Uploaded October 2026 | Updated November 2026, 1 day ago
Ioana Dumitriu (UC San Diego)
https://simons.berkeley.edu/talks/ioana-dumitriu-uc-san-diego-2026-10-15
Probabilistic Phenomena on Graphs, Groups, and Manifolds
Sparse random matrices are a fundamental object of study in a broad range of disciplines, from genomics to machine learning and to numerical linear algebra. The basic random model for sparse matrices is the adjacency matrix of the Erd\H{o}s-R\'enyi random graph, and the spectra and singular values of such matrices have been the subject of intensive study for many decades. Following the work of Benaych-Georges, Bordenave, and Knowles and Alt, Ducatez, and Knowles on the spectra of such matrices above ($np \gg \log n$$) and near criticality ($np \sim \log n$), we have examined the singular values and singular vectors of rectangular 0/1 sparse matrices, corresponding to bipartite Erd\{o}s-R\'enyi random graphs. The intuition is that the singular values stick to the left and right edges of the Mar\v{c}enko-Pastur bulk as long as the graph edge probability is above certain thresholds, under which outliers appear (either to the left or to the right).
In recent work, we have shown this intuition right by proving and characterizing the thresholds for outlier separation from the bulk, as a function of the aspect ratio (subject to one technical condition). We have subsequently extended these results to remove the technical condition, as well as to describe singular vector (de)localization for ranges of corresponding singular values. This is joint with Hai-Xiao Wang, Zhichao Wang, and Yizhe Zhu.
Ioana Dumitriu (UC San Diego)
https://simons.berkeley.edu/talks/ioana-dumitriu-uc-san-diego-2026-10-15
Probabilistic Phenomena on Graphs, Groups, and Manifolds
Sparse random matrices are a fundamental object of study in a broad range of disciplines, from genomics to machine learning and to numerical linear algebra. The basic random model for sparse matrices is the adjacency matrix of the Erd\H{o}s-R\'enyi random graph, and the spectra and singular values of such matrices have been the subject of intensive study for many decades. Following the work of Benaych-Georges, Bordenave, and Knowles and Alt, Ducatez, and Knowles on the spectra of such matrices above ($np \gg \log n$$) and near criticality ($np \sim \log n$), we have examined the singular values and singular vectors of rectangular 0/1 sparse matrices, corresponding to bipartite Erd\{o}s-R\'enyi random graphs. The intuition is that the singular values stick to the left and right edges of the Mar\v{c}enko-Pastur bulk as long as the graph edge probability is above certain thresholds, under which outliers appear (either to the left or to the right).
In recent work, we have shown this intuition right by proving and characterizing the thresholds for outlier separation from the bulk, as a function of the aspect ratio (subject to one technical condition). We have subsequently extended these results to remove the technical condition, as well as to describe singular vector (de)localization for ranges of corresponding singular values. This is joint with Hai-Xiao Wang, Zhichao Wang, and Yizhe Zhu.










