Riemannian Autoencoders Overview

• Riemannian autoencoders are generative frameworks that model the latent space as a differentiable manifold using decoder-induced pull-back metrics to capture intrinsic geometry.
• They integrate geometry-aware priors, reparameterization via exponential maps, and tailored regularization to improve ELBO and produce higher-quality interpolations and samples.
• These models extend to complex, mixed-curvature, and Lie group latent spaces, offering robust representation learning and practical benefits for high-dimensional data.

Riemannian autoencoders are generative learning frameworks in which the latent space is endowed with a Riemannian manifold structure, rather than the standard Euclidean geometry. This geometric formalism leverages decoder-induced or data-driven metrics to respect intrinsic distances, curvature, and topology of the data distribution, yielding enhanced sample quality, more plausible interpolations, improved representation learning, and the ability to model nontrivial latent geometries, including constant or mixed curvature, complex (Kähler), and data-induced metrics. Riemannian autoencoders comprise both variational and non-variational approaches, with core methodologies built upon pull-back metrics (often from the decoder), intrinsic manifold measures, manifold-informed priors and posteriors, and geometry-aware regularization and losses.

1. Mathematical Foundations: Pull-Back Metrics and Manifold Structure

The essential principle of Riemannian autoencoders is to reinterpret the latent space Z of an autoencoder as a differentiable manifold (𝓜, g), where g is a Riemannian metric tensor. This is most classically constructed via the pull-back mechanism: for a differentiable decoder map μₜₕₑₜₐ : 𝓜→ℝᵐ (as in a VAE), the metric at z∈𝓜 is

$$G(z) = J_{μ}(z)^\top J_{μ}(z) + J_{σ}(z)^\top J_{σ}(z)$$

where J_{μ} and J_{σ} are Jacobians of the decoder mean and variance, respectively (Kalatzis et al., 2020). This metric equips the tangent space T_z𝓜 with inner product ⟨v, w⟩_z = v^\top G(z) w. For some models, only the mean map is used; for others, decoder uncertainty or learned metric networks supplement or replace the explicit Jacobian terms (Chadebec et al., 2020).

The Riemannian metric enables definition of geodesic curves γ : [0,1]→𝓜 by minimizing the energy functional:

$$E(\gamma) = \frac{1}{2} \int_0^1 \dot{\gamma}(t)^\top G(\gamma(t)) \dot{\gamma}(t) \, dt,$$

resulting in the Riemannian distance d(z, z′) as the minimal energy path length.

These constructions extend to specialized settings:

• Complex-valued latent spaces endowed with Kähler structure (Fisher information Hermitian metric derived from decoder geometry) (Gracyk, 19 Nov 2025)
• Product or mixed-curvature manifolds (e.g., Euclidean × spherical × hyperbolic) with closed-form Riemannian operators (Skopek et al., 2019)
• Pullback metrics via normalizing flows or score-models, enabling global charts of data-driven geometry (Diepeveen et al., 2024, Diepeveen et al., 26 Jan 2026)

2. Riemannian Priors, Posteriors, and Reparameterization Strategies

To align prior and posterior models with intrinsic geometry, Riemannian autoencoders replace the standard flat Gaussian prior with manifold-adapted distributions:

• Brownian motion/heat kernel priors: On (𝓜, g), the prior is given by the transition density of Riemannian Brownian motion:

$$p_t(z | z_0) = Z_t^{-1} K_t(z_0, z), \qquad K_t(z_0, z) \approx (2\pi t)^{-n/2} H_0(z, z_0) \exp\left(-\frac{d(z, z_0)^2}{2t}\right)$$

where $H_0$ is the volume form ratio. This recovers standard Gaussian when 𝓜 = ℝ^n (Kalatzis et al., 2020).

• Wrapped Gaussians: For homogeneous manifolds (e.g. spheres, hyperboloids), area-preserving diffeomorphisms push Euclidean Gaussians through the exponential (or Lambert/gyro) maps, preserving normalization and simple sampling (Galaz-Garcia et al., 2022, Skopek et al., 2019).
• Product manifolds: Posteriors and priors are constructed as products of wrapped normal distributions over each factor, with full reparameterization via tangent-normal sampling, parallel transport, and exponential maps (Skopek et al., 2019).
• Hamiltonian proposals: For samplers requiring manifold volume correction, the Riemannian Hamiltonian approach defines the auxiliary momentum in T_z𝓜, evolving $(z, \rho)$ under Hamiltonian flow specified by the metric tensor (Chadebec et al., 2020).

Reparameterization on the manifold employs either (i) the exponential map applied to Gaussian tangent noise, or (ii) approximate shooting methods (projected small steps, Jacobian pseudo-inverse back-projection) (Kalatzis et al., 2020).

3. Variational Objectives and Geometry-Aware Regularization

The evidence lower bound (ELBO) for Riemannian autoencoders incorporates densities, volumes, and KL divergences under the manifold measure. When both prior and posterior are written as heat kernels or wrapped Gaussians on 𝓜, normalization factors cancel, and KL is tractably estimated by Monte Carlo (Kalatzis et al., 2020, Galaz-Garcia et al., 2022). The Riemannian structure introduces additional terms:

• Volume corrections: Sampling and loss terms include √det G(z) in density calculations, and geometry-aware sampling exploits the Riemannian measure (Gracyk, 19 Nov 2025, Chadebec et al., 2022).
• Regularization: Flatness (low-bending) and isometry penalties encourage mappings that are locally distance-preserving and extrinsically as flat as possible (Braunsmann et al., 2021, Braunsmann et al., 2022). For example, loss terms penalize deviations of latent-space distances from manifold geodesics and second-difference bending vectors.
• Decoder geometry: Additional auxiliary losses encourage the latent embedding to concentrate near the true data manifold, prevent excessive curvature, and match geometric regularity via curvature or Ricci terms (Gracyk, 19 Nov 2025, Sun et al., 2024).
• Adversarial/geometric membership: For constant-curvature spaces, adversarial objectives force the aggregated posterior to match manifold-valued priors, and explicit constraints penalize deviation from the embedding manifold (Grattarola et al., 2018).

4. Computational Schemes: Geodesic Solvers, Interpolations, and Sampling

Computation of geodesics and interpolations in Riemannian autoencoders is central to their improved generative and representation properties. Techniques span:

• Spline or energy-based geodesic search: Parameterizing γ by splines and minimizing the Riemannian energy by gradient descent, often in latent space with autodifferentiable metrics (Kalatzis et al., 2020, Hartwig et al., 10 Oct 2025, Shamsolmoali et al., 2023). Discrete variational schemes (augmented Lagrangian, shooting) are practical with implicit manifolds defined by denoising-projected latent point clouds (Hartwig et al., 10 Oct 2025).
• Closed-form solutions: Certain score-based or flow-based architectures admit explicit geodesics, exponential/log maps, and charts via invertible neural networks and their Jacobians (Diepeveen et al., 2024, Diepeveen et al., 26 Jan 2026).
• Uniform and volume-weighted sampling: Manifold-aware HMC and Langevin schemes sample from the Riemannian volume density or the aggregate posterior weighted by √det G(z), ensuring that generation and coverage align with the learned geometry (Chadebec et al., 2022, Sun et al., 2024, Gracyk, 19 Nov 2025).
• Curvature and Kähler metrics: In complex or Kähler settings, the Fisher information (as the complex Dolbeault-Hessian of the KL divergence) and its mixture-model approximations guide both geometry-regularization and sampling protocols (Gracyk, 19 Nov 2025).

5. Model Architectures, Implementation, and Empirical Findings

Architectures for Riemannian autoencoders parallel classical autoencoders but with essential geometric augmentations:

• Encoders output manifold-structured parameters, e.g., tangent-space coordinates, means/covariances for wrapped distributions, Lie algebra elements (for group-based models) (Gong et al., 2019).
• Decoders are differentiable immersions (usually MLPs or CNNs) whose Jacobian determines the pull-back metric. In explicit manifold models (sphere/hyperboloid), decoding may map from the manifold to data via little more than an affine or kernel network (Skopek et al., 2019, Galaz-Garcia et al., 2022).
• Projection/denoising heads are used to approximate the manifold by an implicit projection, enabling robust implementation of Riemannian calculus even when the true manifold is not explicitly parameterized (Hartwig et al., 10 Oct 2025).

Empirical evidence demonstrates that geometry-aware priors and interpolations:

• Increase ELBO and reduce negative log-likelihood in low latent dimensions, yielding better generative samples and downstream classification (Kalatzis et al., 2020, Chadebec et al., 2022).
• Enable meaningful morphing/interpolation along the data manifold, with geodesics staying in high-density regions and avoiding low-support regions where Euclidean interpolations fail (Sun et al., 2024, Shamsolmoali et al., 2023, Hartwig et al., 10 Oct 2025).
• Offer superior robustness in low-data regimes (medical imaging, scRNA-seq), with lower FID and higher precision/recall compared to Euclidean or non-geometry-aware baselines (Chadebec et al., 2022, Sun et al., 2024).
• Support scalable learning and accurate intrinsic dimension estimation for high-dimensional data with low-dimensional underlying structure (Diepeveen et al., 2024, Diepeveen et al., 26 Jan 2026).

6. Extensions: Nontrivial Latent Geometries and Theoretical Guarantees

The Riemannian autoencoder framework generalizes to accommodate:

• Complex/Kähler latent spaces: Decoders with complex-valued outputs are equipped with information-theoretic Kähler metrics, with regularization, volume sampling, and analytic curvature (Gracyk, 19 Nov 2025).
• Mixed-curvature and Lie group latent spaces: Latent variables reside in direct products of constant curvature spaces or Lie groups, with modular reparameterization, parallel transport, and geometric regularization (Skopek et al., 2019, Gong et al., 2019).
• Score-based/normalizing-flow models: Pullback metrics and global coordinate charts are constructed via flows or score maps, with provable reconstruction error and dimension-recovery theorems (Diepeveen et al., 2024, Diepeveen et al., 26 Jan 2026).
• Empirical convergence and approximation guarantees: Low-distortion and low-bending regularizers are shown to Γ-converge to corresponding geometric energies, ensuring that learned latents approximate isometric and extrinsically flat embeddings (Braunsmann et al., 2022).

7. Limitations and Open Problems

While Riemannian autoencoders offer clear advantages, key limitations include:

• Computational cost: Geodesic computation and metric inversion can pose scaling issues for high-dimensional latent spaces.
• Approximation error: Short-time heat kernel expansions and area-preserving wraps may become inaccurate on highly curved manifolds or when latent geometry does not align with data distribution (Kalatzis et al., 2020).
• Dependence on accurate volume and curvature estimation: Discrepancies between modeled and true geometry affect volume sampling and interpolation.
• Requirement of ground-truth manifold distances: Some methods depend on precomputed or high-quality diffusion or graph-based distances (Sun et al., 2024).
• Data requirements: Certain approaches (e.g., denoising projections, manifold-aware losses) require sampling or knowledge of small neighborhoods or local averages (Braunsmann et al., 2021, Braunsmann et al., 2022).

Open directions include scalable geodesic solvers for high-dimensional manifolds, joint estimation of latent curvature and dimension, geometry-aware flows for multi-modal or highly anisotropic data, and theoretical bounds on reconstruction and generalization for geometry-regularized autoencoders.

---

References:

• (Kalatzis et al., 2020) Variational Autoencoders with Riemannian Brownian Motion Priors
• (Gracyk, 19 Nov 2025) Complex variational autoencoders admit Kähler structure
• (Gong et al., 2019) Lie Group Auto-Encoder
• (Chadebec et al., 2022) A Geometric Perspective on Variational Autoencoders
• (Galaz-Garcia et al., 2022) Wrapped Distributions on homogeneous Riemannian manifolds
• (Shamsolmoali et al., 2023) VTAE: Variational Transformer Autoencoder with Manifolds Learning
• (Diepeveen et al., 2024) Score-based Pullback Riemannian Geometry: Extracting the Data Manifold Geometry using Anisotropic Flows
• (Grattarola et al., 2018) Adversarial Autoencoders with Constant-Curvature Latent Manifolds
• (Diepeveen et al., 26 Jan 2026) Riemannian AmbientFlow: Towards Simultaneous Manifold Learning and Generative Modeling from Corrupted Data
• (Skopek et al., 2019) Mixed-curvature Variational Autoencoders
• (Braunsmann et al., 2021) Learning low bending and low distortion manifold embeddings
• (Braunsmann et al., 2022) Convergent autoencoder approximation of low bending and low distortion manifold embeddings
• (Sun et al., 2024) Geometry-Aware Generative Autoencoders for Warped Riemannian Metric Learning and Generative Modeling on Data Manifolds
• (Chadebec et al., 2020) Geometry-Aware Hamiltonian Variational Auto-Encoder
• (Hartwig et al., 10 Oct 2025) Geodesic Calculus on Latent Spaces
• (Miolane et al., 2019) Learning Weighted Submanifolds with Variational Autoencoders and Riemannian Variational Autoencoders