Fisher’s lost information

After a post on X validated and a good discussion at work, I came to the conclusion [after many years of sweeping the puzzle under the carpet] that the (a?) Fisher information obtained for the Uniform distribution U(0,θ) as θ⁻¹ is meaningless. Indeed, there are many arguments:

1. The lack of derivability of the indicator function for x=θ is a non-issue since the derivative is defined almost everywhere.
2. In many textbooks, the Fisher information θ⁻² is derived from the Fréchet-Darmois-Cramèr-Rao inequality, which does not apply for the Uniform U(0,θ) distribution.
3. One connected argument for the expression of the Fisher information as the expectation of the squared score is that it is the variance of the score, since its expectation is zero. Except that it is not zero for the Uniform U(0,θ) distribution.
4. For the same reason, the opposite of the second derivative of the log-likelihood is not equal to the expectation of the squared score. It is actually -θ⁻²!
5. Looking at the Taylor expansion justification of the (observed) Fisher information, expanding the log-likelihood around the maximum likelihood estimator does not work since the maximum likelihood estimator does not cancel the score.
6. When computing the Fisher information for an n-sample rather than a 1-sample, the information is n²θ⁻², rather than nθ⁻².
7. Since the speed of convergence of the maximum likelihood estimator is of order n⁻², the central limit theorem does not apply and the limiting variance of the maximum likelihood estimator is not the Fisher information.

Metropolis-Hastings importance sampling

[Warning: As I first got the paper from the authors and sent them my comments, this paper read contains their reply as well.]

In a sort of crazy coincidence, Daniel Rudolf and Björn Sprungk arXived a paper on a Metropolis-Hastings importance sampling estimator that offers similarities with  the one by Ingmar Schuster and Ilja Klebanov posted on arXiv the same day. The major difference in the construction of the importance sampler is that Rudolf and Sprungk use the conditional distribution of the proposal in the denominator of their importance weight, while Schuster and Klebanov go for the marginal (or a Rao-Blackwell representation of the marginal), mostly in an independent Metropolis-Hastings setting (for convergence) and for a discretised Langevin version in the applications. The former use a very functional L² approach to convergence (which reminded me of the early Schervish and Carlin, 1990, paper on the convergence of MCMC algorithms), not all of it necessary in my opinion. As for instance the extension of convergence properties to the augmented chain, namely (current, proposed), is rather straightforward since the proposed chain is a random transform of the current chain. An interesting remark at the end of the proof of the CLT is that the asymptotic variance of the importance sampling estimator is the same as with iid realisations from the target. This is a point we also noticed when constructing population Monte Carlo techniques (more than ten years ago), namely that dependence on the past in sequential Monte Carlo does not impact the validation and the moments of the resulting estimators, simply because “everything cancels” in importance ratios. The mean square error bound on the Monte Carlo error (Theorem 20) is not very surprising as the term ρ(y)²/P(x,y) appears naturally in the variance of importance samplers.

The first illustration where the importance sampler does worse than the initial MCMC estimator for a wide range of acceptance probabilities (Figures 2 and 3, which is which?) and I do not understand the opposite conclusion from the authors.

[Here is an answer from Daniel and Björn about this point:]

Indeed the formulation in our paper is unfortunate. The point we want to stress is that we observed in the numerical experiments certain ranges of step-sizes for which MH importance sampling shows a better performance than the classical MH algorithm with optimal scaling. Meaning that the MH importance sampling with optimal step-size can outperform MH sampling, without using additional computational resources. Surprisingly, the optimal step-size for the MH importance sampling estimator seems to remain constant for an increasing dimension in contrast to the well-known optimal scaling of the MH algorithm (given by a constant optimal acceptance rate).

The second uses the Pima Indian diabetes benchmark, amusingly (?) referring to Chopin and Ridgway (2017) who warn against the recourse to this dataset and to this model! The loss in mean square error due to the importance sampling may again be massive (Figure 5) and setting for an optimisation of the scaling factor in Metropolis-Hastings algorithms sounds unrealistic.

[And another answer from Daniel and Björn about this point:]

Indeed, Chopin and Ridgway suggest more complex problems with a larger number of covariates as benchmarks. However, the well-studied PIMA data set is a sufficient example in order to illustrate the possible benefits but also the limitations of the MH importance sampling approach. The latter are clearly (a) the required knowledge about the optimal step-size—otherwise the performance can indeed be dramatically worse than for the MH algorithm—and (b) the restriction to a small or at most moderate number of covariates. As you are indicating, optimizing the scaling factor is a challenging task. However, the hope is to derive some simple rule of thumb for the MH importance sampler similar to the well-known acceptance rate tuning for the standard MCMC estimator.

a quincunx on NBC

Through Five-Thirty-Eight, I became aware of a TV game call The Wall [so appropriate for Trumpian times!] that is essentially based on Galton’s quincunx! A huge [15m!] high version of Galton’s quincunx, with seven possible starting positions instead of one, which kills the whole point of the apparatus which is to demonstrate by simulation the proximity of the Binomial distribution to the limiting Normal (density) curve.

But the TV game has obvious no interest in the CLT, or in the Beta binomial posterior, only in a visible sequence of binary events that turn out increasing or decreasing the money “earned” by the player, the highest sums being unsurprisingly less likely. The only decision made by the player is to pick one of the seven starting points (meaning the outcome should behave like a weighted sum of seven Normals with drifted means depending on the probabilities of choosing these starting points). I found one blog entry analysing an “idiot” strategy of playing the game, but not the entire game. (Except for this entry on the older Plinko.) And Five-Thirty-Eight surprisingly does not get into the optimal strategies to play this game (maybe because there is none!). Five-Thirty-Eight also reproduces the apocryphal quote of Laplace not requiring this [God] hypothesis.

[Note: When looking for a picture of the Quincunx, I also found this desktop version! Which “allows you to visualize the order embedded in the chaos of randomness”, nothing less. And has even obtain a patent for this “visual aid that demonstrates [sic] a random walk and generates [re-sic] a bell curve distribution”…]

Monte Carlo with determinantal processes [reply from the authors]

[Rémi Bardenet and Adrien Hardy have written a reply to my comments of today on their paper, which is more readable as a post than as comments, so here it is. I appreciate the intention, as well as the perfect editing of the reply, suited for a direct posting!]

Thanks for your comments, Xian. As a foreword, a few people we met also had the intuition that DPPs would be relevant for Monte Carlo, but no result so far was backing this claim. As it turns out, we had to work hard to prove a CLT for importance-reweighted DPPs, using some deep recent results on orthogonal polynomials. We are currently working on turning this probabilistic result into practical algorithms. For instance, efficient sampling of DPPs is indeed an important open question, to which most of your comments refer. Although this question is out of the scope of our paper, note however that our results do not depend on how you sample. Efficient sampling of DPPs, along with other natural computational questions, is actually the crux of an ANR grant we just got, so hopefully in a few years we can write a more detailed answer on this blog! We now answer some of your other points.

“one has to examine the conditions for the result to operate, from the support being within the unit hypercube,”
Any compactly supported measure would do, using dilations, for instance. Note that we don’t assume the support is the whole hypercube.

“to the existence of N orthogonal polynomials wrt the dominating measure, not discussed here”
As explained in Section 2.1.2, it is enough that the reference measure charges some open set of the hypercube, which is for instance the case if it has a density with respect to the Lebesgue measure.

“to the lack of relation between the point process and the integrand,”
Actually, our method depends heavily on the target measure μ. Unlike vanilla QMC, the repulsiveness between the quadrature nodes is tailored to the integration problem.

“changing N requires a new simulation of the entire vector unless I missed the point.”
You’re absolutely right. This is a well-known open issue in probability, see the discussion on Terence Tao’s blog.

“This requires figuring out the upper bounds on the acceptance ratios, a “problem-dependent” request that may prove impossible to implement”
We agree that in general this isn’t trivial. However, good bounds are available for all Jacobi polynomials, see Section 3.

“Even without this stumbling block, generating the N-sized sample for dimension d=N (why d=N, I wonder?)”
This is a misunderstanding: we do not say that d=N in any sense. We only say that sampling from a DPP using the algorithm of [Hough et al] requires the same number of operations as orthonormalizing N vectors of dimension N, hence the cubic cost.

1. “how does it relate to quasi-Monte Carlo?”
So far, the connection to QMC is only intuitive: both rely on well-spaced nodes, but using different mathematical tools.

2. “the marginals of the N-th order determinantal process are far from uniform (see Fig. 1), and seemingly concentrated on the boundaries”
This phenomenon is due to orthogonal polynomials. We are investigating more general constructions that give more flexibility.

3. “Is the variance of the resulting estimator (2.11) always finite?”
Yes. For instance, this follows from the inequality below (5.56) since ƒ(x)/K(x,x) is Lipschitz.

4. and 5. We are investigating concentration inequalities to answer these points.

6. “probabilistic numerics produce an epistemic assessment of uncertainty, contrary to the current proposal.”
A partial answer may be our Remark 2.12. You can interpret DPPs as putting a Gaussian process prior over ƒ and sequentially sampling from the posterior variance of the GP.

Monte Carlo with determinantal processes

Rémi Bardenet and Adrien Hardy have arXived this paper a few months ago but I was a bit daunted by the sheer size of the paper, until I found the perfect opportunity last week..! The approach relates to probabilistic numerics as well as Monte Carlo, in that it can be seen as a stochastic version of Gaussian quadrature. The authors mention in the early pages a striking and recent result by Delyon and Portier that using an importance weight where the sampling density is replaced with the leave-one-out kernel estimate produces faster convergence than the regular Monte Carlo √n! Which reminds me of quasi-Monte Carlo of course, discussed in the following section (§1.3), with the interesting [and new to me] comment that the theoretical rate (and thus the improvement) does not occur until the sample size N is exponential in the dimension. Bardenet and Hardy achieve similar super-efficient convergence by mixing quadrature with repulsive simulation. For almost every integrable function.

The fact that determinantal point processes (on the unit hypercube) and Gaussian quadrature methods are connected is not that surprising once one considers that such processes are associated with densities made of determinants, which matrices are kernel-based, K(x,y), with K expressed as a sum of orthogonal polynomials. An N-th order determinantal process in dimension d satisfies a generalised Central Limit Theorem in that the speed of convergence is

which means faster than √N…  This is more surprising, of course, even though one has to examine the conditions

Paret’oothed importance sampling and infinite variance [guest post]

[Here are some comments sent to me by Aki Vehtari in the sequel of the previous posts.]

The following is mostly based on our arXived paper with Andrew Gelman and the references mentioned  there.

Koopman, Shephard, and Creal (2009) proposed to make a sample based estimate of the existence of the moments using generalized Pareto distribution fitted to the tail of the weight distribution. The number of existing moments is less than 1/k (when k>0), where k is the shape parameter of generalized Pareto distribution.

When k<1/2, the variance exists and the central limit theorem holds. Chen and Shao (2004) show further that the rate of convergence to normality is faster when higher moments exist. When 1/2≤k<1, the variance does not exist (but mean exists), the generalized central limit theorem holds, and we may assume the rate of convergence is faster when k is closer to 1/2.

In the example with “Exp(1) proposal for an Exp(1/2) target”, k=1/2 and we are truly on the border.

In our experiments in the arXived paper and in Vehtari, Gelman, and Gabry (2015), we have observed that Pareto smoothed importance sampling (PSIS) usually converges well also with k>1/2 but k close to 1/2 (let’s say k<0.7). But if k<1 and k is close to 1 (let’s say k>0.7) the convergence is much worse and both naïve importance sampling and PSIS are unreliable.

Two figures are attached, which show the results comparing IS and PSIS in the Exp(1/2) and Exp(1/10) examples. The results were computed with repeating 1000 times a simulation with 10000 samples in each. We can see the bad performance of IS in both examples as you also illustrated. In Exp(1/2) case, PSIS is also to produce much more stable results. In Exp(1/10) case, PSIS is able to reduce the variance of the estimate, but it is not enough to avoid a big bias.

It would be interesting to have more theoretical justification why infinite variance is not so big problem if k is close to 1/2 (e.g. how the convergence rate is related to the amount of fractional moments).

I guess that max ω[t] / ∑ ω[t] in Chaterjee and Diaconis has some connection to the tail shape parameter of the generalized Pareto distribution, but it is likely to be much noisier as it depends on the maximum value instead of a larger number of tail samples as in the approach by Koopman, Shephard, and Creal (2009).A third figure shows an example where the variance is finite, with “an Exp(1) proposal for an Exp(1/1.9) target”, which corresponds to k≈0.475 < 1/2. Although the variance is finite, we are close to the border and the performance of basic IS is bad. There is no sharp change in the practical behaviour with a finite number of draws when going from finite variance to infinite variance. Thus, I think it is not enough to focus on the discrete number of moments, but for example, the Pareto shape parameter k gives us more information. Koopman, Shephard, and Creal (2009) also estimated the Pareto shape k, but they formed a hypothesis test whether the variance is finite and thus discretising the information in k, and assuming that finite variance is enough to get good performance.

anti-séche