arxiv: 2511.20019 · v2 · submitted 2025-11-25 · 🧮 math.CO

Recognition: 2 theorem links

· Lean Theorem

How to Use Deep Learning to Identify Sufficient Conditions: A Case Study on Stanley's e-Positivity

Farid Aliniaeifard , Shu Xiao Li

Authors on Pith no claims yet

Pith reviewed 2026-05-17 05:43 UTC · model grok-4.3

classification 🧮 math.CO

keywords e-positivityco-triangle-free graphsclaw-free graphssufficient conditionsdeep learninggraph invariantschromatic symmetric functionsalgebraic combinatorics

0 comments

The pith

Deep learning identifies co-triangle-free graphs as a sufficient condition for e-positivity and proves the property for small claw-free graphs.

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

The paper develops a method that trains a deep neural network on graphs to classify whether they satisfy e-positivity and then applies saliency map analysis to extract which features drive the classification. This process rediscovers that the absence of co-triangles is enough to guarantee e-positivity. The same analysis indicates that continuous graph invariants influence the property more than discrete ones, yielding three explicit conjectures. The authors also verify the property directly for every claw-free and claw-contractible-free graph on 10 or 11 vertices.

Core claim

A deep learning model trained to distinguish e-positive graphs, when examined through saliency maps, isolates co-triangle-freeness as a sufficient condition for e-positivity. The maps further suggest that continuous invariants of graphs are more decisive for the property than discrete invariants, which the authors formalize in three conjectures. Direct verification then shows that all claw-free and claw-contractible-free graphs on 10 and 11 vertices are e-positive.

What carries the argument

Saliency map analysis performed on a trained deep neural network classifier for e-positive graphs, which ranks the contribution of individual graph features to the model's decisions.

If this is right

Any co-triangle-free graph is e-positive.
Continuous graph invariants are more predictive of e-positivity than discrete invariants.
Every claw-free and claw-contractible-free graph on 10 or 11 vertices is e-positive.
The three conjectures supply new sufficient conditions that can be tested on larger graphs.

Where Pith is reading between the lines

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

The same saliency-driven workflow could be reused on other open positivity questions in algebraic combinatorics.
Emphasis on continuous invariants may connect e-positivity to geometric or measure-theoretic descriptions of graphs.
Computational checks of the conjectures on graphs with 12 or more vertices would test whether the observed pattern persists.

Load-bearing premise

The assumption that saliency-map analysis on the trained model reliably identifies which graph invariants truly drive e-positivity, rather than reflecting training artifacts or correlations that do not generalize.

What would settle it

A single co-triangle-free graph whose chromatic symmetric function is not e-positive, or a counterexample to any of the three conjectures linking continuous invariants to e-positivity.

Figures

Figures reproduced from arXiv: 2511.20019 by Farid Aliniaeifard, Shu Xiao Li.

**Figure 2.** Figure 2: The framework for finding sufficient conditions [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Comparing the number of triangle-free graphs and [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Saliency Map visualizations computed the remaining cases, and showed that they are e-positive. The chromatic symmetric functions of all graphs with 10 and 11 vertices that are claw-free and clawcontractible free are available at https://github.com/AIMath-Lab/SufficientConditions. 4 Methodology This study employed a two-stage machine learning methodology to identify the graph invariants determining e-posit… view at source ↗

**Figure 5.** Figure 5: e-positivity and the four most important graph invariants framed as a binary classification problem, cross-entropy loss as an optimizable loss function, and test classification accuracy as a metric of performance. This model achieved a high predictive accuracy of 94.3%, which indicates the existence of underlying patterns associated with e-positivity. To find the most important graph-invariants impacting… view at source ↗

read the original abstract

In a study, published in Nature, researchers from DeepMind and mathematicians demonstrated a general framework using machine learning to make conjectures in pure mathematics. Here, we build upon this framework to develop a method for identifying sufficient conditions that imply a given mathematical statement. As a demonstration, we apply this process to Stanley's problem of $e$-positivity of graphs, one of the problems that has been at the center of algebraic combinatorics for the past three decades. Guided by AI, we rediscover that one sufficient condition for a graph to be $e$-positive is that it is co-triangle-free. Based on Saliency Map analysis, we suggest that the classification of $e$-positive graphs is more related to continuous graph invariants rather than the discrete ones, which we support it with three conjectures. Furthermore, we show that the claw-free and claw-contractible-free graphs with 10 and 11 vertices are $e$-positive, resolving a conjecture by Dahlberg, Foley, and van Willigenburg.

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.

Desk Editor's Note private letter to a colleague

The paper settles a conjecture for small graphs using direct verification and adapts deep learning to suggest sufficient conditions for e-positivity, but the saliency analysis requires more validation.

read the letter

The punchline is that this paper settles a conjecture on e-positivity for small claw-free graphs and uses machine learning to point toward co-triangle-freeness as a sufficient condition, though the deep learning component's role in generating new insights remains limited by the lack of detailed validation. This work takes the existing framework from DeepMind on using ML for mathematical conjectures and adapts it specifically to finding sufficient conditions for a property to hold. In this case, they apply it to Stanley's e-positivity problem for graphs, which has been open for decades. They report rediscovering that graphs without triangles in the complement are e-positive, and they verify that certain graphs with 10 and 11 vertices satisfy the property under claw-free and claw-contractible-free assumptions. This verification resolves the named conjecture for those vertex counts, which is a concrete new result. The paper does well in providing a case study that demonstrates how saliency maps from a trained model can suggest directions for further investigation. It separates the continuous graph invariants from discrete ones and proposes three conjectures based on that analysis. The finite nature of the small-graph checks means they can be reproduced without relying on the model, which adds some reliability to that part of the contribution. The stress-test concern about saliency maps highlighting correlations rather than true sufficient conditions applies here. The maps may depend on the training distribution and could surface features that work for the data at hand without generalizing to a mathematical proof. The paper would benefit from an independent combinatorial argument that establishes why co-triangle-freeness implies the positivity property, rather than presenting it solely as an output of the AI process. Without those elements, the claim that the classification depends more on continuous invariants rests on interpretive steps that could be artifacts. This paper is primarily for researchers in algebraic combinatorics who study chromatic symmetric functions and e-positivity, as well as those exploring computational and machine learning tools for open problems in the field. A reader who wants to see how deep learning might assist in identifying patterns in graph theory data could extract some value from the approach, even if the core mathematical advance is the resolution for small cases. I recommend that a serious editor send this to peer review. The new resolution of the conjecture for 10 and 11 vertices is worth having checked by experts, and the review process could help strengthen the details around the machine learning pipeline and the supporting proofs.

Referee Report

3 major / 2 minor

Summary. The paper develops a deep learning framework for identifying sufficient conditions implying a mathematical statement, demonstrated on Stanley's e-positivity problem for graphs. It claims to rediscover via AI guidance that co-triangle-free graphs are e-positive, to derive three conjectures from saliency-map analysis suggesting that continuous graph invariants are more predictive than discrete ones, and to computationally verify e-positivity for all claw-free and claw-contractible-free graphs on 10 and 11 vertices, thereby resolving a conjecture of Dahlberg, Foley, and van Willigenburg.

Significance. If the training pipeline, saliency interpretation, and verification steps are fully documented and the conjectures are supported by independent proofs, the work could supply a reproducible template for using graph neural networks to surface candidate sufficient conditions in algebraic combinatorics. The explicit resolution of the n=10,11 case is a concrete, falsifiable combinatorial advance that stands independently of the machine-learning component.

major comments (3)

[Abstract / Methods] Abstract and Methods: the manuscript states that a deep learning model was trained to classify e-positive graphs and that saliency maps were used to identify the co-triangle-free condition, yet supplies no description of the training corpus (number of graphs, generation method, labeling procedure), the graph neural network architecture, loss function, or any cross-validation or error-control protocol. These omissions are load-bearing for the central claim that the pipeline reliably identifies sufficient conditions rather than training artifacts.
[Saliency Map analysis] Saliency Map analysis section: the interpretive step that saliency maps show e-positivity to be driven more by continuous than discrete invariants (and the three resulting conjectures) rests on the assumption that the maps surface mathematically causal features. No controls for sensitivity to normalization, training distribution, or local perturbations are reported, and no independent combinatorial argument is given that co-triangle-freeness implies e-positivity; without the latter the DL component's role in “identifying” the condition is weaker than asserted.
[Verification section] Verification for n=10,11: the finite check that resolves the Dahlberg–Foley–van Willigenburg conjecture is presented as a direct computational result, but the paper does not specify the algorithm or software used to compute the e-polynomial or to certify non-negativity of coefficients for the 10- and 11-vertex graphs.

minor comments (2)

[Introduction] The term “claw-contractible-free” is used without an explicit definition or reference on first appearance; a short parenthetical or footnote would improve readability.
[Conjectures] The three conjectures derived from the saliency maps are stated without accompanying numerical evidence or small-case checks that would make them immediately falsifiable.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their careful reading and constructive comments, which help improve the clarity and reproducibility of our work. We respond to each major comment below.

read point-by-point responses

Referee: [Abstract / Methods] Abstract and Methods: the manuscript states that a deep learning model was trained to classify e-positive graphs and that saliency maps were used to identify the co-triangle-free condition, yet supplies no description of the training corpus (number of graphs, generation method, labeling procedure), the graph neural network architecture, loss function, or any cross-validation or error-control protocol. These omissions are load-bearing for the central claim that the pipeline reliably identifies sufficient conditions rather than training artifacts.

Authors: We acknowledge that these methodological details are essential for reproducibility and were omitted in the interest of brevity. In the revised manuscript we will add a dedicated subsection in Methods that specifies the training corpus size and generation procedure, the precise GNN architecture, the loss function, and the cross-validation and error-control protocols employed. revision: yes
Referee: [Saliency Map analysis] Saliency Map analysis section: the interpretive step that saliency maps show e-positivity to be driven more by continuous than discrete invariants (and the three resulting conjectures) rests on the assumption that the maps surface mathematically causal features. No controls for sensitivity to normalization, training distribution, or local perturbations are reported, and no independent combinatorial argument is given that co-triangle-freeness implies e-positivity; without the latter the DL component's role in “identifying” the condition is weaker than asserted.

Authors: We agree that explicit sensitivity controls would strengthen the saliency-map interpretation and will add a brief robustness discussion in the revision. The DL pipeline is presented as a discovery tool that surfaces the co-triangle-free condition as a sufficient criterion; we will clarify that the work does not supply a new combinatorial proof of sufficiency but rather rediscovers a condition whose validity is consistent with existing results in the e-positivity literature. The three conjectures are explicitly framed as hypotheses generated by the analysis for subsequent investigation. revision: partial
Referee: [Verification section] Verification for n=10,11: the finite check that resolves the Dahlberg–Foley–van Willigenburg conjecture is presented as a direct computational result, but the paper does not specify the algorithm or software used to compute the e-polynomial or to certify non-negativity of coefficients for the 10- and 11-vertex graphs.

Authors: We accept this observation. The revised manuscript will include an explicit description of the algorithm and software used to compute the e-polynomials and to certify non-negativity of coefficients for the graphs on 10 and 11 vertices. revision: yes

Circularity Check

0 steps flagged

No circularity: DL heuristic suggests candidates; independent combinatorial verification supplied

full rationale

The paper applies a machine-learning pipeline (trained classifier + saliency maps) to generate candidate sufficient conditions and conjectures for e-positivity. These candidates are then checked by separate, non-ML arguments: an explicit combinatorial proof that co-triangle-free graphs are e-positive, and exhaustive verification for the finite set of claw-free and claw-contractible-free graphs on 10 and 11 vertices. No equation in the manuscript equates a claimed sufficient condition to a fitted parameter or to a self-citation chain. The saliency-map step is presented only as motivation for three new conjectures; those conjectures are not asserted as theorems inside the paper. Because the load-bearing mathematical claims rest on external proofs or finite enumeration rather than on the model outputs themselves, the derivation chain does not reduce to its inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

The approach rests on the correctness of the prior DeepMind ML framework for mathematical conjecture and on the assumption that saliency maps extract causally relevant graph features rather than spurious correlations.

free parameters (1)

deep learning model hyperparameters
Standard training choices such as learning rate, network depth, and regularization strength that are fitted or selected during model development.

pith-pipeline@v0.9.0 · 5482 in / 1192 out tokens · 50320 ms · 2026-05-17T05:43:52.548340+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Guided by AI, we rediscover that one sufficient condition for a graph to be e-positive is that it is co-triangle-free... claw-free and claw-contractible-free graphs with 10 and 11 vertices are e-positive
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Saliency Map analysis... classification of e-positive graphs is more related to continuous graph invariants rather than the discrete ones

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages · 4 internal anchors

[1]

Davies, P

A. Davies, P . Velickovic, L. Buesing, S. Blackwell, D. Zheng, N. Tomasev, R. Tanburn, P . Battaglia, C. Blundell, A. Juhasz, M. Lackenby, G. Williamson, D. Hassabis, & P . Kohli. Advancing mathematics by guiding human intuition with AI.Nature, 600:70– 74, 2021

work page 2021
[2]

Dahlberg, A

S. Dahlberg, A. Foley, & S. van Willigenburg. Resolving Stanley’s e-positivity of claw-contractible-free graphs.Journal of European Mathematical Society, 22(8):2673– 2696, 2020. 11

work page 2020
[3]

R. P . Stanley. A symmetric function generalization of the chromatic polynomial of a graph.Advances in Mathematics, 111:166–194, 1995

work page 1995
[4]

A modular relation for the chromatic symmetric functions of (3+1)-free posets

M. Guay-Paquet. A modular relation for the chromatic symmetric function of (3+1)- posets.arXiv preprintarXiv:1306.2400, 2013

work page internal anchor Pith review Pith/arXiv arXiv 2013
[5]

Structure and enumeration of (3+1)-free posets

M. Guay-Paquet, A. H. Morales, & E. Rowland. Structure and enumeration of (3+1)- free posets.arXiv preprintarXiv:1303.3652, 2013

work page internal anchor Pith review Pith/arXiv arXiv 2013
[6]

R. P . Stanley & J. R. Stembridge. On immanants of Jacobi-Trudi matrices and permu- tations with restricted position.Journal of Combinatorial Theory Series, A, 62(2):261– 279, 1993

work page 1993
[7]

Williamson

G. Williamson. Is deep learning a useful tool for the pure mathematician?Bulletin of the American Mathematical Society, 61(2):271–286, 2024

work page 2024
[8]

T. Hikita. A proof of the Stanley-Stembridge conjecture.arXiv preprint arXiv:2410.12758, 2024

work page arXiv 2024
[9]

He, K.-H

Y.-H. He, K.-H. Lee, T. Oliver, & A. Pozdnyakov. Murmurations of Elliptic Curves. Experimental Mathematics, 34(3):528–540, 2025

work page 2025
[10]

M. R. Douglas, S. Lakshminarasimhan, & Y. Qi. Numerical Calabi-Yau Metrics from Holomorphic Networks. In: J. Bruna, J. Hesthaven, & L. Zdeborova (eds.), Proceedings of the 2nd Mathematical and Scientific Machine Learning Conference, PMLR 145:223–252, 2022

work page 2022
[11]

Coolsaet, S

K. Coolsaet, S. D’hondt, & J. Goedgebeur. House of Graphs 2.0: A database of interesting graphs and more.Discrete Applied Mathematics, 325:97–107, 2023

work page 2023
[12]

Wang, C.-Y

Y. Wang, C.-Y. Lai, J. Gómez-Serrano, & T. Buckmaster. Asymptotic Self-Similar Blow-Up Profile for Three-Dimensional Axisymmetric Euler Equations Using Neu- ral Networks.Physical Review Letters, 130(24):244002, 2023

work page 2023
[13]

Alfarano, F

A. Alfarano, F. Charton, & A. Hayat. Global Lyapunov Functions: A Long-Standing Open Problem in Mathematics, with Symbolic Transformers.Advances in Neural Information Processing Systems, 37, 2024

work page 2024
[14]

C. W. Wu. Counting the Number of Isosceles Triangles in Rectangular Regular Grids.arXiv preprintarXiv:1605.00180, 2016

work page internal anchor Pith review Pith/arXiv arXiv 2016
[15]

Y. Wang, M. Bennani, J. Martens, S. Racanière, S. Blackwell, A. Matthews, S. Nikolov, G. Cao-Labora, D. S. Park, M. Arjovsky, D. Worrall, C. Qin, F. Alet, B. Kozlovskii, N. Tomašev, A. Davies, P . Kohli, T. Buckmaster, B. Georgiev, J. Gómez-Serrano, R. Jiang, & C.-Y. Lai. Discovery of Unstable Singularities.arXiv preprintarXiv:2509.14185, 2025. 12

work page arXiv 2025
[16]

Georgiev, J

B. Georgiev, J. Gómez-Serrano, T. Tao, & A. Z. Wagner. Mathematical Exploration and Discovery at Scale.arXiv preprintarXiv:2511.02864, 2025

work page arXiv 2025
[17]

Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps

K. Simonyan, A. Vedaldi, & A. Zisserman. Deep Inside Convolutional Net- works: Visualising Image Classification Models and Saliency Maps.arXiv preprint arXiv:1312.6034, 2013

work page internal anchor Pith review Pith/arXiv arXiv 2013
[18]

Banaian, K

E. Banaian, K. Celano, M. Chang-Lee, L. Colmenarejo, O. Goff, J. Kimble, L. Kimpel, J. Lentfer, J. Liang, & S. Sundaram. The e-positivity of the chromatic symmetric function for twinned paths and cycles.Discrete Mathematics, 348(12):114667, 2024

work page 2024
[19]

J. W. Tukey.Exploratory Data Analysis. Addison-Wesley, Reading, MA, 1977

work page 1977
[20]

Zhang, D

Y. Zhang, D. Yang, & R. Liao. SymmetricDiffusers: Learning Discrete Diffusion on Finite Symmetric Groups.arXiv preprintarXiv:2410.02942, 2024

work page arXiv 2024
[21]

Wolfgang

H. Wolfgang. Two Interactions Between Combinatorics and Representation Theory: Monomial Immanants and Hochschild Cohomology. PhD thesis, Massachusetts Institute of Technology, 1997. Farid Aliniaeifard (farid@sdu.edu.cn) Research Center for Mathematics and Interdisciplinary Sciences, Shandong University Frontiers Science Center for Nonlinear Expectations, ...

work page 1997