From Kalman's seminal paper "A New Approach to Linear Filtering and Prediction Problem", it is clear that Kalman's exposition is based on the following fundamental assumptions:

  1. Measurements that are linearly related to the state.
  2. Measurements are corrupted by white noise: 2a: Serially uncorrelated noise. 2b: Zero mean noise.
  3. Measurements follow the Gaussian distribution.

When these assumptions are met, the Kalman Filter is "the best filter of any conceivable form", according to Maybeck (1979) "Stochastic Models, Estimation, and Control, Vol 1", page 7.

I am looking for an overview of sources that deal with the effect on the estimation error of relaxing the stated assumptions, and I am hoping someone here can help me.

  • $\begingroup$ I think this is not quite what you're looking for, but you can rederive the Kalman filter using Bayesian estimation (french-metrology.com/publications/…), and substitute in your relaxed assumptions. $\endgroup$ – pscheidler Jan 4 '18 at 15:32
  • 1
    $\begingroup$ A formulation that is similar to Kalman Filtering but based on sequential Monte Carlo methods (and relaxes the linearity and Gaussian distribution assumptions) is the Particle Filter. The wikipedia page has tons of references. $\endgroup$ – hops Jan 4 '18 at 18:45

HI: I haven't read it but this discusses the non-linear case so it may be helpful atleast to get an idea of what the issues are. http://sites.utexas.edu/renato/files/2017/04/recursiveUpdate_v3.pdf

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.