complexity of the von Neumann algorithm

Posted in Statistics with tags , , , , , , , , , on April 3, 2017 by xi'an

“Without the possibility of computing infimum and supremum of the density f over compact subintervals of the domain of f, sampling absolutely continuous distribution using the rejection method seems to be impossible in total generality.”

The von Neumann algorithm is another name for the rejection method introduced by von Neumann circa 1951. It was thus most exciting to spot a paper by Luc Devroye and Claude Gravel appearing in the latest Statistics and Computing. Assessing the method in terms of random bits and precision. Specifically, assuming that the only available random generator is one of random bits, which necessarily leads to an approximation when the target is a continuous density. The authors first propose a bisection algorithm for distributions defined on a compact interval, which compares random bits with recursive bisections of the unit interval and stops when the interval is small enough.

In higher dimension, for densities f over the unit hypercube, they recall that the original algorithm consisted in simulating uniforms x and u over the hypercube and [0,1], using the uniform as the proposal distribution and comparing the density at x, f(x), with the rescaled uniform. When using only random bits, the proposed method is based on a quadtree that subdivides the unit hypercube into smaller and smaller hypercubes until the selected hypercube is entirely above or below the density. And is small enough for the desired precision. This obviously requires for the computation of the upper and lower bound of the density over the hypercubes to be feasible, with Devroye and Gravel considering that this is a necessary property as shown by the above quote. Densities with non-compact support can be re-expressed as densities on the unit hypercube thanks to the cdf transform. (Actually, this is equivalent to the general accept-reject algorithm, based on the associated proposal.)

“With the oracles introduced in our modification of von Neumann’s method, we believe that it is impossible to design a rejection algorithm for densities that are not Riemann-integrable, so the question of the design of a universally valid rejection algorithm under the random bit model remains open.”

In conclusion, I enjoyed very much reading this paper, especially the reflection it proposes on the connection between Riemann integrability and rejection algorithms. (Actually, I cannot think straight away of a simulation algorithm that would handle non-Riemann-integrable densities, apart from nested sampling. Or of significant non-Riemann-integrable densities.)

ratio-of-uniforms [-1]

Posted in Books, pictures, R, Statistics, University life with tags , , , on December 12, 2016 by xi'an

Luca Martino pointed out to me my own and forgotten review of a 2012 paper of his, “On the Generalized Ratio of Uniforms as a Combination of Transformed Rejection and Extended Inverse of Density Sampling” that obviously discusses a generalised version of Kinderman and Monahan’s (1977) ratio-of-uniform method. And further points out the earlier 1991 paper by Jon Wakefield, Alan Gelfand and Adrian Smith that contains the general form I rediscovered a few posts ago… Called the GRoU in Martino et al.. While the generalisation in the massive arXiv document is in finding Φ such that the above region is bounded and can be explored by uniform sampling over a box.

Neither reference mentions using the cdf transform, though, which does guarantee a bounded ratio-of-uniform set in u. (An apparent contradiction with Martino et al.  statement (34), unless I am confused. Maybe due to using Φ⁻¹ instead of Φ?) But I still wonder at the usefulness of my derivations those past weeks!

Sobol’s Monte Carlo

Posted in Books, pictures, Statistics, University life with tags , , , , , , , , , , on December 10, 2016 by xi'an

The name of Ilya Sobol is familiar to researchers in quasi-Monte Carlo methods for his Sobol’s sequences. I was thus surprised to find in my office a small book entitled The Monte Carlo Method by this author, which is a translation of his 1968 book in Russian. I have no idea how it reached my office and I went to check with the library of Paris-Dauphine around the corner [of my corridor] whether it had been lost: apparently, the library got rid of it among a collection of old books… Now, having read through this 67 pages book (or booklet as Sobol puts it) makes me somewhat agree with the librarians, in that there is nothing of major relevance in this short introduction. It is quite interesting to go through the book and see the basics of simulation principles and Monte Carlo techniques unfolding, from the inverse cdf principle [established by a rather convoluted proof] to importance sampling, but the amount of information is about equivalent to the Wikipedia entry on the topic. From an historical perspective, it is also captivating to see the efforts to connect physical random generators (such as those based on vacuum tube noise) to shift-register pseudo-random generators created by Sobol in 1958. On a Soviet Strela computer.

While Googling the title of that book could not provide any connection, I found out that a 1994 version had been published under the title of A Primer for the Monte Carlo Method, which is mostly the same as my version, except for a few additional sections on pseudo-random generation, from the congruential method (with a FORTRAN code) to the accept-reject method being then called von Neumann’s instead of Neyman’s, to the notion of constructive dimension of a simulation technique, which amounts to demarginalisation, to quasi-Monte Carlo [for three pages]. A funny side note is that the author notes in the preface that the first translation [now in my office] was published without his permission!

Monty Python generator

Posted in Books, Kids, pictures, R, Statistics, University life with tags , , , , , , , on November 23, 2016 by xi'an

By some piece of luck I came across a paper by the late George Marsaglia, genial contributor to the field of simulation, and Wai Wan Tang, entitled The Monty Python method for generating random variables. As shown by the below illustration, the concept is to flip the piece H outside the rectangle back inside the rectangle, exploiting the remaining area above the density. The fantastic part is actually that “since areas of the rectangle add to 1, the slim in-between area is exactly the tail area”! So the tiny bit between G and the flipped H is the remaining tail.In the case of a Gamma Ga(a,1) variate, the authors express this variate as the transform of another variate with a nearly symmetry density, on which the Monty Python method applies. The transform is

$q(x)=(a-1/3)(1 + x/\sqrt{16a})^3$

with -√16a<x. The second nice trick is that the density of x is provided for free by the Gamma Ga(a,1) density and the transform, thanks to the change of variable formula. One lingering question is obviously how to handle the tail part. This is handled separately in the paper, with a rather involved algorithm, but since the area of the tail is tiny, a mere 1.2% in the case of the Gaussian density, this instance occurs rarely. Very clever if highly specialised! (The case of a<1 has to be processed by the indirect of multiplying a Ga(a+1,1) by a uniform variate to the power 1/a.)

I also found out that there exists a Monte Python software, which is an unrelated Monte Carlo code in python [hence the name] for cosmological inference. Including nested sampling, unsurprisingly.

ratio-of-uniforms [#2]

Posted in Books, R, Statistics with tags , , , , , , on October 31, 2016 by xi'an

Following my earlier post on Kinderman’s and Monahan’s (1977) ratio-of-uniform method, I must confess I remain quite puzzled by the approach. Or rather by its consequences. When looking at the set A of (u,v)’s in R⁺×X such that 0≤u²≤ƒ(v/u), as discussed in the previous post, it can be represented by its parameterised boundary

u(x)=√ƒ(x),v(x)=x√ƒ(x)    x in X

Similarly, since the simulation from ƒ(v/u) can also be derived [check Luc Devroye’s Non-uniform random variate generation in the exercise section 7.3] from a uniform on the set B of (u,v)’s in R⁺×X such that 0≤u≤ƒ(v+u), on the set C of (u,v)’s in R⁺×X such that 0≤u³≤ƒ(v/√u)², or on the set D of (u,v)’s in R⁺×X such that 0≤u²≤ƒ(v/u), which is actually exactly the same as A [and presumably many other versions!, for which I would like to guess the generic rule of construction], there are many sets on which one can consider running simulations. And one to pick for optimality?! Here are the three sets for a mixture of two normal densities:

For instance, assuming slice sampling is feasible on every one of those three sets, which one is the most efficient? While I have no clear answer to this question, I found on Sunday night that a generic family of transforms is indexed by a differentiable  monotone function h over the positive half-line, with the uniform distribution being taken over the set

H={(u,v);0≤u≤h(f(v/g(u))}

when the primitive G of g is the inverse of h, i.e., G(h(x))=x. [Here are the slides I gave at the Warwick reading group on Devroye’s book last week:]

GPU-accelerated Gibbs sampling

Posted in Statistics, Travel, University life with tags , , , , , , on August 18, 2016 by xi'an

Alex Terenin told me during the welcoming reception of MCqMC 2016 that he, along with Shawfeng Dong and David Draper, had arXived a paper on GPU implementation of the Gibbs sampler and thanked me profusely for my accept-reject algorithm of the truncated normal distribution. Algorithm that he reprogrammed in CUDA. The paper is mostly a review on the specifics of GPU programming and of the constraints when compared with CPUs.  The type of models considered therein allows for GPU implementation because of a very large number of latent variables that are independent conditional on the parameter θ. Like, e.g., the horseshoe probit regression model, which is how my sampler enters the picture. Accept-reject algorithms are not ideally suited for GPUs because of the while not_accepted in the code, but I did not get [from our discussion] why it is more efficient to wait for the while loop to exit when compared with running more proposals and subset the accepted ones later. Presumably because this is too costly when ensuring at least one is accepted. The paper also mentions the issue of ensuring random generators remain valid when stretched across many threads, advocating block skips as discussed in an earlier (or even ancient) ‘Og post. In line with earlier comparison tests, the proper GPU implementation of the Gibbs sampler in this setting leads to improvements that are order of magnitude faster. Nonetheless, I wonder at the universality of the comparison in that GPUs lack the programming interface that is now available for CPUs. Some authors, like the current ones, have been putting some effort in constructing random generators in CUDA, but the entry cost for newbies like me still sounds overwhelming.

automatic variational ABC

Posted in pictures, Statistics with tags , , , , , , , , , , on July 8, 2016 by xi'an

“Stochastic Variational inference is an appealing alternative to the inefficient sampling approaches commonly used in ABC.”

Moreno et al. [including Ted Meeds and Max Welling] recently arXived a paper merging variational inference and ABC. The argument for turning variational is computational speedup. The traditional (in variational inference) divergence decomposition of the log-marginal likelihood is replaced by an ABC version, parameterised in terms of intrinsic generators (i.e., generators that do not depend on cyber-parameters, like the U(0,1) or the N(0,1) generators). Or simulation code in the authors’ terms. Which leads to the automatic aspect of the approach. In the paper the derivation of the gradient is indeed automated.

“One issue is that even assuming that the ABC likelihood is an unbiased estimator of the true likelihood (which it is not), taking the log introduces a bias, so that we now have a biased estimate of the lower bound and thus biased gradients.”

I wonder how much of an issue this is, since we consider the variational lower bound. To be optimised in terms of the parameters of the variational posterior. Indeed, the endpoint of the analysis is to provide an optimal variational approximation, which remains an approximation whether or not the likelihood estimator is unbiased. A more “severe” limitation may be in the inversion constraint, since it seems to eliminate Beta or Gamma distributions. (Even though calling qbeta(runif(1),a,b) definitely is achievable… And not rejected by a Kolmogorov-Smirnov test.)

Incidentally, I discovered through the paper the existence of the Kumaraswamy distribution, which main appeal seems to be the ability to produce a closed-form quantile function, while bearing some resemblance with the Beta distribution. (Another arXival by Baltasar Trancón y Widemann studies some connections between those, but does not tell how to select the parameters to optimise the similarity.)