Stochastic Riemannian Metric

• Stochastic Riemannian metric is a field of random or learned positive-definite matrices that defines local distances on manifolds and latent spaces.
• It is constructed via probabilistic models and stochastic dynamics, facilitating curvature-adaptive exploration in applications such as generative modeling and sampling.
• The framework underpins stability analyses and contraction proofs in stochastic flows, while also inspiring efficient surrogate metrics for high-dimensional systems.

A stochastic Riemannian metric is a (possibly random or learned) field of positive-definite matrices that defines infinitesimal distances on a manifold or latent space, where the metric itself is constructed, modified, or evolved via stochastic dynamics, probabilistic models, or as a consequence of uncertainty and random perturbations. Its central role is in characterizing geometry for stochastic processes, latent manifolds in generative models, and random flows in physical and statistical systems, often enabling curvature-dependent exploration, inference, or stability guarantees.

1. Mathematical Definition and Constructions

A stochastic Riemannian metric is a mapping that, for each point $x$ on a manifold $M$ (possibly with additional randomness $\omega$ or latent variables $z$), assigns a symmetric, positive-definite matrix $g_{ij}(x)$ (or $g_{ij}(\omega,x,t)$ if time-dependent/random). Infinitesimal distances are measured by $d\ell^2 = dx^T g(x) dx$, and global distances by geodesic length $$d_g(x, y)^2 = \int_0^1 \dot{\gamma}(s)^T g(\gamma(s)) \dot{\gamma}(s) ds$$, where $\gamma(s)$ is a geodesic.

A typical construction arises in latent generative models, where a stochastic mapping $x = g_\theta(z, \epsilon)$ induces a random Jacobian $J_g(z,\epsilon)$, and averaging over the noise gives a metric in latent space: $$M_\theta(z) = \mathbb{E}_\epsilon [J_g(z, \epsilon)^T J_g(z, \epsilon)].$$ This metric, termed the stochastic pull-back metric, incorporates both the mean and variance of the generative process (Arvanitidis et al., 2021).

For stochastic processes on manifolds, randomness is incorporated directly in the metric: $g_{ij}(\omega,x,t)$ defines a random (or time-dependent) metric field, suitable for analysis of SPDEs, random flows, and contraction properties (Elliott et al., 2012, Pham et al., 2013).

2. Stochastic Metrics in Generative Modeling

In generative models, especially VAEs, the mapping from latent $z$ to data $x$ is often stochastic. The resulting latent geometry is captured by the stochastic Riemannian metric $M_\theta(z)$, which combines curvature from the decoder mean and uncertainty from variance: $$M_\theta(z) = J_{\mu_\theta}(z)^T J_{\mu_\theta}(z) + J_{\sigma_\theta}(z)^T J_{\sigma_\theta}(z).$$ Distances in latent space are scaled in regions with high model uncertainty, yielding metrics that inform geodesics, clustering, and statistical inference (Arvanitidis et al., 2021).

Computational challenges in evaluating $M_\theta(z)$ (especially for high-dimensional, heteroscedastic decoders) motivate efficient surrogates. A notable solution is the conformal surrogate metric, which replaces $M_\theta(z)$ with $M_\psi(z) = m(z) I_d$, where the conformal factor is a function of a learned energy-based prior: $m(z) = [\alpha \nu_\psi(z) + \beta]^{-2/d},$ with $\nu_\psi(z)$ an energy-based density, trained jointly with the generative model. This surrogate retains geometric preference for high-density latent regions and dramatically enhances computational stability (Arvanitidis et al., 2021).

3. Stochastic Riemannian Metrics in Stochastic Differential Geometry

Stochastic metrics characterize evolution and contraction properties of stochastic flows on manifolds. A system with state-dependent metric $M(x, t) = \Theta(x,t)^T \Theta(x,t)$ defines infinitesimal distances and stability. Contraction analysis in this setting considers SDEs or difference equations: $da = f(a,t) dt + \sigma(a,t) dW(t)$ where contraction in $M$ is governed by the largest eigenvalue of a generalized Jacobian, and stochastic noise is bounded in trace against the metric (Pham et al., 2013).

Explicit bounds and performance guarantees are derived: $$\mathbb{E}[d_M(T)(a(T), b(T))^2] \leq \frac{C}{\lambda} + e^{-2\lambda T} \mathbb{E}[d_M(0)(a_0, b_0)^2],$$ where $\lambda$ is the contraction rate. These analyses yield sharp mean-square bounds, inform observer design, and regulate convergence in nonlinear stochastic systems.

In the context of flows developed via SDEs on manifolds, a stochastic Riemannian metric is constructed from energetics (e.g., Hessian of an energy function), encoding constraints directly in the induced geometry, and governing dynamics via stochastic development on the orthonormal frame bundle (Mamajiwala et al., 2020).

4. Random Metrics in Infinite-Dimensional Geometries

Stochastic perturbations of Riemannian metrics themselves (rather than processes on a fixed metric) arise in geometric analysis and quantum gravity. Consider the Fréchet manifold $\mathcal{M}$ of all smooth, positive-definite metric fields on a compact manifold $M$ (Cruzeiro et al., 22 Jul 2025). The stochastic evolution equation on $\mathcal{M}$ for a kinetic energy functional yields: $$dg = \sqrt{\nu} \sum_{i=1}^N H_i(x) dW_i(t) + K(t,x) dt,$$ where $\{H_i\}$ are fixed tensor fields and $K$ is a drift. The induced stochastic geodesic equation incorporates both deterministic geodesic flow (for $\nu = 0$) and random perturbations (for $\nu > 0$), featuring a noise-induced correction operator $\mathcal{L}(K)$ analogous to a Laplacian. Solutions exist locally under regularity and boundedness, with invariant properties depending on divergence-free conditions on $H_i$.

This framework characterizes random evolution of metric fields, facilitating analysis in random shape models, stochastic quantization of gravitational geometries, and Bayesian inference over Riemannian structures.

5. Stochastic Metrics and Sampling Algorithms

In stochastic-gradient sampling, the underlying metric $G(\theta)$ dictates both drift and covariance adaptation in Langevin dynamics: $$d\theta = -G(\theta)^{-1} \nabla U(\theta) dt + \sqrt{2} G(\theta)^{-1/2} dW,$$ where $U(\theta)$ is the negative log-posterior. Non-diagonal stochastic Riemannian metrics (Monge, Shampoo) improve posterior exploration by encoding local curvature and correlations — Monge as a rank-one (gradient-outer-product) update, and Shampoo via Kronecker products of per-mode second-moment matrices (Yu et al., 2023).

Monge admits $O(D)$ cost via Sherman–Morrison inversion; Shampoo requires $O(\sum n_{l,i}^3)$ time per update, but both outperform diagonal approximations in high-curvature regimes, challenging priors, and correlated weights. Empirical analyses confirm improved mixing, log-likelihood, and accuracy, often for modest computational overhead (Yu et al., 2023).

6. Stochastic Metrics in Probability Polytopes and Information Geometry

Stochastic matrices (conditional probability polytopes) admit several natural product-type Riemannian metrics, all of Fisher–type. Three canonical constructions (scaled, unscaled, marginal-weighted) emerge from invariance principles under stochastic maps, exponential-family embeddings, or joint-distribution embeddings (Montufar et al., 2014):

• Scaled-product: $$g_{(i,j),(i',j')}(K) = \frac{1}{k} \delta_{ii'}\delta_{jj'}/K_{ij}$$
• Unscaled-product: $$g_{(i,j),(i',j')}(K) = \delta_{ii'}\delta_{jj'}/K_{ij}$$
• Marginal-weighted: $$g_{(i,j),(i',j')}(K) = \delta_{ii'}\rho(i)/K_{ij}\delta_{jj'}$$

These metrics are flat, diagonal in product coordinates, and underlie closed-form natural-gradient updates in statistical learning. The scaling and choice of invariance depend on application and transformation group.

7. SPDEs and Analysis on Manifolds with Stochastic Metrics

Linear SPDEs on manifolds equipped with stochastic (random, time-dependent) metrics require fine regularity and uniform ellipticity hypotheses (Elliott et al., 2012). For metrics $g_{ij}(\omega,x,t)$, all function spaces (Lebesgue, Sobolev) are constructed relative to the evolving metric and volume form. The standard theory delivers existence, uniqueness, and a priori bounds for variational solutions under monotonicity, boundedness, and measurability conditions.

Concrete instances include randomly scaled metrics via geometric Brownian motion, and smooth Gaussian field perturbations of a baseline metric. These frameworks enable study of random evolution of geometry and its impact on analytical and numerical properties of SPDEs.

---

The stochastic Riemannian metric paradigm encompasses tools for modeling manifold-valued uncertainty, latent-space geometry of probabilistic models, random evolution of geometric structures, and curvature-adaptive stochastic optimization. Its operationalization ranges from concrete Jacobian-based metrics in generative models to infinite-dimensional diffusions in metric space, and underpins convergence guarantees and geometric inference in high-dimensional stochastic systems. The field continues to expand at the interface of geometry, probability, and statistical modeling.