Expected Signature on a Riemannian Manifold and Its Geometric Implications

Xi Geng, Hao Ni, C. Y. WangarXiv 2024 · DOI · arXiv

Abstract

On a compact Riemannian manifold $M$, we show that the Riemannian distance function $d(x,y)$ can be explicitly reconstructed from suitable asymptotics of the expected signature of Brownian bridge from $x$ to $y$. In addition, by looking into the asymptotic expansion of the fourth level expected signature of the Brownian loop based at $x \in M$, one can explicitly reconstruct both intrinsic (Ricci curvature) and extrinsic (second fundamental form) curvature properties of $M$ at $x$. As independent interest, we also derive the intrinsic PDE for the expected Brownian signature dynamics on $M$ from the perspective of the Eells-Elworthy-Malliavin horizontal lifting.

Introduction

In the field of rough path theory, the signature of a path is a collection of iterated integrals that acts as a unique descriptor for that path. While signatures are well-understood in flat Euclidean space, extending them to curved surfaces (Riemannian manifolds) is challenging because there is no natural way to "add" vectors at different points on a manifold. This research addresses a fundamental geometric inversion problem: If we only observe the average (expected) signature of random paths on an unknown manifold, can we figure out the shape of the manifold itself?

This is analogous to the famous "Can one hear the shape of a drum?" problem, but instead of using sound waves (eigenvalues of the Laplacian), it uses the statistical "fingerprints" of Brownian motion.

Methodology

The researchers employ several sophisticated mathematical frameworks to bridge stochastic analysis and differential geometry:

  1. Horizontal Lifting: To define Brownian motion on a manifold $M$ rigorously, the process is "lifted" to the orthonormal frame bundle $O(M)$—a higher-dimensional space that tracks both position and orientation. This allows the use of standard Euclidean-like stochastic differential equations (SDEs).
  2. Intrinsic PDEs: The authors derive a new tensor-algebra-valued parabolic partial differential equation (PDE) that describes how the expected signature evolves over time on the manifold.
  3. Small-Time Asymptotics: By analyzing what happens as the "lifetime" of a Brownian bridge (a path starting at $x$ and ending at $y$) approaches zero, they show the path concentrates around the geodesic (the shortest path).
  4. Computer-Assisted Algebra: The formulas for curvature reconstruction are algebraically dense, involving 4th-level signature tensors and 3rd-order time expansions. The authors utilized Wolfram Mathematica to verify the symbolic derivations, handling nearly a hundred distinct cases for the higher-order coefficients.
The complexity of the symbolic verification is evidenced by the high number of remaining cases processed for the $t^2$ and $t^3$ coefficients.

Key Insights and Ideas

Signatures as Geometric DNA

The expected signature of a Brownian bridge encodes the Riemannian distance $d(x, y)$. As the signature degree $n$ goes to infinity and time $t$ goes to zero, the normalized signature converges exactly to the distance. This suggests that the signature acts as a "geometric DNA" for the underlying space.

The Power of the 4th Level

For a Brownian loop (a path that starts and ends at the same point $x$), the first three levels of the expected signature are zero due to symmetry. The fourth level is the first non-trivial component, and it contains the essential information of the manifold's local geometry.

Intrinsic vs. Extrinsic Information

Most geometric methods (like heat kernel expansions) only reveal intrinsic properties—how the manifold looks to an observer within the space. The signature method is unique because it also reveals extrinsic properties—specifically the second fundamental form, which describes how the manifold is curved within a larger ambient space.

Polynomial Dependence

A major technical breakthrough in this paper is showing that the time scale $t$ only needs to be polynomially small relative to the signature degree $n$ ($t \approx n^{-6}$), rather than exponentially small. This makes the reconstruction theoretically more stable and computationally feasible.

Key Results

Distance Reconstruction

The authors provide an explicit limit formula for recovering the Riemannian distance from signature moments: $$\lim_{n \to \infty} (n! \|\mathbb{E}[\pi_n S(X_{t_n}) ]\|)^{1/n} = d(x, y)$$

Curvature Recovery

They derived explicit formulas for the coefficients of the small-time expansion of the 4th-level signature. By applying a "tensor contraction," they show that:

  • The $t^2$ coefficient recovers the metric tensor (the local ruler of the space).
  • The $t^3$ coefficient recovers a mixture of Ricci curvature and the mean curvature vector.

Limitations and Future Work

The current proofs assume the manifold is compact and has no boundary. Extending this to non-compact spaces would require handling potential "escapes" to infinity. Furthermore, the distance reconstruction currently only works for points within half the injectivity radius. While this paper is purely mathematical, these formulas could provide a theoretical foundation for new algorithms in Manifold Learning, where one needs to estimate the geometry of a hidden surface from noisy sequential data.

Summary updated

Abstract

On a compact Riemannian manifold $M,$ we show that the Riemannian distance function $d(x,y)$ can be explicitly reconstructed from suitable asymptotics of the expected signature of Brownian bridge from $x$ to $y$. In addition, by looking into the asymptotic expansion of the fourth level expected signature of the Brownian loop based at $x\in M$, one can explicitly reconstruct both intrinsic (Ricci curvature) and extrinsic (second fundamental form) curvature properties of $M$ at $x$. As independent interest, we also derive the intrinsic PDE for the expected Brownian signature dynamics on $M$ from the perspective of the Eells-Elworthy-Malliavin horizontal lifting.
BibTeX
@misc{geng2024expected,
  author = {Xi Geng and Hao Ni and C. Y. Wang},
  title = {{Expected Signature on a Riemannian Manifold and Its Geometric Implications}},
  year = {2024},
  eprint = {2407.13086},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.13086},
}
Figure