Course materials for Computational Statistics, a PhD-level course at EMAp.

• We will be using the excellent materials from Professor Patrick Rebeschini (Oxford University) as a general guide for our course.

As complementary material,

• These lecture notes by stellar statistician Professor Susan Holmes are also well worth taking a look.

• Monte Carlo theory, methods and examples by Professor Art Owen, gives a nice and complete treatment of all the topics on simulation, including a whole chapter on variance reduction.

Other materials, including lecture notes and slides may be posted here as the course progresses.

Here you can find a nascent annotated bibliography with landmark papers in the field. This review paper by Professor Hedibert Lopes is far better than anything I could conjure, however.

Books marked with [a] are advanced material.

Main

• Gamerman, D., & Lopes, H. F. (2006). Markov chain Monte Carlo: stochastic simulation for Bayesian inference. Chapman and Hall/CRC.
• Robert, C. P., Casella, G. (2004). Monte Carlo Statistical Methods. John Wiley & Sons, Ltd.

Supplementary

• Givens, G. H., & Hoeting, J. A. (2012). Computational Statistics (Vol. 710). John Wiley & Sons.
• [a] Meyn, S. P., & Tweedie, R. L. (2012). Markov chains and stochastic stability. Springer Science & Business Media. [PDF].
• [a] Nummelin, E. (2004). General irreducible Markov chains and non-negative operators (Vol. 83). Cambridge University Press.

An assigment on Gibbs samplers for linear regression with heterokedasticity under conjugate priors is now available.

• Random Number Generation by Pierre L'Ecuyer;
• Non-Uniform Random Variate Generation by the great Luc Devroye;
• Walker's Alias method is a fast way to generate discrete random variables;
• Rejection Control and Sequential importance sampling (1998), by Liu et al. discusses how to improve importance sampling by controlling rejections.
• This is a nice general comment about the role of simulation in numerical integration.

• These notes from David Levin and Yuval Peres are excellent and cover a lot of material one might find interesting on Markov processes.

• Charlie Geyer's website is a treasure trove of material on Statistics in general, MCMC methods in particular. See, for instance, On the Bogosity of MCMC Diagnostics.
• Efficient construction of reversible jump Markov chain Monte Carlo proposal distributions is nice paper on the construction of efficient proposals for reversible jump/transdimensional MCMC.

The two definitive texts on HMC are Neal (2011) and Betancourt (2017). A nice set of notes is Vishnoi (2021). Moreover, Hoffman & Gelman (2014) describes the No-U-turn sampler.

This post by Radford Neal explains why the Harmonic Mean Estimator (HME) is a terrible estimator of the evidence.

• This book by Nicolas Chopin and Omiros Papaspiliopoulos is a great introduction (as it says in the title) about SMC. SMC finds application in many areas, but dynamic (linear) models deserve a special mention. The seminal 1997 book by West and Harrison remains the de facto text on the subject.

• This elementary tutorial is simple but effective.
• The book The EM algorithm and Extensions is a well-cited resource.
• Monte Carlo EM by Bob Carpenter (Columbia).

• The original 1983 paper in Science open link by Kirpatrick et al is a great read.
• These visualisations of the traveling salesman problem might prove useful.
• These notes have a little bit of theory on the cooling scheme.

• Efron (1979) is a great resource and a seminal paper.
• A good introductory book is An introduction to the bootstrap by Efron and Tibshirani (1993). PDF.
• The technical justification of the bootstrap relies on the Glivenko-Cantelli theorem. The proof given in class is taken from here.

• In these notes, Terence Tao gives insights into concentration of measure, which is the reason why integrating with respect to a probability measure in high-dimensional spaces is hard.

• A Primer for the Monte Carlo Method, by the great Ilya Sobol, is one of the first texts on the Monte Carlo method.

• The Harris inequality, E[fg] >= E[f]E[g], for f and g increasing, is a special case of the FKG inequality.

• In Markov Chain Monte Carlo Maximum Likelihood, Charlie Geyer shows how one can use MCMC to do maximum likelihood estimation when the likelihood cannot be written in closed-form. This paper is an example of MCMC methods being used outside of Bayesian statistics.

• This paper discusses the solution of Problem A in assigment 0 (2021).

Sometimes a clever way to make a target distribution easier to compute expectations with respect to is to reparametrise it. Here are some resources:

• A youtube video Introduction of the concepts and a simple example;
• Hamiltonian Monte Carlo for Hierarchical Models from M. J. Betancourt and Mark Girolami;
• A General Framework for the Parametrization of Hierarchical Models from Omiros Papaspiliopoulos, Gareth O. Roberts, and Martin Sköld;
• Efficient parametrisations for normal linear mixed models from Alan E. Gelfand, Sujit K. Sahu and Bradley P. Carlin.

See #4. Contributed by @lucasmoschen.

• Rao-Blackwellisation is a popular technique for obtaining estimators with lower variance. I recommend the recent International Statistical Review article by Christian Robert and Gareth Roberts on the topic.

• A Visualisation of MCMC for various algorithms and targets.

In these blogs and websites you will often find interesting discussions on computational, numerical and statistical aspects of applied Statistics and Mathematics.

• Christian Robert's blog;
• John Cook's website;
• Statisfaction blog.