nested sampling: any prior anytime?!

A recent arXival by Justin Alsing and Will Handley on “nested sampling with any prior you like” caught my attention. If only because I was under the impression that some priors would not agree with nested sampling. Especially those putting positive weight on some fixed levels of the likelihood function, as well as improper priors.

“…nested sampling has largely only been practical for a somewhat restrictive class of priors, which have a readily available representation as a transform from the unit hyper-cube.”

Reading from the paper, it seems that the whole point is to demonstrate that “any proper prior may be transformed onto the unit hypercube via a bijective transformation.” Which seems rather straightforward if the transform is not otherwise constrained: use a logit transform in every direction. The paper gets instead into the rather fashionable direction of normalising flows as density representations. (Which suddenly reminded me of the PhD dissertation of Rob Cornish at Oxford, which I examined last year. Even though nested was not used there in the same understanding.) The purpose appearing later (in the paper) or in fine to express a random variable simulated from the prior as the (generative) transform of a Uniform variate, f(U). Resuscitating the simulation from an arbitrary distribution from first principles.

“One particularly common scenario where this arises is when one wants to use the (sampled) posterior from one experiment as the prior for another”

But I remained uncertain at the requirement for this representation in implementing nested sampling as I do not see how it helps in bypassing the hurdles of simulating from the prior constrained by increasing levels of the likelihood function. It would be helpful to construct normalising flows adapted to the truncated priors but I did not see anything related to this version in the paper.

The cosmological application therein deals with the incorporation of recent measurements in the study of the ΛCDM cosmological model, that is, more recent that the CMB Planck dataset we played with 15 years ago. (Time flies, even if an expanding Universe!) Namely, the Baryon Oscillation Spectroscopic Survey and the SH₀ES collaboration.

information maximising neural networks summaries

After missing the blood moon eclipse last night, I had a meeting today at the Paris observatory (IAP), where we discussed an ABC proposal made by Tom Charnock, Guilhem Lavaux, and Benjamin Wandelt from this institute.

“We introduce a simulation-based machine learning technique that trains artificial neural networks to find non-linear functionals of data that maximise Fisher information : information maximising neural networks.” T. Charnock et al., 2018

The paper is centred on the determination of “optimal” summary statistics. With the goal of finding “transformation which maps the data to compressed summaries whilst conserving Fisher information [of the original data]”. Which sounds like looking for an efficient summary and hence impossible in non-exponential cases. As seen from the description in (2.1), the assumed distribution of the summary is Normal, with mean μ(θ) and covariance matrix C(θ) that are implicit transforms of the parameter θ. In that respect, the approach looks similar to the synthetic likelihood proposal of Wood (2010). From which an unusual form of Fisher information can be derived, as μ(θ)’C(θ)⁻¹μ(θ)… A neural net is trained to optimise this information criterion at a given (so-called fiducial) value of θ, in terms of a set of summaries of the same dimension as the data. Which means the information contained in the whole data (likelihood) is not necessarily recovered, linking with this comment from Edward Ionides (in a set of lectures at Wharton).

“Even summary statistics derived by careful scientific or statistical reasoning have been found surprisingly uninformative compared to the whole data likelihood in both scientific investigations (Shrestha et al., 2011) and simulation experiments (Fasiolo et al., 2016)” E. Ionides, slides, 2017

The maximal Fisher information obtained in this manner is then used in a subsequent ABC step as the natural metric for the distance between the observed and simulated data. (Begging the question as to why being maximal is necessarily optimal.) Another question is about the choice of the fiducial parameter, which choice should be tested by for instance iterating the algorithm a few steps. But having to run simulations for a single value of the parameter is certainly a great selling point!

ABC²DE

A recent arXival on a new version of ABC based on kernel estimators (but one could argue that all ABC versions are based on kernel estimators, one way or another.) In this ABC-CDE version, Izbicki,  Lee and Pospisilz [from CMU, hence the picture!] argue that past attempts failed to exploit the full advantages of kernel methods, including the 2016 ABCDE method (from Edinburgh) briefly covered on this blog. (As an aside, CDE stands for conditional density estimation.) They also criticise these attempts at selecting summary statistics and hence failing in sufficiency, which seems a non-issue to me, as already discussed numerous times on the ‘Og. One point of particular interest in the long list of drawbacks found in the paper is the inability to compare several estimates of the posterior density, since this is not directly ingrained in the Bayesian construct. Unless one moves to higher ground by calling for Bayesian non-parametrics within the ABC algorithm, a perspective which I am not aware has been pursued so far…

The selling points of ABC-CDE are that (a) the true focus is on estimating a conditional density at the observable x⁰ rather than everywhere. Hence, rejecting simulations from the reference table if the pseudo-observations are too far from x⁰ (which implies using a relevant distance and/or choosing adequate summary statistics). And then creating a conditional density estimator from this subsample (which makes me wonder at a double use of the data).

The specific density estimation approach adopted for this is called FlexCode and relates to an earlier if recent paper from Izbicki and Lee I did not read. As in many other density estimation approaches, they use an orthonormal basis (including wavelets) in low dimension to estimate the marginal of the posterior for one or a few components of the parameter θ. And noticing that the posterior marginal is a weighted average of the terms in the basis, where the weights are the posterior expectations of the functions themselves. All fine! The next step is to compare [posterior] estimators through an integrated squared error loss that does not integrate the prior or posterior and does not tell much about the quality of the approximation for Bayesian inference in my opinion. It is furthermore approximated by  a doubly integrated [over parameter and pseudo-observation] squared error loss, using the ABC(ε) sample from the prior predictive. And the approximation error only depends on the regularity of the error, that is the difference between posterior and approximated posterior. Which strikes me as odd, since the Monte Carlo error should take over but does not appear at all. I am thus unclear as to whether or not the convergence results are that relevant. (A difficulty with this paper is the strong dependence on the earlier one as it keeps referencing one version or another of FlexCode. Without reading the original one, I spotted a mention made of the use of random forests for selecting summary statistics of interest, without detailing the difference with our own ABC random forest papers (for both model selection and estimation). For instance, the remark that “nuisance statistics do not affect the performance of FlexCode-RF much” reproduces what we observed with ABC-RF.

The long experiment section always relates to the most standard rejection ABC algorithm, without accounting for the many alternatives produced in the literature (like Li and Fearnhead, 2018. that uses Beaumont et al’s 2002 scheme, along with importance sampling improvements, or ours). In the case of real cosmological data, used twice, I am uncertain of the comparison as I presume the truth is unknown. Furthermore, from having worked on similar data a dozen years ago, it is unclear why ABC is necessary in such context (although I remember us running a test about ABC in the Paris astrophysics institute once).

an interesting identity

Another interesting X validated question, another remembrance of past discussions on that issue. Discussions that took place in the Institut d’Astrophysique de Paris, nearby this painting of Laplace, when working on our cosmostats project. Namely the potential appeal of recycling multidimensional simulations by permuting the individual components in nearly independent settings. As shown by the variance decomposition in my answer, when opposing N iid pairs (X,Y) to the N combinations of √N simulations of X and √N simulations of Y, the comparison

unsurprisingly gives the upper hand to the iid sequence. A sort of converse to Rao-Blackwellisation…. Unless the production of N simulations gets much more costly when compared with the N function evaluations. No wonder we never see this proposal in Monte Carlo textbooks!