Pith. sign in

REVIEW 4 major objections 6 minor 21 references

Solving Sudoku using oscillatory neural networks

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read An oscillatory neural network with one oscillator per Sudoku cell solves puzzles through phase synchronization and outperforms a Hopfield baseline on puzzles with up to 20 unknowns.

desk verdict A genuinely new oscillator mapping for Sudoku, but the printed dynamics invert the claimed repulsion between unknown cells, so the reported results cannot be attributed to the described mechanism. read the letter →

arxiv 2508.02250 v2 pith:4OXTAN7R submitted 2025-08-04 cond-mat.dis-nn cs.ET

Bram F. Haverkort , Federico Sbravati , Stefan Porfir , Aida Todri-Sanial This is my paper · ORCID
classification cond-mat.dis-nncs.ET MSC 34C1568T0790C27
keywords SudokuoscillatoryneuralnetworksKuramotomodelconstraintsatisfactionphaseencodingHopfieldcombinatorialoptimizationnon-vonNeumanncomputing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes encoding a Sudoku puzzle directly into the phase dynamics of a coupled-oscillator network: one oscillator per cell, with each digit represented by a phase evenly spaced on the unit circle. The row, column, and subgrid constraints are written into a single coupling matrix, and the unmodified Kuramoto equation is run until the phases lock, with the locked phases read out as the solved grid. The authors report that this mapping solves 9×9 puzzles with 5–20 unknown cells with high success, outperforming a Hopfield-network baseline, and still recovers some correct assignments at 25–30 unknowns before both methods fail at 35 unknowns. If correct, the result matters because it gives a training-free, analog route from a discrete constraint-satisfaction problem to hardware that computes with oscillator phases.

What carries the argument

The central object is the complex Kuramoto coupling matrix $J$ built from the three constraint subgraphs of a Sudoku: rows, columns, and subgrids. The phase mapping $\theta_{\mathrm{digit}} = 2\pi(n-1)/N_d$ places the $N_d$ digits on the unit circle, and the update law is $\dot{\theta}_i = -\sum_{j\in\Omega_i} J_{ij}\sin(\theta_j-\theta_i)$, where $\Omega_i$ is the set of row, column, and subgrid neighbors of cell $i$. The coupling values carry the argument: known-known edges use $J_{ij} = e^{i(\theta_i-\theta_j)}$, unknown-unknown edges use $J_{ij}=-1$ to push phases apart, and known-unknown edges are unidirectional $-1$ couplings that let known cells steer unknown ones. The machinery turns the discrete constraint-satisfaction puzzle into a continuous phase-locking problem.

What would settle it

Run the two-oscillator case of Eq. (9) with $J=-1$ and initial phases $\theta_1=0$, $\theta_2=\pi/2$; the difference $\delta=\theta_1-\theta_2$ obeys $d\delta/dt=-2\sin\delta$, so $\delta$ decays to $0$ rather than to $\pi$. If the implemented solver drives this pair into phase lock instead of anti-alignment, the printed $J=-1$ coupling does not produce the claimed repulsion, and the mechanism would have to rely on a different sign convention or readout.

Watch

Extended reading notes

Core claim

The central claim is that Sudoku can be solved by the natural dynamics of an oscillatory neural network, without logic operations or learning. Each cell is one oscillator; digits are phases via $\theta_{\mathrm{digit}} = 2\pi(n-1)/N_d$; and the weight matrix is assembled from three edge types: Hermitian couplings between known cells, repulsive $-1$ couplings between unknown cells, and directed $-1$ couplings from known to unknown cells. The total dynamics $\dot{\theta}_i = -\sum_{j\in\Omega_i} J_{ij}\sin(\theta_j-\theta_i)$ combine row, column, and subgrid constraints, and random initial phases evolve to a locked state that is read as a digit assignment. In the paper's simulations of 9×9 puzzles, the ONN achieves near-100% performance at 5–10 unknowns, stays above 90% at 15 unknowns, and still solves or partially solves some puzzles at 25–30 unknowns, while the Hopfield baseline falls to 0% at 25 unknowns. The authors state that the ONN mapping outperforms the established HNN methodology and supports a proof of concept for phase-based non-von Neumann constraint solving.

Load-bearing premise

The load-bearing premise is that the printed coupling signs and phase readout behave as intended, so unknown-unknown edges push phases apart, known-known edges anchor fixed phases, and known-unknown edges steer free cells toward the correct digit.

Editorial extensions

If this is right

  • Standard 9×9 Sudoku needs only 81 oscillators and an 81×81 coupling matrix, compared with 729 neurons and a 729×729 matrix in the Hopfield formulation.
  • Puzzles with up to about 20 unknown cells can be solved by free oscillator dynamics, with no backtracking, SAT search, or training.
  • The ONN continues to produce partially correct assignments at 25–30 unknowns in cases where the Hopfield baseline solves nothing, indicating that phase-based constraint encoding retains useful information under sparse clues.
  • The construction scales naturally from the 4×4 illustration to any $N_d$-valued Sudoku, since the phase mapping and constraint subgraphs are defined for a general number of digits.
  • For highly ambiguous puzzles, the authors propose augmenting the dynamics with a checker module that repeats solves according to the ratio of unknowns to cells; this is an explicit extension the paper offers rather than a demonstrated result.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the phase mapping is not binary, the same construction should transfer to other multi-valued constraint-satisfaction problems, such as graph coloring or Latin-square completion, where each node must take one of $N$ labels and constrained neighbors must differ.
  • A practical variant would rerun the Kuramoto dynamics with different random initial phases until a globally consistent readout appears, using the paper's proposed ratio $r = N_{\mathrm{unknowns}}/N$ to decide how many repeats are worth attempting.
  • In physical hardware, locked-phase spreads caused by oscillator noise and coupling mismatch would be the main failure mode; measuring the distribution of readout phases as a function of $r$ would test how robust the analog solver is beyond the idealized simulation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes an oscillatory neural network (ONN) approach to Sudoku in which each cell is represented by a single Kuramoto oscillator, with the puzzle constraints encoded through row, column, and subgrid coupling subgraphs that are combined into one weight matrix. The authors claim that the natural phase dynamics converge to a valid solution and report simulations comparing the ONN against a self-implemented Hopfield-network baseline, concluding that the ONN outperforms the baseline for puzzles with up to roughly 20 unknown cells.

Significance. If the proposed mapping worked as described, the contribution would be interesting as a low-complexity, training-free analog constraint-satisfaction scheme: the ONN needs O(N_d^2) oscillators versus O(N_d^3) for the Hopfield formulation, and the use of one oscillator per cell with a direct phase encoding is conceptually appealing. The comparison is not circular: no parameters are fitted to the benchmark data, and the known-cell couplings encode the supplied clues rather than derived outputs. However, the paper as written has load-bearing inconsistencies in the equations of motion, an undefined complex coupling term, no explicit phase-to-digit readout rule, and an under-specified baseline, so the reported empirical claims cannot currently be evaluated or attributed to the stated mechanism.

major comments (4)
  1. [III.B, Eqs. (1) and (9)] The dynamics printed in Eqs. (4)-(9) are a sign-flipped version of the Kuramoto equation in Eq. (1). Eq. (1) has dθ_i/dt = ω - Σ J_ij sin(θ_i - θ_j), while Eq. (9) uses dθ_i/dt = -Σ J_ij sin(θ_j - θ_i). For a bidirectional unknown-unknown edge with J = -1, this gives dδ/dt = -2 sin δ, where δ = θ_i - θ_j, so the phase difference decays to zero and the oscillators synchronize. This is exactly opposite to the text's claim in Section III.B that J = -1 maximizes the distance between phases and is repulsive. The load-bearing repulsion that is supposed to force unknown cells to different digits is therefore not present in the printed dynamics. The sign convention must be corrected and the actual implemented equations stated before the reported results can be attributed to the described mechanism.
  2. [III.B, Eq. (3)] Equation (3) defines J_ij^(known) as e^{i(θ_i-θ_j)}, which is complex-valued, but Eq. (9) is a real ODE in which J_ij multiplies sin(θ_j - θ_i). No real-part/imaginary-part convention, phase-shifted-coupling definition, or complex Kuramoto extension is specified. As written, the known-known edges, which are supposed to anchor the solution, do not have well-defined dynamics. This is a central issue because these edges are one of the three basic coupling cases on which the whole mapping rests.
  3. [III.A and IV] The paper never specifies how the continuous oscillator phases are converted into discrete Sudoku digits after integration. Eq. (2) maps digits to phases, but the final phases can take arbitrary values in [0, 2π) (up to drift), and no nearest-digit rule, modulo-2π convention, or boundary handling is given. Without this readout rule, the 'correctly solved' percentages in Fig. 7 and the success/failure classification in the benchmark cannot be interpreted.
  4. [III.C and Fig. 7] The Hopfield-network baseline is not described beyond citing Refs. [13] and [14]. The implementation details of the HNN, the exact role of the seven puzzles per difficulty level, how the 100 independent runs are distributed across puzzles, and the precise definition of 'performance' are all under-specified. No code or data availability statement is provided. Since the central quantitative claim is that the ONN outperforms the HNN, this comparison needs enough detail to be reproduced or independently checked.
minor comments (6)
  1. [III.B] The text refers to 'the general Kuramoto dynamics given in Eq. (II)', but no Eq. (II) exists; the intended reference appears to be Eq. (1).
  2. [Fig. 6] The y-axis of Fig. 6 lists both phase values in radians and Sudoku digits without explaining the correspondence; please add clear axis labels and a legend.
  3. [III.A] The sentence 'one cell is directly represented by one oscillator, or neuron' is redundant and should be reworded for clarity.
  4. [III.C] The paper states that seven distinct puzzles were used for every difficulty level but also reports 100 independent runs; please clarify whether all runs use the same seven puzzles and how the random masking is applied across runs.
  5. [II] The natural frequency ω_i^(0) is said to be neglected in simulations. If all natural frequencies are equal, this is a harmless constant drift that can be removed by a rotating frame, but that transformation should be stated explicitly.
  6. [IV and VI] The conclusion states that the ONN 'correctly solv[es] 57.57 % of the runs' at 20 unknowns, but the results section does not report that number explicitly; please make the numbers in the text and figures consistent.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the Sudoku mapping is constructed from constraints and evaluated by direct simulation; no fitted parameters or load-bearing self-citations.

full rationale

The paper's derivation chain is self-contained: the phase mapping in Eq. (2), the constraint subgraphs in Eqs. (3)-(6), and the combined dynamics in Eq. (9) are explicit constructions rather than fitted outputs. The reported performance is obtained by numerically simulating the defined Kuramoto dynamics and comparing against a self-implemented Hopfield-network baseline; no parameter is fitted to the reported accuracies, and no uniqueness theorem or prior work by the same authors is invoked to force the result. The self-citations in Sections I-II ([5], [7]) motivate the ONN paradigm but are not load-bearing for the novel Sudoku mapping, which is presented as a new contribution. The reader's concern about the sign/convention inconsistency between Eq. (1) and Eq. (9), and the claim that J=-1 is repulsive, is an internal correctness issue rather than circularity, because it does not make the output equivalent to the input. Therefore no circular step can be exhibited and the appropriate finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The reported mapping uses no fitted numerical constants; the simulation sizes (100 runs, 7 puzzles) are experimental choices. The load-bearing unproven elements are the sufficiency of pairwise couplings, the readout rule, and the well-posedness of the complex-coupled ODE. The proposed checker module threshold t and iteration count Nit are not used in the reported results.

assumptions (3)
  • ad hoc to paper Pairwise phase couplings between oscillators in the same row, column, or subgrid are sufficient to enforce the global Sudoku uniqueness constraints.
    Section III.B constructs subgraphs for rows, columns, and subgrids and combines them into one adjacency matrix; no proof or fixed-point analysis shows that these pairwise interactions enforce global row/column/box uniqueness.
  • domain assumption Final oscillator phases can be read out unambiguously as Sudoku digits.
    The paper states that oscillators 'lock in the right phases' but never specifies the readout rule, phase-wrapping handling, or digit assignment criterion, which is needed to convert a phase vector into a completed grid.
  • domain assumption The Kuramoto model with the stated sign convention governs the ONN hardware dynamics.
    Eq. (1) is assumed to model the system, but the complex coupling weights in Eq. (3) are placed into a real ODE without defining the complex multiplication. This assumption is internally inconsistent as printed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Solving Sudoku using oscillatory neural networks." pith.science (2026). https://pith.science/paper/4OXTAN7R

@misc{pith2026250802250,
  author       = {Pith},
  title        = {Pith review of: Solving Sudoku using oscillatory neural networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4OXTAN7R}},
  note         = {Machine review of arXiv:2508.02250}
}
read the original abstract

We explore the capabilities of physical computing with Oscillatory Neural Networks (ONN) to solve combinatorial optimization problems. To solve Sudokus with ONNs, we define a novel mapping strategy that utilizes the unique characteristics of the computation paradigm. The problem is encoded through a puzzle specific graph-embedding, which implements the constraints through different subgraphs. These subgraphs are then combined into a single adjacency matrix, which allows the natural dynamics of the phases of coupled oscillators to find a solution to the puzzle. We model the phase dynamics of the ONN by means of the Kuramoto differential equation. This novel approach is then compared to the well-established iterative method to solve Sudoku already used in binary Hopfield networks (HNN). Solving optimization problems typically requires a large amount of energy to solve on conventional hardware. Therefore, we are motivated to explore the mapping of Sudoku from a theoretical point of view to establish the validity of this approach. The simulation results show that the novel ONN mapping outperforms the established HNN methodology.

Figures

Figures reproduced from arXiv: 2508.02250 by the authors.

Figure 1
Figure 1. Completed 4 × 4 Sudoku grid. be harnessed to solve combinatorial problems in a parallel and low-energy manner. The existing literature has tackled solving Sudoku through Artificial Neural Networks, such as Hopfield Neural Networks (HNN) [13], [14] or Photonic Spiking Neural Networks [15], [16], making it possible to have comparisons with the performance of the method proposed in this paper, namely, solving Sudoku us… view at source ↗
Figure 3
Figure 3. Graph representation of the first Sudoku subgrid, dubbed the subgrid [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Graph representation of the first Sudoku row, dubbed the row and [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Graph representation of the first Sudoku column, dubbed the column [PITH_FULL_IMAGE:figures/full_fig_p003_5.png]
Figure 7
Figure 7. Figure 7: Comparison of ONN and HNN performance across different numbers [PITH_FULL_IMAGE:figures/full_fig_p004_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 19 canonical work pages

  1. [13]

    Searching for memories, Sudoku, implicit check-bits, and the iterative use of not-always-correct rapid neural computation

    J. J. Hopfield, “Searching for memories, Sudoku, implicit check- bits, and the iterative use of not-always-correct rapid neural computation,” Sep. 2006, arXiv:q-bio/0609006. [Online]. Available: http://arxiv.org/abs/q-bio/0609006

    work page Pith review arXiv 2006

  2. [14]

    Solving sudoku puzzles by using hopfield neural networks,

    V . Mladenov, P. Karampelas, C. Pavlatos, and E. Zirintsis, “Solving sudoku puzzles by using hopfield neural networks,” Proc. of ICACM , vol. 11, pp. 174–179, 2011

    work page 2011

  3. [1]

    Can programming be liberated from the von neumann style? a functional style and its algebra of programs,

    J. Backus, “Can programming be liberated from the von neumann style? a functional style and its algebra of programs,” Communications of the ACM , vol. 21, no. 8, pp. 613–641, Aug. 1978. [Online]. Available: https://doi.org/10.1145/359576.359579

    work page arXiv 1978

  4. [2]

    Breaking the von neumann bottleneck: architecture-level processing-in-memory technology,

    X. Zou, S. Xu, X. Chen et al., “Breaking the von neumann bottleneck: architecture-level processing-in-memory technology,” Science China Information Sciences , vol. 64, no. 160404, Apr. 2021. [Online]. Available: https://doi.org/10.1007/s11432-020-3227-1

    doi:10.1007/s11432-020-3227-1 2021

  5. [3]

    Neuromorphic Artificial Intelligence Systems

    D. Ivanov, A. Chezhegov, A. Grunin, M. Kiselev, and D. Larionov, “Neuromorphic artificial intelligence systems,” 2022. [Online]. Available: https://arxiv.org/abs/2205.13037

    work page Pith review arXiv 2022

  6. [4]

    Physics for neuromorphic computing,

    D. Markovi ´c, A. Mizrahi, D. Querlioz, and J. Grollier, “Physics for neuromorphic computing,” Nature Reviews Physics, vol. 2, pp. 499–510, 2020

    work page 2020

  7. [5]

    Neuromorphic computing based on oscillatory neural networks,

    A. Todri-Sanial, “Neuromorphic computing based on oscillatory neural networks,” in SOPHI.A Summit , Sophia Antipolis, France, Nov. 2020, hAL Id: lirmm-03022129. [Online]. Available: https: //hal-lirmm.ccsd.cnrs.fr/lirmm-03022129

    work page 2020

  8. [6]

    Quantum computing,

    A. Steane, “Quantum computing,” Reports on Progress in Physics , vol. 61, no. 2, p. 117, 1998

    work page 1998

Show all 21 references
  1. [7]

    Digital imple- mentation of oscillatory neural network for image recognition applica- tions,

    M. Abernot, T. Gil, M. Jim ´enez, J. N ´u˜nez, M. J. Avellido, B. Linares- Barranco, T. Gonos, T. Hardelin, and A. Todri-Sanial, “Digital imple- mentation of oscillatory neural network for image recognition applica- tions,” Frontiers in Neuroscience, vol. 15, p. 713054, 2021

    2021

  2. [8]

    Computing with oscillators,

    E. M. Izhikevich, “Computing with oscillators,” Neural Networks, vol. 5255, pp. 1–30, 2000

    2000

  3. [9]

    Sudoku as a constraint problem,

    H. Simonis, “Sudoku as a constraint problem,” in CP Workshop on modeling and reformulating Constraint Satisfaction Problems , vol. 12. Citeseer Sitges, Spain, 2005, pp. 13–27

    2005

  4. [10]

    Backtracking search optimization algorithm for numerical optimization problems,

    P. Civicioglu, “Backtracking search optimization algorithm for numerical optimization problems,” Applied Mathematics and computation , vol. 219, no. 15, pp. 8121–8144, 2013

    2013

  5. [11]

    Sudoku as a sat problem

    I. Lynce and J. Ouaknine, “Sudoku as a sat problem.” in AI&M. Citeseer, 2006

    2006

  6. [12]

    Artificial Kuramoto Oscillatory Neurons,

    T. Miyato, S. L ¨owe, A. Geiger, and M. Welling, “Artificial Kuramoto Oscillatory Neurons,” May 2025, arXiv:2410.13821 [cs] version: 3. [Online]. Available: http://arxiv.org/abs/2410.13821

    2025 arXiv

  7. [15]

    All-optical sudoku solver with photonic spiking neural network,

    S. Gao, S. Xiang, Z. Song, Y . Han, and Y . Hao, “All-optical sudoku solver with photonic spiking neural network,” Optics Communications, vol. 495, p. 127068, 2021

    2021

  8. [16]

    Efficient solution validation of constraint satisfaction problems on neuromorphic hardware: the case of Sudoku puzzles,

    R. Pignari, V . Fra, E. Macii, and G. Urgese, “Efficient solution validation of constraint satisfaction problems on neuromorphic hardware: the case of Sudoku puzzles,” IEEE Transactions on Artificial Intelligence , pp. 1–11, 2025. [Online]. Available: https: //ieeexplore.ieee....

    2025

  9. [17]

    Mori and Y

    H. Mori and Y . Kuramoto, Dissipative Structures and Chaos . Berlin, Heidelberg: Springer Berlin Heidelberg, 1998. [Online]. Available: https://link.springer.com/10.1007/978-3-642-80376-5

    1998 doi

  10. [18]

    A family of embedded runge-kutta formulae,

    J. R. Dormand and P. J. Prince, “A family of embedded runge-kutta formulae,” Journal of computational and applied mathematics , vol. 6, no. 1, pp. 19–26, 1980

    1980

  11. [19]

    Some practical runge-kutta formulas,

    L. F. Shampine, “Some practical runge-kutta formulas,” Mathematics of computation, vol. 46, no. 173, pp. 135–150, 1986

    1986

  12. [20]

    Weakly coupled oscillators,

    E. Izhikevich and Y . Kuramoto, “Weakly coupled oscillators,” in Encyclopedia of Mathematical Physics , J.-P. Franc ¸oise, G. L. Naber, and T. S. Tsun, Eds. Oxford: Academic Press, 2006, pp. 448–453. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ B0125...

    2006

  13. [21]

    3 million sudoku puzzles with ratings,

    D. G. Radcliffe, “3 million sudoku puzzles with ratings,” 2020. [Online]. Available: https://www.kaggle.com/dsv/1495975

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.