Vladimir Dragovich - Complexity of Lasso with Normalized Data

Value of the Information & Strength of the Argument

The value of the information is high: the talk provides new results on the complexity of Lasso paths under normalization, a topic with both theoretical and practical implications. The argumentation is solid, built on geometric intuition and rigorous proofs. The speaker clearly delineates their contributions from existing literature, and the use of explicit examples (e.g., the P=3 construction) strengthens the validity of the claims. The connection to binary codes and Hamming distance provides a surprising yet elegant framework, and the reference to Kleitman’s theorem grounds the bound in established mathematics. The discussion of typical behavior versus worst-case scenarios (Lee-Singer vs. Mairal-Yu) adds depth to the argument.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is evident: the talk is based on a publication in Statistical Science and other preprint/paper works, though specific references are only cited verbally. The speaker acknowledges when they rediscovered known results (e.g., Kleitman’s theorem), showing intellectual honesty. The adequacy between title and content is perfect: the talk focuses on the complexity of Lasso with normalized data, and every part of the presentation serves that theme. The geometric setup and coding theory connection are all directly tied to explaining this complexity. No comments were provided for analysis, so no public reception could be assessed.

The title accurately reflects the content, focusing on the complexity of Lasso with normalized data.

9/10

The talk presents rigorous mathematical proofs and connects multiple fields, delivered by an expert at a prestigious institution. The results are novel and build on established literature, though some references are only verbally mentioned.

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

• 01:30Introduction and context: talk on complexity of Lasso with normalized data, joint work with Boris.
• 04:00Lasso introduced 30 years ago, L1 regularization, predictors becoming zero.
• 06:00Geometric formulation: ellipses and shrinking cross-polytopes, tangency point trajectory.
• 10:00Key observation: normalization of predictors changes complexity of the lasso path.
• 18:00General properties: each visited orthant gives one segment, opposite orthants cannot both be visited; previous results by Mairal-Yu and Lee-Singer.
• 24:00Construction of first normalized example in P=3 with a jump between orthants.
• 30:00Theorems for normalized lasso: at most one segment on coordinate axes, and full-dimensional orthants must agree in at least two coordinates.
• 40:00Connection to coding theory: binary words, Hamming distance, Kleitman's theorem, and computation of Hp2.

• carmin.tv — Video platform hosting the talk.
• Ellipsoidal billiards and moments (paper with Boris) — Mentioned as a prior paper published three years ago, related to the geometric foundations.
• Orthogonal and linear regression (published in Statistical Science) — Mentioned as a prior work in top statistical journal.
• Mairal and Yu paper on lasso complexity — Known result for non-normalized data, proving an upper bound of 2^p - 1.
• Lee and Singer (2018) — Proved that typical cases have very low complexity, contrasting with worst-case examples.
• Kleitman's theorem (1966) — Solved the problem of maximal anti-codes, giving HpR values used in the talk.