1ETH Zürich, Department of Mathematics, Rämistrasse 101, 8092 Zürich, Switzerland and Hong Kong Institute for Advanced Study, City University of Hong Kong, Kowloon Tong, Hong Kong Special Administrative Region. Email: habib.ammari@math.ethz.ch.
2School of Science, Institute for Theoretical Sciences, Westlake University, Hangzhou, Zhejiang Province, 310030, P.R. China. Email: 1550862645cf@gmail.com.
3Yau Mathematical Sciences Center, Tsinghua University, Beijing 100084 and Beijing Institute of Mathematical Sciences and Applications, Beijing 101408, P.R. China. Email: wjing@tsinghua.edu.cn.
4ETH Zürich, Department of Mathematics, Rämistrasse 101, 8092 Zürich, Switzerland. Email: jiayu.qiu@math.ethz.ch.
5Department of Applied Mathematics and Theoretical Physics, University of Cambridge. Email: aju27@cam.ac.uk.
Introduction
High-contrast resonators exhibit resonances at wavelengths much larger than their characteristic size. Their collective response is described, to leading order in the subwavelength regime, by a capacitance matrix or operator; see [8 ] . This operator describes long-range coupling between the resonators and retains their geometry. We study Anderson localisation caused by independent random perturbations of the material parameters, keeping the full capacitance interaction both in the infinite crystal and in finite clusters containing an increasing number of resonators.
For Anderson models with short-range hopping, the mathematical theory of localisation is based on multiscale analysis [18 , 30 ] and fractional-moment estimates [2 , 3 ] ; see also [28 ] . Band-edge localisation and associated Wegner and density of states estimates for classes of random acoustic and electromagnetic operators were established in [17 , 15 ] . Recent OpenAI preprints [23 , 24 ] report pure-point spectrum at every positive disorder in two dimensions and an absolutely continuous spectral interval at weak disorder in dimensions at least three, for nearest-neighbour models with independent uniform site potentials.
Several recent works on resonator crystals provide the starting point of the present analysis that gives probabilistic localisation estimates for the infinite random operator and simultaneous concentration estimates for individual modes of the finite matrix corresponding to a finite cluster of resonators. Ammari, Davies, and Hiltunen [6 ] develop a generalised capacitance formulation for resonator arrays with random material parameters. Their capacitance formulation describes the coupling of localised modes and relates mode hybridisation to level repulsion and changes in localisation. The material perturbations enter through a diagonal multiplication operator, while the capacitance operator describes the geometric interactions. Ammari, Du, Fu, Jing, and Melcher [9 ] establish a variational capacitance framework, band-edge homogenisation, and the nonresonant sound-soft limit for finite crystals of subwavelength high-contrast resonators. Their global analyticity and nondegenerate upper band edge provide the deterministic inputs for the infinite-volume localisation argument. Their positive boundary comparison between truncated and finite capacitance matrices is essential for our finite-cluster result. For chains embedded in three dimensions, Ammari, Barandun, Qiu, and Uhlmann [4 ] establish critical long-range decay of the capacitance coefficients, prove convergence of finite operators and their spectra, and give numerical evidence of localisation at arbitrary disorder strength. Ammari and Qiu [11 ] prove absence of absolutely continuous spectrum for one-dimensional capacitance operators with multiplicative disorder and critical long-range interactions. That result extends the one-dimensional long-range approach of [19 ] and motivates the present spectral question. Full periodicity gives exponential decay of the capacitance coefficients. This decay allows fractional-moment methods [2 , 3 ] to establish spectral and dynamical localisation while retaining the long-range interaction.
Let d ∈ { 2 , 3 } d \in\{2,3\} d ∈ { 2 , 3 } , Y = ( − 1 / 2 , 1 / 2 ) d Y=(-1/2,1/2)^d Y = ( − 1/2 , 1/2 ) d , and let D ⋐ Y D \Subset Y D ⋐ Y be simply-connected with C 2 C^2 C 2 boundary. We also assume that Y ∖ D ‾ Y\setminus\overline{D} Y ∖ D is connected. The unperturbed crystal consists of D n : = D + n D_n:=D+n D n := D + n , n ∈ Z d n\in\mathbb{Z}^d n ∈ Z d , all with the same wave speed. We denote by α ∈ Y ∗ = R d / ( 2 π Z ) d \alpha\in Y^*=\mathbb{R}^d/(2\pi\mathbb{Z})^d α ∈ Y ∗ = R d / ( 2 π Z ) d the quasimomentum associated with the crystal when performing the Floquet transform. Let C C C be its static capacitance operator, introduced in § 2.1 , and c ( α ) c(\alpha) c ( α ) its scalar Bloch symbol. Thus c ( α ) = Cap α , D / ∣ D ∣ c(\alpha)=\operatorname{Cap}_{\alpha,D}/|D| c ( α ) = Cap α , D /∣ D ∣ and C n m = c n − m C_{nm}=c_{n-m} C nm = c n − m , where Cap α , D \operatorname{Cap}_{\alpha,D} Cap α , D is the quasiperiodic capacity of D D D (see [8 ] ) and c r c_r c r are the Fourier coefficients of c c c . See § 2.1 for the precise variational definitions.
To fix the physical normalisation, let ρ \rho ρ and κ o u t \kappa_{\mathrm{out}} κ out be the background density and bulk modulus, and set v 2 = κ o u t / ρ v^2=\kappa_{\mathrm{out}}/\rho v 2 = κ out / ρ . Inside the resonators the density is ρ r = δ ρ \rho_r=\delta\rho ρ r = δ ρ , 0 < δ ≪ 1 0<\delta\ll1 0 < δ ≪ 1 , and the squared wave speeds are v r 2 b n v_r^2b_n v r 2 b n , with v r > 0 v_r>0 v r > 0 fixed and b n > 0 b_n>0 b n > 0 . We denote the material configuration by ϖ \varpi ϖ and reserve ω ≥ 0 \omega\geq0 ω ≥ 0 for physical angular frequency. At unit period the transmission problem is
( Δ + ω 2 / v 2 ) u = 0 in R d ∖ ⋃ n D n ‾ , ( Δ + ω 2 / ( v r 2 b n ) ) u = 0 in D n , u ∣ + = u ∣ − , δ ∂ ν u ∣ + = ∂ ν u ∣ − on ∂ D n . (1) \begin{aligned}
(\Delta+\omega^2/v^2)u&=0 &&\text{in }\mathbb{R}^d\setminus\bigcup_n\overline{D_n},\\
(\Delta+\omega^2/(v_r^2b_n))u&=0 &&\text{in }D_n,\\
u|_+&=u|_- ,\qquad\delta\partial_\nu u|_+=\partial_\nu u|_- &&\text{on }\partial D_n. \tag*{(1)}
\end{aligned} ( Δ + ω 2 / v 2 ) u ( Δ + ω 2 / ( v r 2 b n )) u u ∣ + = 0 = 0 = u ∣ − , δ ∂ ν u ∣ + = ∂ ν u ∣ − in R d ∖ n ⋃ D n , in D n , on ∂ D n . (1)
The normal ν \nu ν points out of the resonator; the signs + + + and − - − denote exterior and interior traces, respectively. With q n q_n q n denoting the leading resonator amplitude, integration of the interior equation and matching of the fluxes give
C q = E M ϖ q , B ϖ C q = E q , E = ω 2 δ v r 2 , M ϖ = diag ( b n − 1 ) , B ϖ = M ϖ − 1 . (2) Cq=EM_{\varpi}q,\qquad B_{\varpi}Cq=Eq,\qquad E=\frac{\omega^2}{\delta v_r^2},\qquad M_{\varpi}=\operatorname{diag}(b_n^{-1}),\qquad B_{\varpi}=M_{\varpi}^{-1}. \tag*{(2)} C q = E M ϖ q , B ϖ C q = E q , E = δ v r 2 ω 2 , M ϖ = diag ( b n − 1 ) , B ϖ = M ϖ − 1 . (2)
For the periodic material configuration, the first Bloch frequency has the asymptotic expansion
ω 1 α ( δ ) = v r δ c ( α ) + O ( δ 3 / 2 ) (3) \omega_1^\alpha(\delta)=v_r\sqrt{\delta c(\alpha)}+O(\delta^{3/2}) \tag*{(3)} ω 1 α ( δ ) = v r δ c ( α ) + O ( δ 3/2 ) (3)
for fixed nonzero α \alpha α ; see [8 ] , Theorem 4.4. At physical period 0 < ε ≪ 1 0<\varepsilon\ll1 0 < ε ≪ 1 , the spectral conversion is instead E = ε 2 ω 2 / ( δ v r 2 ) E=\varepsilon^2\omega^2/(\delta v_r^2) E = ε 2 ω 2 / ( δ v r 2 ) . In particular, δ = ε 2 \delta=\varepsilon^2 δ = ε 2 gives E = ω 2 / v r 2 E=\omega^2/v_r^2 E = ω 2 / v r 2 . Throughout, energy means this reduced spectral parameter, proportional to squared frequency, rather than the physical energy of a wave. Unless explicitly stated otherwise, spectral intervals and their widths are expressed in E E E .
We prescribe the inverse material factors by
m n = b n − 1 = 1 + W ξ n , ξ n ∼ i i d Unif [ 0 , 1 ] , W > 0. m_n=b_n^{-1}=1+W\xi_n,\qquad\xi_n\mathrel{\overset{\mathrm{iid}}{\sim}}\operatorname{Unif}[0,1],\qquad W>0. m n = b n − 1 = 1 + W ξ n , ξ n ∼ iid Unif [ 0 , 1 ] , W > 0.
where “iid” means independent and identically distributed. The reduced random operator B ϖ C B_{\varpi}C B ϖ C has the self-adjoint representation on ℓ 2 ( Z d ) \ell^2(\mathbb{Z}^d) ℓ 2 ( Z d ) given by
H ϖ = M ϖ − 1 / 2 C M ϖ − 1 / 2 , ψ = M ϖ 1 / 2 q . H_{\varpi}=M_{\varpi}^{-1/2}CM_{\varpi}^{-1/2},\qquad\psi=M_{\varpi}^{1/2}q. H ϖ = M ϖ − 1/2 C M ϖ − 1/2 , ψ = M ϖ 1/2 q .
Our spectral results concern this reduced problem and its finite-cluster analogue. Transfer to the full transmission field requires an approximation uniform in the number of resonators.
For the infinite crystal, the variational cell problem gives an analytic symbol and exponentially decaying coefficients [9 , 12 ] . Under coordinate reflection symmetry, the symbol c c c is strictly increasing in each coordinate on ( 0 , π ) (0,\pi) ( 0 , π ) for arbitrary transverse phases. Its critical points are the vertices { 0 , π } d \{0,\pi\}^d { 0 , π } d , and its unique maximum c M = c ( α M ) c_M=c(\alpha_M) c M = c ( α M ) , with α M = ( π , … , π ) \alpha_M=(\pi,\ldots,\pi) α M = ( π , … , π ) , is nondegenerate. For every fixed $W>0, {xref}Theorem 6.1` proves spectral and dynamical localisation on I W = ( c M − Δ W , c M ) I_W=(c_M-\Delta_W,c_M) I W = ( c M − Δ W , c M ) for some Δ W ∈ ( 0 , c M ) \Delta_W\in(0,c_M) Δ W ∈ ( 0 , c M ) . Theorem 4.1 and Corollary 6.4 give the same conclusion on prescribed positive intervals under an explicit large-disorder condition. For each fixed 0 < a < 1 / 2 0<a<1/2 0 < a < 1/2 and sufficiently small W W W , Theorem 7.1 gives the explicit interval
I W , a = ( c M 1 + a W , c M ) . I_{W,a}=\left(\frac{c_M}{1+aW},c_M\right). I W , a = ( 1 + aW c M , c M ) .
This bounds the upper endpoint of any absolutely continuous spectral component, without establishing an extended interior regime.
For a finite set of retained indices Λ \Lambda Λ , we distinguish the truncated capacitance matrix C Λ t = ( C n m ) n , m ∈ Λ C^t_\Lambda=(C_{nm})_{n,m\in\Lambda} C Λ t = ( C nm ) n , m ∈ Λ from the finite capacitance matrix C Λ f C^f_\Lambda C Λ f , defined for the retained resonators after all others are removed. With M ϖ , Λ = diag ( m n : n ∈ Λ ) M_{\varpi,\Lambda}=\operatorname{diag}(m_n:n\in\Lambda) M ϖ , Λ = diag ( m n : n ∈ Λ ) , their self-adjoint representations are H ϖ , Λ a = M ϖ , Λ − 1 / 2 C Λ a M ϖ , Λ − 1 / 2 H^a_{\varpi,\Lambda}=M_{\varpi,\Lambda}^{-1/2}C^a_\Lambda M_{\varpi,\Lambda}^{-1/2} H ϖ , Λ a = M ϖ , Λ − 1/2 C Λ a M ϖ , Λ − 1/2 , a ∈ { t , f } a\in\{t,f\} a ∈ { t , f } . The asymptotic-equivalence method of [7 ] extends to this diagonal weighting. The two matrix sequences have the same limiting density of states; we also obtain a surface-order trace comparison and convergence of their spectra.
To control eigenvectors, we combine the positive boundary estimate of [9 Lemma 5.1] with uniform localisation for truncated matrices and an eigenvalue-spacing estimate for material disorder. The boundary correction is exponentially small on vectors supported far from the boundary, while the spacing estimate separates the nearby finite eigenvalues. These bounds match modes of the truncated matrix centred in the interior to individual eigenvectors of the finite matrix.
More precisely, let Ω ⊂ R d \Omega\subset\mathbb{R}^d Ω ⊂ R d be a bounded Lipschitz sample containing N ε ≍ ε − d N_\varepsilon\asymp\varepsilon^{-d} N ε ≍ ε − d cells. For every fixed p > 0 p>0 p > 0 , with probability at least 1 − C W , p N − p 1-C_{W,p}N^{-p} 1 − C W , p N − p , all but O W , p ( ε 1 − d log ( 1 / ε ) ) O_{W,p}(\varepsilon^{1-d}\log(1/\varepsilon)) O W , p ( ε 1 − d log ( 1/ ε )) modes in the localised interval concentrate in balls of radius O W , p ( ε log ( 1 / ε ) ) O_{W,p}(\varepsilon\log(1/\varepsilon)) O W , p ( ε log ( 1/ ε )) , with squared tail at most N − p N^{-p} N − p ; see Theorem 15.3 . The retained modes are close to exponentially localised modes of the truncated matrix. The unmatched family is controlled in cardinality, but its elements are not necessarily boundary-supported. The estimates are uniform for the contrast-dependent family C δ C_\delta C δ defined in § 2.10 . They apply at δ = ε 2 \delta=\varepsilon^2 δ = ε 2 on an energy interval of fixed disorder-dependent width below its upper edge c M , δ c_{M,\delta} c M , δ . Under δ = ε 2 \delta=\varepsilon^2 δ = ε 2 , its frequency width also has a positive limit. This differs from the O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) frequency detuning, and hence O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) reduced-energy detuning, used in deterministic band-edge homogenisation [9 ] .
At regular interior energies, deterministic estimates give lower bounds of order W − 1 W^{-1} W − 1 on eigenfunction extent, compatible with localisation on larger scales. For the unperturbed crystal, [9 Appendix B] derives a formal local transport equation for a selected regular Bloch mode; one-dimensional finite-sample resonance effects are analysed in [9 Appendix A.5] . Boundary matching is needed to obtain a finite-sample scattering model. This deterministic modulation description does not determine the spectral type of the randomly perturbed operator.
§ 5.8 treats square arrays of L 2 L^2 L 2 disks for L = 36 , 51 , 70 L=36,51,70 L = 36 , 51 , 70 , using finite capacitance matrices and 500 realisations at each positive disorder strength. Centred intensity profiles and spectral participation show concentration near the top of the spectrum and, at stronger disorder, in the interior of the unperturbed band. The selected disordered modes occupy a decreasing fraction of the sample as L L L increases.
The paper is organised into three parts. Part I studies the infinite operator, its symbol, and localisation. Part II establishes the estimates for truncated matrices and eigenvalue spacing needed for the finite-cluster problem. Part III treats finite capacitance matrices, spectral convergence, and numerical illustrations. Appendix A discusses the three-dimensional weak-scattering prediction, and Appendix B extends the band-edge absence result to other material distributions.
Part I. The infinite crystal
The random capacitance operator
We use the Euclidean norm ∣ x ∣ = ∣ x ∣ 2 |x|=|x|_2 ∣ x ∣ = ∣ x ∣ 2 , the lattice norms ∣ x ∣ 1 = ∑ j ∣ x j ∣ |x|_1=\sum_j|x_j| ∣ x ∣ 1 = ∑ j ∣ x j ∣ and ∣ x ∣ ∞ = max j ∣ x j ∣ |x|_\infty=\max_j|x_j| ∣ x ∣ ∞ = max j ∣ x j ∣ , and the standard coordinate vectors e j ∈ R d e_j\in\mathbb{R}^d e j ∈ R d . The vector δ n \delta_n δ n is the coordinate vector at n n n in ℓ 2 ( Z d ) \ell^2(\mathbb{Z}^d) ℓ 2 ( Z d ) . Inner products are conjugate-linear in the first argument, T ∗ T^* T ∗ denotes the adjoint, and [ S , T ] = S T − T S [S,T]=ST-TS [ S , T ] = S T − T S . We write Id \operatorname{Id} Id for the identity, 1 A \mathbf{1}_A 1 A for an indicator or its multiplication operator, and r + = max { r , 0 } r_+=\max\{r,0\} r + = max { r , 0 } . For a self-adjoint operator H H H , 1 I ( H ) \mathbf{1}_I(H) 1 I ( H ) is its spectral projection onto an interval I I I . Throughout, the operator norm refers to the induced norm on the corresponding Hilbert space, and we will omit the subscript when it is clear from the context. In particular, the Hilbert–Schmidt norm satisfies ∥ T ∥ H S 2 = Tr ( T ∗ T ) \lVert T\rVert_{\mathrm{HS}}^2=\operatorname{Tr}(T^*T) ∥ T ∥ HS 2 = Tr ( T ∗ T ) . Positive constants in estimates may change from line to line; their parameter dependence is stated when uniformity is required.
Geometry and capacitance normalisation
Recall that d ∈ { 2 , 3 } d\in\{2,3\} d ∈ { 2 , 3 } , Y = ( − 1 / 2 , 1 / 2 ) d Y=(-1/2,1/2)^d Y = ( − 1/2 , 1/2 ) d , and denote by D = ⋃ n ∈ Z d ( D + n ) \mathcal{D}=\bigcup_{n\in\mathbb{Z}^d}(D+n) D = ⋃ n ∈ Z d ( D + n ) . Assume that D ⋐ Y D\Subset Y D ⋐ Y is simply-connected with C 2 C^2 C 2 boundary. We also assume that Y ∖ D ‾ Y\setminus\overline{D} Y ∖ D is connected. The reflection symmetry is described as follows, which will be imposed in particular results of this paper, e.g., the band-edge localisation,
(R) R j D = D ( 1 ≤ j ≤ d ) , R j ( x 1 , … , x d ) = ( x 1 , … , − x j , … , x d ) R_jD=D\quad(1\le j\le d),\qquad R_j(x_1,\ldots,x_d)=(x_1,\ldots,-x_j,\ldots,x_d) R j D = D ( 1 ≤ j ≤ d ) , R j ( x 1 , … , x d ) = ( x 1 , … , − x j , … , x d ) .
The geometry is illustrated in Figure 1 and Figure 2 . Corresponding to the lattice structure, the Brillouin torus associated with the Floquet transform is Y ∗ = R d / ( 2 π Z ) d Y^*=\mathbb{R}^d/(2\pi\mathbb{Z})^d Y ∗ = R d / ( 2 π Z ) d , represented by [ − π , π ] d [-\pi,\pi]^d [ − π , π ] d with opposite faces identified.
Figure 1. Left: The reference cell. Right: A simple cubic sphere crystal, shown schematically. The orange box marks the reference cell in the repeated structure. The model is periodic in all three directions.
Figure 2. Left: a two-dimensional section illustrating the coordinate reflections and the opposite-face trace convention, with F j ± = { x ∈ Y ‾ : x j = ± 1 / 2 } F_j^{\pm}=\{x\in\overline{Y}:x_j=\pm1/2\} F j ± = { x ∈ Y : x j = ± 1/2 } . In three dimensions all three coordinate reflections are required. Right: the positive quadrant of the two-dimensional Brillouin torus, with Γ = ( 0 , 0 ) \Gamma=(0,0) Γ = ( 0 , 0 ) , X = ( π , 0 ) X=(\pi,0) X = ( π , 0 ) , Y = ( 0 , π ) Y=(0,\pi) Y = ( 0 , π ) , and M = ( π , π ) M=(\pi,\pi) M = ( π , π ) . All four vertices are critical; only M M M is the maximum. The arrows indicate coordinate monotonicity.
For α ∈ Y ∗ \alpha\in Y^* α ∈ Y ∗ , let H α 1 ( Y ) H^1_\alpha(Y) H α 1 ( Y ) denote the Sobolev functions in H 1 ( Y ) H^1(Y) H 1 ( Y ) whose face traces satisfy γ j + u = e i α j γ j − u \gamma_j^+u=e^{i\alpha_j}\gamma_j^-u γ j + u = e i α j γ j − u , where γ j ± \gamma_j^\pm γ j ± is the trace on x j = ± 1 / 2 x_j=\pm1/2 x j = ± 1/2 . Define
c ( α ) = C ^ α = 1 ∣ D ∣ min u ∈ H α 1 ( Y ) u = 1 i n D ∫ Y ∣ ∇ u ∣ 2 . (4) c(\alpha)=\widehat{C}^{\alpha}=\frac{1}{|D|}\min_{\substack{u\in H^1_\alpha(Y)\\u=1\ \mathrm{in}\ D}}\int_Y|\nabla u|^2. \tag*{(4)} c ( α ) = C α = ∣ D ∣ 1 u ∈ H α 1 ( Y ) u = 1 in D min ∫ Y ∣∇ u ∣ 2 . (4)
Section 3 proves the existence and uniqueness of the minimiser, together with its operator interpretation. With the discrete Floquet convention of [8 Section 4.8] , set
q ^ ( α ) = ∑ n ∈ Z d q n e i n ⋅ α , c r = 1 ( 2 π ) d ∫ Y ∗ c ( α ) e − i r ⋅ α d α , ( C q ) n = ∑ m c n − m q m . (5) \widehat{q}(\alpha)=\sum_{n\in\mathbb{Z}^d}q_ne^{in\cdot\alpha},\qquad c_r=\frac{1}{(2\pi)^d}\int_{Y^*}c(\alpha)e^{-ir\cdot\alpha}\,d\alpha,\qquad(Cq)_n=\sum_m c_{n-m}q_m. \tag*{(5)} q ( α ) = n ∈ Z d ∑ q n e in ⋅ α , c r = ( 2 π ) d 1 ∫ Y ∗ c ( α ) e − i r ⋅ α d α , ( C q ) n = m ∑ c n − m q m . (5)
For q ∈ ℓ 2 ( Z d ) q\in\ell^2(\mathbb{Z}^d) q ∈ ℓ 2 ( Z d ) , the Fourier series is understood in L 2 ( Y ∗ ) L^2(Y^*) L 2 ( Y ∗ ) . We use q n q_n q n and q ( n ) q(n) q ( n ) interchangeably. With these notions, C n m = c n − m C_{nm}=c_{n-m} C nm = c n − m and C q ^ = c q ^ \widehat{Cq}=c\widehat{q} C q = c q . Write c M = max Y ∗ c > 0 c_M=\max_{Y^*}c>0 c M = max Y ∗ c > 0 . Under (R), c M = c ( α M ) c_M=c(\alpha_M) c M = c ( α M ) with α M = ( π , … , π ) \alpha_M=(\pi,\ldots,\pi) α M = ( π , … , π ) . The geometric capacity of [8 ] is Cap α , D = ∣ D ∣ c ( α ) \operatorname{Cap}_{\alpha,D}=|D|c(\alpha) Cap α , D = ∣ D ∣ c ( α ) . Its generalised one-resonator symbol is δ v r 2 c ( α ) \delta v_r^2c(\alpha) δ v r 2 c ( α ) . The factor δ v r 2 \delta v_r^2 δ v r 2 has been removed in the unit-period spectral variable E E E of (2) . For a physical period ε > 0 \varepsilon>0 ε > 0 and contrast δ > 0 \delta>0 δ > 0 , the conversion between reduced spectral energy and frequency is
E = ε 2 ω 2 δ v r 2 , ω = ω ε , δ ( E ) : = v r δ ε E . (6) E=\frac{\varepsilon^2\omega^2}{\delta v_r^2},\qquad\omega=\omega_{\varepsilon,\delta}(E):=\frac{v_r\sqrt{\delta}}{\varepsilon}\sqrt{E}. \tag*{(6)} E = δ v r 2 ε 2 ω 2 , ω = ω ε , δ ( E ) := ε v r δ E . (6)
Thus δ = ε 2 \delta=\varepsilon^2 δ = ε 2 gives E = ω 2 / v r 2 E=\omega^2/v_r^2 E = ω 2 / v r 2 . This convention is used for the shrinking-cell limit; formulas retaining both ε \varepsilon ε and δ \delta δ allow other scalings.
For the static reduced model, set
ω M : = ω ε , δ ( c M ) = v r δ c M ε , E c M = ( ω ω M ) 2 . (7) \omega_M:=\omega_{\varepsilon,\delta}(c_M)=\frac{v_r\sqrt{\delta c_M}}{\varepsilon},\qquad\frac{E}{c_M}=\left(\frac{\omega}{\omega_M}\right)^2. \tag*{(7)} ω M := ω ε , δ ( c M ) = ε v r δ c M , c M E = ( ω M ω ) 2 . (7)
Here ω M \omega_M ω M is the reduced edge frequency. By rescaling (3) , the full periodic Bloch frequency at M M M differs from it by O ( ε − 1 δ 3 / 2 ) O(\varepsilon^{-1}\delta^{3/2}) O ( ε − 1 δ 3/2 ) , which is O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) when δ = ε 2 \delta=\varepsilon^2 δ = ε 2 . The frequency conversions below are exact for the reduced model; their interpretation for the transmission problem requires a uniform reduction estimate.
Random material factors
Random material factors. The main operator family under consideration has independent variables ξ n \xi_n ξ n , uniform on [ 0 , 1 ] [0,1] [ 0 , 1 ] , and, for each fixed W > 0 W>0 W > 0 ,
m n = 1 + W ξ n , M ϖ = diag ( m n ) , B ϖ = M ϖ − 1 , A ϖ = B ϖ C , H ϖ = M ϖ − 1 / 2 C M ϖ − 1 / 2 . (8) m_n=1+W\xi_n,\qquad M_\varpi=\operatorname{diag}(m_n),\qquad B_\varpi=M_\varpi^{-1},\qquad A_\varpi=B_\varpi C,\qquad H_\varpi=M_\varpi^{-1/2}CM_\varpi^{-1/2}. \tag*{(8)} m n = 1 + W ξ n , M ϖ = diag ( m n ) , B ϖ = M ϖ − 1 , A ϖ = B ϖ C , H ϖ = M ϖ − 1/2 C M ϖ − 1/2 . (8)
In the sequel, P \mathbb{P} P and E \mathbb{E} E denote the probability and expectation with respect to the product measure generated by ξ n \xi_n ξ n , respectively, and μ W : = E m n = 1 + W / 2 \mu_W:=\mathbb{E}m_n=1+W/2 μ W := E m n = 1 + W /2 is the mean inverse material factor. Uniform inverse factors m n m_n m n correspond to direct factors b n = m n − 1 b_n=m_n^{-1} b n = m n − 1 with density ( W b 2 ) − 1 1 [ ( 1 + W ) − 1 , 1 ] ( b ) (Wb^2)^{-1}\mathbf{1}_{[(1+W)^{-1},1]}(b) ( W b 2 ) − 1 1 [( 1 + W ) − 1 , 1 ] ( b ) . Appendix B also treats bounded densities and uniform direct factors.
The operator A ϖ A_\varpi A ϖ is self-adjoint on H ϖ = ℓ 2 ( Z d ; m ϖ ) \mathcal{H}_\varpi=\ell^2(\mathbb{Z}^d;m_\varpi) H ϖ = ℓ 2 ( Z d ; m ϖ ) , with norm ∥ q ∥ ϖ 2 = ∑ n m n ∣ q n ∣ 2 \lVert q\rVert_\varpi^2=\sum_n m_n|q_n|^2 ∥ q ∥ ϖ 2 = ∑ n m n ∣ q n ∣ 2 . The map U ϖ : H ϖ → ℓ 2 ( Z d ) U_\varpi:\mathcal{H}_\varpi\to\ell^2(\mathbb{Z}^d) U ϖ : H ϖ → ℓ 2 ( Z d ) given by
U ϖ q = M ϖ 1 / 2 q , U ϖ A ϖ U ϖ − 1 = H ϖ (9) U_\varpi q=M_\varpi^{1/2}q,\qquad U_\varpi A_\varpi U_\varpi^{-1}=H_\varpi \tag*{(9)} U ϖ q = M ϖ 1/2 q , U ϖ A ϖ U ϖ − 1 = H ϖ (9)
is unitary. We formulate spectral statements for H ϖ H_\varpi H ϖ . This convention is equivalent to the B C BC B C model in [8 Section 4.8] . If a model is written as V C V VCV V C V , as in [11 ] , its diagonal V V V corresponds to our B ϖ 1 / 2 B_\varpi^{1/2} B ϖ 1/2 .
For finite Λ ⊂ Z d \Lambda\subset\mathbb{Z}^d Λ ⊂ Z d , let E Λ : ℓ 2 ( Λ ) → ℓ 2 ( Z d ) E_\Lambda:\ell^2(\Lambda)\to\ell^2(\mathbb{Z}^d) E Λ : ℓ 2 ( Λ ) → ℓ 2 ( Z d ) be the extension by zero; E Λ ∗ E_\Lambda^* E Λ ∗ is the restriction. Write P Λ = E Λ E Λ ∗ P_\Lambda=E_\Lambda E_\Lambda^* P Λ = E Λ E Λ ∗ and M ϖ , Λ = E Λ ∗ M ϖ E Λ M_{\varpi,\Lambda}=E_\Lambda^*M_\varpi E_\Lambda M ϖ , Λ = E Λ ∗ M ϖ E Λ . Put
C Λ t = E Λ ∗ C E Λ , H ϖ , Λ t = M ϖ , Λ − 1 / 2 C Λ t M ϖ , Λ − 1 / 2 , Λ L = [ − L , L ] d ∩ Z d , L ∈ N . (10) C_\Lambda^t=E_\Lambda^*CE_\Lambda,\qquad H_{\varpi,\Lambda}^t=M_{\varpi,\Lambda}^{-1/2}C_\Lambda^tM_{\varpi,\Lambda}^{-1/2},\qquad\Lambda_L=[-L,L]^d\cap\mathbb{Z}^d,\qquad L\in\mathbb{N}. \tag*{(10)} C Λ t = E Λ ∗ C E Λ , H ϖ , Λ t = M ϖ , Λ − 1/2 C Λ t M ϖ , Λ − 1/2 , Λ L = [ − L , L ] d ∩ Z d , L ∈ N . (10)
We call C Λ t C_\Lambda^t C Λ t the truncated capacitance matrix. The superscript f f f denotes the finite capacitance matrix C Λ f C_\Lambda^f C Λ f of the retained resonators, defined variationally in § 5.1 . The superscripts on H H H indicate the corresponding random matrices. For centred cubes, write N L = ∣ Λ L ∣ = ( 2 L + 1 ) d N_L=|\Lambda_L|=(2L+1)^d N L = ∣ Λ L ∣ = ( 2 L + 1 ) d and P L = P Λ L P_L=P_{\Lambda_L} P L = P Λ L .
Spectral measures and correlators
Spectral measures and correlators. For a finite set Λ \Lambda Λ and a Borel interval I ⊂ R I\subset\mathbb{R} I ⊂ R , define
Q ϖ , Λ t ( x , y ; I ) = sup f : R → C B o r e l ∥ f ∥ ∞ ≤ 1 ∣ ⟨ δ x , f ( H ϖ , Λ t ) 1 I ( H ϖ , Λ t ) δ y ⟩ ∣ . (11) Q_{\varpi,\Lambda}^t(x,y;I)=\sup_{\substack{f:\mathbb{R}\to\mathbb{C}\ \mathrm{Borel}\\ \lVert f\rVert_\infty\le1}}\left|\left\langle\delta_x,f(H_{\varpi,\Lambda}^t)\mathbf{1}_I(H_{\varpi,\Lambda}^t)\delta_y\right\rangle\right|. \tag*{(11)} Q ϖ , Λ t ( x , y ; I ) = f : R → C Borel ∥ f ∥ ∞ ≤ 1 sup ⟨ δ x , f ( H ϖ , Λ t ) 1 I ( H ϖ , Λ t ) δ y ⟩ . (11)
Table 1. Principal notation. Spectral energies are denoted by E E E or by the centred variable s = μ W E s=\mu_W E s = μ W E . Fractional-moment exponents are denoted by ϑ \vartheta ϑ .
Write Q ϖ Q_{\varpi} Q ϖ for the infinite-volume analogue. These correlators are the total variations of the spectral measures on I I I and are bounded by one. On an open interval they can equivalently be defined using a countable dense family in the unit ball of C 0 ( I ) C_0(I) C 0 ( I ) , which also proves measurability.
The almost-sure spectrum
We determine the almost-sure spectrum, thereby fixing the spectral edges used in the localisation estimates below.
Proposition 2.2. For every fixed W > 0 W>0 W > 0 , σ ( H ϖ ) = [ 0 , c M ] \sigma(H_{\varpi})=[0,c_M] σ ( H ϖ ) = [ 0 , c M ] almost surely.
Proof. Since 0 ≤ C ≤ c M Id 0\le C\le c_M\operatorname{Id} 0 ≤ C ≤ c M Id and 0 < B ϖ ≤ Id 0<B_{\varpi}\le\operatorname{Id} 0 < B ϖ ≤ Id , 0 ≤ H ϖ ≤ c M Id 0\le H_{\varpi}\le c_M\operatorname{Id} 0 ≤ H ϖ ≤ c M Id . Thus, σ ( H ϖ ) ⊂ [ 0 , c M ] \sigma(H_{\varpi})\subset[0,c_M] σ ( H ϖ ) ⊂ [ 0 , c M ] for every configuration ϖ \varpi ϖ . Conversely, by Weyl’s criterion for the spectrum, to prove that [ 0 , c M ] ⊂ σ ( H ϖ ) [0,c_M]\subset\sigma(H_{\varpi}) [ 0 , c M ] ⊂ σ ( H ϖ ) , it remains to show that, for every E ∈ [ 0 , c M ] E\in[0,c_M] E ∈ [ 0 , c M ] , there are unit vectors u j u_j u j satisfying
∥ ( H ϖ − E ) u j ∥ → 0 as j → ∞ . (12) \lVert(H_{\varpi}-E)u_j\rVert\to0 \qquad\text{as }j\to\infty. \tag*{(12)} ∥( H ϖ − E ) u j ∥ → 0 as j → ∞. (12)
The remainder of the proof is divided into several steps.
Step 1: construct the approximate eigenvectors for C C C . Since c ( Y ∗ ) = [ 0 , c M ] c(Y^*) = [0,c_{\mathrm{M}}] c ( Y ∗ ) = [ 0 , c M ] , for every E ∈ [ 0 , c M ] E \in[0,c_{\mathrm{M}}] E ∈ [ 0 , c M ] , there is an α E \alpha_E α E such that c ( α E ) = E c(\alpha_E) = E c ( α E ) = E . Concentrate the Fourier transform of a unit vector g j g_j g j near α E \alpha_E α E such that
max α ∈ supp g j ^ ∣ c ( α ) − E ∣ → 0 as j → ∞ . (13) \max_{\alpha\in\operatorname{supp}\widehat{g_j}} |c(\alpha)-E| \to0 \qquad\text{as } j\to\infty. \tag*{(13)} α ∈ supp g j max ∣ c ( α ) − E ∣ → 0 as j → ∞. (13)
Since the symbol of C C C is c c c , we have
∥ ( C − E ) g j ∥ 2 = 1 ( 2 π ) d ∫ Y ∗ ∣ c ( α ) − E ∣ 2 ∣ g j ^ ( α ) ∣ 2 d α → 0 as j → ∞ . (14) \lVert(C-E)g_j\rVert^2 = \frac{1}{(2\pi)^d}\int_{Y^*}|c(\alpha)-E|^2|\widehat{g_j}(\alpha)|^2\,\mathrm{d}\alpha\to0 \qquad\text{as } j\to\infty. \tag*{(14)} ∥( C − E ) g j ∥ 2 = ( 2 π ) d 1 ∫ Y ∗ ∣ c ( α ) − E ∣ 2 ∣ g j ( α ) ∣ 2 d α → 0 as j → ∞. (14)
Although these g j g_j g j need not have finite support, finitely supported vectors are dense in ℓ 2 ( Z d ) \ell^2(\mathbb{Z}^d) ℓ 2 ( Z d ) , and C − E C-E C − E is bounded. By truncating and normalising, we obtain finitely supported unit vectors f j f_j f j such that
∥ ( C − E ) f j ∥ → 0 as j → ∞ . (15) \lVert(C-E)f_j\rVert\to0 \qquad\text{as } j\to\infty. \tag*{(15)} ∥( C − E ) f j ∥ → 0 as j → ∞. (15)
Since C f j ∈ ℓ 2 ( Z d ) Cf_j\in\ell^2(\mathbb{Z}^d) C f j ∈ ℓ 2 ( Z d ) , we may choose a finite set B j B_j B j , containing supp f j \operatorname{supp}f_j supp f j , large enough such that ∥ 1 B j c C f j ∥ → 0 \lVert1_{B_j^c}Cf_j\rVert\to0 ∥ 1 B j c C f j ∥ → 0 as j → ∞ j\to\infty j → ∞ .
Step 2: construct the good event. Choose a countable dense subset E ⊂ [ 0 , c M ] \mathcal{E}\subset[0,c_{\mathrm{M}}] E ⊂ [ 0 , c M ] . For any E ∈ E E\in\mathcal{E} E ∈ E , by Step 1, after passing to a subsequence, there is a sequence { ε j } j ∈ N \{\varepsilon_j\}_{j\in\mathbb{N}} { ε j } j ∈ N converging to zero such that
∥ ( C − E ) f j ∥ + ∥ 1 B j c C f j ∥ ≤ ε j , j ∈ N . (16) \lVert(C-E)f_j\rVert+\lVert1_{B_j^c}Cf_j\rVert\leq\varepsilon_j,\qquad j\in\mathbb{N}. \tag*{(16)} ∥( C − E ) f j ∥ + ∥ 1 B j c C f j ∥ ≤ ε j , j ∈ N . (16)
For each j ∈ N j\in\mathbb{N} j ∈ N , since B j B_j B j is finite, there is a sequence { z j k } k ∈ N ⊂ Z d \{z_j^k\}_{k\in\mathbb{N}}\subset\mathbb{Z}^d { z j k } k ∈ N ⊂ Z d such that the translated sets
B j k : = B j + z j k (17) B_j^k:=B_j+z_j^k \tag*{(17)} B j k := B j + z j k (17)
are mutually disjoint. Define the event
A j k : = ⋂ n ∈ B j k { ϖ : ξ n ( ϖ ) ≤ a j } , a j : = min { 1 2 , 2 ε j W } ∈ ( 0 , 1 ) . (18) A_j^k:=\bigcap_{n\in B_j^k}\{\varpi:\xi_n(\varpi)\leq a_j\},\qquad a_j:=\min\left\{\frac{1}{2},\frac{2\varepsilon_j}{W}\right\}\in(0,1). \tag*{(18)} A j k := n ∈ B j k ⋂ { ϖ : ξ n ( ϖ ) ≤ a j } , a j := min { 2 1 , W 2 ε j } ∈ ( 0 , 1 ) . (18)
Denote M ϖ − 1 / 2 M_\varpi^{-1/2} M ϖ − 1/2 by D ϖ D_\varpi D ϖ . By the choice of a j a_j a j , we see ξ n ( ϖ ) ≤ a j \xi_n(\varpi)\leq a_j ξ n ( ϖ ) ≤ a j implies ∣ D ϖ ( n , n ) − 1 ∣ ≤ ε j \lvert D_\varpi(n,n)-1\rvert\leq\varepsilon_j ∣ D ϖ ( n , n ) − 1 ∣ ≤ ε j . Since { ξ n } \{\xi_n\} { ξ n } are independent and uniformly distributed in [ 0 , 1 ] [0,1] [ 0 , 1 ] ,
P [ A j k ] = a j ∣ B j ∣ > 0. (19) \mathbb{P}[A_j^k]=a_j^{|B_j|}>0. \tag*{(19)} P [ A j k ] = a j ∣ B j ∣ > 0. (19)
For fixed j j j , the events { A j k } \{A_j^k\} { A j k } are independent. Therefore,
P [ ⋂ k = 1 ∞ ( A j k ) c ] = lim K → ∞ ( 1 − a j ∣ B j ∣ ) K = 0. (20) \mathbb{P}\left[\bigcap_{k=1}^{\infty}(A_j^k)^c\right]=\lim_{K\to\infty}(1-a_j^{|B_j|})^K=0. \tag*{(20)} P [ k = 1 ⋂ ∞ ( A j k ) c ] = K → ∞ lim ( 1 − a j ∣ B j ∣ ) K = 0. (20)
This implies that the event
G : = ⋂ E ∈ E ⋂ j = 1 ∞ ⋃ k = 1 ∞ A j k (21) \mathcal{G}:=\bigcap_{E\in\mathcal{E}}\bigcap_{j=1}^{\infty}\bigcup_{k=1}^{\infty}A_j^k \tag*{(21)} G := E ∈ E ⋂ j = 1 ⋂ ∞ k = 1 ⋃ ∞ A j k (21)
has probability one.
Step 3: construct the approximate eigenvectors for H ϖ H_\varpi H ϖ . Fix ϖ ∈ G \varpi\in\mathcal{G} ϖ ∈ G and E ∈ E E\in\mathcal{E} E ∈ E . For each j j j , choose an index k j k_j k j such that ϖ ∈ A j k j \varpi\in A_j^{k_j} ϖ ∈ A j k j . Set u j u_j u j by
u j ( n ) : = f j ( n − z j k j ) , n ∈ Z d . (22) u_j(n):=f_j(n-z_j^{k_j}),\qquad n\in\mathbb{Z}^d. \tag*{(22)} u j ( n ) := f j ( n − z j k j ) , n ∈ Z d . (22)
Then, u j u_j u j is still a unit vector and supported on B j k j B_j^{k_j} B j k j . Translation invariance and (16) give
∥ ( C − E ) u j ∥ + ∥ 1 ( B j k j ) c C u j ∥ ≤ ε j , j ∈ N . (23) \lVert(C-E)u_j\rVert+\lVert\mathbf{1}_{(B_j^{k_j})^c}Cu_j\rVert\leq\varepsilon_j,\qquad j\in\mathbb{N}. \tag*{(23)} ∥( C − E ) u j ∥ + ∥ 1 ( B j k j ) c C u j ∥ ≤ ε j , j ∈ N . (23)
Furthermore, since ϖ ∈ A j k j \varpi\in\mathcal{A}_j^{k_j} ϖ ∈ A j k j ,
∣ D ϖ ( n , n ) − 1 ∣ ≤ ε j , n ∈ B j k j . (24) \lvert D_{\varpi}(n,n)-1\rvert\leq\varepsilon_j,\qquad n\in B_j^{k_j}. \tag*{(24)} ∣ D ϖ ( n , n ) − 1 ∣ ≤ ε j , n ∈ B j k j . (24)
We rewrite
( H ϖ − E ) u j = ( C − E ) u j + D ϖ C ( D ϖ − I d ) u j + ( D ϖ − I d ) 1 B j k j C u j + ( D ϖ − I d ) 1 ( B j k j ) c C u j . (25) \begin{aligned}
(H_{\varpi}-E)u_j={}&(C-E)u_j+D_{\varpi}C(D_{\varpi}-\mathrm{Id})u_j\\
&+(D_{\varpi}-\mathrm{Id})\mathbf{1}_{B_j^{k_j}}Cu_j+(D_{\varpi}-\mathrm{Id})\mathbf{1}_{(B_j^{k_j})^c}Cu_j. \tag*{(25)}
\end{aligned} ( H ϖ − E ) u j = ( C − E ) u j + D ϖ C ( D ϖ − Id ) u j + ( D ϖ − Id ) 1 B j k j C u j + ( D ϖ − Id ) 1 ( B j k j ) c C u j . (25)
For the first and last terms in the right-hand side of (25) , we use (2.20) and $\lVert D_{\varpi}-\mathrm{Id}\rVert\leq 1$; for the second and third terms we use $\operatorname{supp}u_j\subset B_j^{k_j}$ and (2.21)`. As a result, we get ∥ ( H ϖ − E ) u j ∥ → 0 \lVert(H_{\varpi}-E)u_j\rVert\to0 ∥( H ϖ − E ) u j ∥ → 0 as j → ∞ j\to\infty j → ∞ . By Weyl’s criterion, E ∈ σ ( H ϖ ) E\in\sigma(H_{\varpi}) E ∈ σ ( H ϖ ) . Since E ∈ E E\in\mathcal{E} E ∈ E is arbitrary, and the spectrum σ ( H ϖ ) \sigma(H_{\varpi}) σ ( H ϖ ) is closed, we obtain [ 0 , c M ] = E ‾ ⊂ σ ( H ϖ ) [0,c_M]=\overline{\mathcal{E}}\subset\sigma(H_{\varpi}) [ 0 , c M ] = E ⊂ σ ( H ϖ ) . The proof is complete. □
Deterministic properties of the capacitance operator
We establish the deterministic properties used in the localisation arguments: analyticity of the Bloch symbol, exponential decay of the capacitance coefficients, and, under coordinate reflection symmetry, uniqueness and nondegeneracy of the maximum of the symbol. Write O = Y ∖ D ‾ \mathcal{O}=Y\setminus\overline{D} O = Y ∖ D for the connected cell exterior.
A cell Poincaré inequality
We begin with a Poincaré inequality that gives coercivity of the cell energy uniformly in the Bloch phase. We use complex-valued Sobolev spaces. Put
C P = 1 + ∣ D ∣ − 1 π . (26) C_P=\frac{\sqrt{1+\lvert D\rvert^{-1}}}{\pi}. \tag*{(26)} C P = π 1 + ∣ D ∣ − 1 . (26)
Lemma 3.1. Every h ∈ H 1 ( Y ) h\in H^1(Y) h ∈ H 1 ( Y ) that vanishes almost everywhere in D D D satisfies
∥ h ∥ L 2 ( Y ) ≤ C P ∥ ∇ h ∥ L 2 ( Y ) . (27) \lVert h\rVert_{L^2(Y)}\leq C_P\lVert\nabla h\rVert_{L^2(Y)}. \tag*{(27)} ∥ h ∥ L 2 ( Y ) ≤ C P ∥ ∇ h ∥ L 2 ( Y ) . (27) No boundary condition on ∂ Y \partial Y ∂ Y is required.
Proof. Let h ˉ = ∫ Y h \bar{h}=\int_Y h h ˉ = ∫ Y h , since ∣ Y ∣ = 1 \lvert Y\rvert=1 ∣ Y ∣ = 1 . The Neumann Poincaré inequality on the unit cube is ∥ h − h ˉ ∥ 2 ≤ π − 1 ∥ ∇ h ∥ 2 \lVert h-\bar{h}\rVert_2\leq\pi^{-1}\lVert\nabla h\rVert_2 ∥ h − h ˉ ∥ 2 ≤ π − 1 ∥ ∇ h ∥ 2 ; it follows, for example, from the cosine expansion and the first nonzero Neumann eigenvalue π 2 \pi^2 π 2 . Since h = 0 h=0 h = 0 in D D D , ∣ D ∣ ∣ h ˉ ∣ 2 ≤ ∥ h − h ˉ ∥ 2 2 \lvert D\rvert\lvert\bar{h}\rvert^2\leq\lVert h-\bar{h}\rVert_2^2 ∣ D ∣ ∣ h ˉ ∣ 2 ≤ ∥ h − h ˉ ∥ 2 2 . Therefore
∥ h ∥ 2 2 = ∥ h − h ˉ ∥ 2 2 + ∣ h ˉ ∣ 2 ≤ ( 1 + ∣ D ∣ − 1 ) π − 2 ∥ ∇ h ∥ 2 2 . \lVert h\rVert_2^2=\lVert h-\bar{h}\rVert_2^2+\lvert\bar{h}\rvert^2\leq(1+\lvert D\rvert^{-1})\pi^{-2}\lVert\nabla h\rVert_2^2. ∥ h ∥ 2 2 = ∥ h − h ˉ ∥ 2 2 + ∣ h ˉ ∣ 2 ≤ ( 1 + ∣ D ∣ − 1 ) π − 2 ∥ ∇ h ∥ 2 2 . The same argument applies to complex functions. □
Bloch cell problem and the infinite operator
Using the trace convention of § 2.1 , set
U α = { u ∈ H 1 ( Y ; C ) : u = 1 in D , γ j + u = e i α j γ j − u ( 1 ≤ j ≤ d ) } . (28) \mathcal{U}_{\alpha}=\{u\in H^1(Y;\mathbb{C}):u=1\text{ in }D,\ \gamma_j^+u=e^{i\alpha_j}\gamma_j^-u\ (1\leq j\leq d)\}. \tag*{(28)} U α = { u ∈ H 1 ( Y ; C ) : u = 1 in D , γ j + u = e i α j γ j − u ( 1 ≤ j ≤ d )} . (28)
Choose a real H 1 H^1 H 1 cutoff χ \chi χ , equal to one in a neighbourhood of D ‾ \overline{D} D and zero in a neighbourhood of ∂ Y \partial Y ∂ Y . Set E χ = ∣ D ∣ − 1 ∫ Y ∣ ∇ χ ∣ 2 E_{\chi}=\lvert D\rvert^{-1}\int_Y\lvert\nabla\chi\rvert^2 E χ = ∣ D ∣ − 1 ∫ Y ∣ ∇ χ ∣ 2 . We see that U α \mathcal{U}_{\alpha} U α is nonempty since χ ∈ U α \chi\in\mathcal{U}_{\alpha} χ ∈ U α .
The minimiser in (4) is denoted by
u α = argmin u ∈ U α ∫ Y ∣ ∇ u ∣ 2 . u_{\alpha}=\operatorname*{argmin}_{u\in\mathcal{U}_{\alpha}}\int_Y\lvert\nabla u\rvert^2. u α = u ∈ U α argmin ∫ Y ∣ ∇ u ∣ 2 .
Differences of functions in U α \mathcal{U}_{\alpha} U α vanish in D D D , so Lemma 3.1 gives coercivity. Lax–Milgram, or strict convexity of the energy, proves existence and uniqueness of the minimiser. The Euler equation is
∫ Y ∇ u α ⋅ ∇ h ‾ = 0 ( h = 0 in D , h has Bloch phase α ) . (29) \int_Y \nabla u_{\alpha}\cdot\nabla\overline{h}=0 \qquad(h=0\text{ in }D,\ h\text{ has Bloch phase }\alpha). \tag*{(29)} ∫ Y ∇ u α ⋅ ∇ h = 0 ( h = 0 in D , h has Bloch phase α ) . (29)
The minimiser is harmonic in O \mathcal{O} O . Its Bloch extension is weakly harmonic across the cell faces: test the Euler equation with the Bloch periodisation of a compactly supported exterior test function. In particular it is smooth near any open face patch away from its edges and its partial derivatives ∂ j u α \partial_j u_{\alpha} ∂ j u α satisfy
γ j + ( ∂ j u α ) = e i α j γ j − ( ∂ j u α ) , j = 1 , … , d (30) \gamma_j^{+}(\partial_j u_{\alpha})=e^{i\alpha_j}\gamma_j^{-}(\partial_j u_{\alpha}),\qquad j=1,\ldots,d \tag*{(30)} γ j + ( ∂ j u α ) = e i α j γ j − ( ∂ j u α ) , j = 1 , … , d (30)
at the cell faces. Integration by parts cancels opposite face terms and gives
c ( α ) = − 1 ∣ D ∣ ∫ ∂ D ∂ ν D + u α d S = 1 ∣ D ∣ ∫ Y ∣ ∇ u α ∣ 2 , (31) c(\alpha)=-\frac{1}{|D|}\int_{\partial D}\partial_{\nu_D}^{+}u_{\alpha}\,\mathrm{d}S=\frac{1}{|D|}\int_Y|\nabla u_{\alpha}|^2, \tag*{(31)} c ( α ) = − ∣ D ∣ 1 ∫ ∂ D ∂ ν D + u α d S = ∣ D ∣ 1 ∫ Y ∣∇ u α ∣ 2 , (31)
where ν D \nu_D ν D points out of D D D and the superscript means the exterior trace.
For finitely supported q = ( q n ) q=(q_n) q = ( q n ) , minimise the global energy
Q ( q ) = 1 ∣ D ∣ min { ∫ R d ∣ ∇ U ∣ 2 : U ∈ H 1 ( R d ) , U = q n in D + n } . (32) Q(q)=\frac{1}{|D|}\min\left\{\int_{\mathbb{R}^d}|\nabla U|^2:U\in H^1(\mathbb{R}^d),\quad U=q_n\text{ in }D+n\right\}. \tag*{(32)} Q ( q ) = ∣ D ∣ 1 min { ∫ R d ∣∇ U ∣ 2 : U ∈ H 1 ( R d ) , U = q n in D + n } . (32)
A sum of translated versions of the cutoff χ \chi χ , given below (28) , gives an admissible lift with norm bounded by a constant times ∥ q ∥ ℓ 2 \|q\|_{\ell^2} ∥ q ∥ ℓ 2 . Summing (27) over cells gives coercivity on the homogeneous space of functions vanishing in every resonator. Thus the minimiser exists uniquely. Polarisation of Q Q Q defines a bounded, nonnegative, self-adjoint operator C C C on ℓ 2 ( Z d ) \ell^2(\mathbb{Z}^d) ℓ 2 ( Z d ) .
With
( B U ) ( α , x ) = ∑ n U ( x + n ) e i n ⋅ α , (BU)(\alpha,x)=\sum_n U(x+n)e^{in\cdot\alpha}, ( B U ) ( α , x ) = n ∑ U ( x + n ) e in ⋅ α ,
the transformed minimiser is q ^ ( α ) u − α ( x ) \widehat{q}(\alpha)u_{-\alpha}(x) q ( α ) u − α ( x ) . Indeed, the transform has phase − α -\alpha − α on opposite faces and value q ^ ( α ) \widehat{q}(\alpha) q ( α ) in D D D , and the transformed Euler equation has a unique solution. Complex conjugation gives c ( − α ) = c ( α ) c(-\alpha)=c(\alpha) c ( − α ) = c ( α ) . Parseval’s identity therefore gives
Q ( q ) = 1 ( 2 π ) d ∫ [ − π , π ] d c ( α ) ∣ q ^ ( α ) ∣ 2 d α . (33) Q(q)=\frac{1}{(2\pi)^d}\int_{[-\pi,\pi]^d}c(\alpha)|\widehat{q}(\alpha)|^2\,\mathrm{d}\alpha. \tag*{(33)} Q ( q ) = ( 2 π ) d 1 ∫ [ − π , π ] d c ( α ) ∣ q ( α ) ∣ 2 d α . (33)
This proves the convolution representation (5) . Real coefficients of the exterior problem and self-adjointness give c r ∈ R c_r\in\mathbb{R} c r ∈ R and c − r = c r c_{-r}=c_r c − r = c r . At phase zero the constant function one is admissible, so c ( 0 ) = 0 c(0)=0 c ( 0 ) = 0 . If c ( α ) = 0 c(\alpha)=0 c ( α ) = 0 , its minimiser must be constant and equal to one; then all phases equal one. Thus c ( α ) > 0 c(\alpha)>0 c ( α ) > 0 for α ≠ 0 \alpha\ne0 α = 0 .
Analyticity and exponential decay
Exponential off-diagonal decay is established for the crystal capacitance operator in [12 Proposition 2.5] . The variational formulation in [9 Proposition 3.5 and Remark 2] proves analyticity of the Floquet symbol associated with the capacitance operator on the whole Brillouin torus and derives the same decay, including a uniform analytic extension for the contrast-dependent family. This is the decay input used in our finite-volume localisation proof. We give a self-contained proof for the static operator, with explicit geometric constants, in both dimensions.
Proposition 3.2. The symbol is real analytic on the whole Brillouin torus, including zero. With C P C_P C P and E χ E_\chi E χ above, set
γ 0 = 1 8 d C P , A 0 = 3 E χ exp ( d 8 C P ) . (34) \gamma_0=\frac{1}{8\sqrt{d}C_P},\qquad A_0=3E_\chi\exp\left(\frac{\sqrt{d}}{8C_P}\right). \tag*{(34)} γ 0 = 8 d C P 1 , A 0 = 3 E χ exp ( 8 C P d ) . (34) Then
∣ c r ∣ ≤ A 0 e − γ 0 ∣ r ∣ 1 , r ∈ Z d . (35) |c_r|\le A_0e^{-\gamma_0|r|_1},\qquad r\in\mathbb{Z}^d. \tag*{(35)} ∣ c r ∣ ≤ A 0 e − γ 0 ∣ r ∣ 1 , r ∈ Z d . (35) In particular all polynomially weighted coefficient sums are finite, and ∑ r c r = c ( 0 ) = 0 \sum_r c_r=c(0)=0 ∑ r c r = c ( 0 ) = 0 .
Proof. Use the fixed space
H = { w ∈ H 1 ( Y ) : w = 0 in D , γ j + w = γ j − w for all j } . \mathcal{H}=\{w\in H^1(Y):w=0\text{ in }D,\ \gamma_j^+w=\gamma_j^-w\text{ for all }j\}. H = { w ∈ H 1 ( Y ) : w = 0 in D , γ j + w = γ j − w for all j } .
For real α \alpha α , write u α = χ + e i α ⋅ x w α u_\alpha=\chi+e^{i\alpha\cdot x}w_\alpha u α = χ + e i α ⋅ x w α . The equation for w α w_\alpha w α is a α ( w α , v ) = b α ( v ) a_\alpha(w_\alpha,v)=b_\alpha(v) a α ( w α , v ) = b α ( v ) for v ∈ H v\in\mathcal{H} v ∈ H , where the following formulas also define holomorphic dependence on a complex vector z z z :
a z ( w , v ) = ∫ Y ( ∇ + i z ) w ⋅ ( ∇ − i z ) v ‾ . (36) a_z(w,v)=\int_Y(\nabla+iz)w\cdot(\nabla-iz)\overline{v}. \tag*{(36)} a z ( w , v ) = ∫ Y ( ∇ + i z ) w ⋅ ( ∇ − i z ) v . (36)
b z ( v ) = − ∫ Y e − i z ⋅ x ∇ χ ⋅ ( ∇ − i z ) v ‾ . (37) b_z(v)=-\int_Y e^{-iz\cdot x}\nabla\chi\cdot(\nabla-iz)\overline{v}. \tag*{(37)} b z ( v ) = − ∫ Y e − i z ⋅ x ∇ χ ⋅ ( ∇ − i z ) v . (37)
Notice that z z z , rather than z ‾ \overline{z} z , occurs in both factors of (36) ; it is the holomorphic continuation of the real-parameter form. For real α \alpha α , a α a_\alpha a α is a positive Hermitian form.
Put D α = ∇ + i α D_\alpha=\nabla+i\alpha D α = ∇ + i α . Applying Lemma 3.1 to e i α ⋅ x w e^{i\alpha\cdot x}w e i α ⋅ x w gives
∥ w ∥ 2 ≤ C P ∥ D α w ∥ 2 . (38) \|w\|_2\le C_P\|D_\alpha w\|_2. \tag*{(38)} ∥ w ∥ 2 ≤ C P ∥ D α w ∥ 2 . (38)
Thus, for fixed α \alpha α , ∥ w ∥ α = ∥ D α w ∥ 2 \|w\|_\alpha=\|D_\alpha w\|_2 ∥ w ∥ α = ∥ D α w ∥ 2 is a norm on H \mathcal{H} H equivalent to its H 1 H^1 H 1 norm. Let z = α + i β z=\alpha+i\beta z = α + i β with real α , β \alpha,\beta α , β , and put τ = C P ∣ β ∣ 2 \tau=C_P|\beta|^2 τ = C P ∣ β ∣ 2 . Expanding the factors in (36) gives
∣ a z ( w , v ) − a α ( w , v ) ∣ ≤ ( 2 τ + τ 2 ) ∥ w ∥ α ∥ v ∥ α . (39) |a_z(w,v)-a_\alpha(w,v)|\le(2\tau+\tau^2)\|w\|_\alpha\|v\|_\alpha. \tag*{(39)} ∣ a z ( w , v ) − a α ( w , v ) ∣ ≤ ( 2 τ + τ 2 ) ∥ w ∥ α ∥ v ∥ α . (39)
If ∣ β j ∣ ≤ γ 0 |\beta_j|\le\gamma_0 ∣ β j ∣ ≤ γ 0 for every j j j , then τ ≤ 1 / 8 \tau\le1/8 τ ≤ 1/8 , so 2 τ + τ 2 ≤ 17 / 64 < 1 2\tau+\tau^2\le17/64<1 2 τ + τ 2 ≤ 17/64 < 1 . The operator associated with a z a_z a z is invertible by a Neumann series relative to a α a_\alpha a α , with inverse norm at most ( 1 − 2 τ − τ 2 ) − 1 (1-2\tau-\tau^2)^{-1} ( 1 − 2 τ − τ 2 ) − 1 . The solution w z w_z w z depends holomorphically on z z z in a neighbourhood of this closed strip.
Since ∣ x ∣ 2 ≤ d / 2 |x|^2\le\sqrt{d}/2 ∣ x ∣ 2 ≤ d /2 in Y Y Y , let b β = d ∣ β ∣ 2 / 2 b_\beta=\sqrt{d}|\beta|^2/2 b β = d ∣ β ∣ 2 /2 . From (37) and (38) ,
∥ w z ∥ α ≤ e b β ( 1 + τ ) 1 − 2 τ − τ 2 ( ∣ D ∣ E χ ) 1 / 2 . \|w_z\|_\alpha\le\frac{e^{b_\beta}(1+\tau)}{1-2\tau-\tau^2}(|D|E_\chi)^{1/2}. ∥ w z ∥ α ≤ 1 − 2 τ − τ 2 e b β ( 1 + τ ) ( ∣ D ∣ E χ ) 1/2 .
For real phases, using the test function u α − χ u_\alpha-\chi u α − χ in (29) shows that c ( α ) = ∣ D ∣ − 1 ∫ Y ∇ u α ⋅ ∇ χ c(\alpha)=|D|^{-1}\int_Y\nabla u_\alpha\cdot\nabla\chi c ( α ) = ∣ D ∣ − 1 ∫ Y ∇ u α ⋅ ∇ χ . Its holomorphic continuation is therefore
c ( z ) = 1 ∣ D ∣ ∫ Y ∇ ( χ + e i z ⋅ x w z ) ⋅ ∇ χ . (40) c(z)=\frac{1}{|D|}\int_Y\nabla(\chi+e^{iz\cdot x}w_z)\cdot\nabla\chi. \tag*{(40)} c ( z ) = ∣ D ∣ 1 ∫ Y ∇ ( χ + e i z ⋅ x w z ) ⋅ ∇ χ . (40)
It is periodic by uniqueness of analytic continuation from the real torus. The preceding estimates give
∣ c ( z ) ∣ ≤ E χ ( 1 + e 2 b β ( 1 + τ ) 2 1 − 2 τ − τ 2 ) ≤ E χ ( 1 + 81 47 e d / ( 8 C P ) ) ≤ A 0 . |c(z)|\le E_\chi\left(1+e^{2b_\beta}\frac{(1+\tau)^2}{1-2\tau-\tau^2}\right)
\le E_\chi\left(1+\frac{81}{47}e^{\sqrt{d}/(8C_P)}\right)\le A_0. ∣ c ( z ) ∣ ≤ E χ ( 1 + e 2 b β 1 − 2 τ − τ 2 ( 1 + τ ) 2 ) ≤ E χ ( 1 + 47 81 e d / ( 8 C P ) ) ≤ A 0 .
Shift each Fourier integration contour in (5) by − i γ 0 sgn ( r j ) -i\gamma_0\operatorname{sgn}(r_j) − i γ 0 sgn ( r j ) ; periodicity cancels the vertical sides. The exponential factor has modulus e − γ 0 ∣ r ∣ 1 e^{-\gamma_0|r|_1} e − γ 0 ∣ r ∣ 1 . This proves (35) . The remaining assertions follow by absolute convergence of the Fourier series and its derivatives. □
Remark 3.3. Reflection symmetry is not needed for Proposition 3.2 . Moreover, the constants in Proposition 3.2 are deliberately conservative. We point out that the full spatial periodicity and the Dirichlet boundary condition on the resonator in every cell are essential to this argument.
Critical points and coordinate monotonicity
In this subsection, we assume the reflection symmetry (R). Centred balls, and centred axis-aligned ellipsoids contained in the cell, satisfy these assumptions and the connected-exterior condition. The reflection method is related to the half-cell variational argument for a resonator chain in [11 Appendix A] . Here we obtain a coordinate derivative identity for the fully periodic cell problem.
Theorem 3.4. Let j ∈ { 1 , … , d } j\in\{1,\ldots,d\} j ∈ { 1 , … , d } and α ∈ Y ∗ \alpha\in Y^* α ∈ Y ∗ with α j ∉ { 0 , π } \alpha_j\notin\{0,\pi\} α j ∈ / { 0 , π } . Then
∂ α j c ( α ) = 1 2 ∣ D ∣ sin α j ∫ Y ∣ ∇ ( u α − u α ∘ R j ) ∣ 2 , (41) \partial_{\alpha_j}c(\alpha)=\frac{1}{2|D|\sin\alpha_j}\int_Y\left|\nabla(u_\alpha-u_\alpha\circ R_j)\right|^2, \tag*{(41)} ∂ α j c ( α ) = 2∣ D ∣ sin α j 1 ∫ Y ∣ ∇ ( u α − u α ∘ R j ) ∣ 2 , (41) and the integral is strictly positive. Hence ∂ α j c \partial_{\alpha_j}c ∂ α j c is positive for 0 < α j < π 0<\alpha_j<\pi 0 < α j < π and negative for − π < α j < 0 -\pi<\alpha_j<0 − π < α j < 0 , and it vanishes on α j ∈ { 0 , π } \alpha_j\in\{0,\pi\} α j ∈ { 0 , π } because c c c is even and 2 π 2\pi 2 π -periodic in α j \alpha_j α j . Consequently
Crit ( c ) : = { α ∈ Y ∗ : ∇ c ( α ) = 0 } = V : = { 0 , π } d . (42) \operatorname{Crit}(c):=\{\alpha\in Y^*:\nabla c(\alpha)=0\}=\mathcal{V}:=\{0,\pi\}^d. \tag*{(42)} Crit ( c ) := { α ∈ Y ∗ : ∇ c ( α ) = 0 } = V := { 0 , π } d . (42) The unique minimum is c ( 0 ) = 0 c(0)=0 c ( 0 ) = 0 and the unique maximum is c M = c ( α M ) c_M=c(\alpha_M) c M = c ( α M ) , α M = ( π , … , π ) \alpha_M=(\pi,\ldots,\pi) α M = ( π , … , π ) . Thus σ ( C ) = [ 0 , c M ] \sigma(C)=[0,c_M] σ ( C ) = [ 0 , c M ] . If J J J is a nonempty compact interval in ( 0 , c M ) (0,c_M) ( 0 , c M ) avoiding T : = { c ( ν ) : ν ∈ V } \mathcal{T}:=\{c(\nu):\nu\in\mathcal{V}\} T := { c ( ν ) : ν ∈ V } , then
min α : c ( α ) ∈ J ∣ ∇ c ( α ) ∣ 2 > 0. (43) \min_{\alpha:\,c(\alpha)\in J}|\nabla c(\alpha)|^2>0. \tag*{(43)} α : c ( α ) ∈ J min ∣∇ c ( α ) ∣ 2 > 0. (43)
Proof. Reflection R j R_j R j preserves Dirichlet energy and reverses only the j j j -th phase. Hence c c c is even in each coordinate separately. Fix the transverse phases α 1 , … , α j − 1 , α j + 1 , … , α d \alpha_1,\ldots,\alpha_{j-1},\alpha_{j+1},\ldots,\alpha_d α 1 , … , α j − 1 , α j + 1 , … , α d and let t = tan ( α j / 2 ) t=\tan(\alpha_j/2) t = tan ( α j /2 ) . For an arbitrary admissible u u u , put
v = u + u ∘ R j 2 , z = u − u ∘ R j 2 i t . v=\frac{u+u\circ R_j}{2},\qquad z=\frac{u-u\circ R_j}{2it}. v = 2 u + u ∘ R j , z = 2 i t u − u ∘ R j .
Then v v v is even, z z z is odd, v = 1 v=1 v = 1 and z = 0 z=0 z = 0 in D D D , and both retain all transverse phases. Their traces satisfy γ j z = γ j v \gamma_j^z=\gamma_j^v γ j z = γ j v . Conversely these constraints, which do not involve t t t , imply that u = v + i t z u=v+itz u = v + i t z has the required j j j th phase: its two traces are ( 1 + i t ) b (1+it)b ( 1 + i t ) b and ( 1 − i t ) b (1-it)b ( 1 − i t ) b , whose ratio is e i α j e^{i\alpha_j} e i α j . Thus the correspondence parametrises the entire admissible space. The functions remain complex-valued; taking real and imaginary parts would not preserve arbitrary transverse phases.
Parity and change of variables under R j R_j R j give ∫ Y ∇ v ⋅ ∇ z ‾ = 0 \int_Y \nabla v\cdot\nabla\overline{z}=0 ∫ Y ∇ v ⋅ ∇ z = 0 . Therefore, on a fixed affine space A \mathcal{A} A of pairs,
c ( α ) = F ( t 2 ) , F ( a ) = 1 ∣ D ∣ min ( v , z ) ∈ A { ∫ Y ∣ ∇ v ∣ 2 + a ∫ Y ∣ ∇ z ∣ 2 } . (44) c(\alpha)=F(t^2),\qquad F(a)=\frac{1}{|D|}\min_{(v,z)\in\mathcal{A}}\left\{\int_Y|\nabla v|^2+a\int_Y|\nabla z|^2\right\}. \tag*{(44)} c ( α ) = F ( t 2 ) , F ( a ) = ∣ D ∣ 1 ( v , z ) ∈ A min { ∫ Y ∣∇ v ∣ 2 + a ∫ Y ∣∇ z ∣ 2 } . (44)
This affine space is nonempty: average a cutoff with its reflection to obtain an even cutoff, and pair it with zero. Its homogeneous space consists of pairs whose two components vanish in D D D . Lemma 3.1 makes the form coercive, locally uniformly for a > 0 a>0 a > 0 . The minimising pair ( v a , z a ) (v_a,z_a) ( v a , z a ) depends continuously in H 1 × H 1 H^1\times H^1 H 1 × H 1 on a a a . To see this directly, subtract the two Euler equations at a a a and a ′ a' a ′ and test by the difference. Coercivity and the bound from the fixed cutoff competitor yield ∥ ( v a , z a ) − ( v a ′ , z a ′ ) ∥ H 1 × H 1 ≤ C ∣ a − a ′ ∣ \|(v_a,z_a)-(v_{a'},z_{a'})\|_{H^1\times H^1}\le C|a-a'| ∥ ( v a , z a ) − ( v a ′ , z a ′ ) ∥ H 1 × H 1 ≤ C ∣ a − a ′ ∣ on every compact subinterval of ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) .
Using each minimiser as a competitor for the other parameter gives, for a 2 > a 1 a_2>a_1 a 2 > a 1 ,
a 2 − a 1 ∣ D ∣ ∥ ∇ z a 2 ∥ 2 2 ≤ F ( a 2 ) − F ( a 1 ) ≤ a 2 − a 1 ∣ D ∣ ∥ ∇ z a 1 ∥ 2 2 . \frac{a_2-a_1}{|D|}\|\nabla z_{a_2}\|_2^2\le F(a_2)-F(a_1)\le\frac{a_2-a_1}{|D|}\|\nabla z_{a_1}\|_2^2. ∣ D ∣ a 2 − a 1 ∥∇ z a 2 ∥ 2 2 ≤ F ( a 2 ) − F ( a 1 ) ≤ ∣ D ∣ a 2 − a 1 ∥∇ z a 1 ∥ 2 2 .
Consequently F ′ ( a ) = ∣ D ∣ − 1 ∥ ∇ z a ∥ 2 2 F'(a)=|D|^{-1}\|\nabla z_a\|_2^2 F ′ ( a ) = ∣ D ∣ − 1 ∥∇ z a ∥ 2 2 .
This derivative is strictly positive. Otherwise z a z_a z a is constant and zero in D D D , hence zero. The corresponding u α u_\alpha u α is even. Its equal face traces and its Bloch condition imply zero trace on the j j j th face because e i α j ≠ 1 e^{i\alpha_j}\ne1 e i α j = 1 . The coordinate derivative is odd under reflection. Combining this with (30) gives zero normal derivative on that face because e i α j ≠ − 1 e^{i\alpha_j}\ne-1 e i α j = − 1 . On an open face patch, the harmonic Bloch extension has both zero value and zero normal derivative. Tangential differentiation and the Laplace equation make every derivative zero there; harmonic analyticity makes the solution zero in a neighbourhood. Unique continuation in the connected exterior O \mathcal{O} O then gives zero everywhere in O \mathcal{O} O , contradicting its trace one on ∂ D \partial D ∂ D . The Sobolev trace and harmonic regularity facts used here are standard; see [16 ] .
Now d ( t 2 ) / d α j = t ( 1 + t 2 ) \mathrm{d}(t^2)/\mathrm{d}\alpha_j=t(1+t^2) d ( t 2 ) / d α j = t ( 1 + t 2 ) and ( 1 + t 2 ) / t = 2 / sin α j (1+t^2)/t=2/\sin\alpha_j ( 1 + t 2 ) / t = 2/ sin α j , (41) follows by the chain rule. Evenness gives the negative derivative on the negative half-period and zero derivative at zero. Evenness together with periodicity gives zero derivative at π \pi π . Every point outside V \mathcal{V} V has at least one coordinate with nonzero derivative, proving (42) .
The extrema follow from strict coordinate monotonicity and from c ( 0 ) = 0 c(0)=0 c ( 0 ) = 0 . Continuity on the connected torus and the multiplier representation give σ ( C ) = [ 0 , c M ] \sigma(\mathcal{C})=[0,c_M] σ ( C ) = [ 0 , c M ] . Finally, the inverse image of J J J is compact and contains no critical point. The positive continuous function ∣ ∇ c ∣ C 2 |\nabla c|_{C^2} ∣∇ c ∣ C 2 has a positive minimum there. □
Remark 3.5. All 2 d 2^d 2 d reflection vertices are critical, including mixed vertices; α M \alpha_M α M is not the only critical quasimonum. On compact energy intervals avoiding their finitely many values, the level surfaces are regular analytic hypersurfaces. This does not assert nonzero Gaussian curvature. The nondegeneracy at the upper vertex, needed below, is proved next.
Nondegeneracy of the upper edge
The uniqueness and nondegeneracy of the upper maximum are proved in [9 ] , Lemmas 3.6–3.7. We recall the nondegeneracy proof with our normalisation. This result is the geometric input that allows the upper-edge deficit to be compared with a discrete Laplacian in the localisation argument.
Proposition 3.6. Under (R), the matrix − ∇ α 2 c ( α M ) -\nabla_{\alpha}^{2}c(\alpha_M) − ∇ α 2 c ( α M ) is positive definite. In particular, there are a ∗ , a ∗ > 0 a_*,a^*>0 a ∗ , a ∗ > 0 and $r_*>0 such that
a ∗ ∣ β ∣ 2 ≤ c M − c ( α M + β ) ≤ a ∗ ∣ β ∣ 2 ( ∣ β ∣ 2 ≤ r ∗ ) . (45) a_*|\beta|^2 \le c_M-c(\alpha_M+\beta)\le a^*|\beta|^2 \qquad(|\beta|^2\le r_*). \tag*{(45)} a ∗ ∣ β ∣ 2 ≤ c M − c ( α M + β ) ≤ a ∗ ∣ β ∣ 2 ( ∣ β ∣ 2 ≤ r ∗ ) . (45) Proof. Uniqueness and reflection show that u α M u_{\alpha_M} u α M is even in every coordinate: reflection changes π \pi π into the equivalent phase − π -\pi − π . Antiperiodicity therefore gives zero trace on every face of Y Y Y . Complex conjugation and uniqueness show that u α M u_{\alpha_M} u α M is real. In the connected exterior it solves the Dirichlet problem with value one on ∂ D \partial D ∂ D and zero on ∂ Y \partial Y ∂ Y . The maximum principle and the Hopf lemma give
0 < u α M < 1 in Y ∖ D ‾ , ∂ ν u α M < 0 on the interior of each cell face . 0<u_{\alpha_M}<1\quad\text{in }Y\setminus\overline{D},\qquad\partial_\nu u_{\alpha_M}<0\quad\text{on the interior of each cell face}. 0 < u α M < 1 in Y ∖ D , ∂ ν u α M < 0 on the interior of each cell face . No regularity at the edges of the cube is needed.
Fix ξ ∈ R d \xi\in\mathbb{R}^d ξ ∈ R d and put X M = { h ∈ H α 1 ( Y ) : h = 0 in D } X_M=\{h\in H^1_\alpha(Y):h=0\text{ in }D\} X M = { h ∈ H α 1 ( Y ) : h = 0 in D } . Since u α M u_{\alpha_M} u α M vanishes on the cell boundary, for real small t t t we can write uniquely
u α M + t ξ = u α M + e i t ξ ⋅ x w t , w t ∈ X M , w 0 = 0. u_{\alpha_M+t\xi}=u_{\alpha_M}+e^{it\xi\cdot x}w_t,\qquad w_t\in X_M,\qquad w_0=0. u α M + t ξ = u α M + e i t ξ ⋅ x w t , w t ∈ X M , w 0 = 0. The gauged variational equation on the fixed space X M X_M X M is coercive near zero. The analytic inverse argument from Proposition 3.2 therefore gives analytic H 1 H^1 H 1 dependence of w t w_t w t on t t t . The Euler equation, tested by the difference from u α M u_{\alpha_M} u α M , gives the exact energy identity
∣ D ∣ { c M − c ( α M + t ξ ) } = ∫ Y ∣ ∇ ( e i t ξ ⋅ x w t ) ∣ 2 . |D|\{c_M-c(\alpha_M+t\xi)\}=\int_Y|\nabla(e^{it\xi\cdot x}w_t)|^2. ∣ D ∣ { c M − c ( α M + t ξ )} = ∫ Y ∣∇ ( e i t ξ ⋅ x w t ) ∣ 2 . Consequently
ξ ⋅ ∇ α 2 c ( α M ) ξ = − 2 ∣ D ∣ ∫ Y ∣ ∇ w 0 ′ ∣ 2 . (46) \xi\cdot\nabla_{\alpha}^{2}c(\alpha_M)\xi=-\frac{2}{|D|}\int_Y|\nabla w'_0|^2. \tag*{(46)} ξ ⋅ ∇ α 2 c ( α M ) ξ = − ∣ D ∣ 2 ∫ Y ∣∇ w 0 ′ ∣ 2 . (46) If this is zero, the cell Poincaré inequality gives w 0 ′ = 0 w'_0=0 w 0 ′ = 0 . Differentiate the Euler equation against e i t ξ ⋅ x h e^{it\xi\cdot x}h e i t ξ ⋅ x h at zero. With w 0 ′ = 0 w'_0=0 w 0 ′ = 0 the result is
∫ Y ∖ D ‾ ∇ u α M ⋅ ∇ ( ( ξ ⋅ x ) h ‾ ) = 0 ( h ∈ X M ) . \int_{Y\setminus\overline{D}}\nabla u_{\alpha_M}\cdot\nabla\bigl((\xi\cdot x)\overline{h}\bigr)=0\qquad(h\in X_M). ∫ Y ∖ D ∇ u α M ⋅ ∇ ( ( ξ ⋅ x ) h ) = 0 ( h ∈ X M ) . Integration by parts and pairing opposite faces yield
∑ j = 1 d ξ j ∫ F j + ∂ ν u α M h ‾ d S = 0 , F j + = { x ∈ ∂ Y : x j = 1 / 2 } . (47) \sum_{j=1}^{d}\xi_j\int_{F_j^+}\partial_\nu u_{\alpha_M}\,\overline{h}\,dS=0,\qquad F_j^+=\{x\in\partial Y:x_j=1/2\}. \tag*{(47)} j = 1 ∑ d ξ j ∫ F j + ∂ ν u α M h d S = 0 , F j + = { x ∈ ∂ Y : x j = 1/2 } . (47) Here the outward normal derivatives agree on the two faces by reflection, the traces of h h h have opposite signs, and the values of ξ ⋅ x \xi\cdot x ξ ⋅ x differ by ξ j \xi_j ξ j . Choose h h h with a nonzero nonnegative smooth trace supported in the interior of F j + F_j^+ F j + , the opposite trace on F j − F_j^- F j − , and zero trace on other faces and on D D D . Such a function is obtained by an H 1 H^1 H 1 extension in collars of the two faces, which are separated from D D D . Equation (47) and the strict Hopf inequality force ξ j = 0 \xi_j=0 ξ j = 0 . This holds for each j j j , so ξ = 0 \xi=0 ξ = 0 . Thus the Hessian is negative definite. Compactness of the unit sphere and Taylor’s theorem give (45) . □
Corollary 3.7. Under (R), the periodic operator C C C has purely absolutely continuous spectrum.
Proof. Its spectral measure at q ∈ ℓ 2 ( Z d ) q\in\ell^2(\mathbb{Z}^d) q ∈ ℓ 2 ( Z d ) is the push-forward under c c c of ( 2 π ) − d ∣ q ^ ( α ) ∣ 2 d α (2\pi)^{-d}|\widehat{q}(\alpha)|^2\,d\alpha ( 2 π ) − d ∣ q ( α ) ∣ 2 d α . Away from the finite critical set, a local coordinate change can use c ( α ) c(\alpha) c ( α ) as one coordinate. The inverse image of a Lebesgue-null energy set therefore has zero d d d -dimensional measure in each such chart, by Fubini’s theorem. A countable chart cover and the nullity of the finite critical set prove absolute continuity of every spectral measure. □
The contrast-dependent family
For δ > 0 \delta>0 δ > 0 , let f D u = ∣ D ∣ − 1 ∫ D u f_Du=|D|^{-1}\int_Du f D u = ∣ D ∣ − 1 ∫ D u and define the family of [9 ] by
c δ ( α ) = C ^ δ α = 1 ∣ D ∣ min u ∈ H α 1 ( Y ) f D u = 1 ∫ Y ( δ − 1 1 D + 1 Y ∖ D ) ∣ ∇ u ∣ 2 , ( C δ q ) n = ∑ m c δ , n − m q m . (48) c_\delta(\alpha)=\widehat{C}_\delta^\alpha=\frac{1}{|D|}\min_{\substack{u\in H_\alpha^1(Y)\\ f_Du=1}}\int_Y\left(\delta^{-1}\mathbf{1}_D+\mathbf{1}_{Y\setminus D}\right)|\nabla u|^2,\qquad(C_\delta q)_n=\sum_m c_{\delta,n-m}q_m. \tag*{(48)} c δ ( α ) = C δ α = ∣ D ∣ 1 u ∈ H α 1 ( Y ) f D u = 1 min ∫ Y ( δ − 1 1 D + 1 Y ∖ D ) ∣∇ u ∣ 2 , ( C δ q ) n = m ∑ c δ , n − m q m . (48)
The Fourier coefficients use (5) . Set C 0 = C C_0=C C 0 = C , c 0 ( α ) = c ( α ) c_0(\alpha)=c(\alpha) c 0 ( α ) = c ( α ) , and c M , δ = c δ ( α M ) c_{M,\delta}=c_\delta(\alpha_M) c M , δ = c δ ( α M ) under (R). Here c 0 ( α ) c_0(\alpha) c 0 ( α ) denotes an element of the symbol family; the coefficient c 0 c_0 c 0 without an argument retains its meaning in (5) . We use c δ , r c_{\delta,r} c δ , r for family coefficients. Also put H δ , ϖ = M ϖ − 1 / 2 C δ M ϖ − 1 / 2 H_{\delta,\varpi}=M_\varpi^{-1/2}C_\delta M_\varpi^{-1/2} H δ , ϖ = M ϖ − 1/2 C δ M ϖ − 1/2 and C δ , Λ = E Λ ∗ C δ E Λ C_{\delta,\Lambda}=E_\Lambda^*C_\delta E_\Lambda C δ , Λ = E Λ ∗ C δ E Λ , with H δ , ϖ , Λ t = M ϖ , Λ − 1 / 2 C δ , Λ t M ϖ , Λ − 1 / 2 H_{\delta,\varpi,\Lambda}^{\mathsf{t}}=M_{\varpi,\Lambda}^{-1/2}C_{\delta,\Lambda}^{\mathsf{t}}M_{\varpi,\Lambda}^{-1/2} H δ , ϖ , Λ t = M ϖ , Λ − 1/2 C δ , Λ t M ϖ , Λ − 1/2 .
Proposition 3.8 (Uniform deterministic estimates). *There are δ 0 > 0 \delta_0>0 δ 0 > 0 and constants A ∗ , γ ∗ , a ∗ , c ∗ , c ∗ > 0 A_*,\gamma_*,a_*,c_*,c^*>0 A ∗ , γ ∗ , a ∗ , c ∗ , c ∗ > 0 such that, for 0 ≤ δ ≤ δ 0 0\leq\delta\leq\delta_0 0 ≤ δ ≤ δ 0 and every fixed nonnegative integer k k k ,
∣ c δ , r ∣ ≤ A ∗ e − γ ∗ ∣ r ∣ 1 , c δ ⟶ c in C k ( Y ∗ ) ( δ ↓ 0 ) . (49) |c_{\delta,r}|\leq A_*e^{-\gamma_*|r|_1},\qquad c_\delta\longrightarrow c\quad\text{in }C^k(Y^*)\quad(\delta\downarrow0). \tag*{(49)} ∣ c δ , r ∣ ≤ A ∗ e − γ ∗ ∣ r ∣ 1 , c δ ⟶ c in C k ( Y ∗ ) ( δ ↓ 0 ) . (49)
Under (R), the unique maximum is at α M \alpha_M α M , c ∗ ≤ c M , δ ≤ c ∗ c_*\leq c_{M,\delta}\leq c^* c ∗ ≤ c M , δ ≤ c ∗ , and
c M , δ − c δ ( α M + β ) ≥ a ∗ ∑ j = 1 d ( 1 − cos β j ) ( β ∈ Y ∗ ) . (50) c_{M,\delta}-c_\delta(\alpha_M+\beta)\geq a_*\sum_{j=1}^{d}(1-\cos\beta_j)\qquad(\beta\in Y^*). \tag*{(50)} c M , δ − c δ ( α M + β ) ≥ a ∗ j = 1 ∑ d ( 1 − cos β j ) ( β ∈ Y ∗ ) . (50)
:::
Proof. Proposition 3.5 of [9 ] gives a bounded holomorphic extension to a common neighbourhood of the whole real torus, and convergence in each C k C^k C k norm. Shrink the complex neighbourhood to a common strip and shift the Fourier contours as in Proposition 3.2; this gives (49) . The cell Poincaré estimate also holds when f D h = 0 f_Dh=0 f D h = 0 : indeed ∣ D ∣ ∣ h ∣ 2 ≤ ∫ D ∣ h − h ˉ ∣ 2 |D||h|^2\leq\int_D|h-\bar h|^2 ∣ D ∣∣ h ∣ 2 ≤ ∫ D ∣ h − h ˉ ∣ 2 , by Cauchy–Schwarz, and the proof of Lemma 3.1 applies. This is the coercivity used in the contrast-dependent construction.
Each symbol is even in every coordinate by reflection of the variational problem, so α M \alpha_M α M is a critical point. The negative definite Hessian in Proposition 3.6 and C 2 C^2 C 2 convergence give strict concavity and a common quadratic deficit in a fixed neighbourhood of α M \alpha_M α M . Away from that neighbourhood, the unique maximum of c c c and uniform convergence give a common positive gap. Thus the maximum remains unique and at α M \alpha_M α M . The positive upper and lower bounds follow by convergence of its value. Comparison with 2 ∑ j ( 1 − cos β j ) 2\sum_j(1-\cos\beta_j) 2 ∑ j ( 1 − cos β j ) near zero and on the remaining compact set proves (50) . This also recovers the edge properties in [9 Lemmas 3.6–3.7] . □
We use this family only where uniformity in the simultaneous period and contrast limit is needed. Statements without a δ \delta δ subscript refer to the static operator. The family is variational and real self-adjoint. It is distinct from the nonzero-frequency outgoing capacitance matrix of [10 Definition 3.9 and Proposition 3.16] , which is generally non-Hermitian and describes resonance shifts. Its spectral parameter is still E E E in (6) . Its reduced edge frequency is ω M , δ : = ω ε , δ ( c M , δ ) \omega_{M,\delta}:=\omega_{\varepsilon,\delta}(c_{M,\delta}) ω M , δ := ω ε , δ ( c M , δ ) . The notation C 0 = C C_0=C C 0 = C labels a static limit and does not set the physical contrast to zero in the frequency conversion.
Localisation estimates on prescribed intervals
We derive explicit sufficient conditions for exponential fractional-moment bounds on prescribed positive energy intervals. For the uniform inverse-material law, these conditions give quantitative thresholds on the disorder strength.
Let P a c ( H ) P_{\mathrm{ac}}(H) P ac ( H ) be the projection onto the absolutely continuous spectral subspace of a self-adjoint operator H H H . For 0 < ϑ < 1 0<\vartheta<1 0 < ϑ < 1 and 0 < μ < ϑ γ 0 0<\mu<\vartheta\gamma_0 0 < μ < ϑ γ 0 , define
T ϑ ( μ ) = ∑ r ≠ 0 ∣ c r ∣ ϑ e μ ∣ r ∣ < ∞ . (51) T_\vartheta(\mu)=\sum_{r\neq0}|c_r|^\vartheta e^{\mu|r|}<\infty. \tag*{(51)} T ϑ ( μ ) = r = 0 ∑ ∣ c r ∣ ϑ e μ ∣ r ∣ < ∞. (51)
The constants γ 0 \gamma_0 γ 0 , A 0 A_0 A 0 are established in Proposition 3.2 . The fixed-interval estimate does not require reflection symmetry.
Theorem 4.1 (A fixed interval at sufficiently large disorder). Let I = ( E 0 , E 1 ) I=(E_0,E_1) I = ( E 0 , E 1 ) with 0 < E 0 < E 1 < c M 0<E_0<E_1<c_M 0 < E 0 < E 1 < c M . Fix ϑ ∈ ( 0 , 1 ) \vartheta\in(0,1) ϑ ∈ ( 0 , 1 ) and μ ∈ ( 0 , ϑ γ 0 ) \mu\in(0,\vartheta\gamma_0) μ ∈ ( 0 , ϑ γ 0 ) . If
W > W ∗ ( I ; ϑ , μ ) : = 1 E 0 ( 2 ϑ ϑ T ϑ ( μ ) 1 − ϑ ) 1 / ϑ , (52) W>W_\ast(I;\vartheta,\mu):=\frac{1}{E_0}\left(\frac{2\vartheta^\vartheta T_\vartheta(\mu)}{1-\vartheta}\right)^{1/\vartheta}, \tag*{(52)} W > W ∗ ( I ; ϑ , μ ) := E 0 1 ( 1 − ϑ 2 ϑ ϑ T ϑ ( μ ) ) 1/ ϑ , (52) then, for every finite Λ \Lambda Λ and every y ∈ Λ y\in\Lambda y ∈ Λ ,
sup E ∈ I , η > 0 ∑ x ∈ Λ e μ ∣ x − y ∣ 1 E ∣ ⟨ δ x , ( H ϖ , Λ t − E − i η ) − 1 δ y ⟩ ∣ ϑ < ∞ , (53) \sup_{E\in I,\ \eta>0}\sum_{x\in\Lambda}e^{\mu|x-y|_1}\mathbb{E}\left|\left\langle\delta_x,(H_{\varpi,\Lambda}^t-E-i\eta)^{-1}\delta_y\right\rangle\right|^\vartheta<\infty, \tag*{(53)} E ∈ I , η > 0 sup x ∈ Λ ∑ e μ ∣ x − y ∣ 1 E ⟨ δ x , ( H ϖ , Λ t − E − i η ) − 1 δ y ⟩ ϑ < ∞ , (53) with one bound independent of Λ \Lambda Λ and y y y . Consequently 1 I ( H ϖ ) P a c ( H ϖ ) = 0 \mathbf{1}_I(H_\varpi)P_{\mathrm{ac}}(H_\varpi)=0 1 I ( H ϖ ) P ac ( H ϖ ) = 0 almost surely. An improved two-step sufficient condition is given in Theorem 4.4 .
We first allow general iid inverse factors m n ∈ [ m − , m + ] m_n\in[m_-,m_+] m n ∈ [ m − , m + ] , where 0 < m − ≤ m + < ∞ 0<m_-\le m_+<\infty 0 < m − ≤ m + < ∞ . The uniform family of (8) will then follow explicitly. Write the Green’s function
G ϖ , Λ ( x , y ; z ) = ⟨ δ x , ( H ϖ , Λ t − z ) − 1 δ y ⟩ . G_{\varpi,\Lambda}(x,y;z)=\left\langle\delta_x,(H_{\varpi,\Lambda}^t-z)^{-1}\delta_y\right\rangle. G ϖ , Λ ( x , y ; z ) = ⟨ δ x , ( H ϖ , Λ t − z ) − 1 δ y ⟩ .
Definition 4.2 (Condition (F)). For an open bounded interval I = ( E 0 , E 1 ) ⋐ ( 0 , ∞ ) I=(E_0,E_1)\Subset(0,\infty) I = ( E 0 , E 1 ) ⋐ ( 0 , ∞ ) , condition (F) means that there is ϑ ∈ ( 0 , 1 ) \vartheta\in(0,1) ϑ ∈ ( 0 , 1 ) such that, for every fixed y y y , there is a finite K y F K_y^F K y F with
sup L : y ∈ Λ L sup 0 < η ≤ 1 ∑ x ∈ Λ L E ∣ G ϖ , Λ L ( x , y ; E + i η ) ∣ ϑ ≤ K y F . \sup_{L:y\in\Lambda_L}\sup_{0<\eta\le1}\sum_{x\in\Lambda_L}\mathbb{E}\left|G_{\varpi,\Lambda_L}(x,y;E+i\eta)\right|^\vartheta\le K_y^F. L : y ∈ Λ L sup 0 < η ≤ 1 sup x ∈ Λ L ∑ E ∣ G ϖ , Λ L ( x , y ; E + i η ) ∣ ϑ ≤ K y F . A stronger sufficient condition is
E ∣ G ϖ , Λ ( x , y ; E + i η ) ∣ ϑ ≤ B e − μ ∣ x − y ∣ 1 , \mathbb{E}\left|G_{\varpi,\Lambda}(x,y;E+i\eta)\right|^\vartheta\le Be^{-\mu|x-y|_1}, E ∣ G ϖ , Λ ( x , y ; E + i η ) ∣ ϑ ≤ B e − μ ∣ x − y ∣ 1 , uniformly in finite volume, sites, E ∈ I E\in I E ∈ I , and η > 0 \eta>0 η > 0 .
Remark 4.3. Uniformity both in volume and as η ↓ 0 \eta\downarrow0 η ↓ 0 is indispensable. For a discrete random law, even one finite-volume eigenvalue occurring with positive probability can make the corresponding supremum diverge. In the sufficient tests below we impose a fractional integrability condition on the inverse-material distribution.
The fractional-moment tests
Define
K ϑ = sup a ∈ R E ∣ m n − a ∣ − ϑ < ∞ , κ = K ϑ E 0 − ϑ . (54) K_\vartheta=\sup_{a\in\mathbb{R}}\mathbb{E}|m_n-a|^{-\vartheta}<\infty,\qquad\kappa=K_\vartheta E_0^{-\vartheta}. \tag*{(54)} K ϑ = a ∈ R sup E ∣ m n − a ∣ − ϑ < ∞ , κ = K ϑ E 0 − ϑ . (54)
If m n m_n m n has density bounded by R m R_m R m , then
K ϑ ≤ ( 2 R m ) ϑ 1 − ϑ . (55) K_\vartheta\le\frac{(2R_m)^\vartheta}{1-\vartheta}. \tag*{(55)} K ϑ ≤ 1 − ϑ ( 2 R m ) ϑ . (55)
Indeed, P { ∣ m n − a ∣ < r } ≤ min ( 1 , 2 R m r ) \mathbb{P}\{|m_n-a|<r\}\le\min(1,2R_mr) P { ∣ m n − a ∣ < r } ≤ min ( 1 , 2 R m r ) ; integrate the distribution function of ∣ m n − a ∣ − ϑ |m_n-a|^{-\vartheta} ∣ m n − a ∣ − ϑ over its positive levels. For 0 < μ < ϑ γ 0 0<\mu<\vartheta\gamma_0 0 < μ < ϑ γ 0 , put
T ϑ = ∑ r ≠ 0 ∣ c r ∣ ϑ e μ ∣ r ∣ 1 , U ϑ = ∑ r ≠ 0 ∣ c r ∣ 2 ϑ e 2 μ ∣ r ∣ 1 . (56) T_\vartheta=\sum_{r\ne0}|c_r|^\vartheta e^{\mu|r|_1},\qquad U_\vartheta=\sum_{r\ne0}|c_r|^{2\vartheta}e^{2\mu|r|_1}. \tag*{(56)} T ϑ = r = 0 ∑ ∣ c r ∣ ϑ e μ ∣ r ∣ 1 , U ϑ = r = 0 ∑ ∣ c r ∣ 2 ϑ e 2 μ ∣ r ∣ 1 . (56)
Their dependence on μ \mu μ is suppressed only within this section.
Theorem 4.4 (Uniform fractional-moment estimates). Either of the following conditions implies ( F ) (F) ( F ) and ( F e x p ) (F_{\mathrm{exp}}) ( F exp ) :
(C1) q 1 = κ T ϑ < 1 , q_1=\kappa T_{\vartheta}<1, q 1 = κ T ϑ < 1 ,
(C2) q 2 = κ 2 ( T ϑ 2 − U ϑ ) < 1. q_2=\kappa^2(T_{\vartheta}^2-U_{\vartheta})<1. q 2 = κ 2 ( T ϑ 2 − U ϑ ) < 1.
More precisely, for every finite Λ \Lambda Λ and every y ∈ Λ y\in\Lambda y ∈ Λ ,
sup η > 0 ∑ x ∈ Λ e μ ∣ x − y ∣ E ∣ G ϖ , Λ ( x , y ; E + i η ) ∣ ϑ ≤ B , (57) \sup_{\eta>0}\sum_{x\in\Lambda}e^{\mu|x-y|}\mathbb{E}\lvert G_{\varpi,\Lambda}(x,y;E+\mathrm{i}\eta)\rvert^{\vartheta}\le B, \tag*{(57)} η > 0 sup x ∈ Λ ∑ e μ ∣ x − y ∣ E ∣ G ϖ , Λ ( x , y ; E + i η ) ∣ ϑ ≤ B , (57) where one may take, respectively,
B 1 = m + ϑ κ 1 − q 1 , B 2 = m + ϑ κ ( 1 + κ T ϑ ) 1 − q 2 . (58) B_1=m_+^{\vartheta}\frac{\kappa}{1-q_1},\qquad B_2=m_+^{\vartheta}\frac{\kappa(1+\kappa T_{\vartheta})}{1-q_2}. \tag*{(58)} B 1 = m + ϑ 1 − q 1 κ , B 2 = m + ϑ 1 − q 2 κ ( 1 + κ T ϑ ) . (58) Condition ( C 1 ) (C1) ( C 1 ) implies ( C 2 ) (C2) ( C 2 ) , but the converse need not hold.
Conditional averaging for the capacitance operator
Define
R Λ ( z ) = ( C Λ t − z M ϖ , Λ ) − 1 . (59) R_{\Lambda}(z)=\left(C_{\Lambda}^{\mathsf{t}}-zM_{\varpi,\Lambda}\right)^{-1}. \tag*{(59)} R Λ ( z ) = ( C Λ t − z M ϖ , Λ ) − 1 . (59)
For ℑ z > 0 \Im z>0 ℑ z > 0 this inverse exists, since
( H ϖ , Λ t − z ) − 1 = M ϖ , Λ 1 / 2 R Λ ( z ) M ϖ , Λ 1 / 2 . (60) \left(H_{\varpi,\Lambda}^{\mathsf{t}}-z\right)^{-1}=M_{\varpi,\Lambda}^{1/2}R_{\Lambda}(z)M_{\varpi,\Lambda}^{1/2}. \tag*{(60)} ( H ϖ , Λ t − z ) − 1 = M ϖ , Λ 1/2 R Λ ( z ) M ϖ , Λ 1/2 . (60)
The advantage is that, in (59) , the randomness appears only in the diagonal entries. Fix y ∈ Λ y\in\Lambda y ∈ Λ and D y = Λ ∖ { y } D_y=\Lambda\setminus\{y\} D y = Λ ∖ { y } . The Schur complement gives
R Λ ( y , y ; z ) = ( c 0 − z m y − Σ y ( z ) ) − 1 , Σ y ( z ) = ∑ u , v ∈ D y c y − u R D y ( u , v ; z ) c v − y . (61) R_{\Lambda}(y,y;z)=\left(c_0-zm_y-\Sigma_y(z)\right)^{-1},\qquad\Sigma_y(z)=\sum_{u,v\in D_y}c_{y-u}R_{D_y}(u,v;z)c_{v-y}. \tag*{(61)} R Λ ( y , y ; z ) = ( c 0 − z m y − Σ y ( z ) ) − 1 , Σ y ( z ) = u , v ∈ D y ∑ c y − u R D y ( u , v ; z ) c v − y . (61)
The self-energy Σ y ( z ) \Sigma_y(z) Σ y ( z ) is independent of m y m_y m y . For the possibly complex a = ( c 0 − Σ y ( z ) ) / z a=(c_0-\Sigma_y(z))/z a = ( c 0 − Σ y ( z )) / z , conditioning on all other variables yields
E m y ∣ R Λ ( y , y ; z ) ∣ ϑ = ∣ z ∣ − ϑ E m y ∣ m y − a ∣ − ϑ ≤ ∣ z ∣ − ϑ K ϑ ≤ κ . (62) \mathbb{E}_{m_y}\lvert R_{\Lambda}(y,y;z)\rvert^{\vartheta}=|z|^{-\vartheta}\mathbb{E}_{m_y}|m_y-a|^{-\vartheta}\le|z|^{-\vartheta}K_{\vartheta}\le\kappa. \tag*{(62)} E m y ∣ R Λ ( y , y ; z ) ∣ ϑ = ∣ z ∣ − ϑ E m y ∣ m y − a ∣ − ϑ ≤ ∣ z ∣ − ϑ K ϑ ≤ κ . (62)
The supremum over real centres suffices because ∣ m − a ∣ ≥ ∣ m − ℜ a ∣ |m-a|\ge|m-\Re a| ∣ m − a ∣ ≥ ∣ m − ℜ a ∣ . This estimate is uniform in the conditioned variables and in every depleted finite set. For x ≠ y x\ne y x = y , block inversion gives
R Λ ( x , y ; z ) = − R Λ ( y , y ; z ) ∑ u ∈ D y R D y ( x , u ; z ) c u − y . (63) R_{\Lambda}(x,y;z)=-R_{\Lambda}(y,y;z)\sum_{u\in D_y}R_{D_y}(x,u;z)c_{u-y}. \tag*{(63)} R Λ ( x , y ; z ) = − R Λ ( y , y ; z ) u ∈ D y ∑ R D y ( x , u ; z ) c u − y . (63)
Subadditivity of ∣ ⋅ ∣ ϑ |\cdot|^{\vartheta} ∣ ⋅ ∣ ϑ , followed by conditional averaging in m y m_y m y , proves
E ∣ R Λ ( x , y ; z ) ∣ ϑ ≤ κ ∑ u ∈ D y ∣ c u − y ∣ ϑ E ∣ R D y ( x , u ; z ) ∣ ϑ . (64) \mathbb{E}|R_{\Lambda}(x,y;z)|^{\vartheta}\le\kappa\sum_{u\in D_y}|c_{u-y}|^{\vartheta}\mathbb{E}|R_{D_y}(x,u;z)|^{\vartheta}. \tag*{(64)} E ∣ R Λ ( x , y ; z ) ∣ ϑ ≤ κ u ∈ D y ∑ ∣ c u − y ∣ ϑ E ∣ R D y ( x , u ; z ) ∣ ϑ . (64)
This is the site-deletion fractional-moment argument associated with [2 , 28 ] , applied directly to the multiplicative model.
Closing the one-step and two-step recursions
Put w r = ∣ c r ∣ e μ ∣ r ∣ w_r = |c_r|e^{\mu|r|} w r = ∣ c r ∣ e μ ∣ r ∣ for r ≠ 0 r \ne0 r = 0 , and
S Λ ( y ) = ∑ x ∈ Λ e μ ∣ x − y ∣ ∣ E [ R Λ ( x , y ; z ) ] ∣ ϑ . S_\Lambda(y)=\sum_{x\in\Lambda}e^{\mu|x-y|}\left|\mathbb{E}\left[R_\Lambda(x,y;z)\right]\right|^\vartheta. S Λ ( y ) = x ∈ Λ ∑ e μ ∣ x − y ∣ ∣ E [ R Λ ( x , y ; z ) ] ∣ ϑ .
Equations (62) –(64) and the triangle inequality imply
S Λ ( y ) ≤ κ + κ ∑ u ∈ Λ ∖ { y } w u − y S Λ ∖ { y } ( u ) . (65) S_\Lambda(y)\le\kappa+\kappa\sum_{u\in\Lambda\setminus\{y\}}w_{u-y}S_{\Lambda\setminus\{y\}}(u). \tag*{(65)} S Λ ( y ) ≤ κ + κ u ∈ Λ ∖ { y } ∑ w u − y S Λ ∖ { y } ( u ) . (65)
If (C1) holds, induction on ∣ Λ ∣ |\Lambda| ∣Λ∣ gives S Λ ( y ) ≤ κ / ( 1 − q 1 ) S_\Lambda(y)\le\kappa/(1-q_1) S Λ ( y ) ≤ κ / ( 1 − q 1 ) . The singleton case follows from (62) ; inserting the induction hypothesis into (65) gives the same bound.
For the improved estimate apply (65) again to the depleted set:
S Λ ( y ) ≤ κ + κ 2 T ϑ + κ 2 ∑ u ∈ Λ ∖ { y } w u − y ∑ v ∈ Λ ∖ { y , u } w v − u S Λ ∖ { y , u } ( v ) . (66) \begin{aligned}
S_\Lambda(y)&\le\kappa+\kappa^2T_\vartheta\\
&\quad+\kappa^2\sum_{u\in\Lambda\setminus\{y\}}w_{u-y}\sum_{v\in\Lambda\setminus\{y,u\}}w_{v-u}S_{\Lambda\setminus\{y,u\}}(v). \tag*{(66)}
\end{aligned} S Λ ( y ) ≤ κ + κ 2 T ϑ + κ 2 u ∈ Λ ∖ { y } ∑ w u − y v ∈ Λ ∖ { y , u } ∑ w v − u S Λ ∖ { y , u } ( v ) . (66)
The second step cannot return to y y y , which has been removed. Since w − r = w r w_{-r}=w_r w − r = w r , the full-lattice sum of two-step weights is
∑ u ≠ y w u − y ∑ v ≠ u , y w v − u = T ϑ 2 − ∑ r ≠ 0 w r w − r = T ϑ 2 − U ϑ . (67) \sum_{u\ne y}w_{u-y}\sum_{v\ne u,y}w_{v-u}=T_\vartheta^2-\sum_{r\ne0}w_rw_{-r}=T_\vartheta^2-U_\vartheta. \tag*{(67)} u = y ∑ w u − y v = u , y ∑ w v − u = T ϑ 2 − r = 0 ∑ w r w − r = T ϑ 2 − U ϑ . (67)
Induction on the cardinality now gives S Λ ( y ) ≤ κ ( 1 + κ T ϑ ) / ( 1 − q 2 ) S_\Lambda(y)\le\kappa(1+\kappa T_\vartheta)/(1-q_2) S Λ ( y ) ≤ κ ( 1 + κ T ϑ ) / ( 1 − q 2 ) under (C2): the empty-set remainder is zero, the singleton bound is immediate, and (66) closes with precisely this constant.
Finally,
∣ G ϖ , Λ ( x , y ; z ) ∣ ϑ = ( m x m y ) ϑ / 2 ∣ R Λ ( x , y ; z ) ∣ ϑ ≤ m + ϑ ∣ R Λ ( x , y ; z ) ∣ ϑ . |G_{\varpi,\Lambda}(x,y;z)|^\vartheta=(m_xm_y)^{\vartheta/2}|R_\Lambda(x,y;z)|^\vartheta\le m_+^\vartheta|R_\Lambda(x,y;z)|^\vartheta. ∣ G ϖ , Λ ( x , y ; z ) ∣ ϑ = ( m x m y ) ϑ /2 ∣ R Λ ( x , y ; z ) ∣ ϑ ≤ m + ϑ ∣ R Λ ( x , y ; z ) ∣ ϑ .
This proves Theorem 4.4. Retaining a single summand gives ( F exp ) (F_{\exp}) ( F e x p ) , and dropping the exponential weight gives ( F ) (F) ( F ) .
Remark 4.5 (Longer paths). The same deletion argument can retain paths of length k k k with no repeated vertices. Let Z 0 = 1 Z_0=1 Z 0 = 1 and let Z k ( μ ) Z_k(\mu) Z k ( μ ) be the sum of ∏ j = 1 k w x j − x j − 1 \prod_{j=1}^k w_{x_j-x_{j-1}} ∏ j = 1 k w x j − x j − 1 over distinct vertices x 0 = 0 , x 1 , … , x k x_0=0,x_1,\ldots,x_k x 0 = 0 , x 1 , … , x k . If κ k Z k ( μ ) < 1 \kappa^kZ_k(\mu)<1 κ k Z k ( μ ) < 1 , then (57) holds with
B = m + ϑ ∑ j = 0 k − 1 κ j Z j ( μ ) 1 − κ k Z k ( μ ) . B=m_+^\vartheta\frac{\sum_{j=0}^{k-1}\kappa^jZ_j(\mu)}{1-\kappa^kZ_k(\mu)}. B = m + ϑ 1 − κ k Z k ( μ ) ∑ j = 0 k − 1 κ j Z j ( μ ) .
Indeed the iterated Schur expansion terminates on every finite set, and Z n + k ≤ Z n Z k Z_{n+k}\le Z_nZ_k Z n + k ≤ Z n Z k by discarding restrictions between a prefix and its translated suffix. Group the resulting nonnegative path series in blocks of length k k k . The cases k = 1 , 2 k=1,2 k = 1 , 2 are the estimates just proved. Related self-avoiding-walk refinements for additive Anderson models are discussed in [26 ] .
From (F) to absence of absolutely continuous spectrum
Theorem 4.6. If ( F ) (F) ( F ) holds for the truncated matrices of H ϖ H_\varpi H ϖ , then
1 I ( H ϖ ) P a c ( H ϖ ) = 0 almost surely . (68) 1_I(H_\varpi)P_{\mathrm{ac}}(H_\varpi)=0\quad\text{almost surely}. \tag*{(68)} 1 I ( H ϖ ) P ac ( H ϖ ) = 0 almost surely . (68)
Equivalently, A ϖ = B ϖ C A_\varpi=B_\varpi C A ϖ = B ϖ C , considered on H ϖ \mathcal{H}_\varpi H ϖ , has no absolutely continuous spectral component in I I I .
Proof of Theorem 4.6. Infinite-volume resolvent. Let P L = E Λ L E Λ L ∗ P_L=E_{\Lambda_L}E_{\Lambda_L}^* P L = E Λ L E Λ L ∗ . Extend the finite restriction by zero outside Λ L \Lambda_L Λ L : H ^ ϖ , L = P L H ϖ P L \widehat{H}_{\varpi,L}=P_LH_\varpi P_L H ϖ , L = P L H ϖ P L on ℓ 2 ( Z d ) \ell^2(\mathbb{Z}^d) ℓ 2 ( Z d ) . The uniform bound ∥ H ϖ ∥ ≤ m − − 1 ∥ C ∥ \|H_\varpi\|\le m_-^{-1}\|C\| ∥ H ϖ ∥ ≤ m − − 1 ∥ C ∥ and P L → Id P_L\to\operatorname{Id} P L → Id strongly imply H ^ ϖ , L → H ϖ \widehat{H}_{\varpi,L}\to H_\varpi H ϖ , L → H ϖ strongly. For ℑ z > 0 \Im z>0 ℑ z > 0 , the resolvent identity and the bound ∥ ( H ^ ϖ , L − z ) − 1 ∥ ≤ ( ℑ z ) − 1 \|(\widehat{H}_{\varpi,L}-z)^{-1}\|\le(\Im z)^{-1} ∥ ( H ϖ , L − z ) − 1 ∥ ≤ ( ℑ z ) − 1 give strong resolvent convergence. For fixed sites and sufficiently large L L L , the matrix elements of this extended resolvent equal those of the truncated matrix. Thus
G ϖ , Λ L ( x , y ; z ) ⟶ G ϖ ( x , y ; z ) : = ⟨ δ x , ( H ϖ − z ) − 1 δ y ⟩ . G_{\varpi,\Lambda_L}(x,y;z) \longrightarrow G_{\varpi}(x,y;z) := \langle\delta_x,(H_{\varpi}-z)^{-1}\delta_y\rangle. G ϖ , Λ L ( x , y ; z ) ⟶ G ϖ ( x , y ; z ) := ⟨ δ x , ( H ϖ − z ) − 1 δ y ⟩ .
Extend finite-volume summands by zero when x ∉ Λ L x \notin\Lambda_L x ∈ / Λ L . Fatou’s lemma over counting measure and probability gives, for each E ∈ I E \in I E ∈ I and 0 < η ≤ 1 0 < \eta\le1 0 < η ≤ 1 ,
E ∑ x ∈ Z d ∣ G ϖ ( x , y ; E + i η ) ∣ ϑ ≤ K y F . (69) \mathbb{E}\sum_{x\in\mathbb{Z}^d}\lvert G_{\varpi}(x,y;E+i\eta)\rvert^{\vartheta} \le K_y^F. \tag*{(69)} E x ∈ Z d ∑ ∣ G ϖ ( x , y ; E + i η ) ∣ ϑ ≤ K y F . (69)
Boundary values. For fixed y y y , put
S k ( ϖ , E ) = ∑ x ∣ G ϖ ( x , y ; E + i / k ) ∣ ϑ . S_k(\varpi,E)=\sum_x\lvert G_{\varpi}(x,y;E+i/k)\rvert^{\vartheta}. S k ( ϖ , E ) = x ∑ ∣ G ϖ ( x , y ; E + i / k ) ∣ ϑ .
These functions are nonnegative and jointly measurable. Integrate (69) over I I I and apply Fatou’s lemma on the product of probability and Lebesgue measure:
E ∫ I lim inf k → ∞ S k ( ϖ , E ) d E ≤ K y F ∣ I ∣ < ∞ . (70) \mathbb{E}\int_I\liminf_{k\to\infty}S_k(\varpi,E)\,dE\le K_y^F\lvert I\rvert<\infty. \tag*{(70)} E ∫ I k → ∞ lim inf S k ( ϖ , E ) d E ≤ K y F ∣ I ∣ < ∞. (70)
Therefore, for almost every realisation and Lebesgue-almost every E ∈ I E\in I E ∈ I , there is a subsequence k j → ∞ k_j\to\infty k j → ∞ for which S k j ( ϖ , E ) ≤ L < ∞ S_{k_j}(\varpi,E)\le L<\infty S k j ( ϖ , E ) ≤ L < ∞ . The subsequence and L L L may depend on ϖ , E , y \varpi,E,y ϖ , E , y . For 0 < ϑ < 2 0<\vartheta<2 0 < ϑ < 2 , any sequence satisfies
∑ x ∣ b x ∣ 2 ≤ ( ∑ x ∣ b x ∣ ϑ ) 2 / ϑ . \sum_x\lvert b_x\rvert^2\le\left(\sum_x\lvert b_x\rvert^{\vartheta}\right)^{2/\vartheta}. x ∑ ∣ b x ∣ 2 ≤ ( x ∑ ∣ b x ∣ ϑ ) 2/ ϑ .
The self-adjoint resolvent identity gives
ℑ G ϖ ( y , y ; E + i η ) = η ∑ x ∣ G ϖ ( x , y ; E + i η ) ∣ 2 . (71) \Im G_{\varpi}(y,y;E+i\eta)=\eta\sum_x\lvert G_{\varpi}(x,y;E+i\eta)\rvert^2. \tag*{(71)} ℑ G ϖ ( y , y ; E + i η ) = η x ∑ ∣ G ϖ ( x , y ; E + i η ) ∣ 2 . (71)
Consequently
0 ≤ ℑ G ϖ ( y , y ; E + i / k j ) ≤ k j − 1 L 2 / ϑ ⟶ 0. (72) 0\le\Im G_{\varpi}(y,y;E+i/k_j)\le k_j^{-1}L^{2/\vartheta}\longrightarrow0. \tag*{(72)} 0 ≤ ℑ G ϖ ( y , y ; E + i / k j ) ≤ k j − 1 L 2/ ϑ ⟶ 0. (72)
Let ν ϖ , y \nu_{\varpi,y} ν ϖ , y be the spectral measure of δ y \delta_y δ y . Its Stieltjes transform is G ϖ ( y , y ; z ) G_{\varpi}(y,y;z) G ϖ ( y , y ; z ) , and the Poisson boundary-value formula states
d ν ϖ , y d E ( E ) = 1 π lim η ↓ 0 ℑ G ϖ ( y , y ; E + i η ) (73) \frac{d\nu_{\varpi,y}}{dE}(E)=\frac{1}{\pi}\lim_{\eta\downarrow0}\Im G_{\varpi}(y,y;E+i\eta) \tag*{(73)} d E d ν ϖ , y ( E ) = π 1 η ↓ 0 lim ℑ G ϖ ( y , y ; E + i η ) (73)
for Lebesgue-almost every E E E . See [29 Theorems 3.26–3.27] . At energies where this full limit exists, it equals the subsequential limit in (72) . Hence ν ϖ , y a c ( I ) = 0 \nu_{\varpi,y}^{\mathrm{ac}}(I)=0 ν ϖ , y ac ( I ) = 0 almost surely. Intersect the probability-one events over the countably many sites. The projection P = 1 I ( H ϖ ) P a c ( H ϖ ) P=1_I(H_{\varpi})P_{\mathrm{ac}}(H_{\varpi}) P = 1 I ( H ϖ ) P ac ( H ϖ ) then satisfies ∥ P δ y ∥ 2 = ν ϖ , y a c ( I ) = 0 \lVert P\delta_y\rVert^2=\nu_{\varpi,y}^{\mathrm{ac}}(I)=0 ∥ P δ y ∥ 2 = ν ϖ , y ac ( I ) = 0 for every y y y . The coordinate vectors span a dense subspace, so P = 0 P=0 P = 0 . Unitary equivalence (9) proves the assertion. □
Remark 4.7. Condition (F) excludes absolutely continuous spectrum. For the uniform inverse-material law, § 3.1 combines the stronger exponential estimates with spectral averaging to prove spectral and dynamical localisation.
Completion of the proof of Theorem 4.1 . For m n m_n m n uniform on [ 1 , 1 + W ] [1,1+W] [ 1 , 1 + W ] , integration about the midpoint gives
K ϑ = sup a ∈ R 1 W ∫ 1 1 + W ∣ m − a ∣ − ϑ d m = 2 ϑ 1 − ϑ W − ϑ . K_{\vartheta}=\sup_{a\in\mathbb{R}}\frac{1}{W}\int_1^{1+W}\lvert m-a\rvert^{-\vartheta}\,dm=\frac{2^{\vartheta}}{1-\vartheta}W^{-\vartheta}. K ϑ = a ∈ R sup W 1 ∫ 1 1 + W ∣ m − a ∣ − ϑ d m = 1 − ϑ 2 ϑ W − ϑ .
Indeed the integral increases up to the midpoint and decreases afterwards, and its value outside the interval is no larger. Substitution into (C1) gives exactly (52) . Theorem 4.4 gives (53) , with B 1 = ( 1 + W ) ϑ κ / ( 1 − κ T ϑ ) B_1=(1+W)^{\vartheta}\kappa/(1-\kappa T_{\vartheta}) B 1 = ( 1 + W ) ϑ κ / ( 1 − κ T ϑ ) ; Theorem 4.6 then gives the absence conclusion. □
Uniform estimates near the upper band edge
We establish fractional-moment estimates near the upper spectral edge for every fixed positive disorder strength, with bounds uniform over finite truncations. Throughout this section we assume (R) and the material law (2.5).
The deficit operator and a Laplacian comparison
Define
( J q ) n = ( − 1 ) n 1 + ⋯ + n d q n , K = c M Id − J C J , k ( α ) = c M − c ( α M + α ) . (74) (Jq)_n = (-1)^{n_1+\cdots+n_d}q_n,\qquad K=c_M\operatorname{Id}-JCJ,\qquad k(\alpha)=c_M-c(\alpha_M+\alpha). \tag*{(74)} ( J q ) n = ( − 1 ) n 1 + ⋯ + n d q n , K = c M Id − J C J , k ( α ) = c M − c ( α M + α ) . (74)
The unitary J J J shifts quasimomentum by α M \alpha_M α M . Thus K ≥ 0 K\ge0 K ≥ 0 , its symbol vanishes only at zero, and its off-diagonal entries have the same absolute values as those of C C C . Let L L L be the positive nearest-neighbour Laplacian, with symbol ℓ \ell ℓ :
( L q ) n = ∑ j = 1 d ( 2 q n − q n + e j − q n − e j ) , ℓ ( α ) = 2 ∑ j = 1 d ( 1 − cos α j ) . (75) (Lq)_n=\sum_{j=1}^{d}(2q_n-q_{n+e_j}-q_{n-e_j}),\qquad\ell(\alpha)=2\sum_{j=1}^{d}(1-\cos\alpha_j). \tag*{(75)} ( L q ) n = j = 1 ∑ d ( 2 q n − q n + e j − q n − e j ) , ℓ ( α ) = 2 j = 1 ∑ d ( 1 − cos α j ) . (75)
By Proposition 3.6, k ( α ) ≥ a L ℓ ( α ) k(\alpha)\ge a_L\ell(\alpha) k ( α ) ≥ a L ℓ ( α ) near zero for some a L > 0 a_L>0 a L > 0 . Away from zero both functions are continuous and k k k is positive. Decreasing a L a_L a L gives the global inequality
k ( α ) ≥ a L ℓ ( α ) , K ≥ a L L . (76) k(\alpha)\ge a_L\ell(\alpha),\qquad K\ge a_LL. \tag*{(76)} k ( α ) ≥ a L ℓ ( α ) , K ≥ a L L . (76)
A uniform initial finite-volume estimate
Set Ξ ϖ = diag ( ξ n ) \Xi_\varpi=\operatorname{diag}(\xi_n) Ξ ϖ = diag ( ξ n ) and consider
L g , ϖ = K + g Ξ ϖ , g ∈ [ g − , g + ] , 0 < g − ≤ g + < ∞ . (77) L_{g,\varpi}=K+g\Xi_\varpi,\qquad g\in[g_-,g_+],\quad0<g_-\le g_+<\infty. \tag*{(77)} L g , ϖ = K + g Ξ ϖ , g ∈ [ g − , g + ] , 0 < g − ≤ g + < ∞. (77)
Throughout, the restriction of an operator to a finite set Λ \Lambda Λ always means the compressed operator, e.g., K Λ = E Λ ∗ K E Λ K_\Lambda=E_\Lambda^*KE_\Lambda K Λ = E Λ ∗ K E Λ and L g , ϖ , Λ = K Λ + g Ξ ϖ , Λ L_{g,\varpi,\Lambda}=K_\Lambda+g\Xi_{\varpi,\Lambda} L g , ϖ , Λ = K Λ + g Ξ ϖ , Λ on ℓ 2 ( Λ ) \ell^2(\Lambda) ℓ 2 ( Λ ) . In particular K Λ ≥ 0 K_\Lambda\ge0 K Λ ≥ 0 . This compression retains the interactions of the infinite operator; the finite capacitance matrix is defined separately in § 5.1 .
Proposition 5.1. Fix β ∈ ( 0 , 1 ) \beta\in(0,1) β ∈ ( 0 , 1 ) and write
e L = L − β , ℓ L = ⌊ L β / 4 ⌋ . e_L=L^{-\beta},\qquad\ell_L=\left\lfloor L^{\beta/4}\right\rfloor. e L = L − β , ℓ L = ⌊ L β /4 ⌋ . There are C , c > 0 C,c>0 C , c > 0 and L 0 < ∞ L_0<\infty L 0 < ∞ such that, for integers L ≥ L 0 L\ge L_0 L ≥ L 0 ,
p L : = sup g ∈ [ g − , g + ] P { λ min ( L g , ϖ , Λ L ) ≤ e L } ≤ C L d exp ( − c L β d / 4 ) . (78) p_L:=\sup_{g\in[g_-,g_+]}\mathbb{P}\{\lambda_{\min}(L_{g,\varpi,\Lambda_L})\le e_L\}\le CL^d\exp(-cL^{\beta d/4}). \tag*{(78)} p L := g ∈ [ g − , g + ] sup P { λ m i n ( L g , ϖ , Λ L ) ≤ e L } ≤ C L d exp ( − c L β d /4 ) . (78) The starting scale L 0 L_0 L 0 may diverge as g − ↓ 0 g_-\downarrow0 g − ↓ 0 .
Proof of Proposition 5.1. Laplacian comparison. Suppose ψ \psi ψ is a normalised eigenvector for L g , ϖ , Λ L L_{g,\varpi,\Lambda_L} L g , ϖ , Λ L with the eigenvalue at most e L e_L e L , and extend it by zero to Z d \mathbb{Z}^d Z d . Nonnegativity gives separately
⟨ ψ , K ψ ⟩ ≤ e L , g ⟨ ψ , Ξ ϖ ψ ⟩ ≤ e L . \langle\psi,K\psi\rangle\le e_L,\qquad g\langle\psi,\Xi_\varpi\psi\rangle\le e_L. ⟨ ψ , K ψ ⟩ ≤ e L , g ⟨ ψ , Ξ ϖ ψ ⟩ ≤ e L .
The comparison K ≥ a L L K\ge a_LL K ≥ a L L gives ⟨ ψ , L ψ ⟩ ≤ e L / a L \langle\psi,L\psi\rangle\le e_L/a_L ⟨ ψ , L ψ ⟩ ≤ e L / a L . Therefore
λ min ( L Λ L D + g Ξ ϖ , Λ L ) ≤ h L : = ( 1 + a L − 1 ) e L . (79) \lambda_{\min}(L^D_{\Lambda_L}+g\Xi_{\varpi,\Lambda_L})\le h_L:=(1+a_L^{-1})e_L. \tag*{(79)} λ m i n ( L Λ L D + g Ξ ϖ , Λ L ) ≤ h L := ( 1 + a L − 1 ) e L . (79)
Here L Λ L D = E Λ L ∗ L E Λ L L^D_{\Lambda_L}=E_{\Lambda_L}^*LE_{\Lambda_L} L Λ L D = E Λ L ∗ L E Λ L is the Dirichlet compression, or equivalently the operator with zero extension outside the cube.
Neumann bracketing. Partition each coordinate interval { − L , … , L } \{-L,\ldots,L\} { − L , … , L } into consecutive subintervals containing between ℓ L \ell_L ℓ L and 2 ℓ L 2\ell_L 2 ℓ L sites. For sufficiently large L L L , this can be done by taking subintervals of ℓ L \ell_L ℓ L sites and merging any shorter final subinterval with its predecessor. Let B L \mathcal{B}_L B L denote the resulting partition of Λ L \Lambda_L Λ L into rectangular blocks, obtained by taking Cartesian products of these coordinate subintervals. Each edge contributes a nonnegative term to the Laplacian energy. Discarding the terms associated with edges connecting distinct blocks, together with those crossing the outer Dirichlet boundary of Λ L \Lambda_L Λ L , therefore gives
L Λ L D + g Ξ ‾ ϖ , Λ L ≥ ⨁ B ∈ B L ( L B N + g Ξ ‾ ϖ , B ) . (80) L_{\Lambda_L}^{D}+g\overline{\Xi}_{\varpi,\Lambda_L}\geq\bigoplus_{B\in\mathcal{B}_L}\left(L_B^{N}+g\overline{\Xi}_{\varpi,B}\right). \tag*{(80)} L Λ L D + g Ξ ϖ , Λ L ≥ B ∈ B L ⨁ ( L B N + g Ξ ϖ , B ) . (80)
where L B N L_B^{N} L B N is the Neumann graph Laplacian on B B B , containing only edges whose endpoints both lie in B B B . The Neumann graph Laplacian L B N L_B^{N} L B N has a normalised constant ground state ϕ B \phi_B ϕ B and first positive eigenvalue
γ B = min j { 2 − 2 cos ( π / N j ) } ≥ ℓ L − 2 , ℓ L ≤ N j ≤ 2 ℓ L . \gamma_B=\min_j\{2-2\cos(\pi/N_j)\}\geq\ell_L^{-2},\qquad\ell_L\leq N_j\leq2\ell_L. γ B = j min { 2 − 2 cos ( π / N j )} ≥ ℓ L − 2 , ℓ L ≤ N j ≤ 2 ℓ L .
The inequality follows from 2 − 2 cos ( π / N ) ≥ 4 / N 2 2-2\cos(\pi/N)\geq4/N^2 2 − 2 cos ( π / N ) ≥ 4/ N 2 . Put
t L = 1 4 ℓ L 2 , X n = 1 { ξ n ≥ 1 / 2 } , X ‾ B = ∣ B ∣ − 1 ∑ n ∈ B X n . (81) t_L=\frac{1}{4\ell_L^2},\qquad X_n=\mathbf{1}_{\{\xi_n\geq1/2\}},\qquad\overline{X}_B=|B|^{-1}\sum_{n\in B}X_n. \tag*{(81)} t L = 4 ℓ L 2 1 , X n = 1 { ξ n ≥ 1/2 } , X B = ∣ B ∣ − 1 n ∈ B ∑ X n . (81)
For large enough L L L , t L ≤ g − / 2 t_L\leq g_-/2 t L ≤ g − /2 , and thus g ξ n ≥ t L X n g\xi_n\geq t_LX_n g ξ n ≥ t L X n .
We record the needed eigenvalue bound with its proof. If a finite self-adjoint matrix T T T has eigenvalues λ 1 ≤ λ 2 ≤ ⋯ \lambda_1\leq\lambda_2\leq\cdots λ 1 ≤ λ 2 ≤ ⋯ and λ 2 ≥ γ \lambda_2\geq\gamma λ 2 ≥ γ , and a normalised trial vector has mean μ = ⟨ T ϕ , ϕ ⟩ < γ \mu=\langle T\phi,\phi\rangle<\gamma μ = ⟨ T ϕ , ϕ ⟩ < γ and variance σ 2 = ∥ ( T − μ ) ϕ ∥ 2 \sigma^2=\|(T-\mu)\phi\|^2 σ 2 = ∥ ( T − μ ) ϕ ∥ 2 , then
λ 1 ≥ μ − σ 2 γ − μ . (82) \lambda_1\geq\mu-\frac{\sigma^2}{\gamma-\mu}. \tag*{(82)} λ 1 ≥ μ − γ − μ σ 2 . (82)
Indeed, the spectral expansion gives ⟨ ( T − λ 1 ) ( T − γ ) ϕ , ϕ ⟩ ≥ 0 \langle(T-\lambda_1)(T-\gamma)\phi,\phi\rangle\geq0 ⟨( T − λ 1 ) ( T − γ ) ϕ , ϕ ⟩ ≥ 0 ; rearranging this inequality proves (82) .
Apply it to T = L B N + t L diag ( X n ) T=L_B^{N}+t_L\operatorname{diag}(X_n) T = L B N + t L diag ( X n ) and ϕ = ϕ B \phi=\phi_B ϕ = ϕ B . The min–max principle gives λ 2 ( T ) ≥ γ B ≥ 4 t L \lambda_2(T)\geq\gamma_B\geq4t_L λ 2 ( T ) ≥ γ B ≥ 4 t L , whereas
μ = t L X ‾ B , σ 2 = t L 2 X ‾ B ( 1 − X ‾ B ) . \mu=t_L\overline{X}_B,\qquad\sigma^2=t_L^2\overline{X}_B(1-\overline{X}_B). μ = t L X B , σ 2 = t L 2 X B ( 1 − X B ) .
Therefore
λ min ( L B N + g Ξ ‾ ϖ , B ) ≥ 2 3 t L X ‾ B . (83) \lambda_{\min}(L_B^{N}+g\overline{\Xi}_{\varpi,B})\geq\frac{2}{3}t_L\overline{X}_B. \tag*{(83)} λ m i n ( L B N + g Ξ ϖ , B ) ≥ 3 2 t L X B . (83)
If X ‾ B ≥ 1 / 4 \overline{X}_B\geq1/4 X B ≥ 1/4 , this is at least t L / 6 t_L/6 t L /6 . On the other hand,
h L = O ( L − β ) , t L ≍ L − β / 2 , h_L=O(L^{-\beta}),\qquad t_L\asymp L^{-\beta/2}, h L = O ( L − β ) , t L ≍ L − β /2 ,
so h L < t L / 6 h_L<t_L/6 h L < t L /6 for all sufficiently large L L L . Equations (79) –(80) imply that a low eigenvalue requires X ‾ B < 1 / 4 \overline{X}_B<1/4 X B < 1/4 on at least one block.
Probability of a bad block. The variables X n X_n X n are independent Bernoulli variables of mean 1 / 2 1/2 1/2 . For N = ∣ B ∣ N=|B| N = ∣ B ∣ , exponential Markov inequality gives
P { ∑ n ∈ B ( X n − 1 / 2 ) < − N / 4 } ≤ e − N / 4 ( cosh ( 1 / 2 ) ) N ≤ e − N / 8 , \mathbb{P}\left\{\sum_{n\in B}(X_n-1/2)<-N/4\right\}\leq e^{-N/4}\left(\cosh(1/2)\right)^N\leq e^{-N/8}, P { n ∈ B ∑ ( X n − 1/2 ) < − N /4 } ≤ e − N /4 ( cosh ( 1/2 ) ) N ≤ e − N /8 ,
where cosh x ≤ e x 2 / 2 \cosh x\leq e^{x^2/2} cosh x ≤ e x 2 /2 was used. Every block has at least ( ℓ L ) d (\ell_L)^d ( ℓ L ) d sites and there are at most ( 2 L + 1 ) d (2L+1)^d ( 2 L + 1 ) d blocks. A union bound yields
p L ≤ ( 2 L + 1 ) d e − ( ℓ L ) d / 8 . p_L\leq(2L+1)^d e^{-(\ell_L)^d/8}. p L ≤ ( 2 L + 1 ) d e − ( ℓ L ) d /8 .
This proves the stated initial-scale estimate. □
Verification of the fractional-moment block criterion
We verify a finite-volume criterion with constants uniform for g ∈ [ g − , g + ] g\in[g_-,g_+] g ∈ [ g − , g + ] .
Spectral averaging and a finite-volume moment bound.
Lemma 5.2. If Λ \Lambda Λ has N N N sites and J 0 J_0 J 0 is a bounded interval, we have the Wegner estimate
E Tr 1 J 0 ( L g , ϖ , Λ ) ≤ N ∣ J 0 ∣ g . (84) \mathbb{E}\operatorname{Tr}1_{J_0}(L_{g,\varpi,\Lambda}) \le\frac{N|J_0|}{g}. \tag*{(84)} E Tr 1 J 0 ( L g , ϖ , Λ ) ≤ g N ∣ J 0 ∣ . (84) Consequently, for every real e e e , 0 < t < 1 0<t<1 0 < t < 1 , and η ≠ 0 \eta\ne0 η = 0 ,
E ∥ ( L g , ϖ , Λ − e − i η ) − 1 ∥ t ≤ D t ( N , g ) : = ( 2 N / g ) t 1 − t . (85) \mathbb{E}\left\|(L_{g,\varpi,\Lambda}-e-i\eta)^{-1}\right\|^t \le D_t(N,g):=\frac{(2N/g)^t}{1-t}. \tag*{(85)} E ( L g , ϖ , Λ − e − i η ) − 1 t ≤ D t ( N , g ) := 1 − t ( 2 N / g ) t . (85)
Proof. Fix n ∈ Λ n\in\Lambda n ∈ Λ . For a deterministic parameter v ∈ [ 0 , g ] v\in[0,g] v ∈ [ 0 , g ] , let L g , ϖ , Λ ( v ) L_{g,\varpi,\Lambda}(v) L g , ϖ , Λ ( v ) denote the matrix obtained from L g , ϖ , Λ L_{g,\varpi,\Lambda} L g , ϖ , Λ by replacing the diagonal potential g ξ n g\xi_n g ξ n at site n n n with v v v , while retaining all other potentials. Define
F n , v ( z ) : = ⟨ δ n , ( L g , ϖ , Λ ( v ) − z ) − 1 δ n ⟩ , ν n , v ( B ) : = ⟨ δ n , 1 B ( L g , ϖ , Λ ( v ) ) δ n ⟩ . F_{n,v}(z):=\left\langle\delta_n,(L_{g,\varpi,\Lambda}(v)-z)^{-1}\delta_n\right\rangle,\qquad\nu_{n,v}(B):=\left\langle\delta_n,1_B(L_{g,\varpi,\Lambda}(v))\delta_n\right\rangle. F n , v ( z ) := ⟨ δ n , ( L g , ϖ , Λ ( v ) − z ) − 1 δ n ⟩ , ν n , v ( B ) := ⟨ δ n , 1 B ( L g , ϖ , Λ ( v )) δ n ⟩ .
Put D n = Λ ∖ { n } D_n=\Lambda\setminus\{n\} D n = Λ ∖ { n } and b n = ( K m n ) m ∈ D n b_n=(K_{mn})_{m\in D_n} b n = ( K mn ) m ∈ D n . For z = E + i η z=E+i\eta z = E + i η , η > 0 \eta>0 η > 0 , the Schur-complement gives
F n , v ( z ) = 1 v − a ( z ) , a ( z ) = z − K n n + b n ∗ ( L g , ϖ , D n − z ) − 1 b n F_{n,v}(z)=\frac{1}{v-a(z)},\qquad a(z)=z-K_{nn}+b_n^*(L_{g,\varpi,D_n}-z)^{-1}b_n F n , v ( z ) = v − a ( z ) 1 , a ( z ) = z − K nn + b n ∗ ( L g , ϖ , D n − z ) − 1 b n
with ℑ a ( z ) > 0 \Im a(z)>0 ℑ a ( z ) > 0 . Thus,
∫ 0 g ℑ F n , v ( z ) d v = ∫ 0 g ℑ 1 v − a ( z ) d v ≤ π . (86) \int_0^g\Im F_{n,v}(z)\,dv=\int_0^g\Im\frac{1}{v-a(z)}\,dv\le\pi. \tag*{(86)} ∫ 0 g ℑ F n , v ( z ) d v = ∫ 0 g ℑ v − a ( z ) 1 d v ≤ π . (86)
We next derive the corresponding spectral-measure estimate. Write a = inf J 0 a=\inf J_0 a = inf J 0 , b = sup J 0 b=\sup J_0 b = sup J 0 , and fix ε > 0 \varepsilon>0 ε > 0 . For every fixed configuration of the other variables and every fixed v v v , Stone’s formula gives
lim k → ∞ 1 π ∫ a − ε b + ε ℑ F n , v ( E + i / k ) d E = ν n , v ( ( a − ε , b + ε ) ) + 1 2 ν n , v ( { a − ε } ) + 1 2 ν n , v ( { b + ε } ) ≥ ν n , v ( J 0 ) . (87) \lim_{k\to\infty}\frac{1}{\pi}\int_{a-\varepsilon}^{b+\varepsilon}\Im F_{n,v}(E+i/k)\,dE
=\nu_{n,v}((a-\varepsilon,b+\varepsilon))+\frac{1}{2}\nu_{n,v}(\{a-\varepsilon\})+\frac{1}{2}\nu_{n,v}(\{b+\varepsilon\})\ge\nu_{n,v}(J_0). \tag*{(87)} k → ∞ lim π 1 ∫ a − ε b + ε ℑ F n , v ( E + i / k ) d E = ν n , v (( a − ε , b + ε )) + 2 1 ν n , v ({ a − ε }) + 2 1 ν n , v ({ b + ε }) ≥ ν n , v ( J 0 ) . (87)
Set
F n : = σ ( ξ m : m ∈ Λ ∖ { n } ) , E n [ ⋅ ] : = E [ ⋅ ∣ F n ] , V n : = g ξ n . \mathcal{F}_n:=\sigma(\xi_m:m\in\Lambda\setminus\{n\}),\qquad\mathbb{E}_n[\cdot]:=\mathbb{E}[\cdot\mid\mathcal{F}_n],\qquad V_n:=g\xi_n. F n := σ ( ξ m : m ∈ Λ ∖ { n }) , E n [ ⋅ ] := E [ ⋅ ∣ F n ] , V n := g ξ n .
The random variable V n V_n V n is independent of F n \mathcal{F}_n F n and is uniformly distributed on [ 0 , g ] [0,g] [ 0 , g ] . Fatou’s lemma and (86) –(87) imply
E n [ ν n , V n ( J 0 ) ] = 1 g ∫ 0 g ν n , v ( J 0 ) d v ≤ lim inf k → ∞ 1 π g ∫ a − ε b + ε ∫ 0 g ℑ F n , v ( E + i / k ) d v d E ≤ ∣ J 0 ∣ + 2 ε g . \begin{aligned}
\mathbb{E}_n[\nu_{n,V_n}(J_0)]=\frac{1}{g}\int_0^g\nu_{n,v}(J_0)\,dv \\
&\le\liminf_{k\to\infty}\frac{1}{\pi g}\int_{a-\varepsilon}^{b+\varepsilon}\int_0^g\Im F_{n,v}(E+i/k)\,dv\,dE \\
&\le\frac{|J_0|+2\varepsilon}{g}.
\end{aligned} E n [ ν n , V n ( J 0 )] = g 1 ∫ 0 g ν n , v ( J 0 ) d v ≤ k → ∞ lim inf π g 1 ∫ a − ε b + ε ∫ 0 g ℑ F n , v ( E + i / k ) d v d E ≤ g ∣ J 0 ∣ + 2 ε .
Letting ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 yields E n [ ν n , V n ( J 0 ) ] ≤ ∣ J 0 ∣ / g \mathbb{E}_n[\nu_{n,V_n}(J_0)]\le|J_0|/g E n [ ν n , V n ( J 0 )] ≤ ∣ J 0 ∣/ g . Using the tower property and summing over n ∈ Λ n\in\Lambda n ∈ Λ , we obtain
E Tr 1 J 0 ( L g , ϖ , Λ ) = ∑ n ∈ Λ E [ ν n , V n ( J 0 ) ] = ∑ n ∈ Λ E [ E n [ ν n , V n ( J 0 ) ] ] ≤ N ∣ J 0 ∣ g . \mathbb{E}\operatorname{Tr}1_{J_0}(L_{g,\varpi,\Lambda})
=\sum_{n\in\Lambda}\mathbb{E}[\nu_{n,V_n}(J_0)]
=\sum_{n\in\Lambda}\mathbb{E}[\mathbb{E}_n[\nu_{n,V_n}(J_0)]]
\le\frac{N|J_0|}{g}. E Tr 1 J 0 ( L g , ϖ , Λ ) = n ∈ Λ ∑ E [ ν n , V n ( J 0 )] = n ∈ Λ ∑ E [ E n [ ν n , V n ( J 0 )]] ≤ g N ∣ J 0 ∣ .
This proves (84) . For r > 0 r>0 r > 0 it implies
P { dist ( e , σ ( L g , ϖ , Λ ) ) < r } ≤ min { 1 , 2 N r / g } . \mathbb{P}\{\operatorname{dist}(\mathbf{e},\sigma(L_{g,\varpi,\Lambda}))<r\}\leq\min\{1,2Nr/g\}. P { dist ( e , σ ( L g , ϖ , Λ )) < r } ≤ min { 1 , 2 N r / g } .
Integrating the tail of the negative t t t -th power of this distance gives ( 2 N / g ) t / ( 1 − t ) (2N/g)^t/(1-t) ( 2 N / g ) t / ( 1 − t ) . Finally the complex resolvent norm is bounded by the inverse real spectral distance. In particular each fixed real e \mathbf{e} e is almost surely outside the spectrum of this truncated matrix. □
An exponential estimate on the good event
Let G g , Λ ( x , y ; z ) G_{g,\Lambda}(x,y;z) G g , Λ ( x , y ; z ) denote the integral kernel of the finite-volume resolvent of L g , ϖ L_{g,\varpi} L g , ϖ . Let γ 0 > 0 \gamma_0>0 γ 0 > 0 be the exponential decay rate of K K K , which may not be the optimal one. As a consequence, the following constant is finite:
T 1 = ∑ r ≠ 0 ∣ K 0 r ∣ ∣ r ∣ 1 e γ 0 ∣ r ∣ 1 / 2 . (88) T_1=\sum_{r\neq0}|K_{0r}|\,|r|_1e^{\gamma_0|r|_1/2}. \tag*{(88)} T 1 = r = 0 ∑ ∣ K 0 r ∣ ∣ r ∣ 1 e γ 0 ∣ r ∣ 1 /2 . (88)
For 0 < ν ≤ γ 0 / 2 0<\nu\leq\gamma_0/2 0 < ν ≤ γ 0 /2 , conjugating by e ν ∣ x − y ∣ 1 e^{\nu|x-y|_1} e ν ∣ x − y ∣ 1 changes the operator by norm at most ν T 1 \nu T_1 ν T 1 , by the row and column sum bounds. The diagonal random potential is unchanged by this conjugation. On the good event { λ min ( L g , ϖ , Λ L ) > e L } \{\lambda_{\min}(L_{g,\varpi,\Lambda_L})>e_L\} { λ m i n ( L g , ϖ , Λ L ) > e L } , for ∣ e ∣ ≤ e L / 2 |\mathbf{e}|\leq e_L/2 ∣ e ∣ ≤ e L /2 the resolvent norm is at most 2 / e L 2/e_L 2/ e L , uniformly in η ≠ 0 \eta\neq0 η = 0 . Choose
b = min { γ 0 / 2 , [ 4 ( T 1 + 1 ) ] − 1 } , ν = b e L , b=\min\{\gamma_0/2,[4(T_1+1)]^{-1}\},\qquad\nu=be_L, b = min { γ 0 /2 , [ 4 ( T 1 + 1 ) ] − 1 } , ν = b e L ,
and take L L L large enough that e L ≤ 1 e_L\leq1 e L ≤ 1 . The conjugation error is at most e L / 4 e_L/4 e L /4 , so a Neumann series proves
∣ G g , Λ L ( x , y ; e + i η ) ∣ ≤ 4 e L e − b e L ∣ x − y ∣ 1 (89) |G_{g,\Lambda_L}(x,y;\mathbf{e}+i\eta)|\leq\frac{4}{e_L}e^{-be_L|x-y|_1} \tag*{(89)} ∣ G g , Λ L ( x , y ; e + i η ) ∣ ≤ e L 4 e − b e L ∣ x − y ∣ 1 (89)
on the good event.
This argument is valid for exponentially decaying infinite-range hopping. Fix ϑ = 1 / 8 \vartheta=1/8 ϑ = 1/8 and t = 1 / 2 t=1/2 t = 1/2 for definiteness; any 0 < ϑ < 1 / 4 0<\vartheta<1/4 0 < ϑ < 1/4 and ϑ < t < 1 \vartheta<t<1 ϑ < t < 1 would work. Let N L = ( 2 L + 1 ) d N_L=(2L+1)^d N L = ( 2 L + 1 ) d . Hölder’s inequality on the bad event { λ min ( L g , ϖ , Λ L ) ≤ e L } \{\lambda_{\min}(L_{g,\varpi,\Lambda_L})\leq e_L\} { λ m i n ( L g , ϖ , Λ L ) ≤ e L } and Lemma 5.2 show that, when ∣ u ∣ ∞ ≥ L / 2 |u|_\infty\geq L/2 ∣ u ∣ ∞ ≥ L /2 ,
E ∣ G g , Λ L ( 0 , u ; e + i η ) ∣ ϑ ≤ 4 ϑ e L − ϑ e − ϑ b e L L / 2 + D t ( N L , g − ) ϑ / t p L 1 − ϑ / t = : M L . (90) \begin{aligned}
\mathbb{E}|G_{g,\Lambda_L}(0,u;\mathbf{e}+i\eta)|^\vartheta\leq4^\vartheta e_L^{-\vartheta}e^{-\vartheta be_LL/2}+D_t(N_L,g_-)^{\vartheta/t}p_L^{1-\vartheta/t} \\
&=: \mathcal{M}_L. \tag*{(90)}
\end{aligned} E ∣ G g , Λ L ( 0 , u ; e + i η ) ∣ ϑ ≤ 4 ϑ e L − ϑ e − ϑ b e L L /2 + D t ( N L , g − ) ϑ / t p L 1 − ϑ / t =: M L . (90)
This bound is uniform in g ∈ [ g − , g + ] g\in[g_-,g_+] g ∈ [ g − , g + ] , ∣ e ∣ ≤ e L / 2 |\mathbf{e}|\leq e_L/2 ∣ e ∣ ≤ e L /2 , and η ≠ 0 \eta\neq0 η = 0 . Every factor outside the exponentials grows at most polynomially in L L L , while e L L = L 1 − β e_LL=L^{1-\beta} e L L = L 1 − β and (78) give stretched-exponential decay.
Put τ r = ∣ K 0 r ∣ \tau_r=|K_{0r}| τ r = ∣ K 0 r ∣ for r ≠ 0 r\neq0 r = 0 and X ϑ ( Λ ) = ∑ u ∈ Λ , v ∉ Λ τ v − u ϑ \mathcal{X}_\vartheta(\Lambda)=\sum_{u\in\Lambda,v\notin\Lambda}\tau_{v-u}^\vartheta X ϑ ( Λ ) = ∑ u ∈ Λ , v ∈ / Λ τ v − u ϑ . Exponential hopping gives constants C ϑ ′ , a ϑ > 0 C'_\vartheta,a_\vartheta>0 C ϑ ′ , a ϑ > 0 such that
X ϑ ( Λ L ) ≤ C ϑ ′ L d − 1 , sup ∣ u ∣ ∞ < L / 2 ∑ v ∉ Λ L τ v − u ϑ ≤ C ϑ ′ e − a ϑ L . (91) \mathcal{X}_\vartheta(\Lambda_L)\leq C'_\vartheta L^{d-1},\qquad
\sup_{|u|_\infty<L/2}\sum_{v\notin\Lambda_L}\tau_{v-u}^\vartheta\leq C'_\vartheta e^{-a_\vartheta L}. \tag*{(91)} X ϑ ( Λ L ) ≤ C ϑ ′ L d − 1 , ∣ u ∣ ∞ < L /2 sup v ∈ / Λ L ∑ τ v − u ϑ ≤ C ϑ ′ e − a ϑ L . (91)
For the first bound, sum the exponential tail according to the integer distance of u u u from the complement; each layer has at most C d L d − 1 C_dL^{d-1} C d L d − 1 sites. For the second, every crossing displacement has length at least L / 2 L/2 L /2 ; polynomial lattice multiplicities can be absorbed by decreasing the exponential rate.
Uniform estimates for truncated matrices
We use [3 Theorem 3.1] , whose criterion is imposed on every subset of a testing cube. This form yields bounds uniform over all truncated matrices, including near their boundaries.
Lemma 5.3. Fix [ g − , g + ] ⋐ ( 0 , ∞ ) [g_-,g_+] \Subset(0,\infty) [ g − , g + ] ⋐ ( 0 , ∞ ) . There are r > 0 r>0 r > 0 , B < ∞ B<\infty B < ∞ , μ > 0 \mu>0 μ > 0 , and ϑ = 1 / 8 \vartheta=1/8 ϑ = 1/8 such that
E ⟨ δ x , ( K Λ + g Ξ ϖ , Λ − e − i η ) − 1 δ y ⟩ ϑ ≤ B e − μ ∣ x − y ∣ 1 (92) \mathbb{E}\left\langle\delta_x,\left(K_\Lambda+g\Xi_{\varpi,\Lambda}-e-\mathrm{i}\eta\right)^{-1}\delta_y\right\rangle^\vartheta\le Be^{-\mu|x-y|_1} \tag*{(92)} E ⟨ δ x , ( K Λ + g Ξ ϖ , Λ − e − i η ) − 1 δ y ⟩ ϑ ≤ B e − μ ∣ x − y ∣ 1 (92) for every finite Λ \Lambda Λ , x , y ∈ Λ x,y\in\Lambda x , y ∈ Λ , g ∈ [ g − , g + ] g\in[g_-,g_+] g ∈ [ g − , g + ] , ∣ e ∣ ≤ r |e|\le r ∣ e ∣ ≤ r , and η ≠ 0 \eta\ne0 η = 0 . It also holds on the infinite lattice.
Proof. Use Proposition 5.1 with β = 1 / 2 \beta=1/2 β = 1/2 . Thus e L = L − 1 / 2 e_L=L^{-1/2} e L = L − 1/2 , ℓ L = ⌊ L 1 / 8 ⌋ \ell_L=\lfloor L^{1/8}\rfloor ℓ L = ⌊ L 1/8 ⌋ , and the bad-event probability is at most C L d e − c L d / 8 CL^d e^{-cL^{d/8}} C L d e − c L d /8 . If V ⊂ Λ L V\subset\Lambda_L V ⊂ Λ L , the matrix K V + g Ξ ϖ , V K_V+g\Xi_{\varpi,V} K V + g Ξ ϖ , V is a compression of K Λ L + g Ξ ϖ , Λ L K_{\Lambda_L}+g\Xi_{\varpi,\Lambda_L} K Λ L + g Ξ ϖ , Λ L . Its lowest eigenvalue is therefore at least that of the full cube. A good event for the cube is consequently good for every V V V simultaneously. This observation does not require independence of the events for different subsets.
The proof of (89) applies to each such V V V : its hopping row and column sums are bounded by those on the full lattice. On the bad event, (85) applies with ∣ V ∣ ≤ N L = ( 2 L + 1 ) d |V|\le N_L=(2L+1)^d ∣ V ∣ ≤ N L = ( 2 L + 1 ) d . Hölder’s inequality, with t = 1 / 2 t=1/2 t = 1/2 and ϑ = 1 / 8 \vartheta=1/8 ϑ = 1/8 , gives
sup V ⊂ Λ L , 0 , u ∈ V ∣ u ∣ ∞ ≥ L / 2 E ∣ G g , V ( 0 , u ; e + i η ) ∣ ϑ ≤ C L b 0 ( e − c L 1 / 2 + e − c L d / 8 ) = : N L (93) \sup_{\substack{V\subset\Lambda_L,\ 0,u\in V\\ |u|_\infty\ge L/2}}\mathbb{E}|G_{g,V}(0,u;e+\mathrm{i}\eta)|^\vartheta\le CL^{b_0}\left(e^{-cL^{1/2}}+e^{-cL^{d/8}}\right)=:\mathcal{N}_L \tag*{(93)} V ⊂ Λ L , 0 , u ∈ V ∣ u ∣ ∞ ≥ L /2 sup E ∣ G g , V ( 0 , u ; e + i η ) ∣ ϑ ≤ C L b 0 ( e − c L 1/2 + e − c L d /8 ) =: N L (93)
for ∣ e ∣ ≤ e L / 2 |e|\le e_L/2 ∣ e ∣ ≤ e L /2 , uniformly in g g g and η ≠ 0 \eta\ne0 η = 0 . Here b 0 b_0 b 0 is a fixed finite exponent. Set a matrix element equal to zero when either of its indices does not belong to V V V .
Let τ r = ∣ K 0 r ∣ \tau_r=|K_{0r}| τ r = ∣ K 0 r ∣ for r ≠ 0 r\ne0 r = 0 and let X ϑ ( Λ L ) \mathcal{X}_\vartheta(\Lambda_L) X ϑ ( Λ L ) be as in Section 5.3. The scalar version of [3 Theorem 3.1] requires
( 1 + C ϑ g − ϑ X ϑ ( Λ L ) ) sup V ⊂ Λ L ∑ u ∈ Λ L v ∉ Λ L τ v − u ϑ E ∣ G g , V ( 0 , u ; z ) ∣ ϑ < 1. (94) \left(1+C_\vartheta g^{-\vartheta}\mathcal{X}_\vartheta(\Lambda_L)\right)\sup_{V\subset\Lambda_L}\sum_{\substack{u\in\Lambda_L\\ v\notin\Lambda_L}}\tau_{v-u}^\vartheta\mathbb{E}|G_{g,V}(0,u;z)|^\vartheta<1. \tag*{(94)} ( 1 + C ϑ g − ϑ X ϑ ( Λ L ) ) V ⊂ Λ L sup u ∈ Λ L v ∈ / Λ L ∑ τ v − u ϑ E ∣ G g , V ( 0 , u ; z ) ∣ ϑ < 1. (94)
The single-site regularity and decoupling assumptions, labelled R1 and R2 in [3 ] , hold for the uniform law: R1 has exponent one, and R2 holds for ϑ < 1 / 4 \vartheta<1/4 ϑ < 1/4 by [3 Lemma C.3] . Exponential summability and translation invariance give the remaining hypotheses; the deterministic diagonal is absorbed into the energy. For the inner half-cube use (85) and the second estimate in (91) ; for the outer part use (93) . The left side of (94) is bounded by
( 1 + C ϑ g − ϑ X ϑ ( Λ L ) ) [ C ϑ ′ N L D ϑ ( N L , g − ) e − a ϑ L + X ϑ ( Λ L ) N L ] . \left(1+C_\vartheta g^{-\vartheta}\mathcal{X}_\vartheta(\Lambda_L)\right)\left[C'_\vartheta N_LD_\vartheta(N_L,g_-)e^{-a_\vartheta L}+\mathcal{X}_\vartheta(\Lambda_L)\mathcal{N}_L\right]. ( 1 + C ϑ g − ϑ X ϑ ( Λ L ) ) [ C ϑ ′ N L D ϑ ( N L , g − ) e − a ϑ L + X ϑ ( Λ L ) N L ] .
It tends to zero. Choose one L ∗ L_* L ∗ for which it is at most 1 / 2 1/2 1/2 . The criterion gives the conclusion uniformly over all restrictions. The common strict margin, the common exponential hopping bound, and the bounded range of g − 1 g^{-1} g − 1 give common B , μ B,\mu B , μ . One may take r = e L ∗ / 4 r=e_{L_*}/4 r = e L ∗ /4 . Strong resolvent convergence and Fatou’s lemma give the infinite-volume estimate with the same constants. □
Take g − = W c M / 2 g_-=Wc_M/2 g − = W c M /2 , g + = W c M g_+=Wc_M g + = W c M , and choose 0 < Δ W < min { c M / 2 , r } 0<\Delta_W<\min\{c_M/2,r\} 0 < Δ W < min { c M /2 , r } using this lemma. Set I W = ( c M − Δ W , c M ) I_W=(c_M-\Delta_W,c_M) I W = ( c M − Δ W , c M ) and put
A ϖ , Λ ( E ) : = E M ϖ , Λ − J Λ C Λ t J Λ = K Λ + E W Ξ ϖ , Λ − ( c M − E ) I d , (95) \mathcal{A}_{\varpi,\Lambda}(E):=EM_{\varpi,\Lambda}-J_\Lambda C_\Lambda^{t}J_\Lambda=K_\Lambda+EW\Xi_{\varpi,\Lambda}-(c_M-E)\mathrm{Id}, \tag*{(95)} A ϖ , Λ ( E ) := E M ϖ , Λ − J Λ C Λ t J Λ = K Λ + E W Ξ ϖ , Λ − ( c M − E ) Id , (95)
where J Λ J_\Lambda J Λ is the restriction of the staggering operator (74) . For every fixed E ∈ I W E\in I_W E ∈ I W , the finite matrix is invertible almost surely. This follows also from (84) , applied with g = E W g=EW g = E W . Taking η → 0 \eta\to0 η → 0 in (92) and using Fatou’s lemma gives
E ∣ A ϖ , Λ ( E ) − 1 ( x , y ) ∣ ϑ ≤ B e − μ ∣ x − y ∣ 1 . (96) \mathbb{E}|\mathcal{A}_{\varpi,\Lambda}(E)^{-1}(x,y)|^\vartheta\le Be^{-\mu|x-y|_1}. \tag*{(96)} E ∣ A ϖ , Λ ( E ) − 1 ( x , y ) ∣ ϑ ≤ B e − μ ∣ x − y ∣ 1 . (96)
In infinite volume write A ϖ ( E ) = E M ϖ − J C J \mathcal{A}_{\varpi}(E)=EM_{\varpi}-JCJ A ϖ ( E ) = E M ϖ − J C J and H ~ ϖ = J H ϖ J \widetilde{H}_{\varpi}=JH_{\varpi}J H ϖ = J H ϖ J . Then
H ~ ϖ − E − i η = − M ϖ − 1 / 2 ( A ϖ ( E ) + i η M ϖ ) M ϖ − 1 / 2 . (97) \widetilde{H}_{\varpi}-E-\mathrm{i}\eta=-M_{\varpi}^{-1/2}\left(\mathcal{A}_{\varpi}(E)+\mathrm{i}\eta M_{\varpi}\right)M_{\varpi}^{-1/2}. \tag*{(97)} H ϖ − E − i η = − M ϖ − 1/2 ( A ϖ ( E ) + i η M ϖ ) M ϖ − 1/2 . (97)
The random imaginary part prevents direct identification with the additive resolvent. § 3.1 instead transfers the fixed-real-energy estimate (96) by averaging the actual material spectral measure.
Spectral and dynamical localisation
We convert the real-energy estimates of § 2.16 into eigenfunction-correlator bounds and deduce spectral and dynamical localisation. The proof combines spectral averaging in one material variable with independent resampling, following the rank-one analysis of [27 ] and the fractional-moment approach of [2 , 3 , 28 ] .
The localisation theorem
We state the result with correlator bounds uniform over finite truncations, which will also be used in the finite-sample analysis.
Theorem 6.1 (Localisation at every positive disorder). Assume the geometry of § 2.1 , including (R), and the law (2.5). For each prescribed W > 0 W>0 W > 0 , use the interval I W I_W I W chosen in § 2.16 . There are constants A W < ∞ A_W<\infty A W < ∞ and ν W > 0 \nu_W>0 ν W > 0 such that
E Q ϖ , Λ ( x , y ; I W ) ≤ A W e − ν W ∣ x − y ∣ 1 , I W = ( c M − Δ W , c M ) , (98) \mathbb{E}Q_{\varpi,\Lambda}(x,y;I_W)\leq A_W e^{-\nu_W|x-y|_1},\qquad I_W=(c_M-\Delta_W,c_M), \tag*{(98)} E Q ϖ , Λ ( x , y ; I W ) ≤ A W e − ν W ∣ x − y ∣ 1 , I W = ( c M − Δ W , c M ) , (98) uniformly in all finite Λ ⊂ Z d \Lambda\subset\mathbb{Z}^d Λ ⊂ Z d and x , y ∈ Λ x,y\in\Lambda x , y ∈ Λ . The same bound holds for Q ϖ Q_\varpi Q ϖ on the infinite lattice. Almost surely, the spectrum of H ϖ H_\varpi H ϖ in I W I_W I W is pure point, its eigenvalues are simple, and every corresponding eigenfunction satisfies
∣ ψ ( n ) ∣ ≤ C ϖ , ψ , y e − ν ∣ n − y ∣ 1 ( n ∈ Z d ) (99) |\psi(n)|\leq C_{\varpi,\psi,y}e^{-\nu|n-y|_1}\qquad(n\in\mathbb{Z}^d) \tag*{(99)} ∣ ψ ( n ) ∣ ≤ C ϖ , ψ , y e − ν ∣ n − y ∣ 1 ( n ∈ Z d ) (99) for every 0 < ν < ν W 0<\nu<\nu_W 0 < ν < ν W and any y y y with ψ ( y ) ≠ 0 \psi(y)\ne0 ψ ( y ) = 0 . The constant in (99) is allowed to depend on the eigenfunction. In particular,
E sup t ∈ R ∣ ⟨ δ x , e − i t H ϖ 1 I W ( H ϖ ) δ y ⟩ ∣ ≤ A W e − ν W ∣ x − y ∣ 1 . (100) \mathbb{E}\sup_{t\in\mathbb{R}}\left|\left\langle\delta_x,e^{-itH_\varpi}\mathbf{1}_{I_W}(H_\varpi)\delta_y\right\rangle\right|\leq A_We^{-\nu_W|x-y|_1}. \tag*{(100)} E t ∈ R sup ⟨ δ x , e − i t H ϖ 1 I W ( H ϖ ) δ y ⟩ ≤ A W e − ν W ∣ x − y ∣ 1 . (100)
The constants and the interval depend on the fixed value of W W W . Unitary equivalence (2.6) gives the same spectral conclusion for A ϖ A_\varpi A ϖ on H ϖ \mathcal{H}_\varpi H ϖ . The correlator bound also controls the reduced wave evolution cos ( t H ϖ ) 1 I W ( H ϖ ) \cos(t\sqrt{H_\varpi})\mathbf{1}_{I_W}(H_\varpi) cos ( t H ϖ ) 1 I W ( H ϖ ) .
Spectral averaging for a material variable
The following spectral-averaging lemma is independent of translation invariance and of the decay of the capacitance coefficients. It will be used to prove Theorem 6.1 .
Lemma 6.2 (Material-variable averaging). Let C ∗ C_* C ∗ be a bounded nonnegative self-adjoint operator on a finite or countable coordinate space. Fix all inverse material factors except m y = t ∈ [ 1 , 1 + W ] m_y=t\in[1,1+W] m y = t ∈ [ 1 , 1 + W ] , and set H t = M t − 1 / 2 C ∗ M t − 1 / 2 H_t=M_t^{-1/2}C_*M_t^{-1/2} H t = M t − 1/2 C ∗ M t − 1/2 . Denote its spectral measure at δ y \delta_y δ y by ν y t \nu_y^t ν y t . If I ⊂ [ E − , ∞ ) I\subset[E_-,\infty) I ⊂ [ E − , ∞ ) , E − > 0 E_->0 E − > 0 , then for every nonnegative Borel function F F F , independent of t t t ,
1 W ∫ 1 1 + W ∫ I F ( E ) ν y t ( d E ) d t ≤ κ W , I ∫ I F ( E ) d E , κ W , I : = 1 + W W E − . (101) \frac{1}{W}\int_1^{1+W}\int_I F(E)\,\nu_y^t(\mathrm{d}E)\,\mathrm{d}t\leq\kappa_{W,I}\int_I F(E)\,\mathrm{d}E,\qquad\kappa_{W,I}:=\frac{1+W}{WE_-}. \tag*{(101)} W 1 ∫ 1 1 + W ∫ I F ( E ) ν y t ( d E ) d t ≤ κ W , I ∫ I F ( E ) d E , κ W , I := W E − 1 + W . (101) Proof. Put b = t − 1 b=t^{-1} b = t − 1 , T t = C ∗ 1 / 2 M t − 1 / 2 T_t=C_*^{1/2}M_t^{-1/2} T t = C ∗ 1/2 M t − 1/2 and ϕ y = C ∗ 1 / 2 δ y \phi_y=C_*^{1/2}\delta_y ϕ y = C ∗ 1/2 δ y . Then
H t = T t ∗ T t , B b : = T t T t ∗ = B 0 + b ∣ ϕ y ⟩ ⟨ ϕ y ∣ , H_t=T_t^*T_t,\qquad B_b:=T_tT_t^*=B_0+b|\phi_y\rangle\langle\phi_y|, H t = T t ∗ T t , B b := T t T t ∗ = B 0 + b ∣ ϕ y ⟩ ⟨ ϕ y ∣ , where B 0 B_0 B 0 is independent of b b b . Let μ ϕ y b \mu_{\phi_y}^b μ ϕ y b be the spectral measure of B b B_b B b at ϕ y \phi_y ϕ y . Intertwining the functional calculi of T t ∗ T t T_t^*T_t T t ∗ T t and T t T t ∗ T_tT_t^* T t T t ∗ gives
b μ ϕ y b ( d E ) = E ν y t ( d E ) . (102) b\,\mu_{\phi_y}^b(\mathrm{d}E)=E\,\nu_y^t(\mathrm{d}E). \tag*{(102)} b μ ϕ y b ( d E ) = E ν y t ( d E ) . (102) Indeed T t δ y = b ϕ y T_t\delta_y=\sqrt{b}\,\phi_y T t δ y = b ϕ y and T t ∗ 1 S ( B b ) T t = H t 1 S ( H t ) T_t^*\mathbf{1}_S(B_b)T_t=H_t\mathbf{1}_S(H_t) T t ∗ 1 S ( B b ) T t = H t 1 S ( H t ) for Borel S S S . No inverse of C ∗ C_* C ∗ is used.
For completeness, rank-one spectral averaging gives ∫ R ν ϕ y b ( d E ) d b = d E \int_{\mathbb{R}}\nu_{\phi_y}^{b}(\mathrm{d}E)\,\mathrm{d}b=\mathrm{d}E ∫ R ν ϕ y b ( d E ) d b = d E when ϕ y ≠ 0 \phi_y\ne0 ϕ y = 0 . If F 0 ( z ) = ⟨ ϕ y , ( B 0 − z ) − 1 ϕ y ⟩ F_0(z)=\langle\phi_y,(B_0-z)^{-1}\phi_y\rangle F 0 ( z ) = ⟨ ϕ y , ( B 0 − z ) − 1 ϕ y ⟩ , its resolvent formula is
F b ( z ) = F 0 ( z ) 1 + b F 0 ( z ) = 1 b + F 0 ( z ) . F_b(z)=\frac{F_0(z)}{1+bF_0(z)}=\frac{1}{b+F_0(z)}. F b ( z ) = 1 + b F 0 ( z ) F 0 ( z ) = b + F 0 ( z ) 1 .
For ℑ z > 0 \Im z>0 ℑ z > 0 , ℑ F 0 ( z ) > 0 \Im F_0(z)>0 ℑ F 0 ( z ) > 0 , and integration of this Poisson kernel over b ∈ R b\in\mathbb{R} b ∈ R gives ∫ R ℑ F b ( z ) d b = π \int_{\mathbb{R}}\Im F_b(z)\,\mathrm{d}b=\pi ∫ R ℑ F b ( z ) d b = π . The spectral-measure boundary formula then proves the averaging identity. For ϕ y = 0 \phi_y=0 ϕ y = 0 , the spectral measure of H t H_t H t at y y y is supported at zero and (101) is immediate.
The law of b b b has density ( W b 2 ) − 1 (Wb^2)^{-1} ( W b 2 ) − 1 on [ ( 1 + W ) − 1 , 1 ] [(1+W)^{-1},1] [( 1 + W ) − 1 , 1 ] . Using (102) , the integrand is weighted by ( W b E ) − 1 ≤ ( 1 + W ) / ( W E − ) (WbE)^{-1}\le(1+W)/(WE_-) ( W b E ) − 1 ≤ ( 1 + W ) / ( W E − ) . Extending the nonnegative integral in b b b to the real line proves the assertion. The argument applies to unbounded nonnegative F F F by monotone convergence. □
Resampling and finite-volume eigenfunction correlators
We work with H ~ ϖ , Λ = J Λ H ϖ , Λ t J Λ \widetilde{H}_{\varpi,\Lambda}=J_\Lambda H^t_{\varpi,\Lambda}J_\Lambda H ϖ , Λ = J Λ H ϖ , Λ t J Λ . Staggering does not change any of the absolute values in (11) . Fix y y y and replace m y m_y m y by an independent copy m y ′ m'_y m y ′ with the same law. A superscript ( y ) (y) ( y ) denotes the resulting reference operator.
Suppose H ~ ϖ , Λ ψ j = E j ψ j \widetilde{H}_{\varpi,\Lambda}\psi_j=E_j\psi_j H ϖ , Λ ψ j = E j ψ j and put a j = M ϖ , Λ − 1 / 2 ψ j a_j=M_{\varpi,\Lambda}^{-1/2}\psi_j a j = M ϖ , Λ − 1/2 ψ j . Then
A Λ ( y ) ( E j ) a j = E j ( m y ′ − m y ) a j ( y ) δ y . (103) \mathcal{A}^{(y)}_\Lambda(E_j)a_j=E_j(m'_y-m_y)a_j(y)\delta_y. \tag*{(103)} A Λ ( y ) ( E j ) a j = E j ( m y ′ − m y ) a j ( y ) δ y . (103)
For fixed other variables and fixed m y ′ m'_y m y ′ , the reference spectrum is a finite set independent of m y m_y m y . Lemma 6.2 implies that ν y m y \nu^{m_y}_y ν y m y assigns this set zero mass for almost every m y m_y m y . Thus, for every eigenvalue with nonzero spectral mass at y y y , the reference matrix is invertible, except on a probability-zero event. At such an energy, (103) also proves simplicity: an eigenvector vanishing at y y y must be zero. Applying this observation at every site shows that all positive eigenvalues of the finite random matrix are almost surely simple.
Let E − : = inf I W > 0 E_-:=\inf I_W>0 E − := inf I W > 0 , and set d W : = c M W 1 + W d_W:=c_MW\sqrt{1+W} d W := c M W 1 + W . Since E j ≤ c M E_j\le c_M E j ≤ c M and 1 ≤ m n ≤ 1 + W 1\le m_n\le1+W 1 ≤ m n ≤ 1 + W , equation (103) gives
∣ ψ j ( x ) ∣ ≤ d W ∣ ψ j ( y ) ∣ ∣ A Λ ( y ) ( E j ) − 1 ( x , y ) ∣ . (104) |\psi_j(x)|\le d_W|\psi_j(y)||\mathcal{A}^{(y)}_\Lambda(E_j)^{-1}(x,y)|. \tag*{(104)} ∣ ψ j ( x ) ∣ ≤ d W ∣ ψ j ( y ) ∣∣ A Λ ( y ) ( E j ) − 1 ( x , y ) ∣. (104)
Define
Q ϖ , Λ , ϑ t ( x , y ; I W ) : = ∑ E j ∈ I W ∣ ψ j ( y ) ∣ 2 − ϑ ∣ ψ j ( x ) ∣ ϑ . Q^t_{\varpi,\Lambda,\vartheta}(x,y;I_W):=\sum_{E_j\in I_W}|\psi_j(y)|^{2-\vartheta}|\psi_j(x)|^\vartheta. Q ϖ , Λ , ϑ t ( x , y ; I W ) := E j ∈ I W ∑ ∣ ψ j ( y ) ∣ 2 − ϑ ∣ ψ j ( x ) ∣ ϑ .
The eigenvectors are normalised in ordinary ℓ 2 ( Λ ) \ell^2(\Lambda) ℓ 2 ( Λ ) . Raising (104) to power ϑ \vartheta ϑ and using ν y m y ( { E j } ) = ∣ ψ j ( y ) ∣ 2 \nu_y^{m_y}(\{E_j\})=|\psi_j(y)|^2 ν y m y ({ E j }) = ∣ ψ j ( y ) ∣ 2 yields
Q ϖ , Λ , ϑ t ( x , y ; I W ) ≤ d W ϑ ∫ I W ∣ A Λ ( y ) ( E ) − 1 ( x , y ) ∣ ϑ ν y m y ( d E ) . (105) Q^t_{\varpi,\Lambda,\vartheta}(x,y;I_W)\le d_W^\vartheta\int_{I_W}|\mathcal{A}^{(y)}_\Lambda(E)^{-1}(x,y)|^\vartheta\nu_y^{m_y}(\mathrm{d}E). \tag*{(105)} Q ϖ , Λ , ϑ t ( x , y ; I W ) ≤ d W ϑ ∫ I W ∣ A Λ ( y ) ( E ) − 1 ( x , y ) ∣ ϑ ν y m y ( d E ) . (105)
The integrand is independent of the original m y m_y m y . Average first in m y m_y m y using (101) , and then in the reference configuration. The latter still has the original iid law, so (96) gives
E Q ϖ , Λ , ϑ t ( x , y ; I W ) ≤ κ W , I W d W ϑ ∣ I W ∣ B e − μ ∣ x − y ∣ 1 . (106) \mathbb{E}Q^t_{\varpi,\Lambda,\vartheta}(x,y;I_W)\le\kappa_{W,I_W}d_W^\vartheta|I_W|Be^{-\mu|x-y|_1}. \tag*{(106)} E Q ϖ , Λ , ϑ t ( x , y ; I W ) ≤ κ W , I W d W ϑ ∣ I W ∣ B e − μ ∣ x − y ∣ 1 . (106)
All exceptional reference energies have zero original spectral mass by the same averaging lemma. This justifies its use even though the eigenvalues E j E_j E j depend on m y m_y m y .
For a , b ≥ 0 a,b\ge0 a , b ≥ 0 and 0 < ϑ < 1 0<\vartheta<1 0 < ϑ < 1 , a b ≤ a ϑ b 2 − ϑ + a 2 − ϑ b ϑ ab\le a^\vartheta b^{2-\vartheta}+a^{2-\vartheta}b^\vartheta ab ≤ a ϑ b 2 − ϑ + a 2 − ϑ b ϑ . The finite eigenvalues are simple, and hence
Q ϖ , Λ t ( x , y ; I W ) = ∑ E j ∈ I W ∣ ψ j ( x ) ψ j ( y ) ∣ ≤ Q ϖ , Λ , ϑ t ( x , y ; I W ) + Q ϖ , Λ , ϑ t ( y , x ; I W ) . (107) \begin{aligned}
Q^t_{\varpi,\Lambda}(x,y;I_W)=\sum_{E_j\in I_W}|\psi_j(x)\psi_j(y)| \\
&\le Q^t_{\varpi,\Lambda,\vartheta}(x,y;I_W)+Q^t_{\varpi,\Lambda,\vartheta}(y,x;I_W). \tag*{(107)}
\end{aligned} Q ϖ , Λ t ( x , y ; I W ) = E j ∈ I W ∑ ∣ ψ j ( x ) ψ j ( y ) ∣ ≤ Q ϖ , Λ , ϑ t ( x , y ; I W ) + Q ϖ , Λ , ϑ t ( y , x ; I W ) . (107)
This proves (98) , for example with A W = 2 κ W , I W ∂ W ϑ ∣ I W ∣ B A_W=2\kappa_{W,I_W}\partial_W^\vartheta|I_W|B A W = 2 κ W , I W ∂ W ϑ ∣ I W ∣ B and ν W = μ \nu_W=\mu ν W = μ .
A range criterion.
The following lemma converts fractional-moment bounds for resolvent columns into almost-sure solvability of the corresponding operator equation.
Lemma 6.3. Let A A A be a bounded self-adjoint random operator on ℓ 2 ( Z d ) \ell^2(\mathbb{Z}^d) ℓ 2 ( Z d ) . If, for a fixed y y y and some ϑ ∈ ( 0 , 2 ) \vartheta\in(0,2) ϑ ∈ ( 0 , 2 ) ,
sup 0 < η ≤ 1 E ∑ x ∣ ⟨ δ x , ( A − i η ) − 1 δ y ⟩ ∣ ϑ < ∞ \sup_{0<\eta\le1}\mathbb{E}\sum_{x}\left|\langle\delta_x,(A-i\eta)^{-1}\delta_y\rangle\right|^\vartheta<\infty 0 < η ≤ 1 sup E x ∑ ⟨ δ x , ( A − i η ) − 1 δ y ⟩ ϑ < ∞ then δ y ∈ Ran A \delta_y\in\operatorname{Ran} A δ y ∈ Ran A almost surely.
Proof. Let S k S_k S k be the sum with η = 1 / k \eta=1/k η = 1/ k . Fatou’s lemma implies lim inf k S k < ∞ \liminf_k S_k<\infty lim inf k S k < ∞ almost surely. Choose a subsequence with S k j ≤ C ( ϖ ) < ∞ S_{k_j}\le C(\varpi)<\infty S k j ≤ C ( ϖ ) < ∞ . Since ∑ x ∣ b x ∣ 2 ≤ ( ∑ x ∣ b x ∣ ϑ ) 2 / ϑ \sum_x|b_x|^2\le\left(\sum_x|b_x|^\vartheta\right)^{2/\vartheta} ∑ x ∣ b x ∣ 2 ≤ ( ∑ x ∣ b x ∣ ϑ ) 2/ ϑ , the vectors u j = ( A − i / k j ) − 1 δ y u_j=(A-i/k_j)^{-1}\delta_y u j = ( A − i / k j ) − 1 δ y are bounded in ℓ 2 \ell^2 ℓ 2 . A subsequence converges weakly to u u u . The identity A u j = δ y + ( i / k j ) u j Au_j=\delta_y+(i/k_j)u_j A u j = δ y + ( i / k j ) u j and boundedness of A A A imply A u = δ y Au=\delta_y A u = δ y . No measurable choice of the subsequence or of u u u is needed for this pointwise conclusion. □
Infinite-volume spectrum and dynamics.
We complete the proof of Theorem 6.1 by passing the correlator estimates to infinite volume and establishing almost-sure simplicity of the eigenvalues.
As the cubes Λ L \Lambda_L Λ L increase to Z d \mathbb{Z}^d Z d , the extended matrices P L H ϖ P L P_LH_\varpi P_L P L H ϖ P L converge strongly to H ϖ H_\varpi H ϖ and are uniformly bounded. Therefore f ( P L H ϖ P L ) → f ( H ϖ ) f(P_LH_\varpi P_L)\to f(H_\varpi) f ( P L H ϖ P L ) → f ( H ϖ ) strongly for continuous f f f . If f f f has compact support in I W I_W I W , then f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 , and the former operator is the zero extension of f ( H ϖ , Λ L t ) f(H^t_{\varpi,\Lambda_L}) f ( H ϖ , Λ L t ) . Take successively a maximum over finitely many functions in a countable dense family in the unit ball of C 0 ( I W ) C_0(I_W) C 0 ( I W ) . Fatou’s lemma followed by monotone convergence proves the infinite-volume form of (98) . Regularity of finite complex spectral measures identifies this supremum with the total variation on I W I_W I W , so it also controls all bounded Borel functions in (11) .
For 0 < ν < ν W 0<\nu<\nu_W 0 < ν < ν W , the bound implies
E ∑ x ∈ Z d e ν ∣ x − y ∣ 1 Q ϖ ( x , y ; I W ) < ∞ . (108) \mathbb{E}\sum_{x\in\mathbb{Z}^d}e^{\nu|x-y|_1}Q_\varpi(x,y;I_W)<\infty. \tag*{(108)} E x ∈ Z d ∑ e ν ∣ x − y ∣ 1 Q ϖ ( x , y ; I W ) < ∞. (108)
Thus the sum is finite almost surely, simultaneously for all y y y and, by a countable choice of weights, for every ν < ν W \nu<\nu_W ν < ν W . Since Q ϖ ≤ 1 Q_\varpi\le1 Q ϖ ≤ 1 , this gives uniformly vanishing spatial tails for the orbit e − i t H ϖ 1 I W ( H ϖ ) δ y e^{-itH_\varpi}1_{I_W}(H_\varpi)\delta_y e − i t H ϖ 1 I W ( H ϖ ) δ y , t ∈ R t\in\mathbb{R} t ∈ R . Every such orbit is precompact in ℓ 2 \ell^2 ℓ 2 . The spectral theorem (equivalently, the RAGE criterion) implies that its initial vector belongs to the pure point subspace; see [28 , 29 ] . The coordinate vectors are total, so the spectrum in I W I_W I W is pure point. Choosing f ( E ) = e − i t E f(E)=e^{-itE} f ( E ) = e − i tE proves (100) .
We give the additional argument that rules out multiplicities and therefore proves (99) for every eigenfunction. In infinite volume, the weighted version of (92) gives, for each fixed E ∈ I W E\in I_W E ∈ I W and y y y ,
sup η > 0 E ∑ x e ν 0 ∣ x − y ∣ 1 ∣ ⟨ δ x , ( A ϖ ( E ) − i η ) − 1 δ y ⟩ ∣ ϑ < ∞ \sup_{\eta>0}\mathbb{E}\sum_x e^{\nu_0|x-y|_1}\left|\langle\delta_x,(\mathcal{A}_\varpi(E)-i\eta)^{-1}\delta_y\rangle\right|^\vartheta<\infty η > 0 sup E x ∑ e ν 0 ∣ x − y ∣ 1 ⟨ δ x , ( A ϖ ( E ) − i η ) − 1 δ y ⟩ ϑ < ∞
for 0 < ν 0 < μ 0<\nu_0<\mu 0 < ν 0 < μ . Fatou’s lemma, followed by the weak-compactness argument in Lemma 6.3, produces a solution of A ϖ ( E ) u = δ y \mathcal{A}_\varpi(E)u=\delta_y A ϖ ( E ) u = δ y in ℓ 2 \ell^2 ℓ 2 . Taking limits of the coordinates also preserves the finite weighted ℓ ϑ \ell^\vartheta ℓ ϑ sum. By Fubini, for almost every reference configuration this holds for Lebesgue-almost every E E E . Also exclude the countable set of eigenvalues of the fixed reference operator H ˙ ϖ ( y ) \dot H_\varpi^{(y)} H ˙ ϖ ( y ) , using the resampling notation of Section 6.3. Indeed,
A ϖ ( y ) ( E ) = ( M ϖ ( y ) ) 1 / 2 ( E Id − H ˙ ϖ ( y ) ) ( M ϖ ( y ) ) 1 / 2 . \mathcal{A}_\varpi^{(y)}(E)=\left(M_\varpi^{(y)}\right)^{1/2}\left(E\operatorname{Id}-\dot H_\varpi^{(y)}\right)\left(M_\varpi^{(y)}\right)^{1/2}. A ϖ ( y ) ( E ) = ( M ϖ ( y ) ) 1/2 ( E Id − H ˙ ϖ ( y ) ) ( M ϖ ( y ) ) 1/2 .
Since M ϖ ( y ) M_{\varpi}^{(y)} M ϖ ( y ) is bounded and boundedly invertible, A ϖ ( y ) ( E ) A_{\varpi}^{(y)}(E) A ϖ ( y ) ( E ) is injective outside this countable set, and the solution constructed above is unique. This additional exceptional set depends only on the reference configuration and is independent of the original material variable m y m_y m y .
The sets in this argument are measurable: solvability of the reference equation and its solution can be characterised through limits of resolvent columns, while nontriviality of the kernel is detected by spectral atoms.
Fix the outside variables and a reference value m y ′ m'_y m y ′ with these properties. Its exceptional energy set is Lebesgue-null and independent of the original m y m_y m y . The infinite-volume version of Lemma 6.2 makes its original spectral mass at y y y zero for almost every m y m_y m y . At any eigenvalue with nonzero spectral mass at y y y , (103) holds with an injective reference operator. An eigenvector with zero y y y coordinate must therefore vanish. The evaluation map at y y y is injective on the eigenspace, whose dimension is at most one. Fubini and a countable intersection over y y y prove simplicity almost surely.
Finally, at a simple eigenvalue E j E_j E j , its spectral atom satisfies
∣ ψ j ( x ) ψ j ( y ) ∣ = ∣ ⟨ δ x , 1 { E j } ( H ϖ ) δ y ⟩ ∣ ≤ Q ϖ ( x , y ; I W ) . |\psi_j(x)\psi_j(y)| = |\langle\delta_x, \mathbb{1}_{\{E_j\}}(H_\varpi)\delta_y\rangle| \le Q_\varpi(x,y;I_W). ∣ ψ j ( x ) ψ j ( y ) ∣ = ∣ ⟨ δ x , 1 { E j } ( H ϖ ) δ y ⟩ ∣ ≤ Q ϖ ( x , y ; I W ) .
Equation (108) bounds the right side by a finite random constant times e − ν ∣ x − y ∣ 1 e^{-\nu|x-y|_1} e − ν ∣ x − y ∣ 1 . Division by ∣ ψ j ( y ) ∣ > 0 |\psi_j(y)|>0 ∣ ψ j ( y ) ∣ > 0 completes the proof of Theorem 6.1 .
Corollary 6.4 (The large-disorder interval). For the uniform inverse-material law, the conclusions of Theorem 6.1 hold on the interval I I I of Theorem 4.1 whenever its large-disorder condition holds. They also hold under either of the sufficient tests in Theorem 4.4 , with this same law.
Proof. In finite volume the real inverse operator is related to the boundary value of the resolvent of H ϖ , Λ t H_{\varpi,\Lambda}^{t} H ϖ , Λ t by
( C Λ t − E M ϖ , Λ ) − 1 = M ϖ , Λ − 1 / 2 ( H ϖ , Λ t − E ) − 1 M ϖ , Λ − 1 / 2 . (C_{\Lambda}^{t}-EM_{\varpi,\Lambda})^{-1}=M_{\varpi,\Lambda}^{-1/2}(H_{\varpi,\Lambda}^{t}-E)^{-1}M_{\varpi,\Lambda}^{-1/2}. ( C Λ t − E M ϖ , Λ ) − 1 = M ϖ , Λ − 1/2 ( H ϖ , Λ t − E ) − 1 M ϖ , Λ − 1/2 .
For each fixed E > 0 E>0 E > 0 the inverse exists almost surely by Lemma 6.2. Fatou’s lemma and (53) , or (57) , give the analogue of (96) ; the factors m n − 1 / 2 m_n^{-1/2} m n − 1/2 are at most one. The resampling and limiting arguments just proved apply with E − : = inf I E_-:=\inf I E − := inf I and d W : = ( sup I ) W 1 + W d_W:=(\sup I)W\sqrt{1+W} d W := ( sup I ) W 1 + W . Simplicity in infinite volume can alternatively be obtained from the range argument for the resolvent of H ϖ H_\varpi H ϖ and the same averaging lemma. □
An explicit edge window and mobility-edge bounds
We refine the band-edge argument to obtain an explicit localisation window whose energy width is proportional to the disorder strength at weak disorder. This gives an upper bound on the upper endpoint of any absolutely continuous spectral component.
Spectral and mean-material edges
7.1. Spectral and mean-material edges. Note that the exact upper spectral edge of the disordered operator is c M c_M c M . We use the unique quadratic maximum established in § 2.6 :
c ( α ) = c M ⟺ α = α M , A M = − 1 2 D α 2 c ( α M ) > 0. (109) c(\alpha)=c_M \Longleftrightarrow\alpha=\alpha_M,\qquad A_M=-\frac{1}{2}D_{\alpha}^{2}c(\alpha_M)>0. \tag*{(109)} c ( α ) = c M ⟺ α = α M , A M = − 2 1 D α 2 c ( α M ) > 0. (109)
The equivalence is on the Brillouin torus. Put
E t o p = c M , E a ( W ) = c M 1 + a W , E m e a n ( W ) = c M 1 + W / 2 , 0 < a ≤ 1 2 . (110) E_{\mathrm{top}}=c_M,\qquad E_a(W)=\frac{c_M}{1+aW},\qquad E_{\mathrm{mean}}(W)=\frac{c_M}{1+W/2},\qquad0<a\leq\frac{1}{2}. \tag*{(110)} E top = c M , E a ( W ) = 1 + aW c M , E mean ( W ) = 1 + W /2 c M , 0 < a ≤ 2 1 . (110)
Here E m e a n = c M / μ W E_{\mathrm{mean}}=c_M/\mu_W E mean = c M / μ W is the upper edge of C / μ W C/\mu_W C / μ W . The threshold E a ( W ) E_a(W) E a ( W ) is a sufficient localisation threshold, not a definition of a mobility edge. Frequency statements use ω M \omega_M ω M from (7) , with δ \delta δ and ε \varepsilon ε fixed while varying W W W . The mean-material edge frequency is ω M / μ W \omega_M/\sqrt{\mu_W} ω M / μ W .
The quantitative theorem
The following theorem makes the localisation interval explicit, while allowing the decay constants to depend on the fixed disorder strength.
Theorem 7.1 (A localisation window proportional to the disorder). Assume the geometry of § 2.1 , including (R), and independent uniform ξ n ∈ [ 0 , 1 ] \xi_n \in[0,1] ξ n ∈ [ 0 , 1 ] . For each fixed a ∈ ( 0 , 1 / 2 ) a \in(0,1/2) a ∈ ( 0 , 1/2 ) there is W a > 0 ∗ s u c h t h a t , f o r e v e r y p r e s c r i b e d ∗ $ 0 < W < W a W_a > 0 *such that, for every prescribed* \$0 < W < W_a W a > 0 ∗ s u c h t ha t , f or e v er y p r escr ib e d ∗ $0 < W < W a , the interval
I W , a = ( E a ( W ) , c M ) , Δ a ( W ) = c M − E a ( W ) = a c M W 1 + a W (111) I_{W,a} = (E_a(W),c_M), \qquad\Delta_a(W) = c_M - E_a(W) = \frac{ac_MW}{1+aW} \tag*{(111)} I W , a = ( E a ( W ) , c M ) , Δ a ( W ) = c M − E a ( W ) = 1 + aW a c M W (111) satisfies
E Q ϖ , Λ t ( x , y ; I W , a ) ≤ A a , W e − ν a , W ∣ x − y ∣ 1 , A a , W < ∞ , ν a , W > 0 , (112) \mathbb{E}Q^t_{\varpi,\Lambda}(x,y;I_{W,a}) \le A_{a,W}e^{-\nu_{a,W}|x-y|_1}, \qquad A_{a,W}<\infty,\quad\nu_{a,W}>0, \tag*{(112)} E Q ϖ , Λ t ( x , y ; I W , a ) ≤ A a , W e − ν a , W ∣ x − y ∣ 1 , A a , W < ∞ , ν a , W > 0 , (112) uniformly in all finite Λ \Lambda Λ and x , y ∈ Λ x,y\in\Lambda x , y ∈ Λ . The same bound holds for Q ϖ Q_{\varpi} Q ϖ . Almost surely the spectral subspace of H ϖ H_{\varpi} H ϖ in I W , a I_{W,a} I W , a is pure point and has an orthonormal basis of exponentially decaying eigenfunctions. Moreover,
E sup t ∈ R ∣ ⟨ δ x , e − i t H ϖ 1 I W , a ( H ϖ ) δ y ⟩ ∣ ≤ A a , W e − ν a , W ∣ x − y ∣ 1 . (113) \mathbb{E}\sup_{t\in\mathbb{R}}\left|\left\langle\delta_x,e^{-itH_{\varpi}}\mathbf{1}_{I_{W,a}}(H_{\varpi})\delta_y\right\rangle\right|\le A_{a,W}e^{-\nu_{a,W}|x-y|_1}. \tag*{(113)} E t ∈ R sup ⟨ δ x , e − i t H ϖ 1 I W , a ( H ϖ ) δ y ⟩ ≤ A a , W e − ν a , W ∣ x − y ∣ 1 . (113) In frequency variables the proved interval is
ω M 1 + a W < ω < ω M . (114) \frac{\omega_M}{\sqrt{1+aW}}<\omega<\omega_M. \tag*{(114)} 1 + aW ω M < ω < ω M . (114) The constants depend on the fixed a a a and W W W ; no uniformity as a ↑ 1 / 2 a \uparrow1/2 a ↑ 1/2 or W ↓ 0 W \downarrow0 W ↓ 0 is asserted. The interval is sufficient for localisation and need not be maximal. The proof builds on the spectral-averaging and resolvent estimates of Subsection 5.3 and the finite-volume fractional-moment criterion stated in Subsection 5.4. Its energy width is Δ a ( W ) ∼ a c M W \Delta_a(W)\sim ac_MW Δ a ( W ) ∼ a c M W , whereas its frequency width is ω M [ 1 − ( 1 + a W ) − 1 / 2 ] ∼ ( a / 2 ) ω M W \omega_M[1-(1+aW)^{-1/2}]\sim(a/2)\omega_MW ω M [ 1 − ( 1 + aW ) − 1/2 ] ∼ ( a /2 ) ω M W .
The deficit comparison
With the notation of (74) –(75) , the comparison (76) reads
K = c M Id − J C J ≥ a L L ≥ 0. (115) K=c_M\operatorname{Id}-JCJ\ge a_LL\ge0. \tag*{(115)} K = c M Id − J C J ≥ a L L ≥ 0. (115)
Compression preserves this quadratic-form inequality. The off-diagonal entries of K K K have the same absolute values as those of C C C . No sign assumption on those entries is used below.
A uniform low-energy probability estimate at weak coupling
Consider the additive family in (77) , with
L g , ϖ = K + g Ξ ϖ , g ∈ [ g 0 , 2 g 0 ] , g 0 > 0. (116) L_{g,\varpi}=K+g\Xi_{\varpi},\qquad g\in[g_0,2g_0],\qquad g_0>0. \tag*{(116)} L g , ϖ = K + g Ξ ϖ , g ∈ [ g 0 , 2 g 0 ] , g 0 > 0. (116)
For finite V V V , L g , ϖ , V L_{g,\varpi,V} L g , ϖ , V always denotes its compression. Fix a < b < 1 / 2 a<b<1/2 a < b < 1/2 , and put η b = ( 1 / 2 − b ) / 2 > 0 \eta_b=(1/2-b)/2>0 η b = ( 1/2 − b ) /2 > 0 . Choose a constant q b > 0 q_b>0 q b > 0 such that
a L q b 2 ≥ 1 + 1 4 η b , ℓ 0 = ⌊ q b g 0 − 1 / 2 ⌋ , L = ⌈ g 0 − 2 ⌉ . (117) \frac{a_L}{q_b^2}\ge1+\frac{1}{4\eta_b},\qquad\ell_0=\left\lfloor q_bg_0^{-1/2}\right\rfloor,\qquad L=\left\lceil g_0^{-2}\right\rceil. \tag*{(117)} q b 2 a L ≥ 1 + 4 η b 1 , ℓ 0 = ⌊ q b g 0 − 1/2 ⌋ , L = ⌈ g 0 − 2 ⌉ . (117)
For sufficiently small g 0 g_0 g 0 , ℓ 0 ≥ 2 \ell_0\ge2 ℓ 0 ≥ 2 and L ≫ ℓ 0 L\gg\ell_0 L ≫ ℓ 0 . Partition the coordinate intervals of Λ L = [ − L , L ] d ∩ Z d \Lambda_L=[-L,L]^d\cap\mathbb{Z}^d Λ L = [ − L , L ] d ∩ Z d into integer intervals of lengths between ℓ 0 \ell_0 ℓ 0 and 2 ℓ 0 2\ell_0 2 ℓ 0 , merging a short remainder if necessary. Their products form blocks B B B .
Lemma 7.2 (Rare low eigenvalues). There are constants C b , c b > 0 C_b,c_b>0 C b , c b > 0 , independent of sufficiently small g 0 g_0 g 0 , such that
p g 0 : = P { λ min ( L g 0 , ϖ , Λ L ) ≤ b g 0 } ≤ C b L d exp ( − c b g 0 − d / 2 ) . (118) p_{g_0}:=\mathbb{P}\{\lambda_{\min}(L_{g_0,\varpi,\Lambda_L})\le bg_0\}\le C_bL^d\exp\left(-c_bg_0^{-d/2}\right). \tag*{(118)} p g 0 := P { λ m i n ( L g 0 , ϖ , Λ L ) ≤ b g 0 } ≤ C b L d exp ( − c b g 0 − d /2 ) . (118) On the complementary event, simultaneously for every g ∈ [ g 0 , 2 g 0 ] ∗ a n d e v e r y n o n e m p t y ∗ g\in[g_0,2g_0] *and every nonempty* g ∈ [ g 0 , 2 g 0 ] ∗ an d e v er y n o n e m pt y ∗ V\subset\Lambda_L$, one has λ min ( L g , ϖ , V ) > b g 0 \lambda_{\min}(L_{g,\varpi,V})>bg_0 λ m i n ( L g , ϖ , V ) > b g 0 .
Proof. Compression of (115) , followed by deletion of nonnegative nearest-neighbour edge energies between blocks and at the outer boundary, gives
L g 0 , ϖ , Λ L ≥ ⨁ B ( a L L B N + g 0 Ξ ‾ ϖ , B ) . (119) L_{g_0,\varpi,\Lambda_L} \ge\bigoplus_B \left(a_L L_B^N + g_0\overline{\Xi}_{\varpi,B}\right). \tag*{(119)} L g 0 , ϖ , Λ L ≥ B ⨁ ( a L L B N + g 0 Ξ ϖ , B ) . (119)
Here L B N L_B^N L B N is the graph Neumann Laplacian. For side lengths N j ∈ [ ℓ 0 , 2 ℓ 0 ] N_j \in[\ell_0,2\ell_0] N j ∈ [ ℓ 0 , 2 ℓ 0 ] , its first positive eigenvalue is min j ( 2 − 2 cos ( π / N j ) ) ≥ ℓ 0 − 2 \min_j(2-2\cos(\pi/N_j)) \ge\ell_0^{-2} min j ( 2 − 2 cos ( π / N j )) ≥ ℓ 0 − 2 . Thus T B = a L L B N + g 0 Ξ ‾ ϖ , B T_B=a_LL_B^N+g_0\overline{\Xi}_{\varpi,B} T B = a L L B N + g 0 Ξ ϖ , B has second eigenvalue at least
γ B = a L ℓ 0 − 2 ≥ ( 1 + ( 4 η b ) − 1 ) g 0 . \gamma_B=a_L\ell_0^{-2}\ge(1+(4\eta_b)^{-1})g_0. γ B = a L ℓ 0 − 2 ≥ ( 1 + ( 4 η b ) − 1 ) g 0 .
In the normalised constant vector its mean and variance satisfy
m B = g 0 ξ ‾ B ≤ g 0 , v B = g 0 2 ( ξ B 2 ‾ − ( ξ ‾ B ) 2 ) ≤ g 0 2 / 4. m_B=g_0\overline{\xi}_B\le g_0,\qquad v_B=g_0^2\left(\overline{\xi_B^2}-(\overline{\xi}_B)^2\right)\le g_0^2/4. m B = g 0 ξ B ≤ g 0 , v B = g 0 2 ( ξ B 2 − ( ξ B ) 2 ) ≤ g 0 2 /4.
The last inequality holds for every collection of numbers in [ 0 , 1 ] [0,1] [ 0 , 1 ] . Temple’s inequality, as used in (82) , gives
λ min ( T B ) ≥ m B − v B γ B − m B ≥ g 0 ξ ‾ B − η b g 0 . (120) \lambda_{\min}(T_B)\ge m_B-\frac{v_B}{\gamma_B-m_B}\ge g_0\overline{\xi}_B-\eta_b g_0. \tag*{(120)} λ m i n ( T B ) ≥ m B − γ B − m B v B ≥ g 0 ξ B − η b g 0 . (120)
If λ min ( T B ) ≤ b g 0 \lambda_{\min}(T_B)\le bg_0 λ m i n ( T B ) ≤ b g 0 , then ξ ‾ B ≤ b + η b = 1 / 2 − η b \overline{\xi}_B\le b+\eta_b=1/2-\eta_b ξ B ≤ b + η b = 1/2 − η b . Independence gives
P { ξ ‾ B ≤ 1 / 2 − η b } ≤ e − 2 η b 2 ∣ B ∣ . (121) \mathbb{P}\{\overline{\xi}_B\le1/2-\eta_b\}\le e^{-2\eta_b^2|B|}. \tag*{(121)} P { ξ B ≤ 1/2 − η b } ≤ e − 2 η b 2 ∣ B ∣ . (121)
One elementary verification uses E e − t ( ξ n − 1 / 2 ) ≤ cosh ( t / 2 ) ≤ e t 2 / 8 \mathbb{E}e^{-t(\xi_n-1/2)}\le\cosh(t/2)\le e^{t^2/8} E e − t ( ξ n − 1/2 ) ≤ cosh ( t /2 ) ≤ e t 2 /8 ; exponential Markov inequality with t = 4 η b t=4\eta_b t = 4 η b gives (121) . For small g 0 g_0 g 0 , ℓ 0 ≥ ( q b / 2 ) g 0 − 1 / 2 \ell_0\ge(q_b/2)g_0^{-1/2} ℓ 0 ≥ ( q b /2 ) g 0 − 1/2 . Hence every block has at least ( q b / 2 ) d g 0 − d / 2 (q_b/2)^d g_0^{-d/2} ( q b /2 ) d g 0 − d /2 sites. There are at most ( 2 L + 1 ) d (2L+1)^d ( 2 L + 1 ) d blocks. A union bound in (119) proves (118) .
Finally, increasing g g g increases the operator, and compressing a positive lower form bound preserves it. The good event for the full cube at g 0 g_0 g 0 is therefore good for all its subsets and all larger couplings simultaneously. No union bound over the 2 ∣ Λ L ∣ 2^{|\Lambda_L|} 2 ∣ Λ L ∣ subsets is taken. □
Verification of a fractional-moment criterion
We give the details needed for uniformity in the coupling and spectral parameters. Lemma 5.2 gives, for any finite V V V , N = ∣ V ∣ N=|V| N = ∣ V ∣ , any interval B 0 ⊂ R B_0\subset\mathbb{R} B 0 ⊂ R , and any g > 0 g>0 g > 0 ,
E Tr 1 B 0 ( L g , ϖ , V ) ≤ N ∣ B 0 ∣ / g . (122) \mathbb{E}\operatorname{Tr}1_{B_0}(L_{g,\varpi,V})\le N|B_0|/g. \tag*{(122)} E Tr 1 B 0 ( L g , ϖ , V ) ≤ N ∣ B 0 ∣/ g . (122)
It implies, for 0 < p < 1 0<p<1 0 < p < 1 , u ∈ R u\in\mathbb{R} u ∈ R , and η ≠ 0 \eta\ne0 η = 0 ,
E ∥ ( L g , ϖ , V − u − i η ) − 1 ∥ p ≤ D p ( N , g ) : = ( 2 N / g ) p 1 − p . (123) \mathbb{E}\left\|(L_{g,\varpi,V}-u-i\eta)^{-1}\right\|^p\le D_p(N,g):=\frac{(2N/g)^p}{1-p}. \tag*{(123)} E ( L g , ϖ , V − u − i η ) − 1 p ≤ D p ( N , g ) := 1 − p ( 2 N / g ) p . (123)
Let G g , V ( x , y ; z ) G_{g,V}(x,y;z) G g , V ( x , y ; z ) denote the resolvent kernel, and use the constant T 1 T_1 T 1 from (88) . For 0 < ν ≤ γ 0 / 2 0<\nu\le\gamma_0/2 0 < ν ≤ γ 0 /2 , conjugation by e ν ∣ x − y ∣ 1 e^{\nu|x-y|_1} e ν ∣ x − y ∣ 1 changes the operator by norm at most ν T 1 \nu T_1 ν T 1 , by the Schur test. The diagonal random potential is unchanged. On the good event of Lemma 7.2, the distance to every real u ∈ [ 0 , a g 0 ] u\in[0,ag_0] u ∈ [ 0 , a g 0 ] is at least ( b − a ) g 0 (b-a)g_0 ( b − a ) g 0 . Choosing
ν = min { γ 0 / 2 , ( b − a ) / ( 2 ( T 1 + 1 ) ) } g 0 , g 0 ≤ 1 , \nu=\min\{\gamma_0/2,(b-a)/(2(T_1+1))\}g_0,\qquad g_0\le1, ν = min { γ 0 /2 , ( b − a ) / ( 2 ( T 1 + 1 ))} g 0 , g 0 ≤ 1 ,
a Neumann series for the conjugated resolvent gives, uniformly over all V ⊂ Λ L V\subset\Lambda_L V ⊂ Λ L , couplings in [ g 0 , 2 g 0 ] [g_0,2g_0] [ g 0 , 2 g 0 ] , and η ≠ 0 \eta\ne0 η = 0 ,
∣ G g , V ( x , y ; u + i η ) ∣ ≤ 2 ( b − a ) g 0 e − c g 0 ∣ x − y ∣ 1 on the good event . (124) |G_{g,V}(x,y;u+i\eta)|\le\frac{2}{(b-a)g_0}e^{-cg_0|x-y|_1}\quad\text{on the good event}. \tag*{(124)} ∣ G g , V ( x , y ; u + i η ) ∣ ≤ ( b − a ) g 0 2 e − c g 0 ∣ x − y ∣ 1 on the good event . (124)
Take ϑ = 1 / 8 \vartheta=1/8 ϑ = 1/8 and p = 1 / 2 p=1/2 p = 1/2 . For ∣ y ∣ ∞ ≥ L / 2 |y|_\infty\ge L/2 ∣ y ∣ ∞ ≥ L /2 , Hölder’s inequality on the bad event and (123) imply
sup V ⊂ Λ L E ∣ G g , V ( 0 , y ; u + i η ) ∣ ϑ ≤ C g 0 − ϑ e − c ϑ g 0 L + D p ( N L , g 0 ) ϑ / p p 1 − ϑ / p = : M ( g 0 ) . (125) \sup_{V\subset\Lambda_L}\mathbb{E}|G_{g,V}(0,y;u+i\eta)|^\vartheta\le Cg_0^{-\vartheta}e^{-c\vartheta g_0L}+D_p(N_L,g_0)^{\vartheta/p}p^{1-\vartheta/p}=:M(g_0). \tag*{(125)} V ⊂ Λ L sup E ∣ G g , V ( 0 , y ; u + i η ) ∣ ϑ ≤ C g 0 − ϑ e − c ϑ g 0 L + D p ( N L , g 0 ) ϑ / p p 1 − ϑ / p =: M ( g 0 ) . (125)
A matrix element is defined to be zero if an indicated site is not in V V V . Use τ z \tau_z τ z and X ϑ \mathsf{X}_{\vartheta} X ϑ from Subsection 5.3. For the uniform law, the scalar specialisation of [3 Theorem 3.1] states that the strict bound
( 1 + C ϑ g − ϑ X ϑ ( Λ L ) ) sup V ⊂ Λ L x ∈ Λ L ∑ y ∉ Λ L τ y − x ϑ E ∣ G g , V ( 0 , x ; z ) ∣ ϑ < 1 (126) \left(1+C_{\vartheta}g^{-\vartheta}\mathsf{X}_{\vartheta}(\Lambda_L)\right)\sup_{\substack{V\subset\Lambda_L\\x\in\Lambda_L}}\sum_{y\notin\Lambda_L}\tau_{y-x}^{\vartheta}\mathbb{E}\left|G_{g,V}(0,x;z)\right|^{\vartheta}<1 \tag*{(126)} ( 1 + C ϑ g − ϑ X ϑ ( Λ L ) ) V ⊂ Λ L x ∈ Λ L sup y ∈ / Λ L ∑ τ y − x ϑ E ∣ G g , V ( 0 , x ; z ) ∣ ϑ < 1 (126)
implies exponentially weighted fractional-moment bounds uniformly over truncated matrices. The preparatory estimates and the verification of the regularity and decoupling hypotheses are given in Subsections 5.3 and 5.4; the smallness condition in (126) is verified below. The constant C ϑ C_{\vartheta} C ϑ is independent of g g g ; the coupling dependence is the factor g − ϑ g^{-\vartheta} g − ϑ .
Exponential hopping also gives constants C , c > 0 C,c>0 C , c > 0 such that
X ϑ ( Λ L ) ≤ C L d − 1 , sup ∣ x ∣ ∞ < L / 2 ∑ y ∉ Λ L τ y − x ϑ ≤ C e − c L . (127) \mathsf{X}_{\vartheta}(\Lambda_L)\le CL^{d-1},\qquad\sup_{\lvert x\rvert_{\infty}<L/2}\sum_{y\notin\Lambda_L}\tau_{y-x}^{\vartheta}\le Ce^{-cL}. \tag*{(127)} X ϑ ( Λ L ) ≤ C L d − 1 , ∣ x ∣ ∞ < L /2 sup y ∈ / Λ L ∑ τ y − x ϑ ≤ C e − c L . (127)
These are the boundary sums estimated in (91) . Split (126) at ∣ x ∣ ∞ = L / 2 \lvert x\rvert_{\infty}=L/2 ∣ x ∣ ∞ = L /2 . Its left side is at most
( 1 + C g 0 − ϑ L d − 1 ) [ C N L D ϑ ( N L , g 0 ) e − c L + C L d − 1 M ( g 0 ) ] ≤ C g 0 − P ( e − c / g 0 + e − c g 0 − d / 2 + e − c g 0 − 2 ) ⟶ 0. (128) \begin{aligned}
\left(1+Cg_0^{-\vartheta}L^{d-1}\right)\left[CN_LD_{\vartheta}(N_L,g_0)e^{-cL}+CL^{d-1}\mathcal{M}(g_0)\right]
&\le Cg_0^{-P}\left(e^{-c/g_0}+e^{-cg_0^{-d/2}}+e^{-cg_0^{-2}}\right)\\
&\longrightarrow0. \tag*{(128)}
\end{aligned} ( 1 + C g 0 − ϑ L d − 1 ) [ C N L D ϑ ( N L , g 0 ) e − c L + C L d − 1 M ( g 0 ) ] ≤ C g 0 − P ( e − c / g 0 + e − c g 0 − d /2 + e − c g 0 − 2 ) ⟶ 0. (128)
Here P < ∞ P<\infty P < ∞ is fixed, depending on d , ϑ , p d,\vartheta,p d , ϑ , p , and the constants can depend on a , b , a L , C a,b,a_L,C a , b , a L , C . We used L = ⌈ g 0 − 2 ⌉ L=\lceil g_0^{-2}\rceil L = ⌈ g 0 − 2 ⌉ , g 0 L ≥ g 0 − 1 g_0L\ge g_0^{-1} g 0 L ≥ g 0 − 1 , and (118) . All prefactors are polynomial in g 0 − 1 g_0^{-1} g 0 − 1 , including the factor L d L^d L d in that probability. This explicitly proves smallness as the coupling tends to zero.
Choose g 0 g_0 g 0 sufficiently small that (128) is at most 1 / 2 1/2 1/2 . For this fixed g 0 g_0 g 0 , the same testing cube and margin work for all g ∈ [ g 0 , 2 g 0 ] g\in[g_0,2g_0] g ∈ [ g 0 , 2 g 0 ] , u ∈ [ 0 , a g 0 ] u\in[0,ag_0] u ∈ [ 0 , a g 0 ] , and both signs of η \eta η . The contraction kernels in the cited criterion have a common exponential tail: their finite-cube resolvent factors have the uniform bounds (123) , and their remaining factors are hopping tails. Their exponentially weighted norms therefore tend to their unweighted norms uniformly as the weight tends to zero. The strict margin 1 / 2 1/2 1/2 gives one common positive weight and one finite bound. We have proved the following lemma.
Lemma 7.3 (Uniform additive estimate). For each 0 < a < 1 / 2 0<a<1/2 0 < a < 1/2 and all sufficiently small fixed g 0 > 0 g_0>0 g 0 > 0 , there are B g 0 < ∞ B_{g_0}<\infty B g 0 < ∞ and ν g 0 > 0 \nu_{g_0}>0 ν g 0 > 0 such that
E ∣ G g , V ( x , y ; u + i η ) ∣ ϑ ≤ B g 0 e − ν g 0 ∣ x − y ∣ 1 (129) \mathbb{E}\left|G_{g,V}(x,y;u+i\eta)\right|^{\vartheta}\le B_{g_0}e^{-\nu_{g_0}\lvert x-y\rvert_1} \tag*{(129)} E ∣ G g , V ( x , y ; u + i η ) ∣ ϑ ≤ B g 0 e − ν g 0 ∣ x − y ∣ 1 (129) for all finite V V V , x , y ∈ V x,y\in V x , y ∈ V , g ∈ [ g 0 , 2 g 0 ] g\in[g_0,2g_0] g ∈ [ g 0 , 2 g 0 ] , 0 ≤ u ≤ a g 0 0\le u\le ag_0 0 ≤ u ≤ a g 0 , and η ≠ 0 \eta\ne0 η = 0 . It also holds in infinite volume. The constants need not stay bounded or positive uniformly as g 0 ↓ 0 g_0\downarrow0 g 0 ↓ 0 .
The infinite-volume assertion follows by strong resolvent convergence and Fatou’s lemma; it is also included in the cited criterion.
Transfer to the material-disorder operator
Fix 0 < a < 1 / 2 0<a<1/2 0 < a < 1/2 and take W W W sufficiently small. Set
E − = E a ( W ) , g 0 = W E − , g + = W c M , g + g 0 = 1 + a W ≤ 2. (130) E_-=E_a(W),\qquad g_0=WE_-,\qquad g_+=Wc_M,\qquad\frac{g_+}{g_0}=1+aW\le2. \tag*{(130)} E − = E a ( W ) , g 0 = W E − , g + = W c M , g 0 g + = 1 + aW ≤ 2. (130)
For E ∈ I W , a E\in I_{W,a} E ∈ I W , a , define
A ϖ , Λ ( E ) = K Λ + E W Ξ ϖ , Λ − ( c M − E ) Id = E M ϖ , Λ − J Λ C Λ t J Λ . (131) \mathcal{A}_{\varpi,\Lambda}(E)=K_\Lambda+EW\Xi_{\varpi,\Lambda}-(c_M-E)\operatorname{Id}=EM_{\varpi,\Lambda}-J_\Lambda C_\Lambda^{t}J_\Lambda. \tag*{(131)} A ϖ , Λ ( E ) = K Λ + E W Ξ ϖ , Λ − ( c M − E ) Id = E M ϖ , Λ − J Λ C Λ t J Λ . (131)
Its parameters satisfy E W ∈ [ g 0 , 2 g 0 ] EW\in[g_0,2g_0] E W ∈ [ g 0 , 2 g 0 ] and
0 < c M − E < c M − E − = a W E − = a g 0 . (132) 0<c_M-E<c_M-E_-=aWE_-=ag_0. \tag*{(132)} 0 < c M − E < c M − E − = aW E − = a g 0 . (132)
This common parameter rectangle gives constants uniform across I W , a I_{W,a} I W , a . For fixed real E E E , (122) implies that the finite operator is invertible almost surely. Lemma 7.3 and Fatou’s lemma as η → 0 \eta\to0 η → 0 yield
E ∣ A ϖ , Λ ( E ) − 1 ( x , y ) ∣ ϑ ≤ B W e − ν W ∣ x − y ∣ 1 , E ∈ I W , a , (133) \mathbb{E}\lvert A_{\varpi,\Lambda}(E)^{-1}(x,y)\rvert^{\vartheta} \le B_W e^{-\nu_W\lvert x-y\rvert_1}, \qquad E \in I_{W,a}, \tag*{(133)} E ∣ A ϖ , Λ ( E ) − 1 ( x , y ) ∣ ϑ ≤ B W e − ν W ∣ x − y ∣ 1 , E ∈ I W , a , (133)
where dependence on the fixed a a a is suppressed in B W , ν W B_W,\nu_W B W , ν W . The real-energy congruence (97) does not identify H ϖ H_\varpi H ϖ with a single additive operator. We transfer (133) by material-variable averaging.
Completion of the proof of Theorem 7.1. The preceding estimate has the same form as (96) , uniformly on I W , a I_{W,a} I W , a and over all truncated matrices. Apply the material-variable averaging and resampling argument of Section 6, with E − = E a ( W ) E_- = E_a(W) E − = E a ( W ) and d W = c M W 1 + W d_W = c_M W\sqrt{1+W} d W = c M W 1 + W . It gives
E Q ϖ , Λ t ( x , y ; I W , a ) ≤ 2 ( 1 + W ) W E − d W ϑ ∣ I W , a ∣ B W e − ν W ∣ x − y ∣ 1 . \mathbb{E}Q_{\varpi,\Lambda}^{t}(x,y;I_{W,a}) \le\frac{2(1+W)}{W E_-}d_W^{\vartheta}\lvert I_{W,a}\rvert B_W e^{-\nu_W\lvert x-y\rvert_1}. E Q ϖ , Λ t ( x , y ; I W , a ) ≤ W E − 2 ( 1 + W ) d W ϑ ∣ I W , a ∣ B W e − ν W ∣ x − y ∣ 1 .
The limiting argument in that section proves the infinite-volume correlator bound, pure point spectrum with an exponentially decaying eigenbasis, and the dynamical bound. Theorem 10.1 then applies on I W , a I_{W,a} I W , a . Finally, ω = ω M E / c M \omega=\omega_M\sqrt{E/c_M} ω = ω M E / c M gives (114) . □
A constraint on any upper mobility edge
Let Σ a c ( W ) \Sigma_{\mathrm{ac}}(W) Σ ac ( W ) be the almost-sure closed support of the absolutely continuous spectrum, which is deterministic by ergodicity. When it is nonempty, set E a c + ( W ) = sup Σ a c ( W ) E_{\mathrm{ac}}^{+}(W)=\sup\Sigma_{\mathrm{ac}}(W) E ac + ( W ) = sup Σ ac ( W ) .
Corollary 7.4 (One-sided mobility-edge constraint). Under the assumptions of Theorem 7.1 , if Σ a c ( W ) ≠ ∅ \Sigma_{\mathrm{ac}}(W)\ne\varnothing Σ ac ( W ) = ∅ , then, for every fixed 0 < a < 1 / 2 0<a<1/2 0 < a < 1/2 and sufficiently small W W W ,
E a c + ( W ) ≤ c M 1 + a W . (134) E_{\mathrm{ac}}^{+}(W)\le\frac{c_M}{1+aW}. \tag*{(134)} E ac + ( W ) ≤ 1 + aW c M . (134) The same bound holds for an upper mobility edge separating an absolutely continuous region below from a localised region above, if such a transition exists. Along any sequence W ↓ 0 W\downarrow0 W ↓ 0 for which Σ a c ( W ) \Sigma_{\mathrm{ac}}(W) Σ ac ( W ) is nonempty,
lim inf W ↓ 0 c M − E a c + ( W ) W ≥ c M 2 . (135) \liminf_{W\downarrow0}\frac{c_M-E_{\mathrm{ac}}^{+}(W)}{W}\ge\frac{c_M}{2}. \tag*{(135)} W ↓ 0 lim inf W c M − E ac + ( W ) ≥ 2 c M . (135) Writing ω a c + ( W ) = ω M E a c + ( W ) / c M \omega_{\mathrm{ac}}^{+}(W)=\omega_M\sqrt{E_{\mathrm{ac}}^{+}(W)/c_M} ω ac + ( W ) = ω M E ac + ( W ) / c M , the corresponding frequency bound is
lim inf W ↓ 0 ω M − ω a c + ( W ) W ≥ ω M 4 . (136) \liminf_{W\downarrow0}\frac{\omega_M-\omega_{\mathrm{ac}}^{+}(W)}{W}\ge\frac{\omega_M}{4}. \tag*{(136)} W ↓ 0 lim inf W ω M − ω ac + ( W ) ≥ 4 ω M . (136) Proof. The absolutely continuous spectral measure vanishes on I W , a I_{W,a} I W , a by the theorem, outside [ 0 , c M ] [0,c_M] [ 0 , c M ] by Proposition 2.2 , and at the singleton c M c_M c M by absolute continuity. Hence its closed support is contained in [ 0 , E a ( W ) ] [0,E_a(W)] [ 0 , E a ( W )] . This proves (134) , including when the absolutely continuous support has several components. For each fixed 0 < a < 1 / 2 0<a<1/2 0 < a < 1/2 ,
c M − E a c + ( W ) W ≥ a c M 1 + a W . \frac{c_M-E_{\mathrm{ac}}^{+}(W)}{W}\ge\frac{ac_M}{1+aW}. W c M − E ac + ( W ) ≥ 1 + aW a c M . First take the lower limit along the sequence in question, then let a ↑ 1 / 2 a\uparrow1/2 a ↑ 1/2 . This order of limits does not require a smallness threshold uniform in a a a . Similarly ω a c + ( W ) ≤ ω M ( 1 + a W ) − 1 / 2 \omega_{\mathrm{ac}}^{+}(W)\le\omega_M(1+aW)^{-1/2} ω ac + ( W ) ≤ ω M ( 1 + aW ) − 1/2 and [ 1 − ( 1 + a W ) − 1 / 2 ] / W → a / 2 [1-(1+aW)^{-1/2}]/W\to a/2 [ 1 − ( 1 + aW ) − 1/2 ] / W → a /2 , giving (136) . □
This bound neither establishes nonempty absolutely continuous spectrum nor gives a matching lower bound on its upper endpoint. A finite matrix has discrete spectrum regardless of localisation, so a mobility edge refers here to an infinite-volume transition.
Interpretation and the weak-scattering prediction
In dimension three a weak-scattering calculation predicts a separation of order W 4 W^{4} W 4 between a mobility edge and an approximate coherent edge. The latter differs from E m e a n ( W ) E_{\mathrm{mean}}(W) E mean ( W ) at order W 2 W^{2} W 2 ; both differ from the exact spectral edge at order W W W for the present uncentred inverse-material law. Appendix A gives the calculation and the rigorous unperturbed density-of-states asymptotic on which it is based. The W 4 W^{4} W 4 prediction is heuristic and is not an error estimate for the bound in Corollary 7.4 . Figure 3 separates the proved and predicted scales.
Figure 3. The proved localisation window and the distinct mobility-edge question. Panel (a) shows the sufficient threshold E a ( W ) E_a(W) E a ( W ) and the mean-material edge; the region below E a ( W ) E_a(W) E a ( W ) is not classified by Theorem 7.1 . Panel (b) separates the exact spectral edge from the sufficient threshold and the three-dimensional heuristic levels. The displayed ordering of the latter is illustrative, not an additional theorem. Both panels are schematic and not to scale.
Eigenfunction extent at regular interior energies
The estimates in this section hold for every material configuration in the prescribed range. We use the operators and Fourier convention of § 2.1 , with
c r = c − r ∈ R , ∣ c r ∣ ≤ A 0 e − γ 0 ∣ r ∣ 1 . (137) c_r=c_{-r}\in\mathbb{R},\qquad|c_r|\le A_0e^{-\gamma_0|r|_1}. \tag*{(137)} c r = c − r ∈ R , ∣ c r ∣ ≤ A 0 e − γ 0 ∣ r ∣ 1 . (137)
We use the normalised measure d α ˉ = ( 2 π ) − d d α \mathrm{d}\bar{\alpha}=(2\pi)^{-d}\,\mathrm{d}\alpha d α ˉ = ( 2 π ) − d d α on Y ∗ Y^* Y ∗ .
Centred material variables
With μ W = E m n \mu_W=\mathbb{E}m_n μ W = E m n as above, set
ζ n = ξ n − 1 2 , Z ϖ = diag ( ζ n ) , λ = W μ W . (138) \zeta_n=\xi_n-\frac{1}{2},\qquad Z_\varpi=\operatorname{diag}(\zeta_n),\qquad\lambda=\frac{W}{\mu_W}. \tag*{(138)} ζ n = ξ n − 2 1 , Z ϖ = diag ( ζ n ) , λ = μ W W . (138)
Then E ξ n = 0 \mathbb{E}\xi_n=0 E ξ n = 0 , E ξ n ξ m = σ 2 δ n m \mathbb{E}\xi_n\xi_m=\sigma^2\delta_{nm} E ξ n ξ m = σ 2 δ nm , σ 2 = 1 / 12 \sigma^2=1/12 σ 2 = 1/12 , where δ n m \delta_{nm} δ nm is the Kronecker delta, and ∥ Z ϖ ∥ ≤ 1 / 2 \lVert Z_\varpi\rVert\le1/2 ∥ Z ϖ ∥ ≤ 1/2 . There are two useful exact identities:
( C − s ) q = E W Z ϖ q , s = μ W E , (139) (C-s)q=\mathbb{E}WZ_\varpi q,\qquad s=\mu_W E, \tag*{(139)} ( C − s ) q = E W Z ϖ q , s = μ W E , (139)
H ~ ϖ : = μ W H ϖ = ( Id + λ Z ϖ ) − 1 / 2 C ( Id + λ Z ϖ ) − 1 / 2 . (140) \widetilde{H}_\varpi:=\mu_W H_\varpi=(\operatorname{Id}+\lambda Z_\varpi)^{-1/2}C(\operatorname{Id}+\lambda Z_\varpi)^{-1/2}. \tag*{(140)} H ϖ := μ W H ϖ = ( Id + λ Z ϖ ) − 1/2 C ( Id + λ Z ϖ ) − 1/2 . (140)
Multiplication of H ϖ H_\varpi H ϖ by the scalar μ W \mu_W μ W does not change its eigenvectors, but rescales every spectral value. In particular,
E = s μ W , ω = v r δ ε s μ W , ω ω M = s μ W c M . (141) E=\frac{s}{\mu_W},\qquad\omega=\frac{v_r\sqrt{\delta}}{\varepsilon}\sqrt{\frac{s}{\mu_W}},\qquad\frac{\omega}{\omega_M}=\sqrt{\frac{s}{\mu_Wc_M}}. \tag*{(141)} E = μ W s , ω = ε v r δ μ W s , ω M ω = μ W c M s . (141)
Almost surely, the upper edge of H ~ ϖ \widetilde{H}_\varpi H ϖ is μ W c M \mu_Wc_M μ W c M ; the reference symbol c c c has upper edge c M c_M c M . The random potential in (139) depends on E E E ; it must not be treated as one fixed additive operator when studying all spectral energies at once.
Regular-energy quasimodes
Let I ⋐ ( 0 , c M ) I\Subset(0,c_M) I ⋐ ( 0 , c M ) be a compact interval and assume there exist a > 0 a>0 a > 0 and b ∗ > 0 b_*>0 b ∗ > 0 such that
∣ ∇ c ( α ) ∣ ≥ b ∗ whenever dist ( c ( α ) , I ) ≤ a . (142) |\nabla c(\alpha)|\ge b_*\qquad\text{whenever }\operatorname{dist}(c(\alpha),I)\le a. \tag*{(142)} ∣∇ c ( α ) ∣ ≥ b ∗ whenever dist ( c ( α ) , I ) ≤ a . (142)
In particular, I I I avoids all critical values of c c c . Regularity at one selected carrier is insufficient: disorder scatters to the whole energy surface. Coordinate reflection symmetry does not remove the critical energies associated with vertices other than M M M . Here I I I is an interval of centred energy s s s , so the corresponding eigenvalues of H ϖ H_\varpi H ϖ lie in I / μ W : = { s / μ W : s ∈ I } I/\mu_W:=\{s/\mu_W:s\in I\} I / μ W := { s / μ W : s ∈ I } . This convention also applies to the finite-sample interior estimates below.
For a normalised q ∈ ℓ 2 ( Z d ) q\in\ell^2(\mathbb{Z}^d) q ∈ ℓ 2 ( Z d ) set
IPR ( q ) = ∑ n ∣ q ( n ) ∣ 4 , P ( q ) = IPR ( q ) − 1 , R 2 ( q ; x 0 ) = ( ∑ n ∣ n − x 0 ∣ 2 ∣ q ( n ) ∣ 2 ) 1 / 2 . (143) \operatorname{IPR}(q)=\sum_n|q(n)|^4,\qquad P(q)=\operatorname{IPR}(q)^{-1},\qquad R_2(q;x_0)=\left(\sum_n|n-x_0|^2|q(n)|^2\right)^{1/2}. \tag*{(143)} IPR ( q ) = n ∑ ∣ q ( n ) ∣ 4 , P ( q ) = IPR ( q ) − 1 , R 2 ( q ; x 0 ) = ( n ∑ ∣ n − x 0 ∣ 2 ∣ q ( n ) ∣ 2 ) 1/2 . (143)
The last quantity is allowed to be infinite. It is a radius measured by the second moment, not the inverse of an exponential decay rate.
Lemma 8.1 (A regular-energy resolvent integral). Under (142) , there are K I < ∞ K_I<\infty K I < ∞ and h ∗ > 0 h_*>0 h ∗ > 0 such that
sup e ∈ I ∫ Y ∗ d α ˉ h 2 + ( c ( α ) − e ) 2 ≤ K I h , 0 < h ≤ h ∗ . (144) \sup_{e\in I}\int_{Y^*}\frac{\mathrm{d}\bar{\alpha}}{h^2+(c(\alpha)-e)^2}\le\frac{K_I}{h},\qquad0<h\le h_*. \tag*{(144)} e ∈ I sup ∫ Y ∗ h 2 + ( c ( α ) − e ) 2 d α ˉ ≤ h K I , 0 < h ≤ h ∗ . (144)
Proof. On the compact inverse image of a slightly enlarged interval, the coarea formula and (142) give a uniformly bounded density
ρ c ( t ) = ∫ { c = t } d S ( α ) ( 2 π ) d ∣ ∇ c ( α ) ∣ . \rho_c(t)=\int_{\{c=t\}}\frac{\mathrm{d}S(\alpha)}{(2\pi)^d|\nabla c(\alpha)|}. ρ c ( t ) = ∫ { c = t } ( 2 π ) d ∣∇ c ( α ) ∣ d S ( α ) .
where d S dS d S is the surface measure. Boundedness follows, for example, from a finite collection of submersion charts. For ∣ c − e ∣ < a / 2 |c-e|<a/2 ∣ c − e ∣ < a /2 , integration in t = c ( α ) t=c(\alpha) t = c ( α ) is bounded by ∥ ρ c ∥ ∞ ∫ R ( h 2 + t 2 ) − 1 d t = π ∥ ρ c ∥ ∞ / h \|\rho_c\|_\infty\int_{\mathbb{R}}(h^2+t^2)^{-1}\,dt=\pi\|\rho_c\|_\infty/h ∥ ρ c ∥ ∞ ∫ R ( h 2 + t 2 ) − 1 d t = π ∥ ρ c ∥ ∞ / h . On the complement the denominator is at least a 2 / 4 a^2/4 a 2 /4 , and ∫ Y ∗ d α ˉ = 1 \int_{Y^*}d\bar{\alpha}=1 ∫ Y ∗ d α ˉ = 1 . Increasing the constant proves (144) for any fixed sufficiently small h ∗ h_* h ∗ . No assumption on critical values outside the enlarged interval is needed. □
Proposition 8.2 (Extent of a regular-energy quasimode). Let ∥ q ∥ 2 = 1 \|q\|_2=1 ∥ q ∥ 2 = 1 , e ∈ I e\in I e ∈ I , and r q = ∥ ( C − e ) q ∥ 2 r_q=\|(C-e)q\|_2 r q = ∥ ( C − e ) q ∥ 2 . For 0 < r q ≤ r ∗ 0<r_q\le r_* 0 < r q ≤ r ∗ , with r ∗ > 0 r_*>0 r ∗ > 0 sufficiently small depending only on c c c and I I I ,
∥ q ∥ ∞ 2 ≤ 2 K I r q , P ( q ) ≥ ( 2 K I r q ) − 1 , (145) \|q\|_\infty^2\le2K_Ir_q,\qquad P(q)\ge(2K_Ir_q)^{-1}, \tag*{(145)} ∥ q ∥ ∞ 2 ≤ 2 K I r q , P ( q ) ≥ ( 2 K I r q ) − 1 , (145) R 2 ( q ; x 0 ) ≥ b ∗ 2 4 b 1 r q − b 2 2 b 1 , x 0 ∈ R d , (146) R_2(q;x_0)\ge\frac{b_*^2}{4b_1r_q}-\frac{b_2}{2b_1},\qquad x_0\in\mathbb{R}^d, \tag*{(146)} R 2 ( q ; x 0 ) ≥ 4 b 1 r q b ∗ 2 − 2 b 1 b 2 , x 0 ∈ R d , (146) where b 1 = sup Y ∗ ∣ ∇ c ∣ > 0 b_1=\sup_{Y^*}|\nabla c|>0 b 1 = sup Y ∗ ∣∇ c ∣ > 0 and b 2 = sup Y ∗ ∣ Δ α c ∣ b_2=\sup_{Y^*}|\Delta_\alpha c| b 2 = sup Y ∗ ∣ Δ α c ∣ . For any η 0 ∈ ( 0 , 1 ] \eta_0\in(0,1] η 0 ∈ ( 0 , 1 ] , a set containing at least the fraction η 0 \eta_0 η 0 of the squared norm has at least η 0 / ( 2 K I r q ) \eta_0/(2K_Ir_q) η 0 / ( 2 K I r q ) lattice sites.
Proof. By Fourier inversion and Cauchy–Schwarz, for each n n n and h > 0 h>0 h > 0 ,
∣ q ( n ) ∣ 2 ≤ ( ∫ d α ˉ h 2 + ( c − e ) 2 ) ( ∫ [ h 2 + ( c − e ) 2 ] ∣ q ^ ∣ 2 d α ˉ ) ≤ K I h ( h 2 + r q 2 ) . \begin{aligned}
|q(n)|^2\le\left(\int\frac{d\bar{\alpha}}{h^2+(c-e)^2}\right)\left(\int[h^2+(c-e)^2]|\widehat q|^2\,d\bar{\alpha}\right) \\
&\le\frac{K_I}{h}(h^2+r_q^2).
\end{aligned} ∣ q ( n ) ∣ 2 ≤ ( ∫ h 2 + ( c − e ) 2 d α ˉ ) ( ∫ [ h 2 + ( c − e ) 2 ] ∣ q ∣ 2 d α ˉ ) ≤ h K I ( h 2 + r q 2 ) .
Taking h = r q h=r_q h = r q proves the first inequality. The participation and cardinality bounds follow from ∑ ∣ q ∣ 4 ≤ ∥ q ∥ ∞ 2 ∥ q ∥ 2 2 \sum|q|^4\le\|q\|_\infty^2\|q\|_2^2 ∑ ∣ q ∣ 4 ≤ ∥ q ∥ ∞ 2 ∥ q ∥ 2 2 and ∑ n ∈ S ∣ q ( n ) ∣ 2 ≤ ∣ S ∣ ∥ q ∥ ∞ 2 \sum_{n\in S}|q(n)|^2\le|S|\|q\|_\infty^2 ∑ n ∈ S ∣ q ( n ) ∣ 2 ≤ ∣ S ∣∥ q ∥ ∞ 2 .
For the radius estimate, the case R 2 = ∞ R_2=\infty R 2 = ∞ is immediate. Otherwise q ^ ∈ H 1 ( Y ∗ ) \widehat q\in H^1(Y^*) q ∈ H 1 ( Y ∗ ) . Let X j q ( n ) = n j q ( n ) X_jq(n)=n_jq(n) X j q ( n ) = n j q ( n ) and let V j V_j V j be the Fourier multiplier ∂ j c \partial_jc ∂ j c . On this domain introduce the symmetric operator
A x 0 = 1 2 ∑ j = 1 d [ V j ( X j − x 0 j ) + ( X j − x 0 j ) V j ] . A_{x_0}=\frac{1}{2}\sum_{j=1}^{d}\left[V_j(X_j-x_{0j})+(X_j-x_{0j})V_j\right]. A x 0 = 2 1 j = 1 ∑ d [ V j ( X j − x 0 j ) + ( X j − x 0 j ) V j ] .
In our Fourier convention X j = − i ∂ α j X_j=-i\partial_{\alpha_j} X j = − i ∂ α j , hence
i [ A x 0 , C ] = ∣ ∇ c ∣ 2 , ∥ A x 0 q ∥ 2 ≤ b 1 R 2 ( q ; x 0 ) + 1 2 b 2 . (147) i[A_{x_0},C]=|\nabla c|^2,\qquad\|A_{x_0}q\|_2\le b_1R_2(q;x_0)+\frac{1}{2}b_2. \tag*{(147)} i [ A x 0 , C ] = ∣∇ c ∣ 2 , ∥ A x 0 q ∥ 2 ≤ b 1 R 2 ( q ; x 0 ) + 2 1 b 2 . (147)
These identities hold first on smooth Fourier functions and then on H 1 H^1 H 1 by density. The spectral weight where ∣ c − e ∣ ≥ a |c-e|\ge a ∣ c − e ∣ ≥ a is at most r q 2 / a 2 r_q^2/a^2 r q 2 / a 2 . Thus (142) gives
⟨ q , i [ A x 0 , C ] q ⟩ = ∫ ∣ ∇ c ∣ 2 ∣ q ^ ∣ 2 d α ˉ ≥ b ∗ 2 ( 1 − r q 2 / a 2 ) . \langle q,i[A_{x_0},C]q\rangle=\int|\nabla c|^2|\widehat q|^2\,d\bar{\alpha}\ge b_*^2(1-r_q^2/a^2). ⟨ q , i [ A x 0 , C ] q ⟩ = ∫ ∣∇ c ∣ 2 ∣ q ∣ 2 d α ˉ ≥ b ∗ 2 ( 1 − r q 2 / a 2 ) .
Writing C q = e q + ( C − e ) q Cq=eq+(C-e)q C q = e q + ( C − e ) q , the same commutator expectation has absolute value at most 2 r q ∥ A x 0 q ∥ 2 2r_q\|A_{x_0}q\|_2 2 r q ∥ A x 0 q ∥ 2 . Choose r ∗ ≤ a / 2 r_*\le a/\sqrt{2} r ∗ ≤ a / 2 and r ∗ ≤ h ∗ r_*\le h_* r ∗ ≤ h ∗ . Then b ∗ 2 / 2 ≤ 2 r q ( b 1 R 2 + b 2 / 2 ) b_*^2/2\le2r_q(b_1R_2+b_2/2) b ∗ 2 /2 ≤ 2 r q ( b 1 R 2 + b 2 /2 ) , which is (146) . If r q = 0 r_q=0 r q = 0 , the first estimate with h ↓ 0 h\downarrow0 h ↓ 0 would imply q = 0 q=0 q = 0 , so there is no nonzero normalised vector of that kind. □
Theorem 8.3 (Weak-disorder lower bounds on eigenfunction extent). Assume (137) and (142) . There are constants W 0 , C I , c I , C I ′ > 0 W_0,C_I,c_I,C'_I>0 W 0 , C I , c I , C I ′ > 0 such that the following holds for every configuration 0 ≤ ξ n ≤ 1 0\le\xi_n\le1 0 ≤ ξ n ≤ 1 and 0 < W ≤ W 0 0<W\le W_0 0 < W ≤ W 0 . If
C q = E M ϖ q , ∥ q ∥ 2 = 1 , μ W E ∈ I , (148) Cq=E M_{\varpi}q,\qquad\|q\|_2=1,\qquad\mu_W E\in I, \tag*{(148)} C q = E M ϖ q , ∥ q ∥ 2 = 1 , μ W E ∈ I , (148) then
∥ q ∥ ∞ 2 ≤ C I W , P ( q ) ≥ c I W , R 2 ( q ; x 0 ) ≥ c I W − C I ′ for every x 0 ∈ R d . (149) \|q\|_\infty^2\le C_IW,\qquad P(q)\ge\frac{c_I}{W},\qquad R_2(q;x_0)\ge\frac{c_I}{W}-C'_I\quad\text{for every }x_0\in\mathbb{R}^d. \tag*{(149)} ∥ q ∥ ∞ 2 ≤ C I W , P ( q ) ≥ W c I , R 2 ( q ; x 0 ) ≥ W c I − C I ′ for every x 0 ∈ R d . (149)
The conclusions, with adjusted uniform constants, hold for the ordinary ℓ 2 \ell^2 ℓ 2 -normalised eigenvector of H ϖ H_{\varpi} H ϖ .
Proof. Identity (139) gives ∥ ( C − μ W E ) q ∥ 2 ≤ E W / 2 \|(C-\mu_W E)q\|_2\le EW/2 ∥ ( C − μ W E ) q ∥ 2 ≤ E W /2 . Since μ W E ∈ I \mu_W E\in I μ W E ∈ I , E E E is bounded above uniformly. For sufficiently small W W W , this residual is at most r ∗ r_* r ∗ . Apply Proposition 8.2 ; replacing the residual by its upper bound weakens the estimates in the required direction. For a normalised eigenvector ψ \psi ψ of H ϖ H_{\varpi} H ϖ , put q = M ϖ − 1 / 2 ψ q=M_{\varpi}^{-1/2}\psi q = M ϖ − 1/2 ψ . Its norm is between ( 1 + W ) − 1 / 2 (1+W)^{-1/2} ( 1 + W ) − 1/2 and one. Normalise q q q and use 1 ≤ m n ≤ 1 + W 1\le m_n\le1+W 1 ≤ m n ≤ 1 + W . Since M ϖ M_{\varpi} M ϖ commutes with position, both pointwise and second-moment estimates transfer with factors uniformly bounded for W ≤ W 0 W\le W_0 W ≤ W 0 . □
The bounds are compatible with exponentially localised eigenfunctions whose extent diverges as W ↓ 0 W\downarrow0 W ↓ 0 . An exponential upper bound with an uncontrolled prefactor does not yield a decay-rate estimate from (149) alone. The participation number measures the number of occupied sites, whereas R 2 R_2 R 2 also measures their separation.
Verification for crystals of spheres
Let D = B r ( 0 ) D=B_r(0) D = B r ( 0 ) with 0 < r < 1 / 2 0<r<1/2 0 < r < 1/2 . In dimension three this is a simple cubic crystal of spheres; in dimension two it is the corresponding square crystal of disks.
Explicit geometric constants
Choose r < R < 1 / 2 r<R<1/2 r < R < 1/2 and an H 1 H^1 H 1 radial cutoff equal to one in B r B_r B r , harmonic in the annulus B R ∖ B r ‾ B_R\setminus\overline{B_r} B R ∖ B r , and zero outside B R B_R B R . Its normalised energy is
E χ = 1 ∣ D ∣ { 2 π / log ( R / r ) , d = 2 , 4 π / ( r − 1 − R − 1 ) , d = 3. (150) E_\chi=\frac{1}{|D|}
\begin{cases}
2\pi/\log(R/r), & d=2,\\
4\pi/(r^{-1}-R^{-1}), & d=3.
\end{cases} \tag*{(150)} E χ = ∣ D ∣ 1 { 2 π / log ( R / r ) , 4 π / ( r − 1 − R − 1 ) , d = 2 , d = 3. (150)
The three-dimensional formula follows by differentiating ( ∣ x ∣ − 1 − R − 1 ) / ( r − 1 − R − 1 ) (|x|^{-1}-R^{-1})/(r^{-1}-R^{-1}) ( ∣ x ∣ − 1 − R − 1 ) / ( r − 1 − R − 1 ) and integrating its squared gradient; in two dimensions use log ( R / ∣ x ∣ 2 ) / log ( R / r ) \log(R/|x|_2)/\log(R/r) log ( R /∣ x ∣ 2 ) / log ( R / r ) . The constants C P C_P C P , γ 0 \gamma_0 γ 0 , A 0 A_0 A 0 in (26) and (34) are now explicit.
At α M \alpha_M α M , all quasiperiodic Fourier wavevectors have squared length at least d π 2 d\pi^2 d π 2 . Since an admissible function equals one in D D D ,
c ∗ : = d π 2 ≤ c M ≤ E χ . (151) c_*:=d\pi^2\le c_M\le E_\chi. \tag*{(151)} c ∗ := d π 2 ≤ c M ≤ E χ . (151)
These are bounds for the normalised symbol. Without volume normalisation they read d π 2 ∣ D ∣ ≤ Cap α M , D ≤ ∣ D ∣ E χ d\pi^2|D|\le\operatorname{Cap}_{\alpha_M,D}\le|D|E_\chi d π 2 ∣ D ∣ ≤ Cap α M , D ≤ ∣ D ∣ E χ . For d = 3 d=3 d = 3 , r = 1 / 5 r=1/5 r = 1/5 , and R = 2 / 5 R=2/5 R = 2/5 , they give 3 π 2 ≤ c M ≤ 150 3\pi^2\le c_M\le150 3 π 2 ≤ c M ≤ 150 .
The fixed-interval hypothesis
For any ϑ ∈ ( 0 , 1 ) \vartheta\in(0,1) ϑ ∈ ( 0 , 1 ) and 0 < μ < ϑ γ 0 0<\mu<\vartheta\gamma_0 0 < μ < ϑ γ 0 , the geometric series gives
T ϑ ( μ ) ≤ A 0 ϑ [ coth ( ϑ γ 0 − μ 2 ) d − 1 ] . (152) T_\vartheta(\mu)\le A_0^\vartheta\left[\coth\left(\frac{\vartheta\gamma_0-\mu}{2}\right)^d-1\right]. \tag*{(152)} T ϑ ( μ ) ≤ A 0 ϑ [ coth ( 2 ϑ γ 0 − μ ) d − 1 ] . (152)
In particular put ϑ = 1 / 2 \vartheta=1/2 ϑ = 1/2 , μ = γ 0 / 4 \mu=\gamma_0/4 μ = γ 0 /4 , and
T ‾ = A 0 [ coth ( γ 0 / 8 ) d − 1 ] , I ∗ = ( c ∗ / 2 , 3 c ∗ / 4 ) , W s p h = 64 T ‾ 2 / c ∗ . (153) \overline{T}=\sqrt{A_0\left[\coth(\gamma_0/8)^d-1\right]},\qquad I_*=(c_*/2,3c_*/4),\qquad W_{\mathrm{sph}}=64\overline{T}^{\,2}/c_*. \tag*{(153)} T = A 0 [ coth ( γ 0 /8 ) d − 1 ] , I ∗ = ( c ∗ /2 , 3 c ∗ /4 ) , W sph = 64 T 2 / c ∗ . (153)
For the inverse-material law of width W = W s p h W=W_{\mathrm{sph}} W = W sph ,
q 1 ≤ 4 T ‾ W c ∗ = 1 2 . q_1\le\frac{4\overline{T}}{\sqrt{W}c_*}=\frac{1}{2}. q 1 ≤ W c ∗ 4 T = 2 1 .
Thus Theorem 4.1 and Corollary 6.4 apply on the fixed interval I ∗ I_* I ∗ ; they also apply for any larger width. These conservative sufficient bounds do not estimate the onset of localisation. Table 2 gives rounded numerical evaluations for r = 1 / 5 r=1/5 r = 1/5 , R = 2 / 5 R=2/5 R = 2/5 .
Table 2. Explicit normalised geometric bounds. The very large sufficient widths reflect the rough analytic upper bound on the interaction sum.
The band-edge hypotheses
Proposition 3.2, Theorem 3.4, and Proposition 3.6 give exponential decay of the coefficients and a unique nondegenerate upper maximum. They imply (76) , so Theorem 6.1 applies for every fixed W > 0 W > 0 W > 0 and 0 < r < 1 / 2 0 < r < 1/2 0 < r < 1/2 . The truncated and finite capacitance matrices satisfy Theorems 10.1 and 15.3, respectively. For fixed a ∈ ( 0 , 1 / 2 ) a \in(0, 1/2) a ∈ ( 0 , 1/2 ) and sufficiently small W W W , Section 7 gives the window of width a c M W / ( 1 + a W ) ac_MW/(1 + aW) a c M W / ( 1 + aW ) ; its smallness threshold is not evaluated numerically here. No large-disorder threshold from Table 2 is required for these edge results.
Coordinate-permutation symmetry makes the edge Hessian a scalar multiple of the identity, as used in Appendix A. The interior estimates of Section 8 apply to intervals avoiding all critical values, including those at the other reflection vertices. For the direct material law, Proposition B.1 gives absence of absolutely continuous spectrum near the rescaled edge.
Part II. Truncated capacitance matrices
The estimates in this part are the intermediate step from the infinite crystal to the finite capacitance matrix. Truncated matrices retain the periodic interaction and enjoy bounds uniform up to their boundary. We also prove a spacing estimate that applies to both finite constructions.
Uniform localisation for truncated matrices
We derive quantitative concentration bounds for eigenvectors of truncated matrices and verify that their constants remain uniform for the contrast-dependent capacitance family at small contrast.
Simultaneous concentration
We formulate the simultaneous eigenvector bounds both on the lattice and in physical coordinates, using the following description of the sample.
Let Ω ⊂ R d \Omega\subset\mathbb{R}^d Ω ⊂ R d be a nonempty bounded Lipschitz domain and set
I ε = { n ∈ Z d : ε ( n + Y ) ⊂ Ω } , N ε = ∣ I ε ∣ ≍ ε − d . (154) I_\varepsilon= \{n \in\mathbb{Z}^d : \varepsilon(n + Y) \subset\Omega\}, \qquad N_\varepsilon= |I_\varepsilon| \asymp\varepsilon^{-d}. \tag*{(154)} I ε = { n ∈ Z d : ε ( n + Y ) ⊂ Ω } , N ε = ∣ I ε ∣ ≍ ε − d . (154)
Define the isometry J ε : ℓ 2 ( I ε ) → L 2 ( Ω ) J_\varepsilon: \ell^2(I_\varepsilon) \to L^2(\Omega) J ε : ℓ 2 ( I ε ) → L 2 ( Ω ) by ( J ε q ) ( x ) = ε − d / 2 q n (J_\varepsilon q)(x) = \varepsilon^{-d/2}q_n ( J ε q ) ( x ) = ε − d /2 q n for x ∈ ε ( n + Y ) x \in\varepsilon(n + Y) x ∈ ε ( n + Y ) , and zero elsewhere.
Theorem 10.1 (Simultaneous localisation of finite-volume edge modes). Assume the geometry above, including (R), and the material law (2.5). For every prescribed W > 0 W > 0 W > 0 there are Δ W ∈ ( 0 , c M / 2 ) \Delta_W \in(0, c_M/2) Δ W ∈ ( 0 , c M /2 ) , A W < ∞ A_W < \infty A W < ∞ , and ν W > 0 \nu_W > 0 ν W > 0 such that, with I W = ( c M − Δ W , c M ) I_W = (c_M - \Delta_W, c_M) I W = ( c M − Δ W , c M ) ,
E Q ω , Λ ( x , y ; I W ) ≤ A W e − ν W ∣ x − y ∣ 1 (155) \mathbb{E}Q_{\omega,\Lambda}(x,y;I_W) \leq A_W e^{-\nu_W |x-y|_1} \tag*{(155)} E Q ω , Λ ( x , y ; I W ) ≤ A W e − ν W ∣ x − y ∣ 1 (155)
for every finite Λ ⊂ Z d \Lambda\subset\mathbb{Z}^{d} Λ ⊂ Z d and x , y ∈ Λ x,y \in\Lambda x , y ∈ Λ . For each p > 0 p>0 p > 0 , there are R W , p , C W , p < ∞ R_{W,p}, C_{W,p}<\infty R W , p , C W , p < ∞ , independent of Λ \Lambda Λ , such that, with probability at least 1 − C W , p N − p 1-C_{W,p}N^{-p} 1 − C W , p N − p , N = ∣ Λ ∣ ≥ 2 N=|\Lambda|\ge2 N = ∣Λ∣ ≥ 2 , every normalised eigenvector ψ j \psi_j ψ j with eigenvalue E j ∈ I W E_j\in I_W E j ∈ I W has a centre n j ∈ Λ n_j\in\Lambda n j ∈ Λ for which
∣ ψ j ( n ) ∣ ≤ N p + 5 / 2 e − ν W ∣ n − n j ∣ 1 / 2 , n ∈ Λ , (156) |\psi_j(n)| \le N^{p+5/2}e^{-\nu_W|n-n_j|^{1/2}}, \qquad n\in\Lambda, \tag*{(156)} ∣ ψ j ( n ) ∣ ≤ N p + 5/2 e − ν W ∣ n − n j ∣ 1/2 , n ∈ Λ , (156)
∑ ∣ n − n j ∣ 1 > R W , p log ( 2 + N ) ∣ ψ j ( n ) ∣ 2 ≤ N − p . (157) \sum_{|n-n_j|_1>R_{W,p}\log(2+N)}|\psi_j(n)|^2 \le N^{-p}. \tag*{(157)} ∣ n − n j ∣ 1 > R W , p l o g ( 2 + N ) ∑ ∣ ψ j ( n ) ∣ 2 ≤ N − p . (157)
Both estimates hold simultaneously for all modes in I W I_W I W . For Λ = I ε \Lambda=I_\varepsilon Λ = I ε , put x j , ε = ε n j x_{j,\varepsilon}=\varepsilon n_j x j , ε = ε n j and r ε = ε R W , p log ( 2 + N ε ) + d ε / 2 r_\varepsilon=\varepsilon R_{W,p}\log(2+N_\varepsilon)+\sqrt{d}\,\varepsilon/2 r ε = ε R W , p log ( 2 + N ε ) + d ε /2 . Then, on the same event,
∫ Ω ∖ B ( x j , ε , r ε ) ∣ J ε ψ j ( x ) ∣ 2 d x ≤ N ε − p , r ε = O W , p ( ε log ( 1 / ε ) ) . (158) \int_{\Omega\setminus B(x_{j,\varepsilon},r_\varepsilon)}|\mathcal{J}_\varepsilon\psi_j(x)|^2\,dx\le N_\varepsilon^{-p}, \qquad r_\varepsilon=O_{W,p}(\varepsilon\log(1/\varepsilon)). \tag*{(158)} ∫ Ω ∖ B ( x j , ε , r ε ) ∣ J ε ψ j ( x ) ∣ 2 d x ≤ N ε − p , r ε = O W , p ( ε log ( 1/ ε )) . (158)
For q j = M ϖ , I ε − 1 / 2 ψ j q_j=M_{\varpi,I_\varepsilon}^{-1/2}\psi_j q j = M ϖ , I ε − 1/2 ψ j , the normalised field J ε q j / ∥ q j ∥ 2 \mathcal{J}_\varepsilon q_j/\|q_j\|_2 J ε q j /∥ q j ∥ 2 satisfies the same concentration estimate with right-hand side ( 1 + W ) N ε − p (1+W)N_\varepsilon^{-p} ( 1 + W ) N ε − p .
The constants are uniform in sample size for fixed W W W . Along ε = k − 1 \varepsilon=k^{-1} ε = k − 1 , the concentration holds eventually almost surely if d p > 1 dp>1 d p > 1 , by Borel–Cantelli. The same conclusion holds on the large-disorder intervals of Corollary 6.4.
Proof of the finite-volume theorem
Proof of Theorem 10.1 . The correlator estimate is (98) . Write N = ∣ Λ ∣ ≥ 2 N=|\Lambda|\ge2 N = ∣Λ∣ ≥ 2 . Markov’s inequality and a union bound show that
Q ϖ , Λ t ( x , y ; I W ) ≤ N p + 2 e − ν W ∣ x − y ∣ 1 / 2 ( x , y ∈ Λ ) Q^t_{\varpi,\Lambda}(x,y;I_W)\le N^{p+2}e^{-\nu_W|x-y|^{1/2}}\qquad(x,y\in\Lambda) Q ϖ , Λ t ( x , y ; I W ) ≤ N p + 2 e − ν W ∣ x − y ∣ 1/2 ( x , y ∈ Λ )
fails with probability at most
A W N − p − 2 ∑ x , y ∈ Λ e − ν W ∣ x − y ∣ 1 / 2 ≤ C W N − p − 1 . A_WN^{-p-2}\sum_{x,y\in\Lambda}e^{-\nu_W|x-y|^{1/2}}\le C_WN^{-p-1}. A W N − p − 2 x , y ∈ Λ ∑ e − ν W ∣ x − y ∣ 1/2 ≤ C W N − p − 1 .
On this common event choose n j n_j n j where ∣ ψ j ∣ |\psi_j| ∣ ψ j ∣ is maximal. Then ∣ ψ j ( n j ) ∣ ≥ N − 1 / 2 |\psi_j(n_j)|\ge N^{-1/2} ∣ ψ j ( n j ) ∣ ≥ N − 1/2 . The finite eigenvalues are simple almost surely by the resampling argument, and therefore ∣ ψ j ( n ) ψ j ( n j ) ∣ ≤ Q ϖ , Λ t ( n , n j ; I W ) |\psi_j(n)\psi_j(n_j)|\le Q^t_{\varpi,\Lambda}(n,n_j;I_W) ∣ ψ j ( n ) ψ j ( n j ) ∣ ≤ Q ϖ , Λ t ( n , n j ; I W ) . This proves (156) for all modes at once. Summing its square outside lattice radius R R R gives
∑ ∣ n − n j ∣ 1 > R ∣ ψ j ( n ) ∣ 2 ≤ C W N 2 p + 5 e − ν W R / 2 . \sum_{|n-n_j|_1>R}|\psi_j(n)|^2\le C_WN^{2p+5}e^{-\nu_WR/2}. ∣ n − n j ∣ 1 > R ∑ ∣ ψ j ( n ) ∣ 2 ≤ C W N 2 p + 5 e − ν W R /2 .
Choose R = R W , p log ( 2 + N ) R=R_{W,p}\log(2+N) R = R W , p log ( 2 + N ) with R W , p R_{W,p} R W , p sufficiently large to obtain (157) . Every point of ε ( n + Y ) \varepsilon(n+Y) ε ( n + Y ) lies at Euclidean distance at most ε ∣ n − n j ∣ 1 + d ε / 2 \varepsilon|n-n_j|_1+\sqrt{d}\,\varepsilon/2 ε ∣ n − n j ∣ 1 + d ε /2 from ε n j \varepsilon n_j ε n j . The isometry J ε \mathcal{J}_\varepsilon J ε consequently gives (158) . Finally, ( 1 + W ) − 1 / 2 ≤ ∥ M − 1 / 2 ψ j ∥ 2 ≤ 1 (1+W)^{-1/2}\le\|M^{-1/2}\psi_j\|_2\le1 ( 1 + W ) − 1/2 ≤ ∥ M − 1/2 ψ j ∥ 2 ≤ 1 and ∣ M − 1 / 2 ψ j ( n ) ∣ ≤ ∣ ψ j ( n ) ∣ |M^{-1/2}\psi_j(n)|\le|\psi_j(n)| ∣ M − 1/2 ψ j ( n ) ∣ ≤ ∣ ψ j ( n ) ∣ . Normalising the amplitudes costs at most the factor 1 + W 1+W 1 + W in the squared tail. Corollary 6.4 proves the same assertions on its large-disorder interval. □
Remark 10.2 (Number of edge modes). The statement is not restricted to a fixed number of modes. By Theorem 14.3 , a fixed nonempty open subinterval of I W I_W I W contains a positive density of eigenvalues in truncated matrices on growing cubes: choose a nonnegative continuous function supported in that interval with positive density-of-states integral. Thus the simultaneous estimate controls a number of modes proportional to the sample size as the cubes grow.
Uniformity in the contrast parameter
The simultaneous period and contrast limit requires constants that remain uniform as δ → 0 \delta\to0 δ → 0 . We obtain this uniformity for the family C δ C_\delta C δ , allowing the choice δ = ε 2 \delta= \varepsilon^2 δ = ε 2 .
Proposition 10.3. Assume (R) and the material law (8) . For the family in Proposition 3.8 and every fixed W > 0 W > 0 W > 0 , there are Δ W > 0 \Delta_W > 0 Δ W > 0 , A W < ∞ A_W < \infty A W < ∞ , and ν W > 0 \nu_W > 0 ν W > 0 , independent of sufficiently small δ \delta δ , such that
E Q δ t , ϖ , Λ ( x , y ; I W , δ ) ≤ A W e − ν W ∣ x − y ∣ 1 , I W , δ = ( c M , δ − Δ W , c M , δ ) , (159) \mathbb{E}Q_{\delta}^{t,\varpi,\Lambda}(x,y;I_{W,\delta}) \leq A_W e^{-\nu_W |x-y|_1}, \qquad I_{W,\delta} = (c_{M,\delta}-\Delta_W,c_{M,\delta}), \tag*{(159)} E Q δ t , ϖ , Λ ( x , y ; I W , δ ) ≤ A W e − ν W ∣ x − y ∣ 1 , I W , δ = ( c M , δ − Δ W , c M , δ ) , (159) for all finite Λ \Lambda Λ and x , y ∈ Λ x,y \in\Lambda x , y ∈ Λ . Here Q δ t Q_{\delta}^{t} Q δ t is (11) with C C C replaced by C δ C_\delta C δ . For every p > 0 p > 0 p > 0 , every deterministic sufficiently small δ ≥ 0 \delta\geq0 δ ≥ 0 , and every finite Λ \Lambda Λ with N = ∣ Λ ∣ ≥ 2 N = |\Lambda| \geq2 N = ∣Λ∣ ≥ 2 , the simultaneous conclusions of Theorem 10.1 hold with probability at least 1 − C W , p N − p 1-C_{W,p}N^{-p} 1 − C W , p N − p , with constants C W , p C_{W,p} C W , p and R W , p R_{W,p} R W , p independent of δ \delta δ and Λ \Lambda Λ . The event may depend on δ \delta δ and Λ \Lambda Λ ; simultaneity concerns all eigenvectors with eigenvalues in I W , δ I_{W,\delta} I W , δ .
On a prescribed bounded interval I = ( E 0 , E 1 ) I=(E_0,E_1) I = ( E 0 , E 1 ) with E 0 > 0 E_0>0 E 0 > 0 , the same conclusions hold, without reflection symmetry, if
W > 1 E 0 ( 2 ϑ T ϑ ∗ ( μ ) 1 − ϑ ) 1 / ϑ , T ϑ ∗ ( μ ) = sup 0 ≤ δ ≤ δ 0 ∑ r ≠ 0 ∣ c δ , r ∣ r ϑ e μ ∣ r ∣ 1 , 0 < μ < ϑ γ ∗ . (160) W > \frac{1}{E_0}\left(\frac{2^{\vartheta}T_{\vartheta}^{*}(\mu)}{1-\vartheta}\right)^{1/\vartheta}, \qquad T_{\vartheta}^{*}(\mu)=\sup_{0\leq\delta\leq\delta_0}\sum_{r\ne0}|c_{\delta,r}|r^{\vartheta}e^{\mu|r|_1}, \qquad0<\mu<\vartheta\gamma_*. \tag*{(160)} W > E 0 1 ( 1 − ϑ 2 ϑ T ϑ ∗ ( μ ) ) 1/ ϑ , T ϑ ∗ ( μ ) = 0 ≤ δ ≤ δ 0 sup r = 0 ∑ ∣ c δ , r ∣ r ϑ e μ ∣ r ∣ 1 , 0 < μ < ϑ γ ∗ . (160) For each fixed 0 < a < 1 / 2 0<a<1/2 0 < a < 1/2 , the quantitative window ( c M , δ / ( 1 + a W ) , c M , δ ) (c_{M,\delta}/(1+aW),c_{M,\delta}) ( c M , δ / ( 1 + aW ) , c M , δ ) likewise has uniform constants for sufficiently small fixed W W W . In particular, these uniform probability bounds apply for each ε \varepsilon ε under the deterministic choice δ = ε 2 \delta=\varepsilon^2 δ = ε 2 .
Proof. Let J J J be the staggering operator in (74) and put K δ = c M , δ Id − J C δ J K_\delta=c_{M,\delta}\operatorname{Id}-JC_\delta J K δ = c M , δ Id − J C δ J . Proposition 3.8 gives K δ ≥ a ∗ L K_\delta\geq a_*L K δ ≥ a ∗ L , a common exponential coefficient bound, and c ∗ ≤ c M , δ ≤ c ∗ c_*\leq c_{M,\delta}\leq c^* c ∗ ≤ c M , δ ≤ c ∗ . In Proposition 5.1 and Lemma 5.3 , take the common coupling range [ W c ∗ / 2 , W c ∗ ] [Wc_*/2,Wc^*] [ W c ∗ /2 , W c ∗ ] . The Neumann-block gap, the Bernoulli probability, the resolvent conjugation constant, and the hopping boundary sums then have constants independent of δ \delta δ . One testing cube and one strict contraction margin work for the whole family. Choose Δ W < min ( c ∗ / 2 , r ) \Delta_W<\min(c_*/2,r) Δ W < min ( c ∗ /2 , r ) , with r r r from that criterion. The real-operator estimate is uniform; material averaging uses the common lower energy c ∗ / 2 c_*/2 c ∗ /2 and resampling uses the common upper energy c ∗ c^* c ∗ . This proves (159) . The proof of Theorem 10.1 uses only this bound.
For (160) , the one-site recursion of Theorem 4.4 has the common contraction factor 2 ϑ T ϑ ∗ ( μ ) / ( ( 1 − ϑ ) ( W E 0 ) ϑ ) < 1 2^{\vartheta}T_{\vartheta}^{*}(\mu)/((1-\vartheta)(WE_0)^{\vartheta})<1 2 ϑ T ϑ ∗ ( μ ) / (( 1 − ϑ ) ( W E 0 ) ϑ ) < 1 . The same averaging argument applies. For the quantitative window, repeat § 3.5 with the common deficit comparison and decay constants. Its small-coupling estimates are uniform for g 0 = W c M , δ / ( 1 + a W ) g_0=Wc_{M,\delta}/(1+aW) g 0 = W c M , δ / ( 1 + aW ) , since c M , δ c_{M,\delta} c M , δ stays in [ c ∗ , c ∗ ] [c_*,c^*] [ c ∗ , c ∗ ] . This proves the final assertion. □ \square □
Eigenvalue spacing under material disorder
Matching individual modes in § 5.6 requires a quantitative bound on the probability of small eigenvalue gaps. The following consequence of Lemma 6.2 applies to any deterministic nonnegative finite matrix. In particular, it does not require exponential decay of the entries of the finite capacitance matrix.
Proposition 11.1 (A material-variable Minami estimate). Let C ∗ C_* C ∗ be a deterministic nonnegative N × N N\times N N × N matrix and N J = Tr 1 J ( M − 1 / 2 C ∗ M − 1 / 2 ) \mathcal{N}_J=\operatorname{Tr}\mathbf{1}_J(M^{-1/2}C_*M^{-1/2}) N J = Tr 1 J ( M − 1/2 C ∗ M − 1/2 ) , with the law (8) and W > 0 W>0 W > 0 . For an interval J ⊆ [ E 0 , ∞ ) J\subseteq[E_0,\infty) J ⊆ [ E 0 , ∞ ) , E 0 > 0 E_0>0 E 0 > 0 ,
E [ N J ] ≤ κ N ∣ J ∣ , E [ N J ( N J − 1 ) ] ≤ κ 2 N 2 ∣ J ∣ 2 , (161) \mathbb{E}[\mathcal{N}_J]\leq\kappa N|J|,\qquad\mathbb{E}[\mathcal{N}_J(\mathcal{N}_J-1)]\leq\kappa^2N^2|J|^2, \tag*{(161)} E [ N J ] ≤ κ N ∣ J ∣ , E [ N J ( N J − 1 )] ≤ κ 2 N 2 ∣ J ∣ 2 , (161)
where κ = ( 1 + W ) / ( W E 0 ) \kappa=(1+W)/(WE_0) κ = ( 1 + W ) / ( W E 0 ) . On a fixed bounded positive-energy region, the probability that two eigenvalues have separation less than h h h is at most C W N 2 h C_WN^2h C W N 2 h for small h > 0 h>0 h > 0 . Positive eigenvalues are almost surely simple.
Proof. The first estimate follows by summing (101) over coordinates. For the second, independently resample m y m_y m y , and write N J ( y ) \mathcal{N}_J^{(y)} N J ( y ) for the reference count. The finite matrices H H H and B = C ∗ 1 / 2 M − 1 C ∗ 1 / 2 B=C_*^{1/2}M^{-1}C_*^{1/2} B = C ∗ 1/2 M − 1 C ∗ 1/2 have identical spectra, including multiplicities. Changing one material variable changes B B B by a sign-definite rank-one matrix. Rank-one interlacing consequently gives, for any original and reference values,
N J − 1 ≤ N J ( y ) . (162) \mathcal{N}_J-1\leq\mathcal{N}_J^{(y)}. \tag*{(162)} N J − 1 ≤ N J ( y ) . (162)
For example, the eigenvalue counting functions of two sign-definite rank-one perturbations differ at each energy by either zero or one with a common sign; taking their difference at the two ends of J J J proves (162) .
Since N J = ∑ y ν y ( J ) \mathcal{N}_J=\sum_y\nu_y(J) N J = ∑ y ν y ( J ) , one has
N J ( N J − 1 ) ≤ ∑ y ν y ( J ) N J ( y ) . \mathcal{N}_J(\mathcal{N}_J-1)\leq\sum_y\nu_y(J)\mathcal{N}_J^{(y)}. N J ( N J − 1 ) ≤ y ∑ ν y ( J ) N J ( y ) .
Condition on all other variables and on the resampled one. The reference count is independent of the original m y m_y m y , so (101) bounds its contribution by κ ∣ J ∣ N J ( y ) \kappa|J|\mathcal{N}_J^{(y)} κ ∣ J ∣ N J ( y ) . Averaging again and applying the first estimate to the reference operator gives κ 2 N 2 ∣ J ∣ 2 \kappa^2N^2|J|^2 κ 2 N 2 ∣ J ∣ 2 . This proves (161) . This is the material-disorder analogue of the quadratic counting estimate in [22 , 14 ] .
Cover the fixed energy region by O ( h − 1 ) O(h^{-1}) O ( h − 1 ) overlapping intervals of length 3 h 3h 3 h , so that every pair of energies at distance less than h h h lies in one interval. Since P { N J ≥ 2 } ≤ 1 2 E [ N J ( N J − 1 ) ] \mathbb{P}\{\mathcal{N}_J\geq2\}\leq\frac{1}{2}\mathbb{E}[\mathcal{N}_J(\mathcal{N}_J-1)] P { N J ≥ 2 } ≤ 2 1 E [ N J ( N J − 1 )] , a union bound gives C W N 2 h C_WN^2h C W N 2 h . Let h ↓ 0 h\downarrow0 h ↓ 0 for each fixed N N N , and apply the argument on the intervals [ 1 / k , k ] [1/k,k] [ 1/ k , k ] , k = 2 , 3 , … k=2,3,\ldots k = 2 , 3 , … , to obtain simplicity at every positive energy. □
Interior estimates for truncated matrices
We extend the interior eigenfunction bounds of § 3.14 to truncated matrices by controlling the residual created at the boundary and bounding the number of exceptional modes.
Boundary residuals and eigenfunction extent
Use the projection P Λ P_\Lambda P Λ and the truncated matrix C Λ t C_\Lambda^t C Λ t from (10) . The zero extension of C Λ t C_\Lambda^t C Λ t is P Λ C P Λ P_\Lambda CP_\Lambda P Λ C P Λ . Extend a normalised eigenvector amplitude q q q by zero to q ~ ∈ ℓ 2 ( Z d ) \widetilde q\in\ell^2(\mathbb{Z}^d) q ∈ ℓ 2 ( Z d ) . If C Λ t q = E M ϖ , Λ q C_\Lambda^tq=EM_{\varpi,\Lambda}q C Λ t q = E M ϖ , Λ q and e = μ W E e=\mu_WE e = μ W E , then
( C − e ) q ~ = E W Z ϖ q ~ + b Λ , b Λ = ( 1 − P Λ ) C q ~ . (163) (C-e)\widetilde q=EWZ_{\varpi}\widetilde q+b_\Lambda,\qquad b_\Lambda=(1-P_\Lambda)C\widetilde q. \tag*{(163)} ( C − e ) q = E W Z ϖ q + b Λ , b Λ = ( 1 − P Λ ) C q . (163)
The two terms on the right have disjoint supports. Hence
∥ ( C − e ) q ~ ∥ 2 2 ≤ E 2 W 2 / 4 + ∥ b Λ ∥ 2 2 . (164) \|(C-e)\widetilde q\|_2^2\leq E^2W^2/4+\|b_\Lambda\|_2^2. \tag*{(164)} ∥ ( C − e ) q ∥ 2 2 ≤ E 2 W 2 /4 + ∥ b Λ ∥ 2 2 . (164)
Proposition 8.2 applies with this full-lattice residual.
Lemma 12.1 (Exponential boundary leakage estimate). Let d Λ ( n ) = dist ℓ ∞ ( n , Z d ∖ Λ ) d_\Lambda(n)=\operatorname{dist}_{\ell^\infty}(n,\mathbb{Z}^d\setminus\Lambda) d Λ ( n ) = dist ℓ ∞ ( n , Z d ∖ Λ ) for n ∈ Λ n\in\Lambda n ∈ Λ . For every a 0 ∈ ( 0 , γ 0 ) a_0\in(0,\gamma_0) a 0 ∈ ( 0 , γ 0 ) there is a constant K a 0 K_{a_0} K a 0 , independent of Λ \Lambda Λ , such that
∥ ( 1 − P Λ ) C q ~ ∥ 2 ≤ K a 0 ∥ e − a 0 d Λ q ∥ 2 ≤ K a 0 ( ∥ 1 { d Λ < ℓ } q ∥ 2 + e − a 0 ℓ ∥ q ∥ 2 ) . (165) \|(1-P_\Lambda)C\widetilde q\|_2\leq K_{a_0}\|e^{-a_0d_\Lambda}q\|_2\leq K_{a_0}\left(\|1_{\{d_\Lambda<\ell\}}q\|_2+e^{-a_0\ell}\|q\|_2\right). \tag*{(165)} ∥ ( 1 − P Λ ) C q ∥ 2 ≤ K a 0 ∥ e − a 0 d Λ q ∥ 2 ≤ K a 0 ( ∥ 1 { d Λ < ℓ } q ∥ 2 + e − a 0 ℓ ∥ q ∥ 2 ) . (165) Proof. For x ∉ Λ x\notin\Lambda x ∈ / Λ and y ∈ Λ y\in\Lambda y ∈ Λ , d Λ ( y ) ≤ ∣ x − y ∣ 1 d_\Lambda(y)\leq|x-y|_1 d Λ ( y ) ≤ ∣ x − y ∣ 1 . The matrix with entries c x − y e a 0 d Λ ( y ) c_{x-y}e^{a_0d_\Lambda(y)} c x − y e a 0 d Λ ( y ) is therefore bounded in absolute value by A 0 e − ( γ 0 − a 0 ) ∣ x − y ∣ A_0e^{-(\gamma_0-a_0)|x-y|} A 0 e − ( γ 0 − a 0 ) ∣ x − y ∣ . Its row and column sums are bounded by A 0 ∑ r ∈ Z d e − ( γ 0 − a 0 ) ∣ r ∣ 1 A_0\sum_{r\in\mathbb{Z}^d}e^{-(\gamma_0-a_0)|r|_1} A 0 ∑ r ∈ Z d e − ( γ 0 − a 0 ) ∣ r ∣ 1 . The Schur test proves the first inequality after factoring out e − a 0 d Λ ( y ) q ( y ) e^{-a_0d_\Lambda(y)}q(y) e − a 0 d Λ ( y ) q ( y ) . Splitting at d Λ = ℓ d_\Lambda=\ell d Λ = ℓ proves the second. □
Corollary 12.2 (Interior finite-sample modes). Let I ⋐ ( 0 , c M ) I \Subset(0,c_M) I ⋐ ( 0 , c M ) be a compact interval such that (142) holds. Suppose that μ W E ∈ I \mu_W E \in I μ W E ∈ I , ∥ q ∥ 2 = 1 \lVert q\rVert_2=1 ∥ q ∥ 2 = 1 , and an eigenvector of the truncated matrix satisfies
∥ 1 { d Λ < ℓ } q ∥ 2 ≤ C 0 W , e − a 0 ℓ ≤ W . (166) \lVert\mathbf{1}_{\{d_\Lambda<\ell\}}q\rVert_2\le C_0W,\qquad e^{-a_0\ell}\le W. \tag*{(166)} ∥ 1 { d Λ < ℓ } q ∥ 2 ≤ C 0 W , e − a 0 ℓ ≤ W . (166) For sufficiently small W W W , the bounds (149) hold uniformly in Λ \Lambda Λ , with constants also depending on C 0 C_0 C 0 .
Proof. Equations (164) and (165) bound the full-lattice residual by C W CW C W . Apply Proposition 8.2. □ \square □
This hypothesis need not hold for every finite eigenvector. It excludes substantial boundary concentration, which should not be mistaken for bulk Anderson localisation. Nevertheless, there is a uniform count of the exceptions.
Theorem 12.3 (A bound on the number of boundary-affected modes). Let I ⋐ ( 0 , c M ) I \Subset(0,c_M) I ⋐ ( 0 , c M ) be a compact interval such that (142) holds. Fix a 0 ∈ ( 0 , γ 0 ) a_0\in(0,\gamma_0) a 0 ∈ ( 0 , γ 0 ) and 0 < W ≤ W 0 < 1 0<W\le W_0<1 0 < W ≤ W 0 < 1 , and set ℓ W = ⌈ a 0 − 1 log ( 1 / W ) ⌉ + 1 \ell_W=\lceil a_0^{-1}\log(1/W)\rceil+1 ℓ W = ⌈ a 0 − 1 log ( 1/ W )⌉ + 1 . Write S Λ , W = { n ∈ Λ : d Λ ( n ) < ℓ W } S_{\Lambda,W}=\{n\in\Lambda:d_\Lambda(n)<\ell_W\} S Λ , W = { n ∈ Λ : d Λ ( n ) < ℓ W } . Let ( ψ j ) j = 1 ∣ Λ ∣ (\psi_j)_{j=1}^{|\Lambda|} ( ψ j ) j = 1 ∣Λ∣ be any orthonormal eigenbasis of H ϖ , Λ t = M ϖ , Λ − 1 / 2 C ϖ , Λ t M ϖ , Λ − 1 / 2 H^t_{\varpi,\Lambda}=M_{\varpi,\Lambda}^{-1/2}C^t_{\varpi,\Lambda}M_{\varpi,\Lambda}^{-1/2} H ϖ , Λ t = M ϖ , Λ − 1/2 C ϖ , Λ t M ϖ , Λ − 1/2 . At most
∣ S Λ , W ∣ W 2 (167) \frac{|S_{\Lambda,W}|}{W^2} \tag*{(167)} W 2 ∣ S Λ , W ∣ (167) of these eigenvectors can have squared mass in S Λ , W S_{\Lambda,W} S Λ , W greater than W 2 W^2 W 2 . Every remaining eigenvector whose energy satisfies μ W E j ∈ I \mu_W E_j\in I μ W E j ∈ I obeys the pointwise, participation, and radius bounds (149) , with uniform constants after normalisation in the chosen variables.
For cubes Λ L = { − L , … , L } d \Lambda_L=\{-L,\ldots,L\}^d Λ L = { − L , … , L } d with L ≥ 2 ℓ W L\ge2\ell_W L ≥ 2 ℓ W , the fraction of possible exceptions among the entire eigenbasis is at most
C log ( 1 / W ) + 1 L W 2 . (168) C\frac{\log(1/W)+1}{LW^2}. \tag*{(168)} C L W 2 log ( 1/ W ) + 1 . (168) Thus for each fixed sufficiently small W > 0 W>0 W > 0 this fraction tends to zero as the sample grows. The assertion is valid for every material configuration in the bounded range, not only in expectation.
Proof. Completeness of the orthonormal basis gives
∑ j = 1 ∣ Λ ∣ ∥ 1 S Λ , W ψ j ∥ 2 2 = Tr 1 S Λ , W = ∣ S Λ , W ∣ . \sum_{j=1}^{|\Lambda|}\lVert\mathbf{1}_{S_{\Lambda,W}}\psi_j\rVert_2^2=\operatorname{Tr}\mathbf{1}_{S_{\Lambda,W}}=|S_{\Lambda,W}|. j = 1 ∑ ∣Λ∣ ∥ 1 S Λ , W ψ j ∥ 2 2 = Tr 1 S Λ , W = ∣ S Λ , W ∣.
Each exceptional vector contributes more than W 2 W^2 W 2 , proving (167) . For a nonexceptional vector, put q j = M − 1 / 2 ψ j / ∥ M − 1 / 2 ψ j ∥ 2 q_j=M^{-1/2}\psi_j/\lVert M^{-1/2}\psi_j\rVert_2 q j = M − 1/2 ψ j / ∥ M − 1/2 ψ j ∥ 2 . Then
∥ 1 S Λ , W q j ∥ 2 2 ≤ ( 1 + W ) ∥ 1 S Λ , W ψ j ∥ 2 2 ≤ ( 1 + W 0 ) W 2 . \lVert\mathbf{1}_{S_{\Lambda,W}}q_j\rVert_2^2\le(1+W)\lVert\mathbf{1}_{S_{\Lambda,W}}\psi_j\rVert_2^2\le(1+W_0)W^2. ∥ 1 S Λ , W q j ∥ 2 2 ≤ ( 1 + W ) ∥ 1 S Λ , W ψ j ∥ 2 2 ≤ ( 1 + W 0 ) W 2 .
Also e − a 0 ℓ W ≤ W e^{-a_0\ell_W}\le W e − a 0 ℓ W ≤ W . Corollary 12.2 applies, and bounded multiplication by M 1 / 2 M^{1/2} M 1/2 transfers the estimates back to ψ j \psi_j ψ j . For cubes, elementary layer counting gives ∣ S Λ L , W ∣ ≤ C d ℓ W ( 2 L + 1 ) d − 1 |S_{\Lambda_L,W}|\le C_d\ell_W(2L+1)^{d-1} ∣ S Λ L , W ∣ ≤ C d ℓ W ( 2 L + 1 ) d − 1 , proving (168) . □ \square □
The denominator in (168) counts all modes. To conclude that a vanishing fraction of the modes with μ W E j ∈ I \mu_W E_j\in I μ W E j ∈ I are exceptions, one also needs their count to be a positive fraction of ∣ Λ L ∣ |\Lambda_L| ∣ Λ L ∣ ; for random growing samples this can be provided by an appropriate density-of-states limit with positive mass in I / μ W I/\mu_W I / μ W . No eigenvalue simplicity assumption is needed for Theorem 12.3; the count holds for any fixed orthonormal eigenbasis.
For finite capacitance matrices, the residual acquires ( C Λ t − C Λ f ) q (C_{\Lambda}^{t}-C_{\Lambda}^{f})q ( C Λ t − C Λ f ) q . Proposition 13.1 bounds this additional term by an exponentially weighted boundary norm; see Corollary 15.4. Frequency-dependent radiating reductions [10 ] require a different argument, since their modes need not form an orthonormal basis of a self-adjoint matrix.
Physical scales and the interior spectral question
Set R 2 , p h y s = ε R 2 R_{2,\mathrm{phys}}=\varepsilon R_2 R 2 , phys = ε R 2 and allow the disorder to depend on the cell size, W = W ε W=W_\varepsilon W = W ε . The individual bound becomes
R 2 , p h y s ≥ c I ε W ε − C I ′ ε . (169) R_{2,\mathrm{phys}} \ge c_I\frac{\varepsilon}{W_\varepsilon}-C'_I\varepsilon. \tag*{(169)} R 2 , phys ≥ c I W ε ε − C I ′ ε . (169)
For fixed W > 0 W>0 W > 0 , this lower bound tends to zero as the cells shrink. At W ε ≃ ε W_\varepsilon\simeq\sqrt{\varepsilon} W ε ≃ ε it is of order ε \sqrt{\varepsilon} ε , and the exceptional fraction in (168) need not be small. Thus these estimates do not establish macroscopic delocalisation in either scaling.
For every fixed sufficiently small W W W , the boundary-exception fraction does vanish in growing cubes. Since the density-of-states measure has full support by Theorem 14.3 , every fixed regular interval I I I with nonempty interior gives a positive density of modes with E j ∈ I / μ W E_j\in I/\mu_W E j ∈ I / μ W . The exceptional fraction also vanishes relative to those modes, by testing the density-of-states limit with a nonnegative continuous function supported inside I / μ W I/\mu_W I / μ W . Exponential localisation on a large lattice length remains compatible with all these lower bounds.
Part III. Finite resonator clusters
Finite capacitance matrices
We now define the capacitance matrix of a finite cluster and quantify the boundary correction caused by removing the exterior resonators. We retain the same material realisation in the two finite constructions. The static model is treated first; all boundary constants below are uniform for the family in (3.23).
Figure 4. Truncated and finite capacitance matrices. The dashed line encloses the retained indices, represented here by a square sample. The variational constraints give 0 ≤ C Λ L f ≤ C Λ L t 0\le C_{\Lambda L}^{f}\le C_{\Lambda L}^{t} 0 ≤ C Λ L f ≤ C Λ L t . Both sequences recover the infinite operator as L → ∞ L\to\infty L → ∞ .
Variational definition
The truncated matrices C Λ t C_{\Lambda}^{t} C Λ t and H ϖ , Λ t H_{\varpi,\Lambda}^{t} H ϖ , Λ t were defined in (10) . To define the finite capacitance matrix, use the energy spaces
E d = { { u ∈ H l o c 1 ( R 2 ) : ∇ u ∈ L 2 ( R 2 ) } , d = 2 , { u ∈ H l o c 1 ( R 3 ) ∩ L 6 ( R 3 ) : ∇ u ∈ L 2 ( R 3 ) } , d = 3. \mathcal{E}_d=\begin{cases}
\{u\in H_{\mathrm{loc}}^1(\mathbb{R}^2):\nabla u\in L^2(\mathbb{R}^2)\}, & d=2,\\
\{u\in H_{\mathrm{loc}}^1(\mathbb{R}^3)\cap L^6(\mathbb{R}^3):\nabla u\in L^2(\mathbb{R}^3)\}, & d=3.
\end{cases} E d = { { u ∈ H loc 1 ( R 2 ) : ∇ u ∈ L 2 ( R 2 )} , { u ∈ H loc 1 ( R 3 ) ∩ L 6 ( R 3 ) : ∇ u ∈ L 2 ( R 3 )} , d = 2 , d = 3.
For a nonempty finite set Λ \Lambda Λ define
q ∗ C Λ f q = 1 ∣ D ∣ min U ∈ E d , U = q n in D + n ( n ∈ Λ ) ∫ R d ∣ ∇ U ∣ 2 , H ϖ , Λ f = B ϖ , Λ 1 / 2 C Λ f B ϖ , Λ 1 / 2 . (170) q^{*}C^{f}_{\Lambda}q=\frac{1}{|D|}\min_{\substack{U\in\mathcal{E}_{d},\ U=q_{n}\ \text{in}\ D+n\\(n\in\Lambda)}}\int_{\mathbb{R}^{d}}|\nabla U|^{2},\qquad H^{f}_{\varpi,\Lambda}=B_{\varpi,\Lambda}^{1/2}C^{f}_{\Lambda}B_{\varpi,\Lambda}^{1/2}. \tag*{(170)} q ∗ C Λ f q = ∣ D ∣ 1 U ∈ E d , U = q n in D + n ( n ∈ Λ ) min ∫ R d ∣∇ U ∣ 2 , H ϖ , Λ f = B ϖ , Λ 1/2 C Λ f B ϖ , Λ 1/2 . (170)
This is the static limit of the variational construction in [9 Section 3.1] . In dimension three, it is the usual free-space capacitance matrix divided by ∣ D ∣ |D| ∣ D ∣ . In dimension two, constants are admissible and C Λ f 1 = 0 C^{f}_{\Lambda}\mathbf{1}=0 C Λ f 1 = 0 , where 1 \mathbf{1} 1 denotes the constant vector with entries one. The direct method gives the minimiser. In dimension three use the Sobolev inequality in the homogeneous energy space; in dimension two an imposed value on one resonator fixes the additive constant, giving local H 1 H^{1} H 1 compactness. The constraints are closed and the energy is strictly convex on functions with zero constrained traces.
For C Λ t C^{t}_{\Lambda} C Λ t , every omitted resonator remains present with prescribed potential zero. In (170) it is absent and no constraint is imposed in its former location. Hence
0 ≤ C Λ f ≤ C Λ t ≤ c M Id , 0 ≤ H ϖ , Λ f ≤ H ϖ , Λ t ≤ c M Id . (171) 0\leq C^{f}_{\Lambda}\leq C^{t}_{\Lambda}\leq c_{M}\operatorname{Id},\qquad0\leq H^{f}_{\varpi,\Lambda}\leq H^{t}_{\varpi,\Lambda}\leq c_{M}\operatorname{Id}. \tag*{(171)} 0 ≤ C Λ f ≤ C Λ t ≤ c M Id , 0 ≤ H ϖ , Λ f ≤ H ϖ , Λ t ≤ c M Id . (171)
for the main law, since 0 < B ϖ ≤ Id 0<B_{\varpi}\leq\operatorname{Id} 0 < B ϖ ≤ Id . The first comparison follows by inserting the globally constrained minimiser as a competitor for the finite-cluster problem.
For δ > 0 \delta>0 δ > 0 the corresponding contrast-dependent matrix is
q ∗ C δ , Λ f q = 1 ∣ D ∣ min U ∈ E d f D + n U = q n ( n ∈ Λ ) ∫ R d a δ , Λ ∣ ∇ U ∣ 2 , a δ , Λ = δ − 1 1 ∪ n ∈ Λ ( D + n ) + 1 R d ∖ ∪ n ∈ Λ ( D + n ) . (172) q^{*}C^{f}_{\delta,\Lambda}q=\frac{1}{|D|}\min_{\substack{U\in\mathcal{E}_{d}\\f_{D+n}U=q_{n}\ (n\in\Lambda)}}\int_{\mathbb{R}^{d}}a_{\delta,\Lambda}|\nabla U|^{2},\qquad a_{\delta,\Lambda}=\delta^{-1}\mathbf{1}_{\cup_{n\in\Lambda}(D+n)}+\mathbf{1}_{\mathbb{R}^{d}\setminus\cup_{n\in\Lambda}(D+n)}. \tag*{(172)} q ∗ C δ , Λ f q = ∣ D ∣ 1 U ∈ E d f D + n U = q n ( n ∈ Λ ) min ∫ R d a δ , Λ ∣∇ U ∣ 2 , a δ , Λ = δ − 1 1 ∪ n ∈ Λ ( D + n ) + 1 R d ∖ ∪ n ∈ Λ ( D + n ) . (172)
Set C 0 , Λ f = C Λ f C^{f}_{0,\Lambda}=C^{f}_{\Lambda} C 0 , Λ f = C Λ f and, for a ∈ { t , f } a\in\{t,f\} a ∈ { t , f } , H δ , ϖ , Λ a = M ϖ , Λ − 1 / 2 C δ , Λ a M ϖ , Λ − 1 / 2 H^{a}_{\delta,\varpi,\Lambda}=M_{\varpi,\Lambda}^{-1/2}C^{a}_{\delta,\Lambda}M_{\varpi,\Lambda}^{-1/2} H δ , ϖ , Λ a = M ϖ , Λ − 1/2 C δ , Λ a M ϖ , Λ − 1/2 . The same variational comparison holds for every sufficiently small δ \delta δ , with a common upper bound. In dimension two the constant vector lies in the kernel of C δ , Λ f C^{f}_{\delta,\Lambda} C δ , Λ f ; all localisation intervals considered below are separated from zero.
The boundary comparison
With the depth d Λ d_{\Lambda} d Λ from Lemma 12.1, set
∂ r Λ = { n ∈ Λ : d Λ ( n ) ≤ r } . (173) \partial_{r}\Lambda=\{n\in\Lambda:d_{\Lambda}(n)\leq r\}. \tag*{(173)} ∂ r Λ = { n ∈ Λ : d Λ ( n ) ≤ r } . (173)
The variational framework of [9 ] gives an exponentially weighted boundary comparison.
Proposition 13.1 (Boundary correction). There are K , γ > 0 K,\gamma>0 K , γ > 0 , independent of finite Λ \Lambda Λ and 0 ≤ δ ≤ δ 0 0\leq\delta\leq\delta_{0} 0 ≤ δ ≤ δ 0 , such that
0 ≤ R δ , Λ : = C δ , Λ t − C δ , Λ f ≤ K diag ( e − γ d Λ ( n ) ) as quadratic forms. (174) 0\leq R_{\delta,\Lambda}:=C^{t}_{\delta,\Lambda}-C^{f}_{\delta,\Lambda}\leq K\operatorname{diag}\left(e^{-\gamma d_{\Lambda}(n)}\right)\quad\text{as quadratic forms.} \tag*{(174)} 0 ≤ R δ , Λ := C δ , Λ t − C δ , Λ f ≤ K diag ( e − γ d Λ ( n ) ) as quadratic forms. (174) Also 0 ≤ C δ , Λ f ≤ C δ , Λ t ≤ K Id 0\leq C^{f}_{\delta,\Lambda}\leq C^{t}_{\delta,\Lambda}\leq K\operatorname{Id} 0 ≤ C δ , Λ f ≤ C δ , Λ t ≤ K Id .
Proof. This is Lemma 5.1 of [9 ] . Its proof, based on Lemma 4.4 there, is uniform in the cell contrast and applies to arbitrary finite cell sets. We recall the argument, writing R = R δ , Λ R=R_{\delta,\Lambda} R = R δ , Λ .
The full-lattice minimiser with zero-extended means is admissible for the finite problem, and the finite coefficient is no larger than the full periodic coefficient. This proves positivity of R R R . Local cutoff lifts equal to one near D n D_{n} D n , with disjoint cell supports and gradient zero in the resonators, give a common upper bound K Id K\operatorname{Id} K Id .
Suppose the prescribed means q q q are supported at depth greater than m m m , and let U U U minimise the finite energy. The prescribed means vanish in the first m m m layers of retained cells. No mean constraint is imposed outside the retained array. Choose layer cutoffs that are constant in each resonator; their gradients are then supported where the coefficient is one. Testing the Euler equation with χ 2 U \chi^{2}U χ 2 U and using the cell Poincaré inequality for zero-mean functions gives
E ( r ) ≤ C k 2 E ( r + k ) , r + k ≤ m , \mathcal{E}(r)\leq\frac{C}{k^{2}}\mathcal{E}(r+k),\qquad r+k\leq m, E ( r ) ≤ k 2 C E ( r + k ) , r + k ≤ m ,
Here E ( r ) \mathcal{E}(r) E ( r ) is the weighted energy in the exterior together with the first r r r layers, and E ( ∞ ) \mathcal{E}(\infty) E ( ∞ ) is the total weighted energy of U U U . The cell Poincaré constant and cutoff bounds depend on D ⋐ Y D \Subset Y D ⋐ Y , not on δ \delta δ or the sample. Choosing a fixed sufficiently large k k k and iterating yields
E ( r ) ≤ C ρ m − r E ( ∞ ) , 0 < ρ < 1. (175) \mathcal{E}(r) \le C\rho^{m-r}\mathcal{E}(\infty), \qquad0<\rho<1. \tag*{(175)} E ( r ) ≤ C ρ m − r E ( ∞ ) , 0 < ρ < 1. (175)
Cut off U U U between depths m / 3 m/3 m /3 and 2 m / 3 2m/3 2 m /3 to make a full-lattice competitor with means E Λ q E_{\Lambda}q E Λ q . The removed part has zero means on all retained resonators, hence is energy-orthogonal to U U U . Equation (175) then gives
⟨ R δ , Λ q , q ⟩ ≤ C ρ m / 3 ∥ q ∥ 2 if supp q ⊂ { d Λ > m } . (176) \langle R_{\delta,\Lambda}q,q\rangle\le C\rho^{m/3}\lVert q\rVert^{2}\quad\text{if supp }q\subset\{d_{\Lambda}>m\}. \tag*{(176)} ⟨ R δ , Λ q , q ⟩ ≤ C ρ m /3 ∥ q ∥ 2 if supp q ⊂ { d Λ > m } . (176)
For arbitrary q q q , write q = ∑ m ≥ 0 Q m q q=\sum_{m\ge0}Q_mq q = ∑ m ≥ 0 Q m q , where Q m Q_m Q m projects onto the sites with m < d Λ ( n ) ≤ m + 1 m<d_{\Lambda}(n)\le m+1 m < d Λ ( n ) ≤ m + 1 . Positivity allows use of R 1 / 2 R^{1/2} R 1/2 . The triangle inequality followed by weighted Cauchy–Schwarz and (176) gives
∥ R 1 / 2 q ∥ 2 ≤ C ( ∑ m ≥ 0 ρ m / 6 ∥ Q m q ∥ ) 2 ≤ C ∑ m ≥ 0 ρ m / 6 ∥ Q m q ∥ 2 . \lVert R^{1/2}q\rVert^{2}\le C\left(\sum_{m\ge0}\rho^{m/6}\lVert Q_mq\rVert\right)^{2}\le C\sum_{m\ge0}\rho^{m/6}\lVert Q_mq\rVert^{2}. ∥ R 1/2 q ∥ 2 ≤ C ( m ≥ 0 ∑ ρ m /6 ∥ Q m q ∥ ) 2 ≤ C m ≥ 0 ∑ ρ m /6 ∥ Q m q ∥ 2 .
This is (174) after changing the constants. The construction of the cutoffs uses the fixed positive gap between D D D and the cell faces; harmless fixed shifts in layer indices do not change the estimate. Passing to the static variational limit gives the same statement at δ = 0 \delta=0 δ = 0 . □ \square □
The bound is a positive quadratic-form estimate; it does not assert exponential decay of every entry of C f C^{f} C f in the separation between its indices. In particular, interactions between sites close to the exterior boundary are not controlled by their mutual distance in (174) .
Corollary 13.2 (Random boundary correction and surface rates). Write S = M ϖ , Λ − 1 / 2 S=M_{\varpi,\Lambda}^{-1/2} S = M ϖ , Λ − 1/2 and suppress δ , ϖ , Λ \delta,\varpi,\Lambda δ , ϖ , Λ within this corollary. For (8) , put T = H t − H f = S R S T=H^{t}-H^{f}=SRS T = H t − H f = S R S . Then, pathwise,
0 ≤ T ≤ K diag ( e − γ d Λ ( n ) ) . (177) 0\le T\le K\operatorname{diag}(e^{-\gamma d_{\Lambda}(n)}). \tag*{(177)} 0 ≤ T ≤ K diag ( e − γ d Λ ( n ) ) . (177) If P r P_r P r projects onto ℓ 2 ( ∂ r Λ ) \ell^{2}(\partial_r\Lambda) ℓ 2 ( ∂ r Λ ) and Q r = Id − P r Q_r=\operatorname{Id}-P_r Q r = Id − P r , then
∥ T Q r ∥ ≤ K 1 e − γ r / 2 , ∥ T − P r T P r ∥ ≤ 2 K 1 e − γ r / 2 , rank ( P r T P r ) ≤ ∣ ∂ r Λ ∣ . (178) \lVert TQ_r\rVert\le K_1e^{-\gamma r/2},\qquad\lVert T-P_rTP_r\rVert\le2K_1e^{-\gamma r/2},\qquad\operatorname{rank}(P_rTP_r)\le\lvert\partial_r\Lambda\rvert. \tag*{(178)} ∥ T Q r ∥ ≤ K 1 e − γ r /2 , ∥ T − P r T P r ∥ ≤ 2 K 1 e − γ r /2 , rank ( P r T P r ) ≤ ∣ ∂ r Λ ∣ . (178) Let E 1 , L a ≤ ⋯ ≤ E N L , L a E^{a}_{1,L}\le\cdots\le E^{a}_{N_L,L} E 1 , L a ≤ ⋯ ≤ E N L , L a be the eigenvalues of H δ , ϖ , Λ L a H^{a}_{\delta,\varpi,\Lambda_L} H δ , ϖ , Λ L a , counted with multiplicity, for a ∈ { t , f } a\in\{t,f\} a ∈ { t , f } . Uniformly in δ \delta δ and the realisation,
1 N L ∑ j = 1 N L ( E j , L t − E j , L f ) = 1 N L Tr T = O ( L − 1 ) , (179) \frac{1}{N_L}\sum_{j=1}^{N_L}(E^{t}_{j,L}-E^{f}_{j,L})=\frac{1}{N_L}\operatorname{Tr}T=O(L^{-1}), \tag*{(179)} N L 1 j = 1 ∑ N L ( E j , L t − E j , L f ) = N L 1 Tr T = O ( L − 1 ) , (179) 1 N L ∑ j = 1 N L ∣ E j , L t − E j , L f ∣ 2 = O ( L − 1 ) . (180) \frac{1}{N_L}\sum_{j=1}^{N_L}\lvert E^{t}_{j,L}-E^{f}_{j,L}\rvert^{2}=O(L^{-1}). \tag*{(180)} N L 1 j = 1 ∑ N L ∣ E j , L t − E j , L f ∣ 2 = O ( L − 1 ) . (180) For complete cells in a bounded Lipschitz domain at scale ε \varepsilon ε , the corresponding right sides are O ( ε ) O(\varepsilon) O ( ε ) .
Proof. Diagonal multiplication commutes with the boundary weight, and ∥ S ∥ ≤ 1 \lVert S\rVert\le1 ∥ S ∥ ≤ 1 , so (177) follows by congruence. For any vector u u u ,
∥ T Q r u ∥ 2 ≤ ∥ T ∥ ⟨ Q r u , T Q r u ⟩ ≤ K 2 e − γ r ∥ u ∥ 2 . \lVert TQ_ru\rVert^{2}\le\lVert T\rVert\langle Q_ru,TQ_ru\rangle\le K^{2}e^{-\gamma r}\lVert u\rVert^{2}. ∥ T Q r u ∥ 2 ≤ ∥ T ∥ ⟨ Q r u , T Q r u ⟩ ≤ K 2 e − γ r ∥ u ∥ 2 . Moreover T − P r T P r = T Q r + Q r T P r T-P_rTP_r=TQ_r+Q_rTP_r T − P r T P r = T Q r + Q r T P r , proving (178) . Ordered eigenvalues satisfy E j f ≤ E j t E^{f}_{j}\le E^{t}_{j} E j f ≤ E j t , and their sum difference is Tr T \operatorname{Tr}T Tr T . The weighted layer sum is O ( L − 1 ) O(L^{-1}) O ( L − 1 ) for cubes, proving (179) . Finally ∥ T ∥ H S 2 = Tr T 2 ≤ ∥ T ∥ Tr T \lVert T\rVert_{\mathrm{HS}}^{2}=\operatorname{Tr}T^{2}\le\lVert T\rVert\operatorname{Tr}T ∥ T ∥ HS 2 = Tr T 2 ≤ ∥ T ∥ Tr T and (182) prove (180) . For a Lipschitz sample, the cell
boundary-layer count is O ( ε 1 − d r ) O(\varepsilon^{1-d_r}) O ( ε 1 − d r ) for 1 ≤ r ≪ ε − 1 1 \le r \ll\varepsilon^{-1} 1 ≤ r ≪ ε − 1 ; summation with the exponential weights proves the asserted rate. □
In particular, the Wasserstein distance of order one between the two empirical eigenvalue measures is O ( ε ) O(\varepsilon) O ( ε ) . This rate compares two models at the same random realisation. It is not a rate of convergence of either empirical measure to the infinite-volume density of states.
Random asymptotic equivalence and spectral convergence
The deterministic comparison in [7 Lemma 3.1] is boundedness in operator norm together with convergence in the normalised Frobenius norm . It is not convergence of the difference in operator norm. The following extension is purely deterministic and hence holds for every realisation.
Theorem 14.1 (Pathwise weighted asymptotic equivalence). Let C N t , C N f C_N^t,C_N^f C N t , C N f be Hermitian N × N N \times N N × N matrices satisfying
sup N ( ∥ C N t ∥ + ∥ C N f ∥ ) < ∞ , a N : = N − 1 / 2 ∥ C N t − C N f ∥ H S ⟶ 0. \sup_N(\lVert C_N^t\rVert+\lVert C_N^f\rVert)<\infty,\qquad a_N:=N^{-1/2}\lVert C_N^t-C_N^f\rVert_{\mathrm{HS}}\longrightarrow0. N sup (∥ C N t ∥ + ∥ C N f ∥) < ∞ , a N := N − 1/2 ∥ C N t − C N f ∥ HS ⟶ 0. For any diagonal matrices M N M_N M N with 0 < m − ≤ ( M N ) n n ≤ m + < ∞ 0<m_-\le(M_N)_{nn}\le m_+<\infty 0 < m − ≤ ( M N ) nn ≤ m + < ∞ , let H N a = M N − 1 / 2 C N a M N − 1 / 2 H_N^a=M_N^{-1/2}C_N^aM_N^{-1/2} H N a = M N − 1/2 C N a M N − 1/2 . Then
N − 1 / 2 ∥ H N t − H N f ∥ H S ≤ m − − 1 a N . (181) N^{-1/2}\lVert H_N^t-H_N^f\rVert_{\mathrm{HS}}\le m_-^{-1}a_N. \tag*{(181)} N − 1/2 ∥ H N t − H N f ∥ HS ≤ m − − 1 a N . (181) Their ordered eigenvalues E j , N a E_{j,N}^a E j , N a , a ∈ { t , f } a\in\{t,f\} a ∈ { t , f } , satisfy
1 N ∑ j = 1 N ∣ E j , N t − E j , N f ∣ 2 ≤ m − − 2 a N 2 . (182) \frac{1}{N}\sum_{j=1}^{N}\lvert E_{j,N}^t-E_{j,N}^f\rvert^2\le m_-^{-2}a_N^2. \tag*{(182)} N 1 j = 1 ∑ N ∣ E j , N t − E j , N f ∣ 2 ≤ m − − 2 a N 2 . (182) Consequently their empirical eigenvalue measures have the same weak limits. No independence or ergodicity is required for this comparison.
Proof. Put S N = M N − 1 / 2 S_N=M_N^{-1/2} S N = M N − 1/2 . Then H N t − H N f = S N ( C N t − C N f ) S N H_N^t-H_N^f=S_N(C_N^t-C_N^f)S_N H N t − H N f = S N ( C N t − C N f ) S N , and the Hilbert–Schmidt ideal inequality gives (181) . The Hermitian eigenvalue inequality of Hoffman–Wielandt gives (182) . Equivalently, it follows by diagonalising both matrices, using that the squared moduli of the change-of-basis entries form a doubly stochastic matrix, and applying the rearrangement inequality to the ordered eigenvalues. For a Lipschitz function F F F ,
∣ 1 N Tr F ( H N t ) − 1 N Tr F ( H N f ) ∣ ≤ Lip ( F ) ( 1 N ∑ j ∣ E j , N t − E j , N f ∣ 2 ) 1 / 2 . \left|\frac{1}{N}\operatorname{Tr}F(H_N^t)-\frac{1}{N}\operatorname{Tr}F(H_N^f)\right|
\le\operatorname{Lip}(F)\left(\frac{1}{N}\sum_j\lvert E_{j,N}^t-E_{j,N}^f\rvert^2\right)^{1/2}. N 1 Tr F ( H N t ) − N 1 Tr F ( H N f ) ≤ Lip ( F ) ( N 1 j ∑ ∣ E j , N t − E j , N f ∣ 2 ) 1/2 .
The spectra stay in a common compact interval. Uniform approximation of continuous functions by Lipschitz functions completes the proof. □
For (8) , m − = 1 m_-=1 m − = 1 , so (181) is independent of W W W . The two matrices with the same material realisation have the common deterministic density-of-states measure of the random model of Theorem 14.3 ; that measure generally depends on W W W . The direct-factor law of Appendix B gives the same comparison with the constant 1 + W 1+W 1 + W .
Strong and spectral-set convergence
We fix W > 0 W>0 W > 0 and write d H d_H d H for the Hausdorff distance between nonempty compact subsets of R \mathbb{R} R .
Theorem 14.2. For every fixed realisation and a ∈ { t , f } a\in\{t,f\} a ∈ { t , f } ,
E Λ L H ϖ , Λ L a E Λ L ∗ ⟶ H ϖ strongly , Λ L = [ − L , L ] d ∩ Z d . (183) E_{\Lambda_L}H_{\varpi,\Lambda_L}^{a}E_{\Lambda_L}^{*}\longrightarrow H_{\varpi}\quad\text{strongly},\qquad\Lambda_L=[-L,L]^d\cap\mathbb{Z}^d. \tag*{(183)} E Λ L H ϖ , Λ L a E Λ L ∗ ⟶ H ϖ strongly , Λ L = [ − L , L ] d ∩ Z d . (183) Resolvents converge strongly off the real axis and local spectral measures converge weakly. Almost surely,
d H ( σ ( H ϖ , Λ L a ) , [ 0 , c M ] ) ⟶ 0. (184) d_H\bigl(\sigma(H_{\varpi,\Lambda_L}^{a}),[0,c_M]\bigr)\longrightarrow0. \tag*{(184)} d H ( σ ( H ϖ , Λ L a ) , [ 0 , c M ] ) ⟶ 0. (184) The same statements hold at each fixed sufficiently small δ \delta δ , with H δ , ϖ H_{\delta,\varpi} H δ , ϖ and [ 0 , c M , δ ] [0,c_{M,\delta}] [ 0 , c M , δ ] . If δ = δ L → 0 \delta=\delta_L\to0 δ = δ L → 0 , the limits are the static ones.
Proof. The zero-extended truncated matrices are P L H ϖ P L P_LH_{\varpi}P_L P L H ϖ P L , where P L → Id P_L\to\operatorname{Id} P L → Id strongly. They converge strongly since H ϖ H_{\varpi} H ϖ is bounded. By (177) and ∥ T q ∥ 2 ≤ ∥ T ∥ ⟨ T q , q ⟩ \lVert Tq\rVert^2\leq\lVert T\rVert\langle Tq,q\rangle ∥ T q ∥ 2 ≤ ∥ T ∥ ⟨ T q , q ⟩ , the difference of the two extended finite matrices tends to zero on every finitely supported vector: its support has depth tending to infinity. Uniform operator bounds extend this convergence to all of ℓ 2 \ell^2 ℓ 2 . This proves (183) . The resolvent identity and polynomial approximation give the asserted resolvent and local-measure limits.
The almost-sure spectrum is [ 0 , c M ] [0,c_M] [ 0 , c M ] by Proposition 2.2 . Every energy in this interval has finitely supported approximate eigenvectors. Strong convergence applied to each such fixed vector puts finite eigenvalues arbitrarily close to that energy. All finite spectra lie in [ 0 , c M ] [0,c_M] [ 0 , c M ] by (171) . A finite net proves Hausdorff convergence. For fixed δ \delta δ , the same spectrum proof applies: c δ c_\delta c δ is continuous, nonnegative, vanishes at zero, and has maximum c M , δ c_{M,\delta} c M , δ , while favourable material boxes approximate C δ C_\delta C δ . For δ L → 0 \delta_L\to0 δ L → 0 use ∥ C δ L − C ∥ = ∥ c δ L − c ∥ ∞ → 0 \lVert C_{\delta_L}-C\rVert=\lVert c_{\delta_L}-c\rVert_\infty\to0 ∥ C δ L − C ∥ = ∥ c δ L − c ∥ ∞ → 0 and the uniform boundary comparison. The finite spectra are contained in [ 0 , c M , δ L ] [0,c_{M,\delta_L}] [ 0 , c M , δ L ] , whose endpoints converge. No general absence of spectral pollution in gaps is being inferred from strong convergence. □
The common density of states
For fixed W > 0 W>0 W > 0 and δ \delta δ , define the probability measure ν W , δ \nu_{W,\delta} ν W , δ by the identity below for bounded Borel functions F F F :
∫ F d ν W , δ = E ⟨ δ 0 , F ( H δ , ϖ ) δ 0 ⟩ . (185) \int F\,\mathrm{d}\nu_{W,\delta}=\mathbb{E}\langle\delta_0,F(H_{\delta,\varpi})\delta_0\rangle. \tag*{(185)} ∫ F d ν W , δ = E ⟨ δ 0 , F ( H δ , ϖ ) δ 0 ⟩ . (185)
Its cumulative distribution is the integrated density of states. Write δ E D \delta_E^D δ E D for the unit point mass at energy E E E .
Theorem 14.3. Along growing cubes, almost surely,
1 ∣ Λ L ∣ ∑ j = 1 ∣ Λ L ∣ δ E j , L a D ⟶ ν W , δ , a ∈ { t , f } . \frac{1}{|\Lambda_L|}\sum_{j=1}^{|\Lambda_L|}\delta_{E_{j,L}^{a}}^D\longrightarrow\nu_{W,\delta},\qquad a\in\{t,f\}. ∣ Λ L ∣ 1 j = 1 ∑ ∣ Λ L ∣ δ E j , L a D ⟶ ν W , δ , a ∈ { t , f } . For δ L → 0 \delta_L\to0 δ L → 0 both limits are ν W , 0 \nu_{W,0} ν W , 0 on one probability-one event for the prescribed sequence. The support of ν W , δ \nu_{W,\delta} ν W , δ is [ 0 , c M , δ ] [0,c_{M,\delta}] [ 0 , c M , δ ] . It has a locally bounded density on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) and no atom at zero. Consequently the normalised eigenvalue-counting functions converge at every energy.
Proof. For truncated matrices, use stationarity and ergodicity of the infinite family. The ergodic theorem applies to the diagonal entries of each polynomial in H δ , ϖ H_{\delta,\varpi} H δ , ϖ ; see [20 ] . Cut off hopping beyond distance R R R . The norm error is bounded uniformly in the realisation by the summable coefficient tail. For a polynomial of degree k k k in this finite-range operator, finite and infinite diagonal entries agree outside a boundary strip of width k R kR k R . Its relative size vanishes. Let L → ∞ L\to\infty L → ∞ , then R → ∞ R\to\infty R → ∞ , and approximate continuous functions by polynomials on the common compact spectral interval. A countable dense family of polynomials gives one probability-one event for weak convergence. The trace estimate (179) transfers it to finite capacitance matrices.
For moving contrast, sup ϖ , Λ ∥ H δ , ϖ , Λ t − H 0 , ϖ , Λ t ∥ ≤ ∥ c δ − c ∥ ∞ → 0 \sup_{\varpi,\Lambda}\lVert H^{t}_{\delta,\varpi,\Lambda}-H^{t}_{0,\varpi,\Lambda}\rVert\leq\lVert c_{\delta}-c\rVert_{\infty}\to0 sup ϖ , Λ ∥ H δ , ϖ , Λ t − H 0 , ϖ , Λ t ∥ ≤ ∥ c δ − c ∥ ∞ → 0 . Apply the static result and the uniform trace comparison. If an open interval has zero density-of-states mass, positivity and stationarity make its spectral projection vanish at every coordinate almost surely, hence vanish as an operator. This identifies the support with the almost-sure spectrum.
Lemma 6.2 gives ν W , δ ( J ) ≤ ( 1 + W ) ∣ J ∣ / ( W E 0 ) \nu_{W,\delta}(J)\leq(1+W)|J|/(WE_{0}) ν W , δ ( J ) ≤ ( 1 + W ) ∣ J ∣/ ( W E 0 ) for J ⊂ [ E 0 , ∞ ) J\subset[E_{0},\infty) J ⊂ [ E 0 , ∞ ) , E 0 > 0 E_{0}>0 E 0 > 0 . The variational principle shows that c δ c_{\delta} c δ vanishes only at zero: zero energy forces the admissible function to be constant and every Bloch phase to be one. Thus ker C δ = { 0 } \ker C_{\delta}=\{0\} ker C δ = { 0 } as a multiplier on L 2 ( Y ∗ ) L^{2}(Y^{*}) L 2 ( Y ∗ ) . Bounded invertibility of M ϖ 1 / 2 M_{\varpi}^{1/2} M ϖ 1/2 implies ker H δ , ϖ = { 0 } \ker H_{\delta,\varpi}=\{0\} ker H δ , ϖ = { 0 } . There is no atom at zero. The distribution function is continuous, which proves the final assertion. □
The same proof applies to tempered Følner sequences, by the pointwise ergodic theorem of [21 ] . For example, the sets I 1 / k I_{1/k} I 1/ k associated with a fixed bounded Lipschitz domain form a tempered Følner sequence: ∣ I 1 / k ∣ ≍ k d |I_{1/k}|\asymp k^{d} ∣ I 1/ k ∣ ≍ k d , fixed-width boundary fractions vanish, and ⋃ j < k ( I 1 / k − I 1 / j ) \bigcup_{j<k}(I_{1/k}-I_{1/j}) ⋃ j < k ( I 1/ k − I 1/ j ) lies in a cube of O ( k d ) O(k^{d}) O ( k d ) sites. The same density-of-states conclusion therefore holds along these shrinking-cell samples, whether or not they are nested. The boundary comparison applies to each bounded Lipschitz sample separately. At δ = ε 2 \delta=\varepsilon^{2} δ = ε 2 , the frequency distribution is the push-forward of the reduced energy distribution under E ↦ v r E E\mapsto v_{r}\sqrt{E} E ↦ v r E . This map is independent of ε \varepsilon ε , so the moving-contrast limit is the push-forward of ν W , 0 \nu_{W,0} ν W , 0 . More generally, for fixed positive ε , δ \varepsilon,\delta ε , δ , if ρ W , δ \rho_{W,\delta} ρ W , δ is the density of ν W , δ \nu_{W,\delta} ν W , δ in energy, the corresponding frequency density at ω > 0 \omega>0 ω > 0 is
2 ε 2 ω δ v r 2 ρ W , δ ( ε 2 ω 2 δ v r 2 ) . (186) \frac{2\varepsilon^{2}\omega}{\delta v_{r}^{2}}\rho_{W,\delta}\left(\frac{\varepsilon^{2}\omega^{2}}{\delta v_{r}^{2}}\right). \tag*{(186)} δ v r 2 2 ε 2 ω ρ W , δ ( δ v r 2 ε 2 ω 2 ) . (186)
The factor is the Jacobian d E / d ω \mathrm{d}E/\mathrm{d}\omega d E / d ω ; an energy density cannot be read directly as a frequency density.
A density-of-states measure can have a density in a localisation regime: averaging atomic random spectral measures need not produce a singular measure. Neither spectral-set convergence nor asymptotic equivalence proves the absence of absolutely continuous spectrum. That conclusion follows from Part I.
Localisation of finite-cluster modes
We use the uniform estimates for truncated matrices from Theorem 6.1, Corollary 6.4, and Proposition 10.3. Let I = ( E − , E + ) I=(E_{-},E_{+}) I = ( E − , E + ) be bounded, E − > 0 E_{-}>0 E − > 0 , and suppose that for the truncated matrices
E Q δ , ϖ , Λ t ( x , y ; I ) ≤ A e − ν ∣ x − y ∣ 1 (187) \mathbb{E}Q^{t}_{\delta,\varpi,\Lambda}(x,y;I)\leq A e^{-\nu|x-y|_{1}} \tag*{(187)} E Q δ , ϖ , Λ t ( x , y ; I ) ≤ A e − ν ∣ x − y ∣ 1 (187)
uniformly in Λ \Lambda Λ and the parameters considered. Here Q t Q^{t} Q t is the correlator in (11) , with C δ C_{\delta} C δ in place of C C C . The estimate holds near the upper edge for every fixed W > 0 W>0 W > 0 and on positive intervals satisfying (160) .
The approximation of interior localised modes is related to the defect-mode comparison in [5 ] . Here one cannot assume a fixed isolation gap: the finite random eigenvalues become dense. Proposition 11.1 provides a sufficiently small gap scale with high probability.
Theorem 15.1 (Localisation outside an exceptional family). Assume (177) , (8) , and (187) , and let N = ∣ Λ ∣ ≥ 2 N=|\Lambda|\geq2 N = ∣Λ∣ ≥ 2 . For every p > 0 p>0 p > 0 , there are C p , R p < ∞ C_{p},R_{p}<\infty C p , R p < ∞ , independent of the finite sample, such that with probability at least 1 − C p N − p 1-C_{p}N^{-p} 1 − C p N − p the following holds. There is an excluded set B δ , ϖ , Λ , p ( I ) \mathcal{B}_{\delta,\varpi,\Lambda,p}(I) B δ , ϖ , Λ , p ( I ) of eigenvalue indices in I I I , with
∣ B δ , ϖ , Λ , p ( I ) ∣ ≤ C p ∣ ∂ R p log ( 2 + N ) Λ ∣ . (188) \left|\mathcal{B}_{\delta,\varpi,\Lambda,p}(I)\right|\leq C_{p}\left|\partial_{R_{p}\log(2+N)}\Lambda\right|. \tag*{(188)} ∣ B δ , ϖ , Λ , p ( I ) ∣ ≤ C p ∂ R p l o g ( 2 + N ) Λ . (188)
such that every normalised eigenvector ψ j f \psi_j^f ψ j f with E j f ∈ I E_j^f \in I E j f ∈ I and j ∉ B δ , ϖ , Λ , p ( I ) j \notin\mathcal{B}_{\delta,\varpi,\Lambda,p}(I) j ∈ / B δ , ϖ , Λ , p ( I ) has a centre n j ∈ Λ n_j \in\Lambda n j ∈ Λ satisfying
∑ ∣ n − n j ∣ 1 > R p log ( 2 + N ) ∣ ψ j f ( n ) ∣ 2 ≤ N − p . (189) \sum_{\lvert n-n_j\rvert_1>R_p\log(2+N)} \lvert\psi_j^f(n)\rvert^2 \le N^{-p}. \tag*{(189)} ∣ n − n j ∣ 1 > R p l o g ( 2 + N ) ∑ ∣ ψ j f ( n ) ∣ 2 ≤ N − p . (189)
Each retained eigenvector is, up to phase, within C p N − ( p + 5 ) C_pN^{-(p+5)} C p N − ( p + 5 ) in ℓ 2 \ell^2 ℓ 2 of a localised eigenvector of the truncated matrix. The excluded set consists of the indices not obtained by the injective matching in the proof. The constants may depend on W W W , the positive-energy bounds, and the constants in (187) . They are uniform for the contrast-dependent family as δ → 0 \delta\to0 δ → 0 . Intervals may vary with δ \delta δ if their endpoints stay in a common bounded subset of ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) and (187) remains uniform.
Proof. Fix p > 0 p>0 p > 0 . In this proof write H t , f = H δ , ϖ , Λ t , f H^{t,f}=H_{\delta,\varpi,\Lambda}^{t,f} H t , f = H δ , ϖ , Λ t , f and T = H t − H f T=H^t-H^f T = H t − H f ; let K K K be the common norm bound in (177) . We use a spectral gap scale g N = N − b g_N=N^{-b} g N = N − b with b = p + 3 b=p+3 b = p + 3 , and an error scale η N = N − q \eta_N=N^{-q} η N = N − q with q = 2 p + 8 q=2p+8 q = 2 p + 8 . These satisfy
N 2 g N ≤ N − p − 1 , N η N ≤ N − p − 1 , η N / g N = N − ( p + 5 ) . (190) N^2g_N\le N^{-p-1}, \qquad N\eta_N\le N^{-p-1}, \qquad\eta_N/g_N=N^{-(p+5)}. \tag*{(190)} N 2 g N ≤ N − p − 1 , N η N ≤ N − p − 1 , η N / g N = N − ( p + 5 ) . (190)
Enlarging fixed constants handles bounded values of N N N .
Step 1: one event for all eigenvectors of the truncated matrix. From (187) , Markov’s inequality and a union bound over pairs of sites give, with probability 1 − O ( N − p ) 1-O(N^{-p}) 1 − O ( N − p ) , an exponential bound with a polynomial prefactor for every eigenvector of H t H^t H t in I I I . Choosing the constant in a logarithmic radius sufficiently large, one obtains simultaneously centres n j n_j n j and
∥ 1 { ∣ n − n j ∣ 1 > ρ N } ψ j t ∥ 2 ≤ η N 8 ( K + 1 ) , ρ N = a p log ( 2 + N ) . (191) \left\|1_{\{\lvert n-n_j\rvert_1>\rho_N\}}\psi_j^t\right\|_2\le\frac{\eta_N}{8(K+1)}, \qquad\rho_N=a_p\log(2+N). \tag*{(191)} 1 {∣ n − n j ∣ 1 > ρ N } ψ j t 2 ≤ 8 ( K + 1 ) η N , ρ N = a p log ( 2 + N ) . (191)
Indeed, simplicity gives ∣ ψ j t ( x ) ψ j t ( y ) ∣ ≤ Q t ( x , y ; I ) \lvert\psi_j^t(x)\psi_j^t(y)\rvert\le Q^t(x,y;I) ∣ ψ j t ( x ) ψ j t ( y )∣ ≤ Q t ( x , y ; I ) ; at a maximum site ∣ ψ j t ( n j ) ∣ ≥ N − 1 / 2 \lvert\psi_j^t(n_j)\rvert\ge N^{-1/2} ∣ ψ j t ( n j )∣ ≥ N − 1/2 . A common estimate Q t ( x , y ; I ) ≤ N p + 3 e − ν ∣ x − y ∣ 1 / 2 Q^t(x,y;I)\le N^{p+3}e^{-\nu\lvert x-y\rvert_1/2} Q t ( x , y ; I ) ≤ N p + 3 e − ν ∣ x − y ∣ 1 /2 has failure probability O ( N − p − 2 ) O(N^{-p-2}) O ( N − p − 2 ) . Summing its resulting squared eigenvector bound outside a p log ( 2 + N ) a_p\log(2+N) a p log ( 2 + N ) gives (191) for any prescribed polynomial error after increasing a p a_p a p .
Step 2: the boundary perturbation is almost finite rank. Choose r N = a p ′ log ( 2 + N ) r_N=a'_p\log(2+N) r N = a p ′ log ( 2 + N ) large enough that, with P = P r N P=P_{r_N} P = P r N ,
∥ T ( Id − P ) ∥ ≤ η N / 8 , ∥ T − P T P ∥ ≤ η N / 4. (192) \lVert T(\operatorname{Id}-P)\rVert\le\eta_N/8,\qquad\lVert T-PTP\rVert\le\eta_N/4. \tag*{(192)} ∥ T ( Id − P )∥ ≤ η N /8 , ∥ T − P T P ∥ ≤ η N /4. (192)
This uses (178) and holds deterministically. Set r = rank P = ∣ ∂ r N Λ ∣ r=\operatorname{rank}P=\lvert\partial_{r_N}\Lambda\rvert r = rank P = ∣ ∂ r N Λ ∣ . Then
H f = H t − P T P + E N , ∥ E N ∥ ≤ η N / 4. H^f=H^t-PTP+\mathcal{E}_N,\qquad\lVert\mathcal{E}_N\rVert\le\eta_N/4. H f = H t − P T P + E N , ∥ E N ∥ ≤ η N /4.
The large part of the perturbation has rank at most r r r .
Step 3: eigenvalue separation and interval endpoints. Apply Proposition 11.1 to C ∗ = C δ , Λ f C_* = C_{\delta,\Lambda}^f C ∗ = C δ , Λ f . In a fixed positive-energy neighbourhood of I ‾ \overline{I} I , all distinct eigenvalues of H f H^f H f are at distance at least g N g_N g N , except with probability O ( N 2 g N ) = O ( N − p − 1 ) O(N^2g_N)=O(N^{-p-1}) O ( N 2 g N ) = O ( N − p − 1 ) . The first estimate in (161) , applied to H t H^t H t , also shows that there is no eigenvalue of H t H^t H t within 2 η N 2\eta_N 2 η N of either endpoint of I I I , except with probability O ( N η N ) O(N\eta_N) O ( N η N ) . Both estimates are uniform in the deterministic capacitance matrix. Take the intersection with the event of Step 1.
Step 4: match each interior-centred eigenvector of the truncated matrix. Consider an eigenvector from (191) with
d Λ ( n j ) > r N + ρ N + 1. (193) d_\Lambda(n_j)>r_N+\rho_N+1. \tag*{(193)} d Λ ( n j ) > r N + ρ N + 1. (193)
Its localisation ball lies in the range of Id − P \operatorname{Id}-P Id − P . By (191) and (192) ,
∥ ( H f − E j t ) ψ j t ∥ = ∥ T ψ j t ∥ ≤ η N / 8 + K η N / [ 8 ( K + 1 ) ] < η N / 4. (194) \left\|(H^f-E_j^t)\psi_j^t\right\|=\left\|T\psi_j^t\right\|\le\eta_N/8+K\eta_N/[8(K+1)]<\eta_N/4. \tag*{(194)} ( H f − E j t ) ψ j t = T ψ j t ≤ η N /8 + K η N / [ 8 ( K + 1 )] < η N /4. (194)
There is an eigenvalue E k f E_k^f E k f within η N / 4 \eta_N/4 η N /4 of E j t E_j^t E j t , and the endpoint event ensures it belongs to I I I . The gap event makes it the unique such eigenvalue. Expanding ψ j t \psi_j^t ψ j t in an eigenbasis of H f H^f H f gives
∥ ( Id − P k f ) ψ j t ∥ ≤ η N g N − η N , \left\|(\operatorname{Id}-P_k^f)\psi_j^t\right\| \le\frac{\eta_N}{g_N-\eta_N}, ( Id − P k f ) ψ j t ≤ g N − η N η N ,
where P k f P_k^f P k f is its one-dimensional projection. Thus, for a suitable phase,
∥ ψ k f − e i φ j ψ j t ∥ 2 ≤ 3 η N g N . (195) \left\|\psi_k^f-e^{i\varphi_j}\psi_j^t\right\|_2 \le\frac{3\eta_N}{g_N}. \tag*{(195)} ψ k f − e i φ j ψ j t 2 ≤ g N 3 η N . (195)
Two orthogonal eigenvectors of H t H^t H t cannot be matched to the same P k f P_k^f P k f : each would have squared projection larger than 1 / 2 1/2 1/2 onto a one-dimensional space, contradicting Bessel’s inequality. The matching is injective. Equations (191) , (195) , and (190) prove (189) for each matched eigenvector, with room in the exponents to absorb fixed constants.
Step 5: count the unmatched finite modes. Eigenvectors of H t H^t H t failing (193) have all but O ( η N 2 ) O(\eta_N^2) O ( η N 2 ) of their mass in ∂ r N + 2 ρ N + 2 Λ \partial_{r_N+2\rho_N+2\Lambda} ∂ r N + 2 ρ N + 2Λ . Orthogonality therefore bounds their number by
2 ∣ ∂ r N + 2 ρ N + 2 Λ ∣ . (196) 2\left|\partial_{r_N+2\rho_N+2\Lambda}\right|. \tag*{(196)} 2 ∣ ∂ r N + 2 ρ N + 2Λ ∣ . (196)
This bounds the number of boundary-centred modes of the truncated matrix without claiming that the eventual exceptional finite modes themselves must be supported in that layer.
A rank-r r r perturbation changes a cumulative eigenvalue count by at most r r r . The additional norm error in Step 2 can move a counting threshold by at most η N / 4 \eta_N/4 η N /4 . Since there are no eigenvalues of H t H^t H t within 2 η N 2\eta_N 2 η N of the endpoints, the eigenvalue counts for H f H^f H f and H t H^t H t on I I I differ by at most 2 r 2r 2 r . Consequently, the number of finite eigenvectors in I I I not reached by the injective matching is at most
2 r + 2 ∣ ∂ r N + 2 ρ N + 2 Λ ∣ ≤ 4 ∣ ∂ r N + 2 ρ N + 2 Λ ∣ . 2r+2\left|\partial_{r_N+2\rho_N+2\Lambda}\right|
\le4\left|\partial_{r_N+2\rho_N+2\Lambda}\right|. 2 r + 2 ∣ ∂ r N + 2 ρ N + 2Λ ∣ ≤ 4 ∣ ∂ r N + 2 ρ N + 2Λ ∣ .
Define B δ , ϖ , Λ , p ( I ) B_{\delta,\varpi,\Lambda,p}(I) B δ , ϖ , Λ , p ( I ) to be precisely these unmatched indices. This proves (188) . More explicitly, if n A ( E ) = Tr 1 ( − ∞ , E ] ( A ) n_A(E)=\operatorname{Tr}\mathbf{1}_{(-\infty,E]}(A) n A ( E ) = Tr 1 ( − ∞ , E ] ( A ) , rank perturbation and norm perturbation give
n H t ( E − η N / 4 ) − r ≤ n H f ( E ) ≤ n H t ( E + η N / 4 ) + r . n_{H^t}(E-\eta_N/4)-r\le n_{H^f}(E)\le n_{H^t}(E+\eta_N/4)+r. n H t ( E − η N /4 ) − r ≤ n H f ( E ) ≤ n H t ( E + η N /4 ) + r .
Apply this at both endpoints. Positivity of P T P PTP P T P actually improves the interval bound to r r r on the endpoint event, but 2 r 2r 2 r suffices. Finally add the three failure probabilities from Steps 1 and 3. This proves the theorem. □
15.1. Localisation as the cells shrink. Use the cell set I ε I_\varepsilon I ε and its cardinality N ε N_\varepsilon N ε from (154) , and take δ = ε 2 \delta=\varepsilon^2 δ = ε 2 . For 1 ≤ r ≪ ε − 1 1\le r\ll\varepsilon^{-1} 1 ≤ r ≪ ε − 1 , its boundary count obeys
∣ ∂ r I ε ∣ ≤ C Ω ε − d ( r + 1 ) . (197) \left|\partial_r I_\varepsilon\right|\le C_\Omega\varepsilon^{-d}(r+1). \tag*{(197)} ∣ ∂ r I ε ∣ ≤ C Ω ε − d ( r + 1 ) . (197)
Indeed a retained cell within r r r lattice steps of an omitted cell lies in a physical tubular neighbourhood of ∂ Ω \partial\Omega ∂ Ω of thickness C ε ( r + 1 ) C\varepsilon(r+1) C ε ( r + 1 ) . Its volume is O ( ε ( r + 1 ) ) O(\varepsilon(r+1)) O ( ε ( r + 1 )) for a bounded Lipschitz domain; divide by the cell volume.
Theorem 15.3 (Localisation in finite resonator clusters). Assume the geometry of § 2.1 , including (R), and the law (8) . For every fixed W > 0 W>0 W > 0 , let
I W , δ = ( c M , δ − Δ W , c M , δ ) (198) I_{W,\delta}=(c_{M,\delta}-\Delta_W,c_{M,\delta}) \tag*{(198)} I W , δ = ( c M , δ − Δ W , c M , δ ) (198) be the interval furnished by Proposition 10.3 . For every p > 0 p>0 p > 0 and all sufficiently small ε \varepsilon ε , with probability at least 1 − C W , p N ε − p 1-C_{W,p}N_\varepsilon^{-p} 1 − C W , p N ε − p , there is an excluded set B ε , p \mathcal{B}_{\varepsilon,p} B ε , p with
∣ B ε , p ∣ ≤ C W , p ε 1 − d log ( 1 / ε ) (199) |\mathcal{B}_{\varepsilon,p}|\le C_{W,p}\varepsilon^{1-d}\log(1/\varepsilon) \tag*{(199)} ∣ B ε , p ∣ ≤ C W , p ε 1 − d log ( 1/ ε ) (199) such that every ℓ 2 \ell^2 ℓ 2 -normalised eigenvector of the random finite matrix H ε 2 , ϖ , I ε f H^f_{\varepsilon^2,\varpi,I_\varepsilon} H ε 2 , ϖ , I ε f with eigenvalue in I W , ε 2 I_{W,\varepsilon^2} I W , ε 2 and index outside B ε , p \mathcal{B}_{\varepsilon,p} B ε , p satisfies (189) . Here B ε , p \mathcal{B}_{\varepsilon,p} B ε , p abbreviates B ε 2 , ϖ , I ε , p ( I W , ε 2 ) \mathcal{B}_{\varepsilon^2,\varpi,I_\varepsilon,p}(I_{W,\varepsilon^2}) B ε 2 , ϖ , I ε , p ( I W , ε 2 ) from Theorem 15.1 . The cell interpolations of the retained eigenvectors concentrate in physical balls of radius
r ε = C W , p ε log ( 1 / ε ) , (200) r_\varepsilon=C_{W,p}\varepsilon\log(1/\varepsilon), \tag*{(200)} r ε = C W , p ε log ( 1/ ε ) , (200) with squared tail at most N ε − p N_\varepsilon^{-p} N ε − p . For the original amplitudes q = M − 1 / 2 ψ q=M^{-1/2}\psi q = M − 1/2 ψ , normalising in ℓ 2 \ell^2 ℓ 2 costs at most a factor 1 + W 1+W 1 + W in that squared tail.
The same conclusion holds on a prescribed bounded positive interval satisfying (160) , without reflection symmetry, and on the quantitative weak-disorder window of Proposition 10.3 . It also holds for the static finite matrices H ϖ , I ε f H^f_{\varpi,I_\varepsilon} H ϖ , I ε f , with the corresponding static intervals. For the family C ε 2 C_{\varepsilon^2} C ε 2 at δ = ε 2 \delta=\varepsilon^2 δ = ε 2 , the edge interval in frequency is
ν r c M , ε 2 − Δ W < ω < ν r c M , ε 2 . (201) \nu_r\sqrt{c_{M,\varepsilon^2}-\Delta_W}<\omega<\nu_r\sqrt{c_{M,\varepsilon^2}}. \tag*{(201)} ν r c M , ε 2 − Δ W < ω < ν r c M , ε 2 . (201) The eigenvectors and their spatial estimates are unchanged by this reparametrisation. The endpoints converge to ν r c M − Δ W \nu_r\sqrt{c_M-\Delta_W} ν r c M − Δ W and ν r c M \nu_r\sqrt{c_M} ν r c M .
Proof. Use Theorem 15.1 , the uniform correlator bound for truncated matrices, and (197) . The isometry J ε J_\varepsilon J ε defined in § 4.1 converts lattice radii R log ( 2 + N ε ) R\log(2+N_\varepsilon) R log ( 2 + N ε ) to Euclidean radii at most ε R log ( 2 + N ε ) + d ε / 2 \varepsilon R\log(2+N_\varepsilon)+\sqrt{d}\varepsilon/2 εR log ( 2 + N ε ) + d ε /2 . Also ( 1 + W ) − 1 / 2 ≤ ∥ M − 1 / 2 ψ ∥ 2 ≤ 1 (1+W)^{-1/2}\leq\lVert M^{-1/2}\psi\rVert_2\leq1 ( 1 + W ) − 1/2 ≤ ∥ M − 1/2 ψ ∥ 2 ≤ 1 , which proves the final amplitude statement. The other intervals follow from Proposition 10.3 and the same transfer theorem. □
The exceptional fraction among all N ε N_\varepsilon N ε modes is
O W , p ( ε log ( 1 / ε ) ) ⟶ 0. O_{W,p}(\varepsilon\log(1/\varepsilon))\longrightarrow0. O W , p ( ε log ( 1/ ε )) ⟶ 0.
When the chosen interval has positive limiting density of states, it is also a vanishing fraction of the modes within that interval . For the fixed-disorder edge interval, this is the case: the almost-sure spectrum of the infinite model is [ 0 , c M ] [0,c_M] [ 0 , c M ] , the support of its density of states is the same interval, and the limiting window ( c M − Δ W , c M ) (c_M-\Delta_W,c_M) ( c M − Δ W , c M ) is nonempty. The moving-contrast eigenvalue-distribution result provides the corresponding positive asymptotic mode count along the sequences of growing samples considered in § 5.5 .
Along any sequence with ∑ k N ε k − p < ∞ \sum_k N_{\varepsilon_k}^{-p}<\infty ∑ k N ε k − p < ∞ , the concentration statement holds eventually almost surely. For example, ε k = 1 / k \varepsilon_k=1/k ε k = 1/ k and p > 1 / d p>1/d p > 1/ d suffice. The constants can deteriorate as W ↓ 0 W\downarrow0 W ↓ 0 ; arbitrary fixed positive disorder is distinct from a disorder strength tending to zero with ε \varepsilon ε .
Interior extent and the scope of the finite-cluster result
The boundary comparison also transfers the interior eigenfunction bounds to finite clusters, with an explicit bound on the number of excluded modes.
Figure 5. Localisation in a fixed spectral window and a bounded physical sample. The dots schematically indicate the mass of one mode. For finite capacitance matrices, with probability tending to one, every mode in I W I_W I W outside the excluded family of boundary-order size has negligible mass outside a ball of radius r ε = O W , p ( ε log ( 1 / ε ) ) r_\varepsilon= O_{W,p}(\varepsilon\log(1/\varepsilon)) r ε = O W , p ( ε log ( 1/ ε )) . For truncated matrices the conclusion holds for all modes in the window. The dashed circles illustrate this shrinking bound. The two centres are different because the theorem does not identify modes or centres across sample sizes. The spectral window stays fixed as the cells shrink.
Corollary 15.4. For the static finite capacitance matrix, the interior eigenfunction bounds of Theorem 8.3 hold outside a family of at most C ∣ ∂ ℓ W Λ ∣ / W 2 C|\partial_{\ell_W}\Lambda|/W^2 C ∣ ∂ ℓ W Λ∣/ W 2 modes, where ℓ W = C ( 1 + log ( 1 / W ) ) \ell_W = C(1+\log(1/W)) ℓ W = C ( 1 + log ( 1/ W )) and 0 < W ≤ W 0 0 < W \le W_0 0 < W ≤ W 0 . The energy condition remains μ W E j ∈ I \mu_W E_j \in I μ W E j ∈ I , with the regular interval I I I of (142) . For each fixed small W W W , the exceptional proportion vanishes along growing regular samples.
Proof. For a normalised finite-cluster amplitude q q q , its zero extension obeys
( C − μ W E ) q ~ = E W Z ϖ q ~ + E Λ ( C Λ t − C Λ f ) q + ( 1 − P Λ ) C q ~ . (C-\mu_W E)\widetilde{q}=EWZ\varpi\widetilde{q}+E_\Lambda(C_\Lambda^t-C_\Lambda^f)q+(1-P_\Lambda)C\widetilde{q}. ( C − μ W E ) q = E W Z ϖ q + E Λ ( C Λ t − C Λ f ) q + ( 1 − P Λ ) C q .
Lemma 12.1 controls the last term, and (174) together with ∥ R q ∥ 2 ≤ ∥ R ∥ ⟨ R q , q ⟩ \|Rq\|^2\leq\|R\|\langle Rq,q\rangle ∥ R q ∥ 2 ≤ ∥ R ∥ ⟨ R q , q ⟩ controls the middle term. Their sum is O ( W ) O(W) O ( W ) whenever the squared ℓ 2 \ell^2 ℓ 2 mass in ∂ ℓ W Λ \partial_{\ell_W}\Lambda ∂ ℓ W Λ is O ( W 2 ) O(W^2) O ( W 2 ) and ℓ W \ell_W ℓ W is chosen as above. Proposition 8.2 applies. For an orthonormal eigenbasis in the self-adjoint variables, the sum of squared boundary masses is ∣ ∂ ℓ W Λ ∣ |\partial_{\ell_W}\Lambda| ∣ ∂ ℓ W Λ∣ . The counting and bounded-change-of-variable argument in Theorem 12.3 proves the assertion. □
The matching theorem controls individual finite-cluster eigenvectors; it requires the spacing estimate as well as the boundary comparison. It leaves the unmatched family unclassified. The variational matrix also differs from the outgoing acoustic problem, whose resonances can be complex. Transfer to that problem requires a reduction and field reconstruction estimate uniform in sample size and compatible with the shrinking-cell, fixed-disorder limit.
Numerical illustrations
We illustrate the static finite-cluster problem with square arrays of disks. The calculations use the finite capacitance matrix of (170) and retain all pair interactions. They use the static limit C 0 = C C_0 = C C 0 = C of the reduced model; physical frequencies are assigned through (6) with positive contrast.
16.1. Geometry, discretisation, and observables. Let D D D be the disk of radius 0.2 0.2 0.2 centred at the origin, and set
S L = { 0 , … , L − 1 } 2 , N = ∣ S L ∣ = L 2 , L ∈ { 36 , 51 , 70 } . S_L = \{0,\ldots,L-1\}^2,\qquad N = |S_L| = L^2,\qquad L \in\{36,51,70\}. S L = { 0 , … , L − 1 } 2 , N = ∣ S L ∣ = L 2 , L ∈ { 36 , 51 , 70 } .
Here L L L counts cells along one side; S L S_L S L is distinguished from the centred boxes Λ L \Lambda_L Λ L used above. The lattice spacing is one. After rescaling by ε = L − 1 \varepsilon= L^{-1} ε = L − 1 and translating the sample, its cells fill the unit square.
We discretise the logarithmic single-layer potential on each disk by a Fourier–Galerkin method, retaining the constant, sine, and cosine modes through order two. The potential at infinity is an unknown constant, and the total charge is constrained to vanish, consistently with the two-dimensional energy convention in (170) . Eliminating the nonconstant harmonics and dividing the resulting potential-to-charge map by ∣ D ∣ |D| ∣ D ∣ gives the numerical matrix C S L f , ( 2 ) C_{S_L}^{f,(2)} C S L f , ( 2 ) . We form
H ϖ , S L f , ( 2 ) = M ϖ , S L − 1 / 2 C S L f , ( 2 ) M ϖ , S L − 1 / 2 , m n = 1 + W ξ n , (202) H_{\varpi,S_L}^{f,(2)} = M_{\varpi,S_L}^{-1/2} C_{S_L}^{f,(2)} M_{\varpi,S_L}^{-1/2},\qquad m_n = 1 + W \xi_n, \tag*{(202)} H ϖ , S L f , ( 2 ) = M ϖ , S L − 1/2 C S L f , ( 2 ) M ϖ , S L − 1/2 , m n = 1 + W ξ n , (202)
with the law (8) . We use 500 independent realisations for each W ∈ { 0.1 , 0.3 , 1 , 4 } W \in\{0.1,0.3,1,4\} W ∈ { 0.1 , 0.3 , 1 , 4 } and each L L L . The W = 0 W=0 W = 0 control is deterministic and is computed once.
Write ( E j , ψ j ) (E_j,\psi_j) ( E j , ψ j ) for the eigenpairs of (202) , with ∥ ψ j ∥ 2 = 1 \|\psi_j\|_2=1 ∥ ψ j ∥ 2 = 1 and positive eigenvalues in increasing order, j = 1 , … , N − 1 j=1,\ldots,N-1 j = 1 , … , N − 1 . The zero eigenvalue, whose eigenvector is proportional to M ϖ , S L 1 / 2 1 M_{\varpi,S_L}^{1/2}\mathbf{1} M ϖ , S L 1/2 1 , is omitted. As in (9) , q j = M ϖ , S L − 1 / 2 ψ j q_j=M_{\varpi,S_L}^{-1/2}\psi_j q j = M ϖ , S L − 1/2 ψ j ; thus the plotted mass ∣ ψ j ( n ) ∣ 2 |\psi_j(n)|^2 ∣ ψ j ( n ) ∣ 2 equals m n ∣ q j ( n ) ∣ 2 m_n|q_j(n)|^2 m n ∣ q j ( n ) ∣ 2 . The reference periodic edge is computed independently from the grounded-cell problem at α M = ( π , π ) \alpha_M=(\pi,\pi) α M = ( π , π ) , giving c M ≃ 50.4013 c_M\simeq50.4013 c M ≃ 50.4013 . All normalised energies in the figures use this numerical reference. For an individual mode, E j / c M = ( ω j / ω M ) 2 E_j/c_M=(\omega_j/\omega_M)^2 E j / c M = ( ω j / ω M ) 2 ; thus a point at E / c M = 0.45 E/c_M=0.45 E / c M = 0.45 corresponds to ω / ω M = 0.45 ≃ 0.671 \omega/\omega_M=\sqrt{0.45}\simeq0.671 ω / ω M = 0.45 ≃ 0.671 .
For a normalised mode ψ \psi ψ , use the participation number P ( ψ ) P(\psi) P ( ψ ) from (143) . Let n ∗ n_\ast n ∗ be a site where its mass is maximal and define
R 90 ( ψ ) = min { r ≥ 0 : ∑ ∣ n − n ∗ ∣ 2 ≤ r ∣ ψ ( n ) ∣ 2 ≥ 0.9 } . (203) R_{90}(\psi)=\min\left\{r\geq0:\sum_{|n-n_\ast|^2\leq r}|\psi(n)|^2\geq0.9\right\}. \tag*{(203)} R 90 ( ψ ) = min ⎩ ⎨ ⎧ r ≥ 0 : ∣ n − n ∗ ∣ 2 ≤ r ∑ ∣ ψ ( n ) ∣ 2 ≥ 0.9 ⎭ ⎬ ⎫ . (203)
Then P / N P/N P / N measures the fraction of sites effectively occupied by the mode, and R 90 / L R_{90}/L R 90 / L is its physical concentration radius. This radius is centred at a maximum site and is distinct from the second-moment radius used in § 3.14 .
For the spatial profiles only, we select modes satisfying
β 2 ( ψ ) : = ∑ n ∈ ∂ 2 S L ∣ ψ ( n ) ∣ 2 ≤ 0.2 , (204) \beta_2(\psi):=\sum_{n\in\partial_2 S_L}|\psi(n)|^2\leq0.2, \tag*{(204)} β 2 ( ψ ) := n ∈ ∂ 2 S L ∑ ∣ ψ ( n ) ∣ 2 ≤ 0.2 , (204)
where ∂ 2 S L \partial_2 S_L ∂ 2 S L consists of the two outermost layers, as in (173) . The upper profile uses the largest eigenvalue among these modes. The interior profile uses the eigenvalue closest to 0.45 c M 0.45c_M 0.45 c M among them. Every realisation supplies one mode to each average. Condition (204) is a numerical selection rule. It does not identify the exceptional family of Theorem 15.1 .
Ensemble averages
Let S = 500 S = 500 S = 500 , and let ψ ~ ( s ) \tilde{\psi}^{(s)} ψ ~ ( s ) be the zero extension of the selected mode in realisation s s s . The spatial plots display
ρ ‾ L ( k / L ) = N S ∑ s = 1 S ∣ ψ ~ ( s ) ( n ∗ ( s ) + k ) ∣ 2 , k ∈ Z 2 . (205) \overline{\rho}_L(k/L) = \frac{N}{S} \sum_{s=1}^{S} \left|\tilde{\psi}^{(s)}\left(n_*^{(s)} + k\right)\right|^2,\qquad k \in\mathbb{Z}^2. \tag*{(205)} ρ L ( k / L ) = S N s = 1 ∑ S ψ ~ ( s ) ( n ∗ ( s ) + k ) 2 , k ∈ Z 2 . (205)
Thus intensities are translated to their maximum sites before averaging, with no rotation or periodic wrapping. Since localisation centres depend on the realisation, this centred mean differs from the intensity averaged at fixed physical sites. Its normalisation is L − 2 ∑ k ρ ‾ L ( k / L ) = 1 L^{-2} \sum_k \overline{\rho}_L(k/L) = 1 L − 2 ∑ k ρ L ( k / L ) = 1 . The quantities displayed in the panel titles are the means of the individual values of P P P and R 90 / L R_{90}/L R 90 / L , not the corresponding statistics of the mean profile. Angle brackets in the figures denote these empirical means.
For spectral participation, all positive modes are retained, including those failing (204) . At each ordered rank j j j we plot
( E ‾ j c M , P ‾ j N ) , E ‾ j = 1 S ∑ s = 1 S E j ( s ) , P ‾ j = 1 S ∑ s = 1 S P ( ψ j ( s ) ) . (206) \left(\frac{\overline{E}_j}{c_M},\frac{\overline{P}_j}{N}\right),\qquad\overline{E}_j = \frac{1}{S}\sum_{s=1}^{S} E_j^{(s)},\qquad\overline{P}_j = \frac{1}{S}\sum_{s=1}^{S}P\left(\psi_j^{(s)}\right). \tag*{(206)} ( c M E j , N P j ) , E j = S 1 s = 1 ∑ S E j ( s ) , P j = S 1 s = 1 ∑ S P ( ψ j ( s ) ) . (206)
These are averages at fixed rank, rather than conditional averages at fixed energy. In frequency terms, E ‾ j / c M = ( ω j / ω M ) 2 \overline{E}_j/c_M = (\omega_j/\omega_M)^2 E j / c M = ( ω j / ω M ) 2 ; its square root is a root-mean-square frequency ratio, not the mean frequency ratio. The ribbons give ± 1.96 \pm1.96 ± 1.96 sample standard errors of the vertical mean.
Concentration near the top of the spectrum
Figures 6–8 show the upper profiles for the three sizes. Table 3 records the associated statistics. Without disorder, the participation number is approximately 0.46 N 0.46N 0.46 N , and the physical radius remains close to 0.39 0.39 0.39 . At fixed positive disorder, the participation numbers grow much more slowly than N N N , and the physical radii decrease with L L L . For example, at W = 0.1 W = 0.1 W = 0.1 , the mean participation number changes from 121.2 to 103.8 between L = 36 L = 36 L = 36 and L = 70 L = 70 L = 70 , while the mean physical radius decreases from 0.277 to 0.131. At W = 1 W = 1 W = 1 , the selected modes have participation numbers between 5.4 and 5.7 and physical radii decreasing from 0.049 to 0.025. The centred profiles display the same increasing concentration.
Table 3. Mean selected-mode statistics for the finite capacitance matrices. Each positive-disorder entry averages all 500 realisations; the first row is deterministic. The selection and observables are defined in (143) , (203) , and (204) .
Figure 9 complements the selected profiles by showing spectral participation for all positive modes. The reduction in participation is most pronounced in the upper part of the spectrum and extends to lower normalised energies as W W W increases. The blue shading marks the comparison interval ( ( 1 + 0.4 W ) − 1 , 1 ) \left((1 + 0.4W)^{-1}, 1\right) ( ( 1 + 0.4 W ) − 1 , 1 ) on the E / c M E/c_M E / c M axis. Its form is motivated by Theorem 7.1 , but its use in the plot does not verify that theorem’s small-disorder threshold. In particular, the selected upper modes need not lie in this interval. At W = 0.1 W = 0.1 W = 0.1 , only 14, 34, and 58 of the 500 samples, respectively, contain a positive eigenvalue in it. Thus these sample sizes resolve concentration of modes near the top of the spectrum more readily than the sparsely populated region closest to the infinite-volume edge.
Figure 9. Spectral participation for L = 36 , 51 , 70 L = 36, 51, 70 L = 36 , 51 , 70 (top to bottom) and W = 0.1 , 0.3 , 1 W = 0.1, 0.3, 1 W = 0.1 , 0.3 , 1 (left to right). Blue curves are the rank averages (206) over all 500 realisations, with ± 1.96 \pm1.96 ± 1.96 standard-error ribbons for the vertical mean. All positive modes are included. Grey curves give the periodic controls. Shading marks ( ( 1 + 0.4 W ) − 1 , 1 ) ((1 + 0.4W)^{-1}, 1) (( 1 + 0.4 W ) − 1 , 1 ) in E / c M E/c_M E / c M , a comparison window rather than a certified localisation interval. The corresponding interval for individual frequencies is ( 1 + 0.4 W ) − 1 / 2 < ω / ω M < 1 (1 + 0.4W)^{-1/2} < \omega/\omega_M < 1 ( 1 + 0.4 W ) − 1/2 < ω / ω M < 1 .
Interior modes at stronger disorder
Figure 10 compares the periodic interior profiles with those at W = 4 W = 4 W = 4 , using the target energy 0.45 c M 0.45c_M 0.45 c M . At this disorder strength, the mean participation number remains between 3.4 and 3.6 for all three sizes, while the mean physical radius decreases from 0.061 to 0.032; see Table 3 . The spectral panels show small participation fractions throughout the displayed comparison interval 0.35 < E / c M < 0.55 0.35 < E/c_M < 0.55 0.35 < E / c M < 0.55 . These observations illustrate concentration in the interior of the unperturbed band at stronger disorder, consistently with the fixed-interval localisation mechanism. They do not evaluate the sufficient disorder threshold in (160) . Here interior refers to the energy band ( 0 , c M ) (0,c_M) ( 0 , c M ) of C C C . At W = 4 W = 4 W = 4 , the target E = 0.45 c M E = 0.45c_M E = 0.45 c M has centred energy s = 1.35 c M s = 1.35c_M s = 1.35 c M and is outside the regular-centred-energy regime of Theorem 8.3 .
Figure 10. Interior modes for L = 36 , 51 , 70 L = 36,51,70 L = 36 , 51 , 70 (top to bottom). The first two columns show the periodic control and the centred mean at W = 4 W = 4 W = 4 , selecting the mode closest to 0.45 c M 0.45c_M 0.45 c M subject to (204) . Each random profile averages all 500 realisations. The right column gives rank-averaged participation for every positive mode at W = 4 W = 4 W = 4 , with ± 1.96 \pm1.96 ± 1.96 standard-error ribbons. Shading marks 0.35 < E / c M < 0.55 0.35 < E/c_M < 0.55 0.35 < E / c M < 0.55 . The profile normalisation is (205) .
Taken together, the figures support the finite-cluster concentration described by Theorem 15.3 . They do not test its simultaneous exceptional-mode count or establish the asymptotic ε log ( 1 / ε ) \varepsilon\log(1/\varepsilon) ε log ( 1/ ε ) radius from three sizes. Nor do participation numbers and 90% radii alone certify exponential tails. Large participation in the interior at weak disorder remains compatible with localisation on scales exceeding the samples. The calculations use a fixed Fourier discretisation of the static matrix; they do not measure the error of the full acoustic reduction.
Figure 7. Selected high-energy profiles for L = 51 L = 51 L = 51 (N = 2601 N = 2601 N = 2601 ), with the same selection, normalisation, and averaging as in Figure 6 . All 500 realisations enter each positive-disorder panel.
Figure 6. Selected high-energy profiles for L = 36 L = 36 L = 36 (N = 1296 N = 1296 N = 1296 ). Each panel uses the highest-energy mode satisfying (204) . Colours show the centred density (205) ; titles give the means of P P P and R 90 / L R_{90}/L R 90 / L . Each positive-disorder panel averages all 500 realisations. The W = 0 W = 0 W = 0 panel is the deterministic control.
Figure 8. Selected high-energy profiles for L = 70 L = 70 L = 70 (N = 4900 N = 4900 N = 4900 ), with the same selection, normalisation, and averaging as in Figure 6 . All 500 realisations enter each positive-disorder panel. The physical displacement is ( n − n ∗ ) / L (n - n_*)/L ( n − n ∗ ) / L .
Concluding remarks
We have established spectral and dynamical Anderson localisation for a random capacitance operator with long-range interactions. The proofs retain the full exponentially decaying capacitance interaction. Localisation holds near the upper band edge at every fixed positive disorder strength and on prescribed positive intervals at sufficiently large disorder. At weak disorder, the explicit edge window has energy width asymptotic to a M W a_{\mathrm{M}}W a M W , and frequency width asymptotic to ( a / 2 ) ω M W (a/2)\omega_{\mathrm{M}}W ( a /2 ) ω M W , for each fixed a ∈ ( 0 , 1 / 2 ) a \in(0,1/2) a ∈ ( 0 , 1/2 ) . It gives an upper bound on the upper endpoint of any absolutely continuous spectral component, if present.
For finite resonator clusters, uniform estimates for truncated matrices combine with the variational boundary comparison and material-variable eigenvalue spacing. With probability tending to one as the sample size increases, the resulting matching proves concentration of individual modes of the finite matrix outside an exceptional family of size bounded by a constant times the number of sites in a logarithmic boundary layer. Their physical radii are O ( ε log ( 1 / ε ) ) O(\varepsilon\log(1/\varepsilon)) O ( ε log ( 1/ ε )) , and the exceptional fraction vanishes in every localised window of positive limiting density of states. The estimates are uniform for the contrast-dependent capacitance family at small contrast. Weighted asymptotic equivalence also gives a common limiting eigenvalue distribution for the two finite constructions.
The computations for 36 2 36^2 3 6 2 , 51 2 51^2 5 1 2 , and 70 2 70^2 7 0 2 disks, with 500 realisations at each positive disorder strength, show decreasing participation fractions and physical concentration radii for selected high-energy modes. At stronger disorder, similar concentration occurs in an interior comparison interval. The periodic controls remain spread through the sample. These observations illustrate the finite-cluster results without determining asymptotic decay rates or a mobility edge.
The localisation properties of the unmatched modes, the interior spectral type at weak disorder, and the passage to the full transmission problem remain open in this analysis. The interior extent bounds quantify spatial scales but do not prove delocalisation. A uniform capacitance reduction is needed to transfer the finite-matrix conclusions to acoustic fields.
The frequency-dependent capacitance formulation of [10 ] suggests an extension beyond the subwavelength regime. Let λ ∗ > 0 \lambda_* > 0 λ ∗ > 0 be a simple Neumann eigenvalue of − Δ -\Delta − Δ in the reference resonator D D D . The corresponding reference frequency is ω ∗ = v r λ ∗ \omega_* = v_r\sqrt{\lambda_*} ω ∗ = v r λ ∗ at unit period, or v r λ ∗ / ε v_r\sqrt{\lambda_*}/\varepsilon v r λ ∗ / ε at physical period ε \varepsilon ε . Near such frequencies, the regular quasiperiodic reduction in [10 ] , § 2.11 describes the bands generated by the interior Neumann mode. We expect localisation near an isolated edge of these bands for every fixed positive disorder strength in the reduced modal problem. A proof would require uniform exterior non-resonance, exponential decay of the modal interactions, an isolated nondegenerate edge, suitable spectral averaging, and uniform control of the random reduction.
At unit period, the relevant spectral parameter is the detuning ( ω − ω ∗ ) / δ (\omega-\omega_*)/\delta ( ω − ω ∗ ) / δ . The disorder scale must also preserve the modal separation. For the inverse material factors used here, one possible scaling is W = δ W m o d W = \delta W_{\mathrm{mod}} W = δ W mod , with any fixed W m o d > 0 W_{\mathrm{mod}} > 0 W mod > 0 , so that m n = 1 + δ W m o d ζ n m_n = 1 + \delta W_{\mathrm{mod}}\zeta_n m n = 1 + δ W mod ζ n . The decoupled frequencies ω ∗ / m n \omega_*/\sqrt{m_n} ω ∗ / m n then differ from ω ∗ \omega_* ω ∗ by O ( δ ) O(\delta) O ( δ ) , the scale of the regular modal band splitting. Arbitrary strength in this prospective statement refers to W m o d W_{\mathrm{mod}} W mod ; unrestricted unscaled material variations can mix distinct modal bands. Localisation beyond the subwavelength regime therefore remains a conditional prospect. For finite outgoing arrays, radiation losses and comparison of the resulting resonance modes require additional estimates.
Funding declaration. The work of W.J. is partially supported by the NSFC Grant No. 12571220 and by the New Cornerstone Investigator Program 100001127.
Conflict of interest. The authors declare no competing financial or non-financial interests relevant to this work.
Declaration on the use of AI. During the preparation of this manuscript, the authors used ChatGPT (OpenAI, GPT-Astra 6) as an interactive writing, editing, and analytical assistant. The system was employed to explore mathematical ideas, assist in the development and checking of proof arguments, improve the organisation and exposition, and assist in preparing the LaTeX source, numerical code, and illustrations. All mathematical results presented in this article originate either from the cited sources or from the authors’ own work. The authors independently verified the mathematical arguments, references, numerical illustrations, and final formulations presented here. The authors assume full responsibility for the content and originality of this manuscript.
Appendix A. The three-dimensional weak-scattering scale
The mobility-edge predictions in this appendix are heuristic. They do not enter the proof of Theorem 7.1 or Corollary 7.4 . We use the centred variables μ W = 1 + W / 2 \mu_W = 1 + W/2 μ W = 1 + W /2 , λ = W / μ W \lambda= W/\mu_W λ = W / μ W , s = μ W E s = \mu_W E s = μ W E , and σ 2 = E ( ξ n − 1 / 2 ) 2 = 1 / 12 \sigma^2 = \mathbb{E}(\xi_n - 1/2)^2 = 1/12 σ 2 = E ( ξ n − 1/2 ) 2 = 1/12 from § 3.14 . The centred operator is H ~ ϖ = μ W H ϖ \widetilde{H}_{\varpi} = \mu_W H_{\varpi} H ϖ = μ W H ϖ as in (140) .
In dimension three, assume for the explicit calculation that A M = a M Id A_M = a_M \operatorname{Id} A M = a M Id , a M > 0 a_M > 0 a M > 0 , as for a cubic geometry with coordinate-permutation symmetry. Near the maximum,
c ( α M + β ) = c M − a M ∣ β ∣ 2 + O ( ∣ β ∣ 4 ) . (207) c(\alpha_M + \beta) = c_M - a_M |\beta|^2 + O(|\beta|^4). \tag*{(207)} c ( α M + β ) = c M − a M ∣ β ∣ 2 + O ( ∣ β ∣ 4 ) . (207)
Reciprocity and α M = − α M \alpha_M = -\alpha_M α M = − α M modulo 2 π Z d 2\pi\mathbb{Z}^d 2 π Z d remove the odd Taylor terms. Coarea in these local coordinates gives the rigorous unperturbed asymptotic
ρ c ( c M − Δ ) = Δ 4 π 2 a M 3 / 2 ( 1 + O ( Δ ) ) , Δ ↓ 0. (208) \rho_c(c_M - \Delta) = \frac{\sqrt{\Delta}}{4\pi^2 a_M^{3/2}}(1 + O(\Delta)), \qquad\Delta\downarrow0. \tag*{(208)} ρ c ( c M − Δ ) = 4 π 2 a M 3/2 Δ ( 1 + O ( Δ )) , Δ ↓ 0. (208)
Indeed the quadratic level set has area 4 π ∣ β ∣ 2 4\pi|\beta|^2 4 π ∣ β ∣ 2 , gradient modulus 2 a M ∣ β ∣ 2a_M|\beta| 2 a M ∣ β ∣ , and ∣ β ∣ = Δ / a M |\beta| = \sqrt{\Delta/a_M} ∣ β ∣ = Δ/ a M ; division by ( 2 π ) 3 (2\pi)^3 ( 2 π ) 3 gives the coefficient. Smooth even corrections give the stated relative remainder. For an anisotropic positive A M A_M A M , the coefficient is instead ( 4 π 2 det A M ) − 1 (4\pi^2\sqrt{\det A_M})^{-1} ( 4 π 2 det A M ) − 1 .
The simplest weak-scattering description replaces the true tail of the averaged spectrum by a coherent edge. In the centred spectral variable it predicts
s c o h ( λ ) = c M + O ( λ 2 ) , E c o h ( W ) = s c o h ( λ ) μ W = E m e a n ( W ) + O ( W 2 ) . (209) s_{\mathrm{coh}}(\lambda) = c_M + O(\lambda^2), \qquad E_{\mathrm{coh}}(W) = \frac{s_{\mathrm{coh}}(\lambda)}{\mu_W} = E_{\mathrm{mean}}(W) + O(W^2). \tag*{(209)} s coh ( λ ) = c M + O ( λ 2 ) , E coh ( W ) = μ W s coh ( λ ) = E mean ( W ) + O ( W 2 ) . (209)
These formulas describe an approximation, not a proven asymptotic for a spectral support endpoint. Almost surely, the upper edge is c M c_M c M in energy E E E , or μ W c M \mu_W c_M μ W c M in centred energy s s s , by Proposition 2.2 .
To specify the Born coefficient, expansion of (140) gives, uniformly in the material configuration,
H ~ ϖ = C + λ V 1 + O ℓ 2 → ℓ 2 ( λ 2 ) , V 1 = − 1 2 ( Z ϖ C + C Z ϖ ) . (210) \widetilde{H}_{\varpi} = C + \lambda V_1 + O_{\ell^2 \to\ell^2}(\lambda^2), \qquad V_1 = -\frac{1}{2}(Z_{\varpi}C + CZ_{\varpi}). \tag*{(210)} H ϖ = C + λ V 1 + O ℓ 2 → ℓ 2 ( λ 2 ) , V 1 = − 2 1 ( Z ϖ C + C Z ϖ ) . (210)
On a periodic lattice with N N N sites, let C C C denote its periodisation and let ϕ α ( n ) = N − 1 / 2 e i n ⋅ α \phi_\alpha(n) = N^{-1/2}e^{in\cdot\alpha} ϕ α ( n ) = N − 1/2 e in ⋅ α for the allowed quasimomenta. Then C ϕ α = c ( α ) ϕ α C\phi_\alpha= c(\alpha)\phi_\alpha C ϕ α = c ( α ) ϕ α and
E ∣ ⟨ ϕ α ′ , V 1 ϕ α ⟩ ∣ 2 = σ 2 4 N ( c ( α ) + c ( α ′ ) ) 2 . (211) \mathbb{E}\left|\langle\phi_{\alpha'},V_1\phi_\alpha\rangle\right|^2 = \frac{\sigma^2}{4N}\left(c(\alpha) + c(\alpha')\right)^2. \tag*{(211)} E ∣ ⟨ ϕ α ′ , V 1 ϕ α ⟩ ∣ 2 = 4 N σ 2 ( c ( α ) + c ( α ′ ) ) 2 . (211)
Indeed the matrix element of Z ϖ Z_{\varpi} Z ϖ is N − 1 ∑ n ( ξ n − 1 / 2 ) e i n ⋅ ( α − α ′ ) N^{-1}\sum_n(\xi_n - 1/2)e^{in\cdot(\alpha-\alpha')} N − 1 ∑ n ( ξ n − 1/2 ) e in ⋅ ( α − α ′ ) , whose variance is σ 2 / N \sigma^2/N σ 2 / N by independence. On the energy surface c ( α ) = c ( α ′ ) = s c(\alpha) = c(\alpha') = s c ( α ) = c ( α ′ ) = s , the coefficient in (A.5) is σ 2 s 2 / N \sigma^2s^2/N σ 2 s 2 / N . For the auxiliary evolution e − i t H ~ ϖ e^{-it\widetilde{H}_{\varpi}} e − i t H ϖ , whose time is conjugate to s s s , using the distributional identity t − 1 ∣ ∫ 0 t e i u τ d τ ∣ 2 → 2 π δ D ( u ) t^{-1}\left|\int_0^t e^{iu\tau}\,d\tau\right|^2 \to2\pi\delta_D(u) t − 1 ∫ 0 t e i uτ d τ 2 → 2 π δ D ( u ) , where δ D \delta_D δ D is the Dirac distribution, in the first Duhamel term gives the golden-rule prediction
Γ B o r n = 2 π λ 2 σ 2 s 2 ρ c ( s ) , ρ c ( s ) = ∫ { c = s } d S ( α ) ( 2 π ) 3 ∣ ∇ c ( α ) ∣ . (212) \Gamma_{\mathrm{Born}}=2\pi\lambda^2\sigma^2s^2\rho_c(s),\qquad\rho_c(s)=\int_{\{c=s\}}\frac{dS(\alpha)}{(2\pi)^3|\nabla c(\alpha)|}. \tag*{(212)} Γ Born = 2 π λ 2 σ 2 s 2 ρ c ( s ) , ρ c ( s ) = ∫ { c = s } ( 2 π ) 3 ∣∇ c ( α ) ∣ d S ( α ) . (212)
Here s s s is a regular energy, and the continuum of quasimomenta is taken before the large-time limit. The fixed-time perturbation expansion does not justify a microscopic kinetic limit.
Replacing the unperturbed density in (A.6) by the parabolic density measured from the coherent edge, with Δ c o h = s c o h − s > 0 \Delta_{\mathrm{coh}}=s_{\mathrm{coh}}-s>0 Δ coh = s coh − s > 0 , gives the further approximation
Γ B o r n ≃ 2 π λ 2 σ 2 c M 2 Δ c o h 4 π 2 a M 3 / 2 , ∣ ∇ c ∣ ≃ 2 a M Δ c o h . (213) \Gamma_{\mathrm{Born}}\simeq2\pi\lambda^2\sigma^2c_M^2\frac{\sqrt{\Delta_{\mathrm{coh}}}}{4\pi^2a_M^{3/2}},\qquad|\nabla c|\simeq2\sqrt{a_M\Delta_{\mathrm{coh}}}. \tag*{(213)} Γ Born ≃ 2 π λ 2 σ 2 c M 2 4 π 2 a M 3/2 Δ coh , ∣∇ c ∣ ≃ 2 a M Δ coh . (213)
Consequently
ℓ m f ≃ ∣ ∇ c ∣ Γ B o r n = 4 π a M 2 λ 2 σ 2 c M 2 . (214) \ell_{\mathrm{mf}}\simeq\frac{|\nabla c|}{\Gamma_{\mathrm{Born}}}=\frac{4\pi a_M^2}{\lambda^2\sigma^2c_M^2}. \tag*{(214)} ℓ mf ≃ Γ Born ∣∇ c ∣ = λ 2 σ 2 c M 2 4 π a M 2 . (214)
This is a mean free path in cell units, not an eigenfunction decay length. To convert this calculation to positive-frequency evolution, use (141) and put
F ( s ) = v r δ ε s μ W , F ′ ( s ) = v r δ 2 ε μ W s . F(s)=\frac{v_r\sqrt{\delta}}{\varepsilon}\sqrt{\frac{s}{\mu_W}},\qquad F'(s)=\frac{v_r\sqrt{\delta}}{2\varepsilon\sqrt{\mu_Ws}}. F ( s ) = ε v r δ μ W s , F ′ ( s ) = 2 ε μ W s v r δ .
On a regular positive energy surface, the group velocity in cell units is F ′ ( s ) ∇ c F'(s)\nabla c F ′ ( s ) ∇ c , and the first-order on-shell perturbation matrix elements acquire the factor F ′ ( s ) F'(s) F ′ ( s ) . In the golden rule, δ D ( F ( s ′ ) − F ( s ) ) = δ D ( s ′ − s ) / F ′ ( s ) \delta_D(F(s')-F(s))=\delta_D(s'-s)/F'(s) δ D ( F ( s ′ ) − F ( s )) = δ D ( s ′ − s ) / F ′ ( s ) , so the collision rate is F ′ ( s ) Γ B o r n F'(s)\Gamma_{\mathrm{Born}} F ′ ( s ) Γ Born . The factor cancels in their ratio. The physical mean free path is therefore ε ℓ m f \varepsilon\ell_{\mathrm{mf}} ε ℓ mf . This change of variables does not justify the kinetic approximation. Extrapolating the heuristic Ioffe–Regel rule ∣ β ∣ ℓ m f = χ I R |\beta|\ell_{\mathrm{mf}}=\chi_{\mathrm{IR}} ∣ β ∣ ℓ mf = χ IR , where χ I R > 0 \chi_{\mathrm{IR}}>0 χ IR > 0 is an undetermined constant of order one, gives
Δ I R ≃ χ I R 2 σ 4 c M 4 16 π 2 a M 3 λ 4 , σ 4 = ( σ 2 ) 2 . (215) \Delta_{\mathrm{IR}}\simeq\frac{\chi_{\mathrm{IR}}^2\sigma^4c_M^4}{16\pi^2a_M^3}\lambda^4,\qquad\sigma^4=(\sigma^2)^2. \tag*{(215)} Δ IR ≃ 16 π 2 a M 3 χ IR 2 σ 4 c M 4 λ 4 , σ 4 = ( σ 2 ) 2 . (215)
At this scale the weak-scattering approximation is itself at its limit of validity. Thus neither the coefficient nor the mobility transition is proved by the extrapolation. The same distinction between a quadratic coherent-edge shift and a fourth-order mobility-edge separation appears in the self-consistent treatment of a different three-dimensional Anderson model in [25 Section IV, equation (35)] .
Writing s m o b = μ W E m o b h e u r s_{\mathrm{mob}}=\mu_WE_{\mathrm{mob}}^{\mathrm{heur}} s mob = μ W E mob heur for the predicted transition in centred energy, these approximations give
s c o h − s m o b ≃ λ 4 , E m o b h e u r ( W ) = c M 1 + W / 2 + O ( W 2 ) , c M − E m o b h e u r ( W ) = 1 2 c M W + O ( W 2 ) . (216) s_{\mathrm{coh}}-s_{\mathrm{mob}}\simeq\lambda^4,\qquad E_{\mathrm{mob}}^{\mathrm{heur}}(W)=\frac{c_M}{1+W/2}+O(W^2),\qquad c_M-E_{\mathrm{mob}}^{\mathrm{heur}}(W)=\frac{1}{2}c_MW+O(W^2). \tag*{(216)} s coh − s mob ≃ λ 4 , E mob heur ( W ) = 1 + W /2 c M + O ( W 2 ) , c M − E mob heur ( W ) = 2 1 c M W + O ( W 2 ) . (216)
Equivalently, ω m o b h e u r ( W ) = ω M ( 1 − W / 4 + O ( W 2 ) ) \omega_{\mathrm{mob}}^{\mathrm{heur}}(W)=\omega_M(1-W/4+O(W^2)) ω mob heur ( W ) = ω M ( 1 − W /4 + O ( W 2 )) . An O ( W 2 ) O(W^2) O ( W 2 ) uncertainty in the coherent edge is larger than the predicted O ( W 4 ) O(W^4) O ( W 4 ) separation; one cannot obtain a fourth-order expansion for the mobility edge by retaining only the leading coherent-edge correction.
These formulas are specific to the inverse-material law M = diag ( 1 + W ξ n ) M=\operatorname{diag}(1+W\xi_n) M = diag ( 1 + W ξ n ) used here. Replacing it by a centred law or by a direct material multiplier changes the mean shift and the exact spectral endpoints. In two dimensions the usual scalar, time-reversal-invariant scaling picture instead predicts eventual localisation even at weak disorder, with very large lengths [1 ] . For the nearest-neighbour Anderson model with Gaussian disorder, time-averaged quantum diffusion on suitable weak-disorder scales and lower bounds on localisation length are established in [13 ] . These results do not predict a conducting bulk separated by a mobility edge of the form above. None of these physical expectations establishes the spectral type of our operator in the remaining interior band.
Appendix B. Other material distributions
The main finite-volume theorem is stated for the uniform inverse-material law. The following absence statement also covers direct uniform material perturbations. The explicit coefficient 1 / 2 1/2 1/2 in the quantitative weak-disorder theorem is specific to the mean of the main material law.
Proposition B.1. Assume the geometry of § 2.1 , including (R R R ). For m n = 1 + W ξ n m_n = 1 + W\xi_n m n = 1 + W ξ n and every fixed W > 0 W > 0 W > 0 , absence of absolutely continuous spectrum on a nonempty upper-edge interval remains valid if the independent variables ξ n \xi_n ξ n have a common bounded density supported in [ 0 , 1 ] [0,1] [ 0 , 1 ] , zero belongs to its support, and, for some a 0 > 0 a_0 > 0 a 0 > 0 , q 0 : = P { ξ n ≥ a 0 } > 0 q_0 := \mathbb{P}\{\xi_n \ge a_0\} > 0 q 0 := P { ξ n ≥ a 0 } > 0 . All constants may depend on that law. Consequently, if instead the direct factors b n b_n b n are independent and uniform on [ 1 , 1 + w ] [1,1+w] [ 1 , 1 + w ] , w > 0 w > 0 w > 0 , the operator B ϖ 1 / 2 C B ϖ 1 / 2 B_\varpi^{1/2}CB_\varpi^{1/2} B ϖ 1/2 C B ϖ 1/2 has no absolutely continuous spectrum on some nonempty interval
( ( 1 + w ) c M − Δ ^ w , ( 1 + w ) c M ) \left((1+w)c_M-\widehat{\Delta}_w,(1+w)c_M\right) ( ( 1 + w ) c M − Δ w , ( 1 + w ) c M ) almost surely, and its spectrum is [ 0 , ( 1 + w ) c M ] [0,(1+w)c_M] [ 0 , ( 1 + w ) c M ] almost surely.
Proof. For the first assertion, modify the block estimate by using X n = 1 { ξ n ≥ a 0 } X_n = \mathbf{1}_{\{\xi_n\ge a_0\}} X n = 1 { ξ n ≥ a 0 } , and take the starting scale so that t L ≤ g − a 0 t_L \le g-a_0 t L ≤ g − a 0 . The bound (83) is unchanged. A low eigenvalue therefore requires X B < q 0 / 2 X_B < q_0/2 X B < q 0 /2 , once h L < t L q 0 / 3 h_L < t_Lq_0/3 h L < t L q 0 /3 . For N = ∣ B ∣ N = |B| N = ∣ B ∣ and t 0 = 1 / 2 t_0 = 1/2 t 0 = 1/2 , the exponential Markov bound gives
P { ∑ n ∈ B X n < q 0 N / 2 } ≤ e t 0 q 0 N / 2 ( 1 − q 0 + q 0 e − t 0 ) N ≤ e − q 0 N / 8 . \mathbb{P}\left\{\sum_{n\in B}X_n < q_0N/2\right\} \le e^{t_0q_0N/2}(1-q_0+q_0e^{-t_0})^N \le e^{-q_0N/8}. P { n ∈ B ∑ X n < q 0 N /2 } ≤ e t 0 q 0 N /2 ( 1 − q 0 + q 0 e − t 0 ) N ≤ e − q 0 N /8 .
Indeed 1 − e − 1 / 2 > 3 / 8 1-e^{-1/2} > 3/8 1 − e − 1/2 > 3/8 and 1 − x ≤ e − x 1-x \le e^{-x} 1 − x ≤ e − x . Thus the same stretched-exponential initial estimate holds with changed constants. If the density is bounded by R R R , spectral averaging gives N R ∣ J 0 ∣ / g NR|J_0|/g N R ∣ J 0 ∣/ g in (84) and ( 2 N R / g ) t / ( 1 − t ) (2NR/g)^t/(1-t) ( 2 N R / g ) t / ( 1 − t ) in (85) . A bounded density with bounded support satisfies the regularity assumptions of the same fractional-moment criterion, with constants depending on the law. The all-subsets block argument gives the analogue of (92) . At fixed E E E in a sufficiently small upper-edge interval, Lemma 6.3 gives δ y ∈ Ran ( E M ϖ − J C J ) \delta_y \in\operatorname{Ran}(EM_\varpi-JCJ) δ y ∈ Ran ( E M ϖ − J C J ) almost surely. Since
J H ϖ J − E = − M ϖ − 1 / 2 ( E M ϖ − J C J ) M ϖ − 1 / 2 , JH_\varpi J-E=-M_\varpi^{-1/2}(EM_\varpi-JCJ)M_\varpi^{-1/2}, J H ϖ J − E = − M ϖ − 1/2 ( E M ϖ − J C J ) M ϖ − 1/2 ,
also δ y ∈ Ran ( J H ϖ J − E ) \delta_y \in\operatorname{Ran}(JH_\varpi J-E) δ y ∈ Ran ( J H ϖ J − E ) . For a bounded self-adjoint A A A , if δ y = ( A − E ) u \delta_y=(A-E)u δ y = ( A − E ) u , then ∥ ( A − E − i η ) − 1 δ y ∥ ≤ ∥ u ∥ \|(A-E-i\eta)^{-1}\delta_y\| \le\|u\| ∥ ( A − E − i η ) − 1 δ y ∥ ≤ ∥ u ∥ and Im ⟨ δ y , ( A − E − i η ) − 1 δ y ⟩ ≤ η ∥ u ∥ 2 → 0 \operatorname{Im}\langle\delta_y,(A-E-i\eta)^{-1}\delta_y\rangle\le\eta\|u\|^2 \to0 Im ⟨ δ y , ( A − E − i η ) − 1 δ y ⟩ ≤ η ∥ u ∥ 2 → 0 . Solvability of ( A − E ) u = δ y (A-E)u=\delta_y ( A − E ) u = δ y in ℓ 2 ( Z d ) \ell^2(\mathbb{Z}^d) ℓ 2 ( Z d ) is a measurable condition, since it is equivalent to
sup k ≥ 1 ∥ ( A − E − i / k ) − 1 δ y ∥ 2 < ∞ . \sup_{k\ge1}\left\|(A-E-i/k)^{-1}\delta_y\right\|_2<\infty. k ≥ 1 sup ( A − E − i / k ) − 1 δ y 2 < ∞.
Fubini and the Poisson boundary formula, as in Theorem 4.6, give the absence assertion. Finally, zero in the support ensures positive probability of the favourable events used in Proposition 2.2 .
For the direct uniform law, put b = 1 + w b=1+w b = 1 + w and define
η n = b / b n − 1 w , M ϖ ′ = diag ( 1 + w η n ) . \eta_n=\frac{b/b_n-1}{w},\qquad M'_\varpi=\operatorname{diag}(1+w\eta_n). η n = w b / b n − 1 , M ϖ ′ = diag ( 1 + w η n ) .
These variables are iid on [ 0 , 1 ] [0,1] [ 0 , 1 ] , with density
ρ n ( t ) = 1 + w ( 1 + w t ) 2 1 [ 0 , 1 ] ( t ) . \rho_n(t)=\frac{1+w}{(1+wt)^2}\mathbf{1}_{[0,1]}(t). ρ n ( t ) = ( 1 + w t ) 2 1 + w 1 [ 0 , 1 ] ( t ) .
It is bounded, has zero in its support, and gives P { η n ≥ 1 / 2 } = 1 / ( 2 + w ) > 0 \mathbb{P}\{\eta_n \ge1/2\} = 1/(2+w) > 0 P { η n ≥ 1/2 } = 1/ ( 2 + w ) > 0 . Since
B ω = b ( M ω ′ ) − 1 , B_\omega= b(M'_\omega)^{-1}, B ω = b ( M ω ′ ) − 1 ,
B ω 1 / 2 C B ω 1 / 2 = b ( M ω ′ ) − 1 / 2 C ( M ω ′ ) − 1 / 2 . B_\omega^{1/2} C B_\omega^{1/2} = b(M'_\omega)^{-1/2} C(M'_\omega)^{-1/2}. B ω 1/2 C B ω 1/2 = b ( M ω ′ ) − 1/2 C ( M ω ′ ) − 1/2 .
Apply the first assertion with disorder width w w w , and multiply the spectral energy by b b b . This proves the second assertion. □ \square □