Recurrent Neural Networks and (Time-Uniform) Universal Approximation of Bayesian Filters

Abstract

We consider approximations of the Bayesian optimal filtering problem, e.g., estimating some conditional statistics of a latent time-series signal from an observation sequence. Classical approaches often rely on the use of assumed or estimated transition and observation models. Instead, we formulate a generic recurrent neural network framework and seek to learn directly a recursive mapping from observational inputs to the desired estimator statistics. The main focus of this article is on the approximation capabilities of this framework. We provide approximation error bounds for filtering in general noncompact domains. Importantly, we also consider strong time-uniform approximation error bounds that guarantee good long-time performance, i.e., we show under certain conditions that filter approximation errors do not accumulate unbounded over time. We discuss and illustrate a number of practical concerns and implications of these results.

Constructive Universal Approximation by Infinite-Dimensional Neural Networks via Gaussian-Weighted Ridgelet Transform
Sho Sonoda, Luca Galimberti, Yuanchao Xu +2
arXiv16 Oct 2026PreprintCC BY 4.0Full textAnalysis & PDEs StatisticsarXiv
Uniform-in-time relative entropy estimates for Kac's approximation of the Landau equation
Xuanrui Feng, Chenguang Liu, Zhenfu Wang
arXiv30 Aug 2026PreprintarXiv licenceFull textAnalysis & PDEs ProbabilityarXiv
Uniform W^2,m approximation of almost isometric maps
Sergio Conti, Georg Dolzmann, Stefan Müller
arXiv29 Sep 2026PreprintarXiv licenceFull textAnalysis & PDEsarXiv
Zero-one laws for uniform approximation via Gaussian and Eisenstein integers
René Pfitscher, Anurag Rao, Shucheng Yu +1
arXiv8 Oct 2026PreprintarXiv licenceFull textAlgebra & number theory Dynamical systemsarXiv
Twisted Diophantine approximation: asymptotic and uniform results
Taehyeong Kim, Vasiliy Neckrasov
arXiv4 Nov 2026PreprintCC BY 4.0Full textAlgebra & number theoryarXiv

Cite

BibTeX RIS

BibTeX

@article{bishop2026recurrent,
  title = {{Recurrent Neural Networks and (Time-Uniform) Universal Approximation of Bayesian Filters}},
  author = {Adrian Bishop and Edwin Bonilla and Pierre Del Moral},
  journal = {IEEE Transactions on Automatic Control},
  year = {2026},
  month = sep,
  doi = {10.1109/tac.2026.3655312},
  url = {https://inria.hal.science/hal-05735221v1},
  note = {CC BY},
}

RIS

TY  - JOUR
TI  - Recurrent Neural Networks and (Time-Uniform) Universal Approximation of Bayesian Filters
AU  - Adrian Bishop
AU  - Edwin Bonilla
AU  - Pierre Del Moral
PY  - 2026
DA  - 2026/09/02
JO  - IEEE Transactions on Automatic Control
DO  - 10.1109/tac.2026.3655312
UR  - https://inria.hal.science/hal-05735221v1
AB  - We consider approximations of the Bayesian optimal filtering problem, e.g., estimating some conditional statistics of a latent time-series signal from an observation sequence. Classical approaches often rely on the use of assumed or estimated transition and observation models. Instead, we formulate a generic recurrent neural network framework and seek to learn directly a recursive mapping from observational inputs to the desired estimator statistics. The main focus of this article is on the approximation capabilities of this framework. We provide approximation error bounds for filtering in general noncompact domains. Importantly, we also consider strong time-uniform approximation error bounds that guarantee good long-time performance, i.e., we show under certain conditions that filter approximation errors do not accumulate unbounded over time. We discuss and illustrate a number of practical concerns and implications of these results.
SN  - 0018-9286
ER  -

Abstract

Related

Cite