A Wildely Applicable Bayesian Information Criterion (WBIC)

Sumio Watanabe Homepage

WBIC approximates Bayes free energy, even if a true distribution is unrealizable by or singular for a statistical model.

If a true distribution is realizable by and regular for a statistical model, then WBIC coincides with BIC.

Application of WBIC to model selection for reduced rank regression, MATLAB file


WBIC Experiment

S. Watanabe, ``A widely applicable Bayesian information criterion," , Journal of Machine Learning Research, Vol. 14, (Mar), pp.867-897, 2013.

Misprint in the above paper.

Page. 869, Line 3 from the bottom,

Misprint : O_{p}(1)

Correct : o_{p}(1)

Note : Theorem 5 is O.K.