Eyring–Kramers Law in Metastable Dynamics

• Eyring–Kramers law is a formula describing rare transitions between metastable states by combining an Arrhenius exponential factor with a subexponential prefactor based on local quadratic approximations.
• It applies to various dynamics including overdamped, hypoelliptic, and non-reversible processes, offering insights into mean exit times, spectral gaps, and probabilistic exit distributions.
• The law incorporates corrections from boundary effects, non-Gibbsianness, symmetry, and infinite-dimensional extensions, linking potential barriers to transition rates in complex systems.

The Eyring–Kramers law is the sharp low-noise or low-temperature asymptotic for rare transitions between metastable states. In its canonical form, it combines an Arrhenius exponential, determined by the relevant communication height or energy barrier, with a subexponential prefactor determined by local quadratic data near minima and transition states. Depending on normalization, it appears as a mean-time asymptotic proportional to $e^{\Delta V/\epsilon}$ or $e^{\beta \Delta V}$, and equivalently as a rate asymptotic proportional to $e^{-\Delta V/\epsilon}$ or $e^{-\beta \Delta V}$. Contemporary work treats not only reversible overdamped diffusions, but also hypoelliptic kinetic processes, non-reversible and non-Gibbsian diffusions, jump processes, and infinite-dimensional SPDEs, with the prefactor varying accordingly (Lelièvre et al., 2018, Lee et al., 16 Mar 2025, Bouchet et al., 2015).

1. Classical reversible form

For overdamped Langevin dynamics

$$dX_t=-\nabla V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dW_t,$$

with a nondegenerate local minimum $x_i$, an index-1 saddle $x_s$, and barrier $\Delta V=V(x_s)-V(x_i)$, the classical Eyring–Kramers rate is

$$k_{i\to j}\approx \frac{|\lambda_-(x_s)|}{2\pi}\, \sqrt{\frac{\det(\nabla^2 V(x_i))}{|\det(\nabla^2 V(x_s))|}}\, e^{-\beta\Delta V},$$

where $\lambda_-(x_s)<0$ is the unique unstable Hessian eigenvalue at the saddle. The reciprocal formulation gives the mean transition time

$e^{\beta \Delta V}$0

In this reversible setting, the law describes not only metastable transition times but also the small spectrum of the generator, with the same communication heights governing the exponentially small eigenvalues (Lelièvre et al., 2018).

The same exponential scale controls functional inequalities for the Gibbs measure. For smooth Morse potentials with a unique dominant critical depth $e^{\beta \Delta V}$1, the spectral gap and log-Sobolev constant satisfy

$e^{\beta \Delta V}$2

which is the mechanism used to obtain quantitative convergence rates for simulated annealing schedules of the form $e^{\beta \Delta V}$3 when $e^{\beta \Delta V}$4 (Tang et al., 2021).

2. Exit problems, quasi-stationarity, and boundary saddles

A complementary formulation treats exit from a metastable domain $e^{\beta \Delta V}$5. For the Dirichlet realization of the overdamped generator on $e^{\beta \Delta V}$6, the principal eigenvalue $e^{\beta \Delta V}$7 is simple, the associated quasi-stationary distribution (QSD) is explicit in terms of the principal eigenfunction, and if the process starts from the QSD then $e^{\beta \Delta V}$8 is exponential with parameter $e^{\beta \Delta V}$9 and independent of the exit location. This gives an exact metastable Markov-jump reduction at the level of the QSD, with Eyring–Kramers asymptotics entering through the exit fluxes (Lelièvre et al., 2018).

When the dominant escape channels are on the boundary rather than through interior saddles, the prefactor changes. If $e^{-\Delta V/\epsilon}$0 is the unique minimum in $e^{-\Delta V/\epsilon}$1 and $e^{-\Delta V/\epsilon}$2 are the relevant local minima of $e^{-\Delta V/\epsilon}$3 with $e^{-\Delta V/\epsilon}$4, then the exit rates in the normalization

$e^{-\Delta V/\epsilon}$5

are

$e^{-\Delta V/\epsilon}$6

for higher saddles, with the obvious modification when several lowest boundary saddles have equal height. Here $e^{-\Delta V/\epsilon}$7 is the normal Hessian eigenvalue at the boundary saddle. The paper also emphasizes that exit rates and transition rates differ by a factor $e^{-\Delta V/\epsilon}$8 in the interior-saddle picture, because upon reaching the separatrix the process has asymptotically probability $e^{-\Delta V/\epsilon}$9 to return and $e^{-\beta \Delta V}$0 to commit to the neighboring basin (Lelièvre et al., 2022).

This boundary theory clarifies a frequent source of confusion: the Eyring–Kramers prefactor is not universally the inverse unstable Hessian eigenvalue. In domain-exit problems it can instead involve outward normal derivatives and tangential Hessians on $e^{-\beta \Delta V}$1, with the exit distribution concentrating on generalized saddles of the boundary gradient flow (Lelièvre et al., 2018).

3. Hypoelliptic and underdamped Langevin law

For the underdamped Langevin process in phase space,

$e^{-\beta \Delta V}$2

with Hamiltonian

$e^{-\beta \Delta V}$3

Lee, Ramil, and Seo derive the sharp Eyring–Kramers asymptotic for the mean transition time between two wells connected by a unique index-1 saddle $e^{-\beta \Delta V}$4. If $e^{-\beta \Delta V}$5 is the initial minimum, $e^{-\beta \Delta V}$6, $e^{-\beta \Delta V}$7, $e^{-\beta \Delta V}$8 is the magnitude of the unique negative eigenvalue of $e^{-\beta \Delta V}$9, and $$dX_t=-\nabla V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dW_t,$$0, then

$$dX_t=-\nabla V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dW_t,$$1

or, in inverse-temperature notation $$dX_t=-\nabla V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dW_t,$$2,

$$dX_t=-\nabla V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dW_t,$$3

with

$$dX_t=-\nabla V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dW_t,$$4

The factor $$dX_t=-\nabla V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dW_t,$$5 is the hypoelliptic unstable frequency at the saddle; friction and kinetic degrees of freedom enter only through this quantity (Lee et al., 16 Mar 2025).

This formula reduces to the overdamped prefactor in the high-friction scaling regime and to the low-friction Kramers regime as $$dX_t=-\nabla V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dW_t,$$6. In particular, $$dX_t=-\nabla V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dW_t,$$7 as $$dX_t=-\nabla V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dW_t,$$8, while $$dX_t=-\nabla V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dW_t,$$9 as $x_i$0 (Lee et al., 16 Mar 2025).

The derivation is structurally different from the elliptic overdamped case. Because diffusion acts only in momentum, the kinetic generator is hypoelliptic and the usual elliptic Dirichlet/Thomson capacity identities are not directly available. The argument instead introduces an equilibrium potential, a capacity-analog defined through a momentum-reversed test identity, and an elliptic regularization

$x_i$1

which yields a “magic formula” relating mean exit times, a harmonic surface measure, and the capacity-analog before passing to the limits $x_i$2 and $x_i$3 (Lee et al., 16 Mar 2025).

4. Irreversibility, non-Gibbsianness, and symmetry corrections

For general irreversible elliptic diffusions

$x_i$4

the barrier is no longer the potential difference but the Freidlin–Wentzell quasipotential $x_i$5. Under uniqueness and nondegeneracy of the instanton $x_i$6 from the well $x_i$7 to the quasipotential saddle $x_i$8, the mean transition time satisfies

$x_i$9

Here $x_s$0 is the unstable eigenvalue of the Jacobian $x_s$1, $x_s$2 is the limiting Hessian of the quasipotential at the saddle, and the factor

$x_s$3

measures non-Gibbsianness along the instanton. In the Gibbs-preserving case this correction is $x_s$4 (Bouchet et al., 2015).

A more restrictive but important non-reversible class preserves the Gibbs invariant density while breaking reversibility. For

$x_s$5

with $x_s$6 and $x_s$7, the double-well prefactor is obtained by replacing the negative Hessian eigenvalue at the saddle by the unique negative eigenvalue of the non-selfadjoint linearization $x_s$8:

$x_s$9

Since $\Delta V=V(x_s)-V(x_i)$0 in the reversible comparison, the non-reversible perturbation accelerates the metastable transition (Lee et al., 2020).

Analogous replacements occur for non-reversible random walks in a potential field. There the exponential barrier is unchanged, but the prefactor contains the unique negative eigenvalue of the Jacobian $\Delta V=V(x_s)-V(x_i)$1 generated by the cycle structure rather than the negative Hessian eigenvalue itself, together with a non-Gibbs correction $\Delta V=V(x_s)-V(x_i)$2 (Landim et al., 2016).

Symmetry produces a different type of correction. For reversible metastable jump processes invariant under a finite group $\Delta V=V(x_s)-V(x_i)$3, the trivial isotypic component has effective prefactors

$\Delta V=V(x_s)-V(x_i)$4

so stabilizers modify the standard Eyring–Kramers prefactor by explicit orbit-combinatorial factors. In nontrivial irreducible representations, additional Arrhenius exponents can appear through successor cycles and Schur complements of the symmetry-reduced generator (Berglund et al., 2013).

The non-Gibbsian boundary-exit problem leads to yet another correction. For $\Delta V=V(x_s)-V(x_i)$5 with an orthogonal decomposition $\Delta V=V(x_s)-V(x_i)$6, $\Delta V=V(x_s)-V(x_i)$7, and non-characteristic boundary, the mean exit time from a bounded domain satisfies

$\Delta V=V(x_s)-V(x_i)$8

with

$\Delta V=V(x_s)-V(x_i)$9

The new factor

$$k_{i\to j}\approx \frac{|\lambda_-(x_s)|}{2\pi}\, \sqrt{\frac{\det(\nabla^2 V(x_i))}{|\det(\nabla^2 V(x_s))|}}\, e^{-\beta\Delta V},$$0

is the explicit non-Gibbsian correction (Peutrec et al., 22 Sep 2025).

5. Infinite-dimensional, constrained, and field-theoretic extensions

In systems with conserved quantities, the naive determinant ratio may diverge because the mobility has zero modes. For reversible gradient diffusions constrained to a manifold $$k_{i\to j}\approx \frac{|\lambda_-(x_s)|}{2\pi}\, \sqrt{\frac{\det(\nabla^2 V(x_i))}{|\det(\nabla^2 V(x_s))|}}\, e^{-\beta\Delta V},$$1, the generalized Eyring–Kramers law replaces full determinants by determinants restricted to the tangent space:

$$k_{i\to j}\approx \frac{|\lambda_-(x_s)|}{2\pi}\, \sqrt{\frac{\det(\nabla^2 V(x_i))}{|\det(\nabla^2 V(x_s))|}}\, e^{-\beta\Delta V},$$2

For a single conserved quantity with unit normal $$k_{i\to j}\approx \frac{|\lambda_-(x_s)|}{2\pi}\, \sqrt{\frac{\det(\nabla^2 V(x_i))}{|\det(\nabla^2 V(x_s))|}}\, e^{-\beta\Delta V},$$3, this is equivalently

$$k_{i\to j}\approx \frac{|\lambda_-(x_s)|}{2\pi}\, \sqrt{\frac{\det(\nabla^2 V(x_i))}{|\det(\nabla^2 V(x_s))|}}\, e^{-\beta\Delta V},$$4

This framework is applied to fluctuating hydrodynamics, including stochastic thin-film rupture and a mass-conserving social-segregation SPDE, where ignoring the conservation law mispredicts the prefactor (Liu et al., 2024).

A rigorous infinite-dimensional spectral-gap version is now available for reversible stochastic gradient systems

$$k_{i\to j}\approx \frac{|\lambda_-(x_s)|}{2\pi}\, \sqrt{\frac{\det(\nabla^2 V(x_i))}{|\det(\nabla^2 V(x_s))|}}\, e^{-\beta\Delta V},$$5

on an abstract Wiener space. In the double-well case, with minima $$k_{i\to j}\approx \frac{|\lambda_-(x_s)|}{2\pi}\, \sqrt{\frac{\det(\nabla^2 V(x_i))}{|\det(\nabla^2 V(x_s))|}}\, e^{-\beta\Delta V},$$6, relevant saddles $$k_{i\to j}\approx \frac{|\lambda_-(x_s)|}{2\pi}\, \sqrt{\frac{\det(\nabla^2 V(x_i))}{|\det(\nabla^2 V(x_s))|}}\, e^{-\beta\Delta V},$$7, unstable eigenvalues $$k_{i\to j}\approx \frac{|\lambda_-(x_s)|}{2\pi}\, \sqrt{\frac{\det(\nabla^2 V(x_i))}{|\det(\nabla^2 V(x_s))|}}\, e^{-\beta\Delta V},$$8 of the linearized operator in the $$k_{i\to j}\approx \frac{|\lambda_-(x_s)|}{2\pi}\, \sqrt{\frac{\det(\nabla^2 V(x_i))}{|\det(\nabla^2 V(x_s))|}}\, e^{-\beta\Delta V},$$9-metric, and Fredholm determinant factors

$\lambda_-(x_s)<0$0

the first nonzero eigenvalue satisfies, in the symmetric-well case,

$\lambda_-(x_s)<0$1

which is an infinite-dimensional Eyring–Kramers law for the spectral gap (Brooks et al., 15 Jun 2026).

Renormalized SPDEs require further determinant regularization. For the two-dimensional stochastic Allen–Cahn equation with Wick renormalization, the Arrhenius factor remains

$\lambda_-(x_s)<0$2

but the finite-dimensional determinant ratio is replaced by the Carleman–Fredholm determinant

$\lambda_-(x_s)<0$3

reflecting the ultraviolet renormalization of the quartic field theory (Berglund et al., 2016).

Hyperbolic field theories admit a transition-state-theory variant. For the damped stochastic nonlinear wave equation with $\lambda_-(x_s)<0$4 potential on $\lambda_-(x_s)<0$5, $\lambda_-(x_s)<0$6, the transition frequency factorizes into a white-noise flux term and a renormalized Gibbs-measure ratio, producing

$\lambda_-(x_s)<0$7

In the two-dimensional unforced hyperbolic model, the transmission coefficient satisfies $\lambda_-(x_s)<0$8 (Barashkov et al., 2024).

Further SPDE variants show that the transition state itself can bifurcate. For the one-dimensional dynamical sine-Gordon stochastic heat equation, the transition state is constant when $\lambda_-(x_s)<0$9 and becomes a kink–antikink pair when $e^{\beta \Delta V}$00, while the Arrhenius barrier remains $e^{\beta \Delta V}$01 (Laarne, 17 Sep 2025).

6. Methods, applications, and conceptual boundaries

Three proof paradigms dominate the subject. The classical overdamped theory uses potential theory, Dirichlet spectral analysis, Witten Laplacians, and Agmon localization to identify the principal Dirichlet eigenfunction, its normal derivative at the boundary, and the corresponding capacity or flux. Non-selfadjoint Fokker–Planck operators with Gibbs stationary states can also be treated semiclassically by Gaussian quasimodes and graded Schur-complement analysis, without supersymmetry or PT symmetry, yielding prefactors involving the unique negative-real-part eigenvalue of

$e^{\beta \Delta V}$02

rather than a purely Hessian quantity (Lelièvre et al., 2018, Bony et al., 2022).

A distinct geometric viewpoint characterizes capacity by two weighted objects: a Gibbs-weighted geodesic distance

$e^{\beta \Delta V}$03

and a Gibbs-weighted minimal separating hypersurface

$e^{\beta \Delta V}$04

For a single saddle one obtains

$e^{\beta \Delta V}$05

which recovers the classical Hessian-based Eyring–Kramers formula in the Morse case and extends it to flat minima, flat saddles, and multiple same-height saddles arranged in series or parallel (Avelin et al., 2022).

Algorithmically, the law underpins metastable coarse-graining. In overdamped dynamics, the QSD makes the exit event exactly representable as a kinetic Monte Carlo step, with exponentially distributed holding time and independent exit channel. This is the basis for jump-process reductions and accelerated dynamics methods such as TAD. In Adaptive Kinetic Monte Carlo, Eyring–Kramers rates are used both to populate the low-temperature event catalog and, when reliable at the search temperature, to estimate the fraction of total reaction rate already discovered (Lelièvre et al., 2018, Aristoff et al., 2015).

The law also supplies quantitative convergence rates for optimization algorithms. In simulated annealing, the Eyring–Kramers asymptotics for the log-Sobolev constant and spectral gap yield polynomial decay bounds for

$e^{\beta \Delta V}$06

under cooling schedules with $e^{\beta \Delta V}$07 and suitable step-size conditions, directly linking metastable barrier structure to optimization complexity (Tang et al., 2021).

A recurring misconception is that the Eyring–Kramers law is a single formula with a universal prefactor. The data instead show a stable hierarchy of modifications. The Arrhenius exponent is robust and is usually determined by a communication height, a quasipotential barrier, or a domain-exit barrier. The prefactor, however, is sensitive to the operator class and geometry: unstable Hessian eigenvalues in the reversible overdamped case, $e^{\beta \Delta V}$08 in the underdamped hypoelliptic case, normal derivatives and tangential Hessians for boundary exit, pseudo-determinants under conservation laws, Fredholm or Carleman–Fredholm determinants in SPDEs, stabilizer factors in symmetric jump processes, and non-Gibbsianness corrections along instantons or boundary characteristics (Lee et al., 16 Mar 2025, Liu et al., 2024, Peutrec et al., 22 Sep 2025).