Manon Michel

I am a CNRS researcher (CR) at the LMBP (Laboratoire de Mathématiques Blaise Pascal - UMR 6620) at Université Clermont - Auvergne (France) in the Probability, Analysis and Statistics team.

I work at the interface between Physics, Mathematics and Computer Science. My main interest is to study and break the theoretical locks which impede the efficiency of algorithms (mainly Monte Carlo methods) in statistical Physics, Bayesian statistics and Bayesian machine learning. I approach these problems by the study of complex systems, notably around phase transitions, and by the analysis of the convergence and coupling of Markov chains. I currently focus on the implementation of irreversible Markov chains in Monte Carlo methods.

You can find a list of my publications on Google Scholar.

pic pick pick pick


Logo Susa The ANR-JCJC project SuSa I am coordinating was selected for funding! More info here !

Positions at different levels (Master, PhD, Postdoc) are available. More details here !

I am co-organizing with A. Durmus, G. Roberts and L. Zdeborova an upcoming workshop on future synergies for stochastic and learning algorithms, September 27th - October 1st 2021 at CIRM .


Recent works

Loop-Cluster Coupling and Algorithm for Classical Statistical Models

LC coupling

The q-state Potts model plays a major role in the study of critical phenomena, mostly thanks to the different representations one can use for its description, the spin, Fortuin-Kasteleyn (FK) and loop representations. Historically, a coupling between the FK and spin representations was derived and was successfully and extensively used for theoretical and numerical purposes for forty years. In the absence of any coupling with the other representations, the loop one has still to be considered independently, whereas it allows direct access to the magnetic susceptibility or correlation quantities. Following an approach based on the decomposition of random clusters into loop ones, our work now completes the construction of this missing theoretical and numerical link and ties together the three representations by the introduction of a joint-model between the FK and loop ones.
Clock Monte Carlo: From computational complexity to energy extensivity

Clock Monte Carlo

This general method allows for an efficient information extraction and a constant-time complexity by only relying on the separation of the energy into its independent components. The obtained acceleration is however limited by the disorder and frustration present in the system and we show how it is directly ruled by the energy extensivity nature, regardless of the system's peculiarities. As previous reduction methods can be seen as special examples of this new class, this work asks the question whether the complexity of energy landscapes is a strict bound for any random walk's complexity.
Forward event-chain Monte Carlo: taming randomness by symmetry exploitation

Metropolis Monte Carlo Bouncy event-chain Monte Carlo Forward event-chain Monte Carlo

Since fifty years, the conservation of the equilibrium probability distribution in Markov chain Monte Carlo was ensured by the artificial introduction of the reversibility symmetry (left). During the 2010s, non-reversible Markov chains (right), breaking the reversibility symmetry, were introduced and shown to be faster. We show how they still rely on an artificial local symmetry and how by replacing this artificial symmetry into a real one (invariance by rotation aroung the energy gradient) we obtain a dramatic acceleration (right). This generalized non-reversible MCMC is called the Forward event-chain Monte Carlo method.

Some historical simulations

2nd-order phase transition in bidimesional Ising spin systems (Onsager, 1944)


Buffon's experiment (1733)

Buffon Buffon's