control functionals for Monte Carlo integration

A paper on control variates by Chris Oates, Mark Girolami (Warwick) and Nicolas Chopin (CREST) appeared in a recent issue of Series B. I had read and discussed the paper with them previously and the following is a set of comments I wrote at some stage, to be taken with enough gains of salt since Chris, Mark and Nicolas answered them either orally or in the paper. Note also that I already discussed an earlier version, with comments that are not necessarily coherent with the following ones! [Thanks to the busy softshop this week, I resorted to publish some older drafts, so mileage can vary in the coming days.]

First, it took me quite a while to get over the paper, mostly because I have never worked with reproducible kernel Hilbert spaces (RKHS) before. I looked at some proofs in the appendix and at the whole paper but could not spot anything amiss. It is obviously a major step to uncover a manageable method with a rate that is lower than √n. When I set my PhD student Anne Philippe on the approach via Riemann sums, we were quickly hindered by the dimension issue and could not find a way out. In the first versions of the nested sampling approach, John Skilling had also thought he could get higher convergence rates before realising the Monte Carlo error had not disappeared and hence was keeping the rate at the same √n speed.

The core proof in the paper leading to the 7/12 convergence rate relies on a mathematical result of Sun and Wu (2009) that a certain rate of regularisation of the function of interest leads to an average variance of order 1/6. I have no reason to mistrust the result (and anyway did not check the original paper), but I am still puzzled by the fact that it almost immediately leads to the control variate estimator having a smaller order variance (or at least variability). On average or in probability. (I am also uncertain on the possibility to interpret the boxplot figures as establishing super-√n speed.)

Another thing I cannot truly grasp is how the control functional estimator of (7) can be both a mere linear recombination of individual unbiased estimators of the target expectation and an improvement in the variance rate. I acknowledge that the coefficients of the matrices are functions of the sample simulated from the target density but still…

Another source of inner puzzlement is the choice of the kernel in the paper, which seems too simple to be able to cover all problems despite being used in every illustration there. I see the kernel as centred at zero, which means a central location must be know, decreasing to zero away from this centre, so possibly missing aspects of the integrand that are too far away, and isotonic in the reference norm, which also seems to preclude some settings where the integrand is not that compatible with the geometry.

I am equally nonplussed by the existence of a deterministic bound on the error, although it is not completely deterministic, depending on the values of the reproducible kernel at the points of the sample. Does it imply anything restrictive on the function to be integrated?

A side remark about the use of intractable in the paper is that, given the development of a whole new branch of computational statistics handling likelihoods that cannot be computed at all, intractable should possibly be reserved for such higher complexity models.

likelihood-free Bayesian inference on the minimum clinically important difference

Last week, Likelihood-free Bayesian inference on the minimum clinically important difference was arXived by Nick Syring and Ryan Martin and I read it over the weekend, slowly coming to the realisation that their [meaning of] “likelihood free” was not my [meaning of] “likelihood free”, namely that it has nothing to do with ABC! The idea therein is to create a likelihood out of a loss function, in the spirit of Bassiri, Holmes and Walker, the loss being inspired here by a clinical trial concept, the minimum clinically important difference, defined as

which defines a loss function per se when considering the empirical version. In clinical trials, Y is a binary outcome and X a vector of explanatory variables. This model-free concept avoids setting a joint distribution  on the pair (X,Y), since creating a distribution on a large vector of covariates is always an issue. As a marginalia, the authors actually mention our MCMC book in connection with a logistic regression (Example 7.11) and for a while I thought we had mentioned MCID therein, realising later it was a standard description of MCMC for logistic models.

The central and interesting part of the paper is obviously defining the likelihood-free posterior as

The authors manage to obtain the rate necessary for the estimation to be asymptotically consistent, which seems [to me] to mean that a better representation of the likelihood-free posterior should be

(even though this rescaling does not appear verbatim in the paper). This is quite an interesting application of the concept developed by Bissiri, Holmes and Walker, even though it also illustrates the difficulty of defining a specific prior, given that the minimised target above can be transformed by an arbitrary increasing function. And the mathematical difficulty in finding a rate.

control functionals for Monte Carlo integration

This new arXival by Chris Oates, Mark Girolami, and Nicolas Chopin (warning: they all are colleagues & friends of mine!, at least until they read those comments…) is a variation on control variates, but with a surprising twist namely that the inclusion of a control variate functional may produce a sub-root-n (i.e., faster than √n) convergence rate in the resulting estimator. Surprising as I did not know one could get to sub-root-n rates..! Now I had forgotten that Anne Philippe and I used the score in an earlier paper of ours, as a control variate for Riemann sum approximations, with faster convergence rates, but this is indeed a new twist, in particular because it produces an unbiased estimator.

The control variate writes

where π is the target density and φ is a free function to be optimised. (Under the constraint that πφ is integrable. Then the expectation of ψ_φ is indeed zero.) The “explanation” for the sub-root-n behaviour is that ψ_φ is chosen as an L₂ regression. When looking at the sub-root-n convergence proof, the explanation is more of a Rao-Blackwellisation type, assuming a first level convergent (or presistent) approximation to the integrand [of the above form ψ_φ can be found. The optimal φ is the solution of a differential equation that needs estimating and the paper concentrates on approximating strategies. This connects with Antonietta Mira’s zero variance control variates, but in a non-parametric manner, adopting a Gaussian process as the prior on the unknown φ. And this is where the huge innovation in the paper resides, I think, i.e. in assuming a Gaussian process prior on the control functional and in managing to preserve unbiasedness. As in many of its implementations, modelling by Gaussian processes offers nice features, like ψ_φ being itself a Gaussian process. Except that it cannot be shown to lead to presistency on a theoretical basis. Even though it appears to hold in the examples of the paper. Apart from this theoretical difficulty, the potential hardship with the method seems to be in the implementation, as there are several parameters and functionals to be calibrated, hence calling for cross-validation which may often be time-consuming. The gains are humongous, so the method should be adopted whenever the added cost in implementing it is reasonable, cost which evaluation is not clearly provided by the paper. In the toy Gaussian example where everything can be computed, I am surprised at the relatively poor performance of a Riemann sum approximation to the integral, wondering at the level of quadrature involved therein. The paper also interestingly connects with O’Hagan’s (1991) Bayes-Hermite [polynomials] quadrature and quasi-Monte Carlo [obviously!].

rate of convergence for ABC

Barber, Voss, and Webster recently posted and arXived a paper entitled The Rate of Convergence for Approximate Bayesian Computation. The paper is essentially theoretical and establishes the optimal rate of convergence of the MSE—for approximating a posterior moment—at a rate of 2/(q+4), where q is the dimension of the summary statistic, associated with an optimal tolerance in n^-1/4. I was first surprised at the role of the dimension of the summary statistic, but rationalised it as being the dimension where the non-parametric estimation takes place. I may have read the paper too quickly as I did not spot any link with earlier convergence results found in the literature: for instance, Blum (2010, JASA) links ABC with standard kernel density non-parametric estimation and find a tolerance (bandwidth) of order n^-1/q+4 and an MSE of order 2/(q+4) as well. Similarly, Biau et al. (2013, Annales de l’IHP) obtain precise convergence rates for ABC interpreted as a k-nearest-neighbour estimator. And, as already discussed at length on this blog, Fearnhead and Prangle (2012, JRSS Series B) derive rates similar to Blum’s with a tolerance of order n^-1/q+4 for the regular ABC and of order n^-1/q+2 for the noisy ABC…