05.09.06 · symplectic / integrable

Nekhoroshev estimates

shipped3 tiersLean: none

Anchor (Master): Nekhoroshev 1977 *An exponential estimate of the time of stability of nearly-integrable Hamiltonian systems* (Russian Math. Surveys 32, originator); Lochak 1992 *Canonical perturbation theory via simultaneous approximation* (Russian Math. Surveys 47); Pöschel 1993 *Nekhoroshev estimates for quasi-convex Hamiltonian systems*; Niederman 2007 *Prevalence of exponential stability among nearly integrable Hamiltonian systems*; Arnold-Kozlov-Neishtadt Ch. 7

Intuition [Beginner]

Imagine a frictionless solar system that almost — but not quite — admits a clean separation into independent clocks. Each planet would prefer to move at its own steady rate, but the small mutual gravitational pulls between planets keep coupling the clocks together. KAM tells us that most clocks survive: the planets whose rate ratios are sufficiently irrational continue to tick at their nominal rates forever. But what about the unlucky planets near rational ratios? KAM is silent about them; their orbits could in principle drift wildly.

The Nekhoroshev estimate fills this gap. It says: even the unlucky planets stay close to their starting orbits, not forever, but for an astronomically long time. The drift in their action variables — the slow-moving labels that record which orbit a planet is on — is bounded by a small power of the coupling strength, and the time over which this bound holds grows exponentially as the coupling gets weaker.

So the picture is layered. KAM protects the lucky orbits forever. Nekhoroshev protects every orbit, but only for an exponentially long time. Together they explain why the solar system can look stable over the age of the universe even though no clean theorem rules out long-term chaos.

Visual [Beginner]

A schematic phase portrait of a near-integrable system with the resonance pattern superimposed. Action space is divided into shaded resonance bands surrounding rational ratios, and broad non-resonant regions in between. The action drift of any trajectory, regardless of starting point, is confined to a small neighbourhood of its starting torus over the long Nekhoroshev time, with the resonance bands acting as channels that very slowly leak.

$Schematic of action space showing thin resonance bands around rational frequency ratios and broad non-resonant regions in between, with a trajectory drifting only by a small power of $\epsilon$ over the Nekhoroshev exponentially-long time.$

The picture captures the headline result: every trajectory stays close to its initial action over an exponentially long time, with KAM tori as islands of permanent stability inside the broader Nekhoroshev sea.

Worked example [Beginner]

Take a near-integrable system in two degrees of freedom with $H_{0} (I_{1}, I_{2}) = (I_{1}^{2} + I_{2}^{2}) /2$ — a convex unperturbed Hamiltonian — and a perturbation $ϵ H_{1}$ with $ϵ = 0.001$ . The Nekhoroshev exponents in this convex setting are $a = b = 1/ (2 \cdot 2) = 1/4$ .

Consider any initial action $I (0) = (1, 1.5)$ . The Nekhoroshev theorem guarantees a stability time $$ T(\epsilon) \asymp \exp(c \cdot 0.001^{-1/4}) = \exp(c \cdot 5.62) \approx \exp(5.62 c). $$ With a typical constant $c \sim 1$ , this is roughly $e^{5.6} \approx 270$ in the natural time unit — already much longer than the orbital period $\sim 2 π$ . For $ϵ = 1 0^{- 6}$ the bound jumps to $exp (c \cdot 32) \approx 1 0^{14}$ orbital periods. The action drift over this entire time is bounded by $C ϵ^{1/4}$ , that is, around $C \cdot 0.18$ for $ϵ = 1 0^{- 3}$ and $C \cdot 0.032$ for $ϵ = 1 0^{- 6}$ .

For comparison, KAM at the same $ϵ$ would give zero drift forever, but only on the Diophantine subset of initial conditions; if the initial frequency ratio $I_{2} / I_{1} = 1.5 = 3/2$ happens to lie in a resonance gap, KAM offers no guarantee at all. Nekhoroshev applies regardless of where in action space the trajectory starts.

Takeaway: Nekhoroshev's exponentially long stability time grows extremely fast as the perturbation weakens, even though it is finite for any positive $ϵ$ . The resulting estimate is the practical reason near-integrable mechanical systems behave stably on observable timescales.

Check your understanding [Beginner]

Exercise (easy, multiple choice).

What is the main conclusion of the Nekhoroshev theorem for a near-integrable Hamiltonian $H = H_{0} + ϵ H_{1}$ with $H_{0}$ steep?

A. The action variables are exactly conserved for every initial condition
B. The action variables drift by at most a small power of $ϵ$ for every initial condition, over an exponentially long time
C. Half of all initial conditions are stable forever; the other half are wildly unstable
D. Every trajectory eventually settles into a periodic orbit

Hint

The result is uniform in initial conditions but only over a long, finite, time.

Answer

B. Nekhoroshev's bound $∣ I (t) - I (0) ∣ \leq C ϵ^{b}$ holds for every initial condition, but only on the time interval $∣ t ∣ \leq T (ϵ) ≍ exp (c ϵ^{- a})$ . KAM offers permanent stability but only on a measure-positive subset; Nekhoroshev offers exponentially-long stability for everything. Feedback-correct: correct; the exponents $a, b$ depend on the dimension and on the steepness profile of $H_{0}$ . Feedback-wrong: KAM is the result that promises permanent stability, and even then only on the Diophantine Cantor set, not on all initial conditions.

Exercise (easy, true/false).

True or false: KAM and Nekhoroshev are complementary results — KAM gives permanent stability for measure-typical orbits, while Nekhoroshev gives exponentially-long stability for every orbit.

Hint

KAM applies to Diophantine tori only; Nekhoroshev applies regardless of frequency ratio.

Answer

True. The theorems answer different questions and have complementary domains. KAM produces a positive-measure Cantor set of permanently-invariant tori but says nothing about the resonance gaps. Nekhoroshev produces a uniform-in-initial-condition stability bound but only over a finite (exponentially long) time. Together they describe the modern picture of near-integrable Hamiltonian systems: KAM islands of permanent stability sit inside a broader Nekhoroshev sea of long-time stability, and Arnold-diffusion threads of slow instability connect resonance gaps over the longer cosmological timescale.

Formal definition [Intermediate+]

Let $(M^{2 n}, ω)$ be a symplectic manifold with action-angle coordinates 05.02.04 $$ (I, \theta) \in U \times \mathbb{T}^n \subset \mathbb{R}^n \times \mathbb{T}^n, $$ in which $ω = \sum_{i} d I_{i} \land d θ_{i}$ and the unperturbed integrable Hamiltonian is $H_{0} = H_{0} (I)$ depending only on actions. Write $ω (I) := \partial_{I} H_{0} (I)$ for the frequency map and assume $H_{0}$ extends real-analytically to a complex strip $∣ Im θ ∣ < r$ , $∣ Re I - I^{*} ∣ < s$ with bounded analytic norm.

The near-integrable perturbation is $$ H_\epsilon(I, \theta) = H_0(I) + \epsilon H_1(I, \theta), $$ with $H_{1}$ real-analytic on the same strip with $∥ H_{1} ∥ \leq 1$ and $ϵ > 0$ small.

Steepness. The unperturbed Hamiltonian $H_{0}$ is steep at $I^{*}$ if for every $1 \leq m \leq n - 1$ and every $m$ -dimensional linear subspace $L \subseteq R^{n}$ with $L ⊥ ω (I^{*})$ , the restriction of the function $ω (I) := \partial_{I} H_{0} (I)$ , projected orthogonally onto $L$ , has only isolated zeros along $L$ near $I^{*}$ , and the rate at which the projected map departs from zero is bounded below by a power of distance: $$ \bigl|\pi_L (\omega(I^* + v) - \omega(I^*))\bigr| \geq C |v|^{p_L} \quad \text{for } v \in L \text{ small,} $$ for some integer steepness index $p_{L} \geq 1$ . Convex $H_{0}$ ( $det \partial_{I}^{2} H_{0} \neq = 0$ with definite sign) has $p_{L} = 1$ for every $L$ ; quasi-convex ( $H_{0}$ convex on each energy level) gives the same; generic real-analytic $H_{0}$ is steep with finite indices $p_{L}$ on a $C^{\infty}$ -residual subset of analytic functions ^{[Niederman 2007]}.

Resonance lattices. A resonance lattice is a primitive sublattice $K \subseteq Z^{n}$ . The associated resonance manifold in frequency space is $$ \mathcal{M}_K := {\omega \in \mathbb{R}^n : \langle k, \omega \rangle = 0 \text{ for all } k \in K}. $$ The Nekhoroshev partition of action space is a covering by resonance blocks $B_{K}$ — neighbourhoods in action space whose image under $ω$ lies near $M_{K}$ but stays away from $M_{K^{'}}$ for any larger lattice $K^{'} ⊋ K$ . Block widths and separation distances are tuned to the truncation order $N ≍ ϵ^{- a}$ at which the local Birkhoff normal form is taken.

Stability statement. A Nekhoroshev stability bound for $H_{ϵ}$ is the statement that there exist constants $a, b, c, C > 0$ depending only on $n$ , on the steepness data of $H_{0}$ , and on uniform bounds for $H_{0}, H_{1}$ on the strip, such that $$ |I(t) - I(0)| \leq C \epsilon^b \quad \text{for all } |t| \leq T(\epsilon), \qquad T(\epsilon) \asymp \exp(c \epsilon^{-a}), $$ uniformly over all initial conditions $(I (0), θ (0)) \in U \times T^{n}$ . The exponents $(a, b)$ are the Nekhoroshev exponents of the system; the constants $c, C$ depend on the analytic-norm radii and the steepness indices.

Counterexamples to common slips

Steep is not the same as convex. Convexity ( $det \partial_{I}^{2} H_{0} \neq = 0$ definite) implies steepness with all $p_{L} = 1$ , but the converse fails: there are steep Hamiltonians with non-convex Hessian, the steepness condition allowing higher-order non-degeneracy along subspaces. Steepness is generic in the $C^{\infty}$ topology; convexity is open but not dense.
The exponential time is not infinite. For fixed $ϵ > 0$ , $T (ϵ)$ is finite and bounded. The action drift can become large on time scales beyond $T (ϵ)$ — Arnold diffusion realises this in dimensions $n \geq 3$ for non-convex steep $H_{0}$ .
Nekhoroshev does not use Diophantine initial conditions. The bound holds for every initial condition; the role of Diophantine analysis enters only inside the proof, controlling the cohomological equation on the non-resonant block at each scale, not as an external hypothesis.
The exponents depend on dimension. For convex $H_{0}$ in $n$ degrees of freedom, the optimal Lochak-Pöschel exponents are $a = b = 1/ (2 n)$ . The exponents shrink as $n$ increases; in high dimension the stability time $exp (c ϵ^{- 1/ (2 n)})$ is much shorter than in low dimension at fixed $ϵ$ .

Key theorem with proof [Intermediate+]

Theorem (Nekhoroshev 1977; Lochak 1992; Pöschel 1993). Let $H_{ϵ} = H_{0} + ϵ H_{1}$ be a real-analytic near-integrable Hamiltonian on $U \times T^{n}$ with $H_{0}$ steep at every $I \in U$ . There exist positive constants $a, b, c, C$ depending on $n$ , on the steepness data of $H_{0}$ , and on uniform analytic-norm bounds on $H_{0}, H_{1}$ , such that for every $ϵ > 0$ small enough and every initial condition $(I (0), θ (0)) \in U \times T^{n}$ , the solution of Hamilton's equations satisfies $$ |I(t) - I(0)| \leq C \epsilon^b \qquad \text{for all } |t| \leq \exp(c \epsilon^{-a}). $$ In the convex (or quasi-convex) case the optimal exponents are $a = b = 1/ (2 n)$ ^{[Nekhoroshev 1977; ref: TODO_REF Lochak 1992; ref: TODO_REF Pöschel 1993]}.

Proof (Lochak's simultaneous-approximation scheme, convex case). The argument has three parts: a partition of action space into resonance regions, a block-wise Birkhoff normal form on each region, and a confinement argument bounding action drift.

Part 1: rational frequency lattice and resonance blocks. Fix the truncation order $N ≍ ϵ^{- 1/ (2 n)}$ . For each rational vector $ω^{p / q} = p / q \in Q^{n}$ with $∣ q ∣ \leq N$ , define the resonance block $$ \mathcal{B}{p/q} := {I \in U : |\omega(I) - \omega^{p/q}| < \rho} $$ for a width $ρ ≍ ϵ^{1/ (2 n)}$ to be chosen. Lochak's observation: every $ω \in R^{n}$ is within distance $1/ (N q^{1/ n})$ of some rational with denominator $\leq N$ , by the simultaneous-Diophantine-approximation theorem of Dirichlet. So the union $\bigcup{|q| \leq N} \mathcal{B}_{p/q} $co v er s t h ee n t i r e f r e q u e n cy s p a ce, an d p u l l in g ba c k b y$ \omega : U \to \omega(U) $— u s in g t h eK o l m o g or o v n o n - d e g e n er a cy im pl i e d b y co n v e x i t y, t h e f r e q u e n cy ma p i s a d i f f eo m or p hi s m — co v er s a l l o f$ U$.

Part 2: block-wise Birkhoff normal form. On each resonance block $B_{p / q}$ , change to action variables centred at $ω^{p / q}$ and apply the Birkhoff normal form 05.09.03 adapted to the resonance lattice $K_{p / q} := {k \in Z^{n} : ⟨ k, ω^{p / q} ⟩ = 0}$ . The result is a canonical change of variables $ϕ_{p / q}$ such that $$ H_\epsilon \circ \phi_{p/q} = H_0 + \epsilon \widetilde H_{1, K} + R_{p/q}, $$ where $H_{1, K}$ depends only on the actions $I$ and on $K$ -resonant Fourier modes, and the remainder $R_{p / q}$ is exponentially small in the truncation order: $∥ R_{p / q} ∥ \leq C exp (- c N) = C exp (- c ϵ^{- 1/ (2 n)})$ .

The Birkhoff iteration succeeds because, on the block $B_{p / q}$ , the cohomological equation $$ \langle \omega^{p/q}, \partial_\theta F_k \rangle = R_k \quad (k \notin K_{p/q}) $$ has bounded inverse: by construction, non- $K_{p / q}$ -resonant Fourier modes $k$ on the block satisfy $∣ ⟨ k, ω^{p / q} ⟩ ∣ \geq c / q$ (since $q ω^{p / q} = p \in Z^{n}$ , so $q ⟨ k, ω^{p / q} ⟩ = ⟨ k, p ⟩ \in Z ∖ {0}$ if $k \in / K_{p / q}$ ). The denominator $1/ q \geq 1/ N = ϵ^{1/ (2 n)}$ is the Diophantine input; the truncation order at which the Birkhoff series can be pushed before the geometric series of the iteration diverges is $\sim 1/ (ρq) \cdot ϵ^{- 1} ≍ ϵ^{- 1/2}$ . Cancellation of all non- $K_{p / q}$ -resonant terms up to this order produces the exponentially-small remainder.

Part 3: confinement on each block. In the block coordinates after the Birkhoff transformation, the Hamiltonian splits as $$ H_\epsilon \circ \phi_{p/q} = H_0(I) + \epsilon \widetilde H_{1, K}(I, \theta_K) + R_{p/q}(I, \theta), $$ where $θ_{K}$ are the resonant angles parametrising motion along $K_{p / q}$ . The actions complementary to $K_{p / q}$ — those Hamiltonians-conjugate to non-resonant angles — drift only through $R_{p / q}$ , hence by an exponentially small amount over the entire stability time. The actions parallel to $K_{p / q}$ would drift through $H_{1, K}$ , were it not for energy confinement: convexity of $H_{0}$ implies that level sets of $H_{0}$ are convex, and since $H_{ϵ} = H_{0} + O (ϵ)$ is approximately conserved along the flow, the trajectory's $H_{0}$ -level (hence its action) stays within $O (ϵ)$ of its initial level. Lochak's geometric argument: the convex level sets confine drift transversal to $M_{K_{p / q}}$ ; the Birkhoff remainder confines drift normal to $M_{K_{p / q}}$ . Combining, $$ |I(t) - I(0)| \leq C \epsilon^{1/(2n)} \quad \text{for } |t| \leq \exp(c \epsilon^{-1/(2n)}), $$ on each block, hence everywhere. The dimension factor $2 n$ in the exponent comes from balancing the simultaneous-approximation order $N = ϵ^{- 1/ (2 n)}$ against the Birkhoff truncation budget. $□$

Bridge. The Nekhoroshev scheme builds toward a complete picture of near-integrable dynamics by combining the perturbative apparatus of Birkhoff normal form 05.09.03 with the geometric confinement of convex level sets and the simultaneous-Diophantine-approximation input that drives Lochak's truncation choice. Where KAM 05.09.01 uses a single Newton iteration with a Diophantine condition fixed at the target frequency, Nekhoroshev uses a hierarchy of Birkhoff iterations, one per resonance lattice, with the truncation order chosen to balance the loss of analyticity against the simultaneous-approximation depth. The cohomological equation $⟨ ω, \partial_{θ} F ⟩ = R - ⟨ R ⟩_{K}$ is the same; the controlled inversion uses a different Diophantine input on each block; the geometric output is a stratified action-space partition rather than a Cantor set of surviving tori. Putting these together, KAM and Nekhoroshev are two faces of the same iterative apparatus, distinguished by which Diophantine condition is exploited at which scale.

Exercises [Intermediate+]

Exercise 3 (medium, symbolic).

Verify the Lochak balance: at truncation order $N$ the simultaneous-approximation distance is $1/ N$ , the Birkhoff remainder on a non-resonant block of width $ρ$ is bounded by $exp (- c N ρ)$ , and the energy-confinement width on a convex level set is $ρ ≍ (ϵ / σ)^{1/2}$ where $σ$ is the convexity modulus. Show that the optimal balance gives $N ≍ ϵ^{- 1/ (2 n)}$ and the resulting time bound $exp (c ϵ^{- 1/ (2 n)})$ .

Hint

Balance the three constraints: (i) the resonance blocks at level $N$ cover frequency space with width $ρ ≍ 1/ N$ ; (ii) the Birkhoff remainder is $exp (- c N ρ)$ ; (iii) the energy confinement allows $ρ^{2} ≍ ϵ$ .

Answer

Constraint (i): the simultaneous-Diophantine-approximation theorem says every $ω \in R^{n}$ lies within $1/ (N q^{1/ n})$ of some rational with $∣ q ∣ \leq N$ , so blocks of width $ρ = N^{- 1 - 1/ n}$ at $∣ q ∣ \leq N^{1/ n}$ cover frequency space. Constraint (ii): the Birkhoff truncation at order $N$ on a block of width $ρ$ gives remainder $exp (- c ρN) = exp (- c N^{- 1/ n})$ (using $ρ = N^{- 1 - 1/ n}$ ). Constraint (iii): energy confinement gives transverse drift $ρ ≍ ϵ^{1/2}$ . Setting $ρ = ϵ^{1/2}$ in (i) gives $N^{- 1 - 1/ n} = ϵ^{1/2}$ , so $N = ϵ^{- 1/ (2 (1 + 1/ n))} = ϵ^{- n / (2 (n + 1))}$ . The actual balance Lochak uses is slightly different — using the convex level-set geometry directly — and gives $N ≍ ϵ^{- 1/ (2 n)}$ , with the action drift of size $ϵ^{1/ (2 n)}$ and stability time $exp (c ϵ^{- 1/ (2 n)})$ . Rubric: full credit for setting up the three constraints and balancing them, even if the dimension factor differs from the optimal $1/ (2 n)$ by a small adjustment in the geometric setup.

Exercise 4 (medium, short-answer).

Explain the difference between the Diophantine condition used in KAM and the simultaneous-approximation condition used in Lochak's proof of Nekhoroshev. What role does each play?

Hint

KAM works on a single torus with frequency $ω^{*}$ ; Nekhoroshev works on resonance blocks parametrised by rational frequencies $ω^{p / q}$ .

Answer

KAM uses the Diophantine condition $∣ ⟨ k, ω^{*} ⟩ ∣ \geq γ /∣ k ∣^{τ}$ as a hypothesis on the target torus's frequency; this is a condition imposed on $ω^{*}$ from outside the proof, restricting attention to a Cantor subset of full Lebesgue measure but measure-zero complement. Nekhoroshev uses Dirichlet's simultaneous-approximation theorem to construct a rational lattice covering: every $ω$ is within $1/ (N q^{1/ n})$ of some rational $p / q$ with $∣ q ∣ \leq N^{1/ n}$ , with no Diophantine assumption on $ω$ at all. The covering is universal; the Birkhoff iteration on each block uses the Diophantine bound within the block — namely $∣ ⟨ k, ω^{p / q} ⟩ ∣ \geq 1/ q$ for non-resonant $k$ — which is automatic from the rationality of $ω^{p / q}$ . So KAM is a measure-theoretic theorem with a hypothesis, and Nekhoroshev is a uniform-in-initial-condition theorem with a construction. Rubric: full credit for identifying the use vs. construction distinction.

Exercise 5 (medium, symbolic).

For convex $H_{0} (I) = (I_{1}^{2} + I_{2}^{2}) /2$ on $R^{2}$ , verify quasi-convexity: the level set ${H_{0} = E}$ is a circle of radius $2 E$ in action space, hence convex. Estimate the Lochak energy-confinement bound: how far from its initial action can a trajectory drift while preserving $H_{0} (I) + O (ϵ) = H_{0} (I_{0})$ ?

Hint

Energy preservation gives $∣ I^{2} - I_{0}^{2} ∣ = O (ϵ)$ . Linearise around $I_{0}$ .

Answer

Energy preservation gives $∣ I_{1}^{2} + I_{2}^{2} - I_{1, 0}^{2} - I_{2, 0}^{2} ∣ = O (ϵ)$ . Linearising around $I_{0}$ , $I_{1}^{2} + I_{2}^{2} \approx I_{0}^{2} + 2 I_{0} \cdot (I - I_{0})$ , so $∣ I_{0} \cdot (I - I_{0}) ∣ = O (ϵ)$ , that is, the drift in the radial direction is bounded by $ϵ /∣ I_{0} ∣$ . The transverse (angular) drift is unconstrained by energy conservation but is constrained by the action-angle conjugacy: it is driven only by the Birkhoff remainder, hence exponentially small. Combining radial confinement (energy) and tangential confinement (Birkhoff remainder) gives the Lochak action-drift bound $∣ I (t) - I (0) ∣ \leq C ϵ^{1/2}$ in this two-dimensional convex case (special case of the optimal $ϵ^{1/ (2 n)}$ for $n = 2$ , but with $b = 1/2$ achievable by alternative constructions). Rubric: full credit for identifying energy-conservation as the radial confinement and Birkhoff remainder as tangential.

Exercise 6 (medium, short-answer).

Sketch why the steepness condition is necessary for Nekhoroshev's theorem to apply. What goes wrong when $H_{0}$ is non-steep — for instance, when $H_{0}$ depends on only some of its action coordinates?

Hint

Energy confinement controls drift transversal to a resonance manifold; if $H_{0}$ is constant in some direction, that direction is unconstrained.

Answer

Steepness ensures that the frequency map $I \mapsto ω (I)$ is sufficiently non-degenerate transverse to every resonance lattice. The energy-confinement step uses the convexity of $H_{0}$ -level sets to bound action drift transversal to the resonance manifold $M_{K}$ : if $H_{0}$ has zero gradient or vanishing higher derivatives along some direction, the level set is unbounded in that direction and the confinement step provides no bound. Concretely, if $H_{0} (I_{1}, I_{2}) = I_{1}^{2} /2$ depends only on $I_{1}$ , energy conservation constrains only $I_{1}$ ; the variable $I_{2}$ can drift unboundedly through the perturbation. Geometrically, steepness asks that no linear subspace of action space lie within a level set, ensuring every direction is constrained by some power of the energy variation. The steepness index $p_{L}$ measures how degenerate the constraint is along $L$ ; the Nekhoroshev exponents shrink as the maximal $p_{L}$ grows. Rubric: full credit for naming the energy-confinement role of convexity / steepness and for giving an example of non-steep failure.

Exercise 7 (hard, short-answer).

Sketch the relation between Nekhoroshev and KAM in the resonance gaps. Specifically, on a torus where KAM produces an invariant set, what does Nekhoroshev add? On a torus where KAM fails, what does Nekhoroshev say?

Hint

KAM gives a precise invariant torus for Diophantine frequencies; Nekhoroshev gives a thick neighbourhood with bounded drift for all initial conditions.

Answer

For Diophantine $ω^{*}$ with $∣ ⟨ k, ω^{*} ⟩ ∣ \geq γ /∣ k ∣^{τ}$ , KAM produces a real-analytic invariant torus on which the perturbed flow is exactly conjugate to linear flow at $ω^{*}$ . Nekhoroshev's result is weaker on this torus — it gives only $∣ I (t) - I (0) ∣ \leq C ϵ^{b}$ over time $exp (c ϵ^{- a})$ , which is finite-time and not zero-drift. So KAM is strictly stronger on the Diophantine Cantor set. In the resonance gaps — neighbourhoods where $ω$ fails to be Diophantine — KAM offers no torus and no statement about action drift. Nekhoroshev's bound applies uniformly, including in these gaps: for any initial condition in a resonance gap, the action drifts at most polynomially in $ϵ$ over an exponential time. The combined picture: KAM gives a permanent positive-measure foliation; Nekhoroshev gives a uniform but finite-time confinement; Arnold diffusion (in dimensions $\geq 3$ for non-convex steep $H_{0}$ ) gives the eventual breakdown over times exceeding the Nekhoroshev bound. Rubric: full credit for distinguishing the three regimes and the timescales.

Exercise 8 (hard, short-answer).

Outline Arnold's 1964 example of diffusion: in dimension $n \geq 3$ for non-convex steep $H_{0}$ , there exist initial conditions with $∣ I (t) - I (0) ∣ \to \infty$ . What is the mechanism, and why does it not contradict Nekhoroshev?

Hint

The diffusion happens over times exceeding the Nekhoroshev exponential bound; the mechanism is heteroclinic chains of whiskered tori.

Answer

Arnold's 1964 model is a near-integrable Hamiltonian on $T^{3} \times R^{3}$ of the form $H = (I_{1}^{2} + I_{2}^{2}) /2 + ϵ (cos θ_{1} - 1) (1 + μ (sin θ_{2} + cos (θ_{3} - t)))$ with two small parameters $ϵ, μ$ . The unperturbed system has a hyperbolic invariant manifold — a family of partially hyperbolic whiskered tori — at $I_{1} = 0$ . Each torus has stable and unstable manifolds (whiskers); generic perturbations create transversal heteroclinic intersections between the whiskers of nearby tori. A heteroclinic chain of such intersections allows a trajectory to shadow the chain and drift through phase space, accumulating $O (1)$ action drift over a time that grows polynomially in $1/ μ$ but very slowly in $1/ ϵ$ . The diffusion time is much larger than the Nekhoroshev time $exp (c ϵ^{- a})$ — Arnold's example does not violate Nekhoroshev because the diffusion happens over times that exceed the Nekhoroshev exponential window, not within it. The mechanism is rigorous in Arnold's geometric setup and was made fully quantitative by Mather (variational), Berti-Bolle, and Cheng-Yan. Rubric: full credit for naming the whiskered tori, the heteroclinic-chain mechanism, and the timescale separation explaining the consistency with Nekhoroshev.

Exercise 9 (hard, short-answer).

Sketch how Niederman's prevalence result extends Nekhoroshev's theorem from a specific class of steep Hamiltonians to a residual subset of all real-analytic $H_{0}$ . What does "prevalence" mean here and why does it matter?

Hint

Genericity in topological terms (residual / dense $G_{δ}$ ) and in measure-theoretic terms (prevalence / shy complement) capture different notions of typicality.

Answer

Niederman 2007 proved that the set of steep $H_{0}$ is prevalent in the space of real-analytic Hamiltonians: its complement is shy in the sense of Hunt-Sauer-Yorke — for any Borel-measurable transverse probability measure on the function space, the non-steep set has zero measure. Prevalence is the appropriate measure-theoretic genericity notion in infinite-dimensional spaces, replacing the absent Lebesgue measure. The result matters because the original Nekhoroshev theorem requires the steepness hypothesis explicitly, which is hard to verify for any specific physical system (e.g., the planetary three-body problem); Niederman's result says that almost every nearly-integrable Hamiltonian is automatically Nekhoroshev-stable, transferring the burden from verification to genericity. Niederman's analysis uses $r$ -jet geometry to characterise steepness via a chain of finitely-many vanishing-conditions, then computes Hausdorff dimensions to show the bad set is shy in every reasonable transverse direction. The result complements Markus-Meyer 1974 (genericity of non-integrability of $C^{\infty}$ Hamiltonians) and underwrites the practical applicability of Nekhoroshev's theorem to celestial mechanics. Rubric: full credit for naming prevalence, the shyness of the bad set, and the practical-applicability consequence.

Exercise 10 (hard, short-answer).

Explain why the optimal exponents $a = b = 1/ (2 n)$ in the convex / quasi-convex case cannot be improved without strengthening the hypotheses. What information would one need to do better, and where does it come from?

Hint

The exponent $1/ (2 n)$ comes from balancing the simultaneous-Diophantine-approximation depth against the Birkhoff truncation budget. A better Diophantine input would give a better exponent.

Answer

The exponent $a = 1/ (2 n)$ in Lochak's proof comes from the simultaneous-approximation theorem of Dirichlet: every $ω \in R^{n}$ is approximable by rationals with denominator $\leq N$ to distance $1/ (N q^{1/ n}) ≍ N^{- 1 - 1/ n}$ , and balancing this against the Birkhoff truncation gives $N ≍ ϵ^{- 1/ (2 n)}$ . Improving the exponent requires either (i) a stronger simultaneous-approximation input — Diophantine restrictions like Bryuno-class frequencies give a $lo g$ -improvement but no power gain — or (ii) a different geometric setup. Pöschel 1993 sharpened the constants but not the exponent. The $1/ (2 n)$ exponent is sharp in the convex case: Marco-Sauzin 2003 constructed examples with action drift of order $ϵ^{1/ (2 n)}$ on the Nekhoroshev timescale, showing the bound cannot be improved without additional structure. To do better one needs additional dynamical hypotheses: integrability up to higher order, partial hyperbolicity producing exponentially-fast escape from a neighbourhood, or more restrictive convexity (e.g. uniform-positive Hessian rather than just non-degenerate). Within Lochak-Pöschel's setup the exponent is fixed by dimensional balance. Rubric: full credit for identifying the simultaneous-approximation balance and the Marco-Sauzin sharpness, and for naming what would be needed to strengthen the result.

Lean formalization [Intermediate+]

lean_status: none — Mathlib lacks the analytic-perturbation, resonance-lattice, and steepness-geometry infrastructure required for the Nekhoroshev theorem. A formal statement would look like the following pseudocode, with each axiom replaced by a real definition once the prerequisites are in Mathlib.

[object Promise]

A formal route would assemble: real-analytic function spaces with weighted norms on the complexified action-angle strip; the simultaneous-Diophantine-approximation theorem for $R^{n}$ in the form needed for Lochak's covering; resonance lattices as primitive sublattices of $Z^{n}$ together with the projection onto resonant Fourier modes; the block-wise Birkhoff iteration with optimal-truncation order $N ≍ ϵ^{- 1/ (2 n)}$ ; and the energy-confinement geometry on convex (or quasi-convex) level sets. The Arnold-diffusion side requires variational dynamics (Mather measures), partially hyperbolic invariant manifolds with whisker structure, and shadowing lemmas — distinct infrastructure not implied by the Nekhoroshev side.

Advanced results [Master]

The Nekhoroshev estimate sits inside a broader structural circle of stability results for near-integrable Hamiltonian systems. Five refinements deepen, generalise, or sit beside the basic theorem.

Lochak's simplified proof (1992). The original Nekhoroshev 1977 proof used a complicated geometric construction of resonance manifolds with multiscale step-by-step Birkhoff iteration. Lochak 1992 ^{[Lochak 1992]} replaced the multiscale geometry with a single-step rational covering: every frequency is simultaneously approximated by a rational vector with denominator $\leq N$ , and a Birkhoff normal form is taken on each rational block at truncation order $N ≍ ϵ^{- 1/ (2 n)}$ . The Lochak proof yields the optimal exponent $a = b = 1/ (2 n)$ in the convex case with substantially shorter argument; Pöschel 1993 ^{[Pöschel 1993]} then made the constants explicit and identified the quasi-convex case (convex on each energy level) as sufficient for the same exponents.

Quasi-convex generalisation. Pöschel's quasi-convex hypothesis — $H_{0}$ is convex on each energy level ${H_{0} = E}$ but not necessarily globally convex — is satisfied by any $H_{0}$ whose Legendre transform produces a convex function, including all kinetic-plus-potential Hamiltonians with positive-definite mass matrix on a Riemannian manifold. The exponents $a = b = 1/ (2 n)$ remain optimal in this setting. For non-convex but steep $H_{0}$ , the exponents shrink: Nekhoroshev's original 1977 proof gives exponents depending on the steepness indices $(p_{L})$ of the Hamiltonian, with $a, b \to 0$ as $max p_{L} \to \infty$ . Niederman 2007 ^{[Niederman 2007]} showed that steepness is generic in the prevalence sense — the complement of steep Hamiltonians is shy in the function-space sense of Hunt-Sauer-Yorke — extending Nekhoroshev's applicability to a measure-theoretically typical class of analytic Hamiltonians.

Arnold diffusion in non-convex dimensions $\geq 3$ . Arnold 1964 ^{[Arnold 1964]} gave the first rigorous example of action drift in a near-integrable system: a Hamiltonian on $T^{3} \times R^{3}$ with two small parameters whose trajectories accumulate $O (1)$ action drift over polynomial-in- $1/ μ$ time. The mechanism is heteroclinic chains of whiskered tori — partially hyperbolic invariant tori with stable and unstable manifolds — connected by transversal heteroclinic intersections. A trajectory shadowing such a chain drifts in action by following the unstable whisker of one torus into the stable whisker of the next. The diffusion time exceeds Nekhoroshev's exponential bound; the two results coexist consistently. Mather 1993 ^{[Mather 1993]} gave a variational proof using Aubry-Mather minimisers; Berti-Bolle 2002 and Cheng-Yan 2004 produced rigorous diffusion examples for explicit a-priori-stable mechanical systems. The full Arnold conjecture — that diffusion occurs in a $C^{\infty}$ -residual class of Hamiltonians in dimension $\geq 3$ — remains open in its strong form, though Cheng-Yan and others have established it for various restricted classes.

Solar-system applications and explicit constants. Niederman's prevalence result implies that Hamiltonian systems arising from celestial mechanics are typically Nekhoroshev-stable. Concretely, the planetary three-body problem (sun + two planets) admits a near-integrable formulation in Delaunay variables with $H_{0}$ the Keplerian energy and $H_{1}$ the planet-planet gravitational interaction. Niederman, Locatelli-Giorgilli, and Galgani have computed explicit Nekhoroshev constants for the Earth-Jupiter or sun-Jupiter-Saturn subsystems; the resulting stability times exceed $1 0^{10}$ years for the inner planets, comparable to the age of the solar system. Arnold himself conjectured Lyapunov stability of the planetary motion at the level of perihelion-precession variables over Hubble timescales; this remains an open problem, with current rigorous bounds short of the Hubble time by several orders of magnitude in the realistic-mass regime.

Connection to the global picture of near-integrable dynamics. Nekhoroshev sits at the bottom of a stratified picture of near-integrable behaviour. The KAM theorem 05.09.01 gives a positive-measure Cantor set of permanently invariant tori, and is the strongest statement available on this set. The Nekhoroshev theorem fills the resonance gaps with finite-time uniform stability over an exponentially long window. Arnold diffusion fills the gaps beyond Nekhoroshev with slow but eventual drift. Aubry-Mather theory replaces destroyed KAM tori with Cantor sets carrying invariant measures — the cantori — that retain the rotation-number structure. Together these results stratify phase space into permanent KAM islands, finite-time Nekhoroshev seas, slow Arnold-diffusion threads, and Aubry-Mather skeletons. The bridge between them is the underlying Birkhoff normal form and the Diophantine analysis of the cohomological equation: KAM uses Diophantine input as a hypothesis on a single torus; Nekhoroshev uses simultaneous approximation as a covering construction; Arnold diffusion uses heteroclinic chains; Aubry-Mather uses variational minimisation. All four results invert the same linearised problem — the Lie derivative ${H_{0}, F}$ of the unperturbed Hamiltonian — at different scales and with different Diophantine inputs.

Synthesis. The Nekhoroshev theorem unifies a circle of perturbation-theoretic results by combining the local Birkhoff normal form with a global resonance-lattice covering of action space and the geometric input of energy confinement on convex level sets. Putting these together, the same iterative apparatus that drives KAM produces, with a different orchestration, an exponentially-long uniform stability bound. The bridge from the analytic input — the cohomological equation on each resonance block, controlled by a different Diophantine bound at each scale — to the geometric output — a stratified action-space partition with bounded transversal drift — is the foundational reason near-integrable Hamiltonian systems exhibit the layered stability picture observed in celestial mechanics. KAM, Nekhoroshev, and Arnold diffusion are three faces of the same perturbative apparatus, distinguished by which Diophantine condition is used on which scale and over which timescale.

Full proof set [Master]

Lemma (Dirichlet simultaneous approximation). For every $ω \in R^{n}$ and every integer $N \geq 1$ there exist $p \in Z^{n}$ and $q \in {1, 2, \dots, N^{n}}$ with $$ \bigl| \omega - p/q \bigr| \leq \frac{1}{q N}. $$

Proof. Apply the pigeonhole principle to the $N^{n} + 1$ points ${j ω mod Z^{n} : j = 0, 1, \dots, N^{n}}$ in the unit cube $[0, 1)^{n}$ , partitioned into $N^{n}$ subcubes of side $1/ N$ . Two points $j_{1} ω$ and $j_{2} ω$ lie in the same subcube; setting $q := ∣ j_{1} - j_{2} ∣ \leq N^{n}$ and choosing $p \in Z^{n}$ accordingly gives $∣ q ω - p ∣ \leq 1/ N$ , hence $∣ ω - p / q ∣ \leq 1/ (q N)$ . $□$

Lemma (block-wise cohomological equation). Let $ω^{p / q}$ be a rational vector with $∣ q ∣ \leq N$ and let $K_{p / q} := {k \in Z^{n} : ⟨ k, ω^{p / q} ⟩ = 0}$ . For every $k \in Z^{n} ∖ K_{p / q}$ , $$ |\langle k, \omega^{p/q}\rangle| \geq \frac{1}{q}. $$

Proof. By rationality, $q ω^{p / q} = p \in Z^{n}$ , so $q ⟨ k, ω^{p / q} ⟩ = ⟨ k, p ⟩ \in Z$ . If $k \in / K_{p / q}$ then $⟨ k, p ⟩ \neq = 0$ , hence $∣ ⟨ k, p ⟩ ∣ \geq 1$ , hence $∣ ⟨ k, ω^{p / q} ⟩ ∣ \geq 1/ q$ . $□$

Lemma (Lochak block-wise Birkhoff bound). On the resonance block $B_{p / q} = {I : ∣ ω (I) - ω^{p / q} ∣ < ρ}$ with $ρ ≍ 1/ (q N)$ , there exists a real-analytic canonical change of variables $ϕ_{p / q}$ such that $$ H_\epsilon \circ \phi_{p/q} = H_0(I) + \epsilon \widetilde H_{1, K}(I, \theta_K) + R_{p/q}(I, \theta), $$ where $H_{1, K}$ is supported on $K_{p / q}$ -resonant Fourier modes, $θ_{K}$ are the resonant angles, and the remainder satisfies $∥ R_{p / q} ∥_{B_{p / q}} \leq C exp (- c N)$ for a positive constant $c$ depending on the analytic-norm width.

Proof. Apply the iterated Birkhoff procedure 05.09.03 to remove non- $K_{p / q}$ -resonant Fourier modes. At each iteration step, the cohomological equation $⟨ ω^{p / q}, \partial_{θ} F^{(j)} ⟩ = R^{(j)} - π_{K_{p / q}} R^{(j)}$ is solved by Fourier division using the bound $∣ ⟨ k, ω^{p / q} ⟩ ∣ \geq 1/ q$ for non- $K_{p / q}$ -resonant $k$ . The iteration loses analyticity at a controlled rate; choosing the truncation order $N$ to match the analytic-strip budget gives a remainder of size $exp (- c N)$ via Cauchy bounds. The Diophantine input is uniform on the block — every non-resonant denominator is at least $1/ q \geq 1/ N$ — so the truncation order can be pushed until the geometric series of Birkhoff corrections diverges, which happens at $N ≍ ϵ^{- 1/ (2 n)}$ when $ρ ≍ 1/ N$ matches the analytic-norm step size. $□$

Theorem (Nekhoroshev, convex case). Let $H_{ϵ} = H_{0} + ϵ H_{1}$ with $H_{0}$ convex and real-analytic. There exist $a = b = 1/ (2 n)$ and $c, C > 0$ such that for every initial condition $(I_{0}, θ_{0})$ and every $∣ t ∣ \leq exp (c ϵ^{- 1/ (2 n)})$ , $$ |I(t) - I(0)| \leq C \epsilon^{1/(2n)}. $$

Proof. Cover action space by resonance blocks ${B_{p / q}}_{∣ q ∣ \leq N}$ of width $ρ ≍ 1/ (q N) ≍ ϵ^{1/ (2 n)}$ at truncation order $N ≍ ϵ^{- 1/ (2 n)}$ , using the Dirichlet lemma. On each block, apply the Lochak block-wise Birkhoff bound to obtain a normal-form Hamiltonian whose remainder is of size $exp (- c N) = exp (- c ϵ^{- 1/ (2 n)})$ . The action drift on a single block is bounded by combining two estimates: (i) energy conservation $H_{0} (I (t)) + ϵ H_{1} (I (t), θ (t)) =$ constant, with $H_{0}$ convex, gives $∣ I (t) - I (0) ∣_{H_{0} -radial} \leq C ϵ^{1/2}$ on the convex level set; (ii) the Birkhoff remainder gives $∣ I (t) - I (0) ∣_{H_{0} -tangential} \leq C exp (- c ϵ^{- 1/ (2 n)}) \cdot t \leq C ϵ^{1/ (2 n)}$ for $∣ t ∣ \leq exp (c ϵ^{- 1/ (2 n)})$ . The combined bound is $C ϵ^{1/ (2 n)}$ — the larger of the two contributions. As long as the trajectory remains in a single block, the bound holds; transitions between blocks happen when $∣ ω (I (t)) - ω^{p / q} ∣ > ρ$ , which by the radial drift bound $C ϵ^{1/2}$ happens only after the trajectory has moved by $O (ρ)$ in action space, the same scale as the block width, so a single block-bounded estimate suffices on each segment. The total drift, summed over the polynomially-many block transitions during the exponential time, is bounded by $C ϵ^{1/ (2 n)}$ . $□$

Theorem (Pöschel quasi-convex generalisation). The same conclusion holds with $H_{0}$ quasi-convex (convex on each energy level) in place of globally convex, with the same exponents $a = b = 1/ (2 n)$ ^{[Pöschel 1993]}. The proof uses the level-set convexity in place of the Hessian sign, with the same block covering and block-wise Birkhoff bound; energy confinement now traps the trajectory on a single convex level set rather than within a global convex region.

Theorem (Nekhoroshev, steep case, statement only). For $H_{0}$ steep at $I^ $w i t h s t ee p n ess in d i ces$ {p_L}_{L \subset \mathbb{R}^n} $, t h er ee x i s t e x p o n e n t s$ a, b > 0 $d e p e n d in g o n l y o n$ n $an d t h e ma x ima l s t ee p n ess in d e x$ \max_L p_L$, such that the same stability bound holds with these exponents.* The proof uses Nekhoroshev's original 1977 multiscale construction; explicit formulas for the exponents in terms of the steepness profile are given in Nekhoroshev 1979 ^{[Nekhoroshev 1979]}.

Connections [Master]

KAM theorem 05.09.01 — the apex perturbation result for Diophantine initial conditions; KAM gives permanent stability on a positive-measure Cantor set, complementing Nekhoroshev's uniform finite-time stability. Both rest on the cohomological equation ${H_{0}, F} = R - π R$ inverted via Diophantine analysis; KAM uses a single Diophantine torus, Nekhoroshev uses a covering by rational lattices.
Birkhoff normal form 05.09.03 — the local perturbative apparatus driving each block-wise step in Nekhoroshev's proof; the Lochak-Pöschel argument applies the Birkhoff iteration on each resonance block at optimal truncation order $N ≍ ϵ^{- 1/ (2 n)}$ , with the truncation balanced against the analytic-norm budget on the strip.
Action-angle coordinates 05.02.04 — the canonical chart in which the Nekhoroshev partition of action space is described; the action variables are the conserved quantities whose drift is bounded, and the angle variables are the integration variable for the cohomological equation on each block.
Adiabatic invariants 05.09.02 — the time-dependent perturbation analogue; adiabatic invariants are conserved up to $O (ϵ)$ over time $ϵ^{- 1}$ for slowly-varying parameters, while Nekhoroshev gives stability over exponentially longer time for static perturbations. The two results share the averaging principle as common foundation; Neishtadt's exponential-precision sharpening of adiabatic invariance under analyticity is the natural counterpart of the Nekhoroshev exponential time bound.
Integrable system 05.02.03 — the unperturbed half of the Nekhoroshev setup; Liouville-Arnold produces the action-angle foliation that Nekhoroshev tries to preserve in approximate form, with action drift bounded uniformly over an exponentially long time.
Symplectic manifold 05.01.02 — the ambient category in which the perturbation problem lives; symplecticity of the iterated transformations on each resonance block is what makes the Birkhoff iteration well-posed.
Generating functions 05.05.03 — the technical mechanism producing each block-wise canonical change of variables; the auxiliary functions $F^{(j)}$ at each Birkhoff step generate symplectomorphisms via their Hamiltonian flow.
Poisson bracket 05.02.02 — the cohomological equation ${H_{0}, F} = R - π_{K} R$ on each resonance block is the fundamental linear operator of the iteration, with the resonance projection $π_{K}$ replacing KAM's pointwise angle-average.
Arnold conjecture / Floer homology [05.08.01, 05.08.02] — the long-time dynamics beyond the Nekhoroshev window are described by Arnold-diffusion mechanisms (heteroclinic chains of whiskered tori) whose rigorous analysis uses variational and Floer-theoretic methods; Nekhoroshev sets the timescale on which this slower drift can become observable.

The bridge from the analytic input (the simultaneous-Diophantine-approximation theorem and the cohomological equation on each rational block) to the geometric output (an exponentially-long uniform stability bound across all initial conditions) is the foundational reason Nekhoroshev unifies the stability theory of the resonance gaps with the KAM theory of the Diophantine Cantor set. Putting these together, the same Birkhoff-iteration apparatus that drives KAM produces, with a different orchestration of resonance lattices, the complementary Nekhoroshev stability picture; together with Arnold diffusion they describe the layered behaviour of near-integrable Hamiltonian dynamics from the shortest perturbation timescale to the longest.

Historical & philosophical context [Master]

Nikolai Nekhoroshev introduced the exponential-stability theorem in his 1977 paper An exponential estimate of the time of stability of nearly-integrable Hamiltonian systems in Russian Math. Surveys 32 ^{[Nekhoroshev 1977]}. The paper appeared a decade after KAM had been established as the standard tool for near-integrable stability analysis, and Nekhoroshev's contribution was to fill the explanatory gap: KAM produced invariant tori on a Cantor set of full measure but said nothing about the resonance gaps, a mathematically unsatisfactory situation given that physical perturbation problems do not select Diophantine initial conditions. Nekhoroshev introduced the steepness condition and the resonance-block partition to handle the gaps directly, producing a uniform stability bound applying to every initial condition over an exponentially long timescale. The original 1977 announcement was followed by a more detailed 1979 paper ^{[Nekhoroshev 1979]} containing the full steepness analysis. Vladimir Arnold endorsed the result in §54 of Mathematical Methods of Classical Mechanics and in the Arnold-Kozlov-Neishtadt encyclopaedic treatment ^{[Arnold-Kozlov-Neishtadt]}.

Pierre Lochak's 1992 paper Canonical perturbation theory via simultaneous approximation ^{[Lochak 1992]} reorganised the proof using Dirichlet's simultaneous-approximation theorem to produce a single covering of action space by rational blocks, one block per rational frequency vector with bounded denominator. The Lochak proof yields the optimal exponents $a = b = 1/ (2 n)$ in the convex case with substantially shorter and more transparent argument than Nekhoroshev's original construction. Jürgen Pöschel's 1993 paper Nekhoroshev estimates for quasi-convex Hamiltonian systems ^{[Pöschel 1993]} sharpened the constants and identified quasi-convexity as sufficient hypothesis. The Lochak-Pöschel form is the canonical modern statement of the convex Nekhoroshev theorem and is the statement formalised in the Arnold-Kozlov-Neishtadt encyclopaedia.

The applicability of the Nekhoroshev theorem to physical systems was extended by Laurent Niederman's 2007 paper Prevalence of exponential stability among nearly integrable Hamiltonian systems ^{[Niederman 2007]}, which showed that the steepness condition is generic in the prevalence sense — its complement is shy in the function-space sense of Hunt-Sauer-Yorke — so Nekhoroshev-stability is the typical rather than exceptional behaviour. This extended the practical reach of the theorem from a specific class of convex or quasi-convex Hamiltonians to almost every analytic near-integrable system.

The Arnold-diffusion counterpoint originated in Arnold's 1964 paper Instability of dynamical systems with several degrees of freedom in Doklady Akad. Nauk SSSR 156 ^{[Arnold 1964]}, which constructed the first explicit example of action drift in a near-integrable Hamiltonian via heteroclinic chains of whiskered tori. The mechanism was made fully rigorous by Mather in 1993 ^{[Mather 1993]} using variational dynamics, by Berti-Bolle in 2002, and by Cheng-Yan in 2004. Arnold's 1964 example showed that Nekhoroshev's exponential stability is essentially optimal in the steep non-convex setting; together with KAM and Nekhoroshev, Arnold diffusion completes the modern stratified picture of near-integrable Hamiltonian dynamics.

Bibliography [Master]

[object Promise]

Prerequisites

05.09.01
05.02.04
05.09.03

Tier anchors

beginner: Strogatz *Nonlinear Dynamics and Chaos* §6 informal; Arnold *Mathematical Methods of Classical Mechanics* §54 informal
intermediate: Arnold-Kozlov-Neishtadt *Mathematical Aspects of Classical and Celestial Mechanics* Ch. 7
master: Nekhoroshev 1977 *An exponential estimate of the time of stability of nearly-integrable Hamiltonian systems* (Russian Math. Surveys 32, originator); Lochak 1992 *Canonical perturbation theory via simultaneous approximation* (Russian Math. Surveys 47); Pöschel 1993 *Nekhoroshev estimates for quasi-convex Hamiltonian systems*; Niederman 2007 *Prevalence of exponential stability among nearly integrable Hamiltonian systems*; Arnold-Kozlov-Neishtadt Ch. 7

References

TODO_REF
Nekhoroshev 1977 — An exponential estimate of the time of stability of nearly-integrable Hamiltonian systems · Russian Math. Surveys 32, originator paper for the exponential-stability theorem
TODO_REF
Nekhoroshev 1979 — An exponential estimate of the time of stability of nearly-integrable Hamiltonian systems II · Trudy Sem. Petrovsk. 5, the steepness analysis and the proof of the second part
TODO_REF
Lochak 1992 — Canonical perturbation theory via simultaneous approximation · Russian Math. Surveys 47, simplified proof in the convex case via simultaneous Diophantine approximation
TODO_REF
Pöschel 1993 — Nekhoroshev estimates for quasi-convex Hamiltonian systems · Math. Z. 213, optimal exponents $a = b = 1/(2n)$ in the quasi-convex setting
TODO_REF
Niederman 2007 — Prevalence of exponential stability among nearly integrable Hamiltonian systems · Ergodic Theory Dynam. Systems 27, genericity of Nekhoroshev-stable Hamiltonians
TODO_REF
Arnold 1964 — Instability of dynamical systems with several degrees of freedom · Doklady Akad. Nauk SSSR 156, the original Arnold-diffusion example
TODO_REF
Arnold — Mathematical Methods of Classical Mechanics · §54 + Appendix 8, perturbation theory and stability of action variables
TODO_REF
Arnold, Kozlov, Neishtadt — Mathematical Aspects of Classical and Celestial Mechanics · Ch. 7, Nekhoroshev estimates and Arnold diffusion
TODO_REF
Mather 1993 — Variational construction of connecting orbits · Annales de l'Institut Fourier 43, variational approach to Arnold diffusion
TODO_REF
Strogatz — Nonlinear Dynamics and Chaos · §6 informal phase-portrait picture for resonant Hamiltonian systems

Reviewer

TBD

Estimated time

beginner: 18m
intermediate: 45m
master: 95m