## generating from a failure rate function [X’ed]

Posted in Books, Kids, Statistics, University life with tags , , , , , , , , on July 4, 2015 by xi'an

While I now try to abstain from participating to the Cross Validated forum, as it proves too much of a time-consuming activity with little added value (in the sense that answers are much too often treated as disposable napkins by users who cannot be bothered to open a textbook and who usually do not exhibit any long-term impact of the provided answer, while clogging the forum with so many questions that the individual entries seem to get so little traffic, when compared say with the stackoverflow forum, to the point of making the analogy with disposable wipes more appropriate!), I came across a truly interesting question the other night. Truly interesting for me in that I had never considered the issue before.

The question is essentially wondering at how to simulate from a distribution defined by its failure rate function, which is connected with the density f of the distribution by

$\eta(t)=\frac{f(t)}{\int_t^\infty f(x)\,\text{d}x}=-\frac{\text{d}}{\text{d}t}\,\log \int_t^\infty f(x)\,\text{d}x$

From a purely probabilistic perspective, defining the distribution through f or through η is equivalent, as shown by the relation

$F(t)=1-\exp\left\{-\int_0^t\eta(x)\,\text{d}x\right\}$

but, from a simulation point of view, it may provide a different entry. Indeed, all that is needed is the ability to solve (in X) the equation

$\int_0^X\eta(x)\,\text{d}x=-\log(U)$

when U is a Uniform (0,1) variable. Which may help in that it does not require a derivation of f. Obviously, this also begs the question as to why would a distribution be defined by its failure rate function.

Posted in Books, Statistics, University life with tags , , , , , , , , , , on July 3, 2015 by xi'an

Heiko Strathmann, Dino Sejdinovic, Samuel Livingstone, Zoltán Szabó, and Arthur Gretton arXived paper about Kamiltonian MCMC generated comments from Michael Betancourt, Dan Simpson and myself, which themselves induced the following reply by Heiko, detailed enough to deserve a post of its own.

We certainly agree that the naive approach of using a non-parametric kernel density estimator on the chain history (as in [Christian’s book, Example 8.8]) as a *proposal* fails spectacularly on simple examples: the probability of proposing in unexplored regions is extremely small, independent of the current position of the MCMC trajectory. This is not what we do though. Instead, we use the gradient of a density estimator, and not the density itself, for our HMC proposal. Just like KAMH, KMC lite in fact falls back to Random Walk Metropolis in previously unexplored regions and therefore inherits geometric ergodicity properties. This in particular includes the ability to explore previously “unseen” regions, even if adaptation has stopped. I implemented a simple illustration and comparison here.

ABC example.
The main point of the ABC example, is that our method does not suffer from the additional bias from Gaussian synthetic likelihoods when being confronted with skewed models. But there is also a computational efficiency aspect. The scheme by Meeds et al. relies on finite differences and requires $2D$ simulations from the likelihood *every time* the gradient is evaluated (i.e. every leapfrog iteration) and H-ABC discards this valuable information subsequently. In contrast, KMC accumulates gradient information from simulations: it only requires to simulate from the likelihood *once* in the accept/reject step after the leapfrog integration (where gradients are available in closed form). The density is only updated then, and not during the leapfrog integration. Similar work on speeding up HMC via energy surrogates can be applied in the tall data scenario.

Approximating HMC when gradients aren’t available is in general a difficult problem. One approach (like surrogate models) may work well in some scenarios while a different approach (i.e. Monte Carlo) may work better in others, and the ABC example showcases such a case. We very much doubt that one size will fit all — but rather claim that it is of interest to find and document these scenarios.
Michael raised the concern that intractable gradients in the Pseudo-Marginal case can be avoided by running an MCMC chain on the joint space (e.g. $(f,\theta)$ for the GP classifier). To us, however, the situation is not that clear. In many cases, the correlations between variables can cause convergence problems (see e.g. here) for the MCMC and have to be addressed by de-correlation schemes (as here), or e.g. by incorporating geometric information, which also needs fixes as Michaels’s very own one. Which is the method of choice with a particular statistical problem at hand? Which method gives the smallest estimation error (if that is the goal?) for a given problem? Estimation error per time? A thorough comparison of these different classes of algorithms in terms of performance related to problem class would help here. Most papers (including ours) only show experiments favouring their own method.

GP estimator quality.
Finally, to address Michael’s point on the consistency of the GP estimator of the density gradient: this is discussed In the original paper on the infinite dimensional exponential family. As Michael points out, higher dimensional problems are unavoidably harder, however the specific details are rather involved. First, in terms of theory: both the well-specified case (when the natural parameter is in the RKHS, Section 4), and the ill-specified case (the natural parameter is in a “reasonable”, larger class of functions, Section 5), the estimate is consistent. Consistency is obtained in various metrics, including the L² error on gradients. The rates depend on how smooth the natural parameter is (and indeed a poor choice of hyper-parameter will mean slower convergence). The key point, in regards to Michael’s question, is that the smoothness requirement becomes more restrictive as the dimension increases: see Section 4.2, “range space assumption”.
Second, in terms of practice: we have found in experiments that the infinite dimensional exponential family does perform considerably better than a kernel density estimator when the dimension increases (Section 6). In other words, our density estimator can take advantage of smoothness properties of the “true” target density to get good convergence rates. As a practical strategy for hyper-parameter choice, we cross-validate, which works well empirically despite being distasteful to Bayesians. Experiments in the KMC paper also indicate that we can scale these estimators up to dimensions in the 100s on Laptop computers (unlike most other gradient estimation techniques in HMC, e.g. the ones in your HMC & sub-sampling note, or the finite differences in Meeds et al).

## Bayesian computation: a summary of the current state, and samples backwards and forwards

Posted in Books, Statistics, University life with tags , , , , , , , , on June 25, 2015 by xi'an

“The Statistics and Computing journal gratefully acknowledges the contributions for this special issue, celebrating 25 years of publication. In the past 25 years, the journal has published innovative, distinguished research by leading scholars and professionals. Papers have been read by thousands of researchers world-wide, demonstrating the global importance of this field. The Statistics and Computing journal looks forward to many more years of exciting research as the field continues to expand.” Mark Girolami, Editor in Chief for The Statistics and Computing journal

Our joint [Peter Green, Krzysztof Łatuszyński, Marcelo Pereyra, and myself] review [open access!] on the important features of Bayesian computation has already appeared in the special 25th anniversary issue of Statistics & Computing! Along with the following papers

which means very good company, indeed! And happy B’day to Statistics & Computing!

## arXiv frenzy

Posted in R, Statistics, University life with tags , , , , , , on June 23, 2015 by xi'an

In the few past days, there has been so many arXiv postings of interest—presumably the NIPS submission effect!—that I cannot hope to cover them in the coming weeks! Hopefully, some will still come out on the ‘Og in a near future:

• Scalable Approximations of Marginal Posteriors in Variable Selection by Willem van den Boom, Galen Reeves, David B. Dunson
• The MCMC split sampler: A block Gibbs sampling scheme for latent Gaussian models by Óli Páll Geirsson, Birgir Hrafnkelsson, Daniel Simpson, Helgi Sigurðarson [also deserves a special mention for gathering only ***son authors!]
• Bayesian Nonparametric Modeling of Higher Order Markov Chains by Abhra Sarkar, David B. Dunson
• Convergence of Sequential Quasi-Monte Carlo Smoothing Algorithms by Mathieu Gerber, Nicolas Chopin
• Robust Bayesian inference via coarsening by Jeffrey W. Miller, David B. Dunson
• Expectation Particle Belief Propagation by Thibaut Lienart, Yee Whye Teh, Arnaud Doucet
• arXiv:1506.05860: Variational Gaussian Copula Inference by Shaobo Han, Xuejun Liao, David B. Dunson, Lawrence Carin
• arXiv:1506.05855: The Frequentist Information Criterion (FIC): The unification of information-based and frequentist inference by Colin H. LaMont, Paul A. Wiggins
• arXiv:1506.05757: Bayesian Inference for the Multivariate Extended-Skew Normal Distribution by Mathieu Gerber, Florian Pelgrin
• arXiv:1506.05741: Accelerated dimension-independent adaptive Metropolis by Yuxin Chen, David Keyes, Kody J.H. Law, Hatem Ltaief
• arXiv:1506.05269: Bayesian Survival Model based on Moment Characterization by Julyan Arbel, Antonio Lijoi, Bernardo Nipoti
• arXiv:1506.04778: Fast sampling with Gaussian scale-mixture priors in high-dimensional regression by Anirban Bhattacharya, Antik Chakraborty, Bani K. Mallick
• arXiv:1506.04416: Bayesian Dark Knowledge by Anoop Korattikara, Vivek Rathod, Kevin Murphy, Max Welling [a special mention for this title!]
• arXiv:1506.03693: Optimization Monte Carlo: Efficient and Embarrassingly Parallel Likelihood-Free Inference by Edward Meeds, Max Welling
• arXiv:1506.03074: Variational consensus Monte Carlo by Maxim Rabinovich, Elaine Angelino, Michael I. Jordan
• arXiv:1506.02564: Gradient-free Hamiltonian Monte Carlo with Efficient Kernel Exponential Families by Heiko Strathmann, Dino Sejdinovic, Samuel Livingstone, Zoltan Szabo, Arthur Gretton [comments coming soon!]

## simulating correlated random variables [cont’ed]

Posted in Books, Kids, Statistics with tags , , , , on May 28, 2015 by xi'an

Following a recent post on the topic, and comments ‘Og’s readers kindly provided on that post, the picture is not as clear as I wished it was… Indeed, on the one hand, non-parametric measures of correlation based on ranks are, as pointed out by Clara Grazian and others, invariant under monotonic transforms and hence producing a Gaussian pair or a Uniform pair with the intended rank correlation is sufficient to return a correlated sample for any pair of marginal distributions by the (monotonic) inverse cdf transform.  On the other hand, if correlation is understood as Pearson linear correlation, (a) it is not always defined and (b) there does not seem to be a generic approach to simulate from an arbitrary triplet (F,G,ρ) [assuming the three entries are compatible]. When Kees pointed out Pascal van Kooten‘s solution by permutation, I thought this was a terrific resolution, but after thinking about it a wee bit more, I am afraid it is only an approximation, i.e., a way to return a bivariate sample with a given empirical correlation. Not the theoretical correlation. Obviously, when the sample is very large, this comes as a good approximation. But when facing a request to simulate a single pair (X,Y), this gets inefficient [and still approximate].

Now, if we aim at exact simulation from a bivariate distribution with the arbitrary triplet (F,G,ρ), why can’t we find a generic method?! I think one fundamental if obvious reason is that the question is just ill-posed. Indeed, there are many ways of defining a joint distribution with marginals F and G and with (linear) correlation ρ. One for each copula. The joint could thus be associated with a Gaussian copula, i.e., (X,Y)=(F⁻¹(Φ(A)),G⁻¹(Φ(B))) when (A,B) is a standardised bivariate normal with the proper correlation ρ’. Or it can be associated with the Archimedian copula

C(u; v) = (u + v − 1)-1/θ,

with θ>0 defined by a (linear) correlation of ρ. Or yet with any other copula… Were the joint distribution perfectly well-defined, it would then mean that ρ’ or θ (or whatever natural parameter is used for that copula) do perfectly parametrise this distribution instead of the correlation coefficient ρ. All that remains then is to simulate directly from the copula, maybe a theme for a future post…

## non-reversible MCMC

Posted in Books, Statistics, University life with tags , , , , , , on May 21, 2015 by xi'an

While visiting Dauphine, Natesh Pillai and Aaron Smith pointed out this interesting paper of Joris Bierkens (Warwick) that had escaped my arXiv watch/monitoring. The paper is about turning Metropolis-Hastings algorithms into non-reversible versions, towards improving mixing.

In a discrete setting, a way to produce a non-reversible move is to mix the proposal kernel Q with its time-reversed version Q’ and use an acceptance probability of the form

$\epsilon\pi(y)Q(y,x)+(1-\epsilon)\pi(x)Q(x,y) \big/ \pi(x)Q(x,y)$

where ε is any weight. This construction is generalised in the paper to any vorticity (skew-symmetric with zero sum rows) matrix Γ, with the acceptance probability

$\epsilon\Gamma(x,y)+\pi(y)Q(y,x)\big/\pi(x)Q(x,y)$

where ε is small enough to ensure all numerator values are non-negative. This is a rather annoying assumption in that, except for the special case derived from the time-reversed kernel, it has to be checked over all pairs (x,y). (I first thought it also implied the normalising constant of π but everything can be set in terms of the unormalised version of π, Γ or ε included.) The paper establishes that the new acceptance probability preserves π as its stationary distribution. An alternative construction is to make the proposal change from Q in H such that H(x,y)=Q(x,y)+εΓ(x,y)/π(x). Which seems more pertinent as not changing the proposal cannot improve that much the mixing behaviour of the chain. Still, the move to the non-reversible versions has the noticeable plus of decreasing the asymptotic variance of the Monte Carlo estimate for any integrable function. Any. (Those results are found in the physics literature of the 2000’s.)

The extension to the continuous case is a wee bit more delicate. One needs to find an anti-symmetric vortex function g with zero integral [equivalent to the row sums being zero] such that g(x,y)+π(y)q(y,x)>0 and with same support as π(x)q(x,y) so that the acceptance probability of g(x,y)+π(y)q(y,x)/π(x)q(x,y) leads to π being the stationary distribution. Once again g(x,y)=ε(π(y)q(y,x)-π(x)q(x,y)) is a natural candidate but it is unclear to me why it should work. As the paper only contains one illustration for the discretised Ornstein-Uhlenbeck model, with the above choice of g for a small enough ε (a point I fail to understand since any ε<1 should provide a positive g(x,y)+π(y)q(y,x)), it is also unclear to me that this modification (i) is widely applicable and (ii) is relevant for genuine MCMC settings.

## Cauchy Distribution: Evil or Angel?

Posted in Books, pictures, Running, Statistics, Travel, University life, Wines with tags , , , , , , , , , , , , on May 19, 2015 by xi'an

Natesh Pillai and Xiao-Li Meng just arXived a short paper that solves the Cauchy conjecture of Drton and Xiao [I mentioned last year at JSM], namely that, when considering two normal vectors with generic variance matrix S, a weighted average of the ratios X/Y remains Cauchy(0,1), just as in the iid S=I case. Even when the weights are random. The fascinating side of this now resolved (!) conjecture is that the correlation between the terms does not seem to matter. Pushing the correlation to one [assuming it is meaningful, which is a suspension of belief!, since there is no standard correlation for Cauchy variates] leads to a paradox: all terms are equal and yet… it works: we recover a single term, which again is Cauchy(0,1). All that remains thus to prove is that it stays Cauchy(0,1) between those two extremes, a weird kind of intermediary values theorem!

Actually, Natesh and XL further prove an inverse χ² theorem: the inverse of the normal vector, renormalised into a quadratic form is an inverse χ² no matter what its covariance matrix. The proof of this amazing theorem relies on a spherical representation of the bivariate Gaussian (also underlying the Box-Müller algorithm). The angles are then jointly distributed as

$\exp\{-\sum_{i,j}\alpha_{ij}\cos(\theta_i-\theta_j)\}$

and from there follows the argument that conditional on the differences between the θ’s, all ratios are Cauchy distributed. Hence the conclusion!

A question that stems from reading this version of the paper is whether this property extends to other formats of non-independent Cauchy variates. Somewhat connected to my recent post about generating correlated variates from arbitrary distributions: using the inverse cdf transform of a Gaussian copula shows this is possibly the case: the following code is meaningless in that the empirical correlation has no connection with a “true” correlation, but nonetheless the experiment seems of interest…

> ro=.999999;x=matrix(rnorm(2e4),ncol=2);y=ro*x+sqrt(1-ro^2)*matrix(rnorm(2e4),ncol=2)
> cor(x[,1]/x[,2],y[,1]/y[,2])
[1] -0.1351967
> ro=.99999999;x=matrix(rnorm(2e4),ncol=2);y=ro*x+sqrt(1-ro^2)*matrix(rnorm(2e4),ncol=2)
> cor(x[,1]/x[,2],y[,1]/y[,2])
[1] 0.8622714
> ro=1-1e-5;x=matrix(rnorm(2e4),ncol=2);y=ro*x+sqrt(1-ro^2)*matrix(rnorm(2e4),ncol=2)
> z=qcauchy(pnorm(as.vector(x)));w=qcauchy(pnorm(as.vector(y)))
> cor(x=z,y=w)
[1] 0.9999732
> ks.test((z+w)/2,"pcauchy")

One-sample Kolmogorov-Smirnov test

data:  (z + w)/2
D = 0.0068, p-value = 0.3203
alternative hypothesis: two-sided
> ro=1-1e-3;x=matrix(rnorm(2e4),ncol=2);y=ro*x+sqrt(1-ro^2)*matrix(rnorm(2e4),ncol=2)
> z=qcauchy(pnorm(as.vector(x)));w=qcauchy(pnorm(as.vector(y)))
> cor(x=z,y=w)
[1] 0.9920858
> ks.test((z+w)/2,"pcauchy")

One-sample Kolmogorov-Smirnov test

data:  (z + w)/2
D = 0.0036, p-value = 0.9574
alternative hypothesis: two-sided