I study spectral properties of random objects like graphs and matrices, and their application to statistical inference problems. The use of low-dimensional spectral embeddings has proven to be a fruitful method in many problems such as matrix completion, community detection, graph alignment, etc. I recently got into spectral convolutions in graph neural networks.
Arxiv link – published in Annales de l'IHP.
I prove an asymptotic upper bound for the second eigenvalue of the transition matrix of the simple random walk, over a random directed graph with given degree sequence. An immediate consequence of this result is a proof of the Alon conjecture for directed regular graphs. The proof is based on a variation of the trace method introduced by Bordenave (2015).
Joint work with Justin Salez.
Arxiv link – published in Annals of Probability.
We confirm the long-standing prediction that is the threshold for the emergence of a non-vanishing absolutely continuous part (extended states) at zero in the limiting spectrum of the Erdös-Renyi random graph with average degree .
We completely describe the extremal elements in the eigendecomposition of some very sparse matrices, with a new and efficient point of view regarding the problem of matrix completion in the very hard regime. We show how non-symmetric matrices can sometimes be quite beneficial in such regimes.
Joint work with Ludovic Stephan.
Arxiv link – submitted.
We prove spectral asymptotics for very sparse inhomogeneous random matrices, as well as limits for eigenvector distributions. We apply these results to clustering in sparse, directed networks and we show that the simplest method based on the eigenvectors of the adjacency matrix provably works well. We provide numerical evidence for the superiority of Gaussian mixture against Kmeans when doing the last step of the spectral clustering pipeline..
Joint work with Yizhe Zhu.
Arxiv link – Published in Random Matrix Theory and Applications.
This is a note on "bulk insider" eigenvalues for the non-backtracking spectrum of SBM. We prove their existence in the regime, which partially answers a question of Dall'Amico et al 2019. The existence is still not proved in the sparse regime (feb. 2021).
Following the recent paper of Bordenave, Chafaï and Garcia-Zelada, I show that when is a random matrix with all entries independent random variables with distribution and is fixed while , then the random polynomial converges weakly in distribution towards a random analytic function on . This function is a Poisson analog of the Gaussian Holomorphic Chaos, see Najnudel, Paquette, Simm 2020. The result is also proved when is allowed to grow to infinity with slowly. In this semi-sparse regime, the limits are more classical Gaussian objects and the statement on the eigenvalues is still valid: in particular, the second eigenvalue sticks to the bulk of the circular distribution.
Work in progress with Yizhe Zhu and Ludovic Stefan: extension for regular graphs, regular digraphs and non-backtracking matrices.
We show that scattering transforms on graphs are continuous with respect to local-weak distance: as a consequences, these graph descriptors are transferable among network models sharing the same local properties and show a remarkable degree of stability, even in very sparse graph models. From an experimental perspective, we examine how these non-learned transforms characterize graph models and graph signals through moment-constrained sampling.
Work in progress with Bartek B. and Bharatt Chowdhuri.
In parallel, I'm interested in the rigidites of random point processes, such as number-rigidities, fluctuations reductions, hyperuniformity, and the possible links between these notions. There are different ways in which point processes in can exhibit a stronger order than the totally chaotic Poisson process; hyperuniformity is when the (random) number of points falling in a large domain of radius has a reduced variance, that is, when
In this survey, I try to give a mathematical overview of this rich domain. Topics: the Fourier caracterization of hyperuniformity, the fluctuation scale, the links with number-rigidity and maximal rigidity for stealthy processes, the example of pertubed lattices.
Here is a version of this survey. It's still work in progress.
Hyperuniformity survey (june 2021: added a paragraph on JLM laws)
Many stationary point processes have recently been shown to be rigid, that is, the number of points of the process inside a disk is a measurable function of the point configuration outside the disk. However, most of these functions are limits of linear statistics of the point process and they frequently have an exponential radius of stabilization, making it nearly impossible to effectively recover the number of points in a small disk by the observation of the configuration in a large window. Can we construct more explicit reconstruction functions ? With a deep learning perspective, one can try to train invariant neural networks to get back this number and evaluate the complexity of the solutions.
Work in progress with Antoine Brochard.
Joint work with Charles Bordenave.
Arxiv link – Published in Journal of Combinatorial Theory (series B).
This is a short note on a generalization of the Erdös-Gallai theorem on graphical sequences.