Speed of Arnold diffusion for analytic Hamiltonian systems (Q648056)

It is well known from KAM theory that, for a small perturbation of a completely integrable Hamiltonian system satisfying some nondegeneracy condition, most of the phase space is occupied by invariant tori of maximal dimension. The actions along the trajectories on these invariant tori can undergo only very small variations. However, when there are more than two degrees of freedom, the system may also display trajectories along which the actions undergo a change of order 1, however small the perturbation might be. This phenomenon was demonstrated by \textit{V. I. Arnol'd} [Sov. Math., Dokl. 5, 581--585 (1964); translation from Dokl. Akad. Nauk SSSR 156, 9--12 (1964; Zbl 0135.42602)]. On the other hand, according to a theorem of \textit{N. N. Nekhoroshev} [Russ. Math. Surv. 32, No. 6, 1--65 (1977; Zbl 0389.70028)], under some condition satisfied in particular by real-analytic Hamiltonians that are quasi-convex in the fiber, the time of diffusion (which is the minimal time required to obtain a variation of the actions larger than a fixed positive constant) is exponentially long with respect to the size \(\varepsilon\) of the perturbation, namely at least of the order of \(\exp(c\varepsilon^{-a})\), \(c,a> 0\). For analytic quasi-convex Hamiltonian systems, several works have improved the value of the Nekhoroshev exponent \(a\), up to \({1\over 2(n-1)}\) for systems with \(n\) degrees of freedom. The author proves that, in the so-called generalized example of Arnol'd with \(n\geq 5\), the diffusion time is \(O(\exp(c\varepsilon^{-b}))\) with \(b= {1\over 2(n-2)}\), generalizing results due to \textit{U. Bessi} [Nonlinear Anal., Theory Methods Appl. 26, No. 6, 1115--1135 (1996; Zbl 0867.70013); Nonlinearity 10, No. 3, 763--781 (1997; Zbl 0912.70015)] for \(n= 3, 4\). More precisely, the author considers the Hamiltonian \[ H(q,\varphi, p,I)= {1\over 2} p^2+ \varepsilon(\cos q-1)+ {1\over 2} I^2+ \varepsilon\mu(1- \cos q)f(\varphi), \] where \(q\in\mathbb{T}= \mathbb{R}/2\pi\mathbb{Z}\), \(\varphi\in\mathbb{T}^d\), \(p\in\mathbb{R}\), \(I\in\mathbb{R}^d\), \(f(\varphi)= \sum_{k\in\mathbb{Z}^d} f_k\cos(k,\varphi)\) and the real sequence \((f_k)\) decays exponentially. He proves that the associated system has, for \(\mu\) suitably chosen exponentially small with respect to \(\varepsilon\), a diffusion orbit in time \(T= O(\exp(c\varepsilon^{-b}))\). Because of the peculiar form of the perturbation, the system has a continuum of invariant \(d\)-dimensional partially hyperbolic tori \({\mathcal T}_\omega\) defined by \(q= 0\), \(p= 0\), \(I=\omega\), that may have stable and unstable manifolds intersecting transversally. Arnol'd diffusion can occur along a (continuous) chain of tori having this property, and more important is the ``splitting'' between the stable and the unstable manifolds of these tori, faster can be the drift of the actions \(I\). The splitting is related to the properties of some function which (for \(\mu\) exponentially small) can be approximated by the Poincaré Melnikov potential. The main idea of the paper is to define ``transition points'', which are frequency vectors \(\omega\in\mathbb{R}^d\) having special arithmetic properties. The author proves that there is a continuous path of such transition points of distant end-points, and that the splitting at the transition points is relatively large. This enables him, using the variational method introduced by Bessi and Mather, to construct a diffusion orbit with a good estimate of the time. The exponent \(b\) in the upper estimate of the diffusion time, however larger than the Nekhoroshev exponent \(a\), is expected to be almost optimal, because the constructed diffusion orbit passes near a double resonance, which makes the drift of the actions slower.