Interpretable Machine Learning for Kronecker Coefficients

The Big Picture

Imagine you’re trying to figure out how Lego sets combine. Given two sets, you want to know what new builds are possible by merging them, and how many ways each result can appear. Now swap “Lego sets” for abstract mathematical objects called representations (formal ways of describing how symmetries act on a structure) and you’re looking at one of the deepest open problems in mathematics: computing Kronecker coefficients.

These numbers describe how fundamental symmetry patterns combine when layered together. They show up across mathematics and physics, from quantum information theory to computational complexity.

Francis Murnaghan first posed the question in 1938: is there a combinatorial formula (a counting recipe built from discrete patterns) for these numbers? Nearly ninety years later, no such formula exists. Even determining whether a single Kronecker coefficient equals zero is NP-hard, meaning no efficient algorithm can handle it in general, regardless of computing power.

A team of IAIFI researchers has now used machine learning to crack this problem open from a new angle. Their models don’t just predict these mysterious numbers; they explain the predictions, producing human-readable formulas for a question that has stumped mathematicians for almost a century.

Key Insight: By pairing interpretable ML with a geometric embedding of partitions, the researchers derived explicit mathematical formulas predicting when a Kronecker coefficient is zero. Their interpretable models hit 83% accuracy with fully readable decision rules; transformer models pushed past 99%.

How It Works

The inputs are triples of partitions, ways of writing a number n as an ordered sum of positive integers. A partition of 6 might be (3,2,1), meaning 3+2+1=6. The Kronecker coefficient g(λ,µ,ν) is indexed by three such partitions, and the binary question is simple: is this coefficient zero or not?

Earlier work showed that convolutional neural networks and gradient boosting could answer this with ~98% accuracy. Impressive, but opaque. This paper pushes further: why do the models work? What features drive the predictions?

Gradient saliency analysis identifies which input features most strongly influence a network’s output. When partitions are represented as n-dimensional vectors, the analysis reveals that the first few and last few entries matter far more than the middle. The extremes of a partition carry the signal.

The second ingredient is the b-loading, a single number assigned to each partition. The researchers build a matrix of pairwise distances between every partition of n, then extract a ranking via standard linear algebra (principal component analysis). Each partition lands on a one-dimensional scale, rescaled to [0, 100]. That position is its b-loading.

For a triple (λ,µ,ν), the b-loading of the triple is just the sum b(λ) + b(µ) + b(ν). One number compressing a three-partition relationship into something workable. What turned up:

• Histograms of summed b-loadings follow gamma distributions
• Separating by whether g(λ,µ,ν) = 0 reveals a clean threshold
• Below a critical value b*, the coefficient is always nonzero, a provable sufficient condition

Three families of interpretable models were trained on b-loadings: Kolmogorov-Arnold Networks (KANs), which learn flexible functions expressible in closed form; small neural networks whose weights can be read directly; and PySR, a symbolic regression tool that searches for algebraic formulas. All three landed near 83% accuracy, close to the theoretical ceiling for b-loadings alone (~85%).

On the black-box side, transformer models (the architecture behind GPT-style language models) trained on raw partition triples hit over 99% accuracy, the highest ever reported for this problem. What’s going on inside those transformers remains an open question.

Why It Matters

An explicit formula for a decision function, even one covering 83% of cases, is real structural insight into a problem that has resisted attack since 1938. The b-loading sufficient condition (g(t) ≠ 0 whenever b(t) < b*) is a theorem, not a heuristic. Machine learning has produced a mathematical result.

The contrast between the two approaches is telling. The 99% transformer result is scientifically striking but mathematically opaque. The 83% symbolic regression result is less accurate but far more informative: it reveals what structure the data actually contains. As ML becomes a routine tool in mathematical research, this kind of translation from learned representations back to human-readable form will be essential for turning predictions into proofs.

The gap between 83% and 99% is itself a research program. What features do transformers find that b-loadings miss? Answering that could clarify both the structure of Kronecker coefficients and how neural networks handle combinatorial algebra.

Bottom Line: Interpretable ML has extracted explicit mathematical formulas predicting Kronecker coefficients, turning a black-box classifier into a source of algebraic insight. A 99%-accurate transformer suggests there’s more structure to find.