Abstract
Determining the vanishing of Kronecker coefficients ($g_{\lambda, \mu}^\nu$) is a computationally difficult problem in representation theory, known to be NP-hard. While deep learning models have recently achieved high accuracy in predicting these coefficients, their "black-box" nature offers little mathematical insight. This paper presents a "white-box" approach using interpretable machine learning techniques, such as Kolmogorov-Arnold Networks (KANs) and symbolic regression. We identify the "$b$-loading"—a single scalar feature—as a powerful predictor, enabling simple mathematical formulas to achieve 83% accuracy. Furthermore, we establish a new state-of-the-art accuracy of over 99% using a custom Transformer architecture, highlighting a significant "interpretability gap" that warrants further mechanistic study.
Introduction
Kronecker coefficients describe the decomposition of tensor products of irreducible representations of the symmetric group. Since their introduction in 1938, a general combinatorial rule for these coefficients has remained elusive. Recent advancements in machine learning have demonstrated that neural networks can predict whether a coefficient vanishes with high precision. However, the lack of interpretability in these models limits their utility for theoretical mathematics. This research aims to extract human-readable mathematical conditions from these models to advance our understanding of representation theory.
Methods
The study utilizes a multi-pronged approach to interpretability:
- Feature Engineering: Beyond raw partition vectors, the authors introduced the $b$-loading, derived from the principal components of a difference matrix between partitions.
- Interpretable Architectures: Kolmogorov-Arnold Networks (KANs) were employed to visualize learned functions as splines on graph edges. Symbolic regression (PySR) was used to search for the simplest analytical expressions fitting the data.
- High-Performance Models: A 12-block Transformer was developed to push the boundaries of predictive accuracy.
Key Insights: The Power of $b$-loadings
Saliency analysis revealed that the "ends" of partition vectors (the first and last few entries) carry the most significant information. This led to the discovery of the $b$-loading as a primary 1D feature. As shown in the distribution analysis, this single value serves as a robust indicator of the coefficient's behavior.
The separation between the two classes in the $b$-loading space (Figure 5) explains why even simple linear models can achieve high accuracy. The authors identified a threshold $b_*$ such that if the sum of $b$-loadings for a triple is below this value, the Kronecker coefficient is guaranteed to be non-zero, providing a new sufficient condition for non-vanishing.
Results and Discussion
The interpretable models achieved approximately 83% accuracy using only the 1D $b$-loading feature. Symbolic regression yielded explicit formulas, such as a sigmoid-based predictor: $$F_2(b) = \sigma(-0.0855 \times b + 7.2797)$$
In contrast, the Transformer model reached a state-of-the-art accuracy of >99%, surpassing previous benchmarks. This 16% accuracy gap suggests that while $b$-loadings capture the "average-case" simplicity of the problem, the remaining complexity involves non-trivial interactions that require more sophisticated attention mechanisms.
Conclusion
This work demonstrates that interpretable ML can bridge the gap between high-performance prediction and mathematical discovery. While simple formulas provide significant insight, the "interpretability gap" points toward deeper, yet-to-be-discovered structures in representation theory. Future research will focus on mechanistic interpretability to decode the logic within the 99%-accurate Transformer models.
Summary updated