betaFeedback

WorksPhysics and Astronomy

All coupling Anderson localization in the quantum kicked rotor model

Abstract

In this paper, we study H=λΔ+tan⁡π(θ+n2α)δn,n′, n∈Z\mathscr{H}=λΔ+\tanπ(θ+n^2α)δ_{n,n'},\ n\in\Bbb Z and prove its Anderson localization for all λ>0λ>0 and Lebesgue almost every (a.e.) (α,θ)∈T2(α, θ)\in\Bbb T^2. This conforms in the full measure sense the main assumption of Fishman-Grempel-Prange [Phys. Rev. Lett. 49(8), 509–512 (1982)] in the study of the quantum suppression of chaos: The pseudo-randomness of the sequence {tan⁡π(θ+n2α)}n∈Z\{\tanπ(θ+n^2α)\}_{n\in\Bbb Z} (for α∈T∖Qα\in\Bbb T\setminus \Bbb Q) should result in all coupling Anderson localization for HH similar to that for the one-dimensional Anderson model. Our proof is based on nonperturbative Green's function estimates approach using Cayley transform, quadratic Weyl sum estimate and Rellich function analysis.

Introduction and Main result

One of the main models in quantum chaos is the so-called 1-dimensional quantum kicked rotor which is given by the Floquet Hamiltonian

−1∂Ψ(x,t)∂t=(∂2∂x2+λcos⁡(2πx)∑n∈Zδ(t+2nα))Ψ(x,t).(1)\sqrt{-1}\frac{\partial\Psi(x,t)}{\partial t}=\left(\frac{\partial^{2}}{\partial x^{2}}+\lambda\cos(2\pi x)\sum_{n\in\mathbb{Z}}\delta(t+2n\alpha)\right)\Psi(x,t). \tag*{(1)}

where (x,t)∈T×R(x,t)\in\mathbb{T}\times\mathbb{R}, α∈T∖Q\alpha\in\mathbb{T}\setminus\mathbb{Q} and λ∈R\lambda\in\mathbb{R}. This model was first introduced by Casati et. al. [9] as a quantum analogue of the usual Chirikov standard map. It was conjectured by the remarkable work [12] of Fishman, Grempel and Prange that the motion in the quantum kicked rotor is generally almost-periodic even though it is typically chaotic in the classical one for large λ\lambda. This is the famous quantum suppression of chaos phenomenon. Indeed, in [12], by introducing an elegant transformation, the authors reduced the quantum kicked rotor to a discrete operator H\mathcal{H} of the form

(Hu)n=∑m∈Zϕn−mum+tan⁡π(θ+n2α)un,n∈Z, θ∈R,(\mathcal{H}u)_n=\sum_{m\in\mathbb{Z}}\phi_{n-m}u_m+\tan\pi\left(\theta+n^{2}\alpha\right)u_n,\quad n\in\mathbb{Z},\ \theta\in\mathbb{R},

where ϕn=−∫Ttan⁡π(λcos⁡(2πx)2)e−2π−1nx dx\phi_n=-\int_{\mathbb{T}}\tan\pi\left(\frac{\lambda\cos(2\pi x)}{2}\right)e^{-2\pi\sqrt{-1}nx}\,dx. Under this reduction, they further argued that the occurrence of the Anderson localization (i.e., pure point spectrum with exponentially decaying eigenfunctions) for H\mathcal{H} will imply the almost-periodicity of solutions for the quantum kicked rotor (1), and then the absence of chaos (cf. [12, 17]). The quantum suppression of chaos problem could eventually be solved [12] under the main assumption (supported by the numerical evidence) that the pseudo-randomness of the sequence {tan⁡π(θ+n2α)}n∈Z\{\tan\pi(\theta+n^{2}\alpha)\}_{n\in\mathbb{Z}} (for α∈T∖Q\alpha\in\mathbb{T}\setminus\mathbb{Q}) should result in all coupling Anderson localization for H\mathcal{H} similar to that for the one-dimensional Anderson model [1].

keywords: Anderson localization; Quantum kicked rotor; Cayley transform; Weyl sum estimate

In the present work, we aim to conform in the full measure sense this main assumption and proved all coupling Anderson localization for the nearest-neighbor version of H\mathcal{H} (also suggested by Fishman et al. [12]).

The popular mathematical models in the study of quantum kicked rotor problems are typically related to lattice Schrödinger operators with potentials given by the skew-shift:

H=λΔ+v(Tαn(θ1,θ2))δn,n′, n∈Z,\mathcal{H}=\lambda\Delta+v(\mathcal{T}_{\alpha}^{n}(\theta_{1},\theta_{2}))\delta_{n,n'},\ n\in\mathbb{Z},

where Δ\Delta is the discrete Laplacian1, vv is a real function2 on T2\mathbb{T}^{2} and T(θ1,θ2)=(θ1+θ2,θ2+α)\mathcal{T}(\theta_{1},\theta_{2})=(\theta_{1}+\theta_{2},\theta_{2}+\alpha). Bourgain-Goldstein-Schlag [4] first proved the Anderson localization for HH for nonconstant analytic vv and small enough λ\lambda provided α\alpha belonging to a set of large Lebesgue measure. The strategy is based on large deviation estimates on Green’s function and Lyapunov exponent analysis in the sense of [3, 13]. Based on multi-scale analysis type method related to [5], Bourgain [6] treated the more general skew-shift operators with Jacobi type hopping and established Anderson localization if the hopping strength is small enough. This leads to the resolution the problem of [Sin91, BDEF+94] that, for typical quantum kicked rotor models with small potentials, its wave function will be almost-periodic in time [6]. Recently, Shi-Yuan [SY21] modified the method of [4, 6] to prove the Anderson localization for skew-shift operators with the tangent potential assuming λ\lambda is small.

All the above localization results need the λ\lambda to be small enough. However, for the cosine skew-shift operators, it was conjectured by Bourgain that all coupling Anderson localization should be expected (cf. e.g., Chapter 15, [8]). For the cosine skew-shift model at a particular phase, Bourgain [7] has made a major breakthrough and proved the existence of point spectrum whose closure has positive measure for all λ≠0\lambda\ne0 and for α\alpha on a set of positive measure. To the best of our knowledge, the present work provides the first result on the all coupling Anderson localization for a skew-shift type self-adjoint Schrödinger operator.

Our proof differs from the approaches of Bourgain-Goldstein-Schlag [4] and Bourgain [6]. It relies on nonperturbative Green’s function estimates and is closer in spirit to the work of Jitomirskaya [14, 15]. More precisely, we exploit the special structure of the tangent potential to obtain, via the Cayley transform, a three-part decomposition of the Dirichlet determinants. One of our main contributions is to establish quantitative equidistribution estimates for the zeros of the relevant determinants using bounds on quadratic Weyl sums. These estimates lead to large-deviation estimates for Green’s functions for every coupling strength (λ≠0\lambda\ne0). Finally, we eliminate double resonances through a transversality analysis of Rellich functions. Combining the resulting Green’s function estimates with Schnol’s lemma then yields Anderson localization.

Main result

1.1. Main result. We study the following skew-shift Schrödinger operators

Hλ,θ,α=λΔ+tan⁡π(θ+n2α)δn,n′, n,n′∈Z,(2)\mathcal{H}_{\lambda,\theta,\alpha}=\lambda\Delta+\tan\pi(\theta+n^{2}\alpha)\delta_{n,n'},\ n,n'\in\mathbb{Z}, \tag*{(2)}

where λ>0\lambda>0, θ,α∈T:=R/Z\theta,\alpha\in\mathbb{T}:=\mathbb{R}/\mathbb{Z} and the discrete Laplacian Δ\Delta is defined as

Δ(n,n′)=δ∣n−n′∣,1.\Delta(n,n')=\delta_{|n-n'|,1}.

Let mes⁡(⋅)\operatorname{mes}(\cdot) denote the Lebesgue measure. Define the set of poles by

P=⋃n∈Z{(θ,α)∈T2:θ+n2α≡12(mod1)}.(3)\mathcal{P}=\bigcup_{n\in\mathbb{Z}}\left\{(\theta,\alpha)\in\mathbb{T}^{2}:\theta+n^{2}\alpha\equiv\frac{1}{2}\pmod{1}\right\}. \tag*{(3)}

Then we have

Theorem 1.1. For fixed λ>0\lambda>0, there is Gλ⊂T2∖P\mathcal{G}_{\lambda}\subset\mathbb{T}^{2}\setminus\mathcal{P} with mes⁡Gλ=1\operatorname{mes}\mathcal{G}_{\lambda}=1 such that, for every (θ,α)∈Gλ(\theta,\alpha)\in\mathcal{G}_{\lambda}, the operator (2) has pure point spectrum. Moreover, for each eigenvalue EE and corresponding eigenfunction ψ\psi, there is a constant C=C(λ,E,ψ,θ,α)>0C=C(\lambda,E,\psi,\theta,\alpha)>0 such that

∣ψ(n)∣≤Ce−mλ∣n∣,mλ=140arsinh⁡12λ>0,(4)|\psi(n)|\leq Ce^{-m_{\lambda}|n|},\qquad m_{\lambda}=\frac{1}{40}\operatorname{arsinh}\frac{1}{2\lambda}>0, \tag*{(4)}

where arsinh⁡x=log⁡(x+x2+1)\operatorname{arsinh}x=\log\left(x+\sqrt{x^{2}+1}\right) for x∈Rx\in\mathbb{R}.

Idea of the proof and new ingredients

Clearing the tangent denominators gives the Cayley factorization

det⁡(HN−EIN)=AN(E)QN(w;E)BN(w),w=e2πiθ.\det(\mathcal{H}_{N}-E\mathcal{I}_{N})=A_{N}(E)\frac{Q_{N}(w;E)}{B_{N}(w)},\qquad w=e^{2\pi i\theta}.

The roots of BNB_{N} are explicit, and quadratic Weyl sums control how evenly they are distributed on the unit circle. For QNQ_{N}, we truncate the free Cayley kernel and approximate the unitary trace Tr⁡(UE,Nk)\operatorname{Tr}(\mathcal{U}_{E,N}^{k}) by a sum over closed paths. The phase along each path has quadratic coefficient kαk\alpha in the starting point. Since the Weyl sum estimate is uniform, it controls the distribution of the roots of both BNB_{N} and QNQ_{N}.

We also show that a root of QNQ_{N} close to the phase ww would produce an approximate eigenvector at energy EE. Thus, if EE stays away from the finite-volume spectrum, then ww must stay away from every root of QNQ_{N}. By working with the matrix obtained after clearing the tangent denominators, this argument also covers roots at tangent poles.

In the cofactor formula, the BB factors cancel outside the interval joining the two points. The remaining BB and QQ factors contribute sublinear logarithmic errors, and the growth of ANA_{N} gives exponential Green function decay on nonresonant intervals.

Structure of the paper

The paper is organized as follows. After collecting the notation, we derive the Cayley factorization and estimate ANA_{N} in § 3. § 4 establishes bounds for BNB_{N} and QNQ_{N}. § 5 combines the three factor estimates to obtain Green function decay on nonresonant intervals. In § 6, we exclude double resonances and complete the proof of Theorem 1.1.

Appendix § 11 collects auxiliary estimates for the determinant factors. Appendix § 12 gives the details of phase-root separation and the pole count in the shear argument.

Notation

We collect the notation used throughout the paper.

  • Let i=−1i=\sqrt{-1} be the imaginary unit.

  • For x∈Rx\in\mathbb{R}, arsinh⁡x=log⁡(x+x2+1)\operatorname{arsinh}x=\log\left(x+\sqrt{x^{2}+1}\right).

  • The determinant of a matrix M\mathcal{M} is denoted by det⁡M\det\mathcal{M}.

  • For A,B⊂RA,B\subset\mathbb{R}, we define

dist⁡(A,B)=inf⁡{∣a−b∣:a∈A, b∈B}.\operatorname{dist}(A,B)=\inf\{|a-b|:a\in A,\ b\in B\}.
  • Let 1A\mathbf{1}_{\mathcal{A}} be the indicator function of set A\mathcal{A}.

  • Let #A\#\mathcal{A} denote the cardinality of set A\mathcal{A}.

  • For x∈Rx \in\mathbb{R}, we define ∥x∥T=inf⁡l∈Z∣l−x∣\lVert x\rVert_{\mathbb{T}} = \inf_{l\in\mathbb{Z}} \lvert l-x\rvert.

  • {δx}x∈Z\{\delta_x\}_{x\in\mathbb{Z}} is the standard basis of ℓ2(Z)\ell^2(\mathbb{Z}).

  • I\mathcal{I} typically denotes the identity operator.

  • For a linear operator T\mathcal{T}, let ∥T∥\lVert\mathcal{T}\rVert, ∥T∥S1\lVert\mathcal{T}\rVert_{\mathcal{S}_1}, rank⁡(T)\operatorname{rank}(\mathcal{T}) and Tr⁡(T)\operatorname{Tr}(\mathcal{T}) denote the operator norm, the trace norm, the rank and the trace of T\mathcal{T}, respectively,

  • PΛ\mathcal{P}_{\Lambda} denotes the orthogonal projection onto ℓ2(Λ)\ell^2(\Lambda), where Λ⊂Z\Lambda\subset\mathbb{Z}.

  • Denote by ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle the standard inner product on ℓ2(Z)\ell^2(\mathbb{Z}).

  • Let T:ℓ2(Z)→ℓ2(Z)\mathcal{T}:\ell^2(\mathbb{Z})\to\ell^2(\mathbb{Z}) be a linear operator. Set T(x,y)=⟨δx,Tδy⟩\mathcal{T}(x,y)=\langle\delta_x,\mathcal{T}\delta_y\rangle. The spectrum of T\mathcal{T} is denoted by σ(T)\sigma(\mathcal{T}). Finally, we define TΛ=PΛTPΛ\mathcal{T}_{\Lambda}=\mathcal{P}_{\Lambda}\mathcal{T}\mathcal{P}_{\Lambda}, where Λ⊂Z\Lambda\subset\mathbb{Z}. In particular, let ΛN=[0,N−1]∩Z\Lambda_N=[0,N-1]\cap\mathbb{Z} and we simplify TΛN\mathcal{T}_{\Lambda_N} to TN\mathcal{T}_N.

  • For E∉σ(HI)E\notin\sigma(\mathcal{H}_{\mathcal{I}}), we define its Green’s function by

GI(x,y;E)=⟨δx,(HI−EII)−1δy⟩,x,y∈I.\mathcal{G}_{\mathcal{I}}(x,y;E)=\left\langle\delta_x,(\mathcal{H}_{\mathcal{I}}-E\mathcal{I}_{\mathcal{I}})^{-1}\delta_y\right\rangle,\quad x,y\in\mathcal{I}.
  • For a finite Borel measure μ\mu on T\mathbb{T}, its Fourier coefficients are

μ^(k)=∫Te2πikt dμ(t),k∈Z\widehat{\mu}(k)=\int_{\mathbb{T}}e^{2\pi i k t}\,d\mu(t),\quad k\in\mathbb{Z}

and for f∈L1(T)f\in L^1(\mathbb{T}) we use

f^(k)=∫Tf(t)e−2πikt dt,k∈Z.\widehat{f}(k)=\int_{\mathbb{T}}f(t)e^{-2\pi i k t}\,dt,\quad k\in\mathbb{Z}.
  • We denote by BV(T)BV(\mathbb{T}) the space of functions f∈L1(T)f\in L^1(\mathbb{T}) with finite total variation. For f∈BV(T)f\in BV(\mathbb{T}), we write Var⁡T(f)\operatorname{Var}_{\mathbb{T}}(f) for the total variation.

Cayley factorization and the free determinant

Separating the three factors

We first remove the tangent denominators and identify the three factors that will be estimated below. On ΛN={0,…,N−1}\Lambda_N=\{0,\ldots,N-1\}, use the phase

qj=θ+βj+αj2.(5)q_j=\theta+\beta j+\alpha j^2. \tag*{(5)}

Locally, the diagonal entries of HN\mathcal{H}_N are given by tan⁡πqj\tan\pi q_j. For an interval starting at b∈Zb\in\mathbb{Z}, we write its points as b+jb+j. Then

θ+α(b+j)2=θb+βbj+αj2,\theta+\alpha(b+j)^2=\theta_b+\beta_bj+\alpha j^2,

where

θb=θ+b2α,βb=2bα.(6)\theta_b=\theta+b^2\alpha,\quad\beta_b=2b\alpha. \tag*{(6)}

For E∈RE\in\mathbb{R}, define

KE,N=λΔN−EIN,DN=Dα,β,N=diag⁡0≤j≤N−1(e2πi(αj2+βj)),(7)\mathcal{K}_{E,N}=\lambda\Delta_N-E\mathcal{I}_N,\quad\mathcal{D}_N=\mathcal{D}_{\alpha,\beta,N}=\operatorname{diag}_{0\leq j\leq N-1}\left(e^{2\pi i(\alpha j^2+\beta j)}\right), \tag*{(7)}

and

CE,N=(KE,N+iIN)−1(KE,N−iIN),UE,N=CE,NDN.(8)\mathcal{C}_{E,N}=(\mathcal{K}_{E,N}+i\mathcal{I}_N)^{-1}(\mathcal{K}_{E,N}-i\mathcal{I}_N),\quad\mathcal{U}_{E,N}=\mathcal{C}_{E,N}\mathcal{D}_N. \tag*{(8)}

Since KE,N\mathcal{K}_{E,N} is self-adjoint, both CE,N\mathcal{C}_{E,N} and UE,N\mathcal{U}_{E,N} are unitary. Set w=e2πiθw=e^{2\pi i\theta} and write

AN(E)=det⁡(KE,N+iIN),A_N(E)=\det(\mathcal{K}_{E,N}+i\mathcal{I}_N),
BN(w)=det⁡(IN+wDN),B_N(w)=\det(\mathcal{I}_N+w\mathcal{D}_N),
QN(w;E)=det⁡(IN+wUE,N).Q_N(w;E)=\det(\mathcal{I}_N+w\mathcal{U}_{E,N}).

Here ANA_N is the free determinant, BNB_N is the product of the tangent denominators, and the roots of QNQ_N are determined by the eigenvalues of UE,N\mathcal{U}_{E,N}.

To compare determinants on different subintervals, we keep the original integer points. For an interval I⊂Z\mathcal{I}\subset\mathbb{Z}, define

KE,I=λΔI−EII,\mathcal{K}_{E,\mathcal{I}}=\lambda\Delta_{\mathcal{I}}-E\mathcal{I}_{\mathcal{I}},
DI=diag⁡n∈I(e2πiαn2),\mathcal{D}_{\mathcal{I}}=\operatorname{diag}_{n\in\mathcal{I}}\left(e^{2\pi i\alpha n^2}\right),
CE,I=(KE,I+iII)−1(KE,I−iII),\mathcal{C}_{E,\mathcal{I}}=(\mathcal{K}_{E,\mathcal{I}}+i\mathcal{I}_{\mathcal{I}})^{-1}(\mathcal{K}_{E,\mathcal{I}}-i\mathcal{I}_{\mathcal{I}}),
UE,I=CE,IDI.\mathcal{U}_{E,\mathcal{I}}=\mathcal{C}_{E,\mathcal{I}}\mathcal{D}_{\mathcal{I}}.

We then set

AI(E)=det⁡(KE,I+iII),A_{\mathcal{I}}(E)=\det(\mathcal{K}_{E,\mathcal{I}}+i\mathcal{I}_{\mathcal{I}}),
BI(w)=det⁡(II+wDI),B_{\mathcal{I}}(w)=\det(\mathcal{I}_{\mathcal{I}}+w\mathcal{D}_{\mathcal{I}}),
QI(w;E)=det⁡(II+wUE,I).Q_{\mathcal{I}}(w;E)=\det(\mathcal{I}_{\mathcal{I}}+w\mathcal{U}_{E,\mathcal{I}}).

Thus all subintervals use the same phase variable w=e2πiθw=e^{2\pi i\theta}. We use the change of coordinates (6) only when applying estimates stated on ΛN\Lambda_N.

Lemma 3.1. Let w=e2πiθw=e^{2\pi i\theta}. If qj≢12(mod1)q_j\not\equiv\frac{1}{2}\pmod{1} for every j∈ΛNj\in\Lambda_N, where qjq_j is defined by (5), then

(HN−EIN)(IN+wDN)=(KE,N+iIN)+(KE,N−iIN)wDN(\mathcal{H}_N-E\mathcal{I}_N)(\mathcal{I}_N+w\mathcal{D}_N)=(\mathcal{K}_{E,N}+i\mathcal{I}_N)+(\mathcal{K}_{E,N}-i\mathcal{I}_N)w\mathcal{D}_N
=(KE,N+iIN)(IN+wUE,N).(9)=(\mathcal{K}_{E,N}+i\mathcal{I}_N)(\mathcal{I}_N+w\mathcal{U}_{E,N}). \tag*{(9)}

Hence

det⁡(HN−EIN)=AN(E)QN(w;E)BN(w).(10)\det(\mathcal{H}_N-E\mathcal{I}_N)=A_N(E)\frac{Q_N(w;E)}{B_N(w)}. \tag*{(10)}

The polynomial QN(⋅;E)Q_N(\cdot;E) has degree NN, and all its roots lie on the unit circle.

Proof. Using the identity

tan⁡πq=i1−e2πiq1+e2πiq,\tan\pi q=i\frac{1-e^{2\pi iq}}{1+e^{2\pi iq}},

we obtain

V(IN+wDN)=i(IN−wDN),\mathcal{V}(\mathcal{I}_N+w\mathcal{D}_N)=i(\mathcal{I}_N-w\mathcal{D}_N),

where V=diag⁡j∈ΛN(tan⁡πqj)\mathcal{V}=\operatorname{diag}_{j\in\Lambda_N}(\tan\pi q_j). From the fact that

HN−EIN=KN+V,\mathcal{H}_N-E\mathcal{I}_N=\mathcal{K}_N+\mathcal{V},

we have

(HN−EEN)(YN+wDN)=(KN+iIN)+V(IN+wDN)+wKNDN−iIN=(KN+iIN)+(KN−iIN)wDN=(KN+iIN)(IN+(KN+iIN)−1(KN−iIN)wDN)=(KE,N+iIN)(IN+wUE,N).\begin{aligned} (\mathcal{H}_N-E\mathcal{E}_N)(\mathcal{Y}_N+w\mathcal{D}_N)=(\mathcal{K}_N+i\mathcal{I}_N)+\mathcal{V}(\mathcal{I}_N+w\mathcal{D}_N)+w\mathcal{K}_N\mathcal{D}_N-i\mathcal{I}_N \\ &=(\mathcal{K}_N+i\mathcal{I}_N)+(\mathcal{K}_N-i\mathcal{I}_N)w\mathcal{D}_N \\ &=(\mathcal{K}_N+i\mathcal{I}_N)\left(\mathcal{I}_N+(\mathcal{K}_N+i\mathcal{I}_N)^{-1}(\mathcal{K}_N-i\mathcal{I}_N)w\mathcal{D}_N\right) \\ &=(\mathcal{K}_{E,N}+i\mathcal{I}_N)(\mathcal{I}_N+w\mathcal{U}_{E,N}). \end{aligned}

Taking determinants then yields (10). If u1,⋯ ,uNu_1,\cdots,u_N are the eigenvalues of UE,N\mathcal{U}_{E,N}, then

QN(w;E)=∏j=1N(1+wuj),∣uj∣=1.Q_N(w;E)=\prod_{j=1}^{N}(1+wu_j),\qquad|u_j|=1.

Its roots are therefore −uj−1-u_j^{-1}, and its leading coefficient det⁡UE,N\det\mathcal{U}_{E,N} is nonzero. We have finished the proof.

Remark 3.1. At a pole-free phase, a zero of QN(⋅;E)Q_N(\cdot;E) corresponds to an eigenvalue EE of HN\mathcal{H}_N. Even at a tangent pole, the matrix polynomial in (9) remains well-defined. We will use its kernel in Lemma 4.10 to compare the corresponding root with nearby pole-free phases.

Exponential growth of ANA_N

We begin with the phase-independent factor. Its explicit recurrence gives the positive exponent that will survive in the cofactor quotient.

For E∈RE\in\mathbb{R}, let ρE\rho_E denote the unique solution of

ρ+ρ−1=E−iλ(11)\rho+\rho^{-1}=\frac{E-i}{\lambda} \tag*{(11)}

satisfying ∣ρE∣>1|\rho_E|>1, and define

g(E,λ)=log⁡∣ρE∣.g(E,\lambda)=\log|\rho_E|.

Set

gλ=arsinh⁡12λ,J(λ)=log⁡21−e−2gλ.g_\lambda=\operatorname{arsinh}\frac{1}{2\lambda},\qquad J(\lambda)=\log\frac{2}{1-e^{-2g_\lambda}}.

Lemma 3.2. For n≥0n\geq0, we have

An(E)=(−λρE)n1−ρE−2n−21−ρE−2.(12)A_n(E)=(-\lambda\rho_E)^n\frac{1-\rho_E^{-2n-2}}{1-\rho_E^{-2}}. \tag*{(12)}

Moreover,

g(E,λ)≥gλ>0(13)g(E,\lambda)\geq g_\lambda>0 \tag*{(13)}

and

∣log⁡∣An(E)∣−n(log⁡λ+g(E,λ))∣≤J(λ).(14)\left|\log|A_n(E)|-n\left(\log\lambda+g(E,\lambda)\right)\right|\leq J(\lambda). \tag*{(14)}

Proof. Expanding the tridiagonal determinant along the last row gives

An=(−E+i)An−1−λ2An−2,A0=1,A1=−E+i.A_n=(-E+i)A_{n-1}-\lambda^2A_{n-2},\qquad A_0=1,\quad A_1=-E+i.

The two corresponding exponential solutions are (−λρE)n(-\lambda\rho_E)^n and (−λρE−1)n(-\lambda\rho_E^{-1})^n. Formula (12) satisfies the recurrence and the initial conditions. Writing ρE=eg+iφ\rho_E=e^{g+i\varphi}, we obtain from (11)

λ−1=2sinh⁡g ∣sin⁡φ∣≤2sinh⁡g,\lambda^{-1}=2\sinh g\,|\sin\varphi|\leq2\sinh g,

which yields (13). Moreover, we have

1−e−2gλ≤∣1−ρE−2n−2∣,∣1−ρE−2∣≤2.1-e^{-2g_\lambda}\leq|1-\rho_E^{-2n-2}|,\qquad|1-\rho_E^{-2}|\leq2.

Consequently, we can get

∣log⁡∣1−ρE−2n−2∣∣1−ρE2∣∣≤log⁡21−e−2gλ\left|\log\frac{\left|1-\rho_E^{-2n-2}\right|}{\left|1-\rho_E^2\right|}\right| \le\log\frac{2}{1-e^{-2g\lambda}}

which gives (14). This completes the proof. □

The factors BNB_N and QNQ_N

The free determinant supplies the linear term in the logarithmic Green function bound. We now show that the two factors contribute smaller errors, starting with the explicit roots of BNB_N and then passing to the spectral roots of QNQ_N.

Quadratic phases and the product BNB_N

The eigenvalue angles of DN\mathcal{D}_N have normalized counting measure

μD,N(A)=1N∑n=0N−11A(αn2+βn) for A⊂T.\mu_{D,N}(A)=\frac{1}{N}\sum_{n=0}^{N-1}\mathbf{1}_A(\alpha n^2+\beta n)\text{ for }A\subset\mathbb{T}.

The roots of BNB_N have angles 12−αn2−βn\frac{1}{2}-\alpha n^2-\beta n. For a probability measure μ\mu on T\mathbb{T}, define its discrepancy by

D(μ)=sup⁡J∣μ(J)−∣J∣∣,\mathcal{D}(\mu)=\sup_J\left|\mu(J)-|J|\right|,

where the supremum is over half-open arcs. Reflection and translation preserve discrepancy, so it suffices to estimate μD,N\mu_{D,N}. The Erdős–Turán inequality [11] gives, for every integer K≥1K\ge1,

D(μ)≤1K+1+2∑k=1K∣μ^(k)∣k.(15)\mathcal{D}(\mu)\le\frac{1}{K+1}+2\sum_{k=1}^{K}\frac{|\widehat{\mu}(k)|}{k}. \tag*{(15)}

In the present case,

μ^D,N(k)=1N∑n=0N−1e2πi(kαn2+kβn),\widehat{\mu}_{D,N}(k)=\frac{1}{N}\sum_{n=0}^{N-1}e^{2\pi i(k\alpha n^2+k\beta n)},

so we first need a quadratic Weyl bound uniform in the linear phase.

For 0<γ<1/20<\gamma<1/2, define

Dγ={α∈T:∥kα∥T≥γ∣k∣−2 for all k∈Z∖{0}}.(16)D_\gamma=\left\{\alpha\in\mathbb{T}:\|k\alpha\|_{\mathbb{T}}\ge\gamma|k|^{-2}\text{ for all }k\in\mathbb{Z}\setminus\{0\}\right\}. \tag*{(16)}

For each k≥1k\ge1, the exceptional set in (16) has measure 2γk−22\gamma k^{-2}. Hence, by the union bound, we obtain

mes⁡(T∖Dγ)≤2γ∑k≥1k−2=π23γ.(17)\operatorname{mes}(\mathbb{T}\setminus D_\gamma)\le2\gamma\sum_{k\ge1}k^{-2}=\frac{\pi^2}{3}\gamma. \tag*{(17)}

We will use the lower bound in (16) after taking differences of the quadratic phase. The following inequality controls the sum through its shifted correlations.

Lemma 4.1. If ∣zn∣≤1|z_n|\le1 for 0≤n<N0\le n<N and 1≤H≤N1\le H\le N, then

∣1N∑n=0N−1zn∣2≤2H+4NH∑h=1H−1∣∑n=0N−h−1zn+hzn‾∣.(18)\left|\frac{1}{N}\sum_{n=0}^{N-1}z_n\right|^2\le\frac{2}{H}+\frac{4}{NH}\sum_{h=1}^{H-1}\left|\sum_{n=0}^{N-h-1}z_{n+h}\overline{z_n}\right|. \tag*{(18)}

The proof is given in Appendix § 11.

Lemma 4.2. Let W(γ)=4+8/(3γ)W(\gamma)=\sqrt{4+8/(3\gamma)}. If α∈Dγ\alpha\in\mathcal{D}_{\gamma}, then, for every N≥1N\geq1 and 1≤k≤N1/161\leq k\leq N^{1/16},

sup⁡b∈R∣1N∑n=0N−1e2πi(kan2+bn)∣≤W(γ)N−1/8.(19)\sup_{b\in\mathbb{R}}\left|\frac{1}{N}\sum_{n=0}^{N-1}e^{2\pi i(kan^{2}+bn)}\right|\leq W(\gamma)N^{-1/8}. \tag*{(19)}

Proof. Set zn=e2πi(kan2+bn)z_n=e^{2\pi i(kan^{2}+bn)}. Then we have

zn+hzn‾=e2πi(kah2+bh)e2πi2kahn.z_{n+h}\overline{z_n}=e^{2\pi i(kah^{2}+bh)}e^{2\pi i2kahn}.

The geometric sum bound and (16) therefore give

∣∑n=0N−h−1zn+hzn‾∣≤min⁡{N,12∥2kahα∥}≤2γ−1k2h2.\left|\sum_{n=0}^{N-h-1}z_{n+h}\overline{z_n}\right|\leq\min\left\{N,\frac{1}{2\lVert2kah\alpha\rVert}\right\}\leq2\gamma^{-1}k^{2}h^{2}.

Substituting into (18), we obtain

∣1N∑n=0N−1zn∣2≤2H+8k2H23γN.\left|\frac{1}{N}\sum_{n=0}^{N-1}z_n\right|^{2}\leq\frac{2}{H}+\frac{8k^{2}H^{2}}{3\gamma N}.

Taking H=⌊N1/4⌋H=\lfloor N^{1/4}\rfloor and using k≤N1/16k\leq N^{1/16} gives

∣1N∑n=0N−1zn∣2≤(4+83γ)N−1/4.\left|\frac{1}{N}\sum_{n=0}^{N-1}z_n\right|^{2}\leq\left(4+\frac{8}{3\gamma}\right)N^{-1/4}.

Taking square roots proves (19). The proof has been finished.

For α∈Dγ\alpha\in\mathcal{D}_{\gamma}, taking K=⌊N1/16⌋K=\lfloor N^{1/16}\rfloor in (15) and using Lemma 4.2, we obtain

D(μD,N)≤DB(N):=N−1/16+2W(γ)(1+log⁡N)N−1/8.(20)\mathcal{D}(\mu_{D,N})\leq D_{B}(N):=N^{-1/16}+2W(\gamma)(1+\log N)N^{-1/8}. \tag*{(20)}

To turn this root count into a product bound, we use the bounded-variation estimate

∣∫Tf dμ−∫Tf(t) dt∣≤D(μ)Var⁡T(f),(21)\left|\int_{\mathbb{T}}f\,d\mu-\int_{\mathbb{T}}f(t)\,dt\right|\leq\mathcal{D}(\mu)\operatorname{Var}_{\mathbb{T}}(f), \tag*{(21)}

valid for f∈BV(T)∩C(T)f\in BV(\mathbb{T})\cap C(\mathbb{T}), which can be found in Appendix § 11.§ 2. The following lemma applies it to a truncated logarithm. We state both bounds here because the lower bound will also be needed for QNQ_N.

Lemma 4.3. Let ξ1,…,ξN∈T\xi_1,\ldots,\xi_N\in\mathbb{T}, and define their normalized counting measure

μN(A)=1N∑j=1N1A(ξj)for A⊂T.\mu_N(A)=\frac{1}{N}\sum_{j=1}^{N}\mathbf{1}_A(\xi_j)\quad\text{for }A\subset\mathbb{T}.

Suppose that μN\mu_N has discrepancy D(μN)≤D\mathcal{D}(\mu_N)\leq D, where 0<D≤1/20<D\leq1/2. For ∣w∣=1|w|=1, set

Q(w)=∏j=1N(1−we−2πiξj).Q(w)=\prod_{j=1}^{N}(1-we^{-2\pi i\xi_j}).

Then

log⁡∣Q(w)∣≤4NDlog⁡(2/D),(22)\log|Q(w)|\leq4ND\log(2/D), \tag*{(22)}

If ∣w−e2πiξj∣≥ϵ|w-e^{2\pi i\xi_j}|\geq\epsilon for every jj, where 0<ϵ≤10<\epsilon\leq1, then

log⁡∣Q(w)∣≥−4ND[log⁡(2/D)+log⁡(1/ϵ)].(23)\log|Q(w)|\geq-4ND\left[\log(2/D)+\log(1/\epsilon)\right]. \tag*{(23)}

Proof. Let w=e2πiθw=e^{2\pi i\theta} and define

μ~N(A)=μN(A+θ),A⊂T,\widetilde{\mu}_N(A)=\mu_N(A+\theta),\quad A\subset\mathbb{T},

then D(μ~N)=D(μN)\mathcal{D}(\widetilde{\mu}_N)=\mathcal{D}(\mu_N). For

ℓ(t)=log⁡∣1−e2πit∣=log⁡(2∣sin⁡πt∣),\ell(t)=\log\lvert1-e^{2\pi i t}\rvert=\log(2\lvert\sin\pi t\rvert),

we have

1Nlog⁡∣Q(w)∣=∫Tℓ dμ~N,∫Tℓ(t) dt=0.\frac{1}{N}\log\lvert Q(w)\rvert=\int_{\mathbb{T}}\ell\,d\widetilde{\mu}_N,\qquad\int_{\mathbb{T}}\ell(t)\,dt=0.

To apply the discrepancy estimate, we truncate the logarithmic singularity by setting

ℓh(t)=max⁡{ℓ(t),log⁡h},0<h≤1.\ell_h(t)=\max\{\ell(t),\log h\},\quad0<h\leq1.

By Appendix § 11.§ 3, we have

0≤∫Tℓh(t) dt≤h,Var⁡T(ℓh)=2log⁡(2/h).0\leq\int_{\mathbb{T}}\ell_h(t)\,dt\leq h,\quad\operatorname{Var}_{\mathbb{T}}(\ell_h)=2\log(2/h).

Thus by (21), we get

∣∫Tℓh dμ~N−∫Tℓh(t) dt∣≤2Dlog⁡(2/h).(24)\left\lvert\int_{\mathbb{T}}\ell_h\,d\widetilde{\mu}_N-\int_{\mathbb{T}}\ell_h(t)\,dt\right\rvert\leq2D\log(2/h). \tag*{(24)}

Taking h=Dh=D and using ℓ≤ℓh\ell\leq\ell_h, we obtain

1Nlog⁡∣Q(w)∣≤D+2Dlog⁡(2/D)≤4Dlog⁡(2/D).\frac{1}{N}\log\lvert Q(w)\rvert\leq D+2D\log(2/D)\leq4D\log(2/D).

If ∣w−e2πiξj∣≥ε\lvert w-e^{2\pi i\xi_j}\rvert\geq\varepsilon for every jj, take h=min⁡{D,ε}≥Dεh=\min\{D,\varepsilon\}\geq D\varepsilon. Then we have

ℓ(ξj−θ)=log⁡∣1−e2πi(ξj−θ)∣=log⁡∣w−e2πiξj∣≥log⁡ε≥log⁡h,\ell(\xi_j-\theta)=\log\lvert1-e^{2\pi i(\xi_j-\theta)}\rvert=\log\lvert w-e^{2\pi i\xi_j}\rvert\geq\log\varepsilon\geq\log h,

which implies ℓh(ξj−θ)=ℓ(ξj−θ)\ell_h(\xi_j-\theta)=\ell(\xi_j-\theta) for all jj. Hence, one has

1Nlog⁡∣Q(w)∣≥−2Dlog⁡(2/h)≥−2D[log⁡(2/D)+log⁡(1/ε)].\frac{1}{N}\log\lvert Q(w)\rvert\geq-2D\log(2/h)\geq-2D\left[\log(2/D)+\log(1/\varepsilon)\right].

This proves (22) and (23). We complete the proof. □

Applying (22) to the roots of BNB_N gives

log⁡∣BN(w)∣≤4NDB(N)log⁡2DB(N),DB(N)≤12.\log\lvert B_N(w)\rvert\leq4ND_B(N)\log\frac{2}{D_B(N)},\quad D_B(N)\leq\frac{1}{2}.

We will put it in a common form with the bound for QNQ_N below.

For the rest of the proof, fix

σ=132,a=1128,ν=1−σ+a=125128<1.(25)\sigma=\frac{1}{32},\quad a=\frac{1}{128},\quad\nu=1-\sigma+a=\frac{125}{128}<1. \tag*{(25)}

With discrepancy N−σN^{-\sigma} and root separation ce−Nace^{-Na}, the logarithmic error is O(Nν)O(N^\nu), where ν=1−σ+a<1\nu=1-\sigma+a<1.

The free Cayley factor in QNQ_N

For QNQ_N, the roots are determined by the eigenvalues e2πiξje^{2\pi i\xi_j} of UE,N\mathcal{U}_{E,N}. Put

μE,N(A)=1N∑j=1N1A(ξj)  for A⊂T,μ^E,N(k)=1NTr⁡(UE,Nk).\mu_{E,N}(A)=\frac{1}{N}\sum_{j=1}^{N}1_A(\xi_j)\ \text{ for } A\subset\mathbb{T},\qquad\widehat{\mu}_{E,N}(k)=\frac{1}{N}\operatorname{Tr}(\mathcal{U}_{E,N}^{k}).

Again the root angles are 12−ξj\frac{1}{2}-\xi_j. To use (15), we must therefore estimate these traces. We first replace the free Cayley factor by a convolution with short jumps, controlling the boundary and truncation errors.

Define

RE=(λΔ+(i−E)I)−1,CE=I−2iRE.(26)\mathcal{R}_E=(\lambda\Delta+(\mathrm{i}-E)\mathcal{I})^{-1},\qquad\mathcal{C}_E=\mathcal{I}-2\mathrm{i}\mathcal{R}_E. \tag*{(26)}

By the spectral theorem, ∥RE∥≤1\lVert\mathcal{R}_E\rVert\leq1 and CE\mathcal{C}_E is unitary.

Lemma 4.4. Let

μλ=log⁡(1+14λ)>0.(27)\mu_\lambda=\log\left(1+\frac{1}{4\lambda}\right)>0. \tag*{(27)}

Then we get

∣RE(n,m)∣≤2e−μλ∣n−m∣.(28)\lvert\mathcal{R}_E(n,m)\rvert\leq2e^{-\mu_\lambda\lvert n-m\rvert}. \tag*{(28)}

Consequently, CE(n,m)=cE(n−m)\mathcal{C}_E(n,m)=c_E(n-m), with ∣cE(r)∣≤4e−μλ∣r∣\lvert c_E(r)\rvert\leq4e^{-\mu_\lambda\lvert r\rvert} for $r\ne 0.

The decay estimate follows from an exponential-weight conjugation of the resolvent, which can be found in Appendix § 11. We next compare CE\mathcal{C}_E with its finite-volume counterpart.

Lemma 4.5. Regard PNCEPN\mathcal{P}_N\mathcal{C}_E\mathcal{P}_N as acting on ℓ2(ΛN)\ell^2(\Lambda_N). Then

∥CE,N−PNCEPN∥S1≤8.(29)\lVert\mathcal{C}_{E,N}-\mathcal{P}_N\mathcal{C}_E\mathcal{P}_N\rVert_{\mathcal{S}_1}\leq8. \tag*{(29)}

Consequently, for every k≥1k\geq1,

∣Tr⁡(UE,Nk)−Tr⁡((PNCEPNDNUN)k)∣≤8k.(30)\left\lvert\operatorname{Tr}(\mathcal{U}_{E,N}^{k})-\operatorname{Tr}\left((\mathcal{P}_N\mathcal{C}_E\mathcal{P}_N\mathcal{D}_N\mathcal{U}_N)^{k}\right)\right\rvert\leq8k. \tag*{(30)}

With P=PN\mathcal{P}=\mathcal{P}_N, Q=I−P\mathcal{Q}=\mathcal{I}-\mathcal{P}, and RE,N=(KE,N+iIN)−1\mathcal{R}_{E,N}=(\mathcal{K}_{E,N}+\mathrm{i}\mathcal{I}_N)^{-1}, we have

RE,N−PREP=RE,NPλΔQREP.(31)\mathcal{R}_{E,N}-\mathcal{P}\mathcal{R}_E\mathcal{P}=\mathcal{R}_{E,N}\mathcal{P}\lambda\Delta\mathcal{Q}\mathcal{R}_E\mathcal{P}. \tag*{(31)}

Since PΔQ\mathcal{P}\Delta\mathcal{Q} has rank at most two, the resolvent difference has rank at most two. The estimates (29)–(30) follow from the trace-norm bound and the telescoping identity, respectively, which can be found in Appendix § 11.§ 5.

Let CE[M]\mathcal{C}_E^{[M]} be the convolution operator with kernel cE(r)1{∣r∣≤M}c_E(r)1_{\{\lvert r\rvert\leq M\}}. By Lemma 4.4, we get

∥CE−CE[M]∥≤∑∣r∣>M∣cE(r)∣≤8e−μλMeμλ−1=32λe−μλM.\lVert\mathcal{C}_E-\mathcal{C}_E^{[M]}\rVert\leq\sum_{\lvert r\rvert>M}\lvert c_E(r)\rvert\leq\frac{8e^{-\mu_\lambda M}}{e^{\mu_\lambda}-1}=32\lambda e^{-\mu_\lambda M}.

Choose

MN=⌈12log⁡(N+2)+log⁡(32λ+1)μλ⌉,M_N=\left\lceil\frac{12\log(N+2)+\log(32\lambda+1)}{\mu_\lambda}\right\rceil,
A(λ)=12μλ+log⁡(32λ+1)μλlog⁡3+2log⁡3.A(\lambda)=\frac{12}{\mu_\lambda}+\frac{\log(32\lambda+1)}{\mu_\lambda\log3}+\frac{2}{\log3}.

Then, we obtain

MN+1≤A(λ)log⁡(N+2),M_N + 1 \le A(\lambda)\log(N+2),
∥EE−EE[MN]∥≤εN:=N−12,\|\mathcal{E}_E-\mathcal{E}_E^{[M_N]}\| \le\varepsilon_N := N^{-12},
∥EE[MN]∥≤1+εN.(32)\|\mathcal{E}_E^{[M_N]}\| \le1+\varepsilon_N. \tag*{(32)}

Traces and the roots of QNQ_N

The trace of UE,Nk\mathcal{U}_{E,N}^k is a sum over closed paths. The quadratic phase along each path separates into a common term depending on the starting point and a path-dependent remainder.

Fix M=MNM=M_N and define the diagonal operator Dα,β\mathcal{D}_{\alpha,\beta} by

Dα,β=e2πi(αn2+βn)δn,n′,n,n′∈Z.\mathcal{D}_{\alpha,\beta}=e^{2\pi i(\alpha n^2+\beta n)}\delta_{n,n'},\quad n,n'\in\mathbb{Z}.

Lemma 4.6. For each k≥1k\ge1, there is a trigonometric polynomial fkf_k of degree at most k2Mk^2M such that, for every n∈Zn\in\mathbb{Z},

[(EE[M]Dα,β)k](n,n)=e2πi(kαn2+kβn)fk(β+2αn).(33)\left[(\mathcal{E}_E^{[M]}\mathcal{D}_{\alpha,\beta})^k\right](n,n)=e^{2\pi i(k\alpha n^2+k\beta n)}f_k(\beta+2\alpha n). \tag*{(33)}

If N≥2N\ge2 and k≤N1/16k\le N^{1/16}, then

∑ℓ∣fk^(ℓ)∣≤4kM+1.(34)\sum_{\ell}|\widehat{f_k}(\ell)|\le4k\sqrt{M+1}. \tag*{(34)}

Proof. Write the visited points as n+b0,…,n+bkn+b_0,\ldots,n+b_k, where b0=bk=0b_0=b_k=0 and ∣bj−bj−1∣≤M|b_j-b_{j-1}|\le M. Then

α(n+bj)2+β(n+bj)=αn2+βn+(β+2αn)bj+αbj2.\alpha(n+b_j)^2+\beta(n+b_j)=\alpha n^2+\beta n+(\beta+2\alpha n)b_j+\alpha b_j^2.

Consequently, (33) holds with

fk(t)=∑b0=bk=0∣bj−bj−1∣≤M[∏j=1kcE(bj−1−bj)]e2πi(t∑j=1kbj+α∑j=1kbj2).(35)f_k(t)=\sum_{\substack{b_0=b_k=0\\ |b_j-b_{j-1}|\le M}}\left[\prod_{j=1}^{k}c_E(b_{j-1}-b_j)\right]e^{2\pi i\left(t\sum_{j=1}^{k}b_j+\alpha\sum_{j=1}^{k}b_j^2\right)}. \tag*{(35)}

Since ∣bj∣≤jM|b_j|\le jM, we have

∣∑j=1kbj∣≤k2M.\left|\sum_{j=1}^{k}b_j\right|\le k^2M.

Thus fkf_k has at most 2k2M+12k^2M+1 nonzero Fourier coefficients.

Moreover, we get

fk(t)=⟨δ0,(EE[M]Dα,t)kδ0⟩.f_k(t)=\langle\delta_0,(\mathcal{E}_E^{[M]}\mathcal{D}_{\alpha,t})^k\delta_0\rangle.

By (32), we obtain

∥fk∥∞≤(1+εN)k≤ekεN≤2.\|f_k\|_\infty\le(1+\varepsilon_N)^k\le e^{k\varepsilon_N}\le2.

From the Parseval’s identity, we have

∑ℓ∈Z∣fk^(ℓ)∣2=∥fk∥L2(T)2≤4.\sum_{\ell\in\mathbb{Z}}|\widehat{f_k}(\ell)|^2=\|f_k\|_{L^2(\mathbb{T})}^2\le4.

Hence, by Cauchy-Schwarz inequality, we obtain

∑ℓ∈Z∣fk^(ℓ)∣≤2k2M+1(∑ℓ∈Z∣fk^(ℓ)∣2)1/2≤22k2M+1≤4kM+1.\sum_{\ell\in\mathbb{Z}} \left|\widehat{f_k}(\ell)\right| \le\sqrt{2k^2M+1}\left(\sum_{\ell\in\mathbb{Z}}\left|\widehat{f_k}(\ell)\right|^2\right)^{1/2} \le2\sqrt{2k^2M+1}\le4k\sqrt{M+1}.

This proves (4.20). We have finished the proof. □

Each Fourier coefficient of fkf_k produces a quadratic Weyl sum with the same quadratic coefficient kαk\alpha and a modified linear coefficient.

Proposition 4.7. If α∈Dγ\alpha\in D_\gamma, then, for every N≥2N\ge2 and 1≤k≤N1/161\le k\le N^{1/16},

∣Tr⁡(UE,Nk)∣N≤4W(γ)A(λ)klog⁡(N+2)N−1/8+8A(λ)klog⁡(N+2)N+2kN−12.(36)\frac{\left|\operatorname{Tr}(\mathcal{U}_{E,N}^{k})\right|}{N}\le4W(\gamma)\sqrt{A(\lambda)}k\sqrt{\log(N+2)}N^{-1/8} +\frac{8A(\lambda)k\log(N+2)}{N}+2kN^{-12}. \tag*{(36)}

Proof. By Lemma 4.5 and the truncation estimate (4.18), together with the boundary path comparison in Appendix A.6, we have

∣Tr⁡(UE,Nk)−∑n=0N−1e2πi(kαn2+kβn)fk(β+2αn)∣≤8k+2NkεN+8kM.(37)\left|\operatorname{Tr}(\mathcal{U}_{E,N}^{k})-\sum_{n=0}^{N-1}e^{2\pi i(k\alpha n^2+k\beta n)}f_k(\beta+2\alpha n)\right|\le8k+2Nk\varepsilon_N+8kM. \tag*{(37)}

Expanding fkf_k in Fourier series, the normalized path sum becomes

∑ℓ∈Zfk^(ℓ)e2πiℓβ1N∑n=0N−1e2πi(kαn2+(kβ+2ℓα)n).\sum_{\ell\in\mathbb{Z}}\widehat{f_k}(\ell)e^{2\pi i\ell\beta}\frac{1}{N}\sum_{n=0}^{N-1}e^{2\pi i(k\alpha n^2+(k\beta+2\ell\alpha)n)}.

By Lemmas 4.2 and 4.6, its absolute value is bounded by

4W(γ)kM+1N−1/8.4W(\gamma)k\sqrt{M+1}N^{-1/8}.

Combining this with (4.23) and using

εN=N−12,M+1≤A(λ)log⁡(N+2),\varepsilon_N=N^{-12},\quad M+1\le A(\lambda)\log(N+2),

gives (4.22). This completes the proof. □

We can now apply (4.1) to μE,N\mu_{E,N}. The following common threshold also includes the quadratic phase measure considered above.

Corollary 4.8. Fix λ>0\lambda>0 and 0<γ<1/20<\gamma<1/2. Set

Cdisc(λ,γ)=5+8W(γ)A(λ)+16A(λ),C_{\mathrm{disc}}(\lambda,\gamma)=5+8W(\gamma)\sqrt{A(\lambda)}+16A(\lambda),
N0(λ,γ)=max⁡{232,⌈(32Cdisc(λ,γ))64⌉}.N_0(\lambda,\gamma)=\max\left\{2^{32},\left\lceil\left(\sqrt{32}C_{\mathrm{disc}}(\lambda,\gamma)\right)^{64}\right\rceil\right\}.

If α∈Dγ\alpha\in D_\gamma and N≥N0(λ,γ)N\ge N_0(\lambda,\gamma), then

D(μE,N),D(μD,N)≤N−σ.(38)\mathcal{D}(\mu_{E,N}),\mathcal{D}(\mu_{D,N})\le N^{-\sigma}. \tag*{(38)}

The root counting measures of QNQ_N and BNB_N satisfy the same discrepancy bound.

Proof. Taking K=⌊N1/16⌋K=\lfloor N^{1/16}\rfloor in (15) and using Proposition 4.7, we obtain

D(μE,N)≤N−1/16+8W(γ)M+1 N−1/16+16(M+1)N−15/16+4N−191/16≤Cdisc(λ,γ)N−1/16log⁡(N+2).\mathcal{D}(\mu_{E,N}) \le N^{-1/16}+8W(\gamma)\sqrt{M+1}\,N^{-1/16} +16(M+1)N^{-15/16}+4N^{-191/16} \le C_{\mathrm{disc}}(\lambda,\gamma)N^{-1/16}\sqrt{\log(N+2)}.

where we used (32) and 1≤log⁡(N+2)≤N1\le\log(N+2)\le N. By (19), the Fourier coefficients of μD,N\mu_{D,N} satisfy

∣μ^D,N(k)∣≤W(γ)N−1/8,1≤k≤K.|\widehat{\mu}_{D,N}(k)|\le W(\gamma)N^{-1/8},\qquad1\le k\le K.

The same application of (15) therefore gives the same discrepancy bound for μD,N\mu_{D,N}. Since log⁡(N+2)≤32N1/32\log(N+2)\le32N^{1/32}, we conclude that

D(μE,N),D(μD,N)≤32 Cdisc(λ,γ)N−3/64≤N−1/32,\mathcal{D}(\mu_{E,N}),\mathcal{D}(\mu_{D,N})\le\sqrt{32}\,C_{\mathrm{disc}}(\lambda,\gamma)N^{-3/64}\le N^{-1/32},

for N≥N0(λ,γ)N\ge N_0(\lambda,\gamma). The root counting measures of QNQ_N and BNB_N inherit the same bounds under ξ↦12−ξ\xi\mapsto\frac{1}{2}-\xi. We have concluded the proof. □

Upper and lower bounds for Q

We now apply Lemma 4.3 to QNQ_N and BNB_N, including intervals of arbitrary length.

Corollary 4.9. Fix λ>0\lambda>0 and 0<γ<1/20<\gamma<1/2, and suppose that α∈Dγ\alpha\in\mathcal{D}_\gamma. Let N0=N0(λ,γ)N_0=N_0(\lambda,\gamma) be as in Corollary 4.8, and define

FN0(n)=N0log⁡2+4n1−σlog⁡(n+2),n≥0.F_{N_0}(n)=N_0\log2+4n^{1-\sigma}\log(n+2),\qquad n\ge0.

Then, for every integer n≥0n\ge0, E∈RE\in\mathbb{R}, ∣w∣=1|w|=1, and β∈T\beta\in\mathbb{T},

log⁡∣Qn(w;E)∣, log⁡∣Bn(w)∣≤FN0(n),(39)\log|Q_n(w;E)|,\ \log|B_n(w)|\le F_{N_0}(n), \tag*{(39)}

where Q0=B0=1Q_0=B_0=1.

Proof. For n≥N0n\ge N_0, Lemma 4.3 with D=n−σD=n^{-\sigma} gives

log⁡∣Qn(w;E)∣, log⁡∣Bn(w)∣≤4n1−σlog⁡(2nσ)≤4n1−σlog⁡(n+2).\log|Q_n(w;E)|,\ \log|B_n(w)|\le4n^{1-\sigma}\log(2n^\sigma)\le4n^{1-\sigma}\log(n+2).

For 1≤n<N01\le n<N_0, the bound follows from ∣1+wu∣≤2|1+wu|\le2 for ∣w∣=∣u∣=1|w|=|u|=1. The case n=0n=0 is immediate. This concludes the proof. □

For a lower bound on QNQ_N, we need to separate ww from its roots. We use the cleared matrix in (9). Set

LE(z)=(KE,N+iIN)+(KE,N−iIN)zDN.\mathcal{L}_E(z)=(\mathcal{K}_{E,N}+i\mathcal{I}_N)+(\mathcal{K}_{E,N}-i\mathcal{I}_N)z\mathcal{D}_N.

Since

det⁡LE(z)=AN(E)QN(z;E),\det\mathcal{L}_E(z)=A_N(E)Q_N(z;E),

every root w0w_0 of QN(⋅;E)Q_N(\cdot;E) admits a unit vector v∈ker⁡LE(w0)v\in\ker\mathcal{L}_E(w_0). Setting u0=(IN+w0DN)vu_0=(\mathcal{I}_N+w_0\mathcal{D}_N)v, we obtain

(KE,N−iIN)u0=−2iv.(40)(\mathcal{K}_{E,N}-i\mathcal{I}_N)u_0=-2iv. \tag*{(40)}

In particular, u0≠0u_0\ne0, even when w0w_0 is a tangent pole. At the current pole-free phase ww, let u=(IN+wDN)vu=(\mathcal{I}_N+w\mathcal{D}_N)v. Subtracting LE(w0)v=0\mathcal{L}_E(w_0)v=0 from (9) applied to vv at ww, we obtain

(KN−EIN)u=(KE,N−iIN)(w−w0)DNv.(\mathcal{K}_N-E\mathcal{I}_N)u=(\mathcal{K}_{E,N}-i\mathcal{I}_N)(w-w_0)\mathcal{D}_Nv.

Moreover, since DN\mathcal{D}_N is unitary and ∥v∥=1\lVert v\rVert= 1, we have

∥u−u0∥=∣w−w0∣.\lVert u-u_0\rVert= \lvert w-w_0\rvert.

The residual is proportional to ∣w−w0∣\lvert w-w_0\rvert, while (40) controls ∥u0∥\lVert u_0\rVert from below.

Lemma 4.10. Suppose that ∣E∣≤E∗\lvert E\rvert\le E_\ast with E∗≥0E_\ast\ge0, and let C=1+E∗+2λC=1+E_\ast+2\lambda. If w=e2πiθw=e^{2\pi i\theta} is pole-free and

dist⁡(E,σ(HN))≥δ,0<δ≤1,\operatorname{dist}(E,\sigma(\mathcal{H}_N))\ge\delta,\qquad0<\delta\le1,

then every root w0w_0 of QN(⋅;E)Q_N(\cdot;E) satisfies

∣w−w0∣≥C−2δ.(41)\lvert w-w_0\rvert\ge C^{-2}\delta. \tag*{(41)}

This includes roots that are also zeros of BNB_N.

The proof is given in Appendix § 12.§ 1. In particular, the bound is independent of the size of the tangent potential at the current phase.

If N≥N0N\ge N_0, α∈Dγ\alpha\in\mathcal{D}_\gamma, ∣E∣≤E∗\lvert E\rvert\le E_\ast and

dist⁡(E,σ(HN))≥e−Nα,\operatorname{dist}(E,\sigma(\mathcal{H}_N))\ge e^{-N^\alpha},

then Lemma 4.10 gives

∣w−w0∣≥C−2e−Nα\lvert w-w_0\rvert\ge C^{-2}e^{-N^\alpha}

for every root w0w_0 of QN(⋅;E)Q_N(\cdot;E). Applying (23) with D=N−σD=N^{-\sigma} and ε=C−2e−Nα\varepsilon=C^{-2}e^{-N^\alpha}, we obtain

log⁡∣QN(w;E)∣≥−4N1−σ[Nα+log⁡(N+2)+2log⁡C]≥−(524+8log⁡C)Nν.(42)\begin{aligned} \log\lvert Q_N(w;E)\rvert\ge-4N^{1-\sigma}\left[N^\alpha+\log(N+2)+2\log C\right] \\ &\ge-(524+8\log C)N^\nu. \tag*{(42)} \end{aligned}

Here we used log⁡(N+2)<130Nα\log(N+2)<130N^\alpha.

Green function decay

The cofactor formula combines the three determinant factors. The BB factors require only upper bounds, whereas QQ requires both upper and lower bounds.

Let I=[b,c]∩Z\mathcal{I}=[b,c]\cap\mathbb{Z} and x≤yx\le y in I\mathcal{I}. Write d=y−xd=y-x, I−=[b,x−1]∩Z\mathcal{I}_-=[b,x-1]\cap\mathbb{Z}, and I+=[y+1,c]∩Z\mathcal{I}_+=[y+1,c]\cap\mathbb{Z}. Direct computation shows

GI(x,y;E)=(−λ)ddet⁡(HI−−EII−)det⁡(HI+−EII+)det⁡(HI−EII).(43)G_{\mathcal{I}}(x,y;E)=(-\lambda)^d\frac{\det(\mathcal{H}_{\mathcal{I}_-}-E\mathcal{I}_{\mathcal{I}_-})\det(\mathcal{H}_{\mathcal{I}_+}-E\mathcal{I}_{\mathcal{I}_+})}{\det(\mathcal{H}_{\mathcal{I}}-E\mathcal{I}_{\mathcal{I}})}. \tag*{(43)}

Applying (10) to the three determinants with the same phase ww, the B\mathcal{B} factors cancel outside [x,y][x,y], leaving

BI(w)BI−(w)BI+(w)=∏j=xy(1+e2πi(θ+αj2)).(44)\frac{B_{\mathcal{I}}(w)}{B_{\mathcal{I}_-}(w)B_{\mathcal{I}_+}(w)} =\prod_{j=x}^{y}\left(1+e^{2\pi i(\theta+\alpha j^2)}\right). \tag*{(44)}

The remaining QQ factors are QI−Q_{\mathcal{I}_-} and QI+Q_{\mathcal{I}_+} in the numerator, and QIQ_{\mathcal{I}} in the denominator.

Proposition 5.1. Fix λ>0\lambda>0, γ>0\gamma>0, and 0≤E∗<∞0\le E_\ast<\infty. Take N0N_0 from Corollary 4.8 and CC from Lemma 4.10. Define

M(λ,γ,E∗)=2084+8log⁡C+3J(λ)+3N0log⁡2,M(\lambda,\gamma,E_\ast)=2084+8\log C+3J(\lambda)+3N_0\log2,
N1(λ,γ,E∗)=max⁡{N0,⌈(20M(λ,γ,E∗)gλ)1/(1−ν)⌉}.N_1(\lambda,\gamma,E_\ast)=\max\left\{N_0,\left\lceil\left(\frac{20M(\lambda,\gamma,E_\ast)}{g_\lambda}\right)^{1/(1-\nu)}\right\rceil\right\}.

If α∈Dγ\alpha\in\mathcal{D}_{\gamma} and the potential is finite on a finite interval I⊂Z\mathcal{I} \subset\mathbb{Z} of length N≥N1N \ge N_{1} satisfies

dist⁡(E,σ(HI))≥e−Na,∣E∣≤E∗,(45)\operatorname{dist}\left(E,\sigma(\mathcal{H}_{\mathcal{I}})\right) \ge e^{-N^{a}}, \qquad|E| \le E_{*}, \tag*{(45)}

then

∣GI(x,y;E)∣≤e−12gλ∣x−y∣,∣x−y∣≥N/10.(46)|\mathcal{G}_{\mathcal{I}}(x,y;E)| \le e^{-\frac{1}{2}g_{\lambda}|x-y|}, \qquad|x-y| \ge N/10. \tag*{(46)}

Proof. By (6), the local estimates apply to I\mathcal{I}. By symmetry, assume x≤yx \le y, and put d=y−xd=y-x and n±=#I±n_{\pm}=\#\mathcal{I}_{\pm}. If the numerator in (43) vanishes, the claim is immediate. Otherwise, apply the Cayley factorization to (43).

For the free factors, write

rn=log⁡∣An(E)∣−n(log⁡λ+g(E,λ)),∣rn∣≤J(λ).r_n=\log|A_n(E)|-n\bigl(\log\lambda+g(E,\lambda)\bigr), \qquad|r_n|\le J(\lambda).

Since n−+n+=N−d−1n_{-}+n_{+}=N-d-1, Lemma 3.2 gives

dlog⁡λ+log⁡∣An−(E)∣+log⁡∣An+(E)∣−log⁡∣AN(E)∣=−log⁡λ−(d+1)g(E,λ)+rn−+rn+−rN≤−gλd+3J(λ).\begin{aligned} d\log\lambda+\log|A_{n_{-}}(E)|+\log|A_{n_{+}}(E)|-\log|A_N(E)| &=-\log\lambda-(d+1)g(E,\lambda)+r_{n_{-}}+r_{n_{+}}-r_N\\ &\le-g_{\lambda}d+3J(\lambda). \end{aligned}

since ∣rn∣≤J(λ)|r_n|\le J(\lambda) and λegλ=1+1+4λ22≥1\lambda e^{g_{\lambda}}=\frac{1+\sqrt{1+4\lambda^2}}{2}\ge1. For the BB factors, (44) and Corollary 4.9 yield

log⁡∣BI(w)BI−(w)BI+(w)∣≤N0log⁡2+4N1−σlog⁡(N+2).\log\left|\frac{B_{\mathcal{I}}(w)}{B_{\mathcal{I}_{-}}(w)B_{\mathcal{I}_{+}}(w)}\right| \le N_0\log2+4N^{1-\sigma}\log(N+2).

For the QQ factors, Corollary 4.9 gives

log⁡∣QI−(w;E)∣+log⁡∣QI+(w;E)∣≤2N0log⁡2+8N1−σlog⁡(N+2).\log|Q_{\mathcal{I}_{-}}(w;E)|+\log|Q_{\mathcal{I}_{+}}(w;E)| \le2N_0\log2+8N^{1-\sigma}\log(N+2).

The denominator is bounded from below by (42). Consequently, the total contribution of the three factors is at most

3N0log⁡2+12N1−σlog⁡(N+2)+(524+8log⁡C)Nν≤(2084+8log⁡C+3N0log⁡2)Nν,\begin{aligned} 3N_0\log2+12N^{1-\sigma}\log(N+2)+(524+8\log C)N^{\nu} &\le(2084+8\log C+3N_0\log2)N^{\nu}, \end{aligned}

where we used log⁡(N+2)<130Na\log(N+2)<130N^{a} and ν=1−σ+a\nu=1-\sigma+a.

Combining these estimates with the free determinant bound gives

log⁡∣GI(x,y;E)∣≤−gλd+M(λ,γ,E∗)Nν.(47)\log|\mathcal{G}_{\mathcal{I}}(x,y;E)|\le-g_{\lambda}d+M(\lambda,\gamma,E_{*})N^{\nu}. \tag*{(47)}

By the choice of N1N_{1}, we have

M(λ,γ,E∗)Nν≤gλN20≤gλd2.M(\lambda,\gamma,E_{*})N^{\nu}\le\frac{g_{\lambda}N}{20}\le\frac{g_{\lambda}d}{2}.

Hence (47) yields (46). The proof is now complete. □

From spectral separation to localization

Excluding double resonances

Since a nonzero solution is uniquely determined by its values at two consecutive sites, at least one of ψ(0)\psi(0) and ψ(1)\psi(1) must be nonzero. We choose c∈{0,1}c \in\{0,1\} so that ψ(c)≠0\psi(c) \ne0. For each sufficiently distant RR, set N=⌊∣R∣/4⌋≥2N = \lfloor|R|/4 \rfloor\ge2. Define

AN={±(4N+j):j=0,1,2,3},(48)\mathcal{A}_N = \{\pm(4N+j):j=0,1,2,3\}, \tag*{(48)}
Ic,N=[c−N,c+N]∩Z,c∈{0,1},\mathcal{I}_{c,N}=[c-N,c+N]\cap\mathbb{Z},\quad c\in\{0,1\},
JR,N=[R−N,R+N]∩Z,R∈AN.(49)\mathcal{J}_{R,N}=[R-N,R+N]\cap\mathbb{Z},\quad R\in\mathcal{A}_N. \tag*{(49)}

If

dist⁡(E,σ(HIc,N))<τ,dist⁡(E,σ(HJR,N))<τ,\operatorname{dist}(E,\sigma(\mathcal{H}_{\mathcal{I}_{c,N}}))<\tau,\quad\operatorname{dist}(E,\sigma(\mathcal{H}_{\mathcal{J}_{R,N}}))<\tau,

then

dist⁡(σ(HIc,N),σ(HJR,N))<2τ.\operatorname{dist}\bigl(\sigma(\mathcal{H}_{\mathcal{I}_{c,N}}),\sigma(\mathcal{H}_{\mathcal{J}_{R,N}})\bigr)<2\tau.

Thus a separation estimate for the two finite-volume spectra excludes simultaneous resonances.

Lemma 6.1. For λ>0\lambda>0, N≥2N\ge2, c∈{0,1}c\in\{0,1\}, R∈ANR\in\mathcal{A}_N, and δ>0\delta>0,

mes⁡{(θ,α)∉P:dist⁡(σ(HIc,N),σ(HJR,N))<δ}≤200N3δ.(50)\operatorname{mes}\{(\theta,\alpha)\notin\mathcal{P}:\operatorname{dist}(\sigma(\mathcal{H}_{\mathcal{I}_{c,N}}),\sigma(\mathcal{H}_{\mathcal{J}_{R,N}}))<\delta\}\le200N^{3}\delta. \tag*{(50)}

The bound is uniform in λ\lambda and requires no arithmetic condition on α\alpha.

Proof. The potential increases with θ\theta on Ic,N\mathcal{I}_{c,N} and JR,N\mathcal{J}_{R,N}. To obtain opposite eigenvalue motion, we vary θ\theta and α\alpha simultaneously. Since n2≤(N+1)2n^{2}\le(N+1)^{2} for n∈Ic,Nn\in\mathcal{I}_{c,N} and n2≥9N2n^{2}\ge9N^{2} for n∈JR,Nn\in\mathcal{J}_{R,N}, choose 4N24N^{2} between these bounds and set

θ=ϑ−4N2t(mod1),α=t.(51)\theta=\vartheta-4N^{2}t\pmod1,\quad\alpha=t. \tag*{(51)}

The phase at point n∈Zn\in\mathbb{Z} becomes ϑ+snt\vartheta+s_nt, where sn=n2−4N2s_n=n^{2}-4N^{2}. Thus, we get

sn≤−N2,n∈Ic,N,(52)s_n\le-N^{2},\quad n\in\mathcal{I}_{c,N}, \tag*{(52)}
sn≥5N2,n∈JR,N.(53)s_n\ge5N^{2},\quad n\in\mathcal{J}_{R,N}. \tag*{(53)}

On each pole-free interval in tt, let ηj(t)\eta_j(t) and ζk(t)\zeta_k(t) denote the analytic eigenvalue branches of HIc,N\mathcal{H}_{\mathcal{I}_{c,N}} and HJR,N\mathcal{H}_{\mathcal{J}_{R,N}}, respectively. By Rellich's theorem and the Feynman–Hellmann formula,

ηj′(t)=π∑n∈Ic,N∣ψj(n)∣2snsec⁡2π(ϑ+snt),\eta_j'(t)=\pi\sum_{n\in\mathcal{I}_{c,N}}|\psi_j(n)|^{2}s_n\sec^{2}\pi(\vartheta+s_nt),
ζk′(t)=π∑n∈JR,N∣ϕk(n)∣2snsec⁡2π(ϑ+snt),\zeta_k'(t)=\pi\sum_{n\in\mathcal{J}_{R,N}}|\phi_k(n)|^{2}s_n\sec^{2}\pi(\vartheta+s_nt),

where ψj\psi_j and ϕk\phi_k are normalized eigenvectors. Consequently, we obtain

ηj′(t)≤−πN2,ζk′(t)≥5πN2,\eta_j'(t)\le-\pi N^{2},\quad\zeta_k'(t)\ge5\pi N^{2},

and hence

(ζk−ηj)′(t)≥6πN2.(\zeta_k-\eta_j)'(t)\ge6\pi N^{2}.

Hence, by the mean value theorem,

mes⁡{t:∣ζk(t)−ηj(t)∣<δ}≤δ3πN2.\operatorname{mes}\{t:|\zeta_k(t)-\eta_j(t)|<\delta\}\le\frac{\delta}{3\pi N^{2}}.

Since sn∈Z∖{0}s_n \in\mathbb{Z} \setminus\{0\}, the map t↦snt(mod1)t \mapsto s_n t \pmod{1} is an ∣sn∣|s_n|-to-one covering of T\mathbb{T}. Hence each point nn contributes exactly ∣sn∣|s_n| poles. Each block has length L=2N+1L=2N+1. Summing over the L2L^2 eigenvalue pairs and all pole-free intervals, and integrating in ϑ\vartheta, gives (50), which can be found in Appendix § 12.§ 2. We close the proof here. □

Set

τN=e−(2N+1)a.(54)\tau_N=e^{-(2N+1)a}. \tag*{(54)}

Let RNλ\mathcal{R}_N^\lambda be the union of the events

⋃c∈{0,1}R∈AN{(θ,α)∉P:dist⁡(σ(HIc,N),σ(HJR,N))<2τN}.\bigcup_{\substack{c\in\{0,1\}\\ R\in\mathcal{A}_N}}\left\{(\theta,\alpha)\notin\mathcal{P}:\operatorname{dist}\left(\sigma(\mathcal{H}_{I_{c,N}}),\sigma(\mathcal{H}_{J_{R,N}})\right)<2\tau_N\right\}.

Since there are 16 pairs (c,R)(c,R), Lemma 6.1 gives

mes⁡(RNλ)≤6400N3e−(2N+1)a,∑N≥2mes⁡(RNλ)<∞.(55)\operatorname{mes}(\mathcal{R}_N^\lambda)\leq6400N^3e^{-(2N+1)a},\qquad\sum_{N\geq2}\operatorname{mes}(\mathcal{R}_N^\lambda)<\infty. \tag*{(55)}

Proposition 6.2 (An energy-independent set of good parameters). Let

Gλ=[T×⋃j≥3D1/j]∖(P∪lim sup⁡N→∞RNλ).(56)\mathcal{G}_\lambda=\left[\mathbb{T}\times\bigcup_{j\geq3}D_{1/j}\right]\setminus\left(\mathcal{P}\cup\limsup_{N\to\infty}\mathcal{R}_N^\lambda\right). \tag*{(56)}

Then mes⁡Gλ=1\operatorname{mes}\mathcal{G}_\lambda=1. Moreover, for every (θ,α)∈Gλ(\theta,\alpha)\in\mathcal{G}_\lambda, there exists N2N_2 such that, for all N≥N2N\geq N_2, c∈{0,1}c\in\{0,1\}, and R∈ANR\in\mathcal{A}_N,

dist⁡(σ(HIc,N),σ(HJR,N))≥2τN.(57)\operatorname{dist}\left(\sigma(\mathcal{H}_{I_{c,N}}),\sigma(\mathcal{H}_{J_{R,N}})\right)\geq2\tau_N. \tag*{(57)}

Proof. By (17), we have

mes⁡(⋃j≥3D1/j)=lim⁡j→∞mes⁡(D1/j)=1.\operatorname{mes}\left(\bigcup_{j\geq3}D_{1/j}\right)=\lim_{j\to\infty}\operatorname{mes}(D_{1/j})=1.

By (55) and the Borel-Cantelli Lemma,

mes⁡(lim sup⁡N→∞RNλ)=0.\operatorname{mes}\left(\limsup_{N\to\infty}\mathcal{R}_N^\lambda\right)=0.

Since mes⁡(P)=0\operatorname{mes}(\mathcal{P})=0, the first assertion follows. The second follows directly from the definition of Gλ\mathcal{G}_\lambda. This concludes the proof. □

In particular, if

dist⁡(E,σ(HIc,N))<τN,\operatorname{dist}(E,\sigma(\mathcal{H}_{I_{c,N}}))<\tau_N,

then

dist⁡(E,σ(HJR,N))>τN,R∈AN.\operatorname{dist}(E,\sigma(\mathcal{H}_{J_{R,N}}))>\tau_N,\qquad R\in\mathcal{A}_N.

We next show that every nonzero polynomially bounded solution produces such a central resonance.

Proof of Theorem 1.1

As an application of Green’s function estimates of the previous section, we prove Theorem 1.1 below.

Proof of Theorem 1.1. Fix λ>0\lambda> 0 and (θ,α)∈Gλ(\theta,\alpha) \in\mathcal{G}_{\lambda}. Let E∈RE \in\mathbb{R} and let ψ≠0\psi\ne0 be a polynomially bounded solution of

Hλ,θ,αψ=Eψ,(58)\mathcal{H}_{\lambda,\theta,\alpha}\psi= E\psi, \tag*{(58)}

with

∣ψ(n)∣≤(1+∣n∣)2.|\psi(n)| \le(1+|n|)^2.

Choose γ>0\gamma> 0 such that α∈Dγ\alpha\in\mathcal{D}_{\gamma}, and fix E∗≥∣E∣E_* \ge|E|. Since λ>0\lambda> 0, the recurrence determines ψ\psi from its values at two consecutive points. Thus, there exists c∈{0,1}c \in\{0,1\} such that ψ(c)≠0\psi(c) \ne0. For all sufficiently large NN, Propositions 5.1 and 6.2 apply. We first claim that

dist⁡(E,σ(HIc,N))<τN.(59)\operatorname{dist}(E,\sigma(\mathcal{H}_{I_{c,N}})) < \tau_N. \tag*{(59)}

Indeed, otherwise the Green’s function estimate of Proposition 5.1 and the Poisson formula give

∣ψ(c)∣≤λe−gλN/2(∣ψ(c−N−1)∣+∣ψ(c+N+1)∣)|\psi(c)| \le\lambda e^{-g_{\lambda}N/2}\left(|\psi(c-N-1)|+|\psi(c+N+1)|\right)
≤2λ(N+3)2e−gλN/2,\le2\lambda(N+3)^2e^{-g_{\lambda}N/2},

which tends to zero as N→∞N \to\infty, contradicting ψ(c)≠0\psi(c) \ne0. Now let R∈ZR \in\mathbb{Z} be sufficiently large and set N=⌊∣R∣/4⌋N=\lfloor|R|/4\rfloor, so that R∈ANR \in\mathcal{A}_N. By (59) and Proposition 6.2, we can get

dist⁡(E,σ(HJR,N))≥τN.\operatorname{dist}(E,\sigma(\mathcal{H}_{J_{R,N}})) \ge\tau_N.

Another application of Proposition 5.1 and the Poisson formula yields

∣ψ(R)∣≤2λ(∣R∣+N+2)2e−gλN/2.(60)|\psi(R)| \le2\lambda(|R|+N+2)^2e^{-g_{\lambda}N/2}. \tag*{(60)}

Since N≥(∣R∣−3)/4N \ge(|R|-3)/4, we obtain

∣ψ(R)∣≤C(λ)e−gλ∣R∣/40.|\psi(R)| \le C(\lambda)e^{-g_{\lambda}|R|/40}.

Enlarging CC to cover the remaining points gives (4), with

mλ=gλ40.m_{\lambda}=\frac{g_{\lambda}}{40}.

This completes the proof of Theorem 1.1. □

Acknowledgments

This work was partially supported by NSFC (12522110).

Data Availability

The manuscript has no associated data.

Declarations

Conflicts of interest The authors state that there is no conflict of interest.

AI statement

The authors declare that the main ideas of this paper are their own. ChatGPT 6.1 sol (OpenAI) was used to assist with literature searches, language editing, checking calculations and mathematical arguments. The authors reviewed and verified all AI-assisted material included in the paper and take full responsibility for the accuracy, originality, and integrity of the entire paper.

Auxiliary estimates for factors

Proof of Lemma 4.1

Proof. Average HH translates of the same sum before squaring. This keeps the differences below HH, where the arithmetic bounds are useful. Extend znz_n by zero, put S=∑n∈ZznS=\sum_{n\in\mathbb{Z}}z_n, and observe that

HS=∑m∈Z∑r=0H−1zm+r.HS=\sum_{m\in\mathbb{Z}}\sum_{r=0}^{H-1}z_{m+r}.

Let Xm=∑r=0H−1zm+rX_m=\sum_{r=0}^{H-1}z_{m+r}. At most N+H−1≤2NN+H-1\le2N of these sums are nonzero. Cauchy–Schwarz inequality gives

H2∣S∣2=∣∑m∈ZXm∣2≤#{m∈Z:Xm≠0}∑m∈Z∣Xm∣2≤2N∑m∈Z∣∑r=0H−1zm+r∣2.H^2|S|^2=\left|\sum_{m\in\mathbb{Z}}X_m\right|^2\le\#\{m\in\mathbb{Z}:X_m\ne0\}\sum_{m\in\mathbb{Z}}|X_m|^2\le2N\sum_{m\in\mathbb{Z}}\left|\sum_{r=0}^{H-1}z_{m+r}\right|^2.

Expanding the square and grouping the terms by h=r−sh=r-s yields

∑m∈Z∣∑r=0H−1zm+r∣2=H∑n∈Z∣zn∣2+2Re⁡∑h=1H−1(H−h)∑n=0N−h−1zn+hzn‾.\sum_{m\in\mathbb{Z}}\left|\sum_{r=0}^{H-1}z_{m+r}\right|^2=H\sum_{n\in\mathbb{Z}}|z_n|^2+2\operatorname{Re}\sum_{h=1}^{H-1}(H-h)\sum_{n=0}^{N-h-1}z_{n+h}\overline{z_n}.

Since ∑n∈Z∣zn∣2≤N\sum_{n\in\mathbb{Z}}|z_n|^2\le N and H−h≤HH-h\le H, we have

H2∣S∣2≤2N[HN+2H∑h=1H−1∣∑n=0N−h−1zn+hzn‾∣].H^2|S|^2\le2N\left[HN+2H\sum_{h=1}^{H-1}\left|\sum_{n=0}^{N-h-1}z_{n+h}\overline{z_n}\right|\right].

Dividing by H2N2H^2N^2 proves (18). We have completed the proof.

The discrepancy bound for bounded-variation functions

Proof of (21). Write ω=μ−dt\omega=\mu-dt and F(t)=ω([0,t))F(t)=\omega([0,t)). Then

F(0)=F(1)=0,∥F∥∞≤D(μ).F(0)=F(1)=0,\quad\|F\|_\infty\le\mathcal{D}(\mu).

For smooth periodic ff, integration by parts gives

∫Tf dω=−∫01F(t)f′(t) dt,\int_{\mathbb{T}}f\,d\omega=-\int_0^1F(t)f'(t)\,dt,

and hence we obtain

∣∫Tf dω∣≤D(μ)Var⁡T(f).\left|\int_{\mathbb{T}}f\,d\omega\right|\le\mathcal{D}(\mu)\operatorname{Var}_{\mathbb{T}}(f).

The result follows by uniform approximation with smooth periodic convolutions, whose variations do not exceed Var⁡T(f)\operatorname{Var}_{\mathbb{T}}(f). We have concluded the proof.

Mean and variation of the logarithmic kernel

Use the logarithmic kernel ℓ\ell and its truncation ℓh\ell_h from Lemma 4.3. Since ℓ∈L1(T)\ell\in L^1(\mathbb{T}), its mean can be computed as follows. Let L=∫0π/2log⁡sin⁡x dxL = \int_0^{\pi/2} \log\sin x\,dx. Then

2L=∫0π/2log⁡(sin⁡xcos⁡x) dx=−π2log⁡2+12∫0πlog⁡sin⁡u du=−π2log⁡2+L.2L = \int_0^{\pi/2} \log(\sin x \cos x)\,dx = -\frac{\pi}{2}\log2 + \frac{1}{2}\int_0^\pi\log\sin u\,du = -\frac{\pi}{2}\log2 + L.

Thus L=−π2log⁡2L=-\frac{\pi}{2}\log2, and consequently

∫Tℓ(t) dt=0.\int_{\mathbb{T}} \ell(t)\,dt=0.

Let ah=π−1arcsin⁡(h/2)a_h=\pi^{-1}\arcsin(h/2). By symmetry and the definition of ℓh\ell_h, we have

Var⁡T(ℓh)=2(log⁡2−log⁡h)=2log⁡(2/h),\operatorname{Var}_{\mathbb{T}}(\ell_h)=2(\log2-\log h)=2\log(2/h),
∫Tℓh(t) dt=2∫0ahlog⁡h2sin⁡πt dt≤2∫0h/4log⁡h4t dt=h2≤h,\int_{\mathbb{T}}\ell_h(t)\,dt=2\int_0^{a_h}\log\frac{h}{2\sin\pi t}\,dt \leq2\int_0^{h/4}\log\frac{h}{4t}\,dt=\frac{h}{2}\leq h,

where we used 2sin⁡πt≥4t2\sin\pi t\geq4t on [0,1/2][0,1/2], which also implies ah≤h/4a_h\leq h/4.

Proof of Lemma 4.4

Proof. We use the Combes-Thomas argument [10]. Fix m∈Zm\in\mathbb{Z} and set

φT(n)=min⁡{∣n−m∣,T},(Wμ,Tu)(n)=eμφT(n)u(n).\varphi_T(n)=\min\{|n-m|,T\},\qquad(W_{\mu,T}u)(n)=e^{\mu\varphi_T(n)}u(n).

Since ∣φT(n+1)−φT(n)∣≤1|\varphi_T(n+1)-\varphi_T(n)|\leq1, the Schur test gives

∥Wμ,TλΔWμ,T−1−λΔ∥≤2λ(eμ−1).\left\|W_{\mu,T}\lambda\Delta W_{\mu,T}^{-1}-\lambda\Delta\right\|\leq2\lambda(e^\mu-1).

With μ=μλ\mu=\mu_\lambda, the right-hand side is 1/21/2. Write

B=Wμλ,TλΔWμλ,T−1−λΔ.\mathcal{B}=W_{\mu_\lambda,T}\lambda\Delta W_{\mu_\lambda,T}^{-1}-\lambda\Delta.

Since ∥RE∥≤1\|\mathcal{R}_E\|\leq1, we have ∥BRE∥≤1/2\|\mathcal{B}\mathcal{R}_E\|\leq1/2. The Neumann series yields

∥Wμλ,TREWμλ,T−1∥=∥RE(I+BRE)−1∥≤2.\left\|W_{\mu_\lambda,T}\mathcal{R}_E W_{\mu_\lambda,T}^{-1}\right\| =\left\|\mathcal{R}_E(I+\mathcal{B}\mathcal{R}_E)^{-1}\right\|\leq2.

Since φT(m)=0\varphi_T(m)=0, the (n,m)(n,m) matrix entry satisfies

eμλφT(n)∣RE(n,m)∣≤2.e^{\mu_\lambda\varphi_T(n)}|\mathcal{R}_E(n,m)|\leq2.

Taking T≥∣n−m∣T\geq|n-m| proves (28).

Translation invariance of λΔ\lambda\Delta makes RE\mathcal{R}_E and CE\mathcal{C}_E convolution operators. The identity in (26) gives

∣cE(r)∣=2∣RE(r,0)∣≤4e−μλ∣r∣,r≠0.|c_E(r)|=2|\mathcal{R}_E(r,0)|\leq4e^{-\mu_\lambda|r|},\quad r\neq0.

We complete the proof. □

Resolvent compression and trace norms

Proof of Lemma 4.5. Use P\mathscr{P} and Q\mathscr{Q} from (31), and set A=RE−1\mathscr{A}=\mathscr{R}_E^{-1}. Compressing ARE=I\mathscr{A}\mathscr{R}_E=\mathscr{I} gives

P=PAREP=(PAP)(PREP)+PλΔQREP,\begin{aligned} \mathscr{P}=\mathscr{P}\mathscr{A}\mathscr{R}_E\mathscr{P} \\ &=(\mathscr{P}\mathscr{A}\mathscr{P})(\mathscr{P}\mathscr{R}_E\mathscr{P})+\mathscr{P}\lambda\Delta\mathscr{Q}\mathscr{R}_E\mathscr{P}, \end{aligned}

where PAQ=PλΔQ\mathscr{P}\mathscr{A}\mathscr{Q}=\mathscr{P}\lambda\Delta\mathscr{Q} since PQ=0\mathscr{P}\mathscr{Q}=0. Multiplying on the left by RE,N\mathscr{R}_{E,N} yields (31). Moreover,

PΔQu=u(−1)δ0+u(N)δN−1,\mathscr{P}\Delta\mathscr{Q}u=u(-1)\delta_0+u(N)\delta_{N-1},

so

rank⁡(RE,N−PREP)≤2.\operatorname{rank}(\mathscr{R}_{E,N}-\mathscr{P}\mathscr{R}_E\mathscr{P})\le2.

Since both resolvents have norm at most one,

∥RE,N−PREP∥S1≤2∥RE,N−PREP∥≤4.\|\mathscr{R}_{E,N}-\mathscr{P}\mathscr{R}_E\mathscr{P}\|_{\mathcal{S}_1}\le2\|\mathscr{R}_{E,N}-\mathscr{P}\mathscr{R}_E\mathscr{P}\|\le4.

The Cayley transform CE=I−2iRE\mathcal{C}_E=\mathscr{I}-2i\mathscr{R}_E therefore gives (29).

For operators X,Y\mathcal{X},\mathcal{Y} of norm at most one, the identity

Xk−Yk=∑j=0k−1Xk−1−j(X−Y)Yj\mathcal{X}^{k}-\mathcal{Y}^{k}=\sum_{j=0}^{k-1}\mathcal{X}^{k-1-j}(\mathcal{X}-\mathcal{Y})\mathcal{Y}^{j}

and the trace-ideal inequality ∥ABC∥S1≤∥A∥⋅∥B∥S1⋅∥C∥\|\mathcal{A}\mathcal{B}\mathcal{C}\|_{\mathcal{S}_1}\le\|\mathcal{A}\|\cdot\|\mathcal{B}\|_{\mathcal{S}_1}\cdot\|\mathcal{C}\| and ∣Tr⁡(T)∣≤∥T∥S1|\operatorname{Tr}(\mathcal{T})|\le\|\mathcal{T}\|_{\mathcal{S}_1} yield

∣Tr⁡(Xk−Yk)∣≤k∥X−Y∥S1.|\operatorname{Tr}(\mathcal{X}^{k}-\mathcal{Y}^{k})|\le k\|\mathcal{X}-\mathcal{Y}\|_{\mathcal{S}_1}.

Taking

X=UE,N,Y=(PCEP)DN,\mathcal{X}=\mathcal{U}_{E,N},\qquad\mathcal{Y}=(\mathscr{P}\mathcal{C}_E\mathscr{P})\mathcal{D}_N,

both have norm at most one. Moreover, we can get

∥X−Y∥S1=∥CE,N−PCEP∥S1≤8,\|\mathcal{X}-\mathcal{Y}\|_{\mathcal{S}_1}=\|\mathcal{C}_{E,N}-\mathscr{P}\mathcal{C}_E\mathscr{P}\|_{\mathcal{S}_1}\le8,

since right multiplication by the unitary DN\mathcal{D}_N preserves the trace norm. This proves (30). The proof is hereby completed. □

Comparison with the path sum

Proof of (37). Let M=MNM=M_N, P=PN\mathscr{P}=\mathscr{P}_N, and define

YN=(PCEP)DN,\mathcal{Y}_N=(\mathscr{P}\mathcal{C}_E\mathscr{P})\mathcal{D}_N,
YN,M=(PCE[M]P)DN,\mathcal{Y}_{N,M}=(\mathscr{P}\mathcal{C}^{[M]}_E\mathscr{P})\mathcal{D}_N,
WM=CE[M]Dα,β.\mathcal{W}_M=\mathcal{C}^{[M]}_E\mathcal{D}_{\alpha,\beta}.

By (30), we have

∣Tr⁡(UE,Nk)−Tr⁡(YNk)∣≤8k.|\operatorname{Tr}(\mathcal{U}_{E,N}^{k})-\operatorname{Tr}(\mathcal{Y}_N^{k})|\le8k.

Moreover, (32) and ∥T∥S1≤N∥T∥\|\mathcal{T}\|_{\mathcal{S}_1}\le N\|\mathcal{T}\| give

∥YN−YN,M∥S1≤NεN.\|\mathcal{Y}_N-\mathcal{Y}_{N,M}\|_{\mathcal{S}_1}\le N\varepsilon_N.

Since ∥YN∥,∥YN,M∥≤1+εN\lVert\mathcal{Y}_N\rVert,\lVert\mathcal{Y}_{N,M}\rVert\le1+\varepsilon_N,

YNk−YN,Mk=∑j=0k−1YNk−1−j(YN−YN,M)YN,Mj\mathcal{Y}_N^k-\mathcal{Y}_{N,M}^k=\sum_{j=0}^{k-1}\mathcal{Y}_N^{k-1-j}(\mathcal{Y}_N-\mathcal{Y}_{N,M})\mathcal{Y}_{N,M}^j

yields

∣Tr⁡(YNk)−Tr⁡(YN,Mk)∣≤NkεN(1+εN)k−1≤2NkεN.\left|\operatorname{Tr}(\mathcal{Y}_N^k)-\operatorname{Tr}(\mathcal{Y}_{N,M}^k)\right|\le Nk\varepsilon_N(1+\varepsilon_N)^{k-1}\le2Nk\varepsilon_N.

The two diagonal entries coincide outside the boundary layers of total size at most 2kM2kM. Hence, we can get

∣Tr⁡(YN,Mk)−∑n=0N−1WMk(n,n)∣≤2kM⋅4=8kM.\left|\operatorname{Tr}(\mathcal{Y}_{N,M}^k)-\sum_{n=0}^{N-1}\mathcal{W}_M^k(n,n)\right|\le2kM\cdot4=8kM.

Combining the three estimates and using (33) gives (37). The proof is hereby completed. □

Appendix B. Spectral separation and infinite-volume details

The quantitative bound in Lemma 4.10

Proof of Lemma 4.10. Since ∥ΔN∥≤2\lVert\Delta_N\rVert\le2 and ∣E∣≤E∗|E|\le E_\ast,

∥KE,N−iIN∥≤2λ+E∗+1=C.\lVert\mathcal{K}_{E,N}-\mathrm{i}\mathcal{I}_N\rVert\le2\lambda+E_\ast+1=C.

By (40) and ∥v∥=1\lVert v\rVert=1, we have

2≤C∥u0∥,∥u∥≥∥u0∥−∣w−w0∣.2\le C\lVert u_0\rVert,\qquad\lVert u\rVert\ge\lVert u_0\rVert-|w-w_0|.

Thus, if ∣w−w0∣≤C−1|w-w_0|\le C^{-1}, then we get

∥u∥≥C−1.\lVert u\rVert\ge C^{-1}.

Using the residual identity and the spectral theorem, we obtain

δ≤dist⁡(E,σ(HN))≤∥(HN−EIN)u∥∥u∥≤C2∣w−w0∣.\delta\le\operatorname{dist}(E,\sigma(\mathcal{H}_N))\le\frac{\lVert(\mathcal{H}_N-E\mathcal{I}_N)u\rVert}{\lVert u\rVert}\le C^2|w-w_0|.

If ∣w−w0∣>C−1|w-w_0|>C^{-1}, then we conclude

∣w−w0∣>C−1≥δC−2.|w-w_0|>C^{-1}\ge\delta C^{-2}.

Hence, in either case, we have

∣w−w0∣≥C−2δ.|w-w_0|\ge C^{-2}\delta.

The argument only requires HN\mathcal{H}_N to be defined at ww, so it also applies when w0w_0 is a pole. We close the proof here. □

Counting the pole-free intervals in Lemma 6.1

Proof. Fix ϑ\vartheta. Since sn≠0s_n \ne0 on Ic,N\mathcal{I}_{c,N} and JR,N\mathcal{J}_{R,N}, the pole equation

snt≡12−ϑ(mod1)s_n t \equiv\frac{1}{2}-\vartheta\pmod{1}

has exactly ∣sn∣|s_n| solutions on T\mathbb{T} for each point n∈Zn \in\mathbb{Z}. Moreover,

#(Ic,N∪JR,N)≤5N,∣n∣≤7N,\#(\mathcal{I}_{c,N} \cup\mathcal{J}_{R,N}) \le5N,\quad|n| \le7N,

so that

∑n∈Ic,N∪JR,N∣sn∣≤5N(49N2+4N2)=265N3.\sum_{n\in\mathcal{I}_{c,N}\cup\mathcal{J}_{R,N}} |s_n| \le5N(49N^2+4N^2)=265N^3.

Thus the pole-free part of T\mathbb{T} consists of at most 265N3+1265N^3+1 intervals.

Let

σ1(t)=σ(HIc,N(ϑ,t)),σ2(t)=σ(HJR,N(ϑ,t)).\sigma_1(t)=\sigma(\mathcal{H}_{\mathcal{I}_{c,N}}(\vartheta,t)),\quad\sigma_2(t)=\sigma(\mathcal{H}_{\mathcal{J}_{R,N}}(\vartheta,t)).

By the derivative bound, the set of parameters for which any fixed pair of eigenvalues differs by less than δ\delta has measure at most

mes⁡{t:∣ζk(t)−ηj(t)∣<δ}≤δ3πN2.\operatorname{mes}\{t:|\zeta_k(t)-\eta_j(t)|<\delta\}\le\frac{\delta}{3\pi N^2}.

Consequently, using L/N≤5/2L/N\le5/2 and N≥2N\ge2, we have

mes⁡{t:dist⁡(σ1(t),σ2(t))<δ}≤(265N3+1)L2δ3πN2≤5302596πN3δ<200N3δ.\begin{aligned} \operatorname{mes}\{t:\operatorname{dist}(\sigma_1(t),\sigma_2(t))<\delta\}\le(265N^3+1)\frac{L^2\delta}{3\pi N^2} \\ &\le\frac{53025}{96\pi}N^3\delta\\ &<200N^3\delta. \end{aligned}

The same bound holds after removing P\mathcal{P}. Integrating over ϑ\vartheta and using the measure preservation of (51) proves (50). This completes the proof. □ (Shi) School of Mathematics, Sichuan University, Chengdu 610064, China Email address: [email protected]

(Wen) School of Mathematics, Sichuan University, Chengdu 610064, China Email address: [email protected]

References

  1. [1][And58] P. W. Anderson. Absence of diffusion in certain random lattices. Physical review, 109(5):1492–1505, 1958.
  2. [3][BG00] J. Bourgain and M. Goldstein. On nonperturbative localization with quasi-periodic potential. Ann. of Math. (2), 152(3):835–879, 2000.
  3. [4][BGS01] J. Bourgain, M. Goldstein, and W. Schlag. Anderson localization for Schrödinger operators on ℤ with potentials given by the skew-shift. Comm. Math. Phys., 220(3):583–621, 2001.DOI
  4. [5][BGS02] J. Bourgain, M. Goldstein, and W. Schlag. Anderson localization for Schrödinger operators on ℤ² with quasi-periodic potential. Acta Math., 188(1):41–86, 2002.DOI
  5. [6][Bou02a] J. Bourgain. Estimates on Green’s functions, localization and the quantum kicked rotor model. Ann. of Math. (2), 156(1):249–294, 2002.DOI
  6. [7][Bou02b] J. Bourgain. On the spectrum of lattice Schrödinger operators with deterministic potential. J. Anal. Math., 87:37–75, 2002. Dedicated to the memory of Thomas H. Wolff.DOI
  7. [8][Bou05] J. Bourgain. Green’s function estimates for lattice Schrödinger operators and applications, volume 158 of Annals of Mathematics Studies. Princeton University Press, Princeton, NJ, 2005.
  8. [9]G. Casati and et al. Stochastic behavior of a quantum pendulum under a periodic perturbation. In Stochastic behavior in classical and quantum Hamiltonian systems, pages 334–352. Springer, 1979.
  9. [10]J. M. Combes and L. Thomas. Asymptotic behaviour of eigenfunctions for multiparticle Schrödinger operators. Comm. Math. Phys., 34:251–270, 1973.DOI
  10. [11]P. Erdős and P. Turán. On a problem in the theory of uniform distribution. I. Nederl. Akad. Weten-sch., Proc., 51:1146–1154, 1948.
  11. [12]S. Fishman, D. R. Grempel, and R. E. Prange. Chaos, quantum recurrences, and Anderson localization. Phys. Rev. Lett., 49(8):509–512, 1982.DOI
  12. [13]M. Goldstein and W. Schlag. Hölder continuity of the integrated density of states for quasi-periodic Schrödinger equations and averages of shifts of subharmonic functions. Ann. of Math. (2), 154(1):155–203, 2001.
  13. [14]S. Jitomirskaya. Anderson localization for the almost Mathieu equation: a nonperturbative proof. Comm. Math. Phys., 165(1):49–57, 1994.DOI
  14. [15]S. Jitomirskaya. Metal-insulator transition for the almost Mathieu operator. Ann. of Math. (2), 150(3):1159–1175, 1999.arxiv.org/abs/math/9911265
  15. [17]M. S. Santhanam, S. Paul, and J. B. Kannan. Quantum kicked rotor and its variants: Chaos, localization and beyond. Phys. Rep., 956:1–87, 2022.DOI

Paper details

vlnc:W_zjjf6t26
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC0-1.0
Usage

views · 0

downloads · 0

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

  • Physics and Astronomy

ChatGPT 6.1 sol

View statement
Looking for related papers…
Contents