math

Instantons in Statistical Physics

Instantons Without Quantum Mechanics

The word “instanton” sounds like it belongs to quantum field theory. It doesn’t. An instanton is a saddle point of an action functional that mediates a rare transition, and the clearest examples are entirely classical.

The pattern is always the same:

You have a system with two (or more) metastable states
Transitions between them are exponentially rare: rate $\sim e^{-\Delta/k_BT}$ for some barrier $\Delta$
The exponential suppression means the transition is dominated by a single optimal path through configuration space
That optimal path is the instanton

The instanton is the most probable way for an improbable thing to happen.

Kramers Escape: The Prototype

A Brownian particle in a double-well potential $V(x)$ at temperature $T$. The overdamped Langevin dynamics:

\[\gamma \dot{x} = -V'(x) + \sqrt{2\gamma k_BT}\,\xi(t)\]

where $\xi$ is white noise. The particle sits near one minimum and occasionally hops to the other. The rate is Kramers’ formula:

\[k \sim \frac{\omega_{\min}\omega_{\max}}{2\pi\gamma}\, e^{-\Delta V/k_BT}\]

where $\Delta V$ is the barrier height, $\omega_{\min}$ is the curvature at the minimum, and $\omega_{\max}$ is the curvature at the saddle.

But where does the particle go during the transition? It takes the path of least noise.

The Onsager-Machlup Action

The probability of a given trajectory $x(t)$ under the Langevin dynamics is:

\[P[x] \propto \exp\left( -\frac{1}{4k_BT}\int_0^T dt\, \frac{(\gamma\dot{x} + V'(x))^2}{\gamma} \right)\]

This is a path integral with action:

\[S[x] = \frac{1}{4k_BT}\int dt\, \frac{(\gamma\dot{x} + V'(x))^2}{\gamma}\]

The most probable transition path minimizes this action. That minimizer is the instanton.

What the Instanton Looks Like

The Euler-Lagrange equation for $S$ is:

\[\gamma\ddot{x} = -V''(x)(\gamma\dot{x} + V'(x)) + V'(x)\frac{V''(x)}{\gamma}(\gamma\dot{x} + V'(x))\]

but there’s a much cleaner way to see it. The instanton satisfies:

\[\gamma\dot{x} = +V'(x)\]

This is the time-reversed deterministic dynamics: the particle climbs up the potential gradient. The noise is used entirely to push the particle uphill along the steepest ascent.

Compare with the noiseless (relaxational) dynamics $\gamma\dot{x} = -V’(x)$, where the particle rolls downhill. The instanton is literally the time-reverse of relaxation.

The action on this path is:

\[S_{\text{inst}} = \frac{\Delta V}{k_BT}\]

exactly the barrier height divided by temperature. This is why Kramers’ rate goes as $e^{-\Delta V/k_BT}$: the instanton action is the Arrhenius exponent.

Nucleation: Instantons in Field Theory

Now upgrade from a particle to a field. Consider a scalar order parameter $\phi(\mathbf{x})$ with free energy:

\[F[\phi] = \int d^d x\left[ \frac{\kappa}{2}(\nabla\phi)^2 + f(\phi) \right]\]

where $f(\phi)$ has two minima: a metastable phase (say $\phi_+$, the “false vacuum”) and the stable phase ($\phi_-$, the “true vacuum”), with $f(\phi_-) < f(\phi_+)$.

The system sits in the metastable phase. It decays by forming a critical droplet of the stable phase. The nucleation rate is:

\[\Gamma \sim e^{-F^*/k_BT}\]

where $F^*$ is the free energy cost of the critical droplet.

The Critical Droplet Is an Instanton

The critical droplet extremizes $F[\phi]$ subject to the boundary condition $\phi \to \phi_+$ at infinity. It satisfies:

\[-\kappa\nabla^2\phi + f'(\phi) = 0\]

For a spherically symmetric droplet of radius $R$ in $d$ dimensions, the free energy has two competing terms:

\[F(R) \sim -\vert\Delta f\vert R^d + \sigma R^{d-1}\]

where $\Delta f = f(\phi_+) - f(\phi_-)$ is the bulk free energy difference and $\sigma$ is the surface tension. The critical radius maximizes $F(R)$:

\[R^* \sim \frac{\sigma}{\vert\Delta f\vert}\]

Droplets smaller than $R^$ shrink (surface tension wins). Droplets larger than $R^$ grow (bulk free energy wins). The critical droplet is the unstable saddle point between these fates.

This is classical nucleation theory, but the language of instantons clarifies what’s going on: the critical droplet is a saddle point of the free energy functional, and it dominates the decay rate because $e^{-F/k_BT}$ is sharply peaked around the saddle.

The Thin-Wall Limit

When $\Delta f$ is small (the two phases are nearly degenerate), the critical droplet is large and the interface is thin. In this limit:

\[F^* = \frac{\sigma^d}{(\Delta f)^{d-1}} \cdot C_d\]

where $C_d$ is a geometric factor ($C_3 = 16\pi/3$ in three dimensions).

The thin-wall instanton has a sharp domain wall profile:

\[\phi(r) \approx \phi_{\text{wall}}(r - R^*)\]

where $\phi_{\text{wall}}$ is the one-dimensional kink interpolating between $\phi_-$ and $\phi_+$.

Large Deviation Theory: The General Framework

The pattern generalizes. Large deviation theory provides the mathematical framework.

Given a stochastic process $X_n$ (or $X(t)$), a large deviation principle says:

\[P(X \in A) \asymp e^{-n\, I(A)}\]

where $I$ is the rate function and $\asymp$ means logarithmic equivalence. The rate function plays the role of the action.

The most probable way to realize an unlikely event minimizes the rate function subject to the constraint that the event occurs. That minimizer is the instanton.

Example: Empirical Averages

Let $X_1, \ldots, X_n$ be i.i.d. with mean $\mu$. The probability that the empirical mean $\bar{X}_n = a \neq \mu$ decays as:

\[P(\bar{X}_n \approx a) \asymp e^{-n\, I(a)}\]

where $I(a) = \sup_\lambda [\lambda a - \log M(\lambda)]$ is the Legendre transform of the log moment generating function (Cram'er’s theorem).

The “instanton” here is the tilted distribution $P_\lambda(x) \propto e^{\lambda x} P(x)$ that makes $a$ the typical value. It’s the least unlikely way to get an unlikely average.

Example: Dynamical Systems

For a stochastic differential equation $\dot{x} = b(x) + \sqrt{\epsilon}\,\sigma(x)\,\xi(t)$, the Freidlin-Wentzell theory gives:

\[P(\text{path near } \phi) \asymp e^{-S[\phi]/\epsilon}\]

with action:

\[S[\phi] = \frac{1}{2}\int_0^T \lVert\dot{\phi} - b(\phi)\rVert^2_{\sigma\sigma^T} \, dt\]

The instanton is the path minimizing $S$ subject to the constraint of connecting two metastable states. This is exactly the Kramers problem in full generality: the most probable escape path from a basin of attraction.

The Instanton as a Saddle Point

In all these examples, the instanton is an unstable extremum: it minimizes the action among all paths connecting the two states, but it maximizes the action in some transverse direction (corresponding to the timing of the transition).

This instability shows up as a negative eigenvalue in the spectrum of fluctuations around the instanton, and a zero eigenvalue from time-translation invariance (you can shift when the transition happens).

The negative mode means the instanton is a saddle, not a minimum. This is essential: it means the instanton describes a transition over a barrier, not a rest at a minimum.

The zero mode becomes a collective coordinate integrated over (just like in the quantum case), producing a factor of time $T$ that converts the amplitude into a rate.

Why This Matters

The instanton viewpoint unifies several things that look different:

Setting	“Action”	“Instanton”	Rate
Kramers escape	Onsager-Machlup	Steepest ascent path	$e^{-\Delta V/k_BT}$
Nucleation	Free energy functional	Critical droplet	$e^{-F^*/k_BT}$
Large deviations	Rate function	Optimal fluctuation	$e^{-nI}$
Freidlin-Wentzell	SDE action	Minimum action path	$e^{-S/\epsilon}$

The exponential suppression always comes from evaluating the action on the instanton. The prefactor comes from Gaussian fluctuations around it (the instanton determinant).

The conceptual payoff: rare events aren’t random in their details. They happen in the least unlikely way possible, which is a well-defined variational problem. The instanton is that solution.

Summary

An instanton is the optimal fluctuation path mediating a rare transition
In Kramers escape, it’s the time-reverse of deterministic relaxation: steepest ascent up the potential
In nucleation, it’s the critical droplet: the saddle point of the free energy functional
Large deviation theory provides the general framework: rate functions are actions, instantons are their minimizers
The instanton always has a zero mode (time translation) and a negative mode (barrier crossing)
Rare events happen in the most probable improbable way, and that way is the instanton

See also: [Asymptotics, Borel, and Stokes], [WKB and Matched Asymptotics], [Extreme Value Theory].