Skip to main content

Hirofumi Shiba

司馬博文
Ph.D. student in Statistical Science
Google Scholar arXiv Github

Institute of Statistical Mathematics (ISM), Tokyo, Japan

Feel free to call me Hiro. I'm advised by Kengo Kamatani.

I work mainly in Bayesian computation, studying how stochastic dynamics enable modern large-scale inference.

Interests

  • Computational Statistics & Machine Learning
  • Monte Carlo Methods
  • Piecewise Deterministic Markov Processes
  • Bayesian Modelling, with a particular interest in applications to the behavioural sciences

I am interested in the dynamical aspects of estimation algorithms, especially how weak or structured noise can improve their statistical performance.

Shiba dog (柴) and I (司馬) are different in kanji but equally charming.

Software

I maintain the Julia package PDMPFlux.jl, and also contribute to the R package YUIMA.

Selected Publications

Sticky Manifold Piecewise Deterministic Monte Carlo for Bayesian Model Averaging

Hirofumi Shiba, Luke Hardcastle, Kengo Kamatani. Submitted to the Journal of American Statistical Association. 2026

We design a framework for combining PDMP dynamics in different dimensions to sample from mixed-dimensional targets, in a rejection-free and non-reversible manner, without requiring additional likelihood evaluations for transitions between dimensions. Examples include Bayesian PCA with unknown rank and survival extrapolation over different parametric families.

Diffusive Scaling Limits of Forward Event-Chain Monte Carlo: Provably Efficient Exploration with Partial Refreshment

Hirofumi Shiba, Kengo Kamatani. Submitted to the Annals of Applied Probability. 2026

Many PDMP methods have been proposed. We develop a high-dimensional scaling framework to compare their asymptotic efficiency (effective sample size per event).
No matching items