Integrability and non-integrability in Hamiltonian mechanics

HAL is a multi-disciplinary open access archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.


L: Rn{q}xRn{q}--+R
Here p = aLjaq E Rn is the generalized momentum and the Hamiltonian function ll = pq-L lp,q is the "total energy" of the mechanical system. "In part he had been anticipated by the great French mathematicians: for Poisson, in 1809, had taken the step introducing a function (!) and expressing it in terms of q 1 , q 2 , ... , Qn, and had actually derived half of Hamilton's equations: Lagrange in 1810 had obtained a particular set of equations (for the variation of elements) in the Hamiltonian form the disturbing function taking the place of H. Moreover, the theory of non-linear partial differential equations of the first-order had led to systems of ordinary differential equations possessing this form: as was shown by Pfaff in 1814-15 and Cauchy in 1819 (completing the earlier work of Lagrange and Monge), the equations of the characteristics of a partial differential equation (2) where are dp2 dpn -8f/8x2 · · · --8fj8xn ' Hamilton's investigation was extended to the cases when the kinetic potential contains the time, etc. by Ostragradskii in 1845-50 and by Donkin in 1854" (Whittaker [55]). < 2 ) 2. The problem of integration of Hamiltonian systems (not then written in canonical form) had already been discussed in works of the brothers Bernoulli, Clairaut, D' Alembert, Euler and, of course, Lagrange, in connection with the application of the ideas and principles of Newton to various problems of mechanics. Only those problems that could be solved by means (l)T is the kinetic energy of the system.
(Z)"It would be rather desirable to make a detailed critical study of the historical development. In fact, the traditional references to the origin of the fundamental mathematical notions in analytical dynamics are almost always incorrect" (Wintner [54]) . of finitely many algebraic operations and "quadratures", the computation of integrals of known functions, were regarded as "soluble" (integrable). However, most of the actual problems of dynamics (say, the n-body problem) turned out to be "non-integrable" (more precisely, not integrated). Only in the simplest cases when the system had just one degree of freedom (n = 1) or, decomposed into several independent one-dimensional systems, did the integration turn out to be possible, due to the presence of integrals of the type of conservation of the total energy (H = const). 3. Hamilton (in 1834) and Jacobi (in 1837) developed a general method of integrating the equations of dynamics, based on the introduction of special canonical coordinates.
The idea of the Hamilton-Jacobi method appears in the work of Pfaff and Cauchy (and, even earlier, in the investigations of Lagrange and Monge) on the theory of characteristics. The essence of this is the following: a transformation of independent variables p, q ~ P, Q of the form (1) reduces to a search for a "generating" function S(P, q), satisfying the non-linear Hamilton-Jacobi equation which is a particular case of (2). If a problem is solved by the Hamilton-Jacobi method, then the functions P 1 (p, q), ... , Pn(p, q) are first integrals, which, as is easy to verify, are in involution, that is, their Poisson brackets {P;, Pi}=~ ( iJP; iJPj _ iJP; aPi) L.J iJqs i.Jps iJps DCJs s are identically zero. This idea was developed by Bour [ 63] and Liouville [71] in 1855. By means of the Hamilton-Jacobi method they proved that a Hamiltonian equation with n degrees of freedom can be integrated if n independent integrals in involution are known. This is essentially an invariant statement of the Hamilton-Jacobi method. Within the framework of this circle of ideas are works of Jacobi, Liouville, Kovalevskaya, Clebsch, and other authors in which a number of new problems in dynamics, some of which are very non-trivial, were solved. In later works the attention was concentrated on the qualitative investigation of the motion of Hamiltonian systems that can be solved by the Hamilton-Jacobi method, first of all by the method of separation of the variables. In scientific usage the "actionangle" variables, specifically for integrable systems, made their appearance. These "were introduced by Delauney (see [ 66] ) for the discussion of astronomical perturbations. Later, they were found to be admirably suited to the older form of quantum mechanics, for the Bohr-Sommerfeld quantization consisted in making each action-variable an integral multiple of Planck's constant" (Synge [52]). Initially conditions for quantization were stated for systems with separated variables [ 11], but it gradually became clear that in the most general case the compatible levels of a complete set of integrals in involution, in the compact case, are homeomorphic to manydimensional tori, that the motion in them in the corresponding "angle" variables is conditionally periodic, and that the "action" variables are the integrals 2~, ~~ Ji dq over independent cycles, covering the tori in various ways (see, for example, [57], [54]; there are modern accounts in the books [7], [ 16] ). Systems with a complete set of integrals in involution are now called completely integrable. 4. On the other hand, the efforts of Clairaut, Lagrange, Poisson, Laplace, and Gauss, directed towards an approximate solution of applied problems of celestial mechanics, lead ultimately to the creation of perturbation theory. It was proposed to search for solutions of the equations of motion in the form of series in powers of a small parameter (for example, in the solar system such a parameter is the ratio of the mass of Jupiter to the mass of the Sun). Afterwards Delauney, Hilden and Lindstedt modified perturbation theory by using the Hamilton-Jacobi method. Let H = H 0 + €H 1 + E 2 H 2 + ... (€ ~ 1) and suppose that the "unperturbed" problem with the Hamiltonian H 0 is integrable. One then looks for a generating function S in the form of a series S 0 + €S 1 + ... satisfying the equation ( 4) H 0 ( ~! , q) + eH 1 ( ~~ , q) + . . . = K 0 ( P) + eK d P) + ... , where the functions K; are for the present unknown. The functions S 0 and K 0 , by the assumptions, can be found from (4) withe= 0. The Ki and Si, i ~ 1, are found consecutively: the resulting arbitrariness in their definition can be removed by a condition on the absence of so-called "secular" terms.
Thus, the perturbed problem can be regarded as '"solved" if the series of perturbation theory are well-defined and convergent. Their convergence would lead to a number of important consequences (in particular, the eternal stability of the solar system). To anticipate, we mention a disappointing result due to Poincare: in general, because of the presence of the so-called small divisors, the series of perturbation theory diverge. Moreover, the series of an improved perturbation theory proposed by Poincare and Bolinom, in which solutions are expanded in power series in y'e not e, also diverge. We note that if the series of perturbation theory do converge then the equations of motion have a complete set of integrals in involution, which can be expressed as convergent power series in e (or y'e).
Subsequently Whittaker, Cherry, andBirkhoff later (in 1916-1927) obtained similar results for Hamiltonian systems in neighbourhoods of equilibrium positions and periodic trajectories. They showed that, in general, there is a canonical transformation specified by a formal power series, after which the Hamiltonian equations integrate simply. Hamiltonian systems with convergent Birkhoff transformations are sometimes called "integrable in the sense of Birkhoff'. In this case also there is a complete set of independent commuting integrals of special form. 5. As we see, each new generation interprets in its own way the essence of the problem of integration of Hamiltonian systems. However, a common feature of the diverse approaches to this problem is the presence in Hamiltonian systems of independent integrals-"conservation laws". Unfortunately, in a typical situation, integrals not only cannot be found, but do not exist at all, since the trajectories of Hamiltonian systems, generally speaking, do not lie on integral manifolds of a small number of dimensions.
The first rigorous results on non-integrability of Hamiltonian systems are due to Poincare. In [ 4 7] (18 90) he proved the non-existence of analytic integrals that can be represented in the form of convergent power series in a small parameter. Hence, in particular, there follows the divergence of the series of the various versions of perturbation theory. Poincare also mentioned qualitative phenomena in the behaviour of phase trajectories that prevent the appearance of new integrals. Among them are the creation of isolated periodic solutions and the bifurcation of asymptotic surfaces. Poincare applied his general method to varous versions of the n-body problem. It turned out that, apart from the known classical conservation laws, the equations of motion do not have new analytic integrals relative to the masses of the planets. The non-integrability of the n-body problem for fixed values of their masses has not yet been proved_(l) Even earlier, in 1887, Bruns proved the absence of new algebraic integrals in the three-body problem (for all values of the point masses). Afterwards similar results were obtained by Husson (1906) and other authors in the dynamics of a rigid body with a fixed point. We can, however, agree with (l)Here we must make two reservations. Firstly, the investigations of Alekseev on final motions in the three-body problem imply the non-integrability of the restricted threebody problem when two of the masses are equal [1]. Secondly, the question is of integrals on the whole phase space of the problem. A complete set of integrals always exists locally and, consequently, may exist in larger domains, where the motion is not recurrent. Apparently, an example is the domain of positive energy in the many-body problem (a conjecture of Alekseev).
Wintner ( [54], § 129), that these "elegant negative results do not have any dynamical significance" in view of their non-invariance under changes of variables.
The truth is that in practically all integrated problems the first integrals turn out to be either rational functions or simply polynomials. Also, solutions, as functions of complex time, often turn out to be meromorphic. As examples we can quote Jacobi's problem on the motion of a point on a triaxial ellipsoid, Kovalevskaya's spinning top and Clebsch's case of the motion of a rigid body in an ideal fluid. In addition, the investigations of Kovalevskaya and Lyapunov on the classical problem of the rotation of a heavy top showed that the general solution of the equations of motion are single-valued functions of time only when there is an additional polynomial integral. In this connection there arose the interesting problem of the relation between the existence of single-valued holomorphic integrals and branching of solutions in the complex time plane. Its formulation dates back to Painleve.
In 1941-1954 Siegel investigated the question of integrability of Hamiltonian systems close to stable positions of equilibrium. He proved that in a typical situation the Hamiltonian equations do not have a complete set of analytic integrals and the Birkhoff transformation diverges. Siegel's proof ofthe divergence of the Birkhoff transformation dates back in principle to the investigations of Poincare: it is based on a careful analysis of the families of non-degenerate long-periodic solutions.
After the work of Poincare it became clear in the 20-th century that the impossibility of extending local integrals to integrals "in the large" is connected with the complex behaviour of phase trajectories on the level sets of those integrals (not unlike the energy integral), which are known but are not present in sufficient numbers. To put it simply, on an integral level there must exist trajectories that are everywhere dense in some domain in it (see the discussion of these problems, for example, in [52] and [54]). Levi-Civita had proposed to call m-imprimitive systems having m but not m + 1 integrals "in the large". A direct application of the idea of complex behaviour of phase trajectories to the problem of integrability can be found in the above papers of Alekseev.
6. Recently some of the possibilities of the Poincare method have been realized, which make it possible to prove non-integrability of a number of important probl~ms of Hamiltonian mechanics, and also to find new phenomena of a qualitative nature that obstruct integrability. As a result an independent part of the theory of Hamiltonian systems has taken shape. In this paper the author wishes to continue the tradition of a "fairly popular account of the proofs of its basic result", of which Alekseev wrote in the preface to (the Russian translation of) Moser's book [ 41] .
In working on this article the author was helped by numerous conversations with Ya.B. Tatarinov and S.V. Bolotin. In addition, the former read the manuscript and made a number of useful remarks. The author expresses his sincere thanks to them.

HAMil TON IAN SYSTEMS
There are various approaches to an exposition of Hamiltonian mechanics. They can be found in the books [3], [7], [55], and [61]. In this chapter we recall the definitions of the fundamental objects of Hamiltonian mechanics, and also we consider several concrete A solution of it is a smooth map m: .:l ~ M (.:lis an interval on R) such that aF ~; (t)) {F, H} (m (t)) Vt E ~.
These equations can be written in more compact form if we introduce the skew-symmetric matrix where E is the n x n unit matrix. If (x, y) = z, then In local symplectic coordinates the canonical condition for 1.{): x, y ~ X, Y may be expressed by either of the two following equivalent conditions: 1) for each closed contour ' Y In the new coordinates (X, Y) = Z, (1.2) again has Hamiltonian form A symplectic structure on M can be specified by a symplectic atlas: a set of mutually compatible charts, where the transition from chart to chart is a smooth canonical map. For example, let M = T*N be the cotangent bundle of a smooth manifold N. A symplectic structure on T*N is specified by a collection of local coordinates x, y, where x are local coordinates on N and y are the components of linear differential forms from Tx*N in the basis dx.
(l)" ... whenever you have to do with a structure endowed entity L, try to determine its group of automorphisms... You can expect to gain deep insight into the constitution of Lin this way." (Weyl "Symmetry".) It is helpful to study canonical diffeomorphisms by the apparatus of generating functions. For example, let det II ax; ax II =I= 0. In this case we can solve (at least locally) the equation X = X(x, y) for x and regard X and y as "independent" coordinates. Then x = x(X, y), Y = Y(X, y). If we put X, y S= I xdy+ YdX Xo, YO (the value of the integral is independent of the path of integration), then as as

x=ay-· Y=ax·
Then S(X, y) is called a generating function of the canonical map 'P· If, for example, 'P is the identity map, then S = Xy.

5.
A natural mechanical system is a triple (N, T, V), where N is a smooth manifold (the state space), Tis a Riemannian metric on N (the kinetic energy), and V is a smooth function on N (the potential of a force field). The motions of this system are smooth maps q(t): R ~ N that are extremals of the action functional: t2 ,\ L (q (t), q (t)) dt, It where q(t) is the tangent vector toN at q(t), L = T+ Vis the Lagrangian. A time change of the local coordinates q on N is described by the Euler-Lagrange equation:  We consider the total energy of the system, H: T*N ~ R, which is defined A similar construction is valid for the more general "seminatural" systems, when the Lagrangian function contains additional terms that are linear in the velocities.
It is often necessary to consider non-autonomous Hamiltonian systems when the Hamiltonian explicitly depends on time. §2. The motion of a rigid body I. In many problems of mechanics the rotation of a rigid body in threedimensional Euclidean space can be described by equations of the following form: . . (2.1) llf=M Xw+exu, e=eXw, where w = aHjaM, u = aHjae, and H(M, e) is a known function on The vectors wand Mare called the angular velocity and the kinetic momentum of the body. The physical meaning of e and u depend on the concrete statement of the problem. For example, let us consider the rotation of a heavy rigid body with a fixed point. In this case e is a vertical unit vector and u = Er is the product of the weight of the body by the radius vector of the centre of mass. The function H, the total energy, has the following form: where F 1 is a positive definite self-adjoint operator. The equations (2.1) are usually written on the following form: . .

Jw=JwXw-f-eeXr, e=exw.
These are called the Euler-Poisson equations ( [ 3] , [ 14] ). Since J is selfadjoint, in some orthogonal frame ~1 , ~2 , ~3 connected with the body its matrix (also denoted by J) can be brought to diagonal form: The eigendirections of J are called the axes of inertia and the eigenvalues, the numbers J 1 , J 2 , J 3 , the principal moments of inertia of the body. This problem contains six parameters Jt. J 2 , J 3 , and Ertt Er 2 , a 3 (r 9 are the coordinates of the centre of mass relative to the axes of inertia).
In the problem of the motion of a rigid body in an infinite ideal liquid, H is a positive definite quadratic form <AM, 111)/2+ <BM, e>+ <Ce, e)/2.
The vectors e and u are usually called the impulsive force and the impulsive momentum and the equations (2.1) are named after Kirchhoff. The matrices A, B, and Care symmetric: without loss of generality we may assume that A = diag(a~> a 2 , a 3 ). Thus, in the general case the quadratic form H contains 15 parameters. If the rigid body has three mutually perpendicular planes of symmetry (say, a triaxial ellipsoid), then B = 0 and C = diag(c 1 , c 2 , c3).
2. The equations (2.1) have three integrals: Fj = H, F 2 = <M. e>, and F 3 = ( e, e >. In the problem of the rotation of a rigid body around a fixed point F 3 = l, obviously. The integral levels / 23 = {F 2 = f 2 , F 3 = f 3 > 0} c R 6 are diffeomorphic to the (co)tangent bundle of the two-dimensional sphere.
We define in R 6 {Af, e} the bracket {, } by putting Taking the operation {, } to be bilinear, skew-symmetric and satisfying Leibniz' rule we can compute the "Poisson bracket" of any two smooth functions on R 6 by using (2.2). The bracket (2.2) satisfies the Jacobi identity. The equations (2.1) can be expressed in the following Hamiltonian form: . This is well-defined (independent of the method of extension) and the bracket {, }* is non-degenerate and gives a symplectic structure on / 23 .   (2.2) holds for the Poisson brackets of M;. ei. The vectors e, p, and M have a simple interpretation: e is the radius vector of a point in threedimensional space, p is its momentum and M is its kinetic momentum (taken with the opposite sign). We emphasize that the coordinates (e~> e 2 , e 3 ) = e are "surplus". When f 2 =I= 0, the change of variables M = p x e must be somewhat "rectified".
As required.
3. On 1 23 we introduce special canonical coordinates L, G; I, g mod 21T (Fig.l ), which are convenient in what follows. For simplicity we restrict ourselves to the case whenf 2 = 0. In R 3 {e} we consider the sphere ( e, e) = f 3 > 0. We introduce the node line, the intersection of the planes passing through e = 0 and perpendicular to the vectors M and ~3 . Let I and g be the angles between ~1 and ~Y and between ~Y and e (~y is the "direction" vector of the node line).
We put, finally, L = (M, ~3 ) and G = IMI. The Hamiltonian K:/ 23 ~ R can be expressed as a function of L, G, I, and g that is 27T-periodic in I and g.
Theorem 3 [ 62]. The functions L, G, l, and g I 1 satisfy the canonical equations We omit the proof of this theorem, which is based on simple formulae of vector analysis.
Let e = ~ eiSi· Then etf"Vfa =cos l cos g-~ sin l sing, e 2 /Yh =sin l cos g + ~ cos l sing, When [ 2 =I= 0, this formula becomes somewhat complicated (details can be found in [ 32] ). 4. The case when the total energy reduces to a quadratic form (M, F 1 M)/2 is called the Euler problem. It is realized, for example, in the rotation of a heavy rigid body around a fixed point, when the centre of mass coincides with the point of suspension. Let w 1 , w 2 , w 3 be the projections of the angular velocity w onto the eigendirections of J. Then Consequently, H=y (Jro, ro}= Let g be the acceleration of free fall. Then the potential energy of the pendulum is The Lagrange equation has the following form: The state space is the circle S 1 { x mod 2n}, and the phase space is the cylinder S 1 X R{p}. For € = 0 we have an integrable problem with one degree of freedom (a mathematical pendulum of constant length).
2. In many problems of mechanics there occur equations resembling (3.1). Let us consider, for example, the planar oscillations of a satellite in an elliptical orbit. The equation of oscillations can be expressed in the following form: Here e is the eccentricity of the orbit and Jl is a parameter characterizing the mass distribution of the satellite. The meaning of the variables o and v is clear from Fig. 3. that the angular velocity of rotation, the sum of the masses of ,ff and ' ¥' , and also the gravitational constant are 1. It is easy to see that then the distance /1'1f is also 1.
The equations of motion of an asteroid .J1; in a moving system of coordinates can be described in the form of two equations · · · av · · av x-2y= ax' y+2x= ay' where V= (x 2 +y 2 )/2+(1-JJ.)/p 1 +JJ./p 2 , Jl is the mass of Jupiter, and p 1 and p 2 are the distances from .Jt to t!f' and '¥· The equations ( 4.1) have the integral . .
the so-called Jacobi integral. These equations can be expressed in canonical form: the Hamiltonian function His the total energy of the asteroid.  [37]. §5. Some problems of mathematical physics 1. From hydromechanics it is known [36] that the motion of n point (cylindrical) vortices in the plane (in space) can be described by the following system of 2n differential equations: ·''*" Here (x 8 ,Ys) are the Cartesian coordinates of the s-th vortex with intensity rs. It is assumed that all the rs are non-zero. The equations (5. In addition to the Hamiltonian H they have another three independent integrals: P "' r P " 1 r If the sum of the intensities of the system of vortices is zero, then Px and Py commute.
2. Kontopoulos in his paper [ 64] on galactic models considered some Hamiltonian systems in neighbourhoods of positions of equilibrium that admit resonance relations between frequences. The simplest such system with the Hamiltonian =~ --y Y1 ,-Y2 xl T x2 -:--"'x1x2-3 xz was investigated in detail by Henon and Heiles by means of numerical calculations [ 69]. In this problem the frequences of small oscillations are equal to each other. In Gustavson's paper [68] there is an interesting discussion of the numerical results of Henon-Heiles in connection with the construction of formal integrals of Hamiltonian systems.
3. The study of homogeneous two-component models of the Yang-Mills equations is connected with the investigation of the Hamiltonian system with the Hamiltonian (see [16], [17]).

INTEGRATION OF HAMILTONIAN SYSTEMS
Differential equations, including Hamiltonian equations, are usually divided into the integrable and the non-integrable. "When, however, one attempts to formulate a precise definition of integrability, many possibilities appear, each with a certain intrinsic theoretic interest."(l) In this chapter we give a brief list of the various approaches to integrability of Hamiltonian systems, "not forgetting the dictum of Poincare, that a system of differential equations is only more or less integrable".< 1 > (l)D. Birkhoff "Dynamical systems". § 1. Ouadratures 1. Integration by quadratures is the search for solutions by "algebraic" operations (including the inverting of functions) and "quadratures", the calculation of the integrals of known functions. This definition of integrability formally has a local character. The solution by quadratures of a differential equation on a manifold means its integration in any local coordinates. We assume that the transition from one system of local coordinates to another is an "algebraic" operation. The following result connects the integration by quadratures of Hamiltonian systems with the existence of a sufficiently large set of first integrals.   [ 30] .
Corollary. If a Hamiltonian system with n degrees of freedom has n independent integrals in involution (the algebra 2! is commutative), then it can be integrated by quadratures.
This result was first proved by Bour for automonous canonical equations [ 63] and later was generalised by Liouville to the non-autonomous case [ 71 ] . Suppose that H and F~o ... , Fn do not depend on time. Then H is also a first integral, for example, H = F 1 . The theorem on integrability by quadratures still holds, of course, in that case (the condition {H, F;} = 0 can be replaced by the weaker condition {H, F;} = A.;H, }..; = const; The proof of Theorem 1 is based on a lemma due to Lie.  [58], [ 60]).
We prove this result in the very simple case n = 2. In the general case the proof is similar.
The equation .X = X 1 (x), x E U, can be integrated if we can find a first integral F(x) such that F'(x) =F 0 in U. We remark that by the straighteningout theorem such a function obviously exists (at least locally). If XIF = 0, then X 2 F is again an integral, since XI( Obviously, It remains to observe that the functions F 1 , ..• , Fn and K are independent.
2. As a simple example we consider the problem on the motion on a line of three points with an attracting force inversely proportional to the cube of the distances between them. Let m 1 be the masses, x 1 the coordinates, and p; = m 1 x 1 the moment of the points. The potential energy of interactions is The functions F 1 = L]pf/2rn 1 + U, F 2 = 1Jpix; and F 3 = lJp; are independent and {F 1 , Fa}= 0, {F 2 , Fa}=-F 3 , {F1, F 2 } = 2F\. Since the corresponding Lie algebra ~ is soluble, the motions on the zero levels of the total energy and the momentum can be found by quadratures. This possibility is not hard to realize directly. We note that in the case of equal masses m; and coefficients a;i (i < j) we can find a complete set of integrals in involution.
3. Let M be a symplectic manifold and F 1 , ... , Fn independent functions on M generating a finite-dimensional subalgebra of the Lie algebra Coo(M) (that is, {F;, Fi} = ~c7iFk, c7i = canst). At each point x EM the vectors LJA-;3Fi, A; E R, form an n-dimensionallinear subspace ll(X) of TxM. The distribution of the planes II( Consequently, by Frobenius' theorem, through each point x EM there passes a maximal integral manifold Nx of fl. The manifolds Nx can be embedded in M in a very complicated way; in particular, they need not be closed. If n = dim M/2, then among the integral manifolds of IT there are In the special case when F 1 , •.. , Fn commute pairwise M is foliated into the closed manifolds M 1 . §2. Complete integrability 2) on Rk x rn-k there are coordinates y" ... , Yk> 1{) 1 , ..• , IPn-k mod 271' such that in these coordinates the Hamiltonian equation .X = 'JFi takes the following form: The proof of this theorem is by now too well known for us to repeat it here (see [ 71 , [ 161 ). Hamiltonian systems with each of the Hamiltonian functions F 1 , •.. , Fn are called completely integrable.
Small neighbourhoods of invariant tori M 1 "' yn in M are diffeomorphic to the direct product D x yn, where D is a small domain in Rn. It turns out that in D x yn one can always introduce symplectic coordinates I, I{) (IE D, I{) E Tn) such that in these variables the Hamiltonian function of a completely integrable system depends only on I (see [7]). Here · aH · aH I =aq' =0, cp = ar=(J) (/).
Consequently, I= I 0 , w(l) = w(/ 0 ) = const. The variables I, which "enumerate" the invariant tori in D X rn' are called "action" variables, and the uniformly changing coordinates '{) "angle" variables. The Hamiltonian system is called non-degenerate (in D x Tn) if In this case almost all invariant tori (in the sense of Lebesgue measure) are non-resonant, while the resonant tori are everywhere dense in D X rn.
The system is called properly degenerate if The reason for degeneracy may be that the number of first integrals on the whole phase space is greater than n (but, of course, not all of them in involution). Such is the case, for example, in Kepler's and in Euler's problem. This situation is described by generalizations of Liouville's theorem. We denote by F 1 , ... , Fn+ k the independent first integrals of a system with Hamiltonian Hand, as before, let M 1 = {F; = /;}. We assume M 1 to be connected and compact.
Theorem on generalized action-angle variables (Nekhoroshev [ 43] In the paper [39] by Mishchenko and Fomenko, where this theorem is proved and applied, there is also the conjecture that the assumption on the algebra of integrals being finite-dimensional can be removed. In fact, shortly afterwards, Strel'tsov generalized the preceeding two results and showed that if {F;, Fi} = f;i(F 1 , ... , Fn + k) and the rank of II {F;, Fi} II is 2k, then in a neighbourhood of Mr there are first integrals G; satisfying Nekhoroshev's generalization. This result was announced in [ 40] . As noted by Tatarinov (unpublished), all of these generalizations of Liouville's theorem fall under the following observation: part of the integrals (2k of them) cut out canonical submanifolds in M of dimension 2(n-k); in each of these a proper Poisson bracket can be specified, for example, by Dirac's formula [ 15] ; then the restrictions of the remaining (n-k) integrals on these submanifolds satisfy the usual Liouville theorem. §3. Examples of completely integrable systems 1. The equations of rotation of a heavy rigid body around a fixed point are Hamiltonian in the integral manifolds / 23 = { F 2 = / 2 , F 3 = 1 }. One integral always exists: the energy integral. Thus, for the complete integrability of the equations on / 23 it is sufficient to know one other independent integral. We list the known cases of integrability. As we have already noted the problem of a heavy top contains 6 parameters: the three eigenvalues of the inertia operator, 1 1 , 1 2 , 1 3 , and the three coordinates of the centre of mass relative to its eigenaxes rv r 2 , r 3 • 1) Euler's case (1750): r 1 where •= er;J 3 , r 2 = r~ + r:. 4) Goryachev-Chaplygin's case (1900): 1 1 = 1 2 = 41 3 , r 3 = 0 and / 2 = <M, e > = 0. In contrast to 1)-3) here we have an integrable case on a single integral level / 23 . We note that all these integrable cases form manifolds in the sixdimensional parameter space 1;, r; of one and the same dimension 3.
2. The equations of motion in the first two cases have been studied in detail from various points of view in the classical works of Euler, Poinsot, Lagrange, Poisson, and Jacobi. The Kovalevskaya case is non-trivial in many ways. She found it from the condition for meromorphicity of the solutions of the Euler-Lagrange equations in the complex time plane. Recently, Perelomov obtained the Kovalevskaya integral by means of a representation of Lax [ 73] . The Goryachev-Chaplygin case is somewhat simpler: it can be integrated by separation of the variables. Let us show this.
In the special canonical coordinates L, G, l, and g the Hamiltonian function has the following form: We consider the canonical transformation In the new symplectic coordinates p and q 3 3 . .
Putting this expression equal to h and multiplying by p 1 -p 2 we see that it separates: We put Here r is a first integral of the equations of motion. In the special canonical variables it has the following form: 3 and in the traditional Euler-Poisson variables w, e We write down a closed system of equations for the change of variables P1• Pz: or, taking account of (3.1 ), We introduce angle variables 'Pt> l{)z mod 2n by the formulae (3.3) In the new variables (3.2) takes the following form: where p;(z) are the real hyperelliptic functions of period 2n, defined by (3.3).
Since the trajectories of (3.4) on T 2 {<p mod 2n} are straight lines, the ratio of the frequencies of the corresponding conditionally-periodic motions is r tfr 2 , the ratio of the periods of the hyperelliptic integral z (' dz This remarkable fact holds even for the equations of Kovalevskaya's problem. Details can be found in [ 32]. 3. The problem of the motion of a rigid body in an ideal fluid is much richer in integrable cases (see [53]). We mention two of them: they were discovered by Clebsch (1871) and Steklov (1893). In Clebsch's case it is assumed that B = 0, C = diag(cv c 2 , c 3 ) and a~' (c 2 -c 3 ) + a; 1 (c 3 -c 1 ) + a; 1 (c 1 -c 2 ) = 0.

An additional integral of the Kirchhoff equations has the form
An additional integral is We consider in detail the special case when the sum of the intensities rs is zero. Then the integrals Px and Py are in involution. If their constants are zero, then the equations of motion of four vortices turn out to be Liouville integrable. The idea of the solution is based on the application of a suitable canonical transformation, which is standard in celestial mechanics in connection with the "exceptional" motions of the centre of mass in the n-body problem. To be definite let, r 1 = r 2 = -r 3 = -r 4 = -1. We consider the linear canonical transformation x, y -+ ex, ~ given by XI= -~4> In the new coordinates Px = ex 2 , Py = ex 4 . Consequently, the Hamiltonian function H does not depend on the conjugate variables ~2 and ~4. Thus, the number of degrees of freedom is reduced by 2: we have obtained a family of Hamiltonian systems with two degrees of freedom depending on the two parameters ex 2 and ex 4 • The variables ex., ex 3 , ~I> ~3 are symplectic coordinates. When ex 2 = ex 4 = 0, M is an integral of the "reduced" system. Consequently, this Hamiltonian system with two degrees of freedom is completely integrable.
In particular, the functions a 1 , a 3 , ~v /3 3 It can be found by quadratures.
The remaining "cyclic" coordinates, {3 2 and /3 4 , in view of the formulae can be found by a simple integration. As far as the author knows, this possibility has not been realized.

5.
Other interesting examples of completely integrable systems can be . found, for example, in Moser's paper [ 42] . In the same place some modern methods of integration of Hamilton's equations are discussed.
According to Poincare, the investigation of the complete system for small values of E is a basic problem of dynamics [ 48] . The idea of classical perturbation theory consists in the following: to find a canonical transformation /, I{) ~ J, 1/1, depending analytically on E, such that where K 1 (J) is, for the present, unknown. We expand the "perturbing" function H 1 in a multiple Fourier series: mEZn If ( 4. 2) has a solution that is periodic in 1{), then

Let
Then Tn Hm (J) In the subsequent analysis a major role is played by the secular set Hm(l) =I= 0 belong to lB. By Bessel's inequality, .2j s~ < oo mEZn and the generating function sl is not defined on the set lB X Tn c D X Tn.
In essence the secular set consists of those tori of the unperturbed integrable problem that split under a perturbation of order e. In a typical situation m is everywhere dense in D and this is connected with a wellknown difficulty, the phenomena of "small divisors", which obstruct not only convergence, but even the formal construction of a number of the classical schemes in perturbation theory.
3. Theorem 1. Suppose that ( 4.1) has n first analytic integrals If the equations ( 4.1) have integrals that are formally analytic in e (power series in e with analytic coefficients in D x Tn) and satisfy the conditions of the theorem, then we can construct (at least formally) the series of perturbation theory defined for (J, I{)) E D X Tn. Let us prove this.
Let F 8 (l, (j), e) = fs(I) + ~ e"Fsk(J, (J)). We consider the system of equations If we put then from (4.5) we obtain a periodic solution S 1 • When k ~ 1 we have for the definition of Sk and fsk an equation of the form ( 4.5) whose right-hand side contains the known functions Sm and fsm (m < k).
In the new canonical coordinates J, 1/J the functions Fh ... , Fn depend only on J and €. Since these functions are first integrals of the Hamiltonian system (4.1) and are independent, the same is true forJ 1  Of special interest is the case when the eigenvalues of the linearized system ; = 'JH~ are purely imaginary and distinct. It is well-known [ 18], that then there is a linear canonical transformation of coordinates p, q ~ x, y that takes the quadratic form H 2 to ( r 1) 1 ""

H(x, y) into a Hamiltonian K(p), a formal power series in Ps
If the series .:6Sm converges, then the equations with Hamiltonian H are completely integrated: Pt> ... , Pn are power series in x andy that form a complete set of independent integrals in involution. The converse is also true.
Normalization of a Hamiltonian system in a neighbourhood of a stable position of equilibrium is closely connected with the classical scheme of perturbation theory. For by introducing a small parameter € by x ~ Ex, y ~ EY and passing to polar coordinates I, c.p by the formulae  If the series lJGmk are formal (not necessarily convergent), then we can find a formal canonical transformation "normalizing" the Hamiltonian H. In particular, under the conditions of the theorem, the Birkhoff transformation exists also for rationally dependent sets of frequencies The transformation to normal form can be carried out not only in neighbourhoods of positions of equilibrium, but also, for example, in neighbourhoods of periodic trajectories. All that has been said above remains valid with necessary changes in that case.

TOPOLOGICAL OBSTRUCTIONS TO COMPLETE INTEGRABILITY OF NATURAL
SYSTEMS § 1. The topology of the state space of an integrable system 1. We consider a mechanical system with two degrees of freedom (see Ch.I, § 1 ). We assume that its state space M is a compact orientable analytic surface. The topological structure of such surfaces is well known: they are spheres with a certain number x of handles attached. The number x is a topological invariant of the surface, it is called its genus. The motions of a natural system are described by the Hamiltonian equations in the cotangent bundle T*M, which is its phase space. The bundle T*M has a natural structure as a four-dimensional analytic manifold.
We assume that the Hamiltonian function H: T*M -+ R is everywhere analytic. Since H = T(p, q)+ U(q) and T(p, q) is a quadratic form in p E r;M for all q E M, the functions T(p, q) (kinetic energy) and U(q) {potential energy) are analytic on T*M and M, respectively. The solutions of the canonical system Numerous examples are know of integrable systems whose configuration spaces are homeomorphic to S 2 or T 2 (say, the motion of an inertial material particle on a "standard" sphere or torus).
In the infinitely differentiable case Theorem 1, generally speaking, is not valid: for any smooth surface M one can give a "natural" Hamiltonian H == T+ U such that Hamilton's equations (1.1) on T*M have an additional infinitely differentiable integral independent of (more precisely, not everywhere dependent on) H. For let us consider the standard sphere ~ in R 3 and suppose that M is obtained from S 2 by attaching any number of handles to some small domain N on S 2 • Let H be the Hamiltonian function for the problem of the motion of an inertial particle (U = 0) onM, embedded in R 3 . Outside N the particle obviously moves along great circles of S 2 . Consequently, in the phase space T*M there is an invariant domain that is diffeomorphic to the direct product D x ~ foliated into twodimensional invariant tori. The points of D "enumerate" these tori. Let f:D ~ R be a smooth function that vanishes outside some subdomain G lying wholly within D. Corresponding to f there is a smooth function F on D x ~ that is constant on the invariant tori of D x ~. It extends to a smooth function on the whole of T*M if we put F == 0 outside G x ~. Obviously, F is a first integral of the canonical equations ( 1.1) and the functions H and F (for suitable f) are not everywhere dependent.
2. Theorem l is a consequence of a stronger result establishing the nonintegrability of the equations of motion for fixed sufficiently large values of the total energy. The precise statement is as follows. For all values h > maxM U the level of total energy I, = {x E T* M: T + U = h} is a three-dimensional analytic manifold having the natural structure of a fibre space with base M and fibre S 1 • Local coordinates on Ih are q, '-{), where q are coordinates on M and '-P is the angular variable on the "fibre" (!)Analytic functions are called independent if they are independent at some point (they are then independent almost everywhere). S~ = {P E T;M : T(p, q) + U(q) = h}, which is a circle in the cotangent plane. Since the initial Hamiltonian vectorfield 'JH' is tangent to Ih, on Ih there arises a certain analytic system of differential equations. In the analytic case the conditions a) and (3) are automatically satisfied. Here condition (3), obviously, holds for all q EM. But a) is non-trivial: a proof can be found in [ 7 5] .
More generally, if a compact orientable smooth surface M is not homeomorphic to the sphere or the torus, then the equations of motion do not have a new integral F(p, q) that is an infinitely differentiable function on T*M, is analytic for fixed q E M on the cotangent plane Tq*M, and has finitely many distinct critical values. For the standard regular double covering N-+ M, where N is an orientable surface, induces a certain mechanical system on N, which has an additional integral if the system on M has a new integral. It remains to remark that when M is not homeomorphic to RP 2 or K, then the genus of N is greater than 1. §2. Proof of the theorem on non-integrability 1. According to the Maupertuis principle of least action, the trajectories of the motions of a mechanical system that lie on integral level surfaces Ih with total energy h > maxMU are geodesic lines of the Riemannian space (M, ds), where the metric ds is defined by the form (ds) 2 We fix a point q E M satisfying {3). Since (M, ds) is a smooth twodimensional compact orientable Riemannian manifold and not homeomorphic to the sphere, by a theorem of Gaidukov [ 121, for any non-trivial class of freely homotopic paths in M there are geodesic semitrajectories r emanating from and approaching asymptotically some closed geodesic from the given homotopy class. The geodesic r itself may be a closed curve. In what follows, the geodesic semitrajectory r is called a rq-geodesic.
Suppose that the reduced system has on Ih an infinitely differentiable first integral F(q, 4p). Any of its non-critical levels is a union of a certain number of two-dimensional invariant tori. In the cotangent plane Tq*M we consider the circle S~ consisting of vectors p such that T( p, q) + U(q) = h.
To each p E S~ there corresponds a unique motion q(t), p(t) with the initial Since M is not homeomorphic to a sphere or a torus, 2x > 4., and from dimension arguments it follows that H 1 (M) cannot be covered by finitely many one-dimensional and two-dimensional subgroups. This contradiction proves that the collection of critical momenta is infinite.
According to ~) the number of distinct critical values of the function F:Ih -+ R is finite. Consequently, for the value of q EM fixed above the function F(q, '()), '() E S~, takes the same value infinitely often. But then, by {3), F(q, '()) is constant on S~ (that is, does not depend on '{)). The surface M is connected and compact, hence, any two of its points can be joined by a minimal geodesic [ 38]. Since F is constant along each motion, it takes the same value at all points q E M satisfying {3). Since, by assumption, the set of such points is everywhere dense in M, F = const by continuity.
This proves the theorem.
2. Another proof of Theorem 1 based on the introduction of a complexanalytic structure on M, is in the paper [ 34] by Kolokol'tsov. There is also a description of the two-dimensional systems with first integrals that are quadratic in the velocity. curvature is negative. If the curvature is negative everywhere, then the dynamical system on Ih is a Y-system, consequently, is ergodic on Ih [2]. This result holds also in the many-dimensional case (we need only require that the curvature is negative in all two-dimensional directions). Here the differential equations of motion on Ih do not even have continuous integrals, since almost all trajectories are everywhere dense on /h. Of course, a curvature that is negative in the mean is by no means always negative everywhere. It would be of interest to study the connection between complete integrability of a natural system and the geometric properties of the Riemannian space (M, ds) (not only with the coarser topologies).

NON-INTEGRABILITY OF NEARLY INTEGRABLE HAMILTONIAN SYSTEMS
In this chapter we investigate the integrability of the "fundamental problem" of dynamics: 1=i!f!l, cp=TI; H=H 0 (l)+eH 1  From the first equation it follows that F 0 is an integral of the unperturbed equation with Hamilton function H 0 • Suppose that the torus I = I* is nonresonant. Then F 0 (/*, '()) does not depend on '(), since any trajectory fills out a non-resonant torus densely. To complete the proof of 1) it remains to take into account that F 0 is continuous and the set of non-resonant tori of a non-degenerate integrable system is everywhere dense.
Let <l>o, <1>1 D X Tn ~ R be continuously differentiable functions, <Do not depending on '(). Then are analytic in U and i!!3 n U is a key set, the functions H 0 and P 0 are dependent throughout U, consequently, in the new coordinates P 0 = P 0 (H 0 ).
Since P-P 0 (H) = E<!>, we see that <P is a formal integral of (1 ) The result is simple to prove by the method of § 3.1.
3. We now consider the non-autonomous canonical system of equations The Hamilton function H is assumed to be analytic and 27T-periodic in ..p and t.
The equations (1.2) arise, for example, in the study of the autonomous system ( 1) when one of the angle coordinates ..p is taken as the new time.  If the Poincare set !ill~, is everywhere dense in D, then the equations (1.2) obviously have no formal integral with continuously differentiable coefficients.
It is interesting to note that for n = 1 a theorem of Kolmogorov on the preservation of conditionally periodic motions [ 4] has the consequence that there exists a first integral, analytic in e, with non-constant continuous coefficients. By way of contrast, in the many-dimensional case, for systems of general form, even a continuous integral seems impossible (see [ 61 ). § 2. The creation of isolated periodic solutions-an obstruction to integrability 1. We recall some facts from the theory of periodic solutions of differential equations. We consider an autonomous system :X= f(x); let x(t, y) be the solution of it with the initial value x(O, y) = y. We assume that the system has an w-periodic solution x(t, x 0 ). Then X (t) =II ;; lx, I I is the fundamental matrix of the linear system in variation • at s = Tx (x (t, x 0 }) S• Obviously, X(O) =E. Here X(w) is. called the monodromy matrix for the w-periodic solution x(t, x 0 ). Its eigenvalues A are called multipliers, and the numbers a: defined by A = exp(a:w) are called characteristic exponents. The multipliers A may be complex, therefore, the characteristic numbers a: are not uniquely determined. Since (X(w)-E)f(x 0 ) = 0 andf(x 0 ) =I= 0, in the autonomous case one of the multipliers A is always equal t? 1. By the theorem of Poincare-Lyapunov [ 7], the characteristic exponents of an autonomous Hamiltonian system are pairwise equal in size and opposite in sign. Two of them are always zero. In the case of two degrees of freedom the remaining two characteristic exponents are either both real or both purely imaginary. If they are non-zero, then the periodic solution is called non-degenerate or isolated: on the corresponding three-dimensional energy level, in a small neighbourhood of the periodic trajectory, there are no other periodic solutions with period close to {J.). A non-degenerate solution with real exponents is called hyperbolic, and with purely imaginary exponents, elliptic. A hyperbolic periodic solution is unstable, and an elliptic solution is stable in a first approximation.
We assume that the Hamiltonian system with two degrees of freedomi~ ='JH' has, in addition to H(z), an integral F(z).

Theorem 1 (Poincare). If~ is on the trajectory of a non-degenerate periodic solution, then the functions H(z) and
au (X (w)-E) -or-= 0.
Poincare's theorem gives us a method of proving non-integrability: if the trajectories of non-degenerate periodic solutions densely fill out the phase space, or at least form a key set, then the Hamiltonian system has no additional analytic integral. Apparently, in Hamiltonian systems in general position the periodic trajectories are, in fact, everywhere dense (Poincare [ 48 ]). This is still unproved. In the context of Poincare's conjecture we mention the following result on geodesic flows on Riemannian manifolds of negative curvature: all periodic solutions are of hyperbolic type and the set of their trajectories densely fills out the phase space [ 2] .
For Hamiltonian systems close to integrable ones, one can prove the existence of a large number of non-degenerate periodic solutions and from this derive the results of § 1.
2. Suppose that for I = JO the frequencies w 1 and w 2 of the unperturbed integrable problem are commensurable and that w 1 =I= 0. Then the perturbing function H 1 (J0, w 1 t, w 2 t +X) is periodic in t with some period T.
We consider its mean value 2) for some X= X* the derivative an.;ax = 0 but a 2 HtfaX 2 =I= 0. wl ai~ -w1wz a I 1 ai 2 + w2 ali · A proof can be found in [ 48], [32]. The function H 1 (/ 0 , X) is periodic in X with period 21T. Hence, there exist at least two values of X for which dH 1 = 0. In general, these critical points are non-degenerate. There are as many local minima (where a 2 H 1 /aX 2 > 0) as local maxima (where a 2 H 1 /aX 2 < 0). In a typical situation for I= ! 0 values of e =I= 0 the perturbed system has as many periodic solutions of elliptic type as of hyperbolic type. In this situation it is usual to say that the disintegration of the unperturbed invariant torus I = I 0 creates pairs of isolated periodic solutions. By results of KAM-theory, the trajectories of typical elliptic periodic solutions "surround" invariant tori. Hyperbolic periodic solutions have two invariant surfaces (separatrices) filled out by solutions that approximate asymptotically to periodic trajectories as t ~ ±co.
Various asymptotic surfaces may intersect, forming a rather tangled network in the intersection (see Fig. 5). The behaviour of the asymptotic surfaces will be discussed in detail in the next chapter.  (1, <p) is equal to 1 at points (/, ..p) E f(O). In particular, at these points To complete the proof it remains to note that the functions H 0 and F 0 do not depend on 1().
We mention that always-~ c \ill, however, in typical cases the sets ~ and m3 coincide. In addition, in the above arguments the integral F was assumed to be analytic in e, but in § 1 it was proved that there are no integrals formally-analytic in e. However, our aim was to clarify the geometry of the analytic computations of § 1.
For small values of the parameter e =F 0 Theorem 2 guarantees the existence of a large but finite number of distinct isolated periodic solutions. Therefore, from this theorem one cannot deduce the non-integrability of perturbed systems for fixed values of e =F 0. True, in the case of two degrees of freedom, which is what we are considering, the following result holds: if the unperturbed system is non-degenerate, then for small fixed values of e =F 0 the perturbed Hamiltonian system has infinitely many distinct periodic trajectories. Unfortunately, nothing can be said about their isolation. This result can be deduced from Kolmogorov's theorem on the preservation of conditionally periodic motions and Poincare's last geometric theorem [ 50] . § 3. Applications of Poincare's method 1. We return to the restricted three-body problem considered in Ch. I, § 4. We assume to begin with that the mass of Jupiter J..l is zero. Then in the "fixed" space an asteroid rotates around a sun of unit mass in Keplerian orbits, say ellipses. Then it is convenient to go over from the rectilinear coordinates to the Delone canonical elements L, G, l, g; if a and e are the major semi-axis and the eccentricity of the orbit, then L = ...ja, G = ...j(a(le 2 )), g is the length of the perihelion and l is the angle defined by the position of the asteroid in its orbit, the eccentric anomaly [ 48] , [59]. It turns out that in the new coordinates the equations of motion of an asteroid are canonical with the Hamiltonian function F 0 = -l/2L 2 . If J..l =F 0, then the complete Hamiltonian F can be expanded in a series of increasing powers of J..L: F = F 0 + J.LFi + ... Since in a moving coordinate system connected with the Sun and Jupiter Keplerian orbits rotate with unit angular velocity, the Hamiltonian function depends on L, G, !, and g-t.
We put x 1 = L, x 2 = G, y 1 = !, Yz = g-t, and H = F-G. Here H depends on X; andY; only, and is 27T-periodic in the angular variables y 1 and y 2 • As a result we have expressed the equations of motion of an asteroid in the form of the following Hamiltonian system: The expansion of the perturbing function in a multiple trigonometric series in y 1 and y 2 was already studied by Lever'e (see, for example, [59]).
It takes the following form: where Ll is any interval on the half-line x > 0.
We note that (3.1) and (3.2) have additional integrals in the form of convergent power series in J.L with continuous (but not differentiable) coefficients.
2. "Let us proceed to another problem: that of the motion of a rigid body around a fixed point ... We can, therefore, ask whether in this problem the presented in this chapter oppose the existence of a single-valued integral other than those of the vis viva and of area" (Poincare [ 48] ). To where H 0 is the kinetic energy (the Hamiltonian function of the integrable Euler problem on the motion of an inertial body), and EH 1 is the potential energy of the body in a homogeneous gravitational force field (€ is the product of the weight of the body by the distance from the centre of mass to the point of suspension). We assume that € is small. This is equivalent to the study of the rapid rotation of a body in a moderate force field. In the unperturbed integrable Euler problem we can introduce actionangle variables I and ..p. The formulae for the transition from the special canonical variables L, G. I, g to the action-angle variables I and '{) can be found, for example, in [32].  The expansion of the perturbing function H 1 in a multiple Fourier series in the angle variables ..p 1 and '{) 2 is, in fact, contained in Jacobi's paper [70] : It follows, in particular, that in this problem the sets m, liD, and ~ coincide.
When J 1 > J 2 > J 3 , the secular set consists of infinitely many lines passing through I= 0 and accumulating at the pair of lines rr 1 and rr 2 • It can be shown that H 0 is non-degenerate in ~. If H were analytic in I throughout ~. then the results of § 1 would be applicable: the points / 0 lying on the lines rr 1 and rr 2 would satisfy the conditions of Theorem 1. The difficulty associated with the analytic singularities of the Hamiltonian function in the action-angle variables can be overcome by considering the problem of an additional analytic integral on the whole integral level set / 23 . Using Poincare's method we can prove the following result. This result gives a negative answer to a question posed by Poincare [ 48] .  H(y, x, t) on A"+ 1 ) is zero. Lagrangian surfaces are invariant under the action of the phase flow of the system ( 1.1) [ 16] . In the autonomous case Lagrangian surfaces A" c T*V are given by the condition If a Lagrangian surface A"+ 1 has a one-to-one projection onto D x R{t}, D c V, then it can be represented as a graph _ as (x, t) H where S: D x R -+ R is a smooth function. In the autonomous case A" is given by the graph xED.

at ax
In this section we are concerned with Lagrangian surfaces consisting of asymptotic trajectories. Naturally, such surfaces are called asymptotic.
2. We assume that the Hamiltonian function is 21r-periodic in t and depends on a further parameter e :H = H(y, x, t, e). Suppose that H(y, x, t, 0) = H 0 (y, x) fore = 0 does not contain the time and satisfies the following conditions: 1) there exist two critical points y_, x_ andY+, x+ of H 0 (y, x) at which the eigenvalues of the linearized Hamiltonian system 3) There is a domain D c V containing x± and such that in T*D c T*V the equation of the surface N = K can be expressed in the following form: It is useful to consider the differential equation In a small neighbourhood of x± its solution tends to x± as t ~ ±oo. 4) In D (1.2) has a doubly-asymptotic solution: x 0 (t) ~ x± as t ~ ±oo ( Fig. 7). The Hamiltonian system with the Hamiltonian function H 0 (y, x) must be regarded as the unperturbed system. In applications it is most frequently completely integrable. Let D+ (or D_) be a subdomain of D containing x+ (or x_) but not x_ (or x+)· For small € the asymptotic surfaces Nand K do not vanish, but go over to the "perturbed" surfaces A~ and A~. More precisely, in D± x R { t} the equation of the asymptotic surface A~ can be written in the following form: as± where S±(x, t, €) is 21r-periodic in t and is defined and analytic for x ED and small e (Poincare [ 48] ). The functions s± must, of course, satisfy the Hamilton-Jacobi equation By hypothesis, forE = 0 the surfaces At and A 0 coincide. However, as Poincare [ 47] first noted, in general, for small E =F 0, regarded as point sets in  (x 0 (t)), x 0 (t), t) dt =I= 0, -oo then for small E =F 0 the perturbed asymptotic surfaces A: and A~ do not coincide [ 47].
Proof We assume that (1.3) has an analytic solution S(s, t, E) that for small E can be expressed as a convergent power series The function S 0 (x, t) must satisfy the equation Clearly, W(x) coincides with the function S 0 (x) by § 1.
Without loss of generality we may assume that H 1 (y±, X:~:, t) = 0 for all t.
If this is not the case, then instead of H 1 we must take H 1 -H 1 (y±, x±, t). The Poisson bracket remains unchanged.
Since the Taylor  Remark. Another proof of Poincare's theorem can be found in [6]. 4. In the autonomous case the condition for bifurcation of asymptotic surfaces situated on a certain fixed energy level can be expressed as follows: This result is easy to derive from the implicit function theorem.
If we investigate the problem of the existence of independent involutive integrals F;(z, t, e), 1 <; i <; n, that are analytic (or formally-analytic) in e, then 2) can be omitted. In particular, if 1) is satisfied, then the series of perturbation theory diverge in a neighbourhood of bifurcated asymptotic surfaces (Poincare [ 4 7] ). In particular, 1) is for n = 1 a sufficient condition for integrability (Siegel [22] ).  2) for small e the perturbed system has a doubly-asymptotic solution t-+ ze(t) close to t-+ z 0 (t), then for small € =I= 0 the Hamiltonian system z = ;'ill' does not have an additional analytic integral [ 65].

Using
Proof We consider the map at a period g of the section t = t 0 into itself.
For small € this map has two fixed hyperbolic points z 1 and z 2 with invariant separatrices W 1 ± and Wf (see Fig. 8). By the conditions of the theorem, for e =I= 0 the separatrices W! and W2 intersect and do not coincide. Fig. 8 Let V be a small neighbourhood of z 1 and Ll a small segment of W2 intersecting Wt For sufficiently large n the segment gn(Ll) lies wholly in V and again intersects Wt By a theorem of Grobman-Hartman [ 44], in V the map g is topologically dual to a linear hyperbolic rotation. Consequently, as n ~ oo the segment gn(Ll) "stretches" along the separatrix Wi and approaches it unboundedly. Obviously, the union 00 (2.1) u g" ( :'!) n=l is a key set for the class of functions that are analytic in the section t = t 0 • Suppose now that the Hamiltonian equation has an analytic integral f(z, t). The function f(z, t 0 ) is invariant under g and constant on W2 (since the sequence gn(z), z E w;-, converges to Zz as n ~ 00 ). Consequently, the analytic function f(z, t 0 ) is constant on the set (2.1) and is therefore constant for any t 0 .
Remark. Poincare divided the doubly-asymptotic solutions into two types: homoclinic (when z+ = z_) and heteroclinic (when z+ =I= z_). If n = 1, then for small e the perturbed problem always has homoclinic solutions (if, of course, it has them for € = 0) [ 47]. § 3. Some applications 1. We consider first the simplest problem of the oscillations of a pendulum with a vibrating point of suspension. The Hamiltonian function His and f(t) is a 2n-periodic function of time. When E = 0, then the upper position of the pendulum is an unstable equilibrium. The unperturbed problem has two families of homoclinic solutions: (3.1) .
The integrals Jn are easily calculated by residues: Consequently, if f(t) =f.= const (that is, fn =f.= 0 for some n =f.= 0), then (1.8) is non-zero on at least one doubly-asymptotic solution of the family (3.1 ).
Thus, if f(t) =f.= const, then by the results of §2 this problem for sufficiently small (but fixed) € =f.= 0 does not have an analytic first integral F( p, x, t) that is 27T-periodic in x and t. Since < M, e ) 2 <;;;;; < M, M >< e, e) and since the functions F 1 , F 2 , F 3 are independent on /123, it follows that a 2 > 0. The stable and unstable asymptotic surfaces of the periodic solutions (3. 2) can be represented as the intersections of the manifold / 123 with the hyperplanes M 1 ..J(a 2 -a 1 ) ± MJ..J(a 3 -a 2 ) = 0. In the Euler problem the asymptotic surfaces are "doubled": they are completely filled out by doubly-asymptotic trajectories, which as t ~ ±oo approximate unboundedly to the periodic trajectories (3.2). The bifurcation of these surfaces was studied in [28], [22]. It turned out that on perturbation the asymptotic surfaces bifurcate always except in the "Hess-Appelrot case": (3.3) r 2 = 0, In this case one pair of separatrices does not bifurcate, and the other does (Fig. 9). The reason for non-bifurcation is that under the condition (3.3) the perturbed problem, for all €, has the "particular" integral • F = Ml'.j(a 2 -a 1 ) ± MJ..J(a 3 -a 2 ) (F = 0 when F = 0). It can be shown that the closed invariant surfaces H = / 1 , F 2 = [ 2 , F 3 = 1, F = 0 , for small € are just a pair of doubled separatrices of the perturbed problem (see [ 28 ]). In the problem of rapid rotation of a heavy asymmetric top the bifurcated separatrices apparently do not always intersect. However, Theorem 2 of § 2 is applicable, and with its help it can be established that there is no Fig. 9 additional analytic integral of the perturbed problem for small, but fixed, e (Siegel [ 22] ).
The behaviour of the solutions of the perturbed problem has been studied numerically in [ 67] . In Fig. 10 the results of the calculations for various values of e are shown. It is fairly clear that the picture of the invariant curves of the unperturbed problem begins to be destroyed exactly in the neighbourhoods of the separatrices.  If B = 0, then the independent analytic integral exists only when C = diag(cv c 2 , c 3 ) and (3.fi) a~1 (c 2 -c 3 ) + a; 1 (c 3 -c 1 ) + a; 1 (ct-C 2 ) = 0.
The matrix B in Steklov's integrable case is defined precisely by the condition (3.5), and (3.6) gives Clebsch's integrable case. It is interesting to note the coincidence that (3.5) and (3.6) are of the same form.
Corollary. In general, the Kirchhoff equations are non-integrable.
The proof of Theorem 1 is based on the phenomenon of bifurcation of the separatrices. We introduce a small parameter e in (3.4), replacing e by ee. On the fixed integral level I 23 we see that where "" JiJk=) MiMte~r.dt.
The integrals ltik satisfy the following linear equations:  a 2 a 3 -a 1 a 2 -a 1 a 3  a 2 a 3 -a 1 a 2 -a 1 a 3 The integral ! 231 can be calculated by means of residues and it can be verified that it is non-zero (Onishchenko). If (3.5) is not satisfied, then J =I= 0 by the obvious equality .! (a 1 a 2 +a 1 a 3 -a 2 a 3 )/2a 1 a 2 a 3 .! 231 = a~1 (b 3 -bzH-a; 1 consequently, the perturbed separatrices are bifurcated. When a 2 a 3 -a 1 a 2 -a 1 a 3 = 0, then J is proportional to ! 123 or ! 132 . By arguments of symmetry and the preservation of the measure on 1 123 generated by the standard measure on R 6 , it follows that the perturbed separatrices intersect. Hence, the Kirchhoff equations are non-integrable on the invariant manifolds 1 123 and, in particular, on the whole phase space R 6 . If (3.5) (or (3.6) forB = 0) does not hold, then one of the pair of separatrices of the Euler problem must bifurcate under perturbation. It is interesting to note that with a suitable choice of B and Cone pair of separatrices remains doubled and the other is bifurcated. For example, suppose that B = 0 and that the elements of the symmetric matrix C satisfy the following conditions: Ct 2 = c23 = 0, Then for all e the Kirchhoff equations have a "Hess-Appelrot particular integral" F = M 1 y'(a 2 -a 1 ) ± Myyi(a 3 -a 2 ). For small e the separatrices of the Euler problem 1 123 n {F = UJ remain separatrices of the perturbed periodic solutions (3.2). 4. By the method of bifurcation of asymptotic surfaces one can establish non-integrability of the problem of the motion of four-point vortices [ 21]. More precisely, we consider this problem in a restricted formulation: a vortex of zero intensity (that is, simply a particle in an ideal fluid) is moving in the "field" of three vortices of unit intensity. It turns out that the equation of motion of the zero vortex can be expressed in Hamiltonian form with a Hamiltonian that is periodic in time: these equations have hyperbolic periodic motions with intersecting separatrices. Therefore, the restricted problem of four vortices is not completely integrable, although (as in the unrestricted formulation) it has four independent integrals. §4. Isolation of the integrable cases l. When a Hamiltonian system depends on a parameter, then in a typical situation the integrable cases correspond to exceptional isolated values of the parameter. The proof of the isolation of the integrable cases in concrete problems may turn out to be a very difficult matter. We investigate this question for the Hamiltonian equation  .
Since p(t) ~ 0 and p(t) =F 0, the multipliers of this solution are positive, one of them being larger than 1, the other smaller than 1 (Lyapunov). Thus, the solution x(t) = 7T is, in fact, hyperbolic. It has two two-dimensional asymptotic surfaces N and K, completely filled out by trajectories that approximate unboundedly to the points x = ±1r as t ~ ±oo. Since the Hamiltonian His analytic, A~ and A~ are regular analytic surfaces in C x T 1 , depending analytically on e.
It turns out that the surfaces A~ and A~ intersect for all e E (-a, a). This result, obviously, is equivalent to the existence of a homoclinic solution x(t) (x(t) ~ ±1r as t ~ ±oo). A proof can be derived, for example, from the following general result. In our case M = Sl, T = x 2 /2, and U = -w 2 (1 + ef)(l +cos x). If -a < e <a, then U(x, t) < U(7T, t) for all 0 ~ x < 27T and all t.
Since the surfaces A~ and A~ do not coincide for small e =I= 0, the values of e, lei E;;a+8(8 > 0), for which A~= A~, are isolated. Since for lei ~a the surfaces A~ and A~ intersect, ( 4.1) is integrable only for isolated values of e. 2. We now give an example of a Hamiltonian system that for everywhere dense sets of values of the parameter is both completely integrable and nonintegrable. Thus, the integrable cases are not always isolated.
ax These equations obviously are integrable by quadratures: We search for a first integral of (4.2) in the form y + g(x, t), where g: T 2 -+ R is an analytic function, which must satisfy the equation the cylindrical cascade T(x, y) = (x+e, y+h(x)) is ergodic [35].
We note that the irrational numbers e satisfying the conditions of Theorem 2 are everywhere dense in R.
Naturally connected with (4.3) is the periodic map We note in conclusion that the equations ( 4.3) were first studied by Poincare in [ 46] .

NON-INTEGRABILITY IN THE NEIGHBOURHOOD OF A POSITION OF EQUILIBRIUM § 1. Siegel's method
We consider a canonical system of differential equations that converge in some neighbourhood of x = y = 0. We introduce the following topology.'/ in ~J: a neighbourhood of a power series H* with coefficients hZs is the set of power series with coefficients hks for which where Eks is an arbitrary sequence of positive numbers. More precisely, Siegel proved the existence of a countably infinite set of analytic independent power series <I>l> <1> 2 , ••• , in infinitely many variables hks, that are absolutely convergent for I hksl < e (for all k, s) and such that if H E ~ is reduced to normal form by a convergent Birkhoff transformation, then at this point almost all <1> 8 (except possibly finitely many) are zero. Since the functions <1> 8 are analytic, their solutions are nowhere dense in.\~.
Consequently, the set of points of.\~ satisfying at least one equation <I>s = 0 is of the first Baire category. If we attempt to investigate the convergence of the Birkhoff transformation in any concrete Hamiltonian system, then we must check infinitely many conditions. There is no known finite method for this, although all the coefficients of the <I>s can be calculated explicitly.
2. Using Siegel's method we can prove the density of non-integrable systems in certain subspaces of Sj. As an example we consider the equation which describes the motion of a material point in a force field with potential U(x ). are everywhere dense.
It seems that the points U E U for which the Birkhoff transformation to normal form converges, form a subset of the first category in U.
3. For simplicity we restrict ourselves to the case of two degrees of freedom (n = 2). Let w 1 = 1 and w 2 = w be irrational.
We consider a canonical equation with Hamiltonian function of the following form: The coefficients hpq may be complex.
Let €pq < 1 be an arbitrary sequence of positive numbers and w an irrational number that can be approximated by rationals sufficiently well: the inequality Bpq (1.5) O<ir-·ffisi<sa, p=(r, 0) q=(O, s) s must have infinitely many solutions in natural numbers r and s. The measure of the set of such numbers is zero, however, they are everywhere dense in R.
Since w is irrational, by Birkhoffs theorem (Ch. II) the canonical equations with the Hamiltonian function (1.4) have a formal integral F (x, y) = x 1 y 1 + ~ fp,p,q,q.x~'•xf·y~·y~·· p+q;;;.3 Lemma 1. In an epq-neighbourhood of each function (1.4) there is a point H such that for the integers r, s in (1.5) the coefficients froos admit the estimate Equating the terms of the 1-th order to zero we arrive at an equation for F 1 : where the right-hand side is some multinomial of degree l whose coefficients can be expressed in terms of the coefficients of the multinomials F 3 , ..• , F1-1 and hpq for p + q < l. For the terms froosX~Y~ of F 1 we obtain the equation Finally, groos can be expressed in terms of the coefficients hpq for p+q < l. Now let rands be natural numbers satisfying (1.5). The coefficients hroos can be changed by not more than Eroos , so that I irhroos -groos I ~ Eroos· Then by (1. 5)  We return to the analysis of the canonical equations (1.3). In this case We make a linear canonical change of variables with complex coefficients: where P; = pf + qr.
If our equations have n integrals E"t .... , Fn then for e = e 0 we obtain then linear equations Since k 0 =F 0, the quadratic forms 1 F~u , .•. , F~n> are dependent for t-= e 0 as required.
Although the proof of the theorem is simple, its use in concrete problems is beset by rather cumbersome calculations associated with the normalization of the Hamiltonians.
2. As a first example we consider the problem of the rotation around a fixed point of a dynamically symmetric rigid body (1 1 = 1 2 ), whose centre of mass lies on the equatorial plane of the ellipsoid of inertia [ 27] . The majority of the integrable cases occur among this kind. The units of measurement of mass and length can be chosen so that 1 1 = 1 2 = I and the parameter e, the product of the weight of the body by the distance from the centre of mass to the point of attachment, is also 1. The natural parameter in this problem is the moment of inertia 1 3 . In all integral manifolds / 23 = <Jw, e > = / 2 , < e, e > = I the reduced Hamiltonian system has two positions of equilibrium; they correspond to the uniform rotations of the body about the vertical axis in which the centre of gravity lies above (below) the point of suspension.
The angular velocity w of such a rotation is connected with the area constant / 2 by the simple relation / 2 = ±1 3 1 w I. Let us consider, to be definite, the case when the centre of mass is below the point of suspension.
In a neighbourhood of this equilibrium position the Hamiltonian function H of the reduced system with two degrees of freedom has the form H 2 + H 4 + ... (terms of degree 3 are missing). The coefficients depend on two parameters x =fl. y = 1?, 1 • It can be shown that the characteristic roots of the secular equation are purely imaginary if y > x/(x + 1). We denote by ~ the subdomain of R 2 {x, y }, where this inequality is satisfied. The ratio of the frequencies is 3 when the parameters x andy are connected by the relation I: 9x 2 -82xy + 9y 2 + 1 j 8x -82y + 9 = 0. This is the equation of a hyperbola: its branches for x > 0, y > 0 lie wholly in 2:.
3. Next we consider the planar circular restricted three-body problem. The equations of motion of an asteroid in a system of coordinates rotating with the Sun and Jupiter can be written in the Hamiltonian form: all all x.=g-. y.=-ax (s=1,2), V(x 1 +ftJ2+x~ V<xl+ft-1J2+x~ This Hamiltonian system has equilibrium positions at the points x 1 == ~-JJ., x 2 == ±y3/2, y 1 == y 2 == 0, which are called the Lagrange solutions or triangular libration points (see Ch. 1). If 0 < 27p(l-JJ.) < 1, then the eigenvalues of the linearized system are purely imaginary and distinct; their ratio is a non-constant function of JJ.. In cases when commensurability of the third and fourth order holds, the coefficients h' t.' <.:. and h'{' 3 have been calculated by Markeev in an investigation of the stability of the triangular libration points [37]. These numbers are non-zero. It would seem that the same is true for all (or almost all) resonance ratios. From the theorem in § 1.1 it follows, in particular, that in a neighbourhood of a libration point there is not even a formal Birkhoff normalizing transformation that is analytic in JJ.. " ... it is so far unknown whether or not the differential equations of the restricted three body problem with fixed mass ratios can be reduced to normal form by a convergent transformation in a neighbourhood of the Lagrange solutions" (Siegel [20] ).

BRANCHING OF SOLUTIONS AND THE ABSENCE OF SINGLE-VALUED INTEGRALS
Let M~ be a complex symplectic analytic manifold (the whole of M is covered by a set of complex charts from C2n{p, q}, where the transition maps from chart to chart are invertible holomorphic canonical transformations). Any complex analytic function H(p, q, t):M 2 n x C ~ C gives a certain complex Hamiltonian system For € =1=-0 the solutions of the "perturbed" equation (    such that the functions «<>b (l ~ s ~ n) are independent.
(l}We again suppose that the formal series F = ~Fieiis an integral of the canonical equations (1.1) if formally (II, F} """' 0. It is easy to see that in this case the composition of the power series (1.2) and SF 1 e 1 is a power series with constant coefficients.
3. Again we consider the problem of a heavy asymmetric rigid body rotating rapidly around a fixed point. Let ..Po = 0 and I 0 E 23 (where 23 is the secular set of the perturbed problem). We consider in the complex plane t E C a closed contour -y, the boundary of a rectangle ABCD (see Fig. 12).  Here T and iT' are, respectively, the real and purely imaginary periods of the elliptic functions [ 8 (1°, w, z), w 1 = aH 0 /ai 1 . The number r is chosen so that these meromorphic functions do not have poles on -y. It can be shown that the function I~(t, / 0 , o) is unbounded along 'Y [ 29] . Consequently, the solutions of the perturbed problem branch in the complex time plane and this situation prevents the appearance of a new single-valued integral.
4. Using the branching of solutions we can establish the absence of singlevalued analytic integrals for small but fixed values of € =I= 0. We quote a result in this direction due to Ziglin [ 23].
Let z = z 0 E C 2 , Im z 0 = 0, be a hyperbolic fixed point of the unperturbed system . z=JH~, dH 0 (z 0 )=0. The eigenvalues ±A of the linearized system have non-zero real parts (Re A> 0). The solution z(t) = z 0 can be regarded as periodic with period 27T. According to Poincare, for sufficiently small I € I the system (1.3) has a 27T-periodic solution z = p(t, €), p(t, 0) = z 0 • Continuing the solutions of (1.3) that are asymptotic to p(t, €) as t -)--oo to functions maximally analytic in t E C (possibly not single-valued), we obtain a two-dimensional complex surface A-;,, which we call the unstable complex asymptotic surface of the hyperbolic periodic solution p(t, €).
We have seen in Ch. VI that the stable and unstable asymptotic surfaces A~ and A-;, may intersect transversally in the real domain, and this leads to the absence of an analytic integral on R 2 x T~ (consequently, on the whole of C 2 x Tb). In this case the complex asymptotic surface A-; (A\), in contrast to the real case, may have transversal self-intersections, which also prevent the existence of a holomorphic integral for (1.3).
We give a sufficient condition for self-intersection. Suppose that the asymptotic solution z = z 0 (t) of the unperturbed system ( lim za(t) = z 0 ) t---+~00 has a single-valued analytic continuation along a closed continuous path 'Y: [ 0, 1] ~ C, 7(0) = 'Y( 1) E R c C. Then for sufficiently small IE I the solution z(t, t 0 , E) of the perturbed system (1.3) with the initial condition z( 7(0) + t 0 , t 0 , E) = z 0 ( 7(0)) also has an analytic (but, in general, not singlevalued) continuation along the "displaced" path ' Y + t 0 . Let h(t 0 , e)= ll 0 (z(-y(1) + t 0 , t 0 , e))-H 0 (za('v(O)))-= Eh 1 (t 0 ) + o(E) be the increment of H 0 (z(t, t 0 , E)) on a circuit oft along 7+t 0 . Theorem 2. If h 1 has a simple zero, then for sufficiently small I El =F 0 the complex surface A-; has a transversal self-intersection, and the system (1.3) has no single-valued analytic first integral in M 3 . We note that h 1 (t 0 ) can be calculated by the formula I 0~1 (z 0 (t), t+t 0 )dt. Let H = < z, A(t)z >/2 be a quadratic form in z E C 2 n, and let A(t) be a given (2 n x 2 n )-matrix whose coefficients are holomorphic functions defined on some Riemann surface X. If, for example, the elements of A(t) are functions meromorphic on C, then X is the complex plane with some points (poles) removed. The linear Hamiltonian equations with the function H have the form Locally, for a given initial condition z(t 0 ) = z 0 , there always exists a uniquely determined holomorphic solution. This can be continued along any curve in X, however, in general, the continuation is no longer a singlevalued function on X. The branching of a solution of (2.1) is described by its monodromy group G: to each element a of the fundamental group 7T 1 (X) there corresponds a (2n x 2n)-matrix Ta such that after a circuit round a closed path of homotopy class a the vhlue of z(t) becomes Taz(t).
If r is another element of the group 1T 1 (X), then Tra = Tr Ta. The correspondence a~ Ta thus defines a group homomorphism 1T 1 (X) ~ G (details can be found, for example, in [ 13], [56]).
A problem of interest to us is the presence of holomorphic integrals F(z, t): C 2 n x X~ C for (2.1 ). Since any integral F(z, t 0 ) is constant on the solutions of (2.1 ), for each t 0 E X the function F(z, t 0 ) is invariant under the action of the monodromy group G. This property imposes severe restrictions on the form of first integrals: if G is sufficiently "rich", then the only invariant functions (integrals) are constants.
Since (2.1) is Hamiltonian, the monodromy transformation group is symplectic. The problem of integrals of groups of symplectic transformations has been studied by Ziglin in [ 24]. We briefly state his results.
For n = 1 this condition means that X is not a root of unity. Let T be the matrix of a non-resonant symplectic map g. Since no eigenvalue of T is 1, the equation Tz = z has the trivial solution z = 0.
It is convenient to go over to a symplectic basis for the map g: if If g is non-resonant, then s is even and fk 1 = 0 for k =I= /. Theorem 1. Let g E G be non-resonant. If the Hamiltonian system has n independent holomorphic integrals F(z, t) : C 2 n X X ~ C, then any transformation g' E G has the same fixed points as g and takes the eigendirections of g into eigendirections. If no k ~ 2 eigenvalues of g' form a regular polygon in the complex plane with centre at zero, g' commutes with g [24].
The latter condition is necessarily satisfied if g' is also non-resonant. We now prove Theorem 1 for the simple case n = 1, which is important for applications. Suppose that the eigenvalues of g are not roots of unity and that (x, y) = z is a symplectic basis for g. The eigendirections of g are the two lines x = 0 and y = 0. Above, it was shown that any homogeneous integral of g is of the form c(xy)S, s E N. Let g' be another map of G.
Since the function (xy ) 9 is invariant under the action of g', the set xy = 0 is fixed by g'. Since g' is a non-degenerate linear map, the point x = y = 0 is fixed and g' either preserves the eigendirections of g or permutes them. In the first case g', obviously, commutes with g, and in the second case it has the form X ---+ay, y --+~X.
Since g' is symplectic, its matrix satisfies the condition hence a{3 = -1. But in this case the eigenvalues of S are ±i. The points ±i form precisely that exceptional regular polygon mentioned in the conclusion of the theorem, as required. We consider the case when the elements of A(t) are homogeneous doublyperiodic meromorphic functions of the time t E C, having only one pole inside the parallelogram of periods. We may take A(t) to be a meromorphic function on the complex torus X obtained from the complex plane C by factoring out the lattice of periods. We consider two symplectic maps g and g' of a period of A(t). We assume that their eigenvalues satisfy the conditions of Theorem 1. Then for (2.1) to have n independent analytic integrals it is necessary that g and g' commute. Consequently, to a circuit of a singular point (the element gg' g-1 g' -t E G) there corresponds the identity map of C 2 n. 3. We apply this argument to the linear differential equation (2.2) z + (w 2 + ej(t))z = 0, where w and E are real constants, f(t) is an elliptic function with the periods 27r and 27f i, having a unique pole of order 2 in the rectangle of periods. We may assume that f for real t takes real values. An example is the Weierstrass function W. We look for linearly independent solutions of (2.2) in the form of a series z(t)=tP ~ c,t", pEC, c 0 =;i=O.
n~O Since ·; (t) =c tP ~ (p-+-n) (p + n-1) cnt"-~, This equation gives us two values p 1 and p 2 to which there correspond two linearly independent solutions of (2.2). After a circuit of the pole these solutions are multiplied, respectively, by e2nip, and e 2 nip,. The corresponding monodromy matrix is the identity if p 1 and p 2 are integers. In particular, Eo: must be an integer. For € = 0 the eigenvalues of the monodromy matrix of (2.2) under the map with period 211' and 21Ti are, respectively, A. 1 • 2 = e±2nwi and fA-1 , 2 = e±4nw.
Obviously, l!l 1 ,2l =I= 1 and A 1 , 2 =I= ±i if w =I= 0 and w =I=-!+ k1r, k E Z. By continuity, if w =I=!+ k1r, then for small values € =I= 0 the eigenvalues J.1. 1 , 2 are not roots of unity and A 1 , 2 =I= ±i (this property in fact holds for almost all w and e). Consequently, by Theorem 2, (2.2) in these cases is not integrable in the complex domain. We note that in the real domain this equation is completely integrable: it has an analytic integral f(z, z, t) that is 211'-periodic in t. The fact is that by a linear canonical change of variables that is 211'-periodic in t the equations (2.2) can be reduced to a linear autonomous Hamiltonian system with one degree of freedom. For f we can take the Hamiltonian function of the autonomous sytems.
We now consider the non-linear equation of the oscillations of a mathematical pendulum z + (w 2 + ej(t)) sin z = 0.
We claim that it can have an analytic integral f(z, z, t) that is doublyperiodic in t E C only for those values of w and e for which the linear equation (2.2) is integrable. To prove this we expand f in a convergent power senes (2.3) whose coefficients f~cz are elliptic functions with periods 211' and 21Ti. The first form in (2.3) (when s = m) is obviously a single-valued integral of (2.2).
Consequently, by hypothesis, it must be constant. But then the next form (s = m + 1) is an integral of (2.2) and therefore also constant, and so on. 4. The last remark can be generalized. Suppose that the non-linear Hamiltonian system . (2.4) z = 'Jll', z E C 2 n has a particular solution z 0 (t) that is single-valued on its Riemann surface X.
With its help we can, for example, reduce the number of degrees of freedom of (2.5) by 1.
We assume that the non-linear equation (2.4) has several independent holomorphic integrals Fs(z) (1 ,;;;:; s ,;;;:; m). Then (2.5) also has first integrals. They are the homogeneous forms of the expansion of Fs in a power series In u: (F;(z 0 (t)), u>+··· These forms are holomorphic functions on the direct product czn X X. We have Lemma. If (2.4) has m independent integrals, then the equation in variations (2. 5) has m independent polynomial integrals [ 24] .
Thus, the problem of the complete integrability of Hamiltonian systems in the complex domain reduces to an investigation of the integrability of linear canonical systems.
By this method Ziglin has proved the integrability of the Hamiltonian systems of Henon-Heile and Yang-Mills (see Ch. 1). He has also applied it to the problem of the rotation of a heavy rigid body around a fixed point. It turned out that an additional holomorphic integral exists only in the three classical cases of Euler, Lagrange, and Kovalevskaya. If the area constant is fixed to be zero, then to these must be added the case of Goryachev-Chaplygin [24].
For the systems of Henon-Heile and Yang-Mills one can prove that there are no integrals even in a real domain. The question of the existence of an additional real analytic integral for an arbitrary mass distribution in a rigid body remains open.