pith. sign in

arxiv: 2508.06670 · v2 · pith:CI5QRHS5new · submitted 2025-08-08 · 🧮 math.NT · cs.LG

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

Kyu-Hwan Lee , Seewoo Lee This is my paper

Pith reviewed 2026-05-21 23:48 UTC · model grok-4.3

classification 🧮 math.NT cs.LG
keywords Galois groupsDedekind zeta coefficientsdecision treesnumber fieldsGalois extensionsmachine learningarithmetic statistics
0
0 comments X

The pith

Decision trees trained on Dedekind zeta coefficients yield new explicit criteria for classifying Galois groups of number fields.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper trains decision trees to classify Galois groups of Galois extensions of the rationals in degrees 4, 6, 8, 9, and 10 using sequences of Dedekind zeta coefficients. Interpreting the splits and features in these trees reveals how the distribution of zeta coefficients changes with the Galois group. From this interpretation the authors derive and prove new mathematical criteria that distinguish the groups without reference to the original models. A sympathetic reader cares because the approach turns a predictive tool into a source of verifiable arithmetic statements. The work thereby supplies another instance of machine learning guiding the discovery of number-theoretic structure.

Core claim

By applying interpretable machine learning methods such as decision trees, we study how simple models can classify the Galois groups of Galois extensions over Q of degrees 4, 6, 8, 9, and 10, using Dedekind zeta coefficients. Our interpretation of the machine learning results allows us to understand how the distribution of zeta coefficients depends on the Galois group, and to prove new criteria for classifying the Galois groups of these extensions.

What carries the argument

Decision trees whose successive splits on Dedekind zeta coefficients correspond to verifiable properties of their distributions and thereby furnish classification rules.

If this is right

  • Explicit, checkable rules exist for determining the Galois group from the first several zeta coefficients in degrees up to 10.
  • The distribution of zeta coefficients carries enough information to separate distinct Galois groups in these degrees.
  • Machine learning outputs can be translated into self-contained arithmetic statements that do not depend on the original algorithms.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same interpretive procedure could be applied to other arithmetic invariants or to extensions of higher degree once suitable training data become available.
  • The derived criteria might reduce the computational cost of Galois group computation for small-degree fields by replacing full resolvent calculations with coefficient checks.

Load-bearing premise

The features and splits found by the decision trees reflect general, provable distinctions in the zeta coefficient distributions that hold independently of the training data and model architecture.

What would settle it

A single Galois extension in one of the listed degrees whose Dedekind zeta coefficients violate one of the derived criteria while possessing the Galois group the criterion claims to exclude.

Figures

Figures reproduced from arXiv: 2508.06670 by Kyu-Hwan Lee, Seewoo Lee.

Figure 1
Figure 1. Figure 1: A decision tree predicts the Galois groups of quartic fields from zeta coefficients up to [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: A decision tree predicts the Galois groups of nonic fields from zeta coefficients up to [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: A decision tree predicts the Galois groups of quartic fields from polynomial coefficients. [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: A decision tree predicts Galois groups of sextic fields from zeta coefficients up to [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: A decision tree predicts Galois groups of decic fields from zeta coefficients up to [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: A decision tree predicts Galois groups of abelian octic fields from zeta coefficients up to [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: A decision tree predicts commutativity of Galois groups of octic fields from zeta coefficients up to [PITH_FULL_IMAGE:figures/full_fig_p030_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Distribution of the absolute values of the logistic regression model weights 40 [PITH_FULL_IMAGE:figures/full_fig_p040_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Distribution of absolute value of logistic regression model weights B.3 Decic fields (a) Decic fields, the whole set of weights (b) Decic fields, indices with square or quintic factors [PITH_FULL_IMAGE:figures/full_fig_p041_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Distribution of absolute value of logistic regression model weights 41 [PITH_FULL_IMAGE:figures/full_fig_p041_10.png] view at source ↗
read the original abstract

By applying interpretable machine learning methods such as decision trees, we study how simple models can classify the Galois groups of Galois extensions over $\mathbb{Q}$ of degrees 4, 6, 8, 9, and 10, using Dedekind zeta coefficients. Our interpretation of the machine learning results allows us to understand how the distribution of zeta coefficients depends on the Galois group, and to prove new criteria for classifying the Galois groups of these extensions. Combined with previous results, this work provides another example of a new paradigm in mathematical research driven by machine learning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript applies interpretable machine learning, specifically decision trees, to classify Galois groups of Galois extensions of Q in degrees 4, 6, 8, 9, and 10 using coefficients of the Dedekind zeta function. Interpretation of the learned decision rules is used to understand the dependence of zeta coefficient distributions on the Galois group and to derive and prove new classification criteria for these degrees. The work is presented as an instance of machine learning driving new mathematical results in number theory.

Significance. If the derived criteria are shown to be general, data-independent, and rigorously proven without post-hoc adjustment to the training distribution, the paper would illustrate a viable workflow from ML pattern discovery to provable statements in algebraic number theory. This could strengthen the case for ML-assisted exploration in the field, particularly when combined with the authors' prior results on related problems.

major comments (3)
  1. [§4.3] §4.3, decision tree for degree-8 fields: the reported split thresholds on the first few zeta coefficients (e.g., a_2 < 1.5) are presented as leading directly to a provable criterion, but the manuscript does not explicitly verify that these thresholds coincide with exact distinctions in the possible splitting types or Artin symbols that hold for all extensions of degree 8, independent of the finite training sample.
  2. [Theorem 6.2] Theorem 6.2 (degree-9 case): the proof invokes an additional case analysis on the possible cycle types that is not visibly entailed by the single decision-tree split shown in Figure 4; it is therefore unclear whether the final criterion is a direct formalization of the ML output or requires independent number-theoretic reasoning that could have been discovered without the tree.
  3. [§5.1] §5.1, general claim of 'parameter-free' criteria: the statement that the new rules are free of training-set dependence is not accompanied by a statement of the precise sample size, the range of discriminants used, or a cross-validation check showing that the same thresholds emerge on disjoint sets of fields.
minor comments (2)
  1. [Table 1] Table 1: the column headings for the Galois-group labels are not defined in the caption; readers must consult the text to learn that 'C4' denotes the cyclic group of order 4.
  2. [Figure 3] Figure 3: the x-axis label 'zeta coeff index' should specify whether the coefficients are normalized or raw, and the vertical lines marking the learned thresholds should be labeled with their numerical values.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments. The points raised help clarify the relationship between the machine-learning outputs and the general mathematical criteria. We respond to each major comment below and indicate the planned revisions.

read point-by-point responses

  1. Referee: [§4.3] §4.3, decision tree for degree-8 fields: the reported split thresholds on the first few zeta coefficients (e.g., a_2 < 1.5) are presented as leading directly to a provable criterion, but the manuscript does not explicitly verify that these thresholds coincide with exact distinctions in the possible splitting types or Artin symbols that hold for all extensions of degree 8, independent of the finite training sample.

    Authors: We agree that the manuscript would benefit from an explicit statement that the thresholds are not artifacts of the training sample. In the revised version we will augment §4.3 with a short general argument showing that the inequalities on the initial zeta coefficients are equivalent to restrictions on the possible cycle types (or Artin symbols) in the Galois group of any degree-8 extension of Q. This argument relies only on the factorization of the Dedekind zeta function and the Chebotarev density theorem, making the criterion independent of any finite dataset. revision: yes

  2. Referee: [Theorem 6.2] Theorem 6.2 (degree-9 case): the proof invokes an additional case analysis on the possible cycle types that is not visibly entailed by the single decision-tree split shown in Figure 4; it is therefore unclear whether the final criterion is a direct formalization of the ML output or requires independent number-theoretic reasoning that could have been discovered without the tree.

    Authors: The single split identified by the decision tree isolates the coefficient whose distribution most sharply separates the Galois groups in the training data. The subsequent case analysis on cycle types is required to obtain a complete, exhaustive criterion. In the revision we will insert a brief explanatory paragraph immediately before Theorem 6.2 that (i) records the precise coefficient and threshold suggested by the tree and (ii) states that the remaining case distinctions follow from standard Galois-theoretic bookkeeping once that coefficient is fixed. This makes transparent both the ML-guided starting point and the additional number-theoretic work needed for a full proof. revision: partial

  3. Referee: [§5.1] §5.1, general claim of 'parameter-free' criteria: the statement that the new rules are free of training-set dependence is not accompanied by a statement of the precise sample size, the range of discriminants used, or a cross-validation check showing that the same thresholds emerge on disjoint sets of fields.

    Authors: We will expand §5.1 to report the exact number of fields generated for each degree, the range of absolute discriminants sampled, and the results of a 5-fold cross-validation experiment performed on disjoint random partitions of the data. In each fold the same thresholds were recovered (within the precision of the floating-point representation), providing empirical evidence that the criteria are stable across different training distributions. These details will be added without altering the theoretical proofs already given. revision: yes

Circularity Check

0 steps flagged

No significant circularity: ML discovery followed by independent proofs

full rationale

The paper trains decision trees on Dedekind zeta coefficients to classify Galois groups for degrees 4,6,8,9,10, then interprets the learned splits to derive mathematical criteria that are subsequently proven rigorously. This separates empirical pattern discovery from formal verification; the proofs do not reduce to the training data or model outputs by construction. No self-citations are load-bearing for the central claims, and no fitted parameters are relabeled as predictions. The derivation chain is self-contained against external number-theoretic benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on standard facts about Dedekind zeta functions and Galois theory plus the assumption that decision-tree splits reveal distribution properties amenable to proof. No new free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption Coefficients of the Dedekind zeta function of a Galois extension encode information sufficient to distinguish Galois groups in low degrees
    This is the core input representation used for both ML training and subsequent proof extraction.

pith-pipeline@v0.9.0 · 5623 in / 1200 out tokens · 30675 ms · 2026-05-21T23:48:52.703477+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

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

    math.NT 2026-05 unverdicted novelty 7.0

    Redundancy within quadratic twist classes in elliptic curve datasets is sufficient to produce highly accurate machine learning predictions of Frobenius traces.

  2. Computer vision and converse theorems

    math.NT 2026-04 unverdicted novelty 6.0

    2D CNNs on image-encoded elliptic curve twists outperform 1D CNNs on Frobenius traces alone at separating conductor families from Sato-Tate random matrices and can predict analytic rank.

  3. AI for Mathematics: Progress, Challenges, and Prospects

    math.HO 2026-01 unverdicted novelty 4.0

    AI for math combines task-specific architectures and general foundation models to support research and advance AI reasoning capabilities.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages · cited by 3 Pith papers

  1. [1]

    Amir, M., He, Y .-H., Lee, K.-H., Oliver, T., and Sultanow , E.Machine learning class numbers of real quadratic fields.International Journal of Data Science in the Mathematical Sciences 1, 02 (2023), 107–134

    work page 2023

  2. [2]

    Irreducible sextic polynomials and their absolute resolvents.Minnesota Journal of Undergraduate Mathematics 1, 1 (2015)

    Awtrey , C., French, R., Jakes, P ., and Russell, A. Irreducible sextic polynomials and their absolute resolvents.Minnesota Journal of Undergraduate Mathematics 1, 1 (2015)

    work page 2015

  3. [3]

    On the Galois group of a reciprocal even octic polynomial

    Awtrey , C., and P atane, F. On the Galois group of a reciprocal even octic polynomial. Communications in Algebra 52, 7 (2024), 3018–3026

    work page 2024

  4. [4]

    On small discriminants of number fields of degree 8 and 9.J

    Battistoni, F. On small discriminants of number fields of degree 8 and 9.J. Théor. Nombres Bordeaux 32, 2 (2020), 489–501

    work page 2020

  5. [5]

    Bosma, W., Cannon, J., and Playoust , C.The Magma algebra system. I. The user language.J. SymbolicComput.24 ,3-4(1997),235–265. Computationalalgebraandnumbertheory(London, 1993)

    work page 1997

  6. [6]

    Enumeration of quartic fields of small discriminant

    Buchmann, J., Ford, D., and Pohst , M. Enumeration of quartic fields of small discriminant. mathematics of computation 61, 204 (1993), 873–879

    work page 1993

  7. [7]

    Carifio, J., Halverson, J., Krioukov , D., and Nelson, B. D. Machine learning in the string landscape. Journal of High Energy Physics 2017, 9 (2017), 157

    work page 2017

  8. [8]

    Finding Galois splitting models to compute local invariants.Journal of Number Theory 261(2024), 241–251

    Carrillo, B. Finding Galois splitting models to compute local invariants.Journal of Number Theory 261(2024), 241–251

    work page 2024

  9. [9]

    Cavallo, An elementary computation of the Galois groups of symmetric sextic trinomials , arXiv:1902.00965v2

    Ca v allo, A.AnelementarycomputationoftheGaloisgroupsofsymmetricsextictrinomials. arXiv preprint arXiv:1902.00965(2019)

    work page arXiv 1902

  10. [10]

    Chen, M. H. W., Chin, A. Y . M., and Tan, T. S.Galoisgroupsofcertainevenocticpolynomials. Journal of Algebra and its Applications 22, 12 (2023), 2350263

    work page 2023

  11. [11]

    Elliptic curve ‘murmurations’ found with AI take flight.Qunata Magazine(2024)

    Chiou, L. Elliptic curve ‘murmurations’ found with AI take flight.Qunata Magazine(2024)

    work page 2024

  12. [12]

    Number theory: Volume II: Analytic and modern tools, vol

    Cohen, H. Number theory: Volume II: Analytic and modern tools, vol. 240. Springer Science & Business Media, 2007

    work page 2007

  13. [13]

    Advancing mathematics by guiding human intuition with AI.Nature 600, 7887 (2021), 70–74

    Da vies, A., Veličković, P ., Buesing, L., Blackwell, S., Zheng, D., Tomašev , N., Tanburn, R., Battaglia, P ., Blundell, C., Juhász, A., et al. Advancing mathematics by guiding human intuition with AI.Nature 600, 7887 (2021), 70–74

    work page 2021

  14. [14]

    R., and Lee, K.-H

    Douglas, M. R., and Lee, K.-H. Mathematical data science. Advances in Theoretical and Mathematical Physics(2025). To appear. 42

    work page 2025

  15. [15]

    S., Foote, R

    Dummit , D. S., Foote, R. M., et al. Abstract algebra, vol. 3. Wiley Hoboken, 2004

    work page 2004

  16. [16]

    LMSJournal of Computation and Mathematics 17, 1 (2014), 141–158

    Fieker, C., and Klüners, J.ComputationofGaloisgroupsofrationalpolynomials. LMSJournal of Computation and Mathematics 17, 1 (2014), 141–158

    work page 2014

  17. [17]

    An elementary proof of the kronecker-weber theorem.The American Mathe- matical Monthly 81, 6 (1974), 601–607

    Greenberg, M. An elementary proof of the kronecker-weber theorem.The American Mathe- matical Monthly 81, 6 (1974), 601–607

    work page 1974

  18. [18]

    He, Y .-H.Machine-learning the string landscape.Physics Letters B 774(2017), 564–568

    work page 2017

  19. [19]

    Mathematics,Computation and Geometry of Data 2, 1 (2022), 49–66

    He, Y .-H., Lee, K.-H., and Oliver, T.Machine-learningnumberfields. Mathematics,Computation and Geometry of Data 2, 1 (2022), 49–66

    work page 2022

  20. [20]

    Machine-learning the Sato-Tate conjecture.J

    He, Y .-H., Lee, K.-H., and Oliver, T. Machine-learning the Sato-Tate conjecture.J. Symbolic Comput. 111(2022), 61–72

    work page 2022

  21. [21]

    Machine learning invariants of arithmetic curves.J

    He, Y .-H., Lee, K.-H., and Oliver, T. Machine learning invariants of arithmetic curves.J. Symbolic Comput. 115(2023), 478–491

    work page 2023

  22. [22]

    Experi- mental Mathematics(2024), 1–13

    He, Y .-H., Lee, K.-H., Oliver, T., and Pozdnyakov , A.Murmurationsofellipticcurves. Experi- mental Mathematics(2024), 1–13

    work page 2024

  23. [23]

    Iwasawa, K.On Galois groups of local fields.Transactions of the American Mathematical Society 80, 2 (1955), 448–469

    work page 1955

  24. [24]

    W., and Roberts, D

    Jones, J. W., and Roberts, D. P .Nonic 3-adic fields. InInternational Algorithmic Number Theory Symposium(2004), Springer, pp. 293–308

    work page 2004

  25. [25]

    W., and Roberts, D

    Jones, J. W., and Roberts, D. P . A database of local fields.Journal of symbolic computation 41, 1 (2006), 80–97

    work page 2006

  26. [26]

    W., and Roberts, D

    Jones, J. W., and Roberts, D. P . Octic 2-adic fields. Journal of Number Theory 128, 6 (2008), 1410–1429

    work page 2008

  27. [27]

    W., and Roberts, D

    Jones, J. W., and Roberts, D. P . A database of number fields.LMS Journal of Computation and Mathematics 17, 1 (2014), 595–618

    work page 2014

  28. [28]

    A database for field extensions of the rationals.LMS Journal of Computation and Mathematics 4(2001), 182–196

    Klüners, J., and Malle, G. A database for field extensions of the rationals.LMS Journal of Computation and Mathematics 4(2001), 182–196

    work page 2001

  29. [29]

    Machine learning of Calabi-Yau volumes.Physical Review D 96, 6 (2017), 066014

    Krefl, D., and Seong, R.-K. Machine learning of Calabi-Yau volumes.Physical Review D 96, 6 (2017), 066014

    work page 2017

  30. [30]

    Data-scientific study of Kronecker coefficients.Experimental Mathematics(2025)

    Lee, K.-H. Data-scientific study of Kronecker coefficients.Experimental Mathematics(2025). 43

    work page 2025

  31. [31]

    The L-functions and modular forms database

    LMFDB Collaboration, T. The L-functions and modular forms database. https://www. lmfdb.org, 2025. [Online; accessed 19 July 2025]

    work page 2025

  32. [32]

    Explicit local reciprocity for tame extensions

    Newton, R. Explicit local reciprocity for tame extensions. InMathematical Proceedings of the Cambridge Philosophical Society(2012), vol. 152, Cambridge University Press, pp. 425–454

    work page 2012

  33. [33]

    The computation of sextic fields with a cubic subfield and no quadratic subfield

    Olivier, M. The computation of sextic fields with a cubic subfield and no quadratic subfield. Math. Comp. 58, 197 (1992), 419–432

    work page 1992

  34. [34]

    Mathematics of Computation 70, 236 (2001), 1641–1659

    P auli, S., and Roblot , X.-F.On the computation of all extensions of a𝑝-adic field of a given degree. Mathematics of Computation 70, 236 (2001), 1641–1659

    work page 2001

  35. [35]

    Pedregosa, F., V aroquaux, G., Gramfort , A., Michel, V ., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P ., Weiss, R., Dubourg, V ., et al.Scikit-learn: Machine learning in Python.the Journal of machine Learning research 12(2011), 2825–2830

    work page 2011

  36. [36]

    Evolving neural networks with genetic algorithms to study the string landscape

    Ruehle, F. Evolving neural networks with genetic algorithms to study the string landscape. Journal of High Energy Physics 2017, 8 (2017), 1–20

    work page 2017

  37. [37]

    InAlgo- rithmic Number Theory: 8th International Symposium, ANTS-VIII Banff, Canada, May 17-22, 2008 Proceedings 8(2008), Springer, pp

    Voight , J.Enumeration of totally real number fields of bounded root discriminant. InAlgo- rithmic Number Theory: 8th International Symposium, ANTS-VIII Banff, Canada, May 17-22, 2008 Proceedings 8(2008), Springer, pp. 268–281

    work page 2008

  38. [38]

    W ang, S.A counter-example to Grunwald’s theorem.Ann. Math. 49, 4 (1948), 1008–1009

    work page 1948

  39. [39]

    W ang, S.On Grunwald’s theorem.Ann. Math. 51, 2 (1950), 471–484. Department of Mathematics, University of Connecticut, Storrs, CT 06269, USA Korea Institute for Advanced Study, Seoul 02455, Republic of Korea khlee@math.uconn.edu Department of Mathematics, University of California—Berkeley, Berkeley, CA 94720, USA seewoo5@berkeley.edu 44

    work page 1950