A Fields Medalist's 2019 NeurIPS paper is a case study in what NeurIPS 2026 now invites

Hong Wang, one of four mathematicians awarded the 2026 Fields Medal on July 23, is also a co-first author on a 2019 NeurIPS paper that applied a 90-year-old harmonic analysis theorem to a classical machine learning problem. The paper is a quiet case study in what NeurIPS, the field's flagship machine learning research conference, now explicitly invites: cross-disciplinary theory work that tightens the analysis of standard algorithms, independent of empirical state-of-the-art benchmarks.

The paper, "Optimal Analysis of Subset-Selection Based L_p Low-Rank Approximation," was authored by Chen Dan, Hong Wang, Hongyang Zhang, Yuchen Zhou, and Pradeep Ravikumar, with Wang and Zhang marked as equal-contribution co-first authors. It studies a procedure called column subset selection: given a large data table, pick a small number of columns whose span is as close as possible to the full table's span. The procedure is a compression step, a giant matrix replaced by a smaller, near-equivalent one. The authors proved a tighter bound on how much error this kind of compression introduces, and showed the bound is essentially best possible.

The technical move was to import Riesz–Thorin interpolation, a classical tool from harmonic analysis dating to the 1930s, into the analysis. The authors proved the bound at the three endpoints p=1, p=2, and p=∞, then interpolated to all intermediate p. This improved the entrywise ℓ_p bound for column subset selection from a uniform O(k+1) to (k+1)^(1/p) for 1≤p≤2 and (k+1)^(1−1/p) for p≥2, with a matching lower bound up to constant 1. The paper was published in NeurIPS 2019 and posted to arXiv as 1910.13618.

For most of the past decade, a paper like this would be an outlier at NeurIPS, a theory paper from a harmonic-analysis tradition rather than a neural-network result. That is changing. According to qbitai's reporting, NeurIPS 2026 appears to have introduced a contribution-type taxonomy that explicitly elevates theory work alongside general, use-inspired, concept and feasibility, and negative-results contributions, and its reviewer guidelines now value theoretical rigor and the introduction of new proof tools from other disciplines on their own terms, independent of empirical state-of-the-art claims.

Hong Wang's paper is the kind of work that taxonomy was built to admit. It is a tight bound on a classical theoretical problem, not a new method or model. The 2019 result was the contribution itself, not a stepping stone to a benchmark. The Simons Foundation's 2026 announcement and NPR's coverage frame Wang's medal in pure-math terms, but her 2019 paper sits squarely inside the new NeurIPS frame.

There is a small editorial hook in the qbitai narrative that surfaced the story: on Wang's Princeton personal page, which lists more than 40 papers and preprints, the 2019 NeurIPS paper is the only entry without a direct link. That observation is journalistic, not primary, but it fits a larger pattern. The paper was published, cited, and then treated as a footnote, even though its mathematical content was already the kind of cross-disciplinary work NeurIPS 2026 now formally invites.

The takeaway is not that AI is suddenly friendly to mathematicians. It is that the boundary between mathematics and machine learning research was never clean, and the venue is now acknowledging it. A 90-year-old theorem delivered a near-tight bound on a standard problem. Seven years later, the conference that published it is rewriting its reviewer guidelines to admit exactly that kind of contribution.