more air for MCMC

Aki Vehtari, Andrew Gelman, Dan Simpson, Bob Carpenter, and Paul-Christian Bürkner have just published a Bayesian Analysis paper about using an improved R factor for MCMC convergence assessment. From the early days of MCMC, convergence assessment has been a recurring (and recurrent!) question in the community. First leading to a flurry of proposals, [which Kerrie, Chantal, and myself reviewwwed in the Valencia 1998 proceedings], and then slowly disintegrating under the onslaughts of reality—i.e. that none could not be 100% foolproof in full generality—…. This included the (possibly now forgotten) single-versus-multiple-chains debate between Charlie Geyer [for single] and Andrew Gelman and Don Rubin [for multiple]. The later introduced an analysis-of-variance R factor, which remains quite popular up to this day, in part for being part of most MCMC software, like BUGS. That this R may fail to identify convergence issues, even in the more recent split version, does not come as a major surprise, since any situation with a long-term influence of the starting distribution may well fail to identify missing (significant) parts of the posterior support. (It is thus somewhat disconcerting to me to see that the main recommendation is to move the bound on R from 1.1 to 1.01, reminding me to some extent of a recent proposal to move the null rejection boundary from 0.05 to 0.005…) Similarly, the ESS may prove a poor signal for convergence or lack thereof, especially because the approximation of the asymptotic variance relies on stationarity assumptions. While multiplying the monitoring tools (as in CODA) helps with identifying convergence issues, looking at a single convergence indicator is somewhat like looking only at a frequentist estimator! (And with greater automation comes greater responsibility—in keeping a critical perspective.)

Looking for a broader perspective, I thus wonder at what we would instead need to assess the lack of convergence of an MCMC chain without much massaging of the said chain. An evaluation of the (Kullback, Wasserstein, or else) distance between the distribution of the chain at iteration n or across iterations, and the true target? A percentage of the mass of the posterior visited so far, which relates to estimating the normalising constant, with a relatively vast array of solutions made available in the recent years? I remain perplexed and frustrated by the fact that, 30 years later, the computed values of the visited likelihoods are not better exploited. Through for instance machine-learning approximations of the target. that could themselves be utilised for approximating the normalising constant and potential divergences from other approximations.

RSS honours recipients for 2020

Just read the news that my friend [and co-author] Arnaud Doucet (Oxford) is the winner of the 2020 Guy Silver Medal award from the Royal Statistical Society. I was also please to learn about David Spiegelhalter‘s Guy Gold medal (I first met David at the fourth Valencia Bayesian meeting in 1991, where he had a poster on the very early stages of BUGS) and Byron Morgan‘s Barnett Award for his indeed remarkable work on statistical ecology and in particular Bayesian capture recapture models. Congrats to all six recipients!

Hastings 50 years later

What is the exact impact of the Metropolis-Hastings algorithm on the field of Bayesian statistics? and what are the new tools of the trade? What I personally find the most relevant and attractive element in a review on the topic is the current role of this algorithm, rather than its past (his)story, since many such reviews have already appeared and will likely continue to appear. What matters most imho is how much the Metropolis-Hastings algorithm signifies for the community at large, especially beyond academia. Is the availability or unavailability of software like BUGS or Stan a help or an hindrance? Was Hastings’ paper the start of the era of approximate inference or the end of exact inference? Are the algorithm intrinsic features like Markovianity a fundamental cause for an eventual extinction because of the ensuing time constraint and the lack of practical guarantees of convergence and the illusion of a fully automated version? Or are emerging solutions like unbiased MCMC and asynchronous algorithms a beacon of hope?

In their Biometrika paper, Dunson and Johndrow (2019) recently wrote a celebration of Hastings’ 1970 paper in Biometrika, where they cover adaptive Metropolis (Haario et al., 1999; Roberts and Rosenthal, 2005), the importance of gradient based versions toward universal algorithms (Roberts and Tweedie, 1995; Neal, 2003), discussing the advantages of HMC over Langevin versions. They also recall the significant step represented by Peter Green’s (1995) reversible jump algorithm for multimodal and multidimensional targets, as well as tempering (Miasojedow et al., 2013; Woodard et al., 2009). They further cover intractable likelihood cases within MCMC (rather than ABC), with the use of auxiliary variables (Friel and Pettitt, 2008; Møller et al., 2006) and pseudo-marginal MCMC (Andrieu and Roberts, 2009; Andrieu and Vihola, 2016). They naturally insist upon the need to handle huge datasets, high-dimension parameter spaces, and other scalability issues, with links to unadjusted Langevin schemes (Bardenet et al., 2014; Durmus and Moulines, 2017; Welling and Teh, 2011). Similarly, Dunson and Johndrow (2019) discuss recent developments towards parallel MCMC and non-reversible schemes such as PDMP as highly promising, with a concluding section on the challenges of automatising and robustifying much further the said procedures, if only to reach a wider range of applications. The paper is well-written and contains a wealth of directions and reflections, including those in my above introduction. Here are some mostly disconnected directions I would have liked to see covered or more covered

1. convergence assessment today, e.g. the comparison of various approximation schemes
2. Rao-Blackwellisation and other post-processing improvements
3. other approximate inference tools than the pseudo-marginal MCMC
4. importance of the parameterisation of the problem for convergence
5. dimension issues and connection with quasi-Monte Carlo
6. constrained spaces of measure zero, as for instance matrix distributions imposing zeros outside a diagonal band
7. given the rise of the machine(-learners), are exploratory and intrinsically slow algorithms like MCMC doomed or can both fields feed one another? The section on optimisation could be expanded in that direction
8. the wasteful nature of the random walk feature of MCMC algorithms, as opposed to non-reversible kernels like HMC and other PDMPs, missing from the gradient based methods section (and can we once again learn from physicists?)
9. finer convergence issues and hence inference difficulties with complex MCMC algorithms like Gibbs samplers with incompatible conditionals
10. use of the Hastings ratio in other algorithms like ABC or EP (in link with the section on generalised Bayes)
11. adapting Metropolis-Hastings methods for emerging computing tools like GPUs and quantum computers

or possibly less covered, namely data augmentation put forward when it is a special case of auxiliary variables as in slice sampling and in earlier physics literature. For instance, both probit and logistic regressions do not truly require data augmentation and are more toy examples than really challenging applications. The approach of Carlin & Chib (1995) is another illustration, which has met with recent interest, despite requiring heavy calibration (just like RJMCMC). As well as a a somewhat awkward opposition between Gibbs and Hastings, in that I am not convinced that Gibbs does not remain ultimately necessary to handle high dimension problems, in the sense that the alternative solutions like Langevin, HMC, or PDMP, or…, are relying on Euclidean assumptions for the entire vector, while a direct product of Euclidean structures may prove more adequate.

improperties on an astronomical scale

As pointed out by Peter Coles on his blog, In the Dark, Hyungsuk Tak, Sujit Ghosh, and Justin Ellis just arXived a review of the unsafe use of improper priors in astronomy papers, 24 out of 75 having failed to establish that the corresponding posteriors are well-defined. And they exhibit such an instance (of impropriety) in a MNRAS paper by Pihajoki (2017), which is a complexification of Gelfand et al. (1990), also used by Jim Hobert in his thesis. (Even though the formal argument used to show the impropriety of the posterior in Pihajoki’s paper does not sound right since it considers divergence at a single value of a parameter β.) Besides repeating this warning about an issue that was rather quickly identified in the infancy of MCMC, if not in the very first publications on the Gibbs sampler, the paper seems to argue against using improper priors due to this potential danger, stating that instead proper priors that include all likely values and beyond are to be preferred. Which reminds me of the BUGS feature of using a N(0,10⁹) prior instead of the flat prior, missing the fact that “very large” variances do impact the resulting inference (if only for the issue of model comparison, remember Lindley-Jeffreys!). And are informative in that sense. However, it is obviously a good idea to advise checking for propriety (!) and using such alternatives may come as a safety button, providing a comparison benchmark to spot possible divergences in the resulting inference.

adaptive rejection sampling with fixed number of nodes

This paper was arXived today by Martino and Louzada. It starts from the adaptive rejection sampling algorithm of Gilks and Wild (1993), which provided an almost automatic random generator for univariate log-concave target probability densities. By creating an envelope of the true target based on convexity. This algorithm was at the core of BUGS in its early days. And makes a good example of accept reject algorithm that I often use in class. It is however not so popular nowadays because of the unidimensional constraint. And because it is not particularly adequate for simulating a single realisation from a given target. As is the case when used in a Gibbs sampler. The authors only consider here the issue of the growing cost of running the adaptive rejection sampling algorithm, which is due to the update of the envelope at each rejection. They however fail to quantify this increase, which involves (a) finding the interval where the current rejection lies, among n existing intervals, which is of order O(n), and (b) creating both modified envelopes on both new intervals, which is of order O(1). The proposal of the authors, called cheap adaptive rejection sampling, settles for a fixed number τ of support points (and intervals), solely swapping a rejected point with the closest support point when this improves the acceptance rate. The test for swapping involves first computing the alternative target (on a pair of intervals) and the corresponding normalising constant, while swapping the rejected point with the closest support point involves finding the closest point, which is of order O(log τ). I am rather surprised that the authors do not even mention the execution time orders, resorting instead to a simulation where the gain brought by their proposal is not overwhelming. There is also an additional cost for figuring out the fixed number τ of support points, another issue not mentioned in the paper.