pith. sign in

arxiv: 2512.17417 · v3 · pith:OXMQ7QEWnew · submitted 2025-12-19 · 🧮 math.OC

Graph Isomorphism: Mixed-Integer Convex Optimization from First-Order Methods

Pith reviewed 2026-05-21 16:52 UTC · model grok-4.3

classification 🧮 math.OC
keywords graph isomorphismmixed-integer convex optimizationfirst-order methodsvariable fixingpolyhedral structurecomputational experimentssymmetry detection
0
0 comments X

The pith

A mixed-integer convex formulation of graph isomorphism can be solved using first-order optimization methods.

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

The paper develops a convex mixed-integer program that exactly encodes whether two graphs are isomorphic. It solves this program with first-order convex optimization techniques and adds variable-fixing presolvers that remove provably fixed variables while keeping the formulation polyhedral. Computational tests on twelve families of hard instances show that the approach beats the next-best integer-feasibility method on six families and performs equally well on symmetric graphs. Symmetry helps the optimization methods, while local-structure presolving is decisive for asymmetric cases. The work therefore supplies a new practical route to an intermediate-complexity problem by importing tools from convex optimization.

Core claim

We propose a convex mixed-integer formulation of the graph isomorphism problem and leverage first-order convex optimization to tackle it, following a stream of recent work on optimization-driven graph isomorphism detection. We strengthen our formulation with variable fixing techniques that prove highly effective while preserving the polyhedral structure. On a collection of challenging GI instances the method outperforms the integer-feasibility baseline on six of twelve graph families and matches it on the symmetric families.

What carries the argument

The mixed-integer convex formulation of graph isomorphism, strengthened by variable-fixing techniques that detect and remove fixed variables without destroying polyhedral structure.

If this is right

  • First-order convex solvers become applicable to an intermediate-complexity problem.
  • Variable fixing reduces instance size while preserving correctness and polyhedrality.
  • Optimization-based methods gain an advantage on graphs with high symmetry.
  • Presolving that detects local substructures becomes essential for asymmetric instances.

Where Pith is reading between the lines

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

  • The same convex encoding might be adapted to related problems such as subgraph isomorphism or graph canonization.
  • Further polyhedral studies could tighten the formulation and reduce the number of variables that require fixing.
  • Hybrid schemes that combine the convex model with spectral heuristics on the hardest asymmetric cases could be tested.

Load-bearing premise

The mixed-integer convex formulation has no spurious solutions and captures every true isomorphism.

What would settle it

Two non-isomorphic graphs that satisfy the integer constraints of the formulation, or two isomorphic graphs that violate them.

Figures

Figures reproduced from arXiv: 2512.17417 by Deborah Hendrych, Mathieu Besan\c{c}on, Patrick Gel{\ss}, Sebastian Pokutta, Stefan Klus, Wenjie Xiao.

Figure 1
Figure 1. Figure 1: Solved instances over time for sts graph family. For both Paley graph families, Figures 3 and 4, the clique and star detec￾tion are nearly equally effective. Our method outperforms all other methods on these sets as well. Interestingly, the spectral method outperforms the feasibility approach on these sets. It, in particular, fares well on the paley_prime family. The iso_r01N family, [PITH_FULL_IMAGE:figu… view at source ↗
Figure 2
Figure 2. Figure 2: Solved instances over time for usr graph family. tection determines non-isomorphism at the root node for all instances. We high￾light however that having identical spectra is a necessary condition for isomor￾phism and is not respected by these graphs. Spectral analysis would therefore also discard them as non-isomorphic. Investigating the performance of optimization￾based methods on isospectral but non-iso… view at source ↗
Figure 3
Figure 3. Figure 3: Solved instances over time for paley_power graph family. 6 Conclusion In this paper, we develop and evaluate mixed-integer convex formulations and algorithms for the graph isomorphism problem. Despite the problem being well￾studied on theoretical and algorithmic aspects, we showed that a mixed-integer convex approach leveraging first-order methods and graph-specific presolving techniques could achieve grea… view at source ↗
Figure 4
Figure 4. Figure 4: Solved instances over time for paley_prime graph family. To effectively solve GI for asymmetric graphs, presolving techniques that detect substructures and infer variable fixings from them are crucial. The matching problem warrants further specific investigation to develop new formulations, algorithms, and understand in which cases various approaches can perform well [PITH_FULL_IMAGE:figures/full_fig_p015… view at source ↗
Figure 5
Figure 5. Figure 5: Solved instances over time for iso_r01N graph family [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
read the original abstract

The graph isomorphism (GI) problem, which asks whether two graphs are structurally identical, occupies a unique position in computational complexity -- it is neither known to be solvable in polynomial time, nor proven to be NP-complete. We propose a convex mixed-integer formulation of the problem and leverage first-order convex optimization to tackle it, following a stream of recent work on optimization-driven graph isomorphism detection. We strengthen our formulation with variable fixing techniques that prove highly effective while preserving the polyhedral structure. We perform extensive computations evaluating the performance of different families of methods including a mixed-integer convex formulation, mixed-integer linear optimization, local search and spectral heuristics over a collection of challenging GI instances. We find that a high level of symmetry is beneficial for optimization-based methods. On the other hand, presolving techniques that detect local substructures to fix variables are crucial for asymmetric instances. The proposed method outperforms the second best approach, the integer feasibility approach, on 6 of the 12 graphs families and is on par with it on symmetric families.

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 paper proposes a mixed-integer convex formulation for the graph isomorphism problem, solved using first-order convex optimization methods. It strengthens the model with variable-fixing presolving techniques that preserve the polyhedral structure and reports computational comparisons against mixed-integer linear optimization, local search, and spectral heuristics across 12 graph families, claiming outperformance on 6 families and parity on symmetric ones, with observations on the benefits of symmetry and presolving for asymmetric instances.

Significance. If the central equivalence holds, the work adds a convex optimization lens to graph isomorphism detection, extending recent optimization-driven approaches. The emphasis on first-order methods and effective variable fixing offers a practical route for instances where symmetry or local structure can be exploited, and the multi-family computational study provides empirical guidance on method selection that could inform future algorithm design in combinatorial optimization.

major comments (3)
  1. [Formulation section (around the definition of the MI convex program)] The manuscript asserts that the mixed-integer convex program exactly captures graph isomorphism (integer solutions iff the graphs are isomorphic, with each solution corresponding to a valid adjacency-preserving bijection). No explicit theorem, proof, or exhaustive verification on small instances is provided to confirm absence of spurious integer solutions or missed isomorphisms. This equivalence is load-bearing for interpreting the reported computational superiority.
  2. [§4] §4 (variable-fixing techniques): The presolver is claimed to fix variables only when they are constant across the feasible set while preserving all valid isomorphisms and the polyhedral structure. No formal argument, invariant proof, or counter-example search is given to establish that no isomorphism is eliminated. Any over-fixing would render the subsequent optimization results unreliable for exact detection.
  3. [Computational results section] Results section / Table summarizing the 12 families: The claim of outperformance on 6 of 12 families lacks reported instance sizes, solver tolerances, termination criteria, number of replicates, or statistical tests for significance. Without these, it is impossible to determine whether the observed differences reflect genuine algorithmic advantage or artifacts of implementation details.
minor comments (2)
  1. [Introduction] The abstract and introduction could more explicitly reference the specific first-order method employed (e.g., projected gradient, ADMM) and its convergence guarantees under the mixed-integer setting.
  2. [Notation and formulation] Notation for the convex relaxation and the integer variables could be introduced with a single consolidated table or diagram to improve readability for readers outside the immediate subfield.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment point by point below, indicating the revisions we will make to strengthen the presentation while preserving the core contributions.

read point-by-point responses
  1. Referee: [Formulation section (around the definition of the MI convex program)] The manuscript asserts that the mixed-integer convex program exactly captures graph isomorphism (integer solutions iff the graphs are isomorphic, with each solution corresponding to a valid adjacency-preserving bijection). No explicit theorem, proof, or exhaustive verification on small instances is provided to confirm absence of spurious integer solutions or missed isomorphisms. This equivalence is load-bearing for interpreting the reported computational superiority.

    Authors: We agree that an explicit theorem and proof would strengthen the manuscript. The formulation is obtained by lifting the standard permutation-matrix encoding of graph isomorphism into a mixed-integer convex program over the adjacency matrices. In the revised version we will insert a formal theorem stating the exact equivalence (integer solutions correspond one-to-one with isomorphisms) together with a concise proof sketch deriving it from the Birkhoff-von Neumann theorem and the definition of adjacency preservation. We will also add a short computational verification on small instances (paths, cycles, and complete graphs of order at most 6) confirming absence of spurious solutions. revision: yes

  2. Referee: [§4] §4 (variable-fixing techniques): The presolver is claimed to fix variables only when they are constant across the feasible set while preserving all valid isomorphisms and the polyhedral structure. No formal argument, invariant proof, or counter-example search is given to establish that no isomorphism is eliminated. Any over-fixing would render the subsequent optimization results unreliable for exact detection.

    Authors: We acknowledge that a formal invariant argument is missing. The fixing rules rely on degree sequences, neighborhood cardinalities, and other local invariants that are necessarily preserved by any isomorphism. In the revision we will supply a short proof that each fixing step eliminates only values that cannot appear in any feasible permutation matrix, thereby preserving the entire set of isomorphisms and the underlying polyhedron. We will also note that the rules were cross-checked against known automorphism groups on the test families. revision: yes

  3. Referee: [Computational results section] Results section / Table summarizing the 12 families: The claim of outperformance on 6 of 12 families lacks reported instance sizes, solver tolerances, termination criteria, number of replicates, or statistical tests for significance. Without these, it is impossible to determine whether the observed differences reflect genuine algorithmic advantage or artifacts of implementation details.

    Authors: We thank the referee for highlighting this omission. The revised manuscript will expand the computational section with an appendix table listing, for each of the 12 families: number of vertices/edges, number of replicates (10 random instances per family where stochastic generation was used), solver tolerances (1e-6 relative gap), termination criteria (3600 s time limit or 10 000 iterations), and summary statistics (mean and standard deviation of run times). We will also add a brief discussion of the observed differences without claiming statistical significance beyond the raw averages. revision: yes

Circularity Check

0 steps flagged

No circularity: formulation derived directly from GI definition

full rationale

The paper proposes a mixed-integer convex formulation for graph isomorphism and strengthens it with variable-fixing techniques that preserve the polyhedral structure. No quoted equations or steps in the provided abstract and description reduce the central claim to a fitted parameter, self-citation chain, or self-definitional loop. The derivation is presented as following from the standard definition of structural identity between graphs, with computational experiments on external instance families serving as validation rather than input to the formulation itself. This is a self-contained construction against the GI problem statement.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The formulation rests on standard convex optimization theory and the assumption that the chosen integer constraints correctly encode vertex bijections; no new entities or fitted constants are introduced in the abstract.

axioms (1)
  • domain assumption The mixed-integer convex program is a correct encoding of graph isomorphism.
    Stated implicitly when the authors propose the formulation as a way to tackle GI.

pith-pipeline@v0.9.0 · 5724 in / 1136 out tokens · 29813 ms · 2026-05-21T16:52:50.461198+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

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

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

39 extracted references · 39 canonical work pages

  1. [1]

    https://pacechallenge.org/2023/ (2023), PACE 2023 challenge on computing low-width contraction sequences; exact and heuristic tracks

    PACE challenge 2023: Twinwidth. https://pacechallenge.org/2023/ (2023), PACE 2023 challenge on computing low-width contraction sequences; exact and heuristic tracks

  2. [2]

    Proceedings of the National Academy of Sciences112(10), 2942–2947 (2015)

    Aflalo, Y., Bronstein, A., Kimmel, R.: On convex relaxation of graph isomorphism. Proceedings of the National Academy of Sciences112(10), 2942–2947 (2015)

  3. [3]

    Journal of Combinatorial Optimization36(3), 965–1006 (2018)

    Aurora, P., Mehta, S.K.: The QAP-polytope and the graph isomorphism problem. Journal of Combinatorial Optimization36(3), 965–1006 (2018)

  4. [4]

    In: Pro- ceedings of the International Congress of Mathematicians: Rio de Janeiro 2018

    Babai, L.: Group, graphs, algorithms: the graph isomorphism problem. In: Pro- ceedings of the International Congress of Mathematicians: Rio de Janeiro 2018. pp. 3319–3336. World Scientific (2018)

  5. [5]

    In: Proceedings of the fourteenth annual ACM symposium on Theory of computing

    Babai, L., Grigoryev, D.Y., Mount, D.M.: Isomorphism of graphs with bounded eigenvalue multiplicity. In: Proceedings of the fourteenth annual ACM symposium on Theory of computing. pp. 310–324 (1982)

  6. [6]

    INFORMS Journal on Computing (2022)

    Besançon, M., Carderera, A., Pokutta, S.: FrankWolfe.jl: A high-performance and flexible toolbox for Frank-Wolfe algorithms and conditional gradients. INFORMS Journal on Computing (2022)

  7. [7]

    ACM Transactions on Mathematical Software (2025)

    Besançon, M., Designolle, S., Halbey, J., Hendrych, D., Kuzinowicz, D., Pokutta, S., Troppens, H., Herrmannsdoerfer, D.V., Wirth, E.: Improved algorithms and novel applications of the FrankWolfe.jl library. ACM Transactions on Mathematical Software (2025)

  8. [8]

    ACM Transactions on Mathe- matical Software49(2), 1–21 (2023)

    Bestuzheva, K., Besançon, M., Chen, W.K., Chmiela, A., Donkiewicz, T., van Doornmalen, J., Eifler, L., Gaul, O., Gamrath, G., Gleixner, A., et al.: Enabling research through the SCIP Optimization Suite 8.0. ACM Transactions on Mathe- matical Software49(2), 1–21 (2023)

  9. [9]

    Technical report, Optimization Online (February 2024), https://optimization-online.org/2024/02/the-scip-optimization-suite-9-0/

    Bolusani, S., Besançon, M., Bestuzheva, K., Chmiela, A., Dionísio, J., Donkiewicz, T., van Doornmalen, J., Eifler, L., Ghannam, M., Gleixner, A., Graczyk, C., Hal- big, K., Hedtke, I., Hoen, A., Hojny, C., van der Hulst, R., Kamp, D., Koch, T., Kofler, K., Lentz, J., Manns, J., Mexi, G., Mühmer, E., Pfetsch, M.E., Schlösser, F., Serrano, F., Shinano, Y., ...

  10. [10]

    MOS-SIAM Series on Optimiza- tion (1 2025)

    Braun, G., Carderera, A., Combettes, C.W., Hassani, H., Karbasi, A., Mokthari, A., Pokutta, S.: Conditional Gradient Methods. MOS-SIAM Series on Optimiza- tion (1 2025)

  11. [11]

    In: International conference on machine learning

    Braun, G., Pokutta, S., Zink, D.: Lazifying conditional gradient algorithms. In: International conference on machine learning. pp. 566–575. PMLR (2017) Graph isomorphism with mixed-integer convex optimization 19

  12. [12]

    Journal of Experimental Algorithmics (JEA)18, 3–1 (2013)

    Eppstein, D., Löffler, M., Strash, D.: Listing all maximal cliques in large sparse real-world graphs. Journal of Experimental Algorithmics (JEA)18, 3–1 (2013)

  13. [13]

    Naval Research Logistics Quarterly3, 95–110 (1956), https://onlinelibrary.wiley.com/doi/abs/10

    Frank, M., Wolfe, P.: An algorithm for quadratic programming. Naval Research Logistics Quarterly3, 95–110 (1956), https://onlinelibrary.wiley.com/doi/abs/10. 1002/nav.3800030109

  14. [14]

    Journal of the ACM (JACM)34(3), 596–615 (1987)

    Fredman, M.L., Tarjan, R.E.: Fibonacci heaps and their uses in improved network optimization algorithms. Journal of the ACM (JACM)34(3), 596–615 (1987)

  15. [15]

    Garber, D., Meshi, O.: Linear-memory and decomposition-invariant linearly con- vergentconditionalgradientalgorithmforstructuredpolytopes.Advancesinneural information processing systems29(2016)

  16. [16]

    Journal of Global Optimization67(4), 731– 757 (2017)

    Gleixner, A.M., Berthold, T., Müller, B., Weltge, S.: Three enhancements for optimization-based bound tightening. Journal of Global Optimization67(4), 731– 757 (2017)

  17. [17]

    Communications of the ACM63(10), 128–134 (10 2020)

    Grohe, M., Schweitzer, P.: The graph isomorphism problem. Communications of the ACM63(10), 128–134 (10 2020). https://doi.org/10.1145/3372123, https:// doi.org/10.1145/3372123

  18. [18]

    In: Forty-second International Conference on Machine Learning (2025)

    Hendrych, D., Besançon, M., Martínez-Rubio, D., Pokutta, S.: Secant line search for Frank-Wolfe algorithms. In: Forty-second International Conference on Machine Learning (2025)

  19. [19]

    In: 22nd International Symposium on Experimental Algorithms

    Hendrych, D., Besançon, M., Pokutta, S.: Solving the optimal experiment design problem with mixed-integer convex methods. In: Proceedings of the Symposium on Experimental Algorithms (2024). https://doi.org/10.4230/LIPIcs.SEA.2024.16

  20. [20]

    Mathematical Programming Computation 17, 731–757 (2025)

    Hendrych, D., Troppens, H., Besançon, M., Pokutta, S.: Convex mixed-integer op- timization with Frank-Wolfe methods. Mathematical Programming Computation 17, 731–757 (2025)

  21. [21]

    In: Complexity of Computer Computations: Proceedings of a symposium on the Complexity of Com- puter Computations, held March 20–22, 1972, at the IBM Thomas J

    Hopcroft, J.E., Tarjan, R.E.: Isomorphism of planar graphs. In: Complexity of Computer Computations: Proceedings of a symposium on the Complexity of Com- puter Computations, held March 20–22, 1972, at the IBM Thomas J. Watson Research Center, Yorktown Heights, New York, and sponsored by the Office of Naval Research, Mathematics Program, IBM World Trade Co...

  22. [22]

    Information and Inference: A Journal of the IMA7, 689–706 (2018)

    Klus, S., Sahai, T.: A spectral assignment approach for the graph isomorphism problem. Information and Inference: A Journal of the IMA7, 689–706 (2018). https://doi.org/10.1093/imaiai/iay001

  23. [23]

    Information and Inference: A Journal of the IMA14(2), iaaf011 (2025)

    Klus, S., Gelß, P.: Continuous optimization methods for the graph isomorphism problem. Information and Inference: A Journal of the IMA14(2), iaaf011 (2025)

  24. [24]

    Naval research logistics quarterly2(1-2), 83–97 (1955)

    Kuhn, H.W.: The Hungarian method for the assignment problem. Naval research logistics quarterly2(1-2), 83–97 (1955)

  25. [25]

    Advances in neural information processing systems28(2015)

    Lacoste-Julien, S., Jaggi, M.: On the global linear convergence of Frank-Wolfe op- timization variants. Advances in neural information processing systems28(2015)

  26. [26]

    USSR Computa- tional Mathematics And Mathematical Physics6, 1–50 (1966), https://www

    Levitin, E., Polyak, B.: Constrained minimization methods. USSR Computa- tional Mathematics And Mathematical Physics6, 1–50 (1966), https://www. sciencedirect.com/science/article/pii/0041555366901145

  27. [27]

    Journal of computer and system sciences25(1), 42–65 (1982)

    Luks, E.M.: Isomorphism of graphs of bounded valence can be tested in polynomial time. Journal of computer and system sciences25(1), 42–65 (1982)

  28. [28]

    arXiv preprint arXiv:2503.08921 (2025)

    Maskan, H., Hou, Y., Sra, S., Yurtsever, A.: Revisiting Frank-Wolfe for structured nonconvex optimization. arXiv preprint arXiv:2503.08921 (2025)

  29. [29]

    https: //pallini.di.uniroma1.it/Graphs.html (2025), collections of benchmark graphs (DI- MACS and dreadnaut formats); current version 2_9_1 (Sep 7, 2025) 20 W

    McKay, B., Piperno, A.: nauty and traces benchmark graph libraries. https: //pallini.di.uniroma1.it/Graphs.html (2025), collections of benchmark graphs (DI- MACS and dreadnaut formats); current version 2_9_1 (Sep 7, 2025) 20 W. Xiao et al

  30. [30]

    Journal of symbolic computation60, 94–112 (2014)

    McKay, B.D., Piperno, A.: Practical graph isomorphism, ii. Journal of symbolic computation60, 94–112 (2014)

  31. [31]

    McKay, B.D., et al.: Practical graph isomorphism (1981)

  32. [32]

    Journal of the society for industrial and applied mathematics5(1), 32–38 (1957)

    Munkres, J.: Algorithms for the assignment and transportation problems. Journal of the society for industrial and applied mathematics5(1), 32–38 (1957)

  33. [33]

    Mathematics of Operations Research44(1), 1–18 (2019)

    Pena,J.,Rodriguez,D.:PolytopeconditioningandlinearconvergenceoftheFrank- Wolfe algorithm. Mathematics of Operations Research44(1), 1–18 (2019)

  34. [34]

    arXiv preprint arXiv:2507.17545 (2025), https://arxiv.org/abs/2507.17545

    Pokutta, S.: Scalable DC optimization via adaptive Frank-Wolfe algorithms. arXiv preprint arXiv:2507.17545 (2025), https://arxiv.org/abs/2507.17545

  35. [35]

    Journal Of Graph Theory1, 339–363 (1977), https://onlinelibrary.wiley.com/doi/abs/10.1002/jgt.3190010410

    Read, R., Corneil, D.: The graph isomorphism disease. Journal Of Graph Theory1, 339–363 (1977), https://onlinelibrary.wiley.com/doi/abs/10.1002/jgt.3190010410

  36. [36]

    Journal of graph theory 1(4), 339–363 (1977)

    Read, R.C., Corneil, D.G.: The graph isomorphism disease. Journal of graph theory 1(4), 339–363 (1977)

  37. [37]

    arXiv preprint arXiv:2110.12650 (2021)

    Tsuji, K., Tanaka, K., Pokutta, S.: Sparser kernel herding with pairwise conditional gradients without swap steps. arXiv preprint arXiv:2110.12650 (2021)

  38. [38]

    Valls, V., Iosifidis, G., Tassiulas, L.: Birkhoff’s decomposition revisited: Sparse schedulingforhigh-speedcircuitswitches.IEEE/ACMTransactionsonNetworking 29(6), 2399–2412 (2021)

  39. [39]

    Journal of Soviet Mathematics29(4), 1426–1481 (1985)

    Zemlyachenko,V.N.,Korneenko,N.M.,Tyshkevich,R.I.:Graphisomorphismprob- lem. Journal of Soviet Mathematics29(4), 1426–1481 (1985)