Outlier singular values and (de)localized singular vectors of sparse random matrices @SimonsInstitute
Outlier singular values and (de)localized singular vectors of sparse random matrices  @SimonsInstitute
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.
Outlier singular values and (de)localized singular vectors of sparse random matricesUnderstanding Language Model Hallucinations through Parametric Knowledge IntegrationUnderstanding the Structure of Reasoning in Language ModelsFunctional Stochastic LocalizationWhy Pretrained Models HallucinatePeaked quantum advantage using error correctionMatrix Concentration and Equivalent Operators on Fock SpacesRandomness Extractors and Related Objects (Part 2)Sampling from spherical spin glasses: diffusions and simulated annealingTalk by Max Hopkins (Institute for Advanced Study)Sequential Statistical Methods for Reliable Machine LearningNodal surplus and the spectral decomposition of the cycle space
Simons Institute for the Theory of Computing |

Outlier singular values and (de)localized singular vectors of sparse random matrices

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER