1:  % amslatex

2: \documentclass[10pt]{amsart}

3: \usepackage{amssymb}

4: \bibliographystyle{amsalpha}

5: \baselineskip=24pt

6: % ***********************************************************

7: % *** The following two lines take out the picture files! ***

8: %\input{psfig}

9: %\def\psfig#1{}

10: 

11: %\marginparwidth .85in

12: %\oddsidemargin -.775in

13: %\evensidemargin .115in

14: %\textwidth 7in

15: 

16: \marginparwidth .85in

17: \oddsidemargin .075in

18: \evensidemargin .075in

19: \textwidth 6in

20: 

21: \usepackage[OT2,OT1]{fontenc}

22: 

23: \def\cyr{\fontencoding{OT2}\fontfamily{wncyr}\selectfont}

24: 

25: \def\Ch{\textrm{\cyr CH}}

26: 

27: %\def\proclaim ##1. ##2\par{\medbreak

28: %  \noindent{\bf##1.\enspace}{\sl##2\par}%

29: %  \ifdim\lastskip<\medskipamount \removelastskip\penalty55\medskip\fi}%

30: 

31: \setcounter{secnumdepth}{4}

32: 

33: \newcommand{\bmu}{{\bar\mu}}

34: \newcommand{\bnu}{{\bar\nu}}

35: 

36: \newcommand{\crit}{{\mathrm{cr}}}

37: 

38: \newcommand{\proclaim}[2]{\medbreak {\bf #1}{\sl #2} \medbreak}

39: 

40: \newcommand{\ntop}[2]{\genfrac{}{}{0pt}{1}{#1}{#2}}

41: 

42: \let\newpf\proof \let\proof\relax \let\endproof\relax

43: \newenvironment{pf}{\newpf[\proofname]}{\qed\endtrivlist}

44: 

45: \def\area{\operatorname {area}}

46: 

47: \def\MD{\operatorname{MD}}

48: 

49: \def\H{\mathbb {H}}

50: 

51: \def\SL{\mathrm {SL}}

52: \def\SO{\mathrm {SO}}

53: 

54: \def\be{\begin{equation}}

55: \def\ee{\end{equation}}

56: 

57: \def\bm{\begin{pmatrix}}

58: \def\em{\end{pmatrix}}

59: 

60: %\def\ba{{\begin{align}}}

61: %\def\ea{{\end{align}}}

62: 

63: \def\LLL{\Lambda}

64: 

65: \def\u{{\mathbb U}}

66: 

67: \def\tA{\tilde {\mathcal {A}}}

68: \def\tB{\tilde {\mathcal {B}}}

69: \def\tC{\tilde {\mathcal {C}}}

70: 

71: \def\d{{\underline d}}

72: \def\j{{\bf j}}

73: \def\n{{\bf n}}

74: \def\r{{\bf r}}

75: \def\l{{\bf l}}

76: \def\g{{\gamma}}

77: \def\x{{\bf x}}

78: 

79: \def\op{\overline\partial}

80: 

81: \def\O{{\mathbb O}}

82: 

83: \def\0{{\mathbf 0}}

84: 

85: \def\hh{{\mathbf h}}

86: 

87: \def\cal{\mathcal}

88: 

89: \def\SSS{{\cal {S}}}

90: 

91: \newcommand{\Per}{\operatorname{Per}}

92: \newcommand{\Fix}{\operatorname{Fix}}

93: \newcommand{\la}{\label}

94: \newtheorem{thm}{Theorem}[section]

95: \newtheorem*{mainthm}{Main Theorem}

96: \newtheorem*{thmA}{Theorem A}%[section]

97: \newtheorem*{thmB}{Theorem B}%[section]

98: \newtheorem*{thmC}{Theorem C}%[section]

99: \newtheorem*{thmD}{Theorem D}%[section]

100: \newtheorem*{thmE}{Theorem E}%[section]

101: \newtheorem{cor}[thm]{Corollary}

102: \newtheorem{conj}[thm]{Conjecture}

103: \newtheorem{lem}[thm]{Lemma}

104: \newtheorem{lemma}[thm]{Lemma}

105: \newtheorem{claim}[thm]{Claim}

106: \newtheorem{prop}[thm]{Proposition}

107: \newtheorem{schw}{Schwarz Lemma} 

108: \newtheorem{sectl}{Sector Lemma}

109: \theoremstyle{remark}

110: \newtheorem{rem}{Remark}[section]

111: \newtheorem{notation}{Notation}

112: \newtheorem{example}{Example}[section]

113: \newtheorem{problem}{Problem}

114: 

115: \numberwithin{equation}{section}

116: \newcommand{\thmref}[1]{Theorem~\ref{#1}}

117: \newcommand{\propref}[1]{Proposition~\ref{#1}}

118: \newcommand{\secref}[1]{\S\ref{#1}}

119: \newcommand{\lemref}[1]{Lemma~\ref{#1}}

120: \newcommand{\corref}[1]{Corollary~\ref{#1}} 

121: \newcommand{\figref}[1]{Fig.~\ref{#1}}

122: 

123: \renewcommand{\theschw}{}

124: \renewcommand{\thesectl}{}

125: 

126: 

127: 

128: 

129: \def\theequation {\thesection.\arabic{equation}}

130: \def\pn {\par\smallskip\noindent}

131: \def \bn {\hfill \\ \smallskip\noindent}

132: 

133: \theoremstyle{definition}

134: \newtheorem{defn}{Definition}[section]

135: \newtheorem{definition}{Definition}[section]

136: \def\proof{\bn {\bf Proof.} }

137: 

138: \def\note#1

139: {\marginpar

140: %{\nt $\leftarrow$

141: {\tiny $\leftarrow$

142: \par

143: \hfuzz=20pt \hbadness=9000 \hyphenpenalty=-100 \exhyphenpenalty=-100

144: \pretolerance=-1 \tolerance=9999 \doublehyphendemerits=-100000

145: \finalhyphendemerits=-100000 \baselineskip=6pt

146: #1}\hfuzz=1pt}

147: 

148: \newcommand{\bignote}[1]{\begin{quote} \sf #1 \end{quote}}

149: 

150: 

151: 

152: \newcommand{\QED}{\rlap{$\sqcup$}$\sqcap$\smallskip}

153: 

154: \newcommand{\di}{\partial}

155: \newcommand{\dibar}{\bar\partial}

156: \newcommand{\ra}{\rightarrow}

157: \newcommand{\hra}{\hookrightarrow}

158: %\def\lra{\longrightarrow}

159: 

160: \def\ssk{\smallskip}

161: \def\msk{\medskip}

162: \def\bsk{\bigskip}

163: \def\noi{\noindent}

164: \def\nin{\noindent}

165: \def\lqq{\lq\lq}

166: %\def\sm{\setminus}

167: \def\sm{\setminus}

168: \def\bolshe{\succ}

169: \def\ssm{\smallsetminus}

170: \def\tr{{\text{tr}}}

171: 

172: %\renewcommand{\thenotation}{} 

173: \newcommand{\diam}{\operatorname{diam}}

174: \newcommand{\depth}{\operatorname{depth}}

175: \newcommand{\dist}{\operatorname{dist}}

176: \newcommand{\distM}{\operatorname{dist}_M}

177: \newcommand{\cl}{\operatorname{cl}}

178: \newcommand{\inter}{\operatorname{int}}

179: \renewcommand{\mod}{\operatorname{mod}}

180: \newcommand{\tl}{\tilde}

181: 

182: %\newcommand{\arg}{\operatorname{arg}}

183: \newcommand{\orb}{\operatorname{orb}}

184: \newcommand{\HD}{\operatorname{HD}}

185: \newcommand{\supp}{\operatorname{supp}}

186: \newcommand{\id}{\operatorname{id}}

187: \newcommand{\length}{\operatorname{length}}

188: \newcommand{\dens}{\operatorname{dens}}

189: \newcommand{\meas}{\operatorname{meas}}

190: \newcommand{\Spec}{\operatorname{Spec}}

191: 

192: \newcommand{\Dil}{\operatorname{Dil}}

193: \newcommand{\Ker}{\operatorname{Ker}}

194: \newcommand{\Lin}{\operatorname{Lin}}

195: \newcommand{\tg}{\operatorname{tg}}

196: \newcommand{\codim}{\operatorname{codim}}

197: \newcommand{\isom}{\approx}

198: \newcommand{\esssup}{\operatorname{ess-sup}}

199: 

200: \newcommand{\SLa}{\underset{\La}{\Subset}}

201: 

202: \newcommand{\const}{\mathrm{const}}

203: \def\loc{{\mathrm{loc}}}

204: 

205: 

206: \newcommand{\eps}{{\epsilon}}

207: \newcommand{\De}{{\Delta}}

208: \newcommand{\de}{{\delta}}

209: %\newcommand{\la}{{\lambda}}

210: \newcommand{\La}{{\Lambda}}

211: \newcommand{\si}{{\sigma}}

212: \newcommand{\Om}{{\Omega}}

213: \newcommand{\om}{{\omega}}

214: 

215: \newcommand{\AAA}{{\cal A}}

216: \newcommand{\BB}{{\cal B}}

217: \newcommand{\CC}{{\cal C}}

218: \newcommand{\DD}{{\cal D}}

219: \newcommand{\EE}{{\cal E}}

220: \newcommand{\EEE}{{\cal O}}

221: \newcommand{\II}{{\cal I}}

222: \newcommand{\FF}{{\cal F}}

223: \newcommand{\GG}{{\cal G}}

224: \newcommand{\JJ}{{\cal J}}

225: \newcommand{\HH}{{\cal H}}

226: \newcommand{\KK}{{\cal K}}

227: \newcommand{\LL}{{\cal L}}

228: \newcommand{\MM}{{\cal M}}

229: \newcommand{\NN}{{\cal N}}

230: \newcommand{\OO}{{\cal O}}

231: \newcommand{\PP}{{\cal P}}

232: \newcommand{\QQ}{{\cal Q}}

233: \newcommand{\RR}{{\cal R}}

234: %\newcommand{\SS}{{\cal S}}

235: \newcommand{\TT}{{\cal T}}

236: \newcommand{\TTT}{{\cal P}}

237: \newcommand{\UU}{{\cal U}}

238: \newcommand{\VV}{{\cal V}}

239: \newcommand{\WW}{{\cal W}}

240: \newcommand{\XX}{{\cal X}}

241: \newcommand{\YY}{{\cal Y}}

242: \newcommand{\ZZ}{{\cal Z}}

243: 

244: 

245: 

246: \newcommand{\A}{{\mathbb A}}

247: \newcommand{\C}{{\mathbb C}}

248: \newcommand{\D}{{\mathbb D}}

249: \newcommand{\E}{{\mathbb E}}

250: \newcommand{\F}{{\mathbb F}}

251: \newcommand{\Hyp}{{\mathbb H}}

252: \newcommand{\J}{{\mathbb J}}

253: \newcommand{\Ll}{{\mathbb L}}

254: \renewcommand{\L}{{\mathbb L}}

255: \newcommand{\M}{{\mathbb M}}

256: \newcommand{\N}{{\mathbb N}}

257: \newcommand{\Q}{{\mathbb Q}}

258: \newcommand{\R}{{\mathbb R}}

259: \newcommand{\T}{{\mathbb T}}

260: \newcommand{\V}{{\mathbb V}}

261: \renewcommand{\U}{{\Upsilon}}

262: \newcommand{\W}{{\mathbb W}}

263: \newcommand{\X}{{\mathbb X}}

264: \newcommand{\Z}{{\mathbb Z}}

265: 

266: \newcommand{\tT}{{\mathrm{T}}}

267: \newcommand{\tD}{{D}}

268: \newcommand{\hyp}{{\mathrm{hyp}}}

269: \newcommand{\qc}{{\mathrm{qc}}}

270: 

271: \newcommand{\f}{{\bf f}}

272: \newcommand{\h}{{\bf h}}

273: \renewcommand{\i}{{\bar i}}

274: %\renewcommand{\j}{{\bar j}}

275: 

276: 

277: \def\Bf{{\bold{f}}}

278: \def\Bg{{\bold{g}}}

279: \def\BG{{\bold{G}}}

280: \def\Bh{{\bold{h}}}

281: \def\BT{{\bold{T}}}

282: \def\Bj{{\bold{j}}}

283: \def\Bphi{{\bold{\Phi}}}

284: \def\Bpsi{{\bold{\Psi}}}

285: \def\B0{{\bold{0}}}

286: 

287: \newcommand{\Com}{\cal Com}

288: \newcommand{\Top}{\cal Top}

289: \newcommand{\QC}{\cal QC}

290: \newcommand{\Def}{\cal Def}

291: \newcommand{\Teich}{\cal Teich}

292: \newcommand{\QL}{{\cal{QG}}}

293: \newcommand{\PPL}{{\cal P}{\cal L}}

294: 

295: \newcommand{\hf}{{\hat f}}

296: \newcommand{\hz}{{\hat z}}

297: \newcommand{\hM}{{\hat M}} 

298: 

299: 

300: \renewcommand{\lq}{``}

301: \renewcommand{\rq}{''}

302: 

303: 

304: 

305: 

306: 

307: %&&&&&&&&&&    Content   &&&&&&&&

308: 

309: 

310: \catcode`\@=12

311: 

312: \def\Empty{}

313: \newcommand\oplabel[1]{

314:   \def\OpArg{#1} \ifx \OpArg\Empty {} \else

315:   	\label{#1}

316:   \fi}

317: 		

318: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

319: % Insert a postscript figure using psfig.

320: % Usage:	\realfig{label}{filename}{caption}

321: %

322: % uses psfig macros: must have \input{psfig} in the preamble to use

323: % it. 

324: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

325: 

326: \long\def\realfig#1#2#3#4{

327: \begin{figure}[htbp]

328: %%%\centerline{\psfig{figure=#3,height=#2}}

329: \centerline{\psfig{figure=#2,width=#4}}

330: \caption[#1]{#3}

331: \oplabel{#1}

332: \end{figure}}

333: 

334: %&&&&&&&&&&&&       List of figures              &&&&&&&&&

335: %

336: %&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&

337: 

338: \newcommand{\comm}[1]{}

339: \newcommand{\comment}[1]{}

340: %\renewcommand{\marginpar}[1]{}

341: 

342: \begin{document}

343: 

344: 

345: \bigskip\bigskip

346: 

347: \title{The Ten Martini Problem}

348: \author{Artur Avila}

349: \author[Svetlana Jitomirskaya] {Svetlana Jitomirskaya$^\dag$}

350: \address{

351: Laboratoire de Probabilit\'es et Mod\`eles al\'eatoires\\

352: Universit\'e Pierre et Marie Curie--Boite courrier 188\\

353: 75252--Paris Cedex 05, France

354: }

355: \email{artur@ccr.jussieu.fr}

356: \address{

357: University of California, Irvine, California

358: }

359: \email{szhitomi@uci.edu}

360: 

361: \thanks{$^\dag$This work was supported in part by NSF, grant DMS-0300974.}

362: \date{\today}

363: 

364: \begin{abstract}

365: We prove the conjecture (known as the ``Ten Martini Problem'' after

366: Kac and Simon) that the spectrum of the almost Mathieu operator is a

367: Cantor set for all non-zero values of the coupling and all irrational

368: frequencies.

369: \end{abstract}

370: 

371: \setcounter{tocdepth}{1}

372: 

373: \maketitle

374: 

375: \section{Introduction}

376: 

377: The almost Mathieu operator is the Schr\"odinger operator on $\ell^2(\Z)$,

378: \be

379: (H_{\lambda,\alpha,\theta} u)_n=u_{n+1}+u_{n-1}+2 \lambda \cos 2 \pi

380: (\theta+n\alpha) u_n,

381: \ee

382: where $\lambda,\alpha,\theta \in \R$ are parameters (called

383: the {\it coupling}, {\it frequency}, and {\it phase}, respectively), and one

384: assumes that $\lambda \neq 0$.  The interest in this particular model is

385: motivated both by its connections to physics and by a remarkable richness of

386: the related spectral theory.  This has made the latter

387: a subject of intense research in

388: the last three decades (see \cite {L3} for a recent historical account and for the

389: physics background).  Here we are concerned

390: with the topological structure of the spectrum.

391: 

392: If $\alpha=\frac {p} {q}$ is rational, it is well known that the spectrum

393: consists of the union of $q$ intervals called {\it bands},

394: possibly touching at the endpoints.

395: In the case of irrational $\alpha$, the spectrum $\Sigma_{\lambda,\alpha}$

396: (which in this case does not depend on $\theta$) has been conjectured for

397: a long time to be a Cantor set (see a 1964 paper of Azbel \cite {Az}).

398: To prove this conjecture has been dubbed The Ten Martini Problem by

399: Barry Simon, after an offer of Mark Kac in 1981, see

400: Problem 4 in \cite {Sim2}.

401: %For a history of this problem, up to roughly 1994, see \cite {L1}.

402: For a history of this problem see \cite {L3}.

403: Earlier partial results include \cite {BS}, \cite {Sin}, \cite {HS},

404: \cite {CEY}, \cite {L},

405: and recent advances include \cite {P} and \cite {AK}.

406: In this paper, we solve the Ten Martini Problem as stated in \cite {Sim2}.

407: 

408: \begin{mainthm}

409: 

410: The spectrum of the almost Mathieu operator is a Cantor set for all

411: irrational $\alpha$ and for all $\lambda \neq 0$.

412: 

413: \end{mainthm}

414: 

415: It is important to emphasize that the previous results mentioned above

416: covered a large set of parameters $(\lambda,\alpha)$,

417: which is both topologically generic (\cite {BS}), and of full Lebesgue

418: measure (\cite {P}).  As it often happens in the analysis of quasiperiodic

419: systems, the ``topologically generic'' behavior is quite distinct from the

420: ``full Lebesgue measure'' behavior, and the narrow set of parameters left

421: behind does indeed lie in the interface of two distinct regimes.

422: Furthermore, our analysis seems to indicate an interesting characteristic

423: of the Ten Martini Problem, that the two regimes do not cover nicely the

424: parameter space and hence there is a non-empty ``critical region'' of

425: parameters in between (see Remarks \ref {remark2},

426: \ref {remark1}, \ref {appro} and the comments after

427: Theorem \ref {open gaps}).

428: 

429: This is to some degree reflected in the structure of the proof.

430: While the reasoning outside of the critical region can be made

431: quite effective, in the sense that one essentially identifies

432: specific gaps in the spectrum\footnote {Related either to gaps of periodic

433: approximations or to eigenvalues of a dual almost Mathieu operator.}, in

434: order to be able to cover the critical region we make use of very indirect

435: arguments.  As an example, we show that absence of Cantor spectrum enables

436: us to ``analytically continuate'' effective

437: solutions of a small divisor problem, and it is the

438: non-effective solutions thus obtained that can be related to gaps in the

439: spectrum.

440: 

441: This paper builds on a large theory.  Especially important for us are

442: \cite {CEY}, \cite {J}, \cite {P}, whose methods we improve, but several

443: other ingredients are needed (such as Kotani Theory \cite {Sim1},

444: the recent estimates on Lyapunov exponents of \cite {BJ1}).  An important

445: new ingredient is the use of analytic continuation techniques in the

446: study of $m$-functions and in extending the reach of the analysis of

447: Anderson localization.

448: 

449: \subsection{Strategy}

450: 

451: In this problem, arithmetics of $\alpha$ rules the game.  When

452: $\alpha$ is not very Liouville, it is reasonable to try to deal with the

453: small divisors.  When $\alpha$ is not very Diophantine, this does not work

454: and we deal instead with rational approximation arguments.

455: Let $\frac {p_n} {q_n}$ be the approximants of $\alpha \in \R \setminus \Q$. 

456: Let

457: \be

458: \beta=\beta(\alpha)=\limsup \frac {\ln q_{n+1}} {q_n}.

459: \ee

460: The relation between $e^\beta$ and $\lambda$ will play an important role in

461: our argument, and will decide whether we approach the problem from the

462: Diophantine side or from the Liouvillian side.

463: As discussed before, our analysis indicates that there are parameters

464: that can not be effectively described from either side,

465: and it is only through the use

466: of indirect arguments that we can enlarge artificially the Diophantine and

467: Liouville regimes to cover all parameters.  It should be noted that even

468: with such tricks, both sides will just about meet in the middle.

469: 

470: Since $\Sigma_{\lambda,\alpha}=\Sigma_{-\lambda,\alpha}$,

471: it is enough to assume $\lambda>0$.  It is known that the behavior of the

472: almost Mathieu operator changes drastically at $\lambda=1$

473: (``metal-insulator'' transition \cite {J}).

474: Aubry duality shows that

475: $\Sigma_{\lambda,\alpha}=\lambda \Sigma_{\lambda^{-1},\alpha}$.  So each

476: $\lambda \neq 1$ admits two lines of attack, and this will be determinant in

477: what follows.  The case    

478: $\lambda=1$ was settled in \cite {AK} (after several partial results 

479: \cite {AvMS}, \cite {HS}, \cite {L}), but it is also recovered in our

480: approach.

481: 

482: We will work on $\lambda<1$ when approaching from the

483: Liouville side.

484: The approach from the Diophantine side is more delicate.  There are actually

485: two classical small divisor problems that apply to the study of the almost

486: Mathieu operator, corresponding to Floquet reducibility (for $\lambda<1$)

487: and Anderson localization (for $\lambda>1$).  An important point

488: is to attack both problems simultaneously, mixing the best of each

489: problem (``soft'' analysis in one case, ``hard'' analysis in the other).

490: 

491: A key idea in this paper is that absence of Cantor spectrum implies

492: improved regularity of $m$-functions in the regime $0<\lambda \leq 1$.

493: This is proved by analytic continuation techniques.  The improved regularity

494: of $m$-functions (which is fictitious, since we will prove Cantor spectrum)

495: will be used both in the Liouville side and in the

496: Diophantine side.  In the Liouville side, it will

497: give improved estimates for the continuity of the spectrum with respect to

498: the frequency.  In the

499: Diophantine side, it will allow us to use (again) analytic continuation

500: techniques to solve some small divisor problems

501: in some situations which are beyond what is expected to be possible.

502: 

503: \begin{rem} \label {remark2}

504: 

505: Since our approach, designed to overcome the difficulties in the interface

506: of the Diophantine and Liouville regimes, works equally well for other

507: ranges of parameters, it will

508: not be necessary in the proof to precisely delimitate a critical region. 

509: For the reasons discussed in Remarks \ref {remark1}, \ref {appro} and in the

510: comments after Theorem \ref {open gaps}, the critical region is believed to

511: contain the parameters such that $\beta>0$ and

512: $\beta \leq |\ln \lambda| \leq 2 \beta$, the parameters such that

513: $\beta=|\ln \lambda|$ (respectively, $2 \beta=|\ln \lambda|$)

514: being seemingly inaccessible (even after artificial extension) by

515: the Diophantine method (respectively, Liouville method).  It is reasonable

516: to expect that something should be different in the indicated critical

517: region.  For instance, it is the natural place to look for possible

518: counterexamples to the ``Dry Ten Martini'' conjecture (for a precise

519: formulation see Section \ref{mainthm}).

520: 

521: \end{rem}

522: 

523: \section{Background}

524: 

525: \subsection{Cocycles, Lyapunov exponents, fibered rotation

526: number} \label {fibered}

527: 

528: A (one-dimensional quasiperiodic $\SL(2,\R)$) {\it cocycle}

529: is a pair $(\alpha,A) \in \R \times C^0(\R/\Z,\SL(2,\R))$,

530: understood as a {\it linear skew-product}:

531: \begin{align}

532: (\alpha,A):&\R/\Z \times \R^2 \to \R/\Z \times \R^2\\

533: \nonumber

534: &(x,w) \mapsto (x+\alpha,A(x) \cdot w).

535: \end{align}

536: For $n \geq 1$, we let

537: \be

538: A_n(x)=A(x+(n-1)\alpha) \cdots A(x)

539: \ee

540: ($\alpha$ is implicit in this notation).

541: 

542: Given two cocycles $(\alpha,A)$ and $(\alpha,A')$, a {\it conjugacy}

543: between them is a continuous $B:\R/\Z \to \SL(2,\R)$ such that

544: \be

545: B(x+\alpha)A(x)B(x)^{-1}=A'(x).

546: \ee

547: 

548: The {\it Lyapunov exponent} is defined by

549: \be

550: \lim \frac {1} {n} \int \ln \|A_n(x)\| dx,

551: \ee

552: so $L(\alpha,A) \geq 0$.  It is invariant under conjugacy.

553: 

554: Assume now that $A:\R/\Z \to \SL(2,\R)$ is homotopic to the identity.  Then

555: there exists $\psi:\R/\Z \times \R/\Z \to \R$ and $u:\R/\Z \times \R/\Z \to

556: \R^+$ such that

557: \be

558: A(x) \cdot \bm \cos 2 \pi y \\ \sin 2 \pi y \em=u(x,y)

559: \bm \cos 2 \pi (y+\psi(x,y)) \\ \sin 2 \pi (y+\psi(x,y)) \em.

560: \ee

561: The function $\psi$ is called a {\it lift} of $A$.  Let $\mu$ be any

562: probability on $\R/\Z \times \R/\Z$ which is invariant by the continuous

563: map $T:(x,y) \mapsto (x+\alpha,y+\psi(x,y))$, projecting over Lebesgue

564: measure on the first coordinate (for instance, take $\mu$ as any

565: accumulation point of $\frac {1} {n} \sum_{k=0}^{n-1} T_*^k \nu$ where

566: $\nu$ is Lebesgue measure on $\R/\Z \times \R/\Z$).  Then the number

567: \be

568: \rho(\alpha,A)=\int \psi d\mu \mod \Z

569: \ee

570: does not depend on the choices of $\psi$ and $\mu$, and is called the

571: {\it fibered rotation number} of

572: $(\alpha,A)$, see \cite {JM} and \cite {H}.

573: It is invariant under conjugacies homotopic to the identity.

574: It immediately follows from the

575: definitions that the fibered rotation number is a continuous function of

576: $(\alpha,A)$.

577: 

578: Notice that if $A,A':\R/\Z \to \SL(2,\R)$ and $B:\R/\Z \to \SL(2,\R)$

579: are continuous such that $A$ is homotopic to the identity and

580: $B(x+\alpha)A(x)B(x)^{-1}=A'(x)$, then

581: $\rho(\alpha,A)=\rho(\alpha,A')-k\alpha$, where $k$ is

582: such that $x \mapsto B(x)$ is homotopic to $x \mapsto R_{kx}$, where

583: \be

584: R_\theta=\bm \cos 2 \pi \theta&-\sin 2 \pi \theta \\ \sin 2 \pi \theta&

585: \cos 2 \pi \theta \em.

586: \ee

587: 

588: \subsection{Almost Mathieu cocycles, integrated density of states, spectrum}

589: \label {ids}

590: 

591: Let

592: \be \la{am}

593: S_{\lambda,E}=\bm E-2\lambda \cos 2 \pi x & -1 \\ 1 & 0 \em.

594: \ee

595: We call $(\alpha,S_{\lambda,E})$, $\lambda,\alpha,E \in \R$, $\lambda \neq

596: 0$ {\it almost Mathieu cocycles}.  A sequence $(u_n)_{n \in \Z}$

597: is a formal solution of the eigenvalue equation

598: $H_{\lambda,\alpha,\theta} u=Eu$ if and only if

599: $S_{\lambda,E}(\theta+n\alpha) \cdot \bm u_n\\u_{n-1} \em=\bm u_{n+1}\\u_n

600: \em$.

601: 

602: Let

603: \be

604: L_{\lambda,\alpha}(E)=L(\alpha,S_{\lambda,E}).

605: \ee

606: It is easy to see that

607: $\rho(\alpha,S_{\lambda,E})$ admits a determination

608: $\rho_{\lambda,\alpha}(E) \in [0,1/2]$.  We let

609: \be

610: N_{\lambda,\alpha}(E)=1-2\rho_{\lambda,\alpha}(E) \in [0,1].

611: \ee

612: It follows that $E \mapsto N_{\lambda,\alpha}(E)$ is a continuous

613: non-decreasing function.  The function $N$ is the usually defined

614: {\it integrated density of states} of $H_{\lambda,\alpha,\theta}$ if

615: $\alpha \in \R \setminus \Q$ (for $\alpha \in \Q$, $N$ is the integral of

616: the density of states over different $\theta$), see \cite {AS} and \cite

617: {JM}.  Thus defining

618: \be

619: \Sigma_{\lambda,\alpha}=\{E \in \R,\, N_{\lambda,\alpha} \text { is not

620: constant in a neighborhood of } E\},

621: \ee

622: we see that (consistently with the introduction)

623: $\Sigma_{\lambda,\alpha}$ is the spectrum of $H_{\lambda,\alpha,\theta}$ for

624: $\alpha \in \R \setminus \Q$ (in this case the spectrum does not depend on

625: $\theta$), while for $\alpha \in \Q$, $\Sigma_{\lambda,\alpha}$ is the union

626: of the spectra of $H_{\lambda,\alpha,\theta}$, $\theta \in \R$.

627: One also has

628: \be

629: \Sigma_{\lambda,\alpha} \subset [-2-2|\lambda|,2+2|\lambda|].

630: \ee

631: 

632: Continuity of the fibered rotation number implies that

633: $N_{\lambda,\alpha}$ depends continuously on $(\lambda,\alpha)$ on

634: $L^\infty(\R)$.

635: 

636: It turns out that there is a relation between $N$ and $L$, the {\it Thouless

637: formula}, see \cite {AS}

638: \be

639: L(E)=\int \ln |E-E'| dN(E').

640: \ee

641: By the Schwarz reflection principle, if $J \subset \R$ is an

642: open interval where the Lyapunov exponent vanishes, then $E \mapsto

643: N_{\lambda,\alpha}(E)$ is an increasing analytic function of

644: $E \in J$\footnote{Since $N+\frac {iL} {\pi}$

645: is holomorphic in upper half plane

646: and real on $J.$ This can also be obtained from the Thouless formula.} (and

647: obviously $J \subset \Sigma_{\lambda,\alpha}$).

648: 

649: We will use several times the following result

650: \cite {BJ1}.

651: 

652: \begin{thm}[\cite {BJ1}, Corollary 2] \label {L}

653: 

654: Let $\alpha \in \R \setminus \Q$, $\lambda \neq 0$.

655: If $E \in \Sigma_{\lambda,\alpha}$ then

656: \be

657: L_{\lambda,\alpha}(E)=\max \{0, \ln |\lambda|\}.

658: \ee

659: 

660: \end{thm}

661: 

662: This result will be mostly important for us for what it says about the range

663: $0<\lambda \leq 1$ (zero Lyapunov exponent on the spectrum).  It will be

664: also very minorly used in our proof of localization when $\lambda>1$.

665: 

666: \subsection{Kotani theory} \label {kota}

667: 

668: Recall the usual action of $\SL(2,\C)$ on the Riemann sphere $\overline \C$:

669: $\bm a&b\\c&d \em \cdot z=\frac {a z+b} {c z+d}$.

670: We can of course define $\SL(2,\C)$ cocycles as pairs $(\alpha,A) \in \R

671: \times C^0(\R/\Z,\SL(2,\C))$, but it is convenient to view a

672: $\SL(2,\C)$ cocycle as acting by Moebius transformations:

673: \begin{align}

674: (\alpha,A):&\R/\Z \times \overline \C \to \R/\Z \times \overline \C\\

675: \nonumber

676: &(x,z) \mapsto (x+\alpha,A(x) \cdot z).

677: \end{align}

678: If one lets $E$ become a complex number in the definition of the almost

679: Mathieu cocycle, we get a $\SL(2,\C)$ cocycle.

680: 

681: Let $\H$ be the upper half plane.  Fix $(\lambda,\alpha)$.

682: It is well known that there exists a continuous function

683: $m=m_{\lambda,\alpha}:\H \times \R/\Z

684: \to \H$ such that $S_{\lambda,E}(x) \cdot m(E,x)=m(E,x+\alpha)$, thus

685: defining an invariant section for the cocycle $(\alpha,S_{\lambda,E})$:

686: \be \label {minvariantsection}

687: (\alpha,S_{\lambda,E}) (x,m(E,x))=(x+\alpha,m(E,x+\alpha)).

688: \ee

689: Moreover, $E \mapsto m(E,x)$ is holomorphic on $\H$.

690: 

691: \begin{rem}

692: 

693: In the litterature (for instance, in \cite {Sim2}),

694: it is more common to find the definition of a pair of

695: $m$-functions, $m_\pm(x,E)$, which is given in

696: terms of non-zero solutions $(u_\pm(n))_{n \in \Z}$ of

697: $H_{\lambda,\alpha,x} u=Eu$

698: which are $\ell^2$ at $\pm \infty$: $m_\pm(x,E)=-

699: \frac {u_\pm(\pm 1)} {u_\pm(0)}$.  In this notation we have

700: $m(x,E)=-\frac {1} {m_-(x,E)}$ (the

701: relation $S_{\lambda,E}(x) \cdot m(E,x)=m(E,x+\alpha)$ is an immediate

702: consequence of the definition of $m_-(x,E)$).

703: 

704: \end{rem}

705: 

706: The following result

707: of Kotani theory \cite {Sim2} will be important in two key parts

708: of this paper.

709: 

710: \begin{thm} \label {kotani}

711: 

712: Let $\alpha \in \R \setminus \Q$, and assume that $L(\alpha,E)=0$ in an

713: open interval $J \subset \R$.  Then for

714: every $x \in \R/\Z$, the functions $E \mapsto m(E,x)$ admit a holomorphic

715: extension to $\C \setminus (\R \setminus J)$, with values in $\H$.  The

716: function $m:\C \setminus (\R \setminus J) \times \R/\Z \to \H$ is continuous

717: in both variables.

718: 

719: \end{thm}

720: 

721: \subsection{Polar sets}

722: 

723: Recall one of the possible definitions of

724: a polar set in $\C$: it is a set of zero {\it

725: logarithmic capacity}.  We will need only

726: some properties of polar sets in $\C$ (see for instance \cite {Ho}):

727: \begin{enumerate}

728: \item A countable union of polar sets is polar,

729: \item The image of a polar set by a non-constant holomorphic

730: function (defined in some domain of $\C$) is a polar set,

731: \item Polar sets have Hausdorff dimension zero,

732: thus their intersections with $\R$ have zero Lebesque measure,

733: \item Let $U \subset \C$ be a domain and let $f_n:U \to \R$

734: be a sequence of subharmonic functions which is uniformly bounded in

735: compacts of $U$.  Then $f:U \to \R$ given by $f=\limsup f_n$ coincides with

736: its (subharmonic) upper regularization

737: $f^*:U \to \R$ (given by $f^*(z)=\limsup_{w \to z}

738: f(w)$) outside a polar set.

739: \end{enumerate}

740: We will say that a subset of $\R$ is polar if it is polar as a subset of

741: $\C$.

742: 

743: The following result on analytic continuation is well known.  We will

744: quickly go through the proof, since a similar idea will play a role later

745: in a small divisor problem.

746: 

747: \begin{lemma} \label {anal}

748: 

749: Let $W \subset \C$ be a domain and let $f:W \times \R/\Z \to \C$ be a

750: continuous function.  If $z \mapsto f(z,w)$ is holomorphic for all $w \in \R/\Z$ and $w

751: \mapsto f(z,w)$ is analytic for some non-polar set of $z \in W$ then $f$

752: is analytic.

753: 

754: \end{lemma}

755: 

756: \begin{pf}

757: 

758: We may assume that $|f(z,w)|<1$, $(z,w) \in W \times \R/\Z$.  Let

759: \be \label {C/n}

760: f(z,w)=\sum \hat f_z(k) e^{2 \pi i k w}.

761: \ee

762: Then $z \mapsto \hat f_z(k)$ is holomorphic and $|\hat f_z(k)|<1$.

763: Using property (1) of polar sets, 

764: %and passing if needed to a compact nonpolar subset 

765: we obtain that there exists a non-polar set 

766: $\Delta \subset W$,

767: $\epsilon>0$, and $k>0$ such that

768: $|\hat f_z(n)| \leq e^{-\epsilon |n|}$ for $z \in \Delta$ and $|n|>k$.

769: Let

770: \be

771: h(z)=\sup_{|n|>k} \frac {1} {|n|} \ln |\hat f_z(n)|.

772: \ee

773: Then, by property (4) of polar sets,

774: $h^*$ is a non-positive subharmonic function satisfying

775: $h^*(z) \leq -\epsilon$, $z \in \Delta \setminus X$, where $X$ is polar. 

776: Since $\Delta$ is non-polar, we conclude that $h^*$ is not identically $0$

777: in $W$.  It follows from the maximum principle that $h^*(z)<0$, $z \in W$.

778: Thus for any domain $U \subset W$ compactly contained in $W$,

779: there exists $\delta=\delta(U)>0$ such that

780: $h(z) \leq -\delta$, $z \in U$.  We conclude that

781: \be

782: \frac {1} {|n|} \ln |\hat f_z(n)| \leq -\delta, \quad

783: |n|>k, z \in U,

784: \ee

785: which implies that (\ref {C/n}) converges uniformly on compacts of

786: $W \times \{w \in \C/\Z,\, 2 \pi |\Im w|<\delta\}$.

787: \end{pf}

788: 

789: \section{Regularity of the $m$-functions}

790: 

791: \begin{thm} \label {mregu}

792: 

793: Let $\alpha \in \R \setminus \Q$, $\lambda>0$.  Let $m=m_{\lambda,\alpha}:\H

794: \times \R/\Z \to \H$ be as in \S \ref {kota}.  Then $m$ is analytic.

795: 

796: \end{thm}

797: 

798: \begin{pf}

799: 

800: Let us show that $m$ has a holomorphic extension to

801: \be

802: \Omega_\lambda=\{(E,x),\, \Im E>0,\, 2 \lambda \sinh |2 \pi \Im x|<

803: \Im E\}.

804: \ee

805: 

806: We have

807: \be

808: S_{\lambda,E}(x) \cdot z=E-2 \lambda \cos (2 \pi x)-\frac {1} {z}.

809: \ee

810: For $(E,t)$ satisfying

811: \be \label {E,t}

812: \Im E>0, \quad 2 \lambda \sinh |2 \pi t|<\Im E,

813: \ee

814: define the half-plane

815: \be

816: K^1_{\lambda,E,t}=\{z,\, \Im z>\Im E-2\lambda \sinh |2 \pi t|\} \subset \H,

817: \ee

818: and the disk

819: \be

820: K^2_{\lambda,E,t}=\left \{|z|<|E|+2 \lambda \cosh |2 \pi t|+\frac

821: {1} {\Im E-2 \lambda \sinh |2 \pi t|} \right \},

822: \ee

823: and let

824: \be

825: K_{\lambda,E,t}=K^1_{\lambda,E,t} \cap K^2_{\lambda,E,t},

826: \ee

827: which is a domain compactly contained in $\H$ depending continuously on

828: $(E,t)$ satisfying (\ref {E,t}).

829: If $(E,x) \in \Omega_\lambda$ then $(E,\Im x)$ satisfies (\ref {E,t}) and

830: one checks directly that

831: \be

832: S_{\lambda,E}(x) \cdot \H \subset K^1_{\lambda,E,\Im x},

833: \ee

834: \be

835: S_{\lambda,E}(x) \cdot K^1_{\lambda,E,\Im x}

836: \subset K^2_{\lambda,E,\Im x}.

837: \ee

838: Since $\Im x=\Im x+\alpha$, we have

839: \be

840: S_{\lambda,E}(x+\alpha) \cdot S_{\lambda,E}(x) \cdot

841: \H \subset K_{\lambda,E,\Im x}.

842: \ee

843: 

844: Thus, by the Schwarz Lemma applied to $\H$, for every $(E,x) \in \Omega_\lambda$,

845: \be

846: S_{\lambda,E}(x-\alpha) \cdots S_{\lambda,E}(x-n\alpha) \cdot

847: \overline \H

848: \ee

849: is a sequence of nested compact sets shrinking to a single point $\hat m(E,x)$.

850: This implies that $\hat m(E,x)$ is the unique solution to

851: (\ref{minvariantsection}) in $\H.$

852: Since $m:\H \times \R/\Z \to \H$ is a continuous function satisfying (\ref

853: {minvariantsection}),

854: %\be \la{311}

855: %S_{\lambda,E} \cdot (x,m(E,x))=(x+\alpha,m(E,x+\alpha)),

856: %\ee

857: we must have $\hat m(E,x)=m(E,x)$ for $(E,x) \in \H \times \R/\Z$.

858: 

859: Since holomorphic functions

860: $m^n:\Omega_\lambda \to \H$ given by

861: \be

862: m^n(E,x)=S_{\lambda,E}(x-\alpha) \cdots S_{\lambda,E}(x-n\alpha) \cdot i,

863: \ee

864: take values in $\H,$ the sequence $m^n$ is normal. Since it converges pointwise to $\hat m$, we conclude that $\hat m$ is

865: holomorphic.

866: \end{pf}

867: 

868: \begin{thm} \label {bjkot}

869: 

870: Let $\alpha \in \R \setminus \Q$, and let $0<\lambda \leq 1$.  Let

871: $m=m_{\lambda,\alpha}:\H \times \R/\Z \to \H$ be as in \S \ref {kota}.

872: If $J \subset \Sigma_{\lambda,\alpha}$ is an open interval then $m$ admits

873: an analytic extension

874: $m:\C \setminus (\R \setminus J) \times \R/\Z \to \H$.

875: 

876: \end{thm}

877: 

878: \begin{pf}

879: 

880: By Theorems \ref {L} and \ref {kotani}, there exists a continuous

881: extension $m:\C \setminus (\R \setminus J) \times \R/\Z \to \H$

882: which is analytic in $E$. 

883: By Theorem \ref {mregu}, $m$ is also analytic in $x$ for $E \in \H$.

884: Analiticity in $(E,x)$ then follows by Lemma \ref {anal}.

885: \end{pf}

886: 

887: \begin{rem}

888: 

889: Notice that the proof of Theorem \ref {mregu} uses strongly that the

890: dynamics in the basis of the almost Mathieu cocycle

891: is a rotation (and not, say, a hyperbolic toral automorphism or the skew

892: shift).  But one may still get weaker results on smoothness of

893: $m$-functions (in the line of Theorem \ref {bjkot}) for those

894: dynamics (via estimates in the line of \cite {AK2}).

895: 

896: \end{rem}

897: 

898: \section{Analytic continuation}

899: 

900: \begin{lemma}

901: 

902: Let $\alpha \in \R \setminus \Q$ and let $\phi:\R/\Z \to \R$ be analytic,

903: and let $\theta=\int_{\R/\Z} \phi(x) dx$.  The following are equivalent:

904: \begin{enumerate}

905: \item There exists an analytic function $O:\R/\Z \to \SL(2,\R)$, homotopic

906: to the identity, such that

907: \be

908: O(x+\alpha) R_{\phi(x)} O(x)^{-1}=R_\theta,

909: \ee

910: \item There exists an analytic function $\psi:\R/\Z \to \R$ such that

911: \be \label {psi}

912: \phi(x)-\theta=\psi(x+\alpha)-\psi(x).

913: \ee

914: \end{enumerate}

915: 

916: \end{lemma}

917: 

918: \begin{pf}

919: 

920: Obviously (2) implies (1): it is enough to take $O(x)=R_{-\psi(x)}$.

921: 

922: Let us show that (1) implies (2).

923: If $O(x) \cdot i=i$ for all $x$ then $O(x) \in \SO(2,\R)$ for all $x$ and

924: since $O$ is homotopic to the identity we have

925: $O(x)=R_{-\psi(x)}$ for some analytic function

926: $\psi:\R/\Z \to \R$ which has to satisfy

927: $\phi(x)-\theta=\psi(x+\alpha)-\psi(x)$.

928: 

929: Thus we may assume that $O(x_0) \cdot i \neq i$ for some $x_0$.

930: Notice that

931: \be

932: O(x_0+n\alpha) \cdot i=R_{n \theta} O(x_0) \cdot i.

933: \ee

934: It follows that if $n_k \alpha \to 0$ in $\R/\Z$ then $2n_k\theta \to 0$ in

935: $\R/\Z$.  This implies that $\theta=\frac {l} {2} \alpha$ for some $l \in

936: \Z$.  We have

937: \be

938: O(x+\alpha)R_{\phi(x)}O(x)^{-1}=R_{\frac {l} {2} (x+\alpha)} R_{-\frac {l}

939: {2} x},

940: \ee

941: which implies

942: \be

943: R_{-\frac {l} {2} (x+\alpha)} O(x+\alpha) R_{\phi(x)}=R_{-\frac {l} {2} x}

944: O(x),

945: \ee

946: and we get

947: \be

948: R_{-\frac {l} {2} (x+\alpha)} O(x+\alpha) \cdot i=R_{-\frac {l} {2} x} O(x)

949: \cdot i.

950: \ee

951:  It follows that $R_{-\frac {l} {2} x} O(x) \cdot i=z$ does not depend on

952: $x$.  Let $Q \in \SL(2,\R)$ be such that $Q \cdot z=i$, and set

953: \be

954: S(x)=R_{\frac {l} {2} x} Q R_{-\frac {l} {2} x} O(x).

955: \ee

956: Since $O,Q:\R/\Z\to \SL(2,\R),$ where $Q(x)=Q$ are homotopic to the identity, we have that  $S:\R/\Z \to \SL(2,\R)$ is homotopic to the identity and, using that $\theta=\frac{l}2 \alpha,$ we have 

957: \be

958: S(x+\alpha)R_{\phi(x)}S(x)^{-1}=R_\theta.

959: \ee

960: Moreover, $S(x) \cdot i=i$, so $S(x) \in \SO(2,\R)$ and we have

961: $S(x)=R_{-\psi(x)}$, $\psi:\R/\Z \to \R$.  It follows that $\psi$ satisfies

962: (\ref {psi}).

963: \end{pf}

964: 

965: For $\alpha \in \R \setminus \Q$, $0<\lambda \leq 1$, let

966: $\Lambda_{\lambda,\alpha}$ be the set of $E$ such that there exists an

967: analytic function $B_E:\R/\Z \to \SL(2,\R)$, homotopic to the identity, and

968: $\theta(E) \in \R$, such that

969: \be

970: B_E(x+\alpha)S_{\lambda,E}(x)B_E(x)^{-1}=R_{\theta(E)}.

971: \ee

972: 

973: \begin{thm} \label {beta}

974: 

975: Let $\alpha \in \R \setminus \Q$, $0<\lambda \leq 1$.  Let $J \subset

976: \Sigma_{\lambda,\alpha}$ be an open interval.  Then

977: \begin{enumerate}

978: \item If $\beta=0$ then $\Lambda_{\lambda,\alpha}\cap J=J$,

979: \item If $\beta<\infty$ then either $\Lambda_{\lambda,\alpha}\cap J$

980: is polar or $\inter \Lambda_{\lambda,\alpha} \cap J \neq \emptyset$.

981: \end{enumerate}

982: 

983: \end{thm}

984: 

985: \begin{pf}

986: 

987: Assume that $J \subset \Sigma_{\lambda,\alpha}$ is an open interval.

988: Let $m=m_{\lambda,\alpha}$ be given

989: by Theorem \ref {bjkot}, so that

990: $m:\C \setminus (\R \setminus J) \times

991: \R/\Z \to \H$ is continuous, $E \mapsto m(E,x)$ is holomorphic and

992: \be

993: S_{\lambda,E} \cdot m(E,x)=m(E,x+\alpha).

994: \ee

995: Let

996: \be \label {C_E definition}

997: C_E(x)=\bm \frac {\Re m(E,x)} {|m(E,x)| (\Im m(E,x))^{1/2}}

998: & -\frac {|m(E,x)|} {(\Im m(E,x))^{1/2}}\\

999: \frac {(\Im m(E,x))^{1/2}} {|m(E,x)|}

1000: & 0 \em.

1001: \ee

1002: Then

1003: \be

1004: C_E(x+\alpha) S_{\lambda,E}(x) C_E(x)^{-1} \in \SO(2,\R)

1005: \ee

1006: for $E \in J$, $x \in \R/\Z$.  Since $x \mapsto C_E(x)$ is easily verified to be homotopic

1007: to the identity for $E \in J$, we have

1008: \be

1009: C_E(x+\alpha) S_{\lambda,E}(x) C_E(x)^{-1}=R_{\phi(E,x)}

1010: \ee

1011: for some real-analytic function $\phi:J \times \R/\Z \to \R$.  It follows

1012: that $\phi$ has a holomorphic extension $\phi:Z \to \C$ where

1013: $Z \subset \C \times \C/\Z$ is some domain containing

1014: $J \times \R/\Z$.  So there exists a domain $\Delta \subset \C$

1015: such that $J \subset \Delta$ and

1016: $\Delta \times \R/\Z \subset Z$.  For $E \in \Delta$, let

1017: \be

1018: \phi(E,x)=\sum \hat \phi_E(k) e^{2 \pi i k x}.

1019: \ee

1020: 

1021: Let $E \in J$ be such that there exists an analytic function

1022: $\psi_E:\R/\Z \to \R$ such that

1023: \be

1024: \int_{\R/\Z} \psi_E(x) dx=0,

1025: \ee

1026: \be

1027: \phi(E,x)-\int_{\R/\Z} \phi(E,x) dx=\psi_E(x+\alpha)-\psi_E(x).

1028: \ee

1029: Then

1030: %By the previous lemma, this implies that

1031: \be \label {psi1}

1032: \psi_E(x)=\sum \hat \psi_E(k) e^{2 \pi i k x}

1033: \ee

1034: where

1035: \be \label {psi2}

1036: \hat \psi_E(k)=\frac {\hat \phi_E(k)} {e^{2 \pi i k \alpha}-1},

1037: \quad k \neq 0,

1038: \ee

1039: \be \label {psi3}

1040: \hat \psi_E(0)=0.

1041: \ee

1042: We can then define an analytic function $B_E:\R/\Z \to \SL(2,\R)$ by

1043: \be

1044: B_E(x)=R_{-\psi_E(x)} C_E(x),

1045: \ee

1046: which satisfies

1047: \be

1048: B_E(x+\alpha) S_{\lambda,E}(x) B_E(x)^{-1}=R_{\theta(E)}, \quad

1049: \theta(E)=\int_{\R/\Z} \phi(E,x) dx.

1050: \ee

1051: 

1052: Reciprocally, if there exists an analytic function

1053: $B_E:\R/\Z \to \SL(2,\R)$ homotopic to the identity such that $B_E(x+\alpha)

1054: S_{\lambda,E}(x) B_E(x)^{-1}=R_{\theta(E)}$ for some $\theta(E) \in \R$,

1055: then we can write

1056: \be

1057: O_E(x+\alpha) R_{\phi(x)} O_E(x)^{-1}=R_{\theta(E)},

1058: \ee

1059: where

1060: \be

1061: O_E(x)=B_E(x) C_E(x)^{-1}.

1062: \ee

1063: By the previous lemma, there exists an analytic

1064: function (having average $0$)

1065: $\psi:\R/\Z \to \R$ satisfying $\phi(x)-\int_{\R/\Z} \phi(x)

1066: dx=\psi(x+\alpha)-\psi(x)$.

1067: 

1068: Notice that

1069: \be

1070: \limsup_{|k| \to \infty} \frac {1} {|k|} \ln \frac {1} {|e^{2 \pi i

1071: k\alpha}-1|}=\beta,

1072: \ee

1073: so that if $\beta=0$ then (\ref {psi1}), (\ref {psi2}), and (\ref {psi3})

1074: really define an analytic function for any $E \in J$, so (1) follows.

1075: 

1076: Let $a:\Delta \to [-\infty,\beta]$ be given by

1077: \be

1078: a(E)=\limsup_{|k| \to \infty} \frac {1} {|k|} \ln \left |

1079: \frac {\hat \phi_E(k)} {e^{2 \pi i k\alpha}-1} \right |.

1080: \ee

1081: By the previous discussion, $\Lambda_{\lambda,\alpha}=\{E \in J,\,

1082: a(E)<0\}$.  If $\beta<\infty$ then $a$ is $\limsup$ of a

1083: sequence of subharmonic functions which are uniformly

1084: bounded on compacts of $\Delta$.  It follows that $a$ coincides with

1085: its upper regularization

1086: \be

1087: a^*(E)=\limsup_{E' \to E} a(E')

1088: \ee

1089: for $E$ outside some exceptional set which is polar.  Thus the set

1090: $\{E \in J,\, a(E)<0\}$ is either polar (contained in the exceptional set)

1091: or it has non-empty interior.

1092: \end{pf}

1093: 

1094: \begin{lemma} \label {interior}

1095: 

1096: Let $\alpha \in \R \setminus \Q$, $\lambda>0$.  Then

1097: $\Lambda_{\lambda,\alpha}$ has empty interior.

1098: 

1099: \end{lemma}

1100: 

1101: \begin{pf}

1102: 

1103: We may assume that $0<\lambda \leq 1$ (otherwise the Lyapunov exponent is

1104: positive on $\Sigma_{\lambda,\alpha}$ which easily implies that

1105: $\Lambda_{\lambda,\alpha}=\emptyset$).

1106: Assume that $J \subset \Lambda_{\lambda,\alpha}$ is an open interval.

1107: Then $J \subset \Sigma_{\lambda,\alpha}$

1108: (since $L_{\lambda,\alpha}(E)=0$ for $E \in J$).  Let $B_E$ be as in the

1109: definition of $\Lambda_{\lambda,\alpha}$.  Then the definition of fibered

1110: rotation number (see \S \ref {fibered}) implies

1111: \be

1112: \rho_{\lambda,\alpha}(E)=\theta(E) (\mod \Z).

1113: \ee

1114: By the analyticity of $\rho$ on $J$ there exists $E \in J$, $l \in \Z$,

1115: such that $\theta(E)=l \alpha (\mod \Z)$.

1116: Let $T_E:\R/\Z \to \SL(2,\R)$ be given by

1117: \be

1118: T_E(x)=R_{-l x} B_E(x).

1119: \ee

1120: Then

1121: \be

1122: T_E(x+\alpha) S_{\lambda,E}(x) T_E(x)^{-1}=\id.

1123: \ee

1124: The conclusion is as in \cite {P}.  For $v \in \R^2$,

1125: \be

1126: S_{\lambda,E}(x) T_E(x)^{-1}v=T_E(x+\alpha)^{-1} v.

1127: \ee

1128: So by (\ref{am}) there exists an analytic $U_v:\R/\Z \to \R$ such that

1129: \be

1130: T_E(x+\alpha)^{-1} \cdot v=\bm U_v(x) \\ U_v(x-\alpha) \em.

1131: \ee

1132: Let

1133: \be

1134: U_v(x)=\sum u^v_n e^{2 \pi inx}.

1135: \ee

1136: It is a standard Aubry duality argument (and can be checked by direct

1137: calculation) that $u_n^v\in\ell^2(\Z)$ is an eigenvector of

1138: $H_{\lambda^{-1},\alpha,0} $ with eigenvalue $\lambda^{-1}E.$  The fact

1139: that we get such an eigenvector for every $v\in\R^2$ contradicts the

1140: simplicity of the point spectrum.

1141: %Then an immediate computation gives that

1142: %\be

1143: %S_{\lambda^{-1},\lambda^{-1} E}(n\alpha)

1144: %\bm u^v_n\\u^v_{n-1}\em=\bm

1145: %u^v_{n+1}\\u^v_n\em.

1146: %\ee

1147: %Choose $v,w \in \R^2$ which are linearly independent.  We have

1148: %\be

1149: %S_{\lambda^{-1},\lambda^{-1} E}(n\alpha)

1150: %\bm u^v_n&u^w_n\\u^v_{n-1}&u^w_{n-1}\em=\bm

1151: %u^v_{n+1}&u^w_{n+1}\\u^v_n&u^w_n\em

1152: %\ee

1153: %which implies that

1154: %\be

1155: %\det \bm u^v_n&u^w_n\\u^v_{n-1}&u^w_{n-1}\em=\lim_{m \to \infty}

1156: %\det \bm u^v_{n+m}&u^w_{n+m}\\u^v_{n+m-1}&u^w_{n+m-1}\em=0,

1157: %\ee

1158: %since $\lim_{n \to \infty} u^v_n=\lim_{n \to \infty} u^w_n=0$.

1159: %Thus $(u^v_n)_{n \in \Z}$ and $(u^w_n)_{n \in

1160: %\Z}$ are not linearly independent.  This implies that

1161: %$T_E(x)^{-1} \cdot v$ and $T_E(x)^{-1} \cdot w$ are not linearly independent

1162: %for $x \in \R/\Z$, which is a contradiction.\footnote{From the spectral

1163: %viewpoint, we have constructed a two-dimensional family

1164: %$(u^v_n)_{n \in \Z}$, $v \in \R^2$ of eigenvectors for

1165: %$H_{\lambda^{-1},\alpha,0}$ with eigenvalue $\lambda^{-1} E$, contradicting

1166: %the simplicity of the point spectrum.}

1167: \end{pf}

1168: 

1169: \begin{rem} \label {blu}

1170: 

1171: Notice that Lemma \ref {interior} and item (1) of

1172: Theorem \ref {beta} already imply the

1173: Ten Martini Problem in the case $\beta=0$, and we did not need any

1174: localization result (the only recent result we used was Theorem \ref {L}).

1175: 

1176: \end{rem}

1177: 

1178: \section{Localization and Cantor spectrum}

1179: 

1180: We say that the operator $H_{\lambda,\alpha,\theta}$ displays {\it Anderson

1181: localization} if it has pure point spectrum with exponentially decaying

1182: eigenvectors.  This requires $\alpha \in \R \setminus \Q$, and implies

1183: that eigenvalues are dense in $\Sigma_{\lambda,\alpha}$.

1184: 

1185: \begin{thm} \label {cantor localization}

1186: 

1187: Let $\alpha \in \R \setminus \Q$, and let $\lambda \geq 1$.  Assume that

1188: $\beta<\infty$.  If $H_{\lambda,\alpha,\theta}$ displays Anderson

1189: localization for a non-polar set

1190: of $\theta \in \R$, then $\Sigma_{\lambda,\alpha}$ is a Cantor set.

1191: 

1192: \end{thm}

1193: 

1194: \begin{pf}

1195: 

1196: Let $\Theta$ be the set of $\theta$ such that $H_{\lambda,\alpha,\theta}$

1197: displays Anderson localization.  If $\theta \in \Theta$, and $E$ is an

1198: eigenvalue for $H_{\lambda,\alpha,\theta}$, let $(u_n)_{n \in \Z}$ be a

1199: non-zero eigenvector.  Then

1200: \be

1201: S_{\lambda^{-1},\lambda^{-1}E} \cdot W(x)=e^{2 \pi i \theta}

1202: W(x+\alpha),

1203: \ee

1204: where

1205: \be

1206: W(x)=\bm U(x) e^{2\pi i \theta}\\

1207: U(x-\alpha)\em,

1208: \ee

1209: and

1210: \be

1211: U(x)=\sum u_n e^{2 \pi i n x}.

1212: \ee

1213: Let $M(x)$ be the matrix with columns $W(x)$ and $\overline {W(x)}$.  Then

1214: \be

1215: S_{\lambda^{-1},\lambda^{-1}E}(x) \cdot M(x)=M(x+\alpha)

1216: \bm e^{2 \pi i \theta}&0\\0&e^{-2 \pi i \theta} \em.

1217: \ee

1218: This implies that $\det M(x)$ is independent of $x$, so

1219: $\det M(x)=c i$ for some $c \in \R$.

1220: Notice that if $c=0$ then

1221: \be

1222: V(x+\alpha)=e^{-4 \pi i \theta} V(x),

1223: \ee

1224: with

1225: \be

1226: V(x)=\frac {U(x)} {\overline {U(x)}}

1227: \ee

1228: (notice that $U(x) \neq 0$ except at finitely many $x$ since $U(x)$ is a

1229: non-constant analytic function)

1230: and in particular, if $n_k \alpha \to 0$ then $2 n_k \theta \to 0$.  So

1231: $2 \theta=k \alpha+l$ for some $k,l \in \Z$.

1232: If $c>0$, we have

1233: \be

1234: S_{\lambda^{-1},\lambda^{-1}E}(x)=Q(x+\alpha) R_\theta Q(x)^{-1}

1235: \ee

1236: where $Q:\R/\Z \to \SL(2,\R)$ is given by

1237: \be

1238: Q(x)=\frac {1} {(2 c)^{1/2}} M(x) \bm 1 & i \\ 1 & -i \em

1239: \ee

1240: and if $c<0$, we have

1241: \be

1242: S_{\lambda^{-1},\lambda^{-1}E}(x)=Q(x+\alpha) R_{-\theta} Q(x)^{-1}

1243: \ee

1244: where $Q:\R/\Z \to \SL(2,\R)$ is given by

1245: \be

1246: Q(x)=\frac {1} {(-2 c)^{1/2}} M(x) \bm 1 & i \\ 1 & -i \em

1247: \bm 1 & 0 \\ 0 & -1 \em.

1248: \ee

1249: It follows that in either case $\lambda^{-1} E

1250: \in \Lambda_{\lambda^{-1},\alpha}$ and moreover,

1251: \be

1252: \rho_{\lambda^{-1},\alpha}(\lambda^{-1}E)=\pm \theta+k \alpha (\mod \Z)

1253: \ee

1254: for some $k \in \Z$.

1255: 

1256: Let $\Theta' \subset \Theta$ be the set of all $\theta$ such that $2 \theta

1257: \neq k\alpha+l$ for all $k,l \in \Z$.

1258: Let $J \subset \Sigma_{\lambda,\alpha}$ be an open interval.

1259: Then for any $\theta \in \Theta'$, there exists some

1260: $E \in J$ such that $E$ is an eigenvalue for

1261: $H_{\lambda,\alpha,\theta}$, and by the previous discussion any such $E$

1262: satisfies

1263: \be

1264: N_{\lambda^{-1},\alpha}(\lambda^{-1}E)=1-2\rho_{\lambda^{-1},\alpha}

1265: (\lambda^{-1}E)=1-2(\varepsilon \theta+k \alpha+l),

1266: \quad \text {for some } k,l \in \Z, \varepsilon \in \{1,-1\},

1267: \ee

1268: \be

1269: \lambda^{-1}E \in \Lambda_{\lambda^{-1},\alpha}.

1270: \ee

1271: It follows that

1272: \be

1273: \Theta' \subset

1274: %\cup_{k,l \in \Z, \varepsilon \in \{1,-1\}}

1275: \left \{\varepsilon \frac{1-N_{\lambda^{-1},\alpha}

1276: (\Lambda_{\lambda^{-1},\alpha} \cap \lambda^{-1}J)}2-k\alpha-l,\quad

1277: k,l \in \Z, \varepsilon \in \{1,-1\} \right \}.

1278: \ee

1279: By item (2) of

1280: Theorem \ref {beta} and Lemma \ref {interior},

1281: $\Lambda_{\lambda^{-1},\alpha} \cap \lambda^{-1}J$ is polar.

1282: Since $N_{\lambda^{-1},\alpha}$ is a non-constant analytic function on

1283: $\lambda^{-1}J$, it

1284: follows that $\Theta'$ is also polar.  Thus $\Theta \subset \Theta'

1285: \cup \left \{\frac {1} {2} (k\alpha+l),\, k,l \in \Z \right \}$

1286: is polar.

1287: \end{pf}

1288: 

1289: \begin{rem} \label {remark1}

1290: 

1291: In \cite {P}, it is shown that if $\alpha \in DC$ then Anderson localization

1292: of $H_{\lambda,\alpha,0}$ implies Cantor spectrum.

1293: We can not however use the argument of Puig

1294: (based on analytic reducibility) 

1295: to conclude Cantor spectrum in the generality we need.

1296: Indeed, we are not able to conclude analytic      

1297: reducibility from localization of $H_{\lambda,\alpha,0}$ in our setting

1298: (in a sense, we spend all our regularity to take care of 

1299: small divisors in the localization result, which is half of analytic

1300: reducibility, and there is nothing left for the other half).  Though this

1301: can be bypassed (using Kotani theory to conclude continuous reducibility

1302: under the assumption of non-Cantor spectrum), there is a much more serious  

1303: difficulty in this approach, see Remark \ref {appro}.

1304: 

1305: \end{rem}

1306: 

1307: The next result gives us a large range of $\lambda$ and $\alpha$ where

1308: Theorem~\ref {cantor localization} can be applied.

1309: 

1310: \begin{thm} \label {localization}

1311: 

1312: Let $\alpha \in \R \setminus \Q$ be such that $\beta=\beta(\alpha)<\infty$,

1313: and let $\lambda>e^{\frac {16} {9} \beta}$.

1314: Then $H_{\lambda,\alpha,\theta}$ displays Anderson localization for almost

1315: every $\theta$.

1316: 

1317: \end{thm}

1318: 

1319: This result improves on \cite {J}, where Anderson localization

1320: was proved under

1321: the assumption that $\alpha$ is Diophantine.  Recall that $\alpha$ is said

1322: to satisfy a Diophantine condition (briefly, $\alpha \in DC$) if

1323: \be

1324: \ln q_{n+1}=O(\ln q_n)

1325: \ee

1326: where $\frac {p_n} {q_n}$ are the rational approximations of $\alpha$.

1327: In particular $\alpha \in DC$ implies (but is strictly stronger than)

1328: $\beta(\alpha)=0$. The proof in \cite{J} with some modifications can be extended to the case $\beta(\alpha)=0$ but not to the case $\beta(\alpha)>0.$

1329: 

1330: The proof of Theorem~\ref {localization} is the most technical part of this

1331: paper, and the considerations involved are independent from our other

1332: arguments.  We will thus postpone its proof to \S~\ref {proof of

1333: localization}.

1334: 

1335: \begin{rem} \label {appro}

1336: 

1337: We expect that the operator $H_{\lambda,\alpha,0}$ does not display Anderson

1338: localization for $1<\lambda \leq e^{2 \beta}$.  The key reason is that in

1339: this regime $0$ is a very resonant phase, and since $\alpha$ is Diophantine 

1340: only in a very weak sense, the compound effect on the small divisors can  

1341: not be compensated by the Lyapunov exponent.  See also Remark~\ref {approx1}.

1342: 

1343: \end{rem}

1344: 

1345: \section{Fictitious results on continuity of the spectrum}

1346: 

1347: The spectrum $\Sigma_{\lambda,\alpha}$ is a continuous function of $\alpha$

1348: in the Hausdorff topology.

1349: There are several results in the literature about quantitative continuity. 

1350: The best general result is due to \cite {AvMS}, $1/2$-H\"older continuity. 

1351: Better estimates can be obtained for $\alpha$ not very Liouville in

1352: the region of positive Lyapunov exponent \cite {JK}.

1353: None of those results are enough for our purposes.

1354: 

1355: The results described above have something in common: they deal with

1356: something that actually happens, and it is not clear if it is possible to

1357: improve them sufficiently (to the level we need).  Thus we will

1358: argue by contradiction: assuming the spectrum is not Cantor, we will get

1359: very good continuity estimates.  This will allow us to proceed the argument,

1360: but obviously, since we will eventually conclude that the spectrum is a

1361: Cantor set, estimates in this section are not valid for any

1362: existing almost Mathieu operator.  Those estimates might be useful also when

1363: analyzing more general Schr\"odinger operators.

1364: 

1365: \begin{thm} \label {continuity}

1366: 

1367: Let $\alpha \in \R \setminus \Q$ and $0<\lambda \leq 1$.  Let

1368: $J \subset \R$ be an open interval such that

1369: $\overline J \subset \inter \Sigma_{\lambda,\alpha}$.  There exists

1370: $K>0$ such that

1371: \be

1372: |N_{\lambda,\alpha}(E)-N_{\lambda,\alpha'}(E)|

1373: \leq K |\alpha-\alpha'|, \quad E \in J.

1374: \ee

1375: 

1376: \end{thm}

1377: 

1378: \begin{pf}

1379: 

1380: Let $m=m_{\lambda,\alpha}$ be as in Theorem \ref {bjkot}.  Define $x \mapsto

1381: C_E(x)$ by (\ref {C_E definition}).

1382: %Let

1383: %\be

1384: %C_E(x)=\bm \frac {\Re m(E,x)} {|m(E,x)| (\Im m(E,x))^{1/2}}

1385: %& -\frac {|m(E,x)|} {(\Im m(E,x))^{1/2}}\\

1386: %\frac {(\Im m(E,x))^{1/2}} {|m(E,x)|}

1387: %& 0 \em.

1388: %\ee

1389: Then, as discussed in the proof of Theorem \ref{beta}, $C_E:\R/\Z \to \SL(2,\R)$ is homotopic to the identity and satisfies

1390: $C_E(x+\alpha)S_{\lambda,E}(x)C_E(x)^{-1} \in \SO(2,\R)$ so

1391: \be

1392: C_E(x+\alpha)S_{\lambda,E}(x)C_E(x)^{-1}=R_{\phi_E(x)},

1393: \ee

1394: where $\phi_E:\R/\Z \to \R$ is analytic.

1395: Recall the definition of the fibered rotation number \S \ref {fibered}.

1396: Then

1397: \be

1398: \rho(\alpha,S_{\lambda,E}(x))=\rho(\alpha,R_{\phi(x)}).

1399: \ee

1400: In this case we can take as lift of $R_{\phi_E(x)}$ the function

1401: $\psi(x,y)=\phi(x)$.

1402: 

1403: Write

1404: \be

1405: \rho(\alpha',S_{\lambda,E})=\rho(\alpha',C_E(x+\alpha')

1406: S_{\lambda,E}(x)C_E(x)^{-1})=\rho(\alpha',C_E(x+\alpha')C_E(x+\alpha)^{-1}

1407: R_{\phi_E(x)}).

1408: \ee

1409: Since $m$ is analytic in $x,$ we can take as lift of $C_E(x+\alpha')C_E(x+\alpha)^{-1} R_{\phi_E(x)}$

1410: a function $\tilde \psi(x,y)$ satisfying $|\tilde \psi(x,y)-\phi(x)| \leq

1411: K|\alpha-\alpha'|$.  Thus

1412: \be

1413: \|\rho(\alpha,S_{\lambda,E})-\rho(\alpha',S_{\lambda,E})\|_{\R/\Z}

1414: \leq \int \sup_y |\phi(x)-\tilde \psi(x,y)| dx \leq K |\alpha-\alpha'|.

1415: \ee

1416: The result now follows, since $N=1-2\rho$ (see \S \ref {ids})

1417: for the determination of $\rho$ in $[0,1/2]$.

1418: \end{pf}

1419: 

1420: \begin{rem}

1421: 

1422: Clearly we also get the fictitious estimate

1423: \be

1424: |L_{\lambda,\alpha'}(E)-L_{\lambda,\alpha}(E)| \leq K|\alpha-\alpha'|, \quad E \in J.

1425: \ee

1426: 

1427: \end{rem}

1428: 

1429: \section{Gaps for rational approximants}

1430: 

1431: It is well known that for any $\lambda \neq 0$ if $\frac {p} {q}$ is a

1432: minimal denomination of a rational number then $\Sigma_{\lambda,\frac {p}

1433: {q}}$ consists of $q$ bands with disjoint interior.  All those bands are

1434: actually disjoint, except if $q$ is even when there are two bands touching

1435: at $0$ \cite {vM}, \cite {CEY}.

1436: The variation of $N_{\lambda,\frac {p} {q}}$

1437: in each band is precisely $1/q$.  The

1438: connected components of

1439: $\R \setminus \Sigma_{\lambda,\frac {p} {q}}$ are called gaps.

1440: Let $M(\lambda,\frac {p} {q})$ be the maximum size of the bands of

1441: $\Sigma_{\lambda,\frac {p} {q}}$.

1442: 

1443: The following result is well known.

1444: 

1445: \begin{lemma} \label {density of states}

1446: 

1447: Let $\alpha \in \R \setminus \Q$ and $\lambda \neq 0$.  If $\frac {p_n}

1448: {q_n} \to \alpha$, then $M(\lambda,p_n/q_n) \to 0$.

1449: In particular (since $N_{\lambda,p_n/q_n} \to N_{\lambda,\alpha}$

1450: uniformly), if one selects a point $a_{n,i}$ in each band of

1451: $\Sigma_{\lambda,p_n/q_n}$ then

1452: \be \la{conv}

1453: \frac {1} {q_n} \sum_i \delta a_{n,i} \to dN_{\lambda,\alpha} \quad \text

1454: {in the weak$^*$ topology.}

1455: \ee

1456: 

1457: \end{lemma}

1458: 

1459: In \cite {CEY}, a lower bound for the size of gaps of $\Sigma_{\lambda,\frac

1460: {p} {q}}$

1461: is derived of the form

1462: $C(\lambda)^{-q}$, where, for instance, $C(1)=8$.  We will need the

1463: following sharpening of this estimate, in the case where $\frac {p} {q}$ are

1464: close to a given irrational number.

1465: 

1466: \begin{thm} \label {cey}

1467: 

1468: Let $\alpha \in \R \setminus \Q$ and let $0<\lambda \leq 1$.

1469: Let $\frac {p_n} {q_n} \to \alpha$.

1470: For every $\epsilon>0$, for every $n$ sufficiently large,

1471: all gaps of $\Sigma_{\lambda,\frac {p_n} {q_n}}$ have size ar least

1472: $e^{-\epsilon q_n} \lambda^{q_n/2}$.

1473: 

1474: \end{thm}

1475: 

1476: \begin{pf}

1477: 

1478: It is known (see the proof of \cite {CEY} Theorem 3.3 for the

1479: case $\lambda=1$, the general case being obtained as described in the

1480: proof of \cite {CEY} Corollary 3.4) that for any bounded gap

1481: $G$ of $\Sigma_{\lambda,\frac {p} {q}}$, one can find a

1482: sequence $a_i$, $1 \leq i \leq q$, with one $a_i$ in each band of

1483: $\Sigma_{\lambda,\frac {p} {q}}$, such that $G=(a_i,a_{i+1})$ and

1484: \be

1485: \prod_{j \neq i} |a_j-a_i| \geq \lambda^m,

1486: \ee

1487: where $q=2m+1$ or $q=2m+2$.

1488: 

1489: Let $G_n$ be a bounded gap of $\Sigma_{\lambda,\frac {p_n} {q_n}}$

1490: of minimal size.

1491: Then

1492: \be

1493: |G_n| \geq \lambda^{q_n/2} \prod_{j \neq i_n,i_n+1}

1494: |a_{n,j}-a_{n,i_n}|^{-1},

1495: \ee

1496: where the $a_{n,i}$ satisfy the hypothesis of the previous lemma.

1497: Passing to a subsequence, we may assume that $a_{n,i_n} \to E \in

1498: \Sigma_{\lambda,\alpha}$ and

1499: $|G_n| \to 0$ (otherwise the result is obvious).

1500: By the previous lemma,

1501: we get the estimate for $0<\delta<1$ and for $n$ large

1502: \be

1503: \frac {1} {q_n} \ln (|G_n| \lambda^{-q_n/2}) \geq -\frac {1} {q_n}

1504: \sum_{j \neq i_n,i_n+1} \ln |a_{n,j}-a_{n,i_n}| \geq -

1505: \frac {1} {q_n} \sum_{|a_{n,j}-a_{n,i_n}|>\delta} \ln |a_{n,j}-a_{n,i_n}|,

1506: \ee

1507: which implies by

1508: %Fatou's Lemma and

1509: %here we do not use Fatou's Lemma, since we have convergence of measures and

1510: %not of functions

1511: (\ref{conv}) and the definition of the weak$^*$ topology that

1512: \be

1513: \liminf \frac {1} {q_n} \ln (|G_n| \lambda^{-q_n/2}) \geq

1514: -\int_{|E'-E|>\delta} \ln |E-E'| dN_{\lambda,\alpha}(E').

1515: \ee

1516: Thus

1517: \be

1518: \liminf \frac {1} {q_n} \ln (|G_n| \lambda^{-q_n/2}) \geq

1519: -\int \ln |E-E'| dN_{\lambda,\alpha}(E').

1520: \ee

1521: By the Thouless formula and Theorem \ref {L}, this gives

1522: $\liminf \frac {1} {q_n} \ln (|G_n| \lambda^{-q_n/2}) \geq

1523: -L_{\lambda,\alpha}(E)=0$.

1524: \end{pf}

1525: 

1526: \begin{rem}

1527: 

1528: It is possible to get an estimate on the convergence rate on Lemma \ref

1529: {density of states} using

1530: \cite {AvMS}.  This implies an estimate on the rate of convergence in

1531: Theorem \ref {cey}.

1532: 

1533: \end{rem}

1534: 

1535: \section{Proof of the Main Theorem}\la{mainthm}

1536: 

1537: We now put together the results of the previous sections.  Recall that it is

1538: enough to consider $\lambda>0$, and that the case $\lambda=1$ follows from

1539: Theorem 1.5 of \cite {AK}.  Moreover, Cantor spectrum for $\lambda$ implies

1540: Cantor spectrum for $\frac {1} {\lambda}$.  Let $\beta=\beta(\alpha)$.

1541: The Main Theorem follows then from the following.

1542: 

1543: \begin{thm} \label {1}

1544: 

1545: Let $\alpha \in \R \setminus \Q$.  Then

1546: \begin{enumerate}

1547: \item If $\beta<\infty$ and

1548: $\lambda>e^{\frac {16} {9} \beta}$,

1549: $\Sigma_{\lambda,\alpha}$ is a Cantor set,

1550: \item If $\beta=\infty$ or if $0<\beta<\infty$ and

1551: $e^{-2\beta}<\lambda \leq 1$, $\Sigma_{\lambda,\alpha}$ is a Cantor set.

1552: \end{enumerate}

1553: 

1554: \end{thm}

1555: 

1556: \begin{pf}

1557: 

1558: Item (1) follows from Theorems \ref {localization} and \ref {cantor

1559: localization}.

1560: 

1561: To get item (2), we argue by contradiction.  Let $J \subset \inter \Sigma$

1562: be a compact interval.  Then the density of states satisfies $\frac {dN}

1563: {dE} \geq c>0$ for $E \in J$ \cite {AS}.

1564: Let $\frac {p} {q}$ be close to

1565: $\alpha$ such that $\frac {1} {q}

1566: \ln |\alpha-\frac {p} {q}|$ is close to $-\beta$.

1567: By Lemma \ref{density of states} and Theorem \ref {cey},

1568: $J \setminus \Sigma_{p/q}$ contains an interval $G=(a,b)$ of size

1569: $e^{-\epsilon q} \lambda^{q/2}$.  Notice that

1570: $N_{\lambda,\frac {p} {q}}(a)=N_{\lambda,\frac {p} {q}}(b)$.

1571: Theorem \ref {continuity}

1572: implies

1573: \be

1574: |N_{\lambda,\alpha}(a)-N_{\lambda,\alpha}(b)| \leq K \left |\alpha-

1575: \frac {p} {q} \right | \leq e^{\epsilon q} e^{-\beta q}.

1576: \ee

1577: Thus

1578: \be

1579: c \leq \frac {N_{\lambda,\alpha}(a)-N_{\lambda,\alpha}(b)} {a-b} \leq

1580: e^{2 \epsilon q} e^{-\beta q} \lambda^{-q/2}.

1581: \ee

1582: By taking $\epsilon \to 0$, we conclude that $\lambda \leq e^{-2\beta}$.

1583: \end{pf}

1584: 

1585: Let us point out that $1/2$-H\"older continuity of the spectrum \cite {AvMS}

1586: (which holds for every $\alpha$ and $\lambda$) together with Theorem \ref

1587: {cey} implies the following improvement of \cite {CEY}.  Let us say that

1588: all gaps of $\Sigma_{\lambda,\alpha}$ are open if whenever $E \in

1589: \Sigma_{\lambda,\alpha}$ is such that

1590: $N_{\lambda,\alpha}(E)=k\alpha+l$ for some $k \in \Z \setminus

1591: \{0\}$, $l \in \Z$ then $E$ is the endpoint of some bounded gap (this

1592: obviously implies Cantor spectrum).  The conjecture that

1593: $\Sigma_{\lambda,\alpha}$ has all gaps open for all $\lambda \neq 0$,

1594: $\alpha \in \R \setminus \Q$ is sometimes called the ``dry'' version of the

1595: Ten Martini Problem.

1596: 

1597: \begin{thm} \label {open gaps}

1598: 

1599: Let $\alpha \in \R \setminus \Q$ and let $\beta=\beta(\alpha)$.

1600: If $\beta=\infty$ or if $0<\beta<\infty$ and

1601: $e^{-\beta}<\lambda<e^\beta$, $\Sigma_{\lambda,\alpha}$ has all gaps

1602: open.

1603: 

1604: \end{thm}

1605: 

1606: The conclusion from Theorem \ref {open gaps}

1607: appears to be the natural boundary of what can be taken honestly from

1608: the Liouvillian method: our computations indicate that although one

1609: can get improved estimates on continuity of the spectrum for

1610: $\lambda>e^\beta$ (following \cite {JK}), things seem to

1611: break up at the precise parameter $\lambda=e^\beta$.  Notice that

1612: $\lambda=e^\beta$ is the expected threshold for localization

1613: (for almost every phase)\footnote{In particular, by the Gordon's argument

1614: enhanced with the Theorem \ref{L}, $H_{\lambda,\alpha,\theta}$ has no 

1615: eigenvalues for $\lambda<e^\beta.$, and no localized eigenfunctions for

1616: $\lambda=e^\beta$.} and 

1617: falls

1618: short of the expected threshold for localization with phase

1619: $\theta=0$, $\lambda=e^{2\beta}$.  Thus the use of fictitious estimates does

1620: not seem to be an artifact of our estimates, but a

1621: rather essential aspect of an

1622: approach that tries to cover all parameters with Diophantine and

1623: Liouvillian techniques.

1624: 

1625: \begin{rem}

1626: 

1627: Notice that we do not actually need the measure-theoretical result of \cite

1628: {AK} to obtain Cantor spectrum for $|\lambda|=1$.  Indeed, Lemma \ref{interior} 

1629: and item (1) of Theorem \ref{beta} imply Cantor spectrum for $\beta=0$ (any $\lambda \neq 0$; see

1630: Remark \ref {blu})

1631: and item (2) of Theorem \ref {1} implies

1632: Cantor spectrum for $\beta>0$ (if $|\lambda|=1$).

1633: 

1634: \end{rem}

1635: 

1636: \input{martini}

1637: 

1638: \begin{thebibliography}{BKNS}

1639: 

1640: \bibitem[AK1]{AK} Avila, A.; Krikorian, R.

1641: Reducibility or non-uniform hyperbolicity for quasiperiodic Schr\"odinger

1642: cocycles.  Preprint (www.arXiv.org).  To appear in Annals of Mathematics.

1643: 

1644: \bibitem[AK2]{AK2} Avila, A.; Krikorian, R.

1645: Quasiperiodic $\SL(2,\R)$ cocycles.  In preparation.

1646: 

1647: \bibitem[AvMS]{AvMS} Avron, J.; van Mouche, P. H. M.; Simon, B. On the

1648: measure of the spectrum for the almost Mathieu operator.  Comm. Math. Phys. 

1649: 132  (1990),  no. 1, 103--118.

1650: 

1651: \bibitem[AS]{AS} Avron, Joseph; Simon, Barry Almost periodic Schrödinger

1652: operators. II. The integrated density of states.  Duke Math. J.  50  (1983),

1653: no. 1, 369--391.

1654: 

1655: \bibitem[Az]{Az} Azbel, M. Ya.

1656: Energy spectrum of a conduction electron in a magnetic field.  Sov. Phys.

1657: JETP 19 (1964), 634-645.

1658: 

1659: \bibitem[Be]{ber} Berezanskii, Ju. M.

1660:    Expansions in eigenfunctions of selfadjoint operators. Translated from

1661:    the Russian by R. Bolstein, J. M. Danskin, J. Rovnyak and L. Shulman. 

1662:    Translations of Mathematical Monographs, Vol. 17 American Mathematical

1663:    Society, Providence, R.I. 1968 ix+809 pp.

1664: 

1665: \bibitem[BJ]{BJ1} Bourgain, J.; Jitomirskaya, S. Continuity of the Lyapunov

1666: exponent for quasiperiodic operators with analytic potential. Dedicated to

1667: David Ruelle and Yasha Sinai on the occasion of their 65th birthdays.  J.

1668: Statist. Phys.  108  (2002),  no. 5-6, 1203--1218.

1669: 

1670: \bibitem[BS]{BS} B\'ellissard, J.; Simon, B.

1671: Cantor spectrum for the almost Mathieu equation.

1672: J. Funct. Anal. 48 (1982), no. 3, 408--419.

1673: 

1674: \bibitem[CEY]{CEY} Choi, Man Duen; Elliott, George A.; Yui, Noriko Gauss

1675: polynomials and the rotation algebra.  Invent. Math.  99  (1990),  no. 2,

1676: 225--246.

1677: 

1678: \bibitem[Fur]{fur} Furman,  On the multiplicative

1679:             ergodic theorem for uniquely ergodic systems. Ann. Inst.

1680:             Henri Poincare 33 (1997), 797-815.

1681: 

1682: \bibitem[HS]{HS} Helffer, B.; Sj\"ostrand, J. Semiclassical analysis for

1683: Harper's equation. III. Cantor structure of the spectrum.  M\'em. Soc. Math.

1684: France (N.S.)  No. 39 (1989), 1--124.

1685: 

1686: \bibitem[H]{H} Herman, Michael-R. Une m\'ethode pour minorer les

1687: exposants de Lyapounov et quelques exemples montrant le caract\`ere local

1688: d'un th\'eor\`eme

1689: d'Arnol'd et de Moser sur le tore de dimension $2$.

1690: Comment. Math. Helv.  58  (1983),  no. 3, 453--502.

1691: 

1692: \bibitem[Ho]{Ho} H\"ormander, Lars Notions of convexity. Progress in

1693: Mathematics, 127. Birkhäuser Boston, Inc., Boston, MA, 1994. viii+414 pp.

1694: 

1695: \bibitem[JKS]{jks} Jitomirskaya, S.; Koslover, D.; Schulteis, M.Localization for a family of one dimensional 

1696: quasiperiodic operators of magnetic origin. Annales Henri Poincare. 6 (2005), 103-121

1697: 

1698: 

1699: \bibitem[JK]{JK} Jitomirskaya, S. Ya.; Krasovsky, I. V. Continuity of the

1700: measure of the spectrum for discrete quasiperiodic operators.  Math. Res.

1701: Lett.  9  (2002),  no. 4, 413--421.

1702: 

1703: \bibitem[J]{J} Jitomirskaya, Svetlana Ya. Metal-insulator transition for the

1704: almost Mathieu operator.  Ann. of Math. (2)  150  (1999),  no. 3,

1705: 1159--1175.

1706: 

1707: \bibitem[JM]{JM} Johnson, R.; Moser, J. The rotation number for almost

1708: periodic potentials.  Comm. Math. Phys.  84  (1982), no. 3, 403--438.

1709: 

1710: \bibitem[L1]{L} Last, Y.

1711: Zero measure spectrum for the almost Mathieu operator.  Comm.

1712: Math. Phys.  164  (1994),  no. 2, 421--432.

1713: 

1714: %\bibitem[L2]{L1} Last, Y. Almost everything about the almost Mathieu

1715: %operator. I.  XIth International Congress of Mathematical Physics (Paris,

1716: %1994),  366--372, Internat. Press, Cambridge, MA, 1995.

1717: 

1718: \bibitem[L2]{L3} Last Y., Spectral theory of Sturm-Liouville operators on

1719: infinite intervals: a review of recent developments. Preprint 2004.

1720: 

1721: \bibitem[P]{P} Puig, Joaquim Cantor spectrum for the almost Mathieu

1722: operator.  Comm. Math. Phys.  244  (2004),  no. 2, 297--309.

1723: 

1724: %\bibitem[Sim1]{semi} Simon, Barry Schr\"odinger semigroups.

1725: %Bull. Amer. Math. Soc. (N.S.) 7 (1982), no. 3, 447--526.

1726: 

1727: \bibitem[Sim1]{Sim2} Simon, Barry Kotani theory for one-dimensional

1728: stochastic Jacobi matrices.  Comm. Math. Phys.  89  (1983),  no. 2,

1729: 227--234.

1730: 

1731: \bibitem[Sim2]{Sim1} Simon, Barry Schr\"odinger operators in the

1732: twenty-first century.

1733: Mathematical physics 2000,  283--288,

1734: Imp. Coll. Press, London, 2000.

1735: 

1736: \bibitem[Sin]{Sin} Sinai, Ya. G.

1737: Anderson localization for one-dimensional difference

1738: Schr\"odinger operator with quasiperiodic potential.  J. Statist. Phys.  46

1739: (1987),  no. 5-6, 861--909.

1740: 

1741: \bibitem[vM]{vM} van Mouche, Pierre The coexistence problem for the discrete

1742: Mathieu operator.  Comm. Math. Phys.  122  (1989),  no. 1, 23--33.

1743: 

1744: \end{thebibliography}

1745: 

1746: \end{document}

1747: 

1748: