COMMUNICATIONS ON doi:10.3934/cpaa.2020240 PURE AND APPLIED ANALYSIS Volume 19, Number 12, December 2020 pp. 5305–5335
LARGE DEVIATION THEOREMS FOR DIRICHLET DETERMINANTS OF ANALYTIC QUASI-PERIODIC JACOBI OPERATORS WITH BRJUNO-RUSSMANN¨ FREQUENCY
Wenmeng Geng and Kai Tao∗ College of Sciences, Hohai University, No.1 Xikang Road Nanjing, Jiangsu, 210098, China
(Communicated by Enrico Valdinoci)
Abstract. In this paper, we first study the strong Birkhoff Ergodic Theorem for subharmonic functions with the Brjuno-R¨ussmannshift on the Torus. Then, we apply it to prove the large deviation theorems for the finite scale Dirichlet determinants of quasi-periodic analytic Jacobi operators with this frequency. It shows that the Brjuno-R¨ussmannfunction, which reflects the irrationality of the frequency, plays the key role in these theorems via the smallest deviation. At last, as an application, we obtain a distribution of the eigenvalues of the Jacobi operators with Dirichlet boundary conditions, which also depends on the smallest deviation, essentially on the irrationality of the frequency.
1. Introduction. We study the following quasi-periodic analytic Jacobi operators H(x, ω) on l2(Z): [H(x, ω)φ](n) = − a(x + (n + 1)ω)φ(n + 1) − a(x + nω)φ(n − 1) + v(x + nω)φ(n), n ∈ Z, (1.1) where v : T → R is a real analytic function called potential, a : T → C is a complex analytic function and not identically zero. The characteristic equations H(x, ω)φ = Eφ can be expressed as φ(n + 1) φ(n) 1 v(x + nω) − E −a(x + nω) φ(n) = . (1.2) a(x + (n + 1)ω) a(x + (n + 1)ω) 0 φ(n − 1) Define 1 v(x) − E −a(x) M(x, E, ω) := (1.3) a(x + ω) a(x + ω) 0 and call a map (ω, M):(x,~v) 7→ (x + ω, M(x)~v)
2020 Mathematics Subject Classification. 37C55, 37F10, 37C40. Key words and phrases. Large deviation theorems; Jacobi operators; finite scale Dirichlet de- terminants; Brjuno-R¨ussmannfrequency; strong Birkhoff ergodic theorem. The second author was supported by the Fundamental Research Funds for the Central Univer- sities(Grant B200202004) and China Postdoctoral Science Foundation (Grant 2019M650094). ∗ Corresponding author.
5305 5306 WENMENG GENG AND KAI TAO a Jacobi cocycle. Due to the fact that an analytic function only has finite zeros, M(x, E, ω) and the n-step transfer matrix 1 Y Mn(x, E, ω) := M(x + kω, E) k=n make sense almost everywhere. By the Kingman’s subadditive ergodic theorem, the Lyapunov exponent
L(E, ω) = lim Ln(E, ω) = inf Ln(E, ω) ≥ 0 (1.4) n→∞ n→∞ always exists, where 1 Z L (E, ω) = log kM (x, E, ω)kdx. n n n T Let H[m,n](x, ω) be the Jaocbi operator defined by (1.1) on a finite interval [m, n] a with Dirichlet boundary conditions, φ(m−1)=0 and φ(n+1)=0. Let f[m,n](x, E, ω) = det(H[m,n](x, ω) − E) be its characteristic polynomial. One has a a f[m,n](x, E, ω) = fn−m+1 x + (m − 1)ω, E, ω , (1.5) where a fn (x, E, ω) = det Hn(x, ω) − E v x + ω − E −a x + 2ω 0 ··· 0
−a x + 2ω v x + 2ω − E −a x + 3ω ··· 0
= ...... 0 0 · · · −a x + nω v x + nω − E
a In this paper, the aim is to study the properties of fn (x, E, ω). To state our conclusions, we first make some introductions to the background of our topic. The operator (1.1) has the following important special case, which is called the Schr¨odingeroperator and has been studied extensively: s [H (x, ω)φ](n) = φ(n + 1) + φ(n − 1) + v(x + nω)φ(n), n ∈ Z. (1.6) s s s s Then, Mn(x, E, ω), L (E, ω), Ln(E, ω) and fn(x, E, ω) have the similar definitions. In [6], Bourgain and Goldstein proved that if Ls(E, ω) > 0, then for almost all ω, the operator Hs(0, ω) has Anderson Localization, which means that it has pure- point spectrum with exponentially decaying eigenfunction. In [11], Goldstein and Schlag obtained the H¨oldercontinuity of Ls(E, ω) in E with the strong Diophantine ω, i.e. for some α > 1 and any integer n, C knωk > ω , (1.7) |n| (log |n| + 1)α where kxk = min |x + n|. n∈Z It is well known that for a fixed α > 1, almost every irrational ω satisfies (1.7). Obviously, if we define the Diophantine number as C knωk > ω , (1.8) |n|α LDT DIRICHLET DETERMINANTS WITH BRJUNO-RUSSMANN¨ FREQUENCY 5307 then it also has a full measure. In these two references, the key lemmas are the s following so-called large deviation theorems (LDTs for short) for matrix Mn(x, E, ω) with these two types of frequencies: for the Diophantine ω, it was proved in [6] that there exists 0 < σ < 1 such that 1 s s −σ σ mes x : log kM (x, E, ω)k − L (E, ω) > n < exp (−n ) ; (1.9) n n n
(n) (log n)A for the strong Diophantine ω, it was proved in [11] that there exists δ0 = n (n) such that for any δ > δ0 1 s s 2 mes x : log kM (x, E, ω)k − L (E, ω) > δ < exp −cδ n . (1.10) n n n (n) Here δ0 is called the smallest deviation in the LDT and very important in our paper. Compared with the Schr¨odingercocycle, one of the distinguishing features of the Jacobi cocycle is that it is not SL(2, C). Then Jitomirskaya, Koslover and Schulteis [17], and Jitomirskaya and Marx [18] proved that the LDT (1.9) for Mn(x, E, ω) and the weak H¨oldercontinuity of the Lyapunov exponent of the analytic GL(2, C) co- cycles hold with the Diophantine frequency. In [20], we showed that (1.10) can hold for Mn(x, E, ω) with the strong Diophantine ω and the continuity of the Lyapunov exponent of the Jacobi cocycles L(E, ω) can be H¨olderin E. For any irrational ω, there exist its continued fraction approximates { ps }∞ , qs s=1 satisfying 1 p 1 < |ω − s | < . (1.11) qs(qs+1 + qs) qs qsqs+1 Define β as the exponential growth exponent of { ps }∞ as follows: qs s=1 log q β(ω) := lim sup s+1 ∈ [0, ∞]. s qs Obviously, both the sets of the strong Diophantine frequency and the Diophantine one are the subsets of {ω : β(ω) = 0}. We say ω is the Liouville number, if β(ω) ≥ 0. Recently, more and more attentions are paid to the question that what will happen to these operators with more generic ω, such as the one satisfying β(ω) = 0, the finite Liouville one satisfying 0 ≤ β(ω) < ∞ and the irrational one. So far, the most striking answers are mainly for the almost Mathieu operators (AMO for short), which is also a special case of the Jacobi ones m [H (x, ω, λ)φ](n) = φ(n + 1) + φ(n − 1) + 2λ cos (2π(x + nω)) φ(n). n ∈ Z. (1.12) The most famous one, the Ten Martini Problem, which was dubbed by Barry Simon and conjectures that for any irrational ω, the spectrum of AMO is a Cantor set, was completely solved by Avila and Jitomirskaya [1]. In that reference, they also proved that Hm(x, ω, λ) has Anderson Localization for almost every x ∈ T with 16 β β λ > e 9 . In [4], Avila, You and Zhou improved it to λ > e . While, the answers for the Schr¨odingeror Jacobi operators in the positive Lya- punov exponent regimes are mainly in the study of the continuity of the Lyapunov exponent. In [5], they proved that the Lyapunov exponent is continuous in E for any irrational ω. The first result that the H¨oldercontinuity holds for some weak Liouville frequency, which means that β(ω) < c, where c is a small constant de- pending only on the analytic potential v(x), is [22]. Recently, Han and Zhang [15] 5308 WENMENG GENG AND KAI TAO ameliorated it to λ > eCβ in the large coupling regimes, where the potential v is of the form λv0 with a general analytic v0 and C is a positive constant also depending only on v0. Our second author also proved the corresponding conclusion for the Ja- cobi operators in [21]. These two results are optimal, because Avila, Last, Shamis and Zhou [3] showed that the continuity of the Lyapunov exponent of the almost Mathieu operators can’t be H¨olderif β > 0 and e−β < λ < eβ. Until now, we do not know much about the spectrum of the Schr¨odingeror Jacobi operators in the positive Lyapunov exponent regimes when the frequency is not strong Diophantine. The main reason is that we do not know much about the finite- a volume determinant fn (x, E, ω). While, for the almost Mathieu operators, it can be handled explicitly via the Lagrange interpolation for the trigonometric polynomial. This method can be applied for the following extend Harper’s operators, which also have the cosine potential, to obtain many spectral conclusions with the generic frequency, such as [2] and [14]: ω ω a(x) =λ exp[−2πi(x + )] + λ + λ exp[2πi(x + )], 0 ≤ λ , 0 ≤ λ + λ , 3 2 2 1 2 2 1 3 v(x) =2 cos(2πx).
However, the Lagrange interpolation can not work for the more general Schr¨odinger operators, since their potentials both are generic analytic functions. Therefore, in [12], Goldstein and Schlag applied the LDT (1.10) and the relationship that
s s s fn(x, E, ω) fn−1(x + ω, E, ω) Mn(x, E, ω) = s s (1.13) fn−1(x, E, ω) fn−2(x + ω, E, ω)
s to estimate the BMO norm of fn(x, E, ω). Then they obtained the following LDT s for fn(x, E, ω) with the strong Diophantine ω by the John-Nirenberg inequality:
−1 s s (n) mes {x ∈ T : |log |fn (x)| − hlog |fn|i| > nδ} ≤ C exp −cδn δ0 . (1.14)
This LDT was applied to get the H¨olderexponent of the H¨oldercontinuity of s s L (E, ω) in E and the upper bound on the number of eigenvalues of Hn(x, ω) contained in an interval of size n−C . What’s more, with its help, the estimation on s the separation of the eigenvalues of Hn(x, ω) and the property that the spectrum of s H (x, ω), denoted by Sω, is a Cantor set were obtained in [13], and the homogeneity of Sω was proved in [10]. In [8] and [9], Binder and Voda applied this method to our analytic Jacobi operators (1.1). It must be noted that the above conclusions all hold s a only for the strong Diophantine ω and the LDTs for fn(x, E, ω) and fn (x, E, ω) are the key lemma in the method. Now, we can declare that the concrete content of our main aim is to obtain a the LDT for the finite-volume determinant fn (x, E, ω) with more generic ω. It is important in our field and is the key preparation for the study of the spectrum problem for discrete quasiperiodic operators of second order in the future. In this paper, we assume that the frequency ω is the Brjuno-R¨ussmannnumber, which is a famous extension of the strong Diophantine number. It says that there exists a monotone increasing and continuous function ∆(t) : [1, ∞) → [1, ∞) such that ∆(1) = 1 and for any positive integer k > 0, C kkωk > ω , (1.15) ∆(k) LDT DIRICHLET DETERMINANTS WITH BRJUNO-RUSSMANN¨ FREQUENCY 5309 and Z ∞ log ∆(t) 2 < +∞. (1.16) 1 t α 1 For example, this Brjuno-R¨ussmannfunction ∆(t) can be t(log t + 1) , exp(t α ), α α t −1 t , exp((log t) ) and exp( (log t)α ) with α > 1. Define Γω(n) = knωk . Due to (1.11), we have that
qs+1 < Γω(qs) < qs + qs+1, and Γω(n) < Γω(qs), ∀n ∈ (qs, qs+1). (1.17) Therefore, there exists another definition of the Brjuno-R¨ussmannnumber as follow: −1 There exists a function Ψω(t) = max{kkωk , ∀0 < k ≤ t, k ∈ Z} satisfying (1.16). ∞ Note that the denominator series {qs}s=1 and the function Ψω depend on ω. Thus, to make almost every irrational number satisfy (1.15), we assume that ∆(t) > t(log t + 1). (1.18)
∆0(t) 0 0 1 It implies that log ∆(t) > log t but it is false that ∆(t) = (log ∆(t)) > (log t) = t . However, it holds for all examples which we have given. So we make the following hypothesis: Hypothesis H.1. ∆(t) > t(log t + 1) and t∆0(t) ≥ ∆(t) for any t ≥ 1. a Then, our first LDT for fn (x, E, ω) is Theorem 1.1. Let ω be the Brjuno-R¨ussmannnumber satisfying Hypothesis H.1 and L(E, ω) > 0. There exist constants c = c(a, v, E, ω) and C˘ = C˘(a, v, ω), and absolute constant C such that for any integer n ≥ 1 and δ > δH.1(n) := C˘ log ∆(n) −1 −1 1− , where ∆ (·) means the inverse function of ∆(·), then (∆ (Cω n)) a a −1 mes {x ∈ T : |log |fn (x)| − hlog |fn |i| > nδ} ≤ C exp −cδ(δH.1(n)) . Remark 1. In this paper, the notation A1− means A1− for any small absolute > 0. And A1+ has the similar definition.
Remark 2. If we assume the potential v is of the form λv0 with a general analytic v0, then the second author [21] proved that there exists λ0 = λ0(v0, a) such that the Lyapunov exponent L(E, ω) is always positive for any E and any irrational ω under the condition λ > λ0.
α C˘ log n If ∆(t) = t(log t + 1) , then δH.1(n) = n1− which is very close to the smallest (log n)A 1 α deviation for the strong Diophantine number n . But if ∆(t) = exp(t ), then 1 − δH.1(n) = Cn˘ α which is too large for us to apply Theorem 1.1 to the research of the spectrum of the analytic quasi-periodic operators (1.1) in our future work. Thus, we need to make some hypothesis to improve this smallest deviation when ∆(t) grows fast: Hypothesis H.2. ω satisfies Hypothesis H.1 and for any t ≥ 1, t ∆(t) < exp . (1.19) log t t From (1.16), ∆(t) has an upper bound of exp( log t ) generally. But it is possible that it grows very fast and exceeds this upper bound in some intervals, and in the rest it grows very slowly and makes the integral converge. Therefore, the aim of this hypothesis is to avoid this extreme possibility. Then, our second LDT for a fn (x, E, ω) is 5310 WENMENG GENG AND KAI TAO
Theorem 1.2. Let ω be the Brjuno-R¨ussmannnumber satisfying Hypothesis H.2 and L(E, ω) > 0. There exist constants c = c(a, v, E, ω) and C˘ = C˘(a, v, ω), and absolute constant C such that for any integer n ≥ 1 and δ > δH.2(n) := C˘ −1 1− , [log(∆ (Cω n))] a a −1 mes {x ∈ T : |log |fn (x)| − hlog |fn |i| > nδ} ≤ C exp −cδ(δH.2(n)) .
1 α t Remark 3. With this hypothesis, no matter ∆(t) equals to exp(t ) or exp( (log t)α ), C˘ the smallest deviation δH.2(n) = [log log n]1− 1 which satisfies what we need for the study of the spectrum, such as Theorem 1.5. As mentioned above, what we want to avoid is the case that ∆(t) grows faster than exp(t) in some intervals. But Hypothesis H.2 only requires that ∆(t) has an upper bound, but has no restriction on its derivative. Thus, we make the following hypothesis, which gives the mutual restriction between ∆(t) and ∆0(t) and looks also very reasonable: log ∆(t) Hypothesis H.3. ω satisfies Hypothesis H.1 and t is non-increasing for any t ≥ 1. Easy computation shows that this hypothesis is equivalent to the inequality ∆(t) log ∆(t) ∆0(t) ≤ . t Combined it with Hypothesis H.1, it shows that the bound of t∆0(t) is determined by ∆(t). Obviously, all examples of functions mentioned above satisfy this hypothesis. With its help, we improve Theorem 1.1 and 1.2 as follows: Theorem 1.3. Let ω be the Brjuno-R¨ussmannnumber satisfying Hypothesis H.3 and L(E, ω) > 0. There exist constants c = c(a, v, E, ω) and C˘ = C˘(a, v, ω), and absolute constant C such that for any integer n ≥ 1 and δ > δH.3(n) := C˘ log(Cω n) −1 1− , [∆ (Cω n)] a a −1 mes {x ∈ T : |log |fn (x)| − hlog |fn |i| > nδ} ≤ C exp −cδ(δH.3(n)) . Remark 4. Since t(log t + 1) < ∆(t), it is obvious that
log ∆(n) log(Cωn) δH.1(n) = δH.3(n) = . −1 1− −1 1− ∆ (Cωn) ∆ (Cωn) On the other hand, due to the fact that 0 < ∆−1(t) < t < ∆(t), n ∆(log2 ∆−1(n)), we have that
1 log(Cωn) δH.2(n) = δH.3(n) = . −1 1− −1 1− log(∆ (Cωn)) ∆ (Cωn) a The key to prove these three LDTs for fn (x, E, ω) is an ergodic theorem for the subharmonic function shifting on T. Specifically, we know that if T : X → X is an ergodic transformation on a measurable space (X, Σ, m) and f is an m−integrable function, then the Birkhoff Ergodic Theorem tells that the time average functions 1 Pn−1 k 1 R fn(x) = n k=0 f(T x) converge to the space average hfi = m(X) X fdm for almost every x ∈ X. But it doesn’t tell us how fast do they converge. So, we call a theorem the strong Birkhoff Ergodic Theorem, if it gives the convergence rate. The following strong Birkhoff Ergodic Theorem for the subharmonic function shifting on T is the key which we just mentioned above: LDT DIRICHLET DETERMINANTS WITH BRJUNO-RUSSMANN¨ FREQUENCY 5311
Theorem 1.4. Let u :Ω → R be a subharmonic function on a domain Ω ⊂ C and ω be the Brjuno-R¨ussmannumber satisfying Hypothesis H.1, or H.2, or H.3. Suppose that ∂Ω consists of finitely many piece-wise C1 curves and T is contained 0 0 in Ω b Ω(i.e., Ω is a compactly contained subregion of Ω). There exist constants ˘ ˘ (n) c = c(ω, u) and C = C(Ω, u) such that for any positive n and δ > δ0 , ( n )! X mes x ∈ T : | u(x + kω) − nhui| > δn ≤ exp (−cδn) , (1.20) k=1 where C˘ log ∆(n) δH.1(n) := −1 1− , if ω satisfies H.1, (∆ (Cω n)) (n) C˘ δ = δH.2(n) := −1 1− , if ω satisfies H.2, (1.21) 0 [log(∆ (Cω n))] ˘ C log(Cω n) δH.3(n) := −1 1− , if ω satisfies H.3. [∆ (Cω n)] A very interesting thing we find is that no matter the irrational frequency is, the convergence rate of the exceptional measure is always exp (−cδn). The only (n) difference is the smallest deviation δ0 . If β(ω) > 0, then our second author (n) obtained in [21] that δ0 = cβ which is proved to be optimal in [3]; if β(ω) = 0, we obtain (1.21) which includes the result for the strong Diophantine number by Goldstein and Schlag. Correspondingly, the three LDTs we obtain in this paper can be unified into the following form: a a (n)−1 (n) mes x∈T:|log |fn (x)|−hlog |fn |i|>nδ ≤C exp −cδ δ0 , ∀δ >δ0 . (1.22)
(n) While, the exceptional measure in (1.22) will not converge when δ0 = cβ! The method created by Goldstein and Schlag and applied in this paper should be im- proved for the Liouville frequency. We think it is a good question for our further research in the future. Here we need to emphasize that our paper is not weaker version of [21]. That (n) shows that the smallest deviation is δ0 = cβ, and then the strong Birkhoff er- godic theorem and the LDTs for matrices hold when the deviation is larger than (n) δ0 . Letting the positive Lyapunov exponent be this deviation, our second author obtained the H¨oldercontinuity of the Lyapunov exponent. However, if we applied these results in our condition that β = 0, then the smallest deviation is 0! It is absurd! So, compared to [21], the main aim of our second section is to find the smallest deviation when β = 0. What’s more, we will find that in technology the P2m−1 qs−j+1 key is to estimate log qs−j+1. It is easy when β > 0: j=1 qs−j 2m−1 2m−1 X qs−j+1 X log q ≤ 2β q ≤ 8βn. q s−j+1 s−j+1 j=1 s−j j=1 While, when β = 0, the fact that { log qs+1 }∞ has different speeds, which depend qs s=1 on ∆(t), to converge to 0 makes this estimation much harder. On the other hand, a the aims of our Section 3 and 4 are to obtain the LDT for fn and its applications, which are nonexistent in [21]. In summary, the focus point of our paper is to show the importance of the smallest deviation of the strong Birkhoff ergodic theorem and calculate it when β = 0. Of course, when we need the LDTs for matrices and the H¨oldercontinuity of the Lyapunov exponent, such as Lemma 3.1 and 4.1, we can use the results from [21] directly. 5312 WENMENG GENG AND KAI TAO
At last, we have an application of our LDTs, which estimates the upper bound (n) 1 on the number of eigenvalues of Hn(x, ω) contained in an interval of size (δ0 ) h , where h is the H¨olderexponent of the H¨oldercontinuity of L(E, ω), see Lemma 4.1. The distribution of the eigenvalues is very important in the further study of the spectrum problem for discrete quasiperiodic operators of second order. With fixed x and ω, the matrix Hn(x, ω) has n eigenvalues. So we have an intuition that these eigenvalues have a more uniform distribution when the frequency ω is “more irrational”. For the Brjuno-R¨ussmannumber, it means that ∆(t) grows more slowly (n) and then δ0 is smaller. The following theorem verifies our intuition: Theorem 1.5. Let ω be the Brjuno-R¨ussmannnumber satisfying Hypothesis H.1, or H.2, or H.3 and L(E, ω) > 0. Then, for any x0 ∈ T and E0 ∈ R, 1 a (n) h (n) # E ∈ R : fn (x0, E, ω) = 0, |E − E0| < δ0 ≤ 13nδ0 .
We organize this paper as follows. In Section 2, we prove Theorem 1.4, the strong Birkhoff Ergodic theorem for the subharmonic function shifting on T with our Brjuno-R¨ussmannfrequency. We apply it to the analytic quasi-periodic Jacobi operator and obtain Theorem 1.1, Theorem 1.2 and Theorem 1.3 in Section 3, a which are all the LDTs for fn (x, E, ω) with different hypothesises. Then, we prove Theorem 1.5, an application of them, in the last section.
2. Strong birkhoff ergodic theorem for subharmonic functions with the Brjuno-R¨ussmannshift. Let {x} = x − [x]. For any positive integer q, complex number ζ = ξ + iη and 0 ≤ x < 1, define
X Z 1 Fq,ζ (x) = log |{x + kω} − ζ| and I(ζ) = log |y − ζ|dy. 0≤k qs Let |{x + k0ω} − ξ| = mink=1 |{x + kω} − ξ|, where qs is the denominator of the continued fraction approximants. In [11], Goldstein and Schlag proved Lemma 3.1 that for any irrational ω, there exists an absolute constant C such that |Fqs,ζ (x) − qsI(ζ)| ≤ C log qs + |log |{x + k0ω} − ζ|| . (2.1) Then Lemma 2.1. For any irrational ω and integer l < qs+1 , qs |Flsqs,ζ (x) − lsqsI(ζ)| < Cls log qs + | log D(x − ξ, ω, lsqs)| + 2ls log qs+1, n−1 where D(x, ω, n) := mink=0 {x + kω}. qs−1 Proof. Define xh = x + hqsω and |{xh + khω} − ξ| = mink=0 |{xh + kω} − ξ|. Due to (2.1), we have l −1 Xs |Flsqs,ζ (x) − lsqsI(ζ)| ≤ |Fqs,ζ (xh) − lsqsI(ζ)| h=0 l −1 Xs ≤ |log |{xh + khω} − ζ|| + Cls log qs. (2.2) h=0 LDT DIRICHLET DETERMINANTS WITH BRJUNO-RUSSMANN¨ FREQUENCY 5313 1 1 We declare that if there exists 0 ≤ j < qs such that |{x + jω} − ξ| ≤ − , 2qs qs+1 1 1 then j = k0. Actually, if |{x+jω}−ξ| ≤ − and j 6= k0, then |{x+k0ω}−ξ| ≤ 2qs qs+1 |{x + jω} − ξ| ≤ 1 − 1 , which implies 2qs qs+1 1 2 {x + k0ω} − {x + jω} ≤ − . qs qs+1 Due to (1.11), it has k kps k 1 < |kω − | < ≤ , 0 < k < qs. (2.3) qs(qs+1 + qs) qs qsqs+1 qs+1 Then, ps ps 1 {x + j } − {x + k0 } < . qs qs qs It is a contraction. Thus, there is at most one integer 0 ≤ k0 < qs such that 1 1 |{x + k0ω} − ξ| < − and 2qs qs+1 1 1 1 |{x + kω} − ξ| > − > , k 6= k0. (2.4) 2qs qs+1 4qs Due to (1.11) again, it has 1 1 1 < < |qsω − ps| < . (2.5) 2qs+1 qs + qs+1 qs+1 qs+1 1 Define Q = [ ] and let j be the number such that |{xj + kjω} − ξ| < . qs 4qs+1 Then by (2.5) and the above declaration, we have for any j − 2Q + 1 ≤ h < j and j < h ≤ j + 2Q − 1, 1 |{xh + khω} − ξ| > . 4qs+1 Thus there are at most one point which is small than 1 . Combining it with 4qs+1 (2.2), we have 1 |Flsqs,ζ (x) − lsqsI(ζ)| ≤| log D(x − ξ, ω, lsqs)| + Cls log qs + ls log 4qs+1 ≤| log D(x − ξ, ω, lsqs)| + Cls log qs + 2ls log qs+1. qs+1 Lemma 2.2. For any 0 < σ < 1, irrational ω and integer ls < , qs Z exp (σ|Flsqs,ζ (x) − lsqsI(ζ)|) dx < exp (5σls log qs+1) . T Proof. We first apply Lemma 3.2 in [11]. It says that if Ω ⊂ T is an arbitrary finite set, then for any 0 < σ < 1, Z 2σ exp (σ| log dist(x, Ω)|) dx ≤ (]Ω)σ. (2.6) 1 − σ T Set Ω = {mω : 0 ≤ m < lsqs}. Then ]Ω = lsqs and dist(x−ξ, Ω) = D(x−ξ, ω, lsqs). Thus, by (2.6), Z Z 2σ exp (σ| log D(x − ξ, ω, l q )|) dx = exp (σ| log dist(x, Ω)|) dx ≤ (l q )σ. s s 1 − σ s s T T By Lemma 2.1, we have Z exp (σ|Flsqs,ζ (x) − lsqsI(ζ)|) dx T 5314 WENMENG GENG AND KAI TAO ≤ exp (2Cσ log(lsqs) + Cσls log qs + 2σls log qs+1) < exp (5σls log qs+1) . n Now for any n, there exist qs and qs+1 such that qs ≤ n < qs+1. Let ls = [ ] and qs qi+1 li = [ ] for i < s. Then there exists rs−2m+1 satisfying 0 ≤ rs−2m+1 < qs−2m+1 qi such that n = lsqs + ls−1qs−1 + ··· + ls−2m+1qs−2m+1 + rs−2m+1. Define n¯ = lsqs + ls−1qs−1 + ··· + ls−2m+1qs−2m+1. (2.7) Then we have Lemma 2.3. For any compact Ω ⊂ C, there exist constants c˜ =c ˜(ω) such that for any 0 < σ ≤ c˜, we have Z 1 ˜ ˘(n) exp(σ|Fn,ζ¯ (x) − nI¯ (ζ)|)dx ≤ exp Cσnδ0 , (2.8) 0 where ˘ log ∆(n) δH.1(n) := 1− , if ω satisfies H.1, [∆−1(C n)] ω (n) ˘ 1 δ˘ = δH.2(n) := , if ω satisfies H.2, (2.9) 0 [log(∆−1(C n))]1− ω ˘ log(Cωn) δH.3(n) := , if ω satisfies H.3. −1 1− [∆ (Cωn)] Proof. Let rs =n ¯ − lsqs. Due to the H¨olderinequality and Lemma 2.2, Z 1 exp(σ|Fn,ζ¯ (x) − nI¯ (ζ)|)dx 0 1 1 Z 1 2 Z 1 2 ≤ exp(2σ|Flsqs,ζ (x) − lsqsI(ζ)|)dx × exp(2σ|Frs,ζ (x) − rsI(ζ)|)dx 0 0 1 Z 1 2 ≤ exp(5σls log qs+1) exp(2σ|Frs,ζ (x) − rsI(ζ)|)dx . 0 rs−i+1 Let rs−i+1 = ls−iqs−i + rs−i, where ls−i = [ ], 0 ≤ rs−i = rs−i+1 − ls−iqs−i < qs−i qs−i. Then 1 Z 1 2i i exp(2 σ|Frs−i+1,ζ (x) − rs−i+1I(ζ)|)dx 0 1 Z 1 2i+1 i+1 ≤ exp(2 σ|Fls−iqs−i,ζ (x) − ls−iqs−iI(ζ)|)dx 0 1 Z 1 2i+1 i+1 × exp(2 σ|Frs−i,ζ (x) − rs−iI(ζ)|)dx 0 1 Z 1 2i+1 i+1 ≤ exp (5σls−i log qs−i+1) × exp(2 σ|Frs−i,ζ (x) − rs−iI(ζ)|)dx . 0 Therefore, Z 1 exp(σ|Fn,ζ¯ (x) − nI¯ (ζ)|)dx 0 ≤ exp [5σ (ls log qs+1 + ls−1 log qs + ··· + ls−2m+1 log qs−2m+2)] LDT DIRICHLET DETERMINANTS WITH BRJUNO-RUSSMANN¨ FREQUENCY 5315 2m−1 n X qs−j+1 ≤ exp 5σ log q + log q . (2.10) q s+1 q s−j+1 s j=1 s−j t log t Assume that ω satisfies Hypothesis H.1. Then, we assert that −1 is mono- ∆ (Cω t) −1 tone increasing. Indeed, let y = ∆ (Cωt). Then, due to the hypothesis that y∆0(y) > ∆(y), it yields that C t ∆−1 (C t) ∆0 ∆−1 (C t) > C t ≥ C t − ω . ω ω ω ω log t + 1 Combining it with the fact that 1 = ∆0 ∆−1 (C t) , −1 0 ω (∆ (Cωt)) we have −1 −10 ∆ (Cωt) (log t + 1) > Cωt log t ∆ (Cωt) . (2.11) Now we finish the proof of the assertion as (2.11) shows that the numerator of the t log t derivative of −1 is positive. ∆ (Cω t) Due to (1.15) and (1.11), −1 Cωqi < ∆(qi−1) and qi−1 > ∆ (Cωqi). (2.12) Therefore, qi qi log qi < −1 log qi. qi−1 ∆ (Cωqi) We apply the assertion and obtain 2m−1 X qs−j+1 n log q < (2m − 1) log n. (2.13) q s−j+1 ∆−1(C n) j=1 s−j ω Recall that ∆(t) is monotone increasing and continuous. And so is ∆−1(t). Com- bining it with (2.12), we have −1 −1 qs > ∆ (Cωqs+1) > ∆ (Cωn), (2.14) and for any n > n0(ω), n ∆(qs) 2n log ∆(qs) 2n log ∆(qs) 2n log ∆(n) ls log qs+1 ≤ log ≤ ≤ −1 ≤ −1 . (2.15) qs Cω qs ∆ (Cωn) ∆ (Cωn) Choosing −1 m = mH.1 = log2 ∆ (Cωn), (2.16) and due to (2.10), (2.13), (2.15), we have Z 1 exp(σ|Fn,ζ¯ (x) − nI¯ (ζ)|)dx 0 2n log ∆(n) −1 n ≤ exp 5σ −1 + 2 log2 ∆ (Cωn) −1 log n ∆ (Cωn) ∆ (Cωn) −1 logC¯ ∆ (Cωn) ≤ exp 20σn log ∆(n) −1 ∆ (Cωn) ( ) log ∆(n) ≤ exp 20σn . (2.17) −1 1− (∆ (Cωn)) 5316 WENMENG GENG AND KAI TAO t Assume that ω satisfies Hypothesis H.2 which implies that ∆(t) < exp( log t ) holds. Then log ∆(qs) 1 log n 1 < and −1 < −1 . (2.18) qs log qs ∆ (Cωn) log ∆ (Cωn) Combining them with (2.13) and (2.14), 2n log ∆(qs) 2n 2n ls log qs+1 < < < −1 , qs log qs log(∆ (Cωn)) and 2m−1 X qs−j+1 qs log q <(2m − 1) log q q s−j+1 ∆−1(C q ) s j=1 s−j ω s n <(2m − 1) −1 log n ∆ (Cωn) n <(2m − 1) −1 . log(∆ (Cωn)) Letting −1 m = mH.2 = log2 log ∆ (Cωn), (2.19) we obtain that Z 1 −1 n exp(σ|Fn,ζ¯ (x) − nI¯ (ζ)|)dx ≤ exp 20σ log2 log ∆ (Cωn) −1 0 log ∆ (Cωn) ( ) n ≤ exp 20σ . (2.20) −1 1− [log(∆ (Cωn))] log ∆(t) Assume the ω satisfies Hypothesis H.3 which implies that t is non-increasing. Due to (2.14) and (2.15), −1 2n log ∆(qs) n log ∆(∆ (Cωn)) n log(Cωn) ls log qs+1 ≤ ≤ −1 = −1 . (2.21) qs ∆ (Cωn) ∆ (Cωn) By (2.13), 2m−1 X qs−j+1 qs log q <(2m − 1) log q q s−j+1 ∆−1(C q ) s j=1 s−j ω s n ≤(2m − 1) −1 log n. (2.22) ∆ (Cωn) Choosing −1 m = mH3 = log2(∆ (Cωn)), (2.23) we have Z 1 ! n log(Cωn) exp(σ|Fn,ζ¯ (x) − nI¯ (ζ)|)dx ≤ exp 20σ . (2.24) −1 1− 0 [∆ (Cωn)] To finish the proof of Theorem 1.4, we need the following Riesz’s theorem proved in [12]: LDT DIRICHLET DETERMINANTS WITH BRJUNO-RUSSMANN¨ FREQUENCY 5317 Lemma 2.4. Let u :Ω → R be a subharmonic function on a domain Ω ⊂ C. Suppose that ∂Ω consists of finitely many piece-wise C1 curves. There exists a positive measure µ on Ω such that for any Ω1 b Ω (i.e., Ω1 is a compactly contained subregion of Ω), Z u(z) = log |z − ζ| dµ(ζ) + h(z), (2.25) Ω1 where h is harmonic on Ω1 and µ is unique with this property. Moreover, µ and h satisfy the bounds µ(Ω1) ≤ C(Ω, Ω1) (sup u − sup u), (2.26) Ω Ω1 ∞ kh − sup ukL (Ω2) ≤ C(Ω, Ω1, Ω2) (sup u − sup u) (2.27) Ω1 Ω Ω1 for any Ω2 b Ω1. What’s more, we can obtain the following strong Birkhoff Ergodic Theorem for this harmonic function h with the Brjuno-R¨ussmanfrequency easily: Lemma 2.5. Let h be a 1−periodic harmonic function defined on a neighborhood of the real axis. There exists a constant C depending only on h such that for any integer n and Brjuno-R¨ussman ω, n X Z 1 C h(x + kω) − n h(y)dy < , Cω k=1 0 where Cω comes from (1.15). Proof. n Z 1 X h(x + kω) − n h(y)dy k=1 0 n ∞ n ! X X 2πij(x+kω) X X 2πijkω = hˆ(j)e − nhˆ(0) ≤ hˆ(j) e , k=1 j=−∞ j6=0 k=1 where hˆ(j) is the j−th Fourier coefficients of h. Due to its harmonicity, there exists two constants Cˇh and ρ depending only on h such that ˆ |h(j)| ≤ Cˇh exp (−ρj) . (2.28) Easy computation shows that n X 2πijkω exp(2πijω) · (1 − exp(2πinjω)) 1 e = ≤ . 1 − exp(2πijω) 2kjωk k=1 P∞ log ∆(j) Therefore, we obtain this lemma by (2.28) and j=1 j2 < ∞ from (1.16). The proof of Theorem 1.4. Notice that the ergodic measure for the shift on the Torus is the Lebesgue measure and m( ) = 1. Then, hui = R u(x)dx, and T T n n X X Z Z u(x + kω) − nhui = log |{x + kω} − ζ|dµ(ζ) − n I(ζ)dµ(ζ) k=1 k=1 Ω1 Ω1 n X Z 1 + h({x + kω}) − n h(y)dy. k=1 0 5318 WENMENG GENG AND KAI TAO Recall that n X Z Z log |{x + kω} − ζ|dµ(ζ) = Fn,ζ (x)dµ(ζ). k=1 Ω1 Ω1 1 Thus, due to Lemma 2.5, it yields that for any n and δ n , ( n ) X x ∈ T : u(x + kω) − nhui > δn k=1 Z δn ⊆ x ∈ T : (Fn,ζ (x) − nI(ζ))dµ(ζ) > . (2.29) Ω1 2 To estimate the measure of the upper set, we use the Markov’s inequality: For any measurable extended real-valued function f(x) and > 0,we have 1 Z mes ({x ∈ : |f(x)| ≥ }) ≤ |f|dx. X X R Let f(x) = exp σ (Fn,ζ (x) − nI(ζ))dµ(ζ) and = exp(σδn/2), then Ω1 Z δn mes x ∈ T : (Fn,ζ (x) − nI(ζ))dµ(ζ) > Ω1 2 Z 1 Z σδn ≤ exp − exp σ (Fn,ζ (x) − nI(ζ))dµ(ζ) . (2.30) 2 0 Ω1 Due to the H¨olderinequality, Z 1 Z exp σ (Fn,ζ (x) − nI(ζ))dµ(ζ) dx 0 Ω1 1 Z 1 Z 2 ≤ exp 2σ (Fn−n,ζ¯ (x) − (n − n¯)I(ζ))dµ(ζ) dx 0 Ω1 1 Z 1 Z 2 × exp 2σ (Fn,ζ¯ (x) − nI¯ (ζ))dµ(ζ) dx , (2.31) 0 Ω1 wheren ¯ comes from (2.7). Since exp(σ·) is a convex function, the Jensen’s inequality and Lemma 2.3 imply that Z 1 Z exp σ (Fn,ζ¯ (x) − nI¯ (ζ))dµ(ζ) dx 0 Ω1 Z 1 Z dµ(ζ) ≤ exp (σµ(Ω1) |Fn,ζ¯ (x) − nI¯ (ζ)|) dx 0 Ω1 µ(Ω1) Z Z 1 dµ(ζ) = exp (σµ(Ω1) |Fn,ζ¯ (x) − nI¯ (ζ)|) dx Ω1 0 µ(Ω1) Z ˘(n) dµ(ζ) ˘(n) ≤ exp 20σµ(Ω1)nδ0 ≤ exp 20σµ(Ω1)nδ0 . (2.32) µ(Ω1) On the other hand, due to the facts that it is integrable for log |z| on the disc |z| < r and |I(ζ)| ≤ |log (Imζ)| if |Imζ| is close to 0, it is obvious that there exists a constant Cˆ = Cˆ(Ω1) such that Z Z dµ(ζ) ˆ (log |x − ζ| − I(ζ)) dµ(ζ) ≤ µ(Ω1) (log |x − ζ| − I(ζ)) ≤ µ(Ω1)C. Ω1 Ω1 µ(Ω1) LDT DIRICHLET DETERMINANTS WITH BRJUNO-RUSSMANN¨ FREQUENCY 5319 Thus, 1 Z 1 Z 2 exp 2σ (Fn−n,ζ¯ (x) − (n − n¯)I(ζ))dµ(ζ) dx 0 Ω1 ≤ exp σµ(Ω1)Cˆ(n − n¯) ≤ exp σµ(Ω1)Crˆ s−2m+1 . (2.33) Note that for any irrational ω, the denominators of its continued fraction approxi- mates satisfy qt+1 = at+1qt + qt−1 > 2qt−1. Thus m −m qt > 2 qt−2m and rs−2m+1 < 2 n. −1 Recall that in the proof of Lemma 2.3, we choose m = mH.1 = log2 ∆ (Cωn) when ω satisfy the hypothesis H.1. Then ˆ ˆ n ˇ ˇ(n) Crs−2m+1 < C −1 nδH.1(n) = nδ0 . ∆ (Cωn) Similarly, Crˆ s−2m+1 also has the same upper bound in the other two hypothesises. Combined it with (2.30), (2.31), (2.32) and (2.33), we have that for any 0<σ ≤ c˜ , µ(Ω1) Z δn mes x ∈ T : (Fn,ζ (x) − nI(ζ))dµ(ζ) > Ω1 2 σδn ≤ exp − + 21σµ(Ω )nδ˘(n) . 2 1 0 Thus, we finish this proof by (2.29) and setting C˘ = 100µ(Ω1) with the fact that (n) 1 δ0 n . a 3. Large deviation theorems for fn (x, E, ω). To apply Theorem 1.4, we first need to define some subharmonic functions. Let n a Y Mn (x, E, ω) := a(x + jω) Mn(x, E, ω) j=1 n Y v(x + jω) − E a(x + jω) = . (3.1) a(x + (j + 1)ω) 0 j=1 Note that a real function f(x) on T has its complex analytic extension f(z) on the complex strip Tρ = {z : |Imz| < ρ} and the complex analytic extension ofa ¯(x) should be defined on Tρ by 1 a˜(z) := a( ). z a Then, the extension of Mn (x, E, ω) is n Y v(z + jω) − E a˜(z + jω) M a(z, E, ω) = , (3.2) n a(z + (j + 1)ω) 0 j=1 where z + ω means z exp (2πiω) here. Moreover, simple computations yield that a a a fn (z, E, ω)a ˜(z)fn−1(z + ω, E, ω) Mn (z, E, ω) = a a , a(z + nω)fn−1(z, E, ω) −a˜(z)a(z + nω)fn−2(z + ω, E, ω) (3.3) 5320 WENMENG GENG AND KAI TAO where a fn (z, E, ω) = det Hn(z, ω) − E v z + ω − E −a z + 2ω 0 ··· 0