Research
Below are links to some additional material:
Supplementary materials for "Unbiased Markov chain Monte Carlo with couplings" (joint work with John O'Leary, Yves F. Atchadé). Corresponds to arXiv v3.
Supplementary materials for "Inference in generative models using the Wasserstein distance" (joint work with Espen Bernton, Mathieu Gerber, Christian Robert). Corresponds to arXiv v2.
Supplementary materials for "Bayesian model comparison with the Hyvärinen score: computation and consistency" (joint work with Stephane Shao, Jie Ding, Vahid Tarokh). Corresponds to arXiv v1.
PhD thesis, "Computational contributions to Bayesian Statistics", 2012.
Erratum:
There are errors in the numerical results of "Coupling-based convergence assessment of some Gibbs samplers for high-dimensional Bayesian regression with shrinkage priors" paper. The code on https://github.com/niloyb/CoupledHalfT has been fixed.
There was an error in the L-lag coupling article, in the implementation of Langevin Monte Carlo. We didn't use the correct scaling of the step size with respect to dimension. Explanations and updated code are here: https://github.com/niloyb/LlagCouplings/tree/master/ula_versus_mala . I am grateful to Tamás Papp and Chris Sherlock for pointing this out.
There were errors in the code that produced some of the figures in "Maximal Couplings of the Metropolis-Hastings Algorithm", as pointed out by Adrien Corenflos (thanks!).
There is an error in the SMC² article. Around Equation (3) , which gives the likelihood estimator obtained by particle filters, we wrongly write that particle filters provide unbiased estimators of p(y_t|y_{1:t-1},theta) for all times t. This is not true. Instead, particle filters provide unbiased estimators of the marginal likelihood p(y_1,...y_t| theta).
There was a bug in the code of the article "Smoothing with Couplings of Conditional Particle Filters" (https://arxiv.org/abs/1701.02002). The bug was fixed and the latest version on arXiv (v3) contains the updated figures.
There was errors in the numerical results of the Unbiased HMC paper in Biometrika, which resulted in a published erratum (https://academic.oup.com/biomet/advance-article-abstract/doi/10.1093/biomet/asaa060/5875370?redirectedFrom=fulltext). The erratum reads: "The values in the fifth column of Table 1, labelled ‘Rel. ineff.’, should be multiplied by a factor of 3.105691. The penultimate sentence in § 5.3 should consequently be changed to: Our guideline for choosing k and m results in a relative inefficiency of 3.26 at an average computational cost of 3518 applications of Kε,L,σ, or approximately 5 minutes of computing time with our implementation. We are grateful to Kai Xu and Hong Ge for pointing out this transcription error."