Introduction
The extended Harper model describes an electron on a square lattice in a uniform perpendicular magnetic field, with hopping between nearest and next-nearest neighbors. It is thus a natural next model after the Harper model in the study of magnetic lattice spectra [14 ] . The latter goes back to the work of Peierls and Harper [28 , 17 ] , and its dependence on the magnetic flux is beautifully depicted by the Hofstadter butterfly [18 ] . Choosing the Landau gauge and taking a partial Fourier transform in one lattice direction gives a family of one-dimensional quasiperiodic Jacobi operators, also referred to as the extended Harper model (EHM). Writing 2 π α 2\pi\alpha 2 π α for the flux through a square and θ ∈ T = R / Z \theta\in\mathbb{T} = \mathbb{R}/\mathbb{Z} θ ∈ T = R / Z for the transverse quasimomentum, for λ = ( λ 1 , λ 2 , λ 3 ) ∈ [ 0 , ∞ ) 3 \lambda= (\lambda_1,\lambda_2,\lambda_3) \in[0,\infty)^3 λ = ( λ 1 , λ 2 , λ 3 ) ∈ [ 0 , ∞ ) 3 the extended Harper operator on ℓ 2 ( Z ) \ell^2(\mathbb{Z}) ℓ 2 ( Z ) is
( H α , θ , λ u ) m = c λ , α ( θ + ( m − 1 ) α ) u m − 1 + c λ , α ( θ + m α ) u m + 1 + 2 cos ( 2 π ( θ + m α ) ) u m , (H_{\alpha,\theta,\lambda}u)_m = c_{\lambda,\alpha}(\theta+ (m-1)\alpha)u_{m-1} + c_{\lambda,\alpha}(\theta+ m\alpha)u_{m+1}
+ 2\cos(2\pi(\theta+ m\alpha))u_m, ( H α , θ , λ u ) m = c λ , α ( θ + ( m − 1 ) α ) u m − 1 + c λ , α ( θ + m α ) u m + 1 + 2 cos ( 2 π ( θ + m α )) u m ,
c λ , α ( t ) = λ 1 e − 2 π i ( t + α / 2 ) + λ 2 + λ 3 e 2 π i ( t + α / 2 ) . (1) c_{\lambda,\alpha}(t) = \lambda_1 e^{-2\pi i(t+\alpha/2)} + \lambda_2 + \lambda_3 e^{2\pi i(t+\alpha/2)}. \tag*{(1)} c λ , α ( t ) = λ 1 e − 2 π i ( t + α /2 ) + λ 2 + λ 3 e 2 π i ( t + α /2 ) . (1)
For irrational α \alpha α , its spectrum is independent of θ \theta θ ; we denote it by Σ α , λ \Sigma_{\alpha,\lambda} Σ α , λ . For λ = ( 0 , λ 2 , 0 ) \lambda= (0,\lambda_2,0) λ = ( 0 , λ 2 , 0 ) , with λ 2 > 0 \lambda_2 > 0 λ 2 > 0 , EHM becomes the rescaled almost Mathieu operator, which arises by allowing only nearest-neighbor hopping in the two-dimensional magnetic Laplacian and which features a basic phase transition at λ 2 = 1 \lambda_2 = 1 λ 2 = 1 . Allowing the additional hoppings leads to a considerably richer phase diagram (cf. Figure 1 ), with several critical boundaries between regions I, II, and III, and the entire region III being self-dual; see [27 , 14 , 2 ] . The triangular lattice is also a special case of extended Harper [19 ] .
Figure 1. Left: the EHM regions projected onto ( Λ , λ 2 ) (\Lambda,\lambda_2) ( Λ , λ 2 ) , where Λ = λ 1 + λ 3 \Lambda= \lambda_1 + \lambda_3 Λ = λ 1 + λ 3 . The blue boundary is Λ = max { 1 , λ 2 } \Lambda= \max\{1,\lambda_2\} Λ = max { 1 , λ 2 } . The purple shading is the projection of I i s o \mathcal{I}_{\mathrm{iso}} I iso ; in the interior of region III, only couplings with λ 1 = λ 3 \lambda_1 = \lambda_3 λ 1 = λ 3 belong to C 0 \mathcal{C}_0 C 0 . Right: projection onto λ 2 = 1 \lambda_2 = 1 λ 2 = 1 . The purple ray λ 1 = λ 3 ≥ 1 / 2 \lambda_1 = \lambda_3 \ge1/2 λ 1 = λ 3 ≥ 1/2 is the slice of I i s o \mathcal{I}_{\mathrm{iso}} I iso . The entire shaded triangle L I I ⊂ C 0 \mathcal{L}_{\mathrm{II}} \subset\mathcal{C}_0 L II ⊂ C 0 has Hausdorff dimension of the spectrum at most 1 / 2 1/2 1/2 for every irrational frequency. At the blue midpoint ( 1 / 2 , 1 , 1 / 2 ) ∈ B i s o (1/2,1,1/2) \in\mathcal{B}_{\mathrm{iso}} ( 1/2 , 1 , 1/2 ) ∈ B iso , the bound improves to 1 / 3 1/3 1/3 . Red circles mark the critical AMO; green squares mark the triangular operator.
In mathematics, spectral theory of the extended Harper model has been a popular domain for the study of validity and limitations of arguments originally developed for the almost Mathieu family, e.g. [20 , 21 , 22 , 15 ] . The main interest in this regard comes from two aspects that differ significantly from almost Mathieu: the nontrivial self-dual region III, where the usual duality between supercritical and subcritical regimes is not available, and the set of parameters where the corresponding Jacobi matrix becomes singular, that is c λ , α ( θ ) c_{\lambda,\alpha}(\theta) c λ , α ( θ ) has a zero on a period, limiting applicability of methods based on nonsingularity.
The phase diagram of spectral types is by now well understood [2 ] . Various questions, in particular, about the geometry of the spectrum, however, remain.
In particular, fractal dimensions of the spectrum have been largely unstudied. The question, of course, only makes sense when the spectrum has measure zero, a regime, that has remained rather mysterious, being not amenable to either reducibility or localization methods. For any irrational α \alpha α , a direct corollary of Proposition 3.1, Theorems 3.4, 3.5, and 4.1 of [2 ] and Aubry duality is that, writing Λ = λ 1 + λ 3 \Lambda= \lambda_1 + \lambda_3 Λ = λ 1 + λ 3 ,
∣ Σ α , λ ∣ = 0 ⟺ λ ∈ C 0 = { λ : Λ = 1 ≥ λ 2 } ∪ { λ : λ 2 = 1 ≥ Λ } ∪ { λ : Λ = λ 2 ≥ 1 } ∪ { λ : λ 1 = λ 3 , Λ ≥ max { 1 , λ 2 } } . \lvert\Sigma_{\alpha,\lambda}\rvert= 0 \quad\Longleftrightarrow\quad\lambda\in\mathcal{C}_0 = \{\lambda: \Lambda= 1 \ge\lambda_2\}
\cup\{\lambda: \lambda_2 = 1 \ge\Lambda\}
\cup\{\lambda: \Lambda= \lambda_2 \ge1\}
\cup\{\lambda: \lambda_1 = \lambda_3,\ \Lambda\ge\max\{1,\lambda_2\}\}. ∣ Σ α , λ ∣ = 0 ⟺ λ ∈ C 0 = { λ : Λ = 1 ≥ λ 2 } ∪ { λ : λ 2 = 1 ≥ Λ } ∪ { λ : Λ = λ 2 ≥ 1 } ∪ { λ : λ 1 = λ 3 , Λ ≥ max { 1 , λ 2 }} .
The first three regions, denoted by L I \mathcal{L}_{\mathrm{I}} L I , L I I \mathcal{L}_{\mathrm{II}} L II , and L I I I \mathcal{L}_{\mathrm{III}} L III , respectively, are precisely the critical boundaries where at least two of the three quantities 1 1 1 , λ 2 \lambda_2 λ 2 , and Λ = λ 1 + λ 3 \Lambda= \lambda_1 + \lambda_3 Λ = λ 1 + λ 3 attain the max ( 1 , λ 2 , Λ ) \max(1,\lambda_2,\Lambda) max ( 1 , λ 2 , Λ ) . The fourth, denoted by I i s o \mathcal{I}_{\mathrm{iso}} I iso is the isotropic1 part of the interior of region III.
We distinguish the isotropic critical boundary
B i s o = { λ : λ 1 = λ 3 , λ 1 + λ 3 = max { 1 , λ 2 } } , \mathcal{B}_{\mathrm{iso}} = \{\lambda: \lambda_1 = \lambda_3,\ \lambda_1 + \lambda_3 = \max\{1,\lambda_2\}\}, B iso = { λ : λ 1 = λ 3 , λ 1 + λ 3 = max { 1 , λ 2 }} ,
which is the union of intersections I i s o ∩ L j \mathcal{I}_{\mathrm{iso}} \cap\mathcal{L}_j I iso ∩ L j , j = I , I I , I I I j = \mathrm{I},\mathrm{II},\mathrm{III} j = I , II , III .
For the almost Mathieu family, λ = ( 0 , λ 2 , 0 ) \lambda= (0,\lambda_2,0) λ = ( 0 , λ 2 , 0 ) , there is only one parameter with zero measure spectrum: λ 2 = 1 \lambda_2 = 1 λ 2 = 1 . Thouless conjectured that Hausdorff dimension is bounded by 1 / 2 1/2 1/2 in this case, for all irrational α \alpha α .
In [5 ] we developed a method to find tight spectral covers for singular Jacobi matrices, that lead to dimension bounds. Historically, singularity has been an obstacle to overcome when extending various almost Mathieu arguments [21 , 22 , 2 , 16 ] . In contrast, in [5 ] we exploit singularity: while critical almost Mathieu is a Schrödinger operator, so certainly not a singular Jacobi matrix, we found what we call a chiral transform that gives it an isospectral singular Jacobi representation, leading to a proof of the Thouless conjecture for every irrational frequency.
Here we show that the method of [5 ] can be extended to the entire zero-measure regime of the EHM, giving the same bound 1 / 2 1/2 1/2 without arithmetic restrictions on α \alpha α . Moreover, on the isotropic critical boundary the bound improves to 1 / 3 1/3 1/3 .
Operators in L I I I \mathcal{L}_{\mathrm{III}} L III and I i s o \mathcal{I}_{\mathrm{iso}} I iso are already singular, and operators in L I \mathcal{L}_{\mathrm{I}} L I are made singular by the standard Aubry duality. For those, the problem reduces to finding good spectral covers.
The region L I I \mathcal{L}_{\mathrm{II}} L II , however, is more challenging. It is the self-dual triangle (see Figure 1 ), that contains critical almost Mathieu and the isotropic triangular model as its boundary points, at ( 0 , 1 , 0 ) (0,1,0) ( 0 , 1 , 0 ) and ( 1 , 1 , 0 ) (1,1,0) ( 1 , 1 , 0 ) , respectively. The additional challenge is to unlock its hidden singularity. We do it by finding a family of transforms that singularize each H α , θ , λ H_{\alpha,\theta,\lambda} H α , θ , λ with λ ∈ L I I \lambda\in\mathcal{L}_{\mathrm{II}} λ ∈ L II . The chiral transform of [5 ] is one element in this family.
Intuitively, the physical question of what survives when additional hopping is allowed is related to the recently established analytic robustness of the almost Mathieu results [13 , 11 , 26 ] , although we note that criticality itself is not robust [1 ] . Our results are global and complete in the parameter class considered here, with no smallness assumptions or other restrictions.
Hausdorff dimension
We write m e a s t \mathrm{meas}_{t} meas t for t t t -dimensional Hausdorff measure, dim H \dim_{\mathrm H} dim H for Hausdorff dimension, and dim ‾ B \underline{\dim}_{\mathrm B} dim B and dim ‾ B \overline{\dim}_{\mathrm B} dim B for lower and upper box dimensions.
Theorem 1.1. For every irrational α \alpha α and every λ ∈ C 0 \lambda\in\mathcal{C}_{0} λ ∈ C 0 ,
m e a s 1 / 2 ( Σ α , λ ) < ∞ , dim H Σ α , λ ≤ dim ‾ B Σ α , λ ≤ 1 2 . \mathrm{meas}_{1/2}(\Sigma_{\alpha,\lambda}) < \infty,\qquad\dim_{\mathrm H}\Sigma_{\alpha,\lambda} \le\underline{\dim}_{\mathrm B}\Sigma_{\alpha,\lambda} \le\frac{1}{2}. meas 1/2 ( Σ α , λ ) < ∞ , dim H Σ α , λ ≤ dim B Σ α , λ ≤ 2 1 . For λ ∈ B i s o \lambda\in\mathcal{B}_{\mathrm{iso}} λ ∈ B iso , these conclusions improve to
m e a s 1 / 3 ( Σ α , λ ) < ∞ , dim H Σ α , λ ≤ dim ‾ B Σ α , λ ≤ 1 3 . \mathrm{meas}_{1/3}(\Sigma_{\alpha,\lambda}) < \infty,\qquad\dim_{\mathrm H}\Sigma_{\alpha,\lambda} \le\underline{\dim}_{\mathrm B}\Sigma_{\alpha,\lambda} \le\frac{1}{3}. meas 1/3 ( Σ α , λ ) < ∞ , dim H Σ α , λ ≤ dim B Σ α , λ ≤ 3 1 .
The almost Mathieu results of Theorem 1.1 provide a strengthening of the Hausdorff dimension bound of [5 ] , following from the covers proved there in Theorem 7.11. The 1 / 3 1/3 1/3 bound concerns the bicritical boundary studied in [14 ] , including the point ( 1 / 2 , 1 , 1 / 2 ) (1/2,1,1/2) ( 1/2 , 1 , 1/2 ) . A different scaling at this boundary was already apparent in the rational bandwidth analysis of [14 ] ; here it leads to a dimension bound for every irrational frequency. In our argument, it is the higher-order zero of the hopping coefficient that produces the improvement.
Box-counting dimension
Of course, a bound on Hausdorff dimension does not in general control box-counting dimension. Already for critical almost Mathieu, the Hausdorff dimension can be zero [25 , 3 ] while the packing dimension is one [23 ] . Our argument also gives bounds on upper box dimension in terms of the arithmetic properties of the frequency.
Let p n / q n p_n/q_n p n / q n be the canonical approximants of α \alpha α , and set
τ ( α ) = lim sup n ln q n + 1 ln q n ∈ [ 1 , ∞ ] , β ( α ) = lim sup n ln q n + 1 q n ∈ [ 0 , ∞ ] . (2) \tau(\alpha)=\limsup_{n}\frac{\ln q_{n+1}}{\ln q_n}\in[1,\infty],\qquad\beta(\alpha)=\limsup_{n}\frac{\ln q_{n+1}}{q_n}\in[0,\infty]. \tag*{(2)} τ ( α ) = n lim sup ln q n ln q n + 1 ∈ [ 1 , ∞ ] , β ( α ) = n lim sup q n ln q n + 1 ∈ [ 0 , ∞ ] . (2)
For almost every α \alpha α , τ ( α ) = 1 \tau(\alpha)=1 τ ( α ) = 1 and β ( α ) = 0 \beta(\alpha)=0 β ( α ) = 0 . We call α \alpha α weakly Liouville when β ( α ) > 0 \beta(\alpha)>0 β ( α ) > 0 (and therefore τ ( α ) = ∞ \tau(\alpha)=\infty τ ( α ) = ∞ ).
Theorem 1.2. For every irrational α \alpha α and every λ ∈ C 0 \lambda\in\mathcal{C}_{0} λ ∈ C 0 ,
dim ‾ P Σ α , λ ≤ dim ‾ B Σ α , λ ≤ τ ( α ) 1 + τ ( α ) . \underline{\dim}_{\mathrm P}\Sigma_{\alpha,\lambda}\le\overline{\dim}_{\mathrm B}\Sigma_{\alpha,\lambda}\le\frac{\tau(\alpha)}{1+\tau(\alpha)}. dim P Σ α , λ ≤ dim B Σ α , λ ≤ 1 + τ ( α ) τ ( α ) . For λ ∈ B i s o \lambda\in\mathcal{B}_{\mathrm{iso}} λ ∈ B iso , the last bound improves to
dim ‾ B Σ α , λ ≤ τ ( α ) 2 + τ ( α ) . \overline{\dim}_{\mathrm B}\Sigma_{\alpha,\lambda}\le\frac{\tau(\alpha)}{2+\tau(\alpha)}. dim B Σ α , λ ≤ 2 + τ ( α ) τ ( α ) .
Here dim P \dim_{\mathrm{P}} dim P denotes packing dimension. When τ ( α ) = 1 \tau(\alpha)=1 τ ( α ) = 1 , which holds for Lebesgue almost every α \alpha α , these are upper box-dimension bounds of 1 / 2 1/2 1/2 and 1 / 3 1/3 1/3 , respectively.
For weakly Liouville frequencies the situation is quite different. Combining our bounds on lower box dimension with the packing estimates of [23 , 16 ] gives
Corollary 1.3. Suppose β ( α ) > 0 \beta(\alpha)>0 β ( α ) > 0 , and
λ ∈ { ( 0 , 1 , 0 ) , ( 1 , 1 , 0 ) } \lambda\in\{(0,1,0),(1,1,0)\} λ ∈ {( 0 , 1 , 0 ) , ( 1 , 1 , 0 )} or λ ∈ I i s o ∪ B i s o \lambda\in\mathcal{I}_{\mathrm{iso}}\cup\mathcal{B}_{\mathrm{iso}} λ ∈ I iso ∪ B iso , λ 2 > 0 \lambda_2>0 λ 2 > 0 .
Then
dim ‾ B Σ α , λ ≤ 1 2 < 1 = dim P Σ α , λ = dim ‾ B Σ α , λ . (3) \underline{\dim}_{\mathrm{B}}\Sigma_{\alpha,\lambda}\leq\frac{1}{2}<1=\dim_{\mathrm{P}}\Sigma_{\alpha,\lambda}=\overline{\dim}_{\mathrm{B}}\Sigma_{\alpha,\lambda}. \tag*{(3)} dim B Σ α , λ ≤ 2 1 < 1 = dim P Σ α , λ = dim B Σ α , λ . (3) For λ ∈ B i s o \lambda\in\mathcal{B}_{\mathrm{iso}} λ ∈ B iso with λ 2 > 0 \lambda_2>0 λ 2 > 0 , the lower box bound improves to 1 / 3 1/3 1/3 . Thus box-counting dimension does not exist in all these cases, including the critical almost Mathieu and triangular models.
This distinction is important when interpreting the dimension suggested by rational approximants or numerical calculations. The small dimension seen along one sequence of scales need not be the box-counting dimension of the spectrum: the limit defining the latter may simply fail to exist.
In Section 2 we find covers by intervals for the spectrum Σ α , λ \Sigma_{\alpha,\lambda} Σ α , λ when α \alpha α is irrational and λ ∈ C 0 \lambda\in\mathcal{C}_0 λ ∈ C 0 . Following the approach of [5 ] , we first reduce EHM for λ ∈ C 0 \lambda\in\mathcal{C}_0 λ ∈ C 0 to singular Jacobi form. Then we need to consider blocks obtained from the Jacobi matrix by replacing by zero hopping coefficients where they are small. The blocks B ( x ) B(x) B ( x ) so obtained will have two dimensions, q n q_n q n and q n − 1 q_{n-1} q n − 1 and will depend on a phase parameter x x x which ranges over a small interval of length of order 1 / q n 1/q_n 1/ q n (with removed hopping coefficients of order 1 / q n 2 1/q_n^2 1/ q n 2 for λ ∈ B i s o \lambda\in\mathcal{B}_{\mathrm{iso}} λ ∈ B iso ). The spectral cover we seek will be a cover of the union of spectra of these blocks with added rank 2 small-norm perturbations over all such perturbations and over x x x . To apply the method of [5 ] , we need to find a matrix K K K such that
E : = d d x B ( x ) − [ K ( x ) , B ( x ) ] \mathcal{E}:=\frac{d}{dx}B(x)-[K(x),B(x)] E := d x d B ( x ) − [ K ( x ) , B ( x )]
has a bounded rank and norm. We find two such matrices: one for λ ∈ L I ∪ L I I ∪ L I I I \lambda\in\mathcal{L}_{I}\cup\mathcal{L}_{II}\cup\mathcal{L}_{III} λ ∈ L I ∪ L I I ∪ L I I I , and another for λ ∈ I i s o \lambda\in\mathcal{I}_{\mathrm{iso}} λ ∈ I iso . The required spectral covers of at most q n + q n − 1 q_n+q_{n-1} q n + q n − 1 intervals of total length O ( 1 / q n ) O(1/q_n) O ( 1/ q n ) (improving to O ( 1 / q n 2 ) O(1/q_n^2) O ( 1/ q n 2 ) for λ ∈ B i s o \lambda\in\mathcal{B}_{\mathrm{iso}} λ ∈ B iso ) then follow. In Section 3 we use these covers to prove the theorems stated above.
Similar results on fractal dimensions hold for the quantum graph graphene [6 ] .
Acknowledgement and AI statement: The research of SJ was supported by NSF DMS-2052899, DMS2155211, and Simons 896624. We used AI to accelerate determination of the explicit formulae for matrices K K K , and to help find the transform of Lemma 2.3. These constructions are ideologically close to those of [5 ] . We also used AI to assist in preparing an initial draft of this manuscript, which the authors subsequently rewrote in full.
Spectral covers for irrational flux
We use the direct-integral notation
M α , λ = ∫ T ⊕ H α , θ , λ d θ , Σ α , λ = spec ( M α , λ ) . M_{\alpha,\lambda}=\int_{\mathbb{T}}^{\oplus}H_{\alpha,\theta,\lambda}\,d\theta,\qquad\Sigma_{\alpha,\lambda}=\operatorname{spec}(M_{\alpha,\lambda}). M α , λ = ∫ T ⊕ H α , θ , λ d θ , Σ α , λ = spec ( M α , λ ) .
For irrational α \alpha α , Σ α , λ \Sigma_{\alpha,\lambda} Σ α , λ coincides with the spectrum of every fiber [4 ] .
In this section we will prove the following:
Theorem 2.1. Let α \alpha α be irrational with canonical approximants p n / q n p_n/q_n p n / q n . For λ ∈ C 0 \lambda\in\mathcal{C}_0 λ ∈ C 0 and every n ≥ 4 n\geq4 n ≥ 4 , there is a cover of Σ α , λ \Sigma_{\alpha,\lambda} Σ α , λ by closed intervals I n , j I_{n,j} I n , j such that
Σ α , λ ⊂ ⋃ j = 1 M n I n , j , M n ≤ q n + q n − 1 , ∑ j ∣ I n , j ∣ ≤ C max { 1 , λ 1 + λ 3 } q n . (4) \Sigma_{\alpha,\lambda}\subset\bigcup_{j=1}^{M_n}I_{n,j},\qquad M_n\leq q_n+q_{n-1},\qquad\sum_j|I_{n,j}|\leq\frac{C\max\{1,\lambda_1+\lambda_3\}}{q_n}. \tag*{(4)} Σ α , λ ⊂ j = 1 ⋃ M n I n , j , M n ≤ q n + q n − 1 , j ∑ ∣ I n , j ∣ ≤ q n C max { 1 , λ 1 + λ 3 } . (4) If λ ∈ L I , L I I I \lambda\in\mathcal{L}_{I},\mathcal{L}_{III} λ ∈ L I , L I I I , a cover with the same interval count satisfies
∑ j ∣ I n , j ∣ ≤ C ( ∣ λ 1 − λ 3 ∣ q n + λ 1 + λ 3 q n 2 ) . (5) \sum_j|I_{n,j}|\leq C\left(\frac{|\lambda_1-\lambda_3|}{q_n}+\frac{\lambda_1+\lambda_3}{q_n^2}\right). \tag*{(5)} j ∑ ∣ I n , j ∣ ≤ C ( q n ∣ λ 1 − λ 3 ∣ + q n 2 λ 1 + λ 3 ) . (5)
If λ = ( λ 1 , 1 , λ 3 ) ∈ L I I \lambda=(\lambda_1,1,\lambda_3)\in\mathcal{L}_{\mathrm{II}} λ = ( λ 1 , 1 , λ 3 ) ∈ L II , the same interval count holds and
∑ j ∣ I n , j ∣ ≤ C ( 1 − 4 λ 1 λ 3 q n + 1 q n 2 ) . (6) \sum_j |I_{n,j}| \le C\left(\frac{\sqrt{1-4\lambda_1\lambda_3}}{q_n}+\frac{1}{q_n^2}\right). \tag*{(6)} j ∑ ∣ I n , j ∣ ≤ C ( q n 1 − 4 λ 1 λ 3 + q n 2 1 ) . (6)
The constant C > 0 C>0 C > 0 above is absolute.
The Hausdorff-measure bounds of Theorem 1.1 now follow immediately from Hölder’s inequality. Indeed, if M n = O ( q n ) M_n=O(q_n) M n = O ( q n ) and ∑ j ∣ I n , j ∣ = O ( q n − ϰ ) \sum_j|I_{n,j}|=O(q_n^{-\varkappa}) ∑ j ∣ I n , j ∣ = O ( q n − ϰ ) , then
∑ j ∣ I n , j ∣ 1 / ( 1 + ϰ ) ≤ M n ϰ / ( 1 + ϰ ) ( ∑ j ∣ I n , j ∣ ) 1 / ( 1 + ϰ ) = O ( 1 ) . \sum_j |I_{n,j}|^{1/(1+\varkappa)}\le M_n^{\varkappa/(1+\varkappa)}\left(\sum_j|I_{n,j}|\right)^{1/(1+\varkappa)}=O(1). j ∑ ∣ I n , j ∣ 1/ ( 1 + ϰ ) ≤ M n ϰ / ( 1 + ϰ ) ( j ∑ ∣ I n , j ∣ ) 1/ ( 1 + ϰ ) = O ( 1 ) .
The largest interval length tends to zero, so ϰ = 1 , 2 \varkappa=1,2 ϰ = 1 , 2 give finite Hausdorff measure at exponents 1 / 2 , 1 / 3 1/2,1/3 1/2 , 1/3 , respectively. The box-dimension bounds follow from Proposition 3.1 below.
For the square and triangular lattice models we use the notation
λ A = ( 0 , 1 , 0 ) , λ T = ( 1 , 1 , 0 ) , Σ α A = Σ α , λ A , Σ α T = Σ α , λ T . (7) \lambda_{\mathrm{A}}=(0,1,0),\qquad\lambda_{\mathrm{T}}=(1,1,0),\qquad\Sigma^\mathrm{A}_\alpha=\Sigma_{\alpha,\lambda_{\mathrm{A}}},\qquad\Sigma^\mathrm{T}_\alpha=\Sigma_{\alpha,\lambda_{\mathrm{T}}}. \tag*{(7)} λ A = ( 0 , 1 , 0 ) , λ T = ( 1 , 1 , 0 ) , Σ α A = Σ α , λ A , Σ α T = Σ α , λ T . (7)
In particular, H α , θ , λ A H_{\alpha,\theta,\lambda_{\mathrm{A}}} H α , θ , λ A is the critical almost Mathieu operator, and M α , λ A = M α M_{\alpha,\lambda_{\mathrm{A}}}=M_\alpha M α , λ A = M α in the notation of [5 ] .
Reduction to singular Jacobi families
Let J J J be the bounded Jacobi operator
( J u ) j = b ‾ j − 1 u j − 1 + v j u j + b j u j + 1 , v j ∈ R , b j ∈ C . (Ju)_j=\overline{b}_{j-1}u_{j-1}+v_ju_j+b_ju_{j+1},\qquad v_j\in\mathbb{R},\quad b_j\in\mathbb{C}. ( J u ) j = b j − 1 u j − 1 + v j u j + b j u j + 1 , v j ∈ R , b j ∈ C .
For continuous sampling functions v : T → R v:\mathbb{T}\to\mathbb{R} v : T → R and b : T → C b:\mathbb{T}\to\mathbb{C} b : T → C , we set v j = v ( θ + j α ) v_j=v(\theta+j\alpha) v j = v ( θ + j α ) and b j = b ( θ + j α ) b_j=b(\theta+j\alpha) b j = b ( θ + j α ) . When b b b is real, J = H v , b , α , θ J=H_{v,b,\alpha,\theta} J = H v , b , α , θ of [5 ] . We call the family singular if b b b has a zero. For irrational α \alpha α , singularity gives
inf j ∈ Z ∣ b ( θ + j α ) ∣ = 0 \inf_{j\in\mathbb{Z}}|b(\theta+j\alpha)|=0 j ∈ Z inf ∣ b ( θ + j α ) ∣ = 0
for every θ \theta θ . Deleting the bond between j j j and j + 1 j+1 j + 1 changes the operator by
J − J c u t = b j e j e j + 1 ∗ + b ‾ j e j + 1 e j ∗ , ∥ J − J c u t ∥ = ∣ b j ∣ . J-J^{\mathrm{cut}}=b_je_je_{j+1}^*+\overline{b}_je_{j+1}e_j^*,\qquad\lVert J-J^{\mathrm{cut}}\rVert=|b_j|. J − J cut = b j e j e j + 1 ∗ + b j e j + 1 e j ∗ , ∥ J − J cut ∥ = ∣ b j ∣.
where e j e_j e j is the standard basis vector. As in [5 ] , we will use continued-fraction returns to choose cuts where these coefficients are small, and construct covers for the spectrum of perturbed blocks. Thus, we first need to see if the family (in θ \theta θ ) of H α , θ , λ H_{\alpha,\theta,\lambda} H α , θ , λ can be reduced to a family of singular Jacobi matrices.
As we discuss below the families for λ ∈ L I I I , I i s o \lambda\in\mathcal{L}_{\mathrm{III}},\mathcal{I}_{\mathrm{iso}} λ ∈ L III , I iso are already singular. We will now reduce the families for λ ∈ L I , L I I \lambda\in\mathcal{L}_{\mathrm{I}},\mathcal{L}_{\mathrm{II}} λ ∈ L I , L II to singular form by using the duality transform for λ ∈ L I \lambda\in\mathcal{L}_{\mathrm{I}} λ ∈ L I , and a chiral-type transform for λ ∈ L I I \lambda\in\mathcal{L}_{\mathrm{II}} λ ∈ L II .
Aubry duality exchanges the potential amplitude and the constant hopping. We therefore introduce a potential amplitude ν ≥ 0 \nu\geq0 ν ≥ 0 in (1):
( H ν , α , θ , λ u ) j = c λ , α ( θ + ( j − 1 ) α ) u j − 1 + c λ , α ( θ + j α ) u j + 1 + 2 ν cos ( 2 π ( θ + j α ) ) u j . (8) (H_{\nu,\alpha,\theta,\lambda}u)_j=c_{\lambda,\alpha}(\theta+(j-1)\alpha)u_{j-1}+c_{\lambda,\alpha}(\theta+j\alpha)u_{j+1}+2\nu\cos(2\pi(\theta+j\alpha))u_j. \tag*{(8)} ( H ν , α , θ , λ u ) j = c λ , α ( θ + ( j − 1 ) α ) u j − 1 + c λ , α ( θ + j α ) u j + 1 + 2 ν cos ( 2 π ( θ + j α )) u j . (8)
We write
S ν , λ ( α ) = ⋃ θ ∈ T spec ( H ν , α , θ , λ ) ) . S_{\nu,\lambda}(\alpha)=\bigcup_{\theta\in\mathbb{T}}\operatorname{spec}(H_{\nu,\alpha,\theta,\lambda)).} S ν , λ ( α ) = θ ∈ T ⋃ spec ( H ν , α , θ , λ )) .
By norm continuity in θ \theta θ , this is also the spectrum of the direct integral. For irrational α \alpha α , application of the unitary transform ( T u ) j = u j + 1 (Tu)_j=u_{j+1} ( T u ) j = u j + 1 shows that it is the spectrum of every fiber. Note that S 1 , λ ( α ) = Σ α , λ S_{1,\lambda}(\alpha)=\Sigma_{\alpha,\lambda} S 1 , λ ( α ) = Σ α , λ .
We first recall the usual Aubry duality, allowing also λ 2 = 0. \lambda_2=0. λ 2 = 0.
Lemma 2.2. For every α ∈ R \alpha\in\mathbb{R} α ∈ R , ν ≥ 0 \nu\ge0 ν ≥ 0 , and λ = ( λ 1 , λ 2 , λ 3 ) ∈ [ 0 , ∞ ) 3 \lambda= (\lambda_1,\lambda_2,\lambda_3) \in[0,\infty)^3 λ = ( λ 1 , λ 2 , λ 3 ) ∈ [ 0 , ∞ ) 3 , the families H ν , α , θ , λ H_{\nu,\alpha,\theta,\lambda} H ν , α , θ , λ and H λ 2 , α , θ , ( λ 3 , ν , λ 1 ) H_{\lambda_2,\alpha,\theta,(\lambda_3,\nu,\lambda_1)} H λ 2 , α , θ , ( λ 3 , ν , λ 1 ) have unitarily equivalent direct integrals. In particular,
S ν , ( λ 1 , λ 2 , λ 3 ) ( α ) = S λ 2 , ( λ 3 , ν , λ 1 ) ( α ) . (9) S_{\nu,(\lambda_1,\lambda_2,\lambda_3)}(\alpha)=S_{\lambda_2,(\lambda_3,\nu,\lambda_1)}(\alpha). \tag*{(9)} S ν , ( λ 1 , λ 2 , λ 3 ) ( α ) = S λ 2 , ( λ 3 , ν , λ 1 ) ( α ) . (9) For ν = 1 \nu=1 ν = 1 and λ 2 > 0 \lambda_2>0 λ 2 > 0 , setting λ ∨ = ( λ 3 / λ 2 , 1 / λ 2 , λ 1 / λ 2 ) \lambda^\vee=(\lambda_3/\lambda_2,1/\lambda_2,\lambda_1/\lambda_2) λ ∨ = ( λ 3 / λ 2 , 1/ λ 2 , λ 1 / λ 2 ) gives Σ α , λ = λ 2 Σ α , λ ∨ \Sigma_{\alpha,\lambda}=\lambda_2\Sigma_{\alpha,\lambda^\vee} Σ α , λ = λ 2 Σ α , λ ∨ .
Proof. This is the same Aubry duality calculation as in [12 Theorem 3.1] . On L 2 ( T ; ℓ 2 ( Z ) ) L^2(\mathbb{T};\ell^2(\mathbb{Z})) L 2 ( T ; ℓ 2 ( Z )) , let
( T φ ) ( θ , j ) = φ ( θ , j + 1 ) , ( S φ ) ( θ , j ) = e 2 π i ( θ + j α ) φ ( θ , j ) . (T\varphi)(\theta,j)=\varphi(\theta,j+1),\qquad(S\varphi)(\theta,j)=e^{2\pi i(\theta+j\alpha)}\varphi(\theta,j). ( T φ ) ( θ , j ) = φ ( θ , j + 1 ) , ( S φ ) ( θ , j ) = e 2 π i ( θ + j α ) φ ( θ , j ) .
These operators satisfy the commutation relation T S = e 2 π i α S T TS=e^{2\pi i\alpha}ST T S = e 2 π i α S T . Applying S − 1 S^{-1} S − 1 to it from both sides yields S − 1 T = e 2 π i α T S − 1 S^{-1}T=e^{2\pi i\alpha}TS^{-1} S − 1 T = e 2 π i α T S − 1 , and similarly applying T − 1 T^{-1} T − 1 we have S T − 1 = e 2 π i α T − 1 S ST^{-1}=e^{2\pi i\alpha}T^{-1}S S T − 1 = e 2 π i α T − 1 S .
The Fourier transform
( U φ ) ( θ , j ) = ∑ k ∈ Z e − 2 π i k ( θ + j α ) ∫ T e − 2 π i j x φ ( x , k ) d x (U\varphi)(\theta,j)=\sum_{k\in\mathbb{Z}}e^{-2\pi ik(\theta+j\alpha)}\int_{\mathbb{T}}e^{-2\pi ijx}\varphi(x,k)\,dx ( U φ ) ( θ , j ) = k ∈ Z ∑ e − 2 π ik ( θ + j α ) ∫ T e − 2 π ij x φ ( x , k ) d x
is unitary, with the sum interpreted in L 2 L^2 L 2 , and satisfies U S = T − 1 U US=T^{-1}U U S = T − 1 U and U T = S U UT=SU U T = S U (and consequently, by taking inverses and applying U U U , U S − 1 = T U US^{-1}=TU U S − 1 = T U , U T − 1 = S − 1 U UT^{-1}=S^{-1}U U T − 1 = S − 1 U ) exactly as in that proof.
We now write (8) in terms of S S S and T T T ,
We have
H ν , α , θ , λ = ν ( S + S − 1 ) + λ 2 ( T + T − 1 ) + λ 1 e − π i α ( S − 1 T + S T − 1 ) + λ 3 e π i α ( S T + S − 1 T − 1 ) . H_{\nu,\alpha,\theta,\lambda}=\nu(S+S^{-1})+\lambda_2(T+T^{-1})
+\lambda_1e^{-\pi i\alpha}(S^{-1}T+ST^{-1})+\lambda_3e^{\pi i\alpha}(ST+S^{-1}T^{-1}). H ν , α , θ , λ = ν ( S + S − 1 ) + λ 2 ( T + T − 1 ) + λ 1 e − π i α ( S − 1 T + S T − 1 ) + λ 3 e π i α ( S T + S − 1 T − 1 ) .
Let Q ν , λ = ∫ T ⊕ H ν , α , θ , λ d θ Q_{\nu,\lambda}=\int_{\mathbb{T}}^\oplus H_{\nu,\alpha,\theta,\lambda}\,d\theta Q ν , λ = ∫ T ⊕ H ν , α , θ , λ d θ . The above commutation relations then give
U Q ν , λ = Q λ 2 , ( λ 3 , ν , λ 1 ) U , UQ_{\nu,\lambda}=Q_{\lambda_2,(\lambda_3,\nu,\lambda_1)}U, U Q ν , λ = Q λ 2 , ( λ 3 , ν , λ 1 ) U ,
which proves the claim. The relation Σ α , λ = λ 2 Σ α , λ ∨ \Sigma_{\alpha,\lambda}=\lambda_2\Sigma_{\alpha,\lambda^\vee} Σ α , λ = λ 2 Σ α , λ ∨ for ν = 1 \nu=1 ν = 1 , λ 2 > 0 \lambda_2>0 λ 2 > 0 , follows by factoring out λ 2 \lambda_2 λ 2 .
As we discuss below, if λ ∈ L I \lambda\in\mathcal{L}_I λ ∈ L I , the family for λ ∨ \lambda^\vee λ ∨ is singular.
The remaining step is to make H H H for λ \lambda λ in the self-dual triangle (see Figure 1 ) of L I I \mathcal{L}_{II} L I I singular. We use a diagonal phase gauge, which allows to replace a hopping by any other hopping with the same modulus; see [2 Section 7, footnote 9] and [29 (1.57) and Lemma 1.6] . The identity below is a simple instance of reflecting a polynomial factor across the unit circle [7 Section 3] . Combined with Aubry duality, it gives the desired singular representative.
Lemma 2.3. Let λ = ( a , 1 , b ) ∈ L I I \lambda=(a,1,b)\in\mathcal{L}_{II} λ = ( a , 1 , b ) ∈ L I I , and set
Λ = a + b , r ± = 1 ± 1 − 4 a b 2 . (10) \Lambda=a+b,\qquad r_\pm=\frac{1\pm\sqrt{1-4ab}}{2}. \tag*{(10)} Λ = a + b , r ± = 2 1 ± 1 − 4 ab . (10) For every α , θ \alpha,\theta α , θ , there is a diagonal unitary G α , θ G_{\alpha,\theta} G α , θ such that
G α , θ ∗ H α , θ , ( a , 1 , b ) G α , θ = H α , θ , ( r + , Λ , r − ) . (11) G_{\alpha,\theta}^*H_{\alpha,\theta,(a,1,b)}G_{\alpha,\theta}=H_{\alpha,\theta,(r_+,\Lambda,r_-)}. \tag*{(11)} G α , θ ∗ H α , θ , ( a , 1 , b ) G α , θ = H α , θ , ( r + , Λ , r − ) . (11) Proof. Since r + + r − = 1 r_++r_-=1 r + + r − = 1 and r + r − = a b r_+r_-=ab r + r − = ab , for ∣ z ∣ = 1 |z|=1 ∣ z ∣ = 1 ,
∣ a z − 1 + 1 + b z ∣ 2 = 1 + a 2 + b 2 + 2 Λ Re z + 2 a b Re ( z 2 ) = ∣ r + z − 1 + Λ + r − z ∣ 2 . |az^{-1}+1+bz|^2=1+a^2+b^2+2\Lambda\operatorname{Re}z+2ab\operatorname{Re}(z^2)
=|r_+z^{-1}+\Lambda+r_-z|^2. ∣ a z − 1 + 1 + b z ∣ 2 = 1 + a 2 + b 2 + 2Λ Re z + 2 ab Re ( z 2 ) = ∣ r + z − 1 + Λ + r − z ∣ 2 . Therefore, the hopping moduli coincide: ∣ c λ , α ( θ + j α ) ∣ = ∣ c ( r + , Λ , r − ) , α ( θ + j α ) ∣ |c_{\lambda,\alpha}(\theta+j\alpha)|=|c_{(r_+,\Lambda,r_-),\alpha}(\theta+j\alpha)| ∣ c λ , α ( θ + j α ) ∣ = ∣ c ( r + , Λ , r − ) , α ( θ + j α ) ∣ . If this modulus is nonzero, denote c j ( 1 ) = exp { i arg c λ , α ( θ + j α ) } c_j^{(1)}=\exp\{i\arg c_{\lambda,\alpha}(\theta+j\alpha)\} c j ( 1 ) = exp { i arg c λ , α ( θ + j α )} , c j ( 2 ) = exp { i arg c ( r + , Λ , r − ) , α ( θ + j α ) } c_j^{(2)}=\exp\{i\arg c_{(r_+,\Lambda,r_-),\alpha}(\theta+j\alpha)\} c j ( 2 ) = exp { i arg c ( r + , Λ , r − ) , α ( θ + j α )} , otherwise set c j ( 1 ) = c j ( 2 ) = 1 c_j^{(1)}=c_j^{(2)}=1 c j ( 1 ) = c j ( 2 ) = 1 . Let
g j + 1 = g j c j ( 2 ) c j ( 1 ) , j ∈ Z ; g 0 = 1. (12) g_{j+1}=g_j\frac{c_j^{(2)}}{c_j^{(1)}},\qquad j\in\mathbb{Z};\qquad g_0=1. \tag*{(12)} g j + 1 = g j c j ( 1 ) c j ( 2 ) , j ∈ Z ; g 0 = 1. (12) Then ( G α , θ u ) j = g j u j (G_{\alpha,\theta}u)_j=g_ju_j ( G α , θ u ) j = g j u j gives (11). The gauges are measurable in θ \theta θ , so they also give unitary equivalence of the direct integrals.
Theorem 2.4. With the notation of (8) and Lemma 2.3, for every α ∈ R \alpha\in\mathbb{R} α ∈ R and λ ∈ L I I \lambda\in\mathcal{L}_{\mathrm{II}} λ ∈ L II , M α , λ M_{\alpha,\lambda} M α , λ is unitarily equivalent to the direct integral of a singular Jacobi family H L , α , θ , ( r − , 1 , r + ) H_{\mathcal{L},\alpha,\theta,(r_{-},1,r_{+})} H L , α , θ , ( r − , 1 , r + ) .
Proof. Apply Lemma 2.3, followed by the Aubry duality (Lemma 2.2). The resulting family is H L , α , θ , ( r − , 1 , r + ) H_{\mathcal{L},\alpha,\theta,(r_{-},1,r_{+})} H L , α , θ , ( r − , 1 , r + ) , whose hopping vanishes at t = 1 / 2 − α / 2 t=1/2-\alpha/2 t = 1/2 − α /2 , since r − + r + = 1 r_{-}+r_{+}=1 r − + r + = 1 .
Corollary 2.5. For every α ∈ R \alpha\in\mathbb{R} α ∈ R and λ ∈ C 0 \lambda\in\mathcal{C}_{0} λ ∈ C 0 , M α , λ M_{\alpha,\lambda} M α , λ is unitarily equivalent to the direct integral of the following singular Jacobi family at frequency α \alpha α with b ( 0 ) = 0 b(0)=0 b ( 0 ) = 0 :
( H v , b , α , θ φ ) ( n ) = b ( θ + ( n − 1 ) α ) φ ( n − 1 ) + b ( θ + n α ) φ ( n + 1 ) + v ( θ + n α ) φ ( n ) , (13) (H_{v,b,\alpha,\theta}\varphi)(n)=b(\theta+(n-1)\alpha)\varphi(n-1)+b(\theta+n\alpha)\varphi(n+1)+v(\theta+n\alpha)\varphi(n), \tag*{(13)} ( H v , b , α , θ φ ) ( n ) = b ( θ + ( n − 1 ) α ) φ ( n − 1 ) + b ( θ + n α ) φ ( n + 1 ) + v ( θ + n α ) φ ( n ) , (13) where for λ ∈ L I , L I I , L I I I \lambda\in\mathcal{L}_{\mathrm{I}},\mathcal{L}_{\mathrm{II}},\mathcal{L}_{\mathrm{III}} λ ∈ L I , L II , L III ,
v ( x ) = − 2 ν cos ( 2 π x − π α ) , b ( x ) = 2 w ( π x ) sin ( π x ) , w ( t ) = s 2 sin 2 t + d 2 cos 2 t , (14) v(x)=-2\nu\cos(2\pi x-\pi\alpha),\qquad b(x)=2w(\pi x)\sin(\pi x),\qquad w(t)=\sqrt{s^{2}\sin^{2}t+d^{2}\cos^{2}t}, \tag*{(14)} v ( x ) = − 2 ν cos ( 2 π x − π α ) , b ( x ) = 2 w ( π x ) sin ( π x ) , w ( t ) = s 2 sin 2 t + d 2 cos 2 t , (14) with parameters
{ ν = λ 2 , s = λ 1 + λ 3 = 1 , d = λ 3 − λ 1 for L I , ν = λ 1 + λ 3 , s = 1 , d = − 1 − 4 λ 1 λ 3 for L I I , ν = 1 , s = λ 1 + λ 3 = λ 2 , d = λ 1 − λ 3 for L I I I . (15) \begin{cases}
\nu=\lambda_{2},\ s=\lambda_{1}+\lambda_{3}=1,\ d=\lambda_{3}-\lambda_{1} & \text{for }\mathcal{L}_{\mathrm{I}},\\
\nu=\lambda_{1}+\lambda_{3},\ s=1,\ d=-\sqrt{1-4\lambda_{1}\lambda_{3}} & \text{for }\mathcal{L}_{\mathrm{II}},\\
\nu=1,\ s=\lambda_{1}+\lambda_{3}=\lambda_{2},\ d=\lambda_{1}-\lambda_{3} & \text{for }\mathcal{L}_{\mathrm{III}}. \tag*{(15)}
\end{cases} ⎩ ⎨ ⎧ ν = λ 2 , s = λ 1 + λ 3 = 1 , d = λ 3 − λ 1 ν = λ 1 + λ 3 , s = 1 , d = − 1 − 4 λ 1 λ 3 ν = 1 , s = λ 1 + λ 3 = λ 2 , d = λ 1 − λ 3 for L I , for L II , for L III . (15) For λ ∈ I i s o \lambda\in\mathcal{I}_{\mathrm{iso}} λ ∈ I iso ,
v ( x ) = − 2 cos ( 2 π x − π α + γ ) , b ( x ) = λ 2 − 2 λ 1 cos ( 2 π x + γ ) , γ = arccos λ 2 2 λ 1 . (16) v(x)=-2\cos(2\pi x-\pi\alpha+\gamma),\qquad b(x)=\lambda_{2}-2\lambda_{1}\cos(2\pi x+\gamma),\qquad\gamma=\arccos\frac{\lambda_{2}}{2\lambda_{1}}. \tag*{(16)} v ( x ) = − 2 cos ( 2 π x − π α + γ ) , b ( x ) = λ 2 − 2 λ 1 cos ( 2 π x + γ ) , γ = arccos 2 λ 1 λ 2 . (16)
Proof. On the three critical boundaries, choose
λ ∈ L I I I : ν = 1 , λ ~ = λ , λ ∈ L I : ν = λ 2 , λ ~ = ( λ 3 , 1 , λ 1 ) , λ = ( a , 1 , b ) ∈ L I I : ν = a + b , λ ~ = ( r − , 1 , r + ) , (17) \begin{aligned}
\lambda\in\mathcal{L}_{\mathrm{III}} &: \quad\nu=1,\quad\widetilde{\lambda}=\lambda,\\
\lambda\in\mathcal{L}_{\mathrm{I}} &: \quad\nu=\lambda_{2},\quad\widetilde{\lambda}=(\lambda_{3},1,\lambda_{1}),\\
\lambda=(a,1,b)\in\mathcal{L}_{\mathrm{II}} &: \quad\nu=a+b,\quad\widetilde{\lambda}=(r_{-},1,r_{+}), \tag*{(17)}
\end{aligned} λ ∈ L III λ ∈ L I λ = ( a , 1 , b ) ∈ L II : ν = 1 , λ = λ , : ν = λ 2 , λ = ( λ 3 , 1 , λ 1 ) , : ν = a + b , λ = ( r − , 1 , r + ) , (17)
where r ± r_{\pm} r ± are defined in Lemma 2.3. These are, respectively, the original family, its Aubry dual, and the family constructed in Theorem 2.4. In each case,
S 1 , λ ( α ) = S ν , λ ~ ( α ) , 0 ≤ ν ≤ λ ~ 2 = λ ~ 1 + λ ~ 3 > 0. (18) S_{1,\lambda}(\alpha)=S_{\nu,\widetilde{\lambda}}(\alpha),\qquad0\leq\nu\leq\widetilde{\lambda}_{2}=\widetilde{\lambda}_{1}+\widetilde{\lambda}_{3}>0. \tag*{(18)} S 1 , λ ( α ) = S ν , λ ( α ) , 0 ≤ ν ≤ λ 2 = λ 1 + λ 3 > 0. (18)
Writing z = e 2 π i ( t + α / 2 ) z=e^{2\pi i(t+\alpha/2)} z = e 2 π i ( t + α /2 ) , we have c ~ λ , α ( t ) = z − 1 ( 1 + z ) ( λ ~ 1 + λ ~ 3 z ) \widetilde{c}_{\lambda,\alpha}(t)=z^{-1}(1+z)(\widetilde{\lambda}_{1}+\widetilde{\lambda}_{3}z) c λ , α ( t ) = z − 1 ( 1 + z ) ( λ 1 + λ 3 z ) , so the hopping vanishes at t = 1 / 2 − α / 2 t=1/2-\alpha/2 t = 1/2 − α /2 . We further denote
s = λ ~ 1 + λ ~ 3 , d = λ ~ 1 − λ ~ 3 , s=\widetilde{\lambda}_{1}+\widetilde{\lambda}_{3},\qquad d=\widetilde{\lambda}_{1}-\widetilde{\lambda}_{3}, s = λ 1 + λ 3 , d = λ 1 − λ 3 ,
and write
b ^ ( t ) : = c ~ λ , α ( t − α / 2 − 1 / 2 ) = 2 w ( π t ) sin ( π t ) γ ( π t ) , γ ( t ) : = e i arg ( s sin t + i d cos t ) . (19) \widehat{b}(t):=\widetilde{c}_{\lambda,\alpha}(t-\alpha/2-1/2)=2w(\pi t)\sin(\pi t)\gamma(\pi t),\qquad\gamma(t):=e^{i\arg(s\sin t+id\cos t)}. \tag*{(19)} b ( t ) := c λ , α ( t − α /2 − 1/2 ) = 2 w ( π t ) sin ( π t ) γ ( π t ) , γ ( t ) := e i a r g ( s s i n t + i d c o s t ) . (19)
At zeros of s sin t + i d cos t s\sin t+id\cos t s sin t + i d cos t , set γ ( t ) = 1 \gamma(t)=1 γ ( t ) = 1 . Now similarly to Lemma 2.3, let
g j + 1 ′ = g j ′ / γ ( π ( θ + α j ) ) , j ∈ Z ; g 0 ′ = 1. g'_{j+1}=g'_{j}/\gamma(\pi(\theta+\alpha j)),\qquad j\in\mathbb{Z};\qquad g'_{0}=1. g j + 1 ′ = g j ′ / γ ( π ( θ + α j )) , j ∈ Z ; g 0 ′ = 1.
Then ( G α , θ ′ u ) j = g j ′ u j (G'_{\alpha,\theta}u)_{j}=g'_{j}u_{j} ( G α , θ ′ u ) j = g j ′ u j is the unitary transform of H H H which removes the factors γ ( π t ) \gamma(\pi t) γ ( π t ) from the off-diagonal matrix elements without changing diagonal elements and thus reduces the family to the form (13) as the same phase shift t → t − α / 2 − 1 / 2 t\to t-\alpha/2-1/2 t → t − α /2 − 1/2 has to be applied to the diagonal elements as well.
For λ ∈ I i s o \lambda\in\mathcal{I}_{\mathrm{iso}} λ ∈ I iso , the original hopping c λ , α ( t ) = λ 2 + 2 λ 1 cos ( 2 π ( t + α / 2 ) ) c_{\lambda,\alpha}(t)=\lambda_{2}+2\lambda_{1}\cos(2\pi(t+\alpha/2)) c λ , α ( t ) = λ 2 + 2 λ 1 cos ( 2 π ( t + α /2 )) already has a zero, since λ 2 ≤ 2 λ 1 \lambda_{2}\leq2\lambda_{1} λ 2 ≤ 2 λ 1 . Choosing b ( t ) = c λ , α ( t − α / 2 + 1 / 2 + γ / ( 2 π ) ) b(t)=c_{\lambda,\alpha}(t-\alpha/2+1/2+\gamma/(2\pi)) b ( t ) = c λ , α ( t − α /2 + 1/2 + γ / ( 2 π )) , γ = arccos λ 2 2 λ 1 \gamma=\arccos\frac{\lambda_{2}}{2\lambda_{1}} γ = arccos 2 λ 1 λ 2 , we complete the proof.
Separation into blocks
Consider the operator H v , b , α , θ H_{v,b,\alpha,\theta} H v , b , α , θ given by (13) . Note that it is singular with b ( 0 ) = 0 b(0)=0 b ( 0 ) = 0 . We now construct covers of its spectrum applying the strategy of our article [5 ] . First, we separate it into blocks and use a submersion lemma.
Let p n / q n p_n/q_n p n / q n be a sequence of canonical approximants to α \alpha α and let
δ n = q n ( α − p n q n ) , J n = [ − ∣ δ n − 1 ∣ , ∣ δ n − 1 ∣ ] n = 2 , 3 , … (20) \delta_n=q_n\left(\alpha-\frac{p_n}{q_n}\right),\qquad J_n=[-|\delta_{n-1}|,|\delta_{n-1}|]\qquad n=2,3,\ldots \tag*{(20)} δ n = q n ( α − q n p n ) , J n = [ − ∣ δ n − 1 ∣ , ∣ δ n − 1 ∣ ] n = 2 , 3 , … (20)
By standard properties of approximants, the length
∣ J n ∣ < 2 / q n . |J_n|<2/q_n. ∣ J n ∣ < 2/ q n .
Fix any n ≥ 3 n\geq3 n ≥ 3 and any phase θ = θ 0 \theta=\theta_0 θ = θ 0 of H v , b , α , θ H_{v,b,\alpha,\theta} H v , b , α , θ . Choose n 0 ∈ Z n_0\in\mathbb{Z} n 0 ∈ Z such that x 0 : = θ 0 + n 0 α x_0:=\theta_0+n_0\alpha x 0 := θ 0 + n 0 α satisfies x 0 ( m o d 1 ) ∈ ( − ∣ δ n ∣ , ∣ δ n ∣ ) x_0\pmod1\in(-|\delta_n|,|\delta_n|) x 0 ( mod 1 ) ∈ ( − ∣ δ n ∣ , ∣ δ n ∣ ) . There is a strictly increasing sequence of integers m k m_k m k , k = … , − 1 , 0 , 1 , … k=\ldots,-1,0,1,\ldots k = … , − 1 , 0 , 1 , … , such that m 0 = 0 m_0=0 m 0 = 0 , N k : = m k + 1 − m k N_k:=m_{k+1}-m_k N k := m k + 1 − m k takes on only two values, N k = q n N_k=q_n N k = q n or N k = q n − 1 N_k=q_{n-1} N k = q n − 1 , and such that x 0 + α m k ( m o d 1 ) ∈ J n x_0+\alpha m_k\pmod1\in J_n x 0 + α m k ( mod 1 ) ∈ J n , k ∈ Z k\in\mathbb{Z} k ∈ Z , and x 0 + α j ( m o d 1 ) ∉ ( − ∣ δ n ∣ , ∣ δ n ∣ ) x_0+\alpha j\pmod1\notin(-|\delta_n|,|\delta_n|) x 0 + α j ( mod 1 ) ∈ / ( − ∣ δ n ∣ , ∣ δ n ∣ ) if j ≠ m k j\ne m_k j = m k for any k k k (this is discussed in more detail in Section 7.1 of [5 ] ).
As in [5 ] , we separate H v , b , α , θ H_{v,b,\alpha,\theta} H v , b , α , θ into blocks removing the matrix elements b ( x k ) b(x_k) b ( x k ) , where x k : = θ 0 + ( n 0 + m k ) α x_k:=\theta_0+(n_0+m_k)\alpha x k := θ 0 + ( n 0 + m k ) α . Then x k ( m o d 1 ) ∈ J n x_k\pmod1\in J_n x k ( mod 1 ) ∈ J n and ∣ b ( x k ) ∣ ≤ C ( b ) ∣ J n ∣ |b(x_k)|\leq C(b)|J_n| ∣ b ( x k ) ∣ ≤ C ( b ) ∣ J n ∣ . The length of the k k k ’th block between the removed bonds b ( x k ) b(x_k) b ( x k ) and b ( x k + 1 ) b(x_{k+1}) b ( x k + 1 ) is N k = m k + 1 − m k ∈ { q n , q n − 1 } N_k=m_{k+1}-m_k\in\{q_n,q_{n-1}\} N k = m k + 1 − m k ∈ { q n , q n − 1 } .
Denote an orthonormal basis of C N \mathbb{C}^N C N by e 0 , e 1 , … , e N − 1 e_0,e_1,\ldots,e_{N-1} e 0 , e 1 , … , e N − 1 , and write a vector in this basis as u = ( u 0 , u 1 , … , u N − 1 ) T u=(u_0,u_1,\ldots,u_{N-1})^T u = ( u 0 , u 1 , … , u N − 1 ) T . Let Γ ^ N : C N → C 2 \widehat{\Gamma}_N:\mathbb{C}^N\to\mathbb{C}^2 Γ N : C N → C 2 and the projection P N : C N → C N P_N:\mathbb{C}^N\to\mathbb{C}^N P N : C N → C N onto the span{ e 0 , e N − 1 } \{e_0,e_{N-1}\} { e 0 , e N − 1 } , be given by
Γ ^ N u = ( u 0 , u N − 1 ) T , P N = Γ ^ N ∗ Γ ^ N . (21) \widehat{\Gamma}_N u=(u_0,u_{N-1})^T,\qquad P_N=\widehat{\Gamma}_N^*\widehat{\Gamma}_N. \tag*{(21)} Γ N u = ( u 0 , u N − 1 ) T , P N = Γ N ∗ Γ N . (21)
Define a tridiagonal matrix B ^ N ( x ) : C N → C N \widehat{B}_N(x):\mathbb{C}^N\to\mathbb{C}^N B N ( x ) : C N → C N using the matrix elements of H v , b , α , θ H_{v,b,\alpha,\theta} H v , b , α , θ as
B ^ N ( x ) j , j = v ( x + ( j + 1 ) α ) , 0 ≤ j ≤ N − 1 , \widehat{B}_N(x)_{j,j}=v(x+(j+1)\alpha),\qquad0\leq j\leq N-1, B N ( x ) j , j = v ( x + ( j + 1 ) α ) , 0 ≤ j ≤ N − 1 ,
B ^ N ( x ) j + 1 , j = B ^ N ( x ) j , j + 1 = b ( x + ( j + 1 ) α ) , 0 ≤ j ≤ N − 2. (22) \widehat{B}_N(x)_{j+1,j}=\widehat{B}_N(x)_{j,j+1}=b(x+(j+1)\alpha),\qquad0\leq j\leq N-2. \tag*{(22)} B N ( x ) j + 1 , j = B N ( x ) j , j + 1 = b ( x + ( j + 1 ) α ) , 0 ≤ j ≤ N − 2. (22)
The following proposition from [5 ] applies here.
Proposition 2.6. [5 Proposition 7.3] Let H v , b , α , θ H_{v,b,\alpha,\theta} H v , b , α , θ be given by (13) with b , v ∈ C 1 ( R ; R ) b,v\in C^1(\mathbb{R};\mathbb{R}) b , v ∈ C 1 ( R ; R ) , v ( x ) v(x) v ( x ) periodic of period 1 1 1 , b ( x ) b(x) b ( x ) either periodic of period 1 1 1 with b ( 0 ) = 0 b(0)=0 b ( 0 ) = 0 , or of period 2 2 2 and b ( 0 ) = b ( 1 ) = 0 b(0)=b(1)=0 b ( 0 ) = b ( 1 ) = 0 . Let α \alpha α be irrational, θ ∈ R \theta\in\mathbb{R} θ ∈ R , Γ ^ N \widehat{\Gamma}_N Γ N given in (21) , B ^ N ( x ) \widehat{B}_N(x) B N ( x ) given in (22) , and J n J_n J n in (20) . Then for N 1 ′ = q n N'_1=q_n N 1 ′ = q n and N 2 ′ = q n − 1 N'_2=q_{n-1} N 2 ′ = q n − 1 , n ≥ 3 n\geq3 n ≥ 3 , τ = sup k ∣ b ( x k ) ∣ \tau=\sup_k|b(x_k)| τ = sup k ∣ b ( x k ) ∣ ,
spec ( H v , b , α , θ ) ⊂ ⋃ ν = 1 2 ⋃ x ∈ J n ⋃ V = V ∗ ∈ M 2 ( C ) ∥ V ∥ ≤ τ σ ( B ^ N ν ′ ( x ) + Γ ^ N ν ′ ∗ V Γ ^ N ν ′ ) . (23) \operatorname{spec}(H_{v,b,\alpha,\theta})\subset\bigcup_{\nu=1}^{2}\ \bigcup_{x\in J_n}\ \bigcup_{\substack{V=V^*\in M_2(\mathbb{C})\\ \|V\|\leq\tau}}\sigma\left(\widehat{B}_{N'_\nu}(x)+\widehat{\Gamma}_{N'_\nu}^{*}V\widehat{\Gamma}_{N'_\nu}\right). \tag*{(23)} spec ( H v , b , α , θ ) ⊂ ν = 1 ⋃ 2 x ∈ J n ⋃ V = V ∗ ∈ M 2 ( C ) ∥ V ∥ ≤ τ ⋃ σ ( B N ν ′ ( x ) + Γ N ν ′ ∗ V Γ N ν ′ ) . (23)
Note that Γ ^ N ν ′ ∗ V Γ ^ N ν ′ \widehat{\Gamma}_{N'_\nu}^{*}V\widehat{\Gamma}_{N'_\nu} Γ N ν ′ ∗ V Γ N ν ′ has only 4 nonzero matrix elements located at upper right, left, and lower right, left corners of the matrix. Thus the spectrum of H v , b , α , θ H_{v,b,\alpha,\theta} H v , b , α , θ is included into the union of spectra of perturbed blocks. The perturbation is of rank at most 2 2 2 and of small (for large n n n ) norm
∥ V ∥ ≤ τ = sup k ∣ b ( x k ) ∣ ≤ C ( b ) ∣ J n ∣ < 2 C ( b ) / q n . \|V\|\leq\tau=\sup_k|b(x_k)|\leq C(b)|J_n|<2C(b)/q_n. ∥ V ∥ ≤ τ = k sup ∣ b ( x k ) ∣ ≤ C ( b ) ∣ J n ∣ < 2 C ( b ) / q n .
We will see that in some cases, the bound on sup k ∣ b ( x k ) ∣ \sup_k|b(x_k)| sup k ∣ b ( x k ) ∣ is smaller.
General cover estimates
A proof of the following proposition is contained in Section 7.2 of [5 ] in the case of constant K K K . We reproduce it here briefly for reader’s convenience and extend to K K K depending on the phase x x x . In the following sections we find the matrices K , E K,\mathcal{E} K , E and λ ∈ C 0 \lambda\in\mathcal{C}_0 λ ∈ C 0 with E \mathcal{E} E of rank no more than 4 4 4 satisfying ( I − P N ) E ( I − P N ) = 0 (I-P_N)\mathcal{E}(I-P_N)=0 ( I − P N ) E ( I − P N ) = 0 , and derive explicit expressions for the estimates in Proposition 2.7 below for λ ∈ C 0 \lambda\in\mathcal{C}_0 λ ∈ C 0 .
We write ∥ A ∥ S 1 = Tr A ∗ A \|A\|_{S_1}=\operatorname{Tr}\sqrt{A^*A} ∥ A ∥ S 1 = Tr A ∗ A , and ∥ A ∥ S 2 2 = Tr ( A ∗ A ) \|A\|_{S_2}^2=\operatorname{Tr}(A^*A) ∥ A ∥ S 2 2 = Tr ( A ∗ A ) for trace and Hilbert–Schmidt norms of a matrix A A A , respectively.
Proposition 2.7. Let N ≥ 2 N \ge2 N ≥ 2 , J J J be a compact interval, let B : J → M N ( C ) B:J \to\mathcal{M}_N(\mathbb{C}) B : J → M N ( C ) be a C 1 C^1 C 1 Hermitian path, and let K : J → M N ( C ) K:J \to\mathcal{M}_N(\mathbb{C}) K : J → M N ( C ) be continuous with K ∗ = − K K^*=-K K ∗ = − K . Let
E ( x ) = d d x B ( x ) − [ K ( x ) , B ( x ) ] . \mathcal{E}(x)=\frac{d}{dx}B(x)-[K(x),B(x)]. E ( x ) = d x d B ( x ) − [ K ( x ) , B ( x )] . Suppose that ( I − P N ) E ( I − P N ) = 0 (I-P_N)\mathcal{E}(I-P_N)=0 ( I − P N ) E ( I − P N ) = 0 , where P N P_N P N is defined in (21) . For β ≥ 0 \beta\ge0 β ≥ 0 , let H ± = B ± β P N H^\pm=B\pm\beta P_N H ± = B ± β P N and define the closed intervals
I j = [ min x ∈ J λ j ( H − ( x ) ) , max x ∈ J λ j ( H + ( x ) ) ] , 1 ≤ j ≤ N , I_j=\left[\min_{x\in J}\lambda_j(H^-(x)),\max_{x\in J}\lambda_j(H^+(x))\right],\qquad1\le j\le N, I j = [ x ∈ J min λ j ( H − ( x )) , x ∈ J max λ j ( H + ( x )) ] , 1 ≤ j ≤ N , where λ 1 ( H ± ) ≤ ⋯ ≤ λ N ( H ± ) \lambda_1(H^\pm)\le\cdots\le\lambda_N(H^\pm) λ 1 ( H ± ) ≤ ⋯ ≤ λ N ( H ± ) are the ordered eigenvalues of H ± H^\pm H ± . Then
⋃ x ∈ J ⋃ V = V ∗ ∈ M 2 ( C ) ∥ V ∥ ≤ β spec ( B ( x ) + Γ ^ N ∗ V Γ ^ N ) ⊂ ⋃ j = 1 N I j , (24) \bigcup_{x\in J}\ \bigcup_{\substack{V=V^*\in\mathcal{M}_2(\mathbb{C})\\ \lVert V\rVert\le\beta}}\operatorname{spec}\left(B(x)+\widehat{\Gamma}_N^*V\widehat{\Gamma}_N\right)\subset\bigcup_{j=1}^{N}I_j, \tag*{(24)} x ∈ J ⋃ V = V ∗ ∈ M 2 ( C ) ∥ V ∥ ≤ β ⋃ spec ( B ( x ) + Γ N ∗ V Γ N ) ⊂ j = 1 ⋃ N I j , (24) and the total length of intervals satisfies
∑ j = 1 N ∣ I j ∣ ≤ 4 β + 4 ∫ J ( ∥ E ( x ) ∥ S 2 + β ∥ [ K ( x ) , P N ] ∥ S 2 ) d x . (25) \sum_{j=1}^{N}|I_j|\le4\beta+4\int_J\left(\lVert\mathcal{E}(x)\rVert_{S_2}+\beta\lVert[K(x),P_N]\rVert_{S_2}\right)\,dx. \tag*{(25)} j = 1 ∑ N ∣ I j ∣ ≤ 4 β + 4 ∫ J ( ∥ E ( x ) ∥ S 2 + β ∥[ K ( x ) , P N ] ∥ S 2 ) d x . (25) Proof. For V = V ∗ V=V^* V = V ∗ with ∥ V ∥ ≤ β \lVert V\rVert\le\beta ∥ V ∥ ≤ β ,
− β P N ≤ Γ ^ N ∗ V Γ ^ N ≤ β P N , H − ≤ B + Γ ^ N ∗ V Γ ^ N ≤ H + . -\beta P_N\le\widehat{\Gamma}_N^*V\widehat{\Gamma}_N\le\beta P_N,\qquad H^-\le B+\widehat{\Gamma}_N^*V\widehat{\Gamma}_N\le H^+. − β P N ≤ Γ N ∗ V Γ N ≤ β P N , H − ≤ B + Γ N ∗ V Γ N ≤ H + . The min–max principle gives, for every x ∈ J x\in J x ∈ J ,
λ j ( H − ( x ) ) ≤ λ j ( B ( x ) + Γ ^ N ∗ V Γ ^ N ) ≤ λ j ( H + ( x ) ) , \lambda_j(H^-(x))\le\lambda_j\left(B(x)+\widehat{\Gamma}_N^*V\widehat{\Gamma}_N\right)\le\lambda_j(H^+(x)), λ j ( H − ( x )) ≤ λ j ( B ( x ) + Γ N ∗ V Γ N ) ≤ λ j ( H + ( x )) , which proves the inclusion (24) .
Now fix x 0 ∈ J x_0\in J x 0 ∈ J . Note that
∑ j = 1 N ( λ j ( H + ( x 0 ) ) − λ j ( H − ( x 0 ) ) ) = Tr ( 2 β P N ) = 4 β . \sum_{j=1}^{N}\left(\lambda_j(H^+(x_0))-\lambda_j(H^-(x_0))\right)=\operatorname{Tr}(2\beta P_N)=4\beta. j = 1 ∑ N ( λ j ( H + ( x 0 )) − λ j ( H − ( x 0 )) ) = Tr ( 2 β P N ) = 4 β . Adding the variations for each j j j gives
∑ j = 1 N ∣ I j ∣ ≤ 4 β + ∑ j = 1 N Var J λ j ( H + ) + ∑ j = 1 N Var J λ j ( H − ) . (26) \sum_{j=1}^{N}|I_j|\le4\beta+\sum_{j=1}^{N}\operatorname{Var}_J\lambda_j(H^+)+\sum_{j=1}^{N}\operatorname{Var}_J\lambda_j(H^-). \tag*{(26)} j = 1 ∑ N ∣ I j ∣ ≤ 4 β + j = 1 ∑ N Var J λ j ( H + ) + j = 1 ∑ N Var J λ j ( H − ) . (26) Note that ( I − P N ) [ K , P N ] ( I − P N ) = 0 (I-P_N)[K,P_N](I-P_N)=0 ( I − P N ) [ K , P N ] ( I − P N ) = 0 , and therefore in particular rank [ K , P N ] ≤ 4 \operatorname{rank}[K,P_N]\le4 rank [ K , P N ] ≤ 4 and so ∥ [ K , P N ] ∥ S 1 ≤ 2 ∥ [ K , P N ] ∥ S 2 \lVert[K,P_N]\rVert_{S_1}\le2\lVert[K,P_N]\rVert_{S_2} ∥[ K , P N ] ∥ S 1 ≤ 2 ∥[ K , P N ] ∥ S 2 . Let E ± = d d x H ± − [ K , H ± ] \mathcal{E}^{\pm}=\frac{d}{dx}H^\pm-[K,H^\pm] E ± = d x d H ± − [ K , H ± ] . Then we clearly have using our assumption on E \mathcal{E} E that
E ± = E ± β [ K , P N ] , ( I − P N ) E ± ( I − P N ) = 0 , \mathcal{E}^{\pm}=\mathcal{E}\pm\beta[K,P_N],\qquad(I-P_N)\mathcal{E}^{\pm}(I-P_N)=0, E ± = E ± β [ K , P N ] , ( I − P N ) E ± ( I − P N ) = 0 , and so rank E ± ≤ 4 \operatorname{rank}\mathcal{E}^{\pm}\le4 rank E ± ≤ 4 . Thus ∥ E ± ∥ S 1 ≤ 2 ∥ E ± ∥ S 2 \lVert\mathcal{E}^{\pm}\rVert_{S_1}\le2\lVert\mathcal{E}^{\pm}\rVert_{S_2} ∥ E ± ∥ S 1 ≤ 2 ∥ E ± ∥ S 2 .
Lemma 7.5 in [5 ] applied to H + H^+ H + and then to H − H^- H − would bound the sum of variations by integrals of ∥ E ± ∥ S 1 \lVert\mathcal{E}^{\pm}\rVert_{S_1} ∥ E ± ∥ S 1 . However, it was proved in [5 ] for a constant K K K . The proof is based on the unitary transform e − ( x − a ) K e^{-(x-a)K} e − ( x − a ) K . To extend the proof to the present case of K ( x ) K(x) K ( x ) depending on x x x , consider instead the unitary transform U ( x ) U(x) U ( x ) which is the solution to the equation
d d x U ( x ) = − U ( x ) K ( x ) \frac{d}{dx}U(x)=-U(x)K(x) d x d U ( x ) = − U ( x ) K ( x ) with initial condition U ( a ) = I U(a)=I U ( a ) = I . (The fact that it is unitary is easy to see differentiating U ( x ) U ( x ) ∗ U(x)U(x)^* U ( x ) U ( x ) ∗ .) Otherwise, the proof remains the same. Then, the lemma yields
∑ j = 1 N ( Var J λ j ( H + ) + Var J λ j ( H − ) ) ≤ ∫ J ( ∥ E + ∥ S 1 + ∥ E − ∥ S 1 ) d x ≤ 4 ∫ J ( ∥ E ∥ S 2 + β ∥ [ K , P N ] ∥ S 2 ) d x . \sum_{j=1}^{N}\left(\operatorname{Var}_J\lambda_j(H^+)+\operatorname{Var}_J\lambda_j(H^-)\right)\le\int_J\left(\lVert\mathcal{E}^+\rVert_{S_1}+\lVert\mathcal{E}^-\rVert_{S_1}\right)\,dx\le4\int_J\left(\lVert\mathcal{E}\rVert_{S_2}+\beta\lVert[K,P_N]\rVert_{S_2}\right)\,dx. j = 1 ∑ N ( Var J λ j ( H + ) + Var J λ j ( H − ) ) ≤ ∫ J ( ∥ E + ∥ S 1 + ∥ E − ∥ S 1 ) d x ≤ 4 ∫ J ( ∥ E ∥ S 2 + β ∥[ K , P N ] ∥ S 2 ) d x . Together with (26) , this proves (25) . □
Spectral covers on the three critical boundaries
We now provide explicit matrices K K K and E \mathcal{E} E for λ ∈ L I , L I I , L I I I \lambda\in\mathcal{L}_{I},\mathcal{L}_{II},\mathcal{L}_{III} λ ∈ L I , L I I , L I I I , such that E \mathcal{E} E and β [ K , P N ] \beta[K,P_N] β [ K , P N ] have bounded Hilbert–Schmidt norm and ( I − P N ) E ( I − P N ) = 0 (I-P_N)\mathcal{E}(I-P_N)=0 ( I − P N ) E ( I − P N ) = 0 . Applying then Proposition 2.7 with J = J n J=J_n J = J n and β = C ( 2 s / q n 2 + ∣ d ∣ / q n ) \beta=C(2s/q_n^2+|d|/q_n) β = C ( 2 s / q n 2 + ∣ d ∣/ q n ) , we obtain smallness of the total length of the cover. In the next subsection we do the same for λ ∈ I i s o \lambda\in\mathcal{I}_{\mathrm{iso}} λ ∈ I iso .
Recall (14), (15). Assume first that d ≠ 0 d\ne0 d = 0 . We then have 0 < ∣ d ∣ ≤ w ( t ) ≤ s 0<|d|\le w(t)\le s 0 < ∣ d ∣ ≤ w ( t ) ≤ s for all t ∈ R t\in\mathbb{R} t ∈ R .
Let ρ = s 2 − ν 2 \rho=\sqrt{s^2-\nu^2} ρ = s 2 − ν 2 and
C ( t ) = ( ν sin t − i d cos t ρ sin t − ρ sin t ν sin t + i d cos t ) , M ( t ) = C ( π t ) w ( π t ) . C(t)=\begin{pmatrix}\nu\sin t-id\cos t&\rho\sin t\\-\rho\sin t&\nu\sin t+id\cos t\end{pmatrix},\qquad M(t)=\frac{C(\pi t)}{w(\pi t)}. C ( t ) = ( ν sin t − i d cos t − ρ sin t ρ sin t ν sin t + i d cos t ) , M ( t ) = w ( π t ) C ( π t ) .
For r ≥ 1 r\ge1 r ≥ 1 , we write
P r ( t ) = M ( t ) M ( t + α ) ⋯ M ( t + ( r − 1 ) α ) , χ r ( t ) = 1 2 Tr P r ( t ) . P_r(t)=M(t)M(t+\alpha)\cdots M(t+(r-1)\alpha),\qquad\chi_r(t)=\frac{1}{2}\operatorname{Tr}P_r(t). P r ( t ) = M ( t ) M ( t + α ) ⋯ M ( t + ( r − 1 ) α ) , χ r ( t ) = 2 1 Tr P r ( t ) .
Denote
t j = x + ( j + 1 ) α , j ∈ Z . (27) t_j=x+(j+1)\alpha,\qquad j\in\mathbb{Z}. \tag*{(27)} t j = x + ( j + 1 ) α , j ∈ Z . (27)
We define the coefficients k r k_r k r and the entries of K N ( x ) K_N(x) K N ( x ) by the formulae
k r ( t ) = π ( − 1 ) r − 1 χ r ( t ) sin ( π α r ) , r ≥ 1 , k_r(t)=\frac{\pi(-1)^{r-1}\chi_r(t)}{\sin(\pi\alpha r)},\qquad r\ge1, k r ( t ) = sin ( π α r ) π ( − 1 ) r − 1 χ r ( t ) , r ≥ 1 ,
( K N ( x ) ) j , j + r = k r ( t j ) , ( K N ( x ) ) j + r , j = − k r ( t j ) , 0 ≤ j < j + r ≤ N − 1 , (K_N(x))_{j,j+r}=k_r(t_j),\qquad(K_N(x))_{j+r,j}=-k_r(t_j),\qquad0\le j<j+r\le N-1, ( K N ( x ) ) j , j + r = k r ( t j ) , ( K N ( x ) ) j + r , j = − k r ( t j ) , 0 ≤ j < j + r ≤ N − 1 ,
( K N ( x ) ) j j = 0 , 0 ≤ j ≤ N − 1. (28) (K_N(x))_{jj}=0,\qquad0\le j\le N-1. \tag*{(28)} ( K N ( x ) ) j j = 0 , 0 ≤ j ≤ N − 1. (28)
Note that sin ( π α r ) ≠ 0 \sin(\pi\alpha r)\ne0 sin ( π α r ) = 0 since α \alpha α is irrational.
Lemma 2.8. Let N > 2 N>2 N > 2 , B N ( x ) = B ^ N ( x ) B_N(x)=\widehat{B}_N(x) B N ( x ) = B N ( x ) be given by (22) with v ( x ) , b ( x ) v(x),b(x) v ( x ) , b ( x ) as defined in (14), and let K N ( x ) K_N(x) K N ( x ) be as defined in (28). Then
E N ( x ) : = d d x B N ( x ) − [ K N ( x ) , B N ( x ) ] \mathcal{E}_N(x):=\frac{d}{dx}B_N(x)-[K_N(x),B_N(x)] E N ( x ) := d x d B N ( x ) − [ K N ( x ) , B N ( x )] has the following edge matrix elements for 1 ≤ j ≤ N − 2 1\le j\le N-2 1 ≤ j ≤ N − 2 ,
( E N ) 0 , j = ( E N ) j , 0 = ( − 1 ) j + 1 2 π sin ( π ( j + 1 ) α ) w ( π t − 1 ) sin ( π t − 1 ) χ j + 1 ( t − 1 ) , (\mathcal{E}_N)_{0,j}=(\mathcal{E}_N)_{j,0}=
\frac{(-1)^{j+1}2\pi}{\sin(\pi(j+1)\alpha)}
w(\pi t_{-1})\sin(\pi t_{-1})\chi_{j+1}(t_{-1}), ( E N ) 0 , j = ( E N ) j , 0 = sin ( π ( j + 1 ) α ) ( − 1 ) j + 1 2 π w ( π t − 1 ) sin ( π t − 1 ) χ j + 1 ( t − 1 ) , ( E N ) j , N − 1 = ( E N ) N − 1 , j = ( − 1 ) N − 1 − j 2 π sin ( π ( N − j ) α ) w ( π t N − 1 ) sin ( π t N − 1 ) χ N − j ( t j ) , (29) (\mathcal{E}_N)_{j,N-1}=(\mathcal{E}_N)_{N-1,j}=
\frac{(-1)^{N-1-j}2\pi}{\sin(\pi(N-j)\alpha)}
w(\pi t_{N-1})\sin(\pi t_{N-1})\chi_{N-j}(t_j), \tag*{(29)} ( E N ) j , N − 1 = ( E N ) N − 1 , j = sin ( π ( N − j ) α ) ( − 1 ) N − 1 − j 2 π w ( π t N − 1 ) sin ( π t N − 1 ) χ N − j ( t j ) , (29) and the corner elements
( E N ) 0 , 0 = − 4 π sin ( π α ) w ( π t − 1 ) sin ( π t − 1 ) χ 1 ( t − 1 ) , (\mathcal{E}_N)_{0,0}=-\frac{4\pi}{\sin(\pi\alpha)}
w(\pi t_{-1})\sin(\pi t_{-1})\chi_1(t_{-1}), ( E N ) 0 , 0 = − sin ( π α ) 4 π w ( π t − 1 ) sin ( π t − 1 ) χ 1 ( t − 1 ) , ( E N ) N − 1 , N − 1 = 4 π sin ( π α ) w ( π t N − 1 ) sin ( π t N − 1 ) χ 1 ( t N − 1 ) , (\mathcal{E}_N)_{N-1,N-1}=
\frac{4\pi}{\sin(\pi\alpha)}
w(\pi t_{N-1})\sin(\pi t_{N-1})\chi_1(t_{N-1}), ( E N ) N − 1 , N − 1 = sin ( π α ) 4 π w ( π t N − 1 ) sin ( π t N − 1 ) χ 1 ( t N − 1 ) , ( E N ) 0 , N − 1 = ( E N ) N − 1 , 0 = π ( − 1 ) N − 1 Tr P N − 1 ( t 0 ) ( ν sin γ − i d cos γ ρ sin γ − ρ sin γ ν sin γ + i d cos γ ) , (\mathcal{E}_N)_{0,N-1}=(\mathcal{E}_N)_{N-1,0}
=\pi(-1)^{N-1}\operatorname{Tr}P_{N-1}(t_0)
\begin{pmatrix}
\nu\sin\gamma-id\cos\gamma&\rho\sin\gamma\\
-\rho\sin\gamma&\nu\sin\gamma+id\cos\gamma
\end{pmatrix}, ( E N ) 0 , N − 1 = ( E N ) N − 1 , 0 = π ( − 1 ) N − 1 Tr P N − 1 ( t 0 ) ( ν sin γ − i d cos γ − ρ sin γ ρ sin γ ν sin γ + i d cos γ ) , γ = π ( t − 1 + t N − 1 ) , (30) \gamma=\pi(t_{-1}+t_{N-1}), \tag*{(30)} γ = π ( t − 1 + t N − 1 ) , (30) and the rest of the matrix elements of E N ( x ) \mathcal{E}_N(x) E N ( x ) are zero.
Proof. We omit the subscript N N N . The identities B T = B B^{T}=B B T = B and K T = − K K^{T}=-K K T = − K give [ K , B ] T = [ K , B ] [K,B]^{T}=[K,B] [ K , B ] T = [ K , B ] , so it suffices to compute the entries of E \mathcal{E} E on and above the diagonal.
First, let 1 ≤ j ≤ N − 2 1 \le j \le N-2 1 ≤ j ≤ N − 2 , ℓ ≥ 2 \ell\ge2 ℓ ≥ 2 , j + ℓ ≤ N − 2 j+\ell\le N-2 j + ℓ ≤ N − 2 . Then using the fact that B B B is tridiagonal, we have
[ K , B ] j , j + ℓ = ∑ k K j , k B k , j + ℓ − B j , k K k , j + ℓ = K j , j + ℓ B j + ℓ , j + ℓ + K j , j + ℓ − 1 B j + ℓ − 1 , j + ℓ + K j , j + ℓ + 1 B j + ℓ + 1 , j + ℓ − K j , j + ℓ B j , j − K j − 1 , j + ℓ B j , j − 1 − K j + 1 , j + ℓ B j , j + 1 = k ℓ ( t j ) v ( t j + ℓ ) + k ℓ − 1 ( t j ) b ( t j + ℓ − 1 ) + k ℓ + 1 ( t j ) b ( t j + ℓ ) − k ℓ ( t j ) v ( t j ) − k ℓ + 1 ( t j − 1 ) b ( t j − 1 ) − k ℓ − 1 ( t j + 1 ) b ( t j ) . (31) \begin{aligned}
[K,B]_{j,j+\ell}=\sum_k K_{j,k}B_{k,j+\ell}-B_{j,k}K_{k,j+\ell} \\
&=K_{j,j+\ell}B_{j+\ell,j+\ell}+K_{j,j+\ell-1}B_{j+\ell-1,j+\ell}+K_{j,j+\ell+1}B_{j+\ell+1,j+\ell}
-K_{j,j+\ell}B_{j,j}-K_{j-1,j+\ell}B_{j,j-1}-K_{j+1,j+\ell}B_{j,j+1} \\
&=k_{\ell}(t_j)v(t_{j+\ell})+k_{\ell-1}(t_j)b(t_{j+\ell-1})+k_{\ell+1}(t_j)b(t_{j+\ell})
-k_{\ell}(t_j)v(t_j)-k_{\ell+1}(t_{j-1})b(t_{j-1})-k_{\ell-1}(t_{j+1})b(t_j). \tag*{(31)}
\end{aligned} [ K , B ] j , j + ℓ = k ∑ K j , k B k , j + ℓ − B j , k K k , j + ℓ = K j , j + ℓ B j + ℓ , j + ℓ + K j , j + ℓ − 1 B j + ℓ − 1 , j + ℓ + K j , j + ℓ + 1 B j + ℓ + 1 , j + ℓ − K j , j + ℓ B j , j − K j − 1 , j + ℓ B j , j − 1 − K j + 1 , j + ℓ B j , j + 1 = k ℓ ( t j ) v ( t j + ℓ ) + k ℓ − 1 ( t j ) b ( t j + ℓ − 1 ) + k ℓ + 1 ( t j ) b ( t j + ℓ ) − k ℓ ( t j ) v ( t j ) − k ℓ + 1 ( t j − 1 ) b ( t j − 1 ) − k ℓ − 1 ( t j + 1 ) b ( t j ) . (31)
Now denoting
A = P ℓ ( t j ) , s j = sin ( π α j ) , A=P_{\ell}(t_j),\qquad s_j=\sin(\pi\alpha_j), A = P ℓ ( t j ) , s j = sin ( π α j ) ,
we remark that
k ℓ ( t j ) = ( − 1 ) ℓ − 1 π 2 s ℓ Tr A , k_{\ell}(t_j)=\frac{(-1)^{\ell-1}\pi}{2s_{\ell}}\operatorname{Tr}A, k ℓ ( t j ) = 2 s ℓ ( − 1 ) ℓ − 1 π Tr A ,
k ℓ − 1 ( t j ) = ( − 1 ) ℓ π 2 s ℓ − 1 Tr A M ( t j + ℓ − 1 ) − 1 , k ℓ − 1 ( t j + 1 ) = ( − 1 ) ℓ π 2 s ℓ − 1 Tr A M ( t j ) − 1 , k_{\ell-1}(t_j)=\frac{(-1)^{\ell}\pi}{2s_{\ell-1}}\operatorname{Tr}AM(t_{j+\ell-1})^{-1},\qquad
k_{\ell-1}(t_{j+1})=\frac{(-1)^{\ell}\pi}{2s_{\ell-1}}\operatorname{Tr}AM(t_j)^{-1}, k ℓ − 1 ( t j ) = 2 s ℓ − 1 ( − 1 ) ℓ π Tr A M ( t j + ℓ − 1 ) − 1 , k ℓ − 1 ( t j + 1 ) = 2 s ℓ − 1 ( − 1 ) ℓ π Tr A M ( t j ) − 1 ,
k ℓ + 1 ( t j ) = ( − 1 ) ℓ π 2 s ℓ + 1 Tr A M ( t j + ℓ ) , k ℓ + 1 ( t j − 1 ) = ( − 1 ) ℓ π 2 s ℓ + 1 Tr A M ( t j − 1 ) , k_{\ell+1}(t_j)=\frac{(-1)^{\ell}\pi}{2s_{\ell+1}}\operatorname{Tr}AM(t_{j+\ell}),\qquad
k_{\ell+1}(t_{j-1})=\frac{(-1)^{\ell}\pi}{2s_{\ell+1}}\operatorname{Tr}AM(t_{j-1}), k ℓ + 1 ( t j ) = 2 s ℓ + 1 ( − 1 ) ℓ π Tr A M ( t j + ℓ ) , k ℓ + 1 ( t j − 1 ) = 2 s ℓ + 1 ( − 1 ) ℓ π Tr A M ( t j − 1 ) ,
and therefore
[ K , B ] j , j + ℓ = ( − 1 ) ℓ π 2 s ℓ s ℓ − 1 s ℓ + 1 Tr A [ ( v ( t j ) − v ( t j + ℓ ) ) s ℓ + 1 s ℓ − 1 I + ( M ( t j + ℓ ) b ( t j + ℓ ) − M ( t j − 1 ) b ( t j − 1 ) ) s ℓ s ℓ − 1 + ( M ( t j + ℓ − 1 ) − 1 b ( t j + ℓ − 1 ) − M ( t j ) − 1 b ( t j ) ) s ℓ s ℓ + 1 ] . (32) \begin{aligned}
[K,B]_{j,j+\ell}
={}&\frac{(-1)^{\ell}\pi}{2s_{\ell}s_{\ell-1}s_{\ell+1}}\operatorname{Tr}A\bigl[(v(t_j)-v(t_{j+\ell}))s_{\ell+1}s_{\ell-1}I\\
&+(M(t_{j+\ell})b(t_{j+\ell})-M(t_{j-1})b(t_{j-1}))s_{\ell}s_{\ell-1}\\
&+(M(t_{j+\ell-1})^{-1}b(t_{j+\ell-1})-M(t_j)^{-1}b(t_j))s_{\ell}s_{\ell+1}\bigr]. \tag*{(32)}
\end{aligned} [ K , B ] j , j + ℓ = 2 s ℓ s ℓ − 1 s ℓ + 1 ( − 1 ) ℓ π Tr A [ ( v ( t j ) − v ( t j + ℓ )) s ℓ + 1 s ℓ − 1 I + ( M ( t j + ℓ ) b ( t j + ℓ ) − M ( t j − 1 ) b ( t j − 1 )) s ℓ s ℓ − 1 + ( M ( t j + ℓ − 1 ) − 1 b ( t j + ℓ − 1 ) − M ( t j ) − 1 b ( t j )) s ℓ s ℓ + 1 ] . (32)
Using trigonometric identities, it is straightforward to verify that with γ j = π ( t j − 1 + t j + ℓ ) \gamma_j=\pi(t_{j-1}+t_{j+\ell}) γ j = π ( t j − 1 + t j + ℓ )
( M ( t j + ℓ ) b ( t j + ℓ ) − M ( t j − 1 ) b ( t j − 1 ) ) s ℓ s ℓ − 1 = 2 s ℓ s ℓ + 1 s ℓ − 1 ( ν sin γ j − i d cos γ j ρ sin γ j − ρ sin γ j ν sin γ j + i d cos γ j ) , (33) (M(t_{j+\ell})b(t_{j+\ell})-M(t_{j-1})b(t_{j-1}))s_{\ell}s_{\ell-1}
=2s_{\ell}s_{\ell+1}s_{\ell-1}
\begin{pmatrix}
\nu\sin\gamma_j-id\cos\gamma_j & \rho\sin\gamma_j\\
-\rho\sin\gamma_j & \nu\sin\gamma_j+id\cos\gamma_j
\end{pmatrix}, \tag*{(33)} ( M ( t j + ℓ ) b ( t j + ℓ ) − M ( t j − 1 ) b ( t j − 1 )) s ℓ s ℓ − 1 = 2 s ℓ s ℓ + 1 s ℓ − 1 ( ν sin γ j − i d cos γ j − ρ sin γ j ρ sin γ j ν sin γ j + i d cos γ j ) , (33)
and since γ j = π ( t j − 1 + t j + ℓ ) = π ( t j + t j + ℓ − 1 ) \gamma_j=\pi(t_{j-1}+t_{j+\ell})=\pi(t_j+t_{j+\ell-1}) γ j = π ( t j − 1 + t j + ℓ ) = π ( t j + t j + ℓ − 1 ) ,
( M ( t j + ℓ − 1 ) − 1 b ( t j + ℓ − 1 ) − M ( t j ) − 1 b ( t j ) ) s ℓ s ℓ + 1 = 2 s ℓ s ℓ + 1 s ℓ − 1 ( ν sin γ j + i d cos γ j − ρ sin γ j ρ sin γ j ν sin γ j − i d cos γ j ) , (34) (M(t_{j+\ell-1})^{-1}b(t_{j+\ell-1})-M(t_j)^{-1}b(t_j))s_{\ell}s_{\ell+1}
=2s_{\ell}s_{\ell+1}s_{\ell-1}
\begin{pmatrix}
\nu\sin\gamma_j+id\cos\gamma_j & -\rho\sin\gamma_j\\
\rho\sin\gamma_j & \nu\sin\gamma_j-id\cos\gamma_j
\end{pmatrix}, \tag*{(34)} ( M ( t j + ℓ − 1 ) − 1 b ( t j + ℓ − 1 ) − M ( t j ) − 1 b ( t j )) s ℓ s ℓ + 1 = 2 s ℓ s ℓ + 1 s ℓ − 1 ( ν sin γ j + i d cos γ j ρ sin γ j − ρ sin γ j ν sin γ j − i d cos γ j ) , (34)
and
v ( t j ) − v ( t j + ℓ ) = − 4 ν s ℓ sin γ j , v(t_j)-v(t_{j+\ell})=-4\nu s_{\ell}\sin\gamma_j, v ( t j ) − v ( t j + ℓ ) = − 4 ν s ℓ sin γ j ,
we obtain that the expression inside the square bracket in (32) vanishes, so that [ K , B ] j , j + ℓ = 0 [K,B]_{j,j+\ell}=0 [ K , B ] j , j + ℓ = 0 .
Using (31) we similarly have for 1 ≤ j ≤ N − 3 1\le j\le N-3 1 ≤ j ≤ N − 3 , ℓ = 1 \ell=1 ℓ = 1 ,
[ K , B ] j , j + 1 = − π 2 s 1 s 2 Tr M ( t j ) [ ( v ( t j ) − v ( t j + 1 ) ) s 2 I + ( M ( t j + 1 ) b ( t j + 1 ) − M ( t j − 1 ) b ( t j − 1 ) ) s 1 ] . (35) [K,B]_{j,j+1}=-\frac{\pi}{2s_1s_2}\operatorname{Tr}M(t_j)\bigl[(v(t_j)-v(t_{j+1}))s_2I
+(M(t_{j+1})b(t_{j+1})-M(t_{j-1})b(t_{j-1}))s_1\bigr]. \tag*{(35)} [ K , B ] j , j + 1 = − 2 s 1 s 2 π Tr M ( t j ) [ ( v ( t j ) − v ( t j + 1 )) s 2 I + ( M ( t j + 1 ) b ( t j + 1 ) − M ( t j − 1 ) b ( t j − 1 )) s 1 ] . (35)
Substituting the explicit expressions for M M M and b b b gives
[ K , B ] j , j + 1 = − π 2 w ( π t j ) ( − 4 s 2 sin ( 2 π t j ) sin ( π t j ) − 4 d 2 cos ( 2 π t j ) cos ( π t j ) ) = d d x b ( t j ) = d d x B ( x ) j , j + 1 . (36) [K,B]_{j,j+1}=-\frac{\pi}{2w(\pi t_j)}\bigl(-4s^2\sin(2\pi t_j)\sin(\pi t_j)-4d^2\cos(2\pi t_j)\cos(\pi t_j)\bigr)
=\frac{d}{dx}b(t_j)=\frac{d}{dx}B(x)_{j,j+1}. \tag*{(36)} [ K , B ] j , j + 1 = − 2 w ( π t j ) π ( − 4 s 2 sin ( 2 π t j ) sin ( π t j ) − 4 d 2 cos ( 2 π t j ) cos ( π t j ) ) = d x d b ( t j ) = d x d B ( x ) j , j + 1 . (36)
Now for 1 ≤ j ≤ N − 2 1\le j\le N-2 1 ≤ j ≤ N − 2 , ℓ = 0 \ell=0 ℓ = 0 ,
[ K , B ] j , j = π s 1 Tr [ M ( t j ) b ( t j ) − M ( t j − 1 ) b ( t j − 1 ) ] = 4 π ν sin ( π ( t j + t j − 1 ) ) = d d x v ( t j ) = d d x B ( x ) j , j . (37) \begin{aligned}
[K,B]_{j,j}=\frac{\pi}{s_1}\operatorname{Tr}\bigl[M(t_j)b(t_j)-M(t_{j-1})b(t_{j-1})\bigr] \\
&=4\pi\nu\sin(\pi(t_j+t_{j-1}))=\frac{d}{dx}v(t_j)=\frac{d}{dx}B(x)_{j,j}. \tag*{(37)}
\end{aligned} [ K , B ] j , j = s 1 π Tr [ M ( t j ) b ( t j ) − M ( t j − 1 ) b ( t j − 1 ) ] = 4 π ν sin ( π ( t j + t j − 1 )) = d x d v ( t j ) = d x d B ( x ) j , j . (37)
Consider now 1 ≤ j ≤ N − 2 1 \le j \le N - 2 1 ≤ j ≤ N − 2 , j + ℓ = N − 1 j + \ell= N - 1 j + ℓ = N − 1 . Then in (31) there is no element K j , N B N , N − 1 K_{j,N}B_{N,N-1} K j , N B N , N − 1 . We can define it by the same formula,
K j , N B N , N − 1 : = k N − j ( t j ) b ( t N − 1 ) , N − j = ℓ + 1. K_{j,N}B_{N,N-1}:=k_{N-j}(t_j)b(t_{N-1}), \qquad N-j=\ell+1. K j , N B N , N − 1 := k N − j ( t j ) b ( t N − 1 ) , N − j = ℓ + 1.
Adding and subtracting it in (31) we obtain
[ K , B ] j , N − 1 = d d x B ( x ) j , N − 1 − E j , N − 1 , (38) [K,B]_{j,N-1}=\frac{d}{dx}B(x)_{j,N-1}-\mathcal{E}_{j,N-1}, \tag*{(38)} [ K , B ] j , N − 1 = d x d B ( x ) j , N − 1 − E j , N − 1 , (38)
with the element E j , N − 1 \mathcal{E}_{j,N-1} E j , N − 1 as given in (29). The other matrix elements, [ K , B ] 0 , ℓ [K,B]_{0,\ell} [ K , B ] 0 , ℓ , [ K , B ] N − 1 , N − 1 [K,B]_{N-1,N-1} [ K , B ] N − 1 , N − 1 of the upper triangular part of [ K , B ] [K,B] [ K , B ] are verified similarly. □
We also have
Lemma 2.9. For d ≠ 0 d \ne0 d = 0 , the matrices and coefficients above satisfy, for every real t t t and every r ≥ 1 r \ge1 r ≥ 1 ,
M ( t ) , P r ( t ) ∈ S U ( 2 ) , χ r ( t ) ∈ [ − 1 , 1 ] . (39) M(t),\mathcal{P}_r(t)\in\mathrm{SU}(2), \qquad\chi_r(t)\in[-1,1]. \tag*{(39)} M ( t ) , P r ( t ) ∈ SU ( 2 ) , χ r ( t ) ∈ [ − 1 , 1 ] . (39) Proof. Multiplication of matrices, using ν 2 + ρ 2 = s 2 \nu^2+\rho^2=s^2 ν 2 + ρ 2 = s 2 , gives
C ( t ) ∗ C ( t ) = ( ( ν 2 + ρ 2 ) sin 2 t + d 2 cos 2 t ) I = w ( t ) 2 I , det C ( t ) = w ( t ) 2 ≥ d 2 ≠ 0. C(t)^*C(t)=\left((\nu^2+\rho^2)\sin^2t+d^2\cos^2t\right)I=w(t)^2I,\qquad\det C(t)=w(t)^2\ge d^2\ne0. C ( t ) ∗ C ( t ) = ( ( ν 2 + ρ 2 ) sin 2 t + d 2 cos 2 t ) I = w ( t ) 2 I , det C ( t ) = w ( t ) 2 ≥ d 2 = 0. Thus M ( t ) , P r ( t ) ∈ S U ( 2 ) M(t),\mathcal{P}_r(t)\in\mathrm{SU}(2) M ( t ) , P r ( t ) ∈ SU ( 2 ) . The eigenvalues of P r ( t ) \mathcal{P}_r(t) P r ( t ) are e i φ r ( t ) e^{i\varphi_r(t)} e i φ r ( t ) and e − i φ r ( t ) e^{-i\varphi_r(t)} e − i φ r ( t ) , so
χ r ( t ) = cos φ r ( t ) ∈ [ − 1 , 1 ] . \chi_r(t)=\cos\varphi_r(t)\in[-1,1]. χ r ( t ) = cos φ r ( t ) ∈ [ − 1 , 1 ] .
Lemma 2.10. For n ≥ 4 n\ge4 n ≥ 4 , let x ( m o d 1 ) , x + α N ( m o d 1 ) ∈ J n x\pmod1,x+\alpha N\pmod1\in J_n x ( mod 1 ) , x + α N ( mod 1 ) ∈ J n , and N ∈ { q n , q n − 1 } N\in\{q_n,q_{n-1}\} N ∈ { q n , q n − 1 } . Then the matrices E N ( x ) \mathcal{E}_N(x) E N ( x ) and K N ( x ) K_N(x) K N ( x ) from Lemma 2.8 satisfy
∥ E N ∥ S 2 ≤ C ( s q n + ∣ d ∣ ) , ∥ [ K N , P N ] ∥ S 2 ≤ C q n , (40) \|\mathcal{E}_N\|_{S_2}\le C\left(\frac{s}{q_n}+|d|\right),\qquad\|[K_N,P_N]\|_{S_2}\le Cq_n, \tag*{(40)} ∥ E N ∥ S 2 ≤ C ( q n s + ∣ d ∣ ) , ∥ [ K N , P N ] ∥ S 2 ≤ C q n , (40) where C > 0 C>0 C > 0 is absolute.
Proof. Using Lemma 2.9 and the formulae for the matrix elements of K N ( x ) K_N(x) K N ( x ) and E N ( x ) \mathcal{E}_N(x) E N ( x ) given by (28), (29), (30), we obtain with an absolute C > 0 C>0 C > 0 ,
∥ E N ( x ) ∥ S 2 2 ≤ C ( ∣ w ( π t − 1 ) sin ( π t − 1 ) ∣ 2 + ∣ w ( π t N − 1 ) sin ( π t N − 1 ) ∣ 2 ) ∑ r = 1 N − 1 1 ∣ sin ( π α r ) ∣ 2 + C ( s 2 sin 2 γ + d 2 ) , (41) \|\mathcal{E}_N(x)\|_{S_2}^2\le C\left(|w(\pi t_{-1})\sin(\pi t_{-1})|^2+|w(\pi t_{N-1})\sin(\pi t_{N-1})|^2\right)\sum_{r=1}^{N-1}\frac{1}{|\sin(\pi\alpha r)|^2}+C(s^2\sin^2\gamma+d^2), \tag*{(41)} ∥ E N ( x ) ∥ S 2 2 ≤ C ( ∣ w ( π t − 1 ) sin ( π t − 1 ) ∣ 2 + ∣ w ( π t N − 1 ) sin ( π t N − 1 ) ∣ 2 ) r = 1 ∑ N − 1 ∣ sin ( π α r ) ∣ 2 1 + C ( s 2 sin 2 γ + d 2 ) , (41) ∥ [ K N , P N ] ∥ S 2 2 ≤ C ∑ r = 1 N − 1 1 ∣ sin ( π α r ) ∣ 2 . \|[K_N,P_N]\|_{S_2}^2\le C\sum_{r=1}^{N-1}\frac{1}{|\sin(\pi\alpha r)|^2}. ∥ [ K N , P N ] ∥ S 2 2 ≤ C r = 1 ∑ N − 1 ∣ sin ( π α r ) ∣ 2 1 . As argued in (87) of [5 ] , ∑ r = 1 L 1 / ∣ sin ( π α r ) ∣ 2 < C q n 2 \sum_{r=1}^{L}1/|\sin(\pi\alpha r)|^2<Cq_n^2 ∑ r = 1 L 1/∣ sin ( π α r ) ∣ 2 < C q n 2 , for L ≤ ⌊ q n / 2 ⌋ L\le\lfloor q_n/2\rfloor L ≤ ⌊ q n /2 ⌋ with absolute C > 0 C>0 C > 0 . However that argument straightforwardly extends for any 1 < L < q n 1<L<q_n 1 < L < q n , and since N ≤ q n N\le q_n N ≤ q n , we conclude that
∑ r = 1 N − 1 1 ∣ sin ( π α r ) ∣ 2 < C q n 2 , (42) \sum_{r=1}^{N-1}\frac{1}{|\sin(\pi\alpha r)|^2}<Cq_n^2, \tag*{(42)} r = 1 ∑ N − 1 ∣ sin ( π α r ) ∣ 2 1 < C q n 2 , (42) which gives the second inequality of the lemma. Since x ( m o d 1 ) , x + α N ( m o d 1 ) ∈ J n x\pmod1,x+\alpha N\pmod1\in J_n x ( mod 1 ) , x + α N ( mod 1 ) ∈ J n , we also have that
∣ w ( π t − 1 ) sin ( π t − 1 ) ∣ ≤ 2 ( s ∣ sin ( π x ) ∣ + ∣ d ∣ ) ∣ sin ( π x ) ∣ ≤ C ( s ∣ J n ∣ 2 + ∣ d ∣ ∣ J n ∣ ) < C ( s q n 2 + ∣ d ∣ q n ) . (43) |w(\pi t_{-1})\sin(\pi t_{-1})|\le2(s|\sin(\pi x)|+|d|)|\sin(\pi x)|\le C(s|J_n|^2+|d||J_n|)<C\left(\frac{s}{q_n^2}+\frac{|d|}{q_n}\right). \tag*{(43)} ∣ w ( π t − 1 ) sin ( π t − 1 ) ∣ ≤ 2 ( s ∣ sin ( π x ) ∣ + ∣ d ∣ ) ∣ sin ( π x ) ∣ ≤ C ( s ∣ J n ∣ 2 + ∣ d ∣∣ J n ∣ ) < C ( q n 2 s + q n ∣ d ∣ ) . (43) ∣ w ( π t N − 1 ) sin ( π t N − 1 ) ∣ ≤ 2 ( s ∣ sin ( π ( x + α N ) ) ∣ + ∣ d ∣ ) ∣ sin ( π ( x + α N ) ) ∣ |w(\pi t_{N-1})\sin(\pi t_{N-1})|\le2(s|\sin(\pi(x+\alpha N))|+|d|)|\sin(\pi(x+\alpha N))| ∣ w ( π t N − 1 ) sin ( π t N − 1 ) ∣ ≤ 2 ( s ∣ sin ( π ( x + α N )) ∣ + ∣ d ∣ ) ∣ sin ( π ( x + α N )) ∣ < C ( s q n 2 + ∣ d ∣ q n ) . (44) <C\left(\frac{s}{q_n^2}+\frac{|d|}{q_n}\right). \tag*{(44)} < C ( q n 2 s + q n ∣ d ∣ ) . (44)
Noting also that ∣ sin γ ∣ ≤ ∣ sin ( π t − 1 ) ∣ + ∣ sin ( π t N − 1 ) ∣ ≤ C / q n |\sin\gamma| \le|\sin(\pi t_{-1})| + |\sin(\pi t_{N-1})| \le C/q_n ∣ sin γ ∣ ≤ ∣ sin ( π t − 1 ) ∣ + ∣ sin ( π t N − 1 ) ∣ ≤ C / q n and substituting (42) , (43) , (44) into (41) , we obtain the first inequality of the lemma. □ \square □
Proposition 2.11. Let λ \lambda λ be in the regions L j \mathcal{L}_j L j , j = 1 , 2 , 3 j=1,2,3 j = 1 , 2 , 3 . For every irrational α \alpha α and every n ≥ 4 n \ge4 n ≥ 4 , there are closed intervals I n , 1 , … , I n , M n I_{n,1},\ldots,I_{n,M_n} I n , 1 , … , I n , M n such that
Σ α , λ = spec ( H α , θ , λ ) ⊂ ⋃ j = 1 M n I n , j , M n ≤ q n + q n − 1 , (45) \Sigma_{\alpha,\lambda}=\operatorname{spec}(H_{\alpha,\theta,\lambda})\subset\bigcup_{j=1}^{M_n} I_{n,j}, \qquad M_n\le q_n+q_{n-1}, \tag*{(45)} Σ α , λ = spec ( H α , θ , λ ) ⊂ j = 1 ⋃ M n I n , j , M n ≤ q n + q n − 1 , (45) and
∑ j = 1 M n ∣ I n , j ∣ ≤ C ( ∣ d ∣ q n + s q n 2 ) , (46) \sum_{j=1}^{M_n}|I_{n,j}|\le C\left(\frac{|d|}{q_n}+\frac{s}{q_n^2}\right), \tag*{(46)} j = 1 ∑ M n ∣ I n , j ∣ ≤ C ( q n ∣ d ∣ + q n 2 s ) , (46) where s s s and d d d are given in (15) above, and C C C is absolute.
Proof. In the singularized representation for H α , θ , λ H_{\alpha,\theta,\lambda} H α , θ , λ , we have by (18) in all 3 cases L j \mathcal{L}_j L j , j = 1 , 2 , 3 j=1,2,3 j = 1 , 2 , 3 ,
s = λ ~ 1 + λ ~ 3 = λ ~ 2 > 0. s=\tilde{\lambda}_1+\tilde{\lambda}_3=\tilde{\lambda}_2>0. s = λ ~ 1 + λ ~ 3 = λ ~ 2 > 0.
We also have d = λ ~ 1 − λ ~ 3 d=\tilde{\lambda}_1-\tilde{\lambda}_3 d = λ ~ 1 − λ ~ 3 and λ ~ = ( s / 2 + d / 2 , s , s / 2 − d / 2 ) \tilde{\lambda}=(s/2+d/2,s,s/2-d/2) λ ~ = ( s /2 + d /2 , s , s /2 − d /2 ) in all 3 cases.
We first suppose that d ≠ 0 d\ne0 d = 0 . In Proposition 2.7, set J = J n J=J_n J = J n , K = K N K=K_N K = K N , ε = ε N \varepsilon=\varepsilon_N ε = ε N ,
β = sup k ∣ b ( x k ) ∣ ≤ C ( 2 s / q n 2 + ∣ d ∣ / q n ) . \beta=\sup_k |b(x_k)|\le C(2s/q_n^2+|d|/q_n). β = k sup ∣ b ( x k ) ∣ ≤ C ( 2 s / q n 2 + ∣ d ∣/ q n ) .
Note that B N ( x ) B_N(x) B N ( x ) , B N ( x + 1 ) B_N(x+1) B N ( x + 1 ) , B N ( x − 1 ) B_N(x-1) B N ( x − 1 ) are unitarily equivalent by the transform
U = diag { 1 , − 1 , 1 , − 1 , … } . U=\operatorname{diag}\{1,-1,1,-1,\ldots\}. U = diag { 1 , − 1 , 1 , − 1 , … } .
so we only need to consider the case x ∈ J n x\in J_n x ∈ J n . Using now Proposition 2.6 with τ = β \tau=\beta τ = β and Lemma 2.10, we immediately obtain (45) and (46) in this case.
Suppose now that d = 0 d=0 d = 0 . Let 0 < ε < s / 2 0<\varepsilon<s/2 0 < ε < s /2 and consider d ′ = 2 ε d'=2\varepsilon d ′ = 2 ε and the corresponding λ ~ = λ ( ε ) = ( s / 2 + ε , s , s / 2 − ε ) \tilde{\lambda}=\lambda^{(\varepsilon)}=(s/2+\varepsilon,s,s/2-\varepsilon) λ ~ = λ ( ε ) = ( s /2 + ε , s , s /2 − ε ) .
With z = e 2 π i ( x + α / 2 ) z=e^{2\pi i(x+\alpha/2)} z = e 2 π i ( x + α /2 ) , the difference of the respective matrix elements given in (1) is
c λ ( ε ) , α ( x ) − c λ ( 0 ) , α ( x ) = ε ( z − 1 − z ) , ∥ c λ ( ε ) , α − c λ ( 0 ) , α ∥ ∞ ≤ 2 ε . c_{\lambda^{(\varepsilon)},\alpha}(x)-c_{\lambda^{(0)},\alpha}(x)=\varepsilon(z^{-1}-z), \qquad\|c_{\lambda^{(\varepsilon)},\alpha}-c_{\lambda^{(0)},\alpha}\|_\infty\le2\varepsilon. c λ ( ε ) , α ( x ) − c λ ( 0 ) , α ( x ) = ε ( z − 1 − z ) , ∥ c λ ( ε ) , α − c λ ( 0 ) , α ∥ ∞ ≤ 2 ε .
Hence, uniformly in phase,
∥ H ν , α , θ , λ ( ε ) − H ν , α , θ , λ ( 0 ) ∥ ≤ 4 ε . \|H_{\nu,\alpha,\theta,\lambda^{(\varepsilon)}}-H_{\nu,\alpha,\theta,\lambda^{(0)}}\|\le4\varepsilon. ∥ H ν , α , θ , λ ( ε ) − H ν , α , θ , λ ( 0 ) ∥ ≤ 4 ε .
Writing S ε = S ν , λ ( ε ) ( α ) S_\varepsilon=S_{\nu,\lambda^{(\varepsilon)}}(\alpha) S ε = S ν , λ ( ε ) ( α ) , we obtain the Hausdorff distance
d H ( S ε , S 0 ) ≤ 4 ε . d_H(S_\varepsilon,S_0)\le4\varepsilon. d H ( S ε , S 0 ) ≤ 4 ε .
Fix n n n and write M = q n + q n − 1 M=q_n+q_{n-1} M = q n + q n − 1 . By (45) and (46) just proved in the case d ′ = 2 ε ≠ 0 d'=2\varepsilon\ne0 d ′ = 2 ε = 0 , adding zero-length intervals if needed, we have
S ε ⊂ U ε : = ⋃ j = 1 M [ a ε , j , b ε , j ] , ∑ j = 1 M ( b ε , j − a ε , j ) ≤ C ( 2 ε q n + s q n 2 ) . S_\varepsilon\subset U_\varepsilon:=\bigcup_{j=1}^{M}[a_{\varepsilon,j},b_{\varepsilon,j}], \qquad\sum_{j=1}^{M}(b_{\varepsilon,j}-a_{\varepsilon,j})\le C\left(\frac{2\varepsilon}{q_n}+\frac{s}{q_n^2}\right). S ε ⊂ U ε := j = 1 ⋃ M [ a ε , j , b ε , j ] , j = 1 ∑ M ( b ε , j − a ε , j ) ≤ C ( q n 2 ε + q n 2 s ) .
Since S ε S_\varepsilon S ε lies in a fixed compact interval, we may intersect the covering intervals with it and replace empty intersections by singletons in it, without increasing the total length. Thus there is a sequence ε k ↓ 0 \varepsilon_k\downarrow0 ε k ↓ 0 on which every endpoint a ε k , j a_{\varepsilon_k,j} a ε k , j , b ε k , j b_{\varepsilon_k,j} b ε k , j converges. We write
a ε k , j ⟶ a j , b ε k , j ⟶ b j , a j ≤ b j , U 0 = ⋃ j = 1 M [ a j , b j ] . a_{\varepsilon_k,j}\longrightarrow a_j,\qquad b_{\varepsilon_k,j}\longrightarrow b_j,\qquad a_j\le b_j,\qquad U_0=\bigcup_{j=1}^{M}[a_j,b_j]. a ε k , j ⟶ a j , b ε k , j ⟶ b j , a j ≤ b j , U 0 = j = 1 ⋃ M [ a j , b j ] .
Then
d H ( U ε k , U 0 ) ≤ max 1 ≤ j ≤ M max { ∣ a ε k , j − a j ∣ , ∣ b ε k , j − b j ∣ } ⟶ 0. d_H(U_{\varepsilon_k},U_0)\le\max_{1\le j\le M}\max\{|a_{\varepsilon_k,j}-a_j|,|b_{\varepsilon_k,j}-b_j|\}\longrightarrow0. d H ( U ε k , U 0 ) ≤ 1 ≤ j ≤ M max max { ∣ a ε k , j − a j ∣ , ∣ b ε k , j − b j ∣ } ⟶ 0.
Since
sup E ∈ S 0 dist ( E , U 0 ) ≤ d H ( S 0 , S ε k ) + d H ( U ε k , U 0 ) ⟶ 0 , \sup_{E\in S_0}\operatorname{dist}(E,U_0)\le d_H(S_0,S_{\varepsilon_k})+d_H(U_{\varepsilon_k},U_0)\longrightarrow0, E ∈ S 0 sup dist ( E , U 0 ) ≤ d H ( S 0 , S ε k ) + d H ( U ε k , U 0 ) ⟶ 0 ,
we have sup E ∈ S 0 dist ( E , U 0 ) = 0 \sup_{E\in\mathcal{S}_0}\operatorname{dist}(E,U_0)=0 sup E ∈ S 0 dist ( E , U 0 ) = 0 . Since U 0 U_0 U 0 is closed, S 0 ⊂ U 0 \mathcal{S}_0\subset U_0 S 0 ⊂ U 0 , and
∑ j = 1 M ( b j − a j ) = lim k → ∞ ∑ j = 1 M ( b ε k , j − a ε k , j ) ≤ C s q n 2 . \sum_{j=1}^{M}(b_j-a_j)=\lim_{k\to\infty}\sum_{j=1}^{M}(b_{\varepsilon_k,j}-a_{\varepsilon_k,j})\le\frac{C_s}{q_n^2}. j = 1 ∑ M ( b j − a j ) = k → ∞ lim j = 1 ∑ M ( b ε k , j − a ε k , j ) ≤ q n 2 C s .
which proves the proposition for d = 0 d=0 d = 0 . □
Spectral covers on I i s o \mathcal{I}_{\mathrm{iso}} I iso .
We now take λ ∈ I i s o \lambda\in\mathcal{I}_{\mathrm{iso}} λ ∈ I iso , so v ( x ) v(x) v ( x ) and b ( x ) b(x) b ( x ) are given by (16) . In this case we have the following commutator matrix K N K_N K N independent of phase. Define
φ = arccos 1 2 λ 1 , κ r = π ( − 1 ) r − 1 cos ( r φ ) sin ( π α r ) , r ≥ 1 , \varphi=\arccos\frac{1}{2\lambda_1},\qquad\kappa_r=\frac{\pi(-1)^{r-1}\cos(r\varphi)}{\sin(\pi\alpha r)},\qquad r\ge1, φ = arccos 2 λ 1 1 , κ r = sin ( π α r ) π ( − 1 ) r − 1 cos ( r φ ) , r ≥ 1 ,
( K N ) j , j + r = κ r , ( K N ) j + r , j = − κ r , ( K N ) j j = 0 , 0 ≤ j < j + r ≤ N − 1. (47) (K_N)_{j,j+r}=\kappa_r,\qquad(K_N)_{j+r,j}=-\kappa_r,\qquad(K_N)_{jj}=0,\qquad0\le j<j+r\le N-1. \tag*{(47)} ( K N ) j , j + r = κ r , ( K N ) j + r , j = − κ r , ( K N ) j j = 0 , 0 ≤ j < j + r ≤ N − 1. (47)
Lemma 2.12. Let N > 2 N>2 N > 2 , B N ( x ) = B ^ N ( x ) B_N(x)=\widehat{B}_N(x) B N ( x ) = B N ( x ) be given by (22) with v ( x ) v(x) v ( x ) , b ( x ) b(x) b ( x ) as defined in (16) , and let K N K_N K N be as defined in (47) , t j t_j t j in (27) . Then
E N ( x ) : = d d x B N ( x ) − [ K N , B N ( x ) ] \mathcal{E}_N(x):=\frac{d}{dx}B_N(x)-[K_N,B_N(x)] E N ( x ) := d x d B N ( x ) − [ K N , B N ( x )] has the following edge matrix elements for 1 ≤ j ≤ N − 2 1\le j\le N-2 1 ≤ j ≤ N − 2 ,
( E N ) 0 , j = ( E N ) j , 0 = ( − 1 ) j + 1 π sin ( π ( j + 1 ) α ) cos ( ( j + 1 ) φ ) b ( t − 1 ) , (\mathcal{E}_N)_{0,j}=(\mathcal{E}_N)_{j,0}=\frac{(-1)^{j+1}\pi}{\sin(\pi(j+1)\alpha)}\cos((j+1)\varphi)b(t_{-1}), ( E N ) 0 , j = ( E N ) j , 0 = sin ( π ( j + 1 ) α ) ( − 1 ) j + 1 π cos (( j + 1 ) φ ) b ( t − 1 ) , ( E N ) j , N − 1 = ( E N ) N − 1 , j = ( − 1 ) N − 1 − j π sin ( π ( N − j ) α ) cos ( ( N − j ) φ ) b ( t N − 1 ) , (48) (\mathcal{E}_N)_{j,N-1}=(\mathcal{E}_N)_{N-1,j}=\frac{(-1)^{N-1-j}\pi}{\sin(\pi(N-j)\alpha)}\cos((N-j)\varphi)b(t_{N-1}), \tag*{(48)} ( E N ) j , N − 1 = ( E N ) N − 1 , j = sin ( π ( N − j ) α ) ( − 1 ) N − 1 − j π cos (( N − j ) φ ) b ( t N − 1 ) , (48) and the corner elements
( E N ) 0 , 0 = − 2 π sin ( π α ) cos ( φ ) b ( t − 1 ) , (\mathcal{E}_N)_{0,0}=-\frac{2\pi}{\sin(\pi\alpha)}\cos(\varphi)b(t_{-1}), ( E N ) 0 , 0 = − sin ( π α ) 2 π cos ( φ ) b ( t − 1 ) , ( E N ) N − 1 , N − 1 = 2 π sin ( π α ) cos ( φ ) b ( t N − 1 ) , (\mathcal{E}_N)_{N-1,N-1}=\frac{2\pi}{\sin(\pi\alpha)}\cos(\varphi)b(t_{N-1}), ( E N ) N − 1 , N − 1 = sin ( π α ) 2 π cos ( φ ) b ( t N − 1 ) , ( E N ) 0 , N − 1 = − κ N − 1 ( v ( t N − 1 ) − v ( t 0 ) ) − κ N − 2 ( b ( t N − 2 ) − b ( t 0 ) ) . (49) (\mathcal{E}_N)_{0,N-1}=-\kappa_{N-1}(v(t_{N-1})-v(t_0))-\kappa_{N-2}(b(t_{N-2})-b(t_0)). \tag*{(49)} ( E N ) 0 , N − 1 = − κ N − 1 ( v ( t N − 1 ) − v ( t 0 )) − κ N − 2 ( b ( t N − 2 ) − b ( t 0 )) . (49) and the rest of the matrix elements of E N ( x ) \mathcal{E}_N(x) E N ( x ) are zero.
Proof. Similarly to Lemma 2.8, we see that E N \mathcal{E}_N E N is symmetric, so it is sufficient to verify its upper-right corner. We proceed in the same way as in the proof of Lemma 2.8 and use the relations (31) . The argument is simpler in this case. In particular, for 1 ≤ j ≤ N − 2 1\le j\le N-2 1 ≤ j ≤ N − 2 , ℓ ≥ 2 \ell\ge2 ℓ ≥ 2 , (31) with the present B B B and K K K gives
[ K , B ] j , j + ℓ = ( − 1 ) ℓ − 1 π sin ( π ℓ α ) cos ( ℓ φ ) ( v ( t j + ℓ ) − v ( t j ) ) + ( − 1 ) ℓ π sin ( π ( ℓ − 1 ) α ) cos ( ( ℓ − 1 ) φ ) ( b ( t j + ℓ − 1 ) − b ( t j ) ) + ( − 1 ) ℓ π sin ( π ( ℓ + 1 ) α ) cos ( ( ℓ + 1 ) φ ) ( b ( t j + ℓ ) − b ( t j − 1 ) ) . (50) \begin{aligned}
[K,B]_{j,j+\ell}
={}&\frac{(-1)^{\ell-1}\pi}{\sin(\pi\ell\alpha)}\cos(\ell\varphi)(v(t_{j+\ell})-v(t_j))\\
&+\frac{(-1)^\ell\pi}{\sin(\pi(\ell-1)\alpha)}\cos((\ell-1)\varphi)(b(t_{j+\ell-1})-b(t_j))\\
&+\frac{(-1)^\ell\pi}{\sin(\pi(\ell+1)\alpha)}\cos((\ell+1)\varphi)(b(t_{j+\ell})-b(t_{j-1})). \tag*{(50)}
\end{aligned} [ K , B ] j , j + ℓ = sin ( π ℓ α ) ( − 1 ) ℓ − 1 π cos ( ℓ φ ) ( v ( t j + ℓ ) − v ( t j )) + sin ( π ( ℓ − 1 ) α ) ( − 1 ) ℓ π cos (( ℓ − 1 ) φ ) ( b ( t j + ℓ − 1 ) − b ( t j )) + sin ( π ( ℓ + 1 ) α ) ( − 1 ) ℓ π cos (( ℓ + 1 ) φ ) ( b ( t j + ℓ ) − b ( t j − 1 )) . (50) Substituting here the explicit formulae for v ( t ) v(t) v ( t ) and b ( t ) b(t) b ( t ) and using the fact that cos φ = 1 / ( 2 λ 1 ) \cos\varphi=1/(2\lambda_1) cos φ = 1/ ( 2 λ 1 ) , we obtain that [ K , B ] j , j + ℓ = 0 [K,B]_{j,j+\ell}=0 [ K , B ] j , j + ℓ = 0 . The other matrix elements are verified similarly. □
Proposition 2.13. Let λ ∈ I i s o \lambda\in\mathcal{I}_{\mathrm{iso}} λ ∈ I iso , α \alpha α be irrational, and n ≥ 4 n\ge4 n ≥ 4 . There are closed intervals I n , 1 , … , I n , M n I_{n,1},\ldots,I_{n,M_n} I n , 1 , … , I n , M n such that
Σ α , λ ⊂ ⋃ j = 1 M n I n , j , M n ≤ q n + q n − 1 , ∑ j = 1 M n ∣ I n , j ∣ ≤ C λ 1 q n , (51) \Sigma_{\alpha,\lambda}\subset\bigcup_{j=1}^{M_n}I_{n,j},\qquad M_n\le q_n+q_{n-1},\qquad\sum_{j=1}^{M_n}|I_{n,j}|\le\frac{C\lambda_1}{q_n}, \tag*{(51)} Σ α , λ ⊂ j = 1 ⋃ M n I n , j , M n ≤ q n + q n − 1 , j = 1 ∑ M n ∣ I n , j ∣ ≤ q n C λ 1 , (51) where C > 0 C>0 C > 0 is absolute.
Proof. Applying similar arguments as in the proof of Lemma 2.10 to E N \mathcal{E}_N E N of Lemma 2.12, we easily see that ∥ E N ∥ S 2 ≤ C λ 1 \|\mathcal{E}_N\|_{S_2} \le C\lambda_1 ∥ E N ∥ S 2 ≤ C λ 1 , ∥ [ K N , P N ] ∥ S 2 ≤ C q n \|[K_N,P_N]\|_{S_2} \le Cq_n ∥ [ K N , P N ] ∥ S 2 ≤ C q n , and therefore Propositions 2.6, 2.7 with J = J n J=J_n J = J n , K = K N K=K_N K = K N , E = E N \mathcal{E}=\mathcal{E}_N E = E N , τ = β = sup k ∣ b ( x k ) ∣ ≤ C λ 1 / q n \tau=\beta=\sup_k|b(x_k)| \le C\lambda_1/q_n τ = β = sup k ∣ b ( x k ) ∣ ≤ C λ 1 / q n , give the desired results. □
Proof of Theorem 2.1. This is the union of Propositions 2.11, 2.13. □
Dimension bounds
The covers in Section 2 give upper bounds for fractal dimensions of the spectrum. For β ( α ) > 0 \beta(\alpha)>0 β ( α ) > 0 (see (2) ), we combine them with lower bounds on packing dimension from [16 ] .
Dimension bounds from spectral covers: proof of Theorems 1.1 and 1.2
For a nonempty compact set Ω ⊂ R \Omega\subset\mathbb{R} Ω ⊂ R , we write N Ω ( ε ) N_\Omega(\varepsilon) N Ω ( ε ) for the least number of intervals of length at most ε \varepsilon ε covering Ω \Omega Ω . Then the lower and upper box dimensions of Ω \Omega Ω are given, respectively, by (see, e.g., [10 ] )
dim ‾ B Ω = lim inf ε ↓ 0 ln N Ω ( ε ) ln ( 1 / ε ) , dim ‾ B Ω = lim sup ε ↓ 0 ln N Ω ( ε ) ln ( 1 / ε ) . \underline{\dim}_{B}\Omega=\liminf_{\varepsilon\downarrow0}\frac{\ln N_\Omega(\varepsilon)}{\ln(1/\varepsilon)},\qquad\overline{\dim}_{B}\Omega=\limsup_{\varepsilon\downarrow0}\frac{\ln N_\Omega(\varepsilon)}{\ln(1/\varepsilon)}. dim B Ω = ε ↓ 0 lim inf ln ( 1/ ε ) ln N Ω ( ε ) , dim B Ω = ε ↓ 0 lim sup ln ( 1/ ε ) ln N Ω ( ε ) .
When these coincide, their common value is the box-counting dimension dim B Ω \dim_B\Omega dim B Ω . The t t t -dimensional Hausdorff measure of Ω \Omega Ω is
meas t ( Ω ) = lim δ ↓ 0 inf { ∑ j ∣ I j ∣ t : Ω ⊂ ⋃ j I j , ∣ I j ∣ ≤ δ } . \operatorname{meas}_t(\Omega)=\lim_{\delta\downarrow0}\inf\left\{\sum_j |I_j|^t:\Omega\subset\bigcup_j I_j,\ |I_j|\le\delta\right\}. meas t ( Ω ) = δ ↓ 0 lim inf { j ∑ ∣ I j ∣ t : Ω ⊂ j ⋃ I j , ∣ I j ∣ ≤ δ } .
where infimum is taken over all finite or countable covers of Ω \Omega Ω by intervals I j I_j I j of length no larger than δ \delta δ . The Hausdorff dimension of Ω \Omega Ω is then dim H ( Ω ) = inf { t > 0 : meas t ( Ω ) < ∞ } \dim_H(\Omega)=\inf\{t>0:\operatorname{meas}_t(\Omega)<\infty\} dim H ( Ω ) = inf { t > 0 : meas t ( Ω ) < ∞ } . The packing dimension is
dim P Ω = inf Ω ⊂ ⋃ j ≥ 1 Ω j sup j dim ‾ B Ω j , \dim_P\Omega=\inf_{\Omega\subset\bigcup_{j\ge1}\Omega_j}\sup_j\overline{\dim}_B\Omega_j, dim P Ω = Ω ⊂ ⋃ j ≥ 1 Ω j inf j sup dim B Ω j ,
where the covering sets Ω j \Omega_j Ω j are bounded. Recall that in contrast with box dimensions, the packing dimension is countably stable: dim P ⋃ j = 1 ∞ Ω j = sup j dim P Ω j \dim_P\bigcup_{j=1}^{\infty}\Omega_j=\sup_j\dim_P\Omega_j dim P ⋃ j = 1 ∞ Ω j = sup j dim P Ω j . Furthermore, recall that the dimensions satisfy the following inequalities:
dim H Ω ≤ dim ‾ B Ω ≤ dim ‾ B Ω , dim H Ω ≤ dim P Ω ≤ dim ‾ B Ω . (52) \dim_H\Omega\le\underline{\dim}_B\Omega\le\overline{\dim}_B\Omega,\qquad\dim_H\Omega\le\dim_P\Omega\le\overline{\dim}_B\Omega. \tag*{(52)} dim H Ω ≤ dim B Ω ≤ dim B Ω , dim H Ω ≤ dim P Ω ≤ dim B Ω. (52)
Proposition 3.1. Let Ω ⊂ R \Omega\subset\mathbb{R} Ω ⊂ R be nonempty and compact, Q n ↑ ∞ Q_n\uparrow\infty Q n ↑ ∞ monotonically as n → ∞ n\to\infty n → ∞ , and κ > 0 \kappa>0 κ > 0 . Suppose that, for n ≥ n 0 n\ge n_0 n ≥ n 0 with some n 0 > 0 n_0>0 n 0 > 0 , there are covers of Ω \Omega Ω by intervals I n , j I_{n,j} I n , j satisfying
Ω ⊂ ⋃ j = 1 M n I n , j , M n ≤ A Q n , ∑ j ∣ I n , j ∣ ≤ B Q n − κ , \Omega\subset\bigcup_{j=1}^{M_n}I_{n,j},\qquad M_n\le AQ_n,\qquad\sum_j|I_{n,j}|\le BQ_n^{-\kappa}, Ω ⊂ j = 1 ⋃ M n I n , j , M n ≤ A Q n , j ∑ ∣ I n , j ∣ ≤ B Q n − κ ,
for some A , B > 0 A,B>0 A , B > 0 independent of n n n . Let d ∗ = 1 / ( 1 + κ ) d_* = 1/(1+\kappa) d ∗ = 1/ ( 1 + κ ) . Then
meas d ∗ ( Ω ) ≤ A 1 − d ∗ B d ∗ , dim ‾ B Ω ≤ d ∗ (53) \operatorname{meas}_{d_*}(\Omega)\le A^{1-d_*}B^{d_*},\qquad\underline{\dim}_B\Omega\le d_* \tag*{(53)} meas d ∗ ( Ω ) ≤ A 1 − d ∗ B d ∗ , dim B Ω ≤ d ∗ (53)
Moreover,
dim ‾ B Ω ≤ lim sup n ln Q n + 1 κ ln Q n + ln Q n + 1 . (54) \overline{\dim}_B\Omega\le\limsup_{n}\frac{\ln Q_{n+1}}{\kappa\ln Q_n+\ln Q_{n+1}}. \tag*{(54)} dim B Ω ≤ n lim sup κ ln Q n + ln Q n + 1 ln Q n + 1 . (54)
Proof. For 0 < t ≤ 1 0<t\le1 0 < t ≤ 1 , Hölder inequality gives
∑ j = 1 M n ∣ I n , j ∣ t ≤ M n 1 − t ( ∑ j = 1 M n ∣ I n , j ∣ ) t ≤ A 1 − t B t Q n 1 − ( 1 + κ ) t . \sum_{j=1}^{M_n}|I_{n,j}|^t\le M_n^{1-t}\left(\sum_{j=1}^{M_n}|I_{n,j}|\right)^t\le A^{1-t}B^tQ_n^{1-(1+\kappa)t}. j = 1 ∑ M n ∣ I n , j ∣ t ≤ M n 1 − t ( j = 1 ∑ M n ∣ I n , j ∣ ) t ≤ A 1 − t B t Q n 1 − ( 1 + κ ) t .
Since max j ∣ I n , j ∣ ≤ B Q n − κ → 0 \max_j|I_{n,j}|\le BQ_n^{-\kappa}\to0 max j ∣ I n , j ∣ ≤ B Q n − κ → 0 , taking t = d ∗ t=d_* t = d ∗ shows that meas d ∗ ( Ω ) ≤ A 1 − d ∗ B d ∗ \operatorname{meas}_{d_*}(\Omega)\le A^{1-d_*}B^{d_*} meas d ∗ ( Ω ) ≤ A 1 − d ∗ B d ∗ . This immediately implies that dim H Ω ≤ d ∗ \dim_H\Omega\le d_* dim H Ω ≤ d ∗ . However, we will now obtain a stronger (in view of (52)) result on dim ‾ B Ω \underline{\dim}_B\Omega dim B Ω . Dividing the intervals I n , j I_{n,j} I n , j into subintervals of length ε \varepsilon ε , we obtain
N Ω ( ε ) ≤ ∑ j = 1 M n ( 1 ε ∣ I n , j ∣ + 1 ) = 1 ε ∑ j = 1 M n ∣ I n , j ∣ + M n ≤ B ε Q n ϰ + A Q n , n ≥ n 0 . (55) N_{\Omega}(\varepsilon) \le\sum_{j=1}^{M_n}\left(\frac{1}{\varepsilon}|I_{n,j}|+1\right)=\frac{1}{\varepsilon}\sum_{j=1}^{M_n}|I_{n,j}|+M_n\le\frac{B}{\varepsilon Q_n^{\varkappa}}+AQ_n,\qquad n\ge n_0. \tag*{(55)} N Ω ( ε ) ≤ j = 1 ∑ M n ( ε 1 ∣ I n , j ∣ + 1 ) = ε 1 j = 1 ∑ M n ∣ I n , j ∣ + M n ≤ ε Q n ϰ B + A Q n , n ≥ n 0 . (55)
Let
ε j = 1 Q j 1 + ϰ , j ≥ n 0 . (56) \varepsilon_j=\frac{1}{Q_j^{1+\varkappa}},\qquad j\ge n_0. \tag*{(56)} ε j = Q j 1 + ϰ 1 , j ≥ n 0 . (56)
We take n 0 n_0 n 0 sufficiently large so that ε n 0 < 1 \varepsilon_{n_0}<1 ε n 0 < 1 . Setting ε = ε n \varepsilon=\varepsilon_n ε = ε n in (55) we obtain that N Ω ( ε n ) ≤ ( A + B ) Q n = ( A + B ) ε n − d ∗ N_{\Omega}(\varepsilon_n)\le(A+B)Q_n=(A+B)\varepsilon_n^{-d_*} N Ω ( ε n ) ≤ ( A + B ) Q n = ( A + B ) ε n − d ∗ , which immediately implies that dim B Ω ≤ d ∗ \dim_B\Omega\le d_* dim B Ω ≤ d ∗ .
It remains to estimate dim B Ω \dim_B\Omega dim B Ω . Let ε n + 1 ≤ ε ≤ ε n \varepsilon_{n+1}\le\varepsilon\le\varepsilon_n ε n + 1 ≤ ε ≤ ε n . First, we note that by (55),
N Ω ( ε ) ≤ B ε Q n + 1 ϰ + A Q n + 1 ≤ B ε n + 1 Q n + 1 ϰ + A Q n + 1 = ( A + B ) Q n + 1 . (57) N_{\Omega}(\varepsilon)\le\frac{B}{\varepsilon Q_{n+1}^{\varkappa}}+AQ_{n+1}\le\frac{B}{\varepsilon_{n+1}Q_{n+1}^{\varkappa}}+AQ_{n+1}=(A+B)Q_{n+1}. \tag*{(57)} N Ω ( ε ) ≤ ε Q n + 1 ϰ B + A Q n + 1 ≤ ε n + 1 Q n + 1 ϰ B + A Q n + 1 = ( A + B ) Q n + 1 . (57)
On the other hand, since
Q n = 1 ε n 1 / ( 1 + ϰ ) ≤ 1 ε 1 / ( 1 + ϰ ) = ε n ϰ / ( 1 + ϰ ) ε ≤ ε n ϰ / ( 1 + ϰ ) ε = 1 ε Q n ϰ , Q_n=\frac{1}{\varepsilon_n^{1/(1+\varkappa)}}\le\frac{1}{\varepsilon^{1/(1+\varkappa)}}=\frac{\varepsilon_n^{\varkappa/(1+\varkappa)}}{\varepsilon}\le\frac{\varepsilon_n^{\varkappa/(1+\varkappa)}}{\varepsilon}=\frac{1}{\varepsilon Q_n^{\varkappa}}, Q n = ε n 1/ ( 1 + ϰ ) 1 ≤ ε 1/ ( 1 + ϰ ) 1 = ε ε n ϰ / ( 1 + ϰ ) ≤ ε ε n ϰ / ( 1 + ϰ ) = ε Q n ϰ 1 ,
we also have
N Ω ( ε ) ≤ B ε Q n ϰ + A Q n ≤ A + B ε Q n ϰ . (58) N_{\Omega}(\varepsilon)\le\frac{B}{\varepsilon Q_n^{\varkappa}}+AQ_n\le\frac{A+B}{\varepsilon Q_n^{\varkappa}}. \tag*{(58)} N Ω ( ε ) ≤ ε Q n ϰ B + A Q n ≤ ε Q n ϰ A + B . (58)
Together, (57) and (58) give
N Ω ( ε ) ≤ ( A + B ) min ( 1 ε Q n ϰ , Q n + 1 ) , ε n + 1 ≤ ε ≤ ε n . (59) N_{\Omega}(\varepsilon)\le(A+B)\min\left(\frac{1}{\varepsilon Q_n^{\varkappa}},Q_{n+1}\right),\qquad\varepsilon_{n+1}\le\varepsilon\le\varepsilon_n. \tag*{(59)} N Ω ( ε ) ≤ ( A + B ) min ( ε Q n ϰ 1 , Q n + 1 ) , ε n + 1 ≤ ε ≤ ε n . (59)
Denote x = ln ( 1 / ε ) x=\ln(1/\varepsilon) x = ln ( 1/ ε ) , x ∗ = ϰ ln Q n + ln Q n + 1 x_*=\varkappa\ln Q_n+\ln Q_{n+1} x ∗ = ϰ ln Q n + ln Q n + 1 . Note that
ln 1 ε n = ( 1 + ϰ ) ln Q n ≤ x ∗ ≤ ( 1 + ϰ ) ln Q n + 1 = ln 1 ε n + 1 \ln\frac{1}{\varepsilon_n}=(1+\varkappa)\ln Q_n\le x_*\le(1+\varkappa)\ln Q_{n+1}=\ln\frac{1}{\varepsilon_{n+1}} ln ε n 1 = ( 1 + ϰ ) ln Q n ≤ x ∗ ≤ ( 1 + ϰ ) ln Q n + 1 = ln ε n + 1 1
and
f ( x ) : = 1 x min ( x − ϰ ln Q n , ln Q n + 1 ) = { 1 − ϰ ln Q n / x , x ≤ x ∗ , ln Q n + 1 / x , x ≥ x ∗ . f(x):=\frac{1}{x}\min(x-\varkappa\ln Q_n,\ln Q_{n+1})=
\begin{cases}
1-\varkappa\ln Q_n/x,&x\le x_*,\\
\ln Q_{n+1}/x,&x\ge x_*.
\end{cases} f ( x ) := x 1 min ( x − ϰ ln Q n , ln Q n + 1 ) = { 1 − ϰ ln Q n / x , ln Q n + 1 / x , x ≤ x ∗ , x ≥ x ∗ .
Clearly, this function reaches its maximum on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) at x = x ∗ x=x_* x = x ∗ , and we have
f ( x ) ≤ ln Q n + 1 ϰ ln Q n + ln Q n + 1 , ε n + 1 ≤ ε ≤ ε n . f(x)\le\frac{\ln Q_{n+1}}{\varkappa\ln Q_n+\ln Q_{n+1}},\qquad\varepsilon_{n+1}\le\varepsilon\le\varepsilon_n. f ( x ) ≤ ϰ ln Q n + ln Q n + 1 ln Q n + 1 , ε n + 1 ≤ ε ≤ ε n .
Therefore (59) implies
ln N Ω ( ε ) ln 1 / ε ≤ ln ( A + B ) ln 1 / ε + f ( x ) ≤ ln ( A + B ) ln 1 / ε + ln Q n + 1 ϰ ln Q n + ln Q n + 1 , (60) \frac{\ln N_{\Omega}(\varepsilon)}{\ln1/\varepsilon}\le\frac{\ln(A+B)}{\ln1/\varepsilon}+f(x)\le\frac{\ln(A+B)}{\ln1/\varepsilon}+\frac{\ln Q_{n+1}}{\varkappa\ln Q_n+\ln Q_{n+1}}, \tag*{(60)} ln 1/ ε ln N Ω ( ε ) ≤ ln 1/ ε ln ( A + B ) + f ( x ) ≤ ln 1/ ε ln ( A + B ) + ϰ ln Q n + ln Q n + 1 ln Q n + 1 , (60)
and taking lim sup \limsup lim sup gives (54). □
Proof of Theorems 1.1,1.2. We apply Theorem 2.1 together with Proposition 3.1 where Ω = Σ α , λ \Omega=\Sigma_{\alpha,\lambda} Ω = Σ α , λ with λ ∈ C 0 \lambda\in C_0 λ ∈ C 0 , and Q n = q n Q_n=q_n Q n = q n . By Theorem 2.1 we can take ϰ = 1 \varkappa=1 ϰ = 1 , improving to ϰ = 2 \varkappa=2 ϰ = 2 on B i s o B_{\mathrm{iso}} B iso . This with (52) proves Theorem 1.1. Theorem 1.2 follows from (54) since for positive C C C , the function t / ( C + t ) t/(C+t) t / ( C + t ) , t > 0 t>0 t > 0 , is continuous and strictly increasing. □
Packing lower bounds and absence of box-counting dimension: Proof of Corollary 1.3.
For a finite nonzero measure μ \mu μ on R \mathbb{R} R , its packing dimension is
dim P μ = inf { dim P F : μ ( R ∖ F ) = 0 } . \dim_{\mathrm{P}} \mu= \inf\{\dim_{\mathrm{P}} F : \mu(\mathbb{R} \setminus F) = 0\}. dim P μ = inf { dim P F : μ ( R ∖ F ) = 0 } .
For a bounded self-adjoint operator H H H on a Hilbert space, the spectral measure associated with f ≠ 0 f \ne0 f = 0 , is defined in terms of the inner product as μ f ( B ) = ( f , 1 B ( H ) f ) \mu_f(B) = (f, \mathbf{1}_B(H)f) μ f ( B ) = ( f , 1 B ( H ) f ) , where B B B is a Borel set in R \mathbb{R} R and 1 B \mathbf{1}_B 1 B its indicator function. The spectral measure satisfies
supp μ f ⊂ spec ( H ) , dim P μ f ≤ dim P spec ( H ) ≤ 1. \operatorname{supp}\mu_f \subset\operatorname{spec}(H), \qquad\dim_{\mathrm{P}}\mu_f \le\dim_{\mathrm{P}}\operatorname{spec}(H) \le1. supp μ f ⊂ spec ( H ) , dim P μ f ≤ dim P spec ( H ) ≤ 1.
We assume β ( α ) > 0 \beta(\alpha) > 0 β ( α ) > 0 and take the parameters in Corollary 1.3:
λ ∈ { λ A , λ T } ∪ ( ( I i s o ∪ B i s o ) ∩ { λ 2 > 0 } ) . \lambda\in\{\lambda_A,\lambda_T\} \cup\left((I_{\mathrm{iso}}\cup B_{\mathrm{iso}}) \cap\{\lambda_2 > 0\}\right). λ ∈ { λ A , λ T } ∪ ( ( I iso ∪ B iso ) ∩ { λ 2 > 0 } ) .
By [16 Corollary 2.6(3) and (2.18)] , in this case there is a spectral measure μ f \mu_f μ f of H α , θ , λ H_{\alpha,\theta,\lambda} H α , θ , λ such that dim P μ f = 1 \dim_{\mathrm{P}}\mu_f = 1 dim P μ f = 1 . Therefore, dim P Σ α , λ = 1 \dim_{\mathrm{P}}\Sigma_{\alpha,\lambda} = 1 dim P Σ α , λ = 1 , and hence (by (52) ) also dim ‾ B Σ α , λ = 1 \overline{\dim}_{\mathrm{B}}\Sigma_{\alpha,\lambda} = 1 dim B Σ α , λ = 1 . For the almost Mathieu operator, the equality dim P Σ α , λ = 1 \dim_{\mathrm{P}}\Sigma_{\alpha,\lambda} = 1 dim P Σ α , λ = 1 also follows from [23 Corollary 4] . By Theorem 1.1 we then obtain (3). In particular, we see that the lower and upper box dimensions are different so that there is no box-counting dimension. This proves Corollary 1.3.
References
Bocconi University, Milan, Italy
University of California, Berkeley, USA, and University of California, Irvine, USA
Imperial College London, UK, and SISSA, Trieste, Italy