---

Periodic Orbits for Hamiltonian systems

arXiv:math/9201297v1 [math.DS] 11 Nov 1991Periodic Orbits for Hamiltonian systemsin Cotangent Bundles1 Christophe Gol´eIMS, SUNY at Stony BrookAbstract: We prove the existence of at least cl(M) periodic orbits for certaintime dependant Hamiltonian systems on the cotangent bundle of an arbitrarycompact manifold M. These Hamiltonians are not necessarily convex but theysatisfy a certain boundary condition given by a Riemannian metric on M. Wediscretize the variational problem by decomposing the time 1 map into a productof “symplectic twist maps”. A second theorem deals with homotopically non trivialorbits in manifolds of negative curvature.0 IntroductionThe celebrated theorem of Poincar´e-Birkhoffstates the existence of at leasttwo fixed points for an area preserving map of the annulus S1 × [0, 1] which“twists” the boundaries in opposite directions.In the 60’s, Arnold proposed a generalization of this theorem for a time1 map F of a time dependent Hamiltonian of Tn × Bn (where Bn is theclosed ball in IRn).

While the Hamiltonian condition naturally generalizesthe preservation of area, a linking of the boundary of each fiber ( a sphere inIRn) with its image by F in the boundary of Tn×Bn was to replace the twistcondition. Arnold [Ar1] conjectured that such a map has at least as manyfixed points as a real valued function Tn has critical points.

The philosophywas that fixed points for symplectic maps should arise from Morse theoryand not, say, from Lefshetz theory.Later, in [Ar2], he explained how fixed points theorems on the annuluscould be derived from theorems on the 2-torus, by glueing carefully twoannuli together (see also [Ch1]). He thus transformed the problem to one offixed points of symplectic maps on a compact symplectic manifold.

Thislast conjecture, which asserts that the number of fixed point for the map is at1 Partially supported by an NSF postdoctoral grant.

2Christophe Gol´eleast equal to the minimum number of critical points of a real valued functionon the manifold, is what got to be well known as the Arnold conjecture.However, it is unclear whether the glueing construction can be done(symplectically) in higher dimensions. Even if it could, one would (if onecould) have to use existing proofs of the Arnold conjecture (e.g.

[F2]), whichwe think are substantially harder than the techniques we use here (and donot deal with homotopically non trivial orbits as our Theorem 2 does).In 1982, Conley and Zehnder [CZ 1] gave a first proof of the Arnoldconjecture for the torus T2n. In the same article, they also gave a directproof of Arnold’s original conjecture on Tn × Bn.However, they were not able to use the linking of spheres in its full gen-erality.

Their result remains crucial since it was the first non perturbationalone in this direction. The boundary condition that they used is expressed onthe Hamiltonian in the following way.

Letting (q, p) be the coordinates onTn × Bn which is endowed with the canonical symplectic structure dp ∧dq,they set:(0.1)H(q, p, t) = ⟨Ap, p⟩+ ⟨b, p⟩for ∥p∥≥K,where A = At is a non degenerate n×n matrix and b ∈IRn . This conditionimplies the linking of spheres at the boundary.We propose here a version of this theorem on the cotangent bundle ofan arbitrary compact manifold.

We also find, in a second theorem, orbits ofall free homotopy classes (and large enough period).The bulk of this work was done as I was on a Postdoctoral position at theForschungsinstitut f¨ur Mathematik, E.T.H. Z¨urich.

I would like to expressmy deep gratitude to Prof. Moser and Prof. Zehnder for inviting me there. Ihad some invaluable discussions with them as well as with my companionsFredy K¨unzle, Boris Hasselblat, Frank Josellis, to whom I extend my thanks.I am very much endebted to Patrice LeCalvez, whose work is the startingpoint of mine.Special thanks to Maciej Wojtkowski, Claude Viterbo, Misha Bialy,Leonid Polterovitch, Phil Boyland and Dusa McDufffor their specific helpon this work.Finally, were it not for the narrow mindedness of the French immigrationoffice, this work would have been joint with Augustin Banyaga.

I dedicatethis work to him.

Periodic Orbits31 Results and basic ideasLet (M, g) be a compact Riemannian manifold. DefineB∗M = {(q, p) ∈T ∗M | g(q)(p, p) = ∥p∥2 ≤C2 < R2},where R is the radius of injectivity of (M, g).

Let π denote the canonicalprojection π : B∗M →M.Theorem 1 Let F : B∗M →B∗M be the time 1 map of a time dependentHamiltonian H on B∗M, where H satisfies the boundary condition:H(q, p, t) = g(q)(p, p) for ∥p∥= C.Then F has cl(M) distinct fixed points and sb(M) if they are all non de-generate. Moreover, these fixed points can all be chosen to to correspond tohomotopically trivial closed orbits of the Hamiltonian flow.We remind the reader that cl(M) is the cup length of M, which is knownto be a lower bound for the number of critical points of any real valuedfunction on M. Non degenerate means that no Floquet multiplier is equalto one.

sb(M) is a lower bound for the number of critical points for a Morsefunction on M.Remark 1.2 It is important to note that, in the case where M has IRnas covering space, Theorem 1 can be expressed for a lift of F. In this case,the radius of injectivity may be ∞(e.g. for a metric close to a flat metricon the torus, or when M has a metric of negative curvature), and the setB∗M can be as big as one wants.

Theorem 1 can then serve as a startingpoint to study Hamiltonian systems with asymptotic boundary conditions.Theorem 2 Let F be as in Theorem 1. If (M,g) is of negative curvature, thenF has at least two periodic orbits of period d in any given free homotopyclass, provided d is big enough.

In particular, F has infinitely many periodicorbits in B∗M.Exactly how big d should be in Theorem 2 depends only on the metric.For a more precise statement, see section 7. Note also that if H is 1–periodicin time, periodic orbits of F correspond to periodic orbits of the Hamiltonianflow of H. Such a Hamiltonian system will then have infinitely periodicsolutions in B∗M.

22If H is not 1–periodic, periodic orbits of F will correspond to orbits of theHamiltonian flow that come back to their starting point, but generally at anangle. One can find infinitely many of these orbits from Theorem 1, by applyingit to time t map, t ∈(0, C), rescaling the metric each time.

4Christophe Gol´eNote the difference in the boundary conditions (1.1) and that of Conley-Zehnder (0.1) : theirs allow basically all pseudo Riemannian metrics thatare completely integrable and constant. Ours only deals with Riemannianmetrics, but with no further condition.

Note also that the orbits they findare homotopically trivial. We refer the reader to [G1,2], [J] for the study ofthe homotopically nontrivial case for M = T ∗Tn (the former with a methodakin to that of this paper, the latter in the spirit of [CZ 1]).The method used to prove Theorem 1 and 2 is quite different from thatof Conley and Zehnder: whereas they use cut-offs on Fourier expansions,we decompose the time 1 map into “symplectic twist maps” to get a finitedimensional variational problem.Symplectic twist maps are a natural generalization of monotone twistmaps of the cotangent bundle of the circle (i.e.

the annulus).In short, a symplectic twist map is a diffeomorphism F from some neigh-borhood U of the zero section of T ∗M onto itself with the property thatF ∗pdq −pdq = dS for some S and that (q, p) →(q, Q) is a change ofcoordinates, where F(q, p) = (Q, P).To give an example, we make the following trivial remark. The shearmap of the annulus :(q, p) →(q + p, p),which is a key model in the twist map theory, is nothing more than the time1 of the Hamiltonian H0(q, p) = 12p2 and in fact, its first coordinate map:T ∗S1 →S1(q, p) →q + pis just the exponential map for the standard (flat) metric on S1×IR = T ∗S1.This suggest that the key model for symplectic twist maps on the cotan-gent bundle T ∗M of a general compact manifold M should be the time onemap of a metric.

The twist condition is given in that case by the fact thatthe exponential map is a diffeomorphism of a neighborhood of zero in eachfiber T ∗q M and a neighborhood of q in M. Of course, most of the time, sucha map is not completely integrable.If F is a symplectic twist map, we have a simple proof of the originalconjecture of Arnold:Theorem 3(Banyaga, Gol´e [BG] ) Let F be a symplectic twist map ofB∗M. Suppose that each sphere ∂B∗qM links with its image by F in ∂B∗M.Then the fixed points of F are given by the critical points of a real valuedfunction on M.In Appendix B, we reproduce the proof of [BG], for the convenienceof the reader.

As in all these questions about fixed points, the major taskis to make the argument global: symplectic twist maps should be seen as

Periodic Orbits5local objects (even though they should not be seen as perturbations) andthe problem is to piece them together to form global ones. Here is onefundamental principle involved in this.Suppose we have two “exact symplectic maps”:F ∗pdq −pdq = dS and G∗pdq −pdq = dS′Then it is simple to see that:(F ◦G)∗pdq −pdq = d(S ◦G + S′)which we express as : generating functions add under compositions of maps.This simple fact is key to the method in this paper: the functional we useis a sum of generating functions of a finite sequence of twist maps thatdecompose the time 1 map we study.This additivity property is the common thread between the methodexposed here and that of “broken geodesics” reintroduced in symplecticgeometry by Chaperon [Ch2].

The essential difference is in the choice ofcoordinates in which one expresses the generating function :(p, Q) in themethod of Chaperon, (q, Q) in the twist map method. In this sense thetwist map method is closer to the original method of broken geodesics asdiscribed in [Mi].

It even coincides with it in the case of the geodesic flow.Whereas Moser [Mo2] noticed that the time 1 map of a two dimensionalconvex Hamiltonian can be decomposed into a product of twist maps, theidea of decomposition of a time 1 map of a general Hamiltonian stems fromthe work of LeCalvez [L] on twist maps of the annulus. We generalize hissimple but extremely efficient construction to any cotangent bundle (Lemma(3.4)).There are various theorems on the suspension of certain classes of sym-plectic twist maps by Hamiltonian flows ([D], [Mo2], [P-B]).

In this sense,one might decide to forget about symplectic twist maps and concentrate onHamiltonian systems instead. In this paper, we take the opposite point ofview: we think that symplectic twist maps are a very useful tool to studyHamiltonian systems on cotangent bundles (see also the work of LeCalvez[L] on the torus).The rest of the paper is organized as follows:In section 2, we review some facts about geodesic flows and exponentialmaps.

We prove a lemma which is crucial for the construction of an isolatingblock in section 4.In section 3 we give a precise definition of symplectic twist maps andprove the Decomposition Lemma (3.4).In section 4, we use this decomposition and the additive property ofgenerating functions to construct a finite dimensional variational problem,i.e. a functional W on a finite dimensional space.

This method is basicallyAubry’s ( [Au], [Ka]) , when seen on maps of the annulus.

6Christophe Gol´eIn section 5, we construct an isolating block for the functional W. Forthis, the boundary condition in the theorem is crucial.In section 6, we make use of a theorem of Floer [F1] on global continua-tion of normally hyperbolic invariant sets: we exhibit such an invariant setfor the time 1 map of H0 whose cohomology survives under a deformationto our H. We then use the Conley-Zehnder Morse theory to finish the proofof Theorem 1.In section 7, we show how to adapt the proof of theorem 1 to the caseof non trivial homotopy classes, and prove theorem 2.In Appendix A, we outline the connection that there is between theindex of the Hessian of W and the Floquet multipliers along a closed orbitof F. This is used in sections 5 and 7 to prove normal hyperbolicity of theinvariant set.In Appendix B, we reproduce the proof of Theorem 3, given in [BG].1 A few facts about the geodesic flowWe start with some notation. Let (M, g) be a Riemannian manifold.

Boththe tangent fiber TqM and the cotangent fiber T ∗q M are endowed withbilinear forms:(v, v′) →g(q)(v, v′) for v, v′ ∈TqM, and(p, p′) →g#(q)(p, p′) for p, p′ ∈T ∗q M.We will denote by∥v∥:=pg(q)(v, v′), and ∥p∥:=qg#(q)(p, p),hoping that the context will make it clear whether we speak about a vectoror a covector.The relation between g and g# is better understood in local coordinates:If A(q) denotes the matrix for g# then A−1(q) is the matrix for g. The termsof these matrices are usually denoted gij and gij respectively. The matrixA(q) also gives the standard ( although g– dependent) isomorphism betweenT ∗q M and TqM, which is an isometry for the above metrics.

We will use thesame notation “A(q)” for this isomorphism, even though it is coordinateindependent, whereas the matrix is not.We want to outline here some connections between the geodesic flowfor the metric g, the exponential map and the Hamiltonian flow for theHamiltonian:H0(q, p) = 12∥p∥2.

Periodic Orbits7Let T ∗M be given the usual symplectic structure dp ∧dq, and canonicalprojection π. Let ht0 denote the time t map of the Hamiltonian flow of H0.Then:expq(tA(q)p) = πht0(q, p),This is basically a rewording of the equivalence of Hamilton’s andLagrange’s equations under the Legendre transformation.

Here H0 andL0(q, ˙q) = 12∥˙q∥2 are Legendre transforms of one another under the changeof coordinate ˙q = ∂H0∂p = A(q)p ([Arnold], section 15 or [Abraham-M]theorem 3.7.1 and 3.6.2). This change of coordinate we will refer to as theLegendre transformation as well.What is usually called the geodesic flow is just the flow ht0 restricted tothe (invariant) energy level {(q, p) ∈T ∗M|H0(q, p) = 1 = ∥p∥} (theunit sphere bundle).Because the exponential map :(2.1)exp : TM →M × M(q, v) →(q, Q) := (q, expq(v))defines a diffeomorphism between a neighborhood of the 0–section in TMand some neighborhood of the diagonal in M × M ([Mi], Lemma 10.3) , wealso have, via the Legendre transformation:exp : T ∗M →M × M(q, p) →(q, Q) := (q, expq(A(q)p)which gives a diffeomorphism between a neighborhood of the 0–section inT ∗M and some neighborhood of the diagonal in M ×M.

Just how big theseneighborhoods are is measured by the radius of injectivity R.Because the Legendre transformation A(q) is an isometry, equation (2.1)gives a relation between distances between points in M and norms of vectorsin T ∗M:(q, Q) = exp(q, p) ⇒∥p∥= Dis(q, Q)It will be of interest for us to know the differential of the map “Dis”.Lemma 2.2 If (q, Q) = exp(q, p), and h10(q, p) = (Q, P), then:(2.2)∂1Dis(q, Q) = −p∥p∥and ∂2Dis(q, Q) =P∥P∥Proof.Let v = A(q)p. We have Dis(q, Q) = ∥v∥. The point expq(−tv) ison the same geodesic as the one running from q to Q, namely {expq(tv)t ∈[0, 1]}.

For all small and positive t we must then have:Dis(expq(−tv), Q) = (1 + t) ∥v∥.

8Christophe Gol´eDifferentiating with respect to t at t = 0 yields:(2.3)−∂1Dis(q, Q).v = ∥v∥.On the other hand, by Gauss’ lemma ([Mi], Lemma 10.5), the geodesicthrough q and Q must be orthogonal to the sphere centered at Q and ofradius Dis(q, Q). This sphere is just the level set of the function:q′ →Dis(q′, Q)whose gradient A(q)∂1Dis(q, Q) at q must be colinear to v. Equation (2.3)yields:A(q)∂1Dis(q, Q) = −v∥v∥which immediately translates to the first equation we wanted to prove.For the proof of the second equation, one must remember that V =A(Q)P is tangent at Q to the geodesic between q and Q and has same normas v. It is , more precisely, the parallel transport of v along this geodesic.Thus:ddtDis (q, expq ((1 + t)v))t=0 = ∥V ∥= ∂2Dis(q, Q).Vand the rest of the reasoning is the same as for the first equation.⊓⊔3 Symplectic twist maps and the decompositionlemmaIf H(q, p, t) is an optical Hamiltonian function (i.e.

Hpp is convex), then itsflow has many similar features to that of H0(q, p) = 12∥p∥2. In particular ifF is its time ǫ , and F(q, p) = (Q, P), the correspondance (q, p) →(q, Q) isa diffeomorphism between suitable neighborhoods of the 0–section in T ∗Mand the diagonal in M × M (compare equation 2.1).

This can be seen ina chart, looking at the Taylor series of the solution with respect to smalltime:Q = q(ǫ) = q(0) + ǫ.Hp + o(ǫ2)P = p(ǫ) = p(0) −ǫ.Hq + o(ǫ2),¿From this we see that ∂Q∂p (z(0)) is non degenerate. This remark wasmade by Moser in the dimension 2 case ([Mo2]).Another feature enjoyed by Hamiltonian flows is that they are exactsymplectic.These two properties put together give us the following:

Periodic Orbits9Definition 3.1 A symplectic twist map F is a diffeomorphism of a neigh-borhood U of the 0–section of T ∗M onto itself satisfying the following:(1) F is exact symplectic: F ∗pdq −pdq = dS for some real function S onU. (2) (Twist) if F(q, p) = (Q, P), then the map ψ : (q, p) →(q, Q) is embed-ding of U in M × M.The function S(q, Q) is then called the generating function for F.Remark 3.2Of course ([G1,2,3]), monotone twist maps of the annulus(i.e.

of T ∗S1)) are symplectic twist maps in the sense of this definition. U isusually taken to be either the whole cylinder, or the subset S1 × [0, 1].

Notethat one way to express the twist condition is by saying that the image byF of a (vertical) fiber in U intersects any fiber in at most one point.To my knowledge, the term symplectic twist map was introduced byMcKay, Meiss and Stark. Their definition ([MMS]) is a little more restric-tive than the above, in that they work on T ∗Tn and ask that ∂Q/∂p bedefinite positive.

Our condition only implies that det(∂Q/∂p) ̸= 0. Similarmaps have also been studied extensively by Herman ([H]): he called themmonotone.

We also have used this terminology ([G1,2]) but in the end foundit misleading because we were also dealing with monotone flows [G2], thetwo concepts being only related in certain cases.Remark 3.3 Equation (2.1) tells us that the time 1 map h10 of H0 isalso a symplectic twist map on some neighborhood U of the 0–section. Notethat for the time 1 of an Hamiltonian H, the function S is (when defined)the action:S(q, Q) =Z (P,Q)(p,q)pdq −Hdttaken along the unique solution of the Hamiltonian flow between (p, q) and(P, Q).

If L is the Legendre transform of H, the above integral is justS(q, Q) =Z 10L(q, ˙q, t)dtalong the solution. In the case where H = H0, L(q, ˙q, t) = 12 ∥˙q∥2, i.e.

S isthe energy of the (unique) geodesic between q and Q.As noted in the introduction, h10 should be our model map, the way theshear map is the model map in the theory of monotone twist maps.The reason why twist maps can be so useful lies in the following funda-mental lemma, due to LeCalvez [L] in the case of diffeomorphisms of theannulus isotopic to the Id:Lemma 3.4 (Decomposition) (LeCalvez, Banyaga, Gol´e):

10Christophe Gol´eLet F be the time 1 map of a (time dependent) Hamiltonian on a compactneighborhood U of the 0–section. Suppose that F leaves U invariant.

Then,F can be decomposed into a finite product of symplectic twist maps:F = F2N ◦. .

. ◦F1Remark 3.5No convexity is assumed of the Hamiltonian, nor any close-ness to an integrable one.Proof.Let Gt be the time t map of the Hamiltonian, starting at t = 0.We can write :F = G1 ◦G−1N−1N◦.

. .

◦G kN ◦G−1k−1N ◦. .

. ◦G 1N ◦Idand each of the G kN ◦G−1k−1N is an exact symplectic map, which we can make asclose as we want to the Id by increasing N. If Hpp is positive definite, each ofthese maps are twist, by Moser’s remark, and we are done (F is the productof N twist maps in this case) .

In general, we do the following. The twistcondition (2) in Definition (3.1) of symplectic twist is an open condition.Hence, if ht0 is the time t map of H0, the map F2k−1 := h−10◦G kN ◦G−1k−1Nmust satisfy (2) for N big enough (here, the compactness of U is needed).We then set F2k = h10 for all k to get the decomposition advertized.⊓⊔Remark 3.6We leave it to the reader to check that Lemma 3.4 is alsovalid for lifts of maps to the covering space of M.4 The discrete variational settingLet F be as in Theorem 1.

From the previous section, we can writeF = F2N ◦. .

. ◦F1,with the further information that F2k restrained to the boundary ∂B∗M ofB∗M is the time 1 map of H0, that we have called h10.

Likewise, F2k−1 ish1N −10on ∂B∗M, by the proof of the decomposition Lemma (3.4) and theboundary condition (1.1) imposed on F.Let Sk be the generating function for the twist map Fk and ψk thediffeomorphism (q, p) →(q, Q) induced by Fk. We can assume that ψk isdefined on a neighborhood U of B∗M in T ∗M.

Let(4.1)O = {q = (q0, . .

., q2N−1) ∈M 2N |(qk, qk+1) ∈ψk(U) and(q2N, q0) ∈ψ2N−1(U)}

Periodic Orbits11O is an open set in M 2N, containing a copy of M (the elements q such thatqk = q0, for all k).Next, define :(4.2)W(q) =2N−1Xq=0Sk(qk, qk+1),where we have set q2N = q0. Let pk be such that ψk(qk, pk) = (qk, qk+1)and let Pk be such that Fk(qk, pk) = (qk+1, Pk).

Pk and pk are well definedfunctions of (qk, qk+1).We claim:Lemma 4.3 The sequence q of O is a critical point of W if and only if thesequence {(qk, pk)}k∈{0,...,2N,0} is an orbit under the successive Fk’s, that isif and only if (q0, p0) is a fixed point for F.Proof.Because the twist maps are exact symplectic and using the defini-tions of pk, Pk, we have:(4.4)Pkdqk+1 −pkdqk = dSk(qk, qk+1),and hencedW(q) =2N−1Xk=0(Pk−1 −pk)dqkwhich is null exactly when Pk−1 = pk, i.e. when Fk(qk−1, pk−1) = (qk, pk).Now remember that we assumed that q2N = q0.⊓⊔Hence, to prove Theorem 1, we need to find enough critical points forW.

For this , we will study the gradient flow of W (where the gradient willbe given in terms of the metric g) and use the boundary condition to findan isolating block.We now indicate how this variational setting is related to the classicalmethod of broken geodesics, and how to modify it to deal with homotopicallynon trivial solutions.Because each Fk is close to htk0 for some positive or negative tk, we havethat:q ∈ψk(B∗Mq)and, since B∗qM →ψk(B∗M) is a diffeomorphism, we can define a pathck(q, Q) between q and a point Q of ψk(B∗qM) by taking the image of theoriented line segment between ψ−1k (q) and ψ−1k (Q) in B∗qM. In the casewhere Fk = h10, this amounts to taking the unique geodesic between q andQ in ψk(B∗qM) .If we look for periodic orbits of period d and of a given homotopy type,we decompose F d into 2Nd twist maps, by decomposing F into 2N.

Anal-ogously to (4.1), we define :

12Christophe Gol´eOd = {q = (q0, . .

., q2Nd−1) ∈M 2Nd |(qk, qk+1) ∈ψk(U) and(q2Nd, q0) ∈ψ2Nd−1(U)},remarking that the ψk’s here correspond to the decomposition of F d into2Nd steps (U is as before a neighborhood of B∗M).To each element q in Od, we can associate a closed curve, made by joiningup each pair (qk, qk+1) by the unique curve ck(qk, qk+1) defined above. Thisloop c(q) is piecewise differentiable and it depends continuously on q, andso does its derivatives (left and right).

In the case of the decomposition ofh10 , taking Fk=h10, this is exactly the construction of the broken geodesics([Mi], §16). Now any closed curve in M belongs to a free homotopy class m.To any d periodic point for F, we can associate a sequence q(x) ∈Odof q coordinates of the orbit of this point under the successive Fk’s in thedecomposition of F d.Definition 4.5Let x be a periodic point of period d for F. Let q be thesequence in Od corresponding to x.

We say that x is an (m, d) point ifc(q(x)) is in the free homotopy class m.To look for (m, d) orbits in (Theorem 2 in section 7), we will work in:(4.6)Om,d = {q ∈O | c(q) ∈m}Since c(q) depends continuously on q ∈O, we see that Om,d is actually aconnected component of O.The functional W will be given this time by:W(q) =2Nd−1Xk=0Sk(qk, qk+1)defined on Om,d. Again, as in Lemma 4.3, critical points of W in Om,dcorrespond to (m, d) periodic points.Remark 4.7 The reader that wants to make sure that, in the proof ofTheorem 1, the orbits found are homotopically trivial, should check thatthroughout the proof, one can work in the component Oe,1 of O1 = O ofsequences q which have c(q) ∈e, where e is the identity element of π1(M).

Periodic Orbits135 The isolating blockIn this section we prove that the set B defined as follows:(5.1)B = {q ∈O | ∥pk(qk, qk+1)∥≤C}is an isolating block for the gradient flow of W, where O is defined in (4.1),C is as in (1.1) and pk is the function defined in the previous section (seebelow (4.2)) . To try to visualize this set in M 2N, the reader should realizethat the twist condition on Fk and the fact that Fk coincides with the time1 or time1n −1 of the hamiltonian H0 at the boundary of B∗M impliesthat:(5.2)Dis(qk, qk+1) = ak ∥pk∥whereak=1if k is evenak=1−NNif k is oddNote that B still contains a copy of M ( the constant sequences).We will define an isolating block for a flow to be a compact neighbor-hood with the property that the solution through each boundary point ofthe block goes immediately out of the block in one or the other time direc-tion ( [C], 3.2 ).

Sometimes, more refined definitions are made, but this oneis sufficient to ensure that the maximal invariant set for the flow containedin the block is actually contained in its interior: a block in this sense is anisolating neighborhood, which is really the only property we need here.Proposition 5.3 B is an isolating block for the gradient flow of W.Proof.Suppose the point q of U is in the boundary of B. this meansthat ∥p∥k = C for at least one k. As noted in (5.2), this means thatDis(qk, qk+1) = akC for some factor ak only depending on the parity ofk. We want to show that this distance increases either in positive or nega-tive time under the gradient flow of W. This flow is given by:(5.4)˙qk = Ak(Pk−1 −pk) = ∇Wk(q)Where Ak = A(qk) is the duality morphism associated to the metric g atthe point qk (see beginning of section 2).

Remember that we have put theproduct metric on O, induced by its inclusion in M 2N.Let us compute the derivative of the distance along the flow at a bound-ary point of B, using Lemma 2.1:(5.5)ddtDis(qk, qk+1)t=0 = ∂1Dis(qk, qk+1).∇Wk(q)+ ∂2Dis(qk, qk+1).∇Wk+1(q)= ( ak|ak|) −pk∥pk∥.Ak. (Pk−1 −pk)+ ( ak|ak|) Pk∥Pk∥.Ak+1.

(Pk −pk+1)

14Christophe Gol´eWe now need a simple linear algebra lemma to treat this equation.Lemma 5.6 Let ⟨, ⟩denote a metric form in IRn, and ∥.∥its correspondingnorm. Suppose that p and p′ are in IRn ,that ∥p∥= C and that ∥p′∥≤C.Then :⟨p , p′ −p ⟩≤0.Moreover, equality occurs if and only if p′ = p.Proof.From the positive definiteness of the metric, we get:⟨p′ −p, p′ −p ⟩≥0,with equality occuring if and only if p′ = p (call this last assertion *).

Fromthis, we get:2⟨p, p′ ⟩≤⟨p′, p′ ⟩+ ⟨p, p ⟩with *. Finally,⟨(p′ −p), p ⟩= ⟨p′, p ⟩−⟨p, p ⟩≤0with *.⊓⊔Applying Lemma 5.6 to each of the right hand side terms in (5.5), wecan deduce thatddtDis(qk, qk+1) is positive when k is pair, negative when kis odd.

Indeed, because of the boundary condition in the hypothesis of thetheorem, we have ∥Pk∥= ∥pk∥whenever ∥pk∥= C: the boundary ∂B∗Mis invariant under F and all the Fk’s. On the other hand q ∈B ⇒∥pl∥≤C and ∥Pl∥≤C, for all C, by invariance of B∗M .

Finally, ak is positivewhen k is even, negative when k is odd.But what we really want is this derivative to be of a definite sign, notzero. It is certainly the case when at least one of ∇Wk(q), ∇Wk+1(q) is notzero.

Suppose they are both zero. Then k is in an interval {l, .

. ., m} suchthat, for all j in this interval, ∥pj∥= C = ∥Pj∥and ∇Wj(q) = 0.It is now crucial to notice that {l, .

. ., m} can not cover all of {0, .

. ., 2N}:this would mean that q is a critical point corresponding to a fixed point of h10in ∂B∗M.

But such a fixed point is forbidden by our choice of C: geodesicsin that energy level can not be fixed loops (C > 0), and they can not closeup in time one either (C is less than the injectivity radius).We now let k = m in (5.5) and see that the flow must definitely escapethe set P at q in either positive or negative time, from the the mth face ofP.⊓⊔Remark 5.7If we have decomposed the time 1 map of a Hamiltonianthat is positive definite into a product of N twist maps, all the Fk’s coincidewith h1N0on the boundary of B∗M. In that case,∥pk∥= 1N Dis(qk, qk+1), for all k

Periodic Orbits15and the ak’s in the above proof are always positive. Following the argumentthrough, we find that B is a repeller block in this case: all points on ∂Bexit in positive time.Remark 5.8LeCalvez ([L]) provides a more detailed analysis of the be-havior of the flow at “corner” points of his analog of the set B.

He indicatesan induction to show that the flow enters or exits the jth face (j is in{l, . .

., m} as in the above proof) at different orders in small time. Such areasoning could be made in our context also, but we find it unnecessary,given our working definition of an isolating block.6 Proof of Theorem 1To finish the proof of Theorem 1 we will be using a refinement of the Con-ley Index continuation proved by Floer ([F1]).

The homology group of theinvariant set Gλ appearing in this lemma bears the germs of what becamelater Floer Cohomology (see e.g. [F2], and also [McD]) , and in the casethat we study, it is probable that it is one and the same thing.

The presentapproach enables us to avoid the problem of infinite dimensionality in [F2],i.e. all the analysis.Lemma 6.1(Floer) Let φtλ be a one parameter family of flows on a C2manifold M. Suppose that G0 is a compact C2 submanifold invariant underthe flow φt0.

Assume moreover that G0 is normally hyperbolic, i.e. thereis a decomposition:TM|G0 = TG0 ⊕E+ ⊕E−which is invariant under the covariant linearization of the vector field V0corresponding to φt0 with respect to some metric ⟨, ⟩, so that for someconstant m > 0:6.1⟨ξ, DV0ξ⟩≤−m⟨ξ, ξ⟩for ξ ∈E−⟨ξ, DV0ξ⟩≥m⟨ξ, ξ⟩for ξ ∈E+Suppose that there is a retraction α : M →G0 and that there is a com-pact neighborhood B which is isolating for all λ.

Then, if Gλ denotes themaximum invariant set for φtλ in B, the map:α|Gλ∗: H∗(G0) →H∗(Gλ)in ˇCech cohomology is injective.In this precise sense, normally hyperbolic invariant sets continue globally:their topology can only get more complicated as the parameter varies awayfrom 0. Note that we have given here a watered down version of Floer’s

16Christophe Gol´etheorem. His uses the notion of Conley continuation of invariant sets.

Healso works in the equivariant case. But the above, taken from his Theorem2 in [F1], is what we need here.The family of flows we consider is ζtλ, the flow solution ofddtq = ∇W λ(q),and W λ is defined as in 4.2 for the map Fλ, time 1 map of the Hamiltonian:Hλ = (1 −λ)H0 + λHWe can assume that this construction is well defined, i.e., that we makethe decomposition in the Decomposition Lemma 3.4 fine enough to fit anyFλ, λ in [0, 1].

The manifold on which we consider these (local) flows is O,an open neighborhood of B in M 2N. Of course, each of the Fλ satisfies thehypothesis of Theorem 1, and thus Proposition 5.3 applies to ζtλ for all λ in[0, 1]: B is an isolating block for each one of these flows.The part of Floer’s lemma that we are missing so far is the normallyhyperbolic invariant manifold for ζt0.Lemma 6.2Let G0 = {q ∈B | qk = q0, ∀k}.

Then G0 is a normallyhyperbolic invariant set for ζt0. It is a retract of O and is the maximalinvariant set in B.Proof.All the Fk’s in the decomposition of h10 are time ak maps of theHamiltonian H0, for ak as in (5.2).

But for this Hamiltonian, the 0–sectionof T ∗M is made out of fixed points. These translate, in terms of sequences,to points in G0.

Moreover, these are the only periodic orbits for the Hamil-tonian flow of H0 in B∗M, by the definition of this set. (e.g.

in the caseM = Sn with the standard metric, the orbits corresponding to great circleswould not be fixed points of h10 in B∗M).This implies that G0 is the maximum invariant set for ζt0 in B. Indeed,since ζt0 is a gradient flow, such an invariant set should be formed by criti-cal points and connections between them. We saw that there are no othercritical points but the points of G0.

If there were a connection orbit entirelylying in B, it would have to connect two points in G0, which is absurd sinceby continuity any two points of G0 give the same value for W 0, whereas W 0should increase along non constant orbits.G0 is a retract of M 2N under the composition of the maps:q = (q1, . .

., q2N) →q1 →(q1, q1, . .

., q1) = α(q)which is obviously continuous and fixes the points of G0.We are left to show that G0 is normally hyperbolic. For this, we aregoing to appeal to a relationship between the linearized flow of ζtλ and that

Periodic Orbits17of Hλ. The following lemma was proven by McKay and Meiss in the twistmap of the annulus case.

We present their proof in Appendix A: it holds inthe setting of general cotangent bundles.Lemma 6.3 ( [M-M]) Let F be the time 1 map of a Hamiltonian and let Wbe its associated functional. If q is a critical point corresponding to the orbitof (q0, p0), the set of eigenvectors of eigenvalue 1 of DF(q0,p0) are in 1–1correspondence with the set of eigenvectors of eigenvalue 0 of HessW(q)To use this lemma, we remark that since G0 is made out of critical points,saying that it is normally hyperbolic is equivalent to saying that HessW 0(q)has exactly n = dimG0 eigenvalues equal to zero for any point q in G0.

Theseeigenvalues have to correspond to eigenvectors in TG0, the normal space ofwhich must be spanned by eigenvectors with non zero eigenvalues (HessW 0is symmetric). Hence, from Lemma 6.3, it is enough to check that at a point(q0, 0) ∈B∗M corresponding to q, 1 is an eigenvalue of multiplicity exactlyn for Dh10(q0, 0).

Let us compute Dh10(q0, 0) in local coordinates. It is thesolution at time 1 of the linearized equation:˙U = JHessH0(q0, 0)Ualong the constant solution (q(t), p(t)) = (q0, 0), where J denotes the usualsymplectic matrix0I−I0.

An operator solution for the above equationis given by exp (tJHessH0(q0, 0)) On the other hand:HessH0(q0, 0) =000A(q0)which we computed from H0(q, p) = A(q)p.p, the zero terms appearing atp = 0 because they are either quadratic or linear in p. ¿from this,Dh10(q0, 0) = exp (JHessH0(q0, 0)) =IA(q0)0Iis easily derived. This matrix has exactly n eigenvectors of eigenvalue 1 ( ithas in fact no other eigenvector).

Hence, from Lemma 6.3, HessW(q) hasexactly n vectors with eigenvalue 0, as was to be shown.⊓⊔We now conclude the proof of Theorem 6.3. We have proved that the flowζt, which is gradient, has an invariant set G = G1 with H∗(M) ֒→H ∗(G).From this we get in particular:cl(G) ≥cl(M) and sb(G) ≥sb(M).The corollary of Theorem 5 in [CZ 1] tells us that ζt must have at leastcl(G) rest points in the set G, whereas the generalized Morse inequalitiesin Theorem 3.3 of [CZ 2] tell us that, if they are all assumed to be non

18Christophe Gol´edegenerate, ζt must have sb(G) rest points. But Lemma 6.3 tells us that nondegeneracy for HessW at a critical point is the same thing as nondegeneracyof a fixed point for F (no eigenvector of eigenvalue 1).As was stated in Remark 4.7, we could have worked in Oe,1 all along toguarantee that the orbits found are homotopically trivial.

The only thingthat one should check is that G0m,d is indeed in this component of O, whichis the case. This concludes the proof of Theorem 1.⊓⊔7 Negative curvature and orbits of different homotopytypes7.1 Setting the problemWe are going to sketch here the changes needed in the proof of Theorem 1in order to prove Theorem 2 on the existence of orbits of different homotopytypes.It is a known (see e.g.

[GHL], 2.98) that on a compact Riemannianmanifold there exists in any nontrivial free homotopy class m a smooth andclosed geodesic which is of length minimal in m. 3Moreover, a theorem of Cartan asserts that if the manifold is of negativecurvature, there is in fact one and only one geodesic in each class m (m notcontaining the point curves ([Kl], Theorem 3.8.14)).Let M be of negative curvature and let l(m) denote the length of thegeodesic in m in that case.In 4.5, we have defined (m, d) orbits by saying that a certain curve thatthe orbit defines in M is of class m. We could also use a favored lift F of Fto the covering space M of M to define such orbits, by asking:Fd(x) = m.xwhere m.x denotes the action of m seen as a deck transformation in T ∗M(the favored lift is the one corresponding to lifting the solution curves ofthe Hamiltonian flow). But definition 4.5 turns out to be more convenientto use here (both are equivalent, of course).

We now restate:Theorem 2 Let (M,g) be a Riemannian manifold of negative curvature. andH be as in Theorem 1.

Then, whenever (M, g) has a geodesic whose classin π1(M) is m, F has at least 2 (m, d) orbits in B∗M when l(m) < dC .3 We remind the reader that free homotopy classes of loops differ from elementsof π1(M) in that no base point is kept fixed under the homotopies. As a result,free homotopy classes can be seen as conjugacy classes in π1(M), and thus cannot be endowed with a natural algebraic structure.

Two elements of a free classgive the same element in H1(M). Hence free homotopy classes form a set smallerthan π1(M), bigger than H1(M).

All these sets coincide if π1(M) is abelian.

Periodic Orbits19Remark 7.1.1The fact that we find two orbits and not a number givenby the topology of the manifold is not an artifact of the proof, but derivesfrom the unicity of the closed geodesic in a given class. Note also that we donot guarantee that an orbit of the form (mk, kd) is not actually an (m, d)orbit.

We should then ask for (m, d) to be prime, in that very sense.Note also, there are a priori more of these pairs (m, d) than there arerational homology directions.The proof of Theorem 2 has the same broad outline as that of Theorem1.We decompose F = F2N ◦. .

. ◦F0 as before, which gives us a decompo-sition of F d into 2Nd twist maps.We would like to claim, in analogy to Proposition 5.3 thatB = {q ∈Om,d | ∥pk(qk, pk)∥≤C}is an isolating block for the gradient flow of W.But this will not be enough for our purpose.

To make sure that twocritical points correspond to points that are actually on 2 distinct orbits,one should do the following: to decompose F d, we have decomposed F in2N steps. Define:σ : Om,d →Om,d by setting (σq)k = qk+2Nwhere we identify: qk+2Nd = qk.

It is clear that 2 critical sequences corre-sponding to points in the same orbit by F get identified in the quotient bythe action of σ. So our candidate for isolating block will be given by thequotient B/σ of B by the action of σ (note that σ leaves B invariant, sothat the quotient makes sense).It can be seen that if m is non trivial, then the action of σ is without fixedpoints.

Since it is also periodic, the action is then properly discontinuous([Gr], Chapter 5), and hence the quotient map Om,d →Om,d/σ is a coveringmap.We now describe the candidate for normally hyperbolic invariant set. Itwill be the quotient by σ of the set G0m,d made of the critical sequences cor-responding to the continuum of (m, d) orbits that form the closed geodesicof class m, parametrized so that the Hamiltonian flow goes through it intime d. Call this orbit γ.Note that G0m,d contains all the possible critical points for W 0 in Om,dsince a critical point for W 0 must be contained in a continuum of criticalpoints: if (q0, p0) is an (m, d) point, so is ht0(q0, p0), for any t. But we know,in the case of H0 that there is one and only one such set, namely γ.Writing F 0k ◦.

. .

◦F 01 = φk, where the F 0k ’s decompose the map hd0, wecan write:G0m,d = {q(t) ∈Om,d | qk = π ◦φk(γ(t))}where π(q, p) = q is the canonical projection. Again, since σ restricted toG0m,d actually corresponds to the action of F on γ, G0m,d is σ invariant

20Christophe Gol´eand hence the quotient G0m,d/σ makes sense. Since the quotient map is acovering map and G0m,d ∼= γ, we have :(7.1.2)G0m,d/σ ∼= S17.2 Proof of Theorem 2Lemma 7.2.1 B/σ is an isolating blockProof.Because we have assumed l(m) < dC, F cannot have any (m, d)orbits confined to ∂B∗M: since F coincides with F0 on this set, such anorbit would have to correspond to a closed geodesic of free homotopy classm, but of length dC, which is absurd.

This in turn implies that W has nocritical points on ∂B, and the reasoning of Proposition 5.3 applies withoutchange to show that points in ∂B must exit B in positive or negative time.Since the covering map is a local diffeomorphism, this is also true in B/σ,which is then an isolating block.⊓⊔Lemma 7.2.2 G0m,d/σ is a normally hyperbolic invariant set for ζt0. It is aretract of Om,d/σ.Proof.We prove the statement “upstairs”, taking the quotient by σ onlyat the end.According to Lemma 6.3, and the reasoning in the proof of Lemma 6.2,it is enough to show that the differential of hd0 on a point of γ has no othereigenvector of eigenvalue 1 than the vector tangent to γ.To compute the differential of hd0 at the point (q0, p0) = γ(0) we are goingto choose a coordinate system (z, t, s) around (q0, p0) in the following way:z, t will be a coordinate system for a tubular neighborhood in the energysurface containing (q0, p0), t being in the direction of γ.

We will define s bythe following: a point (q, p) on the energy level of γ will be assigned coordi-nates (z, t, 1) and the point (q, sp) will be assigned coordinates (z, t, s). It isclear that in an interval s ∈(a, b), a > 0, this gives a system of coordinates.It is interesting to notice that (0, t, 1) is a parametrization of γ, whereasthe cylinder (0, t, s) is foliated by circles s = c invariant under the flow ht0:each one corresponds to a reparametrization of γ, by rescaling the velocityby s.The map hd0 leaves the cylinder invariant and in fact induces a monotonetwist map on it :hd0(0, t, s) = (0, t + (s −1)d, s)Now, remember that the geodesic flow of a manifold with strictly negativecurvature is Anosov.

This translates into: in the subspace tangent at a point

Periodic Orbits21(0, t, s) to the z coordinate, Dhd0 has no eigenvalue equal to 1 ( we can assumethe splitting tangent to t, s to be invariant by Dht0) . Hence, in the (z, t, s)coordinates:Dhd0(0, s, t)A1d01where A has no eigenvalue 1.

Hence Dhd0(0, t, s) has only the vector tangentto γ as eigenvector with eigenvalue 1 (I am endebted to Leonid Polterovitchfor giving me the idea of this argument).We now have to prove that G0m,d is a retract of Om,d. Define the followingmapρ : Om,d(2N) →Om,d(N)(q0, q1, q2, .

. ., q2Nd) →(q0, q2, .

. ., q2k .

. .

, q2Nd)It is not hard to see that ρ induces a diffeomorphism on G0m,d: the projectionon the 0th factor would itself give a diffeomorphism. We claim that the imageG of G0m,d under ρ is a deformation retraction of the image O of Om,d inOm,N.

Call r this retraction. Then ρ−1|G ◦r ◦ρ is a retract of Om,d to G0m,d,as we want to prove.We now construct the map r. Decompose hd0 = (h1N0 )Nd.

Since h1N0is asymplectic twist map , we can rig up the variational setting relative to thisdecomposition. CallW(q) =k=NdXk=1S(qk, qk+1)where S is the generating function of h1N0 .

In this case, the isolating blockB = ρB is a repeller block for the gradient flow of W (see Remark 5.7).Hence W has attains a minimum value, say a, in the interior of B. It has tobe at a point in G, which contains all the critical points of W, as we remarkedabove for G0m,d.

Hence on all of G , W must equal a. Since we can choose Oto be exhausted by an increasing sequence of repeller neighborhoods of thesame type as P, a is actually a global minimum for the function W in O.The same argument as for G0m,d shows that G is normally hyperbolic.

Inparticular, this implies that the set W ≤a+ǫ forms a tubular neighborhoodof G ([DNF],§20). Then a standard argument ([Mi] Theorem 3.1) in Morsetheory shows that, since there are no other critical points but those in thelevel W = a, the set W ≤a + ǫ must be a deformation retraction of O.Finally, G is a deformation retraction of W ≤a + ǫ, since the latter is atubular neighborhood of G. This finishes the construction of r.Finally, we indicate how all these features go through in the quotient byσ.To check that G0m,d/σ is normally hyperbolic, we just note that this no-tion is a local one, in the tangent space, and the quotient map is a local

22Christophe Gol´ediffeomorphism. It can be checked that the above construction of the retrac-tion map is σ invariant.

And, finally, the quotient of our invariant set G0m,d(see 7.3) is a circle, as noted in 7.1.2. This concludes the proof of Lemma7.2.2.⊓⊔To finish the proof of Theorem 2, we use Floer’s Lemma, as in the proofof Theorem 1, to find that there is an invariant set Gm,d for the flow ζt inOm,d/σ which is such that:H∗(G0m,d/σ) = H∗(S1) ֒→H∗(Gm,d)Since cl(S1) = sb(S1) = 2, in all cases, we will get at least 2 distinct orbitsof type (m, d).⊓⊔Appendix A Linearized gradient flow vs.linearizedHamiltonian flowSuppose that (q0, p0) = x0 is a fixed point for F. We want to solve theequation:(A.1)DFx0(v) = λvwith v ∈T(T ∗M)x0.

In terms of Hamiltonian flow , we want to find theFloquet multipliers of the periodic orbit corresponding to x0.In the (qk, qk+1) coordinates, we want to express a condition on the orbit(δqk, δqk+1) of a tangent vector (δq1, δq2) under the successive differentialsof the maps Fk−1 along the given orbit. A way to do it is the following([M-M]): If q corresponds to the orbit of x0 under the the successive Fk’s,it must satisfy:∂W(q)∂qk= ∂2Sk−1(qk−1, qk) + ∂1Sk(qk, qk+1) = 0(see 4.3).

Therefore, a “tangent orbit” δq must satisfy:(A.2)Sk−121 δqk−1 + (Sk11 + Sk−122 )δqk + Sk12δqk+1 = 0where we have abbreviated:Skij = ∂ijSk(qk, qk+1).When q corresponds to a fixed point (q0, p0). Equation A.1 translates,in terms of the δq, to:(A.3)δq2N = λδq0

Periodic Orbits23Equations (A.2) and (A.3) can be put in matrix form as M(λ)δq = 0 whereM(λ) is the following 2Nn × 2Nn tridiagonal matrix:M(λ) =S022 + S111S1120. .

.01λS021S121S122 + S211S212...00S212.........00. . .0S2N−112λS0120.

. .0S2N−121S2N−122+ S011Hence the eigenvalues of DFx0 are in one to one correspondence with thevalues λ for which detM(λ) = 0.

More precisely, to each vector v solutionof (A.1) corresponds one and only one vector δq solution of M(λ)δq = 0.Setting λ = 1, this proves Lemma 6.3.Remark (A.4) This construction can be given a symplectic interpre-tation: the Lagrangian manifolds graph(dW) and graph(F) are related bysymplectic reduction. Lemma 6.3 can then be restated in terms of the in-variance of a certain Maslov index under reduction ([V]).Appendix B: Twist maps and linking of spheresIn this appendix, we present the proof given in [BG] of the original conjec-ture of Arnold in the restrictive case of symplectic twist maps (Theorem 3in the introduction).To that effect, we have to give our interpretation of what linking ofspheres in ∂B∗M is.Call ∆q the fiber of B∗M over q, and ∂∆q its boundary in ∂B∗M.

Then∂∆q is an n dimensional sphere. It make sense to talk about its linking withits image F(∂∆q) in ∂B∗M: the latter set has dimension 2n −1 and thedimensions of the spheres add up to 2n −2.We first restrict ourselves to the case when the two spheres ∂∆q andF(∂∆q) are in a trivializing neighborhood in ∂B∗M, say U × Sn = E.The type of linking of F(∂∆q) with ∂∆q should be given by the class[F(∂∆q)] ∈Hn−1(∂E\∂∆q)More precisely, we have:(B.1)Hn−1(∂E\∂∆q) ∼= Hn−1Sn−1 × (IRn −{0})Kunneth∼=Hn−1(IRn −{0}) ⊕Hn−1(Sn).Thus, taking ∂∆q from ∂E creates a new generator in the n−1st homology,i.e.

the generator b of Hn−1(IRn −{0}).

24Christophe Gol´eDefinition (Linking condition) We say that the spheres F(∂∆q) and ∂∆qlink in ∂E if they do not intersect and if the decomposition of [F(∂∆q)] inthe direct sum in (B.1) has a non zero term in its Hn−1(IRn −{0}) factor.We will say that the symplectic twist map F satisfies the linking conditionif for all q ∈M these spheres link in ∂E for some trivializing neighborhoodE 4If F is a symplectic twist map, it turns out that this is a well definedcharacterization of linking: we can always construct a trivializing neigh-borhood containing both ∂∆q and its image. Indeed, take T ∗(π ◦F(∆q))(homeomorphic to Bn ×IRn since F is twist) if q is in π ◦F(∆q).

If not joinq to this set by a path, and fatten this path. The union of the set and thefattened path is homeomorphic to a ball.

Hence the bundle over this ball istrivial.Moreover, it turns out that if the spheres link in one trivializing neigh-borhood, they do in all of them, as a consequence of the followingLemma B.2 If F is a symplectic twist map, the following are equivalent:a) The spheres ∂∆q and F(∂∆q) link in some trivializing neighborhood in∂B∗Mb) The fiber ∆q and its image F(∆q) intersect in one point of their interior.Remark B.3We can also define the linking condition for a map F ofB∗M which is not necessarily symplectic twist. If the covering space of Mis IRn, we say that F satisfies the linking condition if at least one of its liftsdoes (the trivializing neighborhood is taken to be M × IRn ∼= IR2n in thiscase.) If M is not covered by IRn, Lemma B.2 suggests that we may take asa linking condition that the intersection number ♯(∆q ∩F(∆q)) is ±1.Proof.Suppose that E is a trivializing neighborhood containing the 2spheres.

We complete B.1 into the following commutative diagram:Hn−1(∂E\∂∆q)∼=Hn−1(IRn −{0}) ⊕Hn−1(Sn)y i∗y j∗Hn−1(E\∆q)∼=Hn−1 ((IRn −{0} × Bn))where i, j are inclusion maps. It is clear that j∗b generatesHn−1 ((IRn −{0}) × Bn) ∼= Hn−1 ((IRn −{0}) × IRn) .The last group measures the (usual) linking number of a sphere with thefiber π−1(q) in T ∗E ∼= IR2n.

But it is well known that such a number is4 Here, as a convention, a trivializing neighborhood will always be homeomorphicto Bn × IRn.

Periodic Orbits25the intersection number of any ball bounded by the sphere with the fiberπ−1(q), counted with orientation. In our case, where the sphere considered isF(∂∆q) this number can only be 0 or 1 or -1, because of the twist condition(see Remark 3.2).Conversally, if ∆q and F(∆q) intersect in their interior , then theirbounding spheres must lie on the trivializing neighborhood over F(∆q),and must link.⊓⊔Remark B.4If all fibers intersect their image under a twist map, i.e.if the linking condition is satisfied, then the intersection number must beuniformally 1 or -1: we could call F a positive twist map in the first case,a negative twist map in the second case.

Of course, this corresponds to thesame notion in dimension 2.We can now prove Theorem 3.¿From Lemma 4.3, fixed points of F correspond to critical points ofq →S(q, q). This function only make sense for all q in M if the diagonal inM × M is in the image of U by the embedding ψ (see Definition 3.1).

Thisis exactly the case when q ∈F(∆q) for all q, i.e., from Lemma B.2, exactlywhen the linking condition is satisfied. Hence F has as many fixed pointsas the function q →S(q, q) has critical points on M.⊓⊔To our knowledge, Arnold’s original conjecture is still open, even in thecase M = Tn.References[Ar1] V. Arnold, “ Sur une propri´et´e topologique des applications globalementcannoniques de la m´ecanique classique”, C.R.

Acad.Sc. Paris, t.261 (1965).Groupe 1.

[Ar1] V.I. Arnold: “Mathematical Methods of Classical Mechanics” (Appendix9), Springer-Verlag 1978.[AL]S.

Aubry and P.Y. LeDaeron: “The discrete Frenkel-Kontarova model andits extensions I.

Exact results for ground states”, Physica 8D (1983), 381-422. [BG] A.Banyaga and C. Gol´e : “A remark on a conjecture of Arnold: linkedspheres and fixed points”, to appear in proc.

conf. on Hamiltonian systemsand celestial mechanics, Guanajuato[BK] D. Bernstein and A.B.

Katok: “Birkhoffperiodic orbits for small perturba-tions of completely integrable Hamiltonian systems with convex Hamilto-nians”, Invent. Math.

88 (1987), 225-241.[BP]M. Bialy and L. Polterovitch, “ Hamiltonian diffeomorphisms and La-grangian distributions”, preprint, Tel Aviv University, (1991)[Ch1] M. Chaperon: “Quelques questions de G´eometrie Symplectique”, S´eminaireBourbaki no.

610 (1982/83). [Ch2] M. Chaperon, “Une id´ee du type “g´eod´esiques bris´ees” pour les syst`emeshamiltoniens”, C.R.

Acad. Sc., Paris, 298, S´erie I, no 13, 1984, p 293-296.

26Christophe Gol´e[C]Conley, C.C. : “Isolated invariant sets and the Morse index” , CBMS, Re-gional Conf.

Series in Math., Vol.38 (1978). [CZ1] C.C.

Conley and E. Zehnder: “The Birkhoff-Lewis fixed point theorem anda conjecture of V.I. Arnold”, Invent.

Math. (1983).

[CZ2] C.C. Conley and E. Zehnder: “Morse type index theory for Hamiltonianequations”, Comm.

Pure and Appl. Math.

XXXVII (1984), pp. 207–253.[D]R.

Douady: “Stabilit´e ou instabilit´e des points fixes elliptiques”, Ann. Sci.Ec.

Norm. Sup.

4`eme s´erie, t.21 (1988), pp. 1–46[DNF]B. Doubrovine, S. Novikov, A. Fomenko : “G´eom´etrie contemporaine”, vol3, Editions Mir, Moscow, (1987) (see also english translation in Springer–Verlag)[F1]A. Floer: “A refinement of Conley index and an application to the stabilityof hyperbolic invariant sets”, Ergod.

Th. and Dyn.

Sys., vol 7, ( 1987).[F2]A. Floer: “Morse theory for Lagrangian intersections”, J. Diff.

Geom. 28(1988) pp.

513–547[G1]C. Gol´e, “Periodic points for monotone symplectomorphisms of Tn × IRn”;Ph.D. Thesis, Boston University (1989)[G2]C. Gol´e, “Ghost circles for twist maps”; IMA preprint (1990), to appear inJ. of Diff.

eq. [GH] C. Gol´e and G.R.

Hall: “Poincar´e’s proof of Poincar´e’s last geometric the-orem” To appear in the proceedings for the workshop on twist maps, IMA,Springer[GHL]S. Gallot, D. Hulin and J. Lafontaine: “Riemannian Geometry”, Springer-Verlag, (1987).[Gr]M.J. Greenberg: “Lectures on Algebraic Topology” , Math.

lecture noteseries (5th printing, 1977).[H]M.R. Herman: “Existence et non existence de Tores invariants par desdiff´eomorphismes symplectiques”, S´eminaire sur les ´equations aux d´eriv´espartielles, Ecole Polytechique, Palaiseau (1988).[J]F.W.

Josellis: “Global periodic orbits for Hamiltonian systems on Tn×Rn”Ph.D. Thesis Nr.

9518, ETH Z¨urich, (1991).[Ka]A. Katok: “Some remarks on Birkhoffand Mather twist map theorems”,Ergod.

Th. & Dynam.

Sys. (1982), 2, 185-194.[Kl]W.

Klingenberg: “Riemannian geometry”, de Gruyter studies in Mathemat-ics (1982). [K-M] H. Kook and J. Meiss: “Periodic orbits for Reversible, Symplectic Map-pings”, Physica D 35 (1989) 65–86.[L]P.

LeCalvez: “Existence d’orbites de BirkhoffG´en´eralis´ees pour les diff´eomorphismes conservatifs de l’anneau”, Preprint, Universit´e Paris-Sud, Or-say.[McD]D. McDuff: “Elliptic methods in symplectic geometry”, Bull.

Am. Math.Soc., vol 23, n.2, (1990)[MM]]R.S.

Mackay and J. Meiss: “Linear stability of periodic orbits in Lagrangiansystems”, Phys. Lett.

98A, 92 (1983). [MMS]R.S McKay, J. Meiss and J.Stark: “Converse KAM theory for symplectictwist maps”, Nonlinearity 2 (1989) pp 469–512.[Ma]J.

Mather: “ Action minimizing invariant measures for positive definiteLagrangian systems”, preprint, ETH Z¨urich, (1989)

Periodic Orbits27[Mi,2] J. Milnor: “Morse Theory”, Princeton University Press. [Mo1] J.Moser: “Proof of a generalized form of a fixed point theorem due to G.D.Birkhoff”, Lecture Notes in Mathematics, Vol.

597: Geometry and Topology,pp. 464-494.

Springer (1977)[Mo2] J. Moser: “Monotone twist mappings and the calculus of variations”, Ergod.Th. and Dyn.

Sys., Vol 6 (1986), p 401-413[V]C. Viterbo: “ Intersection de sous vari´et´es Lagrangiennes, fonctionnellesd’action et indice de syst`emes Hamiltoniens”, Bull. Soc.

math. France 115,(1987), pp.

361–390

---

출처: arXiv:9201.297 • 원문 보기