Machines Learn Number Fields, But How? The Case of Galois Groups

Technical Summary: Machines Learn Number Fields, But How? The Case of Galois Groups

Problem Statement
The paper addresses the classification of Galois groups for Galois extensions of the rational numbers $\mathbb{Q}$ with degrees 4, 6, 8, 9, and 10. While classical methods exist for determining these groups (often via discriminants, resolvents, or specific polynomial properties), the authors investigate whether the Galois group can be determined solely from a finite set of coefficients of the Dedekind zeta function, $\zeta_K(s)$. Specifically, they explore the relationship between the distribution of zeta coefficients $a_n(K)$ and the underlying Galois group structure. The central challenge is that while machine learning (ML) models can achieve high accuracy in this classification task, their decision logic is typically opaque ("black-box"). The authors aim to reverse-engineer this logic to derive rigorous mathematical criteria.

Methodology
The authors employ a "machine learning-driven discovery" paradigm:

1. Data Generation: They utilize the LMFDB (L-functions and Modular Forms Database) number field database, restricting their dataset to Galois extensions of degrees 4, 6, 8, 9, and 10. Features are derived from two sources:
   • Zeta Coefficients: The first 1000 coefficients ($a_n(K)$) of the Dedekind zeta function.
   • Polynomial Coefficients: Coefficients of the normalized defining polynomials of the number fields.
2. Model Training: They train interpretable machine learning models, primarily Decision Trees and Logistic Regression, using the scikit-learn library.
   • Decision trees are trained with the Gini criterion to maximize interpretability, allowing the authors to inspect the splitting rules.
   • Logistic regression is used to analyze feature weights, particularly to identify which indices (e.g., square or cube indices) carry the most predictive power.
3. Interpretation and Proof: The core methodology involves analyzing the structure of the trained models. When a model consistently splits on specific zeta coefficients (e.g., $a_{p^k}(K) = 0$), the authors formulate mathematical conjectures based on these patterns. They then use classical algebraic number theory—specifically ramification theory, decomposition groups, inertia groups, and the Chebotarev density theorem—to prove these conjectures rigorously.

Key Contributions and Results

• Quartic and Nonic Fields (Degrees 4 and 9):

  • Observation: Decision trees achieved near-perfect accuracy (100%) by splitting on coefficients with indices that are perfect powers (squares for degree 4, cubes for degree 9).
  • Theorem: For a Galois extension $K/\mathbb{Q}$ of degree $\ell^2$ (where $\ell$ is prime), the Galois group is cyclic ($C_{\ell^2}$) if and only if there exists a prime $p$ such that $a_{p^\ell}(K) = 0$. Conversely, if the group is $C_\ell^2$, then $a_{p^\ell}(K) \neq 0$ for all primes $p$.
  • Proof: This relies on analyzing decomposition types. For $C_\ell^2$ extensions, the decomposition group cannot be cyclic of order $\ell^2$ (which would require $a_{p^\ell}=0$) due to constraints on inertia and Frobenius actions in local fields.

• Sextic Fields (Degree 6):

  • Observation: The distinction between the cyclic group $C_6$ and the symmetric group $S_3$ depends on the values of $a_{p^2}(K)$ and $a_{p^3}(K)$ relative to the congruence class of $p \pmod 6$.
  • Result: The authors prove that if $p \equiv 1 \pmod 6$ and $a_{p^2}(K) = 1$, the group must be $C_6$. If $p \equiv 5 \pmod 6$ and $a_{p^2}(K) = 1$, the group must be $S_3$. Similar criteria are established for $a_{p^3}(K)$.
  • Mechanism: These results are derived from the fact that $S_3$ lacks certain cyclic quotients or subgroups required for specific decomposition types (e.g., totally ramified cyclic extensions of degree 3 or 6) when $p$ satisfies specific congruences.

• Decic Fields (Degree 10):

  • Observation: Distinguishing between the cyclic group $C_{10}$ and the dihedral group $D_5$ relies on coefficients indexed by squares ($p^2$) and fifth powers ($p^5$).
  • Result: Criteria are established where the vanishing or specific values of $a_{p^2}(K)$ and $a_{p^5}(K)$, combined with $p \pmod{10}$, uniquely determine the Galois group. For instance, $a_{p^5}(K)=1$ for $p \neq 5$ implies a cyclic extension.

• Octic Fields (Degree 8):

  • Abelian Case: The authors distinguish between $C_8$, $C_4 \times C_2$, and $C_2^3$. They prove that $a_{p^4}(K) = 0$ for some $p$ if and only if the group is $C_8$. Furthermore, $a_{p^2}(K) \neq 0$ for all $p$ if and only if the group is $C_2^3$.
  • Non-Abelian Case: Distinguishing between $D_4$ and $Q_8$ proved more difficult via zeta coefficients alone, as the possible decomposition types for a given prime are identical for both groups. However, the decision tree models still achieved high accuracy (98.67%) by exploiting subtle statistical differences in the distribution of square-indexed coefficients, though no simple "if-then" theorem was derived for this specific pair.
  • Abelian vs. Non-Abelian: A clear criterion is provided: if $a_{p^4}(K) = 0$, the extension is abelian (specifically $C_8$). If $a_{p^4}(K)=1$ and $a_{p^2}(K)>0$ for $p \equiv 3 \pmod 4$, the extension is non-abelian.

• Polynomial Coefficients:

  • While decision trees performed well using polynomial coefficients, the models were complex and difficult to interpret. The authors noted that the models often relied on the magnitude of coefficients (e.g., $C_6$ extensions having much larger coefficients than $S_3$), but they did not derive new theorems from these patterns, noting that the normalization of polynomials in LMFDB heavily influences these distributions.

Significance and Claims
The paper claims to demonstrate a new paradigm in mathematical research where machine learning serves not merely as a predictive tool but as a generator of conjectures.

• Interpretability: By using interpretable models (decision trees), the authors were able to "reverse-engineer" the machine's logic to find precise mathematical statements that were previously unknown or unexplored in the literature.
• Rigorous Proof: Unlike many ML applications in math that stop at high accuracy, this work completes the cycle by proving the conjectures generated by the models using traditional number-theoretic proofs.
• New Criteria: The paper provides simple, new criteria for classifying Galois groups based on zeta coefficients (e.g., checking if $a_{1000}(K) \leq 4$ implies cyclicity for nonic fields), which could not be formulated without the initial ML analysis.
• Limitations: The authors are modest about the scope, noting that while the approach works well for the specific degrees studied, it does not yet provide a universal algorithm for all degrees. They also highlight that for non-abelian octic groups ($D_4$ vs. $Q_8$), the ML models found statistical patterns that did not immediately translate into a simple theoretical criterion, suggesting that some ML insights may remain statistical rather than structural.

The work concludes by suggesting that this methodology could be extended to non-Galois extensions, higher degrees, and other base fields, potentially leading to more efficient algorithms for computing Galois groups.