A recent paper on the arXiv that caught my attention was this one by entitled Accelerated Parameter Estimation with DALE by Daniel & Linder (2017), in which the authors present a code for determining confidence contours on model parameters for which the likelihood is computationally expensive to evaluate. Important to note is that the authors are focused here on Frequentist confidence contours, rather than Bayesian credible intervals. Hence, their objective is to trace out an n-1 dimensional shell (or set of shells) inside the n dimensional parameter space of a given model corresponding to some specified proportion of the maximum likelihood. This is a fundamentally different objective to that of identifying credible intervals in Bayesian inference in which there is (typically) no simple prescription for the likelihood (or posterior density) that defines a credible interval given the maximum likelihood (or density at the posterior mode). Though the authors suggest an algorithm in an earlier paper that purports to produce approximate credible intervals from a Frequentist driven exploration of the parameter space it is also acknowledged therein that this is a very crude approximation. For this reason I find it rather surprising that much of this recent paper is focused on a comparison against MultiNest which is designed specifically as a Bayesian tool with completely different objectives: evaluation of the model evidence (marginal likelihood) and posterior sampling. This confusion of objectives is compounded by numerous incorrect characterisations of methods like MCMC throughout the article, such as the idea that MCMC is focused “on finding the maximum of the posterior probability”.
All that aside, what is the DALE algorithm? Well, basically it’s a guided stochastic search algorithm designed to efficiently map out a given likelihood contour through a mix of directed and random moves. The direction is through an objective function based on a distance from the target likelihood value and an objective function to maximise distance from previously sampled points. This is achieved through a Nelder-Mead procedure instead of the Gaussian process-based search strategy used in earlier papers by the same authors and collaborators, including Daniel et al. (2014) and Bryan et al. (2007). Interestingly the full author list on those papers contains Jeff Schneider who’s more recently been involved in a lot of progressive work on Bayesian optimisation with Gaussian processes, but in Bryan et al. (2007) they spend a long time dumping on Bayesian methods, both the use of MCMC for posterior sampling and the use of Bayesian credible intervals in general! Despite their dogmatic stance in that paper the authors do actually do something interesting which is to look at a non-parameteric confidence ball around the observed CMB power spectrum (based on even earlier work by Miller and Genovese). In today’s paper though the focus is simply on Neyman-Pearson generalised likelihood ratio contours.
If I was reviewing this paper I would ask for a comparison of the new code not against MultiNest but against the authors’ earlier Gaussian process-based codes. The interest then is under what types of likelihood function costs does one outperform the other. At face value one would imagine that the Gaussian process-based code is more efficient for very expensive likelihoods (say, 1 min or more) while the new version would be more efficient for very cheap likelihoods (say <0.1 sec), but their relative performance in between is more difficult to speculate and likely depends significantly on the dimensionality of the problem. If a comparison against Bayesian methods is desired then I think it would make more sense to also compare against a Bayesian method also designed for genuinely expensive likelihoods such as the Bayesian optimisation codes used for semi-analytic model fitting by Bower et al. (2010); and I would in that case set the target to the level set credible interval identified by first running the Bayesian analysis.