Gradient Flows in Random-Measure Spaces

• Gradient flows in random-measure spaces are the evolution of probability measures under the steepest descent of functionals within a geometric framework.
• The theory leverages Wasserstein metrics along with EVI and EDE formulations to ensure uniqueness, contraction, and quantitative stability even in stochastic settings.
• Key applications include stochastic control, mean-field dynamics, and interacting particle systems, with proven convergence rates and energy dissipation properties.

Gradient flows in random-measure spaces study the evolution of probability measures (or more generally, random measures) under the steepest descent of a functional, with respect to a chosen geometric structure. This theory generalizes classical gradient flows in Euclidean or Hilbert spaces, extending them to the nonlinear, often infinite-dimensional setting of measure spaces equipped with Wasserstein or related metrics, and accommodates randomness either in the underlying space, the measures themselves, or the functionals. The interplay of curvature, convexity, and underlying geometry is central, as are applications to stochastic control, mean-field dynamics, probabilistic sampling, and statistical physics.

1. Geometric Foundations: Metric Measure Spaces and Distortion Distances

The foundational object is the metric measure space (mm-space), denoted $(X,d,\mu)$, where $X$ is a Polish (complete, separable) space, $d$ is a metric, and $\mu$ is a Borel probability measure. To compare two mm-spaces $(X_0, d_0, \mu_0)$ and $(X_1, d_1, \mu_1)$, the L²-distortion distance is defined as

$$D((X_0,d_0,\mu_0), (X_1,d_1,\mu_1)) := \inf_{\pi \in \mathrm{Cpl}(\mu_0,\mu_1)} \left[ \iint_{(X_0 \times X_1)^2} |d_0(x,x') - d_1(y,y')|^2 \, d\pi(x,y)\, d\pi(x',y') \right]^{1/2},$$

where $\mathrm{Cpl}(\mu_0,\mu_1)$ denotes couplings of $\mu_0$ and $\mu_1$ (Sturm, 2012).

This structure induces a metric space $X$0 of mm-spaces, with well-defined geodesics: given an optimal coupling $X$1, the interpolated mm-space $X$2, where

$X$3

yields constant-speed geodesics in $X$4.

The tangent cone at a point $X$5 is identified with equivalence classes $X$6 modulo the symmetry group of mm-space automorphisms, with the tangent metric given by

$X$7

This tangent structure yields an Alexandrov tangent cone of nonnegative curvature (Sturm, 2012), making $X$8 a geodesic space of nonnegative Alexandrov curvature.

2. Classes of Functionals and Semiconvexity on Random-Measure Spaces

Functionals of interest are often semiconvex along geodesics in $X$9 or Wasserstein spaces. Main examples include:

• Polynomial functionals: $d$0, where $d$1 is $d$2 with bounded derivatives. Such functionals are $d$3-Lipschitz if $d$4 is $d$5-Lipschitz and $d$6-convex if $d$7 (Sturm, 2012).
• Nested polynomials (“order-2”): $d$8, with precise convexity/Lipschitz bounds in terms of $d$9 (Sturm, 2012).

Key properties—Lipschitz continuity and $\mu$0-convexity—transfer explicitly from $\mu$1, $\mu$2, and $\mu$3 to $\mu$4. Such functionals include regularized energies, interaction potentials, and integrals of geometrically meaningful quantities, crucial in statistical physics, geometric analysis, and optimal transport.

Lower semicontinuity required for flow well-posedness follows from continuity and tightness of the couplings (Sturm, 2012).

3. Gradient Flow Formulations: EVI and EDE in Random-Measure Spaces

For functionals $\mu$5 that are $\mu$6-semiconvex, gradient flows in metric measure settings are characterized by two equivalent formulations:

• Evolution Variational Inequality (EVI$\mu$7):

$\mu$8

• Energy Dissipation Equality (EDE):

$\mu$9

where $(X_0, d_0, \mu_0)$0 is the metric speed and $(X_0, d_0, \mu_0)$1 is the local descending slope.

In these settings, EVI implies uniqueness, contraction of flows, and quantitative stability. Specifically, if $(X_0, d_0, \mu_0)$2 is $(X_0, d_0, \mu_0)$3-convex, there is exponential contraction $(X_0, d_0, \mu_0)$4 and exponential convergence to the unique minimizer for $(X_0, d_0, \mu_0)$5 (Sturm, 2012). The chain rule and slope-speed identity are verified under $(X_0, d_0, \mu_0)$6-convexity and lower semicontinuity.

4. Gradient Flows of Random Measures and Pathwise Uniqueness

Randomness enters the theory when the measure $(X_0, d_0, \mu_0)$7 defining the mm-space is replaced by a random measure $(X_0, d_0, \mu_0)$8 on a probability space $(X_0, d_0, \mu_0)$9, yielding a random mm-space $(X_1, d_1, \mu_1)$0. Pathwise gradient flows are well-posed under the following conditions:

• Almost sure nonnegative Alexandrov curvature: Ensures existence and uniqueness of geodesics in each realization.
• Measurability: Of both the couplings and the induced flow map with respect to $(X_1, d_1, \mu_1)$1.
• Uniform convexity/Lipschitz constants in $(X_1, d_1, \mu_1)$2: Guarantees stochastic stability and well-definiteness (Sturm, 2012).

Given these, for $(X_1, d_1, \mu_1)$3-almost every $(X_1, d_1, \mu_1)$4, there exists a unique gradient flow satisfying the EVI$(X_1, d_1, \mu_1)$5 and the associated contractivity property almost surely. Expectation bounds on contraction and energy dissipation follow via Jensen's inequality, enabling rigorous control of ensemble-averaged behavior.

A closely related construction defines a random measure space as a stochastic process valued in a Wasserstein space (e.g., in stochastic control and mean-field SDEs), admitting a gradient flow on the space of measure-valued processes under suitable Wasserstein or L²-type metrics (Šiška et al., 2020).

5. Convergence Rates, Uniqueness, and Contractivity

Existence and uniqueness of the downward gradient flow for any $(X_1, d_1, \mu_1)$6-Lipschitz, $(X_1, d_1, \mu_1)$7-convex functional in a complete geodesic mm-space of nonnegative curvature are guaranteed, with the contractive property

$(X_1, d_1, \mu_1)$8

for two solutions $(X_1, d_1, \mu_1)$9 and $$D((X_0,d_0,\mu_0), (X_1,d_1,\mu_1)) := \inf_{\pi \in \mathrm{Cpl}(\mu_0,\mu_1)} \left[ \iint_{(X_0 \times X_1)^2} |d_0(x,x') - d_1(y,y')|^2 \, d\pi(x,y)\, d\pi(x',y') \right]^{1/2},$$0. For strictly positive $$D((X_0,d_0,\mu_0), (X_1,d_1,\mu_1)) := \inf_{\pi \in \mathrm{Cpl}(\mu_0,\mu_1)} \left[ \iint_{(X_0 \times X_1)^2} |d_0(x,x') - d_1(y,y')|^2 \, d\pi(x,y)\, d\pi(x',y') \right]^{1/2},$$1, all flows converge exponentially fast to the unique equilibrium minimizer; for $$D((X_0,d_0,\mu_0), (X_1,d_1,\mu_1)) := \inf_{\pi \in \mathrm{Cpl}(\mu_0,\mu_1)} \left[ \iint_{(X_0 \times X_1)^2} |d_0(x,x') - d_1(y,y')|^2 \, d\pi(x,y)\, d\pi(x',y') \right]^{1/2},$$2, non-expansion holds (Sturm, 2012, Sturm, 2014).

In the random measure case, the expectation contractivity and convergence rates are preserved: for random initial data or coefficients, the same exponential bounds apply almost surely, provided the uniformity conditions above are met. This underpins stability results in stochastic control, mean-field games, and measure-valued Markov processes (Šiška et al., 2020).

6. Applications and Extensions

Key applications include:

• Stochastic Control and Reinforcement Learning: Gradient flows for regularized stochastic control problems are constructed on the space of random admissible relaxed controls, endowed with a path-space $$D((X_0,d_0,\mu_0), (X_1,d_1,\mu_1)) := \inf_{\pi \in \mathrm{Cpl}(\mu_0,\mu_1)} \left[ \iint_{(X_0 \times X_1)^2} |d_0(x,x') - d_1(y,y')|^2 \, d\pi(x,y)\, d\pi(x',y') \right]^{1/2},$$3–Wasserstein metric. The flow yields monotonic decrease of the cost functional, stationarity characterized by Pontryagin-type optimality, and (under convexity) exponential convergence. Posterior distributions are characterized as Gibbs-type formulae, enabling Bayesian interpretations (Šiška et al., 2020).
• Interacting Particle Systems and Infinite-Dimensional Flows: The theory extends to infinite-dimensional configuration spaces, such as the space of locally finite point measures with the $$D((X_0,d_0,\mu_0), (X_1,d_1,\mu_1)) := \inf_{\pi \in \mathrm{Cpl}(\mu_0,\mu_1)} \left[ \iint_{(X_0 \times X_1)^2} |d_0(x,x') - d_1(y,y')|^2 \, d\pi(x,y)\, d\pi(x',y') \right]^{1/2},$$4-matching distance. Evolution variational inequalities characterize the gradient flow of relative entropy (Kullback-Leibler divergence), and the RCD$$D((X_0,d_0,\mu_0), (X_1,d_1,\mu_1)) := \inf_{\pi \in \mathrm{Cpl}(\mu_0,\mu_1)} \left[ \iint_{(X_0 \times X_1)^2} |d_0(x,x') - d_1(y,y')|^2 \, d\pi(x,y)\, d\pi(x',y') \right]^{1/2},$$5 condition, HWI, and Brunn-Minkowski inequalities propagate to these infinite particle limits (Suzuki, 8 Sep 2025).
• Functional Inequalities and Rigidity: Entropy flows in such spaces preserve number-rigidity and tail-triviality of determinantal point processes, with implications for random matrix theory and statistical mechanics (Suzuki, 8 Sep 2025).
• Quantitative Optimization and Convergence: Recent results show that even in the absence of displacement convexity, combining diffusive regularization and linear convexity yields $$D((X_0,d_0,\mu_0), (X_1,d_1,\mu_1)) := \inf_{\pi \in \mathrm{Cpl}(\mu_0,\mu_1)} \left[ \iint_{(X_0 \times X_1)^2} |d_0(x,x') - d_1(y,y')|^2 \, d\pi(x,y)\, d\pi(x',y') \right]^{1/2},$$6 rates for drift-diffusion PDEs as Wasserstein gradient flows, with exponential rates possible under strong convexity. This broadens the class of mean-field Langevin dynamics and sampling algorithms admitting quantitative guarantees (Chizat et al., 16 Jul 2025).
• Stochastic Laws of Large Numbers: Pathwise contraction, well-posedness, and propagation of chaos can be deduced for measure-valued and random initial data, connecting to interacting particle approximations and sampling.

7. Broader Significance and Open Problems

Gradient flow theory in random-measure spaces systematizes and unifies diverse domains: stochastic control, mean-field PDEs, infinite-dimensional dynamics, and optimization over probability distributions. Essential techniques include:

• EVI and EDE formulations for non-smooth and random functionals.
• Explicit Lipschitz/convexity transfer from functionals on spaces of measures.
• Stability, uniqueness, and contraction via curvature and convexity.
• Measure-theoretic and probabilistic techniques for randomizations.

Open challenges remain—extending displacement convexity to non-convex or random functionals, discretization and numerics for high-dimensional or measure-valued settings, handling partial observations or non-entropic regularization, and exploring rigidity phenomena beyond determinantal processes.

The stochastic gradient flow paradigm for random measures is thus a unifying thread linking the analysis of geometric evolution equations, stochastic process theory, and high-dimensional optimization in random environments (Sturm, 2012, Šiška et al., 2020, Chizat et al., 16 Jul 2025, Suzuki, 8 Sep 2025, Sturm, 2014).