1
$\begingroup$

I am struggling with quotes like

The EnKF applies a Markov chain Monte Carlo (MCMC) method ... (1, p. 6)

or

In fact, the Kalman filter is a MCMC algorithm in the case of a linear and Gaussian state space model with known parameters. (2, p. 5)

I cannot find the construction of a Markov chain in the description of neither linear nor ensemble Kalman filter. Could anyone please clarify how this is meant?


[1] Evensen, Geir. "The ensemble Kalman filter for combined state and parameter estimation." Control Systems, IEEE 29.3 (2009): 83-104.

[2] Johannes, Michael, and Nicholas Polson. "MCMC methods for financial econometrics." The Handbook of Financial Econometrics 65 (2002).

$\endgroup$
  • 1
    $\begingroup$ IMHO, the second quote is incorrect. The Kalman filter with backward sampling, usually referred to as the forward filtering backward sampling algorithm, provides a Monte Carlo (but not a Markov chain) approach to jointly sampling the states in a linear Gaussian state space model with known parameters. $\endgroup$ – jaradniemi Apr 7 '16 at 16:47
  • 2
    $\begingroup$ Could you add references to the sources you cite rather than just linking to them as "source"? This would help future-proof your question in case external links change address. $\endgroup$ – Silverfish Apr 7 '16 at 18:37
  • $\begingroup$ For what it's worth, the 2nd source seems to have been "softened" a bit in what looks like an updated/final version: www0.gsb.columbia.edu/faculty/mjohannes/PDFpapers/JP_2006.pdf $\endgroup$ – Matt Krause Apr 7 '16 at 22:16
  • $\begingroup$ @jaradniemi If it does not provide a Markov chain, how is it then a Markov Chain Monte Carlo (MCMC) algorithm? $\endgroup$ – arney Apr 8 '16 at 13:14
1
$\begingroup$

To my knowledge, neither the WEnKF or the KF are MCMC algorithms. I think that most of the ambiguity comes from the fact these algorithms use the markovian structure of the state space to estimate sequentially the filtering distribution $p(x_t|y_{1:t})$ using particle-based methods (so there are indeed Markov chain and Monte-Carlo inside). But in any case, the WEnKF generates any Markov chain having as stationary distribution the filtering distribution (which is the general purpose of a MCMC). If it must be categorized, an appropriate term for it could be a "particle system with mean–field interaction" but definitively not a MCMC.

$\endgroup$

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.