Access the full text.
Sign up today, get DeepDyve free for 14 days.
J. Tits (1977)
Travaux de Margulis sur les sous-groupes discrets de groupes de lie, 18
D. Ruelle (1978)
An inequality for the entropy of differentiable mapsBoletim da Sociedade Brasileira de Matemática - Bulletin/Brazilian Mathematical Society, 9
K. Jacobs (1963)
Lecture notes on ergodic theory (2 vol.)
(1977)
SHUB, Invariant manifolds
R. Bowen, D. Ruelle (1975)
The Ergodic Theory of Axiom A Flows.Inventiones Mathematicae, 29
(1976)
HIRSCH, Differential topology
D. Ruelle (1979)
SENSITIVE DEPENDENCE ON INITIAL CONDITION AND TURBULENT BEHAVIOR OF DYNAMICAL SYSTEMSAnnals of the New York Academy of Sciences, 316
D. Ruelle (1976)
A MEASURE ASSOCIATED WITH AXIOM-A ATTRACTORS.American Journal of Mathematics, 98
Ya. B. Pesin (1976)
Invariant manifold families which correspond to nonvanishing characteristic exponentsIzv. Akad. Nauk SSSR, Ser. Mat., 40
M. W. Hirsch (1976)
Differential topology
Ya. G. Sinai (1972)
Gibbs measures in ergodic theoryUspehi Mat. Nauk, 27
Ya. B. Pesin (1976)
Lyapunov characteristic exponents and ergodic properties of smooth dynamical systems with an invariant measureDokl. Akad. Nauk SSSR, 226
Y. Sinai (1972)
GIBBS MEASURES IN ERGODIC THEORYRussian Mathematical Surveys, 27
Author Kingman (1968)
The Ergodic Theory of Subadditive Stochastic ProcessesJournal of the royal statistical society series b-methodological, 30
S. Smale (1970)
Notes on differentiable dynamical systemsProc. Sympos. Pure Math., 14
Y. Derriennic (1975)
Sur le théorème ergodique sous-additifC.R.A.S. Paris, 281 A
A. Weil (1967)
Basic number theory
Ya. B. Pesin (1977)
Lyapunov characteristic exponents and smooth ergodic theoryUspehi Mat. Nauk, 32
H. Furstenberg, H. Kesten (1960)
Products of Random MatricesAnnals of Mathematical Statistics, 31
M. Akcoglu, L. Sucheston (1978)
A ratio ergodic theorem for superadditive processesZeitschrift für Wahrscheinlichkeitstheorie und Verwandte Gebiete, 44
M. Hirsch, C. Pugh, M. Shub (1977)
Invariant manifolds
J. F. C. Kingman (1976)
École d’été des probabilités de Saint-Flour
V. Oseledec (1968)
A multiplicative ergodic theorem: Lyapunov characteristic num-bers for dynamical systems
Y. Pesin (1977)
CHARACTERISTIC LYAPUNOV EXPONENTS AND SMOOTH ERGODIC THEORYRussian Mathematical Surveys, 32
J. Tits (1976)
Séminaire Bourbaki
K. Jacobs (1963)
Lecture notes on ergodic theory
ERGODIC THEORY OF DIFFERENTIABLE DYNAMICAL SYSTEMS by DAVID RUELLE Dedicated to the memory of Rufus Bowen Abstract. -- Iff is a G TM diffeomorphism of a compact manifold M, we prove the existence of stable manifolds, almost everywhere with respect to every f-invariant probability measure on M. These stable manifolds are smooth but do not in general constitute a continuous family. The proof of this stable manifold theorem (and similar results) is through the study of random matrix products (multiplicative ergodic theorem) and perturbation of such products. o. Introduction. Let M be a smooth compact manifold, f a diffeomorphism, and p an f-invariant probability measure on M. The asymptotic behavior for large n of the tangent map Txf n is determined for p-almost all x by the multiplicative ergodic theorem of Oseledec [I I]. This theorem (see (I. 6) below) is a sort of spectral theorem for random matrix products. It treats the ergodic theory of the diffeomorphismfso to say in linear approximation. The aim of the present paper is to tackle the nonlinear theory, and our main result is an " almost everywhere " stable manifold theorem (see Theorem (6.3)). This theorem says that for p-almost all x, the points y such that the distance offnx and f~y tends to zero at a suitable exponential rate (when n-++oo) form a differentiable manifold (1). The proof goes via a study of perturbations of the matrix products (Theorem (4. i)) occurring in the multiplicative ergodic theorem. The proof of the multiplicative ergodic theorem given by Oseledec is not appropriate for our discussion, and we use a proof due to Raghunathan [15]. A version of this proof is reproduced in Section I. We have included in the present paper some results of general interest, which fitted naturally, but are not needed for the proof of Theorem (6.3). The reader who only wants to get to the stable manifold theorem may thus omit Section 3 and the (1) That something like this should be true was suggested by Smale in [20]. 275 28 DAVID RUELLE Appendices B and C. We have not tried to present all our results in the greatest gene- rality. Since the articulation of the proofs is reasonably simple, the reader should be able to obtain further results without too much work. Our theorem (6.3) is very close to results of Pesin ([i 2], [I3], [14] ) who has a stable manifold theorem almost everywhere with respect to a smooth invariant measure, assuming that such a measure exists. Our techniques are however rather different from those of Pesin. We refer the reader to the monograph of Hirsch, Pugh and Shub [6] for the much studied case where a continuous splitting of the tangent space exists. The present paper originated in an attempt at proving certain conjectures on the asymptotic behavior of differentiable dynamical systems. These conjectures, presented in [I8], generalize results obtained for Axiom A systems (see [I9] , [I6], [2]). The results obtained here constitute a preliminary step towards proving the conjectures of [I8]. Another step is contained in [i7] (see also Katok [8]). Ultimately, this work should serve to determine the measures which describe hydrodynamic turbulence, and more generally the asymptotic behavior of dissipative physical systems. (o. i ) Note on the multiplicative ergodic theorem. Besides its applications to differentiable dynamical systems, the multiplicative ergodic theorem has applications to algebraic groups. The idea is due to Margulis (see Tits [2I]), and involves extending the theorem to local fields. The original proof of the multiplicative ergodic theorem is due to Oseledec, and applies to flows as well as maps. In view of the applications to algebraic groups, Raghunathan [i5] devised a simpler proof, based on a theorem of Furstenberg and Kesten [4]. This theorem in turn is a corollary (Corollary (I.2) below) of Kingman's subadditive ergodic theorem ([9], [io]) (see Theorem (i. i) and Appendix A). An extension of the sub- additive ergodic theorem to quasi-invariant measures has been obtained by Akcoglu and Sucheston [I], and would permit a similar extension of all our results. While Raghunathan's results apply to maps, an extension to flows, following the ideas of Oseledec, is easy, and carried out in Appendix B (1). (o. 2) Terminology. Here are a few definitions which might be helpful for what follows. A class X of subsets of a space M is a a-algebra if OeE, and if X is stable under countable intersections and complementation (X~M\X). A (finite) measure space (M, E, 0) is a space M with a a-algebra E of subsets (measurable sets) and a countably additive function ~:E-+R+. The function p is (1) I am indebted to A. Connes, M. Herman, and D. Sullivan for pointing out to me the literature on the subadditive ergodic theorem, and in general for encouragement in writing the present paper. I also want to thank J. Tits who informed me of the work of Raghunathan. 276 ERGODIC THEORY OF DIFFERENTIABLE DYNAMICAL SYSTEMS a (finite positive) measure. We also assume completeness: if p(X)----o and Y C X then YeN (and p(Y)=o). If p(M)= i, we say that (M, E, p) is aprobability space, and p a probability measure. Let M be a topological space; the elements of the e-algebra generated by the open sets are called Borel sets. In particular, if M is compact metrizable, and p is a positive Radon measure on M, one can define o(X) when X is a Borel set. A measure space (M, E, p) is then defined where the measurable sets are all the sets XuN with NCY, X and YBorel, and p(Y)=o. Let S be a topological space and M a measure space (resp. a topological space). A map q0 : M~S is called measurable (resp. Borel) if ?-~0 is measurable (resp. Borel) for every open ~7 C S. These definitions extend to sections of fiber bundles, using local trivializations. As usual a map from a measure space to a measure space is measurable if the inverse image of a measurable set is measurable. x. Some basic results. In this section (M, Z, p) is a fixed probability space, and z : M-+M is a measurable map preserving p. Almost everywhere means p-almost everywhere. We denote byf + the positive part of a function f:f~(x)=max(o,f(x)). Theorem (x. 9 (Subadditive ergodic theorem). Let (fn),>0 be a sequence of measurable functions M~Rw{-- ov } satisfying the conditions: a) integrability: f+eLl(M, p); b) subadditivity: fm+, < fm § oZ m a.e. Then, there exists a v-invariant measurable function f: M-+Rw{--oe} such that f+ eLl(M, p), lim -f, =f a.e., n-+~n and lirn- x)p(dx) = inf x)p(dx) = x)p(dx). n n This is one version of Kingman's theorem (see [io], Theorem (I.8)). In Appendix A we reduce Theorem (I. I) to another version, for which an easy proof has been given by Derriennic [3]. Corollary (x. 2). -- Let T : M-+M m be a measurable function to the real m � m matrices such that log+ll T() [leL'( M, o). Write T~" = T(z"-lx). -.- . T(-~x). T(x). 277 3 ~ DAVID RUELLE Then there exists a z-invariant measurable function X: M~Rw{--oo} such that )+ eLl(M, p), lim = x(x) n~oo n for almost all x, and limoo Iflogll T:llp(dx ) -= inf~flog[ I T:llo(dx ) = f Z(x)p(dx). This is proved by taking f.(x)=1ogl[%"l[ in Theorem (I. I). Proposition (*.3) (1). -- Let (T.)n> o be a sequence of real m� matrices such that l.l) lim sup -: log[ I T,[150. We write: T"=T......T2.T: and assume that the limits: lim Llogll (W")^'l[ ~l~oo n exist for q= I7 . . ., m. Then: a) lina (T~*T~)I/2'~= A exists, where 9 denotes matrix transposition. b) Let exp X(1)<... <exp X (') be the eigenvalues of A (real X (r), possibly X(1)=--oo), and U (1), ..., U (') the corresponding eigenspaces. Writing V(~ and V (r) =U/t~-t-... -t-U (r), we have: lim I-log [ITnul[ =X It) when ueV(~)\V ('-a) n --4- oo n for r~--- I~ ...~s. (~) If the assumptions of the proposition are satisfied, and det A -# o (i.e. X 1 > -- o0 ), (i. I ) can be replaced by lim X-logtlT,,ll= lira IlogltTy~alr=o. ~ -'-'~- 00 n ~'l "r 00 n In view of a), lim ~ ~ log [det T n ] = log act A, hence lim -~ log I det T n I = o, and since N-*oo IN n = I n-+~ rt I[T;, ~ [I < I[T,~[lm- ~/Idet Tn [, we have litmsup;logtlTnl[]<o ). :-logllT~ll ERGODIC THEORY OF DIFFERENTIABLE DYNAMICAL SYSTEMS 3~ Let t,(a)<... <t, (") be the eigenvalues of (T"*T") ~/2. By assumption, the limits: lim - log lim - log I I (Tn) ^q [] n-~oo R p=m--q + l n--,.oo n exist for q=-I, ..., m, and therefore also the limits: lim I_ log t~ v) -= Z (p) n~r ?/ for p = i, ..., m. Let X(~)<... <X (~) be the distinct Z (p), and U(n ~) be the space spanned by the eigenvectors of (T"*T") 1/z corresponding to the eigenvalues t~ v) such that (z. ~,) lim I log t(.r)= X C'). 7*-->00 n We interrupt now the proof of Proposition (I .3) for a lemma. For simplicity we shall assume that X(~)+--oo. Lemma (I.4). -- Given 8>0, there is K>o such that, for all k>o, (' .3) max{l (u, u')l: ueU~ ~), u'eU(~k, llull = Ilu' fl =I }<K exp(--n(lX(~')--X/')t--3)). We first prove (I. 3) for r<r'. Equivalently, it suffices to prove that, if v,,k is the orthogonal projection of u~ Y' --n lI(t) in ~] -~,lr(t')+k, then t<r t'~r' (I4) l[~,.,[lSKllull exp(- n(X(")-- X(')-- 8)). It will be convenient to assume 8 less than all IX(~')--?dr) 1 for r4=r', and to write 8"=8/s. In view of (I.I) there is C>o such that, for all n, 3" logllTn+al[<C+n-. For large n we have thus: ( (, t[vl,,l[exp (n+I) X(')--~ <ilT~+lu[l < IIT.+llI.IIT"ull <exp(Oq-n~).llu"exp(n(X(')q-~)). 3" 8* If n is so large tlaat C--X(")+~-<n~, this gives: II vL, I I< I1~11 exp(--n(XC~')--X/~)-- 8")) 9 From this we obtain in particular: k--I II 4 ~+1 II <_ ~}o I1 u II exp (-- (n +j)(X ('+ t)-- X(r)-- 8") ) < K1 II u It expC--n(Xr + 1)-- X(~)-- 3")) 279 DAVID RUELLE 3 2 with Kl=(I--exp(--(k(r+l)--X(~)--S*))) -1. Therefore also: k--1 ZollUl[ exp(--(n+j)(X(~+2)--)~(~)--S*)) k--1 + 2 glliull exp(--n(X(~+l)--X(~)--8*)) exp(--(n+j)(X(~+2)--X(~+x)--8*)) J=0 !K~llul[ exp (-- n(X(r + ~)-- X(')-- 23")). In general: If V~r' I I _<; K~'-r [I u I I exp(- n(x (r')- x (r)- (r'- r)8*). Since (r'--r)3*<3, this proves (1.4). Notice that the lemma gives bounds on the elements of the m � m matrix S of scalar products between the eigenvectors of (T'*T") a/2 and those of (T("+k)*T"+k)l/~. We have proved up to now the bounds for the elements on one side of the diagonal of S. The other bounds are readily obtained from the calculation of S* = S- ~ by the minors of S. Allowing for change of ~ and K, it suffices to use the bounds already obtained, and the fact that all matrix elements are bounded by 1 in absolute value. This conclude the proof of the lemma. Lemma (I.4) shows that (u(,r)),>0 is a Cauchy sequence for each r. Part a) of Proposition (I.3) follows from this and (I.2). Let U(r)=lirng(nr); (1.3) then becomes: max{[(u, u')[: ueU (*), u'aU(, ~'), Ilull=lIu'll= i}<_K exp(--n(IX(*')--x(~)[--3)). Therefore we have, for large n, if o+ueU (r), x("/--,:, ~ < !log [1T"u[I < X(r) § 2 8, -n Ilull - hence: lim I-log [IT"ul[ =k (~) if ueU(')\{o} n --~ oo n and part b) of the proposition follows. Corollary (x.5) (of Proposition (I.3)). --Let X(r)<)~<X (~+1) (put X (~+1)= § /f r =s). Then: R={u~R": IIZnull<e for all n>o} is a bounded open neighborhood of o in V (r). That RCV (~) is clear from Proposition (i.3) b). Furthermore, we have: ,!ira II W"ul[ e-"X= o 280 ERGODIC THEORY OF DIFFERENTIABLE DYNAMICAL SYSTEMS uniformly for u in the unit ball B of V (r). Since R C B, there is N finite such that e={ueV(r): llT-ull<e -x for o<n<N} proving the Corollary. Theorem (x. 6) (Multiplicative ergodic theorem). -- Let T : M-*M m be a measurable function to the real m � m matrices such that (1.5) log + [IT(')I[~L~( M, 0)- Write T~=T(z'-~x). ... . T(vx).T(x), and use * to denote matrix transposition. There is FCM such that zFC F, p(F)=I, and the following propert#s hold if xeF: a) lim (T~*T~)I/2~-----A exists. b) Let exp X~)<...<exp X~ / be the eigenvalues of A~ (where s=s(x), the X(~ ~1 are real, and X~ I may be -- or), and UJ 1, ..., U~ "/the corresponding eigenspaces. Let m~ ~/= dim U~ ~/. Writing V~ ) ----{o } and V (r) = II (~) +. d- lr(~) The functions x~X~ ), m(~ ~) are z-invariant. we have: lim - log l[ T~u [I when n-~oo n for r=I~ , , .~ s. According to (i-5) and the ergodic theorem, there is I'lC iV[ such that "~I~IC I'l, o(Fa)=I, and lim !log + IIy(-:"-lx)[[=o if xeF1. n~co n By Corollary (I ,2), there is also P2 such that -:F2C P2, P(P2) =I, and, for q=I, ..., m, lim Ilog If(z" )^ ll n ---> oo n exists, and is a v-invariant function of x. Let I'=Plt~P z. The theorem follows from Proposition (I.3) applied to Tn=T(vn-lx) for xeP. Corollary (z. 7). -- Let x~P, ueR"; then: (x. 6) lim ~-log llT;ull = Z(x, u) n ---~ oo n exists, finite or --oo. If XeR, the linear space V~X=(ueRm: Z(x,u)<X) b a measurable function of xer. We have X(x, u)=X~ ) This is an immediate consequence of Theorem (2.6). if usV(~r/\V~ -t/, and V2= [J{V~rl: X(~r/<X}. 281 DAVID RUELLE Remark (I.8).- (1.6)implies Z('rx, T(x)u)=z(x , u). In particular x x T(x)V~CV,~, T(x)V(~)CV(,~. If x~)+--oe, T(x) is invertible and therefore T(x)V(~ ~/=V(~, T(x)V~ =V,X~. On the other hand, the U~ / do not transform simply under T(x). 2. The spectrum. As in Section I, (~V[, ]~, p) is a probability space, and v : M-+M a measurable map preserving p; T : M-+M m is a measurable function such that log+[[W(.) [[eLI(M, p). We write T~=T(z"-lx).....T(xx).T(x). According to Corollary (1.2) and the multiplicative ergodic theorem (Theorem (I.6) and Corollary (i.7)), there is P C M with -:I'C I', o(P) = i, such that, if xeP, we can define A~; s=s(x) ; X~)<... <X~=Z(x) ; U~/, ...,U(f; {o}=V~/CV~IC...CV(~I=Rm; and the functions u~z(x,u), X~VX~. Let mJ/=dim U~/=dim V(~l--dim V(x ~-~/ The numbers X~ / are called charac- teristic exponents; with the multiplicities m(~ ~/ they constitute the spectrum of (% T), or T, at x. We shall say that V~ ale ... C V~ ~/ is the associated filtration of R m. The spectrum is v-invariant. If p is v-ergodic, the spectrum is almost everywhere constant. In what follows we shall determine the spectrum of (% T^), (z-~, T*) and (% T*-~). (e.I) Spectrum of (~, TA). Let TAP:M~M(,~) be the p-th exterior power of T. We have: TAP('r '*-*x) ...-. TAP(':x). TAP(x) = (T~)Av and lim ((T~) A;* (T~) Ap) 1/~ = Ax/Xp . This determines the spectrum of T Ap and the associated filtration of R (~). Writing TA= @ T:';, we obtain in particular: ~0=0 I log 11 (T~)A][ = Y~ ,~(%(,) lim tb ---~ o0 n r : k(xr) >0 --1 (e. 2 ) Spectrum IkT , T*), of: Suppose that v has a measurable inverse, we shall show that the spectrum of (x-1, T*) is almost everywhere the same as that of (% T). Let X~= lirno (~-~,T~)1/2~ where ~'~ = T*(-:-" +ix). 9 9 9 T*(z-lx). T*(x). Since the spectrum of ,~ is v-invariant it is also v v the limit almost everywhere of the spectra of the ~_~(T~*T n~lj2"_zj , where T~" = T* (x) T* (zx). 9 9 9 T* (-: -~ x). 282 ERGODIC THEORY OF DIFFERENTIABLE DYNAMICAL SYSTEMS v v v v The spectrum of T~ T~ is the same as that of T'T "*~_~ = T~ T~. Therefore the spectrum of _A~ is the same as the spectrum of A~. (2.3) Spectrum of (% T*-I). Suppose that T is almost everywhere invertible and that log + I[T-~(.)lieU(M, p). Define A~= lim , where ~I~-----T -l(v"-lx) . ....T*-l(~x).T*-l(x). We have ~ --> o0 then ~= A~-I. Therefore the spectrum of (% T*-I) is obtained by changing the sign of the spectrum of (% T):~r/=__..~Zds--r+t/. The filtration of R m associated with (% T *-1) is the orthogonal of the filtration associated with (% T) : ~(~/=V~ ~-~1 3" The invertible case. In this section, (M, E, p) is a probability space, and -~ : M~M is a measurable map with measurable inverse preserving p. Theorem (3. I). -- Let T: M-+GL,~ be a measurable function to the invertible real m � m matrices, such that IoN+ ]l T(") II, log+ 11 T-l(") lieU( M, P)- Write: T~= T(-r"-J x). ..- . T(,rx). T(x) T~ -~ = T-l(v-~x)..... T- 1 (.~ - 2x). T-l(v - ix). There is then ACM such that -rA=A, p(A)=I, and a measurable splitting ~,w~l)e. . . ew;'~ of S ~ ooer a (with s=s(~)), such that k~176 IlZ~ull=x~ ~ /f o,u~W~'>. Let again the numbers X~)<... <X(f with multiplicities m~ ), ..., m(~ 8) constitute the spectrum of (% T) at x. Let V~ )C ... CV(~ ~) be the associated filtration of R". From Sections (2.2) and (2.3) we know that the spectrum of ('r- 1, T- 1 o'r- 1) at x consists of the numbers --9~])<... <--X~ ) with multiplicities m(~ *), . .., m~ ). Let: v~-'~ c... c v~-l~ be the associated filtration. Suppose that we can show that (s- ' ) v~'-' n v~-~/= [o} (3.2) V(,'-I) § V~-'/---- R '~ for r=2, ...,s, and almost all x. Then, putting w~rl = v~,l,-, v~-'> ~--x(T~*Tn~l/2n--xJ DAVID RUELLE 3 6 we obtain: It" = v~-'~ n (W~ + v~-~) n (v~ + v~-~) c~... c~ v~ (1) (2) = W,~ | | | ~ and the theorem holds. It remains thus to prove (3. i) and (3.2). Define S as the set of those x such that (3.I) does not hold. Given 8>o and rs[2, s], let S~ be the subset of S such that, if xeS~, (3.3) IIT:ull_<_llull exp n(k~-l) +,3) and (3.4) IIT;-"ull <llull exp n(--X;) + 8) for all e (r-~)c~ (-r/ u V~ V~ . From (3.4) we get, if xez-"Sn, (3.s) II W:u II_>_ [lull exp n(?,(~*)-- 8) for all ueV~-l)c~V(z-*). For x~SJ~z-'S~, (3.3) and (3.5) yield x,lr)--Xx(*--l/<2~. -- n _.~ --X --X Since p(S,c~-: S,) p(S) we have ),(r/--?,(r--~/~23 for almost all xeS and, since is arbitrary, we get p(S)=o. We have proved (3.1); (3.2) follows because dim V~- l/-t- dim V(~ - r) = m. (3.2) Spectrum and associated splitting. The characteristic exponents Z~/<... <Z~ s/ with multiplicities m(~ rl = dim W~ r/ constitute the spectrum of (% T) at x. We call W~I| | / the associated splitting of It". Notice that the X~ r) are all finite, and that: T(x)W~ ) =W(,~ r = I, ..., s. (See Remark (i. 8).) The spectrum of (% T Ap ) at x consists of the numbers ~=~nrX~ r/ with o<nr<m~ rl, and En=p. The subspace corresponding to [z in the associated splitting of R (~) is generated by ul^...^u" where ujeW~i) and ~ x(ri}=[X. (This follows readily from Section (2. I).) j~ 1 The spectrum of (v-1 T-lo.~-1 at x consists of the numbers --X(zs)<... <--k~ ) with multiplicities m(f, ..., m~ ). The associated splitting of It" is W(~")|174 The spectrum of (% T *-1) at x consists of the numbers --X~/<... <--X~ 11 with multiplicities m(f, ..., m~ I. The associated splitting of It" is W~-~/|174 a/ where W(Z r) is the orthogonal complement of ~] Wf / in It". (This follows readily from Section (2.3).) r': ~' *r The spectrum of (.:-1, T.o.~-l) at x is the same as that of (% T). The associated splitting of It m is W(Z~I| | -~/. (This follows from what has been said of (% T *-~) and (-~-~, T-~o-:-~).) 284 ERGODIC THEORY OF DIFFERENTIABLE DYNAMICAL SYSTEMS Corollary (3.3). -- Define: ",AT(r). ~] W (r'), u' v~(x)=max{l(u,u')l: u~vv z u'e Ilull=[l I[=I} r' :r' +r (put y~(x)------o /f s(x)=I). Then: ar(X) = (~ - v,(x)'?/~ = min max{](u,v)l: veWF ~),]Iv][=I} and lirn _I log 8~(~x) = o. k-+ i~176 k Letindeed p=m~ ~), q=m--p, o,w~(W~)) ^p, o,w'e( E WJ>) Aq, then: li((Tb~w) ^ <_ ll(T~)^~w' II and it suffices to apply what has been said on the spectrum of T A in Section (3. @. 4. A perturbation theorem. Theorem (4. x). -- Let T = (T,),~> 0 be a sequence of real m � m matrices such that (1) lim_+sup i log I I T,, I] <~ o. (4" I) We write T"=Tn..... T~.T1 and assume the existence of lim (T"*T")I/2"=- A (4.2) with det A#:o. Denote by k(1)<... <X (") the eigenvalues of log A. Let ~>o be given and, for T'=(T~),>0, write IIW'--Wll=sup llTL--Z.lld" and T'"=T~ ....T~.T~. Then there are 8, A>o and, given ~>o, there are B~>o, g;> I with the following properties: If ]tW'--Wll< a, (4.3) lim o (T'"* T"*)'/2n= A' exists and has the same eigenvalues as A (inclung multiplicity). Furthermore, if P(r)(T') denotes the orthogonal projection of A' corresponding to exp X (r), and I I T"-- Tit <_ a, we have: (4.4) 11 P(')(T')- P(')(T")II<A II T'--T '' II (4.5) B~ exp n(X (,>- e) < I I T'"W)(T')[[ <_ B~ exp n(X (r) -I- ~). (a) Instead of (4. i) one could write: lira .-X_log IlTnll =o. n ---~ oo n See the footnote to Proposition (i.3). ~r(~)11(T})^~wIl. ((T})^"w')l} 3 8 DAVID RUELLE If (4- I) holds, it is known (Proposition (i. 3) a)) that the existence of the limit (4.2) is equivalent to the existence of the limits lim Ilog II(T' )A II ~ ---~ oo n for q=I, ...,m. Since (4. I) and ][T'--TII<-[-oa imply lim sup--log IIW~ll <_ o, n ---~ c~ n (4.3) will follow if we can prove the existence of lim -Ilog II(T"~)^,[I ,tl -~ n for q=i, ..., m. Furthermore these limits determine uniquely the eigenvalues of A'. Therefore, to prove (4.3) and the fact that A' has the same eigenvalues as A, it suffices to show that (4.6) lim -Ilog [I(T TM) II^q= lim -~log II(T")^qll. ~t ---~ co n ~,t -~ oo n Let o<~'<~ and define: I[T'^'--T^q[l=supl T_,~Aq--T~ql[e 3'~'. Then (4. i) implies the existence of Eq>o such that (4-7) [[ Z'^'--TAqll! Eq II r'--Z [1 for 8~i. Therefore, the replacements T~ ~T Aq, T~ ~T~ Aq reduce the proof of (4.6) to the case q=I, i.e.: lim I log [1T'Y'[[ = X ('~>. Equivalently, it suffices to find an open set UCR m such that lim~logllZ'"uIl=x for u u. ~ ---.-- Qo n To see this take u(l/, ..., u ('~! linearly independent in U and notice that the matrix norm [1[. I1[ defined by: [llXlll = II xu<a ll +... + II Xu( )ll is equivalent to ][" [1- The existence of the limit (4.3), and the fact that A and A' have the same eigenvalues, are therefore a consequence of the following result: Lemma (,t.2). -- Let X('(1;'<...<X('(ml}=X ("> be the e@envalues of log A repeated according to multiplicity. Let ~ot, ..., ~1 be unit vectors spanning R m and such that (4.8) lim _I log [1 T" ~(k~ ! [ ----- X ('(k)). n .--~ oo n 286 ERGODIC THEORY OF DIFFERENTIABLE DYNAMICAL SYSTEMS There is then 8>0 such that lira -~ log I[ T'nu ][ = x(~) ~ ---> oo n ueU, where: whenever o<~z<i, lIT T[I<S~, and [ m-- 1 g(O) max,<,, u~l<lu,,I}. The existence of ~[0), ..-, ~) satisfying (4.8) follows from Proposition (I.3) b). The reason for not assuming the ~(k ~ orthogonal will appear in Remark (4.7). (4.3) Proof of the lemma and further inequalities. By Proposition (I.3) a): n (0) ~ )~(r(k)). (4.9) lim Ilog ItTn~i~ ^T ~ tl = Let ~(k ~) be a unit vector proportional to T'~ ~ and write: (4 IO) %~(k n-l)- t(n)g(n) 9 -- vk ~k " Let also ~J~) be the j-th component of ~'~). The matrix ~(") --~-- satisfies I[ ~(')[I <V/m and, because of (4-9), lim ~ log l det ~(~)l ---- o. w --*- oo ~,/ Therefore lira i log I[ ~( ")- ~11 = o n-+oo n and given ~>o, we have: (4. II) D~ =sup e-"~ [ I ~(n)- 1 l I < -~-oo. We write D~=D. In view of proving (4-5) we shall obtain a result somewhat stronger than the lemma. We suppose that I[T'--TI] <_ ~ and estimate the components u~ '~/, ..., ~(') for any u+o in R "~. (~) u (') of T'~u along ~i")/a, ~)-t/a, -,m, Let ~ be the smallest integer such that (4.X2) ('fin) maxluJ"?12maxlu~l I. j_<~ k>~ (In particular if ueU, we have ~x=-rn.) Because of (4. io) and (4. II) we have: I ui.~ I < t~ ~) j u~,~-~t + D~e-~Z I u?-~)[. -- t (~(~)~jkJ DAVID RUELLE 4 ~ We weaken these inequalities if we replace the okt(n) by ts ") such that a) N ,~2 log ts ")* = X (~(~)) for k _< tz and b) t(~)* = t2 *) . In view of (4-8), (4-'~ this can be achieved by multiplying the sequences (t~")), for k<~, by constants ~i. Since ~q>o, a) implies the existence of C>o such that, for all v~o, N>v, and k,t<~, N--2 H t ('')*/ N (4.1a) ~=~+~ t In=H+2 ~k ~(')*~'Ce~'~ ~ 9 One can also choose C independent of ~x. provided In view of the above we have l u(k"liSU(k *~) for n2v, U~)= U (~) = max ]u~ ~) l, and T r(~) > ~(,*)* 11(,, - 2) + D ~ e- 2,,~. m. max U} '~- 2). "-'k -- ~k ""k t_< Using (4.13) we see that this is satisfied by N N (4. i4) U(k~l = ,=~+ ")*1~ tt(k . =~+~IV[ (i + mCD 3e-"~) .U (~). We choose I oo (4"15) 3=mC1)n~l (I -- e-n~)2" In this way mCD3< I, and oo I1 (,+mCD~e -'~) co i+e-~ ~o --e-'~) -2= I (4.16) C'__ n=2 <n ~ 1~ (, . o~ -- = i--e -~'~ n 2 mCD3 II (I--g -n~) n=2 Therefore (4.14) gives: N N (4.17) ].~N),/'rv(N)/~,uk [~w k ~ ~I t~ (')*. I1 (,--e -nn).U ('). n=v+2 n=v+l In view of the definition of ~ by (4-'2), we may choose v such that t u(;) I = %ax Ir = u("). Using (4-13) and (4.17) we obtain then, for N>v: u(n) t(N/ (N-1)I__D3e-2NnE lu~N-2)[ N N--1 (~) (N-l) >t~ [u~ [--mCC'D~e -N~ II t(~ n). n=v+l -- n=v+2 ~-.o~lim v~t(n)*"~-- ERGODIC THEORY'OF DIFFERENTIABLE DYNAMICAL SYSTEMS Using (4. i6) gives: N--1 N--1 lu~> I ]> t,(+s)(luC~n-t)I--e-~ [I t. Cn) . II (t--e-'+~).lu~)l) n=,~+l ~ n=v+l which implies, by induction, N N (4.x8) lukN)12 II t (~) [I (i--e-n~).lu~) 1. n=v+l IJ- "n=v+l From (4. iI), (4.i7) and (4. I8) we obtain: 9 I (4. I9) !1s N log II Z"+u II--x (+(+)>. In particular, if u~U we have r(~x)=s, and the lemma results from (4. I9)- (4.4) Partial proof of (4.4). Suppose that the eigenvalue expX (') is simple, i.e. the corresponding pro- m -- 1 ~(0) jection P(+)(T) is one-dimensional. If U=k~=lUk~+Um~'~m)CU , we have: [iP(+)(T)u 11 <_[u,,,,[ < max .~l<++lt~(O>-lll. [l,,tl. -- k<m Let ~ be a unit vector in the range of P(+)(T). Since the kernel of P(")(T') cannot inter- sect U, we have, in view of the above estimate and triangle similarity: It (t- P(+)(T'))P(+)(T) t1 = II (I--P(+)(T'))~[I <--~ ~(~ hence I I p~+; (T') P(+)(T) It ~ ~I -- ~.211 ~(0)- 1 112. We apply this result to the situation where A' is replaced by A 'As, p being the sum of the multiplicities of the largest eigenvalues of A' corresponding to the projections P(r)(T'), ..., P(+)(T'). Writing :r instead ore, e'-~%l[(~(~ and = W)(T) +... + P(+)(T) ~' = W)(T') +... + P(~)(T') we obtain: PP'P has at most p non zero eigenvalues, with product 2I--~,2, SO each eigenvalue is >1--s 2. Therefore IIP--PP'PI]_ ,or ]I(I--P')PI]<~. Similarly: II (I--P)P'IISs 289 DAVID RUELLE so that II~'-~ll!2~' provided [[T'A'--TA'[I<~p% with ~v determined by the lemma. In view of (4.7), we can take %= ~E, IIT'--TII. This is less than i because we choose S in the theorem < bp/Ep for each p. Thus: I1~'-~ I[ <_2~' = ~EvI[ (K(~ [1T'--TI[. Therefore: (4.2o) 1[ P(~)(T') -- V(r)(T) [I !All T'- T II with max 4 A-- , ~Ev[ I (~,(~ (4-2I) (4.5) Proof of (4-5). Ifu is in the range of P(~)(T'), and u#o, (4. I9) shows that r(~x) = r. In particular, we may use (4.I7) with v=o to obtain liT'null <_g21[u][ exp n(Xr + ~) which is the second half of (4.5). If ueU, then ~x =m, and one can take ~ =o in (4-i8). Therefore (4. II), (4. I7), and (4.18) show that, given ~>o, there are C~, C'>o such that (4 ~2) c~[I. I I exp n(X (')- r <[I T'~u [I <--C'di u [I exp n (x (~) -k z). We shall now prove that ~ may be decreased so that these inequalities hold for all u in the range of P(~)(T') when []T'--T][<S. Let u be a unit vector in the range of P(~)(T), and u' be such that (4.23) [lu'[] <I, 11.'-~ll<(2mIl~(~ Write u ~ ~(o) u' ~ ,/~(o1 Ukqk ~ = ~k"~k " k k Then E ,, ~(o) has norm > I. Therefore [ukl> ~k~k -- -- for some k with r(k)=s and, k: r(k) = s Ir[t by renumbering the ~(o) I sk , we may assume [ u ml>- Since lu~_u/l<ll~(0~_,ll.llu,_.ll ~ i , 2m wehave ]Um[>~. Andsince [[u'll!I wehave luLl![l~(~ for k<m. Therefore 2m u'~g when ~<(2m11~(~ -1. According to (4.2o), every vector in the range of 290 ERGODIC THEORY OF DIFFERENTIABLE DYNAMICAL SYSTEMS P(~)(T') is proportional to u' satisfying (4.23), provided AI]T'--Tll<(2mlI~(~ We also want [IT'--TI[<~8. This is achieved by replacing 8 by (4.24) min ( (2m ][ 4 (0) --1 [[) --18, (2ma][ 4 (0) -1 [1) - t). With this choice (4.22) holds whenever u is in the range of P(~)(T') and IIT'--TI[<8. Let q be the sum of the multiplicities of exp X (~+11, ..., exp X (~), and apply (4.22) with T' replaced by T ''~q and T '^(q+l) One finds, for u in the range of P(')(T'), and v#o in the range of (P(~+I)(T')-k...+P(~)(T'))^q: __ ~'r iiT,.ul[> llT,.u^(T,.)Aqvl I c(q+~l -- II(Z"~)^'vll ->- ~.,2~'('1 " []ull-exp n(X(~l--~)" Therefore: I I T'"W)(T') I1 ~B=exp n(X (~)- ~). This completes the proof of (4.5). (4.6) Proof of (4.4)- The earlier " partial proof of (4.4) " in Section (4.4) yields (4-2o). We obtain (4.4) from (4.2o) by the replacement T~T" ifA can be chosen independent of T". In view of (4-2I) this is achieved if we can replace T by T" in Lemma (4.2) and get bounds on 8~ -t, Ep (defined by (4.7)) and 114(~ uniform in T". Since [ITL'II<IIT, II+8 it is easy to obtain a bound on Ep. We take the vectors 4~ ~ ..., 4~ ~ in the lemma to be orthogonal, so that I1~/~ = I. The choice of 8 made in the proof of the lemma is given by (4. i5). Therefore it suffices that we find upper bounds to C and D independent of T". Remember that C is given by (4.13), and D is given by (4. I ~). In view of (4.5) we can bound C by (B'~/4/Bv4)2.exp(--XI"t). Applying (4.5) to T ''Am we obtain an estimate: B/-,/ [ det ~("/I> ~-~ exp ( -- n (m + i ) ~). Taking ~=~/(m-4-I) yields the desired bound on D. (4.7) Complement to Theorem (4. I). If, instead of (Tn),>0, we consider the sequence T/tl= (T,~+t),>0 , the conditions of Theorem (4-i) are again satisfied. We check here that 8 -a, A, and B" can be chosen to increase with t at most like e 3tn, e eta, and e t~ respectively. This result will be used in Remark (5.2) c) and the proof of Theorem (6.3). First, we replace in Lemma (4-2) the vectors ~0), ..., ~) by 4~ ), 9 9 Then D~ and C are multiplied at most by e t~ and e t~. Therefore 8 is multiplied by a factor not smaller than e -~tn. Replacement of T by T Ap replaces 8 by 8p which is multiplied by a factor not -~m.~(t) DAVID RUELLE smaller than e -zt:/. The Ep (see (4.7)) are multiplied by at most e 3t(*- ~'), and therefore Sp/Ep is multiplied by at least e -t(3~--*') Remember that min bp/Ep is the choice of used to prove the existence of the limit (4-3), and also in Section (4.4). From (4. ii) it follows that the choice of A given by (4.2I) does not grow faster than e t(~*- ~; +~) The choice of ~ in (4.24) therefore does not decrease faster than e t(3~-,,+2~), i.e. e -zt~ if ~'=2z; going over to min 8p/Ep does not change this. In Section (4.5), B'~, C~, C -t~ , B~ -1 do not increase faster than e w for any ~'>o, e.g. z' = ~. Therefore in Section (4.6) we obtain finally that A does not increase faster than e 2t~'. 5. A nonlinear ergodic theorem. In what follows we denote by B(0~) the open unit ball of radius ~ centered at the origin of R m, and by B(a) its closure. We shall say that a map is of class C r'~ if its derivatives up to order r are H61der continuous of exponent 0; similarly for manifolds. Theorem (5. i ). -- Let (M, E, ~) be a probability space and -r : M-~-M a measurable map preserving p. Givenaninteger r~I, and 0~(o, I], let x~Fxma p M to cr'~ m, o). We write F~= F,,-~xO... o F~o Fx and denote by T(x) the derivative Of F~ at o. We assume that x~T(x), [lF~][r,o are measurable and that (5" I) flog+ll FxIlr, 0 p(dx)< + oe. We choose X<o and assume that almost everywhere the spectrum of T at x contains neither X nor --oo (the spectrum is ,finite, in particular T(x) is invertible). There is then a measurable set P C M such that -rF C F, p(P) ----- I, and there aremeasurable functions ~>c~>o, y>I on P with the following properties: a) If xeF the set Z={ueg(~(x)) : IIF2uil<~(xDe "x for all n>o} D x is a C ~'~ submanifold of ~J(e(x)), tangent at o to V 2. b) If u, veuX~, then IlF:u--F;vll <y(x) llu--vlle nx. If p is ergodic, the spectrum may be assumed constant on F. IJ" X' < X and the interval IX', X] is disjoint from the spectrum, there exists y' measurable on F with the property: b') when u, veuX~, then I1 ~n v IlLu-I vll<_ (x)llu-vlle We first study the case r= I; the case r>I will be dealt with later. 292 ERGODIC THEORY OF DIFFERENTIABLE DYNAMICAL SYSTEMS We may take PCM such that -~PCF, p(F)=l, and, if xeP: (5. ~' ) li+rnoo (T~* Tx ~)*'2~ = A~, (5.3) det A~+o, det(A~--eX 1) +o and (5-4) lim _i log+{l F.._~xll~,0 = o. ,O, -~- oo n This follows from Theorem (1.6) for (5.2), by assumption for (5-3) and from (5. i) and the ergodic theorem for (5-4)- Notice that (5.4) implies lim I-log+llT(z"-lx)11 = o. n --+oo Let o<4~q<_--X0; we may then write, using (5-4), { G = sup II F,,-**ill,0 exp(-- mq--x0) < + oo (5.5) I I F,,_1~111,0 exp (n(X0 + 3~) -- x0) <G. Given xeP, we write A~=A and let log A have the eigenvalues X(1)<... <X (~) (characteristic exponents) with the multiplicities m {1], ..., m (*/. Let V (~lC ... C V (*/ be the associated filtration of R m. We assume that X (p) is the largest characteristic z _ V(V) exponent <X. Therefore, with the notation of Corollary (I. 7), V~-- . We write (5.6) ~ = X-- X (p). Given }, o<~<i, we shall use the definitions: S~(~)={ueR~: llFgull<f3e 'a for o<n<v}, (5.7) S([~) ={ueRm: IlF';ull< e "z for all n2o}. (5.8) There are 8, A>o such that Theorem (4-I) holds with ~ as defined above and We can make 8 smaller so that AS< -- (5.9) and then choose ~=~(x) satisfying (5. xo) o<~<I, G~~ The functions x~8, A may be assumed measurable, as follows from their (essentially) explicit construction in the proof of Theorem (4.1). Therefore also x~ may be assumed measurable. Take x>i such that x~5i , G(�176 We shall show that there is me(o, ~) such that, for all v>o, (5" IX) g(0~) nS~([3)n{u~R '~ : 1] Tt"~;-~V;ull< [3e "x for all n>v} = n n 293 46 DAVID RUELLE ueS~(� ~-1 x Let indeed c~ (F~) V~,x. The m � m matrices: %= DF~,-t~(tF2-~u)dt if n< v, f: T,I=T ~ if n>v, are such that T'nu--T ,.....T'lu=F;u if n< v, and using (5.7), (5-5), we have 1[ T'--TII = sup IIZ~--Z, IIe 3n~ ! sup IlDF~-lzl[0 (~)0 exp(n(X0 -t- 3~)--X0)< O(�176 8. Therefore, Theorem (4. I) applies. In particular u is in the space V '(p) CR m spanned by the eigenvectors of log A' corresponding to the eigenvalues X (~), ..., X (pl. Using (5-6), (4.5) gives llZ'"ull:<__g'~e "~ Ilull v --I I, uniformly in v and ueS~(K~)r~(F~) Vt~ ~. If o~----~/B~<~, we see that the right- hand side of (5. I I) is contained in the left-hand side. The converse inclusion is immediate. As for ~, we can assume that x~ is measurable. Let D~(00 be the set defined by (5. I I). Since the boundary of S~(,:~) is disjoint from S~(~), and hence from D~(e), we conclude from (5. I I) that D~(00 is open and closed v --i k in g(~) n(g) V~x. Let now u, veD'(e) or u, veg(e)nS(~)=u~ x (in the latter case, write v=oo). The m� m matrices T~=f2DF~,-~,(tF~-~u+(i--t)F~-~v)dt if n~v, T~ =T. if n>,J, are such that T'n(u--v)=T~..-..T'l(u--v)=F~u--F~v if n_<v, and, using (5.7) or (5.8), and (5.IO), we have ][ T'--T I[ <_sup l[ DF~"-lxl[o D ~ exp(n(X0 + 3~) --X0) n~V < G~~ Therefore Theorem (4. I) applies and, since u--v is in the range of P(P)(T'), (4.5) yields (5. I2) I1F;u--F:v I[ <--Y 1[ u--v [I e"x. In this formula we have written T=B_'>I, and x~T may be assumed measurable. This proves part b) of the theorem. Part b') is also obtained if we take r (vl instead of (5.6). From (4-4) we obtain: (5. I3) [](I- P(')(T))(u--v)[ I = I[(P(')(T')--P(')(T))(u--v)l[ < A8 [lu-vll 294 ERGODIC THEORY OF DIFFERENTIABLE DYNAMICAL SYSTEMS which implies: A8 [[(~- P{')(T))(u--v)[I ~ (5" I4 ) It P<')(T) (u-- ~)ll- l/~ --(A8) ~ Define + : (vw n g(~)) x (v(')~ n g(~)) -~ g(~) by: +(u~, ,,~)-- ulv'~-II,,~ I I" + u.. Let +(u~, u~), +(u;, u~)~D'(,) or g(~)nS(~). Then (5.13) yields II~=ll, Ilu~ll~ Aa~ and, by (5-I4): %/I --(AS) 2 A8 0C + !~/~__ ilu~ll ~ " II ~-u~ll AS~ i ii.~li-II411 i -- %/~. (A8002 A8 --%/I --(A8) = so that A8 %/I----22-A8) -- In view of (5.9) the expression in parenthesis is >o. Since D~(00 is open and closed in B(~)r~(F;)-IV~,,x as a consequence of (5.ii), we conclude from (5.I5) that D~(a) V~,. Furthermore +-ID~(~) is the is the connected component of o in B(00n(F;) -1 x graph of a G 1 function ~: V/r)r~B(~) ---~V(r) with derivative bounded uniformly with respect to v. Let q~ be the limit of a uniformly convergent subsequence of (%). Since +(graph q~)=D~(00C B(~)nS'(~), we have +(graph q0C B(~)nS(~). The converse inclusion follows from (5- I5) applied to B(~)nS(~). Therefore +(graph ~) = B(o~) n S(}) and, by uniqueness of % lira q~= q~ uniformly. Let u, veDa(00 and define m� matrices: T', = DF,,_~,(F,'- ~u), T'/= DF~,-x,(F~-'v) if n~v, T~=TZ=T~ if ~>~. 295 DAVID RUELLE Then [IW'--W[I, I,- IT" - T[[<8. Using (5-I2) we have also I W',-- T,' [I <ll II Fg-'u-- F7-'v !l ~ ~l[F,,-~ilt,07~ -- v'l~ if n<_v, and therefore lIT' -T" ![!G?llg--vll ~ By (4.4) we have then I P(p~(T')- P',P/(T")II < (AGu176 l[ u --vll~ where the ranges of P(P)(T') and P(P)(T") are the tangent spaces to DV(~) at u and v. Letting v-+oo we find that the tangent space to ]J(a)c~S(~)=u~ at w also depends H61der continuously on w, with exponent 0. This tangent space is the range of p~F (T), where T,, --= DF,,_~,(F~ -' w) for all n: to see this notice that we may assume 1' T'-- ]"i ~ ~o as ~oo, and apply (4.4). In particular the tangent space to u~ x at o is P(v~(T), i.e. \;~;. This proves part a) of the theorem when r= i. We prove now that u~ Xis C~,0 by induction on r for r>i. Let F~: g(I) | be the C ~-a,~ map defined by F,(u, v)= (V~u, DF,(u)o). We can apply the results obtained till now to F instead of F. In particular, let S(~) be replaced by g(~) CR"| The above identification of the tangent space to S(~) as the range of p/p~(~) shows that (u, v)~]~(~)c~'S(~) if and onlyif uel3(~)nS(~) and v is tangent to S(~) at u and sufficiently small. Since B(~) ng(~) is C ~-~'~ by induction, the dependence on u of the tangent space to B(~)nS(~) at u is C '-u~ Therefore g(a')nS(~) is C ~'~ if a'<~. Remarks (5.2). -- a) The theorem as we have stated it assumes only the measurability of x~T(x), I]F~[Ir, 0. One could easily give an " abstract " version for a sequence of maps F,,ECr'~ m, o) satisfying conditions corresponding to (5.2), (5.3), (5.4)- On the other hand further measurability properties of xi ~F, would imply x is the, measurability properties of xl ~, ,~. x Such properties follow from the fact that u x C r limit, as v~oo, of the connected component DV(~) of o in B(e)c~(F;)-~V~ (with C ~,0 estimates uniform in ~). b) Let T~=DF~,_~x(F~-'u). The range of P"r for q--I .... ,s, has C r-~'~ dependence on u~uX~. This was shown above when q=p. For general q tile step r= x is the same; the argument used for r> I has to be modified by writing F~(u, v)= (Vzu, eX-X'DVz(u)v) where k' is not in the spectrum and q is the largest characteristic exponent <X'. 296 ERGODIC THEORY OF DIFFERENTIABLF DYNAMICAL SYSTEMS c) From Section (4-7) and (5.9), it follows that we can take 8 at ztx to decrease at most like e -3tn. From (5.5) we see that G increases at most like eel Therefore, by (5. Io), we can take }(ztx) to decrease at most like e -4t~'/~ Since ~=~/B~, ~(=tx) decreases at most like e -~t~'/~ (5.3) The C ~ case. Theorem (5-I) has a C ~ version as we now indicate. Let x~F~ map M to o; R m, o). We assume that x,-,Y(x), IlLll, are measurable and, instead of (5. i), that (5.x6) f log~llF~[Iro(dx)< + oo for every integer r> o. Then the conclusions of Theorem (5. x) hold with u~ a C ~~ submanifold of B(~(x)). Let P,., ~, ~, y~ be a choice of P, a, ~, y according to Theorem (5. i), for r>I and any 0e(o, I], say 0= -I-. Let: o t ', hR.,/ ( n "~ We have o<~1<I (see (5.1o)) and o<~<I; therefore f~ ,',(,,)%(='x)p(dx) 5e ''x, ' U r,(n); then p(P',)--I. If xeI"~, there implying ,-.colim o(M\P,(n))=o. Let P'=,,_>_0 is some n>o_ such that F~ maps u xx,1 (i.e. u,x defined with ~x and ~1) onto a subset of the C * manifold u,,~,x ~ (i.e. ux,, defined with z~ and ~). Since F~ is C ~, and is a C t diffeo- x it is also a C ~ diffeomorphism, and ux. x 1 is therefore CL Let now morphism on u~.,, P~= fi fi v-kF'~. We have ~(P~)=I and vF~C P~; let ~, ~, Y~o be the res- r--lk=0 trictions of ~a, ~1, 5"1 to 1"~. Then the desired C ~~ version of Theorem (5. I) is obtained with Po~, ~, ~, Y~o in place of I', a, ~, y. Notice that we have also shown the following: /f the conditions of Theorem (5. I) are satisfied, the functions ~, ~, y can be determined by considering x~Fx as a map from M to Ct'~ o; R, o) (but F might depend on r). (5.4) The analytic case. Let B(I) denote here the open unit ball centered at o in C" and H(B(I), o, C "~, o) be the space of maps holomorphic in B(I) and continuous on B(I). The holomorphic version of Theorem (5. x) is as follows. Let xP~F~ map M to H(]~(I), o; C'", o). We assume that x~T(x), [IFxl[1 are measurable and, instead of (5.I), that (5.I7) flog-IlF=l!lo(dx)< +oo. ?, Then the conclusions of Theorem (5-I) hold with u x a holomorphic submanifold of B(~(x)). Notice that (5-I7) implies: flog+ll F~II~ o(dx)< + oo 297 DAVID RUELLE 5 ~ where I1" II~ is the C 2 norm on a ball with radius < I, Therefore a C t manifold is defined by Theorem (5. I). By construction, this manifold is a limit of holomorphic mani- folds D~(a), defined by (5. I I), and therefore u, x is holomorphic. In Section 6, this result on holomorphic maps will be used to handle real-analytic maps. 6. Stable manifold theorem. Let M be a compact differentiable manifold, and f: M---~M a C t map. Applying Appendix D with -: -=f, E = TM, T = Tf yields a Borel set F C M with the following properties: I. fI'C 1" and e(l')= i for every f-invariant probability measure ~ on M. II. For each xeP, the spectrum {X~,..., k~ )} of Tf and the associated filtration V~':C... CV(2)--T~M of TxM are defined. We write vx=U{v~:: k~"<X}. III. P is the union of disjoint Borel subsets I'o indexed by the f-ergodic measures, such that vP~C Pc, and lim I n~_~l q~(f kx) =_ p(q0) whenever xc-P~ and q~ : M-+R is continuous. The spectrum is constant on each P~. Theorem (6.i). ---Let M be a compact differentiable manifold and f: M-+M a diffe- rentiable map of class C ~'~ (r integer >I, 0e(o, I]). Let d be a Riemann metric on M and denote by B(x, ~) the open ball of (sufficiently small) radius o: centered at x in M. Given Z<o there are Borel functions ~;>~>o and u on the set FX--{xeP: the spectrum of ~if at x contains neither X nor --oo} with the following properties: a) If xeP x, the set u~'(~(x))--={.),EB(x, 9 c~(x)) : d (fy " ........ .t"'x)<g(x)e ~z for all n " o} is a C ~'~ submanifold ofB(x, ~(x)), tangent at o to VX~. If M and fare C ~ (resp. C '~ i.e. real- analytic), then ,~z~(:~(x)) is C ~~ (resp. C~~ b) If y, zeuX~(~(x)), then d(f"y,f"z) _<T(x)d(y, z)e '~ Give~ 0 ergodic, if X'< X and the interval IX', X] is disjoint from the (constant) spectrum on I'~, there exists a Borel function X~T'~T with the property: b') If y, zeu~(a(x)), then d(f"y,f"z)<T'(x)d(y , z)e "x'. We may assume that M is C ~. There is then a C ~ map (x,u)~+~(u) of TM to M such that +~ maps the open unit ball ofTxM diffeomorphically onto a subset of M 298 ERGODIC THEORY OF DIFFERENTIABLE DYNAMICAL SYSTEMS 5 x and ~(o)=x. With a finite Bore1 partition of M trivializing the tangent bundle we may associate a Borel map (x, u).~ds which is piecewise C ~, is a bijection of M� onto TM, and is such that +~:11 T~M is a linear contraction. We choose 3 so small that the image byf~.~+~ of the closed ball B(3) is contained in +:~:'~B(I) for all x. Define now u:~-=y~+~8 : -+ T ' g(I) M and F,=TtT~ofo~,. Given X<o, the proper- ties (5-2), (5.3), (5.4) hold if xeF x. Therefore Theorem (5.1) holds (with F x replacing F) and we obtain readily the C ''~ version of thc prescnt theorem. The (essentially) cxplicit construction of xebec, ~, y ensures that these functions are Borel. In thc C ~ case, Scction (5.3) should bc used instead of Thcorem (5. i). If M and f are C ~', lct f be a holomorphic extension off to a neighborhood N of M in a complexification 1~I. There is then a C ~ map (x, u)~+~(u) of T~tlQ to such that ~, restricted to the open unit ball of T~IQ is a holomorphic reality preserving diffeomorphism onto a subset of N, and +~(o)=x. With a finite Borel partition of M trivializing T~I we may associate a Borel map (x, u)!~d?~(u) which is piecewise C '~ is a bijection of M � C" onto TM.-Q, and is such that +~ : C"-+%~I is a C-linear realily preserving contraction. Choose now 8 so small that the image byj~+~+;, of the closed ball - B(S) (in C m) is contained in ~t~+:~B(I) for all x. Define ~=y~+~: " ' " g(i)~N /I2-- 1 ~ II? and F~ = ~t~ ~ o ~ ~. Given X< o we may apply section (5.4) and we obtain a t~amily of holomorphic manifolds. Their real parts are the desired C ~ manifolds u2(e(x)). Corollary (6.2). -- If p is ergodic and all the characteristic exponents of Tf are strictly negative on Pp, then p is carried by an attracting periodic orbit. Let the characteristic exponents be <X<o. There is xeFo such that Since we have here u~x(~(x))--i3(x, 0~(x)), we find f"l~(x, ~(x) ) C B(f"x, ~(x)e "z) and we have p(f"B(x, ~(x))) >~. Thus thc set B(f"x, ~(x)e "x) has measurc at least ~. Using compactness, and taking a limit, we find a point with mass ~. Its orbit carries p by crgodicity, and is finite. Clcarly, it is also attracting. Theorem (6.3). -- Let M be a compact differentiable manifold, and f a diffeomorphism of class C" o. We have here .IT = P, and the following properties hold: a) Let X~)<...<Xf be the strictly negative characteristic exponents at xer. Define @It... C u~ ) by _:,P)={yeM : lim sup,,.+~o -n'l~ d(f"x,f"y)<X~"} for p = I, . .., q. Then ~" (p) is the image of V~ ) by an injective C" e immersion I= tangent to the identity at x. 299 DAVID RUELLE b) If xeF and X x ipl < o, then uTI C F~ for some ergodic ~. 7"&filtration V~/C 9 ~ 9 , . CV~ is, has C r-l'~ dependence on I;-lyeV~ ~. One may in the above replace C r'~ (resp. C "-1'~ in b)) by C ~~ or C% With q=q(x) defined above, choose a Borel function ~ on P such that o < - We take q+I numbers Xl, ...,X~,~ such that (6. x ) X~I<XS-X(s <XT!<Xq< -- ~(x), and such that Xl,..., Xq, ~ are constant on a countable family of f-invariant Borel sets forming a partition of I'. On each one of these sets, and for p= I, ..., q, a function ~ is defined by Theorem (6. i) with respect to X=Xp. We call again ~ the minimum over p of these functions. This new function e defined on I' is again Borel, and is such that whenever X is one of the Xp, u2(~(x)) is defined and Theorem (6. I) holds. The number ~ is that appearing in the proof of Theorem (5. i), it satisfies o<4~<-X0 as it should. By reference to Appendix D one sees readily that if xeI'p, then u~(~(x))C Fo. (The main point is to check that u~x(~(x))CF '. This follows from the inequality IIT--T!I<~ in the proof of Theorem (5.i), and application of Theorem (4.I).) In particular, if X= X~, ~2(~(x)) is tangent to V(u p) for each yeu~(~(x)). Also the filtration V(r'C y ..... CV~ ~ has C ~--1'~ dependence ony as noted in Remark (5 2) b). (In the C ~ case, the dependence is C ~ (cf. Section (5.3)); in the C ~ case the dependence is C'~ use a complex extension off and M as in the proof of Theorem (6. I).) We come now to the proof of the theorem. For the fact that fP=P see Remark D.2 a). By Remark (5.2) c), we know that ~(ftx) decreases at most like e -St':~ (This asymptotic behavior is not changed by the mappings in the proof of Theorem (6. I).) Here 5r,/0<~(x) by (6. i), so that ~(ftx) decreases less fast than e -t:(~). Therefore for each k>o there are arbitrarily large integers l>o such that (g. 2 ) he- tv,~', < c,.(ftx). Let xsP, X~P)<o, and X=Xp. If ye,Jx ,(~> , there is k such that, for all n>o, d( ff x, f"y) <_ he "x. In particular there are arbitrarily large t~o such that, for n~o, d( ft :-~,y, ft '"x) < ke-t~(~! e"Z< ~( ft x)e"X< ~( ft x)e"X. Therefore ftye@~(e(ftx)), hence oO 300 ERGODIC THEORY OF DIFFERENTIABLE DYNAMICAL SYSTEMS 53 The argument applied above to yEu~P) also applies uniformly to all yef-'@,(a(f"x)). Thus, for each n>o there is g>n such that f-,u~,x(~(f,x))Cf-t ~ t x)), and we may assume thatf~-"@,~(~(f'x)) is contained in the open ball B(ftx, ~(ffx)). In particular u~ p) is the union of an increasing sequence of " disks " f-t@~(~(ftx)) tangent to V~ p) at x. It readily follows that u~ p) is the image of V~ p) by an injective immersion tangent to the identity at x (see Hirsch [5], Chapter 2, Section 5). This proves a) and b). Remarks (6.4). -- a) If f is replaced byf -1, F will be replaced by a set F-, and there is no reason to expect in general that P and F- will coincide. b) One sees easily that In particular the manifolds u:q)={y~M : limsup-~log d(f"x,f"y)<o} may be called stable manifolds for f. They provide a foliation (discontinuous in general) of the set r. Theorem (6.3) provides a variety of other foliations depending on the choice of an f-invariant function X < o on r. APPENDIX A. Proof of Theorem ( I . I ). The assumptions a) and b) of the theorem imply that f+< f+ +A + or+...-{-f+ovn-leL '. Therefore I,~ff,(x)p(dx) exists, finite or --0% and b) gives Im+,,~Im+l~. The sequence (I,) being subadditive, we have For every positive integer N, define = max{f.(x), -- nN}. It is easy to see that the sequences (f(s)),>. again satisfy the subadditivity condition b) of the theorem. Let us assume that the conclusions of the theorem hold for these sequences. Then -If, (N) has a limitf (~) a.e. for each N, and there is f: 1V[ ~llu{--oo} such that 7/ f(S)(x) = max { f(x), -- N} 301 DAVID RUELLE for almost all x. In particular (A.2) . i h rnnfn= f a.e. and f f(x)p(dx) = inf f f(~l(x)p(dx). Since the conclusions of Theorem (1.I) hold for (f~(N))~>0 by assumption, we have (f(Nl(x)p(dx) = inf f-I f~l(x)p(dx)" j n jn Thus (A.3) =inff lfn(X)O(dx). In view of (A. i), (A. 2), (A. 3), Theorem (I. I) is a consequence of the following result. Theorem (A. x ). -- Let (f~),> 0 be a sequence of real functions such that a) f~eLl(M, p); b) fm+n~fm@'fnO'V m a.e.; c) there is N>o such that fA(x)o(dx)2--nN. Then there is a z-invariant real function fELl(M, p) such that ~ fn tends to f almost every- ii. where. Furthermore: l" I . . I (A.4) lmoon f J,~(x)p(dx)=Infn f f~(x)p(dx) =ff(x) o(dx). For a proof, see Derriennic [3]. APPENDIX B. Semiflows. In this Appendix, (M, 23, p) is a fixed probability space, and (vt)t>0:M~M is a measurable semiflow preserving p. (This means that (x, t)~vtx is measurable 1VI� -:~ the identity, -?+t=-?o-~t, and each vt preserves 0.) Almost every- where means p-almost everywhere. Theorem (B.I) (subadditive ergodic theorem). Let the map (x, t) ~ f(x) : M � R+ -+ R w {--oo} be measurable and satisfy the conditions: a) integrability: qh= sup ffeLl(M, p), ~02= sup f~+_uo-~ueLl(M, p) O<u<l 0<u<l 302 ERGODIC THEORY OF DIEFERENTIABLE DYNAMICAL SYSTEMS b) subadditivity: f,+ t<f~ +f~o z' a.e. Then there exists a (~:t)-invariant measurable function f= M--> R u{--oo} such that f+ cLl(M, p), lim -f~ =f a.e., and t-+oo t lim t fft(x)?(dt ) = inf t f ft(x)p(dx) =f f(x)p(dx). Let n be the integral part of t. We have then and, since q~t, q~2 ~L1, we have _ I n lim i ~~ o T n = lim - % o'~ = o a.e. by the ergodic theorem. The above theorem follows thus from the corresponding theorem (i. I). (B. 2) Cocycles. (x,t)~T~ from M� to M m (the real m� matrices) will be called A map a cocycle if T~+, t = We also assume that the cocycle is measurable M� -~ M., and that the functions %, q~2 defined by (B.x) qh(x)= sup log+ltZ~ll 0<u<l (B.2) q~2(x)= sup log+llT~;"ll 0<u<l are in L~(M, p). From Theorem (B.~) we obtain the existence, for q=I, . .., m, and almost all x, of (B.3) lira -~log II(T )^ II. t -->-~o t We also have, for almost all x, (B,4) lim suPt sup logiiT~t+~li<o. t~co 0<u<l 303 5 6 DAVID RUELLE To see this, write t=n-+-v (n integer, O<V<I) and observe that Tl+u ~ u+v 1--v (if u+v>i) I T"+"~,.+,~ [< IITCE-~[[ . I[T~.+,~I[ yield l~ + [I ~'~t T1 +x u lll[ 5 q)2( 'I'n X) J 7 q)l( Tn +1 X) @- q)l( %.n +2 X). (B.4) follows then from the ergodic theorem. Using the existence of (B.3), and (B.4), the proof of the multiplicative ergodic theorem in Section I is easily adapted to flows. Theorem (B.3) (Multiplicative ergodic theorem). Let (Tt)e_>0 be a measurable cocycle with values in M m (the real m � matrices) such that the functions ~, ~2 defined by (B.I), (B.2) are in L~(M, p). There is I" C M such that ,:tP C F for all t>o, and the following properties hold if x~P: a) lim ~_,(Tt* Tq r/2e_~, = A~ t ---~ oo exists. b) Let expX~)<... <exp X~ / be the eigenvalues of A~ (where s=s(x), the X~ ) are real, and X~ ) may be -- oo), and U~ I, ..., UIJ ~ the corresponding eigenspaces. Let m2 r) = dim U~ r?. The functions x~X~ ~, m~ rl are ('?)-invariant. Writing V~ ) ={o} and V(~ r) = U~I/-F... § U(f, we have lim-I logllT~u][ =),2 ) when usV~\V~(rl (~-~) t-+oo t fo~" Y~ I, . .., S. APPENDIX C. Local fields. The multiplicative ergodic theorem extends to local fields (1), as noticed by Margulis. If R is replaced by C, matrix transposition has to be replaced by Hermitean conjugation in Theorem (i.6). In general, replacing m� complex matrices by ( (i) em� real matrices reduces the complex case to the real case. Let I~ i ' We shall not discuss ultrametric local fields, foi which see Raghunathan [I5]. APPENDIX D. Continuous maps. Let M be a metrizable compact space, z : M-+M a continuous map, rc : E--~M a continuous m-dimensional vector bundle over M, and T : E-+E a continuous vector bundle map over v. (1) For definitions, see WEIL [22]. 304 ERGODIC THEORY OF DIFFERENTIABLE DYNAMICAL SYSTEMS 57 Proposition (D.x). - There is a Borel set PCM such that zFCF, and ~(F)=I for every z-invariant probability measure c on M. Furthermore, for each x~ F, there is a v-ergodic probability measure 9x such that (D.x) lira _I ~ q~(vkx)=9~(~0) n-+m nk=l for all continuous q0:M-+ll, and f fl lim !log I((Z;)^ql[ = l%n logll(W~)^qll p~(dy)=inf -I .log,,(T)Aq ox(dy) n~~176 nd for q=i,...,m, where we have written T~----T(v~-lx).....T(x). Let F 1 consist of those x~M such that Vq(x) = lim ! log [] (T~) Aq [] ~ ---~ oo r~ exists, and (D. 2) Vq(x) = Fq(-~x) for q = i, ..., m. Then Pl is a Borel set and a(Pl) = i for every invariant probability measure ~ by Theorem (i. I). We write I"= ~ "~-"Pl. n>0 Let I" 2 be the set of all xeM for which there is a -:-ergodic probability measure Pc such that n--I vague lim I_ y~ 8~k. = Pc- n--~m nk=0 Then F 2 is a Borel set, vF~C F2, and ~(F2)= i for every "~-invariant probability measure ~. This follows from the Bogoliubov-Krylov theory (see Jacobs [7]). Further- more, if ~ is a z-invariant probability measure we have, by (D.2), for a-almost all xeF'~F~, (D. 3) Fq(x) = f Fe(y) p~(dy). We define continuous functions F~, by Vl.(x) = max{log T "Aq 11(x) I[,-e}. From Theorem (I. I) we get: 'Fq(y) 9,(dy) = lirn I_ [ log [I (T~)Aq l[ p~(dy) nj nj pi N--1 = lim lim lim _I j N --~0 F~n(%'kx)" n-+mt~r N~m n k Therefore the set P of those xeP'c~P 2 for which (D.3) holds isBorel, zFC P, ~(P)= 2, and Proposition (D. i) holds. a0a 8 58 DAVID RUELLE Remarks (D.2).- a) The proof of the proposition gives -:P=r when T(x) is invertible for all x. b) Since E can be trivialized by a finite Borel partition of M, a multiplicative ergodic theorem follows from Proposition (D. I) and Proposition (I. 3). The arbitrariness in the choice of norm on E is without consequence for the definition of the spectrum of (-:, T) at x~P, and the associated filtration of E(x). REFERENCES [I] M. A. AKCOOLU and L. A ratio ergodic theorem for superadditive processes, to appear. [2] R. BowEN and D. RU~LL~, The ergodic theory of Axiom A flows, Inventiones math., 29 (I975) , I8I-2o2. [3] Y. DERgIENmC, Sur le th6or~me ergodique sons-additif, C.R.A.S. Paris, 281 A (1975) , 985-988. [4] H. FURSTENnEgG and H. K~STEN, Products of random matrices, Ann. Math. Statist., 31 (I96O), 457-469 . [5] M. W. HIRSCH, Differential topology, Graduate Texts in Mathematics, n ~ 33, Berlin, Springer, 1976. [6] M. HIRSCH, C. PUGH and M. SHUB, Invariant manifolds, Lecture Notes in Math., n ~ 583, Berlin, Springer, I977. [7] K. JAcoBs, Lecture notes on e~godic theory (2 vol.), Aarhus, Aarhus Universitet, I963. [8] S. KATOK, The estimation from above for the topological entropy of a diffeomorphism, to appear. [9] J. F. C. KINGMAN, The ergodic theory of subadditive stochastic processes, J. Royal Statist. Soc., B 30 (1968), 499-51o. [IO] J. F. C. KINGMAN, Subadditive processes, in l~cole d'6t6 des probabilit6s de Saint-Flour, Lecture Notes in Math., n ~ 589, Berlin, Springer, I976. [i i] V. I. OSELEDEC, A multiplicative ergodic theorem. Ljapunov characteristic numbers for dynamical systems, Trudy Moskov. Mat. Ob$d., 19 (I968), I79-2IO. English transl. Trans. Moscow Math Soc., 19 (I968), 197-23I. [12] Ya. B. PESlN, Lyapunov characteristic exponents and ergodic properties of smooth dynamical systems with an invariant measure, Dokl. Akad. Nauk SSSR, 226, n ~ 4 (I976), 774-777. English transl. Soviet Math. Dokl., 17, n ~ 1 (x976), 196-199. [I3] Ya. B. PESlN, Invariant manifold families which correspond to nonvanishing characteristic exponents, Izv. Akad. Nauk SSSR, Set. Mat. 40, n ~ 6 (I976), I332-I379. English transl. Math. USSR Izvest~]a, 10, n ~ 6 (I976), I26I-I3O 5. [I4] Ya. B. PESlN, Lyapunov characteristic exponents and smooth ergodic theory, Uspehi Mat. Nauk, 32, n ~ 4 (I96) (I977) , 55-II2. English transl., Russian Math. Surveys, 32, n ~ 4 (I977), 55-I14 9 [15] M. S. RAGHUNATHAN, A proof of Oseledec' multiplicative ergodic theorem. Israel. 07. Math., to appear. [I6] D. RUELLE, A measure associated with axiom A attractors, Amer. J. Math., 98 (I976), 619-654. [17] D. RUELLE, An inequality for the entropy of differentiable maps, Bol. Soc. Btas. Mat., 9 (I978), 83-87. [18] D. RUELLE, Sensitive dependence on initial condition and turbulent behavior of dynamical systems, Ann. N.Y. Acad. Sci., to appear. [I9] Ya. G. SINA1, Gibbs measures in ergodic theory, Uspehi Mat. Nauk, 27, n ~ 4 (1972), 21-64. English transl. Russian Math. Surveys, 27, n ~ 4 (I972), 21-69. [2o] S. SMALE, Notes on differentiable dynamical systems, Ptoc. Sympos. Pure Math., 14, A.M.S., Providence, R. I. (I97O), pp. 277-287. [2I] J. TITs, Travaux de Margulis sur les sous-groupes discrets de groupes de Lie, Sdminaite Bourbaki, expos6 n ~ 482 (1976), Lecture Notes in Math., n ~ 567, Berlin, Springer, 1977 . [22] A. WEIL, Basic number theory, Berlin, Springer, I973, 2nd ed. Manuscrit refu le 15 septembre 1978. SUCHESTON,
Publications mathématiques de l'IHÉS – Springer Journals
Published: Aug 7, 2007
You can share this free article with as many people as you like with the url below! We hope you enjoy this feature!
Read and print from thousands of top scholarly journals.
Already have an account? Log in
Bookmark this article. You can see your Bookmarks on your DeepDyve Library.
To save an article, log in first, or sign up for a DeepDyve account if you don’t already have one.
Copy and paste the desired citation format or use the link below to download a file formatted for EndNote
Access the full text.
Sign up today, get DeepDyve free for 14 days.
All DeepDyve websites use cookies to improve your online experience. They were placed on your computer when you launched this website. You can change your cookie settings through your browser.