Empirical Occupation Measure

Updated 22 June 2026

Empirical occupation measure is a construct that encodes time- or population-averaged state distributions in stochastic processes, labor networks, and economic systems.
It underpins methodologies for quantifying convergence rates, controlling multi-agent systems, and robust statistical inference through semidefinite relaxations and moment-based techniques.
Applications range from ergodic theory and large deviations in statistical physics to mean-field games and network economics, providing tractable approximations for complex dynamics.

An empirical occupation measure is a fundamental construct in probability, statistics, stochastic processes, ergodic theory, statistical physics, mean-field games, and labor economics, encoding the empirical distribution of state or event realizations—typically by averaging over time, space, individuals, or both. It provides a data-driven proxy for the time- or population-averaged law associated with a process or ensemble. Modern applications span from quantifying mixing and convergence rates of stochastic dynamics, to constructing semidefinite relaxations for stochastic control, to measuring skill-task matches across entire economies, and directly informing inference in large-scale agent-based or network settings.

1. Mathematical Definition and Variants

The empirical occupation measure is context-dependent but universally takes the form of an empirical distribution associated with process samples or population instances. For a single stochastic process $(X_t)_{t\geq 0}$ on state space $E$ , the time- $T$ occupation measure is defined by

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

where $\delta_{x}$ is the Dirac mass at $x$ . Analogously, for a discrete-time process $(X_n)_{n=1}^N$ , the occupation measure is

$\mu_N := \frac{1}{N}\sum_{n=1}^N \delta_{X_n}.$

For a finite population of $n$ agents or particles with states $\{X^i_t\}_{i=1}^n$ , the empirical occupation measure at time $E$ 0 is

$E$ 1

as in mean-field models or large deviations for interacting particle systems (Béthencourt et al., 2024).

In control and reinforcement learning, if trajectories $E$ 2 are sampled under a control policy, the empirical occupation measure estimates the state-action-time visitation frequency: $E$ 3 (Holtorf et al., 2022, Yu et al., 17 Mar 2026). In network-economics, empirical occupation measures may encode match frequencies in bipartite graphs of workers and jobs, with matrix-valued analogs (see Section 3 below) (Fogel et al., 2023).

2. Law of Large Numbers and Convergence Rates

The empirical occupation measure $E$ 4 or $E$ 5 converges (almost surely or in probability) to an invariant measure or occupation law under weak regularity conditions—a quantitative formulation of the law of large numbers for processes or populations.

Key Results:

For i.i.d. sampling, $E$ 6 almost surely, with $E$ 7 the $E$ 8-Wasserstein metric; rates are governed by metric entropy/covering numbers: $E$ 9 in $T$ 0 (Boissard et al., 2011, Boissard, 2011, Pagès et al., 7 May 2026).
For ergodic Markov chains, occupation measure convergence rates depend on mixing/contractivity constants, and can achieve similar $T$ 1 or $T$ 2 rates under suitable conditions (Boissard et al., 2011, Pagès et al., 7 May 2026).
For ergodic diffusions on compact spaces, Donsker-type CLTs hold: $T$ 3 for classes of smooth test-functions $T$ 4, yielding $T$ 5 in up to three dimensions (Deo, 2022).
In non-stationary or non-Markovian setups with conditional mixing, convergence holds at similar rates provided suitable coupling and moment conditions (Pagès et al., 7 May 2026).
For time-inhomogeneous Markov chains with vanishing "mutation" (annealing/simulated annealing, evolutionary models), the occupation measure converges almost surely to the zero-noise stochastically stable law at an explicit rate: $T$ 6 (Benaïm et al., 6 Feb 2026).

Implications: These convergence results justify empirical occupation measures for plug-in inference and as tractable approximations of deterministic or expectation-based occupation measure formulations in high-dimensional or complex models.

3. Occupation Measures in Large-Scale Economic and Network Models

In labor economics and network theory, the "empirical occupation measure" can generalize far beyond time averages to encode empirical interaction structure in large bipartite graphs. Notably, (Fogel et al., 2023) applies a data-driven assignment of workers to "types" and jobs to "markets" by modeling the observed worker-job match network via a degree-corrected bipartite stochastic block model. The empirical occupation measure here is the (estimated) block-wise matrix $T$ 7 defining match probabilities between latent worker types ( $T$ 8) and market classes ( $T$ 9).

Pipeline:

Nodes: workers, jobs (occupation-establishment pairs).
Edge weights: counts of realized job holdings.
Block modeling: cluster nodes into types and markets that maximize the likelihood of observed links.
Economic interpretation: latent block structure corresponds to a Roy-equilibrium matching of heterogeneous skills and tasks.

Significance: Empirically optimized occupation measures yield low-dimensional sufficient statistics for predicting wage changes, job flows, and shock propagation, outperforming classical occupation-sector classification systems in explanatory power for wage and flow data.

4. Empirical Occupation Measures in Control and Mean Field Games

Occupation measures are foundational in stochastic control, mean field games, and convex relaxation of optimal control problems. Empirical occupation measures are constructed from sampled or simulated trajectories and serve as statistically consistent proxies for theoretical (pre-averaged) occupation measures.

Applications:

Semidefinite programming (SDP) relaxations: Moment-based constraints on empirical occupation measures can be embedded directly into SDP frameworks, yielding data-driven or robust optimization problems (Holtorf et al., 2022).
Frank–Wolfe optimization over measure spaces: Empirical occupation measures enter both the forward population simulation and the aggregation of control policies in large multi-agent systems (e.g., UAV swarm, satellite constellations), converging as $\mu_T := \frac{1}{T}\int_0^T \delta_{X_t}\,dt$ 0 to mean-field limits (Yu et al., 17 Mar 2026).
Empirical moments: For each trajectory, observed empirical occupation moments are

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

and used to enforce affine constraints and to inform lower bounds for optimal values (Holtorf et al., 2022).

Propagation of chaos: As $\mu_T := \frac{1}{T}\int_0^T \delta_{X_t}\,dt$ 2, empirical occupation measures converge to law-determined occupation measures; sampling error scales like $\mu_T := \frac{1}{T}\int_0^T \delta_{X_t}\,dt$ 3 (Yu et al., 17 Mar 2026).

Significance: Empirical occupation measures allow sample-based plug-in control synthesis, verification, and robustification of stochastic optimal control and mean-field equilibrium problems intractable by direct analytic means.

5. Large Deviations and Statistical Physics

Occupation measures in statistical mechanics and interacting particle systems encode spatial and temporal averages of system configurations or trajectories. Large deviations principles (LDPs) for empirical occupation measures provide variational characterizations of rare event probabilities.

Main developments:

For symmetric exclusion and diffusive interacting systems, large deviations rate functionals $\mu_T := \frac{1}{T}\int_0^T \delta_{X_t}\,dt$ 4 quantify the cost to enforce atypical empirical density profiles, often with quadratic Onsager–Machlup structure (Landim et al., 2017).
In Brownian motion, conditioning on extreme trajectory self-intersections or mutual interactions leads to occupation measures concentrating on optimal profiles determined by Sobolev or Gagliardo–Nirenberg extremals (Park, 9 Apr 2026).
Compactification and profile decomposition: By taking quotient spaces under spatial shifts and allowing splitting into multiple "islands", full LDPs are established, with minimizers reflecting system symmetries and variational constraints (Park, 9 Apr 2026).

Significance: Empirical occupation measures become canonical random objects for describing the macroscopic empirical profile of stochastic systems, both in typical and in large deviation regimes.

6. Empirical Occupation Measures in Stochastic Control of Particle Ensembles

For $\mu_T := \frac{1}{T}\int_0^T \delta_{X_t}\,dt$ 5 controlled Brownian particles, the empirical occupation measure $\mu_T := \frac{1}{T}\int_0^T \delta_{X_t}\,dt$ 6 encodes the time- $\mu_T := \frac{1}{T}\int_0^T \delta_{X_t}\,dt$ 7 spatial distribution of all particles. Modern theory (see (Béthencourt et al., 2024)) analyzes controlled dynamics where the cost depends on $\mu_T := \frac{1}{T}\int_0^T \delta_{X_t}\,dt$ 8, and the value function $\mu_T := \frac{1}{T}\int_0^T \delta_{X_t}\,dt$ 9 solves an HJB equation in the Wasserstein space of probability measures: $\delta_{x}$ 0 where $\delta_{x}$ 1 denotes the Lions derivative, and $\delta_{x}$ 2 is a Hamiltonian functional encoding feedback coupling through the measure.

A generalized Itô formula for functionals of the occupation measure allows derivation of both the measure-space HJB and its probabilistic representation via Boué–Dupuis variational formulas and nonlinear Feynman–Kac representations. This machinery underpins control of macroscopic observables, e.g., maximizing the Wiener sausage volume in Brownian ensembles, with empirical occupation measures encoding the ensemble configuration (Béthencourt et al., 2024).

7. Statistical Inference, Robustness, and Empirical Use

Empirical occupation measures undergird statistical inference, estimation, and verification for stochastic process models:

Nonasymptotic deviation inequalities (in Wasserstein metrics, under transport-entropy or contractivity conditions) quantify the sampling error between occupation measures and their limiting laws (Boissard et al., 2011, Boissard, 2011, Pagès et al., 7 May 2026).
Moment-based constraints following from empirical occupation measures can rigorously quantify confidence intervals or robustify solutions in stochastic optimization (Holtorf et al., 2022, Yu et al., 17 Mar 2026).
Bootstrapping and finite-sample scaling: The precise scaling of empirical occupation measures allows robust design and verification even in high-dimensional, non-i.i.d., or networked agent settings.
In ergodic estimation, empirical occupation measures serve as consistent estimators for invariant laws and mixing times, justifying their use in MCMC and SDE simulation output analysis (Pagès et al., 7 May 2026, Deo, 2022).