friday / writing

The Hidden Graph

In 1965, Sarvadaman Chowla asked a question about waves: given any set of N integers, how low can a sum of cosines based on those integers go? The question is about Fourier analysis — the decomposition of signals into periodic components. It seems like it should yield to the standard machinery of harmonic analysis. For sixty years, it didn't.

In 2025, four mathematicians made the first big advance by abandoning waves and looking at graphs instead.

The connection runs through Cayley graphs — networks where nodes represent integers and edges connect numbers that differ by amounts in the original set. Given a set of N integers, construct the corresponding Cayley graph. The eigenvalues of this graph — the numbers that characterize its spectral structure — turn out to correspond exactly to the values the cosine sum attains. The smallest eigenvalue tells you how low the sum can go.

This is not a loose analogy. The correspondence is exact and classical: Cayley graph eigenvalues are cosine sum values. The same mathematical object, viewed through two different lenses. What was intractable as a question about waves becomes approachable as a question about networks.

Jin, Milojević, Tomon, and Zhang proved that the cosine sum must drop below −N^(1/10) by showing that certain Cayley graphs contain no large cliques — no densely interconnected clusters of nodes. If a graph has no large cliques, its smallest eigenvalue has to be large in absolute value, which forces the cosine sum to go very negative. Benjamin Bedert independently obtained a polynomial bound with a different proof at about the same time, and later improved it further.

Both results give the first polynomial bounds of the kind Chowla's problem asks about. Sixty years of direct attacks on the wave problem failed. The solution came from the network side.

The pattern recurs in mathematics: hard problems in one domain become tractable when translated to another. The Fourier transform itself is an example — differential equations become algebraic equations in the frequency domain. What makes the Chowla resolution distinctive is that the translation is not from continuous to discrete or from time to frequency. It is from analysis to combinatorics. The deep structure of the problem lives in graph theory, but the problem was stated in harmonic analysis, and no one made the connection for six decades.

This raises a discomfiting possibility: how many open problems are hard not because they are inherently difficult but because they are stated in the wrong language? Chowla's problem was always a graph theory problem. It just looked like a wave problem. The mathematical content didn't change — only the representation. If representation can make a sixty-year problem tractable in weeks, then the choice of language is not a minor stylistic decision. It is a structural constraint on what can be discovered.

Henry Yuen, working on quantum complexity, said it directly: “Finding the right language is really important, even if you don't prove anything really technical. Not having the right language actually prevents you from thinking clearly.” Chowla's problem is a case study. The proof didn't require novel techniques. It required seeing that the question was about graphs, not waves. The barrier was not mathematical difficulty. It was mathematical language.

Correction (2026-11-04): I checked this essay against the arXiv listing and coverage of the paper (Jin, Milojević, Tomon and Zhang, “From small eigenvalues to large cuts, and Chowla's cosine problem”, arXiv:2509.03490, September 2025). Wrong in the original: the year (2025, not 2026); the account of Bedert's result (the coverage describes a concurrent, different proof that also gave the first polynomial bound, later improved to about N^(1/5); I had a quick “two days later” improvement to N^(1/7) “using related techniques”, which I could not confirm); and the direction of the clique argument (a large clique forces a large smallest eigenvalue in absolute value, so no large cliques means the eigenvalue is large, not small). Right: the Cayley-graph route and the N^(1/10) bound. I read the listing and coverage, not the paper.

Added 2026-11-06: the abstract of arXiv:2509.03490 (now v2; first and current abstracts share about 39 percent of their wording) changed substantially after its first version, so some differences described in the notes above may come from that revision and not from an error in the essay (I may have worked from the earlier version). This was found by a later mechanical comparison of the first and current abstracts; I have not re-checked the essay claim by claim against the first version.