Archive for invariance

optimal choice among MCMC kernels

Posted in Statistics with tags , , , , , , , , , , on March 14, 2019 by xi'an

Last week in Siem Reap, Florian Maire [who I discovered originates from a Norman town less than 10km from my hometown!] presented an arXived joint work with Pierre Vandekerkhove at the Data Science & Finance conference in Cambodia that considers the following problem: Given a large collection of MCMC kernels, how to pick the best one and how to define what best means. Going by mixtures is a default exploration of the collection, as shown in (Tierney) 1994 for instance since this improves on both kernels (esp. when each kernel is not irreducible on its own!). This paper considers a move to local weights in the mixture, weights that are not estimated from earlier simulations, contrary to what I first understood.

As made clearer in the paper the focus is on filamentary distributions that are concentrated nearby lower-dimension sets or manifolds Since then the components of the kernel collections can be restricted to directions of these manifolds… Including an interesting case of a 2-D highly peaked target where converging means mostly simulating in x¹ and covering the target means mostly simulating in x². Exhibiting a schizophrenic tension between the two goals. Weight locally dependent means correction by Metropolis step, with cost O(n). What of Rao-Blackwellisation of these mixture weights, from weight x transition to full mixture, as in our PMC paper? Unclear to me as well [during the talk] is the use in the mixture of basic Metropolis kernels, which are not absolutely continuous, because of the Dirac mass component. But this is clarified by Section 5 in the paper. A surprising result from the paper (Corollary 1) is that the use of local weights ω(i,x) that depend on the current value of the chain does jeopardize the stationary measure π(.) of the mixture chain. Which may be due to the fact that all components of the mixture are already π-invariant. Or that the index of the kernel constitutes an auxiliary (if ancillary)  variate. (Algorithm 1 in the paper reminds me of delayed acceptance. Making me wonder if computing time should be accounted for.) A final question I briefly discussed with Florian is the extension to weights that are automatically constructed from the simulations and the target.

risk-adverse Bayes estimators

Posted in Books, pictures, Statistics with tags , , , , , , , , , , on January 28, 2019 by xi'an

An interesting paper came out on arXiv in early December, written by Michael Brand from Monash. It is about risk-adverse Bayes estimators, which are defined as avoiding the use of loss functions (although why avoiding loss functions is not made very clear in the paper). Close to MAP estimates, they bypass the dependence of said MAPs on parameterisation by maximising instead π(θ|x)/√I(θ), which is invariant by reparameterisation if not by a change of dominating measure. This form of MAP estimate is called the Wallace-Freeman (1987) estimator [of which I never heard].

The formal definition of a risk-adverse estimator is still based on a loss function in order to produce a proper version of the probability to be “wrong” in a continuous environment. The difference between estimator and true value θ, as expressed by the loss, is enlarged by a scale factor k pushed to infinity. Meaning that differences not in the immediate neighbourhood of zero are not relevant. In the case of a countable parameter space, this is essentially producing the MAP estimator. In the continuous case, for “well-defined” and “well-behaved” loss functions and estimators and density, including an invariance to parameterisation as in my own intrinsic losses of old!, which the author calls likelihood-based loss function,  mentioning f-divergences, the resulting estimator(s) is a Wallace-Freeman estimator (of which there may be several). I did not get very deep into the study of the convergence proof, which seems to borrow more from real analysis à la Rudin than from functional analysis or measure theory, but keep returning to the apparent dependence of the notion on the dominating measure, which bothers me.

O’Bayes in action

Posted in Books, Kids, Statistics, University life with tags , , , , , , , , , , , , on November 7, 2017 by xi'an

My next-door colleague [at Dauphine] François Simenhaus shared a paradox [to be developed in an incoming test!] with Julien Stoehr and I last week, namely that, when selecting the largest number between a [observed] and b [unobserved], drawing a random boundary on a [meaning that a is chosen iff a is larger than this boundary] increases the probability to pick the largest number above ½2…

When thinking about it in the wretched RER train [train that got immobilised for at least two hours just a few minutes after I went through!, good luck to the passengers travelling to the airport…] to De Gaulle airport, I lost the argument: if a<b, the probability [for this random bound] to be larger than a and hence for selecting b is 1-Φ(a), while, if a>b, the probability [of winning] is Φ(a). Hence the only case when the probability is ½ is when a is the median of this random variable. But, when discussing the issue further with Julien, I exposed an interesting non-informative prior characterisation. Namely, if I assume a,b to be iid U(0,M) and set an improper prior 1/M on M, the conditional probability that b>a given a is ½. Furthermore, the posterior probability to pick the right [largest] number with François’s randomised rule is also ½, no matter what the distribution of the random boundary is. Now, the most surprising feature of this coffee room derivation is that these properties only hold for the prior 1/M. Any other power of M will induce an asymmetry between a and b. (The same properties hold when a,b are iid Exp(M).)  Of course, this is not absolutely unexpected since 1/M is the invariant prior and since the “intuitive” symmetry only holds under this prior. Power to O’Bayes!

When discussing again the matter with François yesterday, I realised I had changed his wording of the puzzle. The original setting is one with two cards hiding the unknown numbers a and b and of a player picking one of the cards. If the player picks a card at random, there is indeed a probability of ½ of picking the largest number. If the decision to switch or not depends on an independent random draw being larger or smaller than the number on the observed card, the probability to get max(a,b) in the end hits 1 when this random draw falls into (a,b) and remains ½ outside (a,b). Randomisation pays.

Pitman medal for Kerrie Mengersen

Posted in pictures, Statistics, Travel, University life with tags , , , , , , , , , , , , , on December 20, 2016 by xi'an

6831250-3x2-700x467My friend and co-author of many years, Kerrie Mengersen, just received the 2016 Pitman Medal, which is the prize of the Statistical Society of Australia. Congratulations to Kerrie for a well-deserved reward of her massive contributions to Australian, Bayesian, computational, modelling statistics, and to data science as a whole. (In case you wonder about the picture above, she has not yet lost the medal, but is instead looking for jaguars in the Amazon.)

This medal is named after EJG Pitman, Australian probabilist and statistician, whose name is attached to an estimator, a lemma, a measure of efficiency, a test, and a measure of comparison between estimators. His estimator is the best equivariant (or invariant) estimator, which can be expressed as a Bayes estimator under the relevant right Haar measure, despite having no Bayesian motivation to start with. His lemma is the Pitman-Koopman-Darmois lemma, which states that outside exponential families, sufficient is essentially useless (except for exotic distributions like the Uniform distributions). Darmois published the result first in 1935, but in French in the Comptes Rendus de l’Académie des Sciences. And the measure of comparison is Pitman nearness or closeness, on which I wrote a paper with my friends Gene Hwang and Bill Strawderman, paper that we thought was the final paper on the measure as it was pointing out several majors deficiencies with this concept. But the literature continued to grow after that..!

slice sampling revisited

Posted in Books, pictures, Statistics with tags , , , , , , , , on April 15, 2016 by xi'an

Figure 1 (c.) Neal, 2003Thanks to an X validated question, I re-read Radford Neal’s 2003 Slice sampling paper. Which is an Annals of Statistics discussion paper, and rightly so. While I was involved in the editorial processing of this massive paper (!), I had only vague memories left about it. Slice sampling has this appealing feature of being the equivalent of random walk Metropolis-Hastings for Gibbs sampling, without the drawback of setting a scale for the moves.

“These slice sampling methods can adaptively change the scale of changes made, which makes them easier to tune than Metropolis methods and also avoids problems that arise when the appropriate scale of changes varies over the distribution  (…) Slice sampling methods that improve sampling by suppressing random walks can also be constructed.” (p.706)

One major theme in the paper is fighting random walk behaviour, of which Radford is a strong proponent. Even at the present time, I am a bit surprised by this feature as component-wise slice sampling is exhibiting clear features of a random walk, exploring the subgraph of the target by random vertical and horizontal moves. Hence facing the potential drawback of backtracking to previously visited places.

“A Markov chain consisting solely of overrelaxed updates might not be ergodic.” (p.729)

Overrelaxation is presented as a mean to avoid the random walk behaviour by removing rejections. The proposal is actually deterministic projecting the current value to the “other side” of the approximate slice. If it stays within the slice it is accepted. This “reflection principle” [in that it takes the symmetric wrt the centre of the slice] is also connected with antithetic sampling in that it induces rather negative correlation between the successive simulations. The last methodological section covers reflective slice sampling, which appears as a slice version of Hamiltonian Monte Carlo (HMC). Given the difficulty in implementing exact HMC (reflected in the later literature), it is no wonder that Radford proposes an approximation scheme that is valid if somewhat involved.

“We can show invariance of this distribution by showing (…) detailed balance, which for a uniform distribution reduces to showing that the probability density for x¹ to be selected as the next state, given that the current state is x0, is the same as the probability density for x⁰ to be the next state, given that x¹ is the current state, for any states x⁰ and x¹ within [the slice] S.” (p.718)

In direct connection with the X validated question there is a whole section of the paper on implementing single-variable slice sampling that I had completely forgotten, with a collection of practical implementations when the slice

S={x; u < f(x) }

cannot be computed in an exact manner. Like the “stepping out” procedure. The resulting set (interval) where the uniform simulation in x takes place may well miss some connected component(s) of the slice. This quote may sound like a strange argument in that the move may well leave a part of the slice off and still satisfy this condition. Not really since it states that it must hold for any pair of states within S… The very positive side of this section is to allow for slice sampling in cases where the inversion of u < f(x) is intractable. Hence with a strong practical implication. The multivariate extension of the approximation procedure is more (potentially) fraught with danger in that it may fell victim to a curse of dimension, in that the box for the uniform simulation of x may be much too large when compared with the true slice (or slice of the slice). I had more of a memory of the “trail of crumbs” idea, mostly because of the name I am afraid!, which links with delayed rejection, as indicated in the paper, but seems awfully delicate to calibrate.

covariant priors, Jeffreys and paradoxes

Posted in Books, Statistics, University life with tags , , , , , , , , , , , on February 9, 2016 by xi'an

“If no information is available, π(α|M) must not deliver information about α.”

In a recent arXival apparently submitted to Bayesian Analysis, Giovanni Mana and Carlo Palmisano discuss of the choice of priors in metrology. Which reminded me of this meeting I attended at the Bureau des Poids et Mesures in Sèvres where similar debates took place, albeit being led by ferocious anti-Bayesians! Their reference prior appears to be the Jeffreys prior, because of its reparameterisation invariance.

“The relevance of the Jeffreys rule in metrology and in expressing uncertainties in measurements resides in the metric invariance.”

This, along with a second order approximation to the Kullback-Leibler divergence, is indeed one reason for advocating the use of a Jeffreys prior. I at first found it surprising that the (usually improper) prior is used in a marginal likelihood, as it cannot be normalised. A source of much debate [and of our alternative proposal].

“To make a meaningful posterior distribution and uncertainty assessment, the prior density must be covariant; that is, the prior distributions of different parameterizations must be obtained by transformations of variables. Furthermore, it is necessary that the prior densities are proper.”

The above quote is quite interesting both in that the notion of covariant is used rather than invariant or equivariant. And in that properness is indicated as a requirement. (Even more surprising is the noun associated with covariant, since it clashes with the usual notion of covariance!) They conclude that the marginal associated with an improper prior is null because the normalising constant of the prior is infinite.

“…the posterior probability of a selected model must not be null; therefore, improper priors are not allowed.”

Maybe not so surprisingly given this stance on improper priors, the authors cover a collection of “paradoxes” in their final and longest section: most of which makes little sense to me. First, they point out that the reference priors of Berger, Bernardo and Sun (2015) are not invariant, but this should not come as a surprise given that they focus on parameters of interest versus nuisance parameters. The second issue pointed out by the authors is that under Jeffreys’ prior, the posterior distribution of a given normal mean for n observations is a t with n degrees of freedom while it is a t with n-1 degrees of freedom from a frequentist perspective. This is not such a paradox since both distributions work in different spaces. Further, unless I am confused, this is one of the marginalisation paradoxes, which more straightforward explanation is that marginalisation is not meaningful for improper priors. A third paradox relates to a contingency table with a large number of cells, in that the posterior mean of a cell probability goes as the number of cells goes to infinity. (In this case, Jeffreys’ prior is proper.) Again not much of a bummer, there is simply not enough information in the data when faced with a infinite number of parameters. Paradox #4 is the Stein paradox, when estimating the squared norm of a normal mean. Jeffreys’ prior then leads to a constant bias that increases with the dimension of the vector. Definitely a bad point for Jeffreys’ prior, except that there is no Bayes estimator in such a case, the Bayes risk being infinite. Using a renormalised loss function solves the issue, rather than introducing as in the paper uniform priors on intervals, which require hyperpriors without being particularly compelling. The fifth paradox is the Neyman-Scott problem, with again the Jeffreys prior the culprit since the estimator of the variance is inconsistent. By a multiplicative factor of 2. Another stone in Jeffreys’ garden [of forking paths!]. The authors consider that the prior gives zero weight to any interval not containing zero, as if it was a proper probability distribution. And “solve” the problem by avoid zero altogether, which requires of course to specify a lower bound on the variance. And then introducing another (improper) Jeffreys prior on that bound… The last and final paradox mentioned in this paper is one of the marginalisation paradoxes, with a bizarre explanation that since the mean and variance μ and σ are not independent a posteriori, “the information delivered by x̄ should not be neglected”.

mixtures are slices of an orange

Posted in Kids, R, Statistics with tags , , , , , , , , , , , , , , , , on January 11, 2016 by xi'an

licenceDataTempering_mu_pAfter presenting this work in both London and Lenzerheide, Kaniav Kamary, Kate Lee and I arXived and submitted our paper on a new parametrisation of location-scale mixtures. Although it took a long while to finalise the paper, given that we came with the original and central idea about a year ago, I remain quite excited by this new representation of mixtures, because the use of a global location-scale (hyper-)parameter doubling as the mean-standard deviation for the mixture itself implies that all the other parameters of this mixture model [beside the weights] belong to the intersection of a unit hypersphere with an hyperplane. [Hence the title above I regretted not using for the poster at MCMskv!]fitted_density_galaxy_data_500iters2This realisation that using a (meaningful) hyperparameter (μ,σ) leads to a compact parameter space for the component parameters is important for inference in such mixture models in that the hyperparameter (μ,σ) is easily estimated from the entire sample, while the other parameters can be studied using a non-informative prior like the Uniform prior on the ensuing compact space. This non-informative prior for mixtures is something I have been seeking for many years, hence my on-going excitement! In the mid-1990‘s, we looked at a Russian doll type parametrisation with Kerrie Mengersen that used the “first” component as defining the location-scale reference for the entire mixture. And expressing each new component as a local perturbation of the previous one. While this is a similar idea than the current one, it falls short of leading to a natural non-informative prior, forcing us to devise a proper prior on the variance that was a mixture of a Uniform U(0,1) and of an inverse Uniform 1/U(0,1). Because of the lack of compactness of the parameter space. Here, fixing both mean and variance (or even just the variance) binds the mixture parameter to an ellipse conditional on the weights. A space that can be turned into the unit sphere via a natural reparameterisation. Furthermore, the intersection with the hyperplane leads to a closed form spherical reparameterisation. Yay!

While I do not wish to get into the debate about the [non-]existence of “non-informative” priors at this stage, I think being able to using the invariant reference prior π(μ,σ)=1/σ is quite neat here because the inference on the mixture parameters should be location and scale equivariant. The choice of the prior on the remaining parameters is of lesser importance, the Uniform over the compact being one example, although we did not study in depth this impact, being satisfied with the outputs produced from the default (Uniform) choice.

From a computational perspective, the new parametrisation can be easily turned into the old parametrisation, hence leads to a closed-form likelihood. This implies a Metropolis-within-Gibbs strategy can be easily implemented, as we did in the derived Ultimixt R package. (Which programming I was not involved in, solely suggesting the name Ultimixt from ultimate mixture parametrisation, a former title that we eventually dropped off for the paper.)

Discussing the paper at MCMskv was very helpful in that I got very positive feedback about the approach and superior arguments to justify the approach and its appeal. And to think about several extensions outside location scale families, if not in higher dimensions which remain a practical challenge (in the sense of designing a parametrisation of the covariance matrices in terms of the global covariance matrix).