estimation of deformations of densities

Today, Jean-Michel Loubes from Toulouse gave a seminar in Dauphine on the estimation of deformations using Wassertsein distances. This is functional data analysis, where samples from random transforms of the original density are observed towards estimating the baseline (or true) measure

As a neophyte, I found the problem of interest if difficult to evaluate, in particular wrt the identifiability of μ. Esp. when the distribution of the transform φ is unknown. I also wondered about the choice of means over medians, because of the added robustness of the later… In a possible connection with David Dunson’s median estimate of densities. I ran the following simulation based on 150 (centred) location-scale transforms of a normal mixture [in red] with the median of the 150 density estimators [in blue]. It is not such a poor estimate! Now, the problem itself could have implications in ABC where we have replicas of random versions of the ABC density. For instance, DIYABC produces a few copies of the ABC posteriors for the parameters of the model. Jean-Michel also mentioned  connection with transport problems.

O’Bayes 2013 [#3]

A final day for this O’Bayes 2013 conference, where I missed the final session for travelling reasons. Several talks had highly attractive features (for me), from David Dunson’s on his recently arXived paper on parallel MCMC, that provides an alternative to the embarrassingly parallel algorithm I discussed a few weeks ago, to be discussed further in a future post, to Marty Wells hindered by poor weather and delivering by phone a talk on L1 shrinkage estimators (a bit of a paradox since, as discussed by Yuzo Maruyama, most MAP estimators cannot be minimax and, more broadly, since they cannot be expressed as resolutions of loss minimisation), to Malay Ghosh revisiting g-priors from an almost frequentist viewpoint,  to Gonzalo Garci-Donato presenting criteria for objective Bayesian model choice in a vision that was clearly the closest to my own perspective on the topic. Overall, when reflecting upon the diversity and high quality of the talks at this O’Bayes meeting, and also as the incoming chair-elect of the corresponding section of ISBA, I think that what emerges most significantly from those talks is an ongoing pondering on the nature of (objective Bayesian) testing, not only in the works extending the g-priors in various directions, but also in the whole debate between Bayes factors and information criteria, model averaging versus model selection. During the discussion on Gonzalo’s talk, David Draper objected to the search for an automated approach to the comparison of models, but I strongly lean towards Gonzalo’s perspective as we need to provide a reference solution able to tackle less formal and more realistic problems. I do hope to see more of those realistic problems tackled at O’Bayes 2015 (which location is not yet settled). In the meanwhile, a strong thank you! to the local organising committee and most specifically to Jim Berger!