Onsager-Machlup Functional: Theory & Applications

• The Onsager-Machlup functional is a local density analog defined via asymptotic ratios of small-ball probabilities in metric and measure spaces.
• It incorporates drift, divergence, and geometric corrections to capture dynamics in diffusion, fractional, jump, and infinite-dimensional stochastic systems.
• Its computation informs techniques in data assimilation and machine learning by enabling variational path selection and optimized transition path sampling.

The Onsager-Machlup functional is a path-space or metric-space analog of a probability density, defined through asymptotic ratios of small-ball probabilities rather than with respect to a Lebesgue reference measure. In its most general form, it compares the probabilities of infinitesimal neighborhoods of points or paths and thereby encodes local relative likelihood. In diffusion theory it describes the probability of a process remaining in a small tube around a smooth trajectory, while in abstract metric measure spaces it is determined by ratios of measures of small balls. Across the literature, sign conventions differ: some authors define a functional $J$ through $\exp(J)$, whereas others define an action appearing with a negative sign in the exponential tube asymptotics (Selk, 12 Oct 2025, Sugiura, 2017).

1. Small-ball definition and basic examples

For a metric measure space $M=(X,d,\mu)$, the Onsager-Machlup functional is defined by the asymptotic relation

$$\lim_{r\to 0^+}\frac{\mu(B(r,x))}{\mu(B(r,y))}=\exp\!\big(\operatorname{OM}(y)-\operatorname{OM}(x)\big),$$

whenever the limit exists on a subset $Z\subseteq X$. It is determined only up to an additive constant. In path-space formulations, the same principle appears as a ratio of tube probabilities around a deterministic reference path and a reference noise process; for additive fractional noise, for example,

$$e^{J(h)}=\lim_{\epsilon\searrow 0}\frac{P(\|X-h\|<\epsilon)}{P(\|B\|<\epsilon)}.$$

These formulations share the same structural content: the functional records the logarithmic relative asymptotics of small neighborhoods (Selk, 12 Oct 2025, Maayan, 2017).

Several canonical examples fix the interpretation. On $(\mathbb{R}^n,d,e^{-f}\lambda)$, with $d$ the standard metric and $\lambda$ Lebesgue measure, the Onsager-Machlup functional is $f$. For a discrete probability measure, it is the negative logarithm of the probability mass function. For Gaussian measures in infinite dimensions, the functional is finite only on the Cameron-Martin space, even though that space itself has measure zero (Selk, 12 Oct 2025).

This definition supports two closely related readings. First, it is a generalized density concept for spaces where a reference Lebesgue measure does not exist. Second, it is a local mode functional: it identifies the centers of small balls or tubes with largest asymptotic probability. This role underlies its use in path-space MAP formulations and in the analysis of most probable transition paths (Selk, 12 Oct 2025).

2. Diffusion formulas and geometric corrections

For classical Itô diffusions with additive Brownian noise,

$\exp(J)$0

the continuous-time Onsager-Machlup functional takes the form

$\exp(J)$1

The quadratic mismatch term measures deviation from the drift, and the divergence term is the tube-volume correction. In discrete-time estimation, this same structure leads to a cost functional with background and observation penalties plus a time-discretized Onsager-Machlup prior (Sugiura, 2017).

When the diffusion coefficient varies in time,

$\exp(J)$2

the functional generalizes to

$\exp(J)$3

with

$\exp(J)$4

The time-varying amplitude enters directly through $\exp(J)$5, and the analysis requires measurable norms in Hölder spaces together with a Girsanov transformation (Zhang et al., 2024).

On a manifold with time-dependent metric $\exp(J)$6, generated by

$\exp(J)$7

the inhomogeneous geometric Onsager-Machlup integrand is

$\exp(J)$8

Relative to the time-homogeneous formula, the new term is the infinitesimal volume variation $\exp(J)$9. In the Ricci flow case, the resulting action is closely related to Lott’s $M=(X,d,\mu)$0-distance, and the corresponding most probable path equation matches the critical equation for the associated space-time functional (Coulibaly-Pasquier, 2011).

3. Dependence on metric, topology, and renormalization

A central structural fact is that the Onsager-Machlup functional depends on the metric as well as on the measure. For a geodesic metric measure space $M=(X,d,\mu)$1 with Onsager-Machlup functional $M=(X,d,\mu)$2, simultaneous conformal reweightings

$M=(X,d,\mu)$3

lead, under polynomial small-ball scaling,

$M=(X,d,\mu)$4

If $M=(X,d,\mu)$5 is constant, this reduces to $M=(X,d,\mu)$6. In finite-dimensional settings, such formulas imply that the functional can be altered by modifying the geometry; on $M=(X,d,\mu)$7, a suitable metric can “uniformize” any continuous density so that the Onsager-Machlup functional is identically zero. By contrast, under ultrasmall-ball asymptotics of the form

$M=(X,d,\mu)$8

any non-constant conformal change of metric destroys the Onsager-Machlup functional: the limit either vanishes or diverges (Selk, 12 Oct 2025).

This metric and topology dependence becomes sharper in constructive field theory. For the finite-volume $M=(X,d,\mu)$9 measures, the $$\lim_{r\to 0^+}\frac{\mu(B(r,x))}{\mu(B(r,y))}=\exp\!\big(\operatorname{OM}(y)-\operatorname{OM}(x)\big),$$0 Onsager-Machlup functional coincides with the $$\lim_{r\to 0^+}\frac{\mu(B(r,x))}{\mu(B(r,y))}=\exp\!\big(\operatorname{OM}(y)-\operatorname{OM}(x)\big),$$1 action on $$\lim_{r\to 0^+}\frac{\mu(B(r,x))}{\mu(B(r,y))}=\exp\!\big(\operatorname{OM}(y)-\operatorname{OM}(x)\big),$$2. In $$\lim_{r\to 0^+}\frac{\mu(B(r,x))}{\mu(B(r,y))}=\exp\!\big(\operatorname{OM}(y)-\operatorname{OM}(x)\big),$$3, the natural topology is insufficient because the Wick-ordered quartic potential is measurable but not continuous in $$\lim_{r\to 0^+}\frac{\mu(B(r,x))}{\mu(B(r,y))}=\exp\!\big(\operatorname{OM}(y)-\operatorname{OM}(x)\big),$$4. The remedy is to introduce “enhanced” distances controlling $$\lim_{r\to 0^+}\frac{\mu(B(r,x))}{\mu(B(r,y))}=\exp\!\big(\operatorname{OM}(y)-\operatorname{OM}(x)\big),$$5 together with Wick powers such as $$\lim_{r\to 0^+}\frac{\mu(B(r,x))}{\mu(B(r,y))}=\exp\!\big(\operatorname{OM}(y)-\operatorname{OM}(x)\big),$$6 and $$\lim_{r\to 0^+}\frac{\mu(B(r,x))}{\mu(B(r,y))}=\exp\!\big(\operatorname{OM}(y)-\operatorname{OM}(x)\big),$$7, in a manner explicitly described as analogous to rough path metrics; with these enhanced balls, the Onsager-Machlup functional agrees with the renormalized $$\lim_{r\to 0^+}\frac{\mu(B(r,x))}{\mu(B(r,y))}=\exp\!\big(\operatorname{OM}(y)-\operatorname{OM}(x)\big),$$8 action. In $$\lim_{r\to 0^+}\frac{\mu(B(r,x))}{\mu(B(r,y))}=\exp\!\big(\operatorname{OM}(y)-\operatorname{OM}(x)\big),$$9, two natural generalizations are degenerate, with $Z\subseteq X$0 and $Z\subseteq X$1 for nonzero smooth $Z\subseteq X$2, while the $Z\subseteq X$3 action can nevertheless be recovered through a joint small-radius/large-frequency limit under precise coupling conditions (Gasteratos et al., 24 Sep 2025).

These results rule out a common oversimplification: the Onsager-Machlup functional is not generally an intrinsic object of the measure alone. The small-ball geometry, the chosen topology, and, in singular theories, the enhancement or renormalization scheme can be decisive.

4. Fractional noise and degenerate dynamics

For stochastic differential equations driven by fractional Brownian motion with time-varying diffusion,

$Z\subseteq X$4

the Onsager-Machlup functional is defined by a small-ball ratio relative to $Z\subseteq X$5, with the candidate path constrained by $Z\subseteq X$6. The general expression is

$Z\subseteq X$7

The explicit operator form depends on the Hurst index: for $Z\subseteq X$8 it involves left-sided Riemann-Liouville fractional integrals, for $Z\subseteq X$9 left-sided fractional derivatives, and for $$e^{J(h)}=\lim_{\epsilon\searrow 0}\frac{P(\|X-h\|<\epsilon)}{P(\|B\|<\epsilon)}.$$0 ordinary derivatives. The admissible norms also depend on $$e^{J(h)}=\lim_{\epsilon\searrow 0}\frac{P(\|X-h\|<\epsilon)}{P(\|B\|<\epsilon)}.$$1: for $$e^{J(h)}=\lim_{\epsilon\searrow 0}\frac{P(\|X-h\|<\epsilon)}{P(\|B\|<\epsilon)}.$$2, the supremum norm and Hölder norms of order $$e^{J(h)}=\lim_{\epsilon\searrow 0}\frac{P(\|X-h\|<\epsilon)}{P(\|B\|<\epsilon)}.$$3 are valid; for $$e^{J(h)}=\lim_{\epsilon\searrow 0}\frac{P(\|X-h\|<\epsilon)}{P(\|B\|<\epsilon)}.$$4, Hölder norms must satisfy $$e^{J(h)}=\lim_{\epsilon\searrow 0}\frac{P(\|X-h\|<\epsilon)}{P(\|B\|<\epsilon)}.$$5; and for $$e^{J(h)}=\lim_{\epsilon\searrow 0}\frac{P(\|X-h\|<\epsilon)}{P(\|B\|<\epsilon)}.$$6, the Hölder range depends on the spatial regularity of the drift (Zhu et al., 12 Nov 2025).

Second-order and degenerate systems preserve the same fractional structure but require additional analysis. For the stochastic Newton equation

$$e^{J(h)}=\lim_{\epsilon\searrow 0}\frac{P(\|X-h\|<\epsilon)}{P(\|B\|<\epsilon)}.$$7

the Onsager-Machlup functional is derived by applying Girsanov only to the non-degenerate component and by evaluating a limiting conditional expectation using the stochastic Fubini theorem. The resulting action again splits into the three $$e^{J(h)}=\lim_{\epsilon\searrow 0}\frac{P(\|X-h\|<\epsilon)}{P(\|B\|<\epsilon)}.$$8-regimes and contains the derivative $$e^{J(h)}=\lim_{\epsilon\searrow 0}\frac{P(\|X-h\|<\epsilon)}{P(\|B\|<\epsilon)}.$$9 rather than $(\mathbb{R}^n,d,e^{-f}\lambda)$0 (Zhu et al., 24 Feb 2026).

For degenerate systems driven by fractional Brownian motion, the reference path must respect the deterministic coupling of the non-noisy coordinates. One formulation writes

$(\mathbb{R}^n,d,e^{-f}\lambda)$1

with the Onsager-Machlup functional expressed in terms of $(\mathbb{R}^n,d,e^{-f}\lambda)$2 and the divergence $(\mathbb{R}^n,d,e^{-f}\lambda)$3. In the multidimensional, time-dependent case, the derivation uses a Gaussian correlation inequality together with an approximation argument for infinite-dimensional convex sets to decouple conditioning effects. A sufficient condition for preservation of the deterministic most probable path is that $(\mathbb{R}^n,d,e^{-f}\lambda)$4 be constant (Liu et al., 2023, Zhu et al., 5 Jul 2026).

These fractional results show that memory changes the action from a local quadratic penalty into a nonlocal fractional one. A plausible implication is that “most probable path” computations in non-Markovian systems must be formulated in operator-adapted path spaces rather than by direct transplantation of Brownian formulas.

5. Jump, mean-field, and infinite-dimensional stochastic systems

For jump-diffusion processes, the Onsager-Machlup functional must account for the asymmetry of small jumps. In the one-dimensional model

$(\mathbb{R}^n,d,e^{-f}\lambda)$5

with finite first moment of small jumps, one derived expression is

$(\mathbb{R}^n,d,e^{-f}\lambda)$6

or equivalently as a completed square plus an additive constant. The Lévy contribution disappears for symmetric jump distributions and the formula reduces to the Brownian case (Chao et al., 2018).

A later probability-flow approach constructs a diffusion process with modified drift whose marginal laws match those of a jump-diffusion through the corresponding Fokker-Planck equations. For finite jump activity, this yields a closed-form continuous-time Onsager-Machlup functional

$(\mathbb{R}^n,d,e^{-f}\lambda)$7

where $(\mathbb{R}^n,d,e^{-f}\lambda)$8 contains a term from the Lévy intensity at the origin. For infinite jump activity, a continuous-time functional does not exist in that framework, but a time-discrete action can be written explicitly on partitions (Huang et al., 2024).

Infinite-dimensional systems exhibit analogous structures. For SPDEs of the form

$(\mathbb{R}^n,d,e^{-f}\lambda)$9

the Onsager-Machlup action is

$d$0

with $d$1. If $d$2, the Lévy part does not alter the leading Onsager-Machlup action (Hu et al., 2020).

For McKean-Vlasov SDEs, distribution dependence obstructs direct application of classical theory. An Euler-type approximation freezes the law on each time interval,

$d$3

and the limit yields an Onsager-Machlup action involving the Dirac mass $d$4, the modified drift $d$5, the metric $d$6, the geometric divergence, and the scalar curvature term $d$7 (Chao et al., 25 Jun 2026).

A parallel infinite-dimensional extension appears for stochastic lattice dynamical systems in $d$8 with time-varying diagonal diffusion $d$9: $\lambda$0 Its derivation uses an infinite-dimensional Girsanov transform, a Karhunen-Loève expansion, and Brownian ball probability estimates (Zhang et al., 2024).

6. Non-Euclidean and conformally structured examples

On the Heisenberg group, the Onsager-Machlup functional reflects sub-Riemannian rather than Riemannian structure. For horizontal Brownian motion $\lambda$1 and a path $\lambda$2, the asymptotic ratio of small sup-norm balls is

$\lambda$3

provided the horizontal projection $\lambda$4 belongs to the Cameron-Martin space. The functional is finite only for paths whose horizontal projection has finite energy. Because absolute continuity is too restrictive for this hypoelliptic setting, the analysis introduces horizontal continuous curves and a horizontal supremum semi-metric (Carfagnini et al., 2019).

In Schramm-Loewner evolution, the Onsager-Machlup functional appears not as a quadratic drift-divergence expression but as a conformally defined Loewner potential. For multi-chordal SLE, the relevant potential is the sum of the single-curve Loewner potentials plus a Brownian loop interaction term. For multi-radial SLE, the corresponding potential includes the multi-radial energy, pairwise logarithmic sine interactions, and a spiral-rate contribution. The core asymptotic statement is that ratios of SLE probabilities of shrinking admissible neighborhoods around conformally equivalent multicurves converge to exponentials involving differences of these potentials, up to boundary terms that vanish as the neighborhoods shrink. The conformal deformation formula in the multi-radial setting is a key technical step (Fan, 9 Aug 2025).

These examples show that the Onsager-Machlup idea extends beyond elliptic diffusions on Euclidean spaces. It survives in hypoelliptic sub-Riemannian geometry and in conformally invariant random-curve theories, but the functional form is then dictated by the native geometric structure rather than by a universal quadratic template.

7. Variational role, computation, and limitations

The Onsager-Machlup functional is routinely used as a variational principle for most probable paths, but the relevant object is the most probable tube, not the probability of an individual path. This distinction matters computationally. In data assimilation, the path prior for weak-constraint 4D-Var should include the divergence term, while for MCMC path sampling with the Euler scheme the quadratic “energy” functional without divergence is sufficient. In large systems, the divergence and its derivative can be estimated with Hutchinson’s trace estimator,

$\lambda$5

which avoids explicit second-derivative calculations (Sugiura, 2017).

Long-time behavior reveals a limitation. For

$\lambda$6

the Onsager-Machlup functional at fixed transition time $\lambda$7 is well defined, but as $\lambda$8 its infimum becomes unbounded below. The minimizing paths nevertheless have convergent graph subsequences in Fréchet distance, and the graph limit is an extremal of the abbreviated action

$\lambda$9

with optimal energy $f$0. The energy-climbing geometric minimization algorithm (EGMA) is designed to compute this graph limit and the optimal energy simultaneously (Du et al., 2020).

The functional has also become a computational tool in machine learning. In transition path sampling, candidate paths can be interpreted as trajectories under stochastic dynamics induced by the learned score of a pre-trained diffusion or flow-matching model. For a discretized path $f$1, the discrete Onsager-Machlup action is

$f$2

with a smoothness term $f$3, a drift-magnitude term $f$4, and a divergence term $f$5. Minimizing this action with fixed endpoints repurposes pre-trained generative models for zero-shot transition path sampling (Raja et al., 25 Apr 2025).

In stochastic Hamiltonian systems,

$f$6

the Onsager-Machlup functional is the sum of the quadratic deviations of $f$7 and $f$8 from the deterministic Hamiltonian flow, weighted by $f$9 and $\exp(J)$00. The most probable path satisfies the deterministic Hamilton equations, and the associated large deviation principle has a rate function equal to one half of the corresponding Onsager-Machlup action. In the nearly integrable setting, this framework is combined with KAM theory to show that invariant tori remain stable in the most probable sense, with deviations suppressed at an exponential rate determined by the rate function (Zhang et al., 18 Mar 2025).

Several recurring misconceptions are therefore not sustained by the literature. The Onsager-Machlup functional is not a literal probability density of single paths; it depends on the chosen norm, topology, and sometimes the metric; and its direct long-time minimization does not generally define a finite quasi-potential. Its proper scope is more precise: it is a local small-ball asymptotic functional that organizes relative likelihood, variational path selection, and geometry-dependent mode structure across finite-dimensional, infinite-dimensional, degenerate, fractional, jump, and conformally invariant stochastic systems.