Finite mixture models do not reliably learn the number of components

When preparing my talk for Padova, I found that Diana Cai, Trevor Campbell, and Tamara Broderick wrote this ICML / PLMR paper last year on the impossible estimation of the number of components in a mixture.

“A natural check on a Bayesian mixture analysis is to establish that the Bayesian posterior on the number of components increasingly concentrates near the truth as the number of data points becomes arbitrarily large.” Cai, Campbell & Broderick (2021)

Which seems to contradict [my formerly-Glaswegian friend] Agostino Nobile  who showed in his thesis that the posterior on the number of components does concentrate at the true number of components, provided the prior contains that number in its support. As well as numerous papers on the consistency of the Bayes factor, including the one against an infinite mixture alternative, as we discussed in our recent paper with Adrien and Judith. And reminded me of the rebuke I got in 2001 from the late David McKay when mentioning that I did not believe in estimating the number of components, both because of the impact of the prior modelling and of the tendency of the data to push for more clusters as the sample size increased. (This was a most lively workshop Mike Titterington and I organised at ICMS in Edinburgh, where Radford Neal also delivered an impromptu talk to argue against using the Galaxy dataset as a benchmark!)

“In principle, the Bayes factor for the MFM versus the DPM could be used as an empirical criterion for choosing between the two models, and in fact, it is quite easy to compute an approximation to the Bayes factor using importance sampling” Miller & Harrison (2018)

This is however a point made in Miller & Harrison (2018) that the estimation of k logically goes south if the data is not from the assumed mixture model. In this paper, Cai et al. demonstrate that the posterior diverges, even when it depends on the sample size. Or even the sample as in empirical Bayes solutions.

off to Padova??? [for its 800th anniversary]

infinite mixtures are likely to take a while to simulate

Another question on X validated got me highly interested for a while, as I had considered myself the problem in the past, until I realised while discussing with Murray Pollock in Warwick that there was no general answer: when a density f is represented as an infinite series decomposition into weighted densities, some weights being negative, is there an efficient way to generate from such a density? One natural approach to the question is to look at the mixture with positive weights, f⁺, since it gives an upper bound on the target density. Simulating from this upper bound f⁺ and accepting the outcome x with probability equal to the negative part over the sum of the positive and negative parts f⁻(x)/f(x) is a valid solution. Except that it is not implementable if

1.  the positive and negative parts both involve infinite sums with no exploitable feature that can turn them into finite sums or closed form functions,
2.  the sum of the positive weights is infinite, which is the case when the series of the weights is not absolutely converging.

Even when the method is implementable it may be arbitrarily inefficient in the sense that the probability of acceptance is equal to to the inverse of the sum of the positive weights and that simulating from the bounding mixture in the regular way uses the original weights which may be unrelated in size with the actual importance of the corresponding components in the actual target. Hence, when expressed in this general form, the problem cannot allow for a generic solution.

Obviously, if more is known about the components of the mixture, as for instance the sequence of weights being alternated, there exist specialised methods, as detailed in the section of series representations in Devroye’s (1985) simulation bible. For instance, in the case when positive and negative weight densities can be paired, in the sense that their weighted difference is positive, a latent index variable can be included. But I cannot think of a generic method where the initial positive and negative components are used for simulation, as it may on the opposite be the case that no finite sum difference is everywhere positive.