📖
Nonlinear extensions of the gaussian pairwise kalman filter
Clément Fernandes, Zied Bouyahia, Stéphane Derrode

Nonlinear extensions of the gaussian pairwise kalman filter

Science & Mathematics Published 2026 Verification Pending

Overview & Synopsis

The Pairwise Kalman Filter (PKF), built upon the Pairwise Markov Model (PMM) framework, extends the classical Kalman filter by relaxing the conditional independence assumptions between states and observations. This richer dependency structure enables the modelling of phenomena such as measurement back-action and recursive observation dependencies, while preserving the recursive and optimal nature of the classical Kalman paradigm; existing formulations, however, are restricted to linear and Gaussian systems. This paper introduces two nonlinear extensions of the PKF: the Extended Pairwise Kalman Filter (EPKF), based on a firstorder Taylor linearization of the pairwise state-space model, and the Unscented Pairwise Kalman Filter (UPKF), which adapts the sigma-point propagation strategy to the pairwise dependence structure. For both filters, we establish a Joseph form of the covariance update equation, derive when the classical nonlinear models (EKF, UKF) are particular cases of the proposed pairwise formulation, and conversely that any pairwise model can be rewritten as a classical augmented-state system. Numerical experiments on synthetic nonlinear systems with bidirectional state-observation feedback show that both EPKF and UPKF achieve estimation accuracy comparable to a Pairwise Particle Filter while remaining consistent in the NEES and NIS senses. A back-action sweep further reveals that a practitioner who ignores the pairwise structure (and uses a naive nonlinear EKF on a Markov misspecification) incurs a Mean Squared Error two orders of magnitude larger than the proposed pairwise filters, illustrating the practical cost of the modelling shortcut. Finally, the filters are validated on a real financial dataset (S&P 500 daily returns), where the joint dynamics of latent logvolatility and log-squared returns are estimated from training data using a neural network, confirming the practical applicability of the pairwise framework to real-world nonlinear systems. All implementations are freely available.

Indexed from repository: HAL Open Science (hal.science)
Loading document from verified source...
Keyboard Shortcuts: ← / → for pages • + / - for zoom • 0 to fit width Powered by Mozilla PDF.js v3.11

Document Information

Document Title Nonlinear extensions of the gaussian pairwise kalman filter
Author / Organization Clément Fernandes, Zied Bouyahia, Stéphane Derrode
Subject & Discipline Multidisciplinary Academic Open Research
Categories Science & Mathematics
Publication Year 2026
License / Rights Open Access / Creative Commons
Resource Verification Listed PDF (Verification Pending)
Original Host Website HAL Open Science (hal.science)
File Format application/pdf (Portable Document Format)
Language EN
Catalog ID DOC-252517
Original Source Link Visit Source Page on hal.science »