Rolled Gaussian process models for curves on manifolds

1. Problem Statement

Modern statistical applications increasingly generate data in the form of curves $t \mapsto x(t)$ lying on a Riemannian manifold $\mathcal{M}$ (e.g., trajectories on spheres, shapes in shape spaces, covariance matrices, or 3D orientations). While Gaussian Processes (GPs) are the standard tool for modeling curves in Euclidean space ($\mathbb{R}^d$) due to their tractability and rich covariance structures, extending them to nonlinear manifolds is challenging.

Existing approaches face significant limitations:

• Tangent Space Projections: Mapping the entire manifold to a single tangent space (flattening) introduces geometric distortions, especially for large deviations.
• Stochastic Differential Equations (SDEs): Intrinsic Brownian motion models allow for generative modeling but suffer from Markovian constraints (limiting covariance structures) and computational intractability (likelihoods are hard to compute, exact simulation is difficult).
• Descriptive PCA: Principal Component Analysis on manifolds often relies on abstract covariance definitions along a base curve without providing a unified generative framework with clear estimator convergence rates.

The core problem is to define a generative statistical model for manifold-valued curves that preserves the geometric structure of $\mathcal{M}$, allows for flexible covariance modeling, and supports rigorous statistical inference (estimation and hypothesis testing).

2. Methodology: The Rolling Framework

The authors propose a novel approach based on the geometric concept of "rolling" a manifold along a curve without slipping or twisting. This induces a local isometry between the manifold and a tangent space.

The Four Maps

The methodology relies on four specific maps that facilitate moving between the manifold $\mathcal{M}$ and a fixed tangent space $T_b\mathcal{M}$ (identified with $\mathbb{R}^d$):

1. Unrolling ($\downarrow$): Maps a curve $\gamma$ on $\mathcal{M}$ to a curve $\gamma^\downarrow$ in the tangent space $T_b\mathcal{M}$. This is defined via parallel transport of the velocity vector: $\dot{\gamma}^\downarrow(t) = P^{\gamma}_{0 \leftarrow t} \dot{\gamma}(t)$.
2. Rolling ($\uparrow$): The inverse of unrolling; maps a curve in $T_b\mathcal{M}$ back to $\mathcal{M}$.
3. Unwrapping ($\downarrow^\gamma$): Maps a curve $x$ on $\mathcal{M}$ to $T_b\mathcal{M}$ relative to a reference curve $\gamma$. It combines the inverse exponential map (logarithm) at each point with parallel transport to the base tangent space.
4. Wrapping ($\uparrow^\gamma$): The inverse of unwrapping; maps a curve in $T_b\mathcal{M}$ back to $\mathcal{M}$ relative to $\gamma$.

The Rolled Gaussian Process (RGP):
The model is constructed by:

1. Defining a standard Euclidean Gaussian Process $z(t) \sim \text{GP}(m, K)$ in $\mathbb{R}^d$.
2. Mapping $z$ to the tangent space $T_b\mathcal{M}$ to get $y_b(t)$.
3. Defining a reference curve $\gamma$ as the rolling of the mean function $m_b$ (the image of $m$ in $T_b\mathcal{M}$).
4. Generating the final random curve $x(t)$ on $\mathcal{M}$ by wrapping the centered process $y_b(t) - m_b(t)$ around the reference curve $\gamma$.

This construction ensures that the resulting process $x$ is a "covariant Gaussian vector field" along $\gamma$, where the covariance structure is preserved via parallel transport.

3. Key Contributions

A. Theoretical Foundations

• Equivariance: The authors prove that the distribution of the Rolled Gaussian Process is equivariant with respect to the choice of the base point $b$ and the orthonormal frame $U$ in the tangent space. This ensures the model is intrinsic to the manifold geometry.
• Fréchet Mean Coincidence: A critical theoretical result (Theorem 4.1) establishes conditions under which the rolling of the Euclidean mean ($m^\uparrow_b$) equals the Fréchet mean of the resulting manifold process. This holds if:
  • $\mathcal{M}$ is a Hadamard manifold (non-positive curvature, e.g., symmetric positive-definite matrices).
  • $\mathcal{M}$ is a symmetric space and the distribution possesses geodesic symmetry.
    This result is vital because it allows the complex Fréchet mean to be estimated simply by rolling the Euclidean mean.

B. Estimation and Convergence Rates

The paper develops estimators for the mean and covariance parameters based on the "unwrapping" of observed data curves onto a single tangent space.

• Estimators: They propose simple estimators for the mean $m$ and covariance $K$ by unwrapping data relative to the sample Fréchet mean curve.
• $C^1$ Convergence: A major theoretical contribution is proving that the sample Fréchet mean curve converges to the population Fréchet mean in the $C^1$ topology (convergence of both the curve and its derivative) at the optimal Euclidean rate of $O_P(n^{-1/2})$. Previous literature typically only established uniform ($C^0$) convergence.
• Optimal Rates: The estimators for the mean and covariance functions achieve the same optimal convergence rates as in the Euclidean setting, provided the manifold satisfies bounded geometry and curvature conditions.

C. Parametric Models for Discrete Data

Recognizing that real-world data is discrete, the authors introduce a parametric version using basis functions (e.g., B-splines).

• They define a Rolled Matrix Normal (RMN) distribution for discrete-time curves.
• They provide Maximum Likelihood Estimators (MLE) and Least Squares estimators for the parameters, along with iterative algorithms for computation.
• They derive a two-sample test (analogous to Hotelling's $T^2$) for comparing mean curves on manifolds, utilizing permutation tests to handle the non-Euclidean geometry.

4. Results and Applications

The paper validates the methodology through numerical examples and a real-world application:

1. Synthetic Examples:
   • Unit Sphere ($S^2$): Simulating heteroscedastic curves to demonstrate the rolling/wrapping mechanics.
   • Symmetric Positive-Definite Matrices ($\text{Sym}_{>0}(2)$): Demonstrating the model on a non-compact, negatively curved manifold. Numerical results confirm the convergence of estimators as sample size increases.
2. Robotics Application:
   • Data: Orientations of a Franka robot arm's end-effector (elements of $SO(3)$) performing a task 60 times.
   • Analysis: The model successfully captured the mean trajectory and the heteroscedastic variability (increasing variance as the robot moves).
   • Inference: A two-sample test was performed on data collected in two sessions. The test statistic was extreme compared to the null distribution (via permutation), providing conclusive evidence that the two sessions had significantly different mean trajectories.

5. Significance

This work bridges a critical gap in geometric statistics by providing a computationally feasible, generative, and theoretically rigorous framework for modeling curves on manifolds.

• Generative Power: Unlike descriptive PCA, this is a full probabilistic model capable of simulation and likelihood-based inference.
• Geometric Fidelity: By using rolling/unrolling, the model avoids the distortions inherent in global projection methods while maintaining the ability to use standard Euclidean GP machinery.
• Statistical Rigor: The derivation of $C^1$ convergence rates and optimal estimation rates provides a strong theoretical foundation that was previously missing for manifold-valued curve data.
• Practical Utility: The inclusion of discrete-time parametric models and hypothesis testing tools makes the approach immediately applicable to robotics, medical imaging, and shape analysis.

In summary, the paper transforms the abstract concept of "rolling" into a practical statistical engine, enabling the application of powerful Gaussian process tools to complex, non-Euclidean data structures.