Introduction
Quantitative mixing estimates are an important tool in the study of statistical limit laws, including central and local limit theorems, Berry–Esseen estimates, and large deviation principles. For systems with a neutral direction, proving exponential mixing often requires a cancellation mechanism, first exploited by Dolgopyat for Anosov flows [7 ] and compact group extensions of hyperbolic systems [8 ] . For a circle extension ( x , z ) ↦ ( f ( x ) , z + τ ( x ) ) (x,z) \mapsto(f(x),z+\tau(x)) ( x , z ) ↦ ( f ( x ) , z + τ ( x )) over an expanding map, Fourier decomposition in the fibre leads to the twisted transfer operators L ν = L f ( e i ν τ ⋅ ) \mathcal{L}_{\nu}=\mathcal{L}_{f}(e^{i\nu\tau}\cdot) L ν = L f ( e i ν τ ⋅ ) . A Dolgopyat-type inequality gives quantitative bounds on the twisted transfer operators at large frequencies. Together with suitable spectral information at bounded frequencies, these bounds yield exponential mixing and higher-order statistical expansions, as we show below. For results on mixing and related spectral estimates for extensions of expanding maps see [8 , 12 , 9 , 15 , 5 , 3 ] ; for suspension semiflows, see [18 ] . The cancellation method underlying Dolgopyat-type estimates originates in [7 ] and it was used even in the presence of discontinuities [2 ] . However, if the system does not have a skew-product structure, the mixing problem becomes much harder to treat, and even proving existence of invariant measures is highly non-trivial.
The main technical result of this paper is the Dolgopyat estimate (Theorem 2.10 ) for a broad class of two-dimensional partially hyperbolic system for which existence of finitely many invariant mixing densities is known. We consider the strongly dominated vertical partially hyperbolic maps (SVPHs) of [4 ] , restricted The paper was started when RC was working at Scuola Normale Superiore di Pisa, and he was partially supported by the research project ‘Dynamics and Information Research Institute—Quantum Information, Quantum Technologies’ within the agreement between UniCredit Bank and Scuola Normale Superiore and by the research project PRIN 2017S35EHN of the MIUR. The authors acknowledge the UMI Group “DinAmicI” (www.dinamici.org ) and the INdAM group GNFM. Finally, it is a pleasure to thank the three Cs, Claudio Bonanno, Carlo Collari and Carlangelo Liverani, for helpful discussions. to a class of virtually expanding maps in the sense of Tsujii [19 ] . The latter restriction is made to work with the classical Sobolev spaces H s ( T 2 ) H^s(\mathbb{T}^2) H s ( T 2 ) . Extending the argument to the anisotropic spaces of [4 ] is possible, but it requires additional operator and localisation estimates. The main difficulty faced by the paper is that inverse branches need not contract the central direction, so the standard expanding-base arguments do not apply directly. Our result apply in particular to the extension
F ~ ( x , θ , z ) = ( F ( x , θ ) , z + τ ( x , θ ) ) ( m o d Z 3 ) , \widetilde{F}(x,\theta,z) = (F(x,\theta),z+\tau(x,\theta)) \pmod{\mathbb{Z}^3}, F ( x , θ , z ) = ( F ( x , θ ) , z + τ ( x , θ )) ( mod Z 3 ) ,
where F F F is partially hyperbolic, under a quantitative cotangent transversality condition analogous to the partial captivity conditions in [9 , 15 ] .
We give two important applications of Theorem 2.10 . Theorem A establishes Edgeworth expansions and higher order mixing global and local limit expansions for the Birkhoff sums of τ \tau τ over F F F . Several ergodic acips are allowed and the leading distribution is a mixture of centred Gaussians, with weights determined by the initial density. The spectral gap and perturbation near zero give the Gaussian terms, while the Dolgopyat estimate controls the large real frequencies needed for expansions of arbitrary order. This application uses the spectral methods developed in [13 , 10 , 11 ] . Theorem B establishes exponential mixing of F ~ \widetilde{F} F for Hölder observables.
For skew products F ( x , θ ) = ( f ( x ) , θ + ω ( x ) ) F(x,\theta)=(f(x),\theta+\omega(x)) F ( x , θ ) = ( f ( x ) , θ + ω ( x )) with τ = τ ( x ) \tau=\tau(x) τ = τ ( x ) , the extension is a T 2 \mathbb{T}^2 T 2 -extension of f f f with cocycle ( ω , τ ) (\omega,\tau) ( ω , τ ) . In that case exponential mixing is already known from [5 ] when a ω + b τ a\omega+b\tau aω + b τ is not cohomologous to a constant for every ( a , b ) ≠ ( 0 , 0 ) (a,b)\ne(0,0) ( a , b ) = ( 0 , 0 ) . Lemma A.1 identifies (B1) with this condition (see also Remark 2.8 for its necessity for DI at all large real frequencies). Our estimate also applies when F F F is not a skew product, or when τ \tau τ depends on the central coordinate, provided the stated hypotheses hold.
The proof combines Lasota–Yorke inequalities with an L 2 L^2 L 2 estimate at times of order log ∣ ν ∣ \log|\nu| log ∣ ν ∣ . We localise in space at scale ∣ ν ∣ − 1 / 2 |\nu|^{-1/2} ∣ ν ∣ − 1/2 and in frequency. Condition (B1) controls frequencies reached by many inverse branches and non-stationary phase controls the remaining frequencies.
§ 2 gives the assumptions, the technical theorem, and its two applications. § 3 develops the spectral tools. § 4 proves the Dolgopyat estimate, followed by the proofs of Theorems A and B in § 5 and § 6 . Appendix § 7 discuss the main assumption in the skew-product case and provide a simple example to which our results apply and, finally, Appendix § 8 recalls the needed estimates from [4 ] .
Setting and main results
Assumptions on the map
We first introduce the systems F ∈ C ∞ ( T 2 , T 2 ) F\in C^\infty(\mathbb{T}^2,\mathbb{T}^2) F ∈ C ∞ ( T 2 , T 2 ) studied in [4 ] and called strongly dominated vertical partially hyperbolic systems , or SVPHs for short.1 For the convenience of the reader we recall the main assumptions. A set of sufficient conditions for checking whether a system is an SVPH are provided in [4 Lemma 2.2] . We refer to [4 Section 2.2] for a detailed discussion, including examples, of assumptions below, and we refer to the Appendix § 8 for a collection of technical results about SVPHs.
Let F : T 2 → T 2 F:\mathbb{T}^2\to\mathbb{T}^2 F : T 2 → T 2 be a surjective C ∞ C^\infty C ∞ local diffeomorphism of the form
F ( x , θ ) = ( f ( x , θ ) , θ + ω ( x , θ ) ) ( m o d Z 2 ) , ω ∈ C ∞ ( T 2 , R ) . (1) F(x,\theta)=(f(x,\theta),\theta+\omega(x,\theta)) \pmod{\mathbb{Z}^2}, \qquad\omega\in C^\infty(\mathbb{T}^2,\mathbb{R}). \tag*{(1)} F ( x , θ ) = ( f ( x , θ ) , θ + ω ( x , θ )) ( mod Z 2 ) , ω ∈ C ∞ ( T 2 , R ) . (1)
We call F F F an SVPH if F F F satisfies the following properties (A1)–(A5) .
(A1) For all p ∈ T 2 p \in\mathbb{T}^{2} p ∈ T 2 we have det ( D p F ) > 0 \det(D_{p}F)>0 det ( D p F ) > 0 .
(A2) There exist χ u ∈ ( 0 , 1 ) \chi_{u} \in(0,1) χ u ∈ ( 0 , 1 ) , χ c ∈ ( 0 , 1 ] \chi_{c} \in(0,1] χ c ∈ ( 0 , 1 ] and 0 < λ c − < 1 < λ c + < λ ≤ Λ 0<\lambda_{c}^{-}<1<\lambda_{c}^{+}<\lambda\leq\Lambda 0 < λ c − < 1 < λ c + < λ ≤ Λ such that, setting
C u : = { ( ξ , η ) ∈ T z T 2 : ∣ η ∣ ≤ χ u ∣ ξ ∣ } C_{u}:=\{(\xi,\eta)\in T_{z}\mathbb{T}^{2}:|\eta|\leq\chi_{u}|\xi|\} C u := {( ξ , η ) ∈ T z T 2 : ∣ η ∣ ≤ χ u ∣ ξ ∣ }
C c : = { ( ξ , η ) ∈ T z T 2 : ∣ ξ ∣ ≤ χ c ∣ η ∣ } , (2) C_{c}:=\{(\xi,\eta)\in T_{z}\mathbb{T}^{2}:|\xi|\leq\chi_{c}|\eta|\}, \tag*{(2)} C c := {( ξ , η ) ∈ T z T 2 : ∣ ξ ∣ ≤ χ c ∣ η ∣ } , (2)
defining
λ n − ( z ) : = inf v ∈ R 2 ∖ C c m F n ( z , v ) λ n + ( z ) : = sup v ∈ R 2 ∖ C c m F n ( z , v ) , \lambda_{n}^{-}(z):=\inf_{v\in\mathbb{R}^{2}\setminus C_{c}}\mathbf{m}_{F^{n}}(z,v)
\qquad
\lambda_{n}^{+}(z):=\sup_{v\in\mathbb{R}^{2}\setminus C_{c}}\mathbf{m}_{F^{n}}(z,v), λ n − ( z ) := v ∈ R 2 ∖ C c inf m F n ( z , v ) λ n + ( z ) := v ∈ R 2 ∖ C c sup m F n ( z , v ) ,
λ c , n − ( z ) : = inf v ∈ C c ∖ { 0 } m F n ∗ ( z , v ) λ c , n + ( z ) : = sup v ∈ C c ∖ { 0 } m F n ∗ ( z , v ) , (3) \lambda_{c,n}^{-}(z):=\inf_{v\in C_{c}\setminus\{0\}}\mathbf{m}_{F^{n}}^{*}(z,v)
\qquad
\lambda_{c,n}^{+}(z):=\sup_{v\in C_{c}\setminus\{0\}}\mathbf{m}_{F^{n}}^{*}(z,v), \tag*{(3)} λ c , n − ( z ) := v ∈ C c ∖ { 0 } inf m F n ∗ ( z , v ) λ c , n + ( z ) := v ∈ C c ∖ { 0 } sup m F n ∗ ( z , v ) , (3)
where m F ( z , v ) = ∥ D z F v ∥ ∥ v ∥ \mathbf{m}_{F}(z,v)=\frac{\lVert D_{z}Fv\rVert}{\lVert v\rVert} m F ( z , v ) = ∥ v ∥ ∥ D z F v ∥ for v ∈ T z T 2 v\in T_{z}\mathbb{T}^{2} v ∈ T z T 2 , and m F ∗ ( z , v ) = ∥ ( D z F ) − 1 v ∥ ∥ v ∥ \mathbf{m}_{F}^{*}(z,v)=\frac{\lVert(D_{z}F)^{-1}v\rVert}{\lVert v\rVert} m F ∗ ( z , v ) = ∥ v ∥ ∥( D z F ) − 1 v ∥ for v ∈ T F ( z ) T 2 v\in T_{F(z)}\mathbb{T}^{2} v ∈ T F ( z ) T 2 , and letting λ n − = inf z λ n − ( z ) \lambda_{n}^{-}=\inf_{z}\lambda_{n}^{-}(z) λ n − = inf z λ n − ( z ) and λ n + = sup z λ n + ( z ) \lambda_{n}^{+}=\sup_{z}\lambda_{n}^{+}(z) λ n + = sup z λ n + ( z ) we assume the following:
There exist constants C ⋆ ≥ 1 C_{\star}\geq1 C ⋆ ≥ 1 and ι ⋆ ∈ ( 0 , 1 ) \iota_{\star}\in(0,1) ι ⋆ ∈ ( 0 , 1 ) such that, for all z ∈ T 2 z\in\mathbb{T}^{2} z ∈ T 2 and n ∈ N n\in\mathbb{N} n ∈ N ,2
D z F C u ⊂ { ( ξ , η ) : ∣ η ∣ ≤ ι ⋆ χ u ∣ ξ ∣ } ⋐ C u ; D_{z}F C_{u}\subset\{(\xi,\eta):|\eta|\leq\iota_{\star}\chi_{u}|\xi|\}\Subset C_{u}; D z F C u ⊂ {( ξ , η ) : ∣ η ∣ ≤ ι ⋆ χ u ∣ ξ ∣ } ⋐ C u ;
( D z F ) − 1 C c ⊂ { ( ξ , η ) : ∣ ξ ∣ ≤ ι ⋆ χ c ∣ η ∣ } , (4) (D_{z}F)^{-1}C_{c}\subset\{(\xi,\eta):|\xi|\leq\iota_{\star}\chi_{c}|\eta|\}, \tag*{(4)} ( D z F ) − 1 C c ⊂ {( ξ , η ) : ∣ ξ ∣ ≤ ι ⋆ χ c ∣ η ∣ } , (4)
C ⋆ − 1 ( λ c − ) n ≤ λ c , n − ( z ) ≤ λ c , n + ( z ) ≤ C ⋆ ( λ c + ) n ; C_{\star}^{-1}(\lambda_{c}^{-})^{n}\leq\lambda_{c,n}^{-}(z)\leq\lambda_{c,n}^{+}(z)\leq C_{\star}(\lambda_{c}^{+})^{n}; C ⋆ − 1 ( λ c − ) n ≤ λ c , n − ( z ) ≤ λ c , n + ( z ) ≤ C ⋆ ( λ c + ) n ;
C ⋆ − 1 λ n ≤ λ n − ( z ) ≤ λ n + ( z ) ≤ C ⋆ Λ n . (5) C_{\star}^{-1}\lambda^{n}\leq\lambda_{n}^{-}(z)\leq\lambda_{n}^{+}(z)\leq C_{\star}\Lambda^{n}. \tag*{(5)} C ⋆ − 1 λ n ≤ λ n − ( z ) ≤ λ n + ( z ) ≤ C ⋆ Λ n . (5)
(A3) Let Υ \Upsilon Υ be the family of closed curves γ ∈ C ∞ ( T , T 2 ) \gamma\in C^{\infty}(\mathbb{T},\mathbb{T}^{2}) γ ∈ C ∞ ( T , T 2 ) such that3
c0) γ ′ ≠ 0 \gamma'\neq0 γ ′ = 0 ,
c1) γ \gamma γ has homotopy class ( 0 , 1 ) (0,1) ( 0 , 1 ) ,
c2) γ ′ ( t ) ∈ C c \gamma'(t)\in C_{c} γ ′ ( t ) ∈ C c , for each t ∈ T t\in\mathbb{T} t ∈ T .
We assume that for each γ ∈ Υ \gamma\in\Upsilon γ ∈ Υ there exist { γ i } i = 1 N ⊂ Υ \{\gamma_{i}\}_{i=1}^{N}\subset\Upsilon { γ i } i = 1 N ⊂ Υ such that F − 1 ( γ ) = ⋃ i = 1 N γ i F^{-1}(\gamma)=\bigcup_{i=1}^{N}\gamma_{i} F − 1 ( γ ) = ⋃ i = 1 N γ i .
(A4) We assume that F F F satisfies the pinching condition4
max { λ c + , ( λ c − ) − 1 } < λ . (6) \max\{\lambda_{c}^{+},(\lambda_{c}^{-})^{-1}\}<\lambda. \tag*{(6)} max { λ c + , ( λ c − ) − 1 } < λ . (6)
(A5) With the notation (2.1) we require, for each p ∈ T 2 p\in\mathbb{T}^{2} p ∈ T 2 ,
∂ x f ( p ) > max { 2 ( 1 + ∥ ∂ x ω ∥ ∞ ) , ∣ ∂ θ f ( p ) ∣ } . \partial_{x}f(p)>\max\{2(1+\lVert\partial_{x}\omega\rVert_{\infty}),|\partial_{\theta}f(p)|\}. ∂ x f ( p ) > max { 2 ( 1 + ∥ ∂ x ω ∥ ∞ ) , ∣ ∂ θ f ( p ) ∣ } .
From now on we set
λ c : = max { λ c + , 1 / λ c − } ≥ 1 , (7) \lambda_c := \max\{\lambda_c^+,1/\lambda_c^-\} \ge1, \tag*{(7)} λ c := max { λ c + , 1/ λ c − } ≥ 1 , (7)
so that (6) reads λ c < λ \lambda_c < \lambda λ c < λ .
Next, the spectral gap for the transfer operator associated to F F F allows for nontrivial peripheral eigenvalues, corresponding to finite cyclic decompositions of the ergodic components. To state our statistical applications without passing to iterates, we impose the following assumption besides (A1)–(A5).
(A6) Every ergodic absolutely continuous invariant probability measure of F F F is weakly mixing.
Finally, note that since F F F is a local diffeomorphism of the compact torus, so we can fix constants C ∞ ≥ 1 C_\infty\ge1 C ∞ ≥ 1 and Θ ≥ 1 \Theta\ge1 Θ ≥ 1 such that
∥ L F n 1 ∥ ∞ ≤ C ∞ Θ n for all n ∈ N . (8) \|\mathcal{L}_F^n 1\|_\infty\le C_\infty\Theta^n \qquad\text{for all } n \in\mathbb{N}. \tag*{(8)} ∥ L F n 1 ∥ ∞ ≤ C ∞ Θ n for all n ∈ N . (8)
Example 2.1. Notable examples of SVPH are partially expanding systems with neutral or close to neutral central direction which includes:
Skew torus extensions over smooth expanding map (and small perturbations of it): F ( x , θ ) = ( f ( x ) , θ + ω ( x ) ) F(x,\theta)=(f(x),\theta+\omega(x)) F ( x , θ ) = ( f ( x ) , θ + ω ( x )) on T 2 \mathbb{T}^2 T 2 where f f f is expanding.
Fast-slow partially expanding systems: F ε ( x , θ ) = ( f ( x , θ ) , θ + ε ω ( x , θ ) ) F_\varepsilon(x,\theta)=(f(x,\theta),\theta+\varepsilon\omega(x,\theta)) F ε ( x , θ ) = ( f ( x , θ ) , θ + ε ω ( x , θ )) on T 2 \mathbb{T}^2 T 2 where inf ( x , θ ) ∈ T 2 ∂ x f ( x , θ ) > 1 \inf_{(x,\theta)\in\mathbb{T}^2}\partial_x f(x,\theta)>1 inf ( x , θ ) ∈ T 2 ∂ x f ( x , θ ) > 1 and ε > 0 \varepsilon>0 ε > 0 is small, under the further assumptions of [4 Section 10] .
One can check that the first example satisfies the assumptions in [4 Lemma 2.2] and hence is an SVPH, provided the expansion is large enough.5 For a detailed discussion of the second, see [4 Section 10] and [6 , 14 ] .
The main statistical properties of SVPHs (existence of invariant measures, regularity, decay of correlations) are proved in [4 ] under a transversality property of the unstable cones, which is generic [17 ] . We recall it, as it will appear in the case of skew products (Lemma A.1 ).
Definition 2.2. Given n ∈ N n \in\mathbb{N} n ∈ N , y ∈ T 2 y \in\mathbb{T}^2 y ∈ T 2 and z 1 , z 2 ∈ F − n ( y ) z_1,z_2 \in F^{-n}(y) z 1 , z 2 ∈ F − n ( y ) , we say that z 1 z_1 z 1 is transversal to z 2 z_2 z 2 (at time n n n ) if D z 1 F n C u ∩ D z 2 F n C u = { 0 } D_{z_1}F^n C_u \cap D_{z_2}F^n C_u = \{0\} D z 1 F n C u ∩ D z 2 F n C u = { 0 } , and we write z 1 ⋔ n z 2 z_1 \pitchfork_n z_2 z 1 ⋔ n z 2 . We set
N F ( n ) : = sup y ∈ T 2 sup z 1 ∈ F − n ( y ) ∑ z 2 ∈ F − n ( y ) z 2 ̸ ⋔ n z 1 ∣ det D z 2 F n ∣ − 1 , \mathcal{N}_F(n):=\sup_{y\in\mathbb{T}^2}\sup_{z_1\in F^{-n}(y)}\sum_{\substack{z_2\in F^{-n}(y)\\z_2\not\pitchfork_n z_1}}\left|\det D_{z_2}F^n\right|^{-1}, N F ( n ) := y ∈ T 2 sup z 1 ∈ F − n ( y ) sup z 2 ∈ F − n ( y ) z 2 ⋔ n z 1 ∑ ∣ det D z 2 F n ∣ − 1 , and we say that F F F satisfies the transversality condition if
(T) there exist C > 0 C>0 C > 0 , γ 0 > 0 \gamma_0>0 γ 0 > 0 such that N F ( n ) ≤ C e − γ 0 n \mathcal{N}_F(n)\le Ce^{-\gamma_0 n} N F ( n ) ≤ C e − γ 0 n for all n ∈ N n\in\mathbb{N} n ∈ N .
Remark 2.3. The constants in (8) always exist. The relevant restriction is the compatibility condition (P) below, which requires an admissible growth rate Θ \Theta Θ sufficiently close to 1 relative to the spectral and transversality rates. By (B.4), (8) holds for every SVPH with Θ = λ c \Theta=\lambda_c Θ = λ c . For skew products F ( x , θ ) = ( f ( x ) , θ + ω ( x ) ) F(x,\theta)=(f(x),\theta+\omega(x)) F ( x , θ ) = ( f ( x ) , θ + ω ( x )) over expanding maps it holds with Θ = 1 \Theta=1 Θ = 1 , since then ∑ y ∈ F − n ( x , θ ) ∣ det D y F n ∣ − 1 = ∑ x ′ ∈ f − n ( x ) ∣ ( f n ) ′ ( x ′ ) ∣ − 1 \sum_{y\in F^{-n}(x,\theta)}|\det D_yF^n|^{-1}=\sum_{x'\in f^{-n}(x)}|(f^n)'(x')|^{-1} ∑ y ∈ F − n ( x , θ ) ∣ det D y F n ∣ − 1 = ∑ x ′ ∈ f − n ( x ) ∣ ( f n ) ′ ( x ′ ) ∣ − 1 , which is bounded uniformly in n n n . In our results, Θ \Theta Θ enters only through the condition (P) below, which requires Θ \Theta Θ to be close to 1 in terms of the other rates.
Virtually expanding maps
We restrict to the class of systems introduced in [19 ] . This allows us to work directly with the classical Sobolev spaces H s ( T 2 ) H^s(\mathbb{T}^2) H s ( T 2 ) , rather than with the anisotropic Banach spaces used in [4 ] , although the arguments can be adapted to this case.
We consider the push-forward action of F F F on the cotangent space T ∗ T 2 = T 2 × R 2 T^*\mathbb{T}^2=\mathbb{T}^2\times\mathbb{R}^2 T ∗ T 2 = T 2 × R 2 given by
F † ( p , η ) = ( F ( p ) , [ ( D p F ) ∗ ] − 1 η ) . F^\dagger(p,\eta)=\left(F(p),[(D_pF)^*]^{-1}\eta\right). F † ( p , η ) = ( F ( p ) , [( D p F ) ∗ ] − 1 η ) .
where ∗ * ∗ denotes the transpose of a matrix. Note that F † F^\dagger F † is a local diffeomorphism and a finite covering of T ∗ T 2 T^*\mathbb{T}^2 T ∗ T 2 , and ( F n ) † = ( F † ) n (F^n)^\dagger=(F^\dagger)^n ( F n ) † = ( F † ) n .
For s ≥ 0 s\geq0 s ≥ 0 we consider the quantity
B s ( F ) : = sup q ∈ T 2 , ξ ∈ R 2 ∖ { 0 } F † ( p , η ) = ( q , ξ ) ∑ ( ∥ ξ ∥ ∥ η ∥ ) s 1 ∣ det D p F ∣ . (9) B_s(F):=\sup_{\substack{q\in\mathbb{T}^2,\ \xi\in\mathbb{R}^2\setminus\{0\}\\ F^\dagger(p,\eta)=(q,\xi)}}\sum\left(\frac{\|\xi\|}{\|\eta\|}\right)^s\frac{1}{\lvert\det D_pF\rvert}. \tag*{(9)} B s ( F ) := q ∈ T 2 , ξ ∈ R 2 ∖ { 0 } F † ( p , η ) = ( q , ξ ) sup ∑ ( ∥ η ∥ ∥ ξ ∥ ) s ∣ det D p F ∣ 1 . (9)
We assume that there is s > 0 s>0 s > 0 such that B s ( F n ) B_s(F^n) B s ( F n ) decays exponentially fast enough to compensate for the possible expansion of inverse branches in the central direction:
λ c s inf n ≥ 1 ( B s ( F n ) ) 1 / n = λ c s lim n → ∞ ( B s ( F n ) ) 1 / n < 1 , (10) \lambda_c^s\inf_{n\geq1}\left(B_s(F^n)\right)^{1/n}=\lambda_c^s\lim_{n\to\infty}\left(B_s(F^n)\right)^{1/n}<1, \tag*{(10)} λ c s n ≥ 1 inf ( B s ( F n ) ) 1/ n = λ c s n → ∞ lim ( B s ( F n ) ) 1/ n < 1 , (10)
where we have used submultiplicativity. Following the naming convention in [19 ] , we refer to an SVPH F F F which satisfies (10) as s s s -Virtually Expanding (s s s -VE).6 When the role of s s s is not important, we will drop the prefix and call such maps virtually expanding, or simply VE.
Example 2.4. In [19 ] it is shown that suitable skew products F ( x , θ ) = ( f ( x ) , θ + ω ( x ) ) F(x,\theta)=(f(x),\theta+\omega(x)) F ( x , θ ) = ( f ( x ) , θ + ω ( x )) over expanding maps of the circle, with sufficiently large expansion, are virtually expanding. We refer to [19 ] for the precise statement and for further examples.
Assumptions on the observable τ \tau τ
We will consider observables τ ∈ C ∞ ( T 2 , R ) \tau\in C^\infty(\mathbb{T}^2,\mathbb{R}) τ ∈ C ∞ ( T 2 , R ) .7 For q ∈ T 2 q\in\mathbb{T}^2 q ∈ T 2 and n ∈ N n\in\mathbb{N} n ∈ N we denote by H n ( q ) \mathcal{H}^n(q) H n ( q ) the set of inverse branches of F n F^n F n defined in a neighbourhood of q q q , and set J h ( q ) : = ∣ det D h F n ∣ − 1 J_h(q):=\lvert\det D_hF^n\rvert^{-1} J h ( q ) := ∣ det D h F n ∣ − 1 for h ∈ H n ( q ) h\in\mathcal{H}^n(q) h ∈ H n ( q ) . Recall τ n = ∑ k = 0 n − 1 τ ∘ F k \tau_n=\sum_{k=0}^{n-1}\tau\circ F^k τ n = ∑ k = 0 n − 1 τ ∘ F k . For a differentiable function φ \varphi φ , we denote by d φ q d\varphi_q d φ q its differential at q q q . Thus, for u ∈ T q T 2 u\in T_q\mathbb{T}^2 u ∈ T q T 2 ,
d ( τ n ∘ h ) q ( u ) = ∇ ( τ n ∘ h ) ( q ) ⋅ u = d ( τ n ) h ( q ) ( D q h u ) . d(\tau_n\circ h)_q(u)=\nabla(\tau_n\circ h)(q)\cdot u=d(\tau_n)_{h(q)}(D_qh\,u). d ( τ n ∘ h ) q ( u ) = ∇ ( τ n ∘ h ) ( q ) ⋅ u = d ( τ n ) h ( q ) ( D q h u ) .
We assume the following
(B1) There exist C > 0 C>0 C > 0 , γ 1 > 0 \gamma_1>0 γ 1 > 0 and κ ∈ ( 0 , log λ ) \kappa\in(0,\log\lambda) κ ∈ ( 0 , log λ ) such that, for all n ∈ N n\in\mathbb{N} n ∈ N , q ∈ T 2 q\in\mathbb{T}^2 q ∈ T 2 and ζ ∈ ( R 2 ) ∗ \zeta\in(\mathbb{R}^2)^* ζ ∈ ( R 2 ) ∗ ,
∑ h ∈ H n ( q ) : ∃ u ∈ D h ( q ) F n C u , ∥ u ∥ = 1 , ∣ ζ ( u ) − d ( τ n ∘ h ) q ( u ) ∣ ≤ e κ n λ − n J h ( q ) ≤ C e − γ 1 n . (11) \sum_{\substack{h\in\mathcal{H}^n(q):\ \exists u\in D_h(q)F^n C u,\ \|u\|=1,\\ \lvert\zeta(u)-d(\tau_n\circ h)_q(u)\rvert\leq e^{\kappa n}\lambda^{-n}}}J_h(q)\leq Ce^{-\gamma_1 n}. \tag*{(11)} h ∈ H n ( q ) : ∃ u ∈ D h ( q ) F n C u , ∥ u ∥ = 1 , ∣ ζ ( u ) − d ( τ n ∘ h ) q ( u )∣ ≤ e κn λ − n ∑ J h ( q ) ≤ C e − γ 1 n . (11)
(B2) There do not exist an ergodic acip μ k \mu_k μ k of F F F , ν ∈ R ∖ { 0 } \nu\in\mathbb{R}\setminus\{0\} ν ∈ R ∖ { 0 } , c ∈ S 1 c\in S^1 c ∈ S 1 and a measurable ϕ : T 2 → S 1 \phi:\mathbb{T}^2\to S^1 ϕ : T 2 → S 1 such that
e i ν τ = c ϕ ∘ F ϕ μ k -a.e. e^{i\nu\tau}=c\frac{\phi\circ F}{\phi}\qquad\mu_k\text{-a.e.} e i ν τ = c ϕ ϕ ∘ F μ k -a.e.
Limit theorems and higher order corrections
We study the statistical properties of the Birkhoff sums
τ n = ∑ j = 0 n − 1 τ ∘ F j . \tau_n=\sum_{j=0}^{n-1}\tau\circ F^j. τ n = j = 0 ∑ n − 1 τ ∘ F j .
For this application we assume that μ k ( τ ) = 0 \mu_k(\tau)=0 μ k ( τ ) = 0 for every ergodic acip μ k \mu_k μ k , 1 ≤ k ≤ ℓ 1\le k\le\ell 1 ≤ k ≤ ℓ . This is a componentwise assumption: subtracting a single constant achieves it only when all component means agree. The multiplicative non-arithmeticity assumption is (B2) and it excludes observables cohomologous to lattice-valued functions. We do not consider the lattice case here and in particular, the equality L ν + 2 π k = L ν \mathcal{L}_{\nu+2\pi k}=\mathcal{L}_\nu L ν + 2 π k = L ν holds for integer-valued observables.
Write
n σ ( x ) = 1 2 π σ 2 e − x 2 / ( 2 σ 2 ) , Φ σ ( z ) = ∫ − ∞ z n σ ( x ) d x . n_\sigma(x)=\frac{1}{\sqrt{2\pi\sigma^2}}e^{-x^2/(2\sigma^2)},\qquad\Phi_\sigma(z)=\int_{-\infty}^{z}n_\sigma(x)\,dx. n σ ( x ) = 2 π σ 2 1 e − x 2 / ( 2 σ 2 ) , Φ σ ( z ) = ∫ − ∞ z n σ ( x ) d x .
Let Π u = ∑ k = 1 ℓ ρ k ℓ k ( u ) \Pi_u=\sum_{k=1}^{\ell}\rho_k\ell_k(u) Π u = ∑ k = 1 ℓ ρ k ℓ k ( u ) be the spectral projector of L F \mathcal{L}_F L F at 1, with ∫ ρ k d m = 1 \int\rho_k\,\mathrm{d}m=1 ∫ ρ k d m = 1 . Here ℓ k \ell_k ℓ k are positive invariant linear functionals, normalized by ℓ k ( ρ j ) = δ k j \ell_k(\rho_j)=\delta_{kj} ℓ k ( ρ j ) = δ k j and ∑ k ℓ k ( u ) = ∫ u d m \sum_k\ell_k(u)=\int u\,\mathrm{d}m ∑ k ℓ k ( u ) = ∫ u d m . For an initial probability density v ∈ H s v\in H^s v ∈ H s , put c k ( v ) = ℓ k ( v ) c_k(v)=\ell_k(v) c k ( v ) = ℓ k ( v ) . Then c k ( v ) ≥ 0 c_k(v)\ge0 c k ( v ) ≥ 0 and ∑ k c k ( v ) = 1 \sum_k c_k(v)=1 ∑ k c k ( v ) = 1 .
The Edgeworth expansion of order R R R means
P v m ( τ n n ≤ z ) = ∑ k = 1 ℓ c k ( v ) Φ σ k ( z ) + ∑ k = 1 ℓ ∑ j = 1 R n − j / 2 P j , k , v ( z ) n σ k ( z ) + o ( n − R / 2 ) . (12) \mathbb{P}_{vm}\left(\frac{\tau_n}{\sqrt{n}}\le z\right)=\sum_{k=1}^{\ell}c_k(v)\Phi_{\sigma_k}(z)+\sum_{k=1}^{\ell}\sum_{j=1}^{R}n^{-j/2}P_{j,k,v}(z)n_{\sigma_k}(z)+o\left(n^{-R/2}\right). \tag*{(12)} P v m ( n τ n ≤ z ) = k = 1 ∑ ℓ c k ( v ) Φ σ k ( z ) + k = 1 ∑ ℓ j = 1 ∑ R n − j /2 P j , k , v ( z ) n σ k ( z ) + o ( n − R /2 ) . (12)
uniformly in z ∈ R z\in\mathbb{R} z ∈ R . The variances σ k 2 > 0 \sigma_k^2>0 σ k 2 > 0 are defined in Lemma 5.1 below. For an ergodic invariant initial measure only its component occurs in the leading term. In general the limit is a mixture of Gaussians. The order-one expansion gives a Berry–Esseen bound relative to this mixture, with a constant depending on F F F , τ \tau τ , and the initial density.
For endpoint functions ψ ∈ H s \psi\in H^s ψ ∈ H s and ξ ∈ L 2 \xi\in L^2 ξ ∈ L 2 , we use the notation
I n ( g ; ψ , ξ ) : = ∫ T 2 ψ g ( τ n ) ξ ∘ F n d m , c k ( ψ , ξ ) : = ℓ k ( ψ ) ∫ ξ ρ k d m . I_n(g;\psi,\xi):=\int_{\mathbb{T}^2}\psi\,g(\tau_n)\xi\circ F^n\,\mathrm{d}m,\qquad c_k(\psi,\xi):=\ell_k(\psi)\int\xi\rho_k\,\mathrm{d}m. I n ( g ; ψ , ξ ) := ∫ T 2 ψ g ( τ n ) ξ ∘ F n d m , c k ( ψ , ξ ) := ℓ k ( ψ ) ∫ ξ ρ k d m .
Here ψ \psi ψ incorporates both the initial density and the initial observable. More precisely, for an initial probability measure v m vm v m , with v ∈ H s v\in H^s v ∈ H s , and an observable a ∈ C ⌈ s ⌉ a\in C^{\lceil s\rceil} a ∈ C ⌈ s ⌉ , we set ψ = a v \psi=av ψ = a v . The Sobolev multiplication estimate gives
∥ ψ ∥ H s ≤ C ∥ a ∥ C ⌈ s ⌉ ∥ v ∥ H s , \|\psi\|_{H^s}\le C\|a\|_{C^{\lceil s\rceil}}\|v\|_{H^s}, ∥ ψ ∥ H s ≤ C ∥ a ∥ C ⌈ s ⌉ ∥ v ∥ H s ,
so the theorem applies. More generally, it suffices that multiplication by a a a be bounded on H s H^s H s , and we do not require a Banach algebra property.
The mixing global expansion of order R R R will have the form
I n ( g ; ψ , ξ ) = ∑ j = 0 R n − ( j + 1 ) / 2 ∫ R K j , ψ , ξ ( x / n ) g ( x ) d x + o ( n − ( R + 1 ) / 2 ) , (13) I_n(g;\psi,\xi)=\sum_{j=0}^{R}n^{-(j+1)/2}\int_{\mathbb{R}}K_{j,\psi,\xi}(x/\sqrt{n})g(x)\,\mathrm{d}x+o\left(n^{-(R+1)/2}\right), \tag*{(13)} I n ( g ; ψ , ξ ) = j = 0 ∑ R n − ( j + 1 ) /2 ∫ R K j , ψ , ξ ( x / n ) g ( x ) d x + o ( n − ( R + 1 ) /2 ) , (13)
where K j , ψ , ξ = ∑ k R j , k , ψ , ξ n σ k K_{j,\psi,\xi}=\sum_k R_{j,k,\psi,\xi}n_{\sigma_k} K j , ψ , ξ = ∑ k R j , k , ψ , ξ n σ k and the R j , k , ψ , ξ R_{j,k,\psi,\xi} R j , k , ψ , ξ are polynomials of degree at most 3 j 3j 3 j . In particular,
K 0 , ψ , ξ ( x ) = ∑ k c k ( ψ , ξ ) n σ k ( x ) . K_{0,\psi,\xi}(x)=\sum_k c_k(\psi,\xi)n_{\sigma_k}(x). K 0 , ψ , ξ ( x ) = k ∑ c k ( ψ , ξ ) n σ k ( x ) .
The mixing local expansion of order R R R means
n I n ( g ; ψ , ξ ) = ∑ m = 0 ⌊ R / 2 ⌋ n − m ∫ R Q m , ψ , ξ ( x ) g ( x ) d x + o ( n − R / 2 ) , (14) \sqrt{n}I_n(g;\psi,\xi)=\sum_{m=0}^{\lfloor R/2\rfloor}n^{-m}\int_{\mathbb{R}}Q_{m,\psi,\xi}(x)g(x)\,\mathrm{d}x+o\left(n^{-R/2}\right), \tag*{(14)} n I n ( g ; ψ , ξ ) = m = 0 ∑ ⌊ R /2 ⌋ n − m ∫ R Q m , ψ , ξ ( x ) g ( x ) d x + o ( n − R /2 ) , (14)
where deg Q m , ψ , ξ ≤ 2 m \deg Q_{m,\psi,\xi} \le2m deg Q m , ψ , ξ ≤ 2 m and
Q 0 , ψ , ξ = ∑ k c k ( ψ , ξ ) 2 π σ k . Q_{0,\psi,\xi}=\sum_k \frac{c_k(\psi,\xi)}{\sqrt{2\pi}\sigma_k}. Q 0 , ψ , ξ = k ∑ 2 π σ k c k ( ψ , ξ ) .
The coefficients depend on the initial and terminal functions through the spectral amplitudes, as well as on the cumulants of τ n \tau_n τ n . In the stationary ergodic case, ψ = a ρ k \psi=a\rho_k ψ = a ρ k gives c k ( ψ , ξ ) = μ k ( a ) μ k ( ξ ) c_k(\psi,\xi)=\mu_k(a)\mu_k(\xi) c k ( ψ , ξ ) = μ k ( a ) μ k ( ξ ) whenever a ρ k ∈ H s a\rho_k\in H^s a ρ k ∈ H s .
Function spaces for the mixing expansions
We use the Fourier transform
g ^ ( ν ) = ∫ R e − i ν x g ( x ) d x . (15) \widehat{g}(\nu)=\int_{\mathbb{R}}e^{-i\nu x}g(x)\,\mathrm{d}x. \tag*{(15)} g ( ν ) = ∫ R e − i ν x g ( x ) d x . (15)
For p > 1 p>1 p > 1 and an integer k ≥ 0 k\ge0 k ≥ 0 , set
C p ( g ) = sup ν ∈ R ( 1 + ∣ ν ∣ ) p ∣ g ^ ( ν ) ∣ , M k ( g ) = ∫ R ( 1 + ∣ x ∣ k ) ∣ g ( x ) ∣ d x , ∥ g ∥ F k p = C p ( g ) + M k ( g ) . C^p(g)=\sup_{\nu\in\mathbb{R}}(1+|\nu|)^p|\widehat{g}(\nu)|,\qquad M_k(g)=\int_{\mathbb{R}}(1+|x|^k)|g(x)|\,\mathrm{d}x,\qquad\lVert g\rVert_{\mathcal{F}^p_k}=C^p(g)+M_k(g). C p ( g ) = ν ∈ R sup ( 1 + ∣ ν ∣ ) p ∣ g ( ν ) ∣ , M k ( g ) = ∫ R ( 1 + ∣ x ∣ k ) ∣ g ( x ) ∣ d x , ∥ g ∥ F k p = C p ( g ) + M k ( g ) .
We write g ∈ F k p g\in\mathcal{F}^p_k g ∈ F k p when g g g is continuous and this norm is finite. In particular, g ^ ∈ C k \widehat{g}\in C^k g ∈ C k and
max 0 ≤ j ≤ k ∥ g ^ ( j ) ∥ ∞ ≤ M k ( g ) . \max_{0\le j\le k}\lVert\widehat{g}^{(j)}\rVert_\infty\le M_k(g). 0 ≤ j ≤ k max ∥ g ( j ) ∥ ∞ ≤ M k ( g ) .
The explicit absolute-moment requirement ensures that the polynomial integrals in (14) are ordinary Lebesgue integrals. Compactly supported smooth functions belong to all these spaces. For integer p p p , integrable derivatives through order p p p imply C p ( g ) < ∞ C^p(g)<\infty C p ( g ) < ∞ by integration by parts. We use p = q + 2 p=q+2 p = q + 2 with any fixed q > 0 q>0 q > 0 in Theorem A.
Main Results
For ν ∈ R \nu\in\mathbb{R} ν ∈ R let L ν u : = L F ( e i ν τ u ) \mathcal{L}_\nu u:=\mathcal{L}_F(e^{i\nu\tau}u) L ν u := L F ( e i ν τ u ) be the twisted transfer operator, where L F \mathcal{L}_F L F is the transfer operator of F F F (see § 3 for the precise definitions), and for s > 0 s>0 s > 0 let
∥ u ∥ ν : = ∥ u ∥ L 2 + ( 1 + ∣ ν ∣ ) − s ∥ u ∥ H s . \lVert u\rVert_\nu:=\lVert u\rVert_{L^2}+(1+|\nu|)^{-s}\lVert u\rVert_{H^s}. ∥ u ∥ ν := ∥ u ∥ L 2 + ( 1 + ∣ ν ∣ ) − s ∥ u ∥ H s .
Let σ F ∈ ( 0 , 1 ) \sigma_F\in(0,1) σ F ∈ ( 0 , 1 ) be the rate of the Lasota–Yorke inequality for L F \mathcal{L}_F L F on H s H^s H s , given by Corollary 3.3 below, and let Θ , γ 1 , κ \Theta,\gamma_1,\kappa Θ , γ 1 , κ be as in (8) and (B1). We will assume the following compatibility condition:
(P) Θ σ F 2 < 1 , log Θ < γ 1 , 2 log Θ < κ s \Theta\sigma_F^2<1,\qquad\log\Theta<\gamma_1,\qquad2\log\Theta<\kappa s Θ σ F 2 < 1 , log Θ < γ 1 , 2 log Θ < κ s .
Remark 2.9. Condition (P) requires the possible growth of ∥ L F n 1 ∥ ∞ \lVert\mathcal{L}_F^n1\rVert_\infty ∥ L F n 1 ∥ ∞ to be sufficiently slow compared with the Sobolev contraction and the cancellation supplied by (B1). More precisely, it reads
log Θ < min { 2 log ( σ F − 1 ) , γ 1 , κ s 2 } . \log\Theta<\min\left\{2\log(\sigma_F^{-1}),\gamma_1,\frac{\kappa s}{2}\right\}. log Θ < min { 2 log ( σ F − 1 ) , γ 1 , 2 κ s } .
The first inequality allows the Sobolev contraction to absorb the weak-norm growth, while the remaining inequalities ensure that the cancellation and frequency truncation estimates still yield contraction after that growth is taken into account.
This restriction is natural for the systems considered here, whose central direction is neutral or close to neutral. For skew products one can take Θ = 1 \Theta=1 Θ = 1 , and (P) is automatic. More generally, the bound Θ = λ c \Theta=\lambda_c Θ = λ c shows that it suffices for the central rate λ c \lambda_c λ c to be sufficiently close to 1 1 1 relative to the other rates. In a family of systems, this conclusion requires uniform bounds keeping σ F \sigma_F σ F below 1 1 1 and γ 1 , κ s \gamma_1,\kappa s γ 1 , κ s above 0 0 0 . Thus (P) compares central growth with the available contraction and transversality rates; it does not follow from partial hyperbolicity alone. Moreover, if s > 1 s > 1 s > 1 , Sobolev embedding and (20) give
∥ L 1 n ∥ ∞ ≤ C ∥ L 1 n ∥ H s ≤ C . \| \mathcal{L}_1^n \|_{\infty} \le C \| \mathcal{L}_1^n \|_{H^s} \le C. ∥ L 1 n ∥ ∞ ≤ C ∥ L 1 n ∥ H s ≤ C .
In that case one may take Θ = 1 \Theta= 1 Θ = 1 , and (P) is again automatic.
Our principal technical result is the following Dolgopyat-type estimate, proved in Section § 4 .
Theorem 2.10 (Dolgopyat inequality). Let s > 0 s > 0 s > 0 be such that F F F is s s s –VE and satisfies (8) , let τ ∈ C ∞ ( T 2 , R ) \tau\in C^{\infty}(\mathbb{T}^2,\mathbb{R}) τ ∈ C ∞ ( T 2 , R ) satisfy (B1), and assume (P). There exist ν 0 > 1 \nu_0 > 1 ν 0 > 1 , C > 0 C > 0 C > 0 , γ > 0 \gamma> 0 γ > 0 and δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) such that8
∥ L ν n ∥ ν ≤ C ∣ ν ∣ δ e − γ n , for all n ∈ N , ∣ ν ∣ ≥ ν 0 . (16) \| \mathcal{L}_{\nu}^{n} \|_{\nu} \le C|\nu|^{\delta}e^{-\gamma n}, \qquad\text{for all } n \in\mathbb{N},\ |\nu| \ge\nu_0. \tag*{(16)} ∥ L ν n ∥ ν ≤ C ∣ ν ∣ δ e − γ n , for all n ∈ N , ∣ ν ∣ ≥ ν 0 . (16) In particular, ∥ L ν n ∥ H s → H s ≤ C ∣ ν ∣ s + δ e − γ n \| \mathcal{L}_{\nu}^{n} \|_{H^s \to H^s} \le C|\nu|^{s+\delta}e^{-\gamma n} ∥ L ν n ∥ H s → H s ≤ C ∣ ν ∣ s + δ e − γ n for all n ∈ N n \in\mathbb{N} n ∈ N and ∣ ν ∣ ≥ ν 0 |\nu| \ge\nu_0 ∣ ν ∣ ≥ ν 0 .
The above Theorem Theorem 2.10 is the key to prove the following Theorems A and B.
Theorem A (Limit theorems and higher order expansions). Let F F F be an s s s –VE SVPH satisfying (A6), and let τ ∈ C ∞ ( T 2 , R ) \tau\in C^{\infty}(\mathbb{T}^2,\mathbb{R}) τ ∈ C ∞ ( T 2 , R ) satisfy (B1), (B2), and μ k ( τ ) = 0 \mu_k(\tau) = 0 μ k ( τ ) = 0 for every ergodic acip μ k \mu_k μ k . Assume (P). Then all component variances σ k 2 \sigma_k^2 σ k 2 are strictly positive. For every integer R ≥ 1 R \ge1 R ≥ 1 :
(a) Every initial probability measure ν m \nu m ν m with ν ∈ H s \nu\in H^s ν ∈ H s admits the Edgeworth expansion (12) , uniformly in z z z .
(b) For every q > 0 q > 0 q > 0 , g ∈ F 0 q + 2 g \in\mathcal{F}_0^{q+2} g ∈ F 0 q + 2 , ψ ∈ H s \psi\in H^s ψ ∈ H s , and ξ ∈ H s \xi\in H^s ξ ∈ H s , the mixing global expansion (13) holds.
(c) For every q > 0 q > 0 q > 0 , g ∈ F R + 1 q + 2 g \in\mathcal{F}_{R+1}^{q+2} g ∈ F R + 1 q + 2 , ψ ∈ H s \psi\in H^s ψ ∈ H s , and ξ ∈ H s \xi\in H^s ξ ∈ H s , the mixing local expansion (14) holds.
In (b)–(c) one may more generally take ξ ∈ L 2 \xi\in L^2 ξ ∈ L 2 . The coefficients are independent of g g g ; the remainders are understood for fixed ν \nu ν , or fixed g , ψ , ξ g,\psi,\xi g , ψ , ξ , respectively.
The proof of the above theorem is given in Section § 5 .
Remark 2.11. The spectral gap and perturbation near zero frequency provide the cumulants and the Gaussian terms. The additional role of the Dolgopyat estimate is to control large real frequencies, allowing expansions of arbitrary order. The general spectral methods underlying these expansions are developed in [10 , 11 ] ; Section § 5 gives the required argument for the present, possibly non-ergodic, Sobolev setting.
The second application concerns the exponential decay of correlations for the skew product F ~ ∈ C ∞ ( T 3 , T 3 ) \widetilde{F} \in C^{\infty}(\mathbb{T}^3,\mathbb{T}^3) F ∈ C ∞ ( T 3 , T 3 ) given by
F ~ ( x , y , z ) = ( F ( x , y ) , z + τ ( x , y ) ) m o d Z 3 . (17) \widetilde{F}(x,y,z) = (F(x,y),z+\tau(x,y)) \mod\mathbb{Z}^3. \tag*{(17)} F ( x , y , z ) = ( F ( x , y ) , z + τ ( x , y )) mod Z 3 . (17)
It is proved in Section § 6 .
Theorem B (Exponential mixing). Let s > 0 s > 0 s > 0 be such that F F F is s s s –VE and satisfies (A6), let τ ∈ C ∞ ( T 2 , R ) \tau\in C^{\infty}(\mathbb{T}^2,\mathbb{R}) τ ∈ C ∞ ( T 2 , R ) satisfy (B1) and (B2), assume (P), and let μ ∗ \mu_* μ ∗ be an ergodic acip for F F F with density h ∗ ∈ H s h_* \in H^s h ∗ ∈ H s .9 Then the F ~ \widetilde{F} F -invariant absolutely continuous measure d μ ~ ( p , z ) = h ∗ ( p ) d p d z d\widetilde{\mu}(p,z) = h_*(p)\,dp\,dz d μ ( p , z ) = h ∗ ( p ) d p d z is exponentially mixing with respect to Hölder observables; that is, for each α > 0 \alpha> 0 α > 0 there exist γ > 0 \gamma> 0 γ > 0 and C > 0 C > 0 C > 0 such that, for any f , g ∈ C α ( T 3 , C ) f,g \in C^\alpha(\mathbb{T}^3,\mathbb{C}) f , g ∈ C α ( T 3 , C ) ,
∣ ∫ T 3 f ⋅ g ∘ F ~ n d μ ~ − ∫ T 3 f d μ ~ ∫ T 3 g d μ ~ ∣ ≤ C ∥ f ∥ C α ∥ g ∥ C α e − γ n , ∀ n ≥ 1. \left|\int_{\mathbb{T}^3} f\cdot g\circ\widetilde{F}^{n}\,d\widetilde{\mu}-\int_{\mathbb{T}^3}f\,d\widetilde{\mu}\int_{\mathbb{T}^3}g\,d\widetilde{\mu}\right|\le C\|f\|_{C^\alpha}\|g\|_{C^\alpha}e^{-\gamma n}, \qquad\forall n\ge1. ∫ T 3 f ⋅ g ∘ F n d μ − ∫ T 3 f d μ ∫ T 3 g d μ ≤ C ∥ f ∥ C α ∥ g ∥ C α e − γ n , ∀ n ≥ 1.
Remark 2.12. The circle extension is related at the operator level to a suspension semiflow over F F F . To define a roof one must require positivity; for a general real τ \tau τ , one may instead use r = τ + c > 0 r=\tau+c>0 r = τ + c > 0 . On the imaginary axis this constant shift only multiplies the twisted operators by scalars of modulus one. For a positive roof τ \tau τ (see, e.g., [16 , 18 ] ), the Fourier modes in z z z of the correlations of F ~ \widetilde{F} F are governed by the operators L 2 π k \mathcal{L}_{2\pi k} L 2 π k , which are the restriction to the imaginary axis of the operators L F ( e − z τ ⋅ ) \mathcal{L}_{F}(e^{-z\tau}\cdot) L F ( e − z τ ⋅ ) , z ∈ C z\in\mathbb{C} z ∈ C , governing the correlations of the semiflow. As recalled in the introduction, in the skew-product case F ~ ( x , θ , z ) = ( f ( x ) , θ + ω ( x ) , z + τ ( x ) ) \widetilde{F}(x,\theta,z)=(f(x),\theta+\omega(x),z+\tau(x)) F ( x , θ , z ) = ( f ( x ) , θ + ω ( x ) , z + τ ( x )) exponential mixing already follows from [5 ] . Theorem B applies also when the base F F F is a partially expanding map which is not a skew product, or when τ \tau τ depends on θ \theta θ , provided the hypotheses are satisfied.
Transfer operators
This section collects the transfer-operator tools used in the Dolgopyat estimate and both applications. The twisted Lasota–Yorke inequality is proved in § 4.2 .
Spectral gap
The Ruelle transfer operator associated to F F F is defined via
L F u ( x ) = ∑ y ∈ F − 1 ( x ) u ( y ) ∣ det D F ( y ) ∣ , u ∈ L 1 ( T 2 ) , x , y ∈ T 2 . (18) \mathcal{L}_{F}u(x)=\sum_{y\in F^{-1}(x)}\frac{u(y)}{\lvert\det DF(y)\rvert},\qquad u\in L^{1}(\mathbb{T}^{2}),\ x,y\in\mathbb{T}^{2}. \tag*{(18)} L F u ( x ) = y ∈ F − 1 ( x ) ∑ ∣ det D F ( y )∣ u ( y ) , u ∈ L 1 ( T 2 ) , x , y ∈ T 2 . (18)
and used to study the statistical properties of F F F . The main result of [19 ] is the quasi-compactness of L F \mathcal{L}_{F} L F on suitable Sobolev spaces, which will be crucial for our arguments. To state the result, let us recall the definition of the spaces used: for each s ∈ R + s\in\mathbb{R}^{+} s ∈ R + , H s ( T 2 , C ) H^{s}(\mathbb{T}^{2},\mathbb{C}) H s ( T 2 , C ) will denote the standard Hilbert space of L 2 ( T 2 , C ) L^{2}(\mathbb{T}^{2},\mathbb{C}) L 2 ( T 2 , C ) functions endowed with the norm10
∥ u ∥ H s : = ( ∑ k ∈ Z 2 ⟨ k ⟩ 2 s ∣ u ^ ( k ) ∣ 2 ) 1 2 . (19) \lVert u\rVert_{H^{s}}:=\left(\sum_{k\in\mathbb{Z}^{2}}\langle k\rangle^{2s}\lvert\widehat{u}(k)\rvert^{2}\right)^{\frac{1}{2}}. \tag*{(19)} ∥ u ∥ H s := ( k ∈ Z 2 ∑ ⟨ k ⟩ 2 s ∣ u ( k ) ∣ 2 ) 2 1 . (19)
We first relate (10) to the condition used in [19 ] .
Lemma 3.1. Let F F F be s s s -VE. Then, for every t ∈ [ s 2 , s ] t\in\left[\frac{s}{2},s\right] t ∈ [ 2 s , s ] ,
lim n → ∞ B 2 t ( F n ) 1 / n < 1. \lim_{n\to\infty}B_{2t}(F^{n})^{1/n}<1. n → ∞ lim B 2 t ( F n ) 1/ n < 1.
Proof. If ( F n ) ∗ ( p , η ) = ( q , ξ ) (F^{n})^{*}(p,\eta)=(q,\xi) ( F n ) ∗ ( p , η ) = ( q , ξ ) , then ξ = [ ( D p F n ) ∗ ] − 1 η \xi=[(D_{p}F^{n})^{*}]^{-1}\eta ξ = [( D p F n ) ∗ ] − 1 η , hence, by (B.3), ∥ ξ ∥ / ∥ η ∥ ≤ ∥ ( D p F n ) − 1 ∥ ≤ C λ c n \lVert\xi\rVert/\lVert\eta\rVert\leq\lVert(D_{p}F^{n})^{-1}\rVert\leq C\lambda_{c}^{n} ∥ ξ ∥ / ∥ η ∥ ≤ ∥( D p F n ) − 1 ∥ ≤ C λ c n . Since 0 ≤ 2 t − s ≤ s 0\leq2t-s\leq s 0 ≤ 2 t − s ≤ s and λ c ≥ 1 \lambda_{c}\geq1 λ c ≥ 1 ,
( ∥ ξ ∥ ∥ η ∥ ) 2 t ≤ ( C λ c n ) 2 t − s ( ∥ ξ ∥ ∥ η ∥ ) s ≤ C λ c s n ( ∥ ξ ∥ ∥ η ∥ ) s . \left(\frac{\lVert\xi\rVert}{\lVert\eta\rVert}\right)^{2t}\leq(C\lambda_{c}^{n})^{2t-s}\left(\frac{\lVert\xi\rVert}{\lVert\eta\rVert}\right)^{s}\leq C\lambda_{c}^{sn}\left(\frac{\lVert\xi\rVert}{\lVert\eta\rVert}\right)^{s}. ( ∥ η ∥ ∥ ξ ∥ ) 2 t ≤ ( C λ c n ) 2 t − s ( ∥ η ∥ ∥ ξ ∥ ) s ≤ C λ c s n ( ∥ η ∥ ∥ ξ ∥ ) s .
Thus B 2 t ( F n ) ≤ C λ c n B s ( F n ) B_{2t}(F^{n})\leq C\lambda_{c}^{n}B_{s}(F^{n}) B 2 t ( F n ) ≤ C λ c n B s ( F n ) , and the claim follows from (10) . □ \square □
Theorem 3.2 ([19 ] ). Let t > 0 t>0 t > 0 be such that lim n → ∞ B 2 t ( F n ) 1 / n < 1 \lim_{n\to\infty}B_{2t}(F^{n})^{1/n}<1 lim n → ∞ B 2 t ( F n ) 1/ n < 1 . Then L F \mathcal{L}_{F} L F extends to a bounded operator on H t ( T 2 ) H^{t}(\mathbb{T}^{2}) H t ( T 2 ) , which is quasi-compact with essential spectral radius strictly smaller than 1.
Corollary 3.3 (DFLY inequality for L F \mathcal{L}_{F} L F ). Let F F F be s s s -VE for some s > 0 s>0 s > 0 . For any s ′ ∈ R s'\in\mathbb{R} s ′ ∈ R such that max { s 2 , s − 1 2 } ≤ s ′ < s \max\left\{\frac{s}{2},s-\frac{1}{2}\right\}\leq s'<s max { 2 s , s − 2 1 } ≤ s ′ < s there exist σ ∗ ∈ ( 0 , 1 ) \sigma_{*}\in(0,1) σ ∗ ∈ ( 0 , 1 ) and C > 0 C>0 C > 0 such that:
∥ L F n u ∥ H s ′ ≤ C ∥ u ∥ H s ′ \lVert\mathcal{L}_{F}^{n}u\rVert_{H^{s'}}\leq C\lVert u\rVert_{H^{s'}} ∥ L F n u ∥ H s ′ ≤ C ∥ u ∥ H s ′ ∥ L F n u ∥ H s ≤ C σ ∗ n ∥ u ∥ H s + C ∥ u ∥ H s ′ . (20) \lVert\mathcal{L}_{F}^{n}u\rVert_{H^{s}}\leq C\sigma_{*}^{n}\lVert u\rVert_{H^{s}}+C\lVert u\rVert_{H^{s'}}. \tag*{(20)} ∥ L F n u ∥ H s ≤ C σ ∗ n ∥ u ∥ H s + C ∥ u ∥ H s ′ . (20)
for each n ∈ N n \in\mathbb{N} n ∈ N and u ∈ H s ( T 2 , C ) u \in H^s(\mathbb{T}^2,\mathbb{C}) u ∈ H s ( T 2 , C ) .
Proof. Let t ∈ { s ′ , s } t \in\{s',s\} t ∈ { s ′ , s } . By Lemma 3.1 and Theorem 3.2 , L F \mathcal{L}_F L F is quasi-compact on H t H^t H t with essential spectral radius ϑ t < 1 \vartheta_t < 1 ϑ t < 1 . Since H t ⊂ L 1 H^t \subset L^1 H t ⊂ L 1 and ∥ L F n u ∥ L 1 ≤ ∥ u ∥ L 1 \|\mathcal{L}_F^n u\|_{L^1} \le\|u\|_{L^1} ∥ L F n u ∥ L 1 ≤ ∥ u ∥ L 1 , every eigenvalue z z z of L F \mathcal{L}_F L F on H t H^t H t satisfies ∣ z ∣ ≤ 1 |z| \le1 ∣ z ∣ ≤ 1 . Moreover, eigenvalues of modulus one have no Jordan blocks: if L F u = z u \mathcal{L}_F u = zu L F u = z u and L F v = z v + u \mathcal{L}_F v = zv+u L F v = z v + u with ∣ z ∣ = 1 |z|=1 ∣ z ∣ = 1 , then L F n v = z n v + n z n − 1 u \mathcal{L}_F^n v=z^n v+nz^{n-1}u L F n v = z n v + n z n − 1 u , whence n ∥ u ∥ L 1 ≤ 2 ∥ v ∥ L 1 n\|u\|_{L^1} \le2\|v\|_{L^1} n ∥ u ∥ L 1 ≤ 2∥ v ∥ L 1 for all n n n , i.e. u = 0 u=0 u = 0 . It follows that sup n ∥ L F n ∥ H t < ∞ \sup_n\|\mathcal{L}_F^n\|_{H^t}<\infty sup n ∥ L F n ∥ H t < ∞ , which for t = s ′ t=s' t = s ′ is the first inequality.
Choose σ ∗ ∈ ( max { ϑ s , ϑ s ′ } , 1 ) \sigma_* \in(\max\{\vartheta_s,\vartheta_{s'}\},1) σ ∗ ∈ ( max { ϑ s , ϑ s ′ } , 1 ) such that L F \mathcal{L}_F L F has no spectrum on { ∣ z ∣ = σ ∗ } \{|z|=\sigma_*\} { ∣ z ∣ = σ ∗ } , neither on H s H^s H s nor on H s ′ H^{s'} H s ′ , and let P P P , P ′ P' P ′ be the spectral projectors of L F \mathcal{L}_F L F on H s H^s H s , H s ′ H^{s'} H s ′ , respectively, associated with the spectrum in { ∣ z ∣ > σ ∗ } \{|z|>\sigma_*\} { ∣ z ∣ > σ ∗ } ; they have finite rank. For ∣ z ∣ |z| ∣ z ∣ large the resolvents of L F \mathcal{L}_F L F on H s H^s H s and on H s ′ H^{s'} H s ′ are given by the same Neumann series, so they coincide on H s H^s H s ; by analytic continuation they coincide on H s H^s H s for all z z z in { ∣ z ∣ ≥ σ ∗ } \{|z|\ge\sigma_*\} { ∣ z ∣ ≥ σ ∗ } outside a finite set, hence P ′ ∣ H s = P P'|_{H^s}=P P ′ ∣ H s = P . Since H s H^s H s is dense in H s ′ H^{s'} H s ′ and P ′ P' P ′ is continuous with finite rank, P ′ ( H s ′ ) = P ( H s ) ⊂ H s P'(H^{s'})=P(H^s)\subset H^s P ′ ( H s ′ ) = P ( H s ) ⊂ H s . On this finite dimensional space all norms are equivalent, so
∥ P u ∥ H s = ∥ P ′ u ∥ H s ≤ C ∥ P ′ u ∥ H s ′ ≤ C ∥ u ∥ H s ′ \|Pu\|_{H^s}=\|P'u\|_{H^s}\le C\|P'u\|_{H^{s'}}\le C\|u\|_{H^{s'}} ∥ P u ∥ H s = ∥ P ′ u ∥ H s ≤ C ∥ P ′ u ∥ H s ′ ≤ C ∥ u ∥ H s ′
for u ∈ H s u\in H^s u ∈ H s . Finally,
∥ L F n ( Id − P ) ∥ H s ≤ C σ ∗ n , \|\mathcal{L}_F^n(\operatorname{Id}-P)\|_{H^s}\le C\sigma_*^n, ∥ L F n ( Id − P ) ∥ H s ≤ C σ ∗ n ,
and therefore
∥ L F n u ∥ H s ≤ ∥ L F n ( Id − P ) u ∥ H s + ∥ L F n ∥ H s ∥ P u ∥ H s ≤ C σ ∗ n ∥ u ∥ H s + C ∥ u ∥ H s ′ . \|\mathcal{L}_F^n u\|_{H^s}\le\|\mathcal{L}_F^n(\operatorname{Id}-P)u\|_{H^s}+\|\mathcal{L}_F^n\|_{H^s}\|Pu\|_{H^s}\le C\sigma_*^n\|u\|_{H^s}+C\|u\|_{H^{s'}}. ∥ L F n u ∥ H s ≤ ∥ L F n ( Id − P ) u ∥ H s + ∥ L F n ∥ H s ∥ P u ∥ H s ≤ C σ ∗ n ∥ u ∥ H s + C ∥ u ∥ H s ′ .
Fix a number σ F ∈ ( 0 , 1 ) \sigma_F\in(0,1) σ F ∈ ( 0 , 1 ) for which (20) holds for some admissible s ′ s' s ′ . It may be chosen arbitrarily close to the infimum of the admissible rates; we do not assume that this infimum is positive or attained.
Next, we define twisted transfer operators.
Definition 3.4. For any ν ∈ R \nu\in\mathbb{R} ν ∈ R and u ∈ L 1 ( T 2 , C ) u\in L^1(\mathbb{T}^2,\mathbb{C}) u ∈ L 1 ( T 2 , C ) we define the twisted transfer operator L ν \mathcal{L}_\nu L ν associated to F F F by
L ν u ( x , y ) = ∑ ( w 1 , w 2 ) ∈ F − 1 ( x , y ) u ( w 1 , w 2 ) det D F ( w 1 , w 2 ) e i ν τ ( w 1 , w 2 ) . \mathcal{L}_\nu u(x,y)=\sum_{(w_1,w_2)\in F^{-1}(x,y)}\frac{u(w_1,w_2)}{\det DF(w_1,w_2)}e^{i\nu\tau(w_1,w_2)}. L ν u ( x , y ) = ( w 1 , w 2 ) ∈ F − 1 ( x , y ) ∑ det D F ( w 1 , w 2 ) u ( w 1 , w 2 ) e i ν τ ( w 1 , w 2 ) . Iterating the above formula we see that for each n ∈ N n\in\mathbb{N} n ∈ N
L ν n u ( x , y ) = ∑ ( w 1 , w 2 ) ∈ F − n ( x , y ) u ( w 1 , w 2 ) det D F n ( w 1 , w 2 ) e i ν τ n ( w 1 , w 2 ) . (21) \mathcal{L}_\nu^n u(x,y)=\sum_{(w_1,w_2)\in F^{-n}(x,y)}\frac{u(w_1,w_2)}{\det DF^n(w_1,w_2)}e^{i\nu\tau_n(w_1,w_2)}. \tag*{(21)} L ν n u ( x , y ) = ( w 1 , w 2 ) ∈ F − n ( x , y ) ∑ det D F n ( w 1 , w 2 ) u ( w 1 , w 2 ) e i ν τ n ( w 1 , w 2 ) . (21) where τ n : = ∑ k = 0 n − 1 τ ∘ F k \tau_n:=\sum_{k=0}^{n-1}\tau\circ F^k τ n := ∑ k = 0 n − 1 τ ∘ F k . Note also that
L ν f = L F ( e i ν τ f ) . (22) \mathcal{L}_\nu f=\mathcal{L}_F(e^{i\nu\tau}f). \tag*{(22)} L ν f = L F ( e i ν τ f ) . (22) Finally, let us define the norm we will be using. For each ν ∈ R \nu\in\mathbb{R} ν ∈ R and s ∈ R + s\in\mathbb{R}^+ s ∈ R + let
∥ u ∥ ν : = ∥ u ∥ L 2 + ( 1 + ∣ ν ∣ ) − s ∥ u ∥ H s . \|u\|_\nu:=\|u\|_{L^2}+(1+|\nu|)^{-s}\|u\|_{H^s}. ∥ u ∥ ν := ∥ u ∥ L 2 + ( 1 + ∣ ν ∣ ) − s ∥ u ∥ H s . For each fixed ν \nu ν , the above norm is equivalent to ∥ ⋅ ∥ H s \|\cdot\|_{H^s} ∥ ⋅ ∥ H s (not uniformly in ν \nu ν ).
The following is a Lasota-Yorke type inequality with strong norm ∥ ⋅ ∥ ν \|\cdot\|_\nu ∥ ⋅ ∥ ν and weak norm ∥ ⋅ ∥ L 2 \|\cdot\|_{L^2} ∥ ⋅ ∥ L 2 , whose proof is postponed to § 4.2 .
Lemma 3.5 (DFLY inequality for L ν \mathcal{L}_{\nu} L ν ). Let F F F be s s s –VE for some s > 0 s > 0 s > 0 , satisfying (8) , and τ ∈ C ∞ \tau\in C^{\infty} τ ∈ C ∞ . For each σ ∈ ( σ F , 1 ) \sigma\in(\sigma_F,1) σ ∈ ( σ F , 1 ) there exist C σ > 0 C_{\sigma} > 0 C σ > 0 and C τ > 0 C_{\tau} > 0 C τ > 0 such that, for all n ∈ N n \in\mathbb{N} n ∈ N , u ∈ H s u \in H^s u ∈ H s , ν ∈ R \nu\in\mathbb{R} ν ∈ R ,
∥ L ν n u ∥ ν ≤ C σ σ n ∥ u ∥ ν + C τ Θ n / 2 ∥ u ∥ L 2 . (23) \lVert\mathcal{L}_{\nu}^{n}u\rVert_{\nu} \le C_{\sigma}\sigma^{n}\lVert u\rVert_{\nu} + C_{\tau}\Theta^{n/2}\lVert u\rVert_{L^2}. \tag*{(23)} ∥ L ν n u ∥ ν ≤ C σ σ n ∥ u ∥ ν + C τ Θ n /2 ∥ u ∥ L 2 . (23)
Remark 3.6. The proof of Theorem 2.10 specifies admissible γ \gamma γ and δ \delta δ in terms of the Lasota–Yorke constants, the rates in (8) and (B1), and the inverse-branch derivative bounds of Lemma 4.5 . These bounds also depend on finitely many derivatives of F F F and τ \tau τ .
Ergodic properties
Let m \mathrm{m} m denote normalised Lebesgue measure on T 2 \mathbb{T}^2 T 2 . We record the ergodic consequences of quasi-compactness that will be used below. In particular, we distinguish the invariant probabilities from the dual eigenfunctionals of the transfer operator.
Proposition 3.7. Suppose that F F F is s s s –VE for some s > 0 s > 0 s > 0 . Then:
sup n ≥ 0 ∥ L F n ∥ H s → H s < ∞ . \sup_{n \ge0}\lVert\mathcal{L}_F^n\rVert_{H^s \to H^s} < \infty. n ≥ 0 sup ∥ L F n ∥ H s → H s < ∞. Moreover, L F \mathcal{L}_F L F is positive and ∥ L F n ∥ L 1 → L 1 ≤ 1 \lVert\mathcal{L}_F^n\rVert_{L^1 \to L^1} \le1 ∥ L F n ∥ L 1 → L 1 ≤ 1 .
(ii) There are nonnegative functions ρ 1 , … , ρ ℓ ∈ H s \rho_1,\ldots,\rho_{\ell} \in H^s ρ 1 , … , ρ ℓ ∈ H s , satisfying
∫ ρ k d m = 1 , ρ j ρ k = 0 m -a.e. ( j ≠ k ) , \int\rho_k\,\mathrm{d}\mathrm{m} = 1,\qquad\rho_j\rho_k = 0\quad\mathrm{m}\text{-a.e.}\quad(j \ne k), ∫ ρ k dm = 1 , ρ j ρ k = 0 m -a.e. ( j = k ) , which form a basis of ker ( L F − Id ) \ker(\mathcal{L}_F - \operatorname{Id}) ker ( L F − Id ) on H s H^s H s . The probabilities μ k = ρ k m \mu_k = \rho_k\mathrm{m} μ k = ρ k m are precisely the ergodic absolutely continuous invariant probabilities. Every absolutely continuous invariant probability is a convex combination of the μ k \mu_k μ k .
Π u = ∑ k = 1 ℓ ρ k ℓ k ( u ) , \Pi u = \sum_{k=1}^{\ell}\rho_k\ell_k(u), Π u = k = 1 ∑ ℓ ρ k ℓ k ( u ) , and
ℓ k L F = ℓ k , ℓ k ( ρ j ) = δ k j , ∑ k = 1 ℓ ℓ k ( u ) = ∫ u d m . \ell_k\mathcal{L}_F = \ell_k,\qquad\ell_k(\rho_j) = \delta_{kj},\qquad\sum_{k=1}^{\ell}\ell_k(u) = \int u\,\mathrm{d}\mathrm{m}. ℓ k L F = ℓ k , ℓ k ( ρ j ) = δ k j , k = 1 ∑ ℓ ℓ k ( u ) = ∫ u dm . Here positivity means ℓ k ( u ) ≥ 0 \ell_k(u) \ge0 ℓ k ( u ) ≥ 0 whenever u ∈ H s u \in H^s u ∈ H s is real and nonnegative almost everywhere. Furthermore,
1 N ∑ n = 0 N − 1 L F n ⟶ Π in operator norm on H s . \frac{1}{N}\sum_{n=0}^{N-1}\mathcal{L}_F^n \longrightarrow\Pi\quad\text{in operator norm on }H^s. N 1 n = 0 ∑ N − 1 L F n ⟶ Π in operator norm on H s . (iv) The physical basins of μ 1 , … , μ ℓ \mu_1,\ldots,\mu_{\ell} μ 1 , … , μ ℓ cover m \mathrm{m} m -almost every point.
Proof. Write L = L F L = \mathcal{L}_F L = L F . Positivity, preservation of integrals and L 1 L^1 L 1 -contractivity follow from the inverse-branch formula:
∣ L u ∣ ≤ L ∣ u ∣ , ∫ L u d m = ∫ u d m . |Lu| \le L|u|,\qquad\int Lu\,\mathrm{d}\mathrm{m} = \int u\,\mathrm{d}\mathrm{m}. ∣ Lu ∣ ≤ L ∣ u ∣ , ∫ Lu dm = ∫ u dm .
By Lemma 3.1 and Theorem 3.2 , the essential spectral radius of L L L on H s H^s H s is strictly smaller than one. Since H s H^s H s embeds continuously into L 1 L^1 L 1 , contractivity excludes eigenvalues of modulus greater than one.
It also excludes nontrivial Jordan blocks at peripheral eigenvalues. Indeed, if
L u = z u , L v = z v + u , ∣ z ∣ = 1 , Lu=zu,\qquad Lv=zv+u,\qquad|z|=1, Lu = z u , Lv = z v + u , ∣ z ∣ = 1 ,
then
L n v = z n v + n z n − 1 u . L^n v=z^n v+nz^{n-1}u. L n v = z n v + n z n − 1 u .
Consequently,
n ∥ u ∥ L 1 ≤ ∥ L n v ∥ L 1 + ∥ v ∥ L 1 ≤ 2 ∥ v ∥ L 1 , n\lVert u\rVert_{L^1}\leq\lVert L^n v\rVert_{L^1}+\lVert v\rVert_{L^1}\leq2\lVert v\rVert_{L^1}, n ∥ u ∥ L 1 ≤ ∥ L n v ∥ L 1 + ∥ v ∥ L 1 ≤ 2 ∥ v ∥ L 1 ,
so u = 0 u=0 u = 0 . Finally, integration is a nonzero invariant bounded functional on H s H^s H s . Hence 1 belongs to the spectrum of L L L , and, since it lies outside the essential spectrum, it is an eigenvalue. These facts imply that the spectral radius is one and that the powers of L L L are uniformly bounded on H s H^s H s .
Let
A N = 1 N ∑ n = 0 N − 1 L n . A_N=\frac{1}{N}\sum_{n=0}^{N-1}L^n. A N = N 1 n = 0 ∑ N − 1 L n .
The spectral decomposition, semisimplicity of the peripheral eigenvalues, and the fact that N − 1 ∑ n = 0 N − 1 z n → 0 N^{-1}\sum_{n=0}^{N-1}z^n\to0 N − 1 ∑ n = 0 N − 1 z n → 0 for ∣ z ∣ = 1 |z|=1 ∣ z ∣ = 1 , z ≠ 1 z\ne1 z = 1 , give
A N ⟶ Π in operator norm on H s . (24) A_N\longrightarrow\Pi\quad\text{in operator norm on }H^s. \tag*{(24)} A N ⟶ Π in operator norm on H s . (24)
In particular, Π \Pi Π is positive and preserves integrals. Next, set
E = ker ( L − I d ) ⊂ H s . E=\ker(L-\mathrm{Id})\subset H^s. E = ker ( L − Id ) ⊂ H s .
This space is finite-dimensional and therefore closed also as a subspace of L 1 L^1 L 1 . Let v ∈ L 1 v\in L^1 v ∈ L 1 satisfy L v = v Lv=v Lv = v . Choose u j ∈ H s u_j\in H^s u j ∈ H s with u j → v u_j\to v u j → v in L 1 L^1 L 1 . Since A N v = v A_Nv=v A N v = v and ∥ A N ∥ L 1 → L 1 ≤ 1 \lVert A_N\rVert_{L^1\to L^1}\leq1 ∥ A N ∥ L 1 → L 1 ≤ 1 , letting N → ∞ N\to\infty N → ∞ gives
∥ v − Π u j ∥ L 1 ≤ ∥ v − u j ∥ L 1 . \lVert v-\Pi u_j\rVert_{L^1}\leq\lVert v-u_j\rVert_{L^1}. ∥ v − Π u j ∥ L 1 ≤ ∥ v − u j ∥ L 1 .
Thus v v v belongs to the L 1 L^1 L 1 -closure of E E E , hence to E E E itself. We have proved
ker ( L − I d ) ∣ L 1 = ker ( L − I d ) ∣ H s . (25) \ker(L-\mathrm{Id})\big|_{L^1}=\ker(L-\mathrm{Id})\big|_{H^s}. \tag*{(25)} ker ( L − Id ) L 1 = ker ( L − Id ) H s . (25)
The real fixed space E R E_{\mathbb{R}} E R is a vector sublattice of L 1 L^1 L 1 . Indeed, if u ∈ E R u\in E_{\mathbb{R}} u ∈ E R , then
∣ u ∣ = ∣ L u ∣ ≤ L ∣ u ∣ . |u|=|Lu|\leq L|u|. ∣ u ∣ = ∣ Lu ∣ ≤ L ∣ u ∣.
Both sides have the same integral, so L ∣ u ∣ = ∣ u ∣ L|u|=|u| L ∣ u ∣ = ∣ u ∣ . By (25) , ∣ u ∣ ∈ E R |u|\in E_{\mathbb{R}} ∣ u ∣ ∈ E R . The usual formulas for the maximum and minimum of two real functions now give the sublattice property.
A finite-dimensional vector sublattice of L 1 L^1 L 1 has a basis of nonnegative, pairwise disjoint functions: its positive cone is generated by finitely many disjoint atoms. Normalising these atoms by their integrals gives ρ 1 , … , ρ ℓ \rho_1,\ldots,\rho_\ell ρ 1 , … , ρ ℓ . Every nonnegative fixed function consequently has a unique representation
v = ∑ k = 1 ℓ c k ρ k , c k ≥ 0. (26) v=\sum_{k=1}^{\ell}c_k\rho_k,\qquad c_k\geq0. \tag*{(26)} v = k = 1 ∑ ℓ c k ρ k , c k ≥ 0. (26)
For a probability density, ∑ k c k = 1 \sum_k c_k=1 ∑ k c k = 1 .
Each μ k = ρ k m \mu_k=\rho_km μ k = ρ k m is ergodic. Otherwise, an invariant measurable set of μ k \mu_k μ k -measure strictly between zero and one would split μ k \mu_k μ k into two nonzero invariant absolutely continuous measures. Their densities would belong to E E E by (25) , contradicting that ρ k \rho_k ρ k generates an extreme ray of its positive cone. Conversely, (26) shows that every ergodic acip is one of the μ k \mu_k μ k .
Next, define ℓ k \ell_k ℓ k by the unique expansion
Π u = ∑ k ρ k ℓ k ( u ) . \Pi u = \sum_k \rho_k\ell_k(u). Π u = k ∑ ρ k ℓ k ( u ) .
They are bounded because Π \Pi Π is bounded and its range is finite-dimensional. Their positivity follows from positivity of Π \Pi Π and disjointness of the ρ k \rho_k ρ k . The identities
L Π = Π L = Π , Π ρ j = ρ j \mathcal{L}\Pi= \Pi\mathcal{L} = \Pi,\qquad\Pi\rho_j = \rho_j L Π = Π L = Π , Π ρ j = ρ j
give ℓ k L = ℓ k \ell_k\mathcal{L} = \ell_k ℓ k L = ℓ k and ℓ k ( ρ j ) = δ k j \ell_k(\rho_j) = \delta_{kj} ℓ k ( ρ j ) = δ k j . Since Π \Pi Π preserves integrals,
∑ k ℓ k ( u ) = ∫ Π u d m = ∫ u d m . \sum_k \ell_k(u) = \int\Pi u\,\mathrm{d}m = \int u\,\mathrm{d}m. k ∑ ℓ k ( u ) = ∫ Π u d m = ∫ u d m .
Finally, for each k k k , let
B k = { p : 1 N ∑ n = 0 N − 1 φ ( F n p ) ⟶ ∫ φ d μ k for every φ ∈ C 0 ( T 2 ) } . \mathcal{B}_k = \left\{p : \frac{1}{N}\sum_{n=0}^{N-1}\varphi(F^n p) \longrightarrow\int\varphi\,\mathrm{d}\mu_k\text{ for every }\varphi\in C^0(\mathbb{T}^2)\right\}. B k = { p : N 1 n = 0 ∑ N − 1 φ ( F n p ) ⟶ ∫ φ d μ k for every φ ∈ C 0 ( T 2 ) } .
Using a countable uniformly dense subset of C 0 ( T 2 ) C^0(\mathbb{T}^2) C 0 ( T 2 ) , these are measurable sets. Birkhoff’s theorem and ergodicity give μ k ( B k ) = 1 \mu_k(\mathcal{B}_k)=1 μ k ( B k ) = 1 .
Note that F − 1 B k = B k F^{-1}\mathcal{B}_k=\mathcal{B}_k F − 1 B k = B k . Set
B = T 2 ∖ ⋃ k B k . B=\mathbb{T}^2\setminus\bigcup_k\mathcal{B}_k. B = T 2 ∖ k ⋃ B k .
Then F − 1 B = B F^{-1}B=B F − 1 B = B and μ k ( B ) = 0 \mu_k(B)=0 μ k ( B ) = 0 for every k k k . Transfer-operator duality gives
∫ B A N 1 d m = 1 N ∑ n = 0 N − 1 ∫ 1 B ∘ F n d m = m ( B ) . \int_B A_N1\,\mathrm{d}m=\frac{1}{N}\sum_{n=0}^{N-1}\int1_B\circ F^n\,\mathrm{d}m=m(B). ∫ B A N 1 d m = N 1 n = 0 ∑ N − 1 ∫ 1 B ∘ F n d m = m ( B ) .
By (24) , A N 1 → Π 1 A_N1\to\Pi1 A N 1 → Π1 in L 1 L^1 L 1 . Since Π 1 \Pi1 Π1 is a linear combination of the ρ k \rho_k ρ k , it follows that
m ( B ) = ∫ B Π 1 d m = 0. m(B)=\int_B\Pi1\,\mathrm{d}m=0. m ( B ) = ∫ B Π1 d m = 0.
Moreover, μ k ( B k ) = 1 \mu_k(\mathcal{B}_k)=1 μ k ( B k ) = 1 and μ k ≪ m \mu_k\ll m μ k ≪ m imply m ( B k ) > 0 m(\mathcal{B}_k)>0 m ( B k ) > 0 , so each μ k \mu_k μ k is physical. □ \square □
Remark 3.8. The functionals ℓ k \ell_k ℓ k are dual eigenfunctionals of L F \mathcal{L}_F L F ; they should not be identified with the invariant probabilities μ k \mu_k μ k . In fact, for u ∈ H s u\in H^s u ∈ H s ,
ℓ k ( u ) = ∫ B k u d m . \ell_k(u)=\int_{\mathcal{B}_k}u\,\mathrm{d}m. ℓ k ( u ) = ∫ B k u d m .
Indeed, the basins are pairwise disjoint, F − 1 B k = B k F^{-1}\mathcal{B}_k=\mathcal{B}_k F − 1 B k = B k , and μ j ( B k ) = δ j k \mu_j(\mathcal{B}_k)=\delta_{jk} μ j ( B k ) = δ j k . Integrating A N u → Π u A_Nu\to\Pi u A N u → Π u over B k \mathcal{B}_k B k proves the identity.
Under (A6), Lemma 3.9 below excludes peripheral eigenvalues other than 1. Consequently, there exist C > 0 C>0 C > 0 and ϑ ∈ ( 0 , 1 ) \vartheta\in(0,1) ϑ ∈ ( 0 , 1 ) such that
∥ L F n − Π ∥ H s → H s ≤ C ϑ n . \left\|\mathcal{L}_F^n-\Pi\right\|_{H^s\to H^s}\le C\vartheta^n. ∥ L F n − Π ∥ H s → H s ≤ C ϑ n .
Without (A6), the Cesàro convergence in Proposition 3.7 remains valid, but convergence of the iterates themselves is not asserted. Virtually expanding maps are volume expanding (see [19 ] ), in the following sense
lim n → ∞ ( inf p ∈ T 2 ∣ det D p F n ∣ ) 1 / n > 1. (27) \lim_{n\to\infty}\left(\inf_{p\in\mathbb{T}^2}\left|\det D_pF^n\right|\right)^{1/n}>1. \tag*{(27)} n → ∞ lim ( p ∈ T 2 inf ∣ det D p F n ∣ ) 1/ n > 1. (27)
In the present notation, (A6) says that ( F , μ k ) (F,\mu_k) ( F , μ k ) is weakly mixing for each k ∈ { 1 , … , ℓ } k \in\{1,\ldots,\ell\} k ∈ { 1 , … , ℓ } , and (8) reads
∥ L F n 1 ∥ ∞ ≤ C ∞ Θ n . \|\mathcal{L}_F^n 1\|_{\infty} \le C_{\infty}\Theta^n. ∥ L F n 1 ∥ ∞ ≤ C ∞ Θ n .
Lemma 3.9. Assume (A6). Then there are no peripheral eigenvalues11 of L F : H s → H s \mathcal{L}_F:H^s \to H^s L F : H s → H s other than 1.
Proof. Suppose that L F ρ = e i λ ρ \mathcal{L}_F\rho=e^{i\lambda}\rho L F ρ = e iλ ρ with ρ ∈ H s ∖ { 0 } \rho\in H^s\setminus\{0\} ρ ∈ H s ∖ { 0 } and λ ∈ R \lambda\in\mathbb{R} λ ∈ R . Since ∣ ρ ∣ = ∣ L F ρ ∣ ≤ L F ∣ ρ ∣ |\rho|=|\mathcal{L}_F\rho|\le\mathcal{L}_F|\rho| ∣ ρ ∣ = ∣ L F ρ ∣ ≤ L F ∣ ρ ∣ pointwise and ∫ L F ∣ ρ ∣ d m = ∫ ∣ ρ ∣ d m \int\mathcal{L}_F|\rho|\,\mathrm{d}m=\int|\rho|\,\mathrm{d}m ∫ L F ∣ ρ ∣ d m = ∫ ∣ ρ ∣ d m , we have L F ∣ ρ ∣ = ∣ ρ ∣ \mathcal{L}_F|\rho|=|\rho| L F ∣ ρ ∣ = ∣ ρ ∣ m-a.e. Hence ∣ ρ ∣ m |\rho|m ∣ ρ ∣ m is a finite F F F -invariant absolutely continuous measure, so by (ii) ∣ ρ ∣ = ∑ k c k ρ k |\rho|=\sum_k c_k\rho_k ∣ ρ ∣ = ∑ k c k ρ k with c k ≥ 0 c_k\ge0 c k ≥ 0 , and some c k > 0 c_k>0 c k > 0 .
Write ρ = g ∣ ρ ∣ \rho=g|\rho| ρ = g ∣ ρ ∣ with ∣ g ∣ = 1 |g|=1 ∣ g ∣ = 1 on { ∣ ρ ∣ > 0 } \{|\rho|>0\} { ∣ ρ ∣ > 0 } . Then, for m-a.e. x x x ,
∣ ∑ y ∈ F − 1 ( x ) g ( y ) ∣ ρ ∣ ( y ) ∣ det D y F ∣ ∣ = ∑ y ∈ F − 1 ( x ) ∣ ρ ∣ ( y ) ∣ det D y F ∣ . \left|\sum_{y\in F^{-1}(x)}\frac{g(y)|\rho|(y)}{|\det D_yF|}\right|=\sum_{y\in F^{-1}(x)}\frac{|\rho|(y)}{|\det D_yF|}. y ∈ F − 1 ( x ) ∑ ∣ det D y F ∣ g ( y ) ∣ ρ ∣ ( y ) = y ∈ F − 1 ( x ) ∑ ∣ det D y F ∣ ∣ ρ ∣ ( y ) .
and equality in the triangle inequality, together with L F ρ = e i λ g L F ∣ ρ ∣ \mathcal{L}_F\rho=e^{i\lambda}g\mathcal{L}_F|\rho| L F ρ = e iλ g L F ∣ ρ ∣ , yields g ( y ) = e i λ g ( F ( y ) ) g(y)=e^{i\lambda}g(F(y)) g ( y ) = e iλ g ( F ( y )) for all y ∈ F − 1 ( x ) y\in F^{-1}(x) y ∈ F − 1 ( x ) with ∣ ρ ∣ ( y ) > 0 |\rho|(y)>0 ∣ ρ ∣ ( y ) > 0 . Thus g ∘ F = e − i λ g g\circ F=e^{-i\lambda}g g ∘ F = e − iλ g μ k \mu_k μ k -a.e., i.e. g ∈ L 2 ( μ k ) g\in L^2(\mu_k) g ∈ L 2 ( μ k ) is a non-zero eigenfunction of the Koopman operator of ( F , μ k ) (F,\mu_k) ( F , μ k ) with eigenvalue e − i λ e^{-i\lambda} e − iλ . Since ( F , μ k ) (F,\mu_k) ( F , μ k ) is weakly mixing, e − i λ = 1 e^{-i\lambda}=1 e − iλ = 1 . □
Spectral properties of the twisted operators
Lemma 3.10. Let F F F be s-VE and satisfy (A6), and let τ ∈ C ∞ \tau\in C^\infty τ ∈ C ∞ satisfy (B2). Then:
(a) 1 is the only peripheral eigenvalue of L F \mathcal{L}_F L F on H s H^s H s ; the corresponding spectral projector Π \Pi Π satisfies L F Π = Π L F = Π \mathcal{L}_F\Pi=\Pi\mathcal{L}_F=\Pi L F Π = Π L F = Π and ∥ L F n ( Id − Π ) ∥ H s ≤ C θ 0 n \|\mathcal{L}_F^n(\operatorname{Id}-\Pi)\|_{H^s}\le C\theta_0^n ∥ L F n ( Id − Π ) ∥ H s ≤ C θ 0 n for some C > 0 C>0 C > 0 , θ 0 ∈ ( 0 , 1 ) \theta_0\in(0,1) θ 0 ∈ ( 0 , 1 ) ;
(b) for all ν ∈ R ∖ { 0 } \nu\in\mathbb{R}\setminus\{0\} ν ∈ R ∖ { 0 } , the spectral radius of L ν \mathcal{L}_\nu L ν on H s H^s H s is smaller than 1;
(c) for each compact J ⊂ R ∖ { 0 } J\subset\mathbb{R}\setminus\{0\} J ⊂ R ∖ { 0 } there exist C J > 0 C_J>0 C J > 0 and θ J ∈ ( 0 , 1 ) \theta_J\in(0,1) θ J ∈ ( 0 , 1 ) such that
sup ν ∈ J ∥ L ν n ∥ H s ≤ C J θ J n ∀ n ∈ N . (28) \sup_{\nu\in J}\|\mathcal{L}_\nu^n\|_{H^s}\le C_J\theta_J^n \qquad\forall n\in\mathbb{N}. \tag*{(28)} ν ∈ J sup ∥ L ν n ∥ H s ≤ C J θ J n ∀ n ∈ N . (28)
Proof. (a) By Proposition 3.7 (i) and Lemma 3.9 , 1 is the only peripheral eigenvalue and the rest of the spectrum is contained in a disc of radius θ 0 < 1 \theta_0<1 θ 0 < 1 . Since sup n ∥ L F n ∥ H s < ∞ \sup_n\|\mathcal{L}_F^n\|_{H^s}<\infty sup n ∥ L F n ∥ H s < ∞ by (20) , the eigenvalue 1 is semisimple, i.e. L F Π = Π \mathcal{L}_F\Pi=\Pi L F Π = Π .
(b) Fix σ ∈ ( σ F , 1 ) \sigma\in(\sigma_F,1) σ ∈ ( σ F , 1 ) . Lemma 3.5 and the compact embedding H s ↪ L 2 H^s\hookrightarrow L^2 H s ↪ L 2 imply that the essential spectral radius of L ν \mathcal{L}_\nu L ν on H s H^s H s is at most σ < 1 \sigma<1 σ < 1 [13 ] . Consequently any spectral value on or outside the unit circle would be an eigenvalue. Since ∥ L ν n u ∥ L 1 ≤ ∥ L F n ∣ u ∣ ∥ L 1 = ∥ u ∥ L 1 \|\mathcal{L}_\nu^n u\|_{L^1}\le\|\mathcal{L}_F^n|u|\|_{L^1}=\|u\|_{L^1} ∥ L ν n u ∥ L 1 ≤ ∥ L F n ∣ u ∣ ∥ L 1 = ∥ u ∥ L 1 , every eigenvalue z z z satisfies ∣ z ∣ ≤ 1 |z|\le1 ∣ z ∣ ≤ 1 . Suppose L ν u = e i ϑ u \mathcal{L}_\nu u=e^{i\vartheta}u L ν u = e i ϑ u with u ∈ H s ∖ { 0 } u\in H^s\setminus\{0\} u ∈ H s ∖ { 0 } . Then ∣ u ∣ ≤ L F ∣ u ∣ |u|\le\mathcal{L}_F|u| ∣ u ∣ ≤ L F ∣ u ∣ and ∫ L F ∣ u ∣ = ∫ ∣ u ∣ \int\mathcal{L}_F|u|=\int|u| ∫ L F ∣ u ∣ = ∫ ∣ u ∣ , so L F ∣ u ∣ = ∣ u ∣ \mathcal{L}_F|u|=|u| L F ∣ u ∣ = ∣ u ∣ a.e.; by Proposition 3.7 (ii), ∣ u ∣ = ∑ k c k ρ k |u|=\sum_k c_k\rho_k ∣ u ∣ = ∑ k c k ρ k with some c k > 0 c_k>0 c k > 0 . Writing u = g ∣ u ∣ u=g|u| u = g ∣ u ∣ with ∣ g ∣ = 1 |g|=1 ∣ g ∣ = 1 on { ∣ u ∣ > 0 } \{|u|>0\} { ∣ u ∣ > 0 } , the identity
∑ y ∈ F − 1 ( x ) e i ν τ ( y ) g ( y ) ∣ u ∣ ( y ) ∣ det D y F ∣ = e i ϑ g ( x ) ∑ y ∈ F − 1 ( x ) ∣ u ∣ ( y ) ∣ det D y F ∣ \sum_{y\in F^{-1}(x)}\frac{e^{i\nu\tau(y)}g(y)|u|(y)}{|\det D_yF|}=e^{i\vartheta}g(x)\sum_{y\in F^{-1}(x)}\frac{|u|(y)}{|\det D_yF|} y ∈ F − 1 ( x ) ∑ ∣ det D y F ∣ e i ν τ ( y ) g ( y ) ∣ u ∣ ( y ) = e i ϑ g ( x ) y ∈ F − 1 ( x ) ∑ ∣ det D y F ∣ ∣ u ∣ ( y )
forces equality in the triangle inequality, i.e. e i ν τ ( y ) g ( y ) = e i ϑ g ( F ( y ) ) e^{i\nu\tau(y)}g(y)=e^{i\vartheta}g(F(y)) e i ν τ ( y ) g ( y ) = e i ϑ g ( F ( y )) for μ k \mu_k μ k -a.e. y y y . Thus e i ν τ = e i ϑ g ∘ F / g e^{i\nu\tau}=e^{i\vartheta}g\circ F/g e i ν τ = e i ϑ g ∘ F / g μ k \mu_k μ k -a.e., contradicting (B2).
(c) Since τ ∈ C ∞ \tau\in C^\infty τ ∈ C ∞ , multiplication by τ \tau τ is bounded on H s H^s H s , so L ν = ∑ k ≥ 0 ν k k ! L F ( ( i τ ) k ⋅ ) \mathcal{L}_\nu=\sum_{k\ge0}\frac{\nu^k}{k!}\mathcal{L}_F((i\tau)^k\cdot) L ν = ∑ k ≥ 0 k ! ν k L F (( i τ ) k ⋅ ) converges in L ( H s , H s ) L(H^s,H^s) L ( H s , H s ) and ν ↦ L ν \nu\mapsto\mathcal{L}_\nu ν ↦ L ν is continuous in operator norm. By (b), for each ν ∈ J \nu\in J ν ∈ J there is m ν ∈ N m_\nu\in\mathbb{N} m ν ∈ N with ∥ L ν m ν ∥ H s ≤ 1 2 \|\mathcal{L}_\nu^{m_\nu}\|_{H^s}\le\frac12 ∥ L ν m ν ∥ H s ≤ 2 1 ; by continuity the same holds, with 1 2 \frac12 2 1 replaced by 3 4 \frac34 4 3 , for ν ′ \nu' ν ′ in a neighbourhood of ν \nu ν . Covering J J J by finitely many such neighbourhoods, and using sup ν ∈ J , m ≤ M ∥ L ν m ∥ H s < ∞ \sup_{\nu\in J,m\le M}\|\mathcal{L}_\nu^m\|_{H^s}<\infty sup ν ∈ J , m ≤ M ∥ L ν m ∥ H s < ∞ for each M M M , we obtain (28) . □
Proof of the Dolgopyat estimate (Theorem 2.10)
This section is devoted to the proof of Theorem 2.10. Estimates of this type go back to [7 , 8 ] ; for twisted transfer operators of partially expanding maps see also [9 , 2 ] . Throughout, C C C denotes a generic positive constant (with possible subscripts indicating dependence) that does not depend on n ∈ N n \in\mathbb{N} n ∈ N or ν ∈ R \nu\in\mathbb{R} ν ∈ R , and may change from line to line. We fix ρ > 0 \rho> 0 ρ > 0 , to be chosen in § 4.3 , and set
n ( ν ) : = ⌈ ρ log ∣ ν ∣ ⌉ , ∣ ν ∣ > 1. n(\nu) := \lceil\rho\log|\nu| \rceil, \qquad|\nu| > 1. n ( ν ) := ⌈ ρ log ∣ ν ∣ ⌉ , ∣ ν ∣ > 1.
Road map to the proof
The proof follows the logic of [2 ] and it has two main steps. First, we reduce the Dolgopyat estimate to an L 2 L^2 L 2 contraction estimate at a time proportional to log ∣ ν ∣ \log|\nu| log ∣ ν ∣ . We then prove this contraction by localising in space and frequency and applying the transversality assumption (11) .
Section 4.1 establishes the twisted Lasota–Yorke inequality of Lemma 3.5, using the multiplication estimate in Lemma 4.2. This separates two cases. When the Sobolev part of ∥ h ∥ ν \|h\|_\nu ∥ h ∥ ν is large relative to its L 2 L^2 L 2 part, Lemma 4.3 gives contraction after a fixed number of iterates. For the remaining functions, Corollary 4.4 shows that it suffices to obtain the L 2 L^2 L 2 estimate (30) at the logarithmic time n ( ν ) n(\nu) n ( ν ) . The same corollary explains how these two contraction mechanisms combine to yield the estimate for arbitrary times.
Section 4.2 proves the required L 2 L^2 L 2 estimate. After establishing bounds on inverse branches and a non-stationary phase estimate, we truncate the input frequencies and localise in space. For each inverse branch, we distinguish the output frequencies compatible with the phase gradient from those outside this region. Lemma 4.7 uses (11) to bound the total Jacobian weight of branches compatible with any given frequency; Lemma 4.8 converts this bound into an L 2 L^2 L 2 estimate. The complementary contribution is controlled by non-stationary phase in Lemma 4.9. Finally, Proposition 4.10 combines these estimates with the frequency truncation error and verifies the hypotheses of Corollary 4.4.
Proof of Lemma 3.5
We begin with a classic result.
Lemma 4.1 (Sobolev multiplication). For s > 0 s > 0 s > 0 there is C s > 0 C_s > 0 C s > 0 such that
∥ f u ∥ H s ≤ C s ∥ f ∥ C ⌈ s ⌉ ∥ u ∥ H s \|fu\|_{H^s} \le C_s\|f\|_{C^{\lceil s\rceil}}\|u\|_{H^s} ∥ f u ∥ H s ≤ C s ∥ f ∥ C ⌈ s ⌉ ∥ u ∥ H s for f ∈ C ⌈ s ⌉ f \in C^{\lceil s\rceil} f ∈ C ⌈ s ⌉ and u ∈ H s u \in H^s u ∈ H s .
Proof. For integer s s s this follows from the Leibniz rule. For noninteger s s s , apply that rule through order ⌊ s ⌋ \lfloor s\rfloor ⌊ s ⌋ and use the fractional difference norm at the remaining order; see also [1 ] . □
Lemma 4.2 (Twisted multiplication). There exists C 0 > 0 C_0 > 0 C 0 > 0 and, for each N ∈ N N \in\mathbb{N} N ∈ N , a constant C N > 0 C_N > 0 C N > 0 such that, for all ν ∈ R \nu\in\mathbb{R} ν ∈ R and u ∈ H s u \in H^s u ∈ H s ,
∥ e i ν τ N u ∥ H s ≤ C 0 ∥ u ∥ H s + C N ( 1 + ∣ ν ∣ s ) ∥ u ∥ L 2 . \|e^{i\nu\tau_N}u\|_{H^s} \le C_0\|u\|_{H^s}+C_N(1+|\nu|^s)\|u\|_{L^2}. ∥ e i ν τ N u ∥ H s ≤ C 0 ∥ u ∥ H s + C N ( 1 + ∣ ν ∣ s ) ∥ u ∥ L 2 . Proof. Let f : = e i ν τ N f:=e^{i\nu\tau_N} f := e i ν τ N . Then ∥ f u ∥ L 2 = ∥ u ∥ L 2 \|fu\|_{L^2}=\|u\|_{L^2} ∥ f u ∥ L 2 = ∥ u ∥ L 2 and, for each j ≥ 1 j\ge1 j ≥ 1 , ∥ D j f ∥ C 0 ≤ C N ( 1 + ∣ ν ∣ ) j \|D^j f\|_{C^0}\le C_N(1+|\nu|)^j ∥ D j f ∥ C 0 ≤ C N ( 1 + ∣ ν ∣ ) j . We will repeatedly use the interpolation inequality: for 0 ≤ t < s 0\le t<s 0 ≤ t < s and ε > 0 \varepsilon>0 ε > 0 ,
( 1 + ∣ ν ∣ ) s − t ∥ u ∥ H t ≤ ( 1 + ∣ ν ∣ ) s − t ∥ u ∥ L 2 1 − t s ∥ u ∥ H s t s ≤ ε ∥ u ∥ H s + C ε ( 1 + ∣ ν ∣ ) s ∥ u ∥ L 2 , (29) (1+|\nu|)^{s-t}\|u\|_{H^t}\le(1+|\nu|)^{s-t}\|u\|_{L^2}^{1-\frac{t}{s}}\|u\|_{H^s}^{\frac{t}{s}}\le\varepsilon\|u\|_{H^s}+C_\varepsilon(1+|\nu|)^s\|u\|_{L^2}, \tag*{(29)} ( 1 + ∣ ν ∣ ) s − t ∥ u ∥ H t ≤ ( 1 + ∣ ν ∣ ) s − t ∥ u ∥ L 2 1 − s t ∥ u ∥ H s s t ≤ ε ∥ u ∥ H s + C ε ( 1 + ∣ ν ∣ ) s ∥ u ∥ L 2 , (29) which follows from Young’s inequality. Write s = k + σ s=k+\sigma s = k + σ with k ∈ N ∪ { 0 } k\in\mathbb{N}\cup\{0\} k ∈ N ∪ { 0 } and σ ∈ ( 0 , 1 ] \sigma\in(0,1] σ ∈ ( 0 , 1 ] . We use the equivalent norm ∥ u ∥ L 2 + ∑ ∣ β ∣ = k ∥ ∂ β u ∥ H σ \|u\|_{L^2}+\sum_{|\beta|=k}\|\partial^\beta u\|_{H^\sigma} ∥ u ∥ L 2 + ∑ ∣ β ∣ = k ∥ ∂ β u ∥ H σ , where, for σ ∈ ( 0 , 1 ) \sigma\in(0,1) σ ∈ ( 0 , 1 ) , ∥ w ∥ H σ \|w\|_{H^\sigma} ∥ w ∥ H σ is equivalent to ∥ w ∥ L 2 + [ w ] σ \|w\|_{L^2}+[w]_\sigma ∥ w ∥ L 2 + [ w ] σ , with
[ w ] σ 2 : = ∬ ∣ w ( x ) − w ( y ) ∣ 2 d ( x , y ) 2 + 2 σ d x d y . [w]^2_\sigma:=\iint\frac{|w(x)-w(y)|^2}{d(x,y)^{2+2\sigma}}\,dx\,dy. [ w ] σ 2 := ∬ d ( x , y ) 2 + 2 σ ∣ w ( x ) − w ( y ) ∣ 2 d x d y .
Let g ∈ C 1 g \in C^{1} g ∈ C 1 , w ∈ H σ w \in H^{\sigma} w ∈ H σ . If σ ∈ ( 0 , 1 ) \sigma\in(0,1) σ ∈ ( 0 , 1 ) , from
∣ g ( x ) w ( x ) − g ( y ) w ( y ) ∣ ≤ ∣ g ( x ) ∣ ∣ w ( x ) − w ( y ) ∣ + ∣ g ( x ) − g ( y ) ∣ ∣ w ( y ) ∣ , |g(x)w(x)-g(y)w(y)| \le|g(x)||w(x)-w(y)|+|g(x)-g(y)||w(y)|, ∣ g ( x ) w ( x ) − g ( y ) w ( y ) ∣ ≤ ∣ g ( x ) ∣∣ w ( x ) − w ( y ) ∣ + ∣ g ( x ) − g ( y ) ∣∣ w ( y ) ∣ ,
we have
∣ g ( x ) − g ( y ) ∣ ≤ min { 2 ∥ g ∥ C 0 , ∥ D g ∥ C 0 d ( x , y ) } |g(x)-g(y)| \le\min\{2\|g\|_{C^{0}},\|Dg\|_{C^{0}}d(x,y)\} ∣ g ( x ) − g ( y ) ∣ ≤ min { 2∥ g ∥ C 0 , ∥ D g ∥ C 0 d ( x , y )}
and from
∫ min { a 2 , b 2 d ( x , y ) 2 } d ( x , y ) − 2 − 2 σ d x ≤ C a 2 − 2 σ b 2 σ , \int\min\{a^{2},b^{2}d(x,y)^{2}\}d(x,y)^{-2-2\sigma}dx \le Ca^{2-2\sigma}b^{2\sigma}, ∫ min { a 2 , b 2 d ( x , y ) 2 } d ( x , y ) − 2 − 2 σ d x ≤ C a 2 − 2 σ b 2 σ ,
we get
{#eq:4.2}
[gw]_{\sigma} \le\|g\|_{C^{0}}[w]_{\sigma}+C\|g\|_{C^{0}}^{1-\sigma}\|Dg\|_{C^{0}}^{\sigma}\|w\|_{L^{2}}.
For σ = 1 \sigma=1 σ = 1 the same bound follows from ∇ ( g w ) = g ∇ w + w ∇ g \nabla(gw)=g\nabla w+w\nabla g ∇ ( g w ) = g ∇ w + w ∇ g .
For ∣ β ∣ = k |\beta|=k ∣ β ∣ = k , by the Leibniz rule ∂ β ( f u ) = f ∂ β u + ∑ 0 < γ ≤ β c β , γ ∂ γ f ∂ β − γ u \partial^{\beta}(fu)=f\partial^{\beta}u+\sum_{0<\gamma\le\beta}c_{\beta,\gamma}\partial^{\gamma}f\partial^{\beta-\gamma}u ∂ β ( f u ) = f ∂ β u + ∑ 0 < γ ≤ β c β , γ ∂ γ f ∂ β − γ u . By (4.2) with g = f g=f g = f , ∣ f ∣ = 1 |f|=1 ∣ f ∣ = 1 ,
∥ f ∂ β u ∥ H σ ≤ C ∥ ∂ β u ∥ H σ + C N ( 1 + ∣ ν ∣ ) σ ∥ u ∥ H k , \|f\partial^{\beta}u\|_{H^{\sigma}}\le C\|\partial^{\beta}u\|_{H^{\sigma}}+C_{N}(1+|\nu|)^{\sigma}\|u\|_{H^{k}}, ∥ f ∂ β u ∥ H σ ≤ C ∥ ∂ β u ∥ H σ + C N ( 1 + ∣ ν ∣ ) σ ∥ u ∥ H k ,
with C C C depending only on s s s .
For the remaining terms we can argue as follows. For 0 < γ ≤ β 0<\gamma\le\beta 0 < γ ≤ β , ∣ γ ∣ = j ≥ 1 |\gamma|=j\ge1 ∣ γ ∣ = j ≥ 1 , by (4.2) with g = ∂ γ f g=\partial^{\gamma}f g = ∂ γ f ,
∥ ∂ γ f ∂ β − γ u ∥ H σ ≤ C N ( ( 1 + ∣ ν ∣ ) j ∥ u ∥ H s − j + ( 1 + ∣ ν ∣ ) j + σ ∥ u ∥ H k − j ) . \|\partial^{\gamma}f\partial^{\beta-\gamma}u\|_{H^{\sigma}}\le C_{N}\left((1+|\nu|)^{j}\|u\|_{H^{s-j}}+(1+|\nu|)^{j+\sigma}\|u\|_{H^{k-j}}\right). ∥ ∂ γ f ∂ β − γ u ∥ H σ ≤ C N ( ( 1 + ∣ ν ∣ ) j ∥ u ∥ H s − j + ( 1 + ∣ ν ∣ ) j + σ ∥ u ∥ H k − j ) .
Since ( s − j ) + j = ( k − j ) + ( j + σ ) = k + σ = s (s-j)+j=(k-j)+(j+\sigma)=k+\sigma=s ( s − j ) + j = ( k − j ) + ( j + σ ) = k + σ = s and k + σ = s k+\sigma=s k + σ = s , all the terms with a factor C N C_{N} C N are of the form ( 1 + ∣ ν ∣ ) s − t ∥ u ∥ H t (1+|\nu|)^{s-t}\|u\|_{H^{t}} ( 1 + ∣ ν ∣ ) s − t ∥ u ∥ H t with t < s t<s t < s , and by (29) with ε = ( C N ) − 1 \varepsilon=(C_{N})^{-1} ε = ( C N ) − 1 they are bounded by ∥ u ∥ H s + C N ′ ( 1 + ∣ ν ∣ ) s ∥ u ∥ L 2 \|u\|_{H^{s}}+C'_{N}(1+|\nu|)^{s}\|u\|_{L^{2}} ∥ u ∥ H s + C N ′ ( 1 + ∣ ν ∣ ) s ∥ u ∥ L 2 . Summing over β \beta β gives the claim, with C 0 C_{0} C 0 depending only on s s s . □
Note that, by Cauchy–Schwarz, ∣ L ν n u ∣ 2 ≤ L F n 1 ⋅ L F n ∣ u ∣ 2 |\mathcal{L}_{\nu}^{n}u|^{2}\le\mathcal{L}_{F}^{n}1\cdot\mathcal{L}_{F}^{n}|u|^{2} ∣ L ν n u ∣ 2 ≤ L F n 1 ⋅ L F n ∣ u ∣ 2 ; hence, by (8) ,
{#eq:4.3}
\|\mathcal{L}_{\nu}^{n}u\|_{L^{2}}\le\|\mathcal{L}_{F}^{n}1\|_{L^{\infty}}^{1/2}\|u\|_{L^{2}}\le C_{F}^{1/2}\Theta^{n/2}\|u\|_{L^{2}}\qquad\forall n\in\mathbb{N},\ \nu\in\mathbb{R}.
Proof of Lemma 3.5. Let σ ∈ ( σ F , 1 ) \sigma\in(\sigma_{F},1) σ ∈ ( σ F , 1 ) and choose σ ∗ ∈ ( σ F , σ ) \sigma_{*}\in(\sigma_{F},\sigma) σ ∗ ∈ ( σ F , σ ) . Inequality (20) holds with this larger rate and the admissible s ′ s' s ′ chosen above. Fix N ∈ N N\in\mathbb{N} N ∈ N . By (22) , (20) and Lemma 4.2 (applied with s s s and with s ′ s' s ′ ),
∥ L ν N u ∥ H s ≤ C σ ∗ N ( C 0 ∥ u ∥ H s + C N ( 1 + ∣ ν ∣ ) s ∥ u ∥ L 2 ) + C ( C 0 ∥ u ∥ H s ′ + C N ( 1 + ∣ ν ∣ ) s ′ ∥ u ∥ L 2 ) . \|\mathcal{L}_{\nu}^{N}u\|_{H^{s}}\le C\sigma_{*}^{N}(C_{0}\|u\|_{H^{s}}+C_{N}(1+|\nu|)^{s}\|u\|_{L^{2}})+C(C_{0}\|u\|_{H^{s'}}+C_{N}(1+|\nu|)^{s'}\|u\|_{L^{2}}). ∥ L ν N u ∥ H s ≤ C σ ∗ N ( C 0 ∥ u ∥ H s + C N ( 1 + ∣ ν ∣ ) s ∥ u ∥ L 2 ) + C ( C 0 ∥ u ∥ H s ′ + C N ( 1 + ∣ ν ∣ ) s ′ ∥ u ∥ L 2 ) .
By interpolation, for each ε > 0 \varepsilon>0 ε > 0 , ∥ u ∥ H s ′ ≤ ε ∥ u ∥ H s + C ε ∥ u ∥ L 2 \|u\|_{H^{s'}}\le\varepsilon\|u\|_{H^{s}}+C_{\varepsilon}\|u\|_{L^{2}} ∥ u ∥ H s ′ ≤ ε ∥ u ∥ H s + C ε ∥ u ∥ L 2 . Multiplying by ( 1 + ∣ ν ∣ ) − s ≤ 1 (1+|\nu|)^{-s}\le1 ( 1 + ∣ ν ∣ ) − s ≤ 1 ,
( 1 + ∣ ν ∣ ) − s ∥ L ν N u ∥ H s ≤ ( C C 0 σ ∗ N + C C 0 ε ) ( 1 + ∣ ν ∣ ) − s ∥ u ∥ H s + C N , ε ∥ u ∥ L 2 . (1+|\nu|)^{-s}\|\mathcal{L}_{\nu}^{N}u\|_{H^{s}}\le(CC_{0}\sigma_{*}^{N}+CC_{0}\varepsilon)(1+|\nu|)^{-s}\|u\|_{H^{s}}+C_{N,\varepsilon}\|u\|_{L^{2}}. ( 1 + ∣ ν ∣ ) − s ∥ L ν N u ∥ H s ≤ ( C C 0 σ ∗ N + C C 0 ε ) ( 1 + ∣ ν ∣ ) − s ∥ u ∥ H s + C N , ε ∥ u ∥ L 2 .
Since σ ∗ < σ \sigma_{*}<\sigma σ ∗ < σ , we can choose N N N so large that C C 0 σ ∗ N ≤ 1 2 σ N CC_{0}\sigma_{*}^{N}\le\frac{1}{2}\sigma^{N} C C 0 σ ∗ N ≤ 2 1 σ N , and then ε \varepsilon ε so small that C C 0 ε ≤ 1 2 σ N CC_{0}\varepsilon\le\frac{1}{2}\sigma^{N} C C 0 ε ≤ 2 1 σ N . Together with (4.3),
∥ L ν N u ∥ ν ≤ σ N ∥ u ∥ ν + C ′ ∥ u ∥ L 2 . \|\mathcal{L}_{\nu}^{N}u\|_{\nu}\le\sigma^{N}\|u\|_{\nu}+C'\|u\|_{L^{2}}. ∥ L ν N u ∥ ν ≤ σ N ∥ u ∥ ν + C ′ ∥ u ∥ L 2 .
Applying this to L ν k N u \mathcal{L}_{\nu}^{kN}u L ν k N u and using (4.3), we get X k + 1 ≤ σ N X k + C ′ ′ Θ k N / 2 ∥ u ∥ L 2 X_{k+1}\le\sigma^{N}X_{k}+C''\Theta^{kN/2}\|u\|_{L^{2}} X k + 1 ≤ σ N X k + C ′′ Θ k N /2 ∥ u ∥ L 2 for X k : = ∥ L ν k N u ∥ ν X_{k}:=\|\mathcal{L}_{\nu}^{kN}u\|_{\nu} X k := ∥ L ν k N u ∥ ν . Since σ < 1 ≤ Θ 1 / 2 \sigma<1\le\Theta^{1/2} σ < 1 ≤ Θ 1/2 , induction gives X k ≤ σ k N ∥ u ∥ ν + C ′ ′ ′ Θ k N / 2 ∥ u ∥ L 2 X_{k}\le\sigma^{kN}\|u\|_{\nu}+C'''\Theta^{kN/2}\|u\|_{L^{2}} X k ≤ σ k N ∥ u ∥ ν + C ′′′ Θ k N /2 ∥ u ∥ L 2 . The same computation with N N N replaced by m < N m<N m < N gives ∥ L ν m u ∥ ν ≤ C ∥ u ∥ ν \|\mathcal{L}_{\nu}^{m}u\|_{\nu}\le C\|u\|_{\nu} ∥ L ν m u ∥ ν ≤ C ∥ u ∥ ν . Writing n = k N + m n=kN+m n = k N + m yields (23) . □
Lemma 4.3. Let σ \sigma σ , C σ C_{\sigma} C σ and C τ C_{\tau} C τ be as in Lemma 3.5 . There exist n 0 : = n 0 ( σ ) ∈ N n_{0}:=n_{0}(\sigma)\in\mathbb{N} n 0 := n 0 ( σ ) ∈ N and σ 1 ∈ ( 0 , 1 ) \sigma_{1}\in(0,1) σ 1 ∈ ( 0 , 1 ) such that, setting C ‾ : = max { C σ , C τ Θ n 0 / 2 } \overline{C}:=\max\{C_{\sigma},C_{\tau}\Theta^{n_{0}/2}\} C := max { C σ , C τ Θ n 0 /2 } , for each ν ∈ R \nu\in\mathbb{R} ν ∈ R , if h ∈ H s h\in H^{s} h ∈ H s satisfies
2 C ‾ ( 1 + ∣ ν ∣ ) s ∥ h ∥ L 2 ≤ ∥ h ∥ H s , 2\overline{C}(1+|\nu|)^{s}\|h\|_{L^{2}}\le\|h\|_{H^{s}}, 2 C ( 1 + ∣ ν ∣ ) s ∥ h ∥ L 2 ≤ ∥ h ∥ H s ,
then ∥ L ν n 0 h ∥ ν ≤ σ 1 ∥ h ∥ ν \|\mathcal{L}_{\nu}^{n_{0}}h\|_{\nu}\le\sigma_{1}\|h\|_{\nu} ∥ L ν n 0 h ∥ ν ≤ σ 1 ∥ h ∥ ν .
Proof. Fix σ 1 ∈ ( 1 2 , 1 ) \sigma_1 \in\left(\frac{1}{2},1\right) σ 1 ∈ ( 2 1 , 1 ) and choose n 0 ∈ N n_0 \in\mathbb{N} n 0 ∈ N such that C σ σ n 0 ≤ σ 1 − 1 2 C_\sigma\sigma^{n_0} \le\sigma_1-\frac{1}{2} C σ σ n 0 ≤ σ 1 − 2 1 . By Lemma 3.5 ,
∥ L ν n 0 h ∥ ν ≤ C σ σ n 0 ∥ h ∥ ν + C τ Θ n 0 / 2 ∥ h ∥ L 2 . \left\|\mathcal{L}_{\nu}^{n_0}h\right\|_\nu\le C_\sigma\sigma^{n_0}\|h\|_\nu+C_\tau\Theta^{n_0/2}\|h\|_{L^2}. ∥ L ν n 0 h ∥ ν ≤ C σ σ n 0 ∥ h ∥ ν + C τ Θ n 0 /2 ∥ h ∥ L 2 .
Using the assumption on h h h and C ‾ ≥ C τ Θ n 0 / 2 \overline{C}\ge C_\tau\Theta^{n_0/2} C ≥ C τ Θ n 0 /2 ,
C τ Θ n 0 / 2 ∥ h ∥ L 2 ≤ C ‾ ∥ h ∥ L 2 ≤ 1 2 ( 1 + ∣ ν ∣ ) − s ∥ h ∥ H s ≤ 1 2 ∥ h ∥ ν . C_\tau\Theta^{n_0/2}\|h\|_{L^2}\le\overline{C}\|h\|_{L^2}\le\frac{1}{2}(1+|\nu|)^{-s}\|h\|_{H^s}\le\frac{1}{2}\|h\|_\nu. C τ Θ n 0 /2 ∥ h ∥ L 2 ≤ C ∥ h ∥ L 2 ≤ 2 1 ( 1 + ∣ ν ∣ ) − s ∥ h ∥ H s ≤ 2 1 ∥ h ∥ ν .
Therefore,
∥ L ν n 0 h ∥ ν ≤ ( C σ σ n 0 + 1 2 ) ∥ h ∥ ν ≤ σ 1 ∥ h ∥ ν . \left\|\mathcal{L}_{\nu}^{n_0}h\right\|_\nu\le\left(C_\sigma\sigma^{n_0}+\frac{1}{2}\right)\|h\|_\nu\le\sigma_1\|h\|_\nu. ∥ L ν n 0 h ∥ ν ≤ ( C σ σ n 0 + 2 1 ) ∥ h ∥ ν ≤ σ 1 ∥ h ∥ ν .
Corollary 4.4 (L 2 L^2 L 2 -estimate reduction). Fix σ ∈ ( σ F , 1 ) \sigma\in(\sigma_F,1) σ ∈ ( σ F , 1 ) with σ Θ 1 / 2 < 1 \sigma\Theta^{1/2}<1 σ Θ 1/2 < 1 in Lemma 3.5 (possible by (P)), and assume that ρ log Θ < 1 4 \rho\log\Theta<\frac{1}{4} ρ log Θ < 4 1 . Assume there exist ν 0 > 1 \nu_0>1 ν 0 > 1 , C 1 > 0 C_1>0 C 1 > 0 and γ 4 > 1 2 log Θ \gamma_4>\frac{1}{2}\log\Theta γ 4 > 2 1 log Θ such that, for all ∣ ν ∣ ≥ ν 0 |\nu|\ge\nu_0 ∣ ν ∣ ≥ ν 0 and all h ∈ H s h\in H^s h ∈ H s with ∥ h ∥ H s ≤ 2 C ‾ ( 1 + ∣ ν ∣ ) s ∥ h ∥ L 2 \|h\|_{H^s}\le2\overline{C}(1+|\nu|)^s\|h\|_{L^2} ∥ h ∥ H s ≤ 2 C ( 1 + ∣ ν ∣ ) s ∥ h ∥ L 2 ,
∥ L ν n ( ν ) h ∥ L 2 ≤ C 1 e − n ( ν ) γ 4 ∥ h ∥ L 2 . (30) \|\mathcal{L}_{\nu}^{n(\nu)}h\|_{L^2}\le C_1e^{-n(\nu)\gamma_4}\|h\|_{L^2}. \tag*{(30)} ∥ L ν n ( ν ) h ∥ L 2 ≤ C 1 e − n ( ν ) γ 4 ∥ h ∥ L 2 . (30) Then, up to enlarging ν 0 \nu_0 ν 0 , Theorem 2.10 holds, i.e. there exist C > 0 C>0 C > 0 , γ > 0 \gamma>0 γ > 0 and δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) such that (16) holds.
Proof. Fix ∣ ν ∣ ≥ ν 0 |\nu|\ge\nu_0 ∣ ν ∣ ≥ ν 0 and set N : = n ( ν ) N:=n(\nu) N := n ( ν ) ; enlarging ν 0 \nu_0 ν 0 we may assume 2 N ≥ n 0 2N\ge n_0 2 N ≥ n 0 . By Lemma 3.5 and ∥ ⋅ ∥ L 2 ≤ ∥ ⋅ ∥ ν \|\cdot\|_{L^2}\le\|\cdot\|_\nu ∥ ⋅ ∥ L 2 ≤ ∥ ⋅ ∥ ν , ∥ L ν N h ∥ ν ≤ C Θ N / 2 ∥ h ∥ ν \|\mathcal{L}_\nu^N h\|_\nu\le C\Theta^{N/2}\|h\|_\nu ∥ L ν N h ∥ ν ≤ C Θ N /2 ∥ h ∥ ν ; applying Lemma 3.5 once more,
∥ L ν 2 N h ∥ ν ≤ C σ N ∥ L ν N h ∥ ν + C Θ N / 2 ∥ L ν N h ∥ L 2 ≤ C ( ( σ Θ 1 / 2 ) N ∥ h ∥ ν + C Θ N / 2 ∥ L ν N h ∥ L 2 ) . (31) \|\mathcal{L}_\nu^{2N}h\|_\nu\le C\sigma^N\|\mathcal{L}_\nu^N h\|_\nu+C\Theta^{N/2}\|\mathcal{L}_\nu^N h\|_{L^2}\le C\left((\sigma\Theta^{1/2})^N\|h\|_\nu+C\Theta^{N/2}\|\mathcal{L}_\nu^N h\|_{L^2}\right). \tag*{(31)} ∥ L ν 2 N h ∥ ν ≤ C σ N ∥ L ν N h ∥ ν + C Θ N /2 ∥ L ν N h ∥ L 2 ≤ C ( ( σ Θ 1/2 ) N ∥ h ∥ ν + C Θ N /2 ∥ L ν N h ∥ L 2 ) . (31)
Case 1: ∥ h ∥ H s ≥ 2 C ‾ ( 1 + ∣ ν ∣ ) s ∥ h ∥ L 2 \|h\|_{H^s}\ge2\overline{C}(1+|\nu|)^s\|h\|_{L^2} ∥ h ∥ H s ≥ 2 C ( 1 + ∣ ν ∣ ) s ∥ h ∥ L 2 . By Lemma 4.3, ∥ L ν n 0 h ∥ ν ≤ e − γ 6 n 0 ∥ h ∥ ν \|\mathcal{L}_\nu^{n_0}h\|_\nu\le e^{-\gamma_6n_0}\|h\|_\nu ∥ L ν n 0 h ∥ ν ≤ e − γ 6 n 0 ∥ h ∥ ν with γ 6 : = log σ − 1 n 0 \gamma_6:=\frac{\log\sigma^{-1}}{n_0} γ 6 := n 0 l o g σ − 1 .
Case 2: ∥ h ∥ H s ≤ 2 C ‾ ( 1 + ∣ ν ∣ ) s ∥ h ∥ L 2 \|h\|_{H^s}\le2\overline{C}(1+|\nu|)^s\|h\|_{L^2} ∥ h ∥ H s ≤ 2 C ( 1 + ∣ ν ∣ ) s ∥ h ∥ L 2 . By (31) and (30) , ∥ L ν 2 N h ∥ ν ≤ C ( ( σ Θ 1 / 2 ) N + C 1 e − ( γ 4 − 1 2 log Θ ) N ) ∥ h ∥ ν ≤ e − 2 γ 5 N ∥ h ∥ ν \|\mathcal{L}_\nu^{2N}h\|_\nu\le C\left((\sigma\Theta^{1/2})^N+C_1e^{-(\gamma_4-\frac{1}{2}\log\Theta)N}\right)\|h\|_\nu\le e^{-2\gamma_5N}\|h\|_\nu ∥ L ν 2 N h ∥ ν ≤ C ( ( σ Θ 1/2 ) N + C 1 e − ( γ 4 − 2 1 l o g Θ ) N ) ∥ h ∥ ν ≤ e − 2 γ 5 N ∥ h ∥ ν , with γ 5 : = 1 4 min { log ( σ Θ 1 / 2 ) − 1 , γ 4 − 1 2 log Θ } > 0 \gamma_5:=\frac{1}{4}\min\{\log(\sigma\Theta^{1/2})^{-1},\gamma_4-\frac{1}{2}\log\Theta\}>0 γ 5 := 4 1 min { log ( σ Θ 1/2 ) − 1 , γ 4 − 2 1 log Θ } > 0 , up to enlarging ν 0 \nu_0 ν 0 .
Let γ ′ = min { γ 5 , γ 6 } \gamma'=\min\{\gamma_5,\gamma_6\} γ ′ = min { γ 5 , γ 6 } and γ : = min { γ ′ , ( 4 ρ ) − 1 } \gamma:=\min\{\gamma',(4\rho)^{-1}\} γ := min { γ ′ , ( 4 ρ ) − 1 } . We have shown that for each h ∈ H s h\in H^s h ∈ H s there is t ( h ) ∈ { n 0 , 2 N } t(h)\in\{n_0,2N\} t ( h ) ∈ { n 0 , 2 N } with ∥ L ν t ( h ) h ∥ ν ≤ e − γ t ( h ) ∥ h ∥ ν \|\mathcal{L}_\nu^{t(h)}h\|_\nu\le e^{-\gamma t(h)}\|h\|_\nu ∥ L ν t ( h ) h ∥ ν ≤ e − γ t ( h ) ∥ h ∥ ν . Set h 0 = h h_0=h h 0 = h , h j + 1 = L ν t ( h j ) h j h_{j+1}=\mathcal{L}_\nu^{t(h_j)}h_j h j + 1 = L ν t ( h j ) h j and T j = ∑ i < j t ( h i ) T_j=\sum_{i<j}t(h_i) T j = ∑ i < j t ( h i ) , so that ∥ h j ∥ ν ≤ e − γ ′ T j ∥ h ∥ ν ≤ e − γ T j ∥ h ∥ ν \|h_j\|_\nu\le e^{-\gamma'T_j}\|h\|_\nu\le e^{-\gamma T_j}\|h\|_\nu ∥ h j ∥ ν ≤ e − γ ′ T j ∥ h ∥ ν ≤ e − γ T j ∥ h ∥ ν . Given n n n , let j j j be such that T j ≤ n < T j + 1 T_j\le n<T_{j+1} T j ≤ n < T j + 1 ; then n − T j < 2 N n-T_j<2N n − T j < 2 N , so e − γ T j ≤ e 2 γ N e − γ n e^{-\gamma T_j}\le e^{2\gamma N}e^{-\gamma n} e − γ T j ≤ e 2 γ N e − γ n and, since sup m < 2 N ∥ L ν m ∥ ν ≤ C Θ N ≤ C ∣ ν ∣ ρ log Θ \sup_{m<2N}\|\mathcal{L}_\nu^m\|_\nu\le C\Theta^N\le C|\nu|^{\rho\log\Theta} sup m < 2 N ∥ L ν m ∥ ν ≤ C Θ N ≤ C ∣ ν ∣ ρ l o g Θ by (23) ,
∥ L ν n h ∥ ν = ∥ L ν n − T j h j ∥ ν ≤ C ∣ ν ∣ ρ log Θ e 2 γ N e − γ n ∥ h ∥ ν ≤ C ∣ ν ∣ ρ log Θ + 2 γ ρ e − γ n ∥ h ∥ ν . \|\mathcal{L}_\nu^n h\|_\nu=\|\mathcal{L}_\nu^{n-T_j}h_j\|_\nu\le C|\nu|^{\rho\log\Theta}e^{2\gamma N}e^{-\gamma n}\|h\|_\nu\le C|\nu|^{\rho\log\Theta+2\gamma\rho}e^{-\gamma n}\|h\|_\nu. ∥ L ν n h ∥ ν = ∥ L ν n − T j h j ∥ ν ≤ C ∣ ν ∣ ρ l o g Θ e 2 γ N e − γ n ∥ h ∥ ν ≤ C ∣ ν ∣ ρ l o g Θ + 2 γ ρ e − γ n ∥ h ∥ ν .
The claim follows with δ : = ρ log Θ + 2 γ ρ < 3 4 \delta:=\rho\log\Theta+2\gamma\rho<\frac{3}{4} δ := ρ log Θ + 2 γ ρ < 4 3 .
Thanks to Corollary 4.4, we are left to prove (30) for functions satisfying ∥ h ∥ H s ≤ 2 C ‾ ( 1 + ∣ ν ∣ ) s ∥ h ∥ L 2 \|h\|_{H^s}\le2\overline{C}(1+|\nu|)^s\|h\|_{L^2} ∥ h ∥ H s ≤ 2 C ( 1 + ∣ ν ∣ ) s ∥ h ∥ L 2 . This is done in the next subsection.
Proof of (4.4): a captivity estimate
Preliminaries
For q ∈ T 2 q\in\mathbb{T}^2 q ∈ T 2 and n ∈ N n\in\mathbb{N} n ∈ N we denote by H n ( q ) \mathcal{H}^n(q) H n ( q ) the set of inverse branches of F n F^n F n defined on a ball B B B centred at q q q of radius smaller than 1 4 \frac{1}{4} 4 1 (they exist since F n : T 2 → T 2 F^n:\mathbb{T}^2\to\mathbb{T}^2 F n : T 2 → T 2 is a covering map and B B B is simply connected). For h ∈ H n ( q ) h\in\mathcal{H}^n(q) h ∈ H n ( q ) we write
J h ( y ) : = ∣ det D h ( y ) F n ∣ − 1 , c h ( y ) : = d ( τ n ∘ h ) y ∈ ( R 2 ) ∗ , y ∈ B . J_h(y):=\left|\det D_{h(y)}F^n\right|^{-1},\qquad c_h(y):=d(\tau_n\circ h)_y\in(\mathbb{R}^2)^*,\qquad y\in B. J h ( y ) := det D h ( y ) F n − 1 , c h ( y ) := d ( τ n ∘ h ) y ∈ ( R 2 ) ∗ , y ∈ B .
With this notation, for y ∈ B y\in B y ∈ B ,
L ν n u ( y ) = ∑ h ∈ H n ( q ) J h ( y ) e i ν τ n ( h ( y ) ) u ( h ( y ) ) . (32) \mathcal{L}_\nu^n u(y)=\sum_{h\in\mathcal{H}^n(q)}J_h(y)e^{i\nu\tau_n(h(y))}u(h(y)). \tag*{(32)} L ν n u ( y ) = h ∈ H n ( q ) ∑ J h ( y ) e i ν τ n ( h ( y )) u ( h ( y )) . (32)
Lemma 4.5 (A priori bounds). For j ≥ 1 j \ge1 j ≥ 1 and a map φ \varphi φ defined on an open set of T 2 \mathbb{T}^{2} T 2 (with values in T 2 \mathbb{T}^{2} T 2 , lifted locally to R 2 \mathbb{R}^{2} R 2 ) let ∣ φ ∣ j : = max 1 ≤ i ≤ j sup ∥ D i φ ∥ |\varphi|_{j}:=\max_{1\le i\le j}\sup\|D^{i}\varphi\| ∣ φ ∣ j := max 1 ≤ i ≤ j sup ∥ D i φ ∥ ; for a real function ψ \psi ψ let ∥ ψ ∥ C j : = sup ∣ ψ ∣ + ∣ ψ ∣ j \|\psi\|_{C^{j}}:=\sup|\psi|+|\psi|_{j} ∥ ψ ∥ C j := sup ∣ ψ ∣ + ∣ ψ ∣ j . There exist constants C 1 , C 2 , ⋯ > 0 C_{1},C_{2},\cdots>0 C 1 , C 2 , ⋯ > 0 and p 1 ≤ p 2 ≤ ⋯ p_{1}\le p_{2}\le\cdots p 1 ≤ p 2 ≤ ⋯ , with p 1 ≥ log ( Λ λ c Θ ) p_{1}\ge\log(\Lambda\lambda_{c}\Theta) p 1 ≥ log ( Λ λ c Θ ) , such that, for all n ∈ N n\in\mathbb{N} n ∈ N , q ∈ T 2 q\in\mathbb{T}^{2} q ∈ T 2 , h ∈ H n ( q ) h\in\mathcal{H}^{n}(q) h ∈ H n ( q ) (defined on the ball B B B centred at q q q ):
(a) sup B ∥ D h ∥ ≤ C 1 λ c n \sup_{B}\|Dh\|\le C_{1}\lambda_{c}^{n} sup B ∥ D h ∥ ≤ C 1 λ c n ;
(b) if u ∈ D h ( y ) F n C u u\in D_{h(y)}F^{n}C_{u} u ∈ D h ( y ) F n C u , y ∈ B y\in B y ∈ B , is a unit vector, then ∥ D y h u ∥ ≤ C ⋆ λ − n \|D_{y}hu\|\le C_{\star}\lambda^{-n} ∥ D y h u ∥ ≤ C ⋆ λ − n ;
(c) ∣ h ∣ j + ∥ τ n ∘ h ∥ C j + ∥ log J h ∥ C j ≤ C j e p j n |h|_{j}+\|\tau_{n}\circ h\|_{C^{j}}+\|\log J_{h}\|_{C^{j}}\le C_{j}e^{p_{j}n} ∣ h ∣ j + ∥ τ n ∘ h ∥ C j + ∥ log J h ∥ C j ≤ C j e p j n and ∥ D j J h ∥ ≤ C j J h e j p j n \|D^{j}J_{h}\|\le C_{j}J_{h}e^{jp_{j}n} ∥ D j J h ∥ ≤ C j J h e j p j n on B B B ;
(d) # H n ( q ) ≤ C ∞ C ∗ ( Λ λ c Θ ) n \#\mathcal{H}^{n}(q)\le C_{\infty}C_{\ast}(\Lambda\lambda_{c}\Theta)^{n} # H n ( q ) ≤ C ∞ C ∗ ( Λ λ c Θ ) n .
Proof. We divide the proof in steps.
Preliminaries. Since F n F^{n} F n is a local diffeomorphism of a compact manifold, it is a covering map; as B B B is simply connected, F − n ( B ) F^{-n}(B) F − n ( B ) is a disjoint union of open sets, each mapped diffeomorphically onto B B B by F n F^{n} F n , and the elements of H n ( q ) \mathcal{H}^{n}(q) H n ( q ) are the inverses of these restrictions. In particular h ↦ h ( q ) h\mapsto h(q) h ↦ h ( q ) is a bijection between H n ( q ) \mathcal{H}^{n}(q) H n ( q ) and F − n ( q ) F^{-n}(q) F − n ( q ) . For h ∈ H n ( q ) h\in\mathcal{H}^{n}(q) h ∈ H n ( q ) and 0 ≤ k ≤ n 0\le k\le n 0 ≤ k ≤ n set
h ( k ) : = F n − k ∘ h : B → T 2 . h^{(k)}:=F^{n-k}\circ h:B\to\mathbb{T}^{2}. h ( k ) := F n − k ∘ h : B → T 2 .
Since F k ∘ h ( k ) = F n ∘ h = Id B F^{k}\circ h^{(k)}=F^{n}\circ h=\operatorname{Id}_{B} F k ∘ h ( k ) = F n ∘ h = Id B , we have h ( k ) ∈ H k ( q ) h^{(k)}\in\mathcal{H}^{k}(q) h ( k ) ∈ H k ( q ) , h ( 0 ) = Id B h^{(0)}=\operatorname{Id}_{B} h ( 0 ) = Id B and h ( n ) = h h^{(n)}=h h ( n ) = h . Moreover F F F is injective on h ( k + 1 ) ( B ) h^{(k+1)}(B) h ( k + 1 ) ( B ) (because F ∘ h ( k + 1 ) = h ( k ) F\circ h^{(k+1)}=h^{(k)} F ∘ h ( k + 1 ) = h ( k ) is injective), so g k : = ( F ∣ h ( k + 1 ) ( B ) ) − 1 : h ( k ) ( B ) → h ( k + 1 ) ( B ) g_{k}:=(F|_{h^{(k+1)}(B)})^{-1}:h^{(k)}(B)\to h^{(k+1)}(B) g k := ( F ∣ h ( k + 1 ) ( B ) ) − 1 : h ( k ) ( B ) → h ( k + 1 ) ( B ) is well defined and
h ( k + 1 ) = g k ∘ h ( k ) , D g k = ( D F ) − 1 ∘ g k . (33) h^{(k+1)}=g_{k}\circ h^{(k)},\qquad Dg_{k}=(DF)^{-1}\circ g_{k}. \tag*{(33)} h ( k + 1 ) = g k ∘ h ( k ) , D g k = ( D F ) − 1 ∘ g k . (33)
By (A1) and compactness, inf T 2 det D F > 0 \inf_{\mathbb{T}^{2}}\det DF>0 inf T 2 det D F > 0 , so sup ∥ ( D F ) − 1 ∥ < ∞ \sup\|(DF)^{-1}\|<\infty sup ∥ ( D F ) − 1 ∥ < ∞ ; differentiating D g k = ( D F ) − 1 ∘ g k Dg_{k}=(DF)^{-1}\circ g_{k} D g k = ( D F ) − 1 ∘ g k repeatedly, D i g k D^{i}g_{k} D i g k is a polynomial in ( D F ) − 1 ∘ g k (DF)^{-1}\circ g_{k} ( D F ) − 1 ∘ g k and D l F ∘ g k D^{l}F\circ g_{k} D l F ∘ g k , l ≤ i l\le i l ≤ i . Hence there are constants G i G_{i} G i , depending only on ∥ F ∥ C i \|F\|_{C^{i}} ∥ F ∥ C i and sup ∥ ( D F ) − 1 ∥ \sup\|(DF)^{-1}\| sup ∥ ( D F ) − 1 ∥ , with
sup ∥ D i g k ∥ ≤ G i for all i ≥ 1 , k ≥ 0. (34) \sup\|D^{i}g_{k}\|\le G_{i}\qquad\text{for all }i\ge1,\ k\ge0. \tag*{(34)} sup ∥ D i g k ∥ ≤ G i for all i ≥ 1 , k ≥ 0. (34)
(a). Let p ∈ T 2 p\in\mathbb{T}^{2} p ∈ T 2 , k ∈ N k\in\mathbb{N} k ∈ N and w ∈ T F k ( p ) T 2 w\in T_{F^{k}(p)}\mathbb{T}^{2} w ∈ T F k ( p ) T 2 . Since χ u χ c < 1 \chi_{u}\chi_{c}<1 χ u χ c < 1 , there is θ 0 > 0 \theta_{0}>0 θ 0 > 0 , depending only on χ u , χ c \chi_{u},\chi_{c} χ u , χ c , such that the angle between any nonzero vector of C u C_{u} C u and any nonzero vector of C c C_{c} C c is at least θ 0 \theta_{0} θ 0 . By (4) , D p F k C u ⊂ C u D_{p}F^{k}C_{u}\subset C_{u} D p F k C u ⊂ C u ; hence we can write w = e + c w=e+c w = e + c with e ∈ D p F k C u e\in D_{p}F^{k}C_{u} e ∈ D p F k C u , c ∈ C c c\in C_{c} c ∈ C c and ∥ e ∥ , ∥ c ∥ ≤ ∥ w ∥ / sin θ 0 \|e\|,\|c\|\le\|w\|/\sin\theta_{0} ∥ e ∥ , ∥ c ∥ ≤ ∥ w ∥/ sin θ 0 . Write e = D p F k v e=D_{p}F^{k}v e = D p F k v with v ∈ C u v\in C_{u} v ∈ C u ; since C u ∖ { 0 } ⊂ R 2 ∖ C c C_{u}\setminus\{0\}\subset\mathbb{R}^{2}\setminus C_{c} C u ∖ { 0 } ⊂ R 2 ∖ C c , (5) gives ∥ e ∥ ≥ C ∗ − 1 λ k ∥ v ∥ \|e\|\ge C_{\ast}^{-1}\lambda^{k}\|v\| ∥ e ∥ ≥ C ∗ − 1 λ k ∥ v ∥ , i.e. ∥ ( D p F k ) − 1 e ∥ ≤ C ∗ λ − k ∥ e ∥ \|(D_{p}F^{k})^{-1}e\|\le C_{\ast}\lambda^{-k}\|e\| ∥ ( D p F k ) − 1 e ∥ ≤ C ∗ λ − k ∥ e ∥ . Moreover ∥ ( D p F k ) − 1 c ∥ ≤ λ c , k ( p ) ∥ c ∥ ≤ C ∗ ( λ c + ) k ∥ c ∥ \|(D_{p}F^{k})^{-1}c\|\le\lambda_{c,k}(p)\|c\|\le C_{\ast}(\lambda_{c}^{+})^{k}\|c\| ∥ ( D p F k ) − 1 c ∥ ≤ λ c , k ( p ) ∥ c ∥ ≤ C ∗ ( λ c + ) k ∥ c ∥ . Since λ > 1 \lambda>1 λ > 1 and λ c + ≤ λ c \lambda_{c}^{+}\le\lambda_{c} λ c + ≤ λ c , λ c ≥ 1 \lambda_{c}\ge1 λ c ≥ 1 ,
∥ ( D p F k ) − 1 ∥ ≤ 2 C ∗ sin θ 0 λ c k = : C 1 ′ λ c k . (35) \|(D_{p}F^{k})^{-1}\|\le\frac{2C_{\ast}}{\sin\theta_{0}}\lambda_{c}^{k}=:C'_{1}\lambda_{c}^{k}. \tag*{(35)} ∥ ( D p F k ) − 1 ∥ ≤ sin θ 0 2 C ∗ λ c k =: C 1 ′ λ c k . (35)
As D y h ( k ) = ( D h ( k ) ( y ) F k ) − 1 D_{y}h^{(k)}=(D_{h^{(k)}(y)}F^{k})^{-1} D y h ( k ) = ( D h ( k ) ( y ) F k ) − 1 , this gives sup B ∥ D h ( k ) ∥ ≤ C 1 ′ λ c k \sup_{B}\|Dh^{(k)}\|\le C'_{1}\lambda_{c}^{k} sup B ∥ D h ( k ) ∥ ≤ C 1 ′ λ c k for all 0 ≤ k ≤ n 0\le k\le n 0 ≤ k ≤ n , and (a) for k = n k=n k = n .
(b). Write u = D h ( y ) F n v / ∥ D h ( y ) F n v ∥ u=D_{h(y)}F^{n}v/\|D_{h(y)}F^{n}v\| u = D h ( y ) F n v /∥ D h ( y ) F n v ∥ with v ∈ C u v\in C_{u} v ∈ C u . As above, ∥ D h ( y ) F n v ∥ ≥ C ∗ − 1 λ n ∥ v ∥ \|D_{h(y)}F^{n}v\|\ge C_{\ast}^{-1}\lambda^{n}\|v\| ∥ D h ( y ) F n v ∥ ≥ C ∗ − 1 λ n ∥ v ∥ , hence ∥ D y h u ∥ = ∥ v ∥ / ∥ D h ( y ) F n v ∥ ≤ C ∗ λ − n \|D_{y}hu\|=\|v\|/\|D_{h(y)}F^{n}v\|\le C_{\ast}\lambda^{-n} ∥ D y h u ∥ = ∥ v ∥/∥ D h ( y ) F n v ∥ ≤ C ∗ λ − n .
(c), bounds on h h h . Let a k ( i ) : = sup B ∥ D i h ( k ) ∥ a_{k}^{(i)}:=\sup_{B}\|D^{i}h^{(k)}\| a k ( i ) := sup B ∥ D i h ( k ) ∥ . We prove by induction on i i i that there exist c i > 0 c_{i}>0 c i > 0 and π i \pi_{i} π i with
a k ( i ) ≤ c i e π i k for all k ≥ 0. (36) a_{k}^{(i)}\le c_{i}e^{\pi_{i}k}\qquad\text{for all }k\ge0. \tag*{(36)} a k ( i ) ≤ c i e π i k for all k ≥ 0. (36)
Enlarge the bounds G i G_{i} G i so that G i ≥ 1 G_{i}\ge1 G i ≥ 1 , and choose c i c_{i} c i nondecreasing. Here π 1 : = log ( Λ λ c Θ ) \pi_{1}:=\log(\Lambda\lambda_{c}\Theta) π 1 := log ( Λ λ c Θ ) and π i : = i π i − 1 + log G 1 + 1 \pi_{i}:=i\pi_{i-1}+\log G_{1}+1 π i := i π i − 1 + log G 1 + 1 for i ≥ 2 i\ge2 i ≥ 2 . For i = 1 i=1 i = 1 this is (35) , since λ c ≤ Λ λ c \lambda_{c}\le\Lambda\lambda_{c} λ c ≤ Λ λ c . Let i ≥ 2 i\ge2 i ≥ 2 and assume (36) for 1, ..., i − 1 i-1 i − 1 . By (33) and the Faà di Bruno formula,
D i h ( k + 1 ) = D g k ( h ( k ) ) D i h ( k ) + ∑ l = 2 i ∑ b 1 + ⋯ + b l = i b 1 , … , b l ≥ 1 c i , b D l g k ( h ( k ) ) [ D b 1 h ( k ) , … , D b l h ( k ) ] . D^i\mathfrak{h}^{(k+1)}=Dg_k(\mathfrak{h}^{(k)})D^i\mathfrak{h}^{(k)}+\sum_{l=2}^{i}\sum_{\substack{b_1+\cdots+b_l=i\\b_1,\ldots,b_l\geq1}}c_{i,b}D^lg_k(\mathfrak{h}^{(k)})[D^{b_1}\mathfrak{h}^{(k)},\ldots,D^{b_l}\mathfrak{h}^{(k)}]. D i h ( k + 1 ) = D g k ( h ( k ) ) D i h ( k ) + l = 2 ∑ i b 1 + ⋯ + b l = i b 1 , … , b l ≥ 1 ∑ c i , b D l g k ( h ( k ) ) [ D b 1 h ( k ) , … , D b l h ( k ) ] .
with combinatorial coefficients c i , b c_{i,b} c i , b . In the second sum l ≥ 2 l\geq2 l ≥ 2 , so each b m ≤ i − 1 b_m\leq i-1 b m ≤ i − 1 , and by the induction hypothesis and π 1 ≤ ⋯ ≤ π i − 1 \pi_1\leq\cdots\leq\pi_{i-1} π 1 ≤ ⋯ ≤ π i − 1 each product is bounded by c i − 1 e π i − 1 k ≤ c i e π i − 1 k c_{i-1}e^{\pi_{i-1}k}\leq c_i e^{\pi_{i-1}k} c i − 1 e π i − 1 k ≤ c i e π i − 1 k . Using (34) ,
a k + 1 ( i ) ≤ G 1 a k ( i ) + K i e i π i − 1 k , a 0 ( i ) = 0 , a_{k+1}^{(i)}\leq G_1a_k^{(i)}+K_i e^{i\pi_{i-1}k},\qquad a_0^{(i)}=0, a k + 1 ( i ) ≤ G 1 a k ( i ) + K i e i π i − 1 k , a 0 ( i ) = 0 ,
with K i K_i K i depending only on i i i , G 1 , … , G i G_1,\ldots,G_i G 1 , … , G i , c i − 1 c_{i-1} c i − 1 . Iterating,
a k ( i ) ≤ K i ∑ m = 0 k − 1 G 1 k − 1 − m e i π i − 1 m ≤ K i k e ( i π i − 1 + log G 1 ) k ≤ c i e π i k , a_k^{(i)}\leq K_i\sum_{m=0}^{k-1}G_1^{k-1-m}e^{i\pi_{i-1}m}\leq K_i k e^{(i\pi_{i-1}+\log G_1)k}\leq c_i e^{\pi_i k}, a k ( i ) ≤ K i m = 0 ∑ k − 1 G 1 k − 1 − m e i π i − 1 m ≤ K i k e ( i π i − 1 + l o g G 1 ) k ≤ c i e π i k ,
which proves (36) . Note that π i \pi_i π i is nondecreasing in i i i (as π i − 1 > 0 \pi_{i-1}>0 π i − 1 > 0 ) and ∣ h ∣ j = max i ≤ j a n ( i ) ≤ c j e π j n \lvert\mathfrak{h}\rvert_j=\max_{i\leq j}a_n^{(i)}\leq c_j e^{\pi_j n} ∣ h ∣ j = max i ≤ j a n ( i ) ≤ c j e π j n .
(c), bounds on τ n ∘ h \tau_n\circ\mathfrak{h} τ n ∘ h and log J h \log J_{\mathfrak{h}} log J h . Since F m ∘ h = h ( n − m ) F^m\circ\mathfrak{h}=\mathfrak{h}^{(n-m)} F m ∘ h = h ( n − m ) ,
τ n ∘ h = ∑ m = 0 n − 1 τ ∘ F m ∘ h = ∑ k = 1 n τ ∘ h ( k ) , log J h = − ∑ m = 0 n − 1 log det D F ∘ F m ∘ h = ∑ k = 1 n ψ ∘ h ( k ) , \tau_n\circ\mathfrak{h}=\sum_{m=0}^{n-1}\tau\circ F^m\circ\mathfrak{h}=\sum_{k=1}^{n}\tau\circ\mathfrak{h}^{(k)},\qquad\log J_{\mathfrak{h}}=-\sum_{m=0}^{n-1}\log\det DF\circ F^m\circ\mathfrak{h}=\sum_{k=1}^{n}\psi\circ\mathfrak{h}^{(k)}, τ n ∘ h = m = 0 ∑ n − 1 τ ∘ F m ∘ h = k = 1 ∑ n τ ∘ h ( k ) , log J h = − m = 0 ∑ n − 1 log det D F ∘ F m ∘ h = k = 1 ∑ n ψ ∘ h ( k ) ,
where ψ : = − log det D F ∈ C ∞ ( T 2 ) \psi:=-\log\det DF\in C^\infty(\mathbb{T}^2) ψ := − log det D F ∈ C ∞ ( T 2 ) by (A1). Let ϕ ∈ { τ , ψ } \phi\in\{\tau,\psi\} ϕ ∈ { τ , ψ } . Then sup ∣ ϕ ∘ h ( k ) ∣ ≤ ∥ ϕ ∥ C 0 \sup\lvert\phi\circ\mathfrak{h}^{(k)}\rvert\leq\|\phi\|_{C^0} sup ∣ ϕ ∘ h ( k ) ∣ ≤ ∥ ϕ ∥ C 0 and, by Faà di Bruno and (36) , for 1 ≤ i ≤ j 1\leq i\leq j 1 ≤ i ≤ j ,
sup B ∥ D i ( ϕ ∘ h ( k ) ) ∥ ≤ ∑ l = 1 i ∑ b 1 + ⋯ + b l = i c i , b ∥ D l ϕ ∥ C 0 ∏ m = 1 l a k ( b m ) ≤ K j ′ ′ ∥ ϕ ∥ C j e j π j k . \sup_B\|D^i(\phi\circ\mathfrak{h}^{(k)})\|\leq\sum_{l=1}^{i}\sum_{b_1+\cdots+b_l=i}c_{i,b}\|D^l\phi\|_{C^0}\prod_{m=1}^{l}a_k^{(b_m)}\leq K_j''\|\phi\|_{C^j}e^{j\pi_j k}. B sup ∥ D i ( ϕ ∘ h ( k ) ) ∥ ≤ l = 1 ∑ i b 1 + ⋯ + b l = i ∑ c i , b ∥ D l ϕ ∥ C 0 m = 1 ∏ l a k ( b m ) ≤ K j ′′ ∥ ϕ ∥ C j e j π j k .
Summing over 1 ≤ k ≤ n 1\leq k\leq n 1 ≤ k ≤ n gives ∥ τ n ∘ h ∥ C j + ∥ log J h ∥ C j ≤ K j ′ ′ n e j π j n \|\tau_n\circ\mathfrak{h}\|_{C^j}+\|\log J_{\mathfrak{h}}\|_{C^j}\leq K_j''n e^{j\pi_j n} ∥ τ n ∘ h ∥ C j + ∥ log J h ∥ C j ≤ K j ′′ n e j π j n . Setting
p j : = j π j + 1 , p_j:=j\pi_j+1, p j := j π j + 1 ,
which is nondecreasing in j j j and satisfies p 1 = π 1 + 1 ≥ log ( Λ λ c Θ ) p_1=\pi_1+1\geq\log(\Lambda\lambda_c\Theta) p 1 = π 1 + 1 ≥ log ( Λ λ c Θ ) , and choosing C j C_j C j large, we obtain the first estimate in (c). For the second, D j J h = D j ( e log J h ) D^jJ_{\mathfrak{h}}=D^j(e^{\log J_{\mathfrak{h}}}) D j J h = D j ( e l o g J h ) is J h J_{\mathfrak{h}} J h times a polynomial (a complete Bell polynomial) in D log J h , … , D j log J h D\log J_{\mathfrak{h}},\ldots,D^j\log J_{\mathfrak{h}} D log J h , … , D j log J h , which is homogeneous of degree j j j when D i D^i D i is given weight i i i ; hence ∥ D j J h ∥ ≤ C j J h e j p j n \|D^jJ_{\mathfrak{h}}\|\leq C_jJ_{\mathfrak{h}}e^{jp_jn} ∥ D j J h ∥ ≤ C j J h e j p j n .
(d). By (B.2), ∣ det D h ( q ) F n ∣ ≤ C ∗ Λ n λ c n \lvert\det D\mathfrak{h}_{(q)}F^n\rvert\leq C_*\Lambda^n\lambda_c^n ∣ det D h ( q ) F n ∣ ≤ C ∗ Λ n λ c n , i.e. 1 ≤ C ∗ ( Λ λ c ) n J h ( q ) 1\leq C_*(\Lambda\lambda_c)^nJ_{\mathfrak{h}}(q) 1 ≤ C ∗ ( Λ λ c ) n J h ( q ) . Summing over h ∈ H n ( q ) \mathfrak{h}\in\mathcal{H}^n(q) h ∈ H n ( q ) and using that h ↦ h ( q ) \mathfrak{h}\mapsto\mathfrak{h}(q) h ↦ h ( q ) is a bijection onto F − n ( q ) F^{-n}(q) F − n ( q ) ,
# H n ( q ) ≤ C ∗ ( Λ λ c ) n ∑ y ∈ F − n ( q ) ∣ det D y F n ∣ − 1 ≤ C ∞ C ∗ ( Λ λ c Θ ) n \#\mathcal{H}^n(q)\leq C_*(\Lambda\lambda_c)^n\sum_{y\in F^{-n}(q)}\lvert\det D_yF^n\rvert^{-1}\leq C_\infty C_*(\Lambda\lambda_c\Theta)^n # H n ( q ) ≤ C ∗ ( Λ λ c ) n y ∈ F − n ( q ) ∑ ∣ det D y F n ∣ − 1 ≤ C ∞ C ∗ ( Λ λ c Θ ) n
by (8) . □
Lemma 4.6 (Non-stationary phase). Let m ∈ N m\in\mathbb{N} m ∈ N . There exists C m > 0 C_m>0 C m > 0 such that the following holds. Let U ⊂ R 2 U\subset\mathbb{R}^2 U ⊂ R 2 be open, r ∈ ( 0 , 1 ] r\in(0,1] r ∈ ( 0 , 1 ] , a ∈ C c m ( U ) a\in C_c^m(U) a ∈ C c m ( U ) with ∥ D j a ∥ C 0 ≤ A r − j \|D^ja\|_{C^0}\leq Ar^{-j} ∥ D j a ∥ C 0 ≤ A r − j for 0 ≤ j ≤ m 0\leq j\leq m 0 ≤ j ≤ m , and Φ ∈ C m + 1 ( U , R ) \Phi\in C^{m+1}(U,\mathbb{R}) Φ ∈ C m + 1 ( U , R ) with ∥ D j Φ ∥ C 0 ≤ Q \|D^j\Phi\|_{C^0}\leq Q ∥ D j Φ ∥ C 0 ≤ Q for 2 ≤ j ≤ m + 1 2\leq j\leq m+1 2 ≤ j ≤ m + 1 , where Q ≥ 1 Q\geq1 Q ≥ 1 , and ∣ ∇ Φ ∣ ≥ μ ≥ 1 \lvert\nabla\Phi\rvert\geq\mu\geq1 ∣ ∇Φ ∣ ≥ μ ≥ 1 on supp a \operatorname{supp}a supp a . Then
∣ ∫ U e i Φ a ∣ ≤ C m A Leb ( supp a ) ( 1 r μ + Q μ 2 ) m . \left|\int_U e^{i\Phi}a\right|\leq C_mA\operatorname{Leb}(\operatorname{supp}a)\left(\frac{1}{r\mu}+\frac{Q}{\mu^2}\right)^m. ∫ U e i Φ a ≤ C m A Leb ( supp a ) ( r μ 1 + μ 2 Q ) m .
Proof. Set V = ∇ Φ / ( i ∣ ∇ Φ ∣ 2 ) V=\nabla\Phi/(i\lvert\nabla\Phi\rvert^2) V = ∇Φ/ ( i ∣ ∇Φ ∣ 2 ) on a neighbourhood of supp a \operatorname{supp}a supp a , so that V ⋅ ∇ e i Φ = e i Φ V\cdot\nabla e^{i\Phi}=e^{i\Phi} V ⋅ ∇ e i Φ = e i Φ . A cutoff allows an arbitrary smooth extension away from this neighbourhood. Write L T b = − div ( b V ) L^T b=-\operatorname{div}(bV) L T b = − div ( bV ) . Repeated integration by parts gives
∫ U e i Φ a = ∫ U e i Φ ( L T ) m a . \int_U e^{i\Phi}a=\int_U e^{i\Phi}(L^T)^m a. ∫ U e i Φ a = ∫ U e i Φ ( L T ) m a .
On supp a \operatorname{supp}a supp a , the chain rule gives
∣ V ∣ ≤ μ − 1 , ∣ D j V ∣ ≤ C j μ − 1 ( 1 + Q μ ) j ( 1 ≤ j ≤ m ) . \lvert V\rvert\le\mu^{-1},\qquad\lvert D^jV\rvert\le C_j\mu^{-1}\left(1+\frac{Q}{\mu}\right)^j\quad(1\le j\le m). ∣ V ∣ ≤ μ − 1 , ∣ D j V ∣ ≤ C j μ − 1 ( 1 + μ Q ) j ( 1 ≤ j ≤ m ) .
Indeed, derivatives of v ↦ v / ∣ v ∣ 2 v\mapsto v/\lvert v\rvert^2 v ↦ v / ∣ v ∣ 2 of order l l l are bounded by C l ∣ v ∣ − 1 − l C_l\lvert v\rvert^{-1-l} C l ∣ v ∣ − 1 − l ; the remaining factors are derivatives of Φ \Phi Φ of order at least two. Expanding ( L T ) m a (L^T)^m a ( L T ) m a produces products with m m m factors consisting of derivatives of V V V , and total derivative order m m m , distributed between these factors and a a a . Hence
∣ ( L T ) m a ∣ ≤ C m A μ − m ∑ j = 0 m r − j ( 1 + Q μ ) m − j ≤ C m A ( 1 r μ + Q μ 2 ) m . \lvert(L^T)^m a\rvert\le C_mA\mu^{-m}\sum_{j=0}^{m}r^{-j}\left(1+\frac{Q}{\mu}\right)^{m-j}\le C_mA\left(\frac{1}{r\mu}+\frac{Q}{\mu^2}\right)^m. ∣( L T ) m a ∣ ≤ C m A μ − m j = 0 ∑ m r − j ( 1 + μ Q ) m − j ≤ C m A ( r μ 1 + μ 2 Q ) m .
where the last step uses r ≤ 1 r\le1 r ≤ 1 . Integrating over supp a \operatorname{supp}a supp a proves the claim. □ \square □
Localisation in space and frequency
Fix ∣ ν ∣ ≥ ν 0 \lvert\nu\rvert\ge\nu_0 ∣ ν ∣ ≥ ν 0 , set N : = n ( ν ) N:=n(\nu) N := n ( ν ) , and let h ∈ H s h\in H^s h ∈ H s satisfy ∥ h ∥ H s ≤ 2 C ( 1 + ∣ ν ∣ ) s ∥ h ∥ L 2 \lVert h\rVert_{H^s}\le2C(1+\lvert\nu\rvert)^s\lVert h\rVert_{L^2} ∥ h ∥ H s ≤ 2 C ( 1 + ∣ ν ∣ ) s ∥ h ∥ L 2 . We fix the scales
r : = ∣ ν ∣ − 1 / 2 , M : = ∣ ν ∣ 5 / 8 , R : = ∣ ν ∣ ε , r:=\lvert\nu\rvert^{-1/2},\qquad M:=\lvert\nu\rvert^{5/8},\qquad R:=\lvert\nu\rvert^\varepsilon, r := ∣ ν ∣ − 1/2 , M := ∣ ν ∣ 5/8 , R := ∣ ν ∣ ε ,
with ε ∈ ( 0 , 1 16 ) \varepsilon\in(0,\frac{1}{16}) ε ∈ ( 0 , 16 1 ) to be chosen.
Frequency cut-off. Write h = h ≤ + h > h=h_{\le}+h_{>} h = h ≤ + h > , where h ≤ : = ∑ k ∈ Z 2 , 2 π ∣ k ∣ ≤ R ∣ ν ∣ h ^ ( k ) e 2 π i k ⋅ x h_{\le}:=\sum_{k\in\mathbb{Z}^2,\ 2\pi\lvert k\rvert\le R\lvert\nu\rvert}\widehat{h}(k)e^{2\pi i k\cdot x} h ≤ := ∑ k ∈ Z 2 , 2 π ∣ k ∣ ≤ R ∣ ν ∣ h ( k ) e 2 π ik ⋅ x . Then ∥ h ≤ ∥ L 2 ≤ ∥ h ∥ L 2 \lVert h_{\le}\rVert_{L^2}\le\lVert h\rVert_{L^2} ∥ h ≤ ∥ L 2 ≤ ∥ h ∥ L 2 and
∥ h > ∥ L 2 ≤ C ( R ∣ ν ∣ ) − s ∥ h ∥ H s ≤ C R − s ∥ h ∥ L 2 , \lVert h_{>}\rVert_{L^2}\le C(R\lvert\nu\rvert)^{-s}\lVert h\rVert_{H^s}\le CR^{-s}\lVert h\rVert_{L^2}, ∥ h > ∥ L 2 ≤ C ( R ∣ ν ∣ ) − s ∥ h ∥ H s ≤ C R − s ∥ h ∥ L 2 ,
so that, by (4.3),
∥ L ν N h > ∥ L 2 ≤ C Θ N / 2 ∣ ν ∣ − ε s ∥ h ∥ L 2 . (37) \lVert\mathcal{L}^N_\nu h_{>}\rVert_{L^2}\le C\Theta^{N/2}\lvert\nu\rvert^{-\varepsilon s}\lVert h\rVert_{L^2}. \tag*{(37)} ∥ L ν N h > ∥ L 2 ≤ C Θ N /2 ∣ ν ∣ − ε s ∥ h ∥ L 2 . (37)
Space localisation. Let { ψ α } α ∈ A \{\psi_\alpha\}_{\alpha\in\mathcal{A}} { ψ α } α ∈ A be smooth functions with supp ψ α ⊂ U α : = B ( z α , r ) \operatorname{supp}\psi_\alpha\subset U_\alpha:=B(z_\alpha,r) supp ψ α ⊂ U α := B ( z α , r ) , ∑ α ψ α 2 = 1 \sum_\alpha\psi_\alpha^2=1 ∑ α ψ α 2 = 1 , ∥ D j ψ α ∥ C 0 ≤ C j r − j \lVert D^j\psi_\alpha\rVert_{C^0}\le C_jr^{-j} ∥ D j ψ α ∥ C 0 ≤ C j r − j , with intersection multiplicity of { U α } \{U_\alpha\} { U α } bounded by c ♯ c_\sharp c ♯ and # A ≤ C r − 2 = C ∣ ν ∣ \#\mathcal{A}\le Cr^{-2}=C\lvert\nu\rvert # A ≤ C r − 2 = C ∣ ν ∣ . Then
∥ L ν N h ≤ ∥ L 2 2 = ∑ α ∥ ψ α L ν N h ≤ ∥ L 2 2 , ψ α L ν N h ≤ = ∑ h ∈ H N ( z α ) g α , h , \lVert\mathcal{L}^N_\nu h_{\le}\rVert_{L^2}^2=\sum_\alpha\lVert\psi_\alpha\mathcal{L}^N_\nu h_{\le}\rVert_{L^2}^2,\qquad
\psi_\alpha\mathcal{L}^N_\nu h_{\le}=\sum_{\mathfrak h\in\mathcal{H}^N(z_\alpha)}g_{\alpha,\mathfrak h}, ∥ L ν N h ≤ ∥ L 2 2 = α ∑ ∥ ψ α L ν N h ≤ ∥ L 2 2 , ψ α L ν N h ≤ = h ∈ H N ( z α ) ∑ g α , h ,
where, by (32) ,
g α , h ( y ) : = ψ α ( y ) J h ( y ) e i ν τ N ( h ( y ) ) h ≤ ( h ( y ) ) , y ∈ U α . g_{\alpha,\mathfrak h}(y):=\psi_\alpha(y)J_{\mathfrak h}(y)e^{i\nu\tau_N(\mathfrak h(y))}h_{\le}(\mathfrak h(y)),\qquad y\in U_\alpha. g α , h ( y ) := ψ α ( y ) J h ( y ) e i ν τ N ( h ( y )) h ≤ ( h ( y )) , y ∈ U α .
We regard g α , h g_{\alpha,\mathfrak h} g α , h as a function on R 2 \mathbb{R}^2 R 2 (via a chart of U α U_\alpha U α ) and denote by g ^ α , h \widehat{g}_{\alpha,\mathfrak h} g α , h its Fourier transform.
Trapped frequencies. For α ∈ A \alpha\in\mathcal{A} α ∈ A and h ∈ H N ( z α ) \mathfrak h\in\mathcal{H}^N(z_\alpha) h ∈ H N ( z α ) let
Σ α , h : = { ν c h ( y ) + ( D y h ) ∗ η : y ∈ U α , ∣ η ∣ ≤ R ∣ ν ∣ } , \Sigma_{\alpha,\mathfrak h}:=\left\{\nu c_{\mathfrak h}(y)+(D_y\mathfrak h)^*\eta:y\in U_\alpha,\ \lvert\eta\rvert\le R\lvert\nu\rvert\right\}, Σ α , h := { ν c h ( y ) + ( D y h ) ∗ η : y ∈ U α , ∣ η ∣ ≤ R ∣ ν ∣ } ,
S α , h : = { ξ ∈ R 2 : dist ( ξ , Σ α , h ) < M } . S_{\alpha,\mathfrak h}:=\left\{\xi\in\mathbb{R}^2:\operatorname{dist}(\xi,\Sigma_{\alpha,\mathfrak h})<M\right\}. S α , h := { ξ ∈ R 2 : dist ( ξ , Σ α , h ) < M } .
These are the frequencies that g α , h g_{\alpha,\mathfrak h} g α , h can carry: the phase of e i ν τ N ∘ h e 2 π i k ⋅ h e^{i\nu\tau_N\circ\mathfrak h}e^{2\pi i k\cdot\mathfrak h} e i ν τ N ∘ h e 2 π ik ⋅ h has gradient ν c h + ( D h ) ∗ ( 2 π k ) \nu c_{\mathfrak h}+(D\mathfrak h)^*(2\pi k) ν c h + ( D h ) ∗ ( 2 π k ) . We split g ^ α , h = a α , h + b α , h \widehat{g}_{\alpha,\mathfrak h}=a_{\alpha,\mathfrak h}+b_{\alpha,\mathfrak h} g α , h = a α , h + b α , h with a α , h : = 1 S α , h g ^ α , h a_{\alpha,\mathfrak h}:=\mathbb{1}_{S_{\alpha,\mathfrak h}}\widehat{g}_{\alpha,\mathfrak h} a α , h := 1 S α , h g α , h .
The trapped part: captivity.
Lemma 4.7 (Captivity). Assume (B1), and that ρ \rho ρ and ε \varepsilon ε satisfy
ε < κ ρ , ρ ( log λ + p 2 ) + ε < 3 8 . (38) \varepsilon< \kappa\rho,\qquad\rho(\log\lambda+p_2)+\varepsilon< \frac{3}{8}. \tag*{(38)} ε < κ ρ , ρ ( log λ + p 2 ) + ε < 8 3 . (38) Then, for ∣ ν ∣ |\nu| ∣ ν ∣ large enough, for all α ∈ A \alpha\in\mathcal{A} α ∈ A and ξ ∈ R 2 \xi\in\mathbb{R}^2 ξ ∈ R 2 ,
∑ h ∈ H N ( z α ) : ξ ∈ S α , h J h ( z α ) ≤ C e − γ 1 N . \sum_{\substack{h\in\mathcal{H}^N(z_\alpha):\,\xi\in S_{\alpha,h}}} J_h(z_\alpha)\leq Ce^{-\gamma_1N}. h ∈ H N ( z α ) : ξ ∈ S α , h ∑ J h ( z α ) ≤ C e − γ 1 N . Proof. Let ξ ∈ S α , h \xi\in S_{\alpha,h} ξ ∈ S α , h . Then there are y ∈ U α y\in U_\alpha y ∈ U α and η \eta η with ∣ η ∣ ≤ R ∣ ν ∣ |\eta|\leq R|\nu| ∣ η ∣ ≤ R ∣ ν ∣ such that ∣ ξ − ν c h ( y ) − ( D y h ) ∗ η ∣ < M |\xi-\nu c_h(y)-(D_yh)^*\eta|<M ∣ ξ − ν c h ( y ) − ( D y h ) ∗ η ∣ < M . Set ζ : = ξ / ν \zeta:=\xi/\nu ζ := ξ / ν , ζ 0 : = η / ν \zeta_0:=\eta/\nu ζ 0 := η / ν (so ∣ ζ 0 ∣ ≤ R |\zeta_0|\leq R ∣ ζ 0 ∣ ≤ R ), and let u ∈ D h ( z α ) F N C u u\in D_h(z_\alpha)F^N C_u u ∈ D h ( z α ) F N C u be a unit vector. By Lemma 4.5 ,
∣ ζ ( u ) − c h ( z α ) ( u ) ∣ ≤ M ∣ ν ∣ + ∣ c h ( y ) − c h ( z α ) ∣ + ∣ ζ 0 ( D y h u ) ∣ |\zeta(u)-c_h(z_\alpha)(u)|\leq\frac{M}{|\nu|}+|c_h(y)-c_h(z_\alpha)|+|\zeta_0(D_yh\,u)| ∣ ζ ( u ) − c h ( z α ) ( u ) ∣ ≤ ∣ ν ∣ M + ∣ c h ( y ) − c h ( z α ) ∣ + ∣ ζ 0 ( D y h u ) ∣ ≤ ∣ ν ∣ − 3 / 8 + C 2 e p 2 N r + R ( ∥ D z α h u ∥ + C 2 e p 2 N r ) \leq|\nu|^{-3/8}+C_2e^{p_2N}r+R\left(\|D_{z_\alpha}h\,u\|+C_2e^{p_2N}r\right) ≤ ∣ ν ∣ − 3/8 + C 2 e p 2 N r + R ( ∥ D z α h u ∥ + C 2 e p 2 N r ) ≤ ∣ ν ∣ − 3 / 8 + 2 C 2 ∣ ν ∣ ε + ρ p 2 − 1 2 + C ⋆ R λ − N . \leq|\nu|^{-3/8}+2C_2|\nu|^{\varepsilon+\rho p_2-\frac12}+C_\star R\lambda^{-N}. ≤ ∣ ν ∣ − 3/8 + 2 C 2 ∣ ν ∣ ε + ρ p 2 − 2 1 + C ⋆ R λ − N . Since N ≥ ρ log ∣ ν ∣ N\geq\rho\log|\nu| N ≥ ρ log ∣ ν ∣ and ε < κ ρ \varepsilon<\kappa\rho ε < κ ρ , we have C ⋆ R λ − N ≤ 1 2 e κ N λ − N C_\star R\lambda^{-N}\leq\frac12e^{\kappa N}\lambda^{-N} C ⋆ R λ − N ≤ 2 1 e κ N λ − N for ∣ ν ∣ |\nu| ∣ ν ∣ large. Since N ≤ ρ log ∣ ν ∣ + 1 N\leq\rho\log|\nu|+1 N ≤ ρ log ∣ ν ∣ + 1 , e κ N λ − N ≥ λ − 1 ∣ ν ∣ − ρ log λ e^{\kappa N}\lambda^{-N}\geq\lambda^{-1}|\nu|^{-\rho\log\lambda} e κ N λ − N ≥ λ − 1 ∣ ν ∣ − ρ l o g λ , and by (38) the first two terms are bounded by 1 2 e κ N λ − N \frac12e^{\kappa N}\lambda^{-N} 2 1 e κ N λ − N for ∣ ν ∣ |\nu| ∣ ν ∣ large. Hence every h h h in the sum belongs to the set in (11) (with n = N n=N n = N , q = z α q=z_\alpha q = z α and this ζ \zeta ζ ), and the claim follows from (B1). □
Lemma 4.8. Under the assumptions of Lemma 4.7 ,
∑ α ∈ A ∥ ∑ h a α , h ∥ L 2 2 ≤ C e − γ 1 N ∥ h ∥ L 2 2 . \sum_{\alpha\in\mathcal{A}}\left\|\sum_h a_{\alpha,h}\right\|_{L^2}^2\leq Ce^{-\gamma_1N}\|h\|_{L^2}^2. α ∈ A ∑ h ∑ a α , h L 2 2 ≤ C e − γ 1 N ∥ h ∥ L 2 2 . Proof. By Cauchy–Schwarz, for each ξ \xi ξ ,
∣ ∑ h a α , h ( ξ ) ∣ 2 ≤ ( ∑ h ξ ∈ S α , h J h ( z α ) ) ( ∑ h ∣ g ^ α , h ( ξ ) ∣ 2 J h ( z α ) ) . \left|\sum_h a_{\alpha,h}(\xi)\right|^2\leq\left(\sum_{\substack{h\\ \xi\in S_{\alpha,h}}}J_h(z_\alpha)\right)\left(\sum_h\frac{|\widehat{g}_{\alpha,h}(\xi)|^2}{J_h(z_\alpha)}\right). h ∑ a α , h ( ξ ) 2 ≤ h ξ ∈ S α , h ∑ J h ( z α ) ( h ∑ J h ( z α ) ∣ g α , h ( ξ ) ∣ 2 ) . Integrating in ξ \xi ξ , using Lemma 4.7 and Plancherel,
∥ ∑ h a α , h ∥ L 2 2 ≤ C e − γ 1 N ∑ h ∥ g α , h ∥ L 2 2 J h ( z α ) . \left\|\sum_h a_{\alpha,h}\right\|_{L^2}^2\leq Ce^{-\gamma_1N}\sum_h\frac{\|g_{\alpha,h}\|_{L^2}^2}{J_h(z_\alpha)}. h ∑ a α , h L 2 2 ≤ C e − γ 1 N h ∑ J h ( z α ) ∥ g α , h ∥ L 2 2 . By Lemma 4.5 ,
∣ log J h ( y ) − log J h ( z α ) ∣ ≤ C 1 e ρ p 1 N r ≤ log 2 |\log J_h(y)-\log J_h(z_\alpha)|\leq C_1e^{\rho p_1N}r\leq\log2 ∣ log J h ( y ) − log J h ( z α ) ∣ ≤ C 1 e ρ p 1 N r ≤ log 2 for y ∈ U α y\in U_\alpha y ∈ U α and ∣ ν ∣ |\nu| ∣ ν ∣ large (as ρ p 1 < 1 2 \rho p_1<\frac12 ρ p 1 < 2 1 by (38) ). Hence
∑ h ∥ g α , h ∥ L 2 2 J h ( z α ) ≤ 2 ∫ ψ α 2 ∑ h ∣ h ≤ ∘ h ∣ 2 = 2 ∫ ψ α 2 L F N ( ∣ h ≤ 1 ∣ 2 ) = 2 ∫ ( ψ α 2 ∘ F N ) ∣ h ≤ 1 ∣ 2 . \sum_h\frac{\|g_{\alpha,h}\|_{L^2}^2}{J_h(z_\alpha)}
\leq2\int\psi_\alpha^2\sum_h|h_{\leq\circ h}|^2
=2\int\psi_\alpha^2\mathcal{L}^N_F(|h_{\leq1}|^2)
=2\int(\psi_\alpha^2\circ F^N)|h_{\leq1}|^2. h ∑ J h ( z α ) ∥ g α , h ∥ L 2 2 ≤ 2 ∫ ψ α 2 h ∑ ∣ h ≤ ∘ h ∣ 2 = 2 ∫ ψ α 2 L F N ( ∣ h ≤ 1 ∣ 2 ) = 2 ∫ ( ψ α 2 ∘ F N ) ∣ h ≤ 1 ∣ 2 . Summing over α \alpha α and using ∑ α ψ α 2 = 1 \sum_\alpha\psi_\alpha^2=1 ∑ α ψ α 2 = 1 gives the claim. □
The untrapped part: non-stationary phase.
Lemma 4.9. Let m : = 64 m := 64 m := 64 and assume, in addition to (38) , that
ε + ρ p m + 1 ≤ 1 8 . (39) \varepsilon+\rho p_{m+1}\le\frac{1}{8}. \tag*{(39)} ε + ρ p m + 1 ≤ 8 1 . (39) Then, for ∣ ν ∣ |\nu| ∣ ν ∣ large enough,
∑ α ∈ A ∥ ∑ h b α , h ∥ L 2 2 ≤ C ∣ ν ∣ − 4 ∥ h ∥ L 2 2 . \sum_{\alpha\in\mathcal{A}}\left\|\sum_h b_{\alpha,h}\right\|_{L^2}^2\le C|\nu|^{-4}\|h\|_{L^2}^2. α ∈ A ∑ h ∑ b α , h L 2 2 ≤ C ∣ ν ∣ − 4 ∥ h ∥ L 2 2 .
Proof. Write h ≤ = ∑ k h ^ ( k ) e 2 π i k ⋅ x h_{\le}=\sum_k\widehat{h}(k)e^{2\pi ik\cdot x} h ≤ = ∑ k h ( k ) e 2 π ik ⋅ x , the sum running over k ∈ Z 2 k\in\mathbb{Z}^2 k ∈ Z 2 with 2 π ∣ k ∣ ≤ R ∣ ν ∣ 2\pi|k|\le R|\nu| 2 π ∣ k ∣ ≤ R ∣ ν ∣ ; there are at most C ( R ∣ ν ∣ ) 2 C(R|\nu|)^2 C ( R ∣ ν ∣ ) 2 such k k k . Lifting h h h to R 2 \mathbb{R}^2 R 2 ,
g ^ α , h ( ξ ) = ∑ k h ^ ( k ) I k ( ξ ) , I k ( ξ ) : = ∫ e i Φ k ( y ) ψ α ( y ) J h ( y ) d y , \widehat{g}_{\alpha,h}(\xi)=\sum_k\widehat{h}(k)I_k(\xi),\qquad I_k(\xi):=\int e^{i\Phi_k(y)}\psi_\alpha(y)J_h(y)\,dy, g α , h ( ξ ) = k ∑ h ( k ) I k ( ξ ) , I k ( ξ ) := ∫ e i Φ k ( y ) ψ α ( y ) J h ( y ) d y ,
Φ k : = ν τ N ∘ h + 2 π k ⋅ h − ξ ⋅ y . \Phi_k:=\nu\tau_N\circ h+2\pi k\cdot h-\xi\cdot y. Φ k := ν τ N ∘ h + 2 π k ⋅ h − ξ ⋅ y .
We have ∇ Φ k ( y ) = ν c h ( y ) + ( D y h ) ∗ ( 2 π k ) − ξ \nabla\Phi_k(y)=\nu c_h(y)+(D_yh)^*(2\pi k)-\xi ∇ Φ k ( y ) = ν c h ( y ) + ( D y h ) ∗ ( 2 π k ) − ξ , hence ∣ ∇ Φ k ∣ ≥ d ( ξ ) : = dist ( ξ , Σ α , h ) |\nabla\Phi_k|\ge d(\xi):=\operatorname{dist}(\xi,\Sigma_{\alpha,h}) ∣∇ Φ k ∣ ≥ d ( ξ ) := dist ( ξ , Σ α , h ) on U α U_{\alpha} U α , and d ( ξ ) ≥ M d(\xi)\ge M d ( ξ ) ≥ M for ξ ∉ S α , h \xi\notin S_{\alpha,h} ξ ∈ / S α , h . By Lemma 4.5, the amplitude ψ α J h \psi_\alpha J_h ψ α J h satisfies the assumptions of Lemma 4.6 with A = C m J h ( z α ) ≤ C m C ∗ A=C_mJ_h(z_\alpha)\le C_mC_* A = C m J h ( z α ) ≤ C m C ∗ : bounded distortion and (39) imply, for 0 ≤ j ≤ m 0\le j\le m 0 ≤ j ≤ m ,
∥ D j ( ψ α J h ) ∥ ∞ ≤ C m J h ( z α ) r − j , \left\|D^j(\psi_\alpha J_h)\right\|_\infty\le C_mJ_h(z_\alpha)r^{-j}, D j ( ψ α J h ) ∞ ≤ C m J h ( z α ) r − j ,
since e ρ p j N ≤ C ∣ ν ∣ 1 / 8 ≤ C r − 1 e^{\rho p_jN}\le C|\nu|^{1/8}\le Cr^{-1} e ρ p j N ≤ C ∣ ν ∣ 1/8 ≤ C r − 1 . The bound on A A A follows from (B.2) and λ c < λ \lambda_c<\lambda λ c < λ . Moreover,
∥ D j Φ k ∥ ∞ ≤ Q : = C ( 1 + R ) ∣ ν ∣ e ρ p m + 1 N \left\|D^j\Phi_k\right\|_\infty\le Q:=C(1+R)|\nu|e^{\rho p_{m+1}N} D j Φ k ∞ ≤ Q := C ( 1 + R ) ∣ ν ∣ e ρ p m + 1 N
for 2 ≤ j ≤ m + 1 2\le j\le m+1 2 ≤ j ≤ m + 1 . For t ≥ M t\ge M t ≥ M ,
1 r t + Q t 2 ≤ M t ( 1 r M + Q M 2 ) ≤ M t ( ∣ ν ∣ − 1 / 8 + C ∣ ν ∣ ε + ρ p m + 1 − 1 / 4 ) ≤ C M t ∣ ν ∣ − 1 / 8 , \frac{1}{rt}+\frac{Q}{t^2}\le\frac{M}{t}\left(\frac{1}{rM}+\frac{Q}{M^2}\right)\le\frac{M}{t}\left(|\nu|^{-1/8}+C|\nu|^{\varepsilon+\rho p_{m+1}-1/4}\right)\le\frac{CM}{t}|\nu|^{-1/8}, r t 1 + t 2 Q ≤ t M ( r M 1 + M 2 Q ) ≤ t M ( ∣ ν ∣ − 1/8 + C ∣ ν ∣ ε + ρ p m + 1 − 1/4 ) ≤ t C M ∣ ν ∣ − 1/8 ,
by (39) . Lemma 4.6 gives
∣ I k ( ξ ) ∣ ≤ C r 2 ( M ∣ ν ∣ − 1 / 8 / d ( ξ ) ) m |I_k(\xi)|\le Cr^2\left(M|\nu|^{-1/8}/d(\xi)\right)^m ∣ I k ( ξ ) ∣ ≤ C r 2 ( M ∣ ν ∣ − 1/8 / d ( ξ ) ) m
for ξ ∉ S α , h \xi\notin S_{\alpha,h} ξ ∈ / S α , h . The set Σ α , h \Sigma_{\alpha,h} Σ α , h has diameter
D Σ ≤ C R ∣ ν ∣ e ρ p 1 N ≤ C ∣ ν ∣ 2 , D_\Sigma\le CR|\nu|e^{\rho p_1N}\le C|\nu|^2, D Σ ≤ C R ∣ ν ∣ e ρ p 1 N ≤ C ∣ ν ∣ 2 ,
so
Leb ( { d < t } ) ≤ C ( D Σ + t ) 2 . \operatorname{Leb}(\{d<t\})\le C(D_\Sigma+t)^2. Leb ({ d < t }) ≤ C ( D Σ + t ) 2 .
Decomposing { d ≥ M } \{d\ge M\} { d ≥ M } into the shells { 2 j M ≤ d < 2 j + 1 M } \{2^jM\le d<2^{j+1}M\} { 2 j M ≤ d < 2 j + 1 M } , j ≥ 0 j\ge0 j ≥ 0 , on which ∣ I k ∣ 2 ≤ C r 4 ∣ ν ∣ − m / 4 2 − 2 m j |I_k|^2\le Cr^4|\nu|^{-m/4}2^{-2mj} ∣ I k ∣ 2 ≤ C r 4 ∣ ν ∣ − m /4 2 − 2 mj ,
∫ { d ≥ M } ∣ I k ∣ 2 d ξ ≤ C r 4 ∣ ν ∣ − m / 4 ∑ j ≥ 0 2 − 2 m j ( D Σ + 2 j + 1 M ) 2 ≤ C r 4 ∣ ν ∣ − m / 4 ( D Σ 2 + M 2 ) ≤ C ∣ ν ∣ − m / 4 + 4 . \begin{aligned}
\int_{\{d\ge M\}}|I_k|^2d\xi\le Cr^4|\nu|^{-m/4}\sum_{j\ge0}2^{-2mj}(D_\Sigma+2^{j+1}M)^2 \\
&\le Cr^4|\nu|^{-m/4}(D_\Sigma^2+M^2)\le C|\nu|^{-m/4+4}.
\end{aligned} ∫ { d ≥ M } ∣ I k ∣ 2 d ξ ≤ C r 4 ∣ ν ∣ − m /4 j ≥ 0 ∑ 2 − 2 mj ( D Σ + 2 j + 1 M ) 2 ≤ C r 4 ∣ ν ∣ − m /4 ( D Σ 2 + M 2 ) ≤ C ∣ ν ∣ − m /4 + 4 .
By Cauchy–Schwarz, ∑ k ∣ h ^ ( k ) ∣ ≤ C R ∣ ν ∣ ∥ h ∥ L 2 \sum_k|\widehat{h}(k)|\le CR|\nu|\|h\|_{L^2} ∑ k ∣ h ( k ) ∣ ≤ C R ∣ ν ∣∥ h ∥ L 2 , therefore
∥ b α , h ∥ L 2 ≤ ∑ k ∣ h ^ ( k ) ∣ ∥ I k 1 { d ≥ M } ∥ L 2 ≤ C R ∣ ν ∣ ∣ ν ∣ 2 − m / 8 ∥ h ∥ L 2 . \|b_{\alpha,h}\|_{L^2}\le\sum_k|\widehat{h}(k)|\|I_k\mathbf{1}_{\{d\ge M\}}\|_{L^2}\le CR|\nu||\nu|^{2-m/8}\|h\|_{L^2}. ∥ b α , h ∥ L 2 ≤ k ∑ ∣ h ( k ) ∣∥ I k 1 { d ≥ M } ∥ L 2 ≤ C R ∣ ν ∣∣ ν ∣ 2 − m /8 ∥ h ∥ L 2 .
Finally, by Lemma 4.5(d), since ρ p 1 ≤ 1 \rho p_1\le1 ρ p 1 ≤ 1 , we have
# H N ( z α ) ≤ C ( Λ λ c Θ ) N ≤ C ∣ ν ∣ ρ log ( Λ λ c Θ ) ≤ C ∣ ν ∣ \#\mathcal{H}^N(z_\alpha)\le C(\Lambda\lambda_c\Theta)^N\le C|\nu|^{\rho\log(\Lambda\lambda_c\Theta)}\le C|\nu| # H N ( z α ) ≤ C ( Λ λ c Θ ) N ≤ C ∣ ν ∣ ρ l o g ( Λ λ c Θ ) ≤ C ∣ ν ∣
∑ α ∥ ∑ h b α , h ∥ L 2 2 ≤ # A ⋅ ( # H N ) 2 max α , h ∥ b α , h ∥ L 2 2 ≤ C ∣ ν ∣ 3 ∣ ν ∣ 2 + 2 ε + 4 − m / 4 ∥ h ∥ L 2 2 ≤ C ∣ ν ∣ − 4 ∥ h ∥ L 2 2 . \sum_{\alpha}\left\|\sum_h b_{\alpha,h}\right\|_{L^2}^2 \le\#\mathcal{A}\cdot\left(\#\mathcal{H}^N\right)^2\max_{\alpha,h}\|b_{\alpha,h}\|_{L^2}^2 \le C|\nu|^{3}| \nu|^{2+2\varepsilon+4-m/4}\|h\|_{L^2}^2 \le C|\nu|^{-4}\|h\|_{L^2}^2. α ∑ h ∑ b α , h L 2 2 ≤ # A ⋅ ( # H N ) 2 α , h max ∥ b α , h ∥ L 2 2 ≤ C ∣ ν ∣ 3 ∣ ν ∣ 2 + 2 ε + 4 − m /4 ∥ h ∥ L 2 2 ≤ C ∣ ν ∣ − 4 ∥ h ∥ L 2 2 .
Conclusion
Proposition 4.10. Assume (8) , (B1) and (P). Choose ε : = κ ρ / 2 \varepsilon:=\kappa\rho/2 ε := κ ρ /2 and ρ > 0 \rho>0 ρ > 0 so small that (38) and (39) hold and ρ log Θ < 1 4 \rho\log\Theta<\frac{1}{4} ρ log Θ < 4 1 . Then (30) holds with
γ 4 : = min { γ 1 2 , κ s − log Θ 2 , 1 ρ } . \gamma_4:=\min\left\{\frac{\gamma_1}{2},\frac{\kappa s-\log\Theta}{2},\frac{1}{\rho}\right\}. γ 4 := min { 2 γ 1 , 2 κ s − log Θ , ρ 1 } . and γ 4 > 1 2 log Θ \gamma_4>\frac{1}{2}\log\Theta γ 4 > 2 1 log Θ , so that Corollary 4.4 applies.
Proof. By Lemmas 4.8 and 4.9 ,
∥ L ν N h ∥ L 2 2 ≤ C ∑ α ( ∥ ∑ h a α , h ∥ L 2 2 + ∥ ∑ h b α , h ∥ L 2 2 ) ≤ C ( e − γ 1 N + ∣ ν ∣ − 4 ) ∥ h ∥ L 2 2 . \|\mathcal{L}_\nu^N h\|_{L^2}^2\le C\sum_\alpha\left(\left\|\sum_h a_{\alpha,h}\right\|_{L^2}^2+\left\|\sum_h b_{\alpha,h}\right\|_{L^2}^2\right)\le C\left(e^{-\gamma_1N}+|\nu|^{-4}\right)\|h\|_{L^2}^2. ∥ L ν N h ∥ L 2 2 ≤ C α ∑ h ∑ a α , h L 2 2 + h ∑ b α , h L 2 2 ≤ C ( e − γ 1 N + ∣ ν ∣ − 4 ) ∥ h ∥ L 2 2 .
Together with (37) , and since ∣ ν ∣ − 1 ≤ e − ( N − 1 ) / ρ |\nu|^{-1}\le e^{-(N-1)/\rho} ∣ ν ∣ − 1 ≤ e − ( N − 1 ) / ρ and
Θ N / 2 ∣ ν ∣ − ε s = Θ N / 2 ∣ ν ∣ − κ ρ s / 2 ≤ C e − ( κ s − log Θ ) N / 2 , \Theta^{N/2}|\nu|^{-\varepsilon s}=\Theta^{N/2}|\nu|^{-\kappa\rho s/2}\le Ce^{-(\kappa s-\log\Theta)N/2}, Θ N /2 ∣ ν ∣ − ε s = Θ N /2 ∣ ν ∣ − κ ρ s /2 ≤ C e − ( κ s − l o g Θ ) N /2 ,
∥ L ν N h ∥ L 2 2 ≤ C ( e − γ 1 N / 2 + ∣ ν ∣ − 2 + Θ N / 2 ∣ ν ∣ − ε s ) ∥ h ∥ L 2 2 ≤ C 1 e − γ 4 N ∥ h ∥ L 2 2 . \|\mathcal{L}_\nu^N h\|_{L^2}^2\le C\left(e^{-\gamma_1N/2}+|\nu|^{-2}+\Theta^{N/2}|\nu|^{-\varepsilon s}\right)\|h\|_{L^2}^2\le C_1e^{-\gamma_4N}\|h\|_{L^2}^2. ∥ L ν N h ∥ L 2 2 ≤ C ( e − γ 1 N /2 + ∣ ν ∣ − 2 + Θ N /2 ∣ ν ∣ − ε s ) ∥ h ∥ L 2 2 ≤ C 1 e − γ 4 N ∥ h ∥ L 2 2 .
Finally, γ 1 / 2 > 1 2 log Θ \gamma_1/2>\frac{1}{2}\log\Theta γ 1 /2 > 2 1 log Θ and ( κ s − log Θ ) / 2 > 1 2 log Θ (\kappa s-\log\Theta)/2>\frac{1}{2}\log\Theta ( κ s − log Θ ) /2 > 2 1 log Θ by (P), and 1 / ρ > 4 log Θ 1/\rho>4\log\Theta 1/ ρ > 4 log Θ , so γ 4 > 1 2 log Θ \gamma_4>\frac{1}{2}\log\Theta γ 4 > 2 1 log Θ . □
By Corollary 4.4 and Proposition 4.10 , Theorem 2.10 is proved.
Proof of Theorem A
We work with the complex Sobolev space H s H^s H s . Closed subspaces associated with the positivity sets of the invariant densities will provide the componentwise perturbation argument. Their construction does not require multiplication by characteristic functions to be bounded on H s H^s H s , or a full-measure union of the supports.
Small-frequency spectral decomposition
The following lemma implements the Nagaev–Guivarc’h spectral method in our possibly non-ergodic setting. We include a proof sketch, explaining in particular the decomposition associated with the ergodic acips.
Lemma 5.1 (Component branches and variances). Under the hypotheses of Theorem A, there are η > 0 \eta>0 η > 0 , C > 0 C>0 C > 0 , θ ∈ ( 0 , 1 ) \theta\in(0,1) θ ∈ ( 0 , 1 ) , analytic functions λ k \lambda_k λ k and analytic rank-one projections Π k \Pi_k Π k , defined for complex z z z near 0 0 0 , such that
L z n = ∑ k = 1 ℓ λ k ( z ) n Π k ( z ) + N ( z ) n , ∥ N ( z ) n ∥ H s → H s ≤ C θ n . (40) \mathcal{L}_z^n=\sum_{k=1}^{\ell}\lambda_k(z)^n\Pi_k(z)+N(z)^n,\qquad\|N(z)^n\|_{H^s\to H^s}\le C\theta^n. \tag*{(40)} L z n = k = 1 ∑ ℓ λ k ( z ) n Π k ( z ) + N ( z ) n , ∥ N ( z ) n ∥ H s → H s ≤ C θ n . (40) Here L z = L F ( e i z τ ⋅ ) \mathcal{L}_z=\mathcal{L}_F(e^{iz\tau}\cdot) L z = L F ( e i z τ ⋅ ) , λ k ( 0 ) = 1 \lambda_k(0)=1 λ k ( 0 ) = 1 , Π k ( 0 ) u = ρ k ℓ k ( u ) \Pi_k(0)u=\rho_k\ell_k(u) Π k ( 0 ) u = ρ k ℓ k ( u ) , and
λ k ′ ( 0 ) = 0 , λ k ′ ′ ( 0 ) = − σ k 2 , \lambda_k'(0)=0,\qquad\lambda_k''(0)=-\sigma_k^2, λ k ′ ( 0 ) = 0 , λ k ′′ ( 0 ) = − σ k 2 , where
σ k 2 = lim n → ∞ 1 n ∫ τ n 2 d μ k = ∫ τ 2 d μ k + 2 ∑ j = 1 ∞ ∫ τ ( τ ∘ F j ) d μ k > 0. (41) \sigma_k^2=\lim_{n\to\infty}\frac{1}{n}\int\tau_n^2\,d\mu_k=\int\tau^2\,d\mu_k+2\sum_{j=1}^{\infty}\int\tau(\tau\circ F^j)\,d\mu_k>0. \tag*{(41)} σ k 2 = n → ∞ lim n 1 ∫ τ n 2 d μ k = ∫ τ 2 d μ k + 2 j = 1 ∑ ∞ ∫ τ ( τ ∘ F j ) d μ k > 0. (41)
Proof. Multiplication by the smooth function τ \tau τ defines a bounded operator M τ M_\tau M τ on H s H^s H s . Thus L z = L F e i z M τ \mathcal{L}_z=\mathcal{L}_F e^{izM_\tau} L z = L F e i z M τ is entire in operator norm. By Lemma 3.10 (a), 1 1 1 is an isolated semisimple eigenvalue of multiplicity ℓ \ell ℓ , and all remaining spectrum lies strictly inside the unit disk. Let Γ \Gamma Γ be a positively oriented circle enclosing 1 1 1 and no other spectrum of L 0 \mathcal{L}_0 L 0 . For small complex z z z , the spectral projection
P ( z ) = 1 2 π i ∫ Γ ( w Id − L z ) − 1 d w P(z)=\frac{1}{2\pi i}\int_\Gamma(w\operatorname{Id}-\mathcal{L}_z)^{-1}\,dw P ( z ) = 2 π i 1 ∫ Γ ( w Id − L z ) − 1 d w
is analytic and has rank ℓ \ell ℓ . To obtain the individual analytic eigenvalue branches, we restrict L z \mathcal{L}_z L z to invariant subspaces on which 1 1 1 is a simple eigenvalue. Put E k = { ρ k > 0 } E_k=\{\rho_k>0\} E k = { ρ k > 0 } modulo null sets and
B k = { u ∈ H s : u = 0 a.e. on T 2 ∖ E k } . \mathcal{B}_k=\{u\in H^s:u=0\ \text{a.e. on }\mathbb{T}^2\setminus E_k\}. B k = { u ∈ H s : u = 0 a.e. on T 2 ∖ E k } .
This is a closed subspace of H s H^s H s . Positivity and L F ρ k = ρ k \mathcal{L}_F\rho_k=\rho_k L F ρ k = ρ k imply that L F \mathcal{L}_F L F , and hence every L z \mathcal{L}_z L z , preserves B k \mathcal{B}_k B k . On B k \mathcal{B}_k B k the eigenvalue 1 1 1 is simple: the full eigenspace at 1 1 1 is spanned by the densities with disjoint positivity sets, and only ρ k \rho_k ρ k belongs to B k \mathcal{B}_k B k . Moreover, Π ∣ B k u = ρ k ∫ u d m \Pi|_{\mathcal{B}_k}u=\rho_k\int u\,dm Π ∣ B k u = ρ k ∫ u d m and the complementary powers decay exponentially by the full-space estimate. Analytic perturbation on this closed subspace therefore gives a simple analytic branch λ k ( z ) \lambda_k(z) λ k ( z ) with eigenvector v k ( z ) ∈ B k v_k(z)\in\mathcal{B}_k v k ( z ) ∈ B k , normalized by ∫ v k ( z ) d m = 1 \int v_k(z)\,dm=1 ∫ v k ( z ) d m = 1 .
The v k ( z ) v_k(z) v k ( z ) have disjoint supports modulo null sets, so are linearly independent. They belong to ran P ( z ) \operatorname{ran}P(z) ran P ( z ) and form a basis there. Explicitly,
Π k ( z ) u = v k ( z ) ∫ E k P ( z ) u d m . \Pi_k(z)u=v_k(z)\int_{E_k}P(z)u\,dm. Π k ( z ) u = v k ( z ) ∫ E k P ( z ) u d m .
The formula is valid because ∫ E j v k ( z ) d m = δ j k \int_{E_j}v_k(z)\,dm=\delta_{jk} ∫ E j v k ( z ) d m = δ j k . Integration over E k E_k E k is continuous on H s ⊂ L 2 H^s\subset L^2 H s ⊂ L 2 . Thus these projections are analytic, they mutually vanish, commute with L z \mathcal{L}_z L z , and sum to P ( z ) P(z) P ( z ) . Setting N ( z ) = L z ( Id − P ( z ) ) N(z)=\mathcal{L}_z(\operatorname{Id}-P(z)) N ( z ) = L z ( Id − P ( z )) yields (40) , with uniform exponential bounds on a smaller complex neighbourhood by the resolvent formula.
For real z z z , the L 1 L^1 L 1 contraction of L z \mathcal{L}_z L z gives ∣ λ k ( z ) ∣ ≤ 1 |\lambda_k(z)|\leq1 ∣ λ k ( z ) ∣ ≤ 1 . Differentiating at zero and pairing with ℓ k \ell_k ℓ k gives
λ k ′ ( 0 ) = i ℓ k ( L F ( τ ρ k ) ) = i ∫ τ ρ k d m = 0. \lambda_k'(0)=i\ell_k\bigl(\mathcal{L}_F(\tau\rho_k)\bigr)=i\int\tau\rho_k\,dm=0. λ k ′ ( 0 ) = i ℓ k ( L F ( τ ρ k ) ) = i ∫ τ ρ k d m = 0.
Smooth multiplication preserves H s H^s H s , so the spectral gap gives exponential decay of the covariances in (41) . The identity
∫ τ n 2 d μ k = n ∫ τ 2 d μ k + 2 ∑ j = 1 n − 1 ( n − j ) ∫ τ ( τ ∘ F j ) d μ k \int\tau_n^2\,d\mu_k=n\int\tau^2\,d\mu_k+2\sum_{j=1}^{n-1}(n-j)\int\tau(\tau\circ F^j)\,d\mu_k ∫ τ n 2 d μ k = n ∫ τ 2 d μ k + 2 j = 1 ∑ n − 1 ( n − j ) ∫ τ ( τ ∘ F j ) d μ k
proves existence of the limit. Differentiating ∫ L z n ρ k d m = λ k ( z ) n ∫ Π k ( z ) ρ k d m + O ( θ n ) \int\mathcal{L}_z^n\rho_k\,dm=\lambda_k(z)^n\int\Pi_k(z)\rho_k\,dm+O(\theta^n) ∫ L z n ρ k d m = λ k ( z ) n ∫ Π k ( z ) ρ k d m + O ( θ n ) twice and dividing by n n n identifies it with − λ k ′ ′ ( 0 ) -\lambda_k''(0) − λ k ′′ ( 0 ) ; differentiated residual terms also decay exponentially, by analyticity and Cauchy estimates.
If σ k 2 = 0 \sigma_k^2=0 σ k 2 = 0 , exponential covariance decay and the preceding identity give sup n ∥ τ n ∥ L 2 ( μ k ) < ∞ \sup_n\|\tau_n\|_{L^2(\mu_k)}<\infty sup n ∥ τ n ∥ L 2 ( μ k ) < ∞ . Write U u = u ∘ F Uu=u\circ F U u = u ∘ F on L 2 ( μ k ) L^2(\mu_k) L 2 ( μ k ) and set b M = M − 1 ∑ n = 1 M τ n b_M=M^{-1}\sum_{n=1}^M\tau_n b M = M − 1 ∑ n = 1 M τ n . The sequence b M b_M b M is bounded, while
( Id − U ) b M = τ − 1 M ∑ n = 1 M U n τ ⟶ τ in L 2 ( μ k ) (\operatorname{Id}-U)b_M=\tau-\frac{1}{M}\sum_{n=1}^M U^n\tau\longrightarrow\tau\quad\text{in }L^2(\mu_k) ( Id − U ) b M = τ − M 1 n = 1 ∑ M U n τ ⟶ τ in L 2 ( μ k )
by the mean ergodic theorem and centring. A weakly convergent subsequence gives τ = b − b ∘ F \tau=b-b\circ F τ = b − b ∘ F with b ∈ L 2 ( μ k ) b\in L^2(\mu_k) b ∈ L 2 ( μ k ) . Consequently e i ν τ = ϕ ∘ F / ϕ e^{i\nu\tau}=\phi\circ F/\phi e i ν τ = ϕ ∘ F / ϕ for ϕ = e − i ν b \phi=e^{-i\nu b} ϕ = e − i ν b and every real ν ≠ 0 \nu\neq0 ν = 0 , contradicting (B2) . This proves strict positivity.
Lemma 5.2 (Characteristic-function expansion). For ψ ∈ H s \psi\in H^s ψ ∈ H s , ξ ∈ L 2 \xi\in L^2 ξ ∈ L 2 , every integer R ≥ 0 R \ge0 R ≥ 0 , and every fixed b > 0 b > 0 b > 0 , there are polynomials A j , k , ψ , ξ A_{j,k,\psi,\xi} A j , k , ψ , ξ , independent of n n n , with degree at most 3 j 3j 3 j and parity j j j , such that
∫ ψ e i t τ n / n ξ ∘ F n d m = ∑ k = 1 ℓ e − σ k 2 t 2 / 2 ∑ j = 0 R n − j / 2 A j , k , ψ , ξ ( t ) + E R , n ( t ) (42) \int\psi e^{it\tau_n/\sqrt{n}}\xi\circ F^n \, \mathrm{d}m
= \sum_{k=1}^{\ell} e^{-\sigma_k^2t^2/2}\sum_{j=0}^{R}n^{-j/2}A_{j,k,\psi,\xi}(t)+E_{R,n}(t) \tag*{(42)} ∫ ψ e i t τ n / n ξ ∘ F n d m = k = 1 ∑ ℓ e − σ k 2 t 2 /2 j = 0 ∑ R n − j /2 A j , k , ψ , ξ ( t ) + E R , n ( t ) (42) for ∣ t ∣ ≤ b log n |t| \le b\sqrt{\log n} ∣ t ∣ ≤ b log n . Here A 0 , k , ψ , ξ = c k ( ψ , ξ ) A_{0,k,\psi,\xi}=c_k(\psi,\xi) A 0 , k , ψ , ξ = c k ( ψ , ξ ) , and, for some c > 0 c>0 c > 0 , integer M R M_R M R , and θ 1 ∈ ( 0 , 1 ) \theta_1\in(0,1) θ 1 ∈ ( 0 , 1 ) ,
∣ E R , n ( t ) ∣ ≤ C n − ( R + 1 ) / 2 ( 1 + ∣ t ∣ M R ) e − c t 2 + C θ 1 n . (43) |E_{R,n}(t)| \le Cn^{-(R+1)/2}(1+|t|^{M_R})e^{-ct^2}+C\theta_1^n. \tag*{(43)} ∣ E R , n ( t ) ∣ ≤ C n − ( R + 1 ) /2 ( 1 + ∣ t ∣ M R ) e − c t 2 + C θ 1 n . (43) For ψ = v \psi=v ψ = v a probability density and ξ = 1 \xi=1 ξ = 1 , one has A j , k , v , 1 ( 0 ) = 0 A_{j,k,v,1}(0)=0 A j , k , v , 1 ( 0 ) = 0 for j ≥ 1 j\ge1 j ≥ 1 , and the improved estimate
∣ E R , n ( t ) ∣ ≤ C n − ( R + 1 ) / 2 ∣ t ∣ ( 1 + ∣ t ∣ M R ) e − c t 2 + C θ 1 n ∣ t ∣ . (44) |E_{R,n}(t)| \le Cn^{-(R+1)/2}|t|(1+|t|^{M_R})e^{-ct^2}+C\theta_1^n|t|. \tag*{(44)} ∣ E R , n ( t ) ∣ ≤ C n − ( R + 1 ) /2 ∣ t ∣ ( 1 + ∣ t ∣ M R ) e − c t 2 + C θ 1 n ∣ t ∣. (44) For sufficiently small fixed η > 0 \eta>0 η > 0 , the characteristic function in (42) is bounded by C ( e − c t 2 + θ 1 n ) C(e^{-ct^2}+\theta_1^n) C ( e − c t 2 + θ 1 n ) on ∣ t ∣ ≤ η n |t|\le\eta\sqrt n ∣ t ∣ ≤ η n .
Proof. Set a k ( z ) = ∫ ξ Π k ( z ) ψ d m a_k(z)=\int\xi\Pi_k(z)\psi\,\mathrm{d}m a k ( z ) = ∫ ξ Π k ( z ) ψ d m . The scalar spectral decomposition is
∫ ψ e i z τ n ξ ∘ F n d m = ∑ k λ k ( z ) n a k ( z ) + ∫ ξ N ( z ) n ψ d m . \int\psi e^{iz\tau_n}\xi\circ F^n\,\mathrm{d}m
=\sum_k\lambda_k(z)^n a_k(z)+\int\xi N(z)^n\psi\,\mathrm{d}m. ∫ ψ e i z τ n ξ ∘ F n d m = k ∑ λ k ( z ) n a k ( z ) + ∫ ξ N ( z ) n ψ d m .
The analytic logarithm near λ k ( 0 ) = 1 \lambda_k(0)=1 λ k ( 0 ) = 1 has the expansion
log λ k ( z ) = − 1 2 σ k 2 z 2 + ∑ r = 3 R + 2 β r , k z r + O ( ∣ z ∣ R + 3 ) . \log\lambda_k(z)=-\frac{1}{2}\sigma_k^2z^2+\sum_{r=3}^{R+2}\beta_{r,k}z^r+O(|z|^{R+3}). log λ k ( z ) = − 2 1 σ k 2 z 2 + r = 3 ∑ R + 2 β r , k z r + O ( ∣ z ∣ R + 3 ) .
Insert z = t / n z=t/\sqrt n z = t / n , expand a k a_k a k and the exponential, and collect powers of n − 1 / 2 n^{-1/2} n − 1/2 . A log-eigenvalue term of degree r r r has order n − ( r − 2 ) / 2 n^{-(r-2)/2} n − ( r − 2 ) /2 and an amplitude term of degree r r r has order n − r / 2 n^{-r/2} n − r /2 . This proves the degree bound and parity. Taylor remainders on ∣ t ∣ ≤ b log n |t|\le b\sqrt{\log n} ∣ t ∣ ≤ b log n are bounded by a fixed polynomial times n − ( R + 1 ) / 2 e − c t 2 n^{-(R+1)/2}e^{-ct^2} n − ( R + 1 ) /2 e − c t 2 . The spectral residual is exponentially small. This proves (43) .
For ξ = 1 \xi=1 ξ = 1 and probability density v v v , the approximating characteristic function equals 1 1 1 at t = 0 t=0 t = 0 : each a k ( 0 ) = c k ( v ) a_k(0)=c_k(v) a k ( 0 ) = c k ( v ) , ∑ k c k ( v ) = 1 \sum_k c_k(v)=1 ∑ k c k ( v ) = 1 , and every correction polynomial vanishes at zero. The residual integral also vanishes at zero, since ∫ N ( 0 ) n v d m = 0 \int N(0)^n v\,\mathrm{d}m=0 ∫ N ( 0 ) n v d m = 0 . Applying Taylor’s formula to these remainders, with differentiated analytic estimates, inserts a factor ∣ t ∣ |t| ∣ t ∣ and proves (44) . Finally, positivity of the variances gives ∣ λ k ( z ) ∣ ≤ e − c z 2 |\lambda_k(z)|\le e^{-cz^2} ∣ λ k ( z ) ∣ ≤ e − c z 2 for real ∣ z ∣ ≤ η |z|\le\eta ∣ z ∣ ≤ η , proving the last assertion. □
Intermediate and large frequencies
Lemma 5.3 (A uniform scalar bound). There is C > 0 C>0 C > 0 such that for all real ν \nu ν , all n ≥ 0 n\ge0 n ≥ 0 , ψ ∈ H s \psi\in H^s ψ ∈ H s , and ξ ∈ L 2 \xi\in L^2 ξ ∈ L 2 ,
∫ ∣ ψ ∣ ∣ ξ ∘ F n ∣ d m ≤ C ∥ ψ ∥ H s ∥ ξ ∥ L 2 , ∣ ∫ ξ L ν n ψ d m ∣ ≤ C ∥ ψ ∥ H s ∥ ξ ∥ L 2 . (45) \int|\psi|\,|\xi\circ F^n|\,\mathrm{d}m\le C\|\psi\|_{H^s}\|\xi\|_{L^2},
\qquad
\left|\int\xi\mathcal{L}_{\nu}^{n}\psi\,\mathrm{d}m\right|\le C\|\psi\|_{H^s}\|\xi\|_{L^2}. \tag*{(45)} ∫ ∣ ψ ∣ ∣ ξ ∘ F n ∣ d m ≤ C ∥ ψ ∥ H s ∥ ξ ∥ L 2 , ∫ ξ L ν n ψ d m ≤ C ∥ ψ ∥ H s ∥ ξ ∥ L 2 . (45) Proof. If 0 < s ≤ 1 0<s\le1 0 < s ≤ 1 , the modulus map satisfies ∥ ∣ ψ ∣ ∥ H s ≤ C ∥ ψ ∥ H s \||\psi|\|_{H^s}\le C\|\psi\|_{H^s} ∥∣ ψ ∣ ∥ H s ≤ C ∥ ψ ∥ H s , by the fractional difference norm or the weak chain rule. Power boundedness of L F \mathcal{L}_F L F on H s H^s H s therefore gives ∥ L F n ∣ ψ ∣ ∥ 2 ≤ C ∥ ψ ∥ H s \|\mathcal{L}_F^n|\psi|\|_2\le C\|\psi\|_{H^s} ∥ L F n ∣ ψ ∣ ∥ 2 ≤ C ∥ ψ ∥ H s . If s > 1 s>1 s > 1 , Sobolev embedding in dimension two gives instead
∥ L F n ∣ ψ ∣ ∥ 2 ≤ ∥ ψ ∥ ∞ ∥ L F n 1 ∥ 2 ≤ C ∥ ψ ∥ H s , \|\mathcal{L}_F^n|\psi|\|_2\le\|\psi\|_\infty\|\mathcal{L}_F^n1\|_2\le C\|\psi\|_{H^s}, ∥ L F n ∣ ψ ∣ ∥ 2 ≤ ∥ ψ ∥ ∞ ∥ L F n 1 ∥ 2 ≤ C ∥ ψ ∥ H s , again by (20) . In either case, positivity and transfer-operator duality yield
∫ ∣ ψ ∣ ∣ ξ ∘ F n ∣ d m = ∫ ∣ ξ ∣ ∣ L F n ∣ ψ ∣ ∣ d m ≤ C ∥ ψ ∥ H s ∥ ξ ∥ 2 . \int|\psi|\,|\xi\circ F^n|\,\mathrm{d}m
=\int|\xi|\,|\mathcal{L}_F^n|\psi||\,\mathrm{d}m
\le C\|\psi\|_{H^s}\|\xi\|_2. ∫ ∣ ψ ∣ ∣ ξ ∘ F n ∣ d m = ∫ ∣ ξ ∣ ∣ L F n ∣ ψ ∣∣ d m ≤ C ∥ ψ ∥ H s ∥ ξ ∥ 2 .
For unbounded ξ \xi ξ the equality follows by monotone truncation. Finally ∣ L ν n ψ ∣ ≤ L F n ∣ ψ ∣ \lvert L_{\nu}^{n}\psi\rvert\leq L_{F}^{n}\lvert\psi\rvert ∣ L ν n ψ ∣ ≤ L F n ∣ ψ ∣ proves the second bound. □ \square □
For every fixed K > 0 K>0 K > 0 , B > 0 B>0 B > 0 , and integer R ≥ 1 R\geq1 R ≥ 1 , Lemma 3.10 (c) and Theorem 2.10 imply
∫ K < ∣ ν ∣ < B n ( R − 1 ) / 2 ∥ L ν n ∥ H s → H s ∣ ν ∣ d ν = o ( n − R / 2 ) . (46) \int_{K<\lvert\nu\rvert<Bn^{(R-1)/2}}\frac{\lVert\mathcal{L}_{\nu}^{n}\rVert_{H^{s}\to H^{s}}}{\lvert\nu\rvert}\,\mathrm{d}\nu=o\left(n^{-R/2}\right). \tag*{(46)} ∫ K < ∣ ν ∣ < B n ( R − 1 ) /2 ∣ ν ∣ ∥ L ν n ∥ H s → H s d ν = o ( n − R /2 ) . (46)
Indeed, the compact part bounded away from zero is exponentially small, while the remaining integral is at most
C e − γ n ∫ ν 0 max { ν 0 , B n ( R − 1 ) / 2 } ν s + δ − 1 d ν , Ce^{-\gamma n}\int_{\nu_{0}}^{\max\{\nu_{0},Bn^{(R-1)/2}\}}\nu^{s+\delta-1}\,\mathrm{d}\nu, C e − γ n ∫ ν 0 m a x { ν 0 , B n ( R − 1 ) /2 } ν s + δ − 1 d ν ,
which is an exponential times a fixed power of n n n . Note that the restriction away from zero is essential here, since a uniform exponential spectral bound on 0 < ∣ ν ∣ ≤ ν 0 0<\lvert\nu\rvert\leq\nu_{0} 0 < ∣ ν ∣ ≤ ν 0 would contradict λ k ( ν ) → 1 \lambda_{k}(\nu)\to1 λ k ( ν ) → 1 .
For later use, if p > 1 p>1 p > 1 and g ∈ F 0 p g\in\mathcal{F}_{0}^{p} g ∈ F 0 p , the integral over ∣ ν ∣ > n a \lvert\nu\rvert>n^{a} ∣ ν ∣ > n a in Fourier inversion is at most
C ∥ ψ ∥ H s ∥ ξ ∥ 2 C p ( g ) n − a ( p − 1 ) (47) C\lVert\psi\rVert_{H^{s}}\lVert\xi\rVert_{2}C^{p}(g)n^{-a(p-1)} \tag*{(47)} C ∥ ψ ∥ H s ∥ ξ ∥ 2 C p ( g ) n − a ( p − 1 ) (47)
by Lemma 5.3 . On η ≤ ∣ ν ∣ ≤ n a \eta\leq\lvert\nu\rvert\leq n^{a} η ≤ ∣ ν ∣ ≤ n a , the spectral compact-frequency bound and DI make the integral smaller than any inverse power of n n n , for each fixed a a a . Taking a a a large proves that every fixed q > 0 q>0 q > 0 in the test-function classes suffices for every fixed expansion order.
Edgeworth expansions
Let v ∈ H s v\in H^{s} v ∈ H s be nonnegative with ∫ v d m = 1 \int v\,\mathrm{d}m=1 ∫ v d m = 1 , and let χ n ( t ) = ∫ v e i t τ n / n d m \chi_{n}(t)=\int ve^{it\tau_{n}/\sqrt{n}}\,\mathrm{d}m χ n ( t ) = ∫ v e i t τ n / n d m . Define
K j , k , ψ , ξ ( x ) = 1 2 π ∫ R e − i t x e − σ k 2 t 2 / 2 A j , k , ψ , ξ ( t ) d t , K j , ψ , ξ = ∑ k K j , k , ψ , ξ . (48) K_{j,k,\psi,\xi}(x)=\frac{1}{2\pi}\int_{\mathbb{R}}e^{-itx}e^{-\sigma_{k}^{2}t^{2}/2}A_{j,k,\psi,\xi}(t)\,\mathrm{d}t,\qquad K_{j,\psi,\xi}=\sum_{k}K_{j,k,\psi,\xi}. \tag*{(48)} K j , k , ψ , ξ ( x ) = 2 π 1 ∫ R e − i t x e − σ k 2 t 2 /2 A j , k , ψ , ξ ( t ) d t , K j , ψ , ξ = k ∑ K j , k , ψ , ξ . (48)
Each kernel is a polynomial of degree at most 3 j 3j 3 j times n σ k n^{\sigma_{k}} n σ k ; this formula fixes the Fourier signs and normalizations. In the case ( ψ , ξ ) = ( v , 1 ) (\psi,\xi)=(v,1) ( ψ , ξ ) = ( v , 1 ) , ∫ K j , k , v , 1 = 0 \int K_{j,k,v,1}=0 ∫ K j , k , v , 1 = 0 for j ≥ 1 j\geq1 j ≥ 1 . Hence there is a polynomial P j , k , v P_{j,k,v} P j , k , v satisfying
d d x ( P j , k , v ( x ) n σ k ) = K j , k , v , 1 ( x ) . \frac{\mathrm{d}}{\mathrm{d}x}\left(P_{j,k,v}(x)n^{\sigma_{k}} \right)=K_{j,k,v,1}(x). d x d ( P j , k , v ( x ) n σ k ) = K j , k , v , 1 ( x ) .
Define E R , n \mathcal{E}_{R,n} E R , n by the two explicit sums on the right-hand side of (12) . Its derivative is a signed density of total mass one, uniformly bounded in n n n , and its Fourier–Stieltjes transform with the positive characteristic-function sign is
χ R , n ( t ) = ∑ k e − σ k 2 t 2 / 2 ( c k ( v ) + ∑ j = 1 R n − j / 2 A j , k , v , 1 ( t ) ) . \chi_{R,n}(t)=\sum_{k}e^{-\sigma_{k}^{2}t^{2}/2}\left(c_{k}(v)+\sum_{j=1}^{R}n^{-j/2}A_{j,k,v,1}(t)\right). χ R , n ( t ) = k ∑ e − σ k 2 t 2 /2 ( c k ( v ) + j = 1 ∑ R n − j /2 A j , k , v , 1 ( t ) ) .
The Gaussian factor multiplies every correction term. The usual smoothing inequality for comparison with such a signed distribution function gives
sup z ∣ P v m ( τ n n ≤ z ) − E R , n ( z ) ∣ ≤ C ∫ − T T ∣ χ n ( t ) − χ R , n ( t ) ∣ ∣ t ∣ d t + C T ; \sup_{z}\left\lvert\mathbb{P}_{v m}\left(\frac{\tau_{n}}{\sqrt{n}}\leq z\right)-\mathcal{E}_{R,n}(z)\right\rvert\leq C\int_{-T}^{T}\frac{\lvert\chi_{n}(t)-\chi_{R,n}(t)\rvert}{\lvert t\rvert}\,\mathrm{d}t+\frac{C}{T}; z sup P v m ( n τ n ≤ z ) − E R , n ( z ) ≤ C ∫ − T T ∣ t ∣ ∣ χ n ( t ) − χ R , n ( t )∣ d t + T C ;
see the Edgeworth smoothing argument in [10 , 11 ] . Choose T = n ( R + 1 ) / 2 T=n^{(R+1)/2} T = n ( R + 1 ) /2 . On ∣ t ∣ ≤ b log n \lvert t\rvert\leq b\sqrt{\log n} ∣ t ∣ ≤ b log n , (44) makes the integral O ( n − ( R + 1 ) / 2 ) + O ( θ 1 n log n ) O(n^{-(R+1)/2})+O(\theta_{1}^{n}\sqrt{\log n}) O ( n − ( R + 1 ) /2 ) + O ( θ 1 n log n ) . On b log n < ∣ t ∣ ≤ η n b\sqrt{\log n}<\lvert t\rvert\leq\eta\sqrt{n} b log n < ∣ t ∣ ≤ η n , the Gaussian estimates from Lemma 5.2 , and the same estimates for the approximating polynomials, make the integral o ( n − R / 2 ) o(n^{-R/2}) o ( n − R /2 ) when b b b is large enough. On η n < ∣ t ∣ ≤ T \eta\sqrt{n}<\lvert t\rvert\leq T η n < ∣ t ∣ ≤ T , substitute ν = t / n \nu=t/\sqrt{n} ν = t / n . The actual characteristic function is controlled by (46) , with its polynomial upper cutoff n R / 2 n^{R/2} n R /2 (equivalently, use that estimate with order R + 1 R+1 R + 1 ). The approximating characteristic function has an exponentially small Gaussian tail. Finally T − 1 = o ( n − R / 2 ) T^{-1}=o(n^{-R/2}) T − 1 = o ( n − R /2 ) . This proves (12) uniformly in z z z .
Mixing global and local expansions
Let p = q + 2 > 2 p=q+2>2 p = q + 2 > 2 . Since g ^ ∈ L 1 \widehat{g}\in L^{1} g ∈ L 1 and Lemma 5.3 bounds the absolute endpoint integral, Fubini’s theorem and Fourier inversion give
I n ( g ; ψ , ξ ) = 1 2 π ∫ R g ^ ( ν ) ( ∫ ξ L ν n ψ d m ) d ν . (49) I_n(g;\psi,\xi)=\frac{1}{2\pi}\int_{\mathbb{R}}\widehat{g}(\nu)\left(\int\xi\mathcal{L}_{\nu}^{n}\psi\,dm\right)d\nu. \tag*{(49)} I n ( g ; ψ , ξ ) = 2 π 1 ∫ R g ( ν ) ( ∫ ξ L ν n ψ d m ) d ν . (49)
By (47) , the compact-frequency bound and DI, all frequencies outside ∣ ν ∣ ≤ η |\nu|\leq\eta ∣ ν ∣ ≤ η contribute o ( n − L ) o(n^{-L}) o ( n − L ) for any prescribed L L L . After t = n ν t=\sqrt{n}\nu t = n ν , frequencies with b log n < ∣ t ∣ ≤ η n b\sqrt{\log n}<|t|\leq\eta\sqrt{n} b log n < ∣ t ∣ ≤ η n also contribute o ( n − L ) o(n^{-L}) o ( n − L ) if b b b is sufficiently large. On the remaining interval insert (42) . The integrated error in (43) , including the factor n − 1 / 2 n^{-1/2} n − 1/2 from the change of variable, is O ( n − ( R + 2 ) / 2 ) + O ( θ 1 n ) O(n^{-(R+2)/2})+O(\theta_{1}^{n}) O ( n − ( R + 2 ) /2 ) + O ( θ 1 n ) . Extending the Gaussian integrals to R \mathbb{R} R gives
I n ( g ; ψ , ξ ) = ∑ j = 0 R n − ( j + 1 ) / 2 2 π ∑ k ∫ R g ^ ( t / n ) A j , k , ψ , ξ ( t ) e − σ k 2 t 2 / 2 d t + o ( n − ( R + 1 ) / 2 ) . I_n(g;\psi,\xi)=\sum_{j=0}^{R}\frac{n^{-(j+1)/2}}{2\pi}\sum_{k}\int_{\mathbb{R}}\widehat{g}(t/\sqrt{n})A_{j,k,\psi,\xi}(t)e^{-\sigma_k^{2}t^{2}/2}\,dt+o(n^{-(R+1)/2}). I n ( g ; ψ , ξ ) = j = 0 ∑ R 2 π n − ( j + 1 ) /2 k ∑ ∫ R g ( t / n ) A j , k , ψ , ξ ( t ) e − σ k 2 t 2 /2 d t + o ( n − ( R + 1 ) /2 ) .
Using (48) and Fubini, the integral divided by 2 π 2\pi 2 π equals ∫ g ( x ) K j , ψ , ξ ( x / n ) d x \int g(x)K_{j,\psi,\xi}(x/\sqrt{n})\,dx ∫ g ( x ) K j , ψ , ξ ( x / n ) d x . This proves the global expansion (13) . In particular, its leading size is n − 1 / 2 n^{-1/2} n − 1/2 , as required by the local limit theorem.
For the local expansion suppose g ∈ F R + 1 p g\in\mathcal{F}^{p}_{R+1} g ∈ F R + 1 p . Multiply the global expansion by n \sqrt{n} n and Taylor expand each K j ( x / n ) K_j(x/\sqrt{n}) K j ( x / n ) at zero through degree R − j R-j R − j . All derivatives of these Gaussian-polynomial kernels are bounded. Thus the integrated Taylor errors are O ( n − ( R + 1 ) / 2 M R + 1 ( g ) ) O(n^{-(R+1)/2}M_{R+1}(g)) O ( n − ( R + 1 ) /2 M R + 1 ( g )) . The parity in Lemma 5.2 implies K j ( − x ) = ( − 1 ) j K j ( x ) K_j(-x)=(-1)^jK_j(x) K j ( − x ) = ( − 1 ) j K j ( x ) ; hence the terms with j + d j+d j + d odd vanish. Define
Q m , ψ , ξ ( x ) = ∑ j + d = 2 m K j , ψ , ξ ( d ) ( 0 ) d ! x d . (50) Q_{m,\psi,\xi}(x)=\sum_{j+d=2m}\frac{K_{j,\psi,\xi}^{(d)}(0)}{d!}x^d. \tag*{(50)} Q m , ψ , ξ ( x ) = j + d = 2 m ∑ d ! K j , ψ , ξ ( d ) ( 0 ) x d . (50)
Collecting the even powers yields (14) , with deg Q m ≤ 2 m \deg Q_m\leq2m deg Q m ≤ 2 m and the stated leading coefficient. This completes the proof of Theorem A.
Proof of Theorem B
We use the fibrewise Fourier argument of [2 ] . Recall that F ~ ( x , y , z ) = ( F ( x , y ) , z + τ ( x , y ) ) mod Z 3 \widetilde{F}(x,y,z)=(F(x,y),z+\tau(x,y))\ \operatorname{mod}\mathbb{Z}^{3} F ( x , y , z ) = ( F ( x , y ) , z + τ ( x , y )) mod Z 3 .
Thanks to the skew-product structure, we can exploit the fiber-wise Fourier expansion of a function f ∈ C r ( T 3 , C ) f\in C^{r}(\mathbb{T}^{3},\mathbb{C}) f ∈ C r ( T 3 , C ) :
f ( x , y , z ) = ∑ k ∈ Z e 2 π i k z f ^ k ( x , y ) , f ^ k ( x , y ) = ∫ T e − 2 π i k z f ( x , y , z ) d z . f(x,y,z)=\sum_{k\in\mathbb{Z}}e^{2\pi ikz}\widehat{f}_k(x,y),\qquad\widehat{f}_k(x,y)=\int_{\mathbb{T}}e^{-2\pi ikz}f(x,y,z)\,dz. f ( x , y , z ) = k ∈ Z ∑ e 2 π ik z f k ( x , y ) , f k ( x , y ) = ∫ T e − 2 π ik z f ( x , y , z ) d z .
Integrating by parts in z z z , for 0 ≤ j ≤ r 0\leq j\leq r 0 ≤ j ≤ r we have
∥ f ^ k ∥ C j ( T 2 ) ≤ ∥ f ∥ C j , ∥ f ^ k ∥ C 0 ( T 2 ) ≤ ( 2 π ∣ k ∣ ) − r ∥ f ∥ C r ( k ≠ 0 ) . (51) \left\|\widehat{f}_k\right\|_{C^j(\mathbb{T}^{2})}\leq\|f\|_{C^j},\qquad\left\|\widehat{f}_k\right\|_{C^0(\mathbb{T}^{2})}\leq(2\pi|k|)^{-r}\|f\|_{C^r}\quad(k\ne0). \tag*{(51)} f k C j ( T 2 ) ≤ ∥ f ∥ C j , f k C 0 ( T 2 ) ≤ ( 2 π ∣ k ∣ ) − r ∥ f ∥ C r ( k = 0 ) . (51)
For notational convenience we set p = ( x , y ) ∈ T 2 p=(x,y)\in\mathbb{T}^{2} p = ( x , y ) ∈ T 2 and d p = d x d y dp=dxdy d p = d x d y . We will prove the statement for observables in C r C^{r} C r , with an integer r ≥ max { ⌈ s ⌉ , 2 } r\geq\max\{\lceil s\rceil,2\} r ≥ max {⌈ s ⌉ , 2 } , and deduce the conclusion for Hölder observables at the end of the proof. By assumption μ ∗ \mu_* μ ∗ is an ergodic acip of F F F , hence μ ∗ = μ k 0 \mu_*=\mu_{k_0} μ ∗ = μ k 0 and h ∗ = ρ k 0 ∈ H s ( T 2 ) h_*=\rho_{k_0}\in H^{s}(\mathbb{T}^{2}) h ∗ = ρ k 0 ∈ H s ( T 2 ) for some k 0 k_0 k 0 .
Let g , h ∈ C r ( T 3 , C ) g,h\in C^{r}(\mathbb{T}^{3},\mathbb{C}) g , h ∈ C r ( T 3 , C ) and assume that μ ( g ) = 0 \mu(g)=0 μ ( g ) = 0 , where d μ = d μ ∗ × Leb d\mu=d\mu_*\times\operatorname{Leb} d μ = d μ ∗ × Leb ; note in particular that this implies μ ∗ ( g ^ 0 ) = 0 \mu_*(\widehat{g}_0)=0 μ ∗ ( g 0 ) = 0 . Recall that, for u ∈ L 1 u\in L^{1} u ∈ L 1 , w ∈ L ∞ w\in L^{\infty} w ∈ L ∞ and ν ∈ R \nu\in\mathbb{R} ν ∈ R ,
∫ T 2 u ⋅ ( w ∘ F n ) e i ν τ n d p = ∫ T 2 L ν n ( u ) w d p . \int_{\mathbb{T}^{2}}u\cdot(w\circ F^{n})e^{i\nu\tau_n}\,dp=\int_{\mathbb{T}^{2}}\mathcal{L}_{\nu}^{n}(u)w\,dp. ∫ T 2 u ⋅ ( w ∘ F n ) e i ν τ n d p = ∫ T 2 L ν n ( u ) w d p .
We compute for each n ∈ N n \in\mathbb{N} n ∈ N , by linearity and absolute uniform convergence of the Fourier series (as r ≥ 2 r \ge2 r ≥ 2 ),
∫ T 3 ( g ⋅ h ∘ F ~ n ) ( p , z ) h ∗ ( p ) d p d z = ∑ k ∈ Z ∑ j ∈ Z ∫ T 2 ( g ^ k ⋅ h ∗ ) ( p ) h ^ j ( F n ( p ) ) e 2 π i j τ n ( p ) d p ∫ T e 2 π i ( k + j ) z d z = ∑ k ∈ Z ∫ T 2 L − 2 π k n ( g ^ k ⋅ h ∗ ) ( p ) h ^ − k ( p ) d p , \int_{\mathbb{T}^{3}} (g \cdot h \circ\widetilde{F}^{n})(p,z)h_{*}(p)\,dp\,dz
= \sum_{k\in\mathbb{Z}}\sum_{j\in\mathbb{Z}}\int_{\mathbb{T}^{2}}(\widehat{g}_{k}\cdot h_{*})(p)\widehat{h}_{j}(F^{n}(p))e^{2\pi i j\tau_{n}(p)}\,dp\int_{\mathbb{T}}e^{2\pi i(k+j)z}\,dz
= \sum_{k\in\mathbb{Z}}\int_{\mathbb{T}^{2}}\mathcal{L}^{n}_{-2\pi k}(\widehat{g}_{k}\cdot h_{*})(p)\widehat{h}_{-k}(p)\,dp, ∫ T 3 ( g ⋅ h ∘ F n ) ( p , z ) h ∗ ( p ) d p d z = k ∈ Z ∑ j ∈ Z ∑ ∫ T 2 ( g k ⋅ h ∗ ) ( p ) h j ( F n ( p )) e 2 π ij τ n ( p ) d p ∫ T e 2 π i ( k + j ) z d z = k ∈ Z ∑ ∫ T 2 L − 2 π k n ( g k ⋅ h ∗ ) ( p ) h − k ( p ) d p ,
where in the last line we used that ∫ T e 2 π i ( k + j ) z d z = 0 \int_{\mathbb{T}}e^{2\pi i(k+j)z}\,dz=0 ∫ T e 2 π i ( k + j ) z d z = 0 unless j = − k j=-k j = − k .
Set A k : = ∫ T 2 L − 2 π k n ( g ^ k ⋅ h ∗ ) ( p ) h ^ − k ( p ) d p A_k:=\int_{\mathbb{T}^{2}}\mathcal{L}^{n}_{-2\pi k}(\widehat{g}_{k}\cdot h_{*})(p)\widehat{h}_{-k}(p)\,dp A k := ∫ T 2 L − 2 π k n ( g k ⋅ h ∗ ) ( p ) h − k ( p ) d p and write
∑ k ∈ Z A k = A 0 + ∑ 0 < 2 π ∣ k ∣ < ν 0 A k + ∑ 2 π ∣ k ∣ ≥ ν 0 A k . \sum_{k\in\mathbb{Z}}A_k=A_0+\sum_{0<2\pi|k|<\nu_0}A_k+\sum_{2\pi|k|\ge\nu_0}A_k. k ∈ Z ∑ A k = A 0 + 0 < 2 π ∣ k ∣ < ν 0 ∑ A k + 2 π ∣ k ∣ ≥ ν 0 ∑ A k .
where ν 0 \nu_0 ν 0 is given by Theorem 2.10, enlarged so that ν 0 > 2 π \nu_0>2\pi ν 0 > 2 π . We prove that the absolute value of each of these three terms decays exponentially. Below we use Lemma 4.1 in the form ∥ φ u ∥ H s ≤ C ∥ φ ∥ C ⌈ s ⌉ ∥ u ∥ H s \|\varphi u\|_{H^s}\le C\|\varphi\|_{C^{\lceil s\rceil}}\|u\|_{H^s} ∥ φ u ∥ H s ≤ C ∥ φ ∥ C ⌈ s ⌉ ∥ u ∥ H s .
The term A 0 A_0 A 0 . Let Π \Pi Π be the spectral projector of L F \mathcal{L}_F L F on H s H^s H s associated with the eigenvalue 1 1 1 . By Lemma 3.10(a) there exist C > 0 C>0 C > 0 , θ 0 ∈ ( 0 , 1 ) \theta_0\in(0,1) θ 0 ∈ ( 0 , 1 ) such that ∥ L F n ( Id − Π ) ∥ H s ≤ C θ 0 n \|\mathcal{L}_F^n(\operatorname{Id}-\Pi)\|_{H^s}\le C\theta_0^n ∥ L F n ( Id − Π ) ∥ H s ≤ C θ 0 n . We claim that Π ( g ^ 0 h ∗ ) = 0 \Pi(\widehat{g}_0h_*)=0 Π ( g 0 h ∗ ) = 0 . Let E : = { ρ k 0 > 0 } E:=\{\rho_{k_0}>0\} E := { ρ k 0 > 0 } . Since ρ k 0 ≥ 0 \rho_{k_0}\ge0 ρ k 0 ≥ 0 and L F ρ k 0 = ρ k 0 \mathcal{L}_F\rho_{k_0}=\rho_{k_0} L F ρ k 0 = ρ k 0 , for almost every x ∉ E x\notin E x ∈ / E we have ρ k 0 ( y ) = 0 \rho_{k_0}(y)=0 ρ k 0 ( y ) = 0 for all y ∈ F − 1 ( x ) y\in F^{-1}(x) y ∈ F − 1 ( x ) ; hence, if u u u vanishes outside E E E , so does L F u \mathcal{L}_Fu L F u . By Lemma 3.10(a), Π u = lim n L F n u \Pi u=\lim_{n}\mathcal{L}_F^n u Π u = lim n L F n u in H s H^s H s , so Π u \Pi u Π u vanishes outside E E E as well. Since Π u \Pi u Π u is a combination of the ρ j \rho_j ρ j , whose positivity sets are disjoint modulo null sets, Π u = c ρ k 0 \Pi u=c\rho_{k_0} Π u = c ρ k 0 with c = ∫ Π u = lim n ∫ L F n u = ∫ u c=\int\Pi u=\lim_n\int\mathcal{L}_F^n u=\int u c = ∫ Π u = lim n ∫ L F n u = ∫ u . Applying this to u = g ^ 0 h ∗ = g ^ 0 ρ k 0 u=\widehat{g}_0h_*=\widehat{g}_0\rho_{k_0} u = g 0 h ∗ = g 0 ρ k 0 , for which ∫ u = μ ∗ ( g ^ 0 ) = 0 \int u=\mu_*(\widehat{g}_0)=0 ∫ u = μ ∗ ( g 0 ) = 0 , proves the claim. Hence
∣ A 0 ∣ = ∣ ∫ T 2 L F n ( g ^ 0 h ∗ ) ( p ) h ^ 0 ( p ) d p ∣ ≤ ∥ L F n ( Id − Π ) ( g ^ 0 h ∗ ) ∥ L 2 ∥ h ^ 0 ∥ L 2 ≤ C θ 0 n ∥ g ^ 0 h ∗ ∥ H s ∥ h ^ 0 ∥ L 2 ≤ C θ 0 n ∥ g ∥ C r ∥ h ∥ C r ∥ h ∗ ∥ H s . |A_0|=\left|\int_{\mathbb{T}^{2}}\mathcal{L}_F^n(\widehat{g}_0h_*)(p)\widehat{h}_0(p)\,dp\right|
\le\|\mathcal{L}_F^n(\operatorname{Id}-\Pi)(\widehat{g}_0h_*)\|_{L^2}\|\widehat{h}_0\|_{L^2}
\le C\theta_0^n\|\widehat{g}_0h_*\|_{H^s}\|\widehat{h}_0\|_{L^2}
\le C\theta_0^n\|g\|_{C^r}\|h\|_{C^r}\|h_*\|_{H^s}. ∣ A 0 ∣ = ∫ T 2 L F n ( g 0 h ∗ ) ( p ) h 0 ( p ) d p ≤ ∥ L F n ( Id − Π ) ( g 0 h ∗ ) ∥ L 2 ∥ h 0 ∥ L 2 ≤ C θ 0 n ∥ g 0 h ∗ ∥ H s ∥ h 0 ∥ L 2 ≤ C θ 0 n ∥ g ∥ C r ∥ h ∥ C r ∥ h ∗ ∥ H s .
The terms 0 < 2 π ∣ k ∣ < ν 0 0<2\pi|k|<\nu_0 0 < 2 π ∣ k ∣ < ν 0 . These are finitely many, and by (28) with J = { 2 π ≤ ∣ ν ∣ ≤ ν 0 } J=\{2\pi\le|\nu|\le\nu_0\} J = { 2 π ≤ ∣ ν ∣ ≤ ν 0 } ,
max 0 < 2 π ∣ k ∣ < ν 0 ∣ A k ∣ ≤ C θ J n max 0 < 2 π ∣ k ∣ < ν 0 ∥ g ^ k h ∗ ∥ H s ∥ h ^ − k ∥ L 2 ≤ C θ J n ∥ g ∥ C r ∥ h ∥ C r ∥ h ∗ ∥ H s . \max_{0<2\pi|k|<\nu_0}|A_k|
\le C\theta_J^n\max_{0<2\pi|k|<\nu_0}\|\widehat{g}_kh_*\|_{H^s}\|\widehat{h}_{-k}\|_{L^2}
\le C\theta_J^n\|g\|_{C^r}\|h\|_{C^r}\|h_*\|_{H^s}. 0 < 2 π ∣ k ∣ < ν 0 max ∣ A k ∣ ≤ C θ J n 0 < 2 π ∣ k ∣ < ν 0 max ∥ g k h ∗ ∥ H s ∥ h − k ∥ L 2 ≤ C θ J n ∥ g ∥ C r ∥ h ∥ C r ∥ h ∗ ∥ H s .
The terms 2 π ∣ k ∣ ≥ ν 0 2\pi|k|\ge\nu_0 2 π ∣ k ∣ ≥ ν 0 . By (16) , ∥ u ∥ L 2 ≤ ∥ u ∥ ν ≤ 2 ∥ u ∥ H s \|u\|_{L^2}\le\|u\|_{\nu}\le2\|u\|_{H^s} ∥ u ∥ L 2 ≤ ∥ u ∥ ν ≤ 2∥ u ∥ H s , and (51) ,
∣ A k ∣ ≤ ∥ L − 2 π k n ( g ^ k h ∗ ) ∥ − 2 π k ∥ h ^ − k ∥ L 2 ≤ C ∣ k ∣ δ e − γ n ∥ g ^ k h ∗ ∥ H s ∥ h ^ − k ∥ L 2 ≤ C ∣ k ∣ δ − r e − γ n ∥ g ∥ C r ∥ h ∥ C r ∥ h ∗ ∥ H s , |A_k|\le\|\mathcal{L}_{-2\pi k}^n(\widehat{g}_kh_*)\|_{-2\pi k}\|\widehat{h}_{-k}\|_{L^2}
\le C|k|^\delta e^{-\gamma n}\|\widehat{g}_kh_*\|_{H^s}\|\widehat{h}_{-k}\|_{L^2}
\le C|k|^{\delta-r}e^{-\gamma n}\|g\|_{C^r}\|h\|_{C^r}\|h_*\|_{H^s}, ∣ A k ∣ ≤ ∥ L − 2 π k n ( g k h ∗ ) ∥ − 2 π k ∥ h − k ∥ L 2 ≤ C ∣ k ∣ δ e − γ n ∥ g k h ∗ ∥ H s ∥ h − k ∥ L 2 ≤ C ∣ k ∣ δ − r e − γ n ∥ g ∥ C r ∥ h ∥ C r ∥ h ∗ ∥ H s ,
and therefore
∣ ∑ 2 π ∣ k ∣ ≥ ν 0 A k ∣ ≤ C e − γ n ∥ g ∥ C r ∥ h ∥ C r ∥ h ∗ ∥ H s ∑ k ≠ 0 ∣ k ∣ δ − r , \left|\sum_{2\pi|k|\ge\nu_0}A_k\right|
\le Ce^{-\gamma n}\|g\|_{C^r}\|h\|_{C^r}\|h_*\|_{H^s}\sum_{k\ne0}|k|^{\delta-r}, 2 π ∣ k ∣ ≥ ν 0 ∑ A k ≤ C e − γ n ∥ g ∥ C r ∥ h ∥ C r ∥ h ∗ ∥ H s k = 0 ∑ ∣ k ∣ δ − r ,
where the last sum converges since r ≥ 2 > 1 + δ r\ge2>1+\delta r ≥ 2 > 1 + δ . This proves the statement for C r C^r C r observables, with some rate γ ′ > 0 \gamma'>0 γ ′ > 0 and constant C ∥ g ∥ C r ∥ h ∥ C r C\|g\|_{C^r}\|h\|_{C^r} C ∥ g ∥ C r ∥ h ∥ C r .
Hölder observables. Let g , h ∈ C α ( T 3 ) g,h\in C^\alpha(\mathbb{T}^3) g , h ∈ C α ( T 3 ) , α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , and for ε ∈ ( 0 , 1 ) \varepsilon\in(0,1) ε ∈ ( 0 , 1 ) let g ε , h ε g_\varepsilon,h_\varepsilon g ε , h ε be their convolutions with a smooth mollifier at scale ε \varepsilon ε , so that
∥ g − g ε ∥ C 0 ≤ ε α ∥ g ∥ C α , ∥ g ε ∥ C r ≤ C ε α − r ∥ g ∥ C α , \|g-g_\varepsilon\|_{C^0}\le\varepsilon^\alpha\|g\|_{C^\alpha},\qquad\|g_\varepsilon\|_{C^r}\le C\varepsilon^{\alpha-r}\|g\|_{C^\alpha}, ∥ g − g ε ∥ C 0 ≤ ε α ∥ g ∥ C α , ∥ g ε ∥ C r ≤ C ε α − r ∥ g ∥ C α ,
and similarly for h h h . Since μ \mu μ is F ~ \widetilde{F} F -invariant, replacing ( g , h ) (g,h) ( g , h ) by ( g ε , h ε ) (g_{\varepsilon},h_{\varepsilon}) ( g ε , h ε ) changes both μ ( g ⋅ h ∘ F ~ n ) \mu(g\cdot h\circ\widetilde{F}^{n}) μ ( g ⋅ h ∘ F n ) and μ ( g ) μ ( h ) \mu(g)\mu(h) μ ( g ) μ ( h ) by at most C ε α ∥ g ∥ C α ∥ h ∥ C α C\varepsilon^{\alpha}\lVert g\rVert_{C^{\alpha}}\lVert h\rVert_{C^{\alpha}} C ε α ∥ g ∥ C α ∥ h ∥ C α . Applying the C r C^{r} C r estimate to g ε − μ ( g ε ) g_{\varepsilon}-\mu(g_{\varepsilon}) g ε − μ ( g ε ) and h ε h_{\varepsilon} h ε ,
∣ μ ( g ⋅ h ∘ F ~ n ) − μ ( g ) μ ( h ) ∣ ≤ C ( ε α + ε 2 ( α − r ) e − γ ′ n ) ∥ g ∥ C α ∥ h ∥ C α , \left|\mu(g\cdot h\circ\widetilde{F}^{n})-\mu(g)\mu(h)\right|\leq C\left(\varepsilon^{\alpha}+\varepsilon^{2(\alpha-r)}e^{-\gamma' n}\right)\lVert g\rVert_{C^{\alpha}}\lVert h\rVert_{C^{\alpha}}, μ ( g ⋅ h ∘ F n ) − μ ( g ) μ ( h ) ≤ C ( ε α + ε 2 ( α − r ) e − γ ′ n ) ∥ g ∥ C α ∥ h ∥ C α ,
and the choice ε = e − γ ′ n / ( 2 r ) \varepsilon=e^{-\gamma'n/(2r)} ε = e − γ ′ n / ( 2 r ) yields the bound C e − α γ ′ n / ( 2 r ) ∥ g ∥ C α ∥ h ∥ C α Ce^{-\alpha\gamma'n/(2r)}\lVert g\rVert_{C^{\alpha}}\lVert h\rVert_{C^{\alpha}} C e − α γ ′ n / ( 2 r ) ∥ g ∥ C α ∥ h ∥ C α .
For α ≥ 1 \alpha\geq1 α ≥ 1 , the assertion follows by the continuous embedding into C β C^{\beta} C β for any fixed 0 < β < 1 0<\beta<1 0 < β < 1 .
Partial captivity in the case of skew products
In this Appendix we characterise (B1) for skew products and provide an example to which our results apply. Let
F ( x , θ ) = ( f ( x ) , θ + ω ( x ) ) , τ = τ ( x ) . F(x,\theta)=(f(x),\theta+\omega(x)),\qquad\tau=\tau(x). F ( x , θ ) = ( f ( x ) , θ + ω ( x )) , τ = τ ( x ) .
where f ∈ C ∞ ( T , T ) f\in C^{\infty}(\mathbb{T},\mathbb{T}) f ∈ C ∞ ( T , T ) is expanding of degree d ≥ 2 d\geq2 d ≥ 2 , ω , τ ∈ C ∞ ( T , R ) \omega,\tau\in C^{\infty}(\mathbb{T},\mathbb{R}) ω , τ ∈ C ∞ ( T , R ) , and F F F is an SVPH. Let h ∈ C ∞ ( T ) h\in C^{\infty}(\mathbb{T}) h ∈ C ∞ ( T ) be the positive invariant density of f f f (L f h = h \mathcal{L}_{f}h=h L f h = h , ∫ h = 1 \int h=1 ∫ h = 1 ). For x ∈ T x\in\mathbb{T} x ∈ T and n ∈ N n\in\mathbb{N} n ∈ N let H f n ( x ) \mathcal{H}_{f}^{n}(x) H f n ( x ) be the set of inverse branches g g g of f n f^{n} f n near x x x . Writing Φ : = ( ω , τ ) : T → R 2 \Phi:=(\omega,\tau):\mathbb{T}\to\mathbb{R}^{2} Φ := ( ω , τ ) : T → R 2 and Φ n : = ∑ k = 0 n − 1 Φ ∘ f k \Phi_{n}:=\sum_{k=0}^{n-1}\Phi\circ f^{k} Φ n := ∑ k = 0 n − 1 Φ ∘ f k , we set, for g ∈ H f n ( x ) g\in\mathcal{H}_{f}^{n}(x) g ∈ H f n ( x ) ,
P g ( x ) : = ( Φ n ∘ g ) ′ ( x ) = ( ( ω n ∘ g ) ′ ( x ) , ( τ n ∘ g ) ′ ( x ) ) = : ( s g , t g ) ∈ R 2 , P_{g}(x):=(\Phi_{n}\circ g)'(x)=((\omega_{n}\circ g)'(x),(\tau_{n}\circ g)'(x))=:(s_{g},t_{g})\in\mathbb{R}^{2}, P g ( x ) := ( Φ n ∘ g ) ′ ( x ) = (( ω n ∘ g ) ′ ( x ) , ( τ n ∘ g ) ′ ( x )) =: ( s g , t g ) ∈ R 2 ,
p g ( x ) : = g ′ ( x ) h ( g ( x ) ) h ( x ) , p_{g}(x):=\frac{g'(x)h(g(x))}{h(x)}, p g ( x ) := h ( x ) g ′ ( x ) h ( g ( x )) ,
and μ n x : = ∑ g ∈ H f n ( x ) p g ( x ) δ P g ( x ) \mu_{n}^{x}:=\sum_{g\in\mathcal{H}_{f}^{n}(x)}p_{g}(x)\delta_{P_{g}(x)} μ n x := ∑ g ∈ H f n ( x ) p g ( x ) δ P g ( x ) , a probability measure on R 2 \mathbb{R}^{2} R 2 . We say that a function ϕ ∈ C ∞ ( T ) \phi\in C^{\infty}(\mathbb{T}) ϕ ∈ C ∞ ( T ) is cohomologous to a constant if ϕ = C ∘ f − C + k \phi=C\circ f-C+k ϕ = C ∘ f − C + k for some C ∈ C 1 ( T ) C\in C^{1}(\mathbb{T}) C ∈ C 1 ( T ) and k ∈ R k\in\mathbb{R} k ∈ R .12
Lemma A.1. Let F , τ F,\tau F , τ be as above and assume that ω \omega ω is not cohomologous to a constant.13 Then the following are equivalent:
(B1) holds for some κ ∈ ( 0 , log λ ) \kappa\in(0,\log\lambda) κ ∈ ( 0 , log λ ) ;
(B1) holds for every κ ∈ ( 0 , log λ ) \kappa\in(0,\log\lambda) κ ∈ ( 0 , log λ ) , with C C C and γ 1 \gamma_{1} γ 1 allowed to depend on κ \kappa κ ;
for every b ∈ R b\in\mathbb{R} b ∈ R , τ − b ω \tau-b\omega τ − bω is not cohomologous to a constant.
Moreover, if these hold, then F F F satisfies (T).
Condition (iii) with b = 0 b=0 b = 0 is the non-coboundary condition on the fibre map τ \tau τ . For b ≠ 0 b\neq0 b = 0 it excludes a resonance between the two fibre cocycles ω \omega ω and τ \tau τ . Together with the assumption on ω \omega ω , (iii) says that a ω + b τ a\omega+b\tau aω + b τ is not cohomologous to a constant for every ( a , b ) ≠ ( 0 , 0 ) (a,b)\neq(0,0) ( a , b ) = ( 0 , 0 ) , which is the condition under which exponential mixing of the T 2 \mathbb{T}^{2} T 2 -extension F ~ \widetilde{F} F of f f f is proved in [5 ] . Finally, the assumption that ω \omega ω is not cohomologous to a constant is the condition that makes the unstable cones of F F F transversal, see (T).
Proof. Throughout, C C C denotes constants depending only on f , ω , τ f,\omega,\tau f , ω , τ . By (5) applied to v = ( 1 , 0 ) v=(1,0) v = ( 1 , 0 ) , C − 1 Λ − n ≤ g ′ ( x ) ≤ C λ − n C^{-1}\Lambda^{-n}\leq g'(x)\leq C\lambda^{-n} C − 1 Λ − n ≤ g ′ ( x ) ≤ C λ − n for g ∈ H f n ( x ) g\in\mathcal{H}_{f}^{n}(x) g ∈ H f n ( x ) , and C − 1 g ′ ( x ) ≤ p g ( x ) ≤ C g ′ ( x ) C^{-1}g'(x)\leq p_{g}(x)\leq Cg'(x) C − 1 g ′ ( x ) ≤ p g ( x ) ≤ C g ′ ( x ) . Step 1: translation of (11) . Let q = ( x , θ ) q=(x,\theta) q = ( x , θ ) . Since F n ( y , ϕ ) = ( f n ( y ) , ϕ + ω n ( y ) ) F^n(y,\phi)=(f^n(y),\phi+\omega_n(y)) F n ( y , ϕ ) = ( f n ( y ) , ϕ + ω n ( y )) , the inverse branches of F n F^n F n at q q q are h ( x , θ ) = ( g ( x ) , θ − ω n ( g ( x ) ) ) h(x,\theta)=(g(x),\theta-\omega_n(g(x))) h ( x , θ ) = ( g ( x ) , θ − ω n ( g ( x ))) , g ∈ H f n ( x ) g\in\mathcal{H}_{f}^{n}(x) g ∈ H f n ( x ) , and J h ( q ) = g ′ ( x ) J_h(q)=g'(x) J h ( q ) = g ′ ( x ) , τ n ∘ h ( x , θ ) = τ n ( g ( x ) ) \tau_n\circ h(x,\theta)=\tau_n(g(x)) τ n ∘ h ( x , θ ) = τ n ( g ( x )) . With y = g ( x ) y=g(x) y = g ( x ) ,
D h ( q ) F n ( 1 σ ) = ( f n ′ ( y ) ω n ′ ( y ) + σ ) . D_{h(q)}F^n\begin{pmatrix}1\\ \sigma\end{pmatrix}
=
\begin{pmatrix}f^n{}'(y)\\ \omega_n'(y)+\sigma\end{pmatrix}. D h ( q ) F n ( 1 σ ) = ( f n ′ ( y ) ω n ′ ( y ) + σ ) .
so, up to sign (which does not affect the absolute-value condition), the unit vectors of D h ( q ) F n C u D_{h(q)}F^n C_u D h ( q ) F n C u are u = ( 1 , s ) / 1 + s 2 u=(1,s)/\sqrt{1+s^2} u = ( 1 , s ) / 1 + s 2 with s = s g + σ g ′ ( x ) s=s_g+\sigma g'(x) s = s g + σ g ′ ( x ) , ∣ σ ∣ ≤ χ u |\sigma|\leq\chi_u ∣ σ ∣ ≤ χ u ; in particular ∣ s ∣ ≤ χ u < 1 |s|\leq\chi_u<1 ∣ s ∣ ≤ χ u < 1 . Moreover d ( τ n ∘ h ) q ( u ) = t g / 1 + s 2 d(\tau_n\circ h)_q(u)=t_g/\sqrt{1+s^2} d ( τ n ∘ h ) q ( u ) = t g / 1 + s 2 and ζ ( u ) = ( ζ 1 + ζ 2 s ) / 1 + s 2 \zeta(u)=(\zeta_1+\zeta_2s)/\sqrt{1+s^2} ζ ( u ) = ( ζ 1 + ζ 2 s ) / 1 + s 2 for ζ = ( ζ 1 , ζ 2 ) \zeta=(\zeta_1,\zeta_2) ζ = ( ζ 1 , ζ 2 ) . Let δ n : = e κ n λ − n \delta_n:=e^{\kappa n}\lambda^{-n} δ n := e κn λ − n and ℓ ζ : = { ( s , t ) : t = ζ 1 + ζ 2 s } \ell_\zeta:=\{(s,t):t=\zeta_1+\zeta_2s\} ℓ ζ := {( s , t ) : t = ζ 1 + ζ 2 s } . Since dist ( P g , ℓ ζ ) = ∣ ζ 1 + ζ 2 s g − t g ∣ / 1 + ζ 2 2 \operatorname{dist}(P_g,\ell_\zeta)=|\zeta_1+\zeta_2s_g-t_g|/\sqrt{1+\zeta_2^2} dist ( P g , ℓ ζ ) = ∣ ζ 1 + ζ 2 s g − t g ∣/ 1 + ζ 2 2 , we get:
if h h h belongs to the set in (11) , then ∣ ζ 1 + ζ 2 s g − t g ∣ ≤ 2 δ n + ∣ ζ 2 ∣ χ u g ′ ( x ) |\zeta_1+\zeta_2s_g-t_g|\leq\sqrt{2}\delta_n+|\zeta_2|\chi_u g'(x) ∣ ζ 1 + ζ 2 s g − t g ∣ ≤ 2 δ n + ∣ ζ 2 ∣ χ u g ′ ( x ) , hence dist ( P g , ℓ ζ ) ≤ 2 δ n + C λ − n ≤ C δ n \operatorname{dist}(P_g,\ell_\zeta)\leq\sqrt{2}\delta_n+C\lambda^{-n}\leq C\delta_n dist ( P g , ℓ ζ ) ≤ 2 δ n + C λ − n ≤ C δ n ;
if ∣ ζ 1 + ζ 2 s g − t g ∣ ≤ δ n |\zeta_1+\zeta_2s_g-t_g|\leq\delta_n ∣ ζ 1 + ζ 2 s g − t g ∣ ≤ δ n , then h h h belongs to the set in (11) (take σ = 0 \sigma=0 σ = 0 ).
Consequently, with ℓ ( r ) \ell^{(r)} ℓ ( r ) the r r r -neighbourhood of ℓ \ell ℓ ,
∑ g : ∣ ζ 1 + ζ 2 s g − t g ∣ ≤ δ n g ′ ( x ) ≤ (left side of (2.11)) ≤ C μ n x ( ℓ ζ ( C δ n ) ) . (52) \sum_{g:\,|\zeta_1+\zeta_2s_g-t_g|\leq\delta_n} g'(x)
\leq\text{(left side of (2.11))}
\leq C\mu_n^x\left(\ell_\zeta^{(C\delta_n)}\right). \tag*{(52)} g : ∣ ζ 1 + ζ 2 s g − t g ∣ ≤ δ n ∑ g ′ ( x ) ≤ (left side of (2.11)) ≤ C μ n x ( ℓ ζ ( C δ n ) ) . (52)
Step 2: ( i ) ⇒ ( i i i ) (i)\Rightarrow(iii) ( i ) ⇒ ( iii ) . Suppose τ − b ω = C 0 ∘ f − C 0 + k \tau-b\omega=C_0\circ f-C_0+k τ − bω = C 0 ∘ f − C 0 + k with C 0 ∈ C 1 C_0\in C^1 C 0 ∈ C 1 . Then τ n − b ω n = C 0 ∘ f n − C 0 + n k \tau_n-b\omega_n=C_0\circ f^n-C_0+nk τ n − b ω n = C 0 ∘ f n − C 0 + nk , so t g − b s g = ( ( τ n − b ω n ) ∘ g ) ′ ( x ) = C 0 ′ ( x ) − C 0 ′ ( g ( x ) ) g ′ ( x ) t_g-bs_g=((\tau_n-b\omega_n)\circ g)'(x)=C_0'(x)-C_0'(g(x))g'(x) t g − b s g = (( τ n − b ω n ) ∘ g ) ′ ( x ) = C 0 ′ ( x ) − C 0 ′ ( g ( x )) g ′ ( x ) and ∣ C 0 ′ ( x ) + b s g − t g ∣ ≤ C λ − n ≤ δ n |C_0'(x)+bs_g-t_g|\leq C\lambda^{-n}\leq\delta_n ∣ C 0 ′ ( x ) + b s g − t g ∣ ≤ C λ − n ≤ δ n for n n n large. Taking ζ = ( C 0 ′ ( x ) , b ) \zeta=(C_0'(x),b) ζ = ( C 0 ′ ( x ) , b ) in (A.1), the left side of (11) is at least ∑ g g ′ ( x ) = L f n 1 ( x ) ≥ C − 1 \sum_g g'(x)=\mathcal{L}_f^n1(x)\geq C^{-1} ∑ g g ′ ( x ) = L f n 1 ( x ) ≥ C − 1 , so (B1) fails for every κ > 0 \kappa>0 κ > 0 .
Step 3: the limit measures. For g ∈ H f n ( x ) g\in\mathcal{H}_{f}^{n}(x) g ∈ H f n ( x ) and g ~ ∈ H f m ( g ( x ) ) \widetilde{g}\in\mathcal{H}_{f}^{m}(g(x)) g ∈ H f m ( g ( x )) we have g ~ ∘ g ∈ H f n + m ( x ) \widetilde{g}\circ g\in\mathcal{H}_{f}^{n+m}(x) g ∘ g ∈ H f n + m ( x ) and, since Φ n + m ∘ g ~ ∘ g = Φ m ∘ g ~ ∘ g + Φ n ∘ g \Phi_{n+m}\circ\widetilde{g}\circ g=\Phi_m\circ\widetilde{g}\circ g+\Phi_n\circ g Φ n + m ∘ g ∘ g = Φ m ∘ g ∘ g + Φ n ∘ g ,
P g ~ ∘ g ( x ) = P g ( x ) + g ′ ( x ) P g ~ ( g ( x ) ) , p g ~ ∘ g ( x ) = p g ( x ) p g ~ ( g ( x ) ) . (53) P_{\widetilde{g}\circ g}(x)=P_g(x)+g'(x)P_{\widetilde{g}}(g(x)),\qquad
p_{\widetilde{g}\circ g}(x)=p_g(x)p_{\widetilde{g}}(g(x)). \tag*{(53)} P g ∘ g ( x ) = P g ( x ) + g ′ ( x ) P g ( g ( x )) , p g ∘ g ( x ) = p g ( x ) p g ( g ( x )) . (53)
Moreover ∣ P g ( x ) ∣ ≤ ∑ k = 1 n ∥ Φ ′ ∥ ∞ sup ∣ ( f k ) ′ ∣ − 1 ≤ C 0 |P_g(x)|\leq\sum_{k=1}^{n}\|\Phi'\|_\infty\sup|(f^k)'|^{-1}\leq C_0 ∣ P g ( x ) ∣ ≤ ∑ k = 1 n ∥ Φ ′ ∥ ∞ sup ∣ ( f k ) ′ ∣ − 1 ≤ C 0 . Hence
μ n + m x = ∑ g ∈ H f n ( x ) p g ( x ) ( S g ) ∗ μ m g ( x ) , \mu_{n+m}^x=\sum_{g\in\mathcal{H}_{f}^{n}(x)}p_g(x)(S_g)_*\mu_m^{g(x)}, μ n + m x = g ∈ H f n ( x ) ∑ p g ( x ) ( S g ) ∗ μ m g ( x ) ,
where S g ( P ) : = P g ( x ) + g ′ ( x ) P S_g(P):=P_g(x)+g'(x)P S g ( P ) := P g ( x ) + g ′ ( x ) P , and each ( S g ) ∗ μ m g ( x ) (S_g)_*\mu_m^{g(x)} ( S g ) ∗ μ m g ( x ) is supported in the ball of radius C 0 g ′ ( x ) ≤ C λ − n C_0g'(x)\leq C\lambda^{-n} C 0 g ′ ( x ) ≤ C λ − n around P g ( x ) P_g(x) P g ( x ) . Thus ( μ n x ) n (\mu_n^x)_n ( μ n x ) n is Cauchy in the Wasserstein distance, uniformly in x x x , and converges to a probability measure μ x \mu^x μ x , depending continuously on x x x , with
μ x = ∑ g ∈ H f n ( x ) p g ( x ) ( S g ) ∗ μ g ( x ) , μ n x ( A ) ≤ μ x ( A ( C λ − n ) ) (54) \mu^x=\sum_{g\in\mathcal{H}_{f}^{n}(x)}p_g(x)(S_g)_*\mu^{g(x)},\qquad
\mu_n^x(A)\leq\mu^x\left(A^{(C\lambda^{-n})}\right) \tag*{(54)} μ x = g ∈ H f n ( x ) ∑ p g ( x ) ( S g ) ∗ μ g ( x ) , μ n x ( A ) ≤ μ x ( A ( C λ − n ) ) (54)
for all n n n and Borel A ⊂ R 2 A\subset\mathbb{R}^2 A ⊂ R 2 .
Step 4: no line carries μ x \mu^x μ x . We claim that, under (iii) and the assumption on ω \omega ω , there is no x ∈ T x\in\mathbb{T} x ∈ T and no line ℓ ⊂ R 2 \ell\subset\mathbb{R}^2 ℓ ⊂ R 2 with μ x ( ℓ ) = 1 \mu^x(\ell)=1 μ x ( ℓ ) = 1 . Suppose supp μ x 0 ⊂ { α s + β t = c 0 } \operatorname{supp}\mu^{x_0}\subset\{\alpha s+\beta t=c_0\} supp μ x 0 ⊂ { α s + β t = c 0 } , with α 2 + β 2 = 1 \alpha^2+\beta^2=1 α 2 + β 2 = 1 . For y ∈ T y\in\mathbb{T} y ∈ T let c ( y ) : = ∫ ( α s + β t ) d μ y c(y):=\int(\alpha s+\beta t)\,d\mu^y c ( y ) := ∫ ( α s + β t ) d μ y and V ( y ) : = ∫ ( α s + β t − c ( y ) ) 2 d μ y V(y):=\int(\alpha s+\beta t-c(y))^2\,d\mu^y V ( y ) := ∫ ( α s + β t − c ( y ) ) 2 d μ y ; both are continuous and V ( x 0 ) = 0 V(x_0)=0 V ( x 0 ) = 0 . By (A.3) with n = 1 n=1 n = 1 , and since p g > 0 p_g>0 p g > 0 and S g S_g S g is a homothety, V ( x 0 ) = 0 V(x_0)=0 V ( x 0 ) = 0 implies V ( z ) = 0 V(z)=0 V ( z ) = 0 for every z ∈ f − 1 ( x 0 ) z\in f^{-1}(x_0) z ∈ f − 1 ( x 0 ) ; by induction V = 0 V=0 V = 0 on ⋃ n f − n ( x 0 ) \bigcup_n f^{-n}(x_0) ⋃ n f − n ( x 0 ) , which is dense, hence V ≡ 0 V\equiv0 V ≡ 0 : for every y y y , supp μ y ⊂ { α s + β t = c ( y ) } \operatorname{supp}\mu^y\subset\{\alpha s+\beta t=c(y)\} supp μ y ⊂ { α s + β t = c ( y )} . Now let z ∈ T z\in\mathbb{T} z ∈ T , y = f ( z ) y=f(z) y = f ( z ) , and let g ∈ H f 1 ( y ) g\in\mathcal{H}_f^1(y) g ∈ H f 1 ( y ) with g ( y ) = z g(y)=z g ( y ) = z , so that P g ( y ) = Φ ′ ( z ) / f ′ ( z ) P_g(y)=\Phi'(z)/f'(z) P g ( y ) = Φ ′ ( z ) / f ′ ( z ) and g ′ ( y ) = 1 / f ′ ( z ) g'(y)=1/f'(z) g ′ ( y ) = 1/ f ′ ( z ) . For P = ( s , t ) ∈ supp μ z P=(s,t)\in\operatorname{supp}\mu^z P = ( s , t ) ∈ supp μ z we have S g P ∈ supp μ y S_gP\in\operatorname{supp}\mu^y S g P ∈ supp μ y , i.e.
α ( ω ′ ( z ) + s ) + β ( τ ′ ( z ) + t ) f ′ ( z ) = c ( f ( z ) ) , α s + β t = c ( z ) , \frac{\alpha(\omega'(z)+s)+\beta(\tau'(z)+t)}{f'(z)}=c(f(z)),\qquad
\alpha s+\beta t=c(z), f ′ ( z ) α ( ω ′ ( z ) + s ) + β ( τ ′ ( z ) + t ) = c ( f ( z )) , α s + β t = c ( z ) ,
so that α ω ′ + β τ ′ = ( c ∘ f ) f ′ − c \alpha\omega' + \beta\tau' = (c \circ f)f' - c α ω ′ + β τ ′ = ( c ∘ f ) f ′ − c on T \mathbb{T} T . Let f ~ \widetilde f f be a lift of f f f and C ~ \widetilde C C a primitive of c c c on R \mathbb{R} R ; then α ω + β τ = C ~ ∘ f ~ − C ~ + k \alpha\omega+ \beta\tau= \widetilde C \circ\widetilde f - \widetilde C + k α ω + β τ = C ∘ f − C + k for some k ∈ R k \in\mathbb{R} k ∈ R . Since C ~ ( x + 1 ) = C ~ ( x ) + ∫ T c \widetilde C(x+1)=\widetilde C(x)+\int_{\mathbb{T}}c C ( x + 1 ) = C ( x ) + ∫ T c and f ~ ( x + 1 ) = f ~ ( x ) + d \widetilde f(x+1)=\widetilde f(x)+d f ( x + 1 ) = f ( x ) + d , periodicity of the left side forces ( d − 1 ) ∫ T c = 0 (d-1)\int_{\mathbb{T}}c=0 ( d − 1 ) ∫ T c = 0 , i.e. C ~ \widetilde C C is 1-periodic. Hence α ω + β τ \alpha\omega+\beta\tau α ω + β τ is cohomologous to a constant. If β = 0 \beta=0 β = 0 this contradicts the assumption on ω \omega ω ; if β ≠ 0 \beta\ne0 β = 0 it contradicts (iii) with b = − α / β b=-\alpha/\beta b = − α / β .
Step 5. Let ϑ ( δ ) : = sup x ∈ T sup ℓ μ x ( ℓ ( δ ) ) \vartheta(\delta):=\sup_{x\in\mathbb{T}}\sup_{\ell}\mu^x(\ell^{(\delta)}) ϑ ( δ ) := sup x ∈ T sup ℓ μ x ( ℓ ( δ ) ) , the supremum over all lines ℓ ⊂ R 2 \ell\subset\mathbb{R}^2 ℓ ⊂ R 2 . We show that, under (iii), there are C , β 0 > 0 C,\beta_0>0 C , β 0 > 0 with
ϑ ( δ ) ≤ C δ β 0 ∀ δ ∈ ( 0 , 1 ] . (55) \vartheta(\delta)\le C\delta^{\beta_0}\qquad\forall\delta\in(0,1]. \tag*{(55)} ϑ ( δ ) ≤ C δ β 0 ∀ δ ∈ ( 0 , 1 ] . (55)
First, there exist N ∈ N N\in\mathbb{N} N ∈ N and ε > 0 \varepsilon>0 ε > 0 such that μ x ( ℓ ( 3 C λ − N ) ) ≤ 1 − ε \mu^x(\ell^{(3C\lambda^{-N})})\le1-\varepsilon μ x ( ℓ ( 3 C λ − N ) ) ≤ 1 − ε for all x x x and ℓ \ell ℓ . Otherwise there are N k → ∞ N_k\to\infty N k → ∞ , x k ∈ T x_k\in\mathbb{T} x k ∈ T and lines ℓ k \ell_k ℓ k with μ x k ( ℓ k ( 3 C λ − N k ) ) ≥ 1 − 1 k \mu^{x_k}(\ell_k^{(3C\lambda^{-N_k})})\ge1-\frac{1}{k} μ x k ( ℓ k ( 3 C λ − N k ) ) ≥ 1 − k 1 ; since all μ x \mu^x μ x are supported in { ∣ P ∣ ≤ C 0 } \{|P|\le C_0\} { ∣ P ∣ ≤ C 0 } , the lines ℓ k \ell_k ℓ k meet a fixed ball, and along a subsequence x k → x x_k\to x x k → x , ℓ k → ℓ \ell_k\to\ell ℓ k → ℓ . By weak continuity, μ x ( ℓ ( η ) ) = 1 \mu^x(\ell^{(\eta)})=1 μ x ( ℓ ( η ) ) = 1 for every η > 0 \eta>0 η > 0 , i.e. μ x ( ℓ ) = 1 \mu^x(\ell)=1 μ x ( ℓ ) = 1 , contradicting Step 4. Next, fix δ ≤ C λ − N \delta\le C\lambda^{-N} δ ≤ C λ − N , x x x and ℓ \ell ℓ . By (A.3) with n = N n=N n = N ,
μ x ( ℓ ( δ ) ) = ∑ g ∈ H N f ( x ) p g ( x ) μ g ( x ) ( S g − 1 ℓ ( δ ) ) . \mu^x(\ell^{(\delta)})=\sum_{g\in\mathcal{H}_N^f(x)}p_g(x)\mu^{g(x)}\left(S_g^{-1}\ell^{(\delta)}\right). μ x ( ℓ ( δ ) ) = g ∈ H N f ( x ) ∑ p g ( x ) μ g ( x ) ( S g − 1 ℓ ( δ ) ) .
Since ( S g ) ∗ μ g ( x ) (S_g)_*\mu^{g(x)} ( S g ) ∗ μ g ( x ) is supported in B ( P g , C λ − N ) B(P_g,C\lambda^{-N}) B ( P g , C λ − N ) , only the g g g with dist ( P g , ℓ ) ≤ 2 C λ − N \operatorname{dist}(P_g,\ell)\le2C\lambda^{-N} dist ( P g , ℓ ) ≤ 2 C λ − N contribute; for these the ball B ( P g , C λ − N ) B(P_g,C\lambda^{-N}) B ( P g , C λ − N ) lies in ℓ ( 3 C λ − N ) \ell^{(3C\lambda^{-N})} ℓ ( 3 C λ − N ) , so their total weight is at most μ x ( ℓ ( 3 C λ − N ) ) ≤ 1 − ε \mu^x(\ell^{(3C\lambda^{-N})})\le1-\varepsilon μ x ( ℓ ( 3 C λ − N ) ) ≤ 1 − ε . Moreover S g − 1 ℓ ( δ ) = ( ℓ ′ ) ( δ / g ′ ( x ) ) S_g^{-1}\ell^{(\delta)}=(\ell')^{(\delta/g'(x))} S g − 1 ℓ ( δ ) = ( ℓ ′ ) ( δ / g ′ ( x )) for a line ℓ ′ \ell' ℓ ′ parallel to ℓ \ell ℓ , and δ / g ′ ( x ) ≤ C Λ N δ = : K δ \delta/g'(x)\le C\Lambda^N\delta=:K\delta δ / g ′ ( x ) ≤ C Λ N δ =: K δ . Hence ϑ ( δ ) ≤ ( 1 − ε ) ϑ ( K δ ) \vartheta(\delta)\le(1-\varepsilon)\vartheta(K\delta) ϑ ( δ ) ≤ ( 1 − ε ) ϑ ( K δ ) for δ ≤ C λ − N \delta\le C\lambda^{-N} δ ≤ C λ − N , and iterating gives (A.4) with β 0 = log ( 1 − ε ) − 1 / log K \beta_0=\log(1-\varepsilon)^{-1}/\log K β 0 = log ( 1 − ε ) − 1 / log K .
Step 6: (iii)⇒ \Rightarrow ⇒ (ii), and (T). Let κ ∈ ( 0 , log λ ) \kappa\in(0,\log\lambda) κ ∈ ( 0 , log λ ) . By (A.1), (A.3) and (A.4), the left side of (11) is bounded by
C μ n x ( ℓ ( C δ n ) ) ≤ C μ x ( ℓ ( C δ n + C λ − n ) ) ≤ C ϑ ( C δ n ) ≤ C e − β 0 ( log λ − κ ) n , C\mu^x_n(\ell^{(C\delta^n)})\le C\mu^x(\ell^{(C\delta^n+C\lambda^{-n})})\le C\vartheta(C\delta^n)\le Ce^{-\beta_0(\log\lambda-\kappa)n}, C μ n x ( ℓ ( C δ n ) ) ≤ C μ x ( ℓ ( C δ n + C λ − n ) ) ≤ C ϑ ( C δ n ) ≤ C e − β 0 ( l o g λ − κ ) n ,
which is (B1) with γ 1 = β 0 ( log λ − κ ) \gamma_1=\beta_0(\log\lambda-\kappa) γ 1 = β 0 ( log λ − κ ) . For (T): if two branches h , h ~ h,\widetilde h h , h at q q q are not transversal, the slope intervals [ s g ± χ u g ′ ] [s_g\pm\chi u_g'] [ s g ± χ u g ′ ] and [ s g ~ ± χ u g ~ ′ ] [s_{\widetilde g}\pm\chi u_{\widetilde g}'] [ s g ± χ u g ′ ] intersect, so ∣ s g − s g ~ ∣ ≤ C λ − n |s_g-s_{\widetilde g}|\le C\lambda^{-n} ∣ s g − s g ∣ ≤ C λ − n ; hence N F ( n ) ≤ C sup x sup s 0 μ n x ( { ∣ s − s 0 ∣ ≤ C λ − n } ) ≤ C ϑ ( C λ − n ) ≤ C λ − β 0 n \mathcal{N}_F(n)\le C\sup_x\sup_{s_0}\mu^x_n(\{|s-s_0|\le C\lambda^{-n}\})\le C\vartheta(C\lambda^{-n})\le C\lambda^{-\beta_0 n} N F ( n ) ≤ C sup x sup s 0 μ n x ({ ∣ s − s 0 ∣ ≤ C λ − n }) ≤ C ϑ ( C λ − n ) ≤ C λ − β 0 n .
Since (ii)⇒ \Rightarrow ⇒ (i) is trivial, the proof is complete. □
We conclude with a simple example satisfying all our assumptions.
Example A.2. Let f ( x ) = ℓ x f(x)=\ell x f ( x ) = ℓ x with ℓ ≥ 2 \ell\ge2 ℓ ≥ 2 an integer, ω ( x ) = a cos ( 2 π x ) \omega(x)=a\cos(2\pi x) ω ( x ) = a cos ( 2 π x ) with a ≠ 0 a\ne0 a = 0 , and τ ( x ) = sin ( 2 π x ) \tau(x)=\sin(2\pi x) τ ( x ) = sin ( 2 π x ) , where ℓ > 2 ( 1 + 2 π ∣ a ∣ ) \ell>2(1+2\pi|a|) ℓ > 2 ( 1 + 2 π ∣ a ∣ ) , so that (A5) holds, and ℓ \ell ℓ is large enough for F F F to be an SVPH.
Cohomology. If ϕ = C ∘ f − C + k \phi=C\circ f-C+k ϕ = C ∘ f − C + k with C ∈ L 2 ( T ) C\in L^2(\mathbb{T}) C ∈ L 2 ( T ) , then ϕ ^ ( m ) = C ^ ( m / ℓ ) 1 ℓ ∣ m − C ^ ( m ) \widehat\phi(m)=\widehat C(m/\ell)\mathbf{1}_{\ell\mid m}-\widehat C(m) ϕ ( m ) = C ( m / ℓ ) 1 ℓ ∣ m − C ( m ) for m ≠ 0 m\ne0 m = 0 , and therefore ∑ j ≥ 0 ϕ ^ ( ℓ j ) = − lim J → ∞ C ^ ( ℓ J ) = 0 \sum_{j\ge0}\widehat\phi(\ell^j)=-\lim_{J\to\infty}\widehat C(\ell^J)=0 ∑ j ≥ 0 ϕ ( ℓ j ) = − lim J → ∞ C ( ℓ J ) = 0 . For ϕ = τ − b ω \phi=\tau-b\omega ϕ = τ − bω this sum equals ϕ ^ ( 1 ) = 1 2 i − a b 2 ≠ 0 \widehat\phi(1)=\frac{1}{2i}-\frac{ab}{2}\ne0 ϕ ( 1 ) = 2 i 1 − 2 ab = 0 for every b ∈ R b\in\mathbb{R} b ∈ R , and for ϕ = ω \phi=\omega ϕ = ω it equals a 2 ≠ 0 \frac{a}{2}\ne0 2 a = 0 . Hence ω \omega ω is not cohomologous to a constant and (iii) of Lemma A.1 holds, so τ \tau τ satisfies (B1).
Multiplicative cohomology. Let φ ∈ C ∞ ( T , R ) \varphi\in C^\infty(\mathbb{T},\mathbb{R}) φ ∈ C ∞ ( T , R ) and suppose that e i φ = c ψ ∘ f / ψ e^{i\varphi}=c\psi\circ f/\psi e i φ = c ψ ∘ f / ψ Lebesgue-a.e., with c ∈ S 1 c\in S^1 c ∈ S 1 and ψ : T → S 1 \psi:\mathbb{T}\to S^1 ψ : T → S 1 measurable. By Livšic regularity for expanding maps ψ \psi ψ is continuous, so ψ = e i Ψ \psi=e^{i\Psi} ψ = e i Ψ with Ψ \Psi Ψ continuous on R \mathbb{R} R and Ψ ( x + 1 ) = Ψ ( x ) + 2 π p \Psi(x+1)=\Psi(x)+2\pi p Ψ ( x + 1 ) = Ψ ( x ) + 2 π p , p ∈ Z p\in\mathbb{Z} p ∈ Z . Then φ − Ψ ∘ f + Ψ \varphi-\Psi\circ f+\Psi φ − Ψ ∘ f + Ψ is continuous with values in a coset of 2 π Z 2\pi\mathbb{Z} 2 π Z , hence constant; since Ψ ∘ f − Ψ \Psi\circ f-\Psi Ψ ∘ f − Ψ changes by 2 π p ( ℓ − 1 ) 2\pi p(\ell-1) 2 π p ( ℓ − 1 ) under x ↦ x + 1 x\mapsto x+1 x ↦ x + 1 and φ \varphi φ is periodic, p = 0 p=0 p = 0 . Hence φ \varphi φ is cohomologous to a constant.
Uniqueness of the acip and (A6). Let g ∈ L 2 ( T 2 ) g\in L^2(\mathbb{T}^2) g ∈ L 2 ( T 2 ) satisfy g ∘ F = e i α g g\circ F=e^{i\alpha}g g ∘ F = e i α g , and expand g ( x , θ ) = ∑ m g m ( x ) e 2 π i m θ g(x,\theta)=\sum_m g_m(x)e^{2\pi im\theta} g ( x , θ ) = ∑ m g m ( x ) e 2 π im θ . Then g m ∘ f e 2 π i m ω = e i α g m g_m\circ f\,e^{2\pi im\omega}=e^{i\alpha}g_m g m ∘ f e 2 π imω = e i α g m for each m m m . If g m ≠ 0 g_m\ne0 g m = 0 , then ∣ g m ∣ |g_m| ∣ g m ∣ is f f f -invariant, hence constant, and by the previous step 2 π m ω 2\pi m\omega 2 π mω is cohomologous to a constant; thus m = 0 m=0 m = 0 . So g = g 0 ( x ) g=g_0(x) g = g 0 ( x ) with g 0 ∘ f = e i α g 0 g_0 \circ f = e^{i\alpha}g_0 g 0 ∘ f = e i α g 0 , and since f f f is mixing, α = 0 \alpha= 0 α = 0 and g g g is constant. Therefore Lebesgue measure on T 2 \mathbb{T}^2 T 2 is ergodic and weakly mixing for F F F ; being equivalent to Lebesgue, it is the unique acip of F F F , and (A6) holds.
Condition (B2). Suppose e i ν τ = c ϕ ∘ F / ϕ e^{i\nu\tau} = c\phi\circ F/\phi e i ν τ = c ϕ ∘ F / ϕ Lebesgue-a.e., with ν ≠ 0 \nu\ne0 ν = 0 and ϕ \phi ϕ measurable, and expand ϕ ( x , θ ) = ∑ m ϕ m ( x ) e 2 π i m θ \phi(x,\theta) = \sum_m \phi_m(x)e^{2\pi i m\theta} ϕ ( x , θ ) = ∑ m ϕ m ( x ) e 2 π im θ . Comparing the Fourier coefficients in θ \theta θ gives e i ν τ ϕ m = c e 2 π i m ω ϕ m ∘ f e^{i\nu\tau}\phi_m = ce^{2\pi i m\omega}\phi_m \circ f e i ν τ ϕ m = c e 2 π imω ϕ m ∘ f for each m m m . Choose m m m with ϕ m ≠ 0 \phi_m \ne0 ϕ m = 0 ; then ∣ ϕ m ∣ |\phi_m| ∣ ϕ m ∣ is constant and e i ( ν τ − 2 π m ω ) = c ψ ∘ f / ψ e^{i(\nu\tau-2\pi m\omega)} = c\psi\circ f/\psi e i ( ν τ − 2 π mω ) = c ψ ∘ f / ψ with ∣ ψ ∣ = 1 |\psi| = 1 ∣ ψ ∣ = 1 . Hence ν τ − 2 π m ω \nu\tau- 2\pi m\omega ν τ − 2 π mω is cohomologous to a constant, which contradicts (iii) of Lemma A.1 with b = 2 π m / ν b = 2\pi m/\nu b = 2 π m / ν .
Virtual expansion. We also verify the VE assumption. Here J h = ℓ − n J_h = \ell^{-n} J h = ℓ − n and
( D h ( q ) F n ) ∗ ( a , b ) = ( ℓ n ( a + s g b ) , b ) . (D_{h(q)}F^n)^*(a,b) = (\ell^n(a+s_gb),b). ( D h ( q ) F n ) ∗ ( a , b ) = ( ℓ n ( a + s g b ) , b ) .
*The slopes s g s_g s g are uniformly bounded. By the vertical-line case of (A.4) and (A.3), for every unit covector ( a , b ) (a,b) ( a , b ) the total weight of branches satisfying ∣ a + s g b ∣ < ℓ − n / 2 |a+s_gb| < \ell^{-n/2} ∣ a + s g b ∣ < ℓ − n /2 is at most C ℓ − β 0 n / 2 C\ell^{-\beta_0n/2} C ℓ − β 0 n /2 . Indeed, if ∣ b ∣ |b| ∣ b ∣ is sufficiently small there are no such branches for large n n n ; otherwise this is an interval condition on s g s_g s g of width at most C ℓ − n / 2 C\ell^{-n/2} C ℓ − n /2 . The inverse derivatives of this skew product are uniformly bounded. On the other branches, the above covector has norm at least ℓ n / 2 \ell^{n/2} ℓ n /2 . Thus, for every fixed s > 0 s > 0 s > 0 ,
B s ( F n ) ≤ C s ( ℓ − β 0 n / 2 + ℓ − s n / 2 ) . B_s(F^n) \le C_s(\ell^{-\beta_0 n/2}+\ell^{-sn/2}). B s ( F n ) ≤ C s ( ℓ − β 0 n /2 + ℓ − s n /2 ) .
The constants λ c + > 1 \lambda_c^+ > 1 λ c + > 1 and λ c − < 1 \lambda_c^- < 1 λ c − < 1 in (A2) may be chosen arbitrarily close to 1, since the central inverse expansion is bounded above and below uniformly in n n n . Choosing them so that ( λ c + ) s ℓ − min { β 0 , s } / 2 < 1 (\lambda_c^+)^s\ell^{-\min\{\beta_0,s\}/2} < 1 ( λ c + ) s ℓ − m i n { β 0 , s } /2 < 1 proves (2.10). Finally, (2.8) holds with Θ = 1 \Theta= 1 Θ = 1 , so (P) is satisfied. Theorem 2.10 and Theorems A–B therefore apply; centring holds because ∫ τ d Leb = 0 \int\tau\,d\operatorname{Leb} = 0 ∫ τ d Leb = 0 .
Appendix B. Useful estimates for SVPH
We collect some consequences of results from [4 ] for the SVPH F F F and its transfer operator L F \mathcal{L}_F L F . The estimates recalled below require only the domination condition (A4). The stronger pinching imposed in [4 ] is used for higher derivative estimates, which are not needed here.
By [4 Prop. 5.6] , there exists a constant C ∗ ≥ 1 C_* \ge1 C ∗ ≥ 1 such that for every z ∈ T 2 z \in\mathbb{T}^2 z ∈ T 2 , any n ∈ N n \in\mathbb{N} n ∈ N , and any vectors v u ∈ C u v^u \in C_u v u ∈ C u , v c ∈ C c v^c \in C_c v c ∈ C c with ( a , b ) : = D z F n v u ∉ C u (a,b) := D_zF^n v^u \notin C_u ( a , b ) := D z F n v u ∈ / C u ,
C ∗ − 1 ∥ D z F n v u ∥ ∥ v u ∥ ∣ b ∣ ∥ v c ∥ ≤ ∣ det D z F n ∣ ≤ C ∗ ∥ D z F n v u ∥ ∥ v u ∥ ∣ b ∣ ∥ v c ∥ . (56) C_*^{-1}\frac{\lVert D_zF^n v^u\rVert}{\lVert v^u\rVert}\frac{|b|}{\lVert v^c\rVert}
\leq|\det D_zF^n|
\leq C_*\frac{\lVert D_zF^n v^u\rVert}{\lVert v^u\rVert}\frac{|b|}{\lVert v^c\rVert}. \tag*{(56)} C ∗ − 1 ∥ v u ∥ ∥ D z F n v u ∥ ∥ v c ∥ ∣ b ∣ ≤ ∣ det D z F n ∣ ≤ C ∗ ∥ v u ∥ ∥ D z F n v u ∥ ∥ v c ∥ ∣ b ∣ . (56)
In particular, for each n ∈ N n \in\mathbb{N} n ∈ N and z ∈ T 2 z \in\mathbb{T}^2 z ∈ T 2 ,
C ∗ − 1 λ c − n λ n ≤ ∣ det D z F n ∣ ≤ C ∗ Λ n λ c n . (57) C_*^{-1}\lambda_c^{-n}\lambda^n \leq|\det D_zF^n| \leq C_*\Lambda^n\lambda_c^n. \tag*{(57)} C ∗ − 1 λ c − n λ n ≤ ∣ det D z F n ∣ ≤ C ∗ Λ n λ c n . (57)
∥ ( D F n ) − 1 ∥ C 0 ( T 2 ) ≤ C λ c n . (58) \lVert(DF^n)^{-1}\rVert_{C^0(\mathbb{T}^2)} \leq C\lambda_c^n. \tag*{(58)} ∥( D F n ) − 1 ∥ C 0 ( T 2 ) ≤ C λ c n . (58)
∥ L F n 1 ∥ C 0 ( T 2 ) ≤ C λ c n , ∥ L F n 1 ∥ C 1 ( T 2 ) ≤ C λ c 2 n . (59) \lVert\mathcal{L}_F^n1\rVert_{C^0(\mathbb{T}^2)} \leq C\lambda_c^n,\qquad\lVert\mathcal{L}_F^n1\rVert_{C^1(\mathbb{T}^2)} \leq C\lambda_c^{2n}. \tag*{(59)} ∥ L F n 1 ∥ C 0 ( T 2 ) ≤ C λ c n , ∥ L F n 1 ∥ C 1 ( T 2 ) ≤ C λ c 2 n . (59)
References