Manifold-Valued Noise

• Manifold-valued noise is a type of stochastic perturbation that respects the intrinsic geometry of nonlinear manifolds, enabling more accurate modeling and analysis.
• Geometry-aware methods employ tangent-space projections and Riemannian exponential maps to inject or correct noise while preserving manifold structure.
• Robust statistical estimation and denoising techniques on manifolds improve performance in imaging and signal processing by leveraging intrinsic noise models.

Manifold-valued noise refers to stochastic perturbations in data that reside intrinsically on a manifold, as opposed to standard Euclidean (ambient) noise, which typically fails to respect the geometric structure of underlying data manifolds. This concept is central to statistical estimation, geometric learning, denoising, and signal processing where the data are constrained to nonlinear spaces such as spheres, rotation groups, symmetric positive definite matrices, or more general Riemannian manifolds. Modern approaches to modeling, injecting, and correcting noise on manifolds leverage differential-geometric constructions and exploit the manifold's local tangent structure or stochastic processes defined intrinsically on the manifold.

1. Foundational Models of Manifold-Valued Noise

The most prevalent setting assumes that the true signal $y_i \in \mathcal{M}$—with $\mathcal{M}$ a $d$-dimensional, compact, boundaryless $C^2$ manifold embedded in $\mathbb{R}^D$—is corrupted by noise generated either in the ambient space or on the manifold's tangent bundle. The classical ambient model posits

$$x_i = y_i + \xi_i, \quad \xi_i \sim N(0, \sigma^2 I_D)$$

with possibly unbounded Gaussian support. For $k \leq 4$, one has moment and concentration bounds on $\|\xi_i\|$ governed by the variance parameter $\sigma^2$ (Yao et al., 2019). This model is especially prevalent in manifold learning and fitting under high-dimensional noise, but suffers in geometric fidelity as ambient noise typically pushes samples away from $\mathcal{M}$ in directions normal to the manifold.

Intrinsic noise on manifolds is modeled through the exponential map,

$\mathcal{M}$0

where $\mathcal{M}$1 is a zero-mean Gaussian in the tangent space, and $\mathcal{M}$2 is the Riemannian exponential map. The resulting law—Riemannian (or geodesic) Gaussian—has density proportional to $\mathcal{M}$3, where $\mathcal{M}$4 denotes geodesic distance (Bergmann et al., 2018, Laus et al., 2016).

Manifold-valued noise can further be realized through geometry-aware perturbations:

• Tangent-space projection, where ambient Gaussian noise is projected onto $\mathcal{M}$5 and mapped back via $\mathcal{M}$6; and
• Intrinsic manifold stochastic processes such as discrete or continuous Riemannian Brownian motion, which remains within $\mathcal{M}$7 by construction (Jacobsen et al., 24 Sep 2025, McErlean et al., 21 Mar 2026).

2. Geometry-Aware Noise Injection: Theory and Practice

Standard data augmentation with ambient Gaussian noise disrupts the manifold hypothesis in high-dimensional learning: almost all such noise is orthogonal to $\mathcal{M}$8, producing perturbed samples that lie off the data manifold and degrade the structure of learned representations (Jacobsen et al., 24 Sep 2025).

Geometry-aware noise injection addresses this by restricting perturbations to the tangent space $\mathcal{M}$9 (for $d$0), followed by application of the exponential map: $d$1 with $d$2 the orthogonal projector onto $d$3 (Jacobsen et al., 24 Sep 2025). For small $d$4, $d$5 remains close to $d$6 on $d$7. An alternative is to simulate Riemannian Brownian motion via iterative steps

$d$8

which yields samples distributed according to the Laplace–Beltrami operator on $d$9.

For learned or analytic manifolds, the tangent basis may be approximated using the decoder’s Jacobian or local PCA, enabling projection and geometry-aware noise even when the global structure is not explicit (Jacobsen et al., 24 Sep 2025, Yao et al., 2019).

Geometry-aware noise regularizes learned models primarily through tangential gradient penalties. While ambient noise regularizes via the ambient Laplacian, geometry-aware noise yields a penalty proportional to the Laplace–Beltrami operator, thus preserving the manifold structure during training or data augmentation.

3. Statistical Estimation and Denoising under Manifold-Valued Noise

Multiple methods extend classical denoising, regression, and inference to manifold-valued settings through appropriate modeling of noise:

• Intrinsic MMSE estimation: Given noisy linearized observations $C^2$0 (with $C^2$1 and $C^2$2 living in $C^2$3), the minimum mean-square error estimator generalizes via the exponential chart and intrinsic mean and covariance (Laus et al., 2016).
• Nonlocal patch-based denoising: Estimates manifold-valued patches by intrinsic Bayes/MMSE theory. Local statistics are computed in the tangent space at the Karcher mean, and denoised patches are mapped back via the exponential map. This approach improves mean-squared error in $C^2$4, $C^2$5, and $C^2$6 valued images relative to TV-based denoising (Laus et al., 2016).
• Variational regularization (e.g., TV, TGV): Data-fidelity terms use squared geodesic distances, modeling noise as Riemannian Gaussian; manifold total variation and higher-order differences are employed for regularization (Weinmann et al., 2013, Bergmann et al., 2018). Optimization algorithms include subgradient descent, cyclic/parallel proximal-point algorithms, and half-quadratic methods, all adapted to manifold geometry and proven to converge on Hadamard manifolds (Weinmann et al., 2013, Bergmann et al., 2018).

Empirical results consistently demonstrate that manifold-aware noise modeling, estimation, and denoising outperform ambient or naive methods across applications in imaging (Diffusion Tensor Imaging, InSAR, shape analysis) (Weinmann et al., 2013, Laus et al., 2016, Bergmann et al., 2018).

4. Manifold Learning, Fitting, and Recovery under Unbounded Noise

Fitting a manifold to noisy, high-dimensional data under ambient Gaussian noise presents unique challenges, especially when the noise is unbounded. Classical tangent-space based estimators are impaired: tangent-space estimation becomes blurred, leading to inaccuracies in the learned manifold.

The direct tangent-projection approach of Yao & Xia (Yao et al., 2019) introduces a robust estimator:

• At each candidate point $C^2$7 with $C^2$8 small, local weighted means and aggregated normal projectors are computed.
• PCA on neighborhoods gives local normal projections that are weighted and averaged.
• The manifold estimator is the zero-set of the resulting bias map, i.e., points $C^2$9 where the projected displacement from the local mean vanishes in the normal space.

Their method establishes high-probability $\mathbb{R}^D$0 Hausdorff risk under Gaussian noise and demonstrates smoothness (reach) properties for the estimated manifold. Numerical experiments on standard synthetic manifolds and real face-image denoising under large $\mathbb{R}^D$1 confirm the estimator’s stability and convergence without resorting to noise truncation or subsampling.

5. Robustness to Outliers, Gross Errors, and Non-Gaussian Perturbations

Manifold-valued observations are often contaminated not only by regular noise, but also by gross errors or adversarial outliers. Two principal strategies emerge:

• Robust location and regression via extrinsic medians: The extrinsic median—minimizer of the sum of Euclidean distances in the embedding space—achieves a 50% breakdown point and bounded influence. The extrinsic median and its regression analog filter both heavy-tailed and adversarial noise effectively, with efficient Weiszfeld-type solvers (Lee, 2021).
• Correction of gross errors through tangent-space modeling: The PALMR paradigm models each gross error as a large tangent vector $\mathbb{R}^D$2, so $\mathbb{R}^D$3. Joint estimation alternates between correcting corrupted responses by geodesic proximal steps and regression via geodesic-linear models in a block-coordinate minimization framework. The algorithm converges under Kurdyka–Łojasiewicz and Lipschitz assumptions and recovers ground-truth up to 80% corruption (Zhang et al., 2017).

Both approaches reference manifold geometry to achieve robustness against a wide array of perturbations, extending classical robust statistics into nonlinear ambient and intrinsic settings.

6. Stochastic Diffusion and Noise Processes on Manifolds

A powerful class of manifold-valued noise arises from stochastic differential equations (SDEs) on manifolds, especially in time series and biomedical applications. The model

$\mathbb{R}^D$4

defines a manifold-valued Itô diffusion with drift $\mathbb{R}^D$5 and diffusion vector fields $\mathbb{R}^D$6 intrinsic to $\mathbb{R}^D$7. Nonparametric estimation of the drift and diffusion from trajectories employs Nadaraya–Watson kernel smoothing, with bias corrections accounting for curvature and occupation-density estimation. Admissibility and convergence rates depend on bandwidth–interval scaling and recurrence of $\mathbb{R}^D$8 (McErlean et al., 21 Mar 2026).

Empirically, manifold-valued diffusion models capture temporal noise and variability without collapsing onto ambient artifacts, and tangent space estimation from local diffusion matrices supports improved statistical efficiency and geometric fidelity (McErlean et al., 21 Mar 2026).

7. Applications, Comparative Evaluation, and Software Support

Manifold-valued noise and its careful handling have concrete impact in:

• Diffusion MRI and medical imaging: denoising $\mathbb{R}^D$9, $$x_i = y_i + \xi_i, \quad \xi_i \sim N(0, \sigma^2 I_D)$$0, and $$x_i = y_i + \xi_i, \quad \xi_i \sim N(0, \sigma^2 I_D)$$1 images (Weinmann et al., 2013, Laus et al., 2016).
• Machine learning and generative modeling: geometry-aware noise injection for regularization, augmentation, and domain adaptation, extending to learned manifolds via decoder Jacobians (Jacobsen et al., 24 Sep 2025).
• Shape analysis, remote sensing, and directional statistics: robust regression and denoising under heavy-tailed and gross error models (Lee, 2021, Zhang et al., 2017).

Available toolboxes such as Manopt and MVIRT support manifold-optimization schemes, generic exponential/log maps, and proximal algorithms suited to manifold-valued denoising, regression, and fitting (Bergmann et al., 2018).

• (Yao et al., 2019) Yao & Xia, "Manifold Fitting under Unbounded Noise"
• (Jacobsen et al., 24 Sep 2025) "Staying on the Manifold: Geometry-Aware Noise Injection"
• (Lee, 2021) "Robust Extrinsic Regression Analysis for Manifold Valued Data"
• (Zhang et al., 2017) "Multivariate Regression with Gross Errors on Manifold-valued Data"
• (Weinmann et al., 2013) "Total variation regularization for manifold-valued data"
• (McErlean et al., 21 Mar 2026) "Functional Estimation of Manifold-Valued Diffusion Processes"
• (Laus et al., 2016) "A Nonlocal Denoising Algorithm for Manifold-Valued Images Using Second Order Statistics"
• (Bergmann et al., 2018) "Recent Advances in Denoising of Manifold-Valued Images"