News · 2026-10-12

Mathematicians argue that illustration belongs inside research

Kate Stange and Edmund Harriss have argued that mathematical illustration should count as part of research because it can reveal structure, guide discovery and make results understandable. Their October 11 guest post on Terence Tao’s blog connects that argument to AI’s ability to generate code and proofs. Easier production makes the quality of explanation more important, while attractive images still require rigorous checking.

Key facts

• The post contrasts an earlier account of a three-dimensional aperiodic monotile with a later eight-page exposition.
• “Illustration beyond Proof” appeared on Tao’s blog on October 11.
• Its authors are mathematicians Kate Stange and Edmund Harriss, rather than Tao.
• The primary source is the guest post, supported by linked research records.

A diagram can do more than decorate a finished argument. It can expose a pattern that nobody had formulated yet, or provide the mental model that lets another researcher use a difficult theorem. Stange and Harriss put that kind of work inside the research process. Their position is especially relevant when AI makes producing technically complicated artifacts easier than explaining them.

The authors’ related paper, Mathematical Illustration: Proof, Intuition and AI, develops the same subject. The blog reports acceptance in the Notices of the American Mathematical Society; the checked arXiv record is a preprint record. Those are different kinds of publication evidence, so the acceptance statement should remain attributed to the authors.

The post’s historical examples connect drawing and physical models to mathematical knowledge. Its more immediate example concerns pictures of the complex roots of rational cubic polynomials. A shader implementation distorted circles into ellipses. The authors describe how that apparent rendering problem exposed further structure worth investigating. An image was helping generate a research question rather than simply illustrating an already settled one.

That example also explains their insistence on rigor. When a picture invites the viewer to infer a mathematical fact, the maker should establish that the relevant visible property is real. A smooth curve, a gap or a repeated shape may depend on resolution, numerical choices or rendering code. The image can be informative without every apparent feature being a theorem.

A useful analogy is a scientific instrument. A telescope expands what an observer can see, but its optics and calibration affect what appears. A mathematical visualization likewise depends on the program and assumptions that produce it. The right response is to understand what the instrument resolves and which conclusions survive its limitations, rather than choosing between blind trust and rejection of visual evidence.

The post uses the Mandelbrot set to make this tangible. Features smaller than a pixel spacing can disappear. Some countable features can be established at sufficiently fine resolution, but a beautiful picture alone cannot establish all the structure a viewer imagines. The authors’ framework joins two tests: whether the mathematical property is justified, and whether other people actually perceive the intended message.

For the second test, they discuss an art-school critique in which a work is shown with little introduction and the audience reports what it sees. That practice is not a substitute for proof. It helps test communication. A rigorously correct image that systematically conveys the wrong intuition can still be a poor explanation.

The AI connection cuts both ways. A language model can make visualization programming accessible to people who could not readily build the code themselves. But if the user cannot explain that code, access to the image may grow while understanding of its production shrinks. The existing lesson on program synthesis gives a related account of the gap between creating a program and specifying the behavior it should guarantee.

The post’s Chair44 example brings this back to proof. A Strongly Aperiodic Monotile in Three Dimensions describes a machine-checked finite test and formally checked components. Chaim Goodman-Strauss’s later notes give an eight-page exposition built around a diagram of the hierarchical tiling structure. Stange and Harriss use the contrast to show how a result becomes more inspectable and useful through exposition. Their account of the original discoverer’s understanding remains their characterization.

The honest caveat is that this is a methodological argument with examples, not a controlled study showing that illustration improves every reader’s performance. The available Hacker News discussion was also too thin to establish a community verdict. The strongest practical objection is that diagrams can seduce people into accepting structure their construction has not justified. The authors’ demand for rigorous visual claims addresses that concern directly. Their contribution is to make explanatory quality an explicit research responsibility as automated result production expands.

Primary source, verified: read the paper →

Ground Truth. (2026, October 12). Mathematicians argue that illustration belongs inside research. Ground Truth. https://groundtruth.day/news/illustration-beyond-proof-stange-harriss.html

@misc{groundtruth:illustration-beyond-proof-stange-harriss,
  title  = {Mathematicians argue that illustration belongs inside research},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/illustration-beyond-proof-stange-harriss.html}
}

Topics: mathematics · ai-science · explainability · research-culture