Riemannian Gibbs Process

• Riemannian Gibbs process is a stochastic dynamics on manifolds where the invariant law follows a Gibbs measure defined with a potential and the Riemannian volume.
• It employs overdamped Langevin diffusions with explicit Itô-Stratonovich corrections that integrate manifold curvature and metric effects into the drift.
• Applications include Riemannian Langevin and Hamiltonian Monte Carlo methods, fixed-rank PSD samplers, and mixing analyses via logarithmic Sobolev inequalities.

A Riemannian Gibbs process is a stochastic dynamics on a Riemannian manifold whose invariant law is a Gibbs measure of the form

$$\pi(dx)=Z^{-1}e^{-\beta V(x)}\,d\mathrm{vol}_g(x), \qquad Z=\int_M e^{-\beta V(x)}\,d\mathrm{vol}_g(x),$$

with inverse temperature $\beta>0$, potential $V$, and Riemannian volume $d\mathrm{vol}_g$ (Capel et al., 11 Jun 2026). In the papers considered here, this phrase is used most directly for overdamped Langevin diffusions driven by manifold Brownian motion, but it also appears in closely related forms: explicit Riemannian Langevin samplers on the manifold of fixed-rank positive semidefinite matrices (Yu et al., 2023), Riemannian Hamiltonian Monte Carlo (Lee et al., 2017), Gibbs-invariant Hamiltonian PDE dynamics on function spaces (Oh et al., 2015), and graph-based interacting particle systems whose hydrodynamic limits are weighted diffusions on manifolds and principal bundles (Junné et al., 2024).

1. Gibbs structure on a Riemannian manifold

The geometric input is a smooth Riemannian manifold $(M,g)$ together with a smooth potential. The Gibbs measure is defined relative to the Riemannian volume rather than Euclidean Lebesgue measure. In a local chart with metric tensor $G(q)=[g_{ij}(q)]$, the volume form is

$d\mathrm{Vol}_g(q)=\sqrt{\det G(q)}\,dq,$

so the density relative to coordinates acquires the Jacobian factor $\sqrt{\det G(q)}$ (Lee et al., 2017).

For overdamped dynamics, the canonical process targeting $\pi$ is

$$dX_t=-\nabla_g V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dB_t^g,$$

where $\beta>0$0 is Brownian motion on $\beta>0$1, $\beta>0$2 is the Riemannian gradient, and the infinitesimal generator is

$\beta>0$3

on $\beta>0$4 (Capel et al., 11 Jun 2026). The measure $\beta>0$5 is invariant and the process is reversible in $\beta>0$6 (Capel et al., 11 Jun 2026).

A persistent point across the literature is that the geometry enters twice: through the gradient and through the reference volume. In the fixed-rank PSD setting, this dependence is fully explicit because the density factor $\beta>0$7 can be written in eigen-coordinate charts for different metrics on the same manifold (Yu et al., 2023). A common misconception is therefore that a Riemannian Gibbs process is obtained by replacing $\beta>0$8 with $\beta>0$9 while leaving the target density unchanged. The cited constructions show that the metric changes both the drift and the base measure.

2. Langevin dynamics, Itô corrections, and invariant law

In the continuous-time sampling literature, the coordinate-free formulation is the Stratonovich Riemannian Langevin equation

$V$0

with invariant law

$V$1

for a smooth energy $V$2 (Yu et al., 2023). Passing to Itô form introduces an additional drift: $V$3 This Itô–Stratonovich correction encodes curvature and connection terms (Yu et al., 2023).

The fixed-rank PSD case makes this correction especially transparent. Under the embedded metric on

$V$4

the Itô SDE takes the form

$V$5

where the noise is the orthogonal projection of ambient white noise onto the tangent space and $V$6 is the mean curvature vector of the embedding $V$7 (Yu et al., 2023). The paper states that this term pushes the process away from the boundary, where the smallest nonzero eigenvalue tends to zero, thereby stabilizing rank $V$8 (Yu et al., 2023).

This geometry-dependent correction is not special to embedded schemes. In the Bures–Wasserstein formulation, the process is lifted to full-rank factors $V$9 with $d\mathrm{vol}_g$0, and the correction becomes the gradient of a logarithmic orbit-volume, written as a Boltzmann entropy term $d\mathrm{vol}_g$1 (Yu et al., 2023). The general lesson is that the invariant Gibbs law on a manifold is enforced by manifold Brownian noise together with an explicit curvature correction in Itô form.

3. Fixed-rank PSD matrices and explicit samplers

A particularly detailed realization of the Riemannian Gibbs process is given for fixed-rank PSD matrices under two metrics: the embedded metric $d\mathrm{vol}_g$2 and the Bures–Wasserstein metric $d\mathrm{vol}_g$3 (Yu et al., 2023). In eigen-coordinate charts $d\mathrm{vol}_g$4, with $d\mathrm{vol}_g$5, the density factor of the Riemannian volume depends strongly on the metric. For the embedded metric,

$d\mathrm{vol}_g$6

whereas for the Bures–Wasserstein metric,

$d\mathrm{vol}_g$7

(Yu et al., 2023). The two Gibbs laws therefore differ even for the same energy $d\mathrm{vol}_g$8.

Under $d\mathrm{vol}_g$9, the tangent space at $(M,g)$0 is represented by block matrices

$(M,g)$1

with $(M,g)$2 and $(M,g)$3 (Yu et al., 2023). Given a symmetric Euclidean gradient $(M,g)$4, the Riemannian gradient is

$(M,g)$5

with the efficient implementation

$(M,g)$6

$(M,g)$7

(Yu et al., 2023).

The corresponding explicit Euler–Maruyama discretization, Scheme E, updates $(M,g)$8 by adding the Riemannian gradient drift, projected Gaussian noise in the tangent space, and a mean-curvature block term proportional to $(M,g)$9, followed by Euclidean projection onto the rank-$G(q)=[g_{ij}(q)]$0 PSD cone via top-$G(q)=[g_{ij}(q)]$1 eigendecomposition (Yu et al., 2023). The paper states that the mean curvature correction is essential under $G(q)=[g_{ij}(q)]$2 to avoid rank-deficiency (Yu et al., 2023).

For the Bures–Wasserstein metric, the quotient representation

$G(q)=[g_{ij}(q)]$3

is used (Yu et al., 2023). If $G(q)=[g_{ij}(q)]$4, then

$G(q)=[g_{ij}(q)]$5

and Scheme BW updates $G(q)=[g_{ij}(q)]$6 by

$G(q)=[g_{ij}(q)]$7

where $G(q)=[g_{ij}(q)]$8 is computed from the compact SVD of $G(q)=[g_{ij}(q)]$9 through the curvature correction

$d\mathrm{Vol}_g(q)=\sqrt{\det G(q)}\,dq,$0

(Yu et al., 2023). No projection is needed if $d\mathrm{Vol}_g(q)=\sqrt{\det G(q)}\,dq,$1 remains full rank, and the paper gives per-iteration complexity $d\mathrm{Vol}_g(q)=\sqrt{\det G(q)}\,dq,$2, compared with a typically $d\mathrm{Vol}_g(q)=\sqrt{\det G(q)}\,dq,$3 cost for Scheme E (Yu et al., 2023).

Validation is carried out with three energies: $d\mathrm{Vol}_g(q)=\sqrt{\det G(q)}\,dq,$4, the von Neumann entropy $d\mathrm{Vol}_g(q)=\sqrt{\det G(q)}\,dq,$5, and the shifted quadratic $d\mathrm{Vol}_g(q)=\sqrt{\det G(q)}\,dq,$6 for a large center $d\mathrm{Vol}_g(q)=\sqrt{\det G(q)}\,dq,$7 (Yu et al., 2023). For the quadratic case, the scalar statistic $d\mathrm{Vol}_g(q)=\sqrt{\det G(q)}\,dq,$8 has closed-form $d\mathrm{Vol}_g(q)=\sqrt{\det G(q)}\,dq,$9 laws whose empirical CDFs match the predictions under both metrics, and the paper also reports MCMC integration identities with empirical estimates converging at $\sqrt{\det G(q)}$0 (Yu et al., 2023).

4. Mixing theory, entropy decay, and quotient geometry

The analytic theory of rapid mixing for Gibbs measures on Riemannian manifolds is formulated in terms of entropy, Fisher information, logarithmic Sobolev inequalities, and Poincaré inequalities (Capel et al., 11 Jun 2026). For a probability measure $\sqrt{\det G(q)}$1 with density $\sqrt{\det G(q)}$2,

$\sqrt{\det G(q)}$3

Along the Langevin flow, the entropy production identity is

$\sqrt{\det G(q)}$4

and an LSI with constant $\sqrt{\det G(q)}$5 yields

$\sqrt{\det G(q)}$6

hence exponential decay in total variation through the Pinsker–Csiszár–Kullback inequality (Capel et al., 11 Jun 2026).

The paper identifies two regimes. In the high-temperature or strongly convex case, a Bakry–Émery curvature bound

$\sqrt{\det G(q)}$7

implies an LSI with $\sqrt{\det G(q)}$8 (Capel et al., 11 Jun 2026). In the low-temperature regime with symmetry and multiple equivalent minima, the strategy is to quotient by a compact connected Lie group $\sqrt{\det G(q)}$9 acting freely and isometrically, analyze the base manifold $\pi$0, prove a Poincaré inequality there, lift it back to $\pi$1, and then tighten PI to LSI (Capel et al., 11 Jun 2026).

The landscape assumptions on the quotient potential $\pi$2 are strong and explicit: unique global minimum, all other critical points are saddle points, escaping directions from saddles with

$\pi$3

isolated and separated critical points, and a “no barren plateaus” condition

$\pi$4

(Capel et al., 11 Jun 2026). Under these assumptions and for sufficiently large $\pi$5, the paper gives explicit LSI constants: $\pi$6

$\pi$7

(Capel et al., 11 Jun 2026). The resulting mixing time scales polynomially in $\pi$8 when the relevant geometric quantities and $\pi$9 do so (Capel et al., 11 Jun 2026).

A plausible implication is that the geometry of the manifold and the symmetry structure of the energy landscape are not auxiliary technicalities but part of the mixing mechanism itself. In this formulation, rapid mixing depends simultaneously on Ricci curvature, Hessian structure, saddle escape, fiber geometry, and quotient diameter.

5. Hamiltonian variants and geometry-adapted Monte Carlo

Riemannian Hamiltonian Monte Carlo is a discrete-time, Metropolized Riemannian Gibbs process on the cotangent bundle $$dX_t=-\nabla_g V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dB_t^g,$$0 (Lee et al., 2017). For local coordinates $$dX_t=-\nabla_g V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dB_t^g,$$1 and metric tensor $$dX_t=-\nabla_g V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dB_t^g,$$2, the Hamiltonian is

$$dX_t=-\nabla_g V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dB_t^g,$$3

and the momentum is refreshed from $$dX_t=-\nabla_g V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dB_t^g,$$4 (Lee et al., 2017). The $$dX_t=-\nabla_g V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dB_t^g,$$5 term ensures that the marginal in $$dX_t=-\nabla_g V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dB_t^g,$$6 is exactly the Gibbs law $$dX_t=-\nabla_g V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dB_t^g,$$7 (Lee et al., 2017).

The Hamiltonian equations are

$$dX_t=-\nabla_g V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dB_t^g,$$8

and the second-order form is

$$dX_t=-\nabla_g V(X_t)\,dt+\sqrt{\frac{2}{\beta}}\,dB_t^g,$$9

which exhibits geodesic motion modulated by the Riemannian gradient of the potential (Lee et al., 2017).

One RHMC transition consists of momentum refresh, numerical integration of the Hamiltonian ODE for time $\beta>0$00 with a reversible symplectic manifold integrator, and Metropolis acceptance with probability $\beta>0$01 (Lee et al., 2017). The convergence analysis is expressed through conductance and a manifold analog of the KLS conjecture for Gibbs distributions. If $\beta>0$02 is conductance and $\beta>0$03 the $\beta>0$04 spectral gap, then

$\beta>0$05

(Lee et al., 2017).

The main algorithmic application is to polytopes equipped with the log-barrier metric

$\beta>0$06

which inflates distances near the boundary (Lee et al., 2017). For uniform sampling from a polytope, the density relative to $\beta>0$07 becomes

$\beta>0$08

so the potential term in the Hamiltonian cancels and the dynamics reduce to pure geodesic motion with Gaussian momentum up to numerical integration error (Lee et al., 2017). Under the paper’s regularity assumptions, both uniform sampling and polytope volume computation require a total number of RHMC steps

$\beta>0$09

(Lee et al., 2017).

6. Broader usages, neighboring constructions, and terminological scope

The phrase also appears in mathematical physics and stochastic geometry in ways that are related by invariant Gibbs measures but not identical in mechanism.

For the two-dimensional defocusing nonlinear Schrödinger equation on a compact Riemannian manifold without boundary or on a bounded domain in $\beta>0$10, the process is the stationary random evolution generated by the Wick-ordered Hamiltonian

$\beta>0$11

with Gibbs measure

$\beta>0$12

(Oh et al., 2015). The main invariance statement is that if $\beta>0$13, then for all $\beta>0$14, $\beta>0$15 (Oh et al., 2015). Here the “process” is a Gibbs-invariant Hamiltonian PDE flow on an infinite-dimensional function space rather than a sampling diffusion.

In quantum information geometry, the relevant object is the manifold of Gibbs thermal states

$\beta>0$16

equipped with monotone Riemannian metrics classified by operator monotone functions (Tonchev, 2021). The paper gives a spectral representation of the full class of monotone metrics in terms of the dynamical structure factor $\beta>0$17 and the dissipative susceptibility: $\beta>0$18 (Tonchev, 2021). In that presentation, “Riemannian Gibbs process” denotes an end-to-end procedure for obtaining the metric from measurable response functions rather than a stochastic process in the Markov sense (Tonchev, 2021).

A different large-scale construction arises from the symmetric exclusion process on random neighborhood graphs drawn from Poisson point processes with Gibbs intensity on a complete connected Riemannian manifold (Junné et al., 2024). With

$\beta>0$19

and suitably normalized symmetric weights, the hydrodynamic limit is the diffusion with drift induced by the Gibbs potential: $\beta>0$20 and on principal bundles the lifted limit is

$\beta>0$21

(Junné et al., 2024). The operator is reversible with respect to the corresponding Gibbs-weighted volume measure (Junné et al., 2024).

This suggests that the term is not fully standardized. Across these works, the unifying feature is not a single algorithmic template but the pairing of Riemannian geometry with a Gibbs law, together with a dynamics, flow, or geometric construction whose invariant, reversible, or hydrodynamic object is determined by that law.