Riemannian AmbientFlow

• Riemannian AmbientFlow is a mathematical framework that integrates normalizing flows, Riemannian geometry, and manifold learning to extract latent structures from high-dimensional data.
• It employs rigorous variational inference and geometric regularization to manage measurement corruption and ensure that generative models respect the intrinsic manifold properties.
• Empirical results on synthetic and real-world datasets demonstrate superior performance in geodesic recovery, density estimation, and inverse problem solutions.

Riemannian AmbientFlow denotes a set of mathematical frameworks and algorithms that integrate normalizing flows, Riemannian geometry, and manifold learning, leveraging the ambient embedding space for probabilistic modeling, geometric inference, and inverse problems. This class of methods provides principled approaches for both extracting nonlinear manifold structure from (possibly corrupted) high-dimensional data and ensuring that the corresponding generative models respect the differential and topological properties of the latent manifold. The Riemannian AmbientFlow methodology appears in various domains, including generative modeling under measurement corruption, density learning on embedded manifolds, and the geometric generalization of dynamical systems such as optical flow. Foundational works include (Diepeveen et al., 26 Jan 2026, Gemici et al., 2016, Yu et al., 30 May 2025), and (Bauer et al., 2014).

1. Mathematical Foundations and Model Formulation

Riemannian AmbientFlow is predicated on the observation that high-dimensional data often concentrate near (unknown) low-dimensional smooth manifolds embedded in a Euclidean ambient space $\mathbb{R}^D$. The formal model typically consists of:

• A "clean" data space $x \in \mathbb{R}^d$ with unknown density $p_\text{data}(x)$.
• A measurement or corruption model: $y = A x + n$, with $A \in \mathbb{R}^{m \times d}$ and additive noise $n \sim p_\text{noise}$, so $$y \sim p_\text{data}^A := A_\#p_\text{data} * p_\text{noise}$$.
• A parameterized normalizing flow prior $$\phi_\theta : \mathbb{R}^d \rightarrow \mathbb{R}^d$$, typically a diffeomorphism, with latent $z = \phi_\theta(x)$ and base density $z \sim \mathcal{N}(0,I)$, giving induced density

$$p_\theta(x) = (2\pi)^{-d/2} \exp\left(-\frac12 \phi_\theta(x)^T \phi_\theta(x)\right) |\det D_x\phi_\theta|.$$

The corresponding measurement-model density is $\psi_\theta(y) = (A_\#p_\theta)*p_\text{noise}$.

AmbientFlow methods extend beyond Euclidean flows by pulling back Riemannian geometry from the ambient space, equipping the latent or model space with a data-driven metric structure defined by the flow decoder $f = \phi_\theta^{-1}$: $g_z = J_f(z)^T J_f(z)$, where $J_f$ denotes the Jacobian (Diepeveen et al., 26 Jan 2026, Gemici et al., 2016, Yu et al., 30 May 2025).

Normalizing flows on manifolds (Gemici et al., 2016) formalize the mapping between latent coordinates and the embedded manifold by using charts $$\varphi : U \subset \mathbb{R}^n \to M \subset \mathbb{R}^N$$ and adjusting densities with the Riemannian volume correction $[\sqrt{\det G(x)}]^{-1}$, where $G(x) = J_\varphi(x)^T J_\varphi(x)$. The general "AmbientFlow" is then realized by conjugating a flow $T$ in the latent space with $\varphi$.

2. Variational Inference and Training Objectives

Empirical learning in Riemannian AmbientFlow employs a variational framework tailored to corrupted observations. The principal training objective arises from a variational lower bound (ELBO) on the model evidence:

$$L_\mathrm{VLB}(\theta,\phi) = \mathbb{E}_{y \sim p_\text{data}^A} \left[ \log \mathbb{E}_{x^{(a)} \sim q_\phi(\cdot|y)} \Big\{ \exp(\log p_\theta(x^{(a)}) + \log p_\text{noise}(y - Ax^{(a)}) - \log q_\phi(x^{(a)}|y)) \Big\} \right]$$

where $q_\phi(x|y)$ is a flow-based variational posterior.

The complete loss function incorporates the variational bound, a geometric regularizer penalizing curvature in the learned embedding (e.g., $\|D_0\phi_\theta^{-1}\|_F$), and possibly a negative log-likelihood over a small set of clean reference samples:

$$\min_{\theta,\phi} -L_\mathrm{VLB}(\theta,\phi) + \lambda R_\text{geo}(\phi_\theta) - \mu \mathbb{E}_{x \sim p_\text{ref}}[\log p_\theta(x)]$$

(Diepeveen et al., 26 Jan 2026).

In multi-chart constructions designed for manifolds with nontrivial topology, multiple degenerate normalizing flows $f_c: \mathbb{R}^d \to \mathbb{R}^D$ are trained, and densities are combined via soft responsibilities $r_c(x)$ to cover the full manifold (Yu et al., 30 May 2025).

3. Riemannian Geometry and Manifold Structure Extraction

A defining feature of Riemannian AmbientFlow is the extraction and utilization of differential structure from data. The flow decoder $f = \phi_\theta^{-1}$ induces a Riemannian metric in latent space, enabling explicit computations of:

• Pullback inner products: $\langle u, v \rangle_g = u^T J_f(z)^T J_f(z) v$ for $u,v \in T_z \mathbb{R}^d$.
• Geodesic computation, exponential and logarithm maps via the metric.
• Riemannian Principal Components Analysis (PCA) at a basepoint $\overline{x}$, yielding local linearizations.

The Riemannian Autoencoder (RAE) architecture further parameterizes the manifold with an encoder mapping $x \mapsto U^T \log^\phi_{\overline{x}}(x)$ and decoder $z \mapsto \exp^\phi_{\overline{x}}(Uz)$, where the pullback log and exponential maps are those associated to the learned flow embedding (Diepeveen et al., 26 Jan 2026).

In the context of density estimation on arbitrary manifolds, the volume correction term $\sqrt{\det G(x)}$ ensures that pushforward and pullback of densities respect the intrinsic geometry (Gemici et al., 2016).

4. Theoretical Guarantees and Recovery Properties

Riemannian AmbientFlow is equipped with formal guarantees under regularity and measurement assumptions:

1. Distributional Recoverability: If the measurement operator $A$ satisfies a Restricted Isometry Property (RIP) over the RAE-decoder range with constant $\delta \in (0,1)$, any constrained minimizer $(\hat{\theta},\hat{\phi})$ with decoder reconstruction error $\omega$ yields the Wasserstein-1 distance bound:

$$W_1(p_{\hat{\theta}}, p_\text{data}) \le 2\omega \left(1 + \frac{\|A\|}{\sqrt{1 - \delta}}\right)$$

[(Diepeveen et al., 26 Jan 2026), Theorem 3.1].

1. Smoothness and Bi-Lipschitzness: For additive-coupling flow architectures with linear layers, the RAE decoder $D$ is (bi-)Lipschitz and has a Lipschitz Jacobian with explicit bounds on the Lipschitz constants $m_1$, $m_2$, and $M$ in terms of the flow parameters (singular values of component matrices and polynomial bounds) [(Diepeveen et al., 26 Jan 2026), Proposition 4.1].
2. Inverse Problem Guarantees: For new measurements $y = Ax^* + n$, solving $\min_{z \in \mathbb{R}^r} \frac12 \|A D(z) - y\|^2$ via gradient descent converges linearly—under a Range-Restricted Isometry Condition (RRIC) and certain Lipschitz parameter relationships—to a neighborhood of the true $z^*$, up to noise-dependent accuracy [(Diepeveen et al., 26 Jan 2026), Theorem 4.2].

Multi-chart AmbientFlow generalizes these guarantees to manifolds with nontrivial topology by assembling global densities through smoothly glued local flows, ensuring topological correctness and metric accuracy (Yu et al., 30 May 2025).

5. Algorithms and Numerical Realization

The algorithmic realization of Riemannian AmbientFlow spans several regimes:

• Flow Training: Flows are trained either jointly (EM or direct MLE) in the single- or multi-chart setting, with log-likelihoods incorporating manifold volume terms. Numerical differentiation (autodiff) computes required Jacobians.
• Chart Corrections: For embedding manifolds, the formal change-of-variables includes, for a chart map $\varphi$, not only the standard flow Jacobians but also corrections from $\det D\varphi$ and its inverse; in the sphere example, closed-form expressions allow efficient evaluation (Gemici et al., 2016).
• Chart Gluing: For nontrivial topologies, several local flows are "glued" using soft partition-of-unity weights $r_c(x)$, forming a mixture (Yu et al., 30 May 2025).
• Geodesic Computation: Geodesics are numerically integrated in the latent space, transferring through flow Jacobians as prescribed by local Riemannian metrics and mixed-chart Euler schemes (Yu et al., 30 May 2025).
• Inverse Problems: Once the RAE decoder is learned, new inverse problems are solved by optimizing in latent space, using gradient-based algorithms with metric regularity guarantees (Diepeveen et al., 26 Jan 2026).

6. Empirical Results and Applications

Riemannian AmbientFlow frameworks have demonstrated significant empirical performance across manifold learning, generative modeling, and inverse problems:

• Synthetic Manifolds: On low-dimensional manifolds, Riemannian AmbientFlow recovers the clean data distribution and reconstructs geodesic curves. Ablations confirm the necessity of geometric and reference-sample regularization (Diepeveen et al., 26 Jan 2026).
• Image Data (MNIST): On blurred and noisy MNIST datasets, with minimal clean references, the framework generates sharp samples, learns smooth pullback geodesics, and surpasses classical total variation (TV) methods in inverse reconstructions as measured by PSNR and MSE (Diepeveen et al., 26 Jan 2026).
• Topological Manifold Learning: Multi-chart flows recover nontrivial topological invariants, such as $H_1$ cycles on the torus and correct persistence diagrams for the sphere, outperforming single-flow approaches in both geodesic accuracy and homology (Yu et al., 30 May 2025).
• Dynamic Manifolds: In image sequence analysis, Riemannian AmbientFlow enables optical flow computation on evolving surfaces that outperform flat-domain analogues, yielding lower angular error in regions of high curvature or deformation (Bauer et al., 2014).

The following table presents sample empirical outcomes as reported in (Diepeveen et al., 26 Jan 2026) and (Yu et al., 30 May 2025):

Task	Model	Quantitative Outcome
Synthetic $\mathbb{R}^5 \rightarrow \mathbb{R}^3$	Riemannian AmbientFlow	Geodesic recovery, sharp samples
MNIST (14x14, noisy)	Riemannian AmbientFlow	TV outperformed in PSNR/MSE
Sphere $\mathbb{S}^2$	Multi-chart Flow	Geodesic errors $<10^{-3}$, correct $H_2$
Torus	Multi-chart Flow	Recovers two $H_1$ cycles

In all cases, metric and topological fidelity are confirmed by both persistent homology and manifold-geodesic statistics.

7. Generalizations and Connections

The Riemannian AmbientFlow paradigm includes or connects with:

• Normalizing Flows on Manifolds: Generalizing Euclidean flows to arbitrary submanifolds through coordinate charts, Jacobian corrections, and volume forms (Gemici et al., 2016).
• Multi-Chart and Atlas Approaches: Covering nontrivial topology and handling metric singularities by learning an atlas of flows and gluing them with soft weights (Yu et al., 30 May 2025).
• Higher-Order Geometric Flows: In geometric analysis, the "AmbientFlow" concept also refers to flows of Riemannian metrics, such as the Ambient Obstruction Flow evolving along the Fefferman–Graham tensor, with smoothing and singularity analysis for curvature (Lopez, 2015).
• Generalized Dynamic Systems: Applications extend to optical flow on moving manifolds, using ambient metrics, time-dependent geometry, and covariant PDEs capturing both spatial and temporal regularity (Bauer et al., 2014).

This suggests Riemannian AmbientFlow serves as both a methodological and structural bridge linking generative modeling, manifold learning, and geometric analysis in the ambient embedding framework.