Donsker–Varadhan Action Functional

• Donsker–Varadhan action functional is the large-deviation rate functional for occupation measures in Markov processes, capturing long-time behavior.
• It offers generator-based, pressure-based, and entropy-rate variational representations that clarify its role in both reversible and nonequilibrium systems.
• It underpins the analysis of metastability, empirical currents, and convergence properties in finite, infinite-dimensional, and interacting-particle settings.

The Donsker–Varadhan action functional is the large-time rate functional for occupation measures and empirical processes of Markov dynamics. In its classical level-2 form, it governs the asymptotics of empirical measures such as

$\mu_T=\frac{1}{T}\int_0^T \delta_{X_t}\,dt,$

through a large-deviation principle of the form $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$. It admits generator-based, pressure-based, and entropy-rate representations; in reversible settings it reduces to a Dirichlet form, while away from detailed balance it can encode empirical currents, entropy production, or excess dynamical activity (Maes et al., 2011, Zhao, 17 Jun 2025, Hoppenau et al., 2016).

1. Classical definition and level structure

For a finite-state irreducible continuous-time Markov jump process on a finite set $K$ with rates $k(x,y)\ge 0$ and generator

$(Lg)(x)=\sum_{y\in K}k(x,y)\,[g(y)-g(x)],$

the empirical occupation measure of a trajectory $\omega=(x_s,0\le s<T)$ is

$$p_T(\omega,x)=\frac{1}{T}\int_0^T \delta_{x_s,x}\,ds.$$

Under the stationary process $P_\rho$, one has

$$\mathbb P_\rho[p_T\simeq \mu]\asymp \exp\{-T\,I(\mu)\},\qquad T\to\infty,$$

with a strictly convex rate functional $I$ and unique minimum $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$0 in the irreducible finite-state setting (Maes et al., 2011).

The same object appears far beyond finite-state chains. For Markov processes on Polish or infinite-dimensional spaces, the empirical measure

$$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$1

is the canonical level-2 observable, while level-3 formulations lift the large deviations to empirical processes on path space. For the white-forced 2D Navier–Stokes system in a bounded domain, both the level-2 empirical measure and the level-3 empirical process satisfy Donsker–Varadhan type large deviation principles with good rate functions (Zhao, 17 Jun 2025). For discrete-time chains, the analogous empirical measure is

$$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$2

and the action functional is defined by the corresponding discrete-time variational formula (Chen et al., 28 Oct 2025).

A basic structural distinction is between level 2, which concerns one-time occupation statistics, level 2.5, which incorporates empirical currents or flows in addition to empirical densities, and level 3, which is formulated on stationary path-space measures. This hierarchy becomes essential away from detailed balance, because density fluctuations alone no longer capture the full nonequilibrium structure (Hoppenau et al., 2016, Bertini et al., 2022).

2. Variational representations and spectral duality

The classical generator representation for a finite-state continuous-time chain is

$$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$3

and, more generally, for a Markov generator $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$4 on a Polish space,

$$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$5

In discrete time, with one-step Markov operator $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$6, the corresponding formula becomes

$$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$7

These are the canonical Donsker–Varadhan formulas for occupation measures (Maes et al., 2011, Chen et al., 28 Oct 2025).

A complementary representation uses the Feynman–Kac pressure. For bounded continuous potentials $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$8, one defines the pressure by long-time logarithmic growth of the Feynman–Kac semigroup, and the action functional is its Legendre transform: $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$9 For the white-forced 2D Navier–Stokes system this representation is central, and the analysis uses generalized Markov semigroups, principal eigenvalues and eigenvectors, and a resolvent approximation scheme to justify the generator formula in a non-compact infinite-dimensional phase space (Zhao, 17 Jun 2025). In the discrete-time NLS setting, the analogous pressure is

$K$0

and the action functional is

$K$1

with the equivalent entropy formula

$K$2

(Chen et al., 28 Oct 2025).

At level 3, the Donsker–Varadhan functional is an entropy rate on path space. For shift-invariant path measures $K$3, the level-3 functional for the 2D Navier–Stokes system is given by the entropy production per unit time of $K$4 relative to the reference Markov dynamics conditioned on the past, and the level-2 functional is obtained by contraction to one-time marginals (Zhao, 17 Jun 2025). For small-noise diffusions on $K$5, the level-3 Donsker–Varadhan functional at fixed noise strength is likewise the relative entropy per unit time with respect to the stationary path law (Bertini et al., 2022).

Setting	Donsker–Varadhan representation	Source
Continuous-time finite-state chain	$K$6	(Maes et al., 2011)
General continuous-time Markov process	$K$7	(Chen et al., 28 Oct 2025)
Pressure representation	$K$8	(Zhao, 17 Jun 2025)
Discrete-time chain	$K$9	(Chen et al., 28 Oct 2025)
Reversible case	$k(x,y)\ge 0$0	(Maes et al., 2011)
Level 3	entropy production per unit time on path space	(Zhao, 17 Jun 2025)

3. Reversible structure, empirical current, and metastable decomposition

Under detailed balance, the Donsker–Varadhan action functional becomes an explicit Dirichlet form. For a reversible finite-state chain with invariant law $k(x,y)\ge 0$1 and $k(x,y)\ge 0$2,

$k(x,y)\ge 0$3

For reversible diffusions

$k(x,y)\ge 0$4

with Gibbs invariant measure $k(x,y)\ge 0$5, one has

$k(x,y)\ge 0$6

and $k(x,y)\ge 0$7 if $k(x,y)\ge 0$8 (Maes et al., 2011, Landim et al., 16 Sep 2025).

Away from detailed balance, the empirical current must be incorporated. For overdamped Langevin dynamics, the level-2.5 rate functional for the empirical density $k(x,y)\ge 0$9 and empirical current $(Lg)(x)=\sum_{y\in K}k(x,y)\,[g(y)-g(x)],$0 is

$(Lg)(x)=\sum_{y\in K}k(x,y)\,[g(y)-g(x)],$1

with the stationary constraint $(Lg)(x)=\sum_{y\in K}k(x,y)\,[g(y)-g(x)],$2. The level-2 functional $(Lg)(x)=\sum_{y\in K}k(x,y)\,[g(y)-g(x)],$3 is recovered by contraction over divergence-free currents. In the reversible case, this contraction reproduces the Dirichlet-form representation. The same formalism gives a decomposition into adiabatic and non-adiabatic entropy production, and the Donsker–Varadhan functional equals one quarter of the non-adiabatic entropy production rate in equilibrium (Hoppenau et al., 2016).

For small-noise reversible diffusions, the level-2 action itself has a metastable hierarchy. The expansion

$(Lg)(x)=\sum_{y\in K}k(x,y)\,[g(y)-g(x)],$4

decomposes the functional into a microscopic term $(Lg)(x)=\sum_{y\in K}k(x,y)\,[g(y)-g(x)],$5, an order-one term $(Lg)(x)=\sum_{y\in K}k(x,y)\,[g(y)-g(x)],$6 supported on critical points, and exponentially slow contributions $(Lg)(x)=\sum_{y\in K}k(x,y)\,[g(y)-g(x)],$7 associated with effective finite-state Markov chains between wells on metastable time-scales $(Lg)(x)=\sum_{y\in K}k(x,y)\,[g(y)-g(x)],$8 (Landim et al., 16 Sep 2025). In that setting, the Donsker–Varadhan action functional is not merely a single asymptotic cost: it resolves the full hierarchy of inter-well relaxation scales.

4. Dynamical activity and monotone relaxation

For finite-state nonequilibrium Markov jump processes, the Donsker–Varadhan functional has an exact representation in terms of dynamical activity. If $(Lg)(x)=\sum_{y\in K}k(x,y)\,[g(y)-g(x)],$9 is the potential such that $\omega=(x_s,0\le s<T)$0 is stationary for the tilted rates

$\omega=(x_s,0\le s<T)$1

then

$\omega=(x_s,0\le s<T)$2

and therefore

$\omega=(x_s,0\le s<T)$3

The functional is thus the excess of expected dynamical activity of the original dynamics over that of the minimally active tilted dynamics making $\omega=(x_s,0\le s<T)$4 stationary (Maes et al., 2011).

This interpretation is specific to nonequilibrium. Under detailed balance, monotonicity of relative entropy is classical, but away from detailed balance the Donsker–Varadhan functional does not derive from relative entropy and need not be monotone at short times. For the Markov evolution $\omega=(x_s,0\le s<T)$5, eventual monotonicity is nevertheless rigorously available under a sector condition, or “normal linear-response” condition,

$\omega=(x_s,0\le s<T)$6

Under that assumption there exists $\omega=(x_s,0\le s<T)$7, of order the relaxation time, such that

$\omega=(x_s,0\le s<T)$8

The proof expands the attained variational representation along the evolution, identifies a leading quadratic form

$\omega=(x_s,0\le s<T)$9

and shows that the exponentially decaying remainder becomes negligible at sufficiently large times (Maes et al., 2011).

The same work links this sector condition to a susceptibility identity. For a perturbation

$$p_T(\omega,x)=\frac{1}{T}\int_0^T \delta_{x_s,x}\,ds.$$0

the generalized susceptibility satisfies

$$p_T(\omega,x)=\frac{1}{T}\int_0^T \delta_{x_s,x}\,ds.$$1

Hence normal linear response corresponds to initial decay of $$p_T(\omega,x)=\frac{1}{T}\int_0^T \delta_{x_s,x}\,ds.$$2. This relation gives the monotonicity theorem a direct linear-response interpretation (Maes et al., 2011).

The finite-state examples are sharply contrasted. For the totally asymmetric random walk on a homogeneous ring,

$$p_T(\omega,x)=\frac{1}{T}\int_0^T \delta_{x_s,x}\,ds.$$3

and $$p_T(\omega,x)=\frac{1}{T}\int_0^T \delta_{x_s,x}\,ds.$$4 for all $$p_T(\omega,x)=\frac{1}{T}\int_0^T \delta_{x_s,x}\,ds.$$5. By contrast, for an inhomogeneous asymmetric ring, explicit perturbative choices of the centered potential can produce a positive initial derivative, so short-time non-monotonicity occurs even though eventual monotonic decay is recovered later (Maes et al., 2011).

5. Infinite-dimensional, interacting-particle, and controlled extensions

The Donsker–Varadhan action functional persists in genuinely infinite-dimensional stochastic systems. For the white-forced 2D Navier–Stokes equation in a bounded domain, under $$p_T(\omega,x)=\frac{1}{T}\int_0^T \delta_{x_s,x}\,ds.$$6 for all modes and $$p_T(\omega,x)=\frac{1}{T}\int_0^T \delta_{x_s,x}\,ds.$$7, the empirical measure satisfies a uniform level-2 large deviation principle with good rate function

$$p_T(\omega,x)=\frac{1}{T}\int_0^T \delta_{x_s,x}\,ds.$$8

and the empirical process on path space satisfies a level-3 large deviation principle with rate given by entropy production per unit time relative to the Markov dynamics (Zhao, 17 Jun 2025). For the locally damped, randomly forced NLS sampled at integer times, the discrete-time occupation measures satisfy a uniform large deviation principle with a convex good rate function given by the pressure transform and by the discrete entropy formula

$$p_T(\omega,x)=\frac{1}{T}\int_0^T \delta_{x_s,x}\,ds.$$9

and the proof hinges on principal eigenvalues of the Feynman–Kac semigroup and a bootstrap argument yielding Lipschitz estimates (Chen et al., 28 Oct 2025). For path-distribution dependent SDEs and SPDEs, where the original dynamics is non-Markovian, the action functional is identified with the Donsker–Varadhan functional of a reference Markov equation obtained by freezing the law argument at the invariant measure; the transfer from the reference equation to the original one uses exponential equivalence of occupation measures (Ren et al., 2020).

In interacting-particle systems and hydrodynamic limits, the functional survives but changes level and scale. For the weakly asymmetric exclusion process on the $P_\rho$0-dimensional torus, the microscopic level-3 Donsker–Varadhan functional $P_\rho$1 for the empirical process, when projected to empirical density and current and rescaled by $P_\rho$2, converges to the macroscopic quadratic action of macroscopic fluctuation theory (Bertini et al., 2021). For the open one-dimensional symmetric simple exclusion process with reservoirs, the time-averaged two-point correlation field satisfies a large deviation principle with a Donsker–Varadhan type rate functional

$P_\rho$3

where $P_\rho$4 combines a Laplacian term, a nonlocal bilinear operator $P_\rho$5, a diagonal Neumann contribution, and a reservoir source term (Bodineau et al., 2022). In that setting, the functional encodes a variety of scales depending on the observable of interest.

Risk-sensitive control produces a controlled version of the Donsker–Varadhan principle. In discrete time, with controlled kernel $P_\rho$6 and reward $P_\rho$7, the optimal long-run risk-sensitive reward is represented by

$P_\rho$8

where the relative entropy term is the action functional for deviations of the empirical transition kernel from the nominal controlled dynamics (Arapostathis et al., 2019).

6. Degenerate, nonlocal, and conceptual boundary cases

The classical irreducible picture does not persist universally. For degenerate continuous-time jump Markov processes on a compact metric space with absorbing states $P_\rho$9, the joint level-2.5 rate function for empirical measure $$\mathbb P_\rho[p_T\simeq \mu]\asymp \exp\{-T\,I(\mu)\},\qquad T\to\infty,$$0 and flow $$\mathbb P_\rho[p_T\simeq \mu]\asymp \exp\{-T\,I(\mu)\},\qquad T\to\infty,$$1 is

$$\mathbb P_\rho[p_T\simeq \mu]\asymp \exp\{-T\,I(\mu)\},\qquad T\to\infty,$$2

under the equal-marginals constraint $$\mathbb P_\rho[p_T\simeq \mu]\asymp \exp\{-T\,I(\mu)\},\qquad T\to\infty,$$3, and $$\mathbb P_\rho[p_T\simeq \mu]\asymp \exp\{-T\,I(\mu)\},\qquad T\to\infty,$$4 otherwise. By contraction, the level-2 Donsker–Varadhan functional retains the canonical generator form

$$\mathbb P_\rho[p_T\simeq \mu]\asymp \exp\{-T\,I(\mu)\},\qquad T\to\infty,$$5

but its zero level set is no longer a singleton: it contains convex combinations of the stationary measure $$\mathbb P_\rho[p_T\simeq \mu]\asymp \exp\{-T\,I(\mu)\},\qquad T\to\infty,$$6 and measures supported on the absorbing set $$\mathbb P_\rho[p_T\simeq \mu]\asymp \exp\{-T\,I(\mu)\},\qquad T\to\infty,$$7 (Basile et al., 2013).

For the zig-zag process, a nonreversible piecewise deterministic Markov process on position–velocity space, the empirical measure satisfies a Donsker–Varadhan large deviation principle on the torus and in one-dimensional Euclidean space. In the compact case the functional can be written explicitly; if $$\mathbb P_\rho[p_T\simeq \mu]\asymp \exp\{-T\,I(\mu)\},\qquad T\to\infty,$$8 and $$\mathbb P_\rho[p_T\simeq \mu]\asymp \exp\{-T\,I(\mu)\},\qquad T\to\infty,$$9, then finiteness already requires the streaming compatibility condition $I$0. Under that condition, the action involves logarithmic, $I$1, and square-root terms built from $I$2 and the switching rates $I$3, and it is strictly decreasing in the refreshment rate $I$4 away from stationarity (Bierkens et al., 2019). This explicit formula is a nonreversible, non-diffusive analogue of the Dirichlet representation.

For nonlocal elliptic operators $I$5, the action functional

$I$6

admits a nonlocal energy representation. In the pure diffusion case $I$7, if $I$8 with $I$9 sufficiently regular,

$$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$00

For the general transport case, the functional contains the nonlocal energy $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$01, the transport term $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$02, and an explicitly controlled nonnegative remainder. This structure is then used to solve a nonlocal inverse problem: equality of the spectral data $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$03 for all bounded domains and potentials forces equality of the local coefficients $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$04, and under an additional decomposition of the matrices one obtains $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$05 (Dávila et al., 2020).

7. Failures, distinctions, and open directions

The Donsker–Varadhan framework has well-defined boundaries. For renewal processes, the empirical measure of backward and forward recurrence times does satisfy a large deviation principle, but the resulting rate functional differs from the Donsker–Varadhan prediction. In particular, the natural Donsker–Varadhan candidate $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$06 is not good when the waiting-time law has only a finite exponential moment range, and the correct rate functional allows mixtures with a point mass at $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$07, yielding non-strict convexity and non-analyticity (Lefevere et al., 2010). This is an explicit case in which the Donsker–Varadhan theory does not apply and the correct occupation-measure rate functional is not of Donsker–Varadhan form.

A related pathology appears for probabilistic cellular automata viewed as discrete-time Markov chains on infinite product spaces. For a wide class of PCA with synchronous product structure, strict positivity, and finite influence range, finite Donsker–Varadhan action implies boundary invariance. Under spatial shift invariance, this sharpens to a dichotomy: $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$08 so for shift-invariant measures the action functional is effectively $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$09 on time-invariant measures and $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$10 otherwise (Eizenberg, 1 Sep 2025). The finite-action class is therefore extremely narrow.

The Donsker–Varadhan action functional is also distinct from the Freidlin–Wentzell action. In the joint limit of large time and vanishing noise for diffusions on $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$11, the level-3 Donsker–Varadhan functional $$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$12, which is a relative entropy rate per unit time, satisfies

$$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$13

where

$$\mathbb P(\mu_T\simeq \mu)\asymp \exp\{-T\,I(\mu)\}$$14

is the Freidlin–Wentzell action per unit time (Bertini et al., 2022). The distinction is therefore structural: Donsker–Varadhan describes long-time occupation fluctuations of Markov processes, whereas Freidlin–Wentzell describes small-noise path deviations; the two coincide only after a nontrivial joint asymptotic procedure.

Several open directions are explicitly identified in the literature. In nonequilibrium monotonicity theory, the normal linear-response sector condition is sufficient but not necessary, and short-time non-monotonicity may coexist with eventual decay (Maes et al., 2011). In infinite-dimensional SPDEs, extensions beyond the treated non-compact settings require additional functional-analytic control (Zhao, 17 Jun 2025). A plausible implication is that the modern theory of the Donsker–Varadhan action functional is best viewed not as a single closed formula, but as a family of equivalent variational objects whose precise form depends on the level of description, the geometry of the state space, reversibility properties, and the presence or absence of degeneracy.