Hamiltonian Perturbation Theory: Averaging and Adiabatic Invariants

Intuition Beginner

A grandfather clock's pendulum slowly loses energy to friction. As it loses energy, the amplitude decreases and the frequency increases slightly. But the ratio of energy to frequency -- a quantity called the "action" -- stays almost constant. Even as the pendulum's energy changes significantly over many swings, the action barely changes. This is an adiabatic invariant: a quantity that is conserved when the system's parameters change slowly compared to the oscillation period.

More generally, consider a mechanical system whose Hamiltonian depends on a parameter that changes slowly in time. A pendulum whose length is slowly shortened. A charged particle spiralling in a magnetic field that slowly increases. A planet in an orbit that slowly shrinks due to tidal friction. In all these cases, the action variable $J=\oint pdq/(2\pi )$ (the area enclosed by the orbit in phase space, divided by $2\pi$) is approximately conserved. The approximation is excellent: the error is exponentially small in the ratio of the oscillation frequency to the rate of parameter change.

The mathematical framework for studying such slowly-perturbed systems is Hamiltonian perturbation theory. The unperturbed system is integrable (soluble by action-angle variables 09.06.01). The perturbation breaks integrability, but if it is small, the system is nearly integrable: most of phase space is still filled with invariant tori that are only slightly deformed from their unperturbed shapes. The action variables on these tori are approximately conserved, and the deviation from exact conservation can be computed order by order in the perturbation strength.

The central idea is the distinction between slow variables and fast variables. In a near-integrable system, the angles $\theta$ change rapidly (at a rate set by the unperturbed frequencies $\omega$), while the actions $J$ drift slowly (at a rate proportional to the small perturbation strength $ϵ$). Over one fast oscillation, the slow variables barely change. Over many fast oscillations, the slow variables accumulate a net drift. The averaging principle says that this slow drift is well captured by replacing the fast oscillatory forcing with its average over one period. The averaged system is simpler (the fast dependence is gone) but retains the correct slow-time dynamics.

This separation of timescales is what makes perturbation theory work. The fast oscillations cancel out over long times (their net effect is small), while the slow secular changes accumulate. The averaging method extracts the secular drift while discarding the rapid oscillations that average to zero.

Visual Beginner

Picture a phase-space portrait of a harmonic oscillator: nested ellipses (level sets of $H$). Now slowly change the spring constant $k$. The ellipses deform and shift, but the enclosed area (the action $J$) stays nearly constant. As $k$ increases, the ellipses shrink in the $q$-direction but expand in the $p$-direction, preserving area. This is the phase-space manifestation of the adiabatic theorem.

For a faster change in $k$, the area is less well conserved: the orbit does not have time to adjust smoothly, and the phase-space area fluctuates. The key parameter is the ratio $\tau /T$, where $\tau$ is the timescale of the parameter change and $T$ is the oscillation period. When $\tau /T\gg 1$ (many oscillations during one parameter change), the action is an excellent invariant.

Now picture the phase space of a near-integrable system with two degrees of freedom. The unperturbed system has nested invariant tori. The perturbation distorts most tori slightly, but near resonances (where the frequency ratio is a rational number), the tori break up and are replaced by chains of islands and chaotic layers. The KAM theorem guarantees that sufficiently irrational tori survive, while rational ones are destroyed.

Worked example Beginner

A simple pendulum of length $l$ and mass $m$ oscillates with small amplitude. The string is slowly pulled through the support, shortening the length at rate $dl/dt=-ϵ$ where $ϵ$ is much smaller than $\sqrt{g/l}$. How does the amplitude change?

The action is $J=E/\omega$ where $E$ is the energy and $\omega =\sqrt{g/l}$ is the frequency. Adiabatic invariance says $J$ is approximately constant, so $E/\omega$ stays constant, i.e., $E\sqrt{l}$ stays constant.

The energy is $E=\frac{1}{2}m{\omega }^2{A}^2$ where $A$ is the amplitude. So $m{\omega }^2{A}^2\sqrt{l}/2$ is constant. With $\omega =\sqrt{g/l}$, this gives $mg{A}^2/(2\sqrt{l})$ is constant, hence $A\propto l^{1/4}$.

As the pendulum shortens, the amplitude decreases but only as $l^{1/4}$. The energy increases (as $l^{-1/2}$) because the tension does work pulling the string through. The frequency increases (as $l^{-1/2}$) because the pendulum is shorter. The action $J=E/\omega$ stays constant. This result, first derived by Rayleigh, is one of the earliest examples of an adiabatic invariant.

Consider the numbers. Start with $l=1$ m, $A=0.1$ m, $g=9.8$ m/s${}^2$. Then $\omega =3.13$ rad/s, $E=0.049$ J, and $J=0.0157$ J s. Now shorten to $l=0.5$ m. The new amplitude is $A=0.1\times (0.5{)}^{1/4}=0.084$ m. The new frequency is $\omega =4.43$ rad/s. The new energy is $E=J\omega =0.0157\times 4.43=0.0695$ J. The energy increased by 42% while the length halved -- and the action stayed constant to high accuracy.

Check your understanding Beginner

Formal definition Intermediate+

Near-integrable Hamiltonian systems

A Hamiltonian system with $n$ degrees of freedom is integrable 09.06.01 when it possesses $n$ independent constants of motion in involution. By the Liouville-Arnold theorem, the motion is then confined to $n$-dimensional tori parameterised by action-angle variables $(\mathbf{\theta },\mathbf{J})$, with Hamilton's equations taking the simple form ${\dot{\theta }}^{\alpha }={\omega }^{\alpha }(\mathbf{J})$, ${\dot{J}}_{\alpha }=0$.

A near-integrable system is one whose Hamiltonian differs from an integrable one by a small perturbation:

$$H(\mathbf{\theta },\mathbf{J})=H_0(\mathbf{J})+ϵH_1(\mathbf{\theta },\mathbf{J}),$$

where $ϵ\ll 1$ is the perturbation parameter and $H_1$ is $2\pi$-periodic in each angle ${\theta }^{\alpha }$. The full equations of motion are:

$${\dot{\theta }}^{\alpha }={\omega }^{\alpha }(\mathbf{J})+ϵ\frac{\partial H_1}{\partial J_{\alpha }},{\dot{J}}_{\alpha }=-ϵ\frac{\partial H_1}{\partial {\theta }^{\alpha }}.$$

The actions are no longer constant: they drift slowly, at rate $O(ϵ)$. The question is whether this drift remains bounded over long times.

Canonical perturbation theory

The strategy is to find a near-identity canonical transformation $(\mathbf{\theta },\mathbf{J})\to (\mathbf{\varphi },\mathbf{I})$ that eliminates the angle dependence of the perturbation order by order. The generating function is expanded as:

$$S(\mathbf{\theta },\mathbf{I},ϵ)=\mathbf{\theta }\cdot \mathbf{I}+ϵS_1(\mathbf{\theta },\mathbf{I})+{ϵ}^2{S}_2(\mathbf{\theta },\mathbf{I})+\cdots$$

The transformation gives $\mathbf{J}=\mathbf{I}+ϵ\partial S_1/\partial \mathbf{\theta }+O({ϵ}^2)$ and $\mathbf{\varphi }=\mathbf{\theta }+ϵ\partial S_1/\partial \mathbf{I}+O({ϵ}^2)$. Substituting into $H$ and collecting powers of $ϵ$:

• $O(1)$: $K_0(\mathbf{I})=H_0(\mathbf{I})$.
• $O(ϵ)$: $K_1(\mathbf{I})={\omega }^{\alpha }(\mathbf{I})\partial S_1/\partial {\theta }^{\alpha }+H_1(\mathbf{\theta },\mathbf{I})$.

The $O(ϵ)$ balance yields the homological equation:

$$\boxed{{\omega }^{\alpha }(\mathbf{I})\frac{\partial S_1}{\partial {\theta }^{\alpha }}+H_1(\mathbf{\theta },\mathbf{I})=K_1(\mathbf{I})}$$

Decompose $H_1$ into its Fourier series $H_1={\sum }_{\mathbf{k}}{h}_{\mathbf{k}}(\mathbf{I})e^{i\mathbf{k}\cdot \mathbf{\theta }}$. Setting $K_1=h_0(\mathbf{I})={\bar{H}}_1(\mathbf{I})$ (the average over all angles), the equation for $S_1$ becomes:

$$i(\mathbf{k}\cdot \mathbf{\omega })s_{\mathbf{k}}+h_{\mathbf{k}}=0\text{for }\mathbf{k}\ne 0,$$

giving $s_{\mathbf{k}}=-h_{\mathbf{k}}/(i\mathbf{k}\cdot \mathbf{\omega })$. This is solvable provided $\mathbf{k}\cdot \mathbf{\omega }\ne 0$ for all non-zero integer vectors $\mathbf{k}$: the non-resonance condition.

Secular terms and the small-divisor problem

When $\mathbf{k}\cdot \mathbf{\omega }=0$ for some non-zero integer vector $\mathbf{k}$, the system is resonant and the corresponding Fourier mode cannot be eliminated. The generating function $S_1$ has a singularity (division by zero), and the perturbation expansion breaks down.

Even away from exact resonances, the small divisors $|\mathbf{k}\cdot \mathbf{\omega }|$ for large $|\mathbf{k}|$ make the coefficients $s_{\mathbf{k}}$ grow, and the perturbation series may diverge. Poincare showed that the series is generally divergent: the small divisors accumulate and prevent convergence. This is the small-divisor problem, the central obstacle to perturbation theory.

Naive perturbation expansions also produce secular terms: terms that grow without bound in time (e.g., $ϵt\cos \omega t$), making the expansion useless for long times. The canonical perturbation method avoids secular terms at each fixed order by absorbing them into frequency corrections. But the divergence of the full series means that this procedure cannot be continued to arbitrary order.

The averaging principle

The averaging principle replaces the oscillatory perturbation $H_1(\mathbf{\theta },\mathbf{J})$ with its average over the fast angles:

$${\bar{H}}_1(\mathbf{J})=\frac{1}{(2\pi {)}^n}{\int }_0^{2\pi }\cdots {\int }_0^{2\pi }{H}_1(\mathbf{\theta },\mathbf{J})d{\theta }^1\cdots d{\theta }^n=\sum _{\mathbf{k}}{h}_{\mathbf{k}}(\mathbf{J}){\delta }_{\mathbf{k},0}.$$

The averaged equations ${\dot{I}}_{\alpha }=-ϵ\partial {\bar{H}}_1/\partial {\theta }^{\alpha }=0$ predict exact conservation of actions, since ${\bar{H}}_1$ has no angle dependence. The averaging theorem makes this precise:

Theorem (Averaging). Let $H=H_0(\mathbf{J})+ϵH_1(\mathbf{\theta },\mathbf{J})$ be analytic in a domain $D$. Suppose the unperturbed frequencies satisfy no resonances of order $\le K$ in $D$ (i.e., $|\mathbf{k}\cdot \mathbf{\omega }|\ge c>0$ for all integer vectors $\mathbf{k}$ with $0<|\mathbf{k}|\le K$). Then the averaged system approximates the true system with error $O(ϵ)$ over times $O(1/ϵ)$. That is, $|J_{\alpha }(t)-I_{\alpha }(t)|=O(ϵ)$ for $0\le t\le C/ϵ$.

The proof constructs the first-order normal form via the generating function $S_1$ that eliminates the oscillatory part of $H_1$. The remaining Hamiltonian $K_0(\mathbf{I})+ϵK_1(\mathbf{I})+O({ϵ}^2)$ has actions that are constant to $O(ϵ)$ over the stated time interval.

Adiabatic invariants

A parameter-dependent Hamiltonian $H(q,p,\lambda (t))$ depends on a slowly varying parameter $\lambda$ with $d\lambda /dt=ϵ\ll 1$. For fixed $\lambda$, the system is integrable with action $J(\lambda )=\oint pdq/(2\pi )$. The adiabatic theorem states:

Theorem (Adiabatic invariance). If the parameter $\lambda (t)$ changes smoothly from ${\lambda }_0$ to ${\lambda }_1$ over a time $\tau =O(1/ϵ)$, then $|J(\tau )-J(0)|=O(ϵ)$. For analytic systems, $|J(\tau )-J(0)|=O(e^{-c/ϵ})$ for some $c>0$.

The exponential accuracy for analytic systems means that the action is conserved to extraordinary precision even for moderately slow parameter changes. The constant $c$ depends on the analyticity domain of the Hamiltonian.

The physical content is simple: if you change a parameter slowly, the system adjusts quasi-statically, always remaining close to an orbit of the instantaneous Hamiltonian. The action (phase-space area) is preserved during this adjustment because the Hamiltonian flow is area-preserving at each instant. The error comes from the finite (though small) rate of change, which pushes the system slightly off the instantaneous orbit.

Example: Fermi acceleration

A particle bounces elastically between two walls separated by distance $L$. The walls slowly move together at rate $dL/dt=-ϵ$. The action is $J=mvL/\pi$ (where $v$ is the speed between bounces). Adiabatic invariance gives $mvL/\pi$ is constant, so $v\propto 1/L$. As the walls converge, the particle speeds up. Each bounce off an approaching wall adds a small increment to the speed; over many bounces, these increments accumulate. This is the mechanism behind Fermi acceleration of cosmic rays: charged particles bounce between approaching magnetic shock fronts and gain energy with each reflection. Enrico Fermi proposed this mechanism in 1949 to explain the high energies of cosmic rays.

The adiabatic invariant is the phase-space area: the orbit in the $(q,p)$-plane is a rectangle of width $2L$ and height $2mv$. The area $4mvL$ is conserved (divided by $2\pi$ gives the action $J=2mvL/\pi$). As $L$ decreases, $v$ must increase to keep the area constant.

Example: Particle in a magnetic bottle

A charged particle gyrates in a magnetic field $\mathbf{B}$ with cyclotron frequency ${\omega }_c=eB/(mc)$. The magnetic moment $\mu =m{v}_{\perp }^2/(2B)$ (where $v_{\perp }$ is the velocity perpendicular to $\mathbf{B}$) is an adiabatic invariant when $B$ varies slowly over one gyro-orbit.

The action for the cyclotron motion is $J_{\perp }=m{v}_{\perp }^2/(2{\omega }_c)=mc\mu /e$. Adiabatic invariance of $J_{\perp }$ is equivalent to conservation of $\mu$. As the particle moves into a region of stronger $B$, conservation of $\mu$ forces $v_{\perp }^2$ to increase proportionally to $B$, while conservation of total kinetic energy $E=\frac{1}{2}m(v_{\perp }^2+v_{\parallel }^2)$ forces the parallel velocity $v_{\parallel }$ to decrease. At the mirror point, $v_{\parallel }=0$ and the particle reflects. A magnetic bottle traps particles between two regions of strong field. This is the principle behind plasma confinement in magnetic fusion devices and the trapping of radiation belts around planets.

Key derivation Intermediate+

Derivation: First-order normal form and elimination of fast angles.

Seek a generating function $S(\mathbf{\theta },\mathbf{I},ϵ)=\mathbf{\theta }\cdot \mathbf{I}+ϵS_1(\mathbf{\theta },\mathbf{I})$ that transforms $H=H_0(\mathbf{J})+ϵH_1(\mathbf{\theta },\mathbf{J})$ to $K=K_0(\mathbf{I})+ϵK_1(\mathbf{I})+O({ϵ}^2)$ where $K_1$ depends only on $\mathbf{I}$.

The transformation gives $\mathbf{J}=\partial S/\partial \mathbf{\theta }=\mathbf{I}+ϵ\partial S_1/\partial \mathbf{\theta }$ and $\mathbf{\varphi }=\partial S/\partial \mathbf{I}=\mathbf{\theta }+ϵ\partial S_1/\partial \mathbf{I}$. The new Hamiltonian $K(\mathbf{\varphi },\mathbf{I})=H(\mathbf{\theta }(\mathbf{\varphi },\mathbf{I}),\mathbf{J}(\mathbf{\varphi },\mathbf{I}))$ is expanded to $O(ϵ)$ by Taylor expanding $H_0(\mathbf{J})$ around $\mathbf{I}$:

$$H_0(\mathbf{J})=H_0(\mathbf{I})+\frac{\partial H_0}{\partial J_{\alpha }}{|}_{\mathbf{I}}\cdot ϵ\frac{\partial S_1}{\partial {\theta }^{\alpha }}+O({ϵ}^2)=H_0(\mathbf{I})+ϵ{\omega }^{\alpha }(\mathbf{I})\frac{\partial S_1}{\partial {\theta }^{\alpha }}+O({ϵ}^2).$$

Collecting the $O(ϵ)$ terms:

$$K_1(\mathbf{I})={\omega }^{\alpha }(\mathbf{I})\frac{\partial S_1}{\partial {\theta }^{\alpha }}+H_1(\mathbf{\theta },\mathbf{I}).$$

Expand $H_1$ and $S_1$ in Fourier series: $H_1={\sum }_{\mathbf{k}}{h}_{\mathbf{k}}(\mathbf{I})e^{i\mathbf{k}\cdot \mathbf{\theta }}$, $S_1={\sum }_{\mathbf{k}}{s}_{\mathbf{k}}(\mathbf{I})e^{i\mathbf{k}\cdot \mathbf{\theta }}$. The homological equation becomes, for each $\mathbf{k}\ne 0$:

$$i(\mathbf{k}\cdot \mathbf{\omega }(\mathbf{I}))s_{\mathbf{k}}+h_{\mathbf{k}}=0⟹s_{\mathbf{k}}=\frac{-h_{\mathbf{k}}}{i\mathbf{k}\cdot \mathbf{\omega }(\mathbf{I})}.$$

Setting $K_1=h_0={\bar{H}}_1$ (the angle-average of $H_1$) completes the construction. The new Hamiltonian is:

$$K(\mathbf{\varphi },\mathbf{I})=H_0(\mathbf{I})+ϵ{\bar{H}}_1(\mathbf{I})+O({ϵ}^2).$$

Since $K$ is independent of $\mathbf{\varphi }$ to this order, the actions $\mathbf{I}$ are constant up to $O({ϵ}^2)$ corrections, and the angles advance with the corrected frequency $\mathbf{\omega }+ϵ\partial {\bar{H}}_1/\partial \mathbf{I}$.

The non-resonance condition is the key requirement. At resonances ($\mathbf{k}\cdot \mathbf{\omega }=0$ for some $\mathbf{k}$), the denominator vanishes and the transformation fails. The resonant Fourier modes remain in the normal form and drive slow exchange of action between the resonant degrees of freedom.

Bridge. This derivation builds toward the KAM theorem 09.08.01, which addresses what happens when the perturbation is treated globally rather than order by order: the surviving tori are those whose frequencies satisfy a Diophantine condition $|\mathbf{k}\cdot \mathbf{\omega }|\ge c/|\mathbf{k}{|}^{\tau }$ for all $\mathbf{k}$. The averaging method is the first step in a hierarchy of normal forms that culminate in the KAM theorem for eternal stability of sufficiently irrational tori.

Exercises Intermediate+

Lean formalization Intermediate+

The averaging theorem and adiabatic invariants are absent from Mathlib. The specific ingredients needed are:

1. A definition of near-integrable Hamiltonian systems with action-angle variables.
2. The Fourier decomposition of the perturbation and the formal solution of the homological equation.
3. The non-resonance condition as a hypothesis, and the construction of the normal-form transformation.
4. The averaging theorem: a bound on $|J(t)-I(t)|$ over times $O(1/ϵ)$.
5. The adiabatic theorem: approximate conservation of the action under slow parameter variation.
6. The Hannay angle as a connection 1-form on parameter space.

Items (1)--(3) are within reach given Mathlib's existing ODE and symplectic infrastructure. Item (4) requires Gr"onwall-type inequalities and careful estimates of the remainder terms. Items (5)--(6) require parameter-dependent Hamiltonian flows and differential-geometric machinery for connections on fibre bundles. Formalising Nekhoroshev estimates or Lie transform methods would require substantial new infrastructure for exponentially small error bounds and iterative constructions.

Advanced results Master

Resonances and resonance layers

The normal-form transformation fails when $\mathbf{k}\cdot \mathbf{\omega }(\mathbf{J})=0$ for some integer vector $\mathbf{k}\ne 0$. At such resonances, the generating function $S_1$ is singular and the perturbation cannot be fully eliminated. The remaining resonant terms in the normal form couple the actions at the resonant frequency ratio, producing slow exchange of energy between modes.

For $n$ degrees of freedom, the condition $\mathbf{k}\cdot \mathbf{\omega }=0$ defines a hypersurface in action space (the resonance surface). The density of resonance surfaces increases with the order $|\mathbf{k}|$: high-order resonances fill action space densely. Between resonance surfaces, the normal form is well-defined and the actions are approximately conserved. Near resonance surfaces, the dynamics is governed by a resonant normal form that may include pendulum-like phase portraits with separatrices and chaotic layers.

The width of a resonance of order $|\mathbf{k}|$ scales as ${ϵ}^{1/2|\mathbf{k}|}$. Low-order resonances ($|\mathbf{k}|=1,2,3$) are wide and create large chaotic zones. High-order resonances are narrow and have negligible effect. The phase space is a hierarchical structure: large regions of near-integrable motion separated by thin chaotic layers around low-order resonances, with thinner sublayers around higher-order resonances nested within.

The Kolmogorov normal form and KAM theorem preview

Kolmogorov's insight (1954) was to fix the frequency vector ${\mathbf{\omega }}_0$ rather than the action, and construct a sequence of canonical transformations that converge to a torus with exactly the prescribed frequency. The Kolmogorov normal form is:

$$K(\mathbf{\varphi },\mathbf{I})={\mathbf{\omega }}_0\cdot \mathbf{I}+O(|\mathbf{I}{|}^2),$$

where the torus $\mathbf{I}=0$ in the new coordinates is invariant under the flow, and the motion on it is quasiperiodic with frequency ${\mathbf{\omega }}_0$. The construction is a Newton-type iteration (quadratic convergence) rather than a power-series expansion, which is why it avoids the small-divisor divergence.

The Diophantine condition $|\mathbf{k}\cdot {\mathbf{\omega }}_0|\ge c/|\mathbf{k}{|}^{\tau }$ for all $\mathbf{k}\ne 0$ ensures that the frequencies are "sufficiently irrational." This is a stronger condition than mere non-resonance: it requires the frequencies to be bounded away from all rational linear combinations at a controlled rate. The set of Diophantine vectors has full Lebesgue measure in ${\mathbb{R}}^n$, meaning that "almost all" frequency vectors satisfy the condition.

The KAM theorem 38.07.01 then states that for a near-integrable analytic Hamiltonian $H=H_0(\mathbf{J})+ϵH_1(\mathbf{\theta },\mathbf{J})$ with non-degenerate $H_0$ (the frequency map $\mathbf{J}\mapsto \mathbf{\omega }(\mathbf{J})$ has full rank), there exists ${ϵ}_0>0$ such that for $|ϵ|<{ϵ}_0$, a set of invariant tori of positive measure survives the perturbation. The surviving tori carry quasiperiodic motion with Diophantine frequencies, and their total measure approaches full measure as $ϵ\to 0$. The tori that are destroyed are those near rational frequency ratios (low-order resonances).

Nekhoroshev estimates

For systems where KAM does not apply (e.g., the perturbation exceeds the KAM threshold, or $H_0$ is degenerate), Nekhoroshev's theorem (1977) provides long-time stability bounds:

Theorem (Nekhoroshev). Let $H=H_0(\mathbf{J})+ϵH_1(\mathbf{\theta },\mathbf{J})$ be analytic, with $H_0$ satisfying a steepness condition (e.g., $H_0$ is convex or quasi-convex). Then there exist constants $a,b,c,C>0$ such that:

$$|\mathbf{J}(t)-\mathbf{J}(0)|\le C{ϵ}^a\text{for }|t|\le T_0{e}^{c/{ϵ}^b}.$$

The exponents $a,b$ depend on the number of degrees of freedom $n$: for convex $H_0$, one can take $a=1/(2n)$ and $b=1/(2n)$. The stability time is exponentially long in $1/ϵ$: not forever, but for times that are effectively infinite for practical purposes.

Nekhoroshev's theorem explains the long-term stability of the solar system on timescales of billions of years, despite the chaotic dynamics demonstrated by Lyapunov exponent calculations. The steepness condition on $H_0$ is generic (satisfied by most physically relevant Hamiltonians), so the theorem applies broadly. The result is sharp in the sense that Arnold diffusion (slow drift along resonance channels) eventually carries trajectories away from their initial tori, but only on timescales exponentially long in $1/ϵ$.

Lie transform perturbation theory

The canonical perturbation theory using generating functions (the Poincare-von Zeipel method) has a drawback: the transformation is defined implicitly (the new and old variables are mixed in the generating function), and computing higher-order terms requires increasingly complicated algebra. Lie transform perturbation theory (Hori 1966, Deprit 1969, Kamel 1969) provides a cleaner alternative.

A Lie transform is a canonical transformation generated by a Hamiltonian-like function $W(\mathbf{\theta },\mathbf{J},ϵ)$ via the operator $\exp (L_W)$, where $L_{W}f=\{f,W\}$ is the Lie derivative (the Poisson bracket with $W$). The transformed Hamiltonian is:

$$K=\exp (L_W)H=H+\{H,W\}+\frac{1}{2}\{\{H,W\},W\}+\cdots .$$

Expand $W=ϵW_1+{ϵ}^2{W}_2+\cdots$ and collect powers of $ϵ$. At each order, a homological equation determines $W_k$ and $K_k$:

• $O(ϵ)$: $\{H_0,W_1\}+H_1=K_1$ (same as the generating-function method).
• $O({ϵ}^2)$: $\{H_0,W_2\}+\frac{1}{2}\{\{H_0,W_1\},W_1\}+\{H_1,W_1\}=K_2$.

The advantage is that the transformation is explicit (the Lie series gives the new variables directly as functions of the old ones), and the algebraic structure is more transparent: each order involves only Poisson brackets of previously computed functions. The Lie transform is the basis of modern computer-algebra implementations of perturbation theory (e.g., in celestial mechanics and beam physics).

Adiabatic invariants in plasma physics

The adiabatic invariance of the magnetic moment $\mu =m{v}_{\perp }^2/(2B)$ is the workhorse of plasma confinement theory. In a tokamak, charged particles spiral around magnetic field lines with cyclotron frequency ${\omega }_c=eB/(mc)$. If the magnetic field varies slowly over one gyro-orbit (i.e., the Larmor radius ${\rho }_L=m{v}_{\perp }/(eB)$ is small compared to the scale length of $B$), then $\mu$ is conserved to exponential accuracy.

A magnetised particle has three adiabatic invariants corresponding to three timescales:

1. Magnetic moment $\mu =m{v}_{\perp }^2/(2B)$: conserved over the gyro-period $T_c=2\pi /{\omega }_c$.
2. Bounce action $J_{\parallel }=\oint m{v}_{\parallel }dl$: conserved over the bounce period between mirror points.
3. Drift action (magnetic flux) $\Phi =\oint \psi d\alpha$: conserved over the drift period around the torus.

Each invariant is associated with a periodic motion on a successively slower timescale. The gyro-motion is fastest, the bounce motion is intermediate, and the drift motion is slowest. The hierarchy of timescales (${\omega }_c\gg {\omega }_b\gg {\omega }_d$) ensures that each invariant is well-conserved on the timescale of the next faster motion.

Breaking of these invariants (by collisions, fluctuations, or resonances) is the primary mechanism for particle and energy loss from magnetic confinement devices. Understanding when and how invariants break is central to fusion plasma physics.

Synthesis. Hamiltonian perturbation theory provides a unified framework for near-integrable systems through three complementary results: the averaging principle controls short-time dynamics, the KAM theorem guarantees the survival of most invariant tori for all time, and Nekhoroshev estimates bound the drift of actions over exponentially long times when KAM does not apply. The central insight is that the fast oscillatory part of the perturbation can be systematically eliminated by near-identity canonical transformations (whether via generating functions or Lie transforms), leaving a residual Hamiltonian that captures the slow secular evolution. The non-resonance condition is the foundational requirement: away from resonances, the perturbation has no secular effect on the actions. Adiabatic invariants extend this framework to systems with slowly varying parameters, where the action is conserved to exponential accuracy. Together, these results explain why most mechanical systems are far more regular than naive perturbation theory would suggest.

Connections Master

• 09.06.01 Action-angle variables provide the unperturbed integrable framework on which perturbation theory is built. Without action-angle coordinates, the separation into fast angles and slow actions is impossible.
• 09.04.02 Hamilton's equations define the dynamics that perturbation theory analyses. The averaging principle is a systematic approximation to these equations.
• 09.04.01 The Legendre transform produces the Hamiltonian $H(q,p)$ that enters the perturbation expansion. The canonical structure established by the Legendre transform is what makes canonical perturbation theory work.
• 09.08.01 The KAM theorem is the rigorous culmination of the perturbation-theory programme: it upgrades the formal normal forms to convergent constructions.
• 38.07.01 The full treatment of the KAM theorem, including the Diophantine condition, the measure-theoretic survival of tori, and the connection to ergodic theory.
• 09.02.05 The Lagrangian perturbation theory unit uses the method of multiple scales, which is the Lagrangian analogue of the Hamiltonian averaging method. Both eliminate fast oscillations to extract slow secular evolution.

Historical & philosophical context Master

Perturbation theory in classical mechanics originated in the work of Lagrange and Laplace on celestial mechanics (late 18th century), who sought to compute the long-term evolution of planetary orbits under mutual gravitational perturbations. The key difficulty -- the appearance of secular terms (unbounded growth in the perturbation series) -- was recognised by Poincare, who showed in Les Methodes Nouvelles de la Mecanique Celeste (1892--1899) that the perturbation series generally diverges. Poincare's proof of divergence was itself a landmark: it demonstrated that the Newtonian programme of computing the exact future of the solar system by successive approximations is impossible in principle, not merely in practice. The divergence does not mean the approximations are useless -- they are accurate over finite times -- but it means that the infinite series does not converge.

The averaging method was formalised by Krylov, Bogoliubov, and Mitropolsky in the 1930s--1950s. Krylov and Bogoliubov introduced the method of averaging for nonlinear oscillations, providing rigorous error bounds and establishing the connection between averaged equations and the true dynamics. Their work was motivated by engineering problems in radio electronics and nonlinear vibrations. Bogoliubov and Mitropolsky's book Asymptotic Methods in the Theory of Nonlinear Oscillations (1955) became the standard reference.

Kolmogorov's 1954 address at the International Congress of Mathematicians in Amsterdam introduced the idea of fixing the frequency and constructing convergent transformations via a Newton-type iteration. Arnold (1963) developed this into a full proof for Hamiltonian systems, and Moser (1962) proved an analogous result for area-preserving maps (the twist theorem). Together, these results form the KAM theorem. Nekhoroshev's 1977 thesis provided the long-time stability bounds for systems outside the KAM regime.

Lie transform perturbation theory was developed independently by Hori (1966), Deprit (1969), and Kamel (1969). It replaced the older Poincare-von Zeipel method with a more algebraic and computationally efficient framework. The Lie transform is now the standard tool in celestial mechanics, satellite dynamics, and accelerator physics.

The concept of adiabatic invariance was introduced by Ehrenfest (1916) in the context of the old quantum theory. Ehrenfest's hypothesis -- that the action $J=nh$ (Bohr-Sommerfeld quantisation) should be preserved under slow changes to the system -- was used to derive selection rules for quantum transitions. Although the old quantum theory was superseded by Schrodinger's wave mechanics (1926), the adiabatic theorem survived as a rigorous result in classical mechanics. Hannay's 1985 discovery of the geometric phase in classical mechanics (the Hannay angle), and Berry's 1984 discovery of the analogous phase in quantum mechanics (the Berry phase), revealed a deep connection between adiabatic evolution and the topology of the parameter space. The Hannay-Berry connection shows that adiabatic invariants are not merely approximately conserved quantities; they are the classical shadow of a geometric structure (a connection on a fibre bundle) that governs the phase of the wave function in quantum mechanics.

The philosophical significance of perturbation theory lies in the tension between integrability and chaos. The Newtonian worldview suggests that the motion of the solar system is, in principle, exactly predictable. Poincare's discovery that the perturbation series diverges shattered this expectation: exact prediction is impossible, not because of measurement error, but because of the mathematical structure of the equations themselves. The KAM theorem partially restored the picture -- most orbits are regular and predictable -- but the gaps between KAM tori admit chaotic diffusion (Arnold diffusion), so the long-term future of any specific orbit is genuinely uncertain. The Nekhoroshev theorem provides a practical resolution: the uncertainty manifests only on timescales exponentially long in the perturbation strength, which for the solar system is far longer than its age.

Bibliography Master

• Poincare, H., Les Methodes Nouvelles de la Mecanique Celeste, Vols. 1--3 (Gauthier-Villars, 1892--1899). The foundational work on perturbation theory and the discovery of the divergence of the perturbation series.

• Bogoliubov, N. N. & Mitropolsky, Y. A., Asymptotic Methods in the Theory of Nonlinear Oscillations (Hindustan Publishing, 1961). The formalisation of the averaging method for nonlinear oscillations.

• Kolmogorov, A. N., "On conservation of conditionally periodic motions under small perturbations of the Hamiltonian," Dokl. Akad. Nauk SSSR 98 (1954), 527--530. The original statement of the KAM theorem.

• Arnold, V. I., "Proof of a theorem of A. N. Kolmogorov on the invariance of quasi-periodic motions under small perturbations of the Hamiltonian," Russian Math. Surveys 18:5 (1963), 9--36.

• Moser, J., "On invariant curves of area-preserving mappings of an annulus," Nachr. Akad. Wiss. Gottingen Math.-Phys. Kl. II (1962), 1--20.

• Nekhoroshev, N. N., "An exponential estimate of the time of stability of nearly integrable Hamiltonian systems," Russian Math. Surveys 32:6 (1977), 1--65.

• Hori, G., "Theory of general perturbations with unspecified canonical variables," Publ. Astron. Soc. Japan 18 (1966), 287--296.

• Deprit, A., "Canonical transformations depending on a small parameter," Celestial Mechanics 1 (1969), 12--30.

• Hannay, J. H., "Angle variable holonomy in adiabatic excursion of an integrable Hamiltonian," J. Phys. A 18 (1985), 221--230.

• Berry, M. V., "Quantal phase factors accompanying adiabatic changes," Proc. Roy. Soc. A 392 (1984), 45--57.

• Lochak, P. & Meunier, C., Multiphase Averaging for Classical Systems (Springer, 1988).

• Goldstein, H., Poole, C. P. & Safko, J., Classical Mechanics, 3rd ed. (Pearson, 2002), Ch. 11.

• Landau, L. D. & Lifshitz, E. M., Mechanics, 3rd ed. (Pergamon, 1976), §49--51.

• Arnold, V. I., Mathematical Methods of Classical Mechanics, 2nd ed. (Springer GTM 60, 1989), §52.

• Taylor, J. R., Classical Mechanics (University Science Books, 2005), Ch. 12.