About
I am a PhD candidate in mathematics at the Università di Pisa, where I also hold a research fellowship in the project OTODDLE (Optimal Transport and Large Deviations for Deep Learning Theory), led by Dario Trevisan and funded by the Italian Ministry of University and Research.
My work sits between probability theory and applied algebraic topology. I study how large random systems, such as wide neural networks, behave in their scaling limits, and how topological descriptors like persistent homology can be compared, computed and made stable.
Research
Quantitative limits of neural networks
I prove convergence rates for neural networks in large-width regimes. With Andrea Agazzi and Dario Trevisan, I obtained the first quantitative bound, in quadratic Wasserstein distance, on how close a trained wide network is to its Gaussian-process limit at any positive training time. Later work extends this to general architectures through tensor programs, and to deep transformers, where token trajectories converge to a stochastic interacting particle system that synchronizes when the common noise is strong enough.
Multiparameter persistence
I study distances between multiparameter persistence modules. In joint work, we showed that the matching distance is attained on a structured set of parameters described by the extended Pareto grid, a Morse-theoretic invariant, and introduced the convex matching distance, which preserves the smoothness of the input and can be used in gradient-based optimization.
Group equivariant operators
My work on multiparameter persistence fits into the theory of group equivariant non-expansive operators (GENEOs), developed as a way to build prescribed symmetries directly into topological descriptors. The operator families behind both the classical and the convex matching distance are particular cases of GENEOs.
Publications
Published
-
Quantitative convergence of trained neural networks to Gaussian processes
Advances in Neural Information Processing Systems 38 (NeurIPS 2025), 164679–164726. arXiv:2509.24544
-
Matching distance via the extended Pareto grid
SIAM Journal on Applied Algebra and Geometry 9(3), 554–576, 2025.
Preprints
-
Quantitative Gaussian-process limits of tensor programs
2026. arXiv:2607.06290
-
Stochastic scaling limits and synchronization by noise in deep transformer models
2026. arXiv:2604.26898
-
The convex matching distance in multiparameter persistence
2026. arXiv:2512.02944