non-local priors for mixtures

[For some unknown reason, this commentary on the paper by Jairo Fúquene, Mark Steel, David Rossell —all colleagues at Warwick— on choosing mixture components by non-local priors remained untouched in my draft box…]

Choosing the number of components in a mixture of (e.g., Gaussian) distributions is a hard problem. It may actually be an altogether impossible problem, even when abstaining from moral judgements on mixtures. I do realise that the components can eventually be identified as the number of observations grows to infinity, as demonstrated for instance by Judith Rousseau and Kerrie Mengersen (2011). But for a finite and given number of observations, how much can we trust any conclusion about the number of components?! It seems to me that the criticism about the vacuity of point null hypotheses, namely the logical absurdity of trying to differentiate θ=0 from any other value of θ, applies to the estimation or test on the number of components of a mixture. Doubly so, one might argue, since a very small or a very close component is undistinguishable from a non-existing one. For instance, Definition 2 is correct from a mathematical viewpoint, but it does not spell out the multiple contiguities between k and k’ component mixtures.

The paper starts with a comprehensive coverage of l’état de l’art… When using a Bayes factor to compare a k-component and an h-component mixture, the behaviour of the factor is quite different depending on which model is correct. Essentially overfitted mixtures take much longer to detect than underfitted ones, which makes intuitive sense. And BIC should be corrected for overfitted mixtures by a canonical dimension λ between the true and the (larger) assumed number of parameters  into

2 log m(y) = 2 log p(y|θ) – λ log O(n) + O(log log n)

I would argue that this purely invalidates BIG in mixture settings since the canonical dimension λ is unavailable (and DIC does not provide a useful substitute as we illustrated a decade ago…) The criticism about Rousseau and Mengersen (2011) over-fitted mixture that their approach shrinks less than a model averaging over several numbers of components relates to minimaxity and hence sounds both overly technical and reverting to some frequentist approach to testing. Replacing testing with estimating sounds like the right idea.  And I am also unconvinced that a faster rate of convergence of the posterior probability or of the Bayes factor is a relevant factor when conducting

As for non local priors, the notion seems to rely on a specific topology for the parameter space since a k-component mixture can approach a k’-component mixture (when k'<k) in a continuum of ways (even for a given parameterisation). This topology seems to be summarised by the penalty (distance?) d(θ) in the paper. Is there an intrinsic version of d(θ), given the weird parameter space? Like one derived from the Kullback-Leibler distance between the models? The choice of how zero is approached clearly has an impact on how easily the “null” is detected, the more because of the somewhat discontinuous nature of the parameter space. Incidentally, I find it curious that only the distance between means is penalised… The prior also assumes independence between component parameters and component weights, which I think is suboptimal in dealing with mixtures, maybe suboptimal in a poetic sense!, as we discussed in our reparameterisation paper. I am not sure either than the speed the distance converges to zero (in Theorem 1) helps me to understand whether the mixture has too many components for the data’s own good when I can run a calibration experiment under both assumptions.

While I appreciate the derivation of a closed form non-local prior, I wonder at the importance of the result. Is it because this leads to an easier derivation of the posterior probability? I do not see the connection in Section 3, except maybe that the importance weight indeed involves this normalising constant when considering several k’s in parallel. Is there any convergence issue in the importance sampling solution of (3.1) and (3.3) since the simulations are run under the local posterior? While I appreciate the availability of an EM version for deriving the MAP, a fact I became aware of only recently, is it truly bringing an improvement when compared with picking the MCMC simulation with the highest completed posterior?

The section on prior elicitation is obviously of central interest to me! It however seems to be restricted to the derivation of the scale factor g, in the distance, and of the parameter q in the Dirichlet prior on the weights. While the other parameters suffer from being allocated the conjugate-like priors. I would obviously enjoy seeing how this approach proceeds with our non-informative prior(s). In this regard, the illustration section is nice, but one always wonders at the representative nature of the examples and the possible interpretations of real datasets. For instance, when considering that the Old Faithful is more of an HMM than a mixture.

MDL multiple hypothesis testing

“This formulation reveals an interesting connection between multiple hypothesis testing and mixture modelling with the class labels corresponding to the accepted hypotheses in each test.”

After my seminar at Monash University last Friday, David Dowe pointed out to me the recent work by Enes Makalic and Daniel Schmidt on minimum description length (MDL) methods for multiple testing as somewhat related to our testing by mixture paper. Work which appeared in the proceedings of the 4th Workshop on Information Theoretic Methods in Science and Engineering (WITMSE-11), that took place in Helsinki, Finland, in 2011. Minimal encoding length approaches lead to choosing the model that enjoys the smallest coding length. Connected with, e.g., Rissannen‘s approach. The extension in this paper consists in considering K hypotheses at once on a collection of m datasets (the multiple then bears on the datasets rather than on the hypotheses). And to associate an hypothesis index to each dataset. When the objective function is the sum of (generalised) penalised likelihoods [as in BIC], it leads to selecting the “minimal length” model for each dataset. But the authors introduce weights or probabilities for each of the K hypotheses, which indeed then amounts to a mixture-like representation on the exponentiated codelengths. Which estimation by optimal coding was first proposed by Chris Wallace in his book. This approach eliminates the model parameters at an earlier stage, e.g. by maximum likelihood estimation, to return a quantity that only depends on the model index and the data. In fine, the purpose of the method differs from ours in that the former aims at identifying an appropriate hypothesis for each group of observations, rather than ranking those hypotheses for the entire dataset by considering the posterior distribution of the weights in the later. The mixture has somehow more of a substance in the first case, where separating the datasets into groups is part of the inference.

variational Bayes for variable selection

Xichen Huang, Jin Wang and Feng Liang have recently arXived a paper where they rely on variational Bayes in conjunction with a spike-and-slab prior modelling. This actually stems from an earlier paper by Carbonetto and Stephens (2012), the difference being in the implementation of the method, which is less Gibbs-like for the current paper. The approach is not fully Bayesian in that, not only an approximate (variational) representation is used for the parameters of interest (regression coefficient and presence-absence indicators) but also the nuisance parameters are replaced with MAPs. The variational approximation on the regression parameters is an independent product of spike-and-slab distributions. The authors show the approximate approach is consistent in both frequentist and Bayesian terms (under identifiability assumptions). The method is undoubtedly faster than MCMC since it shares many features with EM but I still wonder at the Bayesian interpretability of the outcome, which writes out as a product of estimated spike-and-slab mixtures. First, the weights in the mixtures are estimated by EM, hence fixed. Second, the fact that the variational approximation is a product is confusing in that the posterior distribution on the regression coefficients is unlikely to produce posterior independence.

at CIRM [#2]

Sylvia Richardson gave a great talk yesterday on clustering applied to variable selection, which first raised [in me] a usual worry of the lack of background model for clustering. But the way she used this notion meant there was an infinite Dirichlet process mixture model behind. This is quite novel [at least for me!] in that it addresses the covariates and not the observations themselves. I still wonder at the meaning of the cluster as, if I understood properly, the dependent variable is not involved in the clustering. Check her R package PReMiuM for a practical implementation of the approach. Later, Adeline Samson showed us the results of using pMCM versus particle Gibbs for diffusion processes where (a) pMCMC was behaving much worse than particle Gibbs and (b) EM required very few particles and Metropolis-Hastings steps to achieve convergence, when compared with posterior approximations.

Today Pierre Druilhet explained to the audience of the summer school his measure theoretic approach [I discussed a while ago] to the limit of proper priors via q-vague convergence, with the paradoxical phenomenon that a Be(n⁻¹,n⁻¹) converges to a sum of two Dirac masses when the parameter space is [0,1] but to Haldane’s prior when the space is (0,1)! He also explained why the Jeffreys-Lindley paradox vanishes when considering different measures [with an illustration that came from my Statistica Sinica 1993 paper]. Pierre concluded with the above opposition between two Bayesian paradigms, a [sort of] tale of two sigma [fields]! Not that I necessarily agree with the first paradigm that priors are supposed to have generated the actual parameter. If only because it mechanistically excludes all improper priors…

Darren Wilkinson talked about yeast, which is orders of magnitude more exciting than it sounds, because this is Bayesian big data analysis in action! With significant (and hence impressive) results based on stochastic dynamic models. And massive variable selection techniques. Scala, Haskell, Frege, OCaml were [functional] languages he mentioned that I had never heard of before! And Daniel Rudolf concluded the [intense] second day of this Bayesian week at CIRM with a description of his convergence results for (rather controlled) noisy MCMC algorithms.

maximum likelihood on negative binomial

Estimating both parameters of a negative binomial distribution NB(N,p) by maximum likelihood sounds like an obvious exercise. But it is not because some samples lead to degenerate solutions, namely p=0 and N=∞… This occurs when the mean of the sample is larger than its empirical variance, s²>x̄, not an impossible instance: I discovered this when reading a Cross Validated question asking about the action to take in such a case. A first remark of interest is that this only happens when the negative binomial distribution is defined in terms of failures (since else the number of successes is bounded). A major difference I never realised till now, for estimating N is not a straightforward exercise. A second remark is that a negative binomial NB(N,p) is a Poisson compound of an LSD variate with parameter p, the Poisson being with parameter η=-N log(1-p). And the LSD being a power distribution p^k/k rather than a psychedelic drug. Since this is not an easy framework, Adamidis (1999) introduces an extra auxiliary variable that is a truncated exponential on (0,1) with parameter -log(1-p). A very neat trick that removes the nasty normalising constant on the LSD variate.

“Convergence was achieved in all cases, even when the starting values were poor and this emphasizes the numerical stability of the EM algorithm.” K. Adamidis

Adamidis then constructs an EM algorithm on the completed set of auxiliary variables with a closed form update on both parameters. Unfortunately, the algorithm only works when s²>x̄. Otherwise, it gets stuck at the boundary p=0 and N=∞. I was hoping for a replica of the mixture case where local maxima are more interesting than the degenerate global maximum… (Of course, there is always the alternative of using a Bayesian noninformative approach.)

Bayesian filtering and smoothing [book review]

When in Warwick last October, I met Simo Särkkä, who told me he had published an IMS monograph on Bayesian filtering and smoothing the year before. I thought it would be an appropriate book to review for CHANCE and tried to get a copy from Oxford University Press, unsuccessfully. I thus bought my own book that I received two weeks ago and took the opportunity of my Czech vacations to read it… [A warning pre-empting accusations of self-plagiarism: this is a preliminary draft for a review to appear in CHANCE under my true name!]

“From the Bayesian estimation point of view both the states and the static parameters are unknown (random) parameters of the system.” (p.20)

 Bayesian filtering and smoothing is an introduction to the topic that essentially starts from ground zero. Chapter 1 motivates the use of filtering and smoothing through examples and highlights the naturally Bayesian approach to the problem(s). Two graphs illustrate the difference between filtering and smoothing by plotting for the same series of observations the successive confidence bands. The performances are obviously poorer with filtering but the fact that those intervals are point-wise rather than joint, i.e., that the graphs do not provide a confidence band. (The exercise section of that chapter is superfluous in that it suggests re-reading Kalman’s original paper and rephrases the Monty Hall paradox in a story unconnected with filtering!) Chapter 2 gives an introduction to Bayesian statistics in general, with a few pages on Bayesian computational methods. A first remark is that the above quote is both correct and mildly confusing in that the parameters can be consistently estimated, while the latent states cannot. A second remark is that justifying the MAP as associated with the 0-1 loss is incorrect in continuous settings.  The third chapter deals with the batch updating of the posterior distribution, i.e., that the posterior at time t is the prior at time t+1. With applications to state-space systems including the Kalman filter. The fourth to sixth chapters concentrate on this Kalman filter and its extension, and I find it somewhat unsatisfactory in that the collection of such filters is overwhelming for a neophyte. And no assessment of the estimation error when the model is misspecified appears at this stage. And, as usual, I find the unscented Kalman filter hard to fathom! The same feeling applies to the smoothing chapters, from Chapter 8 to Chapter 10. Which mimic the earlier ones.

how many modes in a normal mixture?

An interesting question I spotted on Cross Validated today: How to tell if a mixture of Gaussians will be multimodal? Indeed, there is no known analytical condition on the parameters of a fully specified k-component mixture for the modes to number k or less than k… Googling around, I immediately came upon this webpage by Miguel Carrera-Perpinan, who studied the issue with Chris Williams when writing his PhD in Edinburgh. And upon this paper, which not only shows that

1. unidimensional Gaussian mixtures with k components have at most k modes;
2. unidimensional non-Gaussian mixtures with k components may have more than k modes;
3. multidimensional mixtures with k components may have more than k modes.

but also provides ways of finding all the modes. Ways which seem to reduce to using EM from a wide variety of starting points (an EM algorithm set in the sampling rather than in the parameter space since all parameters are set!). Maybe starting EM from each mean would be sufficient.  I still wonder if there are better ways, from letting the variances decrease down to zero until a local mode appear, to using some sort of simulated annealing…

Edit: Following comments, let me stress this is not a statistical issue in that the parameters of the mixture are set and known and there is no observation(s) from this mixture from which to estimate the number of modes. The mathematical problem is to determine how many local maxima there are for the function