Introduction
The model
We study the following family of quasi-periodic operators on ℓ 2 ( Z d ) \ell^2(\mathbb{Z}^d) ℓ 2 ( Z d ) (d ≥ 1 d \ge1 d ≥ 1 ):
( H ( θ ) u ) ( n ) = f ( θ + n ⋅ ω ) u ( n ) + ε ∑ m ∈ Z d ϕ ( n − m ) u ( m ) , n ∈ Z d . (1) (H(\theta)u)(n) = f(\theta+ n \cdot\omega)u(n) + \varepsilon\sum_{m \in\mathbb{Z}^d} \phi(n-m)u(m), \qquad n \in\mathbb{Z}^d. \tag*{(1)} ( H ( θ ) u ) ( n ) = f ( θ + n ⋅ ω ) u ( n ) + ε m ∈ Z d ∑ ϕ ( n − m ) u ( m ) , n ∈ Z d . (1)
Here θ ∈ R \theta\in\mathbb{R} θ ∈ R is the phase, ε ∈ R \varepsilon\in\mathbb{R} ε ∈ R is the coupling constant with ∣ ε ∣ ≪ 1 |\varepsilon| \ll1 ∣ ε ∣ ≪ 1 and ω ∈ R d \omega\in\mathbb{R}^d ω ∈ R d is the frequency. The notation “⋅ \cdot ⋅ ” in (1) denotes the Euclidean inner product on R d \mathbb{R}^d R d . We assume ω ∈ D C \omega\in\mathrm{DC} ω ∈ DC , where
D C = { ω ∈ R d : ∥ k ⋅ ω ∥ T ≥ γ ⟨ k ⟩ − τ for all k ∈ Z d ∖ { o } } (2) \mathrm{DC} = \left\{ \omega\in\mathbb{R}^d : \|k \cdot\omega\|_{\mathbb{T}} \ge\gamma\langle k\rangle^{-\tau} \text{ for all } k \in\mathbb{Z}^d \setminus\{o\} \right\} \tag*{(2)} DC = { ω ∈ R d : ∥ k ⋅ ω ∥ T ≥ γ ⟨ k ⟩ − τ for all k ∈ Z d ∖ { o } } (2)
with fixed γ > 0 \gamma> 0 γ > 0 and τ > d \tau> d τ > d .
The symbols o o o , ⟨ k ⟩ \langle k\rangle ⟨ k ⟩ and ∥ ⋅ ∥ T \|\cdot\|_{\mathbb{T}} ∥ ⋅ ∥ T in (2) are defined as follows. For x = ( x 1 , … , x d ) ∈ R d x = (x_1,\ldots,x_d) \in\mathbb{R}^d x = ( x 1 , … , x d ) ∈ R d , set
o = ( 0 , … , 0 ) ∈ R d , ∣ x ∣ = max 1 ≤ i ≤ d ∣ x i ∣ , ⟨ x ⟩ = 1 + ∣ x ∣ . (3) o = (0,\ldots,0) \in\mathbb{R}^d, \qquad|x| = \max_{1 \le i \le d}|x_i|, \qquad\langle x\rangle= 1 + |x|. \tag*{(3)} o = ( 0 , … , 0 ) ∈ R d , ∣ x ∣ = 1 ≤ i ≤ d max ∣ x i ∣ , ⟨ x ⟩ = 1 + ∣ x ∣. (3)
For α ∈ R \alpha\in\mathbb{R} α ∈ R , set
T = R / Z , ∥ α ∥ T = min m ∈ Z ∣ α − m ∣ . (4) \mathbb{T} = \mathbb{R}/\mathbb{Z}, \qquad\|\alpha\|_{\mathbb{T}} = \min_{m \in\mathbb{Z}}|\alpha-m|. \tag*{(4)} T = R / Z , ∥ α ∥ T = m ∈ Z min ∣ α − m ∣. (4)
For some r > d r > d r > d , the hopping ϕ : Z d → C \phi: \mathbb{Z}^d \to\mathbb{C} ϕ : Z d → C in (1) satisfies
ϕ ( o ) = 0 , ϕ ( k ) ‾ = ϕ ( − k ) , ∣ ϕ ( k ) ∣ ≤ C 0 ⟨ k ⟩ − r ( k ∈ Z d ) , (5) \phi(o) = 0, \qquad\overline{\phi(k)} = \phi(-k), \qquad|\phi(k)| \le C_0\langle k\rangle^{-r} \qquad(k \in\mathbb{Z}^d), \tag*{(5)} ϕ ( o ) = 0 , ϕ ( k ) = ϕ ( − k ) , ∣ ϕ ( k ) ∣ ≤ C 0 ⟨ k ⟩ − r ( k ∈ Z d ) , (5)
where
C 0 = C 0 ( r , ϕ ) = max { 1 , sup k ∈ Z d ⟨ k ⟩ r ∣ ϕ ( k ) ∣ } < ∞ . (6) C_0 = C_0(r,\phi) = \max\left\{1,\sup_{k \in\mathbb{Z}^d}\langle k\rangle^r|\phi(k)|\right\} < \infty. \tag*{(6)} C 0 = C 0 ( r , ϕ ) = max { 1 , k ∈ Z d sup ⟨ k ⟩ r ∣ ϕ ( k ) ∣ } < ∞. (6)
For the potential, we assume that f f f is 1 1 1 -periodic,
f : R → [ − ∞ , + ∞ ) , f ( θ + 1 ) = f ( θ ) , f ( θ ) ∈ R ( 0 < θ < 1 ) , (7) f : \mathbb{R} \to[-\infty,+\infty), \qquad f(\theta+1)=f(\theta), \qquad f(\theta)\in\mathbb{R} \quad(0<\theta<1), \tag*{(7)} f : R → [ − ∞ , + ∞ ) , f ( θ + 1 ) = f ( θ ) , f ( θ ) ∈ R ( 0 < θ < 1 ) , (7)
and Lipschitz monotone on [ 0 , 1 ) [0,1) [ 0 , 1 ) : for some L > 0 L>0 L > 0 ,
f ( θ 2 ) − f ( θ 1 ) ≥ L ( θ 2 − θ 1 ) , 0 ≤ θ 1 < θ 2 < 1. (8) f(\theta_2)-f(\theta_1) \ge L(\theta_2-\theta_1), \qquad0 \le\theta_1 < \theta_2 < 1. \tag*{(8)} f ( θ 2 ) − f ( θ 1 ) ≥ L ( θ 2 − θ 1 ) , 0 ≤ θ 1 < θ 2 < 1. (8)
We call L L L the monotonicity constant.
We consider H ( θ ) \mathcal{H}(\theta) H ( θ ) only at phases for which f ( θ + n ⋅ ω ) f(\theta+n\cdot\omega) f ( θ + n ⋅ ω ) is finite for every n ∈ Z d n\in\mathbb{Z}^{d} n ∈ Z d . Denote
Θ = { θ ∈ R : f ( θ + n ⋅ ω ) ∈ R for every n ∈ Z d } . (9) \Theta=\{\theta\in\mathbb{R}:f(\theta+n\cdot\omega)\in\mathbb{R}\text{ for every }n\in\mathbb{Z}^{d}\}. \tag*{(9)} Θ = { θ ∈ R : f ( θ + n ⋅ ω ) ∈ R for every n ∈ Z d } . (9)
Thus Θ = R \Theta=\mathbb{R} Θ = R if f ( 0 ) ∈ R f(0)\in\mathbb{R} f ( 0 ) ∈ R , whereas Θ = R ∖ { m + n ⋅ ω : m ∈ Z , n ∈ Z d } \Theta=\mathbb{R}\setminus\{m+n\cdot\omega:m\in\mathbb{Z},\ n\in\mathbb{Z}^{d}\} Θ = R ∖ { m + n ⋅ ω : m ∈ Z , n ∈ Z d } if f ( 0 ) = − ∞ f(0)=-\infty f ( 0 ) = − ∞ (equivalently, f f f is unbounded below). For every θ ∈ Θ \theta\in\Theta θ ∈ Θ , H ( θ ) \mathcal{H}(\theta) H ( θ ) is self-adjoint.
Motivation
At ε = 0 \varepsilon=0 ε = 0 , the operator (1) has an eigenbasis supported at single lattice points. We ask whether a complete system of localized eigenvectors persists for sufficiently small ε \varepsilon ε . For nearest-neighbor hopping, the Maryland model is an early example ([GFP82]). In 1983, Bellissard, Lima and Scoppola ([BLS83]) established perturbative localization for meromorphic monotone potentials in arbitrary dimension. The restriction to meromorphic potentials led to the question of localization under a quantitative monotonicity condition alone.
In 2019, Jitomirskaya and Kachkovskiy ([JK19]) introduced a general class of bounded Lipschitz monotone potentials. The all-coupling localization theorem applies in one dimension to Diophantine frequencies and almost every phase. Higher-dimensional perturbative results were subsequently obtained for unbounded monotone potentials in 2022 ([KPS22]) and for Hölder monotone potentials in 2025 ([KPS25]). Cao, Shi and Zhang ([CSZ25]) also established localization uniformly in phase for bounded and unbounded Lipschitz monotone potentials.
For the hopping term, perturbative localization also holds under exponential decay ([CSZ25, Remark 1.5]). More recently, Cao ([Cao26]) obtained arithmetic localization for one-dimensional exponentially decaying hopping under an explicit frequency condition and a small-coupling assumption. For power-law hopping, Shi ([Shi23]) and Shi and Wen ([SW23]) proved localization results in 2023. For the Maryland potential, a sufficient pointwise hopping decay condition in [Shi23, SW23] is r > τ + d r>\tau+d r > τ + d . We ask whether perturbative localization holds for the original operator with a general Lipschitz monotone potential and power-law hopping under r > τ r>\tau r > τ .
Main results
Our main result is as follows.
Theorem 1.1. Fix an integer d ≥ 1 d\geq1 d ≥ 1 , numbers τ > d \tau>d τ > d , γ , L > 0 \gamma,L>0 γ , L > 0 and r > τ r>\tau r > τ . Assume (7) , (8) and (5) . There is a constant C 1 = C 1 ( d , r , τ , γ , L , ϕ ) > 0 C_{1}=C_{1}(d,r,\tau,\gamma,L,\phi)>0 C 1 = C 1 ( d , r , τ , γ , L , ϕ ) > 0 such that, for every 0 < ρ < r − τ 0<\rho<r-\tau 0 < ρ < r − τ , there are constants ε 0 = ε 0 ( d , τ , γ , L , r , ρ , ϕ ) > 0 \varepsilon_{0}=\varepsilon_{0}(d,\tau,\gamma,L,r,\rho,\phi)>0 ε 0 = ε 0 ( d , τ , γ , L , r , ρ , ϕ ) > 0 and C 2 = C 2 ( d , r , τ , γ , L , ρ , ϕ ) > 0 C_{2}=C_{2}(d,r,\tau,\gamma,L,\rho,\phi)>0 C 2 = C 2 ( d , r , τ , γ , L , ρ , ϕ ) > 0 with the following property. For every ω ∈ D C \omega\in\mathrm{DC} ω ∈ DC and ∣ ε ∣ < ε 0 |\varepsilon|<\varepsilon_{0} ∣ ε ∣ < ε 0 , there is a real 1-periodic function E ∞ : Θ → R E_{\infty}:\Theta\to\mathbb{R} E ∞ : Θ → R such that
∥ E ∞ − f ∥ ∞ ≤ C 1 ε 2 , E ∞ ( θ 2 ) − E ∞ ( θ 1 ) ≥ L ( θ 2 − θ 1 ) ( θ 1 < θ 2 , θ 1 , θ 2 ∈ Θ ∩ [ 0 , 1 ] ) , \|E_{\infty}-f\|_{\infty}\leq C_{1}\varepsilon^{2},\qquad E_{\infty}(\theta_{2})-E_{\infty}(\theta_{1})\geq L(\theta_{2}-\theta_{1})\quad(\theta_{1}<\theta_{2},\ \theta_{1},\theta_{2}\in\Theta\cap[0,1]), ∥ E ∞ − f ∥ ∞ ≤ C 1 ε 2 , E ∞ ( θ 2 ) − E ∞ ( θ 1 ) ≥ L ( θ 2 − θ 1 ) ( θ 1 < θ 2 , θ 1 , θ 2 ∈ Θ ∩ [ 0 , 1 ]) , where ∥ ⋅ ∥ ∞ \|\cdot\|_{\infty} ∥ ⋅ ∥ ∞ denotes the supremum norm on Θ ∩ [ 0 , 1 ] \Theta\cap[0,1] Θ ∩ [ 0 , 1 ] . For every θ ∈ Θ \theta\in\Theta θ ∈ Θ , the operator H ( θ ) \mathcal{H}(\theta) H ( θ ) has a complete orthonormal eigenbasis ( ψ θ , n ) n ∈ Z d (\psi_{\theta,n})_{n\in\mathbb{Z}^{d}} ( ψ θ , n ) n ∈ Z d satisfying
H ( θ ) ψ θ , n = E ∞ ( θ + n ⋅ ω ) ψ θ , n . \mathcal{H}(\theta)\psi_{\theta,n}=E_{\infty}(\theta+n\cdot\omega)\psi_{\theta,n}. H ( θ ) ψ θ , n = E ∞ ( θ + n ⋅ ω ) ψ θ , n . For m ≠ n m\ne n m = n , E ∞ ( θ + m ⋅ ω ) ≠ E ∞ ( θ + n ⋅ ω ) E_{\infty}(\theta+m\cdot\omega)\ne E_{\infty}(\theta+n\cdot\omega) E ∞ ( θ + m ⋅ ω ) = E ∞ ( θ + n ⋅ ω ) . The vectors ψ θ , n \psi_{\theta,n} ψ θ , n can be chosen so that
sup n ∈ Z d ∑ m ∈ Z d ∣ ψ θ , n ( m ) − δ m n ∣ ∣ m − n ∣ ρ ≤ C 2 ∣ ε ∣ . (10) \sup_{n\in\mathbb{Z}^{d}}\sum_{m\in\mathbb{Z}^{d}}|\psi_{\theta,n}(m)-\delta_{mn}|\lvert m-n\rvert^{\rho}\leq C_{2}|\varepsilon|. \tag*{(10)} n ∈ Z d sup m ∈ Z d ∑ ∣ ψ θ , n ( m ) − δ mn ∣ ∣ m − n ∣ ρ ≤ C 2 ∣ ε ∣. (10) where δ m n = 1 \delta_{mn}=1 δ mn = 1 if m = n m=n m = n and δ m n = 0 \delta_{mn}=0 δ mn = 0 otherwise. Moreover, for the unitary operator exp ( − i u H ( θ ) ) \exp(-\mathrm{i}u\mathcal{H}(\theta)) exp ( − i u H ( θ )) defined by the spectral theorem,
sup u ∈ R sup n ∈ Z d ∑ m ∈ Z d ∣ ⟨ e m , exp ( − i u H ( θ ) ) e n ⟩ ∣ ∣ m − n ∣ ρ ≤ 3 , (11) \sup_{u\in\mathbb{R}}\sup_{n\in\mathbb{Z}^{d}}\sum_{m\in\mathbb{Z}^{d}}\left|\langle e_{m},\exp(-\mathrm{i}u\mathcal{H}(\theta))e_{n}\rangle\right|\lvert m-n\rvert^{\rho}\leq3, \tag*{(11)} u ∈ R sup n ∈ Z d sup m ∈ Z d ∑ ∣ ⟨ e m , exp ( − i u H ( θ )) e n ⟩ ∣ ∣ m − n ∣ ρ ≤ 3 , (11) where ⟨ ⋅ , ⋅ ⟩ \langle\cdot,\cdot\rangle ⟨ ⋅ , ⋅ ⟩ is the ℓ 2 \ell^{2} ℓ 2 inner product, linear in the second argument, and e n ( m ) = δ n m e_{n}(m)=\delta_{nm} e n ( m ) = δ nm .
Related work
The results in Subsection 1.2 arise from two different approaches to localization. Early perturbative constructions for almost-periodic operators were developed by Craig and Pöschel in 1983 ([5 Pös83] ). For monotone potentials, the KAM proof of Bellissard, Lima and Scoppola ([2 ] ) uses meromorphic continuation to control the small denominators. The one-dimensional nonperturbative approach instead relates localization to the Lyapunov exponent. Jitomirskaya and Kachkovskiy ([8 ] ) combined monotonicity estimates with large deviation estimates for finite-volume determinants. This provided a way to treat potentials without the special structure of the Maryland potential.
The nonperturbative theory was extended in several directions. In 2019, Kachkovskiy ([12 ] ) treated unbounded monotone potentials under a logarithmic integrability condition, which ensures that the Lyapunov exponent is finite. The upper Lipschitz bound in the bounded case was also removed ([12 Remark 1.4] ). Meanwhile, arithmetic localization studies the competition between the Lyapunov exponent and rational approximation of the frequency. Jitomirskaya and Liu ([11 ] ) determined the arithmetic spectral transitions for the Maryland model in 2017. Jitomirskaya and Kachkovskiy ([9 ] ) established the sharp arithmetic localization criterion for Lipschitz monotone potentials in 2024. More recently, Jitomirskaya, Levi and Stoiber ([10 ] ) treated a class of monotone potentials with flat regions. The proof in [10 ] recovers quantitative variation of finite-volume eigenvalues by estimating eigenfunction mass outside the flat regions. For exponentially decaying long-range hopping, Cao ([3 ] ) extended the nonperturbative arithmetic method by estimating Green’s functions without a Lyapunov exponent.
In higher dimensions, the recent perturbative proofs use different constructions of the eigenvalues and eigenvectors. Kachkovskiy, Parnovski and Shterenberg ([13 ] ) proved convergence of perturbation series by grouping terms to exploit cancellations. In [14 ] , Kachkovskiy, Parnovski and Shterenberg used local partial diagonalization and finite-volume eigenvalue ordering. The ordered branches recover the monotonicity needed to control small denominators throughout the iteration. Cao, Shi and Zhang ([4 ] ) instead constructed Rellich functions on increasing finite regions and used Green’s function estimates to pass between scales. Monotonicity separates points whose local eigenvalues are close to the same energy. For bounded potentials, a Schur complement comparison across discontinuities ensures that the selected Rellich functions preserve Lipschitz monotonicity.
For power-law hopping, the estimates must also control sums of slowly decaying matrix entries. Shi ([Shi23]) developed a Nash–Moser reducibility scheme in Sobolev matrix norms. Tame product estimates and smoothing compensate for the loss of regularity in the homological equations. The general theorem gives reducibility after a small correction of the diagonal potential ([Shi23, Theorem 2.1]). The theorem for the original operator assumes that the small-denominator bounds persist under all sufficiently small diagonal perturbations in the chosen Banach algebra ([Shi23, Theorem 2.2]). Shi and Wen ([SW23]) used meromorphic monotonicity to obtain diagonalization uniformly in phase. In the present paper, we combine the Rellich-function construction with Schur complement estimates adapted to power-law decay.
Main ideas and new ingredients of the proof
We use the Rellich-function construction of Cao, Shi and Zhang ([4 ] ) as the starting point. The main changes are as follows.
1. Monotonicity of the local eigenvalues. As in [4 ] , we construct local eigenvalues on increasing finite regions and use monotonicity to separate the resonant points. The finite-volume ordering argument in [14 Proposition 4.1 and Remark 4.2] gives eigenvalue branches with the same Lipschitz monotonicity as the potential. By continuation from zero coupling to ε \varepsilon ε , we identify each isolated local eigenvalue with the ordered branch associated with the diagonal entry f ( θ ) f(\theta) f ( θ ) (Proposition 3.4). The branch identification gives the required monotonicity in place of the comparison across jumps used for bounded potentials in [4 ] .
2. Schur estimates under the condition r > τ r > \tau r > τ . Shi’s Nash–Moser iteration controls the conjugating matrices in Sobolev norms ([Shi23]). We estimate the entries of each Schur complement directly and use the separation of the sites retained from earlier steps to bound the sums over coordinates being eliminated. Once the diagonal entries are bounded away from zero, Lemma A.2 gives inverse estimates and preserves the off-diagonal decay exponent r r r . After elimination, however, the diagonal entries depend on the finite region. Corollary 5.3 bounds the change of a diagonal entry when the region is enlarged. We can therefore use local eigenvalues to obtain the diagonal lower bounds needed for elimination in the larger region. The estimates close under r > τ r > \tau r > τ for the prescribed Lipschitz monotone potential.
3. Exponentially growing multiscale lengths. In [4 Definition 2.15] , the spatial scale is squared at each induction step. In our construction, the spatial scale increases by the fixed factor 16. Equation (76) controls the accumulation of the Schur estimates by a geometric series with ratio 16 7 − τ < 1 16^{7-\tau} < 1 1 6 7 − τ < 1 . The summability and the uniform region comparison (5.8) close the induction with a coupling threshold independent of the step.
Outline of this paper
Section 2 collects notation and standard matrix facts. Section 3 constructs the first local eigenvalue and the corresponding eigenvector, then continues with the remaining equations. Section 4 states the induction hypotheses. Section 5 proves the region comparison, performs the next elimination and estimates the changes of the local eigenvalues and corresponding eigenvectors for every n ≥ 1 n \ge1 n ≥ 1 . Section 6 passes to the limit and proves completeness and dynamical localization. Appendix A contains the sparse matrix estimates; Appendix B gives the geometric and domain verifications.
Preliminaries
Notation
1. Distances and phases. With ∣ ⋅ ∣ |\cdot| ∣ ⋅ ∣ as in (3) , define
dist ( x , S ) = inf y ∈ S ∣ x − y ∣ , x ∈ R d , S ⊂ R d , (12) \operatorname{dist}(x,S)=\inf_{y\in S}|x-y|,\qquad x\in\mathbb{R}^{d},\quad S\subset\mathbb{R}^{d}, \tag*{(12)} dist ( x , S ) = y ∈ S inf ∣ x − y ∣ , x ∈ R d , S ⊂ R d , (12)
with dist ( x , ∅ ) = + ∞ \operatorname{dist}(x,\varnothing)=+\infty dist ( x , ∅ ) = + ∞ . For α ∈ R \alpha\in\mathbb{R} α ∈ R , denote the fractional part by
{ α } = α − ⌊ α ⌋ ∈ [ 0 , 1 ) . (13) \{\alpha\}=\alpha-\lfloor\alpha\rfloor\in[0,1). \tag*{(13)} { α } = α − ⌊ α ⌋ ∈ [ 0 , 1 ) . (13)
Unless another domain is specified, phase-dependent functions are considered on the set Θ \Theta Θ in (9) .
2. Lattice sets. For Z , B ⊂ Z d Z,B\subset\mathbb{Z}^{d} Z , B ⊂ Z d and p ∈ Z d p\in\mathbb{Z}^{d} p ∈ Z d , denote
Z c = Z d ∖ Z , p + Z = { p + x : x ∈ Z } , Z + B = { x + y : x ∈ Z , y ∈ B } . (14) Z^{\mathrm{c}}=\mathbb{Z}^{d}\setminus Z,\qquad p+Z=\{p+x:x\in Z\},\qquad Z+B=\{x+y:x\in Z,\ y\in B\}. \tag*{(14)} Z c = Z d ∖ Z , p + Z = { p + x : x ∈ Z } , Z + B = { x + y : x ∈ Z , y ∈ B } . (14)
The cardinality of a finite set Z Z Z is denoted by # Z \#Z # Z . If o ∈ Z o\in Z o ∈ Z , write Z ∘ = Z ∖ { o } Z^{\circ}=Z\setminus\{o\} Z ∘ = Z ∖ { o } . For a set B ⊂ Z d B\subset\mathbb{Z}^{d} B ⊂ Z d containing o o o and p ∈ Z d p\in\mathbb{Z}^{d} p ∈ Z d , denote
B ( p ) = p + B , B ∘ ( p ) = B ( p ) ∖ { p } . (15) B(p)=p+B,\qquad B^{\circ}(p)=B(p)\setminus\{p\}. \tag*{(15)} B ( p ) = p + B , B ∘ ( p ) = B ( p ) ∖ { p } . (15)
For R ≥ 0 R\geq0 R ≥ 0 , let
Λ R = { n ∈ Z d : ∣ n ∣ ≤ R } . (16) \Lambda_R=\{n\in\mathbb{Z}^{d}:|n|\leq R\}. \tag*{(16)} Λ R = { n ∈ Z d : ∣ n ∣ ≤ R } . (16)
For M > 0 M>0 M > 0 , a set X ⊂ Z d X\subset\mathbb{Z}^{d} X ⊂ Z d is M M M -separated if ∣ p − q ∣ ≥ M |p-q|\geq M ∣ p − q ∣ ≥ M for all distinct p , q ∈ X p,q\in X p , q ∈ X .
3. Vectors and matrices. For a vector v v v whose index set contains Z Z Z , write v ∣ Z = ( v ( n ) ) n ∈ Z v|_Z=(v(n))_{n\in Z} v ∣ Z = ( v ( n ) ) n ∈ Z . For 1 ≤ p < ∞ 1\leq p<\infty 1 ≤ p < ∞ , let ∥ v ∥ ℓ p ( Z ) = ( ∑ n ∈ Z ∣ v ( n ) ∣ p ) 1 / p \|v\|_{\ell^p(Z)}=(\sum_{n\in Z}|v(n)|^p)^{1/p} ∥ v ∥ ℓ p ( Z ) = ( ∑ n ∈ Z ∣ v ( n ) ∣ p ) 1/ p ; we write ∥ v ∥ p \|v\|_p ∥ v ∥ p when the index set is fixed. For an operator M \mathcal{M} M , write M ( n , m ) = ⟨ e n , M e m ⟩ \mathcal{M}(n,m)=\langle e_n,\mathcal{M}e_m\rangle M ( n , m ) = ⟨ e n , M e m ⟩ . The adjoint is denoted by M ∗ \mathcal{M}^{*} M ∗ and ∥ M ∥ o p \|\mathcal{M}\|_{\mathrm{op}} ∥ M ∥ op denotes the ℓ 2 \ell^2 ℓ 2 operator norm. For a finite matrix, M ∗ ( n , m ) = M ( m , n ) ‾ \mathcal{M}^{*}(n,m)=\overline{\mathcal{M}(m,n)} M ∗ ( n , m ) = M ( m , n ) ; for a column vector v v v , the symbol v ∗ v^{*} v ∗ denotes the conjugate transpose of v v v . Let I Z I_Z I Z be the identity on ℓ 2 ( Z ) \ell^2(Z) ℓ 2 ( Z ) and let I = I Z d I=I_{\mathbb{Z}^{d}} I = I Z d . For p ∈ Z d p\in\mathbb{Z}^{d} p ∈ Z d , define the translation by
( U ( p ) v ) ( n ) = v ( n + p ) . (17) (\mathcal{U}(p)v)(n)=v(n+p). \tag*{(17)} ( U ( p ) v ) ( n ) = v ( n + p ) . (17)
For a matrix M \mathcal{M} M , let M Z , B = ( M ( p , q ) ) p ∈ Z , q ∈ B \mathcal{M}_{Z,B}=(\mathcal{M}(p,q))_{p\in Z,q\in B} M Z , B = ( M ( p , q ) ) p ∈ Z , q ∈ B and write M Z = M Z , Z \mathcal{M}_Z=\mathcal{M}_{Z,Z} M Z = M Z , Z . A point p p p in a submatrix index denotes the singleton { p } \{p\} { p } ; thus M p , B = M { p } , B \mathcal{M}_{p,B}=\mathcal{M}_{\{p\},B} M p , B = M { p } , B and M Z , q = M Z , { q } \mathcal{M}_{Z,q}=\mathcal{M}_{Z,\{q\}} M Z , q = M Z , { q } . Phase and energy variables are written as arguments.
Operators and finite restrictions
For q = 1 , 2 q=1,2 q = 1 , 2 and a real family ( v n ) n ∈ Z d (v_n)_{n\in\mathbb{Z}^{d}} ( v n ) n ∈ Z d , the symbol diag n ∈ Z d v n \operatorname{diag}_{n\in\mathbb{Z}^{d}}v_n diag n ∈ Z d v n denotes the multiplication operator on ℓ q ( Z d ) \ell^q(\mathbb{Z}^{d}) ℓ q ( Z d ) with the natural domain. The multiplication operator is closed, and finitely supported vectors form a core. On ℓ 2 ( Z d ) \ell^2(\mathbb{Z}^{d}) ℓ 2 ( Z d ) the operator is self-adjoint.
The hopping operator is defined by
( T v ) ( n ) = ∑ m ∈ Z d ϕ ( n − m ) v ( m ) . (18) (\mathcal{T}v)(n)=\sum_{m\in\mathbb{Z}^{d}}\phi(n-m)v(m). \tag*{(18)} ( T v ) ( n ) = m ∈ Z d ∑ ϕ ( n − m ) v ( m ) . (18)
By (5) and r > d r>d r > d , ϕ ∈ ℓ 1 ( Z d ) \phi\in\ell^1(\mathbb{Z}^{d}) ϕ ∈ ℓ 1 ( Z d ) ; hence T \mathcal{T} T is bounded on ℓ 1 ( Z d ) \ell^1(\mathbb{Z}^{d}) ℓ 1 ( Z d ) and ℓ 2 ( Z d ) \ell^2(\mathbb{Z}^{d}) ℓ 2 ( Z d ) . The symmetry in (5) implies that T \mathcal{T} T is self-adjoint on ℓ 2 ( Z d ) \ell^2(\mathbb{Z}^{d}) ℓ 2 ( Z d ) .
The phase set Θ \Theta Θ satisfies Θ + q = Θ \Theta+q=\Theta Θ + q = Θ and Θ + p ⋅ ω = Θ \Theta+p\cdot\omega=\Theta Θ + p ⋅ ω = Θ for q ∈ Z q\in\mathbb{Z} q ∈ Z and p ∈ Z d p\in\mathbb{Z}^{d} p ∈ Z d . Direct substitution into (1) proves
H ( θ + 1 ) = H ( θ ) , H ( θ + p ⋅ ω ) U ( p ) = U ( p ) H ( θ ) , θ ∈ Θ , p ∈ Z d . (19) \mathcal{H}(\theta+1)=\mathcal{H}(\theta),\qquad\mathcal{H}(\theta+p\cdot\omega)\mathcal{U}(p)=\mathcal{U}(p)\mathcal{H}(\theta),\qquad\theta\in\Theta,\quad p\in\mathbb{Z}^{d}. \tag*{(19)} H ( θ + 1 ) = H ( θ ) , H ( θ + p ⋅ ω ) U ( p ) = U ( p ) H ( θ ) , θ ∈ Θ , p ∈ Z d . (19)
Here U ( p ) \mathcal{U}(p) U ( p ) is defined in (17) .
For Z ⊂ Z d Z \subset\mathbb{Z}^d Z ⊂ Z d , let Π Z : ℓ 2 ( Z d ) → ℓ 2 ( Z ) \Pi_Z:\ell^2(\mathbb{Z}^d)\to\ell^2(Z) Π Z : ℓ 2 ( Z d ) → ℓ 2 ( Z ) be restriction to Z Z Z . The adjoint Π Z ∗ \Pi_Z^* Π Z ∗ extends a vector by zero outside Z Z Z . For finite Z Z Z , θ ∈ Θ \theta\in\Theta θ ∈ Θ and z ∈ C z\in\mathbb{C} z ∈ C , let
H Z ( θ ) = Π Z H ( θ ) Π Z ∗ , T Z = Π Z T Π Z ∗ , (20) \mathcal{H}_Z(\theta)=\Pi_Z\mathcal{H}(\theta)\Pi_Z^*, \qquad\mathcal{T}_Z=\Pi_Z\mathcal{T}\Pi_Z^*, \tag*{(20)} H Z ( θ ) = Π Z H ( θ ) Π Z ∗ , T Z = Π Z T Π Z ∗ , (20)
A Z ( z ; θ ) = z I Z − H Z ( θ ) . (21) A_Z(z;\theta)=zI_Z-\mathcal{H}_Z(\theta). \tag*{(21)} A Z ( z ; θ ) = z I Z − H Z ( θ ) . (21)
When A Z ( z ; θ ) A_Z(z;\theta) A Z ( z ; θ ) is invertible, A Z ( z ; θ ) − 1 A_Z(z;\theta)^{-1} A Z ( z ; θ ) − 1 is the finite-volume Green’s function.
Schur complements
Let M \mathcal{M} M be a matrix on a finite disjoint union P ⊔ Q P\sqcup Q P ⊔ Q and assume that M Q \mathcal{M}_Q M Q is invertible. The Schur complement after eliminating Q Q Q is
Sch Q ( M ) = M P − M P , Q M Q − 1 M Q , P . (22) \operatorname{Sch}_Q(\mathcal{M})=\mathcal{M}_P-\mathcal{M}_{P,Q}\mathcal{M}_Q^{-1}\mathcal{M}_{Q,P}. \tag*{(22)} Sch Q ( M ) = M P − M P , Q M Q − 1 M Q , P . (22)
For the decomposition P ⊔ Q P\sqcup Q P ⊔ Q , define W ( M ) : C P → C P ⊔ Q \mathcal{W}(\mathcal{M}):\mathbb{C}^P\to\mathbb{C}^{P\sqcup Q} W ( M ) : C P → C P ⊔ Q by
W ( M ) = ( I P − M Q − 1 M Q , P ) , (23) \mathcal{W}(\mathcal{M})=
\begin{pmatrix}
I_P\\
-\mathcal{M}_Q^{-1}\mathcal{M}_{Q,P}
\end{pmatrix}, \tag*{(23)} W ( M ) = ( I P − M Q − 1 M Q , P ) , (23)
with row blocks in the order P , Q P,Q P , Q and column set P P P . Solving the equations with indices in Q Q Q , we obtain
M v = 0 ⟺ v ∣ Q = − M Q − 1 M Q , P v ∣ P , Sch Q ( M ) v ∣ P = 0 , \mathcal{M}v=0 \Longleftrightarrow v|_Q=-\mathcal{M}_Q^{-1}\mathcal{M}_{Q,P}v|_P,\qquad\operatorname{Sch}_Q(\mathcal{M})v|_P=0, M v = 0 ⟺ v ∣ Q = − M Q − 1 M Q , P v ∣ P , Sch Q ( M ) v ∣ P = 0 ,
ker M = W ( M ) ker Sch Q ( M ) . \ker\mathcal{M}=\mathcal{W}(\mathcal{M})\ker\operatorname{Sch}_Q(\mathcal{M}). ker M = W ( M ) ker Sch Q ( M ) .
Block multiplication also proves
M W ( M ) = ( Sch Q ( M ) 0 ) , det M = det M Q det Sch Q ( M ) . \mathcal{M}\mathcal{W}(\mathcal{M})=
\begin{pmatrix}
\operatorname{Sch}_Q(\mathcal{M})\\
0
\end{pmatrix},
\qquad
\det\mathcal{M}=\det\mathcal{M}_Q\det\operatorname{Sch}_Q(\mathcal{M}). M W ( M ) = ( Sch Q ( M ) 0 ) , det M = det M Q det Sch Q ( M ) .
If M = M ∗ \mathcal{M}=\mathcal{M}^* M = M ∗ , then W ( M ) ∗ M W ( M ) = Sch Q ( M ) \mathcal{W}(\mathcal{M})^*\mathcal{M}\mathcal{W}(\mathcal{M})=\operatorname{Sch}_Q(\mathcal{M}) W ( M ) ∗ M W ( M ) = Sch Q ( M ) .
Let M \mathcal{M} M act on a finite disjoint union P ⊔ Q ⊔ U P\sqcup Q\sqcup U P ⊔ Q ⊔ U . If M Q \mathcal{M}_Q M Q and ( Sch Q ( M ) ) U (\operatorname{Sch}_Q(\mathcal{M}))_U ( Sch Q ( M ) ) U are invertible, then M Q ∪ U \mathcal{M}_{Q\cup U} M Q ∪ U is invertible and
Sch U ( Sch Q ( M ) ) = Sch Q ∪ U ( M ) . (24) \operatorname{Sch}_U\left(\operatorname{Sch}_Q(\mathcal{M})\right)=\operatorname{Sch}_{Q\cup U}(\mathcal{M}). \tag*{(24)} Sch U ( Sch Q ( M ) ) = Sch Q ∪ U ( M ) . (24)
Indeed, block multiplication gives det M Q ∪ U = det M Q det ( Sch Q ( M ) ) U ≠ 0 \det\mathcal{M}_{Q\cup U}=\det\mathcal{M}_Q\det(\operatorname{Sch}_Q(\mathcal{M}))_U\ne0 det M Q ∪ U = det M Q det ( Sch Q ( M ) ) U = 0 , so M Q ∪ U \mathcal{M}_{Q\cup U} M Q ∪ U is invertible. For each prescribed v ∣ P v|_P v ∣ P , solving first for v ∣ Q v|_Q v ∣ Q and then for v ∣ U v|_U v ∣ U produces the unique solution of the equations with indices in Q ∪ U Q\cup U Q ∪ U . Substitution into the equations indexed by P P P proves (24) . For Q = ∅ Q=\varnothing Q = ∅ , sums over Q Q Q are zero, I ∅ I_{\varnothing} I ∅ and M ∅ − 1 \mathcal{M}_{\varnothing}^{-1} M ∅ − 1 denote the unique operator on C ∅ \mathbb{C}^{\varnothing} C ∅ , and Sch ∅ ( M ) = M \operatorname{Sch}_{\varnothing}(\mathcal{M})=\mathcal{M} Sch ∅ ( M ) = M .
For Hermitian matrices M , M ~ \mathcal{M},\widetilde{\mathcal{M}} M , M on the same finite set P ⊔ Q P\sqcup Q P ⊔ Q , assume that M Q \mathcal{M}_Q M Q and M ~ Q \widetilde{\mathcal{M}}_Q M Q are invertible. Then
Sch Q ( M ~ ) − Sch Q ( M ) = W ( M ) ∗ ( M ~ − M ) W ( M ~ ) . (25) \operatorname{Sch}_Q(\widetilde{\mathcal{M}})-\operatorname{Sch}_Q(\mathcal{M})=\mathcal{W}(\mathcal{M})^*(\widetilde{\mathcal{M}}-\mathcal{M})\mathcal{W}(\widetilde{\mathcal{M}}). \tag*{(25)} Sch Q ( M ) − Sch Q ( M ) = W ( M ) ∗ ( M − M ) W ( M ) . (25)
Here W ( M ) \mathcal{W}(\mathcal{M}) W ( M ) and W ( M ~ ) \mathcal{W}(\widetilde{\mathcal{M}}) W ( M ) are defined by (23) with row blocks P , Q P,Q P , Q . By (23) and the Hermitian symmetry of M \mathcal{M} M and M ~ \widetilde{\mathcal{M}} M , block multiplication gives
W ( M ) ∗ M W ( M ~ ) = Sch Q ( M ) , W ( M ) ∗ M ~ W ( M ~ ) = Sch Q ( M ~ ) . \mathcal{W}(\mathcal{M})^*\mathcal{M}\mathcal{W}(\widetilde{\mathcal{M}})=\operatorname{Sch}_Q(\mathcal{M}),\qquad
\mathcal{W}(\mathcal{M})^*\widetilde{\mathcal{M}}\mathcal{W}(\widetilde{\mathcal{M}})=\operatorname{Sch}_Q(\widetilde{\mathcal{M}}). W ( M ) ∗ M W ( M ) = Sch Q ( M ) , W ( M ) ∗ M W ( M ) = Sch Q ( M ) .
Subtracting the identities for W ( M ) ∗ M ~ W ( M ~ ) \mathcal{W}(\mathcal{M})^*\widetilde{\mathcal{M}}\mathcal{W}(\widetilde{\mathcal{M}}) W ( M ) ∗ M W ( M ) and W ( M ) ∗ M W ( M ~ ) \mathcal{W}(\mathcal{M})^*\mathcal{M}\mathcal{W}(\widetilde{\mathcal{M}}) W ( M ) ∗ M W ( M ) proves (25) .
Weighted matrix norms
For a matrix M \mathcal{M} M with rows in X ⊂ Z d X \subset\mathbb{Z}^d X ⊂ Z d and columns in Y ⊂ Z d Y \subset\mathbb{Z}^d Y ⊂ Z d , define
∥ M ∥ ρ = sup y ∈ Y ∑ x ∈ X ∣ M ( x , y ) ∣ ⟨ x − y ⟩ ρ , ρ ≥ 0. (26) \|\mathcal{M}\|_{\rho} = \sup_{y \in Y} \sum_{x \in X} |\mathcal{M}(x,y)|\langle x-y\rangle^{\rho}, \qquad\rho\ge0. \tag*{(26)} ∥ M ∥ ρ = y ∈ Y sup x ∈ X ∑ ∣ M ( x , y ) ∣ ⟨ x − y ⟩ ρ , ρ ≥ 0. (26)
If Y = ∅ Y = \varnothing Y = ∅ , let ∥ M ∥ ρ = 0 \|\mathcal{M}\|_{\rho} = 0 ∥ M ∥ ρ = 0 . When ∥ M ∥ 0 < ∞ \|\mathcal{M}\|_0 < \infty ∥ M ∥ 0 < ∞ , M \mathcal{M} M defines a bounded map from ℓ 1 ( Y ) \ell^1(Y) ℓ 1 ( Y ) to ℓ 1 ( X ) \ell^1(X) ℓ 1 ( X ) with operator norm ∥ M ∥ 0 \|\mathcal{M}\|_0 ∥ M ∥ 0 .
Lemma 2.1. Let ρ ≥ 0 \rho\ge0 ρ ≥ 0 and let X , Y , Z ⊂ Z d X,Y,Z \subset\mathbb{Z}^d X , Y , Z ⊂ Z d . For matrices M : X × Y → C \mathcal{M}: X \times Y \to\mathbb{C} M : X × Y → C and M ~ : Y × Z → C \widetilde{\mathcal{M}}: Y \times Z \to\mathbb{C} M : Y × Z → C with finite norms in (26) ,
∥ M M ~ ∥ ρ ≤ ∥ M ∥ ρ ∥ M ~ ∥ ρ . (27) \|\mathcal{M}\widetilde{\mathcal{M}}\|_{\rho} \le\|\mathcal{M}\|_{\rho}\|\widetilde{\mathcal{M}}\|_{\rho}. \tag*{(27)} ∥ M M ∥ ρ ≤ ∥ M ∥ ρ ∥ M ∥ ρ . (27) For fixed X X X , matrices on X × X X \times X X × X with finite ∥ ⋅ ∥ ρ \|\cdot\|_{\rho} ∥ ⋅ ∥ ρ form a Banach algebra. If ∥ M ∥ ρ < 1 \|\mathcal{M}\|_{\rho} < 1 ∥ M ∥ ρ < 1 for a matrix on X × X X \times X X × X , then
( I X − M ) − 1 = ∑ k = 0 ∞ M k , ∥ ( I X − M ) − 1 ∥ ρ ≤ 1 1 − ∥ M ∥ ρ . (I_X-\mathcal{M})^{-1} = \sum_{k=0}^{\infty}\mathcal{M}^k,\qquad\|(I_X-\mathcal{M})^{-1}\|_{\rho} \le\frac{1}{1-\|\mathcal{M}\|_{\rho}}. ( I X − M ) − 1 = k = 0 ∑ ∞ M k , ∥ ( I X − M ) − 1 ∥ ρ ≤ 1 − ∥ M ∥ ρ 1 .
Proof. The inequality ⟨ x − w ⟩ ≤ ⟨ x − y ⟩ ⟨ y − w ⟩ \langle x-w\rangle\le\langle x-y\rangle\langle y-w\rangle ⟨ x − w ⟩ ≤ ⟨ x − y ⟩ ⟨ y − w ⟩ implies
∑ x ∈ X ∣ ( M M ~ ) ( x , w ) ∣ ⟨ x − w ⟩ ρ ≤ ∑ y ∈ Y ∣ M ~ ( y , w ) ∣ ⟨ y − w ⟩ ρ ∑ x ∈ X ∣ M ( x , y ) ∣ ⟨ x − y ⟩ ρ ≤ ∥ M ∥ ρ ∥ M ~ ∥ ρ , w ∈ Z . \begin{aligned}
\sum_{x \in X}|(\mathcal{M}\widetilde{\mathcal{M}})(x,w)|\langle x-w\rangle^\rho
\le\sum_{y \in Y}|\widetilde{\mathcal{M}}(y,w)|\langle y-w\rangle^\rho\sum_{x \in X}|\mathcal{M}(x,y)|\langle x-y\rangle^\rho\\
&\le\|\mathcal{M}\|_{\rho}\|\widetilde{\mathcal{M}}\|_{\rho},\qquad w \in Z.
\end{aligned} x ∈ X ∑ ∣ ( M M ) ( x , w ) ∣ ⟨ x − w ⟩ ρ ≤ y ∈ Y ∑ ∣ M ( y , w ) ∣ ⟨ y − w ⟩ ρ x ∈ X ∑ ∣ M ( x , y ) ∣ ⟨ x − y ⟩ ρ ≤ ∥ M ∥ ρ ∥ M ∥ ρ , w ∈ Z .
Taking the supremum over w w w proves (27) . If ( M j ) (\mathcal{M}_j) ( M j ) is Cauchy in ∥ ⋅ ∥ ρ \|\cdot\|_{\rho} ∥ ⋅ ∥ ρ , every entry has a limit M ( x , y ) \mathcal{M}(x,y) M ( x , y ) . By Fatou’s lemma,
∥ M j − M ∥ ρ ≤ lim inf k → ∞ ∥ M j − M k ∥ ρ ⟶ 0. \|\mathcal{M}_j-\mathcal{M}\|_{\rho} \le\liminf_{k\to\infty}\|\mathcal{M}_j-\mathcal{M}_k\|_{\rho} \longrightarrow0. ∥ M j − M ∥ ρ ≤ k → ∞ lim inf ∥ M j − M k ∥ ρ ⟶ 0.
This proves completeness. If ∥ M ∥ ρ < 1 \|\mathcal{M}\|_{\rho}<1 ∥ M ∥ ρ < 1 , then (27) gives ∑ k ≥ 0 ∥ M k ∥ ρ ≤ ( 1 − ∥ M ∥ ρ ) − 1 \sum_{k\ge0}\|\mathcal{M}^k\|_{\rho}\le(1-\|\mathcal{M}\|_{\rho})^{-1} ∑ k ≥ 0 ∥ M k ∥ ρ ≤ ( 1 − ∥ M ∥ ρ ) − 1 . The Neumann series in this Banach algebra therefore gives the inverse and the stated norm bound. □
Finite-dimensional spectral facts
For a Hermitian matrix M \mathcal{M} M on a finite set F F F of cardinality m m m , denote the eigenvalues, counted with multiplicity, by
λ 0 ( M ) ≤ ⋯ ≤ λ m − 1 ( M ) , \lambda_0(\mathcal{M}) \le\cdots\le\lambda_{m-1}(\mathcal{M}), λ 0 ( M ) ≤ ⋯ ≤ λ m − 1 ( M ) ,
and denote the spectrum by Spec M \operatorname{Spec}\mathcal{M} Spec M . We recall the min–max principle, Weyl’s inequality and Cauchy interlacing; see [7 ] and [15 Chapter II] . For 0 ≤ j < m 0 \le j < m 0 ≤ j < m ,
λ j ( M ) = min V ⊂ C F dim V = j + 1 max v ∈ V ∥ v ∥ 2 = 1 ⟨ v , M v ⟩ . (28) \lambda_j(\mathcal{M}) =
\min_{\substack{V \subset\mathbb{C}^F\\ \dim V=j+1}}
\max_{\substack{v \in V\\ \|v\|_2=1}}
\langle v,\mathcal{M}v\rangle. \tag*{(28)} λ j ( M ) = V ⊂ C F d i m V = j + 1 min v ∈ V ∥ v ∥ 2 = 1 max ⟨ v , M v ⟩ . (28)
In particular, if M , M ~ \mathcal{M},\widetilde{\mathcal{M}} M , M are Hermitian matrices on the same set F F F and c ∈ R c \in\mathbb{R} c ∈ R , then
M ~ ≥ M + c I F ⟹ λ j ( M ~ ) ≥ λ j ( M ) + c , 0 ≤ j < m , (29) \widetilde{\mathcal{M}} \ge\mathcal{M}+cI_F \quad\Longrightarrow\quad\lambda_j(\widetilde{\mathcal{M}}) \ge\lambda_j(\mathcal{M})+c,\qquad0 \le j < m, \tag*{(29)} M ≥ M + c I F ⟹ λ j ( M ) ≥ λ j ( M ) + c , 0 ≤ j < m , (29)
∣ λ j ( M ) − λ j ( M ~ ) ∣ ≤ ∥ M − M ~ ∥ o p , 0 ≤ j < m . (30) |\lambda_j(\mathcal{M})-\lambda_j(\widetilde{\mathcal{M}})| \le\|\mathcal{M}-\widetilde{\mathcal{M}}\|_{\mathrm{op}},\qquad0 \le j < m. \tag*{(30)} ∣ λ j ( M ) − λ j ( M ) ∣ ≤ ∥ M − M ∥ op , 0 ≤ j < m . (30)
The matrix order in (29) means ⟨ v , ( M ~ − M ) v ⟩ ≥ c ∥ v ∥ 2 2 \langle v,(\widetilde{\mathcal{M}}-\mathcal{M})v\rangle\ge c\|v\|_2^2 ⟨ v , ( M − M ) v ⟩ ≥ c ∥ v ∥ 2 2 for every v ∈ C F v \in\mathbb{C}^F v ∈ C F . If P ⊂ F P \subset F P ⊂ F and # P = m − k ≥ 1 \#P=m-k\ge1 # P = m − k ≥ 1 , then
λ j ( M ) ≤ λ j ( M P ) ≤ λ j + k ( M ) , 0 ≤ j < m − k . \lambda_j(\mathcal{M}) \le\lambda_j(\mathcal{M}_P) \le\lambda_{j+k}(\mathcal{M}),\qquad0 \le j < m-k. λ j ( M ) ≤ λ j ( M P ) ≤ λ j + k ( M ) , 0 ≤ j < m − k .
By (30) , each ordered eigenvalue is continuous under continuous matrix perturbations, including when eigenvalues coincide.
The initial induction step
Assume r > τ r>\tau r > τ . We first solve for the coordinates at which the original diagonal entries are bounded away from zero. In a suitable finite region, the remaining equation involves only v ( o ) v(o) v ( o ) . We select a simple local eigenvalue by a homotopy in the coupling constant and recover the corresponding eigenvector by back substitution.
The equation and the first division of the lattice
Fix θ ∈ Θ \theta\in\Theta θ ∈ Θ , z ∈ R z\in\mathbb{R} z ∈ R and a finite set Ω ⊂ Z d \Omega\subset\mathbb{Z}^{d} Ω ⊂ Z d . For v ∈ C Ω v\in\mathbb{C}^{\Omega} v ∈ C Ω , the equation A Ω ( z ; θ ) v = 0 A_{\Omega}(z;\theta)v=0 A Ω ( z ; θ ) v = 0 reads
( z − f ( θ + p ⋅ ω ) ) v ( p ) − ε ∑ q ∈ Ω φ ( p − q ) v ( q ) = 0 , p ∈ Ω . (31) \left(z-f(\theta+p\cdot\omega)\right)v(p)-\varepsilon\sum_{q\in\Omega}\varphi(p-q)v(q)=0,\qquad p\in\Omega. \tag*{(31)} ( z − f ( θ + p ⋅ ω ) ) v ( p ) − ε q ∈ Ω ∑ φ ( p − q ) v ( q ) = 0 , p ∈ Ω. (31)
Here A Ω A_{\Omega} A Ω is defined in (21) . The diagonal entries z − f ( θ + p ⋅ ω ) z-f(\theta+p\cdot\omega) z − f ( θ + p ⋅ ω ) can be arbitrarily small. We first solve (31) at points satisfying ∣ z − f ( θ + p ⋅ ω ) ∣ ≥ β \lvert z-f(\theta+p\cdot\omega)\rvert\geq\beta ∣ z − f ( θ + p ⋅ ω )∣ ≥ β , where we fix
β = L γ 32 ⋅ 67200 τ , δ 0 = β , E 0 = f . (32) \beta=\frac{L\gamma}{32\cdot67200^{\tau}},\qquad\delta_{0}=\beta,\qquad E_{0}=f. \tag*{(32)} β = 32 ⋅ 6720 0 τ L γ , δ 0 = β , E 0 = f . (32)
Then we define
S 0 = { p ∈ Z d : ∣ z − f ( θ + p ⋅ ω ) ∣ < β } , Q 0 = Z d ∖ S 0 . (33) S_{0}=\{p\in\mathbb{Z}^{d}:\lvert z-f(\theta+p\cdot\omega)\rvert<\beta\},\qquad Q_{0}=\mathbb{Z}^{d}\setminus S_{0}. \tag*{(33)} S 0 = { p ∈ Z d : ∣ z − f ( θ + p ⋅ ω )∣ < β } , Q 0 = Z d ∖ S 0 . (33)
Let
a = ∣ ε ∣ sup k ∈ Z d ⟨ k ⟩ r ∣ φ ( k ) ∣ . (34) a=\lvert\varepsilon\rvert\sup_{k\in\mathbb{Z}^{d}}\langle k\rangle^{r}\lvert\varphi(k)\rvert. \tag*{(34)} a = ∣ ε ∣ k ∈ Z d sup ⟨ k ⟩ r ∣ φ ( k )∣ . (34)
For p , q ∈ Ω p,q\in\Omega p , q ∈ Ω with p ≠ q p\neq q p = q , assumption (5) implies
∣ A Ω ( z ; θ ) ( p , q ) ∣ = ∣ ε φ ( p − q ) ∣ ≤ a ⟨ p − q ⟩ − r . (35) \lvert A_{\Omega}(z;\theta)(p,q)\rvert=\lvert\varepsilon\varphi(p-q)\rvert\leq a\langle p-q\rangle^{-r}. \tag*{(35)} ∣ A Ω ( z ; θ ) ( p , q )∣ = ∣ εφ ( p − q )∣ ≤ a ⟨ p − q ⟩ − r . (35)
Solving for the components in Q0
Let I 0 = Q 0 ∩ Ω I_{0}=Q_{0}\cap\Omega I 0 = Q 0 ∩ Ω and P 0 = S 0 ∩ Ω P_{0}=S_{0}\cap\Omega P 0 = S 0 ∩ Ω . Write A Ω ( − 1 ) = A Ω ( z ; θ ) A_{\Omega}^{(-1)}=A_{\Omega}(z;\theta) A Ω ( − 1 ) = A Ω ( z ; θ ) . Lemma 3.1 expresses v ∣ I 0 v|_{I_{0}} v ∣ I 0 in terms of v ∣ P 0 v|_{P_{0}} v ∣ P 0 and bounds Sch I 0 ( A Ω ( − 1 ) ) \operatorname{Sch}I_{0}(A_{\Omega}^{(-1)}) Sch I 0 ( A Ω ( − 1 ) ) .
Fix a constant C 3 = C 3 ( d , r ) ≥ 16 C_{3}=C_{3}(d,r)\geq16 C 3 = C 3 ( d , r ) ≥ 16 for the sparse convolution and tail estimates in Lemma A.1, chosen also so that
∑ k ∈ Z d ⟨ k ⟩ − r ≤ C 3 . (36) \sum_{k\in\mathbb{Z}^{d}}\langle k\rangle^{-r}\leq C_{3}. \tag*{(36)} k ∈ Z d ∑ ⟨ k ⟩ − r ≤ C 3 . (36)
The proof of Lemma A.1 gives a common finite bound for the convolution, tail and full-lattice sums, depending only on d , r d,r d , r .
Lemma 3.1. If a ≤ ( 2 C 3 ) − 1 β a\leq(2C_{3})^{-1}\beta a ≤ ( 2 C 3 ) − 1 β , then ( A Ω ( − 1 ) ) I 0 (A_{\Omega}^{(-1)})_{I_{0}} ( A Ω ( − 1 ) ) I 0 is invertible for every θ ∈ Θ \theta\in\Theta θ ∈ Θ , z ∈ R z\in\mathbb{R} z ∈ R and finite Ω ⊂ Z d \Omega\subset\mathbb{Z}^{d} Ω ⊂ Z d . Moreover, the equation (31) is equivalent to
v ∣ I 0 = − ( A Ω ( − 1 ) ) I 0 − 1 ( A Ω ( − 1 ) ) I 0 , P 0 v ∣ P 0 , (37) v|_{I_{0}}=-\left(A_{\Omega}^{(-1)}\right)_{I_{0}}^{-1}\left(A_{\Omega}^{(-1)}\right)_{I_{0},P_{0}}v|_{P_{0}}, \tag*{(37)} v ∣ I 0 = − ( A Ω ( − 1 ) ) I 0 − 1 ( A Ω ( − 1 ) ) I 0 , P 0 v ∣ P 0 , (37) A Ω ( 0 ) v ∣ P 0 = 0 , (38) A_{\Omega}^{(0)}v|_{P_{0}}=0, \tag*{(38)} A Ω ( 0 ) v ∣ P 0 = 0 , (38) where
A Ω ( 0 ) = ( A Ω ( − 1 ) ) P 0 − ( A Ω ( − 1 ) ) P 0 , I 0 ( A Ω ( − 1 ) ) I 0 − 1 ( A Ω ( − 1 ) ) I 0 , P 0 . A_{\Omega}^{(0)}=\left(A_{\Omega}^{(-1)}\right)_{P_{0}}-\left(A_{\Omega}^{(-1)}\right)_{P_{0},I_{0}}\left(A_{\Omega}^{(-1)}\right)_{I_{0}}^{-1}\left(A_{\Omega}^{(-1)}\right)_{I_{0},P_{0}}. A Ω ( 0 ) = ( A Ω ( − 1 ) ) P 0 − ( A Ω ( − 1 ) ) P 0 , I 0 ( A Ω ( − 1 ) ) I 0 − 1 ( A Ω ( − 1 ) ) I 0 , P 0 . Furthermore, for p ∈ P 0 p\in P_{0} p ∈ P 0 ,
∣ A Ω ( 0 ) ( p , p ) − ( z − f ( θ + p ⋅ ω ) ) ∣ ≤ 2 C 3 a 2 β − 1 ; (39) \left\lvert A_{\Omega}^{(0)}(p,p)-\left(z-f(\theta+p\cdot\omega)\right)\right\rvert\leq2C_{3}a^{2}\beta^{-1}; \tag*{(39)} A Ω ( 0 ) ( p , p ) − ( z − f ( θ + p ⋅ ω ) ) ≤ 2 C 3 a 2 β − 1 ; (39) and for distinct p , q ∈ P 0 p,q\in P_{0} p , q ∈ P 0 ,
∣ A Ω ( 0 ) ( p , q ) ∣ ≤ a ( 1 + 2 C 3 a β − 1 ) ⟨ p − q ⟩ − r ≤ 2 a ⟨ p − q ⟩ − r . (40) \lvert A_{\Omega}^{(0)}(p,q)\rvert\leq a(1+2C_{3}a\beta^{-1})\langle p-q\rangle^{-r}\leq2a\langle p-q\rangle^{-r}. \tag*{(40)} ∣ A Ω ( 0 ) ( p , q )∣ ≤ a ( 1 + 2 C 3 a β − 1 ) ⟨ p − q ⟩ − r ≤ 2 a ⟨ p − q ⟩ − r . (40)
Proof. By (33) ,
∣ A Ω ( − 1 ) ( p , p ) ∣ = ∣ z − f ( θ + p ⋅ ω ) ∣ ≥ β , p ∈ I 0 . (41) \left|A_{\Omega}^{(-1)}(p,p)\right|=\left|z-f(\theta+p\cdot\omega)\right|\ge\beta,\qquad p\in I_0. \tag*{(41)} A Ω ( − 1 ) ( p , p ) = ∣ z − f ( θ + p ⋅ ω ) ∣ ≥ β , p ∈ I 0 . (41)
Since Ω ⊂ Z d \Omega\subset\mathbb{Z}^d Ω ⊂ Z d is 1-separated and a / β ≤ ( 2 C 3 ) − 1 a/\beta\le(2C_3)^{-1} a / β ≤ ( 2 C 3 ) − 1 , (35) and (41) verify the assumptions of Lemma A.2 with X = Ω X=\Omega X = Ω , P = P 0 P=P_0 P = P 0 , Q = I 0 Q=I_0 Q = I 0 , M = 1 M=1 M = 1 , a ∗ = a a_*=a a ∗ = a and Δ = β \Delta=\beta Δ = β . Thus ( A Ω ( − 1 ) ) I 0 \left(A_{\Omega}^{(-1)}\right)_{I_0} ( A Ω ( − 1 ) ) I 0 is invertible. The equations (31) on I 0 I_0 I 0 imply (37) ; substituting (37) into (31) on P 0 P_0 P 0 proves (38) . Estimates (A.10) and (A.9) prove (39) and the first inequality in (40) . Since 2 C 3 a / β ≤ 1 2C_3a/\beta\le1 2 C 3 a / β ≤ 1 , we obtain the second inequality in (40) .
Construction of the first local eigenvalue and eigenvector
Lemma 3.1 expresses v ∣ I 0 v|_{I_0} v ∣ I 0 in terms of v ∣ P 0 v|_{P_0} v ∣ P 0 , but (38) remains to be solved on P 0 P_0 P 0 . We choose a fixed finite region around o o o in which all other components can be solved in terms of v ( o ) v(o) v ( o ) for energies close to f ( θ ) f(\theta) f ( θ ) . Let
N 0 = 1 , N 1 = 16 , B 0 = { o } , B 1 = Λ 4 N 1 . (42) N_0=1,\qquad N_1=16,\qquad B_0=\{o\},\qquad B_1=\Lambda_{4N_1}. \tag*{(42)} N 0 = 1 , N 1 = 16 , B 0 = { o } , B 1 = Λ 4 N 1 . (42)
The set Λ 4 N 1 \Lambda_{4N_1} Λ 4 N 1 is defined in (16) . For p ∈ B 1 ∘ p\in B_1^\circ p ∈ B 1 ∘ , periodicity and monotonicity in (7) –(8) , together with (2) and (32) , imply
∣ f ( θ + p ⋅ ω ) − f ( θ ) ∣ ≥ L ∥ p ⋅ ω ∥ T ≥ L γ ( 1 + 4 N 1 ) − τ > 8 β . (43) \left|f(\theta+p\cdot\omega)-f(\theta)\right|\ge L\|p\cdot\omega\|_{\mathbb{T}}\ge L\gamma(1+4N_1)^{-\tau}>8\beta. \tag*{(43)} ∣ f ( θ + p ⋅ ω ) − f ( θ ) ∣ ≥ L ∥ p ⋅ ω ∥ T ≥ L γ ( 1 + 4 N 1 ) − τ > 8 β . (43)
Thus, with
J 1 ( θ ) = [ f ( θ ) − 4 β , f ( θ ) + 4 β ] , J_1(\theta)=[f(\theta)-4\beta,f(\theta)+4\beta], J 1 ( θ ) = [ f ( θ ) − 4 β , f ( θ ) + 4 β ] ,
we obtain
∣ z − f ( θ + p ⋅ ω ) ∣ > 4 β , z ∈ J 1 ( θ ) , p ∈ B 1 ∘ . (44) \left|z-f(\theta+p\cdot\omega)\right|>4\beta,\qquad z\in J_1(\theta),\quad p\in B_1^\circ. \tag*{(44)} ∣ z − f ( θ + p ⋅ ω ) ∣ > 4 β , z ∈ J 1 ( θ ) , p ∈ B 1 ∘ . (44)
In particular, B 1 ∘ ⊂ Q 0 B_1^\circ\subset Q_0 B 1 ∘ ⊂ Q 0 for z ∈ J 1 ( θ ) z\in J_1(\theta) z ∈ J 1 ( θ ) .
To select the eigenvalue issuing from f ( θ ) f(\theta) f ( θ ) , we vary the coupling constant. For t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , define
H ( θ ; t ) = diag p ∈ Z d f ( θ + p ⋅ ω ) + t ε T . (45) \mathcal{H}(\theta;t)=\operatorname{diag}_{p\in\mathbb{Z}^d}f(\theta+p\cdot\omega)+t\varepsilon\mathcal{T}. \tag*{(45)} H ( θ ; t ) = diag p ∈ Z d f ( θ + p ⋅ ω ) + tε T . (45)
For a finite set F ⊂ Z d F\subset\mathbb{Z}^d F ⊂ Z d , let
H F ( θ ; t ) = diag p ∈ F f ( θ + p ⋅ ω ) + t ε T F , A F ( z ; θ ; t ) = z I F − H F ( θ ; t ) . (46) \mathcal{H}_F(\theta;t)=\operatorname{diag}_{p\in F}f(\theta+p\cdot\omega)+t\varepsilon\mathcal{T}_F,\qquad A_F(z;\theta;t)=zI_F-\mathcal{H}_F(\theta;t). \tag*{(46)} H F ( θ ; t ) = diag p ∈ F f ( θ + p ⋅ ω ) + tε T F , A F ( z ; θ ; t ) = z I F − H F ( θ ; t ) . (46)
For fixed θ ∈ Θ \theta\in\Theta θ ∈ Θ , the matrix H F ( θ ; t ) \mathcal{H}_F(\theta;t) H F ( θ ; t ) is finite and Hermitian, and
H F ( θ ; t ) − H F ( θ ; s ) = ( t − s ) ε T F , s , t ∈ [ 0 , 1 ] . \mathcal{H}_F(\theta;t)-\mathcal{H}_F(\theta;s)=(t-s)\varepsilon\mathcal{T}_F,\qquad s,t\in[0,1]. H F ( θ ; t ) − H F ( θ ; s ) = ( t − s ) ε T F , s , t ∈ [ 0 , 1 ] .
By (30) , every ordered eigenvalue of H F ( θ ; t ) \mathcal{H}_F(\theta;t) H F ( θ ; t ) is continuous in t t t . At t = 1 t=1 t = 1 , H F ( θ ; 1 ) = H F ( θ ) \mathcal{H}_F(\theta;1)=\mathcal{H}_F(\theta) H F ( θ ; 1 ) = H F ( θ ) and A F ( z ; θ ; 1 ) = A F ( z ; θ ) A_F(z;\theta;1)=A_F(z;\theta) A F ( z ; θ ; 1 ) = A F ( z ; θ ) , with the right-hand sides defined in (20) and (21) . The parameters β \beta β , a a a and the set B 1 B_1 B 1 are fixed throughout t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . We use A Ω ( h ) ( z ; θ ; t ) A_{\Omega}^{(h)}(z;\theta;t) A Ω ( h ) ( z ; θ ; t ) , h = − 1 , 0 h=-1,0 h = − 1 , 0 , for the matrices before and after the first elimination. Since ∣ t ε φ ( p − q ) ∣ ≤ a ⟨ p − q ⟩ − r |t\varepsilon\varphi(p-q)|\le a\langle p-q\rangle^{-r} ∣ tεφ ( p − q ) ∣ ≤ a ⟨ p − q ⟩ − r , Lemma 3.1 yields (37) –(40) for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] whenever a ≤ ( 2 C 3 ) − 1 β a\le(2C_3)^{-1}\beta a ≤ ( 2 C 3 ) − 1 β .
Assume a ≤ ( 2 C 3 ) − 1 β a\le(2C_3)^{-1}\beta a ≤ ( 2 C 3 ) − 1 β for the rest of this subsection. By (44) and Lemma 3.1, A B 1 ∘ ( z ; θ ; t ) A_{B_1^\circ}(z;\theta;t) A B 1 ∘ ( z ; θ ; t ) is invertible for z ∈ J 1 ( θ ) z\in J_1(\theta) z ∈ J 1 ( θ ) and t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . Define
s 1 ( z ; θ ; t ) = z − f ( θ ) − c 1 ( t ) ∗ A B 1 ∘ ( z ; θ ; t ) − 1 c 1 ( t ) , c 1 ( t ) = − t ε Π B 1 ∘ T e o . (47) s_1(z;\theta;t)=z-f(\theta)-c_1(t)^*A_{B_1^\circ}(z;\theta;t)^{-1}c_1(t),\qquad c_1(t)=-t\varepsilon\Pi_{B_1^\circ}\mathcal{T}e_o. \tag*{(47)} s 1 ( z ; θ ; t ) = z − f ( θ ) − c 1 ( t ) ∗ A B 1 ∘ ( z ; θ ; t ) − 1 c 1 ( t ) , c 1 ( t ) = − tε Π B 1 ∘ T e o . (47)
Here s 1 s_1 s 1 is the Schur complement in (22) and c 1 ( t ) c_1(t) c 1 ( t ) is the column of A B 1 ( z ; θ ; t ) A_{B_1}(z;\theta;t) A B 1 ( z ; θ ; t ) with row set B 1 ∘ B_1^\circ B 1 ∘ and column index o o o . We write s 1 ( z ; θ ) = s 1 ( z ; θ ; 1 ) s_1(z;\theta)=s_1(z;\theta;1) s 1 ( z ; θ ) = s 1 ( z ; θ ; 1 ) . For z ∈ J 1 ( θ ) z\in J_1(\theta) z ∈ J 1 ( θ ) , the equation A B 1 ( z ; θ ; t ) v = 0 A_{B_1}(z;\theta;t)v=0 A B 1 ( z ; θ ; t ) v = 0 is equivalent to
v ∣ B 1 ∘ = − A B 1 ∘ ( z ; θ ; t ) − 1 c 1 ( t ) v ( o ) , s 1 ( z ; θ ; t ) v ( o ) = 0. v|_{B_1^\circ}=-A_{B_1^\circ}(z;\theta;t)^{-1}c_1(t)v(o),\qquad s_1(z;\theta;t)v(o)=0. v ∣ B 1 ∘ = − A B 1 ∘ ( z ; θ ; t ) − 1 c 1 ( t ) v ( o ) , s 1 ( z ; θ ; t ) v ( o ) = 0.
Every nonzero solution of A B 1 ( z ; θ ; t ) v = 0 A_{B_1}(z;\theta;t)v=0 A B 1 ( z ; θ ; t ) v = 0 has v ( o ) ≠ 0 v(o)\ne0 v ( o ) = 0 . Hence
z ∈ Spec H B 1 ( θ ; t ) ⟺ s 1 ( z ; θ ; t ) = 0 , z ∈ J 1 ( θ ) . z\in\operatorname{Spec}\mathcal{H}_{B_1}(\theta;t)\quad\Longleftrightarrow\quad s_1(z;\theta;t)=0,\qquad z\in J_1(\theta). z ∈ Spec H B 1 ( θ ; t ) ⟺ s 1 ( z ; θ ; t ) = 0 , z ∈ J 1 ( θ ) .
Block multiplication gives
det A B 1 ( z ; θ ; t ) = det A B 1 ∘ ( z ; θ ; t ) s 1 ( z ; θ ; t ) . (48) \det\mathcal{A}_{B_1}(z;\theta;t)=\det\mathcal{A}_{B_1^\circ}(z;\theta;t)\,s_1(z;\theta;t). \tag*{(48)} det A B 1 ( z ; θ ; t ) = det A B 1 ∘ ( z ; θ ; t ) s 1 ( z ; θ ; t ) . (48)
Since a / β ≤ ( 2 C 3 ) − 1 a/\beta\leq(2C_3)^{-1} a / β ≤ ( 2 C 3 ) − 1 , the bounds (35) and (44) verify the assumptions of Lemma A.2 for A B 1 ( z ; θ ; t ) \mathcal{A}_{B_1}(z;\theta;t) A B 1 ( z ; θ ; t ) with X = B 1 X=B_1 X = B 1 , P = { o } P=\{o\} P = { o } , Q = B 1 ∘ Q=B_1^\circ Q = B 1 ∘ , M = 1 M=1 M = 1 , a ∗ = a a_*=a a ∗ = a and Δ = β \Delta=\beta Δ = β . Thus (A.10) implies
∣ s 1 ( z ; θ ; t ) − ( z − f ( θ ) ) ∣ ≤ 2 C 3 a 2 β − 1 , z ∈ J 1 ( θ ) , t ∈ [ 0 , 1 ] . (49) \left|s_1(z;\theta;t)-(z-f(\theta))\right|\leq2C_3a^2\beta^{-1},\qquad z\in J_1(\theta),\quad t\in[0,1]. \tag*{(49)} ∣ s 1 ( z ; θ ; t ) − ( z − f ( θ )) ∣ ≤ 2 C 3 a 2 β − 1 , z ∈ J 1 ( θ ) , t ∈ [ 0 , 1 ] . (49)
Proposition 3.2. Assume a ≤ ( 2 C 3 ) − 1 β a\leq(2C_3)^{-1}\beta a ≤ ( 2 C 3 ) − 1 β . For θ ∈ Θ \theta\in\Theta θ ∈ Θ , let j = # { p ∈ B 1 : f ( θ + p ⋅ ω ) < f ( θ ) } j=\#\{p\in B_1:f(\theta+p\cdot\omega)<f(\theta)\} j = # { p ∈ B 1 : f ( θ + p ⋅ ω ) < f ( θ )} and define
E 1 ( θ ; t ) = λ j ( H B 1 ( θ ; t ) ) , t ∈ [ 0 , 1 ] , E 1 ( θ ) = E 1 ( θ ; 1 ) . (50) E_1(\theta;t)=\lambda_j\bigl(\mathcal{H}_{B_1}(\theta;t)\bigr),\qquad t\in[0,1],\qquad E_1(\theta)=E_1(\theta;1). \tag*{(50)} E 1 ( θ ; t ) = λ j ( H B 1 ( θ ; t ) ) , t ∈ [ 0 , 1 ] , E 1 ( θ ) = E 1 ( θ ; 1 ) . (50) Then E 1 ( ⋅ ; t ) E_1(\cdot;t) E 1 ( ⋅ ; t ) is real and 1 1 1 -periodic, t ↦ E 1 ( θ ; t ) t\mapsto E_1(\theta;t) t ↦ E 1 ( θ ; t ) is continuous and E 1 ( θ ; 0 ) = f ( θ ) E_1(\theta;0)=f(\theta) E 1 ( θ ; 0 ) = f ( θ ) . For every θ ∈ Θ \theta\in\Theta θ ∈ Θ and t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , E 1 ( θ ; t ) E_1(\theta;t) E 1 ( θ ; t ) is the unique zero of s 1 ( ⋅ ; θ ; t ) s_1(\cdot;\theta;t) s 1 ( ⋅ ; θ ; t ) in J 1 ( θ ) J_1(\theta) J 1 ( θ ) and satisfies
∣ E 1 ( θ ; t ) − f ( θ ) ∣ ≤ 2 C 3 a 2 β − 1 ≤ β / 16. (51) \left|E_1(\theta;t)-f(\theta)\right|\leq2C_3a^2\beta^{-1}\leq\beta/16. \tag*{(51)} ∣ E 1 ( θ ; t ) − f ( θ ) ∣ ≤ 2 C 3 a 2 β − 1 ≤ β /16. (51) The eigenvalue E 1 ( θ ; t ) E_1(\theta;t) E 1 ( θ ; t ) is simple, with
Spec H B 1 ( θ ; t ) ∩ J 1 ( θ ) = { E 1 ( θ ; t ) } , dist ( E 1 ( θ ; t ) , Spec H B 1 ( θ ; t ) ∖ { E 1 ( θ ; t ) } ) ≥ 3 β . (52) \operatorname{Spec}\mathcal{H}_{B_1}(\theta;t)\cap J_1(\theta)=\{E_1(\theta;t)\},\qquad
\operatorname{dist}\left(E_1(\theta;t),\operatorname{Spec}\mathcal{H}_{B_1}(\theta;t)\setminus\{E_1(\theta;t)\}\right)\geq3\beta. \tag*{(52)} Spec H B 1 ( θ ; t ) ∩ J 1 ( θ ) = { E 1 ( θ ; t )} , dist ( E 1 ( θ ; t ) , Spec H B 1 ( θ ; t ) ∖ { E 1 ( θ ; t )} ) ≥ 3 β . (52)
Proof. Fix θ ∈ Θ \theta\in\Theta θ ∈ Θ . By (49) and (36) ,
s 1 ( f ( θ ) − 4 β ; θ ; t ) < 0 < s 1 ( f ( θ ) + 4 β ; θ ; t ) , t ∈ [ 0 , 1 ] . s_1(f(\theta)-4\beta;\theta;t)<0<s_1(f(\theta)+4\beta;\theta;t),\qquad t\in[0,1]. s 1 ( f ( θ ) − 4 β ; θ ; t ) < 0 < s 1 ( f ( θ ) + 4 β ; θ ; t ) , t ∈ [ 0 , 1 ] .
Since A B 1 ∘ ( z ; θ ; t ) \mathcal{A}_{B_1^\circ}(z;\theta;t) A B 1 ∘ ( z ; θ ; t ) is invertible on J 1 ( θ ) J_1(\theta) J 1 ( θ ) , (48) implies
f ( θ ) ± 4 β ∉ Spec H B 1 ( θ ; t ) , t ∈ [ 0 , 1 ] . f(\theta)\pm4\beta\notin\operatorname{Spec}\mathcal{H}_{B_1}(\theta;t),\qquad t\in[0,1]. f ( θ ) ± 4 β ∈ / Spec H B 1 ( θ ; t ) , t ∈ [ 0 , 1 ] .
By (30) and f ( θ ) ± 4 β ∉ Spec H B 1 ( θ ; t ) f(\theta)\pm4\beta\notin\operatorname{Spec}\mathcal{H}_{B_1}(\theta;t) f ( θ ) ± 4 β ∈ / Spec H B 1 ( θ ; t ) , the functions t ↦ λ k ( H B 1 ( θ ; t ) ) − f ( θ ) ± 4 β t\mapsto\lambda_k(\mathcal{H}_{B_1}(\theta;t))-f(\theta)\pm4\beta t ↦ λ k ( H B 1 ( θ ; t )) − f ( θ ) ± 4 β are continuous and nonzero on [ 0 , 1 ] [0,1] [ 0 , 1 ] , hence have constant sign, for every 0 ≤ k < # B 1 0\leq k<\#B_1 0 ≤ k < # B 1 . At t = 0 t=0 t = 0 , (43) places exactly the eigenvalue with index j j j in J 1 ( θ ) J_1(\theta) J 1 ( θ ) . Hence
{ k : λ k ( H B 1 ( θ ; t ) ) ∈ J 1 ( θ ) } = { j } , t ∈ [ 0 , 1 ] . \{k:\lambda_k(\mathcal{H}_{B_1}(\theta;t))\in J_1(\theta)\}=\{j\},\qquad t\in[0,1]. { k : λ k ( H B 1 ( θ ; t )) ∈ J 1 ( θ )} = { j } , t ∈ [ 0 , 1 ] .
The eigenvalues are counted with multiplicity. Consequently, E 1 ( θ ; t ) E_1(\theta;t) E 1 ( θ ; t ) is simple and is the unique eigenvalue of H B 1 ( θ ; t ) \mathcal{H}_{B_1}(\theta;t) H B 1 ( θ ; t ) in J 1 ( θ ) J_1(\theta) J 1 ( θ ) . By (48) , s 1 ( ⋅ ; θ ; t ) s_1(\cdot;\theta;t) s 1 ( ⋅ ; θ ; t ) has the unique zero E 1 ( θ ; t ) E_1(\theta;t) E 1 ( θ ; t ) in J 1 ( θ ) J_1(\theta) J 1 ( θ ) . Taking z = E 1 ( θ ; t ) z=E_1(\theta;t) z = E 1 ( θ ; t ) in (49) proves (51) . Since Spec H B 1 ( θ ; t ) ∖ { E 1 ( θ ; t ) } \operatorname{Spec}\mathcal{H}_{B_1}(\theta;t)\setminus\{E_1(\theta;t)\} Spec H B 1 ( θ ; t ) ∖ { E 1 ( θ ; t )} lies outside J 1 ( θ ) J_1(\theta) J 1 ( θ ) ,
dist ( E 1 ( θ ; t ) , Spec H B 1 ( θ ; t ) ∖ { E 1 ( θ ; t ) } ) ≥ 4 β − 2 C 3 a 2 β − 1 ≥ 3 β . \operatorname{dist}\left(E_1(\theta;t),\operatorname{Spec}\mathcal{H}_{B_1}(\theta;t)\setminus\{E_1(\theta;t)\}\right)\geq4\beta-2C_3a^2\beta^{-1}\geq3\beta. dist ( E 1 ( θ ; t ) , Spec H B 1 ( θ ; t ) ∖ { E 1 ( θ ; t )} ) ≥ 4 β − 2 C 3 a 2 β − 1 ≥ 3 β .
This proves (52) . Since H B 1 ( θ ; t ) \mathcal{H}_{B_1}(\theta;t) H B 1 ( θ ; t ) and j = # { p ∈ B 1 : f ( θ + p ⋅ ω ) < f ( θ ) } j=\#\{p\in B_1:f(\theta+p\cdot\omega)<f(\theta)\} j = # { p ∈ B 1 : f ( θ + p ⋅ ω ) < f ( θ )} are 1 1 1 -periodic in θ \theta θ , (50) proves that E 1 ( ⋅ ; t ) E_1(\cdot;t) E 1 ( ⋅ ; t ) is 1 1 1 -periodic. □
Define the zero-extended vector v 1 , θ ( t ) v_{1,\theta}(t) v 1 , θ ( t ) by
( v 1 , θ ( t ) ) ( o ) = 1 , v 1 , θ ( t ) ∣ B 1 c = 0 , v 1 , θ ( t ) ∣ B 1 ∘ = − A B 1 ∘ ( E 1 ( θ ; t ) ; θ ; t ) − 1 c 1 ( t ) . (53) (v_{1,\theta}(t))(o)=1,\qquad v_{1,\theta}(t)|_{B_1^c}=0,\qquad v_{1,\theta}(t)|_{B_1^\circ}=-\mathcal{A}_{B_1^\circ}(E_1(\theta;t);\theta;t)^{-1}c_1(t). \tag*{(53)} ( v 1 , θ ( t )) ( o ) = 1 , v 1 , θ ( t ) ∣ B 1 c = 0 , v 1 , θ ( t ) ∣ B 1 ∘ = − A B 1 ∘ ( E 1 ( θ ; t ) ; θ ; t ) − 1 c 1 ( t ) . (53)
By (47) and s 1 ( E 1 ( θ ; t ) ; θ ; t ) = 0 s_1(E_1(\theta;t);\theta;t)=0 s 1 ( E 1 ( θ ; t ) ; θ ; t ) = 0 ,
H B 1 ( θ ; t ) v 1 , θ ( t ) ∣ B 1 = E 1 ( θ ; t ) v 1 , θ ( t ) ∣ B 1 . \mathcal{H}_{B_1}(\theta;t)v_{1,\theta}(t)|_{B_1}=E_1(\theta;t)v_{1,\theta}(t)|_{B_1}. H B 1 ( θ ; t ) v 1 , θ ( t ) ∣ B 1 = E 1 ( θ ; t ) v 1 , θ ( t ) ∣ B 1 .
Formula (53) also proves continuity of t ↦ v 1 , θ ( t ) t\mapsto v_{1,\theta}(t) t ↦ v 1 , θ ( t ) and v 1 , θ ( 0 ) = e o v_{1,\theta}(0)=e_o v 1 , θ ( 0 ) = e o . Write v 1 , θ = v 1 , θ ( 1 ) v_{1,\theta}=v_{1,\theta}(1) v 1 , θ = v 1 , θ ( 1 ) and v 0 , θ ( t ) = e o v_{0,\theta}(t)=e_o v 0 , θ ( t ) = e o . Fix 0 < ρ < r − τ 0<\rho<r-\tau 0 < ρ < r − τ and define
∥ u ∥ 1 , ρ = ∑ m ∈ Z d ∣ u ( m ) ∣ ⟨ m ⟩ ρ . (54) \|u\|_{1,\rho}=\sum_{m\in\mathbb{Z}^d}|u(m)|\langle m\rangle^\rho. \tag*{(54)} ∥ u ∥ 1 , ρ = m ∈ Z d ∑ ∣ u ( m ) ∣ ⟨ m ⟩ ρ . (54)
For each 0 ≤ s < r − d 0 \le s < r-d 0 ≤ s < r − d , fix a finite coefficient C 4 ( d , r , s ) > 0 C_4(d,r,s)>0 C 4 ( d , r , s ) > 0 for (A.12) with ρ = s \rho=s ρ = s . For the chosen ρ \rho ρ , set
C 4 = C 4 ( d , r , ρ ) . (55) C_4=C_4(d,r,\rho). \tag*{(55)} C 4 = C 4 ( d , r , ρ ) . (55)
For the decomposition B 1 = { o } ⊔ B 1 ∘ B_1=\{o\}\sqcup B_1^\circ B 1 = { o } ⊔ B 1 ∘ , the assumptions of Lemma A.2 were verified before (49) . Since ρ < r − τ < r − d \rho<r-\tau<r-d ρ < r − τ < r − d , (A.12) implies
sup θ ∈ Θ , t ∈ [ 0 , 1 ] ∥ v 1 , θ ( t ) − v 0 , θ ( t ) ∥ 1 , ρ ≤ C 4 4 a β − 1 . (56) \sup_{\theta\in\Theta,\,t\in[0,1]}\|v_{1,\theta}(t)-v_{0,\theta}(t)\|_{1,\rho}\le\frac{C_4}{4}a\beta^{-1}. \tag*{(56)} θ ∈ Θ , t ∈ [ 0 , 1 ] sup ∥ v 1 , θ ( t ) − v 0 , θ ( t ) ∥ 1 , ρ ≤ 4 C 4 a β − 1 . (56)
We call E 1 E_1 E 1 the first Rellich function. To select the next coordinates to solve in (38) using E 1 ( θ + p ⋅ ω ) E_1(\theta+p\cdot\omega) E 1 ( θ + p ⋅ ω ) , we prove that E 1 E_1 E 1 is Lipschitz monotone in the phase.
Reordered eigenvalue branches
Let F ⊂ Z d F\subset\mathbb{Z}^d F ⊂ Z d be finite with o ∈ F o\in F o ∈ F and m = # F m=\#F m = # F . For t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , order the eigenvalues of H F ( θ ; t ) \mathcal{H}_F(\theta;t) H F ( θ ; t ) increasingly, counting multiplicities, as in Subsection 2.5. Arrange the phases θ ∈ [ 0 , 1 ) \theta\in[0,1) θ ∈ [ 0 , 1 ) for which θ + p ⋅ ω ∈ Z \theta+p\cdot\omega\in\mathbb{Z} θ + p ⋅ ω ∈ Z for some p ∈ F p\in F p ∈ F as
0 = ϑ 0 < ϑ 1 < ⋯ < ϑ m − 1 < ϑ m = 1 , { ϑ 0 , … , ϑ m − 1 } = { { − p ⋅ ω } : p ∈ F } . (57) 0=\vartheta_0<\vartheta_1<\cdots<\vartheta_{m-1}<\vartheta_m=1,\qquad\{\vartheta_0,\ldots,\vartheta_{m-1}\}=\{\{-p\cdot\omega\}:p\in F\}. \tag*{(57)} 0 = ϑ 0 < ϑ 1 < ⋯ < ϑ m − 1 < ϑ m = 1 , { ϑ 0 , … , ϑ m − 1 } = {{ − p ⋅ ω } : p ∈ F } . (57)
For the fractional parts defined in (13) , (2) implies that { − p ⋅ ω } \{-p\cdot\omega\} { − p ⋅ ω } , p ∈ F p\in F p ∈ F , are pairwise distinct, and o ∈ F o\in F o ∈ F implies ϑ 0 = 0 \vartheta_0=0 ϑ 0 = 0 . For 0 ≤ j < m 0\le j<m 0 ≤ j < m , the functions θ ↦ f ( θ + p ⋅ ω ) \theta\mapsto f(\theta+p\cdot\omega) θ ↦ f ( θ + p ⋅ ω ) , p ∈ F p\in F p ∈ F , are Lipschitz monotone with constant L L L on ( ϑ j , ϑ j + 1 ) (\vartheta_j,\vartheta_{j+1}) ( ϑ j , ϑ j + 1 ) . By (28) , each function θ ↦ λ k ( H F ( θ ; t ) ) \theta\mapsto\lambda_k(\mathcal{H}_F(\theta;t)) θ ↦ λ k ( H F ( θ ; t )) , 0 ≤ k < m 0\le k<m 0 ≤ k < m , is Lipschitz monotone with constant L L L on ( ϑ j , ϑ j + 1 ) (\vartheta_j,\vartheta_{j+1}) ( ϑ j , ϑ j + 1 ) . At ϑ j = { − p ⋅ ω } \vartheta_j=\{-p\cdot\omega\} ϑ j = { − p ⋅ ω } , 1 ≤ j < m 1\le j<m 1 ≤ j < m , the fractional part { θ + p ⋅ ω } \{\theta+p\cdot\omega\} { θ + p ⋅ ω } returns to 0 0 0 . We therefore increase the selected eigenvalue index by one at each ϑ j \vartheta_j ϑ j , 1 ≤ j < m 1\le j<m 1 ≤ j < m :
λ ( F ; θ ; t ) = λ j ( H F ( θ ; t ) ) , θ ∈ [ ϑ j , ϑ j + 1 ) ∩ Θ , 0 ≤ j < m . (58) \lambda(F;\theta;t)=\lambda_j(\mathcal{H}_F(\theta;t)),\qquad\theta\in[\vartheta_j,\vartheta_{j+1})\cap\Theta,\quad0\le j<m. \tag*{(58)} λ ( F ; θ ; t ) = λ j ( H F ( θ ; t )) , θ ∈ [ ϑ j , ϑ j + 1 ) ∩ Θ , 0 ≤ j < m . (58)
Extend λ ( F ; ⋅ ; t ) \lambda(F;\cdot;t) λ ( F ; ⋅ ; t ) 1 1 1 -periodically to Θ \Theta Θ and write λ ( F ; θ ) = λ ( F ; θ ; 1 ) \lambda(F;\theta)=\lambda(F;\theta;1) λ ( F ; θ ) = λ ( F ; θ ; 1 ) . If θ ∈ [ ϑ j , ϑ j + 1 ) ∩ Θ \theta\in[\vartheta_j,\vartheta_{j+1})\cap\Theta θ ∈ [ ϑ j , ϑ j + 1 ) ∩ Θ , then exactly j j j diagonal entries of H F ( θ ; 0 ) \mathcal{H}_F(\theta;0) H F ( θ ; 0 ) are smaller than f ( θ ) f(\theta) f ( θ ) : for p ≠ o p\ne o p = o ,
{ θ + p ⋅ ω } < θ ⟺ { − p ⋅ ω } ≤ θ . \{\theta+p\cdot\omega\}<\theta\quad\Longleftrightarrow\quad\{-p\cdot\omega\}\le\theta. { θ + p ⋅ ω } < θ ⟺ { − p ⋅ ω } ≤ θ .
Thus the eigenvalue branch associated with the diagonal entry f ( θ ) f(\theta) f ( θ ) in [14 Remark 4.2] is precisely (58) . Hence
λ ( F ; θ ; 0 ) = f ( θ ) , θ ∈ Θ . (59) \lambda(F;\theta;0)=f(\theta),\qquad\theta\in\Theta. \tag*{(59)} λ ( F ; θ ; 0 ) = f ( θ ) , θ ∈ Θ. (59)
Proposition 3.3 ([14 Proposition 4.1] ). For every finite F ⊂ Z d F\subset\mathbb{Z}^d F ⊂ Z d containing o o o and every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , the function λ ( F ; ⋅ ; t ) \lambda(F;\cdot;t) λ ( F ; ⋅ ; t ) in (58) satisfies
λ ( F ; θ 2 ; t ) − λ ( F ; θ 1 ; t ) ≥ L ( θ 2 − θ 1 ) , θ 1 , θ 2 ∈ Θ ∩ [ 0 , 1 ) , θ 1 < θ 2 . (60) \lambda(F;\theta_2;t)-\lambda(F;\theta_1;t)\ge L(\theta_2-\theta_1),\qquad\theta_1,\theta_2\in\Theta\cap[0,1),\quad\theta_1<\theta_2. \tag*{(60)} λ ( F ; θ 2 ; t ) − λ ( F ; θ 1 ; t ) ≥ L ( θ 2 − θ 1 ) , θ 1 , θ 2 ∈ Θ ∩ [ 0 , 1 ) , θ 1 < θ 2 . (60)
For fixed t t t , the hopping matrix t ε T F t\varepsilon\mathcal{T}_F tε T F is Hermitian and independent of θ \theta θ , as required in [14 Remark 4.2] . Since f / L f/L f / L is Lipschitz monotone with constant 1 1 1 , [14 Proposition 4.1] applied to H F ( θ ; t ) / L \mathcal{H}_F(\theta;t)/L H F ( θ ; t ) / L proves (60) .
Taking F = B 1 F=B_1 F = B 1 in (58) and using (50) , we obtain
E 1 ( θ ; t ) = λ ( B 1 ; θ ; t ) , θ ∈ Θ , t ∈ [ 0 , 1 ] . (61) E_1(\theta;t)=\lambda(B_1;\theta;t),\qquad\theta\in\Theta,\quad t\in[0,1]. \tag*{(61)} E 1 ( θ ; t ) = λ ( B 1 ; θ ; t ) , θ ∈ Θ , t ∈ [ 0 , 1 ] . (61)
Proposition 3.3 thus proves E 1 ( θ 2 ; t ) − E 1 ( θ 1 ; t ) ≥ L ( θ 2 − θ 1 ) E_1(\theta_2;t)-E_1(\theta_1;t)\ge L(\theta_2-\theta_1) E 1 ( θ 2 ; t ) − E 1 ( θ 1 ; t ) ≥ L ( θ 2 − θ 1 ) for θ 1 , θ 2 ∈ Θ ∩ [ 0 , 1 ) \theta_1,\theta_2\in\Theta\cap[0,1) θ 1 , θ 2 ∈ Θ ∩ [ 0 , 1 ) with θ 1 < θ 2 \theta_1<\theta_2 θ 1 < θ 2 .
Homotopy on a moving energy interval
Proposition 3.2 counts the eigenvalues of H B 1 ( θ ; t ) \mathcal{H}_{B_1}(\theta;t) H B 1 ( θ ; t ) in the fixed interval J 1 ( θ ) J_1(\theta) J 1 ( θ ) . For the next local eigenvalue, the energy interval is centered at E 1 ( θ ; t ) E_1(\theta;t) E 1 ( θ ; t ) and therefore varies with t t t . Proposition 3.4 extends the eigenvalue count to an interval with a continuous center h ( θ ; t ) h(\theta;t) h ( θ ; t ) .
Proposition 3.4. Let F ⊂ Z d F \subset\mathbb{Z}^d F ⊂ Z d be finite with o ∈ F o \in F o ∈ F and let η > 0 \eta> 0 η > 0 . For each θ ∈ Θ \theta\in\Theta θ ∈ Θ , let t ↦ h ( θ ; t ) t \mapsto h(\theta;t) t ↦ h ( θ ; t ) be continuous on [ 0 , 1 ] [0,1] [ 0 , 1 ] , with h ( θ ; 0 ) = f ( θ ) h(\theta;0)=f(\theta) h ( θ ; 0 ) = f ( θ ) . Suppose that
∣ f ( θ + p ⋅ ω ) − f ( θ ) ∣ > η , p ∈ F ∘ , θ ∈ Θ , \left|f(\theta+p\cdot\omega)-f(\theta)\right|>\eta,\qquad p\in F^\circ,\qquad\theta\in\Theta, ∣ f ( θ + p ⋅ ω ) − f ( θ ) ∣ > η , p ∈ F ∘ , θ ∈ Θ , h ( θ ; t ) ± η ∉ Spec H F ( θ ; t ) , θ ∈ Θ , t ∈ [ 0 , 1 ] . h(\theta;t)\pm\eta\notin\operatorname{Spec}\mathcal{H}_F(\theta;t),\qquad\theta\in\Theta,\qquad t\in[0,1]. h ( θ ; t ) ± η ∈ / Spec H F ( θ ; t ) , θ ∈ Θ , t ∈ [ 0 , 1 ] . Then [ h ( θ ; t ) − η , h ( θ ; t ) + η ] [h(\theta;t)-\eta,h(\theta;t)+\eta] [ h ( θ ; t ) − η , h ( θ ; t ) + η ] contains exactly one eigenvalue of H F ( θ ; t ) \mathcal{H}_F(\theta;t) H F ( θ ; t ) , counted with multiplicity, namely λ ( F ; θ ; t ) \lambda(F;\theta;t) λ ( F ; θ ; t ) from (58) . The function λ ( F ; θ ; t ) \lambda(F;\theta;t) λ ( F ; θ ; t ) is continuous in t t t , is 1 1 1 -periodic in θ \theta θ and satisfies (60) .
Proof. Fix θ ∈ Θ \theta\in\Theta θ ∈ Θ and let j = # { p ∈ F : f ( θ + p ⋅ ω ) < f ( θ ) } j=\#\{p\in F:f(\theta+p\cdot\omega)<f(\theta)\} j = # { p ∈ F : f ( θ + p ⋅ ω ) < f ( θ )} . For each 0 ≤ k < # F 0\leq k<\#F 0 ≤ k < # F , the functions t ↦ λ k ( H F ( θ ; t ) ) − h ( θ ; t ) ± η t\mapsto\lambda_k(\mathcal{H}_F(\theta;t))-h(\theta;t)\pm\eta t ↦ λ k ( H F ( θ ; t )) − h ( θ ; t ) ± η are continuous and nonzero on [ 0 , 1 ] [0,1] [ 0 , 1 ] . The sign argument in the proof of Proposition 3.2, together with the eigenvalue count at t = 0 t=0 t = 0 , gives
{ k : ∣ λ k ( H F ( θ ; t ) ) − h ( θ ; t ) ∣ < η } = { j } , t ∈ [ 0 , 1 ] . \{k:\left|\lambda_k(\mathcal{H}_F(\theta;t))-h(\theta;t)\right|<\eta\}=\{j\},\qquad t\in[0,1]. { k : ∣ λ k ( H F ( θ ; t )) − h ( θ ; t ) ∣ < η } = { j } , t ∈ [ 0 , 1 ] .
The definition (58) and the count j = # { p ∈ F : f ( θ + p ⋅ ω ) < f ( θ ) } j=\#\{p\in F:f(\theta+p\cdot\omega)<f(\theta)\} j = # { p ∈ F : f ( θ + p ⋅ ω ) < f ( θ )} imply λ j ( H F ( θ ; t ) ) = λ ( F ; θ ; t ) \lambda_j(\mathcal{H}_F(\theta;t))=\lambda(F;\theta;t) λ j ( H F ( θ ; t )) = λ ( F ; θ ; t ) . By (30) , t ↦ λ ( F ; θ ; t ) t\mapsto\lambda(F;\theta;t) t ↦ λ ( F ; θ ; t ) is continuous. The periodic extension in (58) makes λ ( F ; ⋅ ; t ) \lambda(F;\cdot;t) λ ( F ; ⋅ ; t ) 1 1 1 -periodic, and Proposition 3.3 proves (60) . □
To verify h ( θ ; t ) ± η ∉ Spec H F ( θ ; t ) h(\theta;t)\pm\eta\notin\operatorname{Spec}\mathcal{H}_F(\theta;t) h ( θ ; t ) ± η ∈ / Spec H F ( θ ; t ) in Proposition 3.4, we compare the scalar Schur complement at two energies.
Lemma 3.5. Let F ⊂ Z d F\subset\mathbb{Z}^d F ⊂ Z d be finite with o ∈ F o\in F o ∈ F and fix θ ∈ Θ \theta\in\Theta θ ∈ Θ and t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . Suppose that A F ∘ ( z ; θ ; t ) \mathcal{A}_{F^\circ}(z;\theta;t) A F ∘ ( z ; θ ; t ) is invertible for every z z z in a real interval J J J . Denote the scalar Schur complement by
s ( z ; θ ; t ) = Sch F ∘ ( A F ( z ; θ ; t ) ) . s(z;\theta;t)=\operatorname{Sch}_{F^\circ}(\mathcal{A}_F(z;\theta;t)). s ( z ; θ ; t ) = Sch F ∘ ( A F ( z ; θ ; t )) . For z 1 < z 2 z_1<z_2 z 1 < z 2 in J J J ,
s ( z 2 ; θ ; t ) − s ( z 1 ; θ ; t ) ≥ z 2 − z 1 . (62) s(z_2;\theta;t)-s(z_1;\theta;t)\geq z_2-z_1. \tag*{(62)} s ( z 2 ; θ ; t ) − s ( z 1 ; θ ; t ) ≥ z 2 − z 1 . (62) If E ∈ J E\in J E ∈ J and s ( E ; θ ; t ) = 0 s(E;\theta;t)=0 s ( E ; θ ; t ) = 0 , then ∣ s ( z ; θ ; t ) ∣ ≥ ∣ z − E ∣ \left|s(z;\theta;t)\right|\geq\left|z-E\right| ∣ s ( z ; θ ; t ) ∣ ≥ ∣ z − E ∣ for every z ∈ J z\in J z ∈ J .
Proof. Let c = − t ε Π F ∘ T e o c=-t\varepsilon\Pi_{F^\circ}T e_o c = − tε Π F ∘ T e o and omit the fixed arguments θ , t \theta,t θ , t . Since Spec H F ∘ ( θ ; t ) ∩ J = ∅ \operatorname{Spec}\mathcal{H}_{F^\circ}(\theta;t)\cap J=\varnothing Spec H F ∘ ( θ ; t ) ∩ J = ∅ , for every μ ∈ Spec H F ∘ ( θ ; t ) \mu\in\operatorname{Spec}\mathcal{H}_{F^\circ}(\theta;t) μ ∈ Spec H F ∘ ( θ ; t ) we have ( ( z 2 − μ ) ( z 1 − μ ) ) − 1 > 0 ((z_2-\mu)(z_1-\mu))^{-1}>0 (( z 2 − μ ) ( z 1 − μ ) ) − 1 > 0 . Thus A F ∘ ( z 2 ) − 1 A F ∘ ( z 1 ) − 1 \mathcal{A}_{F^\circ}(z_2)^{-1}\mathcal{A}_{F^\circ}(z_1)^{-1} A F ∘ ( z 2 ) − 1 A F ∘ ( z 1 ) − 1 is positive semidefinite, and the resolvent identity gives
s ( z 2 ) − s ( z 1 ) = ( z 2 − z 1 ) ( 1 + c ∗ A F ∘ ( z 2 ) − 1 A F ∘ ( z 1 ) − 1 c ) ≥ z 2 − z 1 . s(z_2)-s(z_1)=(z_2-z_1)\left(1+c^*\mathcal{A}_{F^\circ}(z_2)^{-1}\mathcal{A}_{F^\circ}(z_1)^{-1}c\right)\geq z_2-z_1. s ( z 2 ) − s ( z 1 ) = ( z 2 − z 1 ) ( 1 + c ∗ A F ∘ ( z 2 ) − 1 A F ∘ ( z 1 ) − 1 c ) ≥ z 2 − z 1 .
Taking z 1 = E z_1=E z 1 = E when z > E z>E z > E and z 2 = E z_2=E z 2 = E when z < E z<E z < E in (62) proves ∣ s ( z ; θ ; t ) ∣ ≥ ∣ z − E ∣ \left|s(z;\theta;t)\right|\geq\left|z-E\right| ∣ s ( z ; θ ; t ) ∣ ≥ ∣ z − E ∣ . □
Since Proposition 3.2 identifies E 1 ( θ ; t ) E_1(\theta;t) E 1 ( θ ; t ) as the zero of s 1 ( ⋅ ; θ ; t ) s_1(\cdot;\theta;t) s 1 ( ⋅ ; θ ; t ) in J 1 ( θ ) J_1(\theta) J 1 ( θ ) , Lemma 3.5 with F = B 1 F=B_1 F = B 1 implies
∣ s 1 ( z ; θ ; t ) ∣ ≥ ∣ z − E 1 ( θ ; t ) ∣ , z ∈ J 1 ( θ ) , t ∈ [ 0 , 1 ] . (63) \left|s_1(z;\theta;t)\right|\geq\left|z-E_1(\theta;t)\right|,\qquad z\in J_1(\theta),\qquad t\in[0,1]. \tag*{(63)} ∣ s 1 ( z ; θ ; t ) ∣ ≥ ∣ z − E 1 ( θ ; t ) ∣ , z ∈ J 1 ( θ ) , t ∈ [ 0 , 1 ] . (63)
Continuing the elimination
Fix θ ∈ Θ \theta\in\Theta θ ∈ Θ , z ∈ R z \in\mathbb{R} z ∈ R , t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] and a finite set Ω ⊂ Z d \Omega\subset\mathbb{Z}^d Ω ⊂ Z d , and assume a ≤ ( 2 C 3 ) − 1 β a \le(2C_3)^{-1}\beta a ≤ ( 2 C 3 ) − 1 β . For fixed t t t , we omit t t t from the matrices, Schur complements and eigenvalue functions. By Lemma 3.1 , the remaining equation (38) is
A Ω ( 0 ) v ∣ P 0 = 0 , P 0 = S 0 ∩ Ω . \mathcal{A}_{\Omega}^{(0)}v|_{P_0}=0,\qquad P_0=S_0\cap\Omega. A Ω ( 0 ) v ∣ P 0 = 0 , P 0 = S 0 ∩ Ω.
The diagonal entries A Ω ( 0 ) ( p , p ) \mathcal{A}_{\Omega}^{(0)}(p,p) A Ω ( 0 ) ( p , p ) depend on Ω \Omega Ω . We first examine the equation in B 1 ( p ) B_1(p) B 1 ( p ) , defined in (15) .
For p ∈ S 0 p\in S_0 p ∈ S 0 , we have z ∈ J 1 ( θ + p ⋅ ω ) z\in J_1(\theta+p\cdot\omega) z ∈ J 1 ( θ + p ⋅ ω ) , so (44) implies B 1 ( p ) ∩ S 0 = { p } B_1(p)\cap S_0=\{p\} B 1 ( p ) ∩ S 0 = { p } . By (47) , (19) and (63) ,
A B 1 ( p ) ( 0 ) ( p , p ) = s 1 ( z ; θ + p ⋅ ω ) , ∣ s 1 ( z ; θ + p ⋅ ω ) ∣ ≥ ∣ z − E 1 ( θ + p ⋅ ω ) ∣ . \mathcal{A}_{B_1(p)}^{(0)}(p,p)=s_1(z;\theta+p\cdot\omega),\qquad|s_1(z;\theta+p\cdot\omega)|\ge|z-E_1(\theta+p\cdot\omega)|. A B 1 ( p ) ( 0 ) ( p , p ) = s 1 ( z ; θ + p ⋅ ω ) , ∣ s 1 ( z ; θ + p ⋅ ω ) ∣ ≥ ∣ z − E 1 ( θ + p ⋅ ω ) ∣.
We therefore select the next coordinates using the distance to E 1 ( θ + p ⋅ ω ) E_1(\theta+p\cdot\omega) E 1 ( θ + p ⋅ ω ) . With N 1 = 16 N_1=16 N 1 = 16 as in (42) , define
δ 1 = β N 1 − τ , (64) \delta_1=\beta N_1^{-\tau}, \tag*{(64)} δ 1 = β N 1 − τ , (64)
S 1 = { p ∈ S 0 : ∣ z − E 1 ( θ + p ⋅ ω ) ∣ < δ 1 } , Q 1 = S 0 ∖ S 1 . (65) S_1=\{p\in S_0:|z-E_1(\theta+p\cdot\omega)|<\delta_1\},\qquad Q_1=S_0\setminus S_1. \tag*{(65)} S 1 = { p ∈ S 0 : ∣ z − E 1 ( θ + p ⋅ ω ) ∣ < δ 1 } , Q 1 = S 0 ∖ S 1 . (65)
In particular, ∣ s 1 ( z ; θ + p ⋅ ω ) ∣ ≥ δ 1 |s_1(z;\theta+p\cdot\omega)|\ge\delta_1 ∣ s 1 ( z ; θ + p ⋅ ω ) ∣ ≥ δ 1 for p ∈ Q 1 p\in Q_1 p ∈ Q 1 . The points in S 1 S_1 S 1 satisfy a separation estimate that follows from the next lemma.
Lemma 3.6. Let E : Θ → R E:\Theta\to\mathbb{R} E : Θ → R be 1 1 1 -periodic and satisfy E ( θ 2 ) − E ( θ 1 ) ≥ L ( θ 2 − θ 1 ) E(\theta_2)-E(\theta_1)\ge L(\theta_2-\theta_1) E ( θ 2 ) − E ( θ 1 ) ≥ L ( θ 2 − θ 1 ) for θ 1 , θ 2 ∈ Θ ∩ [ 0 , 1 ) \theta_1,\theta_2\in\Theta\cap[0,1) θ 1 , θ 2 ∈ Θ ∩ [ 0 , 1 ) with θ 1 < θ 2 \theta_1<\theta_2 θ 1 < θ 2 . For distinct p , q ∈ Z d p,q\in\mathbb{Z}^d p , q ∈ Z d and θ ∈ Θ \theta\in\Theta θ ∈ Θ ,
∣ E ( θ + p ⋅ ω ) − E ( θ + q ⋅ ω ) ∣ ≥ L ∥ ( p − q ) ⋅ ω ∥ T ≥ L γ ⟨ p − q ⟩ − τ . (66) |E(\theta+p\cdot\omega)-E(\theta+q\cdot\omega)|\ge L\|(p-q)\cdot\omega\|_{\mathbb{T}}\ge L\gamma\langle p-q\rangle^{-\tau}. \tag*{(66)} ∣ E ( θ + p ⋅ ω ) − E ( θ + q ⋅ ω ) ∣ ≥ L ∥ ( p − q ) ⋅ ω ∥ T ≥ L γ ⟨ p − q ⟩ − τ . (66) Proof. Using the periodicity identity in (7) and the Lipschitz monotonicity inequality (8) , both assumed here for E E E , we obtain
∣ E ( θ + p ⋅ ω ) − E ( θ + q ⋅ ω ) ∣ ≥ L ∣ { θ + p ⋅ ω } − { θ + q ⋅ ω } ∣ ≥ L ∥ ( p − q ) ⋅ ω ∥ T . |E(\theta+p\cdot\omega)-E(\theta+q\cdot\omega)|\ge L|\{\theta+p\cdot\omega\}-\{\theta+q\cdot\omega\}|\ge L\|(p-q)\cdot\omega\|_{\mathbb{T}}. ∣ E ( θ + p ⋅ ω ) − E ( θ + q ⋅ ω ) ∣ ≥ L ∣ { θ + p ⋅ ω } − { θ + q ⋅ ω } ∣ ≥ L ∥ ( p − q ) ⋅ ω ∥ T . The Diophantine condition (2) now proves (66) . □
By (61) , Proposition 3.3 and Lemma 3.6 , distinct p , q ∈ S 1 p,q\in S_1 p , q ∈ S 1 satisfy
⟨ p − q ⟩ > ( L γ 2 β ) 1 / τ N 1 . (67) \langle p-q\rangle>\left(\frac{L\gamma}{2\beta}\right)^{1/\tau}N_1. \tag*{(67)} ⟨ p − q ⟩ > ( 2 β L γ ) 1/ τ N 1 . (67)
Let I 1 = Q 1 ∩ Ω I_1=Q_1\cap\Omega I 1 = Q 1 ∩ Ω and P 1 = S 1 ∩ Ω P_1=S_1\cap\Omega P 1 = S 1 ∩ Ω . The rows indexed by I 1 I_1 I 1 in (38) read
( A Ω ( 0 ) ) I 1 v ∣ I 1 + ( A Ω ( 0 ) ) I 1 , P 1 v ∣ P 1 = 0. (68) (\mathcal{A}_{\Omega}^{(0)})_{I_1}v|_{I_1}+(\mathcal{A}_{\Omega}^{(0)})_{I_1,P_1}v|_{P_1}=0. \tag*{(68)} ( A Ω ( 0 ) ) I 1 v ∣ I 1 + ( A Ω ( 0 ) ) I 1 , P 1 v ∣ P 1 = 0. (68)
To solve (68) for v ∣ I 1 v|_{I_1} v ∣ I 1 , we need ( A Ω ( 0 ) ) I 1 (\mathcal{A}_{\Omega}^{(0)})_{I_1} ( A Ω ( 0 ) ) I 1 to be invertible. The off-diagonal entries are already bounded by (40) . The local lower bound ∣ s 1 ( z ; θ + p ⋅ ω ) ∣ ≥ δ 1 |s_1(z;\theta+p\cdot\omega)|\ge\delta_1 ∣ s 1 ( z ; θ + p ⋅ ω ) ∣ ≥ δ 1 will imply ∣ A Ω ( 0 ) ( p , p ) ∣ ≥ δ 1 / 2 |\mathcal{A}_{\Omega}^{(0)}(p,p)|\ge\delta_1/2 ∣ A Ω ( 0 ) ( p , p ) ∣ ≥ δ 1 /2 if we prove
∣ A Ω ( 0 ) ( p , p ) − s 1 ( z ; θ + p ⋅ ω ) ∣ ≤ δ 1 / 2 , p ∈ I 1 . (69) |\mathcal{A}_{\Omega}^{(0)}(p,p)-s_1(z;\theta+p\cdot\omega)|\le\delta_1/2,\qquad p\in I_1. \tag*{(69)} ∣ A Ω ( 0 ) ( p , p ) − s 1 ( z ; θ + p ⋅ ω ) ∣ ≤ δ 1 /2 , p ∈ I 1 . (69)
Under (69) and 4 a / δ 1 ≤ ( 2 C 3 ) − 1 4a/\delta_1\le(2C_3)^{-1} 4 a / δ 1 ≤ ( 2 C 3 ) − 1 , we verify the assumptions of Lemma A.2 with X = P 0 X=P_0 X = P 0 , P = P 1 P=P_1 P = P 1 , Q = I 1 Q=I_1 Q = I 1 , M = 1 M=1 M = 1 , a ∗ = 2 a a_*=2a a ∗ = 2 a and Δ = δ 1 / 2 \Delta=\delta_1/2 Δ = δ 1 /2 . Thus ( A Ω ( 0 ) ) I 1 (\mathcal{A}_{\Omega}^{(0)})_{I_1} ( A Ω ( 0 ) ) I 1 is invertible, and (38) becomes
v ∣ I 1 = − ( A Ω ( 0 ) ) I 1 − 1 ( A Ω ( 0 ) ) I 1 , P 1 v ∣ P 1 , A Ω ( 1 ) v ∣ P 1 = 0 , A Ω ( 1 ) = Sch I 1 ( A Ω ( 0 ) ) . (70) v|_{I_1}=-(\mathcal{A}_{\Omega}^{(0)})_{I_1}^{-1}(\mathcal{A}_{\Omega}^{(0)})_{I_1,P_1}v|_{P_1},\qquad\mathcal{A}_{\Omega}^{(1)}v|_{P_1}=0,\qquad\mathcal{A}_{\Omega}^{(1)}=\operatorname{Sch}_{I_1}(\mathcal{A}_{\Omega}^{(0)}). \tag*{(70)} v ∣ I 1 = − ( A Ω ( 0 ) ) I 1 − 1 ( A Ω ( 0 ) ) I 1 , P 1 v ∣ P 1 , A Ω ( 1 ) v ∣ P 1 = 0 , A Ω ( 1 ) = Sch I 1 ( A Ω ( 0 ) ) . (70)
To compare A Ω ( 0 ) ( p , p ) \mathcal{A}_{\Omega}^{(0)}(p,p) A Ω ( 0 ) ( p , p ) with A B 1 ( p ) ( 0 ) ( p , p ) \mathcal{A}_{B_1(p)}^{(0)}(p,p) A B 1 ( p ) ( 0 ) ( p , p ) , we require B 1 ( p ) ⊂ Ω B_1(p)\subset\Omega B 1 ( p ) ⊂ Ω . Since B 1 ( p ) ∩ S 0 = { p } B_1(p)\cap S_0=\{p\} B 1 ( p ) ∩ S 0 = { p } , the eliminated coordinates then satisfy
Q 0 ∩ Ω = B 1 o ( p ) ⊔ ( Q 0 ∩ ( Ω ∖ B 1 ( p ) ) ) . Q_0\cap\Omega=B_1^o(p)\sqcup\bigl(Q_0\cap(\Omega\setminus B_1(p))\bigr). Q 0 ∩ Ω = B 1 o ( p ) ⊔ ( Q 0 ∩ ( Ω ∖ B 1 ( p )) ) .
Every point of Q 0 ∩ ( Ω ∖ B 1 ( p ) ) Q_0 \cap(\Omega\setminus B_1(p)) Q 0 ∩ ( Ω ∖ B 1 ( p )) has distance greater than 4 N 1 4N_1 4 N 1 from p p p . The distance bound and (35) will be used to estimate the effect of the additional eliminated coordinates. The comparison is also needed for the remaining coordinates: by (70) , for p ∈ P 1 p \in P_1 p ∈ P 1 ,
A Ω ( 1 ) ( p , p ) = A Ω ( 0 ) ( p , p ) − ( A Ω ( 0 ) ) p , I 1 ( A Ω ( 0 ) ) I 1 − 1 ( A Ω ( 0 ) ) I 1 , p . A_{\Omega}^{(1)}(p,p)=A_{\Omega}^{(0)}(p,p)-(A_{\Omega}^{(0)})_{p,I_1}(A_{\Omega}^{(0)})_{I_1}^{-1}(A_{\Omega}^{(0)})_{I_1,p}. A Ω ( 1 ) ( p , p ) = A Ω ( 0 ) ( p , p ) − ( A Ω ( 0 ) ) p , I 1 ( A Ω ( 0 ) ) I 1 − 1 ( A Ω ( 0 ) ) I 1 , p .
At z = E 1 ( θ + p ⋅ ω ) z=E_1(\theta+p\cdot\omega) z = E 1 ( θ + p ⋅ ω ) , the identity s 1 ( z ; θ + p ⋅ ω ) = 0 s_1(z;\theta+p\cdot\omega)=0 s 1 ( z ; θ + p ⋅ ω ) = 0 allows the region comparison to bound ∣ A Ω ( 0 ) ( p , p ) ∣ \lvert A_{\Omega}^{(0)}(p,p)\rvert ∣ A Ω ( 0 ) ( p , p )∣ in the remaining equation. We therefore impose
B 1 ( p ) ⊂ Ω , p ∈ S 0 ( z ; θ ) ∩ Ω . (71) B_1(p)\subset\Omega,\qquad p\in S_0(z;\theta)\cap\Omega. \tag*{(71)} B 1 ( p ) ⊂ Ω , p ∈ S 0 ( z ; θ ) ∩ Ω. (71)
On a larger finite region, we seek a local eigenvalue near E 1 ( θ ; t ) E_1(\theta;t) E 1 ( θ ; t ) by solving for all coordinates except v ( o ) v(o) v ( o ) . For the next application of Proposition 3.4 with center E 1 ( θ ; t ) E_1(\theta;t) E 1 ( θ ; t ) , the finite region must be fixed as t t t varies. We arrange (71) uniformly for θ ∈ Θ \theta\in\Theta θ ∈ Θ , t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] and ∣ z − E 1 ( θ ; t ) ∣ ≤ 4 δ 1 \lvert z-E_1(\theta;t)\rvert\leq4\delta_1 ∣ z − E 1 ( θ ; t )∣ ≤ 4 δ 1 . For p ∈ S 0 ( z ; θ ) p\in S_0(z;\theta) p ∈ S 0 ( z ; θ ) , (51) gives
L ∥ p ⋅ ω ∥ T ≤ ∣ f ( θ + p ⋅ ω ) − f ( θ ) ∣ < β + 4 δ 1 + β / 16 < 8 β . L\lVert p\cdot\omega\rVert_{\mathbb{T}}\leq\lvert f(\theta+p\cdot\omega)-f(\theta)\rvert<\beta+4\delta_1+\beta/16<8\beta. L ∥ p ⋅ ω ∥ T ≤ ∣ f ( θ + p ⋅ ω ) − f ( θ )∣ < β + 4 δ 1 + β /16 < 8 β .
Define
Z 0 = { p ∈ Z d : L ∥ p ⋅ ω ∥ T ≤ 8 β } . (72) Z_0=\{p\in\mathbb{Z}^d:L\lVert p\cdot\omega\rVert_{\mathbb{T}}\leq8\beta\}. \tag*{(72)} Z 0 = { p ∈ Z d : L ∥ p ⋅ ω ∥ T ≤ 8 β } . (72)
Then
S 0 ( z ; θ ) ⊂ Z 0 , ∣ z − E 1 ( θ ; t ) ∣ ≤ 4 δ 1 , (73) S_0(z;\theta)\subset Z_0,\qquad\lvert z-E_1(\theta;t)\rvert\leq4\delta_1, \tag*{(73)} S 0 ( z ; θ ) ⊂ Z 0 , ∣ z − E 1 ( θ ; t )∣ ≤ 4 δ 1 , (73)
for every θ ∈ Θ \theta\in\Theta θ ∈ Θ and t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . The set Z 0 Z_0 Z 0 is independent of z z z , θ \theta θ and t t t . By (73) , a fixed finite set F ⊃ B 1 F\supset B_1 F ⊃ B 1 satisfies (71) for every θ ∈ Θ \theta\in\Theta θ ∈ Θ , t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] and ∣ z − E 1 ( θ ; t ) ∣ ≤ 4 δ 1 \lvert z-E_1(\theta;t)\rvert\leq4\delta_1 ∣ z − E 1 ( θ ; t )∣ ≤ 4 δ 1 , provided that
p ∈ F ∩ Z 0 ⟹ B 1 ( p ) ⊂ F . p\in F\cap Z_0\Longrightarrow B_1(p)\subset F. p ∈ F ∩ Z 0 ⟹ B 1 ( p ) ⊂ F .
We formulate the resulting induction hypotheses in § 4 .
Inductive hypotheses
At step n ≥ 1 n\geq1 n ≥ 1 , extending the construction in § 3 , the eigenvalues E m E_m E m and corresponding eigenvectors v m , θ v_{m,\theta} v m , θ are known for 0 ≤ m ≤ n 0\leq m\leq n 0 ≤ m ≤ n , and the coordinates in Q 0 , … , Q n − 1 Q_0,\ldots,Q_{n-1} Q 0 , … , Q n − 1 have been solved. The next step solves for the coordinates in Q n Q_n Q n , constructs E n + 1 E_{n+1} E n + 1 and the corresponding eigenvector v n + 1 , θ v_{n+1,\theta} v n + 1 , θ , and estimates E n + 1 − E n E_{n+1}-E_n E n + 1 − E n and v n + 1 , θ − v n , θ v_{n+1,\theta}-v_{n,\theta} v n + 1 , θ − v n , θ .
Scales and the equations at step n n n
Scales and fixed sets. Retain β \beta β from (32) and a a a from (34) . Extend (42) and (64) by
N h = 16 h , δ h = β N h − τ , h ≥ 0 , (74) N_h=16^h,\qquad\delta_h=\beta N_h^{-\tau},\qquad h\geq0, \tag*{(74)} N h = 1 6 h , δ h = β N h − τ , h ≥ 0 , (74)
M 0 = 1 , M h = N h − 1 ( h ≥ 1 ) , η h = a δ h − 1 M h − r ( h ≥ 0 ) . (75) M_0=1,\qquad M_h=N_{h-1}\ (h\geq1),\qquad\eta_h=a\delta_h^{-1}M_h^{-r}\ (h\geq0). \tag*{(75)} M 0 = 1 , M h = N h − 1 ( h ≥ 1 ) , η h = a δ h − 1 M h − r ( h ≥ 0 ) . (75)
The numbers N h N_h N h specify the sizes of the finite regions, δ h \delta_h δ h specify the energy thresholds and M h M_h M h specify the separation used in the matrix estimates. Since r > τ r>\tau r > τ , we have
∑ h ≥ 0 η h = a β − 1 ( 1 + 16 τ ∑ h ≥ 1 N h − 1 τ − r ) < ∞ . (76) \sum_{h\geq0}\eta_h=a\beta^{-1}\left(1+16^\tau\sum_{h\geq1}N_{h-1}^{\tau-r}\right)<\infty. \tag*{(76)} h ≥ 0 ∑ η h = a β − 1 ( 1 + 1 6 τ h ≥ 1 ∑ N h − 1 τ − r ) < ∞. (76)
Extend the set Z 0 Z_0 Z 0 in (72) by defining
Z h = { p ∈ Z d : L ∥ p ⋅ ω ∥ T ≤ 8 δ h } , h ≥ 0. (77) Z_h=\{p\in\mathbb{Z}^d:L\lVert p\cdot\omega\rVert_{\mathbb{T}}\leq8\delta_h\},\qquad h\geq0. \tag*{(77)} Z h = { p ∈ Z d : L ∥ p ⋅ ω ∥ T ≤ 8 δ h } , h ≥ 0. (77)
For every h ≥ 0 h \ge0 h ≥ 0 , the set Z h Z_h Z h is independent of z z z , θ \theta θ and t t t , and satisfies Z h + 1 ⊂ Z h Z_{h+1} \subset Z_h Z h + 1 ⊂ Z h and o ∈ Z h o \in Z_h o ∈ Z h . For distinct p , q ∈ Z h p,q \in Z_h p , q ∈ Z h , the Diophantine condition (2) and ⟨ p − q ⟩ ≤ 2 ∣ p − q ∣ \langle p-q\rangle\le2|p-q| ⟨ p − q ⟩ ≤ 2∣ p − q ∣ imply
∣ p − q ∣ ≥ 1 2 ( L γ 16 β ) 1 / τ N h > 33600 N h . (78) |p-q| \ge\frac{1}{2}\left(\frac{L\gamma}{16\beta}\right)^{1/\tau} N_h > 33600N_h. \tag*{(78)} ∣ p − q ∣ ≥ 2 1 ( 16 β L γ ) 1/ τ N h > 33600 N h . (78)
Here 1 2 ( L γ / ( 16 β ) ) 1 / τ > 33600 \frac{1}{2}(L\gamma/(16\beta))^{1/\tau} > 33600 2 1 ( L γ / ( 16 β ) ) 1/ τ > 33600 follows from the fixed choice (32) .
The remaining equations. We first write the notation for t = 1 t=1 t = 1 . Given E 0 = f , E 1 , … , E n E_0=f,E_1,\ldots,E_n E 0 = f , E 1 , … , E n and finite sets B 0 , … , B n B_0,\ldots,B_n B 0 , … , B n with B 0 = { o } B_0=\{o\} B 0 = { o } , fix θ ∈ Θ \theta\in\Theta θ ∈ Θ and z ∈ R z\in\mathbb{R} z ∈ R . For 0 ≤ h ≤ n 0\le h\le n 0 ≤ h ≤ n , define
Res h = { p ∈ Z d : ∣ E h ( θ + p ⋅ ω ) − z ∣ < δ h } , (79) \operatorname{Res}_h=\{p\in\mathbb{Z}^d: |E_h(\theta+p\cdot\omega)-z|<\delta_h\}, \tag*{(79)} Res h = { p ∈ Z d : ∣ E h ( θ + p ⋅ ω ) − z ∣ < δ h } , (79)
S − 1 = Z d , S h = ⋂ ℓ = 0 h Res ℓ , (80) S_{-1}=\mathbb{Z}^d,\qquad S_h=\bigcap_{\ell=0}^{h}\operatorname{Res}_{\ell}, \tag*{(80)} S − 1 = Z d , S h = ℓ = 0 ⋂ h Res ℓ , (80)
Q h = S h − 1 ∖ S h . (81) Q_h=S_{h-1}\setminus S_h. \tag*{(81)} Q h = S h − 1 ∖ S h . (81)
For h = 0 , 1 h=0,1 h = 0 , 1 , (80) and (81) agree with (33) and (65) . Whenever E h E_h E h satisfies the hypotheses of Lemma 3.6 , distinct p , q ∈ S h p,q\in S_h p , q ∈ S h satisfy
L ∥ ( p − q ) ⋅ ω ∥ T ≤ ∣ E h ( θ + p ⋅ ω ) − E h ( θ + q ⋅ ω ) ∣ < 2 δ h , q − p ∈ Z h . L\|(p-q)\cdot\omega\|_{\mathbb{T}}\le|E_h(\theta+p\cdot\omega)-E_h(\theta+q\cdot\omega)|<2\delta_h,\qquad q-p\in Z_h. L ∥ ( p − q ) ⋅ ω ∥ T ≤ ∣ E h ( θ + p ⋅ ω ) − E h ( θ + q ⋅ ω ) ∣ < 2 δ h , q − p ∈ Z h .
For a finite set Ω \Omega Ω , let
I h = Q h ∩ Ω , P h = S h ∩ Ω , P − 1 = Ω . (82) I_h=Q_h\cap\Omega,\qquad P_h=S_h\cap\Omega,\qquad P_{-1}=\Omega. \tag*{(82)} I h = Q h ∩ Ω , P h = S h ∩ Ω , P − 1 = Ω. (82)
For 0 ≤ h ≤ n 0\le h\le n 0 ≤ h ≤ n , define A Ω ( h ) A_\Omega^{(h)} A Ω ( h ) recursively by
A Ω ( − 1 ) = A Ω ( z ; θ ) , A Ω ( h ) = Sch I h ( A Ω ( h − 1 ) ) , (83) A_\Omega^{(-1)}=A_\Omega(z;\theta),\qquad A_\Omega^{(h)}=\operatorname{Sch}_{I_h}(A_\Omega^{(h-1)}), \tag*{(83)} A Ω ( − 1 ) = A Ω ( z ; θ ) , A Ω ( h ) = Sch I h ( A Ω ( h − 1 ) ) , (83)
provided that ( A Ω ( h − 1 ) ) I h (A_\Omega^{(h-1)})_{I_h} ( A Ω ( h − 1 ) ) I h is invertible. The matrix A Ω ( h ) A_\Omega^{(h)} A Ω ( h ) has row and column set P h P_h P h . The Schur identity (22) makes A Ω ( h − 1 ) v ∣ P h − 1 = 0 A_\Omega^{(h-1)}v|_{P_{h-1}}=0 A Ω ( h − 1 ) v ∣ P h − 1 = 0 equivalent to
v ∣ I h = − ( A Ω ( h − 1 ) ) I h − 1 ( A Ω ( h − 1 ) ) I h , P h v ∣ P h , (84) v|_{I_h}=-(A_\Omega^{(h-1)})_{I_h}^{-1}(A_\Omega^{(h-1)})_{I_h,P_h}v|_{P_h}, \tag*{(84)} v ∣ I h = − ( A Ω ( h − 1 ) ) I h − 1 ( A Ω ( h − 1 ) ) I h , P h v ∣ P h , (84)
A Ω ( h ) v ∣ P h = 0. (85) A_\Omega^{(h)}v|_{P_h}=0. \tag*{(85)} A Ω ( h ) v ∣ P h = 0. (85)
After eliminating I 0 , … , I n − 1 I_0,\ldots,I_{n-1} I 0 , … , I n − 1 , the remaining equation is A Ω ( n − 1 ) v ∣ P n − 1 = 0 A_\Omega^{(n-1)}v|_{P_{n-1}}=0 A Ω ( n − 1 ) v ∣ P n − 1 = 0 . Since P n − 1 = I n ⊔ P n P_{n-1}=I_n\sqcup P_n P n − 1 = I n ⊔ P n , the next elimination requires ( A Ω ( n − 1 ) ) I n (A_\Omega^{(n-1)})_{I_n} ( A Ω ( n − 1 ) ) I n to be invertible.
The finite regions. For 1 ≤ i ≤ n 1\le i\le n 1 ≤ i ≤ n , comparing A Ω , B i ( p ) ( i − 1 ) ( p , p ) A_{\Omega,B_i(p)}^{(i-1)}(p,p) A Ω , B i ( p ) ( i − 1 ) ( p , p ) with A Ω ( i − 1 ) ( p , p ) A_\Omega^{(i-1)}(p,p) A Ω ( i − 1 ) ( p , p ) requires B i ( p ) ⊂ Ω B_i(p)\subset\Omega B i ( p ) ⊂ Ω for p ∈ S i − 1 ∩ Ω p\in S_{i-1}\cap\Omega p ∈ S i − 1 ∩ Ω , as in (71) for i = 1 i=1 i = 1 . To construct E n + 1 E_{n+1} E n + 1 , the preceding eliminations must also be available on the new sets B n + 1 B_{n+1} B n + 1 and B n + 1 ∘ B_{n+1}^{\circ} B n + 1 ∘ . We therefore formulate the estimates for every finite Ω \Omega Ω satisfying the required inclusions for the already constructed B 1 , … , B i B_1,\ldots,B_i B 1 , … , B i ; the region Ω \Omega Ω can then be chosen after the estimates have been proved. Let
Υ 0 = { Ω ⊂ Z d : # Ω < ∞ } . (86) \Upsilon_0=\{\Omega\subset\mathbb{Z}^d:\#\Omega<\infty\}. \tag*{(86)} Υ 0 = { Ω ⊂ Z d : #Ω < ∞ } . (86)
For 1 ≤ i ≤ n 1\le i\le n 1 ≤ i ≤ n , define recursively
Υ i = { Ω ∈ Υ i − 1 : B i ( p ) ⊂ Ω for every p ∈ S i − 1 ∩ Ω } . \Upsilon_i=\{\Omega\in\Upsilon_{i-1}:B_i(p)\subset\Omega\text{ for every }p\in S_{i-1}\cap\Omega\}. Υ i = { Ω ∈ Υ i − 1 : B i ( p ) ⊂ Ω for every p ∈ S i − 1 ∩ Ω } .
Equivalently, Ω ∈ Υ i \Omega\in\Upsilon_i Ω ∈ Υ i if and only if Ω \Omega Ω is finite and
B h ( p ) ⊂ Ω , 1 ≤ h ≤ i , p ∈ S h − 1 ∩ Ω . (87) B_h(p)\subset\Omega,\qquad1\le h\le i,\quad p\in S_{h-1}\cap\Omega. \tag*{(87)} B h ( p ) ⊂ Ω , 1 ≤ h ≤ i , p ∈ S h − 1 ∩ Ω. (87)
For i = 1 i=1 i = 1 , condition (87) reduces to (71) . The condition Ω ∈ Υ i \Omega\in\Upsilon_i Ω ∈ Υ i does not require S i ∩ Ω = ∅ S_i\cap\Omega=\varnothing S i ∩ Ω = ∅ ; the coordinates in S i ∩ Ω S_i\cap\Omega S i ∩ Ω remain in (85) . The sets Res h \operatorname{Res}_h Res h , S h S_h S h , Q h Q_h Q h are defined on Z d \mathbb{Z}^d Z d before Ω \Omega Ω is chosen. Through the sets S h − 1 ( z ; θ ) S_{h-1}(z;\theta) S h − 1 ( z ; θ ) in (87) , the family Υ i = Υ i ( z ; θ ) \Upsilon_i=\Upsilon_i(z;\theta) Υ i = Υ i ( z ; θ ) depends on z z z and θ \theta θ .
The local Schur complement. For 1 ≤ m ≤ n 1 \le m \le n 1 ≤ m ≤ n , let c m = − ε Π B m ∘ T e o c_m=-\varepsilon\Pi_{B_m^\circ}^{T}e_o c m = − ε Π B m ∘ T e o . For z z z such that A B m ∘ ( z ; θ ) A_{B_m^\circ}(z;\theta) A B m ∘ ( z ; θ ) is invertible, define
s m ( z ; θ ) = z − f ( θ ) − c m ∗ A B m ∘ ( z ; θ ) − 1 c m . (88) s_m(z;\theta)=z-f(\theta)-c_m^*A_{B_m^\circ}(z;\theta)^{-1}c_m. \tag*{(88)} s m ( z ; θ ) = z − f ( θ ) − c m ∗ A B m ∘ ( z ; θ ) − 1 c m . (88)
For m = 1 m=1 m = 1 , definition (88) agrees with (47) . As in (48) ,
det A B m ( z ; θ ) = det A B m ∘ ( z ; θ ) s m ( z ; θ ) . (89) \det A_{B_m}(z;\theta)=\det A_{B_m^\circ}(z;\theta)s_m(z;\theta). \tag*{(89)} det A B m ( z ; θ ) = det A B m ∘ ( z ; θ ) s m ( z ; θ ) . (89)
For 1 ≤ m ≤ n 1 \le m \le n 1 ≤ m ≤ n , let
J m ( θ ) = [ E m − 1 ( θ ) − 4 δ m − 1 , E m − 1 ( θ ) + 4 δ m − 1 ] . (90) J_m(\theta)=[E_{m-1}(\theta)-4\delta_{m-1},E_{m-1}(\theta)+4\delta_{m-1}]. \tag*{(90)} J m ( θ ) = [ E m − 1 ( θ ) − 4 δ m − 1 , E m − 1 ( θ ) + 4 δ m − 1 ] . (90)
The homotopy parameter. For t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] , use A F ( z ; θ ; t ) A_F(z;\theta;t) A F ( z ; θ ; t ) from (46) , E h ( θ ; t ) E_h(\theta;t) E h ( θ ; t ) and c m ( t ) = − t ε Π B m ∘ T e o c_m(t)=-t\varepsilon\Pi_{B_m^\circ}^{T}e_o c m ( t ) = − tε Π B m ∘ T e o in (79) –(90) . We write the resulting functions, matrices and sets as s m ( z ; θ ; t ) s_m(z;\theta;t) s m ( z ; θ ; t ) , J m ( θ ; t ) J_m(\theta;t) J m ( θ ; t ) , A Ω ( h ) ( z ; θ ; t ) A_\Omega^{(h)}(z;\theta;t) A Ω ( h ) ( z ; θ ; t ) , Res h ( z ; θ ; t ) \operatorname{Res}_h(z;\theta;t) Res h ( z ; θ ; t ) , S h ( z ; θ ; t ) S_h(z;\theta;t) S h ( z ; θ ; t ) , Q h ( z ; θ ; t ) Q_h(z;\theta;t) Q h ( z ; θ ; t ) and Υ i ( z ; θ ; t ) \Upsilon_i(z;\theta;t) Υ i ( z ; θ ; t ) . The arguments z , θ , t z,\theta,t z , θ , t are omitted when fixed. The sets B h B_h B h , Z h Z_h Z h and the numbers N h N_h N h , δ h \delta_h δ h , M h M_h M h are independent of z , θ , t z,\theta,t z , θ , t . By (34) , ∣ t ε ϕ ( p − q ) ∣ ≤ a ⟨ p − q ⟩ − r |t\varepsilon\phi(p-q)| \le a\langle p-q\rangle^{-r} ∣ tεϕ ( p − q ) ∣ ≤ a ⟨ p − q ⟩ − r for p , q ∈ Z d p,q \in\mathbb{Z}^d p , q ∈ Z d and t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] .
A common choice of constants
The estimates below require three smallness conditions, one for elimination, one for eigenvalue drift and one for eigenvector reconstruction. We fix all coefficients before stating the induction assertions.
Let C 5 = C 5 ( d , r , τ ) C_5=C_5(d,r,\tau) C 5 = C 5 ( d , r , τ ) be a finite constant such that
C 5 ≥ 32 ⋅ 16 2 r C 3 (91) C_5 \ge32\cdot16^{2r}C_3 \tag*{(91)} C 5 ≥ 32 ⋅ 1 6 2 r C 3 (91)
and the diagonal comparison (5.7) holds with coefficient C 5 / 2 C_5/2 C 5 /2 . Under the stated matrix hypotheses, Proposition 5.2 and Corollary 5.3 prove (5.7) with a finite coefficient depending only on d , r , τ d,r,\tau d , r , τ , so C 5 C_5 C 5 can be fixed before the induction. The factor in (91) ensures that the last Schur elimination in (144) also has coefficient at most C 5 / 2 C_5/2 C 5 /2 .
For the fixed 0 < ρ < r − τ 0<\rho<r-\tau 0 < ρ < r − τ , put
G ( s ) = 1 + 16 τ 1 − 16 τ + s − r , 0 ≤ s < r − τ . (92) G(s)=1+\frac{16^\tau}{1-16^{\tau+s-r}},\qquad0\le s<r-\tau. \tag*{(92)} G ( s ) = 1 + 1 − 1 6 τ + s − r 1 6 τ , 0 ≤ s < r − τ . (92)
The scales in (74) –(75) give
∑ h ≥ 0 δ h − 1 M h s − r = β − 1 G ( s ) , 0 ≤ s < r − τ . \sum_{h\ge0}\delta_h^{-1}M_h^{s-r}=\beta^{-1}G(s),\qquad0\le s<r-\tau. h ≥ 0 ∑ δ h − 1 M h s − r = β − 1 G ( s ) , 0 ≤ s < r − τ .
Fix a finite constant C 6 = C 6 ( d , r , τ , γ , L , ρ ) C_6=C_6(d,r,\tau,\gamma,L,\rho) C 6 = C 6 ( d , r , τ , γ , L , ρ ) satisfying
C 6 ≥ 2 C 4 β − 1 G ( ρ ) , C 6 ≥ 2 τ + 4 C 3 L γ . (93) C_6\ge2C_4\beta^{-1}G(\rho),\qquad C_6\ge\frac{2^{\tau+4}C_3}{L\gamma}. \tag*{(93)} C 6 ≥ 2 C 4 β − 1 G ( ρ ) , C 6 ≥ L γ 2 τ + 4 C 3 . (93)
The first inequality controls the product of the reconstruction matrices; the second controls the use of eigenvalue separation in (159) .
We require
C 5 a ∑ h ≥ 0 δ h − 1 M h − r ≤ 1 64 , (94) C_5a\sum_{h\ge0}\delta_h^{-1}M_h^{-r}\le\frac{1}{64}, \tag*{(94)} C 5 a h ≥ 0 ∑ δ h − 1 M h − r ≤ 64 1 , (94)
C 5 a 2 β − 2 1 − 16 τ − 2 r ≤ 1 16 , (95) \frac{C_5a^2\beta^{-2}}{1-16^{\tau-2r}}\le\frac{1}{16}, \tag*{(95)} 1 − 1 6 τ − 2 r C 5 a 2 β − 2 ≤ 16 1 , (95)
C 6 a ≤ 1 4 . (96) C_6a\le\frac{1}{4}. \tag*{(96)} C 6 a ≤ 4 1 . (96)
Choose a threshold
ε 0 = ε 0 ( d , r , τ , γ , L , ρ , C 0 ) > 0 (97) \varepsilon_0=\varepsilon_0(d,r,\tau,\gamma,L,\rho,C_0)>0 \tag*{(97)} ε 0 = ε 0 ( d , r , τ , γ , L , ρ , C 0 ) > 0 (97)
so that (94) –(96) hold with a = C 0 ε 0 a=C_0\varepsilon_0 a = C 0 ε 0 . The left-hand sides tend to zero as a → 0 a\to0 a → 0 , so such a threshold exists. Since a ≤ C 0 ∣ ε ∣ a\le C_0|\varepsilon| a ≤ C 0 ∣ ε ∣ , the same three inequalities hold whenever ∣ ε ∣ < ε 0 |\varepsilon|<\varepsilon_0 ∣ ε ∣ < ε 0 . The dependence in (97) agrees with Theorem 1.1 , since C 0 = C 0 ( r , ϕ ) C_0=C_0(r,\phi) C 0 = C 0 ( r , ϕ ) . Moreover, (94) implies
∑ h ≥ 0 η h ≤ ( 64 C 5 ) − 1 ≤ ( 64 C 3 ) − 1 , a / β ≤ ( 2 C 3 ) − 1 . \sum_{h\ge0}\eta_h\le(64C_5)^{-1}\le(64C_3)^{-1},\qquad a/\beta\le(2C_3)^{-1}. h ≥ 0 ∑ η h ≤ ( 64 C 5 ) − 1 ≤ ( 64 C 3 ) − 1 , a / β ≤ ( 2 C 3 ) − 1 .
Thus the comparison estimates and the initial construction apply. The constants C 3 , … , C 6 C_3,\ldots,C_6 C 3 , … , C 6 and ε 0 \varepsilon_0 ε 0 remain fixed throughout the induction and the localization proof, independently of n n n , θ \theta θ and t t t . Finally, set
κ = ρ 2 τ + ρ > 0. (98) \kappa=\frac{\rho^2}{\tau+\rho}>0. \tag*{(98)} κ = τ + ρ ρ 2 > 0. (98)
The induction assertions
We state assertions (1)–(5) at t = 1 t=1 t = 1 and require validity for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] in assertion (6).
Theorem 4.1 (Multiscale estimates). Assume (5) , the potential assumptions in § 1.1 and ω ∈ D C \omega\in\mathrm{DC} ω ∈ DC , with r > τ > d r>\tau>d r > τ > d . Under (94) , (95) and (96) , there are finite sets ( B m ) m ≥ 0 (B_m)_{m\ge0} ( B m ) m ≥ 0 , with B 0 = { o } B_0=\{o\} B 0 = { o } , for which the functions
E 0 ( θ ; t ) = f ( θ ) , E m ( θ ; t ) = λ ( B m ; θ ; t ) , m ≥ 1 , θ ∈ Θ , t ∈ [ 0 , 1 ] , (99) E_0(\theta;t)=f(\theta),\qquad E_m(\theta;t)=\lambda(B_m;\theta;t),\qquad m\ge1,\quad\theta\in\Theta,\quad t\in[0,1], \tag*{(99)} E 0 ( θ ; t ) = f ( θ ) , E m ( θ ; t ) = λ ( B m ; θ ; t ) , m ≥ 1 , θ ∈ Θ , t ∈ [ 0 , 1 ] , (99) satisfy the following assertions at every step n ≥ 1 n\ge1 n ≥ 1 . Here λ ( B m ; θ ; t ) \lambda(B_m;\theta;t) λ ( B m ; θ ; t ) is the branch defined in (58) for H B m ( θ ; t ) \mathcal{H}_{B_m}(\theta;t) H B m ( θ ; t ) , and we write E m = E m ( ⋅ ; 1 ) E_m=E_m(\cdot;1) E m = E m ( ⋅ ; 1 ) . In particular, each E m ( ⋅ ; t ) E_m(\cdot;t) E m ( ⋅ ; t ) is real and 1 1 1 -periodic.
(1) Geometry. For 1 ≤ m ≤ n 1\le m\le n 1 ≤ m ≤ n ,
B m − 1 ⊂ B m , Λ 4 N m ⊂ B m ⊂ Λ 8 N m , (100) B_{m-1}\subset B_m,\qquad\Lambda_{4N_m}\subset B_m\subset\Lambda_{8N_m}, \tag*{(100)} B m − 1 ⊂ B m , Λ 4 N m ⊂ B m ⊂ Λ 8 N m , (100) p ∈ B m ∩ Z h , 0 ≤ h ≤ m − 2 ⟹ B h + 1 ( p ) + Λ N h + 1 ⊂ B m , (101) p\in B_m\cap Z_h,\quad0\le h\le m-2\quad\Longrightarrow\quad B_{h+1}(p)+\Lambda_{N_{h+1}}\subset B_m, \tag*{(101)} p ∈ B m ∩ Z h , 0 ≤ h ≤ m − 2 ⟹ B h + 1 ( p ) + Λ N h + 1 ⊂ B m , (101) B m ∩ Z m − 1 = { o } . (102) B_m\cap Z_{m-1}=\{o\}. \tag*{(102)} B m ∩ Z m − 1 = { o } . (102) The sets Λ R \Lambda_R Λ R and the addition of sets are defined in (16) and (14) . For 0 ≤ h ≤ n − 1 0\le h\le n-1 0 ≤ h ≤ n − 1 , we also have
p ∈ Z h ∘ ⟹ o ∉ B h + 1 ( p ) + Λ N h + 1 . (103) p\in Z_h^\circ\quad\Longrightarrow\quad o\notin B_{h+1}(p)+\Lambda_{N_{h+1}}. \tag*{(103)} p ∈ Z h ∘ ⟹ o ∈ / B h + 1 ( p ) + Λ N h + 1 . (103) (2) Local spectrum. For 1 ≤ m ≤ n 1\le m\le n 1 ≤ m ≤ n , θ ∈ Θ \theta\in\Theta θ ∈ Θ and z ∈ J m ( θ ) z\in J_m(\theta) z ∈ J m ( θ ) , A B m ∘ ( z ; θ ) A_{B_m}^{\circ}(z;\theta) A B m ∘ ( z ; θ ) is invertible. The function s m ( ⋅ ; θ ) s_m(\cdot;\theta) s m ( ⋅ ; θ ) has exactly one zero in J m ( θ ) J_m(\theta) J m ( θ ) , namely E m ( θ ) E_m(\theta) E m ( θ ) , and satisfies
∣ s m ( z ; θ ) ∣ ≥ ∣ z − E m ( θ ) ∣ , z ∈ J m ( θ ) . (104) |s_m(z;\theta)|\ge|z-E_m(\theta)|,\qquad z\in J_m(\theta). \tag*{(104)} ∣ s m ( z ; θ ) ∣ ≥ ∣ z − E m ( θ ) ∣ , z ∈ J m ( θ ) . (104) The eigenvalue E m ( θ ) E_m(\theta) E m ( θ ) of H B m ( θ ) \mathcal{H}_{B_m}(\theta) H B m ( θ ) is simple and
dist ( E m ( θ ) , Spec H B m ( θ ) ∖ { E m ( θ ) } ) ≥ 3 δ m − 1 . (105) \operatorname{dist}\left(E_m(\theta),\operatorname{Spec}\mathcal{H}_{B_m}(\theta)\setminus\{E_m(\theta)\}\right)\ge3\delta_{m-1}. \tag*{(105)} dist ( E m ( θ ) , Spec H B m ( θ ) ∖ { E m ( θ )} ) ≥ 3 δ m − 1 . (105) (3) Eigenvalue estimates. With the coefficient C 5 C_5 C 5 fixed in Subsection 4.2,
∥ E m − E m − 1 ∥ ∞ ≤ C 5 a 2 β − 1 N m − 1 τ − 2 r , 1 ≤ m ≤ n , (106) \|E_m-E_{m-1}\|_\infty\le C_5a^2\beta^{-1}N_{m-1}^{\tau-2r},\qquad1\le m\le n, \tag*{(106)} ∥ E m − E m − 1 ∥ ∞ ≤ C 5 a 2 β − 1 N m − 1 τ − 2 r , 1 ≤ m ≤ n , (106) ∥ E m − E ℓ ∥ ∞ ≤ δ ℓ / 8 , 0 ≤ ℓ < m ≤ n . (107) \|E_m-E_\ell\|_\infty\le\delta_\ell/8,\qquad0\le\ell<m\le n. \tag*{(107)} ∥ E m − E ℓ ∥ ∞ ≤ δ ℓ /8 , 0 ≤ ℓ < m ≤ n . (107) Here ∥ ⋅ ∥ ∞ \|\cdot\|_\infty ∥ ⋅ ∥ ∞ is the supremum norm specified in Theorem 1.1 . For 1 ≤ m ≤ n 1\le m\le n 1 ≤ m ≤ n , we also have
E m ( θ 2 ) − E m ( θ 1 ) ≥ L ( θ 2 − θ 1 ) , θ 1 < θ 2 , θ 1 , θ 2 ∈ Θ ∩ [ 0 , 1 ) . (108) E_m(\theta_2)-E_m(\theta_1)\ge L(\theta_2-\theta_1),\qquad\theta_1<\theta_2,\quad\theta_1,\theta_2\in\Theta\cap[0,1). \tag*{(108)} E m ( θ 2 ) − E m ( θ 1 ) ≥ L ( θ 2 − θ 1 ) , θ 1 < θ 2 , θ 1 , θ 2 ∈ Θ ∩ [ 0 , 1 ) . (108)
(4) Elimination estimates. For 0 ≤ i ≤ n − 1 0 \le i \le n-1 0 ≤ i ≤ n − 1 , θ ∈ Θ \theta\in\Theta θ ∈ Θ , z ∈ R z \in\mathbb{R} z ∈ R and Ω ∈ Υ i \Omega\in\Upsilon_i Ω ∈ Υ i , the matrices ( A Ω ( h − 1 ) ) Q h ∩ Ω (\mathcal{A}_{\Omega}^{(h-1)})_{Q_h\cap\Omega} ( A Ω ( h − 1 ) ) Q h ∩ Ω are invertible for 0 ≤ h ≤ i 0 \le h \le i 0 ≤ h ≤ i . Moreover,
∣ A Ω ( h − 1 ) ( p , p ) ∣ ≥ δ h / 2 , p ∈ Q h ∩ Ω , 0 ≤ h ≤ i , (109) \left|\mathcal{A}_{\Omega}^{(h-1)}(p,p)\right| \ge\delta_h/2,\qquad p \in Q_h\cap\Omega,\qquad0 \le h \le i, \tag*{(109)} A Ω ( h − 1 ) ( p , p ) ≥ δ h /2 , p ∈ Q h ∩ Ω , 0 ≤ h ≤ i , (109)
∣ A Ω ( h − 1 ) ( p , q ) ∣ ≤ 2 a ⟨ p − q ⟩ − r , p , q ∈ S h − 1 ∩ Ω , p ≠ q , 0 ≤ h ≤ i + 1. (110) \left|\mathcal{A}_{\Omega}^{(h-1)}(p,q)\right| \le2a\langle p-q\rangle^{-r},\qquad p,q \in S_{h-1}\cap\Omega,\qquad p\ne q,\qquad0 \le h \le i+1. \tag*{(110)} A Ω ( h − 1 ) ( p , q ) ≤ 2 a ⟨ p − q ⟩ − r , p , q ∈ S h − 1 ∩ Ω , p = q , 0 ≤ h ≤ i + 1. (110)
Estimate (110) includes A Ω ( i ) \mathcal{A}_{\Omega}^{(i)} A Ω ( i ) , the matrix after eliminating I i I_i I i .
(5) Eigenvectors and increments. For 0 ≤ m ≤ n 0 \le m \le n 0 ≤ m ≤ n and θ ∈ Θ \theta\in\Theta θ ∈ Θ , there is a vector v m , θ v_{m,\theta} v m , θ supported on B m B_m B m , with v 0 , θ = e o v_{0,\theta}=e_o v 0 , θ = e o , such that
H B m ( θ ) v m , θ ∣ B m = E m ( θ ) v m , θ ∣ B m , v m , θ ( o ) = 1. (111) \mathcal{H}_{B_m}(\theta)v_{m,\theta}|_{B_m}=E_m(\theta)v_{m,\theta}|_{B_m},\qquad v_{m,\theta}(o)=1. \tag*{(111)} H B m ( θ ) v m , θ ∣ B m = E m ( θ ) v m , θ ∣ B m , v m , θ ( o ) = 1. (111)
Extend v m , θ v_{m,\theta} v m , θ by zero to Z d \mathbb{Z}^d Z d . Then
∥ v m , θ − e o ∥ 1 , ρ ≤ C 6 a ≤ 1 / 4 , 0 ≤ m ≤ n , (112) \|v_{m,\theta}-e_o\|_{1,\rho}\le C_6a\le1/4,\qquad0\le m\le n, \tag*{(112)} ∥ v m , θ − e o ∥ 1 , ρ ≤ C 6 a ≤ 1/4 , 0 ≤ m ≤ n , (112)
sup θ ∈ Θ ∥ v m , θ − v m − 1 , θ ∥ 1 ≤ 12 C 6 a N m − 1 − κ , 1 ≤ m ≤ n . (113) \sup_{\theta\in\Theta}\|v_{m,\theta}-v_{m-1,\theta}\|_1\le12C_6aN_{m-1}^{-\kappa},\qquad1\le m\le n. \tag*{(113)} θ ∈ Θ sup ∥ v m , θ − v m − 1 , θ ∥ 1 ≤ 12 C 6 a N m − 1 − κ , 1 ≤ m ≤ n . (113)
The norm ∥ ⋅ ∥ 1 , ρ \|\cdot\|_{1,\rho} ∥ ⋅ ∥ 1 , ρ is defined in (54) ; C 6 C_6 C 6 and κ \kappa κ are fixed in (93) and (98) .
(6) Uniformity for t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] . For every t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] , assertions (1)–(5) hold for
H ( ⋅ ; t ) , E m ( ⋅ ; t ) , s m ( ⋅ ; ⋅ ; t ) , J m ( ⋅ ; t ) , v m , θ ( t ) , \mathcal{H}(\cdot;t),\qquad E_m(\cdot;t),\qquad s_m(\cdot;\cdot;t),\qquad J_m(\cdot;t),\qquad v_{m,\theta}(t), H ( ⋅ ; t ) , E m ( ⋅ ; t ) , s m ( ⋅ ; ⋅ ; t ) , J m ( ⋅ ; t ) , v m , θ ( t ) ,
with B m B_m B m , N h N_h N h , δ h \delta_h δ h , M h M_h M h and all constants independent of t t t . In assertion (4), use S h ( z ; θ ; t ) S_h(z;\theta;t) S h ( z ; θ ; t ) , Q h ( z ; θ ; t ) Q_h(z;\theta;t) Q h ( z ; θ ; t ) , Υ i ( z ; θ ; t ) \Upsilon_i(z;\theta;t) Υ i ( z ; θ ; t ) and A Ω ( i ) ( z ; θ ; t ) \mathcal{A}_{\Omega}^{(i)}(z;\theta;t) A Ω ( i ) ( z ; θ ; t ) as defined after (90) . For 0 ≤ m ≤ n 0 \le m \le n 0 ≤ m ≤ n and θ ∈ Θ \theta\in\Theta θ ∈ Θ , equations (99) and (59) give
E m ( θ ; 0 ) = f ( θ ) , E m ( θ ; 1 ) = E m ( θ ) . (114) E_m(\theta;0)=f(\theta),\qquad E_m(\theta;1)=E_m(\theta). \tag*{(114)} E m ( θ ; 0 ) = f ( θ ) , E m ( θ ; 1 ) = E m ( θ ) . (114)
By (30) , t ↦ E m ( θ ; t ) t\mapsto E_m(\theta;t) t ↦ E m ( θ ; t ) is continuous. Moreover, v m , θ ( 0 ) = e o v_{m,\theta}(0)=e_o v m , θ ( 0 ) = e o and t ↦ v m , θ ( t ) t\mapsto v_{m,\theta}(t) t ↦ v m , θ ( t ) is continuous in ℓ 2 \ell^2 ℓ 2 .
We record how assertion (1) gives the regions required in assertion (4). Fix 1 ≤ m ≤ n 1\le m\le n 1 ≤ m ≤ n , θ ∈ Θ \theta\in\Theta θ ∈ Θ , z ∈ R z\in\mathbb{R} z ∈ R and t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . If S h ( z ; θ ; t ) ⊂ Z h S_h(z;\theta;t)\subset Z_h S h ( z ; θ ; t ) ⊂ Z h for 0 ≤ h < m 0\le h<m 0 ≤ h < m , then (101) gives B m ∈ Υ m − 1 ( z ; θ ; t ) B_m\in\Upsilon_{m-1}(z;\theta;t) B m ∈ Υ m − 1 ( z ; θ ; t ) . For p ∈ B m ∘ ∩ S h p\in B_m^\circ\cap S_h p ∈ B m ∘ ∩ S h with 0 ≤ h ≤ m − 2 0\le h\le m-2 0 ≤ h ≤ m − 2 , (103) also gives o ∉ B h + 1 ( p ) o\notin B_{h+1}(p) o ∈ / B h + 1 ( p ) ; hence B m ∘ ∈ Υ m − 1 B_m^\circ\in\Upsilon_{m-1} B m ∘ ∈ Υ m − 1 . Moreover, (101) gives dist ( x , B m c ) ≥ N h + 1 \operatorname{dist}(x,B_m^c)\ge N_{h+1} dist ( x , B m c ) ≥ N h + 1 for p ∈ B m ∩ S h p\in B_m\cap S_h p ∈ B m ∩ S h , 0 ≤ h ≤ m − 2 0\le h\le m-2 0 ≤ h ≤ m − 2 and x ∈ B h + 1 ( p ) x\in B_{h+1}(p) x ∈ B h + 1 ( p ) . Finally, by (102) , the remaining index set after eliminating Q 0 , … , Q m − 1 Q_0,\ldots,Q_{m-1} Q 0 , … , Q m − 1 is B m ∩ S m − 1 ⊂ { o } B_m\cap S_{m-1}\subset\{o\} B m ∩ S m − 1 ⊂ { o } .
We verify Theorem 4.1 at n = 1 n=1 n = 1 . By (42) , B 1 = Λ 4 N 1 B_1=\Lambda_{4N_1} B 1 = Λ 4 N 1 and B 0 = { o } B_0=\{o\} B 0 = { o } . Equation (78) gives ∣ p ∣ > 33600 |p|>33600 ∣ p ∣ > 33600 for p ∈ Z 0 c p\in Z_0^c p ∈ Z 0 c . Since 4 N 1 = 64 4N_1=64 4 N 1 = 64 and 5 N 1 = 80 5N_1=80 5 N 1 = 80 , we have
B 1 ∩ Z 0 = { o } , o ∉ B 1 ( p ) + Λ N 1 ( p ∈ Z 0 c ) . B_1\cap Z_0=\{o\},\qquad o\notin B_1(p)+\Lambda_{N_1}\quad(p\in Z_0^c). B 1 ∩ Z 0 = { o } , o ∈ / B 1 ( p ) + Λ N 1 ( p ∈ Z 0 c ) .
The range 0 ≤ h ≤ m − 2 0\le h\le m-2 0 ≤ h ≤ m − 2 in (101) is empty for m = 1 m=1 m = 1 . This proves assertion (1). Proposition (32) and (63) prove assertion (2). By (51) and C 5 ≥ 2 C 3 C_5\ge2C_3 C 5 ≥ 2 C 3 , the bounds (106) and (107) hold for m = 1 m=1 m = 1 . Equation (61) and Proposition (33) give (99) and (108) , completing assertion (3). Since Υ 0 \Upsilon_0 Υ 0 consists of all finite sets, Lemma (31) , (41) and (40) prove assertion (4) at i = 0 i=0 i = 0 .
The eigenvector in (53) satisfies (111) . By (56) and (93) ,
∥ v 1 , θ ( t ) − e o ∥ 1 , ρ ≤ C 4 a 4 β ≤ C 6 a ≤ 1 / 4. \|v_{1,\theta}(t)-e_o\|_{1,\rho}\le\frac{C_4a}{4\beta}\le C_6a\le1/4. ∥ v 1 , θ ( t ) − e o ∥ 1 , ρ ≤ 4 β C 4 a ≤ C 6 a ≤ 1/4.
Since N 0 = 1 N_0=1 N 0 = 1 , the same bound proves (113) for m = 1 m=1 m = 1 . This proves assertion (5). The first construction holds for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , and (50) and (53) give the continuity and endpoint identities in assertion (6).
The induction step
Assume that the assertions of Theorem 4.1 hold at step n ≥ 1 n \ge1 n ≥ 1 . We follow the equations in Subsection 3.6: compare A Ω ( n − 1 ) ( p , p ) \mathcal{A}_{\Omega}^{(n-1)}(p,p) A Ω ( n − 1 ) ( p , p ) with s n ( z ; θ + p ⋅ ω ) s_n(z;\theta+p\cdot\omega) s n ( z ; θ + p ⋅ ω ) , solve for the coordinates in Q n Q_n Q n , and construct the next local eigenvalue and corresponding eigenvector on a fixed region B n + 1 B_{n+1} B n + 1 . We also estimate E n + 1 − E n E_{n+1}-E_n E n + 1 − E n and v n + 1 , θ − v n , θ v_{n+1,\theta}-v_{n,\theta} v n + 1 , θ − v n , θ . The case n = 1 n=1 n = 1 is included in the argument. Fix t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] and omit t t t from the notation except when checking the homotopy.
Comparison on two regions
Fix θ ∈ Θ \theta\in\Theta θ ∈ Θ , z ∈ R z\in\mathbb{R} z ∈ R and a finite F ∈ Υ n F\in\Upsilon_n F ∈ Υ n , and consider
A F ( n − 1 ) v ∣ S n − 1 ∩ F = 0. \mathcal{A}_{F}^{(n-1)}v|_{S_{n-1}\cap F}=0. A F ( n − 1 ) v ∣ S n − 1 ∩ F = 0.
To solve for v ∣ Q n ∩ F v|_{Q_n\cap F} v ∣ Q n ∩ F , we compare A F ( n − 1 ) ( p , p ) \mathcal{A}_{F}^{(n-1)}(p,p) A F ( n − 1 ) ( p , p ) with s n ( z ; θ + p ⋅ ω ) s_n(z;\theta+p\cdot\omega) s n ( z ; θ + p ⋅ ω ) for p ∈ Q n ∩ F p\in Q_n\cap F p ∈ Q n ∩ F . By (87) , B n ( p ) ⊂ F B_n(p)\subset F B n ( p ) ⊂ F . Thus the application uses ( Ω , Ω ′ ) = ( B n ( p ) , F ) (\Omega,\Omega')=(B_n(p),F) ( Ω , Ω ′ ) = ( B n ( p ) , F ) in the comparison below. The arguments ( z , θ , t ) (z,\theta,t) ( z , θ , t ) are fixed while the finite region varies.
We first compare the matrices after one elimination.
Lemma 5.1. Let Ω ⊂ Ω ′ \Omega\subset\Omega' Ω ⊂ Ω ′ be finite, let Q ⊂ Z d Q\subset\mathbb{Z}^d Q ⊂ Z d , and let M \mathcal{M} M be a Hermitian matrix on Ω ′ \Omega' Ω ′ such that
∣ M ( x , y ) ∣ ≤ 2 a ⟨ x − y ⟩ − r ( x ≠ y ) , ∣ M ( x , x ) ∣ ≥ β / 2 ( x ∈ Q ∩ Ω ′ ) . |\mathcal{M}(x,y)|\le2a\langle x-y\rangle^{-r}\quad(x\ne y),\qquad|\mathcal{M}(x,x)|\ge\beta/2\quad(x\in Q\cap\Omega'). ∣ M ( x , y ) ∣ ≤ 2 a ⟨ x − y ⟩ − r ( x = y ) , ∣ M ( x , x ) ∣ ≥ β /2 ( x ∈ Q ∩ Ω ′ ) . Assume a / β ≤ ( 64 C 3 ) − 1 a/\beta\le(64C_3)^{-1} a / β ≤ ( 64 C 3 ) − 1 and put P = Ω ∖ Q P=\Omega\setminus Q P = Ω ∖ Q , Y = Q ∩ ( Ω ′ ∖ Ω ) Y=Q\cap(\Omega'\setminus\Omega) Y = Q ∩ ( Ω ′ ∖ Ω ) . The Schur complements
B Ω = Sch Q ∩ Ω ( M Ω ) , B Ω ′ = Sch Q ∩ Ω ′ ( M ) \mathcal{B}_{\Omega}=\operatorname{Sch}_{Q\cap\Omega}(\mathcal{M}_{\Omega}),\qquad\mathcal{B}_{\Omega'}=\operatorname{Sch}_{Q\cap\Omega'}(\mathcal{M}) B Ω = Sch Q ∩ Ω ( M Ω ) , B Ω ′ = Sch Q ∩ Ω ′ ( M ) exist and, for p , q ∈ P p,q\in P p , q ∈ P , satisfy
∣ B Ω ′ ( p , q ) − B Ω ( p , q ) ∣ ≤ 72 a 2 β − 1 ∑ u ∈ Y ⟨ p − u ⟩ − r ⟨ q − u ⟩ − r . (115) |\mathcal{B}_{\Omega'}(p,q)-\mathcal{B}_{\Omega}(p,q)|\le72a^2\beta^{-1}\sum_{u\in Y}\langle p-u\rangle^{-r}\langle q-u\rangle^{-r}. \tag*{(115)} ∣ B Ω ′ ( p , q ) − B Ω ( p , q ) ∣ ≤ 72 a 2 β − 1 u ∈ Y ∑ ⟨ p − u ⟩ − r ⟨ q − u ⟩ − r . (115)
Proof. By the assumption a / β ≤ ( 64 C 3 ) − 1 a/\beta\le(64C_3)^{-1} a / β ≤ ( 64 C 3 ) − 1 , 4 a / β ≤ ( 16 C 3 ) − 1 < ( 2 C 3 ) − 1 4a/\beta\le(16C_3)^{-1}<(2C_3)^{-1} 4 a / β ≤ ( 16 C 3 ) − 1 < ( 2 C 3 ) − 1 . Since Ω ′ \Omega' Ω ′ is 1-separated, the assumptions on M \mathcal{M} M verify the hypotheses of Lemma A.2 with X = Ω ′ X=\Omega' X = Ω ′ , P = Ω ′ ∖ ( Q ∩ Ω ) P=\Omega'\setminus(Q\cap\Omega) P = Ω ′ ∖ ( Q ∩ Ω ) , Q = Q ∩ Ω Q=Q\cap\Omega Q = Q ∩ Ω , M = 1 M=1 M = 1 , a ∗ = 2 a a_*=2a a ∗ = 2 a and Δ = β / 2 \Delta=\beta/2 Δ = β /2 . Hence Sch Q ∩ Ω ( M ) \operatorname{Sch}_{Q\cap\Omega}(\mathcal{M}) Sch Q ∩ Ω ( M ) exists. By (22) ,
( Sch Q ∩ Ω ( M ) ) P = Sch Q ∩ Ω ( M Ω ) = B Ω . \left(\operatorname{Sch}_{Q\cap\Omega}(\mathcal{M})\right)_P=\operatorname{Sch}_{Q\cap\Omega}(\mathcal{M}_{\Omega})=\mathcal{B}_{\Omega}. ( Sch Q ∩ Ω ( M ) ) P = Sch Q ∩ Ω ( M Ω ) = B Ω .
The assumption a / β ≤ ( 64 C 3 ) − 1 a/\beta\le(64C_3)^{-1} a / β ≤ ( 64 C 3 ) − 1 implies
16 C 3 a / β ≤ 1 , 16 C 3 ( a / β ) 2 ≤ 1 4 . 16C_3a/\beta\le1,\qquad16C_3(a/\beta)^2\le\frac{1}{4}. 16 C 3 a / β ≤ 1 , 16 C 3 ( a / β ) 2 ≤ 4 1 .
Thus (A.9) gives, for distinct x , y ∈ Ω ′ ∖ ( Q ∩ Ω ) x,y\in\Omega'\setminus(Q\cap\Omega) x , y ∈ Ω ′ ∖ ( Q ∩ Ω ) ,
∣ Sch Q ∩ Ω ( M ) ( x , y ) ∣ ≤ ( 2 + 16 C 3 a / β ) a ⟨ x − y ⟩ − r ≤ 3 a ⟨ x − y ⟩ − r . |\operatorname{Sch}_{Q\cap\Omega}(\mathcal{M})(x,y)|\le(2+16C_3a/\beta)a\langle x-y\rangle^{-r}\le3a\langle x-y\rangle^{-r}. ∣ Sch Q ∩ Ω ( M ) ( x , y ) ∣ ≤ ( 2 + 16 C 3 a / β ) a ⟨ x − y ⟩ − r ≤ 3 a ⟨ x − y ⟩ − r .
For u ∈ Y u\in Y u ∈ Y , (A.10) gives
∣ Sch Q ∩ Ω ( M ) ( u , u ) ∣ ≥ β / 2 − 16 C 3 a 2 / β ≥ β / 4. |\operatorname{Sch}_{Q\cap\Omega}(\mathcal{M})(u,u)|\ge\beta/2-16C_3a^2/\beta\ge\beta/4. ∣ Sch Q ∩ Ω ( M ) ( u , u ) ∣ ≥ β /2 − 16 C 3 a 2 / β ≥ β /4.
The bounds on the entries of Sch Q ∩ Ω ( M ) \operatorname{Sch}_{Q\cap\Omega}(\mathcal{M}) Sch Q ∩ Ω ( M ) and 12 a / β ≤ ( 2 C 3 ) − 1 12a/\beta\le(2C_3)^{-1} 12 a / β ≤ ( 2 C 3 ) − 1 verify the assumptions of Lemma A.2 with X = Ω ′ ∖ ( Q ∩ Ω ) X=\Omega'\setminus(Q\cap\Omega) X = Ω ′ ∖ ( Q ∩ Ω ) , P = Ω ′ ∖ Q P=\Omega'\setminus Q P = Ω ′ ∖ Q , Q = Y Q=Y Q = Y , M = 1 M=1 M = 1 , a ∗ = 3 a a_*=3a a ∗ = 3 a and Δ = β / 4 \Delta=\beta/4 Δ = β /4 . For u ∈ Y u\in Y u ∈ Y and q ∈ P q\in P q ∈ P , (A.11) yields
∣ [ ( Sch Q ∩ Ω ( M ) ) Y − 1 ( Sch Q ∩ Ω ( M ) ) Y , q ] ( u ) ∣ ≤ 24 a β − 1 ⟨ u − q ⟩ − r . \left|\left[\left(\operatorname{Sch}_{Q\cap\Omega}(\mathcal{M})\right)_Y^{-1}\left(\operatorname{Sch}_{Q\cap\Omega}(\mathcal{M})\right)_{Y,q}\right](u)\right|\le24a\beta^{-1}\langle u-q\rangle^{-r}. [ ( Sch Q ∩ Ω ( M ) ) Y − 1 ( Sch Q ∩ Ω ( M ) ) Y , q ] ( u ) ≤ 24 a β − 1 ⟨ u − q ⟩ − r .
Because ( Q ∩ Ω ) ⊔ Y = Q ∩ Ω ′ (Q\cap\Omega)\sqcup Y=Q\cap\Omega' ( Q ∩ Ω ) ⊔ Y = Q ∩ Ω ′ , (24) implies
B Ω ′ = Sch Y ( Sch Q ∩ Ω ( M ) ) . \mathcal{B}_{\Omega'}=\operatorname{Sch}_Y\left(\operatorname{Sch}_{Q\cap\Omega}(\mathcal{M})\right). B Ω ′ = Sch Y ( Sch Q ∩ Ω ( M ) ) .
Hence, for p , q ∈ P p,q\in P p , q ∈ P , (22) gives
∣ B Ω ′ ( p , q ) − B Ω ( p , q ) ∣ ≤ ∑ u ∈ Y 3 a ⟨ p − u ⟩ − r 24 a β − 1 ⟨ u − q ⟩ − r = 72 a 2 β − 1 ∑ u ∈ Y ⟨ p − u ⟩ − r ⟨ q − u ⟩ − r . |B_{\Omega'}(p,q)-B_{\Omega}(p,q)|\le\sum_{u\in Y}3a\langle p-u\rangle^{-r}24a\beta^{-1}\langle u-q\rangle^{-r}=72a^{2}\beta^{-1}\sum_{u\in Y}\langle p-u\rangle^{-r}\langle q-u\rangle^{-r}. ∣ B Ω ′ ( p , q ) − B Ω ( p , q ) ∣ ≤ u ∈ Y ∑ 3 a ⟨ p − u ⟩ − r 24 a β − 1 ⟨ u − q ⟩ − r = 72 a 2 β − 1 u ∈ Y ∑ ⟨ p − u ⟩ − r ⟨ q − u ⟩ − r .
This proves (115) . □
After further eliminations, the matrices on the common indices may already differ. The next proposition estimates the accumulated difference.
Proposition 5.2. Let Ω ⊂ Ω ′ \Omega\subset\Omega' Ω ⊂ Ω ′ be finite and let i ≥ 1 i\ge1 i ≥ 1 . For this proposition, let S − 1 = Z d ⊃ S 0 ⊃ ⋯ ⊃ S i − 1 S_{-1}=\mathbb{Z}^{d}\supset S_{0}\supset\cdots\supset S_{i-1} S − 1 = Z d ⊃ S 0 ⊃ ⋯ ⊃ S i − 1 be any decreasing family of sets. Define Q k Q_k Q k by (81) for 0 ≤ k < i 0\le k<i 0 ≤ k < i . Assume that
∣ x − y ∣ ≥ M k , x , y ∈ S k − 1 , x ≠ y , 0 ≤ k < i . |x-y|\ge M_k,\qquad x,y\in S_{k-1},\qquad x\ne y,\qquad0\le k<i. ∣ x − y ∣ ≥ M k , x , y ∈ S k − 1 , x = y , 0 ≤ k < i . For F ∈ { Ω , Ω ′ } F\in\{\Omega,\Omega'\} F ∈ { Ω , Ω ′ } , define A F ( k − 1 ) A_F^{(k-1)} A F ( k − 1 ) by (83) using S − 1 , … , S i − 1 S_{-1},\ldots,S_{i-1} S − 1 , … , S i − 1 . Suppose that, for each 0 ≤ k < i 0\le k<i 0 ≤ k < i , the matrices A Ω ( k − 1 ) A_{\Omega}^{(k-1)} A Ω ( k − 1 ) and A Ω ′ ( k − 1 ) A_{\Omega'}^{(k-1)} A Ω ′ ( k − 1 ) exist and satisfy the following conditions:
(1) For F ∈ { Ω , Ω ′ } F\in\{\Omega,\Omega'\} F ∈ { Ω , Ω ′ } ,
∣ A F ( k − 1 ) ( x , y ) ∣ ≤ 2 a ⟨ x − y ⟩ − r , x , y ∈ S k − 1 ∩ F , x ≠ y . |A_F^{(k-1)}(x,y)|\le2a\langle x-y\rangle^{-r},\qquad x,y\in S_{k-1}\cap F,\qquad x\ne y. ∣ A F ( k − 1 ) ( x , y ) ∣ ≤ 2 a ⟨ x − y ⟩ − r , x , y ∈ S k − 1 ∩ F , x = y . (2) For F ∈ { Ω , Ω ′ } F\in\{\Omega,\Omega'\} F ∈ { Ω , Ω ′ } ,
∣ A F ( k − 1 ) ( x , x ) ∣ ≥ δ k / 2 , x ∈ Q k ∩ F . |A_F^{(k-1)}(x,x)|\ge\delta_k/2,\qquad x\in Q_k\cap F. ∣ A F ( k − 1 ) ( x , x ) ∣ ≥ δ k /2 , x ∈ Q k ∩ F . (3)
dist ( x , Ω c ) ≥ M k , x ∈ S k − 1 ∩ Ω . \operatorname{dist}(x,\Omega^{c})\ge M_k,\qquad x\in S_{k-1}\cap\Omega. dist ( x , Ω c ) ≥ M k , x ∈ S k − 1 ∩ Ω. If
a ∑ k ≥ 0 δ k − 1 M k − r ≤ ( 64 C 3 ) − 1 , (116) a\sum_{k\ge0}\delta_k^{-1}M_k^{-r}\le(64C_3)^{-1}, \tag*{(116)} a k ≥ 0 ∑ δ k − 1 M k − r ≤ ( 64 C 3 ) − 1 , (116) then, for p , q ∈ S i − 1 ∩ Ω p,q\in S_{i-1}\cap\Omega p , q ∈ S i − 1 ∩ Ω ,
∣ A Ω ′ ( i − 1 ) ( p , q ) − A Ω ( i − 1 ) ( p , q ) ∣ ≤ 32 C 3 a 2 ∑ k = 0 i − 1 δ k − 1 ∑ u ∈ Q k ∩ ( Ω ′ ∖ Ω ) ⟨ p − u ⟩ − r ⟨ q − u ⟩ − r . (117) |A_{\Omega'}^{(i-1)}(p,q)-A_{\Omega}^{(i-1)}(p,q)|\le32C_3a^2\sum_{k=0}^{i-1}\delta_k^{-1}\sum_{u\in Q_k\cap(\Omega'\setminus\Omega)}\langle p-u\rangle^{-r}\langle q-u\rangle^{-r}. \tag*{(117)} ∣ A Ω ′ ( i − 1 ) ( p , q ) − A Ω ( i − 1 ) ( p , q ) ∣ ≤ 32 C 3 a 2 k = 0 ∑ i − 1 δ k − 1 u ∈ Q k ∩ ( Ω ′ ∖ Ω ) ∑ ⟨ p − u ⟩ − r ⟨ q − u ⟩ − r . (117) Proof. We proceed by induction on the number of eliminated sets. Write
Y k = Q k ∩ ( Ω ′ ∖ Ω ) , 0 ≤ k < i . Y_k=Q_k\cap(\Omega'\setminus\Omega),\qquad0\le k<i. Y k = Q k ∩ ( Ω ′ ∖ Ω ) , 0 ≤ k < i . Use I k I_k I k and P k P_k P k from (82) for the region Ω \Omega Ω . By (83) ,
( A Ω ′ ( − 1 ) ) Ω = A Ω ( − 1 ) . \left(A_{\Omega'}^{(-1)}\right)_{\Omega}=A_{\Omega}^{(-1)}. ( A Ω ′ ( − 1 ) ) Ω = A Ω ( − 1 ) . Since δ 0 = β \delta_0=\beta δ 0 = β , conditions (1)–(2) and (116) verify the assumptions of Lemma 5.1 with Q = Q 0 Q=Q_0 Q = Q 0 and M = A Ω ′ ( − 1 ) M=A_{\Omega'}^{(-1)} M = A Ω ′ ( − 1 ) . Thus (115) holds.
The general step. Let 1 ≤ k < i 1\le k<i 1 ≤ k < i and assume that, for p , q ∈ P k − 1 p,q\in P_{k-1} p , q ∈ P k − 1 ,
∣ A Ω ′ ( k − 1 ) ( p , q ) − A Ω ( k − 1 ) ( p , q ) ∣ ≤ 16 C 3 a 2 ∑ h = 0 k − 1 δ h − 1 ∏ h < ℓ < k ( 1 + 16 C 3 η ℓ ) 2 ∑ u ∈ Y h ⟨ p − u ⟩ − r ⟨ q − u ⟩ − r . (118) |A_{\Omega'}^{(k-1)}(p,q)-A_{\Omega}^{(k-1)}(p,q)|\le16C_3a^2\sum_{h=0}^{k-1}\delta_h^{-1}\prod_{h<\ell<k}(1+16C_3\eta_\ell)^2\sum_{u\in Y_h}\langle p-u\rangle^{-r}\langle q-u\rangle^{-r}. \tag*{(118)} ∣ A Ω ′ ( k − 1 ) ( p , q ) − A Ω ( k − 1 ) ( p , q ) ∣ ≤ 16 C 3 a 2 h = 0 ∑ k − 1 δ h − 1 h < ℓ < k ∏ ( 1 + 16 C 3 η ℓ ) 2 u ∈ Y h ∑ ⟨ p − u ⟩ − r ⟨ q − u ⟩ − r . (118) An empty product equals 1. Since 72 ≤ 16 C 3 72\le16C_3 72 ≤ 16 C 3 , (115) proves (118) for k = 1 k=1 k = 1 . At step k k k , we first eliminate I k I_k I k from the common indices P k − 1 = P k ⊔ I k P_{k-1}=P_k\sqcup I_k P k − 1 = P k ⊔ I k and then Y k Y_k Y k in Ω ′ \Omega' Ω ′ . Write M = A Ω ( k − 1 ) \mathcal{M}=A_{\Omega}^{(k-1)} M = A Ω ( k − 1 ) and M ~ = A Ω ′ ( k − 1 ) \widetilde{\mathcal{M}}=A_{\Omega'}^{(k-1)} M = A Ω ′ ( k − 1 ) . Unlike the original matrices at step 0, M ~ P k − 1 \widetilde{\mathcal{M}}_{P_{k-1}} M P k − 1 and M \mathcal{M} M need not agree. We first estimate Sch I k ( M ~ P k − 1 ) − Sch I k ( M ) \operatorname{Sch}_{I_k}(\widetilde{\mathcal{M}}_{P_{k-1}})-\operatorname{Sch}_{I_k}(\mathcal{M}) Sch I k ( M P k − 1 ) − Sch I k ( M ) , then eliminate Y k Y_k Y k in Ω ′ \Omega' Ω ′ .
Estimating Sch I k ( M ~ P k − 1 ) − Sch I k ( M ) \operatorname{Sch}_{I_k}(\widetilde{\mathcal{M}}_{P_{k-1}})-\operatorname{Sch}_{I_k}(\mathcal{M}) Sch I k ( M P k − 1 ) − Sch I k ( M ) . For F ∈ { Ω , Ω ′ } F \in\{\Omega,\Omega'\} F ∈ { Ω , Ω ′ } , conditions (1)–(2) and (116) verify the assumptions of Lemma A.2 for A F ( k − 1 ) \mathcal{A}^{(k-1)}_F A F ( k − 1 ) with X = S k − 1 ∩ F X=S_{k-1}\cap F X = S k − 1 ∩ F , P = ( S k − 1 ∩ F ) ∖ I k P=(S_{k-1}\cap F)\setminus I_k P = ( S k − 1 ∩ F ) ∖ I k , Q = I k Q=I_k Q = I k , M = M k M=M_k M = M k , a ∗ = 2 a a_*=2a a ∗ = 2 a and Δ = δ k / 2 \Delta=\delta_k/2 Δ = δ k /2 . Indeed, S k − 1 S_{k-1} S k − 1 is M k M_k M k -separated and 4 η k ≤ 4 ( 64 C 3 ) − 1 ≤ ( 2 C 3 ) − 1 4\eta_k\leq4(64C_3)^{-1}\leq(2C_3)^{-1} 4 η k ≤ 4 ( 64 C 3 ) − 1 ≤ ( 2 C 3 ) − 1 . Thus M I k \mathcal{M}_{I_k} M I k and M ~ I k \widetilde{\mathcal{M}}_{I_k} M I k are invertible. Let W = W ( M ) \mathcal{W}=\mathcal{W}(\mathcal{M}) W = W ( M ) and W ~ = W ( M ~ P k − 1 ) \widetilde{\mathcal{W}}=\mathcal{W}(\widetilde{\mathcal{M}}_{P_{k-1}}) W = W ( M P k − 1 ) , where the reconstruction matrices in (23) use P = P k P=P_k P = P k and Q = I k Q=I_k Q = I k . By (25) , we have
Sch I k ( M ~ P k − 1 ) − Sch I k ( M ) = W ∗ ( M ~ P k − 1 − M ) W ~ . \operatorname{Sch}_{I_k}(\widetilde{\mathcal{M}}_{P_{k-1}})-\operatorname{Sch}_{I_k}(\mathcal{M})=\mathcal{W}^{*}(\widetilde{\mathcal{M}}_{P_{k-1}}-\mathcal{M})\widetilde{\mathcal{W}}. Sch I k ( M P k − 1 ) − Sch I k ( M ) = W ∗ ( M P k − 1 − M ) W .
Fix u ∈ Y h u\in Y_h u ∈ Y h with 0 ≤ h < k 0\leq h<k 0 ≤ h < k and p ∈ P k p\in P_k p ∈ P k . Since u ∉ Ω u\notin\Omega u ∈ / Ω , condition (3) implies
dist ( u , P k − 1 ) ≥ M k . \operatorname{dist}(u,P_{k-1})\geq M_k. dist ( u , P k − 1 ) ≥ M k .
The set P k − 1 ⊂ S k − 1 P_{k-1}\subset S_{k-1} P k − 1 ⊂ S k − 1 is M k M_k M k -separated and I k ⊂ P k − 1 ∖ { p } I_k\subset P_{k-1}\setminus\{p\} I k ⊂ P k − 1 ∖ { p } . Thus (A.11) and (A.3) imply
∑ x ∈ P k − 1 ∣ W ( x , p ) ∣ ⟨ x − u ⟩ − r ≤ ⟨ p − u ⟩ − r + 8 a δ k − 1 ∑ x ∈ I k ⟨ p − x ⟩ − r ⟨ x − u ⟩ − r ≤ ( 1 + 16 C 3 η k ) ⟨ p − u ⟩ − r . \begin{aligned}
\sum_{x\in P_{k-1}}|\mathcal{W}(x,p)|\langle x-u\rangle^{-r}\leq\langle p-u\rangle^{-r}+8a\delta_k^{-1}\sum_{x\in I_k}\langle p-x\rangle^{-r}\langle x-u\rangle^{-r} \\
&\leq(1+16C_3\eta_k)\langle p-u\rangle^{-r}.
\end{aligned} x ∈ P k − 1 ∑ ∣ W ( x , p ) ∣ ⟨ x − u ⟩ − r ≤ ⟨ p − u ⟩ − r + 8 a δ k − 1 x ∈ I k ∑ ⟨ p − x ⟩ − r ⟨ x − u ⟩ − r ≤ ( 1 + 16 C 3 η k ) ⟨ p − u ⟩ − r .
By (A.11) and (A.3), we also have
∑ y ∈ P k − 1 ∣ W ~ ( y , q ) ∣ ⟨ y − u ⟩ − r ≤ ( 1 + 16 C 3 η k ) ⟨ q − u ⟩ − r , q ∈ P k . \sum_{y\in P_{k-1}}|\widetilde{\mathcal{W}}(y,q)|\langle y-u\rangle^{-r}\leq(1+16C_3\eta_k)\langle q-u\rangle^{-r},\qquad q\in P_k. y ∈ P k − 1 ∑ ∣ W ( y , q ) ∣ ⟨ y − u ⟩ − r ≤ ( 1 + 16 C 3 η k ) ⟨ q − u ⟩ − r , q ∈ P k .
Multiplying the bounds for W \mathcal{W} W and W ~ \widetilde{\mathcal{W}} W , we obtain
∑ x , y ∈ P k − 1 ∣ W ( x , p ) ∣ ⟨ x − u ⟩ − r ⟨ y − u ⟩ − r ∣ W ~ ( y , q ) ∣ ≤ ( 1 + 16 C 3 η k ) 2 ⟨ p − u ⟩ − r ⟨ q − u ⟩ − r . \sum_{x,y\in P_{k-1}}|\mathcal{W}(x,p)|\langle x-u\rangle^{-r}\langle y-u\rangle^{-r}|\widetilde{\mathcal{W}}(y,q)|\leq(1+16C_3\eta_k)^2\langle p-u\rangle^{-r}\langle q-u\rangle^{-r}. x , y ∈ P k − 1 ∑ ∣ W ( x , p ) ∣ ⟨ x − u ⟩ − r ⟨ y − u ⟩ − r ∣ W ( y , q ) ∣ ≤ ( 1 + 16 C 3 η k ) 2 ⟨ p − u ⟩ − r ⟨ q − u ⟩ − r .
Eliminating Y k Y_k Y k . Let M ^ ( k ) = Sch I k ( M ~ ) \widehat{\mathcal{M}}^{(k)}=\operatorname{Sch}_{I_k}(\widetilde{\mathcal{M}}) M ( k ) = Sch I k ( M ) . By (A.10),
∣ M ^ ( k ) ( u , u ) − M ~ ( u , u ) ∣ ≤ 16 C 3 a 2 δ k − 1 M k − 2 r = 16 C 3 η k 2 δ k , u ∈ Y k . |\widehat{\mathcal{M}}^{(k)}(u,u)-\widetilde{\mathcal{M}}(u,u)|\leq16C_3a^2\delta_k^{-1}M_k^{-2r}=16C_3\eta_k^2\delta_k,\qquad u\in Y_k. ∣ M ( k ) ( u , u ) − M ( u , u ) ∣ ≤ 16 C 3 a 2 δ k − 1 M k − 2 r = 16 C 3 η k 2 δ k , u ∈ Y k .
By Proposition 5.2 (2), (A.9) and η k ≤ ( 64 C 3 ) − 1 \eta_k\leq(64C_3)^{-1} η k ≤ ( 64 C 3 ) − 1 , we obtain
∣ M ^ ( k ) ( u , u ) ∣ ≥ δ k / 2 − 16 C 3 η k 2 δ k ≥ δ k / 4 , u ∈ Y k , |\widehat{\mathcal{M}}^{(k)}(u,u)|\geq\delta_k/2-16C_3\eta_k^2\delta_k\geq\delta_k/4,\qquad u\in Y_k, ∣ M ( k ) ( u , u ) ∣ ≥ δ k /2 − 16 C 3 η k 2 δ k ≥ δ k /4 , u ∈ Y k ,
∣ M ^ ( k ) ( x , y ) ∣ ≤ ( 2 + 16 C 3 η k ) a ⟨ x − y ⟩ − r ≤ 3 a ⟨ x − y ⟩ − r , x ≠ y . |\widehat{\mathcal{M}}^{(k)}(x,y)|\leq(2+16C_3\eta_k)a\langle x-y\rangle^{-r}\leq3a\langle x-y\rangle^{-r},\qquad x\neq y. ∣ M ( k ) ( x , y ) ∣ ≤ ( 2 + 16 C 3 η k ) a ⟨ x − y ⟩ − r ≤ 3 a ⟨ x − y ⟩ − r , x = y .
In the second line, x , y ∈ ( S k − 1 ∩ Ω ′ ) ∖ I k x,y\in(S_{k-1}\cap\Omega')\setminus I_k x , y ∈ ( S k − 1 ∩ Ω ′ ) ∖ I k . The entry bounds on M ^ ( k ) \widehat{\mathcal{M}}^{(k)} M ( k ) and 12 η k ≤ 12 ( 64 C 3 ) − 1 ≤ ( 2 C 3 ) − 1 12\eta_k\leq12(64C_3)^{-1}\leq(2C_3)^{-1} 12 η k ≤ 12 ( 64 C 3 ) − 1 ≤ ( 2 C 3 ) − 1 verify the assumptions of Lemma A.2 with X = ( S k − 1 ∩ Ω ′ ) ∖ I k X=(S_{k-1}\cap\Omega')\setminus I_k X = ( S k − 1 ∩ Ω ′ ) ∖ I k , P = S k ∩ Ω ′ P=S_k\cap\Omega' P = S k ∩ Ω ′ , Q = Y k Q=Y_k Q = Y k , M = M k M=M_k M = M k , a ∗ = 3 a a_*=3a a ∗ = 3 a and Δ = δ k / 4 \Delta=\delta_k/4 Δ = δ k /4 . Hence M ^ Y k ( k ) \widehat{\mathcal{M}}^{(k)}_{Y_k} M Y k ( k ) is invertible. By (A.11), we have
∣ ( ( M ^ Y k ( k ) ) − 1 M ^ Y k , q ( k ) ) ( u ) ∣ ≤ 24 a δ k − 1 ⟨ u − q ⟩ − r , u ∈ Y k , q ∈ P k . |((\widehat{\mathcal{M}}^{(k)}_{Y_k})^{-1}\widehat{\mathcal{M}}^{(k)}_{Y_k,q})(u)|\leq24a\delta_k^{-1}\langle u-q\rangle^{-r},\qquad u\in Y_k,\quad q\in P_k. ∣ (( M Y k ( k ) ) − 1 M Y k , q ( k ) ) ( u ) ∣ ≤ 24 a δ k − 1 ⟨ u − q ⟩ − r , u ∈ Y k , q ∈ P k .
It follows that
∣ M ^ p , Y k ( k ) ( M ^ Y k ( k ) ) − 1 M ^ Y k , q ( k ) ∣ ≤ 72 a 2 δ k − 1 ∑ u ∈ Y k ⟨ p − u ⟩ − r ⟨ q − u ⟩ − r . (119) |\widehat{\mathcal{M}}^{(k)}_{p,Y_k}(\widehat{\mathcal{M}}^{(k)}_{Y_k})^{-1}\widehat{\mathcal{M}}^{(k)}_{Y_k,q}|\leq72a^2\delta_k^{-1}\sum_{u\in Y_k}\langle p-u\rangle^{-r}\langle q-u\rangle^{-r}. \tag*{(119)} ∣ M p , Y k ( k ) ( M Y k ( k ) ) − 1 M Y k , q ( k ) ∣ ≤ 72 a 2 δ k − 1 u ∈ Y k ∑ ⟨ p − u ⟩ − r ⟨ q − u ⟩ − r . (119)
Summation over the steps. Since I_k\mathbin{\mathaccent\cdot\cup}Y_k=Q_k\cap\Omega' , we have
Sch Y k ( Sch I k ( M ~ ) ) = Sch Q k ∩ Ω ′ ( M ~ ) = A Ω ′ ( k ) . \operatorname{Sch}_{Y_k}(\operatorname{Sch}_{I_k}(\widetilde{\mathcal{M}}))=\operatorname{Sch}_{Q_k\cap\Omega'}(\widetilde{\mathcal{M}})=\mathcal{A}_{\Omega'}^{(k)}. Sch Y k ( Sch I k ( M )) = Sch Q k ∩ Ω ′ ( M ) = A Ω ′ ( k ) .
Substituting (118) into (25) and adding (119) proves (118) with k k k replaced by k + 1 k+1 k + 1 . Moreover, (116) implies
∏ k < i ( 1 + 16 C 3 η k ) 2 ≤ exp ( 32 C 3 ∑ k ≥ 0 η k ) ≤ e 1 / 2 < 2. (120) \prod_{k<i}(1+16C_3\eta_k)^2\leq\exp\left(32C_3\sum_{k\geq0}\eta_k\right)\leq e^{1/2}<2. \tag*{(120)} k < i ∏ ( 1 + 16 C 3 η k ) 2 ≤ exp ( 32 C 3 k ≥ 0 ∑ η k ) ≤ e 1/2 < 2. (120)
Since Y h = Q h ∩ ( Ω ′ ∖ Ω ) Y_h=Q_h\cap(\Omega'\setminus\Omega) Y h = Q h ∩ ( Ω ′ ∖ Ω ) , combining (120) with (118) at k = i k=i k = i proves (117) . □ \square □
Corollary 5.3 (Diagonal comparison). Under the assumptions of Proposition 5.2, let p ∈ S i − 1 ∩ Ω p \in S_{i-1} \cap\Omega p ∈ S i − 1 ∩ Ω and let R = dist ( p , Ω c ) ≥ 1 R=\operatorname{dist}(p,\Omega^{c}) \ge1 R = dist ( p , Ω c ) ≥ 1 . If M k ≤ R M_k \le R M k ≤ R for 0 ≤ k < i 0 \le k < i 0 ≤ k < i , then
∣ A Ω ′ ( i − 1 ) ( p , p ) − A Ω ( i − 1 ) ( p , p ) ∣ ≤ C 5 2 a 2 β − 1 R τ − 2 r . (121) \left|A_{\Omega'}^{(i-1)}(p,p)-A_{\Omega}^{(i-1)}(p,p)\right| \le\frac{C_5}{2}a^2\beta^{-1}R^{\tau-2r}. \tag*{(121)} A Ω ′ ( i − 1 ) ( p , p ) − A Ω ( i − 1 ) ( p , p ) ≤ 2 C 5 a 2 β − 1 R τ − 2 r . (121) Proof. For 0 ≤ k < i 0 \le k < i 0 ≤ k < i , Q k ⊂ S k − 1 Q_k \subset S_{k-1} Q k ⊂ S k − 1 is M k M_k M k -separated and
∣ p − u ∣ ≥ dist ( p , Ω c ) = R ≥ M k , u ∈ Q k ∩ ( Ω ′ ∖ Ω ) . |p-u| \ge\operatorname{dist}(p,\Omega^{c})=R \ge M_k,\qquad u \in Q_k \cap(\Omega'\setminus\Omega). ∣ p − u ∣ ≥ dist ( p , Ω c ) = R ≥ M k , u ∈ Q k ∩ ( Ω ′ ∖ Ω ) . By (A.4),
∑ u ∈ Q k ∩ ( Ω ′ ∖ Ω ) ⟨ p − u ⟩ − 2 r ≤ C 3 M k − d R d − 2 r . \sum_{u\in Q_k\cap(\Omega'\setminus\Omega)}\langle p-u\rangle^{-2r} \le C_3M_k^{-d}R^{d-2r}. u ∈ Q k ∩ ( Ω ′ ∖ Ω ) ∑ ⟨ p − u ⟩ − 2 r ≤ C 3 M k − d R d − 2 r . By (117) ,
∣ A Ω ′ ( i − 1 ) ( p , p ) − A Ω ( i − 1 ) ( p , p ) ∣ ≤ 32 C 3 2 a 2 R d − 2 r ∑ k < i δ k − 1 M k − d . \left|A_{\Omega'}^{(i-1)}(p,p)-A_{\Omega}^{(i-1)}(p,p)\right| \le32C_3^2a^2R^{d-2r}\sum_{k<i}\delta_k^{-1}M_k^{-d}. A Ω ′ ( i − 1 ) ( p , p ) − A Ω ( i − 1 ) ( p , p ) ≤ 32 C 3 2 a 2 R d − 2 r k < i ∑ δ k − 1 M k − d . Since τ > d \tau>d τ > d and M k ≤ R M_k\le R M k ≤ R , we have
∑ k < i δ k − 1 M k − d = β − 1 ( 1 + 16 τ ∑ k = 1 i − 1 N k τ − d ) ≤ β − 1 ( 1 + 16 τ R τ − d 1 − 16 d − τ ) ≤ ( 1 + 16 τ 1 − 16 d − τ ) β − 1 R τ − d . \sum_{k<i}\delta_k^{-1}M_k^{-d}
=\beta^{-1}\left(1+16^\tau\sum_{k=1}^{i-1}N_k^{\tau-d}\right)
\le\beta^{-1}\left(1+\frac{16^\tau R^{\tau-d}}{1-16^{d-\tau}}\right)
\le\left(1+\frac{16^\tau}{1-16^{d-\tau}}\right)\beta^{-1}R^{\tau-d}. k < i ∑ δ k − 1 M k − d = β − 1 ( 1 + 1 6 τ k = 1 ∑ i − 1 N k τ − d ) ≤ β − 1 ( 1 + 1 − 1 6 d − τ 1 6 τ R τ − d ) ≤ ( 1 + 1 − 1 6 d − τ 1 6 τ ) β − 1 R τ − d . The sum over k ≥ 1 k\ge1 k ≥ 1 is zero when i = 1 i=1 i = 1 . The coefficient 32 C 3 2 ( 1 + 16 τ / ( 1 − 16 d − τ ) ) 32C_3^2(1+16^\tau/(1-16^{d-\tau})) 32 C 3 2 ( 1 + 1 6 τ / ( 1 − 1 6 d − τ )) is finite and depends only on d , r , τ d,r,\tau d , r , τ . This is the coefficient required to be at most C 5 / 2 C_5/2 C 5 /2 when C 5 C_5 C 5 was fixed in Subsection 4.2, and proves (5.7). □
Dividing (5.7) by β R − τ \beta R^{-\tau} β R − τ , we obtain
∣ A Ω ′ ( i − 1 ) ( p , p ) − A Ω ( i − 1 ) ( p , p ) ∣ β R − τ ≤ C 5 2 a 2 β − 2 R 2 τ − 2 r . (122) \frac{\left|A_{\Omega'}^{(i-1)}(p,p)-A_{\Omega}^{(i-1)}(p,p)\right|}{\beta R^{-\tau}}
\le\frac{C_5}{2}a^2\beta^{-2}R^{2\tau-2r}. \tag*{(122)} β R − τ A Ω ′ ( i − 1 ) ( p , p ) − A Ω ( i − 1 ) ( p , p ) ≤ 2 C 5 a 2 β − 2 R 2 τ − 2 r . (122)
Since r > τ r>\tau r > τ , the right-hand side of (5.8) is at most C 5 2 a 2 β − 2 \frac{C_5}{2}a^2\beta^{-2} 2 C 5 a 2 β − 2 for every R ≥ 1 R\ge1 R ≥ 1 . Thus (95) bounds the quotient in (5.8) uniformly for R ≥ 1 R\ge1 R ≥ 1 .
The new diagonal bound and the next elimination
We use (108) only for 1 ≤ h ≤ n 1\le h\le n 1 ≤ h ≤ n ; for h = 0 h=0 h = 0 , the required inequality follows from E 0 = f E_0=f E 0 = f and (8) .
Lemma 5.4. For every θ ∈ Θ \theta\in\Theta θ ∈ Θ , z ∈ R z\in\mathbb{R} z ∈ R and Ω ∈ Υ n \Omega\in\Upsilon_n Ω ∈ Υ n , the matrix ( A Ω ( n − 1 ) ) Q n ∩ Ω (A_{\Omega}^{(n-1)})_{Q_n\cap\Omega} ( A Ω ( n − 1 ) ) Q n ∩ Ω is invertible and
∣ A Ω ( n − 1 ) ( p , p ) ∣ ≥ δ n / 2 , p ∈ Q n ∩ Ω , (123) |A_{\Omega}^{(n-1)}(p,p)|\ge\delta_n/2,\qquad p\in Q_n\cap\Omega, \tag*{(123)} ∣ A Ω ( n − 1 ) ( p , p ) ∣ ≥ δ n /2 , p ∈ Q n ∩ Ω , (123) ∣ A Ω ( n ) ( p , q ) ∣ ≤ 2 a ⟨ p − q ⟩ − r , p , q ∈ S n ∩ Ω , p ≠ q . (124) |A_{\Omega}^{(n)}(p,q)|\le2a\langle p-q\rangle^{-r},\qquad p,q\in S_n\cap\Omega,\quad p\ne q. \tag*{(124)} ∣ A Ω ( n ) ( p , q ) ∣ ≤ 2 a ⟨ p − q ⟩ − r , p , q ∈ S n ∩ Ω , p = q . (124) Consequently, assertion (4) extends to i = n i=n i = n .
Proof. Fix θ ∈ Θ \theta\in\Theta θ ∈ Θ , z ∈ R z\in\mathbb{R} z ∈ R , Ω ∈ Υ n \Omega\in\Upsilon_n Ω ∈ Υ n and p ∈ S n − 1 ∩ Ω p\in S_{n-1}\cap\Omega p ∈ S n − 1 ∩ Ω . By (87) , B n ( p ) ⊂ Ω B_n(p)\subset\Omega B n ( p ) ⊂ Ω . If 1 ≤ h ≤ n 1\le h\le n 1 ≤ h ≤ n and q ∈ B n ( p ) ∩ S h − 1 q\in B_n(p)\cap S_{h-1} q ∈ B n ( p ) ∩ S h − 1 , then p , q ∈ Res h − 1 p,q\in\operatorname{Res}_{h-1} p , q ∈ Res h − 1 by (80) . By (66) ,
L ∥ ( q − p ) ⋅ ω ∥ T ≤ ∣ E h − 1 ( θ + q ⋅ ω ) − E h − 1 ( θ + p ⋅ ω ) ∣ < 2 δ h − 1 . (125) L\lVert(q-p)\cdot\omega\rVert_T\le|E_{h-1}(\theta+q\cdot\omega)-E_{h-1}(\theta+p\cdot\omega)|<2\delta_{h-1}. \tag*{(125)} L ∥( q − p ) ⋅ ω ∥ T ≤ ∣ E h − 1 ( θ + q ⋅ ω ) − E h − 1 ( θ + p ⋅ ω ) ∣ < 2 δ h − 1 . (125)
By (5.11) and (77) , q − p ∈ Z h − 1 q-p\in Z_{h-1} q − p ∈ Z h − 1 . For 1 ≤ h ≤ n − 1 1\le h\le n-1 1 ≤ h ≤ n − 1 , (101) implies B h ( q ) ⊂ B n ( p ) B_h(q)\subset B_n(p) B h ( q ) ⊂ B n ( p ) . This proves B n ( p ) ∈ Υ n − 1 B_n(p)\in\Upsilon_{n-1} B n ( p ) ∈ Υ n − 1 . For h = n h=n h = n , (102) implies q = p q=p q = p . Since p ∈ S n − 1 p\in S_{n-1} p ∈ S n − 1 , we have
B n ( p ) ∩ S n − 1 = { p } . (126) B_n(p)\cap S_{n-1}=\{p\}. \tag*{(126)} B n ( p ) ∩ S n − 1 = { p } . (126)
We verify the assumptions of Proposition 5.2 with i = n i=n i = n and regions B n ( p ) ⊂ Ω B_n(p)\subset\Omega B n ( p ) ⊂ Ω . Since B n ( p ) , Ω ∈ Υ n − 1 B_n(p),\Omega\in\Upsilon_{n-1} B n ( p ) , Ω ∈ Υ n − 1 , (109) –(110) verify conditions (1)–(2) through step n − 1 n-1 n − 1 . For 1 ≤ k < n 1\le k<n 1 ≤ k < n and x ∈ B n ( p ) ∩ S k − 1 x\in B_n(p)\cap S_{k-1} x ∈ B n ( p ) ∩ S k − 1 , (87) implies
B k ( x ) ⊂ B n ( p ) , dist ( x , B n ( p ) c ) ≥ 4 N k ≥ M k . B_k(x)\subset B_n(p),\qquad\operatorname{dist}(x,B_n(p)^c)\ge4N_k\ge M_k. B k ( x ) ⊂ B n ( p ) , dist ( x , B n ( p ) c ) ≥ 4 N k ≥ M k .
@page 22/35 starts_mid=0 ends_mid=1
Corollary 5.3 (Diagonal comparison). Under the assumptions of Proposition 5.2, let p ∈ S i − 1 ∩ Ω p \in S_{i-1} \cap\Omega p ∈ S i − 1 ∩ Ω and let R = dist ( p , Ω c ) ≥ 1 R=\operatorname{dist}(p,\Omega^{c}) \ge1 R = dist ( p , Ω c ) ≥ 1 . If M k ≤ R M_k \le R M k ≤ R for 0 ≤ k < i 0 \le k < i 0 ≤ k < i , then
∣ A Ω ′ ( i − 1 ) ( p , p ) − A Ω ( i − 1 ) ( p , p ) ∣ ≤ C 5 2 a 2 β − 1 R τ − 2 r . (127) \left|A_{\Omega'}^{(i-1)}(p,p)-A_{\Omega}^{(i-1)}(p,p)\right| \le\frac{C_5}{2}a^2\beta^{-1}R^{\tau-2r}. \tag*{(127)} A Ω ′ ( i − 1 ) ( p , p ) − A Ω ( i − 1 ) ( p , p ) ≤ 2 C 5 a 2 β − 1 R τ − 2 r . (127) Proof. For 0 ≤ k < i 0 \le k < i 0 ≤ k < i , Q k ⊂ S k − 1 Q_k \subset S_{k-1} Q k ⊂ S k − 1 is M k M_k M k -separated and
∣ p − u ∣ ≥ dist ( p , Ω c ) = R ≥ M k , u ∈ Q k ∩ ( Ω ′ ∖ Ω ) . |p-u| \ge\operatorname{dist}(p,\Omega^{c})=R \ge M_k,\qquad u \in Q_k \cap(\Omega'\setminus\Omega). ∣ p − u ∣ ≥ dist ( p , Ω c ) = R ≥ M k , u ∈ Q k ∩ ( Ω ′ ∖ Ω ) . By (A.4),
∑ u ∈ Q k ∩ ( Ω ′ ∖ Ω ) ⟨ p − u ⟩ − 2 r ≤ C 3 M k − d R d − 2 r . \sum_{u\in Q_k\cap(\Omega'\setminus\Omega)}\langle p-u\rangle^{-2r} \le C_3M_k^{-d}R^{d-2r}. u ∈ Q k ∩ ( Ω ′ ∖ Ω ) ∑ ⟨ p − u ⟩ − 2 r ≤ C 3 M k − d R d − 2 r . By (117) ,
∣ A Ω ′ ( i − 1 ) ( p , p ) − A Ω ( i − 1 ) ( p , p ) ∣ ≤ 32 C 3 2 a 2 R d − 2 r ∑ k < i δ k − 1 M k − d . \left|A_{\Omega'}^{(i-1)}(p,p)-A_{\Omega}^{(i-1)}(p,p)\right| \le32C_3^2a^2R^{d-2r}\sum_{k<i}\delta_k^{-1}M_k^{-d}. A Ω ′ ( i − 1 ) ( p , p ) − A Ω ( i − 1 ) ( p , p ) ≤ 32 C 3 2 a 2 R d − 2 r k < i ∑ δ k − 1 M k − d . Since τ > d \tau>d τ > d and M k ≤ R M_k\le R M k ≤ R , we have
∑ k < i δ k − 1 M k − d = β − 1 ( 1 + 16 τ ∑ k = 1 i − 1 N k τ − d ) ≤ β − 1 ( 1 + 16 τ R τ − d 1 − 16 d − τ ) ≤ ( 1 + 16 τ 1 − 16 d − τ ) β − 1 R τ − d . \sum_{k<i}\delta_k^{-1}M_k^{-d}
=\beta^{-1}\left(1+16^\tau\sum_{k=1}^{i-1}N_k^{\tau-d}\right)
\le\beta^{-1}\left(1+\frac{16^\tau R^{\tau-d}}{1-16^{d-\tau}}\right)
\le\left(1+\frac{16^\tau}{1-16^{d-\tau}}\right)\beta^{-1}R^{\tau-d}. k < i ∑ δ k − 1 M k − d = β − 1 ( 1 + 1 6 τ k = 1 ∑ i − 1 N k τ − d ) ≤ β − 1 ( 1 + 1 − 1 6 d − τ 1 6 τ R τ − d ) ≤ ( 1 + 1 − 1 6 d − τ 1 6 τ ) β − 1 R τ − d . The sum over k ≥ 1 k\ge1 k ≥ 1 is zero when i = 1 i=1 i = 1 . The coefficient 32 C 3 2 ( 1 + 16 τ / ( 1 − 16 d − τ ) ) 32C_3^2(1+16^\tau/(1-16^{d-\tau})) 32 C 3 2 ( 1 + 1 6 τ / ( 1 − 1 6 d − τ )) is finite and depends only on d , r , τ d,r,\tau d , r , τ . This is the coefficient required to be at most C 5 / 2 C_5/2 C 5 /2 when C 5 C_5 C 5 was fixed in Subsection 4.2, and proves (5.7). □
Dividing (5.7) by β R − τ \beta R^{-\tau} β R − τ , we obtain
∣ A Ω ′ ( i − 1 ) ( p , p ) − A Ω ( i − 1 ) ( p , p ) ∣ β R − τ ≤ C 5 2 a 2 β − 2 R 2 τ − 2 r . (128) \frac{\left|A_{\Omega'}^{(i-1)}(p,p)-A_{\Omega}^{(i-1)}(p,p)\right|}{\beta R^{-\tau}}
\le\frac{C_5}{2}a^2\beta^{-2}R^{2\tau-2r}. \tag*{(128)} β R − τ A Ω ′ ( i − 1 ) ( p , p ) − A Ω ( i − 1 ) ( p , p ) ≤ 2 C 5 a 2 β − 2 R 2 τ − 2 r . (128)
Since r > τ r>\tau r > τ , the right-hand side of (5.8) is at most C 5 2 a 2 β − 2 \frac{C_5}{2}a^2\beta^{-2} 2 C 5 a 2 β − 2 for every R ≥ 1 R\ge1 R ≥ 1 . Thus (95) bounds the quotient in (5.8) uniformly for R ≥ 1 R\ge1 R ≥ 1 .
The new diagonal bound and the next elimination
We use (108) only for 1 ≤ h ≤ n 1\le h\le n 1 ≤ h ≤ n ; for h = 0 h=0 h = 0 , the required inequality follows from E 0 = f E_0=f E 0 = f and (8) .
Lemma 5.4. For every θ ∈ Θ \theta\in\Theta θ ∈ Θ , z ∈ R z\in\mathbb{R} z ∈ R and Ω ∈ Υ n \Omega\in\Upsilon_n Ω ∈ Υ n , the matrix ( A Ω ( n − 1 ) ) Q n ∩ Ω (A_{\Omega}^{(n-1)})_{Q_n\cap\Omega} ( A Ω ( n − 1 ) ) Q n ∩ Ω is invertible and
∣ A Ω ( n − 1 ) ( p , p ) ∣ ≥ δ n / 2 , p ∈ Q n ∩ Ω , (129) |A_{\Omega}^{(n-1)}(p,p)|\ge\delta_n/2,\qquad p\in Q_n\cap\Omega, \tag*{(129)} ∣ A Ω ( n − 1 ) ( p , p ) ∣ ≥ δ n /2 , p ∈ Q n ∩ Ω , (129) ∣ A Ω ( n ) ( p , q ) ∣ ≤ 2 a ⟨ p − q ⟩ − r , p , q ∈ S n ∩ Ω , p ≠ q . (130) |A_{\Omega}^{(n)}(p,q)|\le2a\langle p-q\rangle^{-r},\qquad p,q\in S_n\cap\Omega,\quad p\ne q. \tag*{(130)} ∣ A Ω ( n ) ( p , q ) ∣ ≤ 2 a ⟨ p − q ⟩ − r , p , q ∈ S n ∩ Ω , p = q . (130) Consequently, assertion (4) extends to i = n i=n i = n .
Proof. Fix θ ∈ Θ \theta\in\Theta θ ∈ Θ , z ∈ R z\in\mathbb{R} z ∈ R , Ω ∈ Υ n \Omega\in\Upsilon_n Ω ∈ Υ n and p ∈ S n − 1 ∩ Ω p\in S_{n-1}\cap\Omega p ∈ S n − 1 ∩ Ω . By (87) , B n ( p ) ⊂ Ω B_n(p)\subset\Omega B n ( p ) ⊂ Ω . If 1 ≤ h ≤ n 1\le h\le n 1 ≤ h ≤ n and q ∈ B n ( p ) ∩ S h − 1 q\in B_n(p)\cap S_{h-1} q ∈ B n ( p ) ∩ S h − 1 , then p , q ∈ Res h − 1 p,q\in\operatorname{Res}_{h-1} p , q ∈ Res h − 1 by (80) . By (66) ,
L ∥ ( q − p ) ⋅ ω ∥ T ≤ ∣ E h − 1 ( θ + q ⋅ ω ) − E h − 1 ( θ + p ⋅ ω ) ∣ < 2 δ h − 1 . (131) L\lVert(q-p)\cdot\omega\rVert_T\le|E_{h-1}(\theta+q\cdot\omega)-E_{h-1}(\theta+p\cdot\omega)|<2\delta_{h-1}. \tag*{(131)} L ∥( q − p ) ⋅ ω ∥ T ≤ ∣ E h − 1 ( θ + q ⋅ ω ) − E h − 1 ( θ + p ⋅ ω ) ∣ < 2 δ h − 1 . (131)
By (5.11) and (77) , q − p ∈ Z h − 1 q-p\in Z_{h-1} q − p ∈ Z h − 1 . For 1 ≤ h ≤ n − 1 1\le h\le n-1 1 ≤ h ≤ n − 1 , (101) implies B h ( q ) ⊂ B n ( p ) B_h(q)\subset B_n(p) B h ( q ) ⊂ B n ( p ) . This proves B n ( p ) ∈ Υ n − 1 B_n(p)\in\Upsilon_{n-1} B n ( p ) ∈ Υ n − 1 . For h = n h=n h = n , (102) implies q = p q=p q = p . Since p ∈ S n − 1 p\in S_{n-1} p ∈ S n − 1 , we have
B n ( p ) ∩ S n − 1 = { p } . (132) B_n(p)\cap S_{n-1}=\{p\}. \tag*{(132)} B n ( p ) ∩ S n − 1 = { p } . (132)
We verify the assumptions of Proposition 5.2 with i = n i=n i = n and regions B n ( p ) ⊂ Ω B_n(p)\subset\Omega B n ( p ) ⊂ Ω . Since B n ( p ) , Ω ∈ Υ n − 1 B_n(p),\Omega\in\Upsilon_{n-1} B n ( p ) , Ω ∈ Υ n − 1 , (109) –(110) verify conditions (1)–(2) through step n − 1 n-1 n − 1 . For 1 ≤ k < n 1\le k<n 1 ≤ k < n and x ∈ B n ( p ) ∩ S k − 1 x\in B_n(p)\cap S_{k-1} x ∈ B n ( p ) ∩ S k − 1 , (87) implies
B k ( x ) ⊂ B n ( p ) , dist ( x , B n ( p ) c ) ≥ 4 N k ≥ M k . B_k(x)\subset B_n(p),\qquad\operatorname{dist}(x,B_n(p)^c)\ge4N_k\ge M_k. B k ( x ) ⊂ B n ( p ) , dist ( x , B n ( p ) c ) ≥ 4 N k ≥ M k .
At k = 0 k=0 k = 0 , dist ( x , B n ( p ) c ) ≥ 1 = M 0 \operatorname{dist}(x,B_n(p)^c)\ge1=M_0 dist ( x , B n ( p ) c ) ≥ 1 = M 0 for every x ∈ B n ( p ) x\in B_n(p) x ∈ B n ( p ) . This proves condition (3). The sets S k − 1 ⊂ Res k − 1 S_{k-1}\subset\operatorname{Res}_{k-1} S k − 1 ⊂ Res k − 1 are M k M_k M k -separated for 1 ≤ k < n 1\le k<n 1 ≤ k < n : indeed, (66) and (79) give
∣ x − y ∣ ≥ 1 2 ( L γ 2 β ) 1 / τ N k − 1 − 1 > M k ( x , y ∈ S k − 1 , x ≠ y ) . |x-y|\ge\frac{1}{2}\left(\frac{L\gamma}{2\beta}\right)^{1/\tau}N_{k-1}^{-1}>M_k\qquad(x,y\in S_{k-1},\ x\ne y). ∣ x − y ∣ ≥ 2 1 ( 2 β L γ ) 1/ τ N k − 1 − 1 > M k ( x , y ∈ S k − 1 , x = y ) .
The set S − 1 = Z d S_{-1}=\mathbb{Z}^d S − 1 = Z d is M 0 M_0 M 0 -separated. Finally, (94) verifies (116) . By (5.12), eliminating Q 0 ∩ B n ( p ) , … , Q n − 1 ∩ B n ( p ) Q_0\cap B_n(p),\ldots,Q_{n-1}\cap B_n(p) Q 0 ∩ B n ( p ) , … , Q n − 1 ∩ B n ( p ) leaves the index set { p } \{p\} { p } . Thus (24) and (19) imply
A B n ( p ) ( n − 1 ) ( p , p ) = s n ( z ; θ + p ⋅ ω ) . A_{B_n(p)}^{(n-1)}(p,p)=s_n(z;\theta+p\cdot\omega). A B n ( p ) ( n − 1 ) ( p , p ) = s n ( z ; θ + p ⋅ ω ) .
Here z ∈ J n ( θ + p ⋅ ω ) z\in J_n(\theta+p\cdot\omega) z ∈ J n ( θ + p ⋅ ω ) because p ∈ S n − 1 p\in S_{n-1} p ∈ S n − 1 implies ∣ z − E n − 1 ( θ + p ⋅ ω ) ∣ < δ n − 1 |z-E_{n-1}(\theta+p\cdot\omega)|<\delta_{n-1} ∣ z − E n − 1 ( θ + p ⋅ ω ) ∣ < δ n − 1 . Since dist ( p , B n ( p ) c ) ≥ 4 N n ≥ M k \operatorname{dist}(p,B_n(p)^c)\ge4N_n\ge M_k dist ( p , B n ( p ) c ) ≥ 4 N n ≥ M k for 0 ≤ k < n 0\le k<n 0 ≤ k < n , Corollary 5.3 gives
∣ A Ω ( n − 1 ) ( p , p ) − s n ( z ; θ + p ⋅ ω ) ∣ ≤ C 5 2 a 2 β − 1 N n τ − 2 r . (133) \left|A_{\Omega}^{(n-1)}(p,p)-s_n(z;\theta+p\cdot\omega)\right|\le\frac{C_5}{2}a^2\beta^{-1}N_n^{\tau-2r}. \tag*{(133)} A Ω ( n − 1 ) ( p , p ) − s n ( z ; θ + p ⋅ ω ) ≤ 2 C 5 a 2 β − 1 N n τ − 2 r . (133)
For p ∈ Q n = S n − 1 ∖ Res n p\in Q_n=S_{n-1}\setminus\operatorname{Res}_n p ∈ Q n = S n − 1 ∖ Res n , (104) gives
∣ s n ( z ; θ + p ⋅ ω ) ∣ ≥ ∣ z − E n ( θ + p ⋅ ω ) ∣ ≥ δ n . (134) |s_n(z;\theta+p\cdot\omega)|\ge|z-E_n(\theta+p\cdot\omega)|\ge\delta_n. \tag*{(134)} ∣ s n ( z ; θ + p ⋅ ω ) ∣ ≥ ∣ z − E n ( θ + p ⋅ ω ) ∣ ≥ δ n . (134)
The choice of C 5 C_5 C 5 in Subsection 4.2 and (95) give
C 5 a 2 β − 1 N n τ − 2 r δ n = C 5 a 2 β − 2 N n 2 τ − 2 r ≤ 1 16 . \frac{C_5a^2\beta^{-1}N_n^{\tau-2r}}{\delta_n}
=C_5a^2\beta^{-2}N_n^{2\tau-2r}\le\frac{1}{16}. δ n C 5 a 2 β − 1 N n τ − 2 r = C 5 a 2 β − 2 N n 2 τ − 2 r ≤ 16 1 .
Combining (133) and (134) proves (5.9). For n = 1 n=1 n = 1 , (133) proves the estimate (69) required in Subsection 3.6. The ranges 1 ≤ h ≤ n − 1 1\le h\le n-1 1 ≤ h ≤ n − 1 and 1 ≤ k < n 1\le k<n 1 ≤ k < n are then empty. The membership B 1 ( p ) ∈ Υ 0 B_1(p)\in\Upsilon_0 B 1 ( p ) ∈ Υ 0 is automatic, and Lemma 5.1 gives the comparison for B 1 ( p ) ⊂ Ω B_1(p)\subset\Omega B 1 ( p ) ⊂ Ω .
We now solve for v ∣ I n v|_{I_n} v ∣ I n , where I n = Q n ∩ Ω I_n=Q_n\cap\Omega I n = Q n ∩ Ω and P n = S n ∩ Ω P_n=S_n\cap\Omega P n = S n ∩ Ω as in (82) . The inclusion S n − 1 ⊂ Res n − 1 S_{n-1}\subset\operatorname{Res}_{n-1} S n − 1 ⊂ Res n − 1 and (66) give M n M_n M n -separation of S n − 1 S_{n-1} S n − 1 . Assertion (4) at i = n − 1 i=n-1 i = n − 1 gives
∣ A Ω ( n − 1 ) ( x , y ) ∣ ≤ 2 a ⟨ x − y ⟩ − r , x , y ∈ S n − 1 ∩ Ω , x ≠ y . (135) |A_{\Omega}^{(n-1)}(x,y)|\le2a\langle x-y\rangle^{-r},\qquad x,y\in S_{n-1}\cap\Omega,\ x\ne y. \tag*{(135)} ∣ A Ω ( n − 1 ) ( x , y ) ∣ ≤ 2 a ⟨ x − y ⟩ − r , x , y ∈ S n − 1 ∩ Ω , x = y . (135)
To verify the assumptions of Lemma A.2 for A ( n − 1 ) A^{(n-1)} A ( n − 1 ) , take X = S n − 1 ∩ Ω X=S_{n-1}\cap\Omega X = S n − 1 ∩ Ω , P = P n P=P_n P = P n , Q = I n Q=I_n Q = I n , M = M n M=M_n M = M n , a ∗ = 2 a a_*=2a a ∗ = 2 a and Δ = δ n / 2 \Delta=\delta_n/2 Δ = δ n /2 . The entry bounds are (5.9) and (135) , and
a ∗ Δ − 1 M − r = 4 a δ n − 1 M n − r ≤ ( 16 C 5 ) − 1 ≤ ( 2 C 3 ) − 1 a_*\Delta^{-1}M^{-r}=4a\delta_n^{-1}M_n^{-r}\le(16C_5)^{-1}\le(2C_3)^{-1} a ∗ Δ − 1 M − r = 4 a δ n − 1 M n − r ≤ ( 16 C 5 ) − 1 ≤ ( 2 C 3 ) − 1
by (94) . Lemma A.2 proves that ( A Ω ( n − 1 ) ) I n (A_{\Omega}^{(n-1)})_{I_n} ( A Ω ( n − 1 ) ) I n is invertible. Thus A Ω ( n − 1 ) v ∣ P n − 1 = 0 A_{\Omega}^{(n-1)}v|_{P_{n-1}}=0 A Ω ( n − 1 ) v ∣ P n − 1 = 0 is equivalent to
v ∣ I n = − ( A Ω ( n − 1 ) ) I n − 1 ( A Ω ( n − 1 ) ) I n , P n v ∣ P n , A Ω ( n ) v ∣ P n = 0. v|_{I_n}=-(A_{\Omega}^{(n-1)})_{I_n}^{-1}(A_{\Omega}^{(n-1)})_{I_n,P_n}v|_{P_n},\qquad A_{\Omega}^{(n)}v|_{P_n}=0. v ∣ I n = − ( A Ω ( n − 1 ) ) I n − 1 ( A Ω ( n − 1 ) ) I n , P n v ∣ P n , A Ω ( n ) v ∣ P n = 0.
It remains to prove ∣ A Ω ( n ) ( p , q ) ∣ ≤ 2 a ⟨ p − q ⟩ − r |A_{\Omega}^{(n)}(p,q)|\le2a\langle p-q\rangle^{-r} ∣ A Ω ( n ) ( p , q ) ∣ ≤ 2 a ⟨ p − q ⟩ − r for p , q ∈ S n ∩ Ω p,q\in S_n\cap\Omega p , q ∈ S n ∩ Ω , p ≠ q p\ne q p = q . Since Ω ∈ Υ n ⊂ Υ k \Omega\in\Upsilon_n\subset\Upsilon_k Ω ∈ Υ n ⊂ Υ k for 0 ≤ k ≤ n 0\le k\le n 0 ≤ k ≤ n , assertion (4) verifies the entry assumptions of Lemma A.2 at 0 ≤ k < n 0\le k<n 0 ≤ k < n . At k = n k=n k = n , (5.9) and (135) verify the diagonal and off-diagonal assumptions, respectively. With X = S k − 1 ∩ Ω X=S_{k-1}\cap\Omega X = S k − 1 ∩ Ω , P = S k ∩ Ω P=S_k\cap\Omega P = S k ∩ Ω , Q = Q k ∩ Ω Q=Q_k\cap\Omega Q = Q k ∩ Ω , M = M k M=M_k M = M k , a ∗ = 2 a a_*=2a a ∗ = 2 a and Δ = δ k / 2 \Delta=\delta_k/2 Δ = δ k /2 , the separation follows from (66) for k ≥ 1 k\ge1 k ≥ 1 and from M 0 = 1 M_0=1 M 0 = 1 for k = 0 k=0 k = 0 . Condition (94) verifies the smallness assumption. Summing (A.9) over 0 ≤ k ≤ n 0\le k\le n 0 ≤ k ≤ n , we obtain
∣ A Ω ( n ) ( p , q ) − A Ω ( − 1 ) ( p , q ) ∣ ≤ ∑ k = 0 n ∣ A Ω ( k ) ( p , q ) − A Ω ( k − 1 ) ( p , q ) ∣ ≤ C 5 a 2 ∑ k = 0 n δ k − 1 M k − r ⟨ p − q ⟩ − r ≤ a 64 ⟨ p − q ⟩ − r . (136) \begin{aligned}
\left|A_{\Omega}^{(n)}(p,q)-A_{\Omega}^{(-1)}(p,q)\right|
\le\sum_{k=0}^{n}\left|A_{\Omega}^{(k)}(p,q)-A_{\Omega}^{(k-1)}(p,q)\right| \\
&\le C_5a^2\sum_{k=0}^{n}\delta_k^{-1}M_k^{-r}\langle p-q\rangle^{-r}
\le\frac{a}{64}\langle p-q\rangle^{-r}. \tag*{(136)}
\end{aligned} A Ω ( n ) ( p , q ) − A Ω ( − 1 ) ( p , q ) ≤ k = 0 ∑ n A Ω ( k ) ( p , q ) − A Ω ( k − 1 ) ( p , q ) ≤ C 5 a 2 k = 0 ∑ n δ k − 1 M k − r ⟨ p − q ⟩ − r ≤ 64 a ⟨ p − q ⟩ − r . (136)
Combining (35) and (136) proves (5.10) and extends assertion (4) to i = n i=n i = n .
The next finite region and local eigenvalue
We first construct B n + 1 B_{n+1} B n + 1 satisfying (100) –(103) , which provide the inclusions and distances needed to compare the Schur complements. For a finite B ⊂ Z d B \subset\mathbb{Z}^d B ⊂ Z d , a set X ⊂ Z d X \subset\mathbb{Z}^d X ⊂ Z d and R ≥ 1 R \ge1 R ≥ 1 , define
B ~ = B ∪ ⋃ p ∈ X ( p + Λ R ) ∩ B ≠ ∅ ( p + Λ R ) . (137) \widetilde{B}=B\cup\bigcup_{\substack{p\in X\\(p+\Lambda_R)\cap B\ne\varnothing}}(p+\Lambda_R). \tag*{(137)} B = B ∪ p ∈ X ( p + Λ R ) ∩ B = ∅ ⋃ ( p + Λ R ) . (137)
Start with B = Λ 4 N n + 1 B=\Lambda_{4N_{n+1}} B = Λ 4 N n + 1 and successively apply
B ⟼ B ~ , X = Z h , R = 9 N h + 1 , h = n − 1 , n − 2 , … , 0. (138) B\longmapsto\widetilde{B},\qquad X=Z_h,\qquad R=9N_{h+1},\qquad h=n-1,n-2,\ldots,0. \tag*{(138)} B ⟼ B , X = Z h , R = 9 N h + 1 , h = n − 1 , n − 2 , … , 0. (138)
Denote the final set by B n + 1 B_{n+1} B n + 1 . Since B h + 1 ⊂ Λ 8 N h + 1 B_{h+1}\subset\Lambda_{8N_{h+1}} B h + 1 ⊂ Λ 8 N h + 1 , the ball p + Λ 9 N h + 1 p+\Lambda_{9N_{h+1}} p + Λ 9 N h + 1 contains B h + 1 ( p ) + Λ N h + 1 B_{h+1}(p)+\Lambda_{N_{h+1}} B h + 1 ( p ) + Λ N h + 1 , as required in (101) . The sets Z 0 , … , Z n − 1 Z_0,\ldots,Z_{n-1} Z 0 , … , Z n − 1 and numbers N 1 , … , N n + 1 N_1,\ldots,N_{n+1} N 1 , … , N n + 1 are fixed before t t t , θ \theta θ , z z z are chosen.
Proposition 5.5. The set B n + 1 B_{n+1} B n + 1 defined by (138) is finite and extends assertion (1) of Theorem 4.1 from step n n n to step n + 1 n+1 n + 1 .
The proof of Proposition 5.5 is given in Appendix B.§ 1 .
We now work on B n + 1 B_{n+1} B n + 1 and retain the component v ( o ) v(o) v ( o ) . Define J n + 1 J_{n+1} J n + 1 and s n + 1 s_{n+1} s n + 1 by (90) and (88) with m = n + 1 m=n+1 m = n + 1 , provided that A B n + 1 ∘ ( z ; θ ) A_{B^\circ_{n+1}}(z;\theta) A B n + 1 ∘ ( z ; θ ) is invertible.
Lemma 5.6. For every θ ∈ Θ \theta\in\Theta θ ∈ Θ and z ∈ J n + 1 ( θ ) z\in J_{n+1}(\theta) z ∈ J n + 1 ( θ ) , A B n + 1 ∘ ( z ; θ ) A_{B^\circ_{n+1}}(z;\theta) A B n + 1 ∘ ( z ; θ ) is invertible. The function s n + 1 ( ⋅ ; θ ) s_{n+1}(\cdot;\theta) s n + 1 ( ⋅ ; θ ) has exactly one zero E n + 1 ( θ ) E_{n+1}(\theta) E n + 1 ( θ ) in J n + 1 ( θ ) J_{n+1}(\theta) J n + 1 ( θ ) and
∣ E n + 1 ( θ ) − E n ( θ ) ∣ ≤ C 5 a 2 β − 1 N n τ − 2 r ≤ δ n / 16. (139) \left|E_{n+1}(\theta)-E_n(\theta)\right|\le C_5a^2\beta^{-1}N_n^{\tau-2r}\le\delta_n/16. \tag*{(139)} ∣ E n + 1 ( θ ) − E n ( θ ) ∣ ≤ C 5 a 2 β − 1 N n τ − 2 r ≤ δ n /16. (139) The eigenvalue E n + 1 ( θ ) E_{n+1}(\theta) E n + 1 ( θ ) is simple and satisfies (105) and (104) with m = n + 1 m=n+1 m = n + 1 . For the homotopy (45) , the invertibility, uniqueness and estimates in the lemma hold for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , and E n + 1 ( θ ; t ) = λ ( B n + 1 ; θ ; t ) E_{n+1}(\theta;t)=\lambda(B_{n+1};\theta;t) E n + 1 ( θ ; t ) = λ ( B n + 1 ; θ ; t ) .
Proof. We first prove invertibility on B n + 1 ∘ B^\circ_{n+1} B n + 1 ∘ . We then estimate s n + 1 ( E n ( θ ) ; θ ) s_{n+1}(E_n(\theta);\theta) s n + 1 ( E n ( θ ) ; θ ) by comparing B n B_n B n with B n + 1 B_{n+1} B n + 1 and eliminating Q n ∩ B n + 1 Q_n\cap B_{n+1} Q n ∩ B n + 1 . Finally, Proposition 3.4 identifies the new local eigenvalue for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] .
Let z ∈ J n + 1 ( θ ) z\in J_{n+1}(\theta) z ∈ J n + 1 ( θ ) . By (107) , for 0 ≤ h < n 0\le h<n 0 ≤ h < n ,
∣ z − E h ( θ ) ∣ ≤ 4 δ n + δ h / 8 < δ h . (140) \left|z-E_h(\theta)\right|\le4\delta_n+\delta_h/8<\delta_h. \tag*{(140)} ∣ z − E h ( θ ) ∣ ≤ 4 δ n + δ h /8 < δ h . (140)
The strict inequality follows from δ n ≤ 16 − τ δ h \delta_n\le16^{-\tau}\delta_h δ n ≤ 1 6 − τ δ h and τ > d ≥ 1 \tau>d\ge1 τ > d ≥ 1 . Thus o ∈ S n − 1 o\in S_{n-1} o ∈ S n − 1 throughout J n + 1 ( θ ) J_{n+1}(\theta) J n + 1 ( θ ) . For p ∈ Res h p\in\operatorname{Res}_h p ∈ Res h with 0 ≤ h < n 0\le h<n 0 ≤ h < n , (66) gives
L ∥ p ⋅ ω ∥ T ≤ ∣ E h ( θ + p ⋅ ω ) − E h ( θ ) ∣ < δ h + 4 δ n + δ h / 8 < 8 δ h . L\lVert p\cdot\omega\rVert_{\mathbb{T}}\le\left|E_h(\theta+p\cdot\omega)-E_h(\theta)\right|<\delta_h+4\delta_n+\delta_h/8<8\delta_h. L ∥ p ⋅ ω ∥ T ≤ ∣ E h ( θ + p ⋅ ω ) − E h ( θ ) ∣ < δ h + 4 δ n + δ h /8 < 8 δ h .
For p ∈ Res n p\in\operatorname{Res}_n p ∈ Res n , (66) and ∣ z − E n ( θ ) ∣ ≤ 4 δ n \left|z-E_n(\theta)\right|\le4\delta_n ∣ z − E n ( θ ) ∣ ≤ 4 δ n give
L ∥ p ⋅ ω ∥ T ≤ ∣ E n ( θ + p ⋅ ω ) − E n ( θ ) ∣ < 5 δ n < 8 δ n . L\lVert p\cdot\omega\rVert_{\mathbb{T}}\le\left|E_n(\theta+p\cdot\omega)-E_n(\theta)\right|<5\delta_n<8\delta_n. L ∥ p ⋅ ω ∥ T ≤ ∣ E n ( θ + p ⋅ ω ) − E n ( θ ) ∣ < 5 δ n < 8 δ n .
Hence Res h ⊂ Z h \operatorname{Res}_h\subset Z_h Res h ⊂ Z h for 0 ≤ h ≤ n 0\le h\le n 0 ≤ h ≤ n and z ∈ J n + 1 ( θ ) z\in J_{n+1}(\theta) z ∈ J n + 1 ( θ ) . The geometric inclusion (101) and the exclusion (103) imply, respectively,
B n + 1 ∈ Υ n , B n + 1 ∘ ∈ Υ n . (141) B_{n+1}\in\Upsilon_n,\qquad B^\circ_{n+1}\in\Upsilon_n. \tag*{(141)} B n + 1 ∈ Υ n , B n + 1 ∘ ∈ Υ n . (141)
Indeed, for p ∈ B n + 1 ∘ ∩ S h − 1 p\in B^\circ_{n+1}\cap S_{h-1} p ∈ B n + 1 ∘ ∩ S h − 1 and 1 ≤ h ≤ n 1\le h\le n 1 ≤ h ≤ n , we have p ∈ Z h − 1 ∘ p\in Z^\circ_{h-1} p ∈ Z h − 1 ∘ , so B h ( p ) ⊂ B n + 1 B_h(p)\subset B_{n+1} B h ( p ) ⊂ B n + 1 and o ∉ B h ( p ) o\notin B_h(p) o ∈ / B h ( p ) . Also, (102) gives B n + 1 ∘ ∩ S n = ∅ B^\circ_{n+1}\cap S_n=\varnothing B n + 1 ∘ ∩ S n = ∅ . Therefore
B n + 1 ∘ = ⋃ h = 0 n ( Q h ∩ B n + 1 ∘ ) . B^\circ_{n+1}=\bigcup_{h=0}^{n}(Q_h\cap B^\circ_{n+1}). B n + 1 ∘ = h = 0 ⋃ n ( Q h ∩ B n + 1 ∘ ) .
By assertion (4) and Lemma 5.4, ( A B n + 1 ∘ ( h − 1 ) ) Q h ∩ B n + 1 ∘ \left(A_{B_{n+1}^{\circ}}^{(h-1)}\right)_{Q_h\cap B_{n+1}^{\circ}} ( A B n + 1 ∘ ( h − 1 ) ) Q h ∩ B n + 1 ∘ is invertible for 0 ≤ h ≤ n 0\le h\le n 0 ≤ h ≤ n . Since B n + 1 ∘ = ⋃ h = 0 n ( Q h ∩ B n + 1 ∘ ) B_{n+1}^{\circ}=\bigcup_{h=0}^{n}(Q_h\cap B_{n+1}^{\circ}) B n + 1 ∘ = ⋃ h = 0 n ( Q h ∩ B n + 1 ∘ ) , back substitution for h = n , n − 1 , … , 0 h=n,n-1,\ldots,0 h = n , n − 1 , … , 0 in A B n + 1 ∘ ( z ; θ ) v = 0 A_{B_{n+1}^{\circ}}(z;\theta)v=0 A B n + 1 ∘ ( z ; θ ) v = 0 gives v = 0 v=0 v = 0 . Thus A B n + 1 ∘ ( z ; θ ) A_{B_{n+1}^{\circ}}(z;\theta) A B n + 1 ∘ ( z ; θ ) is invertible for every z ∈ J n + 1 ( θ ) z\in J_{n+1}(\theta) z ∈ J n + 1 ( θ ) . The equation on B n + 1 B_{n+1} B n + 1 is equivalent to
v ∣ B n + 1 ∘ = − A B n + 1 ∘ ( z ; θ ) − 1 c n + 1 v ( o ) , s n + 1 ( z ; θ ) v ( o ) = 0. v\big|_{B_{n+1}^{\circ}}=-A_{B_{n+1}^{\circ}}(z;\theta)^{-1}c_{n+1}v(o),\qquad s_{n+1}(z;\theta)v(o)=0. v B n + 1 ∘ = − A B n + 1 ∘ ( z ; θ ) − 1 c n + 1 v ( o ) , s n + 1 ( z ; θ ) v ( o ) = 0.
By Lemma 3.5, s n + 1 ( z 2 ; θ ) − s n + 1 ( z 1 ; θ ) ≥ z 2 − z 1 s_{n+1}(z_2;\theta)-s_{n+1}(z_1;\theta)\ge z_2-z_1 s n + 1 ( z 2 ; θ ) − s n + 1 ( z 1 ; θ ) ≥ z 2 − z 1 for z 1 < z 2 z_1<z_2 z 1 < z 2 in J n + 1 ( θ ) J_{n+1}(\theta) J n + 1 ( θ ) . When o ∉ S n o\notin S_n o ∈ / S n , the definition (88) still retains v ( o ) v(o) v ( o ) : the elimination at step n n n uses Q n ∩ B n + 1 ∘ Q_n\cap B_{n+1}^{\circ} Q n ∩ B n + 1 ∘ and leaves the index set { o } \{o\} { o } .
Take z = E n ( θ ) z=E_n(\theta) z = E n ( θ ) and let M = A B n + 1 ( n − 1 ) ( E n ( θ ) ; θ ) \mathcal{M}=A_{B_{n+1}}^{(n-1)}(E_n(\theta);\theta) M = A B n + 1 ( n − 1 ) ( E n ( θ ) ; θ ) . By (140) , o ∈ S n − 1 o\in S_{n-1} o ∈ S n − 1 , and (141) gives B n + 1 ∈ Υ n B_{n+1}\in\Upsilon_n B n + 1 ∈ Υ n . Since s n ( E n ( θ ) ; θ ) = 0 s_n(E_n(\theta);\theta)=0 s n ( E n ( θ ) ; θ ) = 0 by assertion (2), (133) with p = o p=o p = o and Ω = B n + 1 \Omega=B_{n+1} Ω = B n + 1 yields
∣ M ( o , o ) − s n ( E n ( θ ) ; θ ) ∣ ≤ C 5 2 a 2 β − 1 N n τ − 2 r . (142) \left|\mathcal{M}(o,o)-s_n(E_n(\theta);\theta)\right|\le\frac{C_5}{2}a^{2}\beta^{-1}N_n^{\tau-2r}. \tag*{(142)} ∣ M ( o , o ) − s n ( E n ( θ ) ; θ ) ∣ ≤ 2 C 5 a 2 β − 1 N n τ − 2 r . (142)
Now o ∈ S n o\in S_n o ∈ S n , since z = E n ( θ ) z=E_n(\theta) z = E n ( θ ) . Also B n + 1 ∩ S n = { o } B_{n+1}\cap S_n=\{o\} B n + 1 ∩ S n = { o } by (102) . Let I n = Q n ∩ B n + 1 I_n=Q_n\cap B_{n+1} I n = Q n ∩ B n + 1 , so o ∉ I n o\notin I_n o ∈ / I n . Solving for v ∣ I n v|_{I_n} v ∣ I n gives
s n + 1 ( E n ( θ ) ; θ ) = [ M ( o , o ) − s n ( E n ( θ ) ; θ ) ] − M o , I n M I n − 1 M I n , o . (143) s_{n+1}(E_n(\theta);\theta)=\left[\mathcal{M}(o,o)-s_n(E_n(\theta);\theta)\right]-\mathcal{M}_{o,I_n}\mathcal{M}_{I_n}^{-1}\mathcal{M}_{I_n,o}. \tag*{(143)} s n + 1 ( E n ( θ ) ; θ ) = [ M ( o , o ) − s n ( E n ( θ ) ; θ ) ] − M o , I n M I n − 1 M I n , o . (143)
The first term in (143) is the change after replacing B n B_n B n by B n + 1 B_{n+1} B n + 1 , bounded in (142) . The second term comes from eliminating I n I_n I n in B n + 1 B_{n+1} B n + 1 . Lemma 5.4, (135) and (94) verify the assumptions of Lemma A.2 for M \mathcal{M} M with X = { o } ⊔ I n X=\{o\}\sqcup I_n X = { o } ⊔ I n , P = { o } P=\{o\} P = { o } , Q = I n Q=I_n Q = I n , M = M n = N n − 1 \mathcal{M}=M_n=N_{n-1} M = M n = N n − 1 , a ∗ = 2 a a_*=2a a ∗ = 2 a and Δ = δ n / 2 \Delta=\delta_n/2 Δ = δ n /2 ; the set { o } ⊔ I n ⊂ S n − 1 \{o\}\sqcup I_n\subset S_{n-1} { o } ⊔ I n ⊂ S n − 1 is M n M_n M n -separated. Equation (A.10) therefore implies
∣ M o , I n M I n − 1 M I n , o ∣ ≤ 16 C 3 a n 2 δ n − 1 M n − 2 r = 16 2 r + 1 C 3 a 2 β − 1 N n τ − 2 r ≤ C 5 2 a 2 β − 1 N n τ − 2 r . (144) \left|\mathcal{M}_{o,I_n}\mathcal{M}_{I_n}^{-1}\mathcal{M}_{I_n,o}\right|\le16C_3a_n^2\delta_n^{-1}M_n^{-2r}=16^{2r+1}C_3a^{2}\beta^{-1}N_n^{\tau-2r}\le\frac{C_5}{2}a^{2}\beta^{-1}N_n^{\tau-2r}. \tag*{(144)} M o , I n M I n − 1 M I n , o ≤ 16 C 3 a n 2 δ n − 1 M n − 2 r = 1 6 2 r + 1 C 3 a 2 β − 1 N n τ − 2 r ≤ 2 C 5 a 2 β − 1 N n τ − 2 r . (144)
Substituting (142) and (144) into (143) , and using (91) , we obtain
∣ s n + 1 ( E n ( θ ) ; θ ) ∣ ≤ C 5 a 2 β − 1 N n τ − 2 r . \left|s_{n+1}(E_n(\theta);\theta)\right|\le C_5a^{2}\beta^{-1}N_n^{\tau-2r}. ∣ s n + 1 ( E n ( θ ) ; θ ) ∣ ≤ C 5 a 2 β − 1 N n τ − 2 r .
By (95) , the right-hand side is at most δ n / 16 \delta_n/16 δ n /16 . By Lemma 3.5,
s n + 1 ( E n ( θ ) − 4 δ n ; θ ) ≤ − 63 δ n 16 < 0 , s n + 1 ( E n ( θ ) + 4 δ n ; θ ) ≥ 63 δ n 16 > 0. s_{n+1}(E_n(\theta)-4\delta_n;\theta)\le-\frac{63\delta_n}{16}<0,\qquad s_{n+1}(E_n(\theta)+4\delta_n;\theta)\ge\frac{63\delta_n}{16}>0. s n + 1 ( E n ( θ ) − 4 δ n ; θ ) ≤ − 16 63 δ n < 0 , s n + 1 ( E n ( θ ) + 4 δ n ; θ ) ≥ 16 63 δ n > 0.
The construction (138) is independent of t t t , and replacing ε \varepsilon ε by t ε t\varepsilon tε preserves the bounds from (140) through (144) for t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . Restoring t t t in the notation, (89) implies
E n ( θ ; t ) ± 4 δ n ∉ Spec H B n + 1 ( θ ; t ) , t ∈ [ 0 , 1 ] . E_n(\theta;t)\pm4\delta_n\notin\operatorname{Spec}\mathcal{H}_{B_{n+1}}(\theta;t),\qquad t\in[0,1]. E n ( θ ; t ) ± 4 δ n ∈ / Spec H B n + 1 ( θ ; t ) , t ∈ [ 0 , 1 ] .
At t = 0 t=0 t = 0 , E n ( θ ; 0 ) = f ( θ ) E_n(\theta;0)=f(\theta) E n ( θ ; 0 ) = f ( θ ) by (114) . Since B n + 1 ∩ Z n = { o } B_{n+1}\cap Z_n=\{o\} B n + 1 ∩ Z n = { o } ,
∣ f ( θ + p ⋅ ω ) − f ( θ ) ∣ ≥ L ∥ p ⋅ ω ∥ T > 8 δ n , p ∈ B n + 1 ∘ . \left|f(\theta+p\cdot\omega)-f(\theta)\right|\ge L\lVert p\cdot\omega\rVert_{\mathbb{T}}>8\delta_n,\qquad p\in B_{n+1}^{\circ}. ∣ f ( θ + p ⋅ ω ) − f ( θ ) ∣ ≥ L ∥ p ⋅ ω ∥ T > 8 δ n , p ∈ B n + 1 ∘ .
By Proposition 3.4 with F = B n + 1 F=B_{n+1} F = B n + 1 , h ( θ ; t ) = E n ( θ ; t ) h(\theta;t)=E_n(\theta;t) h ( θ ; t ) = E n ( θ ; t ) and η = 4 δ n \eta=4\delta_n η = 4 δ n , the function E n + 1 ( θ ; t ) = λ ( B n + 1 ; θ ; t ) E_{n+1}(\theta;t)=\lambda(B_{n+1};\theta;t) E n + 1 ( θ ; t ) = λ ( B n + 1 ; θ ; t ) is continuous in t t t and is the unique eigenvalue in J n + 1 ( θ ; t ) J_{n+1}(\theta;t) J n + 1 ( θ ; t ) , counted with multiplicity. Equation (89) identifies E n + 1 ( θ ; t ) E_{n+1}(\theta;t) E n + 1 ( θ ; t ) with the unique zero of s n + 1 ( ⋅ ; θ ; t ) s_{n+1}(\cdot;\theta;t) s n + 1 ( ⋅ ; θ ; t ) in J n + 1 ( θ ; t ) J_{n+1}(\theta;t) J n + 1 ( θ ; t ) .
Fixing t t t once more, Lemma 3.5 gives
∣ E n + 1 ( θ ) − E n ( θ ) ∣ ≤ ∣ s n + 1 ( E n ( θ ) ; θ ) ∣ ≤ C 5 a 2 β − 1 N n τ − 2 r ≤ δ n / 16 , \left|E_{n+1}(\theta)-E_n(\theta)\right|\le\left|s_{n+1}(E_n(\theta);\theta)\right|\le C_5a^{2}\beta^{-1}N_n^{\tau-2r}\le\delta_n/16, ∣ E n + 1 ( θ ) − E n ( θ ) ∣ ≤ ∣ s n + 1 ( E n ( θ ) ; θ ) ∣ ≤ C 5 a 2 β − 1 N n τ − 2 r ≤ δ n /16 ,
which is (139) . Lemma 3.5 also proves (104) at scale n + 1 n+1 n + 1 . Since Spec H B n + 1 ( θ ) ∩ J n + 1 ( θ ) = { E n + 1 ( θ ) } \operatorname{Spec}\mathcal{H}_{B_{n+1}}(\theta)\cap J_{n+1}(\theta)=\{E_{n+1}(\theta)\} Spec H B n + 1 ( θ ) ∩ J n + 1 ( θ ) = { E n + 1 ( θ )} , we have
dist ( E n + 1 ( θ ) , Spec H B n + 1 ( θ ) ∖ { E n + 1 ( θ ) } ) ≥ 4 δ n − ∣ E n + 1 ( θ ) − E n ( θ ) ∣ ≥ 3 δ n . \operatorname{dist}\left(E_{n+1}(\theta),\operatorname{Spec}\mathcal{H}_{B_{n+1}}(\theta)\setminus\{E_{n+1}(\theta)\}\right)\ge4\delta_n-\left|E_{n+1}(\theta)-E_n(\theta)\right|\ge3\delta_n. dist ( E n + 1 ( θ ) , Spec H B n + 1 ( θ ) ∖ { E n + 1 ( θ )} ) ≥ 4 δ n − ∣ E n + 1 ( θ ) − E n ( θ ) ∣ ≥ 3 δ n .
This proves (105) for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] .
Eigenvector estimates and increments
At z = E j ( θ ; t ) z = E_j(\theta;t) z = E j ( θ ; t ) , we recover the coordinates on B j o B_j^o B j o from the value v j , θ ( t ) ( o ) = 1 v_{j,\theta(t)}(o) = 1 v j , θ ( t ) ( o ) = 1 . We then compare the zero-extended eigenvectors at scales j j j and j + 1 j+1 j + 1 .
Lemma 5.7. Let j ≥ 1 j \ge1 j ≥ 1 and suppose that assertions (1)–(4) of Theorem 4.1 hold through step j j j at a fixed t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] . Then there is a unique eigenvector v j , θ ( t ) v_{j,\theta(t)} v j , θ ( t ) of H B j ( θ ; t ) \mathcal{H}_{B_j}(\theta;t) H B j ( θ ; t ) corresponding to E j ( θ ; t ) E_j(\theta;t) E j ( θ ; t ) and satisfying v j , θ ( t ) ( o ) = 1 v_{j,\theta(t)}(o) = 1 v j , θ ( t ) ( o ) = 1 . Extended by zero outside B j B_j B j , the vector satisfies (112) , uniformly in θ \theta θ . If assertions (1)–(4) hold for every t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] , the bound is uniform in t t t as well.
Proof. Fix θ \theta θ and t t t , suppress t t t , and put z = E j ( θ ) z = E_j(\theta) z = E j ( θ ) . For 0 ≤ h < j 0 \le h < j 0 ≤ h < j , (107) gives o ∈ S h o \in S_h o ∈ S h . For p ∈ S h p \in S_h p ∈ S h , (66) gives
L ∥ p ⋅ ω ∥ T ≤ ∣ E h ( θ + p ⋅ ω ) − E h ( θ ) ∣ < δ h + δ h / 8 < 8 δ h . L\lVert p \cdot\omega\rVert_{\mathsf{T}} \le\lvert E_h(\theta+ p \cdot\omega) - E_h(\theta)\rvert< \delta_h + \delta_h/8 < 8\delta_h. L ∥ p ⋅ ω ∥ T ≤ ∣ E h ( θ + p ⋅ ω ) − E h ( θ )∣ < δ h + δ h /8 < 8 δ h .
Thus S h ( E j ( θ ) ; θ ) ⊂ Z h S_h(E_j(\theta);\theta) \subset Z_h S h ( E j ( θ ) ; θ ) ⊂ Z h . By (101) , (102) and (86) –(87) ,
B j ∈ Υ j − 1 , B j ∩ S j − 1 = { o } . (145) B_j \in\Upsilon_{j-1}, \qquad B_j \cap S_{j-1} = \{o\}. \tag*{(145)} B j ∈ Υ j − 1 , B j ∩ S j − 1 = { o } . (145)
Use I i , P i I_i,P_i I i , P i from (82) with Ω = B j \Omega= B_j Ω = B j . Then o ∈ P i o \in P_i o ∈ P i for − 1 ≤ i < j -1 \le i < j − 1 ≤ i < j and P j − 1 = { o } P_{j-1} = \{o\} P j − 1 = { o } . The final equation is s j ( E j ( θ ) ; θ ) v ( o ) = 0 s_j(E_j(\theta);\theta)v(o) = 0 s j ( E j ( θ ) ; θ ) v ( o ) = 0 , so set v ( j − 1 ) = e o v^{(j-1)} = e_o v ( j − 1 ) = e o . Assertion (4) permits the following back substitution, for i = j − 1 , … , 0 i = j-1,\ldots,0 i = j − 1 , … , 0 :
v ( i − 1 ) = W B j ( i ) v ( i ) , W B j ( i ) = ( I P i − ( A B j ( i − 1 ) ) I i − 1 ( A B j ( i − 1 ) ) I i , P i ) : C P i ⟶ C P i − 1 . (146) v^{(i-1)} = \mathcal{W}_{B_j}^{(i)}v^{(i)}, \qquad\mathcal{W}_{B_j}^{(i)} = \begin{pmatrix} \mathcal{I}_{P_i} \\ -(\mathcal{A}_{B_j}^{(i-1)})^{-1}_{I_i}(\mathcal{A}_{B_j}^{(i-1)})_{I_i,P_i} \end{pmatrix} : \mathbb{C}^{P_i} \longrightarrow\mathbb{C}^{P_{i-1}}. \tag*{(146)} v ( i − 1 ) = W B j ( i ) v ( i ) , W B j ( i ) = ( I P i − ( A B j ( i − 1 ) ) I i − 1 ( A B j ( i − 1 ) ) I i , P i ) : C P i ⟶ C P i − 1 . (146)
The row blocks are ordered as P i − 1 = P i ⊔ I i P_{i-1} = P_i \sqcup I_i P i − 1 = P i ⊔ I i . Equations (84) and (85) give
A B j ( i ) v ( i ) = 0 ⟹ A B j ( i − 1 ) v ( i − 1 ) = 0. \mathcal{A}_{B_j}^{(i)}v^{(i)} = 0 \quad\Longrightarrow\quad\mathcal{A}_{B_j}^{(i-1)}v^{(i-1)} = 0. A B j ( i ) v ( i ) = 0 ⟹ A B j ( i − 1 ) v ( i − 1 ) = 0.
Since o ∈ P i o \in P_i o ∈ P i , the back substitution preserves v ( i ) ( o ) = 1 v^{(i)}(o) = 1 v ( i ) ( o ) = 1 . Consequently, the zero extension of W B j ( 0 ) ⋯ W B j ( j − 1 ) e o \mathcal{W}_{B_j}^{(0)}\cdots\mathcal{W}_{B_j}^{(j-1)}e_o W B j ( 0 ) ⋯ W B j ( j − 1 ) e o satisfies (111) . The invertibility of A B j ∘ ( E j ( θ ) ; θ ) \mathcal{A}_{B_j^\circ}(E_j(\theta);\theta) A B j ∘ ( E j ( θ ) ; θ ) gives uniqueness with this normalization.
For i ≥ 1 i \ge1 i ≥ 1 , (66) shows that P i − 1 P_{i-1} P i − 1 is M i M_i M i -separated; for i = 0 i = 0 i = 0 , use M 0 = 1 M_0 = 1 M 0 = 1 . The bounds (109) –(110) and (94) verify the assumptions of Lemma A.2 for A B j ( i − 1 ) \mathcal{A}_{B_j}^{(i-1)} A B j ( i − 1 ) with X = P i − 1 X = P_{i-1} X = P i − 1 , P = P i P = P_i P = P i , Q = I i Q = I_i Q = I i , M = M i M = M_i M = M i , a ∗ = 2 a a_* = 2a a ∗ = 2 a and Δ = δ i / 2 \Delta= \delta_i/2 Δ = δ i /2 . Since ρ < r − τ < r − d \rho< r-\tau< r-d ρ < r − τ < r − d , (A.12) gives
∥ W B j ( 0 ) ∥ ρ ≤ 1 + C 4 a β − 1 , (147) \lVert\mathcal{W}_{B_j}^{(0)}\rVert_\rho\le1+C_4a\beta^{-1}, \tag*{(147)} ∥ W B j ( 0 ) ∥ ρ ≤ 1 + C 4 a β − 1 , (147)
∥ W B j ( i ) ∥ ρ ≤ 1 + C 4 a δ i − 1 − 1 N i − 1 ρ − r = 1 + C 4 a β − 1 16 τ N i − 1 τ + ρ − r , 1 ≤ i < j . (148) \lVert\mathcal{W}_{B_j}^{(i)}\rVert_\rho\le1+C_4a\delta_{i-1}^{-1}N_{i-1}^{\rho-r}
= 1+C_4a\beta^{-1}16^\tau N_{i-1}^{\tau+\rho-r}, \qquad1 \le i < j. \tag*{(148)} ∥ W B j ( i ) ∥ ρ ≤ 1 + C 4 a δ i − 1 − 1 N i − 1 ρ − r = 1 + C 4 a β − 1 1 6 τ N i − 1 τ + ρ − r , 1 ≤ i < j . (148)
The equality v j , θ ( o ) = 1 v_{j,\theta}(o)=1 v j , θ ( o ) = 1 implies ∥ v j , θ − e o ∥ 1 , ρ = ∥ v j , θ ∥ 1 , ρ − 1 \lVert v_{j,\theta}-e_o\rVert_{1,\rho}=\lVert v_{j,\theta}\rVert_{1,\rho}-1 ∥ v j , θ − e o ∥ 1 , ρ = ∥ v j , θ ∥ 1 , ρ − 1 . Combining (147) , (148) and (27) , we obtain
∥ v j , θ − e o ∥ 1 , ρ ≤ ∏ i = 0 j − 1 ∥ W B j ( i ) ∥ ρ − 1 ≤ exp ( C 4 a β − 1 G ( ρ ) ) − 1 ≤ C 6 a . (149) \lVert v_{j,\theta}-e_o\rVert_{1,\rho}
\le\prod_{i=0}^{j-1}\lVert\mathcal{W}_{B_j}^{(i)}\rVert_\rho-1
\le\exp(C_4a\beta^{-1}G(\rho))-1
\le C_6a. \tag*{(149)} ∥ v j , θ − e o ∥ 1 , ρ ≤ i = 0 ∏ j − 1 ∥ W B j ( i ) ∥ ρ − 1 ≤ exp ( C 4 a β − 1 G ( ρ )) − 1 ≤ C 6 a . (149)
Here (93) and (96) give C 4 a β − 1 G ( ρ ) ≤ 1 / 8 C_4a\beta^{-1}G(\rho)\le1/8 C 4 a β − 1 G ( ρ ) ≤ 1/8 , so exp ( x ) − 1 ≤ 2 x \exp(x)-1\le2x exp ( x ) − 1 ≤ 2 x for 0 ≤ x ≤ 1 / 8 0\le x\le1/8 0 ≤ x ≤ 1/8 proves the last inequality. All bounds are independent of θ \theta θ and t t t . □
Proposition 5.8. Let j ≥ 1 j \ge1 j ≥ 1 and suppose that assertions (1)–(4) of Theorem 4.1 hold through step j + 1 j+1 j + 1 for every t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] , with constants independent of t t t . The eigenvectors constructed in Lemma 5.7 satisfy
sup θ ∈ Θ , t ∈ [ 0 , 1 ] ∥ v j + 1 , θ ( t ) − v j , θ ( t ) ∥ 1 ≤ 12 C 6 a N j − κ , \sup_{\theta\in\Theta,\ t\in[0,1]} \lVert v_{j+1,\theta}(t)-v_{j,\theta}(t)\rVert_1 \le12C_6aN_j^{-\kappa}, θ ∈ Θ , t ∈ [ 0 , 1 ] sup ∥ v j + 1 , θ ( t ) − v j , θ ( t ) ∥ 1 ≤ 12 C 6 a N j − κ , where C 6 C_6 C 6 and κ \kappa κ are fixed in (93) and (98) .
Proof. Fix θ ∈ Θ \theta\in\Theta θ ∈ Θ and t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] , and suppress t t t in v i , θ v_{i,\theta} v i , θ and E i E_i E i . We compare v j + 1 , θ v_{j+1,\theta} v j + 1 , θ with v j , θ v_{j,\theta} v j , θ by expanding v j + 1 , θ v_{j+1,\theta} v j + 1 , θ in the translates v j , θ + k ⋅ ω ( ⋅ − k ) v_{j,\theta+k\cdot\omega}(\cdot-k) v j , θ + k ⋅ ω ( ⋅ − k ) , k ∈ Z d k \in\mathbb{Z}^d k ∈ Z d . For i = j , j + 1 i=j,j+1 i = j , j + 1 and m , k ∈ Z d m,k \in\mathbb{Z}^d m , k ∈ Z d , define
V ( i ) ( m , k ) = v i , θ + k ⋅ ω ( m − k ) , D ( i ) = diag k ∈ Z d E i ( θ + k ⋅ ω ) , (150) \mathcal{V}^{(i)}(m,k)=v_{i,\theta+k\cdot\omega}(m-k), \qquad\mathcal{D}^{(i)}=\operatorname{diag}_{k\in\mathbb{Z}^d}E_i(\theta+k\cdot\omega), \tag*{(150)} V ( i ) ( m , k ) = v i , θ + k ⋅ ω ( m − k ) , D ( i ) = diag k ∈ Z d E i ( θ + k ⋅ ω ) , (150)
R ( i ) ( m , k ) = t ε 1 { m − k ∉ B i } ∑ ℓ ∈ Z d ϕ ( m − ℓ ) V ( i ) ( ℓ , k ) . (151) \mathcal{R}^{(i)}(m,k)=t\varepsilon\mathbf{1}_{\{m-k\notin B_i\}}\sum_{\ell\in\mathbb{Z}^d}\phi(m-\ell)\mathcal{V}^{(i)}(\ell,k). \tag*{(151)} R ( i ) ( m , k ) = tε 1 { m − k ∈ / B i } ℓ ∈ Z d ∑ ϕ ( m − ℓ ) V ( i ) ( ℓ , k ) . (151)
By (19) and (111) ,
H ( θ ; t ) V ( i ) = V ( i ) D ( i ) + R ( i ) (152) \mathcal{H}(\theta;t)\mathcal{V}^{(i)}=\mathcal{V}^{(i)}\mathcal{D}^{(i)}+\mathcal{R}^{(i)} \tag*{(152)} H ( θ ; t ) V ( i ) = V ( i ) D ( i ) + R ( i ) (152)
on finitely supported vectors. Lemma 5.7 , (26) and Lemma 2.1 give
∥ V ( i ) − I ∥ ρ ≤ C 6 a ≤ 1 / 4 , ∥ ( V ( i ) ) − 1 ∥ ρ ≤ ( 1 − C 6 a ) − 1 ≤ 4 / 3. (153) \lVert\mathcal{V}^{(i)}-\mathcal{I}\rVert_{\rho}\le C_6a\le1/4,\qquad\lVert(\mathcal{V}^{(i)})^{-1}\rVert_{\rho}\le(1-C_6a)^{-1}\le4/3. \tag*{(153)} ∥ V ( i ) − I ∥ ρ ≤ C 6 a ≤ 1/4 , ∥( V ( i ) ) − 1 ∥ ρ ≤ ( 1 − C 6 a ) − 1 ≤ 4/3. (153)
Set x = ( V ( j ) ) − 1 v j + 1 , θ x=(\mathcal{V}^{(j)})^{-1}v_{j+1,\theta} x = ( V ( j ) ) − 1 v j + 1 , θ . Since V ( j ) e o = v j , θ \mathcal{V}^{(j)}e_o=v_{j,\theta} V ( j ) e o = v j , θ , we have
v j + 1 , θ − v j , θ = V ( j ) ( x − e o ) . (154) v_{j+1,\theta}-v_{j,\theta}=\mathcal{V}^{(j)}(x-e_o). \tag*{(154)} v j + 1 , θ − v j , θ = V ( j ) ( x − e o ) . (154)
For a radius 1 ≤ R ≤ N j 1\le R\le N_j 1 ≤ R ≤ N j to be chosen below, eigenvalue separation will bound ∑ 0 < ∣ k ∣ ≤ R ∣ x ( k ) ∣ \sum_{0<|k|\le R}|x(k)| ∑ 0 < ∣ k ∣ ≤ R ∣ x ( k ) ∣ , while the weighted norm will bound ∑ ∣ k ∣ > R ∣ x ( k ) ∣ \sum_{|k|>R}|x(k)| ∑ ∣ k ∣ > R ∣ x ( k ) ∣ . We first estimate R ( i ) \mathcal{R}^{(i)} R ( i ) and verify that (152) holds for x x x . By translation covariance, it suffices to consider the column k = o k=o k = o . The inclusion Λ 4 N i ⊂ B i \Lambda_{4N_i}\subset B_i Λ 4 N i ⊂ B i implies
∑ m ∉ B i ⟨ m − ℓ ⟩ − r ≤ C 3 N i d − r , ∣ ℓ ∣ ≤ 2 N i . \sum_{m\notin B_i}\langle m-\ell\rangle^{-r}\le C_3N_i^{d-r},\qquad|\ell|\le2N_i. m ∈ / B i ∑ ⟨ m − ℓ ⟩ − r ≤ C 3 N i d − r , ∣ ℓ ∣ ≤ 2 N i .
For the remaining summation indices, (112) gives
∑ ∣ ℓ ∣ > 2 N i ∣ v i , θ ( ℓ ) ∣ ≤ C 6 a ( 1 + 2 N i ) − ρ . \sum_{|\ell|>2N_i}|v_{i,\theta}(\ell)|\le C_6a(1+2N_i)^{-\rho}. ∣ ℓ ∣ > 2 N i ∑ ∣ v i , θ ( ℓ ) ∣ ≤ C 6 a ( 1 + 2 N i ) − ρ .
Splitting the sum in (151) at ∣ ℓ ∣ = 2 N i |\ell|=2N_i ∣ ℓ ∣ = 2 N i and using ∑ m ⟨ m ⟩ − r ≤ C 3 \sum_m\langle m\rangle^{-r}\le C_3 ∑ m ⟨ m ⟩ − r ≤ C 3 and ρ < r − d \rho<r-d ρ < r − d , we obtain
∥ R ( i ) ∥ 0 ≤ C 3 a ( ( 1 + C 6 a ) N i d − r + C 6 a N i − ρ ) ≤ 2 C 3 a N i − ρ . (155) \lVert\mathcal{R}^{(i)}\rVert_0\le C_3a\bigl((1+C_6a)N_i^{d-r}+C_6aN_i^{-\rho}\bigr)\le2C_3aN_i^{-\rho}. \tag*{(155)} ∥ R ( i ) ∥ 0 ≤ C 3 a ( ( 1 + C 6 a ) N i d − r + C 6 a N i − ρ ) ≤ 2 C 3 a N i − ρ . (155)
To use (152) for x x x , set
F = diag k ∈ Z d f ( θ + k ⋅ ω ) (156) \mathcal{F}=\operatorname{diag}_{k\in\mathbb{Z}^d}f(\theta+k\cdot\omega) \tag*{(156)} F = diag k ∈ Z d f ( θ + k ⋅ ω ) (156)
on ℓ 1 \ell^1 ℓ 1 , with the natural domain. By (107) , D ( i ) − F \mathcal{D}^{(i)}-\mathcal{F} D ( i ) − F is bounded on ℓ 1 \ell^1 ℓ 1 , so Dom ( D ( i ) ) = Dom ( F ) \operatorname{Dom}(\mathcal{D}^{(i)})=\operatorname{Dom}(\mathcal{F}) Dom ( D ( i ) ) = Dom ( F ) . The identity (152) extends to Dom ( F ) \operatorname{Dom}(\mathcal{F}) Dom ( F ) , and
V ( i ) Dom ( F ) = Dom ( F ) , i = j , j + 1. (157) \mathcal{V}^{(i)}\operatorname{Dom}(\mathcal{F})=\operatorname{Dom}(\mathcal{F}),\qquad i=j,j+1. \tag*{(157)} V ( i ) Dom ( F ) = Dom ( F ) , i = j , j + 1. (157)
The domain assertion and the extension of (152) are proved in Appendix B.§ 2 . By (153) , (112) and (96) ,
∥ x ∥ 1 ≤ 1 + C 6 a 1 − C 6 a , ∥ x − e o ∥ 1 , ρ ≤ 2 C 6 a 1 − C 6 a ≤ 3 C 6 a . (158) \lVert x\rVert_1\le\frac{1+C_6a}{1-C_6a},\qquad\lVert x-e_o\rVert_{1,\rho}\le\frac{2C_6a}{1-C_6a}\le3C_6a. \tag*{(158)} ∥ x ∥ 1 ≤ 1 − C 6 a 1 + C 6 a , ∥ x − e o ∥ 1 , ρ ≤ 1 − C 6 a 2 C 6 a ≤ 3 C 6 a . (158)
The vector v j + 1 , θ v_{j+1,\theta} v j + 1 , θ has finite support; hence (157) implies x ∈ Dom ( F ) x\in\operatorname{Dom}(\mathcal{F}) x ∈ Dom ( F ) . Using (152) for i = j , j + 1 i=j,j+1 i = j , j + 1 , we obtain
( D ( j ) − E j + 1 ( θ ) ) x = ( V ( j ) ) − 1 ( R ( j + 1 ) e o − R ( j ) x ) . (159) (\mathcal{D}^{(j)}-E_{j+1}(\theta))x=(\mathcal{V}^{(j)})^{-1}(\mathcal{R}^{(j+1)}e_o-\mathcal{R}^{(j)}x). \tag*{(159)} ( D ( j ) − E j + 1 ( θ )) x = ( V ( j ) ) − 1 ( R ( j + 1 ) e o − R ( j ) x ) . (159)
Equations (153) , (155) and (158) , together with N j + 1 > N j N_{j+1}>N_j N j + 1 > N j , give
∥ ( V ( j ) ) − 1 ( R ( j + 1 ) e o − R ( j ) x ) ∥ 1 ≤ 8 C 3 a N j − ρ . \lVert(\mathcal{V}^{(j)})^{-1}(\mathcal{R}^{(j+1)}e_o-\mathcal{R}^{(j)}x)\rVert_1\le8C_3aN_j^{-\rho}. ∥( V ( j ) ) − 1 ( R ( j + 1 ) e o − R ( j ) x ) ∥ 1 ≤ 8 C 3 a N j − ρ .
For 1 ≤ R ≤ N j 1\le R\le N_j 1 ≤ R ≤ N j and 0 < ∣ k ∣ ≤ R 0<|k|\le R 0 < ∣ k ∣ ≤ R , (66) , (106) , (95) and (32) give
∣ E j ( θ + k ⋅ ω ) − E j + 1 ( θ ) ∣ ≥ L γ ⟨ k ⟩ − τ − δ j / 16 ≥ L γ 2 ( 1 + R ) − τ . |E_j(\theta+k\cdot\omega)-E_{j+1}(\theta)|\ge L\gamma\langle k\rangle^{-\tau}-\delta_j/16\ge\frac{L\gamma}{2}(1+R)^{-\tau}. ∣ E j ( θ + k ⋅ ω ) − E j + 1 ( θ ) ∣ ≥ L γ ⟨ k ⟩ − τ − δ j /16 ≥ 2 L γ ( 1 + R ) − τ .
Consequently, (159) implies
∑ 0 < ∣ k ∣ ≤ R ∣ x ( k ) ∣ ≤ 2 τ + 4 C 3 L γ a R τ N j − ρ ≤ C 6 a R τ N j − ρ . \sum_{0<|k|\le R} |x(k)| \le\frac{2^{\tau+4}C_3}{L\gamma}aR^\tau N_j^{-\rho} \le C_6aR^\tau N_j^{-\rho}. 0 < ∣ k ∣ ≤ R ∑ ∣ x ( k ) ∣ ≤ L γ 2 τ + 4 C 3 a R τ N j − ρ ≤ C 6 a R τ N j − ρ .
For ∣ k ∣ > R |k|>R ∣ k ∣ > R , (158) gives
∑ ∣ k ∣ > R ∣ x ( k ) ∣ ≤ 3 C 6 a R − ρ . \sum_{|k|>R}|x(k)|\le3C_6aR^{-\rho}. ∣ k ∣ > R ∑ ∣ x ( k ) ∣ ≤ 3 C 6 a R − ρ .
Adding the estimates for 0 < ∣ k ∣ ≤ R 0<|k|\le R 0 < ∣ k ∣ ≤ R and ∣ k ∣ > R |k|>R ∣ k ∣ > R , we obtain
∑ k ≠ o ∣ x ( k ) ∣ ≤ 3 C 6 a ( R τ N j − ρ + R − ρ ) . (160) \sum_{k\ne o}|x(k)|\le3C_6a\left(R^\tau N_j^{-\rho}+R^{-\rho}\right). \tag*{(160)} k = o ∑ ∣ x ( k ) ∣ ≤ 3 C 6 a ( R τ N j − ρ + R − ρ ) . (160)
Taking R = N j ρ / ( τ + ρ ) ∈ [ 1 , N j ] R=N_j^{\rho/(\tau+\rho)}\in[1,N_j] R = N j ρ / ( τ + ρ ) ∈ [ 1 , N j ] in (160) , we obtain ∑ k ≠ o ∣ x ( k ) ∣ ≤ 6 C 6 a N j − κ \sum_{k\ne o}|x(k)|\le6C_6aN_j^{-\kappa} ∑ k = o ∣ x ( k ) ∣ ≤ 6 C 6 a N j − κ .
Finally, v j + 1 , θ ( o ) = v j , θ ( o ) = 1 v_{j+1,\theta}(o)=v_{j,\theta}(o)=1 v j + 1 , θ ( o ) = v j , θ ( o ) = 1 and v j + 1 , θ = V ( j ) x v_{j+1,\theta}=\mathcal{V}^{(j)}x v j + 1 , θ = V ( j ) x imply
x ( o ) − 1 = − ∑ k ≠ o V ( j ) ( o , k ) x ( k ) , ∣ x ( o ) − 1 ∣ ≤ C 6 a ∑ k ≠ o ∣ x ( k ) ∣ . x(o)-1=-\sum_{k\ne o}\mathcal{V}^{(j)}(o,k)x(k),\qquad|x(o)-1|\le C_6a\sum_{k\ne o}|x(k)|. x ( o ) − 1 = − k = o ∑ V ( j ) ( o , k ) x ( k ) , ∣ x ( o ) − 1∣ ≤ C 6 a k = o ∑ ∣ x ( k ) ∣.
Using (154) , we conclude that
∥ v j + 1 , θ − v j , θ ∥ 1 ≤ ( 1 + C 6 a ) 2 ∑ k ≠ o ∣ x ( k ) ∣ ≤ 12 C 6 a N j − κ . \|v_{j+1,\theta}-v_{j,\theta}\|_1\le(1+C_6a)^2\sum_{k\ne o}|x(k)|\le12C_6aN_j^{-\kappa}. ∥ v j + 1 , θ − v j , θ ∥ 1 ≤ ( 1 + C 6 a ) 2 k = o ∑ ∣ x ( k ) ∣ ≤ 12 C 6 a N j − κ .
The estimates are uniform in θ \theta θ and t t t and apply to j = 1 j=1 j = 1 .
Completion of the induction
Proposition 5.9. Let n ≥ 1 n\ge1 n ≥ 1 . Suppose that B m B_m B m , E m E_m E m and v m , θ v_{m,\theta} v m , θ , 0 ≤ m ≤ n 0\le m\le n 0 ≤ m ≤ n , satisfy assertions (1)–(6) of Theorem 4.1 at step n n n , with the fixed constants. Then there exist B n + 1 B_{n+1} B n + 1 , E n + 1 E_{n+1} E n + 1 and v n + 1 , θ v_{n+1,\theta} v n + 1 , θ such that assertions (1)–(6) of Theorem 4.1 hold at step n + 1 n+1 n + 1 , with the constants fixed in Subsection 4.2 and without changing ( B m , E m , v m , θ ) (B_m,E_m,v_{m,\theta}) ( B m , E m , v m , θ ) for 0 ≤ m ≤ n 0\le m\le n 0 ≤ m ≤ n .
Proof. Proposition 5.5 proves assertion (1) for B n + 1 B_{n+1} B n + 1 . The proof of Lemma 5.4 applies for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] and proves assertion (4) at i = n i=n i = n . By Lemma 5.6 , assertion (2) and (106) hold at the new step for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] .
By (99) and Proposition 3.3 ,
E n + 1 ( θ 2 ; t ) − E n + 1 ( θ 1 ; t ) ≥ L ( θ 2 − θ 1 ) , θ 1 < θ 2 , θ 1 , θ 2 ∈ Θ ∩ [ 0 , 1 ) , t ∈ [ 0 , 1 ] . E_{n+1}(\theta_2;t)-E_{n+1}(\theta_1;t)\ge L(\theta_2-\theta_1),\qquad\theta_1<\theta_2,\quad\theta_1,\theta_2\in\Theta\cap[0,1),\quad t\in[0,1]. E n + 1 ( θ 2 ; t ) − E n + 1 ( θ 1 ; t ) ≥ L ( θ 2 − θ 1 ) , θ 1 < θ 2 , θ 1 , θ 2 ∈ Θ ∩ [ 0 , 1 ) , t ∈ [ 0 , 1 ] .
This proves (108) at the new step.
For 0 ≤ ℓ < m ≤ n + 1 0\le\ell<m\le n+1 0 ≤ ℓ < m ≤ n + 1 , summing (106) and using (95) and N ℓ τ − 2 r ≤ 1 N_\ell^{\tau-2r}\le1 N ℓ τ − 2 r ≤ 1 , we obtain
∥ E m ( ⋅ ; t ) − E ℓ ( ⋅ ; t ) ∥ ∞ ≤ C 5 a 2 β − 1 ∑ h = ℓ m − 1 N h τ − 2 r ≤ C 5 a 2 β − 1 N ℓ τ − 2 r 1 − 16 τ − 2 r ≤ δ ℓ / 16. \|E_m(\cdot;t)-E_\ell(\cdot;t)\|_\infty
\le C_5a^2\beta^{-1}\sum_{h=\ell}^{m-1}N_h^{\tau-2r}
\le\frac{C_5a^2\beta^{-1}N_\ell^{\tau-2r}}{1-16^{\tau-2r}}
\le\delta_\ell/16. ∥ E m ( ⋅ ; t ) − E ℓ ( ⋅ ; t ) ∥ ∞ ≤ C 5 a 2 β − 1 h = ℓ ∑ m − 1 N h τ − 2 r ≤ 1 − 1 6 τ − 2 r C 5 a 2 β − 1 N ℓ τ − 2 r ≤ δ ℓ /16.
This proves (107) and completes assertion (3). By Lemma 5.7 with j = n + 1 j=n+1 j = n + 1 , there is an eigenvector v n + 1 , θ ( t ) v_{n+1,\theta}(t) v n + 1 , θ ( t ) satisfying (112) . By Proposition 5.8 with j = n j=n j = n , v n + 1 , θ ( t ) v_{n+1,\theta}(t) v n + 1 , θ ( t ) also satisfies (113) , which proves assertion (5). The normalization is v n + 1 , θ ( t ) ( o ) = 1 v_{n+1,\theta}(t)(o)=1 v n + 1 , θ ( t ) ( o ) = 1 , and
v n + 1 , θ ( t ) ∣ B n + 1 ∘ = − A B n + 1 ∘ ( E n + 1 ( θ ; t ) ; θ ; t ) − 1 c n + 1 ( t ) . v_{n+1,\theta}(t)|_{B^\circ_{n+1}}=-\mathcal{A}_{B^\circ_{n+1}}(E_{n+1}(\theta;t);\theta;t)^{-1}c_{n+1}(t). v n + 1 , θ ( t ) ∣ B n + 1 ∘ = − A B n + 1 ∘ ( E n + 1 ( θ ; t ) ; θ ; t ) − 1 c n + 1 ( t ) .
Assertion (2) gives the invertibility of A B n + 1 ∘ ( E n + 1 ( θ ; t ) ; θ ; t ) \mathcal{A}_{B^\circ_{n+1}}(E_{n+1}(\theta;t);\theta;t) A B n + 1 ∘ ( E n + 1 ( θ ; t ) ; θ ; t ) for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . The identity E n + 1 ( θ ; t ) = λ ( B n + 1 ; θ ; t ) E_{n+1}(\theta;t)=\lambda(B_{n+1};\theta;t) E n + 1 ( θ ; t ) = λ ( B n + 1 ; θ ; t ) from Lemma 5.6 , together with (59) and (30) , gives E n + 1 ( θ ; 0 ) = f ( θ ) E_{n+1}(\theta;0)=f(\theta) E n + 1 ( θ ; 0 ) = f ( θ ) and continuity in t t t . The matrix A B n + 1 ∘ ( E n + 1 ( θ ; t ) ; θ ; t ) \mathcal{A}_{B^\circ_{n+1}}(E_{n+1}(\theta;t);\theta;t) A B n + 1 ∘ ( E n + 1 ( θ ; t ) ; θ ; t ) and c n + 1 ( t ) c_{n+1}(t) c n + 1 ( t ) are therefore continuous in t t t , and continuity of matrix inversion proves continuity of v n + 1 , θ ( t ) v_{n+1,\theta}(t) v n + 1 , θ ( t ) . Moreover, $c_{n+1}(0)=0 gives v n + 1 , θ ( 0 ) = e o v_{n+1,\theta}(0)=e_o v n + 1 , θ ( 0 ) = e o . The set B n + 1 B_{n+1} B n + 1 and the constants in Subsection 4.2 are independent of t t t , so assertion (6) holds at step n + 1 n+1 n + 1 .
Proof of Theorem 4.1 . The verification at the end of § 4 establishes all six assertions for n = 1 n=1 n = 1 . Proposition 5.9 extends ( B m , E m , v m , θ ) 0 ≤ m ≤ n (B_m,E_m,v_{m,\theta})_{0\le m\le n} ( B m , E m , v m , θ ) 0 ≤ m ≤ n from each n ≥ 1 n\ge1 n ≥ 1 to n + 1 n+1 n + 1 . Induction therefore constructs the eigenvalues E m ( ⋅ ; t ) E_m(\mathord\cdot;t) E m ( ⋅ ; t ) and corresponding eigenvectors v m , θ ( t ) v_{m,\theta}(t) v m , θ ( t ) for all m ≥ 0 m\ge0 m ≥ 0 and t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . The common choice (97) guarantees all the smallness assumptions, including (96) . □
Localization
We use the summable increments in Theorem 4.1 to construct eigenvectors of H ( θ ) \mathcal{H}(\theta) H ( θ ) . We then prove that the normalized translates of v θ + n ⋅ ω v_{\theta+n\cdot\omega} v θ + n ⋅ ω , n ∈ Z d n\in\mathbb{Z}^d n ∈ Z d , form an orthonormal basis and establish (11) . Throughout this section, use ρ \rho ρ and ε 0 \varepsilon_0 ε 0 fixed in Subsection 4.2, assume ∣ ε ∣ < ε 0 |\varepsilon|<\varepsilon_0 ∣ ε ∣ < ε 0 and take t = 1 t=1 t = 1 . We omit the parameter t = 1 t=1 t = 1 from the notation.
Limiting eigenvectors
By (106) ,
∑ j ≥ 1 ∥ E j − E j − 1 ∥ ∞ ≤ C 5 a 2 β τ − 1 ∑ j ≥ 1 N j τ − 2 r < ∞ . \sum_{j\ge1}\lVert E_j-E_{j-1}\rVert_\infty\le C_5a^2\beta^{\tau-1}\sum_{j\ge1}N_j^{\tau-2r}<\infty. j ≥ 1 ∑ ∥ E j − E j − 1 ∥ ∞ ≤ C 5 a 2 β τ − 1 j ≥ 1 ∑ N j τ − 2 r < ∞.
Since E 0 = f E_0=f E 0 = f , the functions E j − f E_j-f E j − f converge uniformly on Θ \Theta Θ . We define
E ∞ = f + lim j → ∞ ( E j − f ) . E_\infty=f+\lim_{j\to\infty}(E_j-f). E ∞ = f + j → ∞ lim ( E j − f ) .
The function E ∞ : Θ → R E_\infty:\Theta\to\mathbb{R} E ∞ : Θ → R is 1 1 1 -periodic. Summing (106) over j ≥ 1 j\ge1 j ≥ 1 , we obtain
∥ E ∞ − f ∥ ∞ ≤ C 5 a 2 β ( 1 − 16 τ − 2 r ) ≤ C 1 ∣ ε ∣ 2 . \lVert E_\infty-f\rVert_\infty\le\frac{C_5a^2}{\beta(1-16\tau^{-2r})}\le C_1|\varepsilon|^2. ∥ E ∞ − f ∥ ∞ ≤ β ( 1 − 16 τ − 2 r ) C 5 a 2 ≤ C 1 ∣ ε ∣ 2 .
Here C 1 C_1 C 1 depends only on d d d , r r r , τ \tau τ , γ \gamma γ , L L L , ϕ \phi ϕ and is independent of ρ \rho ρ , as required in Theorem 1.1 . By (108) ,
E ∞ ( θ 2 ) − E ∞ ( θ 1 ) = lim j → ∞ ( E j ( θ 2 ) − E j ( θ 1 ) ) ≥ L ( θ 2 − θ 1 ) , θ 1 , θ 2 ∈ Θ ∩ [ 0 , 1 ) , θ 1 < θ 2 . E_\infty(\theta_2)-E_\infty(\theta_1)=\lim_{j\to\infty}(E_j(\theta_2)-E_j(\theta_1))\ge L(\theta_2-\theta_1),\qquad\theta_1,\theta_2\in\Theta\cap[0,1),\quad\theta_1<\theta_2. E ∞ ( θ 2 ) − E ∞ ( θ 1 ) = j → ∞ lim ( E j ( θ 2 ) − E j ( θ 1 )) ≥ L ( θ 2 − θ 1 ) , θ 1 , θ 2 ∈ Θ ∩ [ 0 , 1 ) , θ 1 < θ 2 .
By (66) and (2) , we obtain
∣ E ∞ ( θ + n ⋅ ω ) − E ∞ ( θ + m ⋅ ω ) ∣ ≥ L ∥ ( n − m ) ⋅ ω ∥ T ≥ L γ ⟨ n − m ⟩ − τ , n ≠ m . (161) \lvert E_\infty(\theta+n\cdot\omega)-E_\infty(\theta+m\cdot\omega)\rvert\ge L\lVert(n-m)\cdot\omega\rVert_{\mathbb{T}}\ge L\gamma\langle n-m\rangle^{-\tau},\qquad n\ne m. \tag*{(161)} ∣ E ∞ ( θ + n ⋅ ω ) − E ∞ ( θ + m ⋅ ω )∣ ≥ L ∥( n − m ) ⋅ ω ∥ T ≥ L γ ⟨ n − m ⟩ − τ , n = m . (161)
We now use (112) and (113) to construct an eigenvector of H ( θ ) \mathcal{H}(\theta) H ( θ ) with eigenvalue E ∞ ( θ ) E_\infty(\theta) E ∞ ( θ ) .
Lemma 6.1. For every θ ∈ Θ \theta\in\Theta θ ∈ Θ , there is a vector v θ ∈ Dom ( H ( θ ) ) v_\theta\in\operatorname{Dom}(\mathcal{H}(\theta)) v θ ∈ Dom ( H ( θ )) such that
H ( θ ) v θ = E ∞ ( θ ) v θ , v θ ( o ) = 1 , ∑ m ≠ o ∣ v θ ( m ) ∣ ⟨ m ⟩ ρ ≤ C 6 a ≤ 1 4 . \mathcal{H}(\theta)v_\theta=E_\infty(\theta)v_\theta,\qquad v_\theta(o)=1,\qquad\sum_{m\ne o}\lvert v_\theta(m)\rvert\langle m\rangle^\rho\le C_6a\le\frac14. H ( θ ) v θ = E ∞ ( θ ) v θ , v θ ( o ) = 1 , m = o ∑ ∣ v θ ( m )∣ ⟨ m ⟩ ρ ≤ C 6 a ≤ 4 1 . Proof. Fix θ ∈ Θ \theta\in\Theta θ ∈ Θ and use the vectors in assertion (5) of Theorem 4.1 . Since N j = 16 j N_j=16^j N j = 1 6 j and κ > 0 \kappa>0 κ > 0 , (113) gives
∑ j ≥ 1 sup θ ∈ Θ ∥ v j + 1 , θ − v j , θ ∥ 1 < ∞ . \sum_{j\ge1}\sup_{\theta\in\Theta}\lVert v_{j+1,\theta}-v_{j,\theta}\rVert_1<\infty. j ≥ 1 ∑ θ ∈ Θ sup ∥ v j + 1 , θ − v j , θ ∥ 1 < ∞. Thus v j , θ v_{j,\theta} v j , θ converges in ℓ 1 \ell^1 ℓ 1 , uniformly in θ \theta θ , to a vector v θ v_\theta v θ with v θ ( o ) = 1 v_\theta(o)=1 v θ ( o ) = 1 , and
sup θ ∈ Θ ∥ v θ − v j , θ ∥ 1 ≤ 12 C 6 1 − 16 − κ a N j − κ . \sup_{\theta\in\Theta}\lVert v_\theta-v_{j,\theta}\rVert_1\le\frac{12C_6}{1-16^{-\kappa}}aN_j^{-\kappa}. θ ∈ Θ sup ∥ v θ − v j , θ ∥ 1 ≤ 1 − 1 6 − κ 12 C 6 a N j − κ .
In particular, the convergence also holds in ℓ 2 \ell^2 ℓ 2 . By Fatou’s lemma, (112) and (96) ,
∑ m ≠ o ∣ v θ ( m ) ∣ ⟨ m ⟩ ρ ≤ lim inf j → ∞ ∑ m ≠ o ∣ v j , θ ( m ) ∣ ⟨ m ⟩ ρ ≤ C 6 a ≤ 1 4 . \sum_{m\ne o} |v_\theta(m)|\langle m\rangle^\rho\le\liminf_{j\to\infty}\sum_{m\ne o} |v_{j,\theta}(m)|\langle m\rangle^\rho\le C_6 a \le\frac{1}{4}. m = o ∑ ∣ v θ ( m ) ∣ ⟨ m ⟩ ρ ≤ j → ∞ lim inf m = o ∑ ∣ v j , θ ( m ) ∣ ⟨ m ⟩ ρ ≤ C 6 a ≤ 4 1 .
To verify the eigenvalue equation, fix m ∈ Z d m\in\mathbb{Z}^d m ∈ Z d . By (100) , m ∈ B j m\in B_j m ∈ B j for all sufficiently large j j j . Equation (111) gives
f ( θ + m ⋅ ω ) v j , θ ( m ) + ε ( T v j , θ ) ( m ) = E j ( θ ) v j , θ ( m ) . f(\theta+m\cdot\omega)v_{j,\theta}(m)+\varepsilon(\mathcal{T}v_{j,\theta})(m)=E_j(\theta)v_{j,\theta}(m). f ( θ + m ⋅ ω ) v j , θ ( m ) + ε ( T v j , θ ) ( m ) = E j ( θ ) v j , θ ( m ) .
Since θ ∈ Θ \theta\in\Theta θ ∈ Θ , the number f ( θ + m ⋅ ω ) f(\theta+m\cdot\omega) f ( θ + m ⋅ ω ) is finite. The operator T \mathcal{T} T defined in (18) is bounded on ℓ 2 \ell^2 ℓ 2 by (5) . Hence v j , θ → v θ v_{j,\theta}\to v_\theta v j , θ → v θ in ℓ 2 \ell^2 ℓ 2 and E j ( θ ) → E ∞ ( θ ) E_j(\theta)\to E_\infty(\theta) E j ( θ ) → E ∞ ( θ ) imply
f ( θ + m ⋅ ω ) v θ ( m ) = E ∞ ( θ ) v θ ( m ) − ε ( T v θ ) ( m ) . f(\theta+m\cdot\omega)v_\theta(m)=E_\infty(\theta)v_\theta(m)-\varepsilon(\mathcal{T}v_\theta)(m). f ( θ + m ⋅ ω ) v θ ( m ) = E ∞ ( θ ) v θ ( m ) − ε ( T v θ ) ( m ) .
Since E ∞ ( θ ) v θ − ε T v θ ∈ ℓ 2 ( Z d ) E_\infty(\theta)v_\theta-\varepsilon\mathcal{T}v_\theta\in\ell^2(\mathbb{Z}^d) E ∞ ( θ ) v θ − ε T v θ ∈ ℓ 2 ( Z d ) , we have v θ ∈ Dom ( H ( θ ) ) v_\theta\in\operatorname{Dom}(\mathcal{H}(\theta)) v θ ∈ Dom ( H ( θ )) . Thus H ( θ ) v θ = E ∞ ( θ ) v θ \mathcal{H}(\theta)v_\theta=E_\infty(\theta)v_\theta H ( θ ) v θ = E ∞ ( θ ) v θ . □
Completeness and dynamical localization
We prove completeness by estimating the matrix of normalized eigenvectors in the weighted column norm.
Proof of Theorem 1.1 . Use ε 0 \varepsilon_0 ε 0 from (97) and fix θ ∈ Θ \theta\in\Theta θ ∈ Θ . For each n ∈ Z d n\in\mathbb{Z}^d n ∈ Z d , let v θ + n ⋅ ω v_{\theta+n\cdot\omega} v θ + n ⋅ ω be the eigenvector supplied by Lemma 6.1 at phase θ + n ⋅ ω \theta+n\cdot\omega θ + n ⋅ ω , and define
ψ θ , n = U ( − n ) v θ + n ⋅ ω ∥ v θ + n ⋅ ω ∥ 2 . \psi_{\theta,n}=\frac{\mathcal{U}(-n)v_{\theta+n\cdot\omega}}{\|v_{\theta+n\cdot\omega}\|_2}. ψ θ , n = ∥ v θ + n ⋅ ω ∥ 2 U ( − n ) v θ + n ⋅ ω .
Here U ( − n ) \mathcal{U}(-n) U ( − n ) is the translation defined in (17) . By (19) , ψ θ , n \psi_{\theta,n} ψ θ , n is a normalized eigenvector of H ( θ ) \mathcal{H}(\theta) H ( θ ) with eigenvalue E ∞ ( θ + n ⋅ ω ) E_\infty(\theta+n\cdot\omega) E ∞ ( θ + n ⋅ ω ) and ψ θ , n ( n ) > 0 \psi_{\theta,n}(n)>0 ψ θ , n ( n ) > 0 . By (161) , E ∞ ( θ + n ⋅ ω ) ≠ E ∞ ( θ + m ⋅ ω ) E_\infty(\theta+n\cdot\omega)\ne E_\infty(\theta+m\cdot\omega) E ∞ ( θ + n ⋅ ω ) = E ∞ ( θ + m ⋅ ω ) for n ≠ m n\ne m n = m . Self-adjointness of H ( θ ) \mathcal{H}(\theta) H ( θ ) therefore implies
⟨ ψ θ , n , ψ θ , m ⟩ = δ n m , n , m ∈ Z d . \langle\psi_{\theta,n},\psi_{\theta,m}\rangle=\delta_{nm},\qquad n,m\in\mathbb{Z}^d. ⟨ ψ θ , n , ψ θ , m ⟩ = δ nm , n , m ∈ Z d .
We next estimate the change caused by normalization. Since v θ + n ⋅ ω ( o ) = 1 v_{\theta+n\cdot\omega}(o)=1 v θ + n ⋅ ω ( o ) = 1 , we have
0 ≤ 1 − ∥ v θ + n ⋅ ω ∥ 2 − 1 ≤ 1 2 ( ∥ v θ + n ⋅ ω ∥ 2 2 − 1 ) ≤ 1 2 ( ∑ m ≠ o ∣ v θ + n ⋅ ω ( m ) ∣ ) 2 ≤ C 6 2 2 a 2 . (162) 0\le1-\|v_{\theta+n\cdot\omega}\|_2^{-1}\le\frac{1}{2}\left(\|v_{\theta+n\cdot\omega}\|_2^2-1\right)\le\frac{1}{2}\left(\sum_{m\ne o}|v_{\theta+n\cdot\omega}(m)|\right)^2\le\frac{C_6^2}{2}a^2. \tag*{(162)} 0 ≤ 1 − ∥ v θ + n ⋅ ω ∥ 2 − 1 ≤ 2 1 ( ∥ v θ + n ⋅ ω ∥ 2 2 − 1 ) ≤ 2 1 m = o ∑ ∣ v θ + n ⋅ ω ( m ) ∣ 2 ≤ 2 C 6 2 a 2 . (162)
Since ψ θ , n ( n ) = ∥ v θ + n ⋅ ω ∥ 2 − 1 \psi_{\theta,n}(n)=\|v_{\theta+n\cdot\omega}\|_2^{-1} ψ θ , n ( n ) = ∥ v θ + n ⋅ ω ∥ 2 − 1 and ∑ m ≠ o ∣ v θ + n ⋅ ω ( m ) ∣ ≤ 1 / 4 \sum_{m\ne o}|v_{\theta+n\cdot\omega}(m)|\le1/4 ∑ m = o ∣ v θ + n ⋅ ω ( m ) ∣ ≤ 1/4 , (162) and Lemma 6.1 imply
sup n ∑ m ∣ ψ θ , n ( m ) − δ m n ∣ ⟨ m − n ⟩ ρ ≤ 2 C 6 a . (163) \sup_n\sum_m|\psi_{\theta,n}(m)-\delta_{mn}|\langle m-n\rangle^\rho\le2C_6a. \tag*{(163)} n sup m ∑ ∣ ψ θ , n ( m ) − δ mn ∣ ⟨ m − n ⟩ ρ ≤ 2 C 6 a . (163)
By (163) , (34) and (6) , we obtain (10) with a finite constant C 2 C_2 C 2 depending only on d d d , r r r , τ \tau τ , γ \gamma γ , L L L , ρ \rho ρ and ϕ \phi ϕ .
Define the matrix Ψ ( θ ) \Psi(\theta) Ψ ( θ ) by
Ψ ( θ ) ( m , n ) = ψ θ , n ( m ) , Ψ ( θ ) e n = ψ θ , n . (164) \Psi(\theta)(m,n)=\psi_{\theta,n}(m),\qquad\Psi(\theta)e_n=\psi_{\theta,n}. \tag*{(164)} Ψ ( θ ) ( m , n ) = ψ θ , n ( m ) , Ψ ( θ ) e n = ψ θ , n . (164)
For a finitely supported vector x x x , orthonormality gives ∥ Ψ ( θ ) x ∥ 2 = ∥ x ∥ 2 \|\Psi(\theta)x\|_2=\|x\|_2 ∥Ψ ( θ ) x ∥ 2 = ∥ x ∥ 2 . Hence (164) extends to an isometry on ℓ 2 ( Z d ) \ell^2(\mathbb{Z}^d) ℓ 2 ( Z d ) . By Lemma 6.1 , (162) and (26) , we obtain
∥ Ψ ( θ ) − I ∥ ρ ≤ 1 4 + 1 2 ( 1 4 ) 2 = 9 32 < 1 2 . (165) \|\Psi(\theta)-\mathcal{I}\|_\rho\le\frac{1}{4}+\frac{1}{2}\left(\frac{1}{4}\right)^2=\frac{9}{32}<\frac{1}{2}. \tag*{(165)} ∥Ψ ( θ ) − I ∥ ρ ≤ 4 1 + 2 1 ( 4 1 ) 2 = 32 9 < 2 1 . (165)
By Lemma 2.1 , Ψ ( θ ) \Psi(\theta) Ψ ( θ ) is then invertible on ℓ 1 \ell^1 ℓ 1 , with
Ψ ( θ ) − 1 = ∑ k = 0 ∞ ( I − Ψ ( θ ) ) k , \Psi(\theta)^{-1}=\sum_{k=0}^{\infty}(\mathcal{I}-\Psi(\theta))^k, Ψ ( θ ) − 1 = k = 0 ∑ ∞ ( I − Ψ ( θ ) ) k ,
where the series converges in the weighted column norm. Moreover,
∥ Ψ ( θ ) − 1 ∥ ρ ≤ ( 1 − ∥ Ψ ( θ ) − I ∥ ρ ) − 1 ≤ 2. (166) \lVert\Psi(\theta)^{-1}\rVert_{\rho} \leq\left(1-\lVert\Psi(\theta)-I\rVert_{\rho}\right)^{-1}\leq2. \tag*{(166)} ∥ Ψ ( θ ) − 1 ∥ ρ ≤ ( 1 − ∥ Ψ ( θ ) − I ∥ ρ ) − 1 ≤ 2. (166)
The actions of Ψ ( θ ) \Psi(\theta) Ψ ( θ ) on ℓ 1 \ell^1 ℓ 1 and ℓ 2 \ell^2 ℓ 2 agree on ℓ 1 \ell^1 ℓ 1 : approximate an ℓ 1 \ell^1 ℓ 1 vector by finitely supported vectors and use the continuous inclusion ℓ 1 ⊂ ℓ 2 \ell^1\subset\ell^2 ℓ 1 ⊂ ℓ 2 . For every n ∈ Z d n\in\mathbb{Z}^d n ∈ Z d , we therefore have Ψ ( θ ) e n ∈ ℓ 1 ⊂ ℓ 2 \Psi(\theta)e_n\in\ell^1\subset\ell^2 Ψ ( θ ) e n ∈ ℓ 1 ⊂ ℓ 2 and Ψ ( θ ) ( Ψ ( θ ) − 1 e n ) = e n \Psi(\theta)(\Psi(\theta)^{-1}e_n)=e_n Ψ ( θ ) ( Ψ ( θ ) − 1 e n ) = e n in ℓ 2 \ell^2 ℓ 2 . The range of the isometry Ψ ( θ ) \Psi(\theta) Ψ ( θ ) is closed and contains every e n e_n e n , so Ran Ψ ( θ ) = ℓ 2 ( Z d ) \operatorname{Ran}\Psi(\theta)=\ell^2(\mathbb{Z}^d) Ran Ψ ( θ ) = ℓ 2 ( Z d ) . Hence Ψ ( θ ) \Psi(\theta) Ψ ( θ ) is unitary and ( ψ θ , n ) n ∈ Z d (\psi_{\theta,n})_{n\in\mathbb{Z}^d} ( ψ θ , n ) n ∈ Z d is a complete orthonormal basis. In particular, H ( θ ) \mathcal{H}(\theta) H ( θ ) has pure point spectrum. For every x ∈ ℓ 1 x\in\ell^1 x ∈ ℓ 1 , the identity Ψ ( θ ) Ψ ( θ ) − 1 x = x \Psi(\theta)\Psi(\theta)^{-1}x=x Ψ ( θ ) Ψ ( θ ) − 1 x = x in ℓ 2 \ell^2 ℓ 2 now implies Ψ ( θ ) − 1 x = Ψ ( θ ) ∗ x \Psi(\theta)^{-1}x=\Psi(\theta)^*x Ψ ( θ ) − 1 x = Ψ ( θ ) ∗ x .
To prove (11) , the spectral theorem gives, for u ∈ R u\in\mathbb{R} u ∈ R ,
exp ( − i u H ( θ ) ) = Ψ ( θ ) diag n ∈ Z d ( exp ( − i u E ∞ ( θ + n ⋅ ω ) ) ) Ψ ( θ ) ∗ . (167) \exp(-iu\mathcal{H}(\theta))=\Psi(\theta)\operatorname{diag}_{n\in\mathbb{Z}^d}\left(\exp(-iuE_{\infty}(\theta+n\cdot\omega))\right)\Psi(\theta)^*. \tag*{(167)} exp ( − i u H ( θ )) = Ψ ( θ ) diag n ∈ Z d ( exp ( − i u E ∞ ( θ + n ⋅ ω )) ) Ψ ( θ ) ∗ . (167)
For every u ∈ R u\in\mathbb{R} u ∈ R ,
∥ diag n ∈ Z d ( exp ( − i u E ∞ ( θ + n ⋅ ω ) ) ) ∥ ρ = 1. \left\lVert\operatorname{diag}_{n\in\mathbb{Z}^d}\left(\exp(-iuE_{\infty}(\theta+n\cdot\omega))\right)\right\rVert_{\rho}=1. ∥ diag n ∈ Z d ( exp ( − i u E ∞ ( θ + n ⋅ ω )) ) ∥ ρ = 1.
On ℓ 1 \ell^1 ℓ 1 , (167) therefore holds with Ψ ( θ ) − 1 \Psi(\theta)^{-1} Ψ ( θ ) − 1 in place of Ψ ( θ ) ∗ \Psi(\theta)^* Ψ ( θ ) ∗ . By (27) , (165) and (166) , we obtain
∥ exp ( − i u H ( θ ) ) ∥ ρ ≤ ∥ Ψ ( θ ) ∥ ρ ∥ Ψ ( θ ) − 1 ∥ ρ ≤ 3 2 ⋅ 2 = 3. \left\lVert\exp(-iu\mathcal{H}(\theta))\right\rVert_{\rho}\leq\lVert\Psi(\theta)\rVert_{\rho}\lVert\Psi(\theta)^{-1}\rVert_{\rho}\leq\frac{3}{2}\cdot2=3. ∥ exp ( − i u H ( θ )) ∥ ρ ≤ ∥ Ψ ( θ ) ∥ ρ ∥ Ψ ( θ ) − 1 ∥ ρ ≤ 2 3 ⋅ 2 = 3.
This proves (11) and completes the proof. □
Appendix A. Sparse sums and Schur estimates
We estimate sums over M M M -separated sets and use the resulting bounds to control one Schur elimination.
Sparse sums
Lemma A.1 (Sparse sums). Let M ≥ 1 M\geq1 M ≥ 1 , let X ⊂ Z d X\subset\mathbb{Z}^d X ⊂ Z d be M M M -separated and let r > d r>d r > d . Use C 3 ( d , r ) C_3(d,r) C 3 ( d , r ) from (36) . For every s > d s>d s > d and p ∈ X p\in X p ∈ X , we have
∑ x ∈ X ∖ { p } ⟨ p − x ⟩ − s ≤ 5 d 1 − 2 d − s M − s . (168) \sum_{x\in X\setminus\{p\}}\langle p-x\rangle^{-s}\leq\frac{5^d}{1-2^{d-s}}M^{-s}. \tag*{(168)} x ∈ X ∖ { p } ∑ ⟨ p − x ⟩ − s ≤ 1 − 2 d − s 5 d M − s . (168) For distinct p , q ∈ X p,q\in X p , q ∈ X , we have
∑ x ∈ X ∖ { p , q } ⟨ p − x ⟩ − r ⟨ x − q ⟩ − r ≤ C 3 ( d , r ) M − r ⟨ p − q ⟩ − r . (169) \sum_{x\in X\setminus\{p,q\}}\langle p-x\rangle^{-r}\langle x-q\rangle^{-r}\leq C_3(d,r)M^{-r}\langle p-q\rangle^{-r}. \tag*{(169)} x ∈ X ∖ { p , q } ∑ ⟨ p − x ⟩ − r ⟨ x − q ⟩ − r ≤ C 3 ( d , r ) M − r ⟨ p − q ⟩ − r . (169) If p ∈ X p\in X p ∈ X , Y ⊂ X ∖ { p } Y\subset X\setminus\{p\} Y ⊂ X ∖ { p } , y ∈ R d y\in\mathbb{R}^d y ∈ R d and dist ( y , X ) ≥ M \operatorname{dist}(y,X)\geq M dist ( y , X ) ≥ M , then
∑ x ∈ Y ⟨ p − x ⟩ − r ⟨ x − y ⟩ − r ≤ C 3 ( d , r ) M − r ⟨ p − y ⟩ − r . (170) \sum_{x\in Y}\langle p-x\rangle^{-r}\langle x-y\rangle^{-r}\leq C_3(d,r)M^{-r}\langle p-y\rangle^{-r}. \tag*{(170)} x ∈ Y ∑ ⟨ p − x ⟩ − r ⟨ x − y ⟩ − r ≤ C 3 ( d , r ) M − r ⟨ p − y ⟩ − r . (170) For every p ∈ Z d p\in\mathbb{Z}^d p ∈ Z d and R ≥ M R\geq M R ≥ M , we also have
∑ x ∈ X ∣ x − p ∣ ≥ R ⟨ p − x ⟩ − 2 r ≤ C 3 M − d R d − 2 r . (171) \sum_{\substack{x\in X\\ \lvert x-p\rvert\geq R}}\langle p-x\rangle^{-2r}\leq C_3M^{-d}R^{d-2r}. \tag*{(171)} x ∈ X ∣ x − p ∣ ≥ R ∑ ⟨ p − x ⟩ − 2 r ≤ C 3 M − d R d − 2 r . (171) Proof. The half-open cubes x + [ − M / 2 , M / 2 ) d x+[-M/2,M/2)^d x + [ − M /2 , M /2 ) d , x ∈ X x\in X x ∈ X , are pairwise disjoint. Comparison of volumes gives, for p ∈ R d p\in\mathbb{R}^d p ∈ R d and R ≥ M R\geq M R ≥ M ,
# { x ∈ X : ∣ x − p ∣ < R } ≤ ( 2 R M + 1 ) d ≤ 3 d ( R / M ) d . (172) \#\{x\in X:\lvert x-p\rvert<R\}\leq\left(\frac{2R}{M}+1\right)^d\leq3^d(R/M)^d. \tag*{(172)} # { x ∈ X : ∣ x − p ∣ < R } ≤ ( M 2 R + 1 ) d ≤ 3 d ( R / M ) d . (172)
For j ≥ 0 j \ge0 j ≥ 0 , (A.5) implies
# { x ∈ X : 2 j R ≤ ∣ x − p ∣ < 2 j + 1 R } ≤ ( 2 j + 2 R M + 1 ) d ≤ 5 d 2 j d ( R / M ) d . \#\{x \in X : 2^j R \le|x-p| < 2^{j+1}R\} \le\left(2^{j+2}\frac{R}{M}+1\right)^d \le5^d 2^{jd}(R/M)^d. # { x ∈ X : 2 j R ≤ ∣ x − p ∣ < 2 j + 1 R } ≤ ( 2 j + 2 M R + 1 ) d ≤ 5 d 2 j d ( R / M ) d .
Consequently, for every s > d s>d s > d ,
∑ x ∈ X ∣ x − p ∣ ≥ R ⟨ p − x ⟩ − s ≤ 5 d ( R / M ) d ∑ j ≥ 0 2 j d ( 2 j R ) − s = 5 d 1 − 2 d − s M − d R d − s . (173) \sum_{\substack{x\in X\\|x-p|\ge R}}\langle p-x\rangle^{-s}
\le5^d(R/M)^d\sum_{j\ge0}2^{jd}(2^jR)^{-s}
=\frac{5^d}{1-2^{d-s}}M^{-d}R^{d-s}. \tag*{(173)} x ∈ X ∣ x − p ∣ ≥ R ∑ ⟨ p − x ⟩ − s ≤ 5 d ( R / M ) d j ≥ 0 ∑ 2 j d ( 2 j R ) − s = 1 − 2 d − s 5 d M − d R d − s . (173)
Taking s = 2 r s=2r s = 2 r gives the tail bound (A.4). For p ∈ X p\in X p ∈ X , all points of X ∖ { p } X\setminus\{p\} X ∖ { p } have distance at least M M M from p p p , so (A.6) with R = M R=M R = M proves (A.1). Equation (A.6) with p = y p=y p = y , R = M R=M R = M and s = r s=r s = r gives
∑ x ∈ X ⟨ x − y ⟩ − r ≤ 5 d 1 − 2 d − r M − r , dist ( y , X ) ≥ M . (174) \sum_{x\in X}\langle x-y\rangle^{-r}
\le\frac{5^d}{1-2^{d-r}}M^{-r},
\qquad\operatorname{dist}(y,X)\ge M. \tag*{(174)} x ∈ X ∑ ⟨ x − y ⟩ − r ≤ 1 − 2 d − r 5 d M − r , dist ( y , X ) ≥ M . (174)
For any p , x , q ∈ R d p,x,q\in\mathbb{R}^d p , x , q ∈ R d , the triangle inequality implies ∣ p − x ∣ ≥ ∣ p − q ∣ / 2 |p-x|\ge|p-q|/2 ∣ p − x ∣ ≥ ∣ p − q ∣/2 or ∣ x − q ∣ ≥ ∣ p − q ∣ / 2 |x-q|\ge|p-q|/2 ∣ x − q ∣ ≥ ∣ p − q ∣/2 . Hence
⟨ p − x ⟩ − r ⟨ x − q ⟩ − r ≤ 2 r ⟨ p − q ⟩ − r ( ⟨ p − x ⟩ − r + ⟨ x − q ⟩ − r ) . \langle p-x\rangle^{-r}\langle x-q\rangle^{-r}
\le2^r\langle p-q\rangle^{-r}\left(\langle p-x\rangle^{-r}+\langle x-q\rangle^{-r}\right). ⟨ p − x ⟩ − r ⟨ x − q ⟩ − r ≤ 2 r ⟨ p − q ⟩ − r ( ⟨ p − x ⟩ − r + ⟨ x − q ⟩ − r ) .
For distinct p , q ∈ X p,q\in X p , q ∈ X , summing over x ∈ X ∖ { p , q } x\in X\setminus\{p,q\} x ∈ X ∖ { p , q } and applying (A.1) with s = r s=r s = r proves (A.2), with coefficient 2 r + 1 5 d / ( 1 − 2 d − r ) 2^{r+1}5^d/(1-2^{d-r}) 2 r + 1 5 d / ( 1 − 2 d − r ) . Replacing q q q by y y y and summing over x ∈ Y x\in Y x ∈ Y , we obtain (A.3) from (A.1) and (A.7). Also, (A.1) with X = Z d X=\mathbb{Z}^d X = Z d , M = 1 M=1 M = 1 , s = r s=r s = r and p = o p=o p = o bounds ∑ k ⟨ k ⟩ − r \sum_{k}\langle k\rangle^{-r} ∑ k ⟨ k ⟩ − r by 1 + 5 d / ( 1 − 2 d − r ) 1+5^d/(1-2^{d-r}) 1 + 5 d / ( 1 − 2 d − r ) . The convolution, tail and full-lattice coefficients are finite and depend only on d , r d,r d , r , which justifies the common choice of C 3 C_3 C 3 in (36) . □ \square □
One Schur elimination
Use C 3 C_3 C 3 from (36) and the coefficients C 4 ( d , r , s ) C_4(d,r,s) C 4 ( d , r , s ) fixed before (55) .
Lemma A.2. Let r > d r>d r > d and let X = P ⊔ Q ⊂ Z d X=P\sqcup Q\subset\mathbb{Z}^d X = P ⊔ Q ⊂ Z d be finite and M M M -separated, where M ≥ 1 M\ge1 M ≥ 1 . Let M \mathcal{M} M be a Hermitian matrix on X X X such that
∣ M ( p , q ) ∣ ≤ a ∗ ⟨ p − q ⟩ − r ( p ≠ q ) , ∣ M ( q , q ) ∣ ≥ Δ ( q ∈ Q ) , |\mathcal{M}(p,q)|\le a_*\langle p-q\rangle^{-r}\quad(p\ne q),\qquad|\mathcal{M}(q,q)|\ge\Delta\quad(q\in Q), ∣ M ( p , q ) ∣ ≤ a ∗ ⟨ p − q ⟩ − r ( p = q ) , ∣ M ( q , q ) ∣ ≥ Δ ( q ∈ Q ) , where a ∗ ≥ 0 a_*\ge0 a ∗ ≥ 0 and Δ > 0 \Delta>0 Δ > 0 . If
a ∗ Δ − 1 M − r ≤ ( 2 C 3 ) − 1 , (175) a_*\Delta^{-1}M^{-r}\le(2C_3)^{-1}, \tag*{(175)} a ∗ Δ − 1 M − r ≤ ( 2 C 3 ) − 1 , (175) then M Q \mathcal{M}_Q M Q is invertible and
∣ Sch Q ( M ) ( p , q ) − M ( p , q ) ∣ ≤ 2 C 3 a ∗ 2 Δ − 1 M − r ⟨ p − q ⟩ − r , p , q ∈ P , p ≠ q , (176) |\operatorname{Sch}_Q(\mathcal{M})(p,q)-\mathcal{M}(p,q)|
\le2C_3a_*^2\Delta^{-1}M^{-r}\langle p-q\rangle^{-r},
\qquad p,q\in P,\ p\ne q, \tag*{(176)} ∣ Sch Q ( M ) ( p , q ) − M ( p , q ) ∣ ≤ 2 C 3 a ∗ 2 Δ − 1 M − r ⟨ p − q ⟩ − r , p , q ∈ P , p = q , (176) ∣ Sch Q ( M ) ( p , p ) − M ( p , p ) ∣ ≤ 2 C 3 a ∗ 2 Δ − 1 M − 2 r , p ∈ P , (177) |\operatorname{Sch}_Q(\mathcal{M})(p,p)-\mathcal{M}(p,p)|
\le2C_3a_*^2\Delta^{-1}M^{-2r},
\qquad p\in P, \tag*{(177)} ∣ Sch Q ( M ) ( p , p ) − M ( p , p ) ∣ ≤ 2 C 3 a ∗ 2 Δ − 1 M − 2 r , p ∈ P , (177) ∣ W ( M ) ( q , p ) ∣ ≤ 2 a ∗ Δ − 1 ⟨ q − p ⟩ − r , q ∈ Q , p ∈ P . (178) |\mathcal{W}(\mathcal{M})(q,p)|
\le2a_*\Delta^{-1}\langle q-p\rangle^{-r},
\qquad q\in Q,\ p\in P. \tag*{(178)} ∣ W ( M ) ( q , p ) ∣ ≤ 2 a ∗ Δ − 1 ⟨ q − p ⟩ − r , q ∈ Q , p ∈ P . (178) Here Sch Q \operatorname{Sch}_Q Sch Q and W ( M ) \mathcal{W}(\mathcal{M}) W ( M ) are defined in (22) and (23) , respectively. For 0 ≤ ρ < r − d 0\le\rho<r-d 0 ≤ ρ < r − d , we also have
∥ W ( M ) ∥ ρ ≤ 1 + C 4 ( d , r , ρ ) 4 a ∗ Δ − 1 M ρ − r . (179) \|\mathcal{W}(\mathcal{M})\|_\rho
\le1+\frac{C_4(d,r,\rho)}{4}a_*\Delta^{-1}M^{\rho-r}. \tag*{(179)} ∥ W ( M ) ∥ ρ ≤ 1 + 4 C 4 ( d , r , ρ ) a ∗ Δ − 1 M ρ − r . (179) Proof. Let D = diag q ∈ Q M ( q , q ) \mathcal{D}=\operatorname{diag}_{q\in Q}\mathcal{M}(q,q) D = diag q ∈ Q M ( q , q ) and M o f f = M Q − D \mathcal{M}^{\mathrm{off}}=\mathcal{M}_Q-\mathcal{D} M off = M Q − D . By (A.1) and (A.8),
sup q ∈ Q ∑ x ∈ Q ∣ ( D − 1 M o f f ) ( q , x ) ∣ ≤ C 3 a ∗ Δ − 1 M − r ≤ 1 2 . \sup_{q\in Q}\sum_{x\in Q}|(\mathcal{D}^{-1}\mathcal{M}^{\mathrm{off}})(q,x)|
\le C_3a_*\Delta^{-1}M^{-r}\le\frac12. q ∈ Q sup x ∈ Q ∑ ∣ ( D − 1 M off ) ( q , x ) ∣ ≤ C 3 a ∗ Δ − 1 M − r ≤ 2 1 . It follows that
M Q − 1 = ∑ ℓ = 0 ∞ ( − D − 1 M o f f ) ℓ D − 1 . \mathcal{M}_Q^{-1}
=\sum_{\ell=0}^{\infty}(-\mathcal{D}^{-1}\mathcal{M}^{\mathrm{off}})^\ell\mathcal{D}^{-1}. M Q − 1 = ℓ = 0 ∑ ∞ ( − D − 1 M off ) ℓ D − 1 .
Multiplication by D − 1 M o f f D^{-1}\mathcal{M}^{\mathrm{off}} D − 1 M off involves a sum over x ∈ Q ∖ { q } x \in Q \setminus\{q\} x ∈ Q ∖ { q } . For p ∈ P p \in P p ∈ P and q ∈ Q q \in Q q ∈ Q , (A.2) gives
∑ x ∈ Q ∖ { q } ∣ ( D − 1 M o f f ) ( q , x ) ∣ ⟨ x − p ⟩ − r ≤ C 3 a ∗ Δ − 1 M − r ⟨ q − p ⟩ − r ≤ 1 2 ⟨ q − p ⟩ − r . (180) \sum_{x\in Q\setminus\{q\}} \left|(D^{-1}\mathcal{M}^{\mathrm{off}})(q,x)\right|\langle x-p\rangle^{-r} \le C_{3}a_{*}\Delta^{-1}M^{-r}\langle q-p\rangle^{-r} \le\frac{1}{2}\langle q-p\rangle^{-r}. \tag*{(180)} x ∈ Q ∖ { q } ∑ ( D − 1 M off ) ( q , x ) ⟨ x − p ⟩ − r ≤ C 3 a ∗ Δ − 1 M − r ⟨ q − p ⟩ − r ≤ 2 1 ⟨ q − p ⟩ − r . (180)
Starting from ∣ [ D − 1 M Q , p ] ( q ) ∣ ≤ a ∗ Δ − 1 ⟨ q − p ⟩ − r \left|[D^{-1}\mathcal{M}_{Q,p}](q)\right| \le a_{*}\Delta^{-1}\langle q-p\rangle^{-r} [ D − 1 M Q , p ] ( q ) ≤ a ∗ Δ − 1 ⟨ q − p ⟩ − r , induction using (A.13) gives
∣ [ ( D − 1 M o f f ) ℓ D − 1 M Q , p ] ( q ) ∣ ≤ 2 − ℓ a ∗ Δ − 1 ⟨ q − p ⟩ − r , q ∈ Q , ℓ ≥ 0. \left|[(D^{-1}\mathcal{M}^{\mathrm{off}})^{\ell}D^{-1}\mathcal{M}_{Q,p}](q)\right| \le2^{-\ell}a_{*}\Delta^{-1}\langle q-p\rangle^{-r}, \qquad q\in Q,\ \ell\ge0. [( D − 1 M off ) ℓ D − 1 M Q , p ] ( q ) ≤ 2 − ℓ a ∗ Δ − 1 ⟨ q − p ⟩ − r , q ∈ Q , ℓ ≥ 0.
After summing in ℓ \ell ℓ , we obtain ∣ W ( M ) ( q , p ) ∣ ≤ 2 a ∗ Δ − 1 ⟨ q − p ⟩ − r \left|\mathcal{W}(\mathcal{M})(q,p)\right| \le2a_{*}\Delta^{-1}\langle q-p\rangle^{-r} ∣ W ( M ) ( q , p ) ∣ ≤ 2 a ∗ Δ − 1 ⟨ q − p ⟩ − r , which proves (A.11). By (22) and (A.11),
∣ Sch Q ( M ) ( p , p ′ ) − M ( p , p ′ ) ∣ ≤ 2 a ∗ 2 Δ − 1 ∑ q ∈ Q ⟨ p − q ⟩ − r ⟨ q − p ′ ⟩ − r , p , p ′ ∈ P . \left|\operatorname{Sch}_{Q}(\mathcal{M})(p,p')-\mathcal{M}(p,p')\right| \le2a_{*}^{2}\Delta^{-1}\sum_{q\in Q}\langle p-q\rangle^{-r}\langle q-p'\rangle^{-r}, \qquad p,p'\in P. ∣ Sch Q ( M ) ( p , p ′ ) − M ( p , p ′ ) ∣ ≤ 2 a ∗ 2 Δ − 1 q ∈ Q ∑ ⟨ p − q ⟩ − r ⟨ q − p ′ ⟩ − r , p , p ′ ∈ P .
For p ≠ p ′ p\ne p' p = p ′ , (A.2) proves (A.9). For p = p ′ p=p' p = p ′ , (A.1) with s = 2 r s=2r s = 2 r proves (A.10). Finally, multiply (A.11) by ⟨ q − p ⟩ ρ \langle q-p\rangle^{\rho} ⟨ q − p ⟩ ρ and sum over q q q . Since ρ < r − d \rho<r-d ρ < r − d , (A.1) with s = r − ρ s=r-\rho s = r − ρ proves (A.12) for a finite coefficient C 4 ( d , r , ρ ) C_{4}(d,r,\rho) C 4 ( d , r , ρ ) , as fixed before (55) . □
Appendix B. Geometric and domain verifications
The finite-region construction
B.1. The finite-region construction. We prove Proposition 5.5, using the construction (137) –(138) . Fix n ≥ 1 n\ge1 n ≥ 1 and assume the geometric assertion (1) of Theorem 4.1 through step n n n .
Proof. Since (78) gives ∣ p − q ∣ > 33600 N h > 18 N h + 1 |p-q|>33600N_h>18N_{h+1} ∣ p − q ∣ > 33600 N h > 18 N h + 1 for distinct p , q ∈ Z h p,q\in\mathbb{Z}_h p , q ∈ Z h , the balls p + Λ 9 N h + 1 p+\Lambda_{9N_{h+1}} p + Λ 9 N h + 1 , p ∈ Z h p\in\mathbb{Z}_h p ∈ Z h , are pairwise disjoint. For finite B B B , only finitely many centers p ∈ Z h p\in\mathbb{Z}_h p ∈ Z h satisfy ( p + Λ 9 N h + 1 ) ∩ B ≠ ∅ (p+\Lambda_{9N_{h+1}})\cap B\ne\varnothing ( p + Λ 9 N h + 1 ) ∩ B = ∅ . Moreover, (137) implies
B ⊂ B ~ ⊂ B + Λ 18 N h + 1 , ( p + Λ 9 N h + 1 ) ∩ B ~ ≠ ∅ ⟹ p + Λ 9 N h + 1 ⊂ B ~ , p ∈ Z h . B\subset\widetilde{B}\subset B+\Lambda_{18N_{h+1}}, \qquad(p+\Lambda_{9N_{h+1}})\cap\widetilde{B}\ne\varnothing\Longrightarrow p+\Lambda_{9N_{h+1}}\subset\widetilde{B}, \qquad p\in\mathbb{Z}_h. B ⊂ B ⊂ B + Λ 18 N h + 1 , ( p + Λ 9 N h + 1 ) ∩ B = ∅ ⟹ p + Λ 9 N h + 1 ⊂ B , p ∈ Z h .
Consequently,
Λ 4 N n + 1 ⊂ B n + 1 ⊂ Λ 4 N n + 1 + 18 ∑ ℓ = 1 n N ℓ ⊂ Λ ( 26 / 5 ) N n + 1 ⊂ Λ 8 N n + 1 . \Lambda_{4N_{n+1}}\subset B_{n+1}\subset\Lambda_{4N_{n+1}+18\sum_{\ell=1}^{n}N_{\ell}}\subset\Lambda_{(26/5)N_{n+1}}\subset\Lambda_{8N_{n+1}}. Λ 4 N n + 1 ⊂ B n + 1 ⊂ Λ 4 N n + 1 + 18 ∑ ℓ = 1 n N ℓ ⊂ Λ ( 26/5 ) N n + 1 ⊂ Λ 8 N n + 1 .
Since B n ⊂ Λ 8 N n B_n\subset\Lambda_{8N_n} B n ⊂ Λ 8 N n and 8 N n < 4 N n + 1 8N_n<4N_{n+1} 8 N n < 4 N n + 1 , we obtain (100) with m = n + 1 m=n+1 m = n + 1 .
Let 0 ≤ h ≤ n − 1 0\le h\le n-1 0 ≤ h ≤ n − 1 and p ∈ B n + 1 ∩ Z h p\in B_{n+1}\cap\mathbb{Z}_h p ∈ B n + 1 ∩ Z h . Let F F F be the set immediately after the enlargement with centers in Z h \mathbb{Z}_h Z h . The subsequent enlargements give
B n + 1 ⊂ F + Λ 18 ∑ ℓ = 1 h N ℓ , 18 ∑ ℓ = 1 h N ℓ ≤ 6 5 N h + 1 < 9 N h + 1 . B_{n+1}\subset F+\Lambda_{18\sum_{\ell=1}^{h}N_{\ell}}, \qquad18\sum_{\ell=1}^{h}N_{\ell}\le\frac{6}{5}N_{h+1}<9N_{h+1}. B n + 1 ⊂ F + Λ 18 ∑ ℓ = 1 h N ℓ , 18 ℓ = 1 ∑ h N ℓ ≤ 5 6 N h + 1 < 9 N h + 1 .
Thus p + Λ 9 N h + 1 p+\Lambda_{9N_{h+1}} p + Λ 9 N h + 1 meets F F F and is contained in F F F . By (100) , B h + 1 ⊂ Λ 8 N h + 1 B_{h+1}\subset\Lambda_{8N_{h+1}} B h + 1 ⊂ Λ 8 N h + 1 , so
B h + 1 ( p ) + Λ N h + 1 ⊂ p + Λ 9 N h + 1 ⊂ F ⊂ B n + 1 . B_{h+1}(p)+\Lambda_{N_{h+1}}\subset p+\Lambda_{9N_{h+1}}\subset F\subset B_{n+1}. B h + 1 ( p ) + Λ N h + 1 ⊂ p + Λ 9 N h + 1 ⊂ F ⊂ B n + 1 .
This proves (101) for m = n + 1 m=n+1 m = n + 1 . If p ∈ Z n o p\in\mathbb{Z}_{n}^{o} p ∈ Z n o , then ∣ p ∣ > 33600 N n > 9 N n + 1 |p|>33600N_n>9N_{n+1} ∣ p ∣ > 33600 N n > 9 N n + 1 by (78) . Since B n + 1 ⊂ Λ 8 N n + 1 B_{n+1}\subset\Lambda_{8N_{n+1}} B n + 1 ⊂ Λ 8 N n + 1 , we have B n + 1 ∩ Z n = { o } B_{n+1}\cap\mathbb{Z}_n=\{o\} B n + 1 ∩ Z n = { o } and o ∉ B n + 1 ( p ) + Λ N n + 1 o\notin B_{n+1}(p)+\Lambda_{N_{n+1}} o ∈ / B n + 1 ( p ) + Λ N n + 1 . Thus (102) holds with m = n + 1 m=n+1 m = n + 1 , and (103) holds with h = n h=n h = n . □
Preservation of the multiplication domain
Fix j ≥ 1 j \ge1 j ≥ 1 and assume assertions (1)–(4) of Theorem 4.1 through step j + 1 j+1 j + 1 , as in Proposition 5.8 . Fix θ ∈ Θ \theta\in\Theta θ ∈ Θ , t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] and i ∈ { j , j + 1 } i \in\{j,j+1\} i ∈ { j , j + 1 } . Use the matrices in (150) –(151) and the multiplication operator F \mathcal{F} F in (156) . We prove (157) and extend (152) from finitely supported vectors to Dom ( F ) \operatorname{Dom}(\mathcal{F}) Dom ( F ) .
Proof. Since θ ∈ Θ \theta\in\Theta θ ∈ Θ , every diagonal entry of F \mathcal{F} F is finite. Each column V ( i ) ( ⋅ , k ) \mathcal{V}^{(i)}(\cdot,k) V ( i ) ( ⋅ , k ) is supported on the finite set B i ( k ) B_i(k) B i ( k ) , so V ( i ) \mathcal{V}^{(i)} V ( i ) maps finitely supported vectors to finitely supported vectors. By (107) , D ( i ) − F \mathcal{D}^{(i)}-\mathcal{F} D ( i ) − F is bounded on ℓ 1 \ell^1 ℓ 1 . Equation (152) gives
[ F , V ( i ) ] = V ( i ) ( D ( i ) − F ) + R ( i ) − t ε T V ( i ) [\mathcal{F},\mathcal{V}^{(i)}]=\mathcal{V}^{(i)}(\mathcal{D}^{(i)}-\mathcal{F})+\mathcal{R}^{(i)}-t\varepsilon\mathcal{T}\mathcal{V}^{(i)} [ F , V ( i ) ] = V ( i ) ( D ( i ) − F ) + R ( i ) − tε T V ( i )
on finitely supported vectors. The right-hand side is bounded on ℓ 1 \ell^1 ℓ 1 by (153) , (155) and the boundedness of T \mathcal{T} T . For u ∈ Dom ( F ) u \in\operatorname{Dom}(\mathcal{F}) u ∈ Dom ( F ) , truncation to u R = Π Λ R ∗ Π Λ R u u_R=\Pi_{\Lambda_R}^{*}\Pi_{\Lambda_R}u u R = Π Λ R ∗ Π Λ R u yields
u R ⟶ u , F u R ⟶ F u , F V ( i ) u R ⟶ V ( i ) F u + [ F , V ( i ) ] u in ℓ 1 . u_R \longrightarrow u,\qquad\mathcal{F}u_R \longrightarrow\mathcal{F}u,\qquad\mathcal{F}\mathcal{V}^{(i)}u_R \longrightarrow\mathcal{V}^{(i)}\mathcal{F}u+[\mathcal{F},\mathcal{V}^{(i)}]u\quad\text{in }\ell^1. u R ⟶ u , F u R ⟶ F u , F V ( i ) u R ⟶ V ( i ) F u + [ F , V ( i ) ] u in ℓ 1 .
Since F \mathcal{F} F is closed, V ( i ) u ∈ Dom ( F ) \mathcal{V}^{(i)}u \in\operatorname{Dom}(\mathcal{F}) V ( i ) u ∈ Dom ( F ) . Passing to the limit along u R u_R u R therefore proves (152) for every u ∈ Dom ( F ) u \in\operatorname{Dom}(\mathcal{F}) u ∈ Dom ( F ) . For K ( i ) = V ( i ) − I \mathcal{K}^{(i)}=\mathcal{V}^{(i)}-\mathcal{I} K ( i ) = V ( i ) − I and m ≥ 1 m \ge1 m ≥ 1 , the commutator identity for ( K ( i ) ) m (\mathcal{K}^{(i)})^m ( K ( i ) ) m gives
∥ F ( K ( i ) ) m u ∥ 1 ≤ ( C 6 α ) m ∥ F u ∥ 1 + m ( C 6 α ) m − 1 ∥ [ F , V ( i ) ] ∥ ℓ 1 → ℓ 1 ∥ u ∥ 1 . \|\mathcal{F}(\mathcal{K}^{(i)})^m u\|_1 \le(C_6\alpha)^m\|\mathcal{F}u\|_1+m(C_6\alpha)^{m-1}\|[\mathcal{F},\mathcal{V}^{(i)}]\|_{\ell^1\to\ell^1}\|u\|_1. ∥ F ( K ( i ) ) m u ∥ 1 ≤ ( C 6 α ) m ∥ F u ∥ 1 + m ( C 6 α ) m − 1 ∥ [ F , V ( i ) ] ∥ ℓ 1 → ℓ 1 ∥ u ∥ 1 .
Consequently, ∑ m ≥ 0 ( − K ( i ) ) m u \sum_{m\ge0}(-\mathcal{K}^{(i)})^m u ∑ m ≥ 0 ( − K ( i ) ) m u converges in the graph norm of F \mathcal{F} F . Thus ( V ( i ) ) − 1 (\mathcal{V}^{(i)})^{-1} ( V ( i ) ) − 1 also preserves Dom ( F ) \operatorname{Dom}(\mathcal{F}) Dom ( F ) , which proves (157) . □
References
(Cao) School of Mathematical Sciences, Peking University, Beijing 100871, China Email address: chyyy@stu.pku.edu.cn
(Shi) School of Mathematics, Sichuan University, Chengdu 610064, China Email address: yunfengshi@scu.edu.cn
(Wang) School of Mathematics, Sichuan University, Chengdu 610064, China Email address: pzw@scu.edu.cn