## simulation fodder for future exams

Posted in Books, Kids, R, Statistics with tags , , , , on February 20, 2019 by xi'an

Here are two nice exercises for a future simulation exam, seen and solved on X validated.The first one is about simulating a Gibbs sampler associated with the joint target

exp{-|x|-|y|-a|y-x|}

defined over IR² for a≥0 (or possibly a>-1). The conditionals are identical and non-standard, but a simple bound on the conditional density is the corresponding standard double exponential density, which makes for a straightforward accept-reject implementation. However it is also feasible to break the full conditional into three parts, depending on the respective positions of x, y, and 0, and to obtain easily invertible cdfs on the three intervals.The second exercise is about simulating from the cdf

$F(x)=1-\exp\{-ax-bx^{p+1}/(p+1)\}$

which can be numerically inverted. It is however more fun to call for an accept-reject algorithm by bounding the density with a ½ ½ mixture of an Exponential Exp(a) and of the 1/(p+1)-th power of an Exponential Exp(b/(p+1)). Since no extra constant appears in the solution,  I suspect the (p+1) in b/(p+1) was introduced on purpose. As seen in the above fit for 10⁶ simulations (and a=1,b=2,p=3), there is no deviation from the target! There is nonetheless an even simpler and rather elegant resolution to the exercise: since the tail function (1-F(x)) appears as the product of two tail functions, exp(-ax) and the other one, the cdf is the distribution of the minimum of two random variates, one with the Exp(a) distribution and the other one being the 1/(p+1)-th power of an Exponential Exp(b/(p+1)) distribution. Which of course returns a very similar histogram fit:

## ziggurat algorithm

Posted in Books, pictures, Statistics, University life with tags , , , , , , , , , , , on October 30, 2018 by xi'an

A ziggurat (Akkadian: ziqqurat, D-stem of zaqāru “to build on a raised area”) is a type of massive stone structure built in ancient Mesopotamia. It has the form of a terraced compound of successively receding stories or levels. Wikipedia

In a recent arXival, Jalalvand and Charsooghi revisit the ziggurat algorithm that simulates from a univariate distribution by finding horizontal strips that pile up on top of the target as in a ziggurat or a pyramid, hence the name. Which George Marsaglia introduced in 1963. When finely tuned the method is quite efficient. Maybe because it designs an accept-reject move for each strip of the ziggurat rather than globally. For instance, versions constructed for a Normal target are more efficient [3½ times faster] than the Box-Muller algorithm. The generalisation found in the paper divides the target into strips of equal area, rather than dominating rectangular strips of equal area, which requires some work when the target density is non-standard. For targets with unbounded support or unbounded values, a function g transforming the tail into (0,1) has to be selected. A further constraint is that the inverse cdf of the transformed g(X) has to be known. And a large part of the paper examines several scenarii towards simulating from the tail region. For unbounded densities, a similarly minute analysis is undertaken, again with requests about the target like its algebraic order.

“…the result of division of a random integer by its range is a fixed-point number which unlike a floating-point number does not enjoy increased precision near 0. When such random numbers are used in the tail algorithm they cause premature termination of the tail and large gaps between produced random numbers near the termination point.”

The paper further discusses the correction of an error common to earlier ziggurat algorithms, due to the  conversion from fixed-point to floating-point numbers, as indicated in the above quote. Although this had already been addressed by George Marsaglia in the early 1990’s.

“Ziggurat algorithm has a high setup time, so it’s not suitable for applications that require variates with frequently changing shape parameters.”

When testing the algorithm against different methods (in STL and Boost), and different distributions, the gains are between two and seven times faster, except for the Exponential target where the original ziggurat algorithm performs better. Interestingly, the gains (and the computing time) increase with the degrees of freedom for the Gamma target, in relation with Devroye’s (1986) remark on the absence of uniformly bounded execution times for this distribution. Same thing for the Weibull variates, obviously. Reflecting upon the usually costly computation of cdfs and inverse cdfs on machines and software, the inverse cdf method is systematically left behind! In conclusion, a good Sunday morning read if not of direct consequences for MCMC implementation, as warned by the authors.

## uniform on the sphere [or not]

Posted in pictures, R, Statistics with tags , , , , , , , , , , , , on March 8, 2018 by xi'an

While looking at X validated questions, I came upon this comment that simulating a uniform distribution on a d-dimensional unit sphere does not proceed from generating angles at random on (0,2π) and computing spherical coordinates… Which I must confess would have been my initial suggestion! This is obvious, nonetheless, when computing the Jacobian of the spherical coordinate transform, which involves powers of the sines of the angles, in a decreasing sequence from d-1 to zero. This means that the angles should be simulated according to their respective sine-power densities. However, except for the d=3 case, where simulating from the density sin(φ) is straightforward by inverse cdf, i.e. φ=acos(1-2u), the cdfs for the higher powers are combinations of sines and cosines, and as such are not easily inverted. Take for instance the eighth power:

F⁸(φ)=(840 φ – 672 sin(2 φ) + 168 sin(4 φ) – 32 sin(6 φ) + 3 sin(8 φ))/3072

While the densities are bounded by sin(φ), up to a constant, and hence an accept-reject can be easily derived, the efficiency decreases with the dimension according to the respective ratio of the Wallis’ integrals, unsurprisingly. A quick check for d=4 shows that the Normal simulation+projection-by-division-by-its-norm is faster.

Puzzling a bit further about this while running, I wondered at the simultaneous simulations from sin(φ), sin(φ)², sin(φ)³, &tc., but cannot see a faster way to recycle simulations from sin(φ). Points (φ,u) located in-between two adjacent power curves are acceptable simulations from the corresponding upper curve but they need be augmented by points (φ,u) under the lower curve to constitute a representative sample. In the end, this amounts to multiplying simulations from the highest power density as many times as there are powers. No gain in sight… Sigh!

However, a few days later, while enjoying the sunset over Mont Blanc(!), I figured out that there exists a direct and efficient way to simulate from these powers of the sine function. Indeed, when looking at the density of cos(φ), it happens to be the signed root of a Beta(½,(d-1)/2), which avoids the accept-reject step. Presumably this is well-known, but I have not seen this proposal associated with the uniform distribution on the sphere.

## R package truncnorm

Posted in Statistics with tags , , , , , on November 8, 2017 by xi'an

This week in Warwick, thanks to a (rather incomprehensible) X validated question, I came across the CRAN R package truncnorm, which provides the “density, distribution function, quantile function, random generation and expected value function for the truncated normal distribution”. The short description of the sampler states that the method follows the accept-reject solution of John Geweke (1991), which I reproduced [independently!] a few years later. I may have missed the right code, but checking on the Github depository associated with this package, I did not find in the C code a trace of our optimal solution via a translated exponential proposal, since the exponential proosal, when used, relies on a scale equal to the left truncation point, a in the above picture. Obviously, this does not make a major difference in the execution time (and the algorithm is still correct!).

## Why is it necessary to sample from the posterior distribution if we already KNOW the posterior distribution?

Posted in Statistics with tags , , , , , , , , on October 27, 2017 by xi'an

I found this question on X validated somewhat hilarious, the more because of the shouted KNOW! And the confused impression that because one can write down π(θ|x) up to a constant, one KNOWS this distribution… It is actually one of the paradoxes of simulation that, from a mathematical perspective, once π(θ|x) is available as a function of (θ,x), all other quantities related with this distribution are mathematically perfectly and uniquely defined. From a numerical perspective, this does not help. Actually, when starting my MCMC course at ENSAE a few days later, I had the same question from a student who thought facing a density function like

f(x) ∞ exp{-||x||²-||x||⁴-||x||⁶}

was enough to immediately produce simulations from this distribution. (I also used this example to show the degeneracy of accept-reject as the dimension d of x increases, using for instance a Gamma proposal on y=||x||. The acceptance probability plunges to zero with d, with 9 acceptances out of 10⁷ for d=20.)