betaFeedback

WorksMathematics

Mobility Edge for Sparse Complex Gaussian Matrices

Abstract

We consider the sparse Hermitian random matrix supported on the Erdős–Rényi graph G(n,p/n)G(n,p/n), whose nonzero entries are independent complex Gaussian variables of variance 1/p1/p, together with its local limit, a weighted Poisson–Galton–Watson tree. We prove that for every fixed and sufficiently large pp, the limiting operator on the tree conditioned on survival has exactly one positive mobility edge Ec(p)E_c(p): almost surely, its spectrum is purely absolutely continuous on (−Ec(p),0)∪(0,Ec(p))(-E_c(p),0)\cup(0,E_c(p)) and pure point outside [−Ec(p),Ec(p)][-E_c(p),E_c(p)]. Moreover, Ec(p)/pE_c(p)/\sqrt p converges to π/2\sqrt\pi/2 as pp tends to infinity. Thus the mobility edge lies far in the tail of the spectrum, rather than near the edge 22 of the limiting semicircle law, as had been predicted in the physics literature.

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) 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 pp, 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, 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>2p > 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-stable law with α∈(0,1)\alpha\in(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 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 Zd\mathbb{Z}^{d} for every fixed d≥3d \ge3 [104], and pure-point spectrum throughout the spectrum at every positive disorder strength on Z2\mathbb{Z}^{2} [107]. For random regular graphs of every fixed degree d≥3d \ge3, 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) has exactly one positive mobility edge in its local limit, for every fixed sufficiently large pp, and we determine its location to leading order. Its location, of order p\sqrt{p}, is far in the tail of the spectrum, which contradicts the large-pp 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>1p>1. For each integer n>pn>p, let GnG_n be an Erdős–Rényi random graph on the vertex set {1,…,n}\{1,\ldots,n\} in which each edge is present independently with probability p/np/n. Conditional on GnG_n, let (wij)1≤i<j≤n(w_{ij})_{1\leq i<j\leq n} be independent standard complex Gaussian random variables, so that each wijw_{ij} has density π−1e−∣z∣2\pi^{-1}e^{-|z|^2} with respect to Lebesgue measure on C\mathbb{C}. We consider the n×nn\times n Hermitian matrix HnH_n defined by

Hn(i,j)=1{ij∈E(Gn)}pwij,Hn(j,i)=Hn(i,j)‾,Hn(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)}

for 1≤i<j≤n1\leq i<j\leq n. Throughout the paper, pp is fixed while nn tends to infinity, and statements about large pp concern the resulting limiting objects.

Since p>1p>1, the graph GnG_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 gc\mathrm{gc} for the vertex set of the giant component, that is, of the largest connected component of GnG_n (if there are several, the one containing the smallest vertex; see Definition 3.1), Nn=∣gc∣N_n=|\mathrm{gc}| for its size, and HngcH_n^{\mathrm{gc}} for the restriction of HnH_n to gc\mathrm{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 HnH_n is described by a random tree. Let T\mathcal{T} be a Poisson–Galton–Watson tree with mean offspring pp and root oo, and attach to each edge ee an independent weight he=we/ph_e=w_e/\sqrt{p}, where the wew_e are standard complex Gaussian variables; we set hvu=huv‾h_{vu}=\overline{h_{uv}}. Let H\mathcal{H} be the weighted adjacency operator of T\mathcal{T},

(Hf)(v)=∑u∼vhvuf(u).(\mathcal{H}f)(v)=\sum_{u\sim v}h_{vu}f(u).

Although the degrees in T\mathcal{T} are unbounded, H\mathcal{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 GnG_n converges locally to the root of (T,(he))(\mathcal{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 SS that it is infinite. We write P\mathbb{P} for the law of the weighted tree, ϱp=P(S)\varrho_p=\mathbb{P}(S) for its survival probability, and P∞=P(⋅∣S)\mathbb{P}^{\infty}=\mathbb{P}(\mathord{\cdot}\mid S) for the survival-conditioned law.

For an interval II, we say that H\mathcal{H} has purely absolutely continuous spectrum (respectively pure point spectrum) in II if, for every vector ff, the restriction to II of the spectral measure of ff 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∣=π2,w∼NC(0,1).(2)\mu_* = \mathbb{E}|w| = \frac{\sqrt{\pi}}{2}, \qquad w \sim\mathcal{N}_{\mathbb{C}}(0,1). \tag*{(2)}

Theorem 1.1. There exists p0<∞p_0 < \infty such that for every p≥p0p \ge p_0 there exists Ec(p)>0E_c(p) > 0 with the following properties. Almost surely with respect to P∞\mathbb{P}^\infty, the operator H\mathcal{H} has purely absolutely continuous spectrum in

(−Ec(p),0)∪(0,Ec(p))\left(-E_c(p),0\right) \cup\left(0,E_c(p)\right)

and pure point spectrum in

(−∞,−Ec(p))∪(Ec(p),∞).\left(-\infty,-E_c(p)\right) \cup\left(E_c(p),\infty\right).

Moreover,

lim⁡p→∞Ec(p)p=μ∗=π2.(3)\lim_{p \to\infty} \frac{E_c(p)}{\sqrt{p}} = \mu_* = \frac{\sqrt{\pi}}{2}. \tag*{(3)}

Finally, P∞\mathbb{P}^\infty-almost surely, H\mathcal{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 pp, the positive half-line contains exactly one mobility edge. The theorem makes no statement about the spectral type at the two energies ±Ec(p)\pm 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, the limiting empirical spectral measure of HnH_n converges to the semicircle law on [−2,2][-2,2], as a computation with the moment method shows, while Ec(p)E_c(p) is of order p\sqrt{p}. Eigenvalues near Ec(p)E_c(p) therefore come from rare local configurations with unusually large edge weights. The constant μ∗\mu_* arises from a first-moment computation, which we explain in Section 2.5: at the energy E=ξpE = \xi\sqrt{p}, a typical path of length LL in the tree contributes a product of about LL factors ∣we∣/(ξp)|w_e|/(\xi p) to the resolvent, and there are about pLp^L such paths. Their total contribution therefore grows or decays like (μ∗/ξ)L(\mu_*/\xi)^L.

Our second result identifies Ec(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=⟨δo,(H−z)−1δo⟩\Gamma= \langle\delta_o,(\mathcal{H}-z)^{-1}\delta_o\rangle for zz in the upper half-plane. Its law satisfies the cavity equation

Γ=d−(z+1p∑j=1NγjΓj)−1,N∼Pois⁡(p),γj∼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),

where the variables on the right are independent and the Γj\Gamma_j are distributed as Γ\Gamma. For E>2E>2, let g(E)=(E−E2−4)/2g(E)=(E-\sqrt{E^2-4})/2 and θ=g(E)2\theta=g(E)^2. For a real energy EE, the variable B=−g(E)/ΓB=-g(E)/\Gamma formally satisfies

B=d1+θ−θp∑j=1NγjBj.(4)B\overset{\mathrm{d}}{=}1+\theta-\frac{\theta}{p}\sum_{j=1}^{N}\frac{\gamma_j}{B_j}. \tag*{(4)}

For every real solution of this equation we define in § 6.2 a function φ(s)\varphi(s), s∈(0,2)s\in(0,2), as the logarithm of the Perron–Frobenius eigenvalue λ(s)\lambda(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), φ(s)\varphi(s) is the exponential growth rate of the expected generation sums of the ss-th powers of the formal path products. Its limiting value φ(1)=lim⁡s↑1φ(s)\varphi(1)=\lim_{s\uparrow1}\varphi(s) decides the spectral type.

References

No references were extracted for this paper.

Paper details

vlnc:W_c56ja8gm
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Anthropic Claude Opus 5.5, OpenAI ChatGPT 5.6, OpenAI ChatGPT 6, OpenAI ChatGPT 6.1

View statement
Looking for related papers…
Contents