— Theoretical physics · Machine learning

Most strongly correlated systems are intractable. I look for the exceptions, the ones where correlations are driven by a handful of hidden variables.

Marco Biroli
Marco Biroli Research Scholar (postdoctoral) · University of Chicago, since 2025.
Fig. 1 — resetting brownian gas
FIG. 1 N non-interacting diffusers on the real line, reset together at Poisson times (vertical marks). The reset events alone — no direct interaction — couple them; the apparent independence is, by construction, false. Biroli, Larralde, Majumdar & Schehr, PRL 130, 207101 (2023).
— Lines of work

Three threads, one question.

01 — MACHINE LEARNING

The physics of learning

Where statistical mechanics meets generative modeling. Why a VAE is structurally a finite-size mean-field model, and what that costs when the data isn't — with the 2D Ising transition as a clean test case.

→ VAEs as finite-size mean-field models
02 — CORRELATED SYSTEMS

Dynamically emergent correlations

Non-interacting particles coupled only by a shared reset event — enough to correlate a gas that would otherwise never meet. Exact results for extremes, gaps, and order statistics follow from conditioning on a single hidden variable.

→ Dynamically emergent correlations in switching traps→ Experimental evidence for emergent correlations→ Extreme statistics in a resetting Brownian gas→ Exact extreme, order & sum statistics
03 — RANDOM MATRICES

Spectra under resetting

Dyson Brownian motion of eigenvalues, reset simultaneously. Stationary density, extreme-eigenvalue statistics, and the crossover between repulsion- and reset-dominated regimes.

→ Resetting Dyson Brownian motion
— Currently

What I'm working on.

A VAE is a mean-field model in disguise.

The conditional-independence assumption baked into every VAE decoder — pθ(x | z) = Πi p(xi | z) — is formally equivalent to a finite-size mean-field factorization.

The consequence: a VAE perfectly recovers Curie–Weiss systems, but structurally fails on genuinely correlated ones. Trained on 2D Ising samples, it smears out the sharp singularity at Tc ≈ 2.27. No amount of training recovers it.

— Publications

Twelve papers, reverse chronological.

2026
First-passage resetting gas
Marco Biroli, Satya N. Majumdar, Grégory Schehr
Europhysics Letters 153, 31002
arXiv DOI
2026
Dynamically emergent correlations in Brownian particles subject to simultaneous non-Poissonian resetting protocols
Gabriele de Mauro, Marco Biroli, Satya N. Majumdar, Grégory Schehr
Physical Review E 113, 014120
arXiv DOI
2025
Experimental evidence for strong emergent correlations between particles in a switching trap
Marco Biroli, Sergio Ciliberto, Manas Kulkarni, Satya N. Majumdar, Artyom Petrosyan, Grégory Schehr
arXiv preprint
arXiv
2025
Resetting Dyson Brownian motion
Marco Biroli, Satya N. Majumdar, Grégory Schehr
Physical Review E 112, 014101
arXiv DOI
2025
Strongly correlated stochastic systems
Marco Biroli
arXiv preprint (PhD thesis / review)
arXiv
2024
Exact extreme, order, and sum statistics in a class of strongly correlated systems
Marco Biroli, Hernán Larralde, Satya N. Majumdar, Grégory Schehr
Physical Review E 109, 014101
arXiv DOI
2024
Resetting by rescaling: exact results for a diffusing particle in one dimension
Marco Biroli, Yannick Feld, Alexander K. Hartmann, Satya N. Majumdar, Grégory Schehr
Physical Review E 110, 044142
arXiv DOI
2024
Dynamically emergent correlations between particles in a switching harmonic trap
Marco Biroli, Manas Kulkarni, Satya N. Majumdar, Grégory Schehr
Physical Review E 109, L032106
arXiv DOI
2023
Critical number of walkers for diffusive search processes with resetting
Marco Biroli, Satya N. Majumdar, Grégory Schehr
Physical Review E 107, 064141
arXiv DOI
2023
Extreme statistics and spacing distribution in a Brownian gas correlated by resetting
Marco Biroli, Hernán Larralde, Satya N. Majumdar, Grégory Schehr
Physical Review Letters 130, 207101
arXiv DOI
2022
Number of distinct sites visited by a resetting random walker
Marco Biroli, Francesco Mori, Satya N. Majumdar
Journal of Physics A 55, 244001
arXiv DOI
2022
Time, privacy, robustness, accuracy: trade-offs for the open vote network protocol
Fatima-Ezzahra El Orche, Rémi Géraud-Stewart, Peter B. Rønne, Gergei Bana, David Naccache, Peter YA Ryan, Marco Biroli, Megi Dervishi, Hugo Waltsburger
E-Vote-ID 2022 (LNCS 13553)
DOI

Get in touch.

I'm happy to hear from anyone working on correlated noise, exact results, or the statistical physics of learning. Email is fastest.

marco.biroli@gmail.com
Google Scholar  ·  GitHub