Geometric Milstein Scheme for Stochastic Differential Equations on SO(n) and SE(n)

Technical Summary: Geometric Milstein Scheme for Stochastic Differential Equations on SO(n) and SE(n)

Problem Statement
The paper addresses the numerical integration of nonlinear autonomous Itô stochastic differential equations (SDEs) evolving on the Lie groups $SO(n)$ (special orthogonal group) and $SE(n)$ (special Euclidean group). These equations take the form:
$$dX(t) = X B_I(X(t))dt + \frac{1}{2}X \Sigma_N(X)dt + \sum_{j=1}^d B_j(X(t))dW_j(t)$$
where $B_I$ and $B_j$ are drift and diffusion fields in the tangent space, and $\Sigma_N$ is a "pinning-drift" term in the normal space arising from the Itô correction.

The core challenge is developing Geometry-Preserving Numerical Schemes (GPNS) that achieve a Strong Convergence Rate (SCR) greater than $1/2$. Existing methods face a trade-off:

1. Euler-Maruyama variants (e.g., G-EM, TaSP-EM) preserve geometry but are limited to an SCR of $1/2$ (or $1/2 - \epsilon$).
2. Higher-order methods (e.g., S-RKMK based on Magnus expansion) can achieve higher orders but often rely on Stratonovich formulations (which fail for nonlinear diffusion) or involve computationally intractable infinite series truncations.
3. Standard Euclidean Milstein schemes achieve SCR 1 but fail to preserve the manifold constraint, causing trajectories to drift off the Lie group due to normal components in the drift and diffusion increments.

Methodology: Tangent Space Parametrization–Corrected Milstein (TaSP-CM)
The authors propose the TaSP-CM scheme, which extends the Euclidean Milstein method to Lie groups by combining a corrected Milstein increment in the Lie algebra with a geometric reconstruction step.

1. Tangent Space Parametrization (TaSP):
   Instead of evolving directly on the manifold, the method represents the local solution $R(t)$ on $SO(n)$ as:
   $$R(t) = R_0 + R_0 Z(t) + R_0 C(Z(t))$$
   where $Z(t) \in \mathfrak{so}(n)$ is a stochastic displacement in the Lie algebra, and $C(Z) = \sqrt{I - Z^\top Z} - I$ is a normal-space adjustment term that ensures $R(t)^\top R(t) = I$. This decouples the geometry preservation from the stochastic integration.

2. Corrected Milstein Increment:
   The scheme computes the increment $Z(t)$ in the Lie algebra using a Milstein-type update. Crucially, it replaces the standard Euclidean derivative $\nabla B_j(B_{j'})$ with the intrinsic Covariant Derivative (CovD) $D^O_{B_j}(B_{j'})$.
   $$D^O_{B_j}(B_{j'}) = \nabla_{B_j} B_{j'} + \frac{1}{2}[B_j, B_{j'}]$$
   This correction accounts for the non-commutative structure of the Lie algebra (via the Lie bracket) and ensures the increment remains in the tangent space, avoiding the introduction of spurious normal components that would violate the manifold constraint.

3. Handling Non-Commutative Noise:
   For non-commutative noise (where $[B_j, B_{j'}] \neq 0$), the scheme requires the approximation of double Itô integrals. The authors utilize a known approximation method (Algorithm 2) involving truncated Fourier series of Gaussian variables to maintain the required accuracy without intractable computational costs.

4. Extension to SE(n):
   Leveraging the semidirect-product structure $SE(n) = SO(n) \ltimes \mathbb{R}^n$, the method extends naturally. The rotational component follows the $SO(n)$ dynamics, while the translational component evolves as a standard Euclidean SDE, coupled through the rotation matrix.

Key Contributions

• First Computationally Tractable SCR 1 Scheme: The paper introduces the first computationally feasible GPNS for nonlinear SDEs on $SO(n)$ and $SE(n)$ that achieves a strong convergence rate of 1 under both commutative and non-commutative noise.
• Theoretical Proof of Convergence: The authors provide rigorous proofs (Theorems 1 and 2) demonstrating that the TaSP-CM scheme attains SCR 1. The analysis decomposes the error into the difference between the TaSP-CM trajectory and a "frozen-field" Euclidean Milstein series, showing that the geometric correction term introduces only higher-order errors ($O(\delta^2)$ in mean-square).
• Generalization to SE(n): The framework is successfully extended to $SE(n)$ (Theorem 4), maintaining the SCR 1 guarantee for rigid-body pose estimation problems.
• Algorithmic Implementation: The paper details the specific algorithms (Algorithms 1 and 3) for implementation, including the handling of the square-root adjustment and double integral approximations.

Results
Numerical experiments were conducted on $SO(n)$ and $SE(3)$ under both commutative and non-commutative noise settings:

• Convergence Rate: The TaSP-CM scheme consistently demonstrated an empirical SCR of 1, matching the theoretical prediction. In contrast, the TaSP-EM and S-RKMK methods exhibited SCR 1/2, and the standard Euclidean Milstein (Eu-Mil) scheme failed to converge due to geometric drift.
• Geometry Preservation: The scheme maintained orthogonality (for $SO(n)$) with geometric deviations on the order of $10^{-11}$ to $10^{-13}$, comparable to other geometry-preserving methods, while Eu-Mil drifted significantly off the manifold.
• Efficiency: The computational cost scales as $O(\delta^{-3/2})$ Gaussian calls for non-commutative noise, matching the Euclidean Milstein method and offering a significant improvement over the $O(\delta^{-2})$ cost required by lower-order geometric methods to achieve similar accuracy.

Significance
The paper claims to bridge a critical gap in the numerical analysis of SDEs on Lie groups. Prior to this work, there was no computationally tractable method that simultaneously preserved the underlying geometric constraints and achieved a strong convergence order greater than $1/2$ for nonlinear SDEs. By successfully extending the Milstein scheme to curved manifolds, the TaSP-CM method provides a practical tool for applications requiring high numerical accuracy and strict geometric fidelity, such as satellite control, robot pose estimation, object tracking, and stochastic control on manifolds. The authors note that this approach opens the door for future extensions to other matrix manifolds (e.g., Stiefel, affine groups) and applications in Riemannian diffusion models.