Riemannian Brownian Motion Prior

• Riemannian Brownian Motion Prior is a probabilistic model defined on paths over Riemannian manifolds, generalizing Euclidean Brownian motion by incorporating manifold geometry via the heat kernel.
• It underpins nonparametric Bayesian regression and generative modeling, ensuring that inference on manifold-valued data respects intrinsic curvature and topology.
• Discretized implementations use piecewise-geodesic paths and retraction-based approximations, guaranteeing convergence to the continuous Brownian motion as the mesh size decreases.

A Riemannian Brownian motion prior is a probability measure on paths or random variables valued in a Riemannian manifold, canonically induced by the Brownian motion associated with the Laplace–Beltrami operator. Such priors generalize the classical Gaussian prior (Euclidean Brownian motion) to nonlinear spaces and are governed by the manifold’s geometry. They underlie inference for manifold-valued data, nonparametric Bayesian regression with geometric targets, and generative modeling in machine learning, such as manifold-structured variational autoencoders. The prior is fundamentally determined by the heat kernel on the manifold, encoding curvature, topology, and local volume form.

1. Mathematical Definition and Construction

Let $(M, g)$ be a compact $D$-dimensional Riemannian manifold with Riemannian volume $\mu$ and Laplace–Beltrami operator $\Delta$. The Riemannian Brownian motion $X_t$ on $M$ is the diffusion whose generator is $\tfrac12 \Delta$ and which solves the Stratonovich stochastic differential equation (SDE)

$$dX_t = \sum_{i=1}^D V_i(X_t) \circ dB^i_t, \quad X_0 = x,$$

where $\{V_1, \ldots, V_D\}$ is any local orthonormal frame and $B^i_t$ are independent standard Brownian motions (Wang et al., 2015, Lee et al., 22 Oct 2025).

The fundamental solution to the heat equation,

$$\partial_t u(t,x) = \tfrac12\,\Delta_x u(t,x), \quad u(0,x) = \delta_y(x),$$

is the heat kernel $p_t(x,y)$, which is the transition density at time $t$ of Brownian motion from $x$ to $y$ (Wang et al., 2015). Spectrally,

$$p_t(x, y) = \sum_{k=0}^\infty e^{-\lambda_k t} \phi_k(x)\phi_k(y),$$

where $\{\phi_k\}$ are the eigenfunctions and $\{\lambda_k\}$ the eigenvalues of $-\Delta$.

The canonical Brownian motion prior on the path-space $C([0,1],M)$ is constructed by prescribing finite-dimensional distributions using $p_t$, then extending uniquely to a measure via Kolmogorov’s extension theorem (Wang et al., 2015).

2. Discretized Priors and Practical Implementation

In inference settings, continuous Brownian motion is discretized for computation. A piecewise-geodesic path with mesh size $h>0$ is represented by evaluating $f(0), f(h), \ldots, f(1)$. The discretized Brownian prior assigns a density with respect to $\mu^{\otimes (1/h+1)}$: $$\pi_h(f) = \frac{1}{\mu(M)} \prod_{k=1}^{1/h} p_h(f((k-1)h), f(kh)).$$ This construction is equivalent to sampling a geodesic random walk with transition kernels given by the heat kernel at each step (Wang et al., 2015, Schwarz et al., 2022).

Efficient sampling can be achieved by replacing the exact exponential map with second-order retractions, providing a local approximation at a fraction of the computational cost. When the retraction is accurate to $O(\|v\|^3)$, the resulting random walk converges in law (in the Skorokhod topology) to the Brownian motion as $h \to 0$ (Schwarz et al., 2022).

For embedded or implicit manifolds, projection-based retractions allow for fast computation of proposal points, and for Lie groups, group-exponential or Cayley-type retractions offer efficient alternatives.

Theoretical results ensure that as the mesh size $h \to 0$, the law of the piecewise-geodesic path converges in distribution to the law of the continuous Brownian path on $M$ (Wang et al., 2015, Schwarz et al., 2022).

3. Posterior Consistency and Contraction for Manifold-Valued Regression

In nonparametric Bayesian regression on manifolds, consider data $(t_i, x_i)$, where predictors $t_i \in [0,1]$ and responses $x_i$ in a compact Riemannian manifold $M$, modeled as

$x_i \mid t_i \sim p_{\sigma^2}(f_0(t_i), x_i),$

for some (unknown) Lipschitz function $f_0 : [0,1] \to M$ and fixed $\sigma^2$ (Wang et al., 2015).

The Riemannian Brownian motion prior is used on the space of regression functions: the prior on $C([0,1], M)$ is the law of a Brownian path. For practical inference, one employs a discretized Brownian motion prior on piecewise-geodesic paths, as above.

The main statistical result establishes that the posterior contracts around the true regression function $f_0$ at rate $O(n^{-1/4+\epsilon})$ (for any small $\epsilon>0$) in $L_q$ metrics: $d_q(f, g) = \left( \int_0^1 \dist_M(f(t), g(t))^q p(t) dt \right)^{1/q}.$ Consistency is verified for all $1 \le q < \infty$ for discretized priors, and weak consistency for the continuous path-space prior (Wang et al., 2015).

Proof techniques rely on metric-entropy for Hölder sieves, concentration of prior mass (via small-heat-kernel balls), and Kullback–Leibler neighborhood arguments, leveraging the equivalence between Hellinger/KL distances for induced data densities and sup-distances in path space.

4. Geometric and Algorithmic Aspects

The stochastic differential equation on $M$ can be interpreted in both Stratonovich and Itô forms. On an intrinsic Riemannian manifold, the SDE

$$d x_t = -\tfrac12 \sum_{i=1}^n (\nabla_{e_i} e_i)(x_t) dt + \sum_{i=1}^n e_i(x_t) \circ dW^i_t$$

generates Brownian motion with generator $\tfrac12\Delta_{\rm LB}$, where $\nabla$ is the Levi–Civita connection (Lee et al., 22 Oct 2025). The Stratonovich drift is geometrically tied to the divergence of the vector fields spanning the tangent space. Conversion to the Itô form removes the explicit drift; Itô Brownian motion lacks a drift but accumulates "curvature-induced" drift through the Stratonovich-to-Itô correction.

In the context of embedded submanifolds or specific geometric structures (Lie groups, submanifolds) the drift and frame-fields encode the response of Brownian motion to curvature, mean-curvature, or group coadjoint structure (Lee et al., 22 Oct 2025).

5. Extensions to Manifold-Learned and Data-Driven Settings

When the ambient prior is not Euclidean but manifold-valued (e.g., in latent variable models), one assigns a Brownian motion prior based on the induced metric $g(z)$ from a decoder mapping $f(z)$ in a VAE architecture. The Brownian motion prior generalizes the classic Gaussian prior via the Riemannian heat kernel: $$p(z) \propto (2 \pi t_p)^{-N/2} \exp\left(-\frac{d_g(\mu_p,z)^2}{2t_p}\right) \sqrt{\det g(z)},$$ where $d_g(\cdot,\cdot)$ is the geodesic distance for the pull-back metric (Kalatzis et al., 2020).

In variational inference, differences in log-densities render normalization constants unnecessary, and Riemannian reparameterization enables gradient-based optimization. Computational cost arises mainly from metric inversion and geodesic computations per sample, but the approach is practical for moderate dimensions. Empirical results on image datasets show improved likelihood estimates and latent space regularization compared to Euclidean Gaussian or learned mixture priors, notably at low latent dimensions (Kalatzis et al., 2020).

6. Generalizations and Convergence of Random Walks

Beyond Riemannian cases, geodesic random walks in Finsler or more general geometric settings converge to limiting diffusions governed by a Riemannian-type Laplace operator. For a Finsler manifold $(M, F)$ of bounded geometry and a smooth family of tangent-space distributions $\{\nu_p\}$, geodesic random walks (scaled and centered) converge to a limiting diffusion with generator

$A = \frac12 \Delta^{\mathbf{g}_A} + B,$

where $\mathbf{g}_A$ is the symmetric part of the covariance induced by $\{\nu_p\}$, and $B$ is a drift term. When $\nu_p$ are centered and isotropic, the limit is classical Riemannian Brownian motion. This establishes the robustness of the Riemannian Brownian prior as the canonical limit in a broad class of geometric random walks (Ma et al., 2021).

7. Applications and Empirical Properties

Riemannian Brownian motion priors play a key role in:

• Nonparametric regression on manifolds, yielding posterior contraction rates and weak/strong consistency results in metric function spaces (Wang et al., 2015).
• Variational autoencoders with manifold-valued latent spaces, resulting in improved model capacity, better fit to latent data manifolds, and sharper, geometrically valid generative samples (Kalatzis et al., 2020).
• Latent-trajectory and time-series modeling, where the prior ensures sample paths stay on the manifold, and the law is efficiently sampled via discrete or retraction-based geodesic walks (Schwarz et al., 2022).
• Bayesian smoothing and filtering in systems evolving on $M$, where the prior is combined with data likelihood via state-space models, and inference leverages the Brownian heat kernel or its discretization (Lee et al., 22 Oct 2025).

A plausible implication is that the ubiquity of the Riemannian Brownian motion prior in modeling arises from its invariance property, geometric consistency, and the limiting behavior of a wide range of discrete random walks even in non-Riemannian small-time regimes (Ma et al., 2021). The prior’s contraction properties scale with the mesh size and ambient geometry, with computational bottlenecks primarily in geodesic or metric inversion, but feasible via well-adapted numerical schemes.