ICM 2026 Invited Section Lecture - Richard Samworth

Value of the Information & Strength of the Argument

The lecture offers a comprehensive synthesis of shape-constrained inference, providing both historical context and cutting-edge results. The argumentation is rigorous and well-structured, with clear definitions, precise statements of theorems, and even a complete proof of Marshall’s lemma (simple and elegant). The value lies in unifying diverse shape constraints (monotonicity, convexity, log-concavity) under a coherent framework, highlighting the trade-offs between flexibility and adaptivity. The presenter successfully explains why log-concavity emerges as a natural infinite-dimensional generalization of Gaussian distributions, and supports this with theoretical results on projection properties and convergence rates. The discussion is technical but accessible to a mathematically trained audience, and the presentation is authoritative, reflecting deep expertise.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high: the lecture cites key results with proper attribution (Grenander, Marshall, etc.) and presents proofs for central statements. However, no external links or specific references are provided in the video description, limiting the ability to verify sources directly. The title accurately matches the content, which covers past developments (Grenander), present state (log-concave MLE), and future directions (open problems). The video is produced by the Simons Foundation, a reputable institution, and the lecturer is a recognized statistician, contributing to reliability. No user comments were available for analysis.

The title accurately reflects the content: a lecture on nonparametric inference under shape constraints, covering past, present and future perspectives.

9/10

Lecture by a leading statistician at an ICM invited session, presenting established results and recent advances with rigorous definitions and proofs. High level of expertise.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:30Introduction: the core question of density estimation and the two broad categories of methods.
• 04:30Kernel density estimation: role of bandwidth and bias-variance trade-off.
• 09:30Introduction to shape constraints and the Grenander estimator for decreasing densities.
• 12:30Characterization of the Grenander estimator as the left derivative of the least concave majorant.
• 15:30Marshall's lemma: proof that the Grenander distribution function dominates the empirical CDF in sup-norm error.
• 18:30Minimax lower/upper bounds for Grenander estimator, highlighting the n^{-1/3} rate.
• 22:30Introduction to log-concave densities: examples and closure properties.
• 27:30Existence and uniqueness of the log-concave maximum likelihood estimator, with population projection.
• 32:30Properties of the projection: convex ordering, continuity, and computational challenges (tent functions).
• 37:30Minimax rates for log-concave estimation: elbow at n^{-4/5} (1D) and n^{-2/(d+1)} (d≥2).