Closed-Form Bayesian Updates on Rigid-Body Motion Groups — No Linearization Required

A new conjugate distribution family on SE_n(3) enables exact Bayesian fusion for UAV pose/velocity estimation, beating Lie-group Kalman filters in simulation.

Imagine you're tracking a package on a conveyor belt, but the belt curves through 3D space, rotates, and tilts — and your only tools are rulers designed for flat tables. Every time you try to measure uncertainty about where the package is headed, you have to pretend the curved belt is locally flat, introducing error. This paper builds a ruler that works natively on the curved surface. The committed claim: a new family of probability distributions — Bayesian Conjugate Distributions (BCDs) — defined directly on the extended special Euclidean group SE_n(3), which admit exact closed-form Bayesian updates for a broad class of measurement models arising in motion estimation. No linearization, no local-coordinate singularities. The conjugacy is the load-bearing novelty: when a measurement arrives, the posterior stays in the same distribution family as the prior, so you never need to approximate the update step. The architecture is elegant in its simplicity. BCDs are constructed by restricting matrix Gaussian distributions from the ambient Euclidean space down to SE_n(3). This avoids the singularity problems that plague Euler-angle or quaternion parameterizations. The paper derives two optimal point estimates from these distributions: a maximum-a-posteriori (MAP) estimate and a minimum-mean-squared-error (MMSE) estimate. It further shows that moment matching within the BCD family minimizes KL divergence, and provides maximum likelihood estimation — both in closed form. The entire pipeline lives on the Lie group, not in a linearized tangent space. On the ladder, the paper compares against two established Kalman filters on Lie groups: the Invariant Extended Kalman Filter (InEKF) and the Unscented Kalman Filter on Lie Groups (UKF-LG). In a simulated UAV flight scenario tracking pose and velocity with landmark measurements, the BCD-based estimator outperforms both, especially under large initial uncertainties and high measurement noise — precisely the regimes where linearization-based methods degrade. The improvement is most pronounced in attitude (rotation) estimation, where the nonlinearity bites hardest. Integrity is a mixed bag typical of theoretical statistics papers. The mathematical derivations appear rigorous — proofs of conjugacy, optimality of moment matching, and closed-form update equations are provided. But all validation is numerical simulation on a single UAV scenario. No hardware experiments, no public benchmark, no independent replication. The comparison baselines (InEKF, UKF-LG) are legitimate and current, not straw men, which is a plus. However, the simulated environment is relatively benign — no sensor dropouts, no outlier measurements, no model mismatch. The milestone question maps to real-world deployment. The paper demonstrates SE_2(3) (pose + velocity, a 5×5 matrix group) in simulation. The next concrete gate is a hardware demonstration on a real UAV with real IMU/camera data, where model mismatch and non-Gaussian noise would stress-test whether BCD's closed-form elegance survives contact with reality. Beyond that, scaling to SE_n(3) for n > 2 (incorporating accelerometer biases, gyroscope biases — the full navigation state) would be the bridge to replacing extended Kalman filters in production INS/GPS systems. The obvious experiment the authors didn't run: real sensor data. Given that the entire pipeline is analytical and the paper is 17 pages of dense theory with 4 figures, the most likely read is (a) — the authors are mathematicians/statisticians who built the theory and ran the minimum viable simulation, and hardware validation is either a collaboration they haven't secured or the next paper. The absence of any discussion of computational cost per update (wall-clock time vs. InEKF) also stands out — if closed-form updates are expensive matrix operations on 5×5 groups, the practical advantage over a fast EKF might narrow.