Score-Based Riemannian Metrics

Updated 10 March 2026

Score-Based Riemannian Metrics are data-adaptive geometric constructs that use the gradient of the log density to unveil intrinsic structure in probability distributions and manifold learning.
They guide interpolation, sampling, and optimization by defining spatially varying inner products through methods like rank-one, score-Hessian, and Fisher metrics across Euclidean and non-Euclidean spaces.
These metrics provide practical insights into high-dimensional generative tasks by emphasizing local curvature and normal directions, thereby aligning sampling and optimization with underlying data geometry.

Score-based Riemannian metrics provide a principled, data-adaptive means to capture the intrinsic geometry of probability distributions, data manifolds, and parameter landscapes in machine learning. They leverage the “score”---the gradient of the log density---to define spatially varying inner products, enabling geometric reasoning for generative modeling, manifold exploration, optimization, and representation learning. Recent advances have extended these constructions from classical natural gradient methods and energy-based models to high-dimensional generative models such as diffusion models, both in Euclidean settings and on general Riemannian manifolds.

1. Fundamental Concepts and Metric Definitions

The score function $s(x) = \nabla_x \log p(x)$ is central in defining a data-driven Riemannian metric. Several canonical constructions arise:

Rank-One Score Metric: $g(x) = I + \lambda\, s(x) s(x)^\top$ , with $\lambda>0$ , stretches ambient distances in the normal direction to the data manifold, strongly penalizing off-manifold motion while preserving on-manifold geometry. This metric is positive definite and directly encodes local curvature and normal directions of $p(x)$ , as the score is nearly normal to the data manifold at typical points (Azeglio et al., 16 May 2025).
Score Hessian/Gramian Metric: For a diffusion model with marginal $p_t(x)$ and score $s_t(x)$ , define $J_t(x) = \nabla_x s_t(x)$ (the Jacobian matrix of the score). The corresponding metric at fixed $t$ is

$g^{(t)}_x(v,w) = v^\top [J_t(x)^\top J_t(x)] w$

which informs the local geometry by the sensitivity of the score field to infinitesimal displacements (Saito et al., 28 Apr 2025).

Fisher Information Metric: For a parametric model $p(x|\theta)$ , the Fisher information

$G(\theta) = \mathbb{E}_{x\sim p}\big[s(x;\theta)s(x;\theta)^\top\big]$

furnishes a natural coordinate-invariant Riemannian metric over parameter space, widely used in optimization as the foundation of natural-gradient and information-geometric updates (Ollivier, 2013).

Energy-based Conformal Metrics: For energy functions $E_\theta(x)$ , derive conformal metrics of the form $G_1(x) = [\alpha E_\theta(x) + \beta]I$ (log-energy metric) or $G_2(x) = [\alpha e^{-E_\theta(x)} + \beta]^{-1}I$ (inverse-density metric), scaling spatial distances according to energy and thus probability (Béthune et al., 23 May 2025).

All of these approaches impose a data- or model-dependent geometry, which guides geodesics, optimization, and sampling to adhere to underlying statistical structure.

2. Score-based Metrics on Riemannian Manifolds

When data reside on non-Euclidean spaces $(\mathcal{M},g)$ ---compact spheres, Lie groups, or symmetric spaces---the geometry is intrinsically Riemannian. Generative models must respect the ambient metric structure:

Score-based SDEs on Manifolds: The forward noising SDE on $(\mathcal{M},g)$ is

$dX_t = b(t, X_t) dt + \sigma(t) dW^g_t, \qquad X_0\sim p_0$

where $W^g_t$ is Brownian motion intrinsic to $(\mathcal{M},g)$ (generator $\frac12\Delta_g$ ) and $\Delta_g$ is the Laplace--Beltrami operator. The time-reversal "denoising" SDE utilizes the Riemannian score field $s^* = \nabla^g \log p_t(x)$ (Bortoli et al., 2022, Lou et al., 2023).

Heat Kernel and Score Matching: The transition kernel $K_{\mathcal{M}}(x|x_0,t)$ solves the heat equation. The Riemannian score-matching loss generalizes Euclidean score matching, using geometric norms and the divergence induced by the metric (Lou et al., 2023).
Symmetric Space Computations: For symmetric spaces $M=G/K$ , efficient expressions for heat kernels, geodesic distances, and score fields are available via eigenfunction expansions, radial reductions, and sum-over-paths representations. This permits closed-form or highly accurate approximations even in high-dimensional contexts such as hyperspheres or compact Lie groups (Lou et al., 2023).
Algorithmic Implementations: Discretization of SDEs uses geodesic random walks (exponential map updates), and parameterization of score fields leverages global or local coordinate frames. These approaches enable score-based generative modeling ("Riemannian Score-based Generative Models"/RSGMs) on nontrivial manifolds, with empirical performance superior to extrinsic or flow-based methods in various scientific and synthetic datasets (Bortoli et al., 2022, Lou et al., 2023).

3. Score-based Riemannian Geometry of Data Manifolds

Diffusion models and energy-based models implicitly learn a manifold structure in the high-dimensional ambient space. Riemannian metrics derived from the score reveal and exploit this learned geometry:

Score-induced Normal Emphasis: $g(x) = I + \lambda\, s(x)s(x)^\top$ strongly stretches the metric along directions orthogonal to the manifold, thus making geodesics "hug" the probability mass. This leads to more realistic and data-conforming paths for both interpolation and extrapolation tasks (Azeglio et al., 16 May 2025).
Geodesic Computation: Due to the complexity of high-dimensional learned manifolds, closed-form geodesics are generally unavailable, so discrete variational optimization or neural interpolant networks are used for practical computation. The path energy or kinetic loss incorporates the score-based metric, and optimization is performed under manifold-aware gradients (Saito et al., 28 Apr 2025, Béthune et al., 23 May 2025, Azeglio et al., 16 May 2025).
Pullback Metrics in Latent Spaces: For a diffeomorphic generative mapping $f\colon \mathcal{Z}\to\mathcal{X}$ (e.g., normalizing flows), the metric is pulled back from the data space, naturally incorporating the local Jacobian and the score or Hessian of the data density to yield a latent metric:

$g_z = (Jf(z))^\top A(f(z)) Jf(z)$

with $A(x)$ derived from the score structure or the Hessian of $\log p(x)$ . This enables principled dimension estimation and interpretable representation learning (Diepeveen et al., 2024).

4. Applications: Interpolation, Sampling, and Optimization

Score-based Riemannian metrics enable multiple geometric and practical advances:

Manifold-constrained Interpolation: Image and representation interpolation using metric geodesics yields smooth, data-respecting transitions, outperforming traditional linear or slerp paths on perceptual quality (LPIPS, FID, KID) and semantic faithfulness. Diffusion and EBM-induced metrics both achieve geodesics that closely remain on the data manifold (Azeglio et al., 16 May 2025, Saito et al., 28 Apr 2025, Béthune et al., 23 May 2025).
Extrapolation Along the Manifold: The same geometric principles extend naturally to plausible extrapolation, offering meaningful transformations beyond the support of observed data (Azeglio et al., 16 May 2025).
Sampler Guidance: In generative models, Riemannian geometry induced by the score guides denoising processes and SDE sampling, especially under manifold constraints, improving sample quality and alignment to data distributions (Bortoli et al., 2022, Lou et al., 2023).
Optimization in Parameter Space: The Fisher information metric enables natural-gradient optimization, being intrinsic to the statistical structure and invariant to parameterization. Quasi-diagonal and backpropagated metrics provide scalable approximations with block-wise invariance, undergirding effective training of neural networks and probabilistic models (Ollivier, 2013).
Intrinsic Dimension Estimation and Autoencoders: Riemannian pullback metrics furnish tools for detecting the intrinsic dimensionality of data and constructing Riemannian autoencoders with dimension guarantees and closed-form geodesics along the learned manifold (Diepeveen et al., 2024).

5. Empirical Results and Comparative Analyses

Empirical studies on synthetic, Earth science, and image datasets highlight the practical impact of score-based Riemannian metrics:

Riemannian Diffusion Models: On manifolds such as $\mathbb{S}^2$ , $SO(3)$ , and tori, RSGMs achieve higher likelihoods and more efficient sampling compared to wrapped-Gaussian EM mixtures, Riemannian CNFs, and Moser flows. Notably, high-dimensional scalability is attained via symmetric-space reductions and precise kernel computations (Bortoli et al., 2022, Lou et al., 2023).
Score-based Interpolation: In MNIST and Stable Diffusion benchmarks, geodesics under the score-Hessian or rank-one score metrics enable continuous, semantically meaningful, and low-noise interpolations between images---outperforming standard approaches such as LERP, SLERP, and NoiseDiffusion across quantitative and qualitative criteria (Saito et al., 28 Apr 2025, Azeglio et al., 16 May 2025).
EBM-derived Metrics: Geodesics under EBM-conformal metrics produce interpolants with higher probability-density accumulation, lower off-manifold deviation, and superior Fréchet Inception Distance, especially in high-dimensional latent spaces (Béthune et al., 23 May 2025).
Pullback Riemannian Geometry: Anisotropic flow models equipped with score-based pullback metrics recover lower geodesic and variation errors, avoid spurious detours, and provide accurate intrinsic dimension estimation via Riemannian autoencoders (Diepeveen et al., 2024).

6. Computational and Theoretical Considerations

The adoption of score-based Riemannian metrics entails algorithmic and mathematical factors:

Computing metric tensors and their inverses in high-dimensional settings is challenging; practical schemes employ efficient Jacobian-vector and vector-Jacobian products, symmetries (spheres, Lie groups), or neural surrogates for geodesics (Saito et al., 28 Apr 2025, Lou et al., 2023).
Rank-one score metrics offer fast Sherman--Morrison inversion; score-Hessian metrics circumvents explicit metric inversion by path-length optimization; conformal metrics from EBMs and pullbacks support closed-form or fast approximate geodesics (Azeglio et al., 16 May 2025, Béthune et al., 23 May 2025, Diepeveen et al., 2024).
Theoretical guarantees include guaranteed loss decrease under natural-gradient updates, invariance properties, and dimension-recovery bounds for Riemannian autoencoders (Ollivier, 2013, Diepeveen et al., 2024).
Limitations currently include computational overhead for geodesic computation compared to classic interpolants, dependence on accurate score estimation, and limited closed-form results outside specific manifold classes or synthetic densities (Azeglio et al., 16 May 2025, Lou et al., 2023).

7. Significance, Open Issues, and Outlook

Score-based Riemannian metrics have established themselves as foundational tools for translating probabilistic structure into geometric constructs across data, parameter, and latent spaces. Their capacity to encode local normal and curvature information of learned data manifolds enables improved generative modeling, interpretable interpolation, scalable sampling, and principled representation learning.

Challenges remain in further reducing computational overhead, extending closed-form geometric solutions to broader manifold and multimodal settings, and advancing theoretical analysis of curvature and local spectrum of the induced metrics. Directions for future work include accelerating geodesic computations via neural surrogates, developing semantic geodesic editing, and leveraging metric spectrum analysis for manifold diagnostics (Azeglio et al., 16 May 2025, Diepeveen et al., 2024).

Collectively, the theory and methodology of score-based Riemannian metrics unify geometric insight with modern generative modeling, opening new avenues for mathematically grounded machine learning and data analysis (Bortoli et al., 2022, Lou et al., 2023, Béthune et al., 23 May 2025, Saito et al., 28 Apr 2025, Azeglio et al., 16 May 2025, Ollivier, 2013, Diepeveen et al., 2024).