Joint state-parameter estimation — simultaneously tracking a system's state (where it is) and its parameters (how it behaves) — is hard because the two are coupled but have different characteristics. States evolve quickly and often linearly in the local approximation. Parameters change slowly or not at all, but their distribution can be highly non-Gaussian.
The authors (arXiv:2603.23039) split the problem using Rao-Blackwellization. Conditioned on the parameters, the state is Gaussian — so use a Kalman filter for it (exact, fast). The parameters are non-Gaussian — so use Stein Variational Gradient Descent, a particle-based method that approximates arbitrary distributions by transporting a set of particles along the steepest descent of the KL divergence.
The split is computationally efficient because each half uses the right tool. Kalman filtering handles the easy part exactly, leaving the hard part (parameter estimation) to the more expensive particle method. The theoretical analysis shows how errors in the parameter posterior propagate to the state estimates — the coupling is quantified, not ignored.
Validation includes a bioreactor system and a neural-network-augmented dynamical model. The latter is notable: the “parameters” being estimated are neural network weights, which means the method does online training of a neural network embedded in a dynamical model.
The through-claim: the right decomposition of a hard problem can be more valuable than a better algorithm for the whole problem. Splitting state from parameter estimation and using the analytically tractable part to simplify the intractable part is not an approximation — it's a Rao-Blackwellization, which reduces variance without introducing bias.