omega)(T_\omega) ( ω , u ) ↦ μ u ( ω ) ( T ω ) is A ⊗ B ( R ) \mathcal{A}\otimes\mathcal{B}(\mathbb{R}) A ⊗ B ( R ) -measurable, where T ω = { λ : ( ω , λ ) ∈ T } T_\omega=\{\lambda:(\omega,\lambda)\in T\} T ω = { λ : ( ω , λ ) ∈ T } .
The identity (7.8) is the Aronszajn–Krein formula (see [27]), and (7.9) is spectral averaging; see [123]. Part (iv) is what allows the null sets in (ii) and (iii) to be used inside conditional Fubini arguments in which L L L and v v v are random.
Proof. We prove the four parts in the order (i), (iv), (ii), (iii).
Proof of (i). Fix z ∈ H z \in\mathsf{H} z ∈ H . Since Ran ( L − z ) − 1 = Dom ( L ) = Dom ( L u ) \operatorname{Ran}(L-z)^{-1}=\operatorname{Dom}(L)=\operatorname{Dom}(L_u) Ran ( L − z ) − 1 = Dom ( L ) = Dom ( L u ) , the second resolvent identity yields
( L − z ) − 1 − ( L u − z ) − 1 = ( L u − z ) − 1 ( L u − L ) ( L − z ) − 1 = u ( L u − z ) − 1 ∣ v ⟩ ⟨ v ∣ ( L − z ) − 1 . (L-z)^{-1}-(L_u-z)^{-1}=(L_u-z)^{-1}(L_u-L)(L-z)^{-1}=u(L_u-z)^{-1}|v\rangle\langle v|(L-z)^{-1}. ( L − z ) − 1 − ( L u − z ) − 1 = ( L u − z ) − 1 ( L u − L ) ( L − z ) − 1 = u ( L u − z ) − 1 ∣ v ⟩ ⟨ v ∣ ( L − z ) − 1 .
Applying both sides to v v v and taking the inner product with v v v , we obtain F 0 ( z ) − F u ( z ) = u F u ( z ) F 0 ( z ) F_0(z)-F_u(z)=uF_u(z)F_0(z) F 0 ( z ) − F u ( z ) = u F u ( z ) F 0 ( z ) , that is, F u ( z ) ( 1 + u F 0 ( z ) ) = F 0 ( z ) F_u(z)(1+uF_0(z))=F_0(z) F u ( z ) ( 1 + u F 0 ( z )) = F 0 ( z ) . If u ≠ 0 u\ne0 u = 0 and 1 + u F 0 ( z ) = 0 1+uF_0(z)=0 1 + u F 0 ( z ) = 0 , then F 0 ( z ) = − 1 / u F_0(z)=-1/u F 0 ( z ) = − 1/ u would be real, which is impossible because Im F 0 ( z ) > 0 \operatorname{Im}F_0(z)>0 Im F 0 ( z ) > 0 . This proves (i).
Proof of (iv). We first show that ( ω , z ) ↦ F 0 ( ω , z ) (\omega,z)\mapsto F_0(\omega,z) ( ω , z ) ↦ F 0 ( ω , z ) is A ⊗ B ( H ) \mathcal{A}\otimes\mathcal{B}(\mathsf{H}) A ⊗ B ( H ) -measurable. For n ∈ N n\in\mathbb{N} n ∈ N , partition H \mathsf{H} H into countably many Borel sets Q n , k Q_{n,k} Q n , k of diameter at most 1 / n 1/n 1/ n , choose z n , k ∈ Q n , k z_{n,k}\in Q_{n,k} z n , k ∈ Q n , k , and set F ( n ) ( ω , z ) = F 0 ( ω , z n , k ) F^{(n)}(\omega,z)=F_0(\omega,z_{n,k}) F ( n ) ( ω , z ) = F 0 ( ω , z n , k ) for z ∈ Q n , k z\in Q_{n,k} z ∈ Q n , k . Each F ( n ) F^{(n)} F ( n ) is jointly measurable, and F ( n ) → F 0 F^{(n)}\to F_0 F ( n ) → F 0 pointwise because F 0 ( ω , ⋅ ) F_0(\omega,\cdot) F 0 ( ω , ⋅ ) is continuous. By (7.8), the function ( ω , u , z ) ↦ F u ( ω , z ) (\omega,u,z)\mapsto F_u(\omega,z) ( ω , u , z ) ↦ F u ( ω , z ) is then jointly measurable as well.
Let C C C be the set of ( ω , λ ) (\omega,\lambda) ( ω , λ ) satisfying the rational Cauchy criterion
∀ k ∈ N ∃ m ∈ N ∀ η , η ′ ∈ Q ∩ ( 0 , 1 / m ) : ∣ F 0 ( ω , λ + i η ) − F 0 ( ω , λ + i η ′ ) ∣ < 1 / k . (111) \forall k\in\mathbb{N}\ \exists m\in\mathbb{N}\ \forall\eta,\eta'\in\mathbb{Q}\cap(0,1/m):\ \lvert F_0(\omega,\lambda+i\eta)-F_0(\omega,\lambda+i\eta')\rvert<1/k. \tag*{(111)} ∀ k ∈ N ∃ m ∈ N ∀ η , η ′ ∈ Q ∩ ( 0 , 1/ m ) : ∣ F 0 ( ω , λ + i η ) − F 0 ( ω , λ + i η ′ )∣ < 1/ k . (111)
It is obtained from jointly measurable sets by countable unions and intersections, so C ∈ A ⊗ B ( R ) C\in\mathcal{A}\otimes\mathcal{B}(\mathbb{R}) C ∈ A ⊗ B ( R ) . If F 0 ( ω , λ + i 0 ) F_0(\omega,\lambda+i0) F 0 ( ω , λ + i 0 ) exists, then (7.10) holds. Conversely, if (7.10) holds, then by continuity of η ↦ F 0 ( ω , λ + i η ) \eta\mapsto F_0(\omega,\lambda+i\eta) η ↦ F 0 ( ω , λ + i η ) the same inequality, with ≤ \le ≤ in place of < < < , holds for all real η , η ′ ∈ ( 0 , 1 / m ) \eta,\eta'\in(0,1/m) η , η ′ ∈ ( 0 , 1/ m ) , and completeness of C \mathbb{C} C shows that the limit exists. It follows that { ( ω , λ ) : λ ∈ N ( ω ) } \{(\omega,\lambda):\lambda\in N(\omega)\} {( ω , λ ) : λ ∈ N ( ω )} is the complement of C C C . On C C C we set F 0 ∗ ( ω , λ ) = lim n → ∞ F 0 ( ω , λ + i / n ) F_0^*(\omega,\lambda)=\lim_{n\to\infty}F_0(\omega,\lambda+i/n) F 0 ∗ ( ω , λ ) = lim n → ∞ F 0 ( ω , λ + i / n ) , and we set F 0 ∗ = 0 F_0^*=0 F 0 ∗ = 0 off C C C ; this function is jointly measurable, as a pointwise limit of measurable functions multiplied by 1 C 1_C 1 C . Then
{ ( ω , u , λ ) : λ ∈ Z u ( ω ) } = { ( ω , u , λ ) : ( ω , λ ) ∈ C , Im F 0 ∗ ( ω , λ ) = 0 , 1 + u F 0 ∗ ( ω , λ ) = 0 } \{(\omega,u,\lambda):\lambda\in Z_u(\omega)\}=\{(\omega,u,\lambda):(\omega,\lambda)\in C,\ \operatorname{Im}F_0^*(\omega,\lambda)=0,\ 1+uF_0^*(\omega,\lambda)=0\} {( ω , u , λ ) : λ ∈ Z u ( ω )} = {( ω , u , λ ) : ( ω , λ ) ∈ C , Im F 0 ∗ ( ω , λ ) = 0 , 1 + u F 0 ∗ ( ω , λ ) = 0 }
is measurable. The measurability of R μ 0 ( ω ) ( λ ) R_{\mu_0(\omega)}(\lambda) R μ 0 ( ω ) ( λ ) follows from (7.3).
For the last assertion, let P η ( x ) = π − 1 η / ( x 2 + η 2 ) P_\eta(x)=\pi^{-1}\eta/(x^2+\eta^2) P η ( x ) = π − 1 η / ( x 2 + η 2 ) , so that π − 1 Im F u ( λ + i η ) = ∫ R P η ( λ − t ) μ u ( d t ) \pi^{-1}\operatorname{Im}F_u(\lambda+i\eta)=\int_{\mathbb{R}}P_\eta(\lambda-t)\,\mu_u(\mathrm{d}t) π − 1 Im F u ( λ + i η ) = ∫ R P η ( λ − t ) μ u ( d t ) by (7.1). For a continuous function f f f with compact support, we have by Tonelli’s theorem
1 π ∫ R f ( λ ) Im F u ( λ + i η ) d λ = ∫ R ( P η ∗ f ) ( t ) μ u ( d t ) . \frac{1}{\pi}\int_{\mathbb{R}}f(\lambda)\operatorname{Im}F_u(\lambda+i\eta)\,\mathrm{d}\lambda=\int_{\mathbb{R}}(P_\eta*f)(t)\,\mu_u(\mathrm{d}t). π 1 ∫ R f ( λ ) Im F u ( λ + i η ) d λ = ∫ R ( P η ∗ f ) ( t ) μ u ( d t ) .
Since P η ∗ f → f P_\eta*f\to f P η ∗ f → f uniformly as η ↓ 0 \eta\downarrow0 η ↓ 0 and μ u \mu_u μ u is a probability measure, the right side converges to ∫ f d μ u \int f\,\mathrm{d}\mu_u ∫ f d μ u . The left side, with η = 1 / n \eta=1/n η = 1/ n , is a measurable function of ( ω , u ) (\omega,u) ( ω , u ) by Fubini’s theorem, because its integrand is jointly measurable and bounded by n ∣ f ( λ ) ∣ n|f(\lambda)| n ∣ f ( λ ) ∣ . This shows that ( ω , u ) ↦ ∫ f d μ u ( ω ) (\omega,u)\mapsto\int f\,\mathrm{d}\mu_u(\omega) ( ω , u ) ↦ ∫ f d μ u ( ω ) is measurable. Approximating the indicator of an open interval from below by such f f f and applying the π \pi π –λ \lambda λ theorem, we find that ( ω , u ) ↦ μ u ( ω ) ( B ) (\omega,u)\mapsto\mu_u(\omega)(B) ( ω , u ) ↦ μ u ( ω ) ( B ) is measurable for every Borel set B B B . Finally, the class of T ∈ A ⊗ B ( R ) T\in\mathcal{A}\otimes\mathcal{B}(\mathbb{R}) T ∈ A ⊗ B ( R ) for which ( ω , u ) ↦ μ u ( ω ) ( T ω ) (\omega,u)\mapsto\mu_u(\omega)(T_\omega) ( ω , u ) ↦ μ u ( ω ) ( T ω ) is measurable contains the rectangles A 0 × B A_0\times B A 0 × B , for which the map is 1 A 0 ( ω ) μ u ( ω ) ( B ) \mathbf{1}_{A_0}(\omega)\mu_u(\omega)(B) 1 A 0 ( ω ) μ u ( ω ) ( B ) , and it is closed under complements and increasing unions because each μ u ( ω ) \mu_u(\omega) μ u ( ω ) is a probability measure. By the π \pi π –λ \lambda λ theorem it is all of A ⊗ B ( R ) \mathcal{A}\otimes\mathcal{B}(\mathbb{R}) A ⊗ B ( R ) ; see also [81, Chapter 1].
Proof of (ii). By (iv), applied with a one-point space Ω ′ \Omega' Ω ′ , the set N N N is Borel, and it is Lebesgue-null by Lemma 7.1 (i). We now fix u ∈ R u\in\mathbb{R} u ∈ R and λ ∉ N ∪ Z u \lambda\notin N\cup Z_u λ ∈ / N ∪ Z u , and write a = F 0 ( λ + i 0 ) a=F_0(\lambda+\mathrm{i}0) a = F 0 ( λ + i 0 ) . If 1 + u a ≠ 0 1+ua\ne0 1 + u a = 0 , then (109) implies that F u ( λ + i η ) → a / ( 1 + u a ) ∈ C F_u(\lambda+\mathrm{i}\eta)\to a/(1+ua)\in\mathbb{C} F u ( λ + i η ) → a / ( 1 + u a ) ∈ C ; in particular, Im F u ( λ + i η ) \operatorname{Im}F_u(\lambda+\mathrm{i}\eta) Im F u ( λ + i η ) stays bounded as η ↓ 0 \eta\downarrow0 η ↓ 0 . If 1 + u a = 0 1+ua=0 1 + u a = 0 , then u ≠ 0 u\ne0 u = 0 and a = − 1 / u a=-1/u a = − 1/ u is real, and λ ∈ Z u \lambda\in Z_u λ ∈ Z u , which was excluded. Hence R ∖ ( N ∪ Z u ) \mathbb{R}\setminus(N\cup Z_u) R ∖ ( N ∪ Z u ) does not meet the set on which Im F u ( λ + i η ) → ∞ \operatorname{Im}F_u(\lambda+\mathrm{i}\eta)\to\infty Im F u ( λ + i η ) → ∞ , and Lemma 7.1 (ii), applied to μ u \mu_u μ u , gives ( μ u ) s i n g ( R ∖ ( N ∪ Z u ) ) = 0 (\mu_u)_{\mathrm{sing}}(\mathbb{R}\setminus(N\cup Z_u))=0 ( μ u ) sing ( R ∖ ( N ∪ Z u )) = 0 .
Proof of (iii). The measurability of u ↦ μ u ( B ) u\mapsto\mu_u(B) u ↦ μ u ( B ) is the case of (iv) with a one-point space Ω ′ \Omega' Ω ′ . Fix z ∈ H z\in\mathbb{H} z ∈ H and write F 0 ( z ) = a + i b F_0(z)=a+\mathrm{i}b F 0 ( z ) = a + i b with b > 0 b>0 b > 0 . By (109) ,
Im F u ( z ) = Im ( F 0 ( z ) ( 1 + u F 0 ( z ) ) ) ∣ 1 + u F 0 ( z ) ∣ 2 = b ( 1 + u a ) 2 + u 2 b 2 . \operatorname{Im}F_u(z)=\frac{\operatorname{Im}\bigl(F_0(z)(1+uF_0(z))\bigr)}{|1+uF_0(z)|^2}=\frac{b}{(1+ua)^2+u^2b^2}. Im F u ( z ) = ∣1 + u F 0 ( z ) ∣ 2 Im ( F 0 ( z ) ( 1 + u F 0 ( z )) ) = ( 1 + u a ) 2 + u 2 b 2 b .
With c = a 2 + b 2 c=a^2+b^2 c = a 2 + b 2 , we have ( 1 + u a ) 2 + u 2 b 2 = c ( u + a / c ) 2 + b 2 / c (1+ua)^2+u^2b^2=c(u+a/c)^2+b^2/c ( 1 + u a ) 2 + u 2 b 2 = c ( u + a / c ) 2 + b 2 / c , and therefore
∫ R Im F u ( z ) d u = b ∫ R d u c ( u + a / c ) 2 + b 2 / c = b ⋅ π b = π . (112) \int_{\mathbb{R}}\operatorname{Im}F_u(z)\,\mathrm{d}u=b\int_{\mathbb{R}}\frac{\mathrm{d}u}{c(u+a/c)^2+b^2/c}=b\cdot\frac{\pi}{b}=\pi. \tag*{(112)} ∫ R Im F u ( z ) d u = b ∫ R c ( u + a / c ) 2 + b 2 / c d u = b ⋅ b π = π . (112)
Define the measure ν ( B ) = ∫ R μ u ( B ) d u \nu(B)=\int_{\mathbb{R}}\mu_u(B)\,\mathrm{d}u ν ( B ) = ∫ R μ u ( B ) d u on Borel sets B ⊂ R B\subset\mathbb{R} B ⊂ R . For z = x + i y ∈ H z=x+\mathrm{i}y\in\mathbb{H} z = x + i y ∈ H , Tonelli’s theorem and (112) give
∫ R y ν ( d t ) ( t − x ) 2 + y 2 = ∫ R Im F u ( x + i y ) d u = π = ∫ R y d t ( t − x ) 2 + y 2 . (113) \int_{\mathbb{R}}\frac{y\nu(\mathrm{d}t)}{(t-x)^2+y^2}=\int_{\mathbb{R}}\operatorname{Im}F_u(x+\mathrm{i}y)\,\mathrm{d}u=\pi=\int_{\mathbb{R}}\frac{y\,\mathrm{d}t}{(t-x)^2+y^2}. \tag*{(113)} ∫ R ( t − x ) 2 + y 2 y ν ( d t ) = ∫ R Im F u ( x + i y ) d u = π = ∫ R ( t − x ) 2 + y 2 y d t . (113)
The measure ν \nu ν is locally finite: for t ∈ [ x − y , x + y ] t\in[x-y,x+y] t ∈ [ x − y , x + y ] we have y / ( ( t − x ) 2 + y 2 ) ≥ 1 / ( 2 y ) y/((t-x)^2+y^2)\ge1/(2y) y / (( t − x ) 2 + y 2 ) ≥ 1/ ( 2 y ) , and (113) implies that ν ( [ x − y , x + y ] ) ≤ 2 π y \nu([x-y,x+y])\le2\pi y ν ([ x − y , x + y ]) ≤ 2 π y . Taking z = i z=\mathrm{i} z = i in (113) , we also obtain
∫ R ( 1 + t 2 ) − 1 ν ( d t ) = π . \int_{\mathbb{R}}(1+t^2)^{-1}\nu(\mathrm{d}t)=\pi. ∫ R ( 1 + t 2 ) − 1 ν ( d t ) = π .
It remains to show that a locally finite measure ν \nu ν with ∫ R ( 1 + t 2 ) − 1 ν ( d t ) < ∞ \int_{\mathbb{R}}(1+t^2)^{-1}\nu(\mathrm{d}t)<\infty ∫ R ( 1 + t 2 ) − 1 ν ( d t ) < ∞ is determined by its Poisson transform ( P y ∗ ν ) ( x ) = π − 1 ∫ y ( ( t − x ) 2 + y 2 ) − 1 ν ( d t ) (P_y*\nu)(x)=\pi^{-1}\int y((t-x)^2+y^2)^{-1}\nu(\mathrm{d}t) ( P y ∗ ν ) ( x ) = π − 1 ∫ y (( t − x ) 2 + y 2 ) − 1 ν ( d t ) . Let f f f be continuous with support in [ − r , r ] [-r,r] [ − r , r ] . For 0 < y ≤ 1 0<y\le1 0 < y ≤ 1 and ∣ t ∣ ≥ 2 r |t|\ge2r ∣ t ∣ ≥ 2 r , we have ∣ ( P y ∗ f ) ( t ) ∣ ≤ ∥ f ∥ ∞ ⋅ 2 r ⋅ π − 1 y ( t / 2 ) − 2 |(P_y*f)(t)|\le\|f\|_\infty\cdot2r\cdot\pi^{-1}y(t/2)^{-2} ∣ ( P y ∗ f ) ( t ) ∣ ≤ ∥ f ∥ ∞ ⋅ 2 r ⋅ π − 1 y ( t /2 ) − 2 , and ∣ ( P y ∗ f ) ( t ) ∣ ≤ ∥ f ∥ ∞ |(P_y*f)(t)|\le\|f\|_\infty ∣ ( P y ∗ f ) ( t ) ∣ ≤ ∥ f ∥ ∞ for all t t t ; hence ∣ ( P y ∗ f ) ( t ) ∣ ≤ C f ( 1 + t 2 ) − 1 |(P_y*f)(t)|\le C_f(1+t^2)^{-1} ∣ ( P y ∗ f ) ( t ) ∣ ≤ C f ( 1 + t 2 ) − 1 for some C f < ∞ C_f<\infty C f < ∞ . By Fubini’s theorem,
∫ R f ( x ) ( P y ∗ ν ) ( x ) d x = ∫ R ( P y ∗ f ) ( t ) ν ( d t ) . \int_{\mathbb{R}}f(x)(P_y*\nu)(x)\,\mathrm{d}x=\int_{\mathbb{R}}(P_y*f)(t)\,\nu(\mathrm{d}t). ∫ R f ( x ) ( P y ∗ ν ) ( x ) d x = ∫ R ( P y ∗ f ) ( t ) ν ( d t ) .
As y ↓ 0 y \downarrow0 y ↓ 0 , we have P y ∗ f → f P_y * f \to f P y ∗ f → f pointwise, and the dominated convergence theorem (with the dominating function C f ( 1 + t 2 ) − 1 C_f(1+t^2)^{-1} C f ( 1 + t 2 ) − 1 ) shows that the right side tends to ∫ f d ν \int f\,d\nu ∫ f d ν . The same holds with ν \nu ν replaced by Lebesgue measure. Since the two Poisson transforms agree by (113) , we conclude that ∫ f d ν = ∫ f ( t ) d t \int f\,d\nu=\int f(t)\,dt ∫ f d ν = ∫ f ( t ) d t for every continuous f f f with compact support. By the uniqueness part of the Riesz representation theorem, ν \nu ν is Lebesgue measure, which is (7.9).
The following lemma gives the two spectral criteria in the form needed below. Part (i) is the criterion of Simon and Wolff [124], and part (ii) is its absolutely continuous counterpart [123]. We state both for an individual coupling u u u , under a condition that holds for Lebesgue-almost every u u u by spectral averaging; this is the form that combines with conditional Fubini arguments.
Lemma 7.4. In the setting of (107) and (108) , let S ⊂ R S \subset\mathbb{R} S ⊂ R be a Borel set.
(i) Let T = { λ ∈ S : R μ 0 ( λ ) = ∞ } T=\{\lambda\in S:R_{\mu_0}(\lambda)=\infty\} T = { λ ∈ S : R μ 0 ( λ ) = ∞ } . For every u ∈ R u\in\mathbb{R} u ∈ R , the restriction of μ u \mu_u μ u to S ∖ T S\setminus T S ∖ T is carried by the set of eigenvalues of L u L_u L u , which is countable. In particular, the restriction of μ u \mu_u μ u to S S S is pure point for every u u u with μ u ( T ) = 0 \mu_u(T)=0 μ u ( T ) = 0 . If ∣ T ∣ = 0 |T|=0 ∣ T ∣ = 0 , this holds for Lebesgue-almost every u ∈ R u\in\mathbb{R} u ∈ R .
(ii) Let T ′ T' T ′ be the set of λ ∈ S \lambda\in S λ ∈ S such that F 0 ( λ + i 0 ) F_0(\lambda+i0) F 0 ( λ + i 0 ) does not exist or Im F 0 ( λ + i 0 ) = 0 \operatorname{Im}F_0(\lambda+i0)=0 Im F 0 ( λ + i 0 ) = 0 . For every u ∈ R u\in\mathbb{R} u ∈ R , we have ( μ u ) s i n g ( S ∖ T ′ ) = 0 (\mu_u)_{\mathrm{sing}}(S\setminus T')=0 ( μ u ) sing ( S ∖ T ′ ) = 0 . In particular, the restriction of μ u \mu_u μ u to S S S is absolutely continuous for every u u u with μ u ( T ′ ) = 0 \mu_u(T')=0 μ u ( T ′ ) = 0 . If ∣ T ′ ∣ = 0 |T'|=0 ∣ T ′ ∣ = 0 , this holds for Lebesgue-almost every u ∈ R u\in\mathbb{R} u ∈ R .
Proof. We begin with (i). Fix u ∈ R u\in\mathbb{R} u ∈ R and λ ∈ S ∖ T \lambda\in S\setminus T λ ∈ S ∖ T . By Lemma 7.1(iv), F 0 ( λ + i 0 ) = a F_0(\lambda+i0)=a F 0 ( λ + i 0 ) = a exists and is real. Suppose first that 1 + u a ≠ 0 1+ua\ne0 1 + u a = 0 . Then (109) implies that F u ( λ + i η ) → a / ( 1 + u a ) ∈ R F_u(\lambda+i\eta)\to a/(1+ua)\in\mathbb{R} F u ( λ + i η ) → a / ( 1 + u a ) ∈ R . Then Im F u ( λ + i η ) → 0 \operatorname{Im}F_u(\lambda+i\eta)\to0 Im F u ( λ + i η ) → 0 , and λ \lambda λ belongs to the set A u = { λ : lim η ↓ 0 Im F u ( λ + i η ) = 0 } A_u=\{\lambda:\lim_{\eta\downarrow0}\operatorname{Im}F_u(\lambda+i\eta)=0\} A u = { λ : lim η ↓ 0 Im F u ( λ + i η ) = 0 } . Suppose next that 1 + u a = 0 1+ua=0 1 + u a = 0 . We claim that λ \lambda λ is an eigenvalue of L u L_u L u . To that end, set ψ η = ( L − λ − i η ) − 1 v \psi_\eta=(L-\lambda-i\eta)^{-1}v ψ η = ( L − λ − i η ) − 1 v for η > 0 \eta>0 η > 0 . By the spectral theorem,
∥ ψ η − ψ η ′ ∥ 2 = ∫ R ∣ 1 t − λ − i η − 1 t − λ − i η ′ ∣ 2 μ 0 ( d t ) . \lVert\psi_\eta-\psi_{\eta'}\rVert^2=\int_{\mathbb{R}}\left|\frac{1}{t-\lambda-i\eta}-\frac{1}{t-\lambda-i\eta'}\right|^2\mu_0(dt). ∥ ψ η − ψ η ′ ∥ 2 = ∫ R t − λ − i η 1 − t − λ − i η ′ 1 2 μ 0 ( d t ) .
The integrand is bounded by 4 ( t − λ ) − 2 4(t-\lambda)^{-2} 4 ( t − λ ) − 2 , which is μ 0 \mu_0 μ 0 -integrable because R μ 0 ( λ ) < ∞ R_{\mu_0}(\lambda)<\infty R μ 0 ( λ ) < ∞ , and it tends to zero for t ≠ λ t\ne\lambda t = λ as η , η ′ ↓ 0 \eta,\eta'\downarrow0 η , η ′ ↓ 0 ; since μ 0 ( { λ } ) = 0 \mu_0(\{\lambda\})=0 μ 0 ({ λ }) = 0 , the dominated convergence theorem shows that ( ψ η ) (\psi_\eta) ( ψ η ) is Cauchy as η ↓ 0 \eta\downarrow0 η ↓ 0 , with a limit ψ \psi ψ . We have ( L − λ ) ψ η = v + i η ψ η → v (L-\lambda)\psi_\eta=v+i\eta\psi_\eta\to v ( L − λ ) ψ η = v + i η ψ η → v . Since L L L is closed, we obtain ψ ∈ Dom ( L ) \psi\in\operatorname{Dom}(L) ψ ∈ Dom ( L ) and ( L − λ ) ψ = v (L-\lambda)\psi=v ( L − λ ) ψ = v . In particular ψ ≠ 0 \psi\ne0 ψ = 0 . Moreover, ⟨ v , ψ ⟩ = lim η ↓ 0 F 0 ( λ + i η ) = a \langle v,\psi\rangle=\lim_{\eta\downarrow0}F_0(\lambda+i\eta)=a ⟨ v , ψ ⟩ = lim η ↓ 0 F 0 ( λ + i η ) = a . Then
( L u − λ ) ψ = v + u ⟨ v , ψ ⟩ v = ( 1 + u a ) v = 0. (L_u-\lambda)\psi=v+u\langle v,\psi\rangle v=(1+ua)v=0. ( L u − λ ) ψ = v + u ⟨ v , ψ ⟩ v = ( 1 + u a ) v = 0.
which proves the claim. We conclude that S ∖ T ⊂ A u ∪ σ p ( L u ) S\setminus T\subset A_u\cup\sigma_p(L_u) S ∖ T ⊂ A u ∪ σ p ( L u ) , where σ p ( L u ) \sigma_p(L_u) σ p ( L u ) is the set of eigenvalues of L u L_u L u . By Lemma 7.1(iii), μ u ( A u ) = 0 \mu_u(A_u)=0 μ u ( A u ) = 0 . This implies that the restriction of μ u \mu_u μ u to S ∖ T S \setminus T S ∖ T is carried by σ p ( L u ) \sigma_p(L_u) σ p ( L u ) . This set is countable, because eigenvectors of distinct eigenvalues are orthogonal and H \mathcal{H} H is separable. If moreover μ u ( T ) = 0 \mu_u(T)=0 μ u ( T ) = 0 , the restriction of μ u \mu_u μ u to S S S is carried by a countable set, that is, it is pure point. Finally, if ∣ T ∣ = 0 |T|=0 ∣ T ∣ = 0 , then (7.9) gives ∫ R μ u ( T ) d u = ∣ T ∣ = 0 \int_{\mathbb{R}} \mu_u(T)\,\mathrm{d}u=|T|=0 ∫ R μ u ( T ) d u = ∣ T ∣ = 0 , and μ u ( T ) = 0 \mu_u(T)=0 μ u ( T ) = 0 for Lebesgue-almost every u u u .
For (ii), let λ ∈ S ∖ T ′ \lambda\in S \setminus T' λ ∈ S ∖ T ′ . Then λ ∉ N \lambda\notin N λ ∈ / N and Im F 0 ( λ + i 0 ) > 0 \operatorname{Im}F_0(\lambda+\mathrm{i}0)>0 Im F 0 ( λ + i 0 ) > 0 , and λ ∉ Z u \lambda\notin Z_u λ ∈ / Z u for every u u u . By Lemma 7.3 (ii), we have ( μ u ) s i n g ( S ∖ T ′ ) = 0 (\mu_u)_{\mathrm{sing}}(S \setminus T')=0 ( μ u ) sing ( S ∖ T ′ ) = 0 . If μ u ( T ′ ) = 0 \mu_u(T')=0 μ u ( T ′ ) = 0 , then ( μ u ) s i n g ( S ) ≤ μ u ( T ′ ) = 0 (\mu_u)_{\mathrm{sing}}(S)\leq\mu_u(T')=0 ( μ u ) sing ( S ) ≤ μ u ( T ′ ) = 0 . The last assertion follows from (7.9) as in (i).
Measurability of resolvent entries
We use the realization of the marked tree in Definition 3.2 : vertices are words in the Ulam–Harris set U \mathcal{U} U , so that every vertex is a deterministic word, and the marked tree is built from the independent families ( N u ) u ∈ U (N_u)_{u\in\mathcal{U}} ( N u ) u ∈ U and ( w u ) u ∈ U ∗ (w_u)_{u\in\mathcal{U}_*} ( w u ) u ∈ U ∗ . A quantity indexed by words that are not vertices of the relevant graph is set equal to zero; with this convention, events such as { x ∈ V } \{x\in V\} { x ∈ V } and { x ∼ y } \{x\sim y\} { x ∼ y } are measurable. For a vertex y y y , we write H ( y ) \mathcal{H}^{(y)} H ( y ) for the operator with the vertex y y y deleted and μ x ( y ) \mu_x^{(y)} μ x ( y ) for the spectral measure of δ x \delta_x δ x for H ( y ) \mathcal{H}^{(y)} H ( y ) , as in Theorem 3.10 and Definition 3.15. By Lemma 3.5 (ii), for a neighbor x x x of y y y , the component of x x x in V ∖ { y } V\setminus\{y\} V ∖ { y } is the branch V x → y V_{x\to y} V x → y , so that
G x x ( y ) ( z ) = Γ x → y ( z ) , z ∈ C ∖ R , x ∼ y . (114) G_{xx}^{(y)}(z)=\Gamma_{x\to y}(z),\qquad z\in\mathbb{C}\setminus\mathbb{R},\quad x\sim y. \tag*{(114)} G xx ( y ) ( z ) = Γ x → y ( z ) , z ∈ C ∖ R , x ∼ y . (114)
Lemma 7.5. Let ( F , D ) (F,D) ( F , D ) be a finite removal given by a fixed finite set F F F of words and a fixed finite set D D D of edges, identified with words as in Definition 3.2 (on the event that these are vertices and edges of T T T ), and let u , x u,x u , x be words. Then the map ( ω , z ) ↦ G u x F , D ( z ) (\omega,z)\mapsto G_{ux}^{F,D}(z) ( ω , z ) ↦ G ux F , D ( z ) is measurable on Ω × H \Omega\times\mathbb{H} Ω × H . In particular, for all words u , w u,w u , w , the map ( ω , z ) ↦ Γ u → w ( z ) (\omega,z)\mapsto\Gamma_{u\to w}(z) ( ω , z ) ↦ Γ u → w ( z ) is measurable.
Proof. For fixed z z z , the map ω ↦ G u x F , D ( z ) \omega\mapsto G_{ux}^{F,D}(z) ω ↦ G ux F , D ( z ) is measurable, as explained after Definition 3.15: by Theorem 3.10 (iii), it is the limit of entries of inverses of finite matrices whose entries are measurable functions of finitely many of the variables N v N_v N v and w v w_v w v . It is also continuous in z ∈ H z\in\mathbb{H} z ∈ H . For n ∈ N n\in\mathbb{N} n ∈ N , partition H \mathbb{H} H into countably many Borel sets Q n , k Q_{n,k} Q n , k of diameter at most 1 / n 1/n 1/ n , choose z n , k ∈ Q n , k z_{n,k}\in Q_{n,k} z n , k ∈ Q n , k , and set G u x ( n ) ( ω , z ) = G u x F , D ( z n , k ) G_{ux}^{(n)}(\omega,z)=G_{ux}^{F,D}(z_{n,k}) G ux ( n ) ( ω , z ) = G ux F , D ( z n , k ) for z ∈ Q n , k z\in Q_{n,k} z ∈ Q n , k . Each G ( n ) G^{(n)} G ( n ) is jointly measurable, and G u x ( n ) → G u x F , D G_{ux}^{(n)}\to G_{ux}^{F,D} G ux ( n ) → G ux F , D pointwise by continuity. The last assertion holds because Γ u → w = G u u ∅ , { e } \Gamma_{u\to w}=G_{uu}^{\varnothing,\{e\}} Γ u → w = G uu ∅ , { e } , where e e e is the edge between u u u and w w w .
Row deletion and the square blocks
We now recall the row structure of Proposition 3.23, for C = V C=V C = V , in a form that treats both parities at once. Let y y y be a vertex with deg ( y ) ≥ 1 \deg(y)\geq1 deg ( y ) ≥ 1 , and let ϵ ∈ { 0 , 1 } \epsilon\in\{0,1\} ϵ ∈ { 0 , 1 } be the parity of its neighbors, so that y ∈ V 1 − ϵ y\in V_{1-\epsilon} y ∈ V 1 − ϵ . Let P ϵ P_\epsilon P ϵ be the orthogonal projection onto ℓ 2 ( V ϵ ) \ell^2(V_\epsilon) ℓ 2 ( V ϵ ) , and let S ϵ S_\epsilon S ϵ denote the square block of H 2 \mathcal{H}^2 H 2 on ℓ 2 ( V ϵ ) \ell^2(V_\epsilon) ℓ 2 ( V ϵ ) , that is, S 0 = A ∗ A S_0=A^*A S 0 = A ∗ A and S 1 = A A ∗ S_1=AA^* S 1 = A A ∗ . The row vector, squared row norm and row direction of y y y are
a y = ∑ x ∼ y h x y δ x ∈ ℓ 2 ( V ϵ ) , U y = ∥ a y ∥ 2 = ∑ x ∼ y ∣ h x y ∣ 2 , v y = U y − 1 / 2 a y . (115) a_y=\sum_{x\sim y}h_{xy}\delta_x\in\ell^2(V_\epsilon),\qquad U_y=\lVert a_y\rVert^2=\sum_{x\sim y}|h_{xy}|^2,\qquad v_y=U_y^{-1/2}a_y. \tag*{(115)} a y = x ∼ y ∑ h x y δ x ∈ ℓ 2 ( V ϵ ) , U y = ∥ a y ∥ 2 = x ∼ y ∑ ∣ h x y ∣ 2 , v y = U y − 1/2 a y . (115)
Almost surely every edge weight is nonzero, and then U y > 0 U_y>0 U y > 0 . Set
K ( y ) = S ϵ − U y ∣ v y ⟩ ⟨ v y ∣ . (116) K^{(y)}=S_\epsilon-U_y|v_y\rangle\langle v_y|. \tag*{(116)} K ( y ) = S ϵ − U y ∣ v y ⟩ ⟨ v y ∣. (116)
By Proposition 3.23(iv) and (vii), K ( y ) K^{(y)} K ( y ) is a nonnegative self-adjoint operator with Dom ( K ( y ) ) = Dom ( S ϵ ) \operatorname{Dom}(K^{(y)})=\operatorname{Dom}(S_\epsilon) Dom ( K ( y ) ) = Dom ( S ϵ ) ; for y ∈ V 1 y\in V_1 y ∈ V 1 it is the operator K 0 ( y ) K_0^{(y)} K 0 ( y ) of Proposition 3.23. Further, for z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R ,
P ϵ ( H ( y ) − z ) − 1 P ϵ = z ( K ( y ) − z 2 ) − 1 , (117) P_\epsilon(\mathcal{H}^{(y)}-z)^{-1}P_\epsilon=z(K^{(y)}-z^2)^{-1}, \tag*{(117)} P ϵ ( H ( y ) − z ) − 1 P ϵ = z ( K ( y ) − z 2 ) − 1 , (117)
where both sides are regarded as operators on ℓ 2 ( V ϵ ) \ell^2(V_\epsilon) ℓ 2 ( V ϵ ) ; this is meaningful because y ∉ V ϵ y\notin V_\epsilon y ∈ / V ϵ . We set
ν y = μ v y K ( y ) , m y ( w ) = F ν y ( w ) = ⟨ v y , ( K ( y ) − w ) − 1 v y ⟩ , w ∈ H . (118) \nu_y=\mu_{v_y}^{K^{(y)}},\qquad m_y(w)=F_{\nu_y}(w)=\langle v_y,(K^{(y)}-w)^{-1}v_y\rangle,\qquad w\in\mathbb{H}. \tag*{(118)} ν y = μ v y K ( y ) , m y ( w ) = F ν y ( w ) = ⟨ v y , ( K ( y ) − w ) − 1 v y ⟩ , w ∈ H . (118)
Lemma 7.6. Almost surely, the following holds for every vertex y y y with deg ( y ) ≥ 1 \deg(y)\geq1 deg ( y ) ≥ 1 .
(i) For every z ∈ C z\in\mathbb{C} z ∈ C with Re z > 0 \operatorname{Re}z>0 Re z > 0 and Im z > 0 \operatorname{Im}z>0 Im z > 0 ,
m y ( z 2 ) = 1 z U y ∑ x ∼ y ∣ h x y ∣ 2 Γ x → y ( z ) . (119) m_y(z^2)=\frac{1}{zU_y}\sum_{x\sim y}|h_{xy}|^2\Gamma_{x\to y}(z). \tag*{(119)} m y ( z 2 ) = z U y 1 x ∼ y ∑ ∣ h x y ∣ 2 Γ x → y ( z ) . (119) (ii) For every neighbor x x x of y y y , the measure μ x ( y ) \mu_{x}^{(y)} μ x ( y ) is invariant under λ ↦ − λ \lambda\mapsto-\lambda λ ↦ − λ . Moreover, ν y \nu_y ν y is the image of the measure U y − 1 ∑ x ∼ y ∣ h x y ∣ 2 μ x ( y ) U_y^{-1}\sum_{x\sim y}|h_{xy}|^2\mu_x^{(y)} U y − 1 ∑ x ∼ y ∣ h x y ∣ 2 μ x ( y ) under the map λ ↦ λ 2 \lambda\mapsto\lambda^2 λ ↦ λ 2 .
(iii) For every E > 0 E>0 E > 0 ,
R ν y ( E 2 ) ≤ 2 E 2 U y ∑ x ∼ y ∣ h x y ∣ 2 R μ x ( y ) ( E ) . (120) R_{\nu_y}(E^2)\leq\frac{2}{E^2U_y}\sum_{x\sim y}|h_{xy}|^2R_{\mu_x^{(y)}}(E). \tag*{(120)} R ν y ( E 2 ) ≤ E 2 U y 2 x ∼ y ∑ ∣ h x y ∣ 2 R μ x ( y ) ( E ) . (120)
Proof. We work on the probability-one event on which the conclusions of Theorem 3.10 and Proposition 3.23 hold and all edge weights are nonzero. Part (i) is the identity (32) (for y ∈ V 1 y\in V_1 y ∈ V 1 ; for y ∈ V 0 y\in V_0 y ∈ V 0 we use Proposition 3.23(vii)), divided by z z z , since ∣ v y ( x ) ∣ 2 = ∣ h x y ∣ 2 / U y |v_y(x)|^2=|h_{xy}|^2/U_y ∣ v y ( x ) ∣ 2 = ∣ h x y ∣ 2 / U y .
For (ii), write μ v y ( y ) \mu_{v_y}^{(y)} μ v y ( y ) for the spectral measure of v y v_y v y for H ( y ) \mathcal{H}^{(y)} H ( y ) . By Lemma 3.5 (ii), the neighbors of y y y lie in distinct components of the graph V ∖ { y } V\setminus\{y\} V ∖ { y } , and H ( y ) \mathcal{H}^{(y)} H ( y ) is the orthogonal sum of the operators of these components (Theorem 3.10 (iv)). It follows that ⟨ δ x , f ( H ( y ) ) δ x ′ ⟩ = 0 \langle\delta_x,f(\mathcal{H}^{(y)})\delta_{x'}\rangle=0 ⟨ δ x , f ( H ( y ) ) δ x ′ ⟩ = 0 for distinct neighbors x , x ′ x,x' x , x ′ of y y y and every bounded Borel function f f f , and by (7.14),
μ v y ( y ) = 1 U y ∑ x ∼ y ∣ h x y ∣ 2 μ x ( y ) . (121) \mu_{v_y}^{(y)}=\frac{1}{U_y}\sum_{x\sim y}|h_{xy}|^2\mu_x^{(y)}. \tag*{(121)} μ v y ( y ) = U y 1 x ∼ y ∑ ∣ h x y ∣ 2 μ x ( y ) . (121)
By Proposition 3.21 (i), applied to the components of V ∖ { y } V \setminus\{y\} V ∖ { y } , each μ x ( y ) \mu_{x}^{(y)} μ x ( y ) is invariant under λ ↦ − λ \lambda\mapsto-\lambda λ ↦ − λ , and by (7.20) so is μ v y ( y ) \mu_{v_y}^{(y)} μ v y ( y ) . Let ν ′ \nu' ν ′ be the image of μ v y ( y ) \mu_{v_y}^{(y)} μ v y ( y ) under λ ↦ λ 2 \lambda\mapsto\lambda^2 λ ↦ λ 2 . For z z z with Re z > 0 \operatorname{Re} z > 0 Re z > 0 and Im z > 0 \operatorname{Im} z > 0 Im z > 0 , the symmetry and the identity ( λ − z ) − 1 + ( − λ − z ) − 1 = 2 z ( λ 2 − z 2 ) − 1 (\lambda-z)^{-1}+(-\lambda-z)^{-1}=2z(\lambda^2-z^2)^{-1} ( λ − z ) − 1 + ( − λ − z ) − 1 = 2 z ( λ 2 − z 2 ) − 1 give
∫ R μ v y ( y ) ( d λ ) λ − z = ∫ R z μ v y ( y ) ( d λ ) λ 2 − z 2 = z ∫ R ν ′ ( d t ) t − z 2 . \int_{\mathbb{R}} \frac{\mu_{v_y}^{(y)}(\mathrm{d}\lambda)}{\lambda-z}=\int_{\mathbb{R}} \frac{z\mu_{v_y}^{(y)}(\mathrm{d}\lambda)}{\lambda^2-z^2}=z\int_{\mathbb{R}} \frac{\nu'(\mathrm{d}t)}{t-z^2}. ∫ R λ − z μ v y ( y ) ( d λ ) = ∫ R λ 2 − z 2 z μ v y ( y ) ( d λ ) = z ∫ R t − z 2 ν ′ ( d t ) .
By (7.16), the left side equals z m y ( z 2 ) = z ∫ ( t − z 2 ) − 1 ν y ( d t ) z m_y(z^2)=z\int(t-z^2)^{-1}\nu_y(\mathrm{d}t) z m y ( z 2 ) = z ∫ ( t − z 2 ) − 1 ν y ( d t ) . Since z ↦ z 2 z\mapsto z^2 z ↦ z 2 maps the open first quadrant onto H \mathbb{H} H , the Borel transforms of ν ′ \nu' ν ′ and ν y \nu_y ν y agree on H \mathbb{H} H . By the Stieltjes inversion formula (or by the uniqueness argument in the proof of Lemma 7.3 (iii)), ν ′ = ν y \nu'=\nu_y ν ′ = ν y . This proves (ii).
For (iii), we use (ii) and the factorization λ 2 − E 2 = ( λ − E ) ( λ + E ) \lambda^2-E^2=(\lambda-E)(\lambda+E) λ 2 − E 2 = ( λ − E ) ( λ + E ) :
R v y ( E 2 ) = ∫ R μ v y ( y ) ( d λ ) ( λ − E ) 2 ( λ + E ) 2 ≤ 2 ∫ [ 0 , ∞ ) μ v y ( y ) ( d λ ) ( λ − E ) 2 ( λ + E ) 2 ≤ 2 E 2 R μ v y ( y ) ( E ) . R_{v_y}(E^2)=\int_{\mathbb{R}}\frac{\mu_{v_y}^{(y)}(\mathrm{d}\lambda)}{(\lambda-E)^2(\lambda+E)^2}\le2\int_{[0,\infty)}\frac{\mu_{v_y}^{(y)}(\mathrm{d}\lambda)}{(\lambda-E)^2(\lambda+E)^2}\le\frac{2}{E^2}R_{\mu_{v_y}^{(y)}}(E). R v y ( E 2 ) = ∫ R ( λ − E ) 2 ( λ + E ) 2 μ v y ( y ) ( d λ ) ≤ 2 ∫ [ 0 , ∞ ) ( λ − E ) 2 ( λ + E ) 2 μ v y ( y ) ( d λ ) ≤ E 2 2 R μ v y ( y ) ( E ) .
In the first inequality, we used that the integrand is even in λ \lambda λ and that μ v y ( y ) \mu_{v_y}^{(y)} μ v y ( y ) is symmetric; in the second, we used that ( λ + E ) 2 ≥ E 2 (\lambda+E)^2\ge E^2 ( λ + E ) 2 ≥ E 2 for λ ≥ 0 \lambda\ge0 λ ≥ 0 . Inserting (7.20), we obtain (7.19).
The next lemma passes from the spectral measures of the row directions to a square block. Together with Proposition 3.23(vi), which passes from the two square blocks to H \mathcal{H} H , it reduces the spectral type of H \mathcal{H} H in I ∪ ( − I ) I\cup(-I) I ∪ ( − I ) to that of the row measures in I 2 I^2 I 2 .
Lemma 7.7. Let K K K be a self-adjoint operator on a separable Hilbert space h \mathfrak{h} h , let B ⊂ R B\subset\mathbb{R} B ⊂ R be a Borel set with 0 ∉ B 0\notin B 0 ∈ / B , and let ( φ j ) j ∈ N (\varphi_j)_{j\in\mathbb{N}} ( φ j ) j ∈ N be vectors whose closed linear span contains ( ker K ) ⊥ (\ker K)^\perp ( ker K ) ⊥ . If the restriction of μ φ j K \mu_{\varphi_j}^{K} μ φ j K to B B B is pure point (respectively absolutely continuous) for every j j j , then K K K has pure point spectrum (respectively is purely absolutely continuous) in B B B .
Proof. We treat the pure point case; the other is identical. Since 0 ∉ B 0\notin B 0 ∈ / B , we have 1 B ( K ) ψ = 1 B ( 0 ) ψ = 0 \mathbf{1}_B(K)\psi=\mathbf{1}_B(0)\psi=0 1 B ( K ) ψ = 1 B ( 0 ) ψ = 0 for ψ ∈ ker K \psi\in\ker K ψ ∈ ker K . For each j j j , the spectral measure of 1 B ( K ) φ j \mathbf{1}_B(K)\varphi_j 1 B ( K ) φ j is the restriction of μ φ j K \mu_{\varphi_j}^{K} μ φ j K to B B B , so 1 B ( K ) φ j ∈ h p p ( K ) \mathbf{1}_B(K)\varphi_j\in\mathfrak{h}_{\mathrm{pp}}(K) 1 B ( K ) φ j ∈ h pp ( K ) . Since 1 B ( K ) \mathbf{1}_B(K) 1 B ( K ) is bounded and h p p ( K ) \mathfrak{h}_{\mathrm{pp}}(K) h pp ( K ) is closed, 1 B ( K ) \mathbf{1}_B(K) 1 B ( K ) maps the closed span of the φ j \varphi_j φ j , and hence ( ker K ) ⊥ (\ker K)^\perp ( ker K ) ⊥ , into h p p ( K ) \mathfrak{h}_{\mathrm{pp}}(K) h pp ( K ) . Together with 1 B ( K ) ker K = { 0 } \mathbf{1}_B(K)\ker K=\{0\} 1 B ( K ) ker K = { 0 } , this gives Ran 1 B ( K ) ⊂ h p p ( K ) \operatorname{Ran}\mathbf{1}_B(K)\subset\mathfrak{h}_{\mathrm{pp}}(K) Ran 1 B ( K ) ⊂ h pp ( K ) .
Finally, we record the conditional structure of a single row. For a word y y y , let G y \mathcal{G}^y G y be the σ \sigma σ -algebra of Proposition 3.23(iv) (with C = V C=V C = V , and with the parities exchanged as in Proposition 3.23(vii) if ∣ y ∣ |y| ∣ y ∣ is even). It is generated by the genealogy and auxiliary variables G 0 \mathcal{G}_0 G 0 , by the row direction v y v_y v y , and by the marks of all edges not incident to y y y . We write Ω y = { y ∈ V , deg ( y ) ≥ 1 } \Omega_y=\{y\in V,\ \deg(y)\ge1\} Ω y = { y ∈ V , deg ( y ) ≥ 1 } , which belongs to G 0 \mathcal{G}_0 G 0 . A probability kernel from ( Ω , G y ) (\Omega, \mathcal{G}^{y}) ( Ω , G y ) to R \mathbb{R} R is a map ω ↦ κ ( ω ) \omega\mapsto\kappa(\omega) ω ↦ κ ( ω ) into the Borel probability measures on R \mathbb{R} R such that ω ↦ κ ( ω ) ( B ) \omega\mapsto\kappa(\omega)(B) ω ↦ κ ( ω ) ( B ) is G y \mathcal{G}^{y} G y -measurable for every Borel set B ⊂ R B \subset\mathbb{R} B ⊂ R . For d ≥ 1 d \ge1 d ≥ 1 , let
g d ( u ) = p d Γ ( d ) u d − 1 e − p u 1 ( 0 , ∞ ) ( u ) . (122) g_d(u) = \frac{p^d}{\Gamma(d)}u^{d-1}e^{-pu}1_{(0,\infty)}(u). \tag*{(122)} g d ( u ) = Γ ( d ) p d u d − 1 e − p u 1 ( 0 , ∞ ) ( u ) . (122)
Lemma 7.8. Let Q ∈ { P , P ∞ } Q \in\{\mathbb{P}, \mathbb{P}^{\infty}\} Q ∈ { P , P ∞ } and let y y y be a word.
(i) There exists a probability kernel ν y ′ \nu'_y ν y ′ from ( Ω , G y ) (\Omega, \mathcal{G}^{y}) ( Ω , G y ) to R \mathbb{R} R such that, Q Q Q -almost surely on Ω y \Omega_y Ω y , we have ν y ′ = ν y \nu'_y=\nu_y ν y ′ = ν y .
(ii) For every G y ⊗ B ( R ) \mathcal{G}^{y} \otimes\mathcal{B}(\mathbb{R}) G y ⊗ B ( R ) -measurable function Φ : Ω × R → [ 0 , ∞ ] \Phi:\Omega\times\mathbb{R} \to[0,\infty] Φ : Ω × R → [ 0 , ∞ ] ,
E Q [ Φ ( ⋅ , U y ) 1 Ω y ] = E Q [ 1 Ω y ∫ 0 ∞ Φ ( ⋅ , u ) g deg ( y ) ( u ) d u ] . (123) \mathbb{E}_{Q}\left[\Phi(\cdot,U_y)1_{\Omega_y}\right]=\mathbb{E}_{Q}\left[1_{\Omega_y}\int_{0}^{\infty}\Phi(\cdot,u)g_{\operatorname{deg}(y)}(u)\,\mathrm{d}u\right]. \tag*{(123)} E Q [ Φ ( ⋅ , U y ) 1 Ω y ] = E Q [ 1 Ω y ∫ 0 ∞ Φ ( ⋅ , u ) g deg ( y ) ( u ) d u ] . (123)
Proof. We begin with (i). The operator H ( y ) \mathcal{H}^{(y)} H ( y ) depends only on the genealogy and on the marks of edges not incident to y y y . For a neighbor x x x of y y y , let G ^ x ( z ) \widehat{G}_x(z) G x ( z ) be the limit of the diagonal entries at x x x of the resolvents at z z z of the Dirichlet truncations of H ( y ) \mathcal{H}^{(y)} H ( y ) to the balls of radius n n n around x x x in the component of x x x , when this limit exists, and zero otherwise. The argument of Lemma 7.5 shows that G ^ x \widehat{G}_x G x is G y ⊗ B ( H ) \mathcal{G}^{y} \otimes\mathcal{B}(\mathbb{H}) G y ⊗ B ( H ) -measurable, and G ^ x = G x x ( y ) \widehat{G}_x=G_{xx}^{(y)} G x = G xx ( y ) on Ω s a \Omega_{\mathrm{sa}} Ω sa by Theorem 3.10 (iii). For w ∈ H w \in\mathbb{H} w ∈ H , let w \sqrt{w} w be the square root in the open first quadrant, and set
M y ( w ) = 1 w ∑ x ∼ y ∣ v y ( x ) ∣ 2 G ^ x ( w ) M_y(w)=\frac{1}{\sqrt{w}}\sum_{x\sim y}|v_y(x)|^2\widehat{G}_x(\sqrt{w}) M y ( w ) = w 1 x ∼ y ∑ ∣ v y ( x ) ∣ 2 G x ( w )
on Ω y \Omega_y Ω y , and M y ( w ) = − 1 / w M_y(w)=-1/w M y ( w ) = − 1/ w off Ω y \Omega_y Ω y . This function is G y ⊗ B ( H ) \mathcal{G}^{y} \otimes\mathcal{B}(\mathbb{H}) G y ⊗ B ( H ) -measurable, and by (7.18) and (7.13), almost surely on Ω y \Omega_y Ω y we have M y = m y M_y=m_y M y = m y on H \mathbb{H} H .
We recover ν y \nu_y ν y from M y M_y M y by the Stieltjes inversion formula: for a finite measure ν \nu ν and a < b a<b a < b , the limit as η ↓ 0 \eta\downarrow0 η ↓ 0 of π − 1 ∫ a b Im F ν ( t + i η ) d t \pi^{-1}\int_a^b \operatorname{Im}F_\nu(t+i\eta)\,\mathrm{d}t π − 1 ∫ a b Im F ν ( t + i η ) d t is ( ν ( ( a , b ) ) + ν ( [ a , b ] ) ) / 2 (\nu((a,b))+\nu([a,b]))/2 ( ν (( a , b )) + ν ([ a , b ])) /2 [123]. Taking a = − m a=-m a = − m and b = q + 1 / m b=q+1/m b = q + 1/ m and letting m → ∞ m\to\infty m → ∞ , we obtain ν ( ( − ∞ , q ] ) \nu((-\infty,q]) ν (( − ∞ , q ]) . For rational q q q , set
f q = lim sup m → ∞ lim sup n → ∞ 1 π ∫ − m q + 1 / m Im M y ( t + i / n ) d t . f_q=\limsup_{m\to\infty}\limsup_{n\to\infty}\frac{1}{\pi}\int_{-m}^{q+1/m}\operatorname{Im}M_y(t+i/n)\,\mathrm{d}t. f q = m → ∞ lim sup n → ∞ lim sup π 1 ∫ − m q + 1/ m Im M y ( t + i / n ) d t .
By Fubini’s theorem, f q f_q f q is G y \mathcal{G}^{y} G y -measurable (the integrand is jointly measurable and bounded on [ − m , q + 1 / m ] [-m,q+1/m] [ − m , q + 1/ m ] , since ∣ G ^ x ( z ) ∣ ≤ 1 / Im z |\widehat{G}_x(z)|\le1/\operatorname{Im}z ∣ G x ( z ) ∣ ≤ 1/ Im z ), and almost surely on Ω y \Omega_y Ω y we have f q = ν y ( ( − ∞ , q ] ) f_q=\nu_y((-\infty,q]) f q = ν y (( − ∞ , q ]) for every rational q q q . Let A A A be the set of ω ∈ Ω y \omega\in\Omega_y ω ∈ Ω y such that q ↦ f q ( ω ) q\mapsto f_q(\omega) q ↦ f q ( ω ) is nondecreasing on Q \mathbb{Q} Q , with limits 0 0 0 at − ∞ -\infty − ∞ and 1 1 1 at + ∞ +\infty + ∞ . Then A ∈ G y A\in\mathcal{G}^{y} A ∈ G y , and Q ( Ω y ∖ A ) = 0 Q(\Omega_y\setminus A)=0 Q ( Ω y ∖ A ) = 0 . For ω ∈ A \omega\in A ω ∈ A , let ν y ′ ( ω ) \nu'_y(\omega) ν y ′ ( ω ) be the probability measure with distribution function λ ↦ inf { f q ( ω ) : q ∈ Q , q > λ } \lambda\mapsto\inf\{f_q(\omega):q\in\mathbb{Q},\ q>\lambda\} λ ↦ inf { f q ( ω ) : q ∈ Q , q > λ } , and for ω ∉ A \omega\notin A ω ∈ / A let ν y ′ ( ω ) = δ 0 \nu'_y(\omega)=\delta_0 ν y ′ ( ω ) = δ 0 . Then ω ↦ ν y ′ ( ω ) ( ( − ∞ , λ ] ) \omega\mapsto\nu'_y(\omega)((-\infty,\lambda]) ω ↦ ν y ′ ( ω ) (( − ∞ , λ ]) is G y \mathcal{G}^{y} G y -measurable for every λ \lambda λ , and since the Borel sets B B B for which ω ↦ ν y ′ ( ω ) ( B ) \omega\mapsto\nu'_y(\omega)(B) ω ↦ ν y ′ ( ω ) ( B ) is measurable form a Dynkin system, ν y ′ \nu'_y ν y ′ is a probability kernel. Almost surely on Ω y \Omega_y Ω y , the distribution functions of ν y ′ \nu'_y ν y ′ and ν y \nu_y ν y agree at every λ \lambda λ , by right continuity, and so ν y ′ = ν y \nu'_y=\nu_y ν y ′ = ν y .
For (ii), Proposition 3.23(iv) (and (vii)) states that, under P \mathbb{P} P and under P ∞ \mathbb{P}^{\infty} P ∞ , conditionally on G y \mathcal{G}^{y} G y , the squared row norm U y U_y U y has the density g deg ( y ) g_{\deg(y)} g d e g ( y ) on Ω y \Omega_y Ω y ; recall that the number of neighbors of y y y in V V V is deg ( y ) \deg(y) deg ( y ) , which is G y \mathcal{G}^{y} G y -measurable. The identity (7.22) is the disintegration formula for this conditional law [81, Theorem 6.4], applied with the G y \mathcal{G}^{y} G y -measurable random element ω ↦ ω \omega\mapsto\omega ω ↦ ω of ( Ω , G y ) (\Omega,\mathcal{G}^{y}) ( Ω , G y ) .
We now attach a rank-one family to the kernel ν y ′ \nu'_y ν y ′ . For ω ∈ Ω \omega\in\Omega ω ∈ Ω , let L y ( ω ) L^{y}(\omega) L y ( ω ) be the operator of multiplication by the variable on L 2 ( ν y ′ ( ω ) ) L^{2}(\nu'_y(\omega)) L 2 ( ν y ′ ( ω )) , and let 1 \mathbf{1} 1 be the constant function, whose spectral measure for L y L^{y} L y is ν y ′ \nu'_y ν y ′ . Let ( μ u y ) u ∈ R (\mu^{y}_{u})_{u\in\mathbb{R}} ( μ u y ) u ∈ R be the family (7.6) with L = L y L=L^{y} L = L y and v = 1 v=\mathbf{1} v = 1 , so that μ 0 y = ν y ′ \mu^{y}_{0}=\nu'_y μ 0 y = ν y ′ . The Borel transform of ν y ′ \nu'_y ν y ′ is G y \mathcal{G}^{y} G y -measurable, so Lemma 7.3(iv) applies with ( Ω ′ , A ′ ) = ( Ω , G y ) (\Omega',\mathcal{A}')=(\Omega,\mathcal{G}^{y}) ( Ω ′ , A ′ ) = ( Ω , G y ) . By (7.8), the Borel transform of μ u y \mu^{y}_{u} μ u y is determined by u u u and the Borel transform of μ 0 y \mu^{y}_{0} μ 0 y , and the same holds for the spectral measures of v y v_y v y for the operators K ( y ) + u ∣ v y ⟩ ⟨ v y ∣ K^{(y)}+u|v_y\rangle\langle v_y| K ( y ) + u ∣ v y ⟩ ⟨ v y ∣ , whose member with u = 0 u=0 u = 0 is ν y \nu_y ν y . Since a finite measure is determined by its Borel transform, we conclude from Lemma 7.8(i) and (7.15) that, Q \mathbb{Q} Q -almost surely on Ω y \Omega_y Ω y ,
μ u y = μ v y K ( y ) + u ∣ v y ⟩ ⟨ v y ∣ for every u ∈ R , in particular μ U y y = μ v y S ε (124) \mu^{y}_{u}=\mu_{v_y}^{K^{(y)}+u|v_y\rangle\langle v_y|}
\quad\text{for every }u\in\mathbb{R},\qquad\text{in particular}\qquad
\mu^{y}_{U_y}=\mu_{v_y}^{S_{\varepsilon}} \tag*{(124)} μ u y = μ v y K ( y ) + u ∣ v y ⟩ ⟨ v y ∣ for every u ∈ R , in particular μ U y y = μ v y S ε (124)
where ε \varepsilon ε is the parity of the neighbors of y y y .
The pure point criterion
Proposition 7.9. Let Q ∈ { P , P ∞ } \mathbb{Q}\in\{\mathbb{P},\mathbb{P}^{\infty}\} Q ∈ { P , P ∞ } and let I ⊂ ( 0 , ∞ ) I\subset(0,\infty) I ⊂ ( 0 , ∞ ) be a Borel set (for example, a compact interval). Suppose that Q \mathbb{Q} Q -almost surely, for Lebesgue-almost every E ∈ I E\in I E ∈ I , for every vertex y y y and every neighbor x x x of y y y ,
∫ R μ x ( y ) ( d λ ) ( λ − E ) 2 < ∞ . (125) \int_{\mathbb{R}}\frac{\mu^{(y)}_{x}(\mathrm{d}\lambda)}{(\lambda-E)^2}<\infty. \tag*{(125)} ∫ R ( λ − E ) 2 μ x ( y ) ( d λ ) < ∞. (125) Then Q \mathbb{Q} Q -almost surely, H \mathcal{H} H has pure point spectrum in I ∪ ( − I ) I\cup(-I) I ∪ ( − I ) . If the hypothesis holds P \mathbb{P} P -almost surely, then it holds P ∞ \mathbb{P}^{\infty} P ∞ -almost surely.
The order of quantifiers in the hypothesis is the following: there exists an event of Q \mathbb{Q} Q -probability one on which, outside a Lebesgue-null set of energies E ∈ I E\in I E ∈ I that may depend on the realization, (7.24) holds simultaneously for all pairs of adjacent vertices. The integral in (7.24) is R μ x ( y ) ( E ) R_{\mu^{(y)}_{x}}(E) R μ x ( y ) ( E ) in the notation of (7.2); in particular, it is infinite if E E E is an eigenvalue of the component of x x x in V ∖ { y } V\setminus\{y\} V ∖ { y } with μ x ( y ) ( { E } ) > 0 \mu^{(y)}_{x}(\{E\})>0 μ x ( y ) ({ E }) > 0 .
Proof. The last assertion holds because P ∞ = P ( ⋅ ∣ S ) \mathbb{P}^{\infty}=\mathbb{P}(\,\cdot\mid S) P ∞ = P ( ⋅ ∣ S ) is absolutely continuous with respect to P \mathbb{P} P . We now fix Q \mathbb{Q} Q and assume the hypothesis under Q \mathbb{Q} Q .
The proof isolates one row y y y at a time. Conditional on all variables except the squared row norm U y U_y U y , the spectral measure of v y v_y v y for the square block is the member at coupling U y U_y U y of a rank-one family, and Lemma 7.6 converts the hypothesis into the finiteness of the reciprocal-square integral of the unperturbed member at almost every point of I 2 I^2 I 2 . The Simon–Wolff criterion and spectral averaging then show that the perturbed measure is pure point on I 2 I^2 I 2 for almost every coupling, and the density of U y U_y U y transfers this to the realized coupling. The row directions span the orthogonal complement of the kernel of each square block, which gives the conclusion for the square blocks and then for H \mathcal{H} H .
Step 1. Reduction to a single row. We claim that it suffices to prove the following statement for every word y y y : Q-almost surely on { y ∈ V , deg ( y ) ≥ 1 } \{y \in V,\ \deg(y) \ge1\} { y ∈ V , deg ( y ) ≥ 1 } , the restriction of μ v y ε ε \mu_{v_y^\varepsilon}^{\varepsilon} μ v y ε ε to I 2 I^2 I 2 is pure point, where ε \varepsilon ε is the parity of the neighbors of y y y . Indeed, the set of words is countable, and this then holds Q-almost surely for all such y y y simultaneously. Fix ε ∈ { 0 , 1 } \varepsilon\in\{0,1\} ε ∈ { 0 , 1 } . The vertices whose neighbors lie in V ε V_\varepsilon V ε are those in V 1 − ε V_{1-\varepsilon} V 1 − ε , and A ∗ δ y = a y A^*\delta_y = a_y A ∗ δ y = a y for y ∈ V 1 y \in V_1 y ∈ V 1 , while A δ x = a x A\delta_x = a_x A δ x = a x for x ∈ V 0 x \in V_0 x ∈ V 0 . By Proposition 3.23(ii) and (vii), the row directions v y v_y v y , y ∈ V 1 − ε y \in V_{1-\varepsilon} y ∈ V 1 − ε with deg ( y ) ≥ 1 \deg(y) \ge1 deg ( y ) ≥ 1 , span ( ker A ) ⊥ (\ker A)^\perp ( ker A ) ⊥ if ε = 0 \varepsilon= 0 ε = 0 and ( ker A ∗ ) ⊥ (\ker A^*)^\perp ( ker A ∗ ) ⊥ if ε = 1 \varepsilon= 1 ε = 1 . Since ker A ∗ A = ker A \ker A^*A = \ker A ker A ∗ A = ker A and ker A A ∗ = ker A ∗ \ker AA^* = \ker A^* ker A A ∗ = ker A ∗ , they span ( ker S ε ) ⊥ (\ker S_\varepsilon)^\perp ( ker S ε ) ⊥ . Since 0 ∉ I 2 0 \notin I^2 0 ∈ / I 2 , Lemma 7.7 shows that S ε S_\varepsilon S ε has pure point spectrum in I 2 I^2 I 2 . By Proposition 3.23(vi), H \mathcal{H} H has pure point spectrum in I ∪ ( − I ) I \cup(-I) I ∪ ( − I ) .
Step 2. The exceptional set of the unperturbed row measure. Fix a word y y y , let ν y ′ \nu'_y ν y ′ be the kernel of Lemma 7.8(i), and set
T y = { λ ∈ I 2 : R ν y ′ ( λ ) = ∞ } . T_y = \{\lambda\in I^2 : R_{\nu'_y}(\lambda) = \infty\}. T y = { λ ∈ I 2 : R ν y ′ ( λ ) = ∞ } .
We claim that { ( ω , λ ) : λ ∈ T y } \{(\omega,\lambda) : \lambda\in T_y\} {( ω , λ ) : λ ∈ T y } is G y ⊗ B ( R ) \mathcal{G}^y \otimes\mathcal{B}(\mathbb{R}) G y ⊗ B ( R ) -measurable and that ∣ T y ∣ = 0 |T_y| = 0 ∣ T y ∣ = 0 , Q-almost surely on Ω y \Omega_y Ω y . The measurability follows from Lemma 7.3(iv), applied with ( Ω ′ , A ) = ( Ω , G y ) (\Omega',\mathcal{A}) = (\Omega,\mathcal{G}^y) ( Ω ′ , A ) = ( Ω , G y ) , L = L y L = L^y L = L y and v = 1 v = 1 v = 1 . For the second assertion, consider a realization in Ω y \Omega_y Ω y and in the probability-one event of the hypothesis, of Lemma 7.6 and of Lemma 7.8(i), so that ν y ′ = v y \nu'_y = v_y ν y ′ = v y . Let E ∈ I E \in I E ∈ I be an energy at which (7.24) holds for all x ∼ y x \sim y x ∼ y . Then (7.19) implies that R ν y ′ ( E 2 ) < ∞ R_{\nu'_y}(E^2) < \infty R ν y ′ ( E 2 ) < ∞ . We conclude that T y ⊂ q ( N ) T_y \subset q(\mathcal{N}) T y ⊂ q ( N ) , where q ( E ) = E 2 q(E) = E^2 q ( E ) = E 2 and N ⊂ I \mathcal{N} \subset I N ⊂ I is the Lebesgue-null set of energies at which the hypothesis fails. Since q q q is locally Lipschitz, ∣ T y ∣ = 0 |T_y| = 0 ∣ T y ∣ = 0 .
Step 3. Spectral averaging in the squared row norm. Let ( μ u y ) u ∈ R (\mu_u^y)_{u\in\mathbb{R}} ( μ u y ) u ∈ R be the rank-one family of (7.23), so that μ 0 y = ν y ′ \mu_0^y = \nu'_y μ 0 y = ν y ′ and, almost surely on Ω y \Omega_y Ω y , μ u y = μ v y ε ε \mu_u^y = \mu_{v_y^\varepsilon}^{\varepsilon} μ u y = μ v y ε ε . By Lemma 7.4(i) with L = L y L = L^y L = L y and S = I 2 S = I^2 S = I 2 , the restriction of μ u y \mu_u^y μ u y to I 2 I^2 I 2 is pure point whenever μ u y ( T y ) = 0 \mu_u^y(T_y) = 0 μ u y ( T y ) = 0 . Set
Φ ( ω , u ) = 1 { μ u y ( T y ) > 0 } . \Phi(\omega,u) = 1\{\mu_u^y(T_y) > 0\}. Φ ( ω , u ) = 1 { μ u y ( T y ) > 0 } .
By Lemma 7.3(iv), with T = { ( ω , λ ) : λ ∈ T y } T = \{(\omega,\lambda) : \lambda\in T_y\} T = {( ω , λ ) : λ ∈ T y } , the function Φ \Phi Φ is G y ⊗ B ( R ) \mathcal{G}^y \otimes\mathcal{B}(\mathbb{R}) G y ⊗ B ( R ) -measurable. For every ω \omega ω with ∣ T y ∣ = 0 |T_y| = 0 ∣ T y ∣ = 0 , spectral averaging (7.9) gives ∫ R μ u y ( T y ) d u = ∣ T y ∣ = 0 \int_{\mathbb{R}} \mu_u^y(T_y)\,du = |T_y| = 0 ∫ R μ u y ( T y ) d u = ∣ T y ∣ = 0 , and Φ ( ω , u ) = 0 \Phi(\omega,u)=0 Φ ( ω , u ) = 0 for Lebesgue-almost every u u u . Using (7.22) and Step 2, we obtain
Q ( { μ u y y ( T y ) > 0 } ∩ Ω y ) ≤ Q ( { ∣ T y ∣ > 0 } ∩ Ω y ) + E Q [ 1 Ω y 1 { ∣ T y ∣ = 0 } ∫ 0 ∞ Φ ( ⋅ , u ) g deg ( y ) ( u ) d u ] = 0. \mathbb{Q}\left(\left\{\mu_{u_y}^{y}(T_y)>0\right\}\cap\Omega_y\right)\leq\mathbb{Q}\left(\left\{|T_y|>0\right\}\cap\Omega_y\right)+\mathbb{E}_{\mathbb{Q}}\left[\mathbf{1}_{\Omega_y}\mathbf{1}_{\{|T_y|=0\}}\int_{0}^{\infty}\Phi(\cdot,u)g_{\deg(y)}(u)\,du\right]=0. Q ( { μ u y y ( T y ) > 0 } ∩ Ω y ) ≤ Q ( { ∣ T y ∣ > 0 } ∩ Ω y ) + E Q [ 1 Ω y 1 { ∣ T y ∣ = 0 } ∫ 0 ∞ Φ ( ⋅ , u ) g d e g ( y ) ( u ) d u ] = 0.
Then, Q \mathbb{Q} Q -almost surely on Ω y \Omega_y Ω y , the restriction of μ y ε = μ u y y \mu_{y}^{\varepsilon}=\mu_{u_y}^{y} μ y ε = μ u y y to I 2 I^2 I 2 is pure point. By Step 1, this completes the proof.
Remark 7.10. The hypothesis of Proposition 7.9 concerns spectral measures of the vertex-deleted operators at the real axis. In Section 8, it is verified by constructing, for almost every energy E E E , a solution f ∈ ℓ 2 f\in\ell^2 f ∈ ℓ 2 of the equation ( H ( y ) − E ) f = δ x (\mathcal{H}^{(y)}-E)f=\delta_x ( H ( y ) − E ) f = δ x (see Lemma 8.10).
The absolutely continuous criterion
Proposition 7.11. Let I ⊂ ( 0 , ∞ ) I\subset(0,\infty) I ⊂ ( 0 , ∞ ) be a Borel set (for example, an interval). Suppose that for Lebesgue-almost every E ∈ I E\in I E ∈ I ,
P ( Im Γ o ( E + i 0 ) > 0 ) > 0 (126) \mathbb{P}\left(\operatorname{Im}\Gamma_o(E+\mathrm{i}0)>0\right)>0 \tag*{(126)} P ( Im Γ o ( E + i 0 ) > 0 ) > 0 (126) where the event is that the vertical boundary value Γ o ( E + i 0 ) \Gamma_o(E+\mathrm{i}0) Γ o ( E + i 0 ) exists and has positive imaginary part. Then P ∞ \mathbb{P}^{\infty} P ∞ -almost surely, H \mathcal{H} H is purely absolutely continuous in I ∪ ( − I ) I\cup(-I) I ∪ ( − I ) .
By Proposition 5.11(i), at every E ∈ E p E\in\mathcal{E}_p E ∈ E p with E > 0 E>0 E > 0 the probability in (126) is either 0 0 0 or ϱ p \varrho_p ϱ p . Before the proof, we record two deterministic lemmas. The first propagates positivity of imaginary parts along paths, and the second describes the infinite components left after deleting a vertex. They are the pathwise form, at all energies outside a null set, of statements proved at a fixed regular energy in Section 5.
Lemma 7.12. Fix a realization in Ω s a \Omega_{\mathrm{sa}} Ω sa with all edge weights nonzero, and fix E ∈ R E\in\mathbb{R} E ∈ R such that Γ u → w ( E + i 0 ) \Gamma_{u\to w}(E+\mathrm{i}0) Γ u → w ( E + i 0 ) exists for every oriented edge ( u , w ) (u,w) ( u , w ) . If ( u , w ) (u,w) ( u , w ) is an oriented edge and Im Γ x → u ( E + i 0 ) > 0 \operatorname{Im}\Gamma_{x\to u}(E+\mathrm{i}0)>0 Im Γ x → u ( E + i 0 ) > 0 for some neighbor x ≠ w x\neq w x = w of u u u , then Im Γ u → w ( E + i 0 ) > 0 \operatorname{Im}\Gamma_{u\to w}(E+\mathrm{i}0)>0 Im Γ u → w ( E + i 0 ) > 0 .
Proof. By (3.12) with C = V C=V C = V , we have Γ u → w ( z ) = − Δ ( z ) − 1 \Gamma_{u\to w}(z)=-\Delta(z)^{-1} Γ u → w ( z ) = − Δ ( z ) − 1 with
Δ ( z ) = z + ∑ x ′ ∼ u x ′ ≠ w ∣ h u x ′ ∣ 2 Γ x ′ → u ( z ) . \Delta(z)=z+\sum_{\substack{x'\sim u\\x'\neq w}}|h_{ux'}|^2\Gamma_{x'\to u}(z). Δ ( z ) = z + x ′ ∼ u x ′ = w ∑ ∣ h u x ′ ∣ 2 Γ x ′ → u ( z ) .
Since the sum is finite, the boundary value Δ ( E + i 0 ) \Delta(E+\mathrm{i}0) Δ ( E + i 0 ) exists. Every Γ x ′ → u \Gamma_{x'\to u} Γ x ′ → u maps H \mathbb{H} H to H \mathbb{H} H , and we have Im Γ x ′ → u ( E + i 0 ) ≥ 0 \operatorname{Im}\Gamma_{x'\to u}(E+\mathrm{i}0)\geq0 Im Γ x ′ → u ( E + i 0 ) ≥ 0 and Im Δ ( E + i 0 ) ≥ ∣ h u x ∣ 2 Im Γ x → u ( E + i 0 ) > 0 \operatorname{Im}\Delta(E+\mathrm{i}0)\geq|h_{ux}|^2\operatorname{Im}\Gamma_{x\to u}(E+\mathrm{i}0)>0 Im Δ ( E + i 0 ) ≥ ∣ h ux ∣ 2 Im Γ x → u ( E + i 0 ) > 0 . In particular Δ ( E + i 0 ) ≠ 0 \Delta(E+\mathrm{i}0)\neq0 Δ ( E + i 0 ) = 0 , so Γ u → w ( E + i 0 ) = − Δ ( E + i 0 ) − 1 \Gamma_{u\to w}(E+\mathrm{i}0)=-\Delta(E+\mathrm{i}0)^{-1} Γ u → w ( E + i 0 ) = − Δ ( E + i 0 ) − 1 , and Im Γ u → w ( E + i 0 ) = Im Δ ( E + i 0 ) / ∣ Δ ( E + i 0 ) ∣ 2 > 0 \operatorname{Im}\Gamma_{u\to w}(E+\mathrm{i}0)=\operatorname{Im}\Delta(E+\mathrm{i}0)/|\Delta(E+\mathrm{i}0)|^2>0 Im Γ u → w ( E + i 0 ) = Im Δ ( E + i 0 ) /∣Δ ( E + i 0 ) ∣ 2 > 0 .
Lemma 7.13. Fix a realization as in Lemma 7.12, and fix E ∈ R E \in\mathbb{R} E ∈ R such that Γ u → w ( E + i 0 ) \Gamma_{u\to w}(E+\mathrm{i}0) Γ u → w ( E + i 0 ) exists for every oriented edge ( u , w ) (u,w) ( u , w ) and Im Γ v ( E + i 0 ) > 0 \operatorname{Im}\Gamma_v(E+\mathrm{i}0)>0 Im Γ v ( E + i 0 ) > 0 for every vertex v v v whose forward subtree T v T_v T v is infinite. Let y y y be a vertex and x x x a neighbor of y y y whose component in V ∖ { y } V\setminus\{y\} V ∖ { y } is infinite. Then Im Γ x → y ( E + i 0 ) > 0 \operatorname{Im}\Gamma_{x\to y}(E+\mathrm{i}0)>0 Im Γ x → y ( E + i 0 ) > 0 .
Proof. If x x x is a child of y y y , then the component of x x x is V ( T x ) V(T_x) V ( T x ) and Γ x → y = Γ x \Gamma_{x\to y}=\Gamma_x Γ x → y = Γ x ; the claim is then part of the hypothesis. Otherwise y ≠ o y\ne o y = o and x = y − x=y^- x = y − . Let o = a 0 , a 1 , … , a k = y − o=a_0,a_1,\ldots,a_k=y^- o = a 0 , a 1 , … , a k = y − be the path from the root to y − y^- y − , and set a k + 1 = y a_{k+1}=y a k + 1 = y . The component of y − y^- y − is V ∖ V ( T y ) V\setminus V(T_y) V ∖ V ( T y ) , and it is the union of the path { a 0 , … , a k } \{a_0,\ldots,a_k\} { a 0 , … , a k } and the sets V ( T w ) V(T_w) V ( T w ) , where w w w ranges over the children of the a j a_j a j (0 ≤ j ≤ k 0\le j\le k 0 ≤ j ≤ k ) with w ≠ a j + 1 w\ne a_{j+1} w = a j + 1 . There are finitely many such w w w , since the path is finite and degrees are finite. As the component is infinite, one of these forward subtrees, say T w T_w T w with w w w a child of a j a_j a j and w ≠ a j + 1 w\ne a_{j+1} w = a j + 1 , is infinite, and Im Γ w ( E + i 0 ) > 0 \operatorname{Im}\Gamma_w(E+\mathrm{i}0)>0 Im Γ w ( E + i 0 ) > 0 by hypothesis.
We show by induction on i = j , … , k i=j,\ldots,k i = j , … , k that Im Γ a i → a i + 1 ( E + i 0 ) > 0 \operatorname{Im}\Gamma_{a_i\to a_{i+1}}(E+\mathrm{i}0)>0 Im Γ a i → a i + 1 ( E + i 0 ) > 0 . For i = j i=j i = j , the vertex w w w is a neighbor of a j a_j a j different from a j + 1 a_{j+1} a j + 1 , and Γ w → a j = Γ w \Gamma_{w\to a_j}=\Gamma_w Γ w → a j = Γ w , so Lemma 7.12 applies. For j ≤ i < k j\le i<k j ≤ i < k , the vertex a i a_i a i is a neighbor of a i + 1 a_{i+1} a i + 1 different from a i + 2 a_{i+2} a i + 2 , and Lemma 7.12 together with the induction hypothesis gives the claim for i + 1 i+1 i + 1 . The case i = k i=k i = k is Im Γ y − → y ( E + i 0 ) > 0 \operatorname{Im}\Gamma_{y^-\to y}(E+\mathrm{i}0)>0 Im Γ y − → y ( E + i 0 ) > 0 .
Proof of Proposition 7.11. The proof establishes, at almost every energy of I I I , that the row measures ν y \nu_y ν y have Borel transforms with positive imaginary boundary values, and then applies the absolutely continuous criterion of Lemma 7.4(ii) with spectral averaging in U y U_y U y . Positivity on every infinite forward subtree comes from (126) and the zero–one law of Proposition 5.11; Lemma 7.13 transfers it to the attachment points of the infinite components of the vertex-deleted tree; and Lemma 7.6(i) together with Lemma 7.2 transfers it to the Borel transform of ν y \nu_y ν y at the point E 2 E^2 E 2 .
Step 1. Positivity on infinite forward subtrees. Let B 1 \mathcal{B}_1 B 1 be the set of ( ω , E ) ∈ Ω × I (\omega,E)\in\Omega\times I ( ω , E ) ∈ Ω × I such that, for every vertex v v v with T v T_v T v infinite, the boundary value Γ v ( E + i 0 ) \Gamma_v(E+\mathrm{i}0) Γ v ( E + i 0 ) exists and has positive imaginary part. By Lemma 7.5 and the rational Cauchy criterion (111) , the set of ( ω , E ) (\omega,E) ( ω , E ) at which Γ v ( E + i 0 ) \Gamma_v(E+\mathrm{i}0) Γ v ( E + i 0 ) exists is measurable and the boundary value is a measurable function on it. Since the event that T v T_v T v is infinite is measurable and there are countably many words v v v , the set B 1 \mathcal{B}_1 B 1 is jointly measurable. Let E ∈ I ∩ E p E\in I\cap E_p E ∈ I ∩ E p be an energy at which (126) holds; this is the case for Lebesgue-almost every E ∈ I E\in I E ∈ I , since the complement of E p E_p E p is Lebesgue-null (Definition 5.7). By Proposition 5.11(i), the probability in (126) equals q p q_p q p , and P ∞ ( ( ω , E ) ∈ B 1 ) = 1 \mathbb{P}^{\infty}((\omega,E)\in\mathcal{B}_1)=1 P ∞ (( ω , E ) ∈ B 1 ) = 1 . By Tonelli’s theorem,
E ∞ { ∣ E ∈ I : ( ω , E ) ∉ B 1 ∣ } = ∫ I P ∞ ( ( ω , E ) ∉ B 1 ) d E = 0. \mathbb{E}^{\infty}\left\{\left|E\in I:(\omega,E)\notin\mathcal{B}_1\right|\right\}=\int_I\mathbb{P}^{\infty}\left((\omega,E)\notin\mathcal{B}_1\right)\,\mathrm{d}E=0. E ∞ { ∣ E ∈ I : ( ω , E ) ∈ / B 1 ∣ } = ∫ I P ∞ ( ( ω , E ) ∈ / B 1 ) d E = 0.
Further, each Γ u → w \Gamma_{u\to w} Γ u → w is the Borel transform of a spectral measure. By Lemma 7.1(i), for every realization and every oriented edge, Γ u → w ( E + i 0 ) \Gamma_{u\to w}(E+\mathrm{i}0) Γ u → w ( E + i 0 ) exists for Lebesgue-almost every E E E ; there are countably many oriented edges. We conclude that there exists an event Ω 1 \Omega_1 Ω 1 with P ∞ ( Ω 1 ) = 1 \mathbb{P}^{\infty}(\Omega_1)=1 P ∞ ( Ω 1 ) = 1 , contained in Ω s a \Omega_{\mathrm{sa}} Ω sa , in the event of Lemma 7.6, and in the event that all edge weights are nonzero and T \mathcal{T} T is infinite, with the following property. For every ω ∈ Ω 1 \omega\in\Omega_1 ω ∈ Ω 1 there exists a Lebesgue-null set N ( ω ) ⊂ I \mathcal{N}(\omega) \subset I N ( ω ) ⊂ I such that for every E ∈ I ∖ N ( ω ) E \in I \setminus\mathcal{N}(\omega) E ∈ I ∖ N ( ω ) , all boundary values Γ u → w ( E + i 0 ) \Gamma_{u\to w}(E+i0) Γ u → w ( E + i 0 ) exist, and Im Γ v ( E + i 0 ) > 0 \operatorname{Im}\Gamma_v(E+i0)>0 Im Γ v ( E + i 0 ) > 0 for every v v v with T v \mathcal{T}_v T v infinite.
Step 2. Positivity for the row measures. Fix a word y y y , and let ν y ′ \nu'_y ν y ′ be the kernel of Lemma 7.8(i). Let T y ′ T'_y T y ′ be the set of λ ∈ I 2 \lambda\in I^2 λ ∈ I 2 such that the boundary value F ν y ′ ( λ + i 0 ) F_{\nu'_y}(\lambda+i0) F ν y ′ ( λ + i 0 ) does not exist or has imaginary part zero. By Lemma 7.3(iv), applied with ( Ω ′ , A ) = ( Ω , G y ) (\Omega',\mathcal{A})=(\Omega,\mathcal{G}^y) ( Ω ′ , A ) = ( Ω , G y ) , L = L y L=L^y L = L y and v = 1 v=1 v = 1 , the set { ( ω , λ ) : λ ∈ T y ′ } \{(\omega,\lambda):\lambda\in T'_y\} {( ω , λ ) : λ ∈ T y ′ } is G y ⊗ B ( R ) \mathcal{G}^y\otimes\mathcal{B}(\mathbb{R}) G y ⊗ B ( R ) -measurable. We claim that ∣ T y ′ ∣ = 0 |T'_y|=0 ∣ T y ′ ∣ = 0 , P ∞ \mathbb{P}^{\infty} P ∞ -almost surely on Ω y \Omega_y Ω y .
To that end, fix ω ∈ Ω 1 ∩ Ω y \omega\in\Omega_1\cap\Omega_y ω ∈ Ω 1 ∩ Ω y such that ν y ′ = ν y \nu'_y=\nu_y ν y ′ = ν y , so that F ν y ′ = m y F_{\nu'_y}=m_y F ν y ′ = m y , and fix E ∈ I ∖ N ( ω ) E\in I\setminus\mathcal{N}(\omega) E ∈ I ∖ N ( ω ) . The set V ∖ { y } V\setminus\{y\} V ∖ { y } is the union of the components of the deg ( y ) < ∞ \deg(y)<\infty deg ( y ) < ∞ neighbors of y y y , and V V V is infinite; hence some neighbor x 0 x_0 x 0 of y y y has an infinite component. By Lemma 7.13, Im Γ x 0 → y ( E + i 0 ) > 0 \operatorname{Im}\Gamma_{x_0\to y}(E+i0)>0 Im Γ x 0 → y ( E + i 0 ) > 0 . By (7.18) with z = E + i η z=E+i\eta z = E + i η , the limit
ℓ = lim η ↓ 0 m y ( ( E + i η ) 2 ) = 1 E U y ∑ x ∼ y ∣ h x y ∣ 2 Γ x → y ( E + i 0 ) \ell=\lim_{\eta\downarrow0}m_y\left((E+i\eta)^2\right)=\frac{1}{\mathbb{E}U_y}\sum_{x\sim y}|h_{xy}|^2\Gamma_{x\to y}(E+i0) ℓ = η ↓ 0 lim m y ( ( E + i η ) 2 ) = E U y 1 x ∼ y ∑ ∣ h x y ∣ 2 Γ x → y ( E + i 0 )
exists, and Im ℓ ≥ ( E U y ) − 1 ∣ h x 0 y ∣ 2 Im Γ x 0 → y ( E + i 0 ) > 0 \operatorname{Im}\ell\geq(\mathbb{E}U_y)^{-1}|h_{x_0y}|^2\operatorname{Im}\Gamma_{x_0\to y}(E+i0)>0 Im ℓ ≥ ( E U y ) − 1 ∣ h x 0 y ∣ 2 Im Γ x 0 → y ( E + i 0 ) > 0 , because all terms have nonnegative imaginary part and h x 0 y ≠ 0 h_{x_0y}\ne0 h x 0 y = 0 . By Lemma 7.2, applied to ν y \nu_y ν y , the boundary value m y ( E 2 + i 0 ) m_y(E^2+i0) m y ( E 2 + i 0 ) exists and equals ℓ \ell ℓ , and E 2 ∉ T y ′ E^2\notin T'_y E 2 ∈ / T y ′ . In particular, T y ′ T'_y T y ′ is contained in { E 2 : E ∈ N ( ω ) } \{E^2:E\in\mathcal{N}(\omega)\} { E 2 : E ∈ N ( ω )} , which is Lebesgue-null. Since P ∞ ( Ω 1 ) = 1 \mathbb{P}^{\infty}(\Omega_1)=1 P ∞ ( Ω 1 ) = 1 and ν y ′ = ν y \nu'_y=\nu_y ν y ′ = ν y almost surely on Ω y \Omega_y Ω y , this proves the claim.
Step 3. Spectral averaging. Fix a word y y y , let ϵ \epsilon ϵ be the parity of its neighbors, and let ( μ u y ϵ ) (\mu^\epsilon_{u_y}) ( μ u y ϵ ) be the rank-one family of (7.23), so that μ u y 0 = ν y ′ \mu^0_{u_y}=\nu'_y μ u y 0 = ν y ′ and, almost surely on Ω y \Omega_y Ω y , μ u y ϵ y = μ u y y \mu^{\epsilon_y}_{u_y}=\mu^y_{u_y} μ u y ϵ y = μ u y y . By Lemma 7.4(ii) with L = L y L=L^y L = L y and S = I 2 S=I^2 S = I 2 , the restriction of μ u y \mu^y_u μ u y to I 2 I^2 I 2 is absolutely continuous whenever μ u y ( T y ′ ) = 0 \mu^y_u(T'_y)=0 μ u y ( T y ′ ) = 0 . As in Step 3 of the proof of Proposition 7.9, the function Φ ( ω , u ) = 1 { μ u y ( T y ′ ) > 0 } \Phi(\omega,u)=\mathbf{1}\{\mu^y_u(T'_y)>0\} Φ ( ω , u ) = 1 { μ u y ( T y ′ ) > 0 } is G y ⊗ B ( R ) \mathcal{G}^y\otimes\mathcal{B}(\mathbb{R}) G y ⊗ B ( R ) -measurable, it vanishes for Lebesgue-almost every u u u when ∣ T y ′ ∣ = 0 |T'_y|=0 ∣ T y ′ ∣ = 0 , and (7.22) with Q = P ∞ Q=\mathbb{P}^{\infty} Q = P ∞ gives P ∞ ( { μ u y ( T y ′ ) > 0 } ∩ Ω y ) = 0 \mathbb{P}^{\infty}(\{\mu^y_u(T'_y)>0\}\cap\Omega_y)=0 P ∞ ({ μ u y ( T y ′ ) > 0 } ∩ Ω y ) = 0 . It follows that, P ∞ \mathbb{P}^{\infty} P ∞ -almost surely, the restriction of μ u y ϵ y \mu^{\epsilon_y}_{u_y} μ u y ϵ y to I 2 I^2 I 2 is absolutely continuous for every vertex y y y with deg ( y ) ≥ 1 \deg(y)\geq1 deg ( y ) ≥ 1 .
Step 4. Conclusion. As in Step 1 of the proof of Proposition 7.9, the row directions of the vertices in V 1 − ϵ V_{1-\epsilon} V 1 − ϵ span ( ker S ϵ ) ⊥ (\ker S_\epsilon)^\perp ( ker S ϵ ) ⊥ , and 0 ∉ I 2 0\notin I^2 0 ∈ / I 2 . By Lemma 7.7, each square block S ϵ S_\epsilon S ϵ is purely absolutely continuous in I 2 I^2 I 2 , and by Proposition 3.23(vi), H \mathcal{H} H is purely absolutely continuous in I ∪ ( − I ) I\cup(-I) I ∪ ( − I ) , P ∞ \mathbb{P}^{\infty} P ∞ -almost surely. □
Remark 7.14. The hypothesis of Proposition 7.11 is a statement about the unconditioned law at almost every energy of I I I , and the conclusion is a statement about the survival-conditioned operator. On the extinction event the tree is finite and the spectrum is pure point, and the conclusion does not hold under P \mathbb{P} P . The energy 0 is excluded: H \mathcal{H} H has finitely supported eigenvectors with eigenvalue zero (Proposition 3.25).
Negative free energy implies localization
In this section, we prove that a real solution of the cavity equation with negative free energy at s = 1 s = 1 s = 1 forces pure point spectrum. The precise statement is Theorem 8.4 : if for almost every energy E E E in a set U ⊂ ( 0 , ∞ ) U \subset(0,\infty) U ⊂ ( 0 , ∞ ) there exists a real solution π \pi π at E E E with φ π ( 1 ) < 0 \varphi_{\pi}(1) < 0 φ π ( 1 ) < 0 , then H \mathcal{H} H has pure point spectrum in U ∪ ( − U ) U \cup(-U) U ∪ ( − U ) , almost surely under P \mathbb{P} P and under P ∞ \mathbb{P}^{\infty} P ∞ . The real solution is not assumed to be related to the boundary values of the resolvent of H \mathcal{H} H , and the conclusion concerns only the operator.
We now describe the argument. Fix E > 0 E > 0 E > 0 and a real solution π \pi π at E E E . By Lemma 5.13 , there exists a family of real denominators ( D v ) (D_v) ( D v ) on the tree, with marginal law π \pi π , satisfying the cavity recursion D v = E − p − 1 ∑ i Y v , v i / D v i D_v = E - p^{-1} \sum_i Y_{v,v_i}/D_{v_i} D v = E − p − 1 ∑ i Y v , v i / D v i at every vertex. It is defined on an extension of the probability space of the marked tree. Setting γ v = − 1 / D v \gamma_v = -1/D_v γ v = − 1/ D v , the products
F ( v ) = γ o ∏ j = 1 L ( − h v j v j − 1 γ v j ) , o = v 0 , v 1 , … , v L = v , F(v) = \gamma_o \prod_{j=1}^{L}(-h_{v_jv_{j-1}}\gamma_{v_j}), \qquad o = v_0,v_1,\ldots,v_L = v, F ( v ) = γ o j = 1 ∏ L ( − h v j v j − 1 γ v j ) , o = v 0 , v 1 , … , v L = v ,
define a function F F F on the vertices that solves ( H − E ) F = δ o (\mathcal{H} - E)F = \delta_o ( H − E ) F = δ o pointwise. We call it the formal inverse column. By the many-to-one formula, the expected generation-L L L sum of ∣ F ( v ) ∣ s \lvert F(v)\rvert^s ∣ F ( v ) ∣ s is the path moment Φ L ( s ; E ) \Phi_L(s;E) Φ L ( s ; E ) of Definition 6.24 , whose exponential growth rate is the free energy φ π ( s ) \varphi_{\pi}(s) φ π ( s ) by Proposition 6.27 . If φ π ( 1 ) < 0 \varphi_{\pi}(1) < 0 φ π ( 1 ) < 0 , we may choose s < 1 s < 1 s < 1 with φ π ( s ) < 0 \varphi_{\pi}(s) < 0 φ π ( s ) < 0 , and then F ∈ ℓ s ⊂ ℓ 2 F \in\ell^s \subset\ell^2 F ∈ ℓ s ⊂ ℓ 2 almost surely. Using the maximal domain characterization of Theorem 3.10 , we then obtain δ o = ( H − E ) F \delta_o = (\mathcal{H} - E)F δ o = ( H − E ) F , and this implies that ∫ μ o ( d λ ) / ( λ − E ) 2 < ∞ \int\mu_o(\mathrm{d}\lambda)/(\lambda- E)^2 < \infty ∫ μ o ( d λ ) / ( λ − E ) 2 < ∞ . The Simon–Wolff criterion of Proposition 7.9 requires the same statement for every neighbor of every removed vertex. We obtain it by rerooting: on a component of the forest obtained by removing a vertex or an edge, rerooted at any of its vertices, the inverse column is computed from finitely many real cavity marks along an ancestral path together with finitely many forward columns hanging off that path.
In § 8.2.1 , we define the formal inverse columns and state the main results, Proposition 8.3 and Theorem 8.4 . § 8.2.2 proves the many-to-one identity. § 8.2.3 proves summability of the forward columns, and § 8.2.4 constructs the rerooted columns. § 8.2.5 completes the proofs.
Throughout this section, p > 1 p > 1 p > 1 is fixed.
The formal inverse column
We begin with a deterministic lemma, which is the real-energy analogue of the Schur complement formula of Proposition 3.16 read as an equation for a column of the resolvent.
Lemma 8.1. Let C C C be a countable, connected, locally finite tree with edge weights h x y = h y x ‾ h_{xy}=\overline{h_{yx}} h x y = h y x , let r ∈ C r\in C r ∈ C , and let E ∈ R E\in\mathbb{R} E ∈ R . For x ∈ C ∖ { r } x\in C\setminus\{r\} x ∈ C ∖ { r } , let b ( x ) b(x) b ( x ) be the neighbor of x x x on the path from x x x to r r r , and let ch ( x ) \operatorname{ch}(x) ch ( x ) be the set of neighbors x ′ x' x ′ of x x x with x ′ ≠ b ( x ) x'\ne b(x) x ′ = b ( x ) ; let ch ( r ) \operatorname{ch}(r) ch ( r ) be the set of all neighbors of r r r . Suppose that real numbers ( γ x ) x ∈ C (\gamma_x)_{x\in C} ( γ x ) x ∈ C satisfy, for every x ∈ C x\in C x ∈ C ,
Δ x = E + ∑ x ′ ∈ ch ( x ) ∣ h x x ′ ∣ 2 γ x ′ ≠ 0 , γ x = − 1 Δ x . (127) \Delta_x=E+\sum_{x'\in\operatorname{ch}(x)}\lvert h_{xx'}\rvert^2\gamma_{x'}\ne0,\qquad\gamma_x=-\frac{1}{\Delta_x}. \tag*{(127)} Δ x = E + x ′ ∈ ch ( x ) ∑ ∣ h x x ′ ∣ 2 γ x ′ = 0 , γ x = − Δ x 1 . (127) Define f ( r ) = γ r f(r)=\gamma_r f ( r ) = γ r , and f ( x ) = − γ x h x b ( x ) f ( b ( x ) ) f(x)=-\gamma_xh_{xb(x)}f(b(x)) f ( x ) = − γ x h x b ( x ) f ( b ( x )) for x ≠ r x\ne r x = r , recursively along the path from r r r to x x x . Then for every x ∈ C x\in C x ∈ C ,
∑ x ′ ∼ x h x x ′ f ( x ′ ) − E f ( x ) = δ r ( x ) . (128) \sum_{x'\sim x}h_{xx'}f(x')-Ef(x)=\delta_r(x). \tag*{(128)} x ′ ∼ x ∑ h x x ′ f ( x ′ ) − E f ( x ) = δ r ( x ) . (128)
Proof. Fix x ∈ C x\in C x ∈ C . For x ′ ∈ ch ( x ) x'\in\operatorname{ch}(x) x ′ ∈ ch ( x ) we have b ( x ′ ) = x b(x')=x b ( x ′ ) = x , hence f ( x ′ ) = − γ x ′ h x ′ x f ( x ) f(x')=-\gamma_{x'}h_{x'x}f(x) f ( x ′ ) = − γ x ′ h x ′ x f ( x ) and h x x ′ f ( x ′ ) = − ∣ h x x ′ ∣ 2 γ x ′ f ( x ) h_{xx'}f(x')=-\lvert h_{xx'}\rvert^2\gamma_{x'}f(x) h x x ′ f ( x ′ ) = − ∣ h x x ′ ∣ 2 γ x ′ f ( x ) . The set ch ( x ) \operatorname{ch}(x) ch ( x ) is finite, and by (8.1),
∑ x ′ ∈ ch ( x ) h x x ′ f ( x ′ ) − E f ( x ) = − Δ x f ( x ) = f ( x ) γ x . \sum_{x'\in\operatorname{ch}(x)}h_{xx'}f(x')-Ef(x)=-\Delta_xf(x)=\frac{f(x)}{\gamma_x}. x ′ ∈ ch ( x ) ∑ h x x ′ f ( x ′ ) − E f ( x ) = − Δ x f ( x ) = γ x f ( x ) .
If x = r x=r x = r , then ch ( r ) \operatorname{ch}(r) ch ( r ) contains all neighbors of r r r , and the right side equals γ r / γ r = 1 \gamma_r/\gamma_r=1 γ r / γ r = 1 . If x ≠ r x\ne r x = r , the only neighbor of x x x outside ch ( x ) \operatorname{ch}(x) ch ( x ) is b ( x ) b(x) b ( x ) , and f ( x ) / γ x = − h x b ( x ) f ( b ( x ) ) f(x)/\gamma_x=-h_{xb(x)}f(b(x)) f ( x ) / γ x = − h x b ( x ) f ( b ( x )) ; the left side of (8.2) then equals − h x b ( x ) f ( b ( x ) ) + h x b ( x ) f ( b ( x ) ) = 0 -h_{xb(x)}f(b(x))+h_{xb(x)}f(b(x))=0 − h x b ( x ) f ( b ( x )) + h x b ( x ) f ( b ( x )) = 0 . □
We now fix E > 0 E>0 E > 0 and a real solution π \pi π at energy E E E (Definition 5.1), and we let ( ω , ( D v ) v ∈ U ) (\omega,(D_v)_{v\in\mathbb{U}}) ( ω , ( D v ) v ∈ U ) be a recursive tree process with marginal π \pi π , as in Lemma 5.13. We use the realization of the marked tree of Definition 4.1 and Section 5.3: the edge joining v v v and v i vi v i carries the mark w v , v i = w v i w_{v,vi}=w_{vi} w v , v i = w v i , the weight h v , v i = w v , v i / p h_{v,vi}=w_{v,vi}/\sqrt{p} h v , v i = w v , v i / p and the squared modulus Y v , v i = ∣ w v , v i ∣ 2 Y_{v,vi}=\lvert w_{v,vi}\rvert^2 Y v , v i = ∣ w v , v i ∣ 2 . As in the proof of Lemma 5.13, the innovation at a word v v v is ζ ~ v = ( N v , ζ v , ( w v , v i ) i ≥ 1 ) \widetilde{\zeta}_v=(N_v,\zeta_v,(w_{v,vi})_{i\ge1}) ζ v = ( N v , ζ v , ( w v , v i ) i ≥ 1 ) , and for every word w w w the marked tree T w \mathcal{T}_w T w rooted at w w w is built from the innovations at the words in w U = { w u : u ∈ U } w\mathbb{U}=\{wu:u\in\mathbb{U}\} w U = { w u : u ∈ U } . For w ∈ V w\in V w ∈ V , it is the forward subtree of w w w . By Lemma 5.13, the configuration ω \omega ω has the law P \mathbb{P} P , so the marked tree T \mathcal{T} T has the law of Definition 3.2. We keep writing P \mathbb{P} P and E \mathbb{E} E for the probability measure and the expectation on the space of Lemma 5.13. We may do so because events concerning only the marked tree have the same probability there as under P \mathbb{P} P . For a word w w w , we write
Θ w = ( θ w ω , ( D w u ) u ∈ U ) \Theta_w=\left(\theta_w\omega,(D_{wu})_{u\in\mathbb{U}}\right) Θ w = ( θ w ω , ( D w u ) u ∈ U )
for the process seen from w w w , where θ w \theta_w θ w is the shift of Definition 4.1; it is in measurable bijection with ( ζ ~ w u , D w u ) u ∈ U (\widetilde{\zeta}_{wu},D_{wu})_{u\in\mathbb{U}} ( ζ w u , D w u ) u ∈ U , and Θ o = ( ω , ( D u ) u ∈ U ) \Theta_o=(\omega,(D_u)_{u\in\mathbb{U}}) Θ o = ( ω , ( D u ) u ∈ U ) is the process itself. We use the following three properties of the process, which are the conclusions of Lemma 5.13. (R1) Almost surely, every D v D_v D v is a nonzero real number, and the recursion D v = E − p − 1 ∑ i ≤ N v Y v , v i / D v i D_v = E - p^{-1}\sum_{i\le N_v} Y_{v,v_i}/D_{v_i} D v = E − p − 1 ∑ i ≤ N v Y v , v i / D v i holds at every word v v v .
(R2) For every n ≥ 0 n \ge0 n ≥ 0 , the variables D v D_v D v , ∣ v ∣ = n |v|=n ∣ v ∣ = n , are independent with law π \pi π , and they are independent of F n = σ ( ζ u : u < n ) \mathcal{F}_n=\sigma(\zeta_u:u<n) F n = σ ( ζ u : u < n ) .
(R3) If v 1 , … , v k v_1,\ldots,v_k v 1 , … , v k are words none of which is a prefix of another, then Θ v 1 , … , Θ v k \Theta_{v_1},\ldots,\Theta_{v_k} Θ v 1 , … , Θ v k are independent of each other and of the coordinates N u , ζ u N_u,\zeta_u N u , ζ u with u ∉ ⋃ i v i U u\notin\bigcup_i v_i\mathbb{U} u ∈ / ⋃ i v i U and w u w_u w u with u ∉ ⋃ i v i U ∗ u\notin\bigcup_i v_i\mathbb{U}_* u ∈ / ⋃ i v i U ∗ , where v U ∗ = v U ∖ { v } v\mathbb{U}_*=v\mathbb{U}\setminus\{v\} v U ∗ = v U ∖ { v } , and each Θ v i \Theta_{v_i} Θ v i has the law of Θ o \Theta_o Θ o .
We set
γ v = − 1 D v , v ∈ U . (129) \gamma_v=-\frac{1}{D_v},\qquad v\in\mathbb{U}. \tag*{(129)} γ v = − D v 1 , v ∈ U . (129)
Definition 8.2. For a word w w w , the forward formal column F w F_w F w is the function on V ( T w ) V(\mathcal{T}_w) V ( T w ) defined by F w ( w ) = γ w F_w(w)=\gamma_w F w ( w ) = γ w and
F w ( v i ) = − γ v i h v i , v F w ( v ) , v i ∈ V ( T w ) , v ∈ V ( T w ) , F_w(vi)=-\gamma_{vi}h_{vi,v}F_w(v),\qquad vi\in V(\mathcal{T}_w),\ v\in V(\mathcal{T}_w), F w ( v i ) = − γ v i h v i , v F w ( v ) , v i ∈ V ( T w ) , v ∈ V ( T w ) , and extended by zero to the other words. Equivalently, if w = u 0 , u 1 , … , u m = u w=u_0,u_1,\ldots,u_m=u w = u 0 , u 1 , … , u m = u is the path from w w w to u ∈ V ( T w ) u\in V(\mathcal{T}_w) u ∈ V ( T w ) , then
F w ( u ) = γ w ∏ j = 1 m ( − h u j u j − 1 γ u j ) . (130) F_w(u)=\gamma_w\prod_{j=1}^{m}(-h_{u_j u_{j-1}}\gamma_{u_j}). \tag*{(130)} F w ( u ) = γ w j = 1 ∏ m ( − h u j u j − 1 γ u j ) . (130)
The formal inverse column is F = F o F=F_o F = F o .
The function F w F_w F w is obtained from Θ w \Theta_w Θ w by the same rule by which F F F is obtained from Θ o \Theta_o Θ o . By the recursion (R1), for every x ∈ V ( T w ) x\in V(\mathcal{T}_w) x ∈ V ( T w ) we have
E + ∑ i = 1 N x ∣ h x , x i ∣ 2 γ x i = E − 1 p ∑ i = 1 N x Y x , x i D x i = D x = − 1 γ x . E+\sum_{i=1}^{N_x}|h_{x,x_i}|^2\gamma_{x_i}=E-\frac{1}{p}\sum_{i=1}^{N_x}\frac{Y_{x,x_i}}{D_{x_i}}=D_x=-\frac{1}{\gamma_x}. E + i = 1 ∑ N x ∣ h x , x i ∣ 2 γ x i = E − p 1 i = 1 ∑ N x D x i Y x , x i = D x = − γ x 1 .
This means that the numbers γ x \gamma_x γ x , x ∈ V ( T w ) x\in V(\mathcal{T}_w) x ∈ V ( T w ) , satisfy (8.1) on the tree T w \mathcal{T}_w T w with root r = w r=w r = w , for which b ( x ) = x − b(x)=x^- b ( x ) = x − and ch ( x ) \operatorname{ch}(x) ch ( x ) is the set of children of x x x . By Lemma 8.1, almost surely, for every word w w w ,
( H w , f o r m F w ) ( x ) − E F w ( x ) = δ w ( x ) , x ∈ V ( T w ) , (131) (\mathcal{H}_{w,\mathrm{form}}F_w)(x)-EF_w(x)=\delta_w(x),\qquad x\in V(\mathcal{T}_w), \tag*{(131)} ( H w , form F w ) ( x ) − E F w ( x ) = δ w ( x ) , x ∈ V ( T w ) , (131)
where H w , f o r m \mathcal{H}_{w,\mathrm{form}} H w , form is the formal weighted adjacency operator of T w \mathcal{T}_w T w . In general F w F_w F w need not belong to ℓ 2 \ell^2 ℓ 2 . The point of the next proposition is that it does when φ π ( 1 ) < 0 \varphi_\pi(1)<0 φ π ( 1 ) < 0 , that is, when the free energy of π \pi π at s = 1 s=1 s = 1 is negative.
Recall the path moments of Definition 6.24: for s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) , L ≥ 0 L\ge0 L ≥ 0 and a scaling g > 0 g>0 g > 0 ,
Φ L ( s ; E ) = g s ∫ R ( A s L e s ) ( x ) π x ( d x ) . (132) \Phi_L(s;E)=g^s\int_{\mathbb{R}}(\mathcal{A}_s^L e_s)(x)\,\pi^x(dx). \tag*{(132)} Φ L ( s ; E ) = g s ∫ R ( A s L e s ) ( x ) π x ( d x ) . (132)
where e s ( x ) = ∣ x ∣ s e_s(x)=|x|^s e s ( x ) = ∣ x ∣ s , π x \pi^x π x is the law of X = 1 / ( g D ) X=1/(gD) X = 1/ ( g D ) with D ∼ π D\sim\pi D ∼ π , and A s = A s ( g ) \mathcal{A}_s=\mathcal{A}_s^{(g)} A s = A s ( g ) is the operator of Definition 6.1 for the law of B = g D B=gD B = g D and the rate K = p / g 2 K=p/g^2 K = p / g 2 , so that β = g 2 γ / p \beta=g^2\gamma/p β = g 2 γ / p with γ ∼ Exp ( 1 ) \gamma\sim\operatorname{Exp}(1) γ ∼ Exp ( 1 ) . By Lemma 6.23 (ii), Φ L ( s ; E ) \Phi_L(s;E) Φ L ( s ; E ) does not depend on g g g . For a self-adjoint operator T T T and a vector ψ \psi ψ , we use the reciprocal-square integral R μ ψ T ( E ) = ∫ μ ψ T ( d λ ) / ( λ − E ) 2 R_{\mu_\psi^T}(E)=\int\mu_\psi^T(\mathrm{d}\lambda)/(\lambda-E)^2 R μ ψ T ( E ) = ∫ μ ψ T ( d λ ) / ( λ − E ) 2 of (103) , which is infinite if μ ψ T ( { E } ) > 0 \mu_\psi^T(\{E\})>0 μ ψ T ({ E }) > 0 . For a component C C C of a forest obtained by a finite removal and r ∈ C r\in C r ∈ C , we write μ r C \mu_r^C μ r C for the spectral measure of δ r \delta_r δ r for H C \mathcal{H}_C H C (Theorem 3.10 ).
Proposition 8.3. Let E > 0 E>0 E > 0 , let π \pi π be a real solution at energy E E E , let ( ζ v , D v ) v ∈ U (\zeta_v,D_v)_{v\in\mathcal{U}} ( ζ v , D v ) v ∈ U be a recursive tree process with marginal π \pi π (Lemma 5.13 ), and let F F F be the formal inverse column of Definition 8.2 .
(i) For every s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) and L ≥ 0 L\geq0 L ≥ 0 ,
E ∑ v ∈ V : ∣ v ∣ = L ∣ F ( v ) ∣ s = Φ L ( s ; E ) . (133) \mathbb{E}\sum_{v\in V:\,|v|=L}|F(v)|^s=\Phi_L(s;E). \tag*{(133)} E v ∈ V : ∣ v ∣ = L ∑ ∣ F ( v ) ∣ s = Φ L ( s ; E ) . (133) Suppose in addition that φ π ( 1 ) < 0 \varphi_\pi(1)<0 φ π ( 1 ) < 0 . Then the following hold.
(ii) There exists s 0 ∈ ( 0 , 1 ) s_0\in(0,1) s 0 ∈ ( 0 , 1 ) such that φ π ( s ) < 0 \varphi_\pi(s)<0 φ π ( s ) < 0 for s ∈ [ s 0 , 1 ) s\in[s_0,1) s ∈ [ s 0 , 1 ) . For every such s s s , almost surely F ∈ ℓ s ( V ) F\in\ell^s(V) F ∈ ℓ s ( V ) ; moreover F ∈ Dom ( H ) F\in\operatorname{Dom}(\mathcal{H}) F ∈ Dom ( H ) and ( H − E ) F = δ o (\mathcal{H}-E)F=\delta_o ( H − E ) F = δ o .
(iii) Almost surely under P \mathbb{P} P and under P ∞ \mathbb{P}^\infty P ∞ , the following holds. For every vertex v v v , we have δ v ∈ Ran ( H − E ) \delta_v\in\operatorname{Ran}(\mathcal{H}-E) δ v ∈ Ran ( H − E ) and ∫ μ v ( d λ ) / ( λ − E ) 2 < ∞ \int\mu_v(\mathrm{d}\lambda)/(\lambda-E)^2<\infty ∫ μ v ( d λ ) / ( λ − E ) 2 < ∞ . For every vertex y y y and every edge e e e of T \mathcal{T} T , every component C C C of the reduced forest obtained by removing y y y (or e e e ), and every r ∈ C r\in C r ∈ C , we have δ r ∈ Ran ( H C − E ) \delta_r\in\operatorname{Ran}(\mathcal{H}_C-E) δ r ∈ Ran ( H C − E ) and ∫ μ r C ( d λ ) / ( λ − E ) 2 < ∞ \int\mu_r^C(\mathrm{d}\lambda)/(\lambda-E)^2<\infty ∫ μ r C ( d λ ) / ( λ − E ) 2 < ∞ . In particular, for every vertex y y y and every neighbor x x x of y y y ,
∫ R μ x ( y ) ( d λ ) ( λ − E ) 2 < ∞ . (134) \int_{\mathbb{R}}\frac{\mu_x^{(y)}(\mathrm{d}\lambda)}{(\lambda-E)^2}<\infty. \tag*{(134)} ∫ R ( λ − E ) 2 μ x ( y ) ( d λ ) < ∞. (134)
We note that the assertions in (iii) concern only the marked tree and do not refer to π \pi π . Only their proof uses the recursive tree process. Combining Proposition 8.3 with the Simon–Wolff criterion of Proposition 7.9, we obtain the main result of this section.
Theorem 8.4. Let U ⊂ ( 0 , ∞ ) U\subset(0,\infty) U ⊂ ( 0 , ∞ ) be a Borel set (for example, an open set). Suppose that for Lebesgue-almost every E ∈ U E\in U E ∈ U there exists a real solution π E \pi_E π E at energy E E E with φ π E ( 1 ) < 0 \varphi_{\pi_E}(1)<0 φ π E ( 1 ) < 0 . Then, almost surely under P \mathbb{P} P and under P ∞ \mathbb{P}^\infty P ∞ , the operator H \mathcal{H} H has pure point spectrum in U ∪ ( − U ) U\cup(-U) U ∪ ( − U ) .
The many-to-one identity
We prove Proposition 8.3 (i). For a vertex v v v with ∣ v ∣ = L |v|=L ∣ v ∣ = L and path o = v 0 , … , v L = v o=v_0,\ldots,v_L=v o = v 0 , … , v L = v , and for s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) , set
W L ( v ) = ∏ j = 1 L ( γ v j − 1 , v j p ) s / 2 ∣ D v j ∣ − s , W 0 ( o ) = 1. (135) W_L(v)=\prod_{j=1}^{L}\left(\frac{\gamma_{v_{j-1},v_j}}{p}\right)^{s/2}|D_{v_j}|^{-s},\qquad W_0(o)=1. \tag*{(135)} W L ( v ) = j = 1 ∏ L ( p γ v j − 1 , v j ) s /2 ∣ D v j ∣ − s , W 0 ( o ) = 1. (135)
Since ∣ h v j v j − 1 ∣ s = ( Y v j − 1 , v j / p ) s / 2 |h_{v_jv_{j-1}}|^s=(Y_{v_{j-1},v_j}/p)^{s/2} ∣ h v j v j − 1 ∣ s = ( Y v j − 1 , v j / p ) s /2 and ∣ γ v ∣ = ∣ D v ∣ − 1 |\gamma_v|=|D_v|^{-1} ∣ γ v ∣ = ∣ D v ∣ − 1 , (8.4) gives
∣ F ( v ) ∣ s = ∣ D o ∣ − s W L ( v ) . (136) |F(v)|^s=|D_o|^{-s}W_L(v). \tag*{(136)} ∣ F ( v ) ∣ s = ∣ D o ∣ − s W L ( v ) . (136)
Lemma 8.5. Let s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) and g > 0 g>0 g > 0 , and set X o = 1 / ( g D o ) X_o=1/(gD_o) X o = 1/ ( g D o ) . For every L ≥ 0 L\geq0 L ≥ 0 and every Borel function k : R → [ 0 , ∞ ) k:\mathbb{R}\to[0,\infty) k : R → [ 0 , ∞ ) ,
E ∑ v ∈ V : ∣ v ∣ = L k ( X o ) W L ( v ) = ∫ R ( A g L k ) ( x ) π x ( d x ) . (137) \mathbb{E}\sum_{v\in V:|v|=L}k(X_o)W_L(v)=\int_{\mathbb{R}}(\mathcal{A}_g^Lk)(x)\,\pi^x(\mathrm{d}x). \tag*{(137)} E v ∈ V : ∣ v ∣ = L ∑ k ( X o ) W L ( v ) = ∫ R ( A g L k ) ( x ) π x ( d x ) . (137)
The identity (8.11) is the many-to-one formula of Proposition 4.8 for the recursive tree process; compare Corollary 4.11 and Remark 4.13. We give a direct proof by induction on L L L , using the Mecke formula at the root, which works with the consistent process of Lemma 5.13 on the whole tree and so requires no truncation. (Alternatively, the left side of (8.11) is a function of ( ζ u ) ∣ u ∣ < L (\zeta_u)_{|u|<L} ( ζ u ) ∣ u ∣ < L and ( D u ) ∣ u ∣ = L (D_u)_{|u|=L} ( D u ) ∣ u ∣ = L only, by the recursion (R1). By (R2), the variables ( D u ) ∣ u ∣ = L (D_u)_{|u|=L} ( D u ) ∣ u ∣ = L are independent with law π \pi π and independent of ( ζ u ) ∣ u ∣ < L (\zeta_u)_{|u|<L} ( ζ u ) ∣ u ∣ < L , and after integrating them out, Proposition 4.8 applies to the resulting function of the first L L L generations. The offspring numbers N u N_u N u with ∣ u ∣ = L |u|=L ∣ u ∣ = L do not enter.)
Proof. We prove (8.11) by induction on L L L , simultaneously for all k k k . For L = 0 L=0 L = 0 , the left side is E k ( X o ) = ∫ k d π x \mathbb{E}k(X_o)=\int k\,\mathrm{d}\pi^x E k ( X o ) = ∫ k d π x , since D o ∼ π D_o\sim\pi D o ∼ π .
Let L ≥ 1 L\geq1 L ≥ 1 and assume the claim for L − 1 L-1 L − 1 . For a process Θ = ( ω ′ , ( D u ′ ) u ∈ U ) \Theta=(\omega', (D'_u)_{u\in U}) Θ = ( ω ′ , ( D u ′ ) u ∈ U ) with the law of Θ o \Theta_o Θ o , let d ( Θ ) = D o ′ d(\Theta)=D'_o d ( Θ ) = D o ′ be its root mark and
S L − 1 ( Θ ) = ∑ u ∈ V ( Θ ) : ∣ u ∣ = L − 1 W L − 1 ( u ) ( Θ ) . S_{L-1}(\Theta)=\sum_{\substack{u\in V(\Theta):\\ |u|=L-1}}W_{L-1}(u)(\Theta). S L − 1 ( Θ ) = u ∈ V ( Θ ) : ∣ u ∣ = L − 1 ∑ W L − 1 ( u ) ( Θ ) .
where V ( Θ ) V(\Theta) V ( Θ ) and W L − 1 ( u ) ( Θ ) W_{L-1}(u)(\Theta) W L − 1 ( u ) ( Θ ) are the vertex set and the weight (8.9) computed from Θ \Theta Θ . The vertices of generation L L L are the words i u iu i u with 1 ≤ i ≤ N o 1\leq i\leq N_o 1 ≤ i ≤ N o and u ∈ V ( Θ i ) u\in V(\Theta_i) u ∈ V ( Θ i ) , ∣ u ∣ = L − 1 |u|=L-1 ∣ u ∣ = L − 1 , and W L ( i u ) = ( Y o , i / p ) s / 2 ∣ D i ∣ − s W L − 1 ( u ) ( Θ i ) W_L(iu)=(Y_{o,i}/p)^{s/2}|D_i|^{-s}W_{L-1}(u)(\Theta_i) W L ( i u ) = ( Y o , i / p ) s /2 ∣ D i ∣ − s W L − 1 ( u ) ( Θ i ) . We obtain
∑ v ∈ V : ∣ v ∣ = L k ( X o ) W L ( v ) = ∑ i = 1 N o k ( X o ) ( Y o , i p ) s / 2 ∣ D i ∣ − s S L − 1 ( Θ i ) . (138) \sum_{v\in V:|v|=L}k(X_o)W_L(v)
=\sum_{i=1}^{N_o}k(X_o)\left(\frac{Y_{o,i}}{p}\right)^{s/2}|D_i|^{-s}S_{L-1}(\Theta_i). \tag*{(138)} v ∈ V : ∣ v ∣ = L ∑ k ( X o ) W L ( v ) = i = 1 ∑ N o k ( X o ) ( p Y o , i ) s /2 ∣ D i ∣ − s S L − 1 ( Θ i ) . (138)
where D i = d ( Θ i ) D_i=d(\Theta_i) D i = d ( Θ i ) and, by (R1),
X o = 1 g D o ′ , D o = E − 1 p ∑ j = 1 N o Y o , j d ( Θ j ) . X_o=\frac{1}{gD'_o},\qquad
D_o=E-\frac{1}{p}\sum_{j=1}^{N_o}\frac{Y_{o,j}}{d(\Theta_j)}. X o = g D o ′ 1 , D o = E − p 1 j = 1 ∑ N o d ( Θ j ) Y o , j .
By (R3), applied to the words 1 , … , n 1,\ldots,n 1 , … , n for every n n n , and since the coordinates of ω \omega ω are independent under P \mathbb{P} P , the variables N o N_o N o , ( w o , i ) i ≥ 1 (w_{o,i})_{i\geq1} ( w o , i ) i ≥ 1 and ( Θ i ) i ≥ 1 (\Theta_i)_{i\geq1} ( Θ i ) i ≥ 1 are independent, the w o , i w_{o,i} w o , i are standard complex Gaussians, and each Θ i \Theta_i Θ i has the law of Θ o \Theta_o Θ o . The pairs m i = ( w o , i , Θ i ) m_i=(w_{o,i},\Theta_i) m i = ( w o , i , Θ i ) are independent with law μ = Law ( w o , 1 ) ⊗ Law ( Θ o ) \mu=\operatorname{Law}(w_{o,1})\otimes\operatorname{Law}(\Theta_o) μ = Law ( w o , 1 ) ⊗ Law ( Θ o ) , and they are independent of N 0 ∼ Pois ( p ) N_0 \sim\operatorname{Pois}(p) N 0 ∼ Pois ( p ) . Then Π = ∑ i = 1 N 0 δ m i \Pi= \sum_{i=1}^{N_0} \delta_{m_i} Π = ∑ i = 1 N 0 δ m i is a Poisson process with intensity measure p μ p\mu p μ [94, Proposition 3.5]. The right side of (138) equals ∫ Ψ ( m , Π − δ m ) Π ( d m ) \int\Psi(m,\Pi-\delta_m)\Pi(\mathrm{d}m) ∫ Ψ ( m , Π − δ m ) Π ( d m ) , where for m = ( w , Θ ) m=(w,\Theta) m = ( w , Θ ) and a finite counting measure ϱ \varrho ϱ ,
Ψ ( m , ϱ ) = k ( ( g E − g ∣ w ∣ 2 p d ( Θ ) − g p ∫ ∣ w ′ ∣ 2 d ( Θ ′ ) ϱ ( d w ′ d Θ ′ ) ) − 1 ) ( ∣ w ∣ 2 p ) s / 2 ∣ d ( Θ ) ∣ − s S L − 1 ( Θ ) . \Psi(m,\varrho)=k\left(\left(gE-\frac{g|w|^2}{p\,d(\Theta)}-\frac{g}{p}\int\frac{|w'|^2}{d(\Theta')}\varrho(\mathrm{d}w'\,\mathrm{d}\Theta')\right)^{-1}\right)\left(\frac{|w|^2}{p}\right)^{s/2}|d(\Theta)|^{-s}S_{L-1}(\Theta). Ψ ( m , ϱ ) = k ( ( g E − p d ( Θ ) g ∣ w ∣ 2 − p g ∫ d ( Θ ′ ) ∣ w ′ ∣ 2 ϱ ( d w ′ d Θ ′ ) ) − 1 ) ( p ∣ w ∣ 2 ) s /2 ∣ d ( Θ ) ∣ − s S L − 1 ( Θ ) .
with Ψ = 0 \Psi=0 Ψ = 0 if a denominator vanishes, which happens with probability zero. By the Mecke formula (4.1) (Lemma 4.4), we have
E ∑ v ∈ V : ∣ v ∣ = L k ( X o ) W L ( v ) = p ∫ E Ψ ( m , Π ) μ ( d m ) . (139) \mathbb{E}\sum_{v\in V:|v|=L}k(X_o)W_L(v)=p\int\mathbb{E}\Psi(m,\Pi)\,\mu(\mathrm{d}m). \tag*{(139)} E v ∈ V : ∣ v ∣ = L ∑ k ( X o ) W L ( v ) = p ∫ E Ψ ( m , Π ) μ ( d m ) . (139)
We now evaluate the right side. Fix m = ( w , Θ ) m=(w,\Theta) m = ( w , Θ ) with d = d ( Θ ) ≠ 0 d=d(\Theta)\ne0 d = d ( Θ ) = 0 . The points of Π \Pi Π carry independent pairs ( ∣ w ′ ∣ 2 , d ( Θ ′ ) ) (|w'|^2,d(\Theta')) ( ∣ w ′ ∣ 2 , d ( Θ ′ )) with law Exp ( 1 ) ⊗ π \operatorname{Exp}(1)\otimes\pi Exp ( 1 ) ⊗ π , and their number is Pois ( p ) \operatorname{Pois}(p) Pois ( p ) , so by Definition 5.1 the variable D ′ = E − p − 1 ∫ ∣ w ′ ∣ 2 d ( Θ ′ ) − 1 Π ( d w ′ d Θ ′ ) D'=E-p^{-1}\int|w'|^2d(\Theta')^{-1}\Pi(\mathrm{d}w'\,\mathrm{d}\Theta') D ′ = E − p − 1 ∫ ∣ w ′ ∣ 2 d ( Θ ′ ) − 1 Π ( d w ′ d Θ ′ ) has law π \pi π . Set x = 1 / ( g d ) x=1/(gd) x = 1/ ( g d ) and β = g 2 ∣ w ∣ 2 / p \beta=g^2|w|^2/p β = g 2 ∣ w ∣ 2 / p . Then g ∣ w ∣ 2 / ( p d ) = β x g|w|^2/(pd)=\beta x g ∣ w ∣ 2 / ( p d ) = β x and ( ∣ w ∣ 2 / p ) s / 2 ∣ d ∣ − s = β s / 2 ∣ x ∣ s (|w|^2/p)^{s/2}|d|^{-s}=\beta^{s/2}|x|^s ( ∣ w ∣ 2 / p ) s /2 ∣ d ∣ − s = β s /2 ∣ x ∣ s , and therefore
E Ψ ( m , Π ) = β s / 2 ∣ x ∣ s E [ k ( ( g D ′ − β x ) − 1 ) ] S L − 1 ( Θ ) . \mathbb{E}\Psi(m,\Pi)=\beta^{s/2}|x|^s\mathbb{E}\left[k\left((gD'-\beta x)^{-1}\right)\right]S_{L-1}(\Theta). E Ψ ( m , Π ) = β s /2 ∣ x ∣ s E [ k ( ( g D ′ − β x ) − 1 ) ] S L − 1 ( Θ ) .
Under μ \mu μ , the variable ∣ w ∣ 2 |w|^2 ∣ w ∣ 2 has law Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) and is independent of Θ \Theta Θ , and β \beta β has the exponential law of rate p / g 2 p/g^2 p / g 2 and is independent of g D ′ gD' g D ′ , whose law is that of g D gD g D . Integrating over w w w and recalling the definition of A s \mathcal{A}_s A s and of f B , β ( x ) = ( B − β x ) − 1 f_{B,\beta}(x)=(B-\beta x)^{-1} f B , β ( x ) = ( B − β x ) − 1 , we obtain
p ∫ E Ψ ( m , Π ) μ ( d m ) = E [ ( A s k ) ( 1 g d ( Θ o ) ) S L − 1 ( Θ o ) ] = E ∑ u ∈ V : ∣ u ∣ = L − 1 ( A s k ) ( X o ) W L − 1 ( u ) . \begin{aligned}
p\int\mathbb{E}\Psi(m,\Pi)\,\mu(\mathrm{d}m)=\mathbb{E}\left[(\mathcal{A}_s k)\left(\frac{1}{g\,d(\Theta_o)}\right)S_{L-1}(\Theta_o)\right] \\
&=\mathbb{E}\sum_{u\in V:|u|=L-1}(\mathcal{A}_s k)(X_o)W_{L-1}(u).
\end{aligned} p ∫ E Ψ ( m , Π ) μ ( d m ) = E [ ( A s k ) ( g d ( Θ o ) 1 ) S L − 1 ( Θ o ) ] = E u ∈ V : ∣ u ∣ = L − 1 ∑ ( A s k ) ( X o ) W L − 1 ( u ) .
By the induction hypothesis applied to the Borel function A s k : R → [ 0 , ∞ ] \mathcal{A}_s k:\mathbb{R}\to[0,\infty] A s k : R → [ 0 , ∞ ] , the right side equals ∫ ( A s L − 1 A s k ) d π x \int(\mathcal{A}_s^{L-1}\mathcal{A}_s k)\,\mathrm{d}\pi^x ∫ ( A s L − 1 A s k ) d π x . Together with (8.13), this proves (8.11) for L L L .
Proof of Proposition 8.3(i). By (8.10), we have ∣ F ( v ) ∣ s = ∣ D o ∣ − s W L ( v ) = g s ∣ X o ∣ s W L ( v ) |F(v)|^s=|D_o|^{-s}W_L(v)=g^s|X_o|^sW_L(v) ∣ F ( v ) ∣ s = ∣ D o ∣ − s W L ( v ) = g s ∣ X o ∣ s W L ( v ) . Applying Lemma 8.5 with k = g s e s k=g^s e_s k = g s e s , we obtain E ∑ ∣ v ∣ = L ∣ F ( v ) ∣ s = g s ∫ A s L e s d π x = Φ L ( s ; E ) \mathbb{E}\sum_{|v|=L}|F(v)|^s=g^s\int\mathcal{A}_s^L e_s\,\mathrm{d}\pi^x=\Phi_L(s;E) E ∑ ∣ v ∣ = L ∣ F ( v ) ∣ s = g s ∫ A s L e s d π x = Φ L ( s ; E ) , by (8.6). (The right side is also the spine expectation (6.22) of Lemma 6.23, which is the form in which the many-to-one formula would produce it.)
Summability of the forward columns
Lemma 8.6. Let E > 0 E>0 E > 0 and let π \pi π be a real solution at energy E E E with φ π ( 1 ) < 0 \varphi_\pi(1)<0 φ π ( 1 ) < 0 . There exists s 0 ∈ ( 0 , 1 ) s_0\in(0,1) s 0 ∈ ( 0 , 1 ) such that φ π ( s ) < 0 \varphi_\pi(s)<0 φ π ( s ) < 0 for every s ∈ [ s 0 , 1 ) s\in[s_0,1) s ∈ [ s 0 , 1 ) . For every such s s s , almost surely, for every word w w w ,
∑ u ∈ V ( T w ) ∣ F w ( u ) ∣ s < ∞ . \sum_{u\in V(T_w)}|F_w(u)|^s<\infty. u ∈ V ( T w ) ∑ ∣ F w ( u ) ∣ s < ∞.
Proof. By Proposition 6.15 (i), the free energy φ π \varphi_{\pi} φ π is real-analytic, and in particular continuous, on ( 0 , 2 ) (0,2) ( 0 , 2 ) . Since φ π ( 1 ) < 0 \varphi_{\pi}(1)<0 φ π ( 1 ) < 0 , there exists s 0 ∈ ( 0 , 1 ) s_0\in(0,1) s 0 ∈ ( 0 , 1 ) with φ π < 0 \varphi_{\pi}<0 φ π < 0 on [ s 0 , 1 ) [s_0,1) [ s 0 , 1 ) . (By Proposition 6.15 (v), φ π \varphi_{\pi} φ π is strictly decreasing on ( 0 , 1 ] (0,1] ( 0 , 1 ] , so the set of s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) with φ π ( s ) < 0 \varphi_{\pi}(s)<0 φ π ( s ) < 0 is an interval with right endpoint 1 1 1 .)
Fix s ∈ [ s 0 , 1 ) s\in[s_0,1) s ∈ [ s 0 , 1 ) . By Proposition 6.27 (iii), there exists C < ∞ C<\infty C < ∞ such that Φ L ( s ; E ) ≤ C e L φ π ( s ) \Phi_L(s;E)\le Ce^{L\varphi_{\pi}(s)} Φ L ( s ; E ) ≤ C e L φ π ( s ) for every L ≥ 1 L\ge1 L ≥ 1 , and Φ 0 ( s ; E ) = E ∣ D ∣ − s < ∞ \Phi_0(s;E)=\mathbb{E}|D|^{-s}<\infty Φ 0 ( s ; E ) = E ∣ D ∣ − s < ∞ by Proposition 5.5 . By Proposition 8.3 (i) and Tonelli’s theorem,
E ∑ v ∈ V ∣ F ( v ) ∣ s = ∑ L ≥ 0 Φ L ( s ; E ) ≤ E ∣ D ∣ − s + C ∑ L ≥ 1 e L φ π ( s ) < ∞ . \mathbb{E}\sum_{v\in V}|F(v)|^s=\sum_{L\ge0}\Phi_L(s;E)\le\mathbb{E}|D|^{-s}+C\sum_{L\ge1}e^{L\varphi_{\pi}(s)}<\infty. E v ∈ V ∑ ∣ F ( v ) ∣ s = L ≥ 0 ∑ Φ L ( s ; E ) ≤ E ∣ D ∣ − s + C L ≥ 1 ∑ e L φ π ( s ) < ∞.
since φ π ( s ) < 0 \varphi_{\pi}(s)<0 φ π ( s ) < 0 . We conclude that ∑ v ∣ F ( v ) ∣ s < ∞ \sum_v|F(v)|^s<\infty ∑ v ∣ F ( v ) ∣ s < ∞ almost surely. For a word w w w , the column F w F_w F w is the same function of Θ w \Theta_w Θ w as F F F is of Θ o \Theta_o Θ o , and Θ w \Theta_w Θ w has the law of Θ o \Theta_o Θ o by (R3). Then ∑ u ∣ F w ( u ) ∣ s < ∞ \sum_u|F_w(u)|^s<\infty ∑ u ∣ F w ( u ) ∣ s < ∞ almost surely, and a countable intersection over words w w w completes the proof.
Rerooted columns
We now construct inverse columns at arbitrary vertices of the full tree and of the components left after removing one vertex or one edge. A removed set is one of the following: the empty set; a word y y y , standing for the vertex y y y ; or a word y ≠ o y\ne o y = o , standing for the edge { y − , y } \{y^-,y\} { y − , y } . Fix a removed set and a word r r r . On the event that r ∈ V r\in V r ∈ V , that y ∈ V y\in V y ∈ V (if the removed set is not empty), and that r ≠ y r\ne y r = y (for a removed vertex), let C C C be the component containing r r r of the reduced forest obtained from T \mathcal{T} T by the removal (with C = V C=V C = V for the empty removed set). For a word x x x , let anc ( x ) \operatorname{anc}(x) anc ( x ) be the set of its prefixes, including o o o and x x x . We define
S = anc ( r ) ∪ anc ( y ) S=\operatorname{anc}(r)\cup\operatorname{anc}(y) S = anc ( r ) ∪ anc ( y )
(with S = anc ( r ) S=\operatorname{anc}(r) S = anc ( r ) for the empty removed set), S C = S ∩ C S_C=S\cap C S C = S ∩ C , and
W C = { w ∈ C ∖ S : w − ∈ S } . W_C=\{w\in C\setminus S:w^-\in S\}. W C = { w ∈ C ∖ S : w − ∈ S } .
The set S S S is a finite set of words which contains o o o and all prefixes of its elements, and it is determined by the removed set and r r r . For x ∈ C ∖ { r } x\in C\setminus\{r\} x ∈ C ∖ { r } , let b ( x ) b(x) b ( x ) be the neighbor of x x x on the path in C C C from x x x to r r r , and let ch C ( x ) \operatorname{ch}_C(x) ch C ( x ) be the set of neighbors x ′ ≠ b ( x ) x'\ne b(x) x ′ = b ( x ) of x x x in C C C ; let ch C ( r ) \operatorname{ch}_C(r) ch C ( r ) be the set of all neighbors of r r r in C C C .
Lemma 8.7. On the event above, the following hold.
(i) The set S C S_C S C contains r r r and induces a finite connected subtree of C C C .
(ii) The set W C W_C W C is finite. For every w ∈ W C w\in W_C w ∈ W C , the set V ( T w ) V(\mathcal{T}_w) V ( T w ) is contained in C ∖ S C\setminus S C ∖ S , and C ∖ S C C\setminus S_C C ∖ S C is the disjoint union of the sets V ( T w ) V(\mathcal{T}_w) V ( T w ) , w ∈ W C w\in W_C w ∈ W C .
(iii) Let w ∈ W C w\in W_C w ∈ W C and x ∈ V ( T w ) x\in V(\mathcal{T}_w) x ∈ V ( T w ) . Then b ( x ) = x − b(x)=x^- b ( x ) = x − , and ch C ( x ) \operatorname{ch}_C(x) ch C ( x ) is the set of children of x x x in T \mathcal{T} T .
(iv) If x ∈ S C x \in S_C x ∈ S C and x ′ ∈ ch C ( x ) ∖ S C x' \in\operatorname{ch}_C(x) \setminus S_C x ′ ∈ ch C ( x ) ∖ S C , then x ′ x' x ′ is a child of x x x and x ′ ∈ W C x' \in W_C x ′ ∈ W C .
Proof. For (i), we have r ∈ S ∩ C r \in S \cap C r ∈ S ∩ C . The set S S S is a union of ancestral paths starting at o o o , and it induces a subtree of T \mathcal{T} T . If x , x ′ ∈ S C x,x' \in S_C x , x ′ ∈ S C , then the path in T \mathcal{T} T between them lies in S S S , and it lies in C C C because C C C induces a subtree of T \mathcal{T} T (Lemma 3.5(i)). This shows that S C S_C S C is connected. It is finite because S S S is.
We next prove (ii). Let w ∈ W C w \in W_C w ∈ W C . If some u ∈ V ( T w ) u \in V(\mathcal{T}_w) u ∈ V ( T w ) belonged to S S S , then its prefix w w w would belong to S S S , which is false. We find that V ( T w ) ∩ S = ∅ V(\mathcal{T}_w) \cap S = \varnothing V ( T w ) ∩ S = ∅ . In particular, V ( T w ) V(\mathcal{T}_w) V ( T w ) contains neither the removed vertex y y y nor an endpoint of the removed edge, since both y y y and y − y^- y − lie in S S S . Every edge of T w \mathcal{T}_w T w is therefore retained, and V ( T w ) V(\mathcal{T}_w) V ( T w ) is connected in the reduced forest. Since it contains w ∈ C w \in C w ∈ C , we obtain V ( T w ) ⊂ C ∖ S V(\mathcal{T}_w) \subset C \setminus S V ( T w ) ⊂ C ∖ S . Conversely, let u ∈ C ∖ S C u \in C \setminus S_C u ∈ C ∖ S C , so that u ∉ S u \notin S u ∈ / S . Since o ∈ S o \in S o ∈ S , there exists a shortest prefix w w w of u u u with w ∉ S w \notin S w ∈ / S . Then w ≠ o w \ne o w = o , w − ∈ S w^- \in S w − ∈ S , and u ∈ V ( T w ) u \in V(\mathcal{T}_w) u ∈ V ( T w ) . By the previous argument, V ( T w ) V(\mathcal{T}_w) V ( T w ) lies in one component, which is C C C because it contains u u u . In particular w ∈ C w \in C w ∈ C . It remains to check that w − ∈ C w^- \in C w − ∈ C , that is, that the edge { w − , w } \{w^-,w\} { w − , w } is retained and w − ≠ y w^- \ne y w − = y . If the edge { w − , w } \{w^-,w\} { w − , w } were the removed edge, then w = y ∈ S w = y \in S w = y ∈ S . If w − = y w^- = y w − = y were the removed vertex, then the component of w w w would be V ( T w ) V(\mathcal{T}_w) V ( T w ) , which contains r r r ; then w w w would be a prefix of r r r and would lie in S S S . Both are impossible, and we conclude that w − ∈ S C w^- \in S_C w − ∈ S C and w ∈ W C w \in W_C w ∈ W C . If w , w ′ ∈ W C w,w' \in W_C w , w ′ ∈ W C are distinct and w w w is a prefix of w ′ w' w ′ , then w ′ − ∈ V ( T w ) ∩ S w'^- \in V(\mathcal{T}_w) \cap S w ′ − ∈ V ( T w ) ∩ S , which is impossible. We conclude that the sets V ( T w ) V(\mathcal{T}_w) V ( T w ) , w ∈ W C w \in W_C w ∈ W C , are disjoint. Finally, W C W_C W C is finite, since its elements are children of the finitely many vertices in S S S .
For (iii), we have r ∈ S r \in S r ∈ S and r ∉ V ( T w ) r \notin V(\mathcal{T}_w) r ∈ / V ( T w ) . The path from x x x to r r r therefore leaves V ( T w ) V(\mathcal{T}_w) V ( T w ) , and since V ( T w ) V(\mathcal{T}_w) V ( T w ) is joined to the rest of T \mathcal{T} T only by the edge { w − , w } \{w^-,w\} { w − , w } , the path first moves from x x x to x − x^- x − . All children of x x x belong to V ( T w ) ⊂ C V(\mathcal{T}_w) \subset C V ( T w ) ⊂ C , and the edges to them are retained. The neighbors of x x x in C C C other than x − x^- x − are its children.
For (iv), the parent of x x x belongs to S S S (if x ≠ o x \ne o x = o ), since S S S is closed under prefixes. It follows that x ′ x' x ′ is a child of x x x , and x ′ ∈ C ∖ S x' \in C \setminus S x ′ ∈ C ∖ S with x ′ − = x ∈ S x'^- = x \in S x ′ − = x ∈ S , that is, x ′ ∈ W C x' \in W_C x ′ ∈ W C .
We now define real cavity marks on C C C , adapted to the root r r r . For x ∈ C ∖ S C x \in C \setminus S_C x ∈ C ∖ S C , we set γ x C = γ x \gamma_x^C = \gamma_x γ x C = γ x . For x ∈ S C x \in S_C x ∈ S C , we define, recursively from the leaves of the finite tree S C S_C S C rooted at r r r toward r r r ,
Δ x C = E + ∑ x ′ ∈ ch C ( x ) ∣ h x x ′ ∣ 2 γ x ′ C , γ x C = − 1 Δ x C . (140) \Delta_x^C = E + \sum_{x' \in\operatorname{ch}_C(x)} |h_{xx'}|^2 \gamma_{x'}^C,\qquad\gamma_x^C = -\frac{1}{\Delta_x^C}. \tag*{(140)} Δ x C = E + x ′ ∈ ch C ( x ) ∑ ∣ h x x ′ ∣ 2 γ x ′ C , γ x C = − Δ x C 1 . (140)
This is well defined as long as the denominators Δ x C \Delta_x^C Δ x C do not vanish: the elements of ch C ( x ) \operatorname{ch}_C(x) ch C ( x ) are either in S C S_C S C , where they are closer to the leaves, or in W C W_C W C by Lemma 8.7(iv). For the empty removed set and r = o r = o r = o , we have S C = { o } S_C = \{o\} S C = { o } , W C W_C W C is the set of children of o o o , and γ o C = γ o \gamma_o^C = \gamma_o γ o C = γ o . In general, the numbers γ x C \gamma_x^C γ x C with x ∈ S C x \in S_C x ∈ S C are real counterparts of the cavity resolvents Γ x → b ( x ) \Gamma_{x \to b(x)} Γ x → b ( x ) of the rerooted component (compare Remark 3.17).
Lemma 8.8. Fix a removed set and a word r r r . Almost surely, on the event above, every denominator Δ x C \Delta_x^C Δ x C , x ∈ S C x \in S_C x ∈ S C , is finite and nonzero; hence every γ x C \gamma_x^C γ x C , x ∈ C x \in C x ∈ C , is finite and nonzero.
Proof. The proof uses that the squared modulus of the mark of each edge of the skeleton S C S_C S C is a fresh exponential variable for the denominator at its endpoint farther from the leaves. Let W ( S ) = { s i : s ∈ S , i ∈ N } ∖ S W(S)=\{si:s\in S,\ i\in\mathbb{N}\}\setminus S W ( S ) = { s i : s ∈ S , i ∈ N } ∖ S . Its elements are not prefixes of each other: if w ∈ W ( S ) w\in W(S) w ∈ W ( S ) were a proper prefix of w ′ = s i ∈ W ( S ) w'=si\in W(S) w ′ = s i ∈ W ( S ) , then w w w would be a prefix of s ∈ S s\in S s ∈ S and would belong to S S S . By (R3), applied to finite subsets of W ( S ) W(S) W ( S ) , and since the coordinates of ω \omega ω are independent under P \mathbb{P} P , we see that the following random elements are independent: the variables N s N_s N s and w s , s i w_{s,si} w s , s i for s ∈ S s\in S s ∈ S and i ∈ N i\in\mathbb{N} i ∈ N , and the processes Θ w \Theta_w Θ w for w ∈ W ( S ) w\in W(S) w ∈ W ( S ) . The event in question, the sets S C S_C S C and W C ⊂ W ( S ) W_C\subset W(S) W C ⊂ W ( S ) , and the maps b b b and ch C \operatorname{ch}_C ch C are functions of ( N s ) s ∈ S (N_s)_{s\in S} ( N s ) s ∈ S . The weights of the edges with an endpoint in S C S_C S C are functions of the variables w s , s i w_{s,si} w s , s i , and γ w \gamma_w γ w is a function of Θ w \Theta_w Θ w .
For x ∈ S C x\in S_C x ∈ S C and x ′ ∈ ch C ( x ) x'\in\operatorname{ch}_C(x) x ′ ∈ ch C ( x ) , let C x ′ → x C_{x'\to x} C x ′ → x be the branch of C C C at x ′ x' x ′ away from x x x . By induction from the leaves, γ x C \gamma^C_x γ x C is a function of ( N s ) s ∈ S (N_s)_{s\in S} ( N s ) s ∈ S , of the marks of the edges with both endpoints in C x ′ → x C_{x'\to x} C x ′ → x , and of the processes Θ w \Theta_w Θ w with w ∈ W C ∩ C x ′ → x w\in W_C\cap C_{x'\to x} w ∈ W C ∩ C x ′ → x .
We now prove by induction from the leaves of S C S_C S C toward r r r that, almost surely, Δ x C \Delta^C_x Δ x C is finite and nonzero for every x ∈ S C x\in S_C x ∈ S C . Fix x ∈ S C x\in S_C x ∈ S C and suppose the claim holds for the elements of ch C ( x ) ∩ S C \operatorname{ch}_C(x)\cap S_C ch C ( x ) ∩ S C . Then γ x ′ C \gamma^C_{x'} γ x ′ C is finite and nonzero for every x ′ ∈ ch C ( x ) x'\in\operatorname{ch}_C(x) x ′ ∈ ch C ( x ) : for x ′ ∈ S C x'\in S_C x ′ ∈ S C by the induction hypothesis, and for x ′ ∈ W C x'\in W_C x ′ ∈ W C because γ x ′ C = − 1 / D x ′ \gamma^C_{x'}=-1/D_{x'} γ x ′ C = − 1/ D x ′ with D x ′ ∈ R ∖ { 0 } D_{x'}\in\mathbb{R}\setminus\{0\} D x ′ ∈ R ∖ { 0 } . In particular Δ x C \Delta^C_x Δ x C is finite. If ch C ( x ) \operatorname{ch}_C(x) ch C ( x ) is empty, then Δ x C = E ≠ 0 \Delta^C_x=E\ne0 Δ x C = E = 0 . Otherwise, let x ∗ x_* x ∗ be the element of ch C ( x ) \operatorname{ch}_C(x) ch C ( x ) that comes first in a fixed enumeration of U U U , which is a function of ( N s ) s ∈ S (N_s)_{s\in S} ( N s ) s ∈ S . Since there are countably many words, it suffices to show that Δ x C ≠ 0 \Delta^C_x\ne0 Δ x C = 0 almost surely on the event { x ∗ = x ′ } \{x_*=x'\} { x ∗ = x ′ } , for a fixed word x ′ x' x ′ . On this event, the edge e = { x , x ′ } e=\{x,x'\} e = { x , x ′ } joins two vertices of S ∪ W ( S ) S\cup W(S) S ∪ W ( S ) , one of which is the parent of the other and lies in S S S . Its mark is therefore one of the variables w s , s i w_{s,si} w s , s i , with s s s and i i i determined by x x x and x ′ x' x ′ ; we denote it by ω e \omega_e ω e . We write
Δ x C = A + ∣ ω e ∣ 2 p γ x ′ C , A = E + ∑ x ′ ′ ∈ ch C ( x ) ∖ { x ′ } ∣ h x x ′ ′ ∣ 2 γ x ′ ′ C . \Delta^C_x=A+\frac{|\omega_e|^2}{p}\gamma^C_{x'},\qquad A=E+\sum_{x''\in\operatorname{ch}_C(x)\setminus\{x'\}}|h_{xx''}|^2\gamma^C_{x''}. Δ x C = A + p ∣ ω e ∣ 2 γ x ′ C , A = E + x ′′ ∈ ch C ( x ) ∖ { x ′ } ∑ ∣ h x x ′′ ∣ 2 γ x ′′ C .
Let R e \mathcal{R}_e R e be the σ \sigma σ -algebra generated by all the random elements listed in the first paragraph except ω e \omega_e ω e . Since the edge e e e has an endpoint at x x x , it is not an edge of any branch C x ′ ′ → x C_{x''\to x} C x ′′ → x . By the previous paragraph, the event { x ∗ = x ′ } \{x_*=x'\} { x ∗ = x ′ } and the variables A A A and γ x ′ C \gamma^C_{x'} γ x ′ C are R e \mathcal{R}_e R e -measurable. The variable ω e \omega_e ω e is independent of R e \mathcal{R}_e R e , and ∣ ω e ∣ 2 |\omega_e|^2 ∣ ω e ∣ 2 has, conditionally on R e \mathcal{R}_e R e , the Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) law, which has no atoms. Since γ x ′ C ≠ 0 \gamma^C_{x'}\ne0 γ x ′ C = 0 almost surely on { x ∗ = x ′ } \{x_*=x'\} { x ∗ = x ′ } , we have
P ( { Δ x C = 0 } ∩ { x ∗ = x ′ } ∣ R e ) = 1 { x ∗ = x ′ } P ( ∣ ω e ∣ 2 = − p A / γ x ′ C ∣ R e ) = 0 \mathbb{P}\left(\{\Delta^C_x=0\}\cap\{x_*=x'\}\mid\mathcal{R}_e\right)=\mathbf{1}_{\{x_*=x'\}}\mathbb{P}\left(|\omega_e|^2=-pA/\gamma^C_{x'}\mid\mathcal{R}_e\right)=0 P ( { Δ x C = 0 } ∩ { x ∗ = x ′ } ∣ R e ) = 1 { x ∗ = x ′ } P ( ∣ ω e ∣ 2 = − p A / γ x ′ C ∣ R e ) = 0
almost surely. This completes the induction, and the last assertion follows from (140) and γ x C = γ x \gamma^C_x=\gamma_x γ x C = γ x for x ∉ S C x\notin S_C x ∈ / S C .
Lemma 8.9. Let s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) be such that the conclusion of Lemma 8.6 holds. Fix a removed set and a word r r r , and work on the intersection of the almost sure events of Lemma 8.6, Lemma 8.8 and (R1) with the event above. Define F C , r F^{C,r} F C , r on C C C by F C , r ( r ) = γ r C F^{C,r}(r)=\gamma_r^C F C , r ( r ) = γ r C and F C , r ( x ) = − γ x C h x b ( x ) F C , r ( b ( x ) ) F^{C,r}(x)=-\gamma_x^C h_{x b(x)}F^{C,r}(b(x)) F C , r ( x ) = − γ x C h x b ( x ) F C , r ( b ( x )) for x ≠ r x\ne r x = r . Then
( H C , f o r m F C , r ) ( x ) − E F C , r ( x ) = δ r ( x ) , x ∈ C , (141) (\mathcal{H}_{C,\mathrm{form}}F^{C,r})(x)-EF^{C,r}(x)=\delta_r(x),\qquad x\in C, \tag*{(141)} ( H C , form F C , r ) ( x ) − E F C , r ( x ) = δ r ( x ) , x ∈ C , (141)
and ∑ x ∈ C ∣ F C , r ( x ) ∣ s < ∞ \sum_{x\in C}|F^{C,r}(x)|^s<\infty ∑ x ∈ C ∣ F C , r ( x ) ∣ s < ∞ .
Proof. We verify the hypotheses of Lemma 8.1 on the tree C C C with root r r r and the numbers γ x C \gamma_x^C γ x C . The neighbors of x x x in C C C are b ( x ) b(x) b ( x ) (if x ≠ r x\ne r x = r ) and the elements of ch C ( x ) \operatorname{ch}_C(x) ch C ( x ) . For x ∈ S C x\in S_C x ∈ S C , (8.1) is the definition (8.14), with nonzero denominators by Lemma 8.8. For x ∈ V ( T w ) x\in V(\mathcal{T}_w) x ∈ V ( T w ) with w ∈ W C w\in W_C w ∈ W C , Lemma 8.7(iii) shows that ch C ( x ) \operatorname{ch}_C(x) ch C ( x ) is the set of children of x x x , all of which lie in V ( T w ) V(\mathcal{T}_w) V ( T w ) , so that γ x ′ C = γ x ′ \gamma_{x'}^C=\gamma_{x'} γ x ′ C = γ x ′ for x ′ ∈ ch C ( x ) x'\in\operatorname{ch}_C(x) x ′ ∈ ch C ( x ) . The relation (8.1) then holds by the computation after Definition 8.2. By Lemma 8.7(ii), these two cases cover C C C . Then Lemma 8.1 implies (8.15).
For summability, let w ∈ W C w\in W_C w ∈ W C . For x ∈ V ( T w ) x\in V(\mathcal{T}_w) x ∈ V ( T w ) we have b ( x ) = x − b(x)=x^- b ( x ) = x − and γ x C = γ x \gamma_x^C=\gamma_x γ x C = γ x , so F C , r F^{C,r} F C , r and F w F_w F w satisfy the same recursion f ( x ) = − γ x h x x − f ( x − ) f(x)=-\gamma_xh_{xx^-}f(x^-) f ( x ) = − γ x h x x − f ( x − ) on V ( T w ) ∖ { w } V(\mathcal{T}_w)\setminus\{w\} V ( T w ) ∖ { w } . Comparing their values at w w w , we find
F C , r ( x ) = F C , r ( w ) γ w F w ( x ) , x ∈ V ( T w ) . F^{C,r}(x)=\frac{F^{C,r}(w)}{\gamma_w}F_w(x),\qquad x\in V(\mathcal{T}_w). F C , r ( x ) = γ w F C , r ( w ) F w ( x ) , x ∈ V ( T w ) .
By Lemma 8.7(ii),
∑ x ∈ C ∣ F C , r ( x ) ∣ s = ∑ x ∈ S C ∣ F C , r ( x ) ∣ s + ∑ w ∈ W C ∣ F C , r ( w ) γ w ∣ s ∑ x ∈ V ( T w ) ∣ F w ( x ) ∣ s . \sum_{x\in C}|F^{C,r}(x)|^s=\sum_{x\in S_C}|F^{C,r}(x)|^s+\sum_{w\in W_C}\left|\frac{F^{C,r}(w)}{\gamma_w}\right|^s\sum_{x\in V(\mathcal{T}_w)}|F_w(x)|^s. x ∈ C ∑ ∣ F C , r ( x ) ∣ s = x ∈ S C ∑ ∣ F C , r ( x ) ∣ s + w ∈ W C ∑ γ w F C , r ( w ) s x ∈ V ( T w ) ∑ ∣ F w ( x ) ∣ s .
The sets S C S_C S C and W C W_C W C are finite, all values of F C , r F^{C,r} F C , r are finite, γ w ≠ 0 \gamma_w\ne0 γ w = 0 , and each inner sum is finite by Lemma 8.6. This proves that the left side is finite.
Proofs of the main results
We use the following characterization of the range of T − E T-E T − E for a self-adjoint operator T T T .
Lemma 8.10. Let T T T be a self-adjoint operator on a Hilbert space, let ψ \psi ψ be a vector, and let E ∈ R E\in\mathbb{R} E ∈ R . If f ∈ Dom ( T ) f\in\operatorname{Dom}(T) f ∈ Dom ( T ) and ( T − E ) f = ψ (T-E)f=\psi ( T − E ) f = ψ , then μ ψ T ( { E } ) = 0 \mu_\psi^T(\{E\})=0 μ ψ T ({ E }) = 0 and R μ ψ T ( E ) ≤ ∥ f ∥ 2 R_{\mu_\psi^T}(E)\le\lVert f\rVert^2 R μ ψ T ( E ) ≤ ∥ f ∥ 2 . Conversely, if R μ ψ T ( E ) < ∞ R_{\mu_\psi^T}(E)<\infty R μ ψ T ( E ) < ∞ , then ψ ∈ Ran ( T − E ) \psi\in\operatorname{Ran}(T-E) ψ ∈ Ran ( T − E ) .
Proof. Suppose ( T − E ) f = ψ (T-E)f=\psi ( T − E ) f = ψ . By the spectral theorem, for every bounded Borel function k k k ,
⟨ ψ , k ( T ) ψ ⟩ = ∫ R k ( λ ) ( λ − E ) 2 μ f T ( d λ ) , \langle\psi,k(T)\psi\rangle=\int_{\mathbb{R}}k(\lambda)(\lambda-E)^2\,\mu_f^T(\mathrm{d}\lambda), ⟨ ψ , k ( T ) ψ ⟩ = ∫ R k ( λ ) ( λ − E ) 2 μ f T ( d λ ) ,
that is, μ ψ T ( d λ ) = ( λ − E ) 2 μ f T ( d λ ) \mu_\psi^T(\mathrm{d}\lambda)=(\lambda-E)^2\mu_f^T(\mathrm{d}\lambda) μ ψ T ( d λ ) = ( λ − E ) 2 μ f T ( d λ ) . In particular, μ ψ T ( { E } ) = 0 \mu_\psi^T(\{E\})=0 μ ψ T ({ E }) = 0 and R μ ψ T ( E ) = μ f T ( R ∖ { E } ) ≤ ∥ f ∥ 2 R_{\mu_\psi^T}(E)=\mu_f^T(\mathbb{R}\setminus\{E\})\le\lVert f\rVert^2 R μ ψ T ( E ) = μ f T ( R ∖ { E }) ≤ ∥ f ∥ 2 . Conversely, suppose R μ ψ T ( E ) < ∞ R_{\mu_\psi^T}(E)<\infty R μ ψ T ( E ) < ∞ . Then μ ψ T ( { E } ) = 0 \mu_\psi^T(\{E\})=0 μ ψ T ({ E }) = 0 . Let k ( λ ) = k(\lambda)= k ( λ ) = ( λ − E ) − 1 (\lambda-E)^{-1} ( λ − E ) − 1 for λ ≠ E \lambda\ne E λ = E and k ( E ) = 0 k(E)=0 k ( E ) = 0 . Then ∫ ∣ k ∣ 2 d μ ψ T = R μ ψ T ( E ) < ∞ \int|k|^2\,\mathrm{d}\mu_\psi^T=R_{\mu_\psi}^T(E)<\infty ∫ ∣ k ∣ 2 d μ ψ T = R μ ψ T ( E ) < ∞ , and since ∣ λ k ( λ ) ∣ ≤ 1 + ∣ E ∣ ∣ k ( λ ) ∣ |\lambda k(\lambda)|\le1+|E||k(\lambda)| ∣ λk ( λ ) ∣ ≤ 1 + ∣ E ∣∣ k ( λ ) ∣ , also ∫ ∣ λ k ( λ ) ∣ 2 d μ ψ T < ∞ \int|\lambda k(\lambda)|^2\,\mathrm{d}\mu_\psi^T<\infty ∫ ∣ λk ( λ ) ∣ 2 d μ ψ T < ∞ . By the functional calculus, f = k ( T ) ψ f=k(T)\psi f = k ( T ) ψ belongs to Dom ( T ) \operatorname{Dom}(T) Dom ( T ) , and ( T − E ) f = 1 R ∖ { E } ( T ) ψ = ψ (T-E)f=\mathbf{1}_{\mathbb{R}\setminus\{E\}}(T)\psi=\psi ( T − E ) f = 1 R ∖ { E } ( T ) ψ = ψ , since ∥ 1 { E } ( T ) ψ ∥ 2 = μ ψ T ( { E } ) = 0 \|\mathbf{1}_{\{E\}}(T)\psi\|^2=\mu_\psi^T(\{E\})=0 ∥ 1 { E } ( T ) ψ ∥ 2 = μ ψ T ({ E }) = 0 .
Proof of Proposition 8.3(ii) and (iii). Part (i) was proved in § 8.2.2 . We assume that φ π ( 1 ) < 0 \varphi_\pi(1)<0 φ π ( 1 ) < 0 , let s 0 s_0 s 0 be as in Lemma 8.6, and fix s ∈ [ s 0 , 1 ) s\in[s_0,1) s ∈ [ s 0 , 1 ) .
Step 1. The root column. For the empty removed set and r = o r=o r = o , we have S C = { o } S_C=\{o\} S C = { o } , γ o C = γ o \gamma_o^C=\gamma_o γ o C = γ o and γ x C = γ x \gamma_x^C=\gamma_x γ x C = γ x for x ≠ o x\ne o x = o , and F V , o = F F^{V,o}=F F V , o = F . By Lemma 8.9, almost surely F ∈ ℓ s ( V ) F\in\ell^s(V) F ∈ ℓ s ( V ) , and F ∈ ℓ 2 ( V ) F\in\ell^2(V) F ∈ ℓ 2 ( V ) because ∥ f ∥ 2 ≤ ∥ f ∥ s \|f\|_2\le\|f\|_s ∥ f ∥ 2 ≤ ∥ f ∥ s for s ≤ 2 s\le2 s ≤ 2 . By (8.15), H f o r m F = E F + δ o ∈ ℓ 2 ( V ) \mathcal{H}_{\mathrm{form}}F=EF+\delta_o\in\ell^2(V) H form F = E F + δ o ∈ ℓ 2 ( V ) . By the maximal domain characterization Theorem 3.10 (ii), in the form of Remark 3.11, F ∈ Dom ( H ) F\in\operatorname{Dom}(\mathcal{H}) F ∈ Dom ( H ) and ( H − E ) F = δ o (\mathcal{H}-E)F=\delta_o ( H − E ) F = δ o . This proves (ii).
Step 2. All columns. Fix a removed set and a word r r r . On the intersection of the almost sure events of Lemma 8.9 and Theorem 3.10 with the event on which C C C is defined (see § 8.2.4 ), Lemma 8.9 yields F C , r ∈ ℓ s ( C ) ⊂ ℓ 2 ( C ) F^{C,r}\in\ell^s(C)\subset\ell^2(C) F C , r ∈ ℓ s ( C ) ⊂ ℓ 2 ( C ) and H C , f o r m F C , r = E F C , r + δ r ∈ ℓ 2 ( C ) \mathcal{H}_{C,\mathrm{form}}F^{C,r}=EF^{C,r}+\delta_r\in\ell^2(C) H C , form F C , r = E F C , r + δ r ∈ ℓ 2 ( C ) . By Remark 3.11, F C , r ∈ Dom ( H C ) F^{C,r}\in\operatorname{Dom}(\mathcal{H}_C) F C , r ∈ Dom ( H C ) and ( H C − E ) F C , r = δ r (\mathcal{H}_C-E)F^{C,r}=\delta_r ( H C − E ) F C , r = δ r . In particular δ r ∈ Ran ( H C − E ) \delta_r\in\operatorname{Ran}(\mathcal{H}_C-E) δ r ∈ Ran ( H C − E ) , and Lemma 8.10 gives R μ r C ( E ) ≤ ∥ F C , r ∥ 2 < ∞ R_{\mu_r^C}(E)\le\|F^{C,r}\|^2<\infty R μ r C ( E ) ≤ ∥ F C , r ∥ 2 < ∞ . There are countably many removed sets and words r r r , so these conclusions hold almost surely for all of them simultaneously. For the empty removed set, C = V C=V C = V and μ r C = μ r \mu_r^C=\mu_r μ r C = μ r . For the removal of a vertex y y y , the operator H ( y ) \mathcal{H}^{(y)} H ( y ) is the orthogonal sum of the operators of the components (Theorem 3.10(iv)}), and μ x ( y ) = μ x C \mu_x^{(y)}=\mu_x^C μ x ( y ) = μ x C for x ∈ C x\in C x ∈ C ; this gives (8.8).
Step 3. Transfer to P \mathbb{P} P and P ∞ \mathbb{P}^{\infty} P ∞ . Let A E A_E A E be the event that R μ r C C ( E ) < ∞ R_{\mu_r^C}^C(E)<\infty R μ r C C ( E ) < ∞ for every removed set, every word r r r and the corresponding component C C C (whenever these are defined). By (7.3), R μ r C ( E ) = sup n n Im G r r C ( E + i / n ) R_{\mu_r^C}(E)=\sup_n n\operatorname{Im}G_{rr}^C(E+\mathrm{i}/n) R μ r C ( E ) = sup n n Im G r r C ( E + i / n ) , and by Lemma 7.5 this is a measurable function of the marked tree. It follows that A E A_E A E is an event of the marked tree. By Step 2, it has probability one on the space of Lemma 5.13 , where the marked tree has the law P \mathbb{P} P , and we have P ( A E ) = 1 \mathbb{P}(A_E)=1 P ( A E ) = 1 . Since P ∞ ≤ ϱ p − 1 P \mathbb{P}^{\infty}\le\varrho_p^{-1}\mathbb{P} P ∞ ≤ ϱ p − 1 P , also P ∞ ( A E ) = 1 \mathbb{P}^{\infty}(A_E)=1 P ∞ ( A E ) = 1 . By Lemma 8.10 , on A E A_E A E every δ r \delta_r δ r belongs to Ran ( H C − E ) \operatorname{Ran}(\mathcal{H}_C-E) Ran ( H C − E ) . This proves (iii).
Proof of Theorem 8.4. Let N 0 ⊂ U N_0\subset U N 0 ⊂ U be a Lebesgue-null set such that a real solution π \pi π with φ π ( 1 ) < 0 \varphi_\pi(1)<0 φ π ( 1 ) < 0 exists at every E ∈ U ∖ N 0 E\in U\setminus N_0 E ∈ U ∖ N 0 . Let B \mathcal{B} B be the set of pairs ( ω , E ) ∈ Ω × U (\omega,E)\in\Omega\times U ( ω , E ) ∈ Ω × U such that, for every vertex y y y and every neighbor x x x of y y y , R μ x ( y ) ( E ) < ∞ R_{\mu_x^{(y)}}(E)<\infty R μ x ( y ) ( E ) < ∞ , where Ω \Omega Ω is the probability space of the marked tree. For fixed words y , x y,x y , x , we have R μ x ( y ) ( E ) = sup n n Im G x x ( y ) ( E + i / n ) R_{\mu_x^{(y)}}(E)=\sup_n n\operatorname{Im}G_{xx}^{(y)}(E+\mathrm{i}/n) R μ x ( y ) ( E ) = sup n n Im G xx ( y ) ( E + i / n ) by (7.3), which is jointly measurable in ( ω , E ) (\omega,E) ( ω , E ) by Lemma 7.5 . Since there are countably many pairs of words, B \mathcal{B} B is jointly measurable.
For E ∈ U ∖ N 0 E\in U\setminus N_0 E ∈ U ∖ N 0 , Proposition 8.3 (iii), applied with a real solution π \pi π at E E E with φ π ( 1 ) < 0 \varphi_\pi(1)<0 φ π ( 1 ) < 0 , gives P ( ( ω , E ) ∈ B ) = 1 \mathbb{P}((\omega,E)\in\mathcal{B})=1 P (( ω , E ) ∈ B ) = 1 . The event B \mathcal{B} B does not involve the real solution, and no measurable dependence of the solution on E E E is needed. By Tonelli’s theorem,
E [ ∣ { E ∈ U : ( ω , E ) ∉ B } ∣ ] = ∫ U P ( ( ω , E ) ∉ B ) d E = 0 , \mathbb{E}\left[\left|\left\{E \in U : (\omega,E) \notin\mathcal{B}\right\}\right|\right] = \int_U \mathbb{P}\left((\omega,E) \notin\mathcal{B}\right)\,\mathrm{d}E = 0, E [ ∣ { E ∈ U : ( ω , E ) ∈ / B } ∣ ] = ∫ U P ( ( ω , E ) ∈ / B ) d E = 0 ,
since the integrand vanishes outside the null set N 0 \mathcal{N}_0 N 0 . This implies that, P \mathbb{P} P -almost surely, for Lebesgue-almost every E ∈ U E \in U E ∈ U , we have R μ x ( y ) ( E ) < ∞ R_{\mu_x^{(y)}}(E) < \infty R μ x ( y ) ( E ) < ∞ for every vertex y y y and every neighbor x x x of $y. This is the hypothesis of Proposition 7.9 with I = U I = U I = U and Q = P Q = \mathbb{P} Q = P , and it also holds P ∞ \mathbb{P}^{\infty} P ∞ -almost surely. By Proposition 7.9, H \mathcal{H} H has pure point spectrum in U ∪ ( − U ) U \cup(-U) U ∪ ( − U ) , almost surely under P \mathbb{P} P and under P ∞ \mathbb{P}^{\infty} P ∞ .
Remark 8.11. The energy 0 0 0 is excluded in Theorem 8.4 , and indeed H \mathcal{H} H has finitely supported eigenvectors with eigenvalue zero (Proposition 3.25). In Proposition 8.3 , the real solution π \pi π is arbitrary, and the recursive tree process need not be endogenous, that is, the denominators D v D_v D v need not be functions of the marked tree. This causes no difficulty, because the assertions of Proposition 8.3 (iii) are events of the marked tree; see Step 3 of the proof. If E ∈ E p E \in\mathcal{E}_p E ∈ E p , the no-ac hypothesis holds at E E E , and π \pi π is the real boundary law, then the boundary denominators of Proposition 5.11 (ii) form an endogenous recursive tree process with marginal π \pi π (Remark 5.15). For this process, (29) shows that F ( v ) F(v) F ( v ) is the limit of G v o ( E + i η ) G_{v o}(E + i\eta) G v o ( E + i η ) as η ↓ 0 \eta\downarrow0 η ↓ 0 .
Resonances along paths
This section and the next prove that the boundary values of the cavity variables cannot all be real at an energy where the real boundary law has positive free energy at s = 1 s = 1 s = 1 . The argument follows the resonance strategy of Aizenman and Warzel [12]; variants of it appear in [4, 7]. We look for vertices x x x at a large depth whose Green function entry G o x G_{ox} G o x is exponentially large, and we use the Ward identity to convert such a vertex into a large imaginary part at the root. In this section, we collect the deterministic identities and the estimates along one or two paths that this argument uses. The counting argument itself is given in Section 10.
In § 8.3.1 , we express the Green function between the endpoints of a path through a product of transfer matrices, and we record the Schur complement formula for two vertices. In Section 9.2, we study the end of a path: we show that two vertices with a single child each bring the rootward cavity variable into a fixed disc, uniformly in its incoming value, and we estimate the probability that a pendant leaf creates a large diagonal entry of the Green function, at one vertex and at two vertices simultaneously. Section 9.3 selects blocks of transfer matrices whose truncated free energy approximates the free energy and estimates the probability of a regular growth event, on which the growth of the transfer products along a path stays close to its typical value. Section 9.4 averages bilinear forms of products of transfer matrices over four groups of three independent factors. Finally, Section 9.5 records a semicontinuity property of probabilities used to pass from the real axis to a small positive imaginary part.
The results of Sections 9.1 and 9.2 are deterministic, or concern one or two squared moduli with an explicit density, and they hold under the hypotheses on p p p and z z z stated in each lemma; in particular, they do not use a real solution. From Section 9.3 on, we fix p > 1 p > 1 p > 1 , an energy E > 0 E > 0 E > 0 , and a real solution π \pi π at energy E E E ; in Section 10, π \pi π is the real boundary law. Whenever an energy E E E is fixed, we write z = E + i η z = E + i\eta z = E + i η with η ≥ 0 \eta\ge0 η ≥ 0 . Statements about the tree hold on the almost sure event Ω s a \Omega_{\mathrm{sa}} Ω sa of Theorem 3.10 , intersected with the almost sure event on which all edge weights are nonzero.
Transfer matrices along a path
In this section, we prove two deterministic identities for the resolvent of a component of a forest obtained by a finite removal. The first, Lemma 9.3 , expresses the Green function between the endpoints of a path, and its ratio to the diagonal entry at the far endpoint, through the product of the transfer matrices along the path. In Section 10, it is applied in Lemma 10.8 to the path from the root to a resonant vertex, where the ratio converts a large diagonal entry into a large entry G o x G_{ox} G o x , and in Lemma 10.11 to the path between the parents of two resonant vertices, where the Green function measures the interaction between them. The second, Lemma 9.4, is the Schur complement formula for the 2 × 2 2 \times2 2 × 2 block of the resolvent at two vertices; it is used in Lemma 10.12 to write the two diagonal entries at a pair of resonant vertices as functions of the weights of their pendant leaves.
We use the transfer matrices of Definition 6.1 with the scaling g = 1 g=1 g = 1 , in which the denominator is B = D B=D B = D and β = Γ / p \beta=\Gamma/p β = Γ/ p is the squared weight, and we allow complex denominators, as in Proposition 6.19.
Definition 9.1. (i) For d ∈ C d \in\mathbb{C} d ∈ C and t > 0 t>0 t > 0 , the factor with denominator d d d and squared weight t t t is the matrix
M ( d , t ) = t − 1 / 2 ( 0 1 − t d ) M(d,t)=t^{-1/2}\begin{pmatrix} 0 & 1 \\ -t & d \end{pmatrix} M ( d , t ) = t − 1/2 ( 0 − t 1 d ) For real d d d , this is the matrix M ( b , β ) M(b,\beta) M ( b , β ) of Definition 6.1 with b = d b=d b = d and β = t \beta=t β = t . We write e 2 = ( 0 , 1 ) T e_2=(0,1)^{\mathsf{T}} e 2 = ( 0 , 1 ) T .
(ii) Let C C C be a component of a forest obtained by a finite removal, let w 0 , w 1 , … , w r w_0,w_1,\ldots,w_r w 0 , w 1 , … , w r be a path in C C C with r ≥ 0 r\geq0 r ≥ 0 , and let z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R . For 0 ≤ k ≤ r 0\leq k\leq r 0 ≤ k ≤ r , the side denominator at w k w_k w k is
d k = z + ∑ a ∣ h w k a ∣ 2 G a → w k C ( z ) , d_k=z+\sum_a |h_{w_k a}|^2G^C_{a\to w_k}(z), d k = z + a ∑ ∣ h w k a ∣ 2 G a → w k C ( z ) , where the sum runs over the neighbors a a a of w k w_k w k in C C C other than w k − 1 w_{k-1} w k − 1 and w k + 1 w_{k+1} w k + 1 (only those of them that exist). For 1 ≤ k ≤ r 1\leq k\leq r 1 ≤ k ≤ r , we set t k = ∣ h w k − 1 w k ∣ 2 t_k=|h_{w_{k-1}w_k}|^2 t k = ∣ h w k − 1 w k ∣ 2 . The rootward denominators are defined by R 0 = d 0 R_0=d_0 R 0 = d 0 and R k = d k − t k / R k − 1 R_k=d_k-t_k/R_{k-1} R k = d k − t k / R k − 1 for 1 ≤ k ≤ r 1\leq k\leq r 1 ≤ k ≤ r , and the transfer product of the path is
A = M ( d 0 , t 1 ) M ( d 1 , t 2 ) ⋯ M ( d r − 1 , t r ) , A=M(d_0,t_1)M(d_1,t_2)\cdots M(d_{r-1},t_r), A = M ( d 0 , t 1 ) M ( d 1 , t 2 ) ⋯ M ( d r − 1 , t r ) , with A = I A=I A = I if $r=0.
In words, the k k k -th factor of a path pairs the side denominator at w k w_k w k with the squared weight of the edge leaving w k w_k w k towards w k + 1 w_{k+1} w k + 1 . Every factor has determinant one, since det M ( d , t ) = t − 1 ( 0 ⋅ d + t ) = 1 \det M(d,t)=t^{-1}(0\cdot d+t)=1 det M ( d , t ) = t − 1 ( 0 ⋅ d + t ) = 1 , and so does every product of factors. Let A A A be a complex 2 × 2 2\times2 2 × 2 matrix with det A = 1 \det A=1 det A = 1 , and let σ 1 ≥ σ 2 ≥ 0 \sigma_1\geq\sigma_2\geq0 σ 1 ≥ σ 2 ≥ 0 be its singular values. Then σ 1 σ 2 = ∣ det A ∣ = 1 \sigma_1\sigma_2=|\det A|=1 σ 1 σ 2 = ∣ det A ∣ = 1 , so that σ 2 > 0 \sigma_2>0 σ 2 > 0 , σ 1 = σ 2 − 1 \sigma_1=\sigma_2^{-1} σ 1 = σ 2 − 1 , and σ 1 ≥ 1 \sigma_1\geq1 σ 1 ≥ 1 . Since ∥ A ∥ = σ 1 \lVert A\rVert=\sigma_1 ∥ A ∥ = σ 1 and ∥ A − 1 ∥ = σ 2 − 1 \lVert A^{-1}\rVert=\sigma_2^{-1} ∥ A − 1 ∥ = σ 2 − 1 , we have
∥ A − 1 ∥ = ∥ A ∥ ≥ 1. (142) \lVert A^{-1}\rVert=\lVert A\rVert\geq1. \tag*{(142)} ∥ A − 1 ∥ = ∥ A ∥ ≥ 1. (142)
Bounding the operator norm by the Hilbert–Schmidt norm, we have
∥ M ( d , t ) ∥ ≤ t − 1 / 2 ( 1 + t 2 + ∣ d ∣ 2 ) 1 / 2 . (143) \lVert M(d,t)\rVert\leq t^{-1/2}(1+t^2+|d|^2)^{1/2}. \tag*{(143)} ∥ M ( d , t )∥ ≤ t − 1/2 ( 1 + t 2 + ∣ d ∣ 2 ) 1/2 . (143)
The following algebraic identity relates products of factors to continuants.
Lemma 9.2. Let r ≥ 1 r \ge1 r ≥ 1 , let d 0 , … , d r ∈ C d_0,\ldots,d_r \in\mathbb{C} d 0 , … , d r ∈ C , and let t 1 , … , t r > 0 t_1,\ldots,t_r > 0 t 1 , … , t r > 0 . Define K − 2 = 0 K_{-2}=0 K − 2 = 0 , K − 1 = 1 K_{-1}=1 K − 1 = 1 , K 0 = d 0 K_0=d_0 K 0 = d 0 , and K k = d k K k − 1 − t k K k − 2 K_k=d_kK_{k-1}-t_kK_{k-2} K k = d k K k − 1 − t k K k − 2 for 1 ≤ k ≤ r 1 \le k \le r 1 ≤ k ≤ r . Let A = M ( d 0 , t 1 ) ⋯ M ( d r − 1 , t r ) A=M(d_0,t_1)\cdots M(d_{r-1},t_r) A = M ( d 0 , t 1 ) ⋯ M ( d r − 1 , t r ) and τ = ( t 1 ⋯ t r ) 1 / 2 \tau=(t_1\cdots t_r)^{1/2} τ = ( t 1 ⋯ t r ) 1/2 . Then
τ e 2 T A = ( − t r K r − 2 , K r − 1 ) . \tau e_2^{\mathsf{T}}A=(-t_rK_{r-2},K_{r-1}). τ e 2 T A = ( − t r K r − 2 , K r − 1 ) . In particular, τ e 2 T A e 2 = K r − 1 \tau e_2^{\mathsf{T}}Ae_2=K_{r-1} τ e 2 T A e 2 = K r − 1 and τ e 2 T A ( 1 , d r ) T = K r \tau e_2^{\mathsf{T}}A(1,d_r)^{\mathsf{T}}=K_r τ e 2 T A ( 1 , d r ) T = K r . If the numbers R 0 = d 0 R_0=d_0 R 0 = d 0 and R k = d k − t k / R k − 1 R_k=d_k-t_k/R_{k-1} R k = d k − t k / R k − 1 , 1 ≤ k ≤ r 1\le k\le r 1 ≤ k ≤ r , are well defined and nonzero, then K k = R 0 R 1 ⋯ R k K_k=R_0R_1\cdots R_k K k = R 0 R 1 ⋯ R k for 0 ≤ k ≤ r 0\le k\le r 0 ≤ k ≤ r .
Proof. Let N k = t k + 1 1 / 2 M ( d k , t k + 1 ) N_k=t_{k+1}^{1/2}M(d_k,t_{k+1}) N k = t k + 1 1/2 M ( d k , t k + 1 ) , which has the rows ( 0 , 1 ) (0,1) ( 0 , 1 ) and ( − t k + 1 , d k ) (-t_{k+1},d_k) ( − t k + 1 , d k ) . We show by induction on l ∈ { 1 , … , r } l\in\{1,\ldots,r\} l ∈ { 1 , … , r } that e 2 T N 0 ⋯ N l − 1 = ( − t l K l − 2 , K l − 1 ) e_2^{\mathsf{T}}N_0\cdots N_{l-1}=(-t_lK_{l-2},K_{l-1}) e 2 T N 0 ⋯ N l − 1 = ( − t l K l − 2 , K l − 1 ) . For l = 1 l=1 l = 1 , we have e 2 T N 0 = ( − t 1 , d 0 ) = ( − t 1 K − 1 , K 0 ) e_2^{\mathsf{T}}N_0=(-t_1,d_0)=(-t_1K_{-1},K_0) e 2 T N 0 = ( − t 1 , d 0 ) = ( − t 1 K − 1 , K 0 ) . If the claim holds for l < r l<r l < r , then
e 2 T N 0 ⋯ N l = ( − t l K l − 2 , K l − 1 ) ( 0 1 − t l + 1 d l ) = ( − t l + 1 K l − 1 , d l K l − 1 − t l K l − 2 ) , e_2^{\mathsf{T}}N_0\cdots N_l=(-t_lK_{l-2},K_{l-1})\begin{pmatrix}0&1\\-t_{l+1}&d_l\end{pmatrix}=(-t_{l+1}K_{l-1},d_lK_{l-1}-t_lK_{l-2}), e 2 T N 0 ⋯ N l = ( − t l K l − 2 , K l − 1 ) ( 0 − t l + 1 1 d l ) = ( − t l + 1 K l − 1 , d l K l − 1 − t l K l − 2 ) ,
which is ( − t l + 1 K l − 1 , K l ) (-t_{l+1}K_{l-1},K_l) ( − t l + 1 K l − 1 , K l ) . Since N 0 ⋯ N r − 1 = τ A N_0\cdots N_{r-1}=\tau A N 0 ⋯ N r − 1 = τ A , the case l = r l=r l = r is the first claim. Multiplying it by e 2 e_2 e 2 and by ( 1 , d r ) T (1,d_r)^{\mathsf{T}} ( 1 , d r ) T gives K r − 1 K_{r-1} K r − 1 and − t r K r − 2 + d r K r − 1 = K r -t_rK_{r-2}+d_rK_{r-1}=K_r − t r K r − 2 + d r K r − 1 = K r . For the last claim, we have K 0 = R 0 K_0=R_0 K 0 = R 0 , and if K k − 1 = R 0 ⋯ R k − 1 K_{k-1}=R_0\cdots R_{k-1} K k − 1 = R 0 ⋯ R k − 1 and K k − 2 = R 0 ⋯ R k − 2 K_{k-2}=R_0\cdots R_{k-2} K k − 2 = R 0 ⋯ R k − 2 (with the empty product equal to 1), then K k = R 0 ⋯ R k − 1 ( d k − t k / R k − 1 ) = R 0 ⋯ R k K_k=R_0\cdots R_{k-1}(d_k-t_k/R_{k-1})=R_0\cdots R_k K k = R 0 ⋯ R k − 1 ( d k − t k / R k − 1 ) = R 0 ⋯ R k . □
We now identify the rootward denominators with cavity variables and obtain the path formulas.
Lemma 9.3. In the setting of Definition 9.1 (ii), suppose that r ≥ 1 r\ge1 r ≥ 1 . Then the rootward denominators are well defined and nonzero, and
R k = − 1 Γ w k → w k + 1 C ( z ) ( 0 ≤ k < r ) , R r = − 1 Γ w r , w r C ( z ) . (144) R_k=-\frac{1}{\Gamma^{\mathsf{C}}_{w_k\to w_{k+1}}(z)}\quad(0\le k<r),\qquad R_r=-\frac{1}{\Gamma^{\mathsf{C}}_{w_r,w_r}(z)}. \tag*{(144)} R k = − Γ w k → w k + 1 C ( z ) 1 ( 0 ≤ k < r ) , R r = − Γ w r , w r C ( z ) 1 . (144) Further,
∣ Γ w 0 w r C ( z ) ∣ = ∣ e 2 T A ( 1 , d r ) T ∣ − 1 , (145) \left|\Gamma^{\mathsf{C}}_{w_0w_r}(z)\right|=\left|e_2^{\mathsf{T}}A(1,d_r)^{\mathsf{T}}\right|^{-1}, \tag*{(145)} Γ w 0 w r C ( z ) = e 2 T A ( 1 , d r ) T − 1 , (145) ∣ Γ w 0 w r C ( z ) ∣ ∣ Γ w r w r C ( z ) ∣ = ∣ e 2 T A e 2 ∣ − 1 ≥ ∥ A ∥ − 1 . (146) \frac{\left|\Gamma^{\mathsf{C}}_{w_0w_r}(z)\right|}{\left|\Gamma^{\mathsf{C}}_{w_rw_r}(z)\right|}=\left|e_2^{\mathsf{T}}Ae_2\right|^{-1}\ge\|A\|^{-1}. \tag*{(146)} Γ w r w r C ( z ) Γ w 0 w r C ( z ) = e 2 T A e 2 − 1 ≥ ∥ A ∥ − 1 . (146)
Proof. We prove (9.3) by induction on k k k . By (22) for the oriented edge from w 0 w_0 w 0 to w 1 w_1 w 1 , the number − 1 / Γ w 0 → w 1 C ( z ) -1/\Gamma^{\mathsf{C}}_{w_0\to w_1}(z) − 1/ Γ w 0 → w 1 C ( z ) equals z + ∑ a ∣ h w 0 a ∣ 2 Γ a → w 0 C ( z ) z+\sum_a|h_{w_0a}|^2\Gamma^{\mathsf{C}}_{a\to w_0}(z) z + ∑ a ∣ h w 0 a ∣ 2 Γ a → w 0 C ( z ) , where a a a runs over the neighbors of w 0 w_0 w 0 in C other than w 1 w_1 w 1 ; this is d 0 = R 0 d_0=R_0 d 0 = R 0 . Let 1 ≤ k < r 1\le k<r 1 ≤ k < r , and suppose that R k − 1 = − 1 / Γ w k − 1 → w k C ( z ) R_{k-1}=-1/\Gamma^{\mathsf{C}}_{w_{k-1}\to w_k}(z) R k − 1 = − 1/ Γ w k − 1 → w k C ( z ) . The neighbors of w k w_k w k in C other than w k + 1 w_{k+1} w k + 1 are w k − 1 w_{k-1} w k − 1 and the neighbors entering the definition of d k d_k d k . Then (22) implies that
− 1 Γ w k → w k + 1 C ( z ) = d k + ∣ h w k w k − 1 ∣ 2 Γ w k − 1 → w k C ( z ) = d k − t k R k − 1 = R k . -\frac{1}{\Gamma^{\mathsf{C}}_{w_k\to w_{k+1}}(z)}=d_k+|h_{w_kw_{k-1}}|^2\Gamma^{\mathsf{C}}_{w_{k-1}\to w_k}(z)=d_k-\frac{t_k}{R_{k-1}}=R_k. − Γ w k → w k + 1 C ( z ) 1 = d k + ∣ h w k w k − 1 ∣ 2 Γ w k − 1 → w k C ( z ) = d k − R k − 1 t k = R k .
The same computation with (21) at w r w_r w r , whose neighbors are w r − 1 w_{r-1} w r − 1 and the neighbors entering d r d_r d r , gives the second identity in (144) . All these numbers are finite and nonzero, since by Proposition 3.16 the expressions in parentheses in (21) and (22) have imaginary parts of absolute value at least ∣ Im z ∣ > 0 \lvert\operatorname{Im} z\rvert> 0 ∣ Im z ∣ > 0 .
By (29) , applied to the path from u = w r u=w_r u = w r to v = w 0 v=w_0 v = w 0 , that is, to the vertices v l = w r − l v_l=w_{r-l} v l = w r − l , we have
G w 0 w r C ( z ) = ( − 1 ) r G w r w r C ( z ) ∏ k = 0 r − 1 h w k w k + 1 Γ w k → w k + 1 C ( z ) . G^{C}_{w_0w_r}(z)=(-1)^rG^{C}_{w_rw_r}(z)\prod_{k=0}^{r-1}h_{w_kw_{k+1}}\Gamma^{C}_{w_k\to w_{k+1}}(z). G w 0 w r C ( z ) = ( − 1 ) r G w r w r C ( z ) k = 0 ∏ r − 1 h w k w k + 1 Γ w k → w k + 1 C ( z ) .
Taking absolute values and using (144) , we find
∣ G w 0 w r C ( z ) ∣ = ( t 1 ⋯ t r ) 1 / 2 ∣ R 0 R 1 ⋯ R r ∣ , ∣ G w 0 w r C ( z ) G w r w r C ( z ) ∣ = ( t 1 ⋯ t r ) 1 / 2 ∣ R 0 ⋯ R r − 1 ∣ . \left\lvert G^{C}_{w_0w_r}(z)\right\rvert=\frac{(t_1\cdots t_r)^{1/2}}{\left\lvert R_0R_1\cdots R_r\right\rvert},\qquad
\left\lvert\frac{G^{C}_{w_0w_r}(z)}{G^{C}_{w_rw_r}(z)}\right\rvert=\frac{(t_1\cdots t_r)^{1/2}}{\left\lvert R_0\cdots R_{r-1}\right\rvert}. G w 0 w r C ( z ) = ∣ R 0 R 1 ⋯ R r ∣ ( t 1 ⋯ t r ) 1/2 , G w r w r C ( z ) G w 0 w r C ( z ) = ∣ R 0 ⋯ R r − 1 ∣ ( t 1 ⋯ t r ) 1/2 .
By Lemma 9.2, R 0 ⋯ R r = K r = τ e 2 T A ( 1 , d r ) T R_0\cdots R_r=K_r=\tau e_2^{\mathsf T}A(1,d_r)^{\mathsf T} R 0 ⋯ R r = K r = τ e 2 T A ( 1 , d r ) T and R 0 ⋯ R r − 1 = K r − 1 = τ e 2 T A e 2 R_0\cdots R_{r-1}=K_{r-1}=\tau e_2^{\mathsf T}Ae_2 R 0 ⋯ R r − 1 = K r − 1 = τ e 2 T A e 2 , where τ = ( t 1 ⋯ t r ) 1 / 2 \tau=(t_1\cdots t_r)^{1/2} τ = ( t 1 ⋯ t r ) 1/2 ; here the weights t k t_k t k are positive, since the edge weights are nonzero. This proves (145) and the identity in (146) ; in particular, e 2 T A e 2 = τ − 1 R 0 ⋯ R r − 1 ≠ 0 e_2^{\mathsf T}Ae_2=\tau^{-1}R_0\cdots R_{r-1}\ne0 e 2 T A e 2 = τ − 1 R 0 ⋯ R r − 1 = 0 . The inequality in (146) follows from the Cauchy–Schwarz inequality ∣ e 2 T A e 2 ∣ = ∣ ⟨ e 2 , A e 2 ⟩ ∣ ≤ ∥ A e 2 ∥ ≤ ∥ A ∥ \lvert e_2^{\mathsf T}Ae_2\rvert=\lvert\langle e_2,Ae_2\rangle\rvert\le\lVert Ae_2\rVert\le\lVert A\rVert ∣ e 2 T A e 2 ∣ = ∣⟨ e 2 , A e 2 ⟩∣ ≤ ∥ A e 2 ∥ ≤ ∥ A ∥ , which holds for every complex 2 × 2 2\times2 2 × 2 matrix (the condition det A = 1 \det A=1 det A = 1 is not needed here).
In (146) , the side denominator d r d_r d r at the endpoint does not enter; the ratio depends only on the side denominators d 0 , … , d r − 1 d_0,\ldots,d_{r-1} d 0 , … , d r − 1 and on the weights of the path. In Section 10, we apply (146) to a path from the root to a vertex x x x of the full tree, and (145) to the path between the parents of two vertices x x x and y y y , in the tree from which x x x and y y y have been deleted.
The next lemma is the Schur complement formula for the 2 × 2 2\times2 2 × 2 block of the resolvent at two vertices. For a component C C C of the forest obtained by a finite removal ( F , D ) (F,D) ( F , D ) and x , y ∈ C x,y\in C x , y ∈ C , we write G a b C ∖ { x , y } ( z ) G^{C\setminus\{x,y\}}_{ab}(z) G ab C ∖ { x , y } ( z ) for the entries of the resolvent of the reduced operator of the finite removal ( F ∪ { x , y } , D ) (F\cup\{x,y\},D) ( F ∪ { x , y } , D ) at vertices a , b ∈ C ∖ { x , y } a,b\in C\setminus\{x,y\} a , b ∈ C ∖ { x , y } . These vanish if a a a and b b b lie in different components of the corresponding reduced forest.
Lemma 9.4. Let C C C be a component of a forest obtained by a finite removal, let x , y ∈ C x,y\in C x , y ∈ C be distinct and not adjacent, and let z ∈ C ∖ R z\in C\setminus\mathbb{R} z ∈ C ∖ R . For u , v ∈ { x , y } u,v\in\{x,y\} u , v ∈ { x , y } , set
S u v ( z ) = − z 1 { u = v } − ∑ a ∼ C u ∑ b ∼ C v h u a G a b C ∖ { x , y } ( z ) h b v , S_{uv}(z)=-z1_{\{u=v\}}-\sum_{a\sim_Cu}\sum_{b\sim_Cv}h_{ua}G^{C\setminus\{x,y\}}_{ab}(z)h_{bv}, S uv ( z ) = − z 1 { u = v } − a ∼ C u ∑ b ∼ C v ∑ h u a G ab C ∖ { x , y } ( z ) h b v , where a ∼ C u a\sim_Cu a ∼ C u means that a a a is a neighbor of u u u in C C C . Then the matrix S ( z ) = ( S u v ( z ) ) u , v ∈ { x , y } S(z)=(S_{uv}(z))_{u,v\in\{x,y\}} S ( z ) = ( S uv ( z ) ) u , v ∈ { x , y } is invertible, and
\begin{pmatrix}
G^C_{xx}(z) & G^C_{xy}(z) \\
G^C_{yx}(z) & G^C_{yy}(z)
\end{pmatrix}=S(z)^{-1} \tag*{(147)}\end{pmatrix}
Let x ′ x' x ′ and y ′ y' y ′ be the neighbors of x x x and y y y on the path from x x x to y y y in C C C . Then
S x y ( z ) = − h x x ′ G x ′ y ′ C ∖ { x , y } ( z ) h y ′ y , ∣ S x y ( z ) S y x ( z ) ∣ = ∣ h x x ′ ∣ 2 ∣ h y y ′ ∣ 2 ∣ G x ′ y ′ C ∖ { x , y } ( z ) ∣ 2 , (148) S_{xy}(z)=-h_{xx'}G_{x'y'}^{C\setminus\{x,y\}}(z)h_{y'y},\qquad\lvert S_{xy}(z)S_{yx}(z)\rvert=\lvert h_{xx'}\rvert^2\lvert h_{yy'}\rvert^2\left\lvert G_{x'y'}^{C\setminus\{x,y\}}(z)\right\rvert^2, \tag*{(148)} S x y ( z ) = − h x x ′ G x ′ y ′ C ∖ { x , y } ( z ) h y ′ y , ∣ S x y ( z ) S y x ( z )∣ = ∣ h x x ′ ∣ 2 ∣ h y y ′ ∣ 2 G x ′ y ′ C ∖ { x , y } ( z ) 2 , (148)
and S x x ( z ) = − z − ∑ a ∼ C x ∣ h x a ∣ 2 G a a C ∖ { x , y } ( z ) S_{xx}(z)=-z-\sum_{a\sim_C x}\lvert h_{xa}\rvert^2G_{aa}^{C\setminus\{x,y\}}(z) S xx ( z ) = − z − ∑ a ∼ C x ∣ h x a ∣ 2 G aa C ∖ { x , y } ( z ) ; the analogous formula holds for S y y ( z ) S_{yy}(z) S y y ( z ) . Finally, if Im z > 0 \operatorname{Im}z>0 Im z > 0 , then the Hermitian matrix Im S ( z ) = ( S ( z ) − S ( z ) ∗ ) / ( 2 i ) \operatorname{Im}S(z)=(S(z)-S(z)^*)/(2i) Im S ( z ) = ( S ( z ) − S ( z ) ∗ ) / ( 2 i ) satisfies Im S ( z ) ≤ − ( Im z ) I \operatorname{Im}S(z)\leq-(\operatorname{Im}z)I Im S ( z ) ≤ − ( Im z ) I .
Proof. Let D ′ D' D ′ be the set of edges of C C C with an endpoint in { x , y } \{x,y\} { x , y } . By Lemma 3.13, the components contained in C C C of the reduced forest obtained by adding D ′ D' D ′ to the removed edges are { x } \{x\} { x } , { y } \{y\} { y } , and the components of C ∖ { x , y } C\setminus\{x,y\} C ∖ { x , y } . With respect to ℓ 2 ( C ) = C { x , y } ⊕ h \ell^2(C)=\mathbb{C}^{\{x,y\}}\oplus\mathfrak{h} ℓ 2 ( C ) = C { x , y } ⊕ h , where h = ℓ 2 ( C ∖ { x , y } ) \mathfrak{h}=\ell^2(C\setminus\{x,y\}) h = ℓ 2 ( C ∖ { x , y }) , the same lemma gives H C = ( 0 ⊕ K ) + W D ′ \mathcal{H}_C=(0\oplus K)+W_{D'} H C = ( 0 ⊕ K ) + W D ′ and Dom ( H C ) = C { x , y } ⊕ Dom ( K ) \operatorname{Dom}(\mathcal{H}_C)=\mathbb{C}^{\{x,y\}}\oplus\operatorname{Dom}(K) Dom ( H C ) = C { x , y } ⊕ Dom ( K ) , where K K K is the direct sum of the operators of the components of C ∖ { x , y } C\setminus\{x,y\} C ∖ { x , y } . These components are also the components contained in C C C of the reduced forest of the finite removal ( F ∪ { x , y } , D ) (F\cup\{x,y\},D) ( F ∪ { x , y } , D ) , and in both reduced forests the operator of such a component is the closure of the weighted adjacency operator of the same induced subtree (Lemma 3.5(i)). Then ⟨ δ a , ( K − z ) − 1 δ b ⟩ = G a b C ∖ { x , y } ( z ) \langle\delta_a,(K-z)^{-1}\delta_b\rangle=G_{ab}^{C\setminus\{x,y\}}(z) ⟨ δ a , ( K − z ) − 1 δ b ⟩ = G ab C ∖ { x , y } ( z ) for a , b ∈ C ∖ { x , y } a,b\in C\setminus\{x,y\} a , b ∈ C ∖ { x , y } . Define the linear map J : C { x , y } → h J:\mathbb{C}^{\{x,y\}}\to\mathfrak{h} J : C { x , y } → h by J δ u = ∑ a ∼ C u h a u δ a J\delta_u=\sum_{a\sim_C u}h_{au}\delta_a J δ u = ∑ a ∼ C u h a u δ a ; it is bounded, since x x x and y y y have finitely many neighbors. Since x x x and y y y are not adjacent, we have, for f ∈ C { x , y } f\in\mathbb{C}^{\{x,y\}} f ∈ C { x , y } and g ∈ Dom ( K ) g\in\operatorname{Dom}(K) g ∈ Dom ( K ) ,
H C ( f ⊕ g ) = J ∗ g ⊕ ( J f + K g ) . \mathcal{H}_C(f\oplus g)=J^*g\oplus(Jf+Kg). H C ( f ⊕ g ) = J ∗ g ⊕ ( J f + K g ) .
Indeed, the coordinate of W D ′ g W_{D'}g W D ′ g at u ∈ { x , y } u\in\{x,y\} u ∈ { x , y } is ∑ a ∼ C u h u a g ( a ) = ⟨ J δ u , g ⟩ \sum_{a\sim_C u}h_{ua}g(a)=\langle J\delta_u,g\rangle ∑ a ∼ C u h u a g ( a ) = ⟨ J δ u , g ⟩ , because h u a = h a u ‾ h_{ua}=\overline{h_{au}} h u a = h a u .
Fix f ∈ C { x , y } f\in\mathbb{C}^{\{x,y\}} f ∈ C { x , y } and set g = − ( K − z ) − 1 J f ∈ Dom ( K ) g=-(K-z)^{-1}Jf\in\operatorname{Dom}(K) g = − ( K − z ) − 1 J f ∈ Dom ( K ) . Then f ⊕ g ∈ Dom ( H C ) f\oplus g\in\operatorname{Dom}(\mathcal{H}_C) f ⊕ g ∈ Dom ( H C ) and
( H C − z ) ( f ⊕ g ) = ( − z f − J ∗ ( K − z ) − 1 J f ) ⊕ 0 = ( S ( z ) f ) ⊕ 0 , (\mathcal{H}_C-z)(f\oplus g)=(-zf-J^*(K-z)^{-1}Jf)\oplus0=(S(z)f)\oplus0, ( H C − z ) ( f ⊕ g ) = ( − z f − J ∗ ( K − z ) − 1 J f ) ⊕ 0 = ( S ( z ) f ) ⊕ 0 ,
where we used that the matrix of J ∗ ( K − z ) − 1 J J^*(K-z)^{-1}J J ∗ ( K − z ) − 1 J has the entries ⟨ J δ u , ( K − z ) − 1 J δ v ⟩ = ∑ a , b h u a G a b C ∖ { x , y } ( z ) h b v \langle J\delta_u,(K-z)^{-1}J\delta_v\rangle=\sum_{a,b}h_{ua}G_{ab}^{C\setminus\{x,y\}}(z)h_{bv} ⟨ J δ u , ( K − z ) − 1 J δ v ⟩ = ∑ a , b h u a G ab C ∖ { x , y } ( z ) h b v . Since H C \mathcal{H}_C H C is self-adjoint, H C − z \mathcal{H}_C-z H C − z is a bijection from Dom ( H C ) \operatorname{Dom}(\mathcal{H}_C) Dom ( H C ) onto ℓ 2 ( C ) \ell^2(C) ℓ 2 ( C ) with inverse ( H C − z ) − 1 (\mathcal{H}_C-z)^{-1} ( H C − z ) − 1 . Then ( H C − z ) − 1 ( ( S ( z ) f ) ⊕ 0 ) = f ⊕ g (\mathcal{H}_C-z)^{-1}((S(z)f)\oplus0)=f\oplus g ( H C − z ) − 1 (( S ( z ) f ) ⊕ 0 ) = f ⊕ g . Projecting onto C { x , y } \mathbb{C}^{\{x,y\}} C { x , y } , we find Π ( H C − z ) − 1 Π S ( z ) f = f \Pi(\mathcal{H}_C-z)^{-1}\Pi S(z)f=f Π ( H C − z ) − 1 Π S ( z ) f = f , where Π \Pi Π is the orthogonal projection onto C { x , y } \mathbb{C}^{\{x,y\}} C { x , y } . In particular, S ( z ) S(z) S ( z ) is injective, hence invertible, and its inverse is the matrix on the left side of (9.6).
For (9.7), let a ∼ C x a\sim_C x a ∼ C x and b ∼ C y b\sim_C y b ∼ C y lie in the same component of C ∖ { x , y } C\setminus\{x,y\} C ∖ { x , y } . A path from a a a to b b b in this component, preceded by x x x and followed by y y y , is a path from x x x to y y y in the tree C C C . Since paths in a tree are unique, we have a = x ′ a=x' a = x ′ and b = y ′ b=y' b = y ′ . All other terms in the sum defining S x y ( z ) S_{xy}(z) S x y ( z ) vanish, which gives the first identity. The same argument shows that S y x ( z ) = − h y y ′ G y ′ x ′ C ∖ { x , y } ( z ) h x ′ x S_{yx}(z)=-h_{yy'}G_{y'x'}^{C\setminus\{x,y\}}(z)h_{x'x} S y x ( z ) = − h y y ′ G y ′ x ′ C ∖ { x , y } ( z ) h x ′ x and Lemma 3.20, applied to the component of C ∖ { x , y } C\setminus\{x,y\} C ∖ { x , y } containing x ′ x' x ′ and y ′ y' y ′ , gives ∣ G y ′ x ′ C ∖ { x , y } ( z ) ∣ = ∣ G x ′ y ′ C ∖ { x , y } ( z ) ∣ \left\lvert G_{y'x'}^{C\setminus\{x,y\}}(z)\right\rvert=\left\lvert G_{x'y'}^{C\setminus\{x,y\}}(z)\right\rvert G y ′ x ′ C ∖ { x , y } ( z ) = G x ′ y ′ C ∖ { x , y } ( z ) . This proves the second identity. Finally, two distinct neighbors of x x x lie in different components of ( F ∪ { x } , D ) (F \cup\{x\},D) ( F ∪ { x } , D ) , by Lemma 3.5 (ii), and a fortiori in different components of C ∖ { x , y } C \setminus\{x,y\} C ∖ { x , y } , since deleting y y y as well only splits components further; the cross terms in the sum defining S x x ( z ) S_{xx}(z) S xx ( z ) vanish, and we obtain the formula for S x x ( z ) S_{xx}(z) S xx ( z ) .
For the last claim, let Im z > 0 \operatorname{Im} z > 0 Im z > 0 and f ∈ C { x , y } f \in\mathbb{C}^{\{x,y\}} f ∈ C { x , y } , and set g = J f g=Jf g = J f . Since K K K is self-adjoint, we have Im ⟨ g , ( K − z ) − 1 g ⟩ = Im z ∥ ( K − z ) − 1 g ∥ 2 ≥ 0 \operatorname{Im}\langle g,(K-z)^{-1}g\rangle=\operatorname{Im}z\lVert(K-z)^{-1}g\rVert^{2}\ge0 Im ⟨ g , ( K − z ) − 1 g ⟩ = Im z ∥( K − z ) − 1 g ∥ 2 ≥ 0 . We have ⟨ f , Im S ( z ) f ⟩ = Im ⟨ f , S ( z ) f ⟩ = − Im z ∥ f ∥ 2 − Im ⟨ g , ( K − z ) − 1 g ⟩ ≤ − Im z ∥ f ∥ 2 \langle f,\operatorname{Im}S(z)f\rangle=\operatorname{Im}\langle f,S(z)f\rangle=-\operatorname{Im}z\lVert f\rVert^{2}-\operatorname{Im}\langle g,(K-z)^{-1}g\rangle\le-\operatorname{Im}z\lVert f\rVert^{2} ⟨ f , Im S ( z ) f ⟩ = Im ⟨ f , S ( z ) f ⟩ = − Im z ∥ f ∥ 2 − Im ⟨ g , ( K − z ) − 1 g ⟩ ≤ − Im z ∥ f ∥ 2 .
Endpoint estimates
In this section, we study the last few vertices of a path. The path to a resonant vertex of Section 10 ends with a single-child segment, that is, a sequence of vertices with exactly one child each, and its endpoint has a pendant leaf and a surviving child. Lemmas 9.5 and 9.6 show that, with a probability bounded below uniformly in the rest of the tree, the single-child segment places the denominator with the leaf removed at the endpoint in a fixed region; they are used in Lemmas 10.8 and 10.9. Lemma 9.7 bounds the probability that the weight of the pendant leaf makes the diagonal resolvent entry at the endpoint large, from above and from below; the lower bound is used in the first moment (Lemma 10.9). Lemma 9.8 bounds the probability that this happens at two vertices simultaneously; it is used in the second moment (Lemma 10.12). Finally, Lemma 9.9 records the properties of the quantile ζ α ( η ) \zeta_{\alpha}(\eta) ζ α ( η ) of Im Γ o \operatorname{Im}\Gamma_o Im Γ o on which the contradiction argument of Section 10.6 is based.
We fix E > 0 E>0 E > 0 and write z = E + i η z=E+\mathrm{i}\eta z = E + i η . We first show that two consecutive vertices without side children bring the rootward cavity variable into a fixed disc, uniformly in its incoming value. Consider a vertex w w w with a single neighbor w ′ w' w ′ other than the next vertex of a path, and let X = − Γ w ′ → w ( z ) X=-\Gamma_{w'\to w}(z) X = − Γ w ′ → w ( z ) be the incoming rootward variable. If the squared weight of the edge w w ′ ww' w w ′ is t t t , then by (22) the outgoing variable is ( z − t X ) − 1 (z-tX)^{-1} ( z − tX ) − 1 . We set
ρ 1 = e − 1 − e − 2 , ε E = ρ 1 E 16 , b 2 = ε E E 8 , b 3 = 3 E 2 32 . (149) \rho_1=e^{-1}-e^{-2},\qquad\varepsilon_E=\frac{\rho_1 E}{16},\qquad b_2=\frac{\varepsilon_E E}{8},\qquad b_3=\frac{3E^2}{32}. \tag*{(149)} ρ 1 = e − 1 − e − 2 , ε E = 16 ρ 1 E , b 2 = 8 ε E E , b 3 = 32 3 E 2 . (149)
and we let D E = { X ∈ C : ∣ X − E − 1 ∣ ≤ ( 3 E ) − 1 } \mathcal{D}_E=\{X\in\mathbb{C}:|X-E^{-1}|\le(3E)^{-1}\} D E = { X ∈ C : ∣ X − E − 1 ∣ ≤ ( 3 E ) − 1 } . Every X ∈ D E X\in\mathcal{D}_E X ∈ D E satisfies ∣ X ∣ ≤ 4 / ( 3 E ) |X|\le4/(3E) ∣ X ∣ ≤ 4/ ( 3 E ) .
Lemma 9.5. Let E > 0 E>0 E > 0 and 0 ≤ η ≤ E / 8 0\le\eta\le E/8 0 ≤ η ≤ E /8 .
(i) Let X ∈ C X\in\mathbb{C} X ∈ C , and let t t t be a random variable with density p e − p t pe^{-pt} p e − pt on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . Then
P ( 1 p < t < 2 p , ∣ z − t X ∣ > ε E ) ≥ ρ 1 2 . \mathbb{P}\left(\frac{1}{p}<t<\frac{2}{p},\ |z-tX|>\varepsilon_E\right)\ge\frac{\rho_1}{2}. P ( p 1 < t < p 2 , ∣ z − tX ∣ > ε E ) ≥ 2 ρ 1 . (ii) If ∣ X ∣ < ε E − 1 |X|<\varepsilon_E^{-1} ∣ X ∣ < ε E − 1 and 0 < t < b 2 0<t<b_2 0 < t < b 2 , then z − t X ≠ 0 z-tX\ne0 z − tX = 0 and ( z − t X ) − 1 ∈ D E (z-tX)^{-1}\in\mathcal{D}_E ( z − tX ) − 1 ∈ D E .
(iii) If X ∈ D E X\in\mathcal{D}_E X ∈ D E and 0 < t ≤ b 3 0<t\le b_3 0 < t ≤ b 3 , then z − t X ≠ 0 z-tX\ne0 z − tX = 0 and ( z − t X ) − 1 ∈ D E (z-tX)^{-1}\in\mathcal{D}_E ( z − tX ) − 1 ∈ D E .
Proof. We first note that if w ∈ C w \in\mathbb{C} w ∈ C satisfies ∣ w − E ∣ ≤ E / 4 |w-E| \le E/4 ∣ w − E ∣ ≤ E /4 , then ∣ w ∣ ≥ 3 E / 4 |w| \ge3E/4 ∣ w ∣ ≥ 3 E /4 , and
∣ 1 w − 1 E ∣ = ∣ E − w ∣ ∣ w ∣ E ≤ E / 4 3 E 2 / 4 = 1 3 E . (150) \left|\frac{1}{w}-\frac{1}{E}\right|=\frac{|E-w|}{|w|E}\le\frac{E/4}{3E^{2}/4}=\frac{1}{3E}. \tag*{(150)} w 1 − E 1 = ∣ w ∣ E ∣ E − w ∣ ≤ 3 E 2 /4 E /4 = 3 E 1 . (150)
so that w − 1 ∈ D E w^{-1} \in\mathcal{D}_{E} w − 1 ∈ D E . For (ii), we have ∣ z − t X − E ∣ ≤ η + t ∣ X ∣ < E / 8 + b 2 / ε E = E / 4 |z-tX-E| \le\eta+t|X| < E/8+b_{2}/\varepsilon_{E}=E/4 ∣ z − tX − E ∣ ≤ η + t ∣ X ∣ < E /8 + b 2 / ε E = E /4 , and (9.9) applies. For (iii), we have t ∣ X ∣ ≤ b 3 ⋅ 4 / ( 3 E ) = E / 8 t|X| \le b_{3}\cdot4/(3E)=E/8 t ∣ X ∣ ≤ b 3 ⋅ 4/ ( 3 E ) = E /8 , so that ∣ z − t X − E ∣ ≤ E / 4 |z-tX-E| \le E/4 ∣ z − tX − E ∣ ≤ E /4 , and (9.9) applies again.
For (i), we have P ( 1 / p < t < 2 / p ) = e − 1 − e − 2 = ρ 1 \mathbb{P}(1/p<t<2/p)=e^{-1}-e^{-2}=\rho_{1} P ( 1/ p < t < 2/ p ) = e − 1 − e − 2 = ρ 1 , and it suffices to show that the event { 1 / p < t < 2 / p , ∣ z − t X ∣ ≤ ε E } \{1/p<t<2/p,\ |z-tX|\le\varepsilon_{E}\} { 1/ p < t < 2/ p , ∣ z − tX ∣ ≤ ε E } has probability at most ρ 1 / 2 \rho_{1}/2 ρ 1 /2 . If ∣ X ∣ ≤ p E / 4 |X|\le pE/4 ∣ X ∣ ≤ pE /4 , then for t < 2 / p t<2/p t < 2/ p we have ∣ z − t X ∣ ≥ Re z − t ∣ X ∣ > E − E / 2 = E / 2 > ε E |z-tX|\ge\operatorname{Re}z-t|X|>E-E/2=E/2>\varepsilon_{E} ∣ z − tX ∣ ≥ Re z − t ∣ X ∣ > E − E /2 = E /2 > ε E , and the event is empty. If ∣ X ∣ > p E / 4 |X|>pE/4 ∣ X ∣ > pE /4 , then ∣ z − t X ∣ = ∣ X ∣ ∣ t − z / X ∣ |z-tX|=|X||t-z/X| ∣ z − tX ∣ = ∣ X ∣∣ t − z / X ∣ , and the set of real t t t with ∣ z − t X ∣ ≤ ε E |z-tX|\le\varepsilon_{E} ∣ z − tX ∣ ≤ ε E is contained in an interval of length 2 ε E / ∣ X ∣ < 8 ε E / ( p E ) 2\varepsilon_{E}/|X|<8\varepsilon_{E}/(pE) 2 ε E /∣ X ∣ < 8 ε E / ( pE ) . Since the density of t t t is at most p p p , the event has probability at most 8 ε E / E = ρ 1 / 2 8\varepsilon_{E}/E=\rho_{1}/2 8 ε E / E = ρ 1 /2 .
In applications, X X X is random and t = Y / p t=Y/p t = Y / p , where Y Y Y is the squared modulus of an edge mark. Part (i) then holds conditionally, in the following sense. Let G \mathcal{G} G be a σ \sigma σ -field, let X X X be a G \mathcal{G} G -measurable complex random variable, and suppose that Y Y Y is fresh given G \mathcal{G} G in the sense of Definition 4.16, with M = 1 M=1 M = 1 ; that is, Y Y Y has the law Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) conditionally on G \mathcal{G} G . We apply the substitution identity (4.14) with m = 1 m=1 m = 1 , c = X c=X c = X , a 1 = 0 a_{1}=0 a 1 = 0 , and
Φ ( c , a 1 , y ) = 1 { 1 < y < 2 , ∣ z − y p c ∣ > ε E } . \Phi(c,a_{1},y)=\mathbf{1}\left\{1<y<2,\ \left|z-\frac{y}{p}c\right|>\varepsilon_{E}\right\}. Φ ( c , a 1 , y ) = 1 { 1 < y < 2 , z − p y c > ε E } .
For each fixed value of X X X , the integral of Φ ( X , 0 , ⋅ ) \Phi(X,0,\cdot) Φ ( X , 0 , ⋅ ) against Exp ( 1 ) \operatorname{Exp}(1) Exp ( 1 ) is the probability in (i) with t = Y / p t=Y/p t = Y / p , which has density p e − p t pe^{-pt} p e − pt . We conclude that
P ( 1 p < t < 2 p , ∣ z − t X ∣ > ε E | G ) ≥ ρ 1 2 \mathbb{P}\left(\frac{1}{p}<t<\frac{2}{p},\ |z-tX|>\varepsilon_{E}\,\middle|\,\mathcal{G}\right)\ge\frac{\rho_{1}}{2} P ( p 1 < t < p 2 , ∣ z − tX ∣ > ε E G ) ≥ 2 ρ 1
almost surely.
In the same way, Lemmas 9.7 and 9.8 below hold conditionally on a σ \sigma σ -field G \mathcal{G} G given which the squared moduli of the marks of the leaf edges are fresh and the parameters are revealed: the parameter p p p in Lemma 9.7, and in Lemma 9.8 the three numbers B x x B_{xx} B xx , B y y B_{yy} B y y , and B x y B y x B_{xy}B_{yx} B x y B y x , through which alone the event and the bound depend. In each case, (4.14) is applied with c c c and a 1 , … , a m a_{1},\ldots,a_{m} a 1 , … , a m equal to these revealed numbers.
The next lemma describes the end of a path. In Section 10, the endpoint x x x of a path has three neighbors: its parent, whose rootward variable X X X has been brought into D E \mathcal{D}_{E} D E by the single-child segment, as in Lemma 9.5; a child y y y with a bounded cavity variable; and a pendant leaf. The denominator with the leaf removed at x x x is the quantity Δ \Delta Δ below.
Lemma 9.6. Let E > 0 E>0 E > 0 , 0 ≤ η ≤ E / 8 0\le\eta\le E/8 0 ≤ η ≤ E /8 , and R ′ > 0 R^{\prime}>0 R ′ > 0 . Let X ∈ D E X\in\mathcal{D}_{E} X ∈ D E , 0 < t ≤ b 3 0<t\le b_{3} 0 < t ≤ b 3 , X ′ ∈ C X^{\prime}\in\mathbb{C} X ′ ∈ C with ∣ X ′ ∣ ≤ R ′ |X^{\prime}|\le R^{\prime} ∣ X ′ ∣ ≤ R ′ , and 0 < t ′ ≤ E / ( 8 R ′ ) 0<t^{\prime}\le E/(8R^{\prime}) 0 < t ′ ≤ E / ( 8 R ′ ) , and set Δ = z − t X − t ′ X ′ \Delta=z-tX-t^{\prime}X^{\prime} Δ = z − tX − t ′ X ′ . Then ∣ Δ − E ∣ ≤ 3 E / 8 |\Delta-E|\le3E/8 ∣Δ − E ∣ ≤ 3 E /8 and
E 2 2 ≤ Re ( z Δ ) ≤ 3 E 2 2 . \frac{E^{2}}{2}\le\operatorname{Re}(z\Delta)\le\frac{3E^{2}}{2}. 2 E 2 ≤ Re ( z Δ ) ≤ 2 3 E 2 .
Proof. We have ∣ Δ − E ∣ ≤ η + t ∣ X ∣ + t ′ ∣ X ′ ∣ ≤ E / 8 + E / 8 + E / 8 = 3 E / 8 |\Delta-E|\leq\eta+t|X|+t'|X'|\leq E/8+E/8+E/8=3E/8 ∣Δ − E ∣ ≤ η + t ∣ X ∣ + t ′ ∣ X ′ ∣ ≤ E /8 + E /8 + E /8 = 3 E /8 , where we used t ∣ X ∣ ≤ b 3 ⋅ 4 / ( 3 E ) = E / 8 t|X|\leq b_3\cdot4/(3E)=E/8 t ∣ X ∣ ≤ b 3 ⋅ 4/ ( 3 E ) = E /8 . In particular, 5 E / 8 ≤ Re Δ ≤ 11 E / 8 5E/8\leq\operatorname{Re}\Delta\leq11E/8 5 E /8 ≤ Re Δ ≤ 11 E /8 and ∣ Im Δ ∣ ≤ 3 E / 8 |\operatorname{Im}\Delta|\leq3E/8 ∣ Im Δ∣ ≤ 3 E /8 . Since Re ( z Δ ) = E Re Δ − η Im Δ \operatorname{Re}(z\Delta)=E\operatorname{Re}\Delta-\eta\operatorname{Im}\Delta Re ( z Δ ) = E Re Δ − η Im Δ and η ≤ E / 8 \eta\leq E/8 η ≤ E /8 , we have
Re ( z Δ ) ≥ 5 E 2 8 − 3 E 2 64 ≥ E 2 2 , Re ( z Δ ) ≤ 11 E 2 8 + 3 E 2 64 ≤ 3 E 2 2 . \operatorname{Re}(z\Delta)\geq\frac{5E^2}{8}-\frac{3E^2}{64}\geq\frac{E^2}{2},\qquad\operatorname{Re}(z\Delta)\leq\frac{11E^2}{8}+\frac{3E^2}{64}\leq\frac{3E^2}{2}. Re ( z Δ ) ≥ 8 5 E 2 − 64 3 E 2 ≥ 2 E 2 , Re ( z Δ ) ≤ 8 11 E 2 + 64 3 E 2 ≤ 2 3 E 2 .
We next estimate the probability that a pendant leaf creates a large diagonal entry. If x x x has a leaf child ℓ \ell ℓ with squared weight t = ∣ h x ℓ ∣ 2 t=|h_{x\ell}|^2 t = ∣ h x ℓ ∣ 2 , then Γ ℓ → x ( z ) = − 1 / z \Gamma_{\ell\to x}(z)=-1/z Γ ℓ → x ( z ) = − 1/ z , and (21) gives
G x x ( z ) = − ( Δ − t z ) − 1 = z t − z Δ . (151) G_{xx}(z)=-\left(\Delta-\frac{t}{z}\right)^{-1}=\frac{z}{t-z\Delta}. \tag*{(151)} G xx ( z ) = − ( Δ − z t ) − 1 = t − z Δ z . (151)
where Δ = − 1 / G x x { ℓ } ( z ) \Delta=-1/G_{xx}^{\{\ell\}}(z) Δ = − 1/ G xx { ℓ } ( z ) is the denominator with the leaf removed, computed in the tree from which ℓ \ell ℓ is deleted. In particular, G x x ( z ) G_{xx}(z) G xx ( z ) is a Möbius function of t t t , with a pole at the pole parameter p = z Δ \mathfrak{p}=z\Delta p = z Δ . If η > 0 \eta>0 η > 0 , then Im Δ ≥ η > 0 \operatorname{Im}\Delta\geq\eta>0 Im Δ ≥ η > 0 and Im p = E Im Δ + η Re Δ \operatorname{Im}\mathfrak{p}=E\operatorname{Im}\Delta+\eta\operatorname{Re}\Delta Im p = E Im Δ + η Re Δ .
Lemma 9.7. Let p > 0 p>0 p > 0 , E > 0 E>0 E > 0 , 0 < η ≤ 1 0<\eta\leq1 0 < η ≤ 1 , and T ≥ 4 ( E + 1 ) / E 2 T\geq4(E+1)/E^2 T ≥ 4 ( E + 1 ) / E 2 , and let t t t be a random variable with density p e − p t pe^{-pt} p e − pt on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . Let p ∈ C ∖ R \mathfrak{p}\in\mathbb{C}\setminus\mathbb{R} p ∈ C ∖ R .
(i) We have P ( ∣ z ∣ ≥ T ∣ t − p ∣ ) ≤ 2 p ∣ z ∣ / T \mathbb{P}(|z|\geq T|t-\mathfrak{p}|)\leq2p|z|/T P ( ∣ z ∣ ≥ T ∣ t − p ∣ ) ≤ 2 p ∣ z ∣/ T .
(ii) If E 2 / 2 ≤ Re p ≤ 3 E 2 / 2 E^2/2\leq\operatorname{Re}\mathfrak{p}\leq3E^2/2 E 2 /2 ≤ Re p ≤ 3 E 2 /2 and ∣ Im p ∣ ≤ ∣ z ∣ / ( 2 T ) |\operatorname{Im}\mathfrak{p}|\leq|z|/(2T) ∣ Im p ∣ ≤ ∣ z ∣/ ( 2 T ) , then
P ( ∣ z ∣ ≥ T ∣ t − p ∣ ) ≥ c t u n e T , c t u n e = p E e − 2 p E 2 . \mathbb{P}(|z|\geq T|t-\mathfrak{p}|)\geq\frac{c_{\mathrm{tune}}}{T},\qquad c_{\mathrm{tune}}=pEe^{-2pE^2}. P ( ∣ z ∣ ≥ T ∣ t − p ∣ ) ≥ T c tune , c tune = pE e − 2 p E 2 .
Proof. The set of real t t t with ∣ t − p ∣ ≤ ∣ z ∣ / T |t-\mathfrak{p}|\leq|z|/T ∣ t − p ∣ ≤ ∣ z ∣/ T is empty if ∣ Im p ∣ > ∣ z ∣ / T |\operatorname{Im}\mathfrak{p}|>|z|/T ∣ Im p ∣ > ∣ z ∣/ T , and otherwise it is the interval of center Re p \operatorname{Re}\mathfrak{p} Re p and half-length ϱ = ( ( ∣ z ∣ / T ) 2 − ( Im p ) 2 ) 1 / 2 ≤ ∣ z ∣ / T \varrho=((|z|/T)^2-(\operatorname{Im}\mathfrak{p})^2)^{1/2}\leq|z|/T ϱ = (( ∣ z ∣/ T ) 2 − ( Im p ) 2 ) 1/2 ≤ ∣ z ∣/ T . Since the density of t t t is at most p p p , this proves (i). Under the assumptions of (ii), we have ϱ ≥ ( 3 / 2 ) ∣ z ∣ / T ≥ ( 3 / 2 ) E / T \varrho\geq(\sqrt{3}/2)|z|/T\geq(\sqrt{3}/2)E/T ϱ ≥ ( 3 /2 ) ∣ z ∣/ T ≥ ( 3 /2 ) E / T . Further, ϱ ≤ ∣ z ∣ / T ≤ ( E + 1 ) / T ≤ E 2 / 4 \varrho\leq|z|/T\leq(E+1)/T\leq E^2/4 ϱ ≤ ∣ z ∣/ T ≤ ( E + 1 ) / T ≤ E 2 /4 , so that the interval is contained in [ E 2 / 4 , 7 E 2 / 4 ] [E^2/4,7E^2/4] [ E 2 /4 , 7 E 2 /4 ] , where the density of t t t is at least p e − 7 p E 2 / 4 pe^{-7pE^2/4} p e − 7 p E 2 /4 . The probability in (ii) is therefore at least 2 ϱ p e − 7 p E 2 / 4 ≥ 3 p E e − 7 p E 2 / 4 / T ≥ c t u n e / T 2\varrho pe^{-7pE^2/4}\geq\sqrt{3}\,pEe^{-7pE^2/4}/T\geq c_{\mathrm{tune}}/T 2 ϱp e − 7 p E 2 /4 ≥ 3 pE e − 7 p E 2 /4 / T ≥ c tune / T .
By (151) , the event in Lemma 9.7 with p = z Δ \mathfrak{p}=z\Delta p = z Δ is the event ∣ G x x ( z ) ∣ ≥ T |G_{xx}(z)|\geq T ∣ G xx ( z ) ∣ ≥ T .
For two vertices x x x and y y y with pendant leaves, Lemma 9.4 expresses the 2 × 2 2\times2 2 × 2 block of the resolvent at { x , y } \{x,y\} { x , y } as a function of the squared weights of the two leaf edges. The following lemma bounds the probability that both diagonal entries are large.
Lemma 9.8. Let p ≥ 1 p\geq1 p ≥ 1 , z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R , and T > 0 T>0 T > 0 , and let B = ( B u v ) u , v ∈ { x , y } \mathrm{B}=(B_{uv})_{u,v\in\{x,y\}} B = ( B uv ) u , v ∈ { x , y } be a complex 2 × 2 2\times2 2 × 2 matrix such that diag ( t x , t y ) − B \operatorname{diag}(t_x,t_y)-\mathrm{B} diag ( t x , t y ) − B is invertible for all t x , t y > 0 t_x,t_y>0 t x , t y > 0 . Set
R ( t x , t y ) = z ( diag ( t x , t y ) − B ) − 1 , \mathrm{R}(t_x,t_y)=z\left(\operatorname{diag}(t_x,t_y)-\mathrm{B}\right)^{-1}, R ( t x , t y ) = z ( diag ( t x , t y ) − B ) − 1 , and let t x , t y t_x,t_y t x , t y be independent random variables with density p e − p t pe^{-pt} p e − pt on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . Then
P ( ∣ R x x ∣ ≥ T , ∣ R y y ∣ ≥ T ) ≤ 8 p 2 ∣ z ∣ 2 T 2 + 8 p 2 ∣ z ∣ T min { 1 , ∣ B x y B y x ∣ 1 / 2 } . \mathbb{P}(|R_{xx}|\geq T,\ |R_{yy}|\geq T)\leq\frac{8p^2|z|^2}{T^2}+\frac{8p^2|z|}{T}\min\{1,|B_{xy}B_{yx}|^{1/2}\}. P ( ∣ R xx ∣ ≥ T , ∣ R y y ∣ ≥ T ) ≤ T 2 8 p 2 ∣ z ∣ 2 + T 8 p 2 ∣ z ∣ min { 1 , ∣ B x y B y x ∣ 1/2 } .
Proof. Set θ = ∣ z ∣ / T \theta= |z|/T θ = ∣ z ∣/ T , X = t x − B x x X = t_x - B_{xx} X = t x − B xx , Y = t y − B y y Y = t_y - B_{yy} Y = t y − B y y , and c = B x y B y x c = B_{xy}B_{yx} c = B x y B y x . The determinant X Y − c XY-c X Y − c is nonzero, and
R x x = z Y X Y − c , R y y = z X X Y − c . R_{xx} = \frac{zY}{XY-c}, \qquad R_{yy} = \frac{zX}{XY-c}. R xx = X Y − c z Y , R y y = X Y − c z X .
Let E \mathcal{E} E be the event in the statement. We first bound the probability of { ∣ R x x ∣ ≥ T } \{|R_{xx}| \ge T\} { ∣ R xx ∣ ≥ T } , which contains E \mathcal{E} E . Fix t y t_y t y . If Y = 0 Y=0 Y = 0 , then R x x = 0 R_{xx}=0 R xx = 0 . If Y ≠ 0 Y\ne0 Y = 0 , then R x x = z / ( X − c / Y ) R_{xx}=z/(X-c/Y) R xx = z / ( X − c / Y ) , and ∣ R x x ∣ ≥ T |R_{xx}|\ge T ∣ R xx ∣ ≥ T holds if and only if ∣ t x − B x x − c / Y ∣ ≤ θ |t_x-B_{xx}-c/Y|\le\theta ∣ t x − B xx − c / Y ∣ ≤ θ , which confines t x t_x t x to an interval of length at most 2 θ 2\theta 2 θ . By Tonelli’s theorem and the bound p p p on the density of t x t_x t x , we obtain P ( E ) ≤ 2 p θ \mathbb{P}(\mathcal{E})\le2p\theta P ( E ) ≤ 2 pθ .
Next, ∣ R x x ∣ ≥ T |R_{xx}|\ge T ∣ R xx ∣ ≥ T and ∣ R y y ∣ ≥ T |R_{yy}|\ge T ∣ R y y ∣ ≥ T mean that ∣ X Y − c ∣ ≤ θ ∣ Y ∣ |XY-c|\le\theta|Y| ∣ X Y − c ∣ ≤ θ ∣ Y ∣ and ∣ X Y − c ∣ ≤ θ ∣ X ∣ |XY-c|\le\theta|X| ∣ X Y − c ∣ ≤ θ ∣ X ∣ , so on E \mathcal{E} E we have ∣ X Y − c ∣ ≤ θ min { ∣ X ∣ , ∣ Y ∣ } |XY-c|\le\theta\min\{|X|,|Y|\} ∣ X Y − c ∣ ≤ θ min { ∣ X ∣ , ∣ Y ∣ } . Consider the part of E \mathcal{E} E on which ∣ X ∣ ≤ ∣ Y ∣ |X|\le|Y| ∣ X ∣ ≤ ∣ Y ∣ . There X ≠ 0 X\ne0 X = 0 : otherwise the last inequality would give ∣ X Y − c ∣ ≤ 0 |XY-c|\le0 ∣ X Y − c ∣ ≤ 0 , contradicting X Y − c ≠ 0 XY-c\ne0 X Y − c = 0 . Dividing by ∣ X ∣ |X| ∣ X ∣ , we find ∣ Y − c / X ∣ ≤ θ |Y-c/X|\le\theta ∣ Y − c / X ∣ ≤ θ . Then ∣ X ∣ ≤ ∣ Y ∣ ≤ ∣ c ∣ / ∣ X ∣ + θ |X|\le|Y|\le|c|/|X|+\theta ∣ X ∣ ≤ ∣ Y ∣ ≤ ∣ c ∣/∣ X ∣ + θ , that is, ∣ X ∣ 2 − θ ∣ X ∣ ≤ ∣ c ∣ |X|^2-\theta|X|\le|c| ∣ X ∣ 2 − θ ∣ X ∣ ≤ ∣ c ∣ , which implies ∣ X ∣ ≤ θ + ∣ c ∣ 1 / 2 |X|\le\theta+|c|^{1/2} ∣ X ∣ ≤ θ + ∣ c ∣ 1/2 ; indeed, if ∣ X ∣ > θ + ∣ c ∣ 1 / 2 |X|>\theta+|c|^{1/2} ∣ X ∣ > θ + ∣ c ∣ 1/2 , then ∣ X ∣ ( ∣ X ∣ − θ ) > ( θ + ∣ c ∣ 1 / 2 ) ∣ c ∣ 1 / 2 ≥ ∣ c ∣ |X|(|X|-\theta)>(\theta+|c|^{1/2})|c|^{1/2}\ge|c| ∣ X ∣ ( ∣ X ∣ − θ ) > ( θ + ∣ c ∣ 1/2 ) ∣ c ∣ 1/2 ≥ ∣ c ∣ . Thus t x t_x t x lies in an interval of length at most 2 ( θ + ∣ c ∣ 1 / 2 ) 2(\theta+|c|^{1/2}) 2 ( θ + ∣ c ∣ 1/2 ) , and for each such t x t_x t x , the condition ∣ t y − B y y − c / X ∣ ≤ θ |t_y-B_{yy}-c/X|\le\theta ∣ t y − B y y − c / X ∣ ≤ θ confines t y t_y t y to an interval of length at most 2 θ 2\theta 2 θ . The part of E \mathcal{E} E on which ∣ X ∣ ≤ ∣ Y ∣ |X|\le|Y| ∣ X ∣ ≤ ∣ Y ∣ is therefore contained in a set of area at most 4 θ ( θ + ∣ c ∣ 1 / 2 ) 4\theta(\theta+|c|^{1/2}) 4 θ ( θ + ∣ c ∣ 1/2 ) , and by symmetry the same holds for the part on which ∣ Y ∣ ≤ ∣ X ∣ |Y|\le|X| ∣ Y ∣ ≤ ∣ X ∣ . Since the joint density of ( t x , t y ) (t_x,t_y) ( t x , t y ) is at most p 2 p^2 p 2 , we find P ( E ) ≤ 8 p 2 θ ( θ + ∣ c ∣ 1 / 2 ) \mathbb{P}(\mathcal{E})\le8p^2\theta(\theta+|c|^{1/2}) P ( E ) ≤ 8 p 2 θ ( θ + ∣ c ∣ 1/2 ) .
Combining the two bounds, and using 2 p θ ≤ 8 p 2 θ 2p\theta\le8p^2\theta 2 pθ ≤ 8 p 2 θ for p ≥ 1 p\ge1 p ≥ 1 when ∣ c ∣ ≥ 1 |c|\ge1 ∣ c ∣ ≥ 1 , we find
P ( E ) ≤ min { 2 p θ , 8 p 2 θ 2 + 8 p 2 θ ∣ c ∣ 1 / 2 } ≤ 8 p 2 θ 2 + 8 p 2 θ min { 1 , ∣ c ∣ 1 / 2 } . \mathbb{P}(\mathcal{E})\le\min\left\{2p\theta,\;8p^2\theta^2+8p^2\theta|c|^{1/2}\right\}\le8p^2\theta^2+8p^2\theta\min\{1,|c|^{1/2}\}. P ( E ) ≤ min { 2 pθ , 8 p 2 θ 2 + 8 p 2 θ ∣ c ∣ 1/2 } ≤ 8 p 2 θ 2 + 8 p 2 θ min { 1 , ∣ c ∣ 1/2 } .
which is the claim. □
Finally, we record the properties of the quantiles of the imaginary part of the root cavity variable that we use. For p > 1 p>1 p > 1 , E > 0 E>0 E > 0 , η > 0 \eta>0 η > 0 , and α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , we define
ζ α ( η ) = sup { t ≥ 0 : P ∞ ( Im Γ o ( E + i η ) ≥ t ) ≥ α } . (152) \zeta_\alpha(\eta)=\sup\left\{t\ge0:\mathbb{P}^{\infty}\left(\operatorname{Im}\Gamma_o(E+\mathrm{i}\eta)\ge t\right)\ge\alpha\right\}. \tag*{(152)} ζ α ( η ) = sup { t ≥ 0 : P ∞ ( Im Γ o ( E + i η ) ≥ t ) ≥ α } . (152)
Lemma 9.9. Let p > 1 p>1 p > 1 , E > 0 E>0 E > 0 , and α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) .
(i) For every η > 0 \eta>0 η > 0 , we have 0 < ζ α ( η ) < ∞ 0<\zeta_\alpha(\eta)<\infty 0 < ζ α ( η ) < ∞ , P ∞ ( Im Γ o ( E + i η ) ≥ ζ α ( η ) ) ≥ α \mathbb{P}^{\infty}(\operatorname{Im}\Gamma_o(E+\mathrm{i}\eta)\ge\zeta_\alpha(\eta))\ge\alpha P ∞ ( Im Γ o ( E + i η ) ≥ ζ α ( η )) ≥ α , and P ∞ ( Im Γ o ( E + i η ) ≥ t ) < α \mathbb{P}^{\infty}(\operatorname{Im}\Gamma_o(E+\mathrm{i}\eta)\ge t)<\alpha P ∞ ( Im Γ o ( E + i η ) ≥ t ) < α for every t > ζ α ( η ) t>\zeta_\alpha(\eta) t > ζ α ( η ) .
(ii) Let r ∈ ( 0 , 1 ) r\in(0,1) r ∈ ( 0 , 1 ) and C r = sup 0 < η ≤ 1 E ∞ ∣ Γ o ( E + i η ) ∣ r C_r=\sup_{0<\eta\le1}\mathbb{E}^{\infty}|\Gamma_o(E+\mathrm{i}\eta)|^r C r = sup 0 < η ≤ 1 E ∞ ∣ Γ o ( E + i η ) ∣ r , which is finite by Lemma 5.17 . Set R a u x = ( 4 C r / α ) 1 / r R_{\mathrm{aux}}=(4C_r/\alpha)^{1/r} R aux = ( 4 C r / α ) 1/ r . Then for every η ∈ ( 0 , 1 ] \eta\in(0,1] η ∈ ( 0 , 1 ] ,
P ∞ ( ∣ Γ o ( E + i η ) ∣ < R a u x , Im Γ o ( E + i η ) ≥ ζ α ( η ) / 2 ) ≥ 3 α 4 . \mathbb{P}^{\infty}\left(|\Gamma_o(E+\mathrm{i}\eta)|<R_{\mathrm{aux}},\operatorname{Im}\Gamma_o(E+\mathrm{i}\eta)\ge\zeta_\alpha(\eta)/2\right)\ge\frac{3\alpha}{4}. P ∞ ( ∣ Γ o ( E + i η ) ∣ < R aux , Im Γ o ( E + i η ) ≥ ζ α ( η ) /2 ) ≥ 4 3 α .
Proof. Fix η > 0 \eta>0 η > 0 , write I = Im Γ o ( E + i η ) I=\operatorname{Im}\Gamma_o(E+\mathrm{i}\eta) I = Im Γ o ( E + i η ) , and let ψ ( t ) = P ∞ ( I ≥ t ) \psi(t)=\mathbb{P}^{\infty}(I\ge t) ψ ( t ) = P ∞ ( I ≥ t ) . By Lemma 3.19 (i), I > 0 I>0 I > 0 almost surely, so ψ ( t ) → 1 > α \psi(t)\to1>\alpha ψ ( t ) → 1 > α as t ↓ 0 t\downarrow0 t ↓ 0 , and the set in (9.11) contains some $t>0. Since ψ ( t ) ≤ P ∞ ( ∣ Γ o ( E + i η ) ∣ ≥ t ) → 0 \psi(t) \leq\mathbb{P}^{\infty}(|\Gamma_o(E+i\eta)| \geq t) \to0 ψ ( t ) ≤ P ∞ ( ∣ Γ o ( E + i η ) ∣ ≥ t ) → 0 as t → ∞ t \to\infty t → ∞ , the set is bounded. We conclude that 0 < ζ α ( η ) < ∞ 0 < \zeta_\alpha(\eta) < \infty 0 < ζ α ( η ) < ∞ . The function ψ \psi ψ is nonincreasing, and it is left-continuous because { I ≥ t } = ⋃ t ′ < t { I ≥ t ′ } \{I \geq t\} = \bigcup_{t'<t}\{I \geq t'\} { I ≥ t } = ⋃ t ′ < t { I ≥ t ′ } . Then the set in (9.11) is the interval [ 0 , ζ α ( η ) ] [0,\zeta_\alpha(\eta)] [ 0 , ζ α ( η )] . This proves (i). For (ii), Markov’s inequality implies that P ∞ ( ∣ Γ o ( E + i η ) ∣ ≥ R a u x ) ≤ C r R a u x − r = α / 4 \mathbb{P}^{\infty}(|\Gamma_o(E+i\eta)| \geq R_{\mathrm{aux}}) \leq C_rR_{\mathrm{aux}}^{-r} = \alpha/4 P ∞ ( ∣ Γ o ( E + i η ) ∣ ≥ R aux ) ≤ C r R aux − r = α /4 , and by (i), ψ ( ζ α ( η ) / 2 ) ≥ ψ ( ζ α ( η ) ) ≥ α \psi(\zeta_\alpha(\eta)/2) \geq\psi(\zeta_\alpha(\eta)) \geq\alpha ψ ( ζ α ( η ) /2 ) ≥ ψ ( ζ α ( η )) ≥ α . The claim follows.
Blocks and regular growth
The path from the root to a resonant vertex is required to have a transfer product whose norm grows at a prescribed exponential rate γ ^ \widehat{\gamma} γ , with fluctuations of order b n = n 3 / 4 b_n = n^{3/4} b n = n 3/4 . In this section, we estimate the probability of this requirement for independent factors. To control the norm of a product by the norms of its parts, and to have increments that are bounded, we group the factors into blocks of a fixed length m m m and keep only blocks whose denominators and weights lie in a fixed compact set. Lemma 9.11 chooses m m m and the compact set so that the free energy of the truncated blocks is close to the free energy φ \varphi φ , together with its derivative at the exponent s s s of Lemma 6.16 . Lemma 9.12 then shows, by an exponential change of measure, that the regular growth event , on which the i i i -th partial sum of the block growths stays within distance b b b of i m γ ^ im\widehat{\gamma} im γ for every i i i , has probability e − i m k ± s b e^{-imk\pm sb} e − imk ± s b . The lower bound enters the first moment (Lemma 10.9), and the upper bound enters the second moment (Lemma 10.11).
We fix p > 1 p > 1 p > 1 , E > 0 E > 0 E > 0 , and a real solution π \pi π at energy E E E , and we write φ = φ π \varphi= \varphi_\pi φ = φ π for its free energy. The factor law ν f a c \nu_{\mathrm{fac}} ν fac is the law of ( d , t ) (d,t) ( d , t ) , where d d d has law π \pi π , t = Y / p t = Y/p t = Y / p with Y ∼ Exp ( 1 ) Y \sim\operatorname{Exp}(1) Y ∼ Exp ( 1 ) , and d d d and Y Y Y are independent. Factors M ( d i , t i ) M(d_i,t_i) M ( d i , t i ) with independent ( d i , t i ) ∼ ν f a c (d_i,t_i) \sim\nu_{\mathrm{fac}} ( d i , t i ) ∼ ν fac are independent copies of the matrix M ( B , β ) M(B,\beta) M ( B , β ) of § 6.2 with the scaling g = 1 g = 1 g = 1 , in which B = D B = D B = D has law π \pi π and β \beta β has the exponential law of rate K = p K = p K = p . For u ∈ ( 0 , 1 ) u \in(0,1) u ∈ ( 0 , 1 ) and r ≥ 0 r \geq0 r ≥ 0 , we let
f r ( u ) = E ∥ M ( d 1 , t 1 ) ⋯ M ( d r , t r ) ∥ − u f_r(u) = \mathbb{E}\left\|M(d_1,t_1)\cdots M(d_r,t_r)\right\|^{-u} f r ( u ) = E ∥ M ( d 1 , t 1 ) ⋯ M ( d r , t r ) ∥ − u
be the norm moments of Definition 6.14 , computed with independent factors of law ν f a c \nu_{\mathrm{fac}} ν fac ; the order of the factors does not affect the law of the product. By Proposition 5.5 (iii), E ∣ d ∣ − u < ∞ \mathbb{E}|d|^{-u} < \infty E ∣ d ∣ − u < ∞ , and Corollary 6.22 gives
f r ( u ) ≤ e r ( φ ( u ) − log p ) , r ≥ 0 , u ∈ ( 0 , 1 ) . (153) f_r(u) \leq e^{r(\varphi(u)-\log p)}, \qquad r \geq0,\ u \in(0,1). \tag*{(153)} f r ( u ) ≤ e r ( φ ( u ) − l o g p ) , r ≥ 0 , u ∈ ( 0 , 1 ) . (153)
Definition 9.10. Let m ≥ 1 m \geq1 m ≥ 1 and R > 1 R > 1 R > 1 .
(i) A block of length m m m is a family ( d i , t i ) i = 1 m (d_i,t_i)_{i=1}^{m} ( d i , t i ) i = 1 m with d i ∈ C d_i \in\mathbb{C} d i ∈ C and t i > 0 t_i > 0 t i > 0 . Its growth is
Z = log ∥ M ( d 1 , t 1 ) M ( d 2 , t 2 ) ⋯ M ( d m , t m ) ∥ ≥ 0. Z = \log\left\|M(d_1,t_1)M(d_2,t_2)\cdots M(d_m,t_m)\right\| \geq0. Z = log ∥ M ( d 1 , t 1 ) M ( d 2 , t 2 ) ⋯ M ( d m , t m ) ∥ ≥ 0. The block is retained if R − 1 ≤ ∣ d i ∣ ≤ R R^{-1} \leq|d_i| \leq R R − 1 ≤ ∣ d i ∣ ≤ R and R − 1 ≤ t i ≤ R R^{-1} \leq t_i \leq R R − 1 ≤ t i ≤ R for 1 ≤ i ≤ m 1 \leq i \leq m 1 ≤ i ≤ m , and strictly retained if these inequalities are strict. We write H R H_R H R and H R ∘ H_R^\circ H R ∘ for the corresponding events.
(ii) For independent factors ( d i , t i ) ∼ ν f a c (d_i,t_i)\sim\nu_{\mathrm{fac}} ( d i , t i ) ∼ ν fac , the block free energy is
φ m , R ( u ) = log p + 1 m log E [ 1 H R e − u Z ] , u ∈ R . \varphi_{m,R}(u)=\log p+\frac{1}{m}\log\mathbb{E}\left[1_{H_R}e^{-uZ}\right],\qquad u\in\mathbb{R}. φ m , R ( u ) = log p + m 1 log E [ 1 H R e − u Z ] , u ∈ R .
The growth is nonnegative by (142) . By (143) , a factor with R − 1 ≤ ∣ d ∣ ≤ R R^{-1}\leq|d|\leq R R − 1 ≤ ∣ d ∣ ≤ R and R − 1 ≤ t ≤ R R^{-1}\leq t\leq R R − 1 ≤ t ≤ R has norm at most C R = R 1 / 2 ( 1 + 2 R 2 ) 1 / 2 C_R=R^{1/2}(1+2R^2)^{1/2} C R = R 1/2 ( 1 + 2 R 2 ) 1/2 , and a retained block satisfies
0 ≤ Z ≤ m log C R . (154) 0\leq Z\leq m\log C_R. \tag*{(154)} 0 ≤ Z ≤ m log C R . (154)
The probability P ( H R ) \mathbb{P}(H_R) P ( H R ) is positive. Indeed, since π \pi π has full support (Proposition 5.5 (iv)), the set { R − 1 ≤ ∣ d ∣ ≤ R } \{R^{-1}\leq|d|\leq R\} { R − 1 ≤ ∣ d ∣ ≤ R } , which contains the open interval ( R − 1 , R ) (R^{-1},R) ( R − 1 , R ) , has positive π \pi π -probability; the event { R − 1 ≤ t ≤ R } \{R^{-1}\leq t\leq R\} { R − 1 ≤ t ≤ R } has positive probability since R > 1 R>1 R > 1 ; and H R H_R H R is an intersection of 2 m 2m 2 m independent such events. Let Λ m , R ( u ) = log E [ 1 H R e − u Z ] \Lambda_{m,R}(u)=\log\mathbb{E}[1_{H_R}e^{-uZ}] Λ m , R ( u ) = log E [ 1 H R e − u Z ] , so that φ m , R ( u ) = log p + m − 1 Λ m , R ( u ) \varphi_{m,R}(u)=\log p+m^{-1}\Lambda_{m,R}(u) φ m , R ( u ) = log p + m − 1 Λ m , R ( u ) . Since 0 ≤ Z ≤ m log C R 0\leq Z\leq m\log C_R 0 ≤ Z ≤ m log C R on H R H_R H R and P ( H R ) > 0 \mathbb{P}(H_R)>0 P ( H R ) > 0 , the function u ↦ E [ 1 H R e − u Z ] u\mapsto\mathbb{E}[1_{H_R}e^{-uZ}] u ↦ E [ 1 H R e − u Z ] is finite and positive, and it is smooth on R \mathbb{R} R , with derivatives E [ 1 H R ( − Z ) k e − u Z ] \mathbb{E}[1_{H_R}(-Z)^k e^{-uZ}] E [ 1 H R ( − Z ) k e − u Z ] obtained by differentiating under the expectation (the integrands are bounded uniformly for u u u in compact sets). Then φ m , R \varphi_{m,R} φ m , R is finite and smooth on R \mathbb{R} R . It is convex by Hölder’s inequality: for u 0 , u 1 ∈ R u_0,u_1\in\mathbb{R} u 0 , u 1 ∈ R and τ ∈ ( 0 , 1 ) \tau\in(0,1) τ ∈ ( 0 , 1 ) , applying Hölder’s inequality with the exponents 1 / ( 1 − τ ) 1/(1-\tau) 1/ ( 1 − τ ) and 1 / τ 1/\tau 1/ τ to 1 H R e − ( 1 − τ ) u 0 Z ⋅ 1 H R e − τ u 1 Z 1_{H_R}e^{-(1-\tau)u_0Z}\cdot1_{H_R}e^{-\tau u_1Z} 1 H R e − ( 1 − τ ) u 0 Z ⋅ 1 H R e − τ u 1 Z , we find
Λ m , R ( ( 1 − τ ) u 0 + τ u 1 ) ≤ ( 1 − τ ) Λ m , R ( u 0 ) + τ Λ m , R ( u 1 ) . \Lambda_{m,R}\bigl((1-\tau)u_0+\tau u_1\bigr)\leq(1-\tau)\Lambda_{m,R}(u_0)+\tau\Lambda_{m,R}(u_1). Λ m , R ( ( 1 − τ ) u 0 + τ u 1 ) ≤ ( 1 − τ ) Λ m , R ( u 0 ) + τ Λ m , R ( u 1 ) .
The next lemma replaces the free energy by a block free energy, with a small loss in the quantities of Lemma 6.16 .
Lemma 9.11. Suppose that φ ( 1 ) > 0 \varphi(1)>0 φ ( 1 ) > 0 , and let s ∈ ( 1 / 2 , 1 ) s\in(1/2,1) s ∈ ( 1/2 , 1 ) , γ = − φ ′ ( s ) > 0 \gamma=-\varphi'(s)>0 γ = − φ ′ ( s ) > 0 , and F = φ ( s ) − ( 1 − s ) γ > 0 F=\varphi(s)-(1-s)\gamma>0 F = φ ( s ) − ( 1 − s ) γ > 0 be as in Lemma 6.16 . There exist m ≥ 1 m\geq1 m ≥ 1 and R > max { E , E − 1 } R>\max\{E,E^{-1}\} R > max { E , E − 1 } with the following properties. Write φ ^ = φ m , R \widehat{\varphi}=\varphi_{m,R} φ = φ m , R , and set
γ ^ = − φ ^ ′ ( s ) , F ^ = φ ^ ( s ) − ( 1 − s ) γ ^ , ι = log p − φ ^ ( s ) − s γ ^ . \widehat{\gamma}=-\widehat{\varphi}'(s),\qquad\widehat{F}=\widehat{\varphi}(s)-(1-s)\widehat{\gamma},\qquad\iota=\log p-\widehat{\varphi}(s)-s\widehat{\gamma}. γ = − φ ′ ( s ) , F = φ ( s ) − ( 1 − s ) γ , ι = log p − φ ( s ) − s γ . (i) φ ^ ( s ) ≤ φ ( s ) \widehat{\varphi}(s)\leq\varphi(s) φ ( s ) ≤ φ ( s ) .
(ii) γ ^ ≥ γ / 2 \widehat{\gamma}\geq\gamma/2 γ ≥ γ /2 , F ^ ≥ F / 2 \widehat{F}\geq F/2 F ≥ F /2 , and ( 2 s − 1 ) γ ^ − 2 ( φ ( s ) − φ ^ ( s ) ) ≥ ( 2 s − 1 ) γ / 2 (2s-1)\widehat{\gamma}-2(\varphi(s)-\widehat{\varphi}(s))\geq(2s-1)\gamma/2 ( 2 s − 1 ) γ − 2 ( φ ( s ) − φ ( s )) ≥ ( 2 s − 1 ) γ /2 .
(iii) ι ≥ 0 \iota\geq0 ι ≥ 0 and log p − ι = φ ^ ( s ) + s γ ^ > 0 \log p-\iota=\widehat{\varphi}(s)+s\widehat{\gamma}>0 log p − ι = φ ( s ) + s γ > 0 .
(iv) P ( H R ∖ H R ∘ ) = 0 \mathbb{P}(H_R\setminus H_R^\circ)=0 P ( H R ∖ H R ∘ ) = 0 .
Proof. The proof has three steps. We first let m → ∞ m\to\infty m → ∞ without truncation, then let R → ∞ R\to\infty R → ∞ at fixed m m m , and finally verify the inequalities.
Step 1. Long blocks. Let Z Z Z be the growth of a block of m m m independent factors, and set φ ( m ) ( u ) = log p + m − 1 log E e − u Z = log p + m − 1 log f m ( u ) \varphi^{(m)}(u)=\log p+m^{-1}\log\mathbb{E}e^{-uZ}=\log p+m^{-1}\log f_m(u) φ ( m ) ( u ) = log p + m − 1 log E e − u Z = log p + m − 1 log f m ( u ) for u ∈ ( 0 , 1 ) u\in(0,1) u ∈ ( 0 , 1 ) . By Corollary 6.22, m − 1 log f m ( u ) → φ ( u ) − log p m^{-1}\log f_m(u)\to\varphi(u)-\log p m − 1 log f m ( u ) → φ ( u ) − log p as m → ∞ m\to\infty m → ∞ , that is, φ ( m ) ( u ) → φ ( u ) \varphi^{(m)}(u)\to\varphi(u) φ ( m ) ( u ) → φ ( u ) for every u ∈ ( 0 , 1 ) u\in(0,1) u ∈ ( 0 , 1 ) ; by (153) , φ ( m ) ( u ) ≤ φ ( u ) \varphi^{(m)}(u)\leq\varphi(u) φ ( m ) ( u ) ≤ φ ( u ) . By Hölder’s inequality, as in the discussion after (154) , φ ( m ) \varphi^{(m)} φ ( m ) is convex on ( 0 , 1 ) (0,1) ( 0 , 1 ) . It is differentiable, and the derivative of u ↦ E e − u Z u \mapsto\mathbb{E}e^{-uZ} u ↦ E e − u Z is − E [ Z e − u Z ] -\mathbb{E}[Ze^{-uZ}] − E [ Z e − u Z ] . Indeed, for u , u + h ∈ ( 0 , 1 ) u,u+h \in(0,1) u , u + h ∈ ( 0 , 1 ) , the mean value theorem gives ∣ e − ( u + h ) Z − e − u Z ∣ ≤ ∣ h ∣ Z |e^{-(u+h)Z}-e^{-uZ}| \le|h|Z ∣ e − ( u + h ) Z − e − u Z ∣ ≤ ∣ h ∣ Z , since Z ≥ 0 Z \ge0 Z ≥ 0 , and we may differentiate under the expectation by dominated convergence with the dominating function Z Z Z . This function is integrable, since Z ≤ ∑ i = 1 m log ∥ M ( d i , t i ) ∥ Z \le\sum_{i=1}^{m} \log\lVert M(d_i,t_i)\rVert Z ≤ ∑ i = 1 m log ∥ M ( d i , t i )∥ by submultiplicativity, and E log ∥ M ( d 1 , t 1 ) ∥ < ∞ \mathbb{E}\log\lVert M(d_1,t_1)\rVert< \infty E log ∥ M ( d 1 , t 1 )∥ < ∞ by Proposition 6.15(iii). For 0 < h < min { s , 1 − s } 0<h<\min\{s,1-s\} 0 < h < min { s , 1 − s } , convexity gives
φ ( m ) ( s ) − φ ( m ) ( s − h ) h ≤ ( φ ( m ) ) ′ ( s ) ≤ φ ( m ) ( s + h ) − φ ( m ) ( s ) h . \frac{\varphi^{(m)}(s)-\varphi^{(m)}(s-h)}{h}
\leq(\varphi^{(m)})'(s)
\leq\frac{\varphi^{(m)}(s+h)-\varphi^{(m)}(s)}{h}. h φ ( m ) ( s ) − φ ( m ) ( s − h ) ≤ ( φ ( m ) ) ′ ( s ) ≤ h φ ( m ) ( s + h ) − φ ( m ) ( s ) .
Letting m → ∞ m\to\infty m → ∞ , we find that lim inf m ( φ ( m ) ) ′ ( s ) \liminf_m(\varphi^{(m)})'(s) lim inf m ( φ ( m ) ) ′ ( s ) and lim sup m ( φ ( m ) ) ′ ( s ) \limsup_m(\varphi^{(m)})'(s) lim sup m ( φ ( m ) ) ′ ( s ) lie between ( φ ( s ) − φ ( s − h ) ) / h (\varphi(s)-\varphi(s-h))/h ( φ ( s ) − φ ( s − h )) / h and ( φ ( s + h ) − φ ( s ) ) / h (\varphi(s+h)-\varphi(s))/h ( φ ( s + h ) − φ ( s )) / h . Letting h ↓ 0 h\downarrow0 h ↓ 0 , and using that φ \varphi φ is differentiable at s s s (Proposition 6.15(i)), we conclude that ( φ ( m ) ) ′ ( s ) → φ ′ ( s ) (\varphi^{(m)})'(s)\to\varphi'(s) ( φ ( m ) ) ′ ( s ) → φ ′ ( s ) .
Step 2. Truncation. Fix m m m . Since d i ≠ 0 d_i\ne0 d i = 0 and t i ∈ ( 0 , ∞ ) t_i\in(0,\infty) t i ∈ ( 0 , ∞ ) almost surely, we have 1 H R ↑ 1 \mathbf{1}_{H_R}\uparrow1 1 H R ↑ 1 almost surely as R → ∞ R\to\infty R → ∞ . By monotone convergence, E [ 1 H R e − s Z ] ↑ E e − s Z \mathbb{E}[\mathbf{1}_{H_R}e^{-sZ}]\uparrow\mathbb{E}e^{-sZ} E [ 1 H R e − s Z ] ↑ E e − s Z and E [ 1 H R Z e − s Z ] ↑ E [ Z e − s Z ] \mathbb{E}[\mathbf{1}_{H_R}Ze^{-sZ}]\uparrow\mathbb{E}[Ze^{-sZ}] E [ 1 H R Z e − s Z ] ↑ E [ Z e − s Z ] , and both limits are finite and the first is positive. Since φ m , R ′ ( s ) = − m − 1 E [ 1 H R Z e − s Z ] / E [ 1 H R e − s Z ] \varphi'_{m,R}(s)=-m^{-1}\mathbb{E}[\mathbf{1}_{H_R}Ze^{-sZ}]/\mathbb{E}[\mathbf{1}_{H_R}e^{-sZ}] φ m , R ′ ( s ) = − m − 1 E [ 1 H R Z e − s Z ] / E [ 1 H R e − s Z ] , we obtain φ m , R ( s ) → φ ( m ) ( s ) \varphi_{m,R}(s)\to\varphi^{(m)}(s) φ m , R ( s ) → φ ( m ) ( s ) and φ m , R ′ ( s ) → ( φ ( m ) ) ′ ( s ) \varphi'_{m,R}(s)\to(\varphi^{(m)})'(s) φ m , R ′ ( s ) → ( φ ( m ) ) ′ ( s ) as R → ∞ R\to\infty R → ∞ . Further, φ m , R ( s ) ≤ φ ( m ) ( s ) ≤ φ ( s ) \varphi_{m,R}(s)\le\varphi^{(m)}(s)\le\varphi(s) φ m , R ( s ) ≤ φ ( m ) ( s ) ≤ φ ( s ) , because 1 H R ≤ 1 \mathbf{1}_{H_R}\le1 1 H R ≤ 1 .
Step 3. Choice of m m m and R R R . Let ε = min { γ / 2 , F / 4 , ( 2 s − 1 ) γ / 8 } \varepsilon=\min\{\gamma/2,F/4,(2s-1)\gamma/8\} ε = min { γ /2 , F /4 , ( 2 s − 1 ) γ /8 } . By Steps 1 and 2, we may choose m m m and then R > max { E , E − 1 } R>\max\{E,E^{-1}\} R > max { E , E − 1 } such that ∣ φ ^ ( s ) − φ ( s ) ∣ ≤ ε |\widehat{\varphi}(s)-\varphi(s)|\le\varepsilon ∣ φ ( s ) − φ ( s ) ∣ ≤ ε and ∣ φ ^ ′ ( s ) − φ ′ ( s ) ∣ ≤ ε |\widehat{\varphi}'(s)-\varphi'(s)|\le\varepsilon ∣ φ ′ ( s ) − φ ′ ( s ) ∣ ≤ ε . Then (i) holds by Step 2, and
γ ^ ≥ γ − ε ≥ γ 2 , F ^ ≥ φ ( s ) − ε − ( 1 − s ) ( γ + ε ) ≥ F − 2 ε ≥ F 2 , \widehat{\gamma}\ge\gamma-\varepsilon\ge\frac{\gamma}{2},\qquad
\widehat{F}\ge\varphi(s)-\varepsilon-(1-s)(\gamma+\varepsilon)\ge F-2\varepsilon\ge\frac{F}{2}, γ ≥ γ − ε ≥ 2 γ , F ≥ φ ( s ) − ε − ( 1 − s ) ( γ + ε ) ≥ F − 2 ε ≥ 2 F ,
( 2 s − 1 ) γ ^ − 2 ( φ ( s ) − φ ^ ( s ) ) ≥ ( 2 s − 1 ) ( γ − ε ) − 2 ε ≥ ( 2 s − 1 ) γ − 3 ε ≥ ( 2 s − 1 ) γ 2 , (2s-1)\widehat{\gamma}-2\bigl(\varphi(s)-\widehat{\varphi}(s)\bigr)
\ge(2s-1)(\gamma-\varepsilon)-2\varepsilon
\ge(2s-1)\gamma-3\varepsilon
\ge\frac{(2s-1)\gamma}{2}, ( 2 s − 1 ) γ − 2 ( φ ( s ) − φ ( s ) ) ≥ ( 2 s − 1 ) ( γ − ε ) − 2 ε ≥ ( 2 s − 1 ) γ − 3 ε ≥ 2 ( 2 s − 1 ) γ ,
where we used ( 2 s − 1 ) ε ≤ ε (2s-1)\varepsilon\le\varepsilon ( 2 s − 1 ) ε ≤ ε . This proves (ii). For (iii), let Λ = Λ m , R \Lambda=\Lambda_{m,R} Λ = Λ m , R for the chosen m m m and R R R , so that Λ ( u ) = m ( φ ^ ( u ) − log p ) \Lambda(u)=m(\widehat{\varphi}(u)-\log p) Λ ( u ) = m ( φ ( u ) − log p ) . It is convex and differentiable on R \mathbb{R} R , with Λ ′ ( s ) = − m γ ^ \Lambda'(s)=-m\widehat{\gamma} Λ ′ ( s ) = − m γ , and
m ι = m log p − m φ ^ ( s ) − s m γ ^ = − Λ ( s ) + s Λ ′ ( s ) . m\iota=m\log p-m\widehat{\varphi}(s)-sm\widehat{\gamma}=-\Lambda(s)+s\Lambda'(s). m ι = m log p − m φ ( s ) − s m γ = − Λ ( s ) + s Λ ′ ( s ) .
A convex differentiable function lies above its tangent lines; at the point 0 0 0 , the tangent line at s s s gives Λ ( 0 ) ≥ Λ ( s ) + Λ ′ ( s ) ( 0 − s ) \Lambda(0)\ge\Lambda(s)+\Lambda'(s)(0-s) Λ ( 0 ) ≥ Λ ( s ) + Λ ′ ( s ) ( 0 − s ) . Then m ι ≥ − Λ ( 0 ) = − log P ( H R ) ≥ 0 m\iota\ge-\Lambda(0)=-\log\mathbb{P}(H_R)\ge0 m ι ≥ − Λ ( 0 ) = − log P ( H R ) ≥ 0 . (Equivalently, m ι m\iota m ι is the relative entropy of the tilted block law μ ^ \widehat{\mu} μ of the proof of Lemma 9.12 with respect to the law of a block, and − log P ( H R ) -\log\mathbb{P}(H_R) − log P ( H R ) is a lower bound for it.) Further, log p − ι = φ ^ ( s ) + s γ ^ = F ^ + γ ^ > 0 \log p-\iota=\widehat{\varphi}(s)+s\widehat{\gamma}=\widehat{F}+\widehat{\gamma}>0 log p − ι = φ ( s ) + s γ = F + γ > 0 . For (iv), the only atom of π \pi π is at E E E (Proposition 5.5(i)), and R , R − 1 ≠ E R,R^{-1}\ne E R , R − 1 = E ; thus P ( ∣ d i ∣ ∈ { R , R − 1 } ) = 0 \mathbb{P}(|d_i|\in\{R,R^{-1}\})=0 P ( ∣ d i ∣ ∈ { R , R − 1 }) = 0 , and P ( t i ∈ { R , R − 1 } ) = 0 \mathbb{P}(t_i\in\{R,R^{-1}\})=0 P ( t i ∈ { R , R − 1 }) = 0 since t i t_i t i has a density.
From now on in this section, m m m , R R R , φ ^ \widehat{\varphi} φ , γ ^ \widehat{\gamma} γ , F ^ \widehat{F} F , and ι \iota ι are as in Lemma 9.11. The number ι \iota ι is the exponential cost per factor of the regular growth events of the next lemma, on which the partial sums of the block growths stay close to m γ ^ m\widehat{\gamma} m γ per block, their typical value under the exponentially tilted block law.
Lemma 9.12. Let ( d i , t i ) i ≥ 1 (d_i,t_i)_{i\geq1} ( d i , t i ) i ≥ 1 be independent factors with law ν f a c \nu_{\mathrm{fac}} ν fac , and for l ≥ 1 l\geq1 l ≥ 1 let Z l Z_l Z l , H l H_l H l , and H l ∘ H_l^\circ H l ∘ be the growth and the retention events of the block ( d i , t i ) ( l − 1 ) m < i ≤ l m (d_i,t_i)_{(l-1)m<i\leq lm} ( d i , t i ) ( l − 1 ) m < i ≤ l m . Let S k = Z 1 + ⋯ + Z k S_k=Z_1+\cdots+Z_k S k = Z 1 + ⋯ + Z k , let k ≥ 1 k\geq1 k ≥ 1 , and let b > 0 b>0 b > 0 .
(i) We have
P ( H 1 ∩ ⋯ ∩ H k ∩ { S k ≤ k m γ ^ + b } ) ≤ e − ı m k + s b . \mathbb{P}\left(H_1\cap\cdots\cap H_k\cap\{S_k\leq km\widehat{\gamma}+b\}\right)\leq e^{-\imath mk+sb}. P ( H 1 ∩ ⋯ ∩ H k ∩ { S k ≤ k m γ + b } ) ≤ e − mk + s b . (ii) We have
P ( H 1 ∘ ∩ ⋯ ∩ H k ∘ ∩ { ∣ S i − i m γ ^ ∣ < b for 1 ≤ i ≤ k } ) ≥ e − ı m k − s b ( 1 − k ( m log C R ) 2 b 2 ) . \mathbb{P}\left(H_1^\circ\cap\cdots\cap H_k^\circ\cap\{|S_i-im\widehat{\gamma}|<b\text{ for }1\leq i\leq k\}\right)\geq e^{-\imath mk-sb}\left(1-\frac{k(m\log C_R)^2}{b^2}\right). P ( H 1 ∘ ∩ ⋯ ∩ H k ∘ ∩ { ∣ S i − im γ ∣ < b for 1 ≤ i ≤ k } ) ≥ e − mk − s b ( 1 − b 2 k ( m log C R ) 2 ) .
Proof. Let μ \mu μ be the law of a block, that is, the law of ( d i , t i ) i = 1 m (d_i,t_i)_{i=1}^m ( d i , t i ) i = 1 m on ( C × ( 0 , ∞ ) ) m (\mathbb{C}\times(0,\infty))^m ( C × ( 0 , ∞ ) ) m , let Λ = Λ m , R \Lambda=\Lambda_{m,R} Λ = Λ m , R be as in the proof of Lemma 9.11, and let μ ^ \widehat{\mu} μ be the tilted law μ ^ ( d b ) = e − Λ ( s ) 1 H R ( b ) e − s Z ( b ) μ ( d b ) \widehat{\mu}(\mathrm{d}\mathbf{b})=e^{-\Lambda(s)}\mathbf{1}_{H_R}(\mathbf{b})e^{-sZ(\mathbf{b})}\mu(\mathrm{d}\mathbf{b}) μ ( d b ) = e − Λ ( s ) 1 H R ( b ) e − s Z ( b ) μ ( d b ) , where Z ( b ) Z(\mathbf{b}) Z ( b ) is the growth of the block b \mathbf{b} b . It is a probability measure by the definition of Λ ( s ) \Lambda(s) Λ ( s ) . Let P ^ \widehat{\mathbb{P}} P be a probability measure under which the blocks b 1 , b 2 , … \mathbf{b}_1,\mathbf{b}_2,\ldots b 1 , b 2 , … are independent with law μ ^ \widehat{\mu} μ , and write E ^ \widehat{\mathbb{E}} E for its expectation. For every measurable Φ ≥ 0 \Phi\geq0 Φ ≥ 0 of the first k k k blocks, we have
E [ Φ ∏ l = 1 k 1 H l ] = e k Λ ( s ) E ^ [ Φ e s S k ] . (155) \mathbb{E}\left[\Phi\prod_{l=1}^{k}\mathbf{1}_{H_l}\right]=e^{k\Lambda(s)}\widehat{\mathbb{E}}\left[\Phi e^{sS_k}\right]. \tag*{(155)} E [ Φ l = 1 ∏ k 1 H l ] = e k Λ ( s ) E [ Φ e s S k ] . (155)
Indeed, since the blocks are independent with law μ \mu μ under P \mathbb{P} P , the right side equals
e k Λ ( s ) ∫ Φ ( b 1 , … , b k ) e s ∑ l Z ( b l ) ∏ l = 1 k e − Λ ( s ) 1 H R ( b l ) e − s Z ( b l ) μ ( d b l ) = ∫ Φ ( b 1 , … , b k ) ∏ l = 1 k 1 H R ( b l ) μ ( d b l ) , \begin{aligned}
e^{k\Lambda(s)}\int\Phi(\mathbf{b}_1,\ldots,\mathbf{b}_k)e^{s\sum_l Z(\mathbf{b}_l)}\prod_{l=1}^{k}e^{-\Lambda(s)}\mathbf{1}_{H_R}(\mathbf{b}_l)e^{-sZ(\mathbf{b}_l)}\mu(\mathrm{d}\mathbf{b}_l) \\
&=\int\Phi(\mathbf{b}_1,\ldots,\mathbf{b}_k)\prod_{l=1}^{k}\mathbf{1}_{H_R}(\mathbf{b}_l)\mu(\mathrm{d}\mathbf{b}_l),
\end{aligned} e k Λ ( s ) ∫ Φ ( b 1 , … , b k ) e s ∑ l Z ( b l ) l = 1 ∏ k e − Λ ( s ) 1 H R ( b l ) e − s Z ( b l ) μ ( d b l ) = ∫ Φ ( b 1 , … , b k ) l = 1 ∏ k 1 H R ( b l ) μ ( d b l ) ,
which is the left side. Note that k Λ ( s ) + s k m γ ^ = k m ( φ ^ ( s ) − log p + s γ ^ ) = − ı m k k\Lambda(s)+skm\widehat{\gamma}=km(\widehat{\varphi}(s)-\log p+s\widehat{\gamma})=-\imath mk k Λ ( s ) + s k m γ = k m ( φ ( s ) − log p + s γ ) = − mk . Applying (9.14) with Φ = 1 { S k ≤ k m γ ^ + b } \Phi=\mathbf{1}_{\{S_k\leq km\widehat{\gamma}+b\}} Φ = 1 { S k ≤ k m γ + b } and bounding e s S k ≤ e s ( k m γ ^ + b ) e^{sS_k}\leq e^{s(km\widehat{\gamma}+b)} e s S k ≤ e s ( k m γ + b ) on this event, we conclude (i).
For (ii), let A A A be the event in (ii). Since μ ^ ≪ μ \widehat{\mu}\ll\mu μ ≪ μ and μ ( H R ∖ H R ∘ ) = 0 \mu(H_R\setminus H_R^\circ)=0 μ ( H R ∖ H R ∘ ) = 0 by Lemma 9.11(iv), under E ^ \widehat{\mathbb{E}} E every block lies in H R ∘ H_R^\circ H R ∘ almost surely. Applying (9.14) with Φ = 1 A \Phi=\mathbf{1}_A Φ = 1 A , and using e s S k ≥ e s ( k m γ ^ − b ) e^{sS_k}\geq e^{s(km\widehat{\gamma}-b)} e s S k ≥ e s ( k m γ − b ) on A A A , we find
P ( A ) ≥ e − ı m k − s b P ^ ( ∣ S i − i m γ ^ ∣ < b for 1 ≤ i ≤ k ) . \mathbb{P}(A)\geq e^{-\imath mk-sb}\widehat{\mathbb{P}}\left(|S_i-im\widehat{\gamma}|<b\text{ for }1\leq i\leq k\right). P ( A ) ≥ e − mk − s b P ( ∣ S i − im γ ∣ < b for 1 ≤ i ≤ k ) .
Under P ^ \widehat{\mathbb{P}} P , the variables Z l Z_l Z l are independent and identically distributed, and they take values in [ 0 , m log C R ] [0,m\log C_R] [ 0 , m log C R ] by (9.13). Their mean is
E ^ Z 1 = e − Λ ( s ) E [ 1 H R Z e − s Z ] = − Λ ′ ( s ) = m γ ^ , \widehat{\mathbb{E}}Z_1=e^{-\Lambda(s)}\mathbb{E}\left[\mathbf{1}_{H_R}Ze^{-sZ}\right]=-\Lambda'(s)=m\widehat{\gamma}, E Z 1 = e − Λ ( s ) E [ 1 H R Z e − s Z ] = − Λ ′ ( s ) = m γ ,
where we used the formula for the derivative of Λ \Lambda Λ from the discussion after (154) . In particular, their variance satisfies Var ^ ( Z 1 ) ≤ E ^ Z 1 2 ≤ ( m log C R ) 2 \widehat{\operatorname{Var}}(Z_1) \le\widehat{\mathbb{E}}Z_1^2 \le(m\log C_R)^2 Var ( Z 1 ) ≤ E Z 1 2 ≤ ( m log C R ) 2 . It remains to show that
P ^ ( max 1 ≤ i ≤ k ∣ S i ′ ∣ ≥ b ) ≤ k Var ^ ( Z 1 ) b 2 , S i ′ = S i − i m γ ^ , (156) \widehat{\mathbb{P}}\left(\max_{1\le i\le k}|S_i'|\ge b\right)\le\frac{k\widehat{\operatorname{Var}}(Z_1)}{b^2},\qquad S_i'=S_i-im\widehat{\gamma}, \tag*{(156)} P ( 1 ≤ i ≤ k max ∣ S i ′ ∣ ≥ b ) ≤ b 2 k Var ( Z 1 ) , S i ′ = S i − im γ , (156)
which is Kolmogorov’s maximal inequality (see [64, Section 2.5]); since the right side is at most k ( m log C R ) 2 / b 2 k(m\log C_R)^2/b^2 k ( m log C R ) 2 / b 2 , this proves (ii). For completeness, we include the proof. Let σ \sigma σ be the smallest i ∈ { 1 , … , k } i\in\{1,\ldots,k\} i ∈ { 1 , … , k } with ∣ S i ′ ∣ ≥ b |S_i'|\ge b ∣ S i ′ ∣ ≥ b , and σ = ∞ \sigma=\infty σ = ∞ if there is none. For 1 ≤ i ≤ k 1\le i\le k 1 ≤ i ≤ k , the event { σ = i } \{\sigma=i\} { σ = i } and the variable S i ′ S_i' S i ′ are functions of the first i i i blocks, while S k ′ − S i ′ = ∑ l = i + 1 k ( Z l − m γ ^ ) S_k'-S_i'=\sum_{l=i+1}^{k}(Z_l-m\widehat{\gamma}) S k ′ − S i ′ = ∑ l = i + 1 k ( Z l − m γ ) is a function of the blocks i + 1 , … , k i+1,\ldots,k i + 1 , … , k , which are independent of the first i i i blocks; further, E ^ [ S k ′ − S i ′ ] = 0 \widehat{\mathbb{E}}[S_k'-S_i']=0 E [ S k ′ − S i ′ ] = 0 . All these variables are bounded. Expanding the square, we find
E ^ [ ( S k ′ ) 2 1 { σ = i } ] = E ^ [ ( S i ′ ) 2 1 { σ = i } ] + 2 E ^ [ S i ′ 1 { σ = i } ] E ^ [ S k ′ − S i ′ ] + E ^ [ ( S k ′ − S i ′ ) 2 1 { σ = i } ] ≥ b 2 P ^ ( σ = i ) , \widehat{\mathbb{E}}\left[(S_k')^2\mathbf{1}_{\{\sigma=i\}}\right]=\widehat{\mathbb{E}}\left[(S_i')^2\mathbf{1}_{\{\sigma=i\}}\right]+2\widehat{\mathbb{E}}\left[S_i'\mathbf{1}_{\{\sigma=i\}}\right]\widehat{\mathbb{E}}\left[S_k'-S_i'\right]+\widehat{\mathbb{E}}\left[(S_k'-S_i')^2\mathbf{1}_{\{\sigma=i\}}\right]\ge b^2\widehat{\mathbb{P}}(\sigma=i), E [ ( S k ′ ) 2 1 { σ = i } ] = E [ ( S i ′ ) 2 1 { σ = i } ] + 2 E [ S i ′ 1 { σ = i } ] E [ S k ′ − S i ′ ] + E [ ( S k ′ − S i ′ ) 2 1 { σ = i } ] ≥ b 2 P ( σ = i ) ,
where we used the independence in the middle term, which therefore vanishes, that the third term is nonnegative, and that ∣ S i ′ ∣ ≥ b |S_i'|\ge b ∣ S i ′ ∣ ≥ b on { σ = i } \{\sigma=i\} { σ = i } in the first term. Summing over 1 ≤ i ≤ k 1\le i\le k 1 ≤ i ≤ k , and using that the events { σ = i } \{\sigma=i\} { σ = i } are disjoint with union { max i ≤ k ∣ S i ′ ∣ ≥ b } \{\max_{i\le k}|S_i'|\ge b\} { max i ≤ k ∣ S i ′ ∣ ≥ b } , we obtain
b 2 P ^ ( max 1 ≤ i ≤ k ∣ S i ′ ∣ ≥ b ) ≤ E ^ [ ( S k ′ ) 2 ] = ∑ l = 1 k Var ^ ( Z l ) = k Var ^ ( Z 1 ) , b^2\widehat{\mathbb{P}}\left(\max_{1\le i\le k}|S_i'|\ge b\right)\le\widehat{\mathbb{E}}[(S_k')^2]=\sum_{l=1}^{k}\widehat{\operatorname{Var}}(Z_l)=k\widehat{\operatorname{Var}}(Z_1), b 2 P ( 1 ≤ i ≤ k max ∣ S i ′ ∣ ≥ b ) ≤ E [( S k ′ ) 2 ] = l = 1 ∑ k Var ( Z l ) = k Var ( Z 1 ) ,
where in the first equality we used that the centered variables Z l − m γ ^ Z_l-m\widehat{\gamma} Z l − m γ are independent. This proves (156) .
The upper bound in (i) requires only an upper bound on the final partial sum. In Section 10, it is applied to the blocks of the common part of two paths and to the blocks of each of their two arms separately.
Averaging over fresh factors
In the second moment, the interaction between two resonant vertices is measured by the Green function between their parents, in the tree from which both vertices are deleted. By (145) , its inverse is a bilinear form of the transfer product along the path between the parents. This path runs from the first parent up to the last common ancestor, which we call the fork, and down to the second parent. Its transfer product consists of the factors of the two arms and the factor at the fork. In this section, we bound negative moments of such bilinear forms by the norms of the products along the two arms. The factor at the fork depends on the environment of the common part of the two paths. This environment is constrained by the regular growth condition, so we cannot average over it, and we treat the factor at the fork as an arbitrary matrix of determinant one. To separate the two arms across this factor, and to remove the two endpoint vectors, we average over four groups of three independent factors, which we call averaging triples . Two of them are next to the endpoints of the path, and two are next to the fork. The remaining factors are arbitrary, provided they are independent of the averaging triples. Lemmas 9.13 and 9.14 are applied in the proof of Lemma 10.11. Recall the constant C ∗ C_\ast C ∗ of Proposition 6.18 .
Lemma 9.13. Let s ∈ ( 0 , 1 ) s \in(0,1) s ∈ ( 0 , 1 ) , K > 0 K > 0 K > 0 , and a + , a − ≥ 0 a_+,a_- \ge0 a + , a − ≥ 0 , and let C ∗ C_\ast C ∗ be the constant of Proposition 6.18 . Let G \mathcal{G} G be a σ \sigma σ -algebra, and let G 1 , G 2 , G 3 , G 4 G_1,G_2,G_3,G_4 G 1 , G 2 , G 3 , G 4 be random matrices such that G , G 1 , G 2 , G 3 , G 4 \mathcal{G},G_1,G_2,G_3,G_4 G , G 1 , G 2 , G 3 , G 4 are independent, and each G i G_i G i is a product of three factors as in Proposition 6.19 , with denominators satisfying the moment bounds with a + a_+ a + and a − a_- a − , and with parameters β \beta β of the exponential law of rate K K K . Let A 1 , A 2 , A 3 A_1,A_2,A_3 A 1 , A 2 , A 3 be G \mathcal{G} G -measurable random complex 2 × 2 2 \times2 2 × 2 matrices with determinant one, and let w w w be a G \mathcal{G} G -measurable random vector in C 2 ∖ { 0 } \mathbb{C}^2 \setminus\{0\} C 2 ∖ { 0 } . Then, almost surely,
E [ ∣ e 2 T G 1 A 1 G 2 A 2 G 3 A 3 G 4 w ∣ − s | G ] ≤ C ∗ 4 ∥ A 1 ∥ − s ∥ A 3 ∥ − s ∥ w ∥ − s . \mathbb{E}\left[\left|e_2^{\mathsf{T}}G_1A_1G_2A_2G_3A_3G_4w\right|^{-s}\,\middle|\,\mathcal{G}\right] \le C_\ast^4\left\|A_1\right\|^{-s}\left\|A_3\right\|^{-s}\left\|w\right\|^{-s}. E [ e 2 T G 1 A 1 G 2 A 2 G 3 A 3 G 4 w − s G ] ≤ C ∗ 4 ∥ A 1 ∥ − s ∥ A 3 ∥ − s ∥ w ∥ − s . Proof. The proof averages first over the averaging triples G 1 G_1 G 1 and G 4 G_4 G 4 next to the endpoints, with everything else fixed, and then over the averaging triples G 2 G_2 G 2 and G 3 G_3 G 3 next to the fork.
Let G ′ \mathcal{G}' G ′ be the σ \sigma σ -algebra generated by G \mathcal{G} G , G 2 G_2 G 2 , and G 3 G_3 G 3 . Since G , G 1 , … , G 4 \mathcal{G},G_1,\ldots,G_4 G , G 1 , … , G 4 are independent, the σ \sigma σ -algebra G ′ \mathcal{G}' G ′ and the matrices G 1 G_1 G 1 and G 4 G_4 G 4 are independent (by grouping independent families); in particular, G ′ \mathcal{G}' G ′ is independent of ( G 1 , G 4 ) (G_1,G_4) ( G 1 , G 4 ) , and G 1 G_1 G 1 and G 4 G_4 G 4 are independent. The matrix A = A 1 G 2 A 2 G 3 A 3 A=A_1G_2A_2G_3A_3 A = A 1 G 2 A 2 G 3 A 3 is G ′ \mathcal{G}' G ′ -measurable and has determinant one, so it is nonzero, and the vector w w w is G ′ \mathcal{G}' G ′ -measurable and nonzero. By Corollary 6.21 (i) and (iii), applied with the σ \sigma σ -algebra G ′ \mathcal{G}' G ′ , with u = e 2 u=e_2 u = e 2 , v = w v=w v = w , and with the averaging triples G = G 1 G=G_1 G = G 1 and G ′ = G 4 G'=G_4 G ′ = G 4 , we have
E [ ∣ e 2 T G 1 A G 4 w ∣ − s | G ′ ] ≤ C ∗ 2 ∥ A ∥ − s ∥ w ∥ − s . \mathbb{E}\left[\left|e_2^{\mathsf{T}}G_1AG_4w\right|^{-s}\,\middle|\,\mathcal{G}'\right] \le C_\ast^2\|A\|^{-s}\|w\|^{-s}. E [ e 2 T G 1 A G 4 w − s G ′ ] ≤ C ∗ 2 ∥ A ∥ − s ∥ w ∥ − s . Here we used ∥ e 2 ∥ = 1 \|e_2\|=1 ∥ e 2 ∥ = 1 . Next, G \mathcal{G} G is independent of ( G 2 , G 3 ) (G_2,G_3) ( G 2 , G 3 ) , and G 2 G_2 G 2 and G 3 G_3 G 3 are independent. By Corollary 6.21 (ii) and (iii), applied with the σ \sigma σ -algebra G \mathcal{G} G , with C = A 1 C=A_1 C = A 1 , G = G 2 G=G_2 G = G 2 , F = A 2 F=A_2 F = A 2 , G ′ = G 3 G'=G_3 G ′ = G 3 , and A = A 3 A=A_3 A = A 3 , and since ∥ A 2 ∥ ≥ 1 \|A_2\|\ge1 ∥ A 2 ∥ ≥ 1 by (9.1), we have
E [ ∥ A 1 G 2 A 2 G 3 A 3 ∥ − s | G ] ≤ C ∗ 2 ∥ A 1 ∥ − s ∥ A 3 ∥ − s . \mathbb{E}\left[\left\|A_1G_2A_2G_3A_3\right\|^{-s}\,\middle|\,\mathcal{G}\right] \le C_\ast^2\|A_1\|^{-s}\|A_3\|^{-s}. E [ ∥ A 1 G 2 A 2 G 3 A 3 ∥ − s G ] ≤ C ∗ 2 ∥ A 1 ∥ − s ∥ A 3 ∥ − s . Since G ⊂ G ′ \mathcal{G}\subset\mathcal{G}' G ⊂ G ′ , the tower property gives
E [ ∣ e 2 T G 1 A G 4 w ∣ − s | G ] = E [ E [ ∣ e 2 T G 1 A G 4 w ∣ − s | G ′ ] | G ] ≤ C ∗ 2 ∥ w ∥ − s E [ ∥ A ∥ − s | G ] , \mathbb{E}\left[\left|e_2^{\mathsf{T}}G_1AG_4w\right|^{-s}\,\middle|\,\mathcal{G}\right]
=
\mathbb{E}\left[\mathbb{E}\left[\left|e_2^{\mathsf{T}}G_1AG_4w\right|^{-s}\,\middle|\,\mathcal{G}'\right]\,\middle|\,\mathcal{G}\right]
\le C_\ast^2\|w\|^{-s}\mathbb{E}\left[\|A\|^{-s}\,\middle|\,\mathcal{G}\right], E [ e 2 T G 1 A G 4 w − s G ] = E [ E [ e 2 T G 1 A G 4 w − s G ′ ] G ] ≤ C ∗ 2 ∥ w ∥ − s E [ ∥ A ∥ − s G ] , where we used that w w w is G \mathcal{G} G -measurable. Inserting the second bound, we obtain the claim. □
The factors A 1 A_1 A 1 and A 3 A_3 A 3 will be products of factors along the two arms of a pair of paths. Some of these factors have a deterministic denominator, and the next lemma removes them at a fixed cost.
Lemma 9.14. Let u ∈ ( 0 , 1 ) u \in(0,1) u ∈ ( 0 , 1 ) , r , k ≥ 0 r,k \ge0 r , k ≥ 0 , and let c 1 , … , c k ∈ C c_1,\ldots,c_k \in\mathbb{C} c 1 , … , c k ∈ C with 0 < ∣ c i ∣ ≤ E + 1 0 < |c_i| \le E+1 0 < ∣ c i ∣ ≤ E + 1 . Let V V V be a product of r r r independent factors of law ν f a c \nu_{\mathrm{fac}} ν fac (with V = I V=I V = I if r = 0 r=0 r = 0 ), and let H = M ( c 1 , t 1 ′ ) ⋯ M ( c k , t k ′ ) H=M(c_1,t'_1)\cdots M(c_k,t'_k) H = M ( c 1 , t 1 ′ ) ⋯ M ( c k , t k ′ ) , where t 1 ′ , … , t k ′ t'_1,\ldots,t'_k t 1 ′ , … , t k ′ are independent with density p e − p t p e^{-pt} p e − pt and independent of V V V . Set C a r m = E [ ( t 1 ′ ) − u / 2 ( 1 + ( t 1 ′ ) 2 + ( E + 1 ) 2 ) u / 2 ] C_{\mathrm{arm}}=\mathbb{E}\left[(t'_1)^{-u/2}\left(1+(t'_1)^2+(E+1)^2\right)^{u/2}\right] C arm = E [ ( t 1 ′ ) − u /2 ( 1 + ( t 1 ′ ) 2 + ( E + 1 ) 2 ) u /2 ] , which is finite and at least one. Then
E ∥ H V ∥ − u ≤ C a r m k e r ( φ ( u ) − log p ) , E ∥ V H ∥ − u ≤ C a r m k e r ( φ ( u ) − log p ) . \mathbb{E}\|HV\|^{-u}\le C_{\mathrm{arm}}^k e^{r(\varphi(u)-\log p)},\qquad\mathbb{E}\|VH\|^{-u}\le C_{\mathrm{arm}}^k e^{r(\varphi(u)-\log p)}. E ∥ H V ∥ − u ≤ C arm k e r ( φ ( u ) − l o g p ) , E ∥ V H ∥ − u ≤ C arm k e r ( φ ( u ) − l o g p ) .
Proof. By (9.1), ∥ V ∥ = ∥ H − 1 H V ∥ ≤ ∥ H ∥ ∥ H V ∥ \|V\|=\|H^{-1}HV\|\le\|H\|\|HV\| ∥ V ∥ = ∥ H − 1 H V ∥ ≤ ∥ H ∥∥ H V ∥ , which implies that ∥ H V ∥ − u ≤ ∥ H ∥ u ∥ V ∥ − u \|HV\|^{-u}\le\|H\|^u\|V\|^{-u} ∥ H V ∥ − u ≤ ∥ H ∥ u ∥ V ∥ − u . By submultiplicativity and (9.2), ∥ H ∥ u ≤ ∏ i = 1 k ( t i ′ ) − u / 2 ( 1 + ( t i ′ ) 2 + ( E + 1 ) 2 ) u / 2 \|H\|^u\le\prod_{i=1}^k (t'_i)^{-u/2}\left(1+(t'_i)^2+(E+1)^2\right)^{u/2} ∥ H ∥ u ≤ ∏ i = 1 k ( t i ′ ) − u /2 ( 1 + ( t i ′ ) 2 + ( E + 1 ) 2 ) u /2 . Taking expectations, and using that t 1 ′ , … , t k ′ t'_1,\ldots,t'_k t 1 ′ , … , t k ′ and V V V are independent together with (9.12), we obtain the first bound. The constant C a r m C_{\mathrm{arm}} C arm is finite, since ( 1 + ( t 1 ′ ) 2 + ( E + 1 ) 2 ) u / 2 ≤ 1 + ( t 1 ′ ) u + ( E + 1 ) u \left(1+(t'_1)^2+(E+1)^2\right)^{u/2}\le1+(t'_1)^u+(E+1)^u ( 1 + ( t 1 ′ ) 2 + ( E + 1 ) 2 ) u /2 ≤ 1 + ( t 1 ′ ) u + ( E + 1 ) u for u / 2 ≤ 1 u/2\le1 u /2 ≤ 1 and since E ( t 1 ′ ) − u / 2 \mathbb{E}(t'_1)^{-u/2} E ( t 1 ′ ) − u /2 and E ( t 1 ′ ) u / 2 \mathbb{E}(t'_1)^{u/2} E ( t 1 ′ ) u /2 are finite; it is at least one, since ( t 1 ′ ) − 1 ( 1 + ( t 1 ′ ) 2 ) ≥ 2 (t'_1)^{-1}\left(1+(t'_1)^2\right)\ge2 ( t 1 ′ ) − 1 ( 1 + ( t 1 ′ ) 2 ) ≥ 2 . The second bound follows in the same way from ∥ V ∥ ≤ ∥ V H ∥ ∥ H − 1 ∥ \|V\|\le\|VH\|\|H^{-1}\| ∥ V ∥ ≤ ∥ V H ∥∥ H − 1 ∥ .
Passage to the real axis
In Section 10, the events whose probabilities we bound are defined at E + i η E+i\eta E + i η with a small η > 0 \eta>0 η > 0 , while the probabilities are computed for the real recursive tree process at η = 0 \eta=0 η = 0 . The passage between the two uses only almost sure convergence, through the following form of the portmanteau theorem. Part (i) gives lower bounds for events defined by strict inequalities, and it is used in Lemmas 10.4 and 10.5 and in the proof of Lemma 10.9. Part (ii) gives upper bounds for events defined by non-strict inequalities, weighted by a bounded factor, and it is used in Lemma 10.4 and in the proof of Lemma 10.12. Neither part requires the limiting law to give zero mass to the boundary of the event, and neither requires uniform integrability.
Lemma 9.15. Let ( Ω ′ , A , Q ) (\Omega',\mathcal{A},\mathbb{Q}) ( Ω ′ , A , Q ) be a probability space, let k ≥ 1 k\ge1 k ≥ 1 , and for 0 ≤ η ≤ 1 0\le\eta\le1 0 ≤ η ≤ 1 let V η V_\eta V η be random vectors in C k \mathbb{C}^k C k such that V η → V 0 V_\eta\to V_0 V η → V 0 almost surely as η ↓ 0 \eta\downarrow0 η ↓ 0 .
(i) For every open set O ⊂ C k O\subset\mathbb{C}^k O ⊂ C k , we have lim inf η ↓ 0 Q ( V η ∈ O ) ≥ Q ( V 0 ∈ O ) \liminf_{\eta\downarrow0}\mathbb{Q}(V_\eta\in O)\ge\mathbb{Q}(V_0\in O) lim inf η ↓ 0 Q ( V η ∈ O ) ≥ Q ( V 0 ∈ O ) .
(ii) Let F ⊂ C k F\subset\mathbb{C}^k F ⊂ C k be closed, and let h η h_\eta h η , 0 ≤ η ≤ 1 0\le\eta\le1 0 ≤ η ≤ 1 , be random variables with 0 ≤ h η ≤ C 0\le h_\eta\le C 0 ≤ h η ≤ C for a constant C C C and h η → h 0 h_\eta\to h_0 h η → h 0 almost surely. Then
lim sup η ↓ 0 E Q [ 1 F ( V η ) h η ] ≤ E Q [ 1 F ( V 0 ) h 0 ] . \limsup_{\eta\downarrow0}\mathbb{E}_{\mathbb{Q}}\left[\mathbf{1}_F(V_\eta)h_\eta\right]\le\mathbb{E}_{\mathbb{Q}}\left[\mathbf{1}_F(V_0)h_0\right]. η ↓ 0 lim sup E Q [ 1 F ( V η ) h η ] ≤ E Q [ 1 F ( V 0 ) h 0 ] .
Proof. It suffices to consider a sequence η j ↓ 0 \eta_j\downarrow0 η j ↓ 0 . For (i), if V 0 ∈ O V_0\in O V 0 ∈ O and V η j → V 0 V_{\eta_j}\to V_0 V η j → V 0 , then V η j ∈ O V_{\eta_j}\in O V η j ∈ O for all large j j j , since O O O is open. This implies that lim inf j 1 O ( V η j ) ≥ 1 O ( V 0 ) \liminf_j\mathbf{1}_O(V_{\eta_j})\ge\mathbf{1}_O(V_0) lim inf j 1 O ( V η j ) ≥ 1 O ( V 0 ) almost surely, and (i) follows from Fatou’s lemma. For (ii), if V 0 ∉ F V_0\notin F V 0 ∈ / F , then V η j ∉ F V_{\eta_j}\notin F V η j ∈ / F for all large j j j , since the complement of F F F is open. Hence lim sup j 1 F ( V η j ) h η j ≤ 1 F ( V 0 ) h 0 \limsup_j\mathbf{1}_F(V_{\eta_j})h_{\eta_j}\le\mathbf{1}_F(V_0)h_0 lim sup j 1 F ( V η j ) h η j ≤ 1 F ( V 0 ) h 0 almost surely, and Fatou’s lemma applied to C − 1 F ( V η j ) h η j ≥ 0 C-\mathbf{1}_F(V_{\eta_j})h_{\eta_j}\ge0 C − 1 F ( V η j ) h η j ≥ 0 gives (ii).
We apply Lemma 9.15 under the spine laws of Definition 4.6, with V η V_\eta V η a finite collection of coordinates of the depth-m m m data L m ( E + i η ) \mathcal{L}_m(E+i\eta) L m ( E + i η ) of Definition 5.19, or of continuous functions of them. Under the no-ac hypothesis, Proposition 5.25(i) provides the required almost sure convergence. Events defined by strict inequalities between continuous functions of the coordinates are of the form { V η ∈ O } \{V_\eta\in O\} { V η ∈ O } with O O O open, and events defined by non-strict inequalities are of the form { V η ∈ F } \{V_\eta\in F\} { V η ∈ F } with F F F closed. The growth of a block is a continuous function of the factors only where the weights t i t_i t i are positive. In Section 10, it appears only together with the retention conditions of Definition 9.10: a strict inequality for partial sums of growths together with strict retention defines an open set, since the set of strictly retained blocks is open and the growths are continuous on it; and a non-strict inequality together with retention defines a closed set, since the set of retained blocks is closed and contained in the open set { t i > 0 for all i } \{t_i > 0 \text{ for all } i\} { t i > 0 for all i } , on which the growths are continuous. Further, if A η A_\eta A η , 0 ≤ η ≤ 1 0 \leq\eta\leq1 0 ≤ η ≤ 1 , are events with 1 A η → 1 A 0 \mathbf{1}_{A_\eta} \to\mathbf{1}_{A_0} 1 A η → 1 A 0 almost surely as η ↓ 0 \eta\downarrow0 η ↓ 0 (for instance, events that do not depend on η \eta η ), we may add 1 A η \mathbf{1}_{A_\eta} 1 A η as a coordinate of V η V_\eta V η . Since 1 A η ∈ { 0 , 1 } \mathbf{1}_{A_\eta} \in\{0,1\} 1 A η ∈ { 0 , 1 } , the event A η ∩ { V η ∈ O } A_\eta\cap\{V_\eta\in O\} A η ∩ { V η ∈ O } is then of the form { ( V η , 1 A η ) ∈ O × { w ∈ C : Re w > 1 / 2 } } \{(V_\eta,\mathbf{1}_{A_\eta}) \in O \times\{w \in\mathbb{C} : \operatorname{Re} w > 1/2\}\} {( V η , 1 A η ) ∈ O × { w ∈ C : Re w > 1/2 }} , with an open set, and A η ∩ { V η ∈ F } A_\eta\cap\{V_\eta\in F\} A η ∩ { V η ∈ F } is of the form { ( V η , 1 A η ) ∈ F × { w ∈ C : Re w ≥ 1 / 2 } } \{(V_\eta,\mathbf{1}_{A_\eta}) \in F \times\{w \in\mathbb{C} : \operatorname{Re} w \geq1/2\}\} {( V η , 1 A η ) ∈ F × { w ∈ C : Re w ≥ 1/2 }} , with a closed set. In Section 10, lower bounds are proved for strict inequalities and upper bounds for non-strict ones, and no condition on the boundaries of the events is needed.
Positive free energy implies delocalization
In this section, we prove that the boundary values of the cavity variables cannot all be real at a regular energy where the real boundary law has positive free energy at s = 1 s = 1 s = 1 . Together with the spectral averaging criterion of Proposition 7.11 , this gives purely absolutely continuous spectrum on intervals where, for almost every regular energy, all possible real boundary laws have positive free energy at s = 1 s = 1 s = 1 . Recall from Definition 5.12 that, for E ∈ E p E \in\mathcal{E}_p E ∈ E p with E > 0 E > 0 E > 0 , the no-ac hypothesis holds at E E E if P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 \mathbb{P}^{\infty}(\operatorname{Im}\Gamma_o(E + \mathrm{i}0) = 0) = 1 P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 ; in this case, the law π E \pi_E π E of D o = − 1 / Γ o ( E + i 0 ) D_o = -1/\Gamma_o(E + \mathrm{i}0) D o = − 1/ Γ o ( E + i 0 ) under P \mathbb{P} P is a real solution at energy E E E , by Proposition 5.11 (ii), and we call it the real boundary law.
Theorem 10.1. There exists a numerical constant p ∗ > 1 p_* > 1 p ∗ > 1 such that the following holds for every p ≥ p ∗ p \ge p_* p ≥ p ∗ . Let E ∈ E p E \in\mathcal{E}_p E ∈ E p with E > 0 E > 0 E > 0 , and suppose that P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 \mathbb{P}^{\infty}(\operatorname{Im}\Gamma_o(E + \mathrm{i}0) = 0) = 1 P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 . Then the real boundary law π E \pi_E π E satisfies φ π E ( 1 ) ≤ 0 \varphi_{\pi_E}(1) \le0 φ π E ( 1 ) ≤ 0 .
Corollary 10.2. Let p ≥ p ∗ p \ge p_* p ≥ p ∗ , where p ∗ p_* p ∗ is as in Theorem 10.1 , and let I ⊂ ( 0 , ∞ ) I \subset(0,\infty) I ⊂ ( 0 , ∞ ) be an interval. Suppose that for Lebesgue-almost every E ∈ I ∩ E p E \in I \cap\mathcal{E}_p E ∈ I ∩ E p , either P ∞ ( Im Γ o ( E + i 0 ) = 0 ) < 1 \mathbb{P}^{\infty}(\operatorname{Im}\Gamma_o(E + \mathrm{i}0) = 0) < 1 P ∞ ( Im Γ o ( E + i 0 ) = 0 ) < 1 , or P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 \mathbb{P}^{\infty}(\operatorname{Im}\Gamma_o(E + \mathrm{i}0) = 0) = 1 P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 and the real boundary law π E \pi_E π E satisfies φ π E ( 1 ) > 0 \varphi_{\pi_E}(1) > 0 φ π E ( 1 ) > 0 . Then P ∞ \mathbb{P}^{\infty} P ∞ -almost surely, H \mathcal{H} H is purely absolutely continuous in I ∪ ( − I ) I \cup(-I) I ∪ ( − I ) .
In particular, if P ∞ ( Im Γ o ( E + i 0 ) = 0 ) < 1 \mathbb{P}^{\infty}(\operatorname{Im}\Gamma_o(E + \mathrm{i}0) = 0) < 1 P ∞ ( Im Γ o ( E + i 0 ) = 0 ) < 1 for every E ∈ I ∩ E p E \in I \cap\mathcal{E}_p E ∈ I ∩ E p , then the conclusion of Corollary 10.2 holds.
Proof of Corollary 10.2, assuming Theorem 10.1. By Lemma 5.8 , the set E p \mathcal{E}_p E p has full Lebesgue measure. Let E ∈ I ∩ E p E \in I \cap\mathcal{E}_p E ∈ I ∩ E p be an energy at which the hypothesis holds; this is the case for almost every E ∈ I E \in I E ∈ I . If P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 \mathbb{P}^{\infty}(\operatorname{Im}\Gamma_o(E + \mathrm{i}0) = 0) = 1 P ∞ ( Im Γ o ( E + i 0 ) = 0 ) = 1 , then Theorem 10.1 implies that φ π E ( 1 ) ≤ 0 \varphi_{\pi_E}(1) \le0 φ π E ( 1 ) ≤ 0 , contradicting the hypothesis. Thus P ∞ ( Im Γ o ( E + i 0 ) = 0 ) < 1 \mathbb{P}^{\infty}(\operatorname{Im}\Gamma_o(E + \mathrm{i}0) = 0) < 1 P ∞ ( Im Γ o ( E + i 0 ) = 0 ) < 1 . By Proposition 5.11 (i), this probability equals 1 − u / ϱ p ∈ { 0 , 1 } 1 - u/\varrho_p \in\{0,1\} 1 − u / ϱ p ∈ { 0 , 1 } , where u = P ( Im Γ o ( E + i 0 ) > 0 ) ∈ { 0 , ϱ p } u = \mathbb{P}(\operatorname{Im}\Gamma_o(E + \mathrm{i}0) > 0) \in\{0,\varrho_p\} u = P ( Im Γ o ( E + i 0 ) > 0 ) ∈ { 0 , ϱ p } . Hence it equals zero, and P ( Im Γ o ( E + i 0 ) > 0 ) = u = ϱ p > 0 \mathbb{P}(\operatorname{Im}\Gamma_o(E + \mathrm{i}0) > 0) = u = \varrho_p > 0 P ( Im Γ o ( E + i 0 ) > 0 ) = u = ϱ p > 0 . This verifies the hypothesis of Proposition 7.11 at almost every E ∈ I E \in I E ∈ I , and Proposition 7.11 gives the conclusion.
Outline of the proof
The proof of Theorem 10.1 follows the resonance strategy of Aizenman and Warzel, who used it in [12] to prove absolutely continuous spectrum for the Anderson model on regular trees beyond the band of the hopping operator. Versions of this strategy for tree operators were developed by Aggarwal, Bordenave, and Lopatto for the local limits of Lévy matrices [7], and by Aggarwal and Lopatto for the Anderson model on the Bethe lattice [4]. We adapt it to the Poisson tree with complex Gaussian weights. The argument uses only the properties of real solutions established in Sections 5 and § 6.2 , so it applies to an arbitrary real boundary law. We now describe its steps.
The quantile. Fix p ≥ p ∗ p \ge p_* p ≥ p ∗ and E ∈ E p E \in\mathcal{E}_p E ∈ E p with E > 0 E > 0 E > 0 , suppose that the no-ac hypothesis holds at E E E , and suppose, for a contradiction, that φ ( 1 ) > 0 \varphi(1) > 0 φ ( 1 ) > 0 , where φ = φ π E \varphi= \varphi_{\pi_E} φ = φ π E . Let α = q ∗ / 2 \alpha= q_*/2 α = q ∗ /2 be the numerical level fixed in (10.5) below. For η > 0 \eta> 0 η > 0 , let ζ α ( η ) \zeta_\alpha(\eta) ζ α ( η ) be the quantile of Im Γ o ( E + i η ) \operatorname{Im}\Gamma_o(E+i\eta) Im Γ o ( E + i η ) under P ∞ \mathbb{P}^{\infty} P ∞ defined in (9.11). By Lemma 9.9 (i), we have 0 < ζ α ( η ) < ∞ 0 < \zeta_\alpha(\eta) < \infty 0 < ζ α ( η ) < ∞ , and P ∞ ( Im Γ o ( E + i η ) ≥ t ) < α \mathbb{P}^{\infty}(\operatorname{Im}\Gamma_o(E+i\eta) \ge t) < \alpha P ∞ ( Im Γ o ( E + i η ) ≥ t ) < α for every t > ζ α ( η ) t > \zeta_\alpha(\eta) t > ζ α ( η ) . Under the no-ac hypothesis, Im Γ o ( E + i η ) → 0 \operatorname{Im}\Gamma_o(E+i\eta) \to0 Im Γ o ( E + i η ) → 0 almost surely, so that ζ α ( η ) → 0 \zeta_\alpha(\eta) \to0 ζ α ( η ) → 0 as η ↓ 0 \eta\downarrow0 η ↓ 0 , at an unknown rate. For this reason, we always compare Im Γ o ( E + i η ) \operatorname{Im}\Gamma_o(E+i\eta) Im Γ o ( E + i η ) with the quantile at the same imaginary part η \eta η . We show that there exist δ > 0 \delta> 0 δ > 0 and C < ∞ C < \infty C < ∞ such that, for every large n n n , there exists η n > 0 \eta_n > 0 η n > 0 with
P ∞ ( Im Γ o ( E + i η n ) ≥ e 2 δ n − C n 3 / 4 ζ α ( η n ) ) ≥ q ∗ > α . (157) \mathbb{P}^{\infty}\left(\operatorname{Im}\Gamma_o(E+i\eta_n) \ge e^{2\delta n-Cn^{3/4}}\zeta_\alpha(\eta_n)\right) \ge q_* > \alpha. \tag*{(157)} P ∞ ( Im Γ o ( E + i η n ) ≥ e 2 δ n − C n 3/4 ζ α ( η n ) ) ≥ q ∗ > α . (157)
For large n n n , the threshold in (157) is larger than ζ α ( η n ) \zeta_\alpha(\eta_n) ζ α ( η n ) , which contradicts Lemma 9.9 (i).
Exponents. By Lemma 6.16 , there exists s ∈ ( 1 / 2 , 1 ) s \in(1/2,1) s ∈ ( 1/2 , 1 ) such that γ = − φ ′ ( s ) > 0 \gamma= -\varphi'(s) > 0 γ = − φ ′ ( s ) > 0 and F = φ ( s ) − ( 1 − s ) γ > 0 F = \varphi(s) - (1-s)\gamma> 0 F = φ ( s ) − ( 1 − s ) γ > 0 . The number F F F is the value at u = 1 u=1 u = 1 of the tangent line to φ \varphi φ at s s s . Since φ ′ ( 1 ) = 0 \varphi'(1) = 0 φ ′ ( 1 ) = 0 , it tends to φ ( 1 ) > 0 \varphi(1) > 0 φ ( 1 ) > 0 as s ↑ 1 s \uparrow1 s ↑ 1 , and this is how s s s is chosen. In § 8.3.3 , the free energy is replaced by the free energy φ ^ \widehat{\varphi} φ of blocks of m m m factors with entries in a compact set, with the corresponding numbers γ ^ ≈ γ \widehat{\gamma} \approx\gamma γ ≈ γ and F ^ ≈ F \widehat{F} \approx F F ≈ F (Lemma 9.11 ). We fix δ > 0 \delta> 0 δ > 0 small in terms of F F F and ( 2 s − 1 ) γ (2s-1)\gamma ( 2 s − 1 ) γ , and we set N = n + 10 N = n+10 N = n + 10 , b n = n 3 / 4 b_n = n^{3/4} b n = n 3/4 , and T n = e ( γ ^ + δ ) n T_n = e^{(\widehat{\gamma}+\delta)n} T n = e ( γ + δ ) n .
The gain from a resonance. Let x x x be a vertex of generation N N N , and let A A A be the product of the transfer matrices (factors, in the sense of Definition 9.1 ) along the path from o o o to x x x , evaluated at z = E + i η z = E+i\eta z = E + i η . By Lemma 9.3 , we have ∣ G o x ( z ) ∣ ≥ ∣ G x x ( z ) ∣ / ∥ A ∥ |G_{ox}(z)| \ge|G_{xx}(z)|/\lVert A\rVert ∣ G o x ( z ) ∣ ≥ ∣ G xx ( z ) ∣/ ∥ A ∥ . If y y y is a child of x x x , the Ward identity of Lemma 3.19 (iii) gives Im Γ o ( z ) ≥ ∣ G o x ( z ) ∣ 2 ∣ h x y ∣ 2 Im Γ y ( z ) \operatorname{Im}\Gamma_o(z) \ge|G_{ox}(z)|^2|h_{xy}|^2\operatorname{Im}\Gamma_y(z) Im Γ o ( z ) ≥ ∣ G o x ( z ) ∣ 2 ∣ h x y ∣ 2 Im Γ y ( z ) . Suppose that ∥ A ∥ ≤ e γ ^ n + O ( b n ) \lVert A\rVert\le e^{\widehat{\gamma}n+O(b_n)} ∥ A ∥ ≤ e γ n + O ( b n ) , that ∣ G x x ( z ) ∣ ≥ T n |G_{xx}(z)| \ge T_n ∣ G xx ( z ) ∣ ≥ T n , that ∣ h x y ∣ 2 |h_{xy}|^2 ∣ h x y ∣ 2 is bounded below, and that Im Γ y ( z ) ≥ ζ α ( η ) / 2 \operatorname{Im}\Gamma_y(z) \ge\zeta_\alpha(\eta)/2 Im Γ y ( z ) ≥ ζ α ( η ) /2 . Then Im Γ o ( z ) ≥ e 2 δ n − O ( b n ) ζ α ( η ) \operatorname{Im}\Gamma_o(z) \ge e^{2\delta n-O(b_n)}\zeta_\alpha(\eta) Im Γ o ( z ) ≥ e 2 δ n − O ( b n ) ζ α ( η ) . We call the factor e 2 δ n − O ( b n ) e^{2\delta n-O(b_n)} e 2 δ n − O ( b n ) by which Im Γ o ( z ) \operatorname{Im}\Gamma_o(z) Im Γ o ( z ) exceeds the quantile the gain ; the precise statement is Lemma 10.8.
Resonant vertices. We count the vertices x x x of generation N N N for which the conditions of the previous paragraph hold; they are called resonant (Definition 10.7). The conditions are the following. The factors of the first n n n generations satisfy a regular growth condition: the logarithms of the norms of the products of consecutive blocks of m m m factors have partial sums within b n b_n b n of their typical values (R3). The last ten vertices of the path have no side children (R1), and the squared weights of their edges lie in fixed intervals (R4).
By Lemma 9.5, this forces the rootward cavity variable at the parent of x x x into a fixed disc, whatever its value at generation n n n . The vertex x x x has exactly one leaf child and one surviving child y y y (R2), and the cavity variable of y y y is bounded and has imaginary part at least ζ ^ α ( η ) / 2 \widehat{\zeta}_{\alpha}(\eta)/2 ζ α ( η ) /2 (R5); we call y y y the seed , since its imaginary part is the one that the resonance amplifies. By (9.10), the diagonal entry G x x ( z ) G_{xx}(z) G xx ( z ) is then a Möbius function of the squared weight t t t of the leaf edge, with a pole whose real part lies in [ E 2 / 2 , 3 E 2 / 2 ] [E^2/2,3E^2/2] [ E 2 /2 , 3 E 2 /2 ] (Lemma 9.6) and whose imaginary part, which we call the pole width , is small (R6). Since t t t has the density p e − p t pe^{-pt} p e − pt , the event ∣ G x x ( z ) ∣ ≥ T n |G_{xx}(z)|\geq T_n ∣ G xx ( z ) ∣ ≥ T n (R7) has probability of order T n − 1 T_n^{-1} T n − 1 (Lemma 9.7).
The first moment. By the many-to-one formula for the backbone, the expected number of resonant vertices is p N p^N p N times the probability that the endpoint of a spine of length N N N is resonant. The structural conditions (R1), (R2) and the conditions (R4), (R5) cost a factor independent of n n n , the tuning (R7) costs a factor of order T n − 1 T_n^{-1} T n − 1 , and the regular growth condition (R3) costs a factor e − ι n + O ( b n ) e^{-\iota n+O(b_n)} e − ι n + O ( b n ) , where ι = log p − φ ^ ( s ) − s γ ^ \iota=\log p-\widehat{\varphi}(s)-s\widehat{\gamma} ι = log p − φ ( s ) − s γ (Lemma 9.12, which computes this probability after an exponential tilt of the law of a block at the exponent s s s ). Since log p − ι − γ ^ − δ = F ^ − δ \log p-\iota-\widehat{\gamma}-\delta=\widehat{F}-\delta log p − ι − γ − δ = F − δ , we obtain in Lemma 10.9 that
E ∞ N n ( η ) ≥ e ( F ^ − δ ) n − O ( b n ) . \mathbb{E}^{\infty}N_n(\eta)\geq e^{(\widehat{F}-\delta)n-O(b_n)}. E ∞ N n ( η ) ≥ e ( F − δ ) n − O ( b n ) .
The probability of the regular growth event is computed for the one-spine real recursive tree process, at η = 0 \eta=0 η = 0 , where the factors along a spine are independent with the law ν f a c \nu_{\mathrm{fac}} ν fac determined by π E \pi_E π E (Lemma 5.21). The growth of their products is then governed by φ \varphi φ . The bound is transferred to a small imaginary part η > 0 \eta>0 η > 0 , with n n n fixed, by Lemma 9.15: lower bounds are proved for events defined by strict inequalities, which are open, and upper bounds for events defined by non-strict inequalities, which are closed.
The second moment. By the many-to-two formula, the expected number of ordered pairs ( u , v ) (u,v) ( u , v ) of resonant vertices whose paths split at generation j j j is ϱ p p 2 N − j \varrho_p p^{2N-j} ϱ p p 2 N − j times a probability under a two-spine law. The two leaf weights are fresh, and by Lemma 9.4 the $2\times 2$ block of the resolvent at { u , v } \{u,v\} { u , v } is a function of them. By Lemma 9.8, the probability that both diagonal entries exceed T n T_n T n is at most of order T n − 2 + T n − 1 min { 1 , D } T_n^{-2}+T_n^{-1}\min\{1,\mathcal{D}\} T n − 2 + T n − 1 min { 1 , D } , where D \mathcal{D} D measures the interaction of u u u and v v v through the path that joins their parents. The first term is the one for independent vertices. With the upper bound of Lemma 9.12(i) for regular growth events on the common part and on the two arms of the paths, it contributes e O ( b n ) ( E ∞ N n ) 2 e^{O(b_n)}(\mathbb{E}^{\infty}N_n)^2 e O ( b n ) ( E ∞ N n ) 2 after summation over j j j . For the second term, we bound E D \mathbb{E}\mathcal{D} E D by averaging over four groups of three fresh factors, placed next to the two ends and next to the fork of the joining path (Lemma 9.13). What remains are the negative moments of the norms of the products along the two arms, which decay like e ( n − j ) ( φ ( s ) − log p ) e^{(n-j)(\varphi(s)-\log p)} e ( n − j ) ( φ ( s ) − l o g p ) , where φ ( s ) < log p \varphi(s)<\log p φ ( s ) < log p . The resulting exponent is affine in j j j , and its values at j = 0 j=0 j = 0 and j = n j=n j = n are negative by the choice of δ \delta δ . We obtain $\mathbb{E}^{\infty}[N_n(N_n-1)]\leq e^{O(b_n)}(\mathbb{E}^{\infty}N_n)^2 (Lemma 10.12), and the Paley–Zygmund inequality gives P ∞ ( N n ≥ 1 ) ≥ e − O ( b n ) \mathbb{P}^{\infty}(N_n \ge1) \ge e^{-O(b_n)} P ∞ ( N n ≥ 1 ) ≥ e − O ( b n ) . On this event, Im Γ o ( E + i η ) ≥ e 2 δ n − O ( b n ) ζ α ( η ) \operatorname{Im}\Gamma_o(E+i\eta) \ge e^{2\delta n-O(b_n)}\zeta_\alpha(\eta) Im Γ o ( E + i η ) ≥ e 2 δ n − O ( b n ) ζ α ( η ) .
Amplification. It remains to raise the probability e − O ( b n ) e^{-O(b_n)} e − O ( b n ) to the fixed level q ∗ > α q_* > \alpha q ∗ > α . In § 8.5.2 , we show that if Im Γ o ≥ y \operatorname{Im}\Gamma_o \ge y Im Γ o ≥ y with probability at least e − a e^{-a} e − a , then Im Γ o ≥ e − C 3 ( 1 + a ) y \operatorname{Im}\Gamma_o \ge e^{-C_3(1+a)}y Im Γ o ≥ e − C 3 ( 1 + a ) y with probability at least q ∗ q_* q ∗ (Corollary 10.6). The main step uses the Poisson structure of the children of the root. A random set of children, whose number has a Poisson law with mean about 64, carries independent copies of an event of probability q q q , while the other children keep ∣ D o ∣ |D_o| ∣ D o ∣ in a fixed annulus. If one of the copies occurs, then so does the event at the root, with a fixed loss in the imaginary part, and this happens with probability at least 4 q 4q 4 q . After O ( 1 + a ) O(1+a) O ( 1 + a ) iterations, the probability reaches q ∗ q_* q ∗ . This is the only step that requires p ≥ p ∗ p \ge p_* p ≥ p ∗ , and p ∗ p_* p ∗ does not depend on E E E , since it uses only Lemma 5.6, which holds uniformly over all real solutions. With a = O ( b n ) a=O(b_n) a = O ( b n ) , we obtain (157) .
Order of choices. The constants are fixed in the order (a)–(f) listed in Section 10.3. The numerical constants come first; then the constants that depend on p p p and E E E ; then the depth n n n ; and finally the imaginary part η n \eta_n η n , which depends on n n n . Each estimate at positive η \eta η is obtained from an estimate at η = 0 \eta=0 η = 0 with n n n fixed, and the two limits n → ∞ n\to\infty n → ∞ and η ↓ 0 \eta\downarrow0 η ↓ 0 are never taken in the same estimate. The quantile ζ α ( η ) \zeta_\alpha(\eta) ζ α ( η ) enters only through the seed condition (R5), whose probability is bounded at each fixed η \eta η by Lemma 9.9 (ii), and through the gain.
Organization. § 8.5.2 contains the amplification. Section 10.3 fixes the constants, defines the resonant vertices, and proves the gain. Sections 10.4 and 10.5 contain the first and second moments, and Section 10.6 completes the proof.
Amplification at the root
In this section, we show that an event of small probability on which Im Γ o \operatorname{Im}\Gamma_o Im Γ o is large can be converted into an event of a fixed probability, at the cost of a moderate factor in the imaginary part. We begin with the numerical constants. Let
a 0 = 1 32 2 , (158) a_0 = \frac{1}{32\sqrt{2}}, \tag*{(158)} a 0 = 32 2 1 , (158)
so that 4 a 0 = a ( 1 / 32 ) 4a_0=a(1/32) 4 a 0 = a ( 1/32 ) in the notation of Lemma 5.6. Let N 0 ∈ N N_0\in\mathbb{N} N 0 ∈ N be such that a Pois ( 128 ) \operatorname{Pois}(128) Pois ( 128 ) variable exceeds N 0 − 1 N_0-1 N 0 − 1 with probability at most 1 / 32 1/32 1/32 . Let λ ∈ [ 32 , 128 ] \lambda\in[32,128] λ ∈ [ 32 , 128 ] and N ∼ Pois ( λ ) N\sim\operatorname{Pois}(\lambda) N ∼ Pois ( λ ) . Since a Pois ( 128 ) \operatorname{Pois}(128) Pois ( 128 ) variable is the sum of N N N and an independent Pois ( 128 − λ ) \operatorname{Pois}(128-\lambda) Pois ( 128 − λ ) variable, we have P ( N ≥ N 0 ) ≤ 1 / 32 \mathbb{P}(N\ge N_0)\le1/32 P ( N ≥ N 0 ) ≤ 1/32 , and the identity E [ N 1 { N > N 0 } ] = λ P ( N ≥ N 0 ) \mathbb{E}[N\mathbf{1}_{\{N>N_0\}}]=\lambda\mathbb{P}(N\ge N_0) E [ N 1 { N > N 0 } ] = λ P ( N ≥ N 0 ) gives
P ( N > N 0 ) ≤ 1 32 , E [ N 1 { N ≤ N 0 } ] ≥ λ 2 ≥ 16. (159) \mathbb{P}(N>N_0)\le\frac{1}{32},\qquad\mathbb{E}[N\mathbf{1}_{\{N\le N_0\}}]\ge\frac{\lambda}{2}\ge16. \tag*{(159)} P ( N > N 0 ) ≤ 32 1 , E [ N 1 { N ≤ N 0 } ] ≥ 2 λ ≥ 16. (159)
Set y + = a 0 2 / ( 4 N 0 ) y_{+}=a_{0}^{2}/(4N_{0}) y + = a 0 2 / ( 4 N 0 ) , y − = y + / 2 y_{-}=y_{+}/2 y − = y + /2 , J = [ y − , y + ] J=[y_{-},y_{+}] J = [ y − , y + ] , and ρ J = P ( Y ∈ J ) \rho_{J}=\mathbb{P}(Y\in J) ρ J = P ( Y ∈ J ) for Y ∼ Exp ( 1 ) Y\sim\operatorname{Exp}(1) Y ∼ Exp ( 1 ) . We let p ∗ ≥ 32 log 64 + 1 p_{*}\ge32\log64+1 p ∗ ≥ 32 log 64 + 1 be the smallest number such that
7 16 p ∗ ρ J ≥ 64 , e − p ∗ / 2 + 128 ≤ 1 16 . (160) \frac{7}{16}p_{*}\rho_{J}\ge64,\qquad e^{-p_{*}/2+128}\le\frac{1}{16}. \tag*{(160)} 16 7 p ∗ ρ J ≥ 64 , e − p ∗ /2 + 128 ≤ 16 1 . (160)
and we set
q 0 = 1 2 N 0 , q ∗ = 1 4 N 0 , α = q ∗ 2 . (161) q_{0}=\frac{1}{2N_{0}},\qquad q_{*}=\frac{1}{4N_{0}},\qquad\alpha=\frac{q_{*}}{2}. \tag*{(161)} q 0 = 2 N 0 1 , q ∗ = 4 N 0 1 , α = 2 q ∗ . (161)
All these constants are numerical. Let p ≥ p ∗ p\ge p_{*} p ≥ p ∗ . Since 1 − e − p / 2 > 1 / 2 1-e^{-p/2}>1/2 1 − e − p /2 > 1/2 , the concave function ϱ ↦ 1 − e − p ϱ − ϱ \varrho\mapsto1-e^{-p\varrho}-\varrho ϱ ↦ 1 − e − pϱ − ϱ is positive at ϱ = 1 / 2 \varrho=1/2 ϱ = 1/2 , and therefore ϱ p > 1 / 2 \varrho_{p}>1/2 ϱ p > 1/2 . By Lemma 5.6 with ε = 1 / 32 \varepsilon=1/32 ε = 1/32 , every real solution at every energy E > 0 E>0 E > 0 satisfies P ( ∣ D ∣ ≤ 4 a 0 / p ) < 1 / 32 \mathbb{P}(|D|\le4a_{0}/\sqrt{p})<1/32 P ( ∣ D ∣ ≤ 4 a 0 / p ) < 1/32 . By (10.4), we have p ( ϱ p − 1 / 16 ) ρ J ≥ 64 p(\varrho_{p}-1/16)\rho_{J}\ge64 p ( ϱ p − 1/16 ) ρ J ≥ 64 , e − ( p ϱ p − 128 ) ≤ 1 / 16 e^{-(p\varrho_{p}-128)}\le1/16 e − ( p ϱ p − 128 ) ≤ 1/16 , and e − ( p − 1 ) ϱ p ≤ 1 / 16 e^{-(p-1)\varrho_{p}}\le1/16 e − ( p − 1 ) ϱ p ≤ 1/16 .
Standing assumptions. From now on until the end of this section, we fix p ≥ p ∗ p\ge p_{*} p ≥ p ∗ and E ∈ E p E\in\mathcal{E}_{p} E ∈ E p with E > 0 E>0 E > 0 , and we assume that the no-ac hypothesis holds at E E E . We write π = π E \pi=\pi_{E} π = π E for the real boundary law, D o ( z ) = − 1 / Γ o ( z ) D_{o}(z)=-1/\Gamma_{o}(z) D o ( z ) = − 1/ Γ o ( z ) for z ∈ C ∖ R z\in\mathbb{C}\setminus\mathbb{R} z ∈ C ∖ R , and D o ( E ) = − 1 / Γ o ( E + i 0 ) D_{o}(E)=-1/\Gamma_{o}(E+i0) D o ( E ) = − 1/ Γ o ( E + i 0 ) . By Proposition 5.11 (ii), P \mathbb{P} P -almost surely Γ v ( E + i η ) → Γ v ( E + i 0 ) ∈ R ∖ { 0 } \Gamma_{v}(E+i\eta)\to\Gamma_{v}(E+i0)\in\mathbb{R}\setminus\{0\} Γ v ( E + i η ) → Γ v ( E + i 0 ) ∈ R ∖ { 0 } as η ↓ 0 \eta\downarrow0 η ↓ 0 for every vertex v v v . The law of D o ( E ) D_{o}(E) D o ( E ) is π \pi π under P \mathbb{P} P and π E ∞ \pi_{E}^{\infty} π E ∞ under P ∞ \mathbb{P}^{\infty} P ∞ , and π E ∞ \pi_{E}^{\infty} π E ∞ has no atoms (see the discussion after Definition 5.12 ). We set
d − = a 0 p , h 0 = d − 2 4 . (162) d_{-}=\frac{a_{0}}{\sqrt{p}},\qquad h_{0}=\frac{d_{-}^{2}}{4}. \tag*{(162)} d − = p a 0 , h 0 = 4 d − 2 . (162)
we let d + ≥ 16 d − d_{+}\ge16d_{-} d + ≥ 16 d − be such that P ( ∣ D o ( E ) ∣ > d + / 2 ) ≤ 1 / 32 \mathbb{P}(|D_{o}(E)|>d_{+}/2)\le1/32 P ( ∣ D o ( E ) ∣ > d + /2 ) ≤ 1/32 , and we set c 0 = min { 1 , y − / ( p d + 2 ) } c_{0}=\min\{1,y_{-}/(pd_{+}^{2})\} c 0 = min { 1 , y − / ( p d + 2 )} . Then P ( ∣ D o ( E ) ∣ ≤ 4 d − ) < 1 / 32 \mathbb{P}(|D_{o}(E)|\le4d_{-})<1/32 P ( ∣ D o ( E ) ∣ ≤ 4 d − ) < 1/32 by the choice of a 0 a_{0} a 0 , and N 0 y + / ( p d − ) = d − / 4 N_{0}y_{+}/(pd_{-})=d_{-}/4 N 0 y + / ( p d − ) = d − /4 .
Definition 10.3. For η ∈ ( 0 , 1 ] \eta\in(0,1] η ∈ ( 0 , 1 ] and y > 0 y>0 y > 0 , let A y ( η ) A_{y}(\eta) A y ( η ) be the event that
d − ≤ ∣ D o ( E + i η ) ∣ ≤ d + and y ≤ Im Γ o ( E + i η ) ≤ 1. d_{-}\le|D_{o}(E+i\eta)|\le d_{+}\qquad\text{and}\qquad y\le\operatorname{Im}\Gamma_{o}(E+i\eta)\le1. d − ≤ ∣ D o ( E + i η ) ∣ ≤ d + and y ≤ Im Γ o ( E + i η ) ≤ 1.
Recall from Lemma 4.2 that the children of the root form the Poisson process Π o = ∑ k δ ( w k , θ k ω ) \Pi_{o}=\sum_{k}\delta_{(w_{k},\theta_{k}\omega)} Π o = ∑ k δ ( w k , θ k ω ) on M = C × Ω M=\mathbb{C}\times\Omega M = C × Ω with intensity measure p m p\,m p m , where m = γ ⊗ P m=\gamma\otimes\mathbb{P} m = γ ⊗ P . The auxiliary mark ζ k \zeta_{k} ζ k of the child k k k is the root coordinate of θ k ω \theta_{k}\omega θ k ω ; its coordinates other than the rank ζ k ( 1 ) \zeta_{k}(1) ζ k ( 1 ) are independent uniform variables, independent of the marked tree (Remark 4.3). We use them to color children.
Lemma 10.4. There exists η c h ∈ ( 0 , 1 ] \eta_{\mathrm{ch}}\in(0,1] η ch ∈ ( 0 , 1 ] , depending on p p p and E E E , such that the following holds for every η ∈ ( 0 , η c h ] \eta\in(0,\eta_{\mathrm{ch}}] η ∈ ( 0 , η ch ] , every y > 0 y>0 y > 0 , and every q ∈ ( 0 , q 0 ] q\in(0,q_{0}] q ∈ ( 0 , q 0 ] . If P ∞ ( A y ( η ) ) ≥ q \mathbb{P}^{\infty}(A_{y}(\eta))\ge q P ∞ ( A y ( η )) ≥ q , then P ∞ ( A c 0 y ( η ) ) ≥ 4 q \mathbb{P}^{\infty}(A_{c_{0}y}(\eta))\ge4q P ∞ ( A c 0 y ( η )) ≥ 4 q .
Proof. We write z = E + i η z=E+i\eta z = E + i η . The proof splits the children of the root into a bounded number of amplifying children , which carry copies of the event A y ( η ) A_{y}(\eta) A y ( η ) , and the remaining reserve children , which keep ∣ D o ∣ |D_{o}| ∣ D o ∣ in the interval [ d − , d + ] [d_{-},d_{+}] [ d − , d + ] of Definition 10.3.