This morning, Clara Grazian and I arXived a paper about Jeffreys priors for mixtures. This is a part of Clara’s PhD dissertation between Roma and Paris, on which she has worked for the past year. Jeffreys priors cannot be computed analytically for mixtures, which is such a drag that it led us to devise the delayed acceptance algorithm. However, the main message from this detailed study of Jeffreys priors is that they mostly do not work for Gaussian mixture models, in that the posterior is almost invariably improper! This is a definite death knell for Jeffreys priors in this setting, meaning that alternative reference priors, like the one we advocated with Kerrie Mengersen and Mike Titterington, or the similar solution in Roeder and Wasserman, have to be used. [Disclaimer: the title has little to do with the paper, except that posterior means are off for mixtures…]
Archive for Université Paris Dauphine
“Recently, El Moselhy et al. proposed a method to construct a map that pushed forward the prior measure to the posterior measure, casting Bayesian inference as an optimal transport problem. Namely, the constructed map transforms a random variable distributed according to the prior into another random variable distributed according to the posterior. This approach is conceptually different from previous methods, including sampling and approximation methods.”
Yesterday, Kim et al. arXived a paper with the above title, linking transport theory with Bayesian inference. Rather strangely, they motivate the transport theory with Galton’s quincunx, when the apparatus is a discrete version of the inverse cdf transform… Of course, in higher dimensions, there is no longer a straightforward transform and the paper shows (or recalls) that there exists a unique solution with positive Jacobian for log-concave posteriors. For instance, log-concave priors and likelihoods. This solution remains however a virtual notion in practice and an approximation is constructed via a (finite) functional polynomial basis. And minimising an empirical version of the Kullback-Leibler distance.
I am somewhat uncertain as to how and why apply such a transform to simulations from the prior (which thus has to be proper). Producing simulations from the posterior certainly is a traditional way to approximate Bayesian inference and this is thus one approach to this simulation. However, the discussion of the advantage of this approach over, say, MCMC, is quite limited. There is no comparison with alternative simulation or non-simulation methods and the computing time for the transport function derivation. And on the impact of the dimension of the parameter space on the computing time. In connection with recent discussions on probabilistic numerics and super-optimal convergence rates, Given that it relies on simulations, I doubt optimal transport can do better than O(√n) rates. One side remark about deriving posterior credible regions from (HPD) prior credible regions: there is no reason the resulting region is optimal in volume (HPD) given that the transform is non-linear.
Yesterday, I took part in the thesis defence of James Ridgway [soon to move to the University of Bristol[ at Université Paris-Dauphine. While I have already commented on his joint paper with Nicolas on the Pima Indians, I had not read in any depth another paper in the thesis, “On the properties of variational approximations of Gibbs posteriors” written jointly with Pierre Alquier and Nicolas Chopin.
PAC stands for probably approximately correct and starts with an empirical form of posterior, called the Gibbs posterior, where the log-likelihood is replaced with an empirical error
that is rescaled by a factor λ. Factor that is called the learning rate, to be optimised as the (Kullback) closest approximation to the true unknown distribution, by Peter Grünwald (2012) in his SafeBayes approach. In the paper of James, Pierre and Nicolas, there is no visible Bayesian perspective, since the pseudo-posterior is used to define a randomised estimator that achieves optimal oracle bounds. When λ is of order n. The purpose of the paper is rather to produce an efficient approximation to the Gibbs posterior, by using variational Bayes techniques. And to derive point estimators. With the added appeal that the approximation also achieves the oracle bounds. (Surprisingly, the authors do not leave the Pima Indians alone as they use this benchmark for a ranking model.) Since there is no discussion on the choice of the learning rate λ, as opposed to Bissiri et al. (2013) I discussed around Bayes.250, I have difficulties perceiving the possible impact of this representation on Bayesian analysis. Except maybe as an ABC device, as suggested by Christophe Andrieu.
Ingmar Schuster, who visited Paris-Dauphine last Spring (and is soon to return here as a postdoc funded by Fondation des Sciences Mathématiques de Paris) has arXived last week a paper on gradient importance sampling. In this paper, he builds a sequential importance sampling (or population Monte Carlo) algorithm that exploits the additional information contained in the gradient of the target. The proposal or importance function being essentially the MALA move as its proposal, mixed across the elements of the previous population. When compared with our original PMC mixture of random walk proposals found in e.g. this paper, each term in the mixture thus involves an extra gradient, with a scale factor that decreases to zero as 1/t√t. Ingmar compares his proposal with an adaptive Metropolis, an adaptive MALTa and an HM algorithms, for two mixture distributions and the banana target of Haario et al. (1999) we also used in our paper. As well as a logistic regression. In each case, he finds both a smaller squared error and a smaller bias for the same computing time (evaluated as the number of likelihood evaluations). While we discussed this scheme when he visited, I remain intrigued as to why it works so well when compared with the other solutions. One possible explanation is that the use of the gradient drift is more efficient on a population of particles than on a single Markov chain, provided the population covers all modes of importance on the target surface: the “fatal” attraction of the local model is then much less of an issue…
While visiting Dauphine, Natesh Pillai and Aaron Smith pointed out this interesting paper of Joris Bierkens (Warwick) that had escaped my arXiv watch/monitoring. The paper is about turning Metropolis-Hastings algorithms into non-reversible versions, towards improving mixing.
In a discrete setting, a way to produce a non-reversible move is to mix the proposal kernel Q with its time-reversed version Q’ and use an acceptance probability of the form
where ε is any weight. This construction is generalised in the paper to any vorticity (skew-symmetric with zero sum rows) matrix Γ, with the acceptance probability
where ε is small enough to ensure all numerator values are non-negative. This is a rather annoying assumption in that, except for the special case derived from the time-reversed kernel, it has to be checked over all pairs (x,y). (I first thought it also implied the normalising constant of π but everything can be set in terms of the unormalised version of π, Γ or ε included.) The paper establishes that the new acceptance probability preserves π as its stationary distribution. An alternative construction is to make the proposal change from Q in H such that H(x,y)=Q(x,y)+εΓ(x,y)/π(x). Which seems more pertinent as not changing the proposal cannot improve that much the mixing behaviour of the chain. Still, the move to the non-reversible versions has the noticeable plus of decreasing the asymptotic variance of the Monte Carlo estimate for any integrable function. Any. (Those results are found in the physics literature of the 2000’s.)
The extension to the continuous case is a wee bit more delicate. One needs to find an anti-symmetric vortex function g with zero integral [equivalent to the row sums being zero] such that g(x,y)+π(y)q(y,x)>0 and with same support as π(x)q(x,y) so that the acceptance probability of g(x,y)+π(y)q(y,x)/π(x)q(x,y) leads to π being the stationary distribution. Once again g(x,y)=ε(π(y)q(y,x)-π(x)q(x,y)) is a natural candidate but it is unclear to me why it should work. As the paper only contains one illustration for the discretised Ornstein-Uhlenbeck model, with the above choice of g for a small enough ε (a point I fail to understand since any ε<1 should provide a positive g(x,y)+π(y)q(y,x)), it is also unclear to me that this modification (i) is widely applicable and (ii) is relevant for genuine MCMC settings.