How Twist Class Redundancy Drives the Prediction of Traces of Frobenius of Elliptic Curves

Technical Summary: Twist Class Redundancy Drives the Prediction of Traces of Frobenius of Elliptic Curves

Problem Statement
Recent applications of machine learning (ML) to number theory have yielded models capable of predicting invariants of arithmetic objects, such as the traces of Frobenius ($a_p(E)$) for elliptic curves. Notably, transformer-based models have achieved high Matthews Correlation Coefficients (MCC) in predicting these traces. However, the "black-box" nature of these models raises a critical question: are they learning deep arithmetic properties, or are they exploiting statistical redundancies inherent in the training datasets?

The authors identify a specific source of such redundancy: quadratic twist classes. For an elliptic curve $E$ and its quadratic twist $E_\chi$, the traces of Frobenius satisfy $a_p(E_\chi) = \chi(p)a_p(E)$ for all but finitely many primes $p$. Consequently, the magnitudes $|a_p(E)|$ are identical for all curves within a quadratic twist family. The paper posits that high-performing ML models may simply be learning to identify the quadratic twist family of a curve and recovering signs from previously observed data, rather than discovering new arithmetic patterns.

Methodology
To test this hypothesis, the authors propose a baseline algorithm that explicitly exploits twist class redundancies to predict $a_p(E)$. The methodology involves two primary experiments using the ECQ6 dataset (elliptic curves with conductor $N(E) < 10^6$):

1. Explicit Twist Hash Matching (Algorithm 1):

   • Curves are grouped by their twist hash $h(E)$, a unique identifier for quadratic twist classes derived from the sum of weighted absolute traces.
   • The algorithm precomputes sign patterns between curves sharing the same twist hash in the training set.
   • For a test curve, the algorithm matches its twist hash to the training set, retrieves the relative sign patterns, and uses majority voting to predict the sign of the missing trace $a_p(E)$.
   • This serves as an upper bound for performance driven purely by twist identification.

2. Proxy-Based Twist Identification (Algorithm 2):

   • To mimic the input constraints of previous ML studies (which did not use twist hashes as explicit features), the authors construct a proxy for the twist class using the absolute values of the last $k$ traces of Frobenius: $(|a_{p_1}|, \dots, |a_{p_k}|)$.
   • The authors evaluate different choices for $k$ (the number of primes used) and whether to use the smallest or largest primes.
   • They utilize clustering metrics, specifically the Adjusted Rand Index (ARI), to measure how well these absolute-value tuples partition the data compared to the true twist-hash partition.
   • Algorithm 2 replaces the twist hash in Algorithm 1 with these absolute-value tuples as grouping keys.

Key Results

• Baseline Performance: The explicit twist-matching algorithm (Algorithm 1) achieves an MCC of 0.7905 for predicting $a_{97}(E)$. This significantly outperforms the transformer models from prior work [4], which achieved MCCs of approximately 0.51–0.58 on the same task.
• Proxy Effectiveness: The proxy algorithm (Algorithm 2) using absolute values of the 8 largest primes ($k=8$) closely tracks the performance of the explicit twist-hash model. The Mean Absolute Error (MAE) between the per-prime MCCs of the proxy and the twist-hash model is minimal (0.0019).
• Clustering Analysis: Using the largest primes to form the proxy partition yields an ARI peaking near 0.85, whereas using the smallest primes yields an ARI near 0.28. This confirms that larger primes provide a more robust indicator of the twist class.
• Model Disagreement: While the proxy model and the twist-hash model agree on the vast majority of instances (disagreement drops below 3% for $p \ge 29$), they diverge more frequently on smaller primes, suggesting the proxy is highly faithful for larger primes.
• Unseen Twist Classes: When transformer models are evaluated on test sets containing curves from twist classes not present in the training set, their performance drops significantly, though they retain statistically significant predictive power for parity ($a_p \mod 2$) and modulo 4 features. This suggests that while twist redundancy drives the bulk of the performance, some structural signal remains.

Contributions

1. Demonstration of Redundancy: The paper provides empirical evidence that the high predictive accuracy of previous ML models on elliptic curve traces is largely driven by the dataset's redundancy within quadratic twist classes, rather than the discovery of novel arithmetic laws.
2. Baseline Algorithm: The authors introduce a simple, non-neural algorithm that outperforms complex transformer models by explicitly leveraging twist class identification.
3. Refined Benchmark Dataset: To prevent future models from exploiting these artifacts, the authors introduce a new dataset [2] consisting of unique representatives per quadratic twist class (curves with conductor $N(E) < 10^7$). This dataset removes the ability to predict via simple look-up of twist families, forcing models to learn genuine arithmetic properties.

Significance and Claims
The authors conclude that the "surprisingly strong performance" of previous ML models in predicting traces of Frobenius is an artifact of the data structure. They argue that to advance the field and discover new mathematical phenomena (such as the "murmurations" previously identified), future research must utilize datasets that eliminate quadratic twist redundancies. The paper does not claim that ML cannot find new patterns, but rather that current benchmarks are insufficient for distinguishing between learning deep arithmetic properties and exploiting dataset artifacts. The proposed unique-twist dataset is intended to serve as a rigorous benchmark for future studies in this domain.