pith. sign in

arxiv: 2508.18717 · v3 · submitted 2025-08-26 · 💻 cs.LG · cs.CV· cs.IT· math.AT· math.IT

Natural Image Classification via Quasi-Cyclic Graph Ensembles and Random-Bond Ising Models at the Nishimori Temperature

V.S. Usatyuk , D.A. Sapozhnikov , S.I. Egorov This is my paper

Pith reviewed 2026-05-18 21:51 UTC · model grok-4.3

classification 💻 cs.LG cs.CVcs.ITmath.ATmath.IT
keywords image classificationrandom bond Ising modelLDPC graphsNishimori temperaturegraph ensemblesfeature compressionquasi-cyclic graphsspectral methods
0
0 comments X

The pith

Mapping CNN features to Ising spins on quasi-cyclic graphs at the Nishimori temperature enables accurate compressed image classification.

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

High-dimensional features from networks such as MobileNetV2 are expensive to store and process for multi-class image tasks. The paper establishes that these features can be treated as spins in a random-bond Ising model built on a quasi-cyclic LDPC graph. Running the model at the Nishimori temperature, located where the Bethe-Hessian matrix has a vanishing eigenvalue, allows suppression of harmful graph substructures through a link to the Ihara-Bass zeta function. This produces classifiers that reduce feature dimension to 32 or 64 while reaching 98.7 percent top-1 accuracy on ImageNet-10 and 84.92 percent on ImageNet-100. Ensembles derived from the method also lower computational cost relative to the original network.

Core claim

Frozen MobileNetV2 features are interpreted as Ising spins on a sparse multi-edge quasi-cyclic LDPC graph to define a random-bond Ising model. The model is solved at its Nishimori temperature, identified by the vanishing of the smallest eigenvalue of the Bethe-Hessian matrix. A fast quadratic-Newton estimator computes this temperature in approximately nine Arnoldi iterations. The approach uses a spectral-topological correspondence based on the Ihara-Bass zeta function to suppress trapping sets that degrade accuracy. The resulting three-graph soft ensemble achieves 98.7% top-1 accuracy on ImageNet-10 and 84.92% on ImageNet-100, while a hard ensemble slightly exceeds MobileNetV2 accuracy at 2.

What carries the argument

The random-bond Ising model on a quasi-cyclic multi-edge LDPC graph at the Nishimori temperature, with the Bethe-Hessian eigenvalue condition and Ihara-Bass zeta function providing a link between trapping sets and topological defects for graph optimization.

If this is right

  • 98.7% top-1 accuracy is obtained on ImageNet-10 with features compressed to 32 dimensions.
  • 84.92% top-1 accuracy is obtained on ImageNet-100 with 64-dimensional features.
  • A hard ensemble improves accuracy by 0.10% over MobileNetV2 while using 2.67 times fewer FLOPs.
  • The soft ensemble reduces FLOPs by a factor of 29 relative to ResNet-50 with only a 1.09% accuracy drop.

Where Pith is reading between the lines

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

  • Applying the same graph embedding to features from other CNN architectures could yield comparable compression benefits across different backbones.
  • The use of LDPC-inspired graphs and statistical physics temperatures may transfer to improving generalization in other high-dimensional classification problems.
  • Further work could test whether the topological defect suppression improves performance under distribution shift or adversarial attacks.
  • Scaling the method to the full ImageNet dataset with 1000 classes would test the limits of the dimension reduction.

Load-bearing premise

High-dimensional MobileNetV2 features can be interpreted directly as Ising spins on a quasi-cyclic LDPC graph such that the Nishimori temperature defined by the vanishing Bethe-Hessian eigenvalue yields near-optimal classification without task-specific retraining.

What would settle it

If classification accuracy on ImageNet-10 falls substantially when the operating temperature is chosen away from the point where the Bethe-Hessian eigenvalue vanishes, while all other elements of the pipeline remain fixed, the central role of the Nishimori temperature would be called into question.

Figures

Figures reproduced from arXiv: 2508.18717 by D.A. Sapozhnikov, S.I. Egorov, V.S. Usatyuk.

Figure 1
Figure 1. Figure 1: In (a) and (b) we see cluster structures (community) from phase transaction described by [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (Left) Tanner graph corresponding to the parity-check matrix [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Graphical representation of TS(4, 2) (left), TS(4, 6) (center) and TS(9, 2) (right) [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Dependence of the smallest eigenvalue 𝜆min on the temperature parameter 𝛽. The red curve shows the polynomial part, the blue curve the tail, and the black dashed line is a quadratic approximation to the polynomial part. The coefficient of determination for the fit is 𝑅2 = 0.9998. 4.6. Quadratic–Newton estimation of the Nishimori temperature 𝛽𝑁 The root condition (1) can be solved efficiently because the ma… view at source ↗
Figure 5
Figure 5. Figure 5: Pipeline: MobileNetV2 feature extraction followed by graph spectral embedding. 5.1. Influence of Topological Invariants on Spectral Embedding For the purpose of illustration we consider six frequently occurring trapping sets 𝑇 𝑆(4, 2), 𝑇 𝑆(4, 6), 𝑇 𝑆(9, 2), 𝑇 𝑆(13, 6), 𝑇 𝑆(26, 20), 𝑇 𝑆(28, 22) [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Spectral embedding histogram with adjacency matrix of clusters (visualized by colors) under estimated Nishimori temperature 𝛽𝑁 . (left) MET QC-LDPC Torical graph. (right) QC Spherical graph. where Arb(·) denotes the output of the arbiter network. Instead of hard voting we can average the class-wise posterior probabilities supplied by each graph 𝑝𝑐 = 1 3 ∑︀3 𝑖=1 𝑝 (𝑖) 𝑐 , 𝑦̂︀soft = arg max𝑐 𝑝𝑐, with 𝑝 (𝑖) 𝑐… view at source ↗
Figure 7
Figure 7. Figure 7: (Left) Confusion matrix heatmap for ImageNet100. (Right) Per-class top-1 accuracy heatmap. [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Spectral embedding histogram for ’cock’ and ’hen’ classes [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
read the original abstract

Modern multi-class image classification uses high-dimensional CNN features that incur large memory and computational costs and obscure the data manifold's geometry. Existing graph-based spectral classifiers work on synthetic or binary tasks but degrade on natural images with many classes because feature manifolds have non-trivial topology. We introduce a physics-inspired pipeline where frozen MobileNetV2 features are interpreted as Ising spins on a sparse multi-edge type quasi-cyclic LDPC graph, defining a Random-Bond Ising Model (RBIM). The model is operated at its Nishimori temperature -- where the smallest eigenvalue of the Bethe-Hessian matrix vanishes. A spectral-topological correspondence links trapping sets in the Tanner graph to topological invariants via poles of the Ihara-Bass zeta function, enabling systematic suppression of harmful substructures that otherwise reduce top-1 accuracy by more than a factor of four. A fast quadratic-Newton estimator finds the Nishimori temperature in $\sim 9$ Arnoldi iterations, a sixfold speed-up over bisection. The resulting ensembles compress the original $1280$-dimensional MobileNetV2 representation to $32$ dimensions (ImageNet-10) or $64$ dimensions (ImageNet-100). We achieve $98.7\%$ top-1 accuracy on ImageNet-10 and $84.92\%$ on ImageNet-100 using a three-graph soft ensemble. Relative to MobileNetV2, our hard ensemble increases accuracy by $0.10\%$ while reducing FLOPs by a factor of $2.67$. Against ResNet-50, the soft ensemble drops only 1.09% accuracy yet cuts FLOPs by $29\times$. The novelty lies in (a) establishing a rigorous link between graph trapping sets and algebraic-topological defects, (b) an efficient Nishimori-temperature estimator, and (c) demonstrating topology-guided LDPC graph embedding for highly compressed classifiers.

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

2 major / 2 minor

Summary. The manuscript proposes interpreting frozen 1280-dimensional MobileNetV2 features as Ising spins on a sparse quasi-cyclic multi-edge LDPC graph to define a Random-Bond Ising Model (RBIM), which is then solved at the Nishimori temperature located by the vanishing of the smallest Bethe-Hessian eigenvalue. A three-graph soft ensemble is reported to reach 98.7% top-1 accuracy on ImageNet-10 and 84.92% on ImageNet-100 while compressing the representation to 32 or 64 dimensions; a hard ensemble is claimed to improve accuracy by 0.10% over MobileNetV2 with 2.67× fewer FLOPs. The work also introduces a quadratic-Newton estimator for the Nishimori temperature (∼9 Arnoldi iterations) and links trapping sets in the Tanner graph to topological invariants via poles of the Ihara-Bass zeta function.

Significance. If the central mapping and temperature choice are shown to be robust, the approach could supply a largely parameter-free route to extreme feature compression for multi-class image tasks by exploiting spectral properties of LDPC graphs and the Nishimori line. The algebraic-topological correspondence between trapping sets and zeta-function poles, together with the fast temperature estimator, would constitute a genuine contribution at the interface of statistical physics and graph-based machine learning.

major comments (2)
  1. [Methods (RBIM construction and temperature selection)] The central claim that the Nishimori temperature defined by the vanishing Bethe-Hessian eigenvalue yields near-optimal classification performance rests on an unverified transfer from the RBIM decoding literature to high-dimensional image-feature manifolds. No accuracy-versus-temperature curves or ablation over nearby temperatures are referenced in the methods or results sections to confirm that this operating point is a performance maximum rather than a convenient algebraic choice.
  2. [Experiments (ImageNet-10/100 results and ensemble tables)] The reported 98.7% / 84.92% accuracies and 2.67× FLOP reduction are given without error bars, without ablation on graph sparsity, edge multiplicity, or embedding dimension (32 vs. 64), and without comparison to the same ensemble operated at a temperature chosen by direct validation accuracy. These omissions make it impossible to assess whether the gains survive modest changes in the free parameters listed in the axiom ledger.
minor comments (2)
  1. [Temperature estimator] The abstract states a sixfold speedup for the quadratic-Newton estimator; the corresponding section should include the exact iteration counts and a direct wall-clock comparison against bisection on the same hardware.
  2. [Graph construction] Notation for the multi-edge-type quasi-cyclic LDPC construction and the precise mapping from 1280-dimensional features to spin variables should be made fully explicit, including any normalization or discretization steps.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive report. We address each major comment point by point below, indicating where revisions will be made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Methods (RBIM construction and temperature selection)] The central claim that the Nishimori temperature defined by the vanishing Bethe-Hessian eigenvalue yields near-optimal classification performance rests on an unverified transfer from the RBIM decoding literature to high-dimensional image-feature manifolds. No accuracy-versus-temperature curves or ablation over nearby temperatures are referenced in the methods or results sections to confirm that this operating point is a performance maximum rather than a convenient algebraic choice.

    Authors: The Nishimori temperature is selected because the vanishing of the smallest Bethe-Hessian eigenvalue marks the point where the model is on the Nishimori line, a property that guarantees the absence of a ferromagnetic phase transition and is known to be optimal for decoding in the RBIM literature. This algebraic criterion is not merely convenient but follows directly from the spectral properties of the quasi-cyclic LDPC graph and the mapping of features to spins. Nevertheless, we agree that explicit empirical confirmation is valuable. In the revised manuscript we will add accuracy-versus-temperature curves for both ImageNet-10 and ImageNet-100 together with ablations at nearby temperatures to demonstrate that the chosen operating point is indeed a performance maximum. revision: yes

  2. Referee: [Experiments (ImageNet-10/100 results and ensemble tables)] The reported 98.7% / 84.92% accuracies and 2.67× FLOP reduction are given without error bars, without ablation on graph sparsity, edge multiplicity, or embedding dimension (32 vs. 64), and without comparison to the same ensemble operated at a temperature chosen by direct validation accuracy. These omissions make it impossible to assess whether the gains survive modest changes in the free parameters listed in the axiom ledger.

    Authors: We acknowledge that the initial submission presented point estimates without error bars or exhaustive parameter ablations. The reported figures reflect the specific three-graph soft ensemble and the hard-ensemble comparison described in the text. To address the concern, the revised version will include (i) error bars obtained from multiple random seeds, (ii) additional ablation tables varying graph sparsity, edge multiplicity, and embedding dimension, and (iii) a direct comparison of the ensemble performance when the temperature is instead chosen by maximizing validation accuracy. These additions will allow readers to evaluate the robustness of the reported gains with respect to the free parameters. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation remains self-contained

full rationale

The paper defines the operating point via the algebraic condition that the smallest Bethe-Hessian eigenvalue vanishes, a standard property imported from RBIM/LDPC literature rather than fitted to classification accuracy. Graph construction (quasi-cyclic multi-edge LDPC) and embedding dimensions are chosen explicitly and then evaluated empirically on ImageNet-10/100; no equation or claim reduces the reported accuracies or FLOP reductions to a tautological renaming of the inputs. The quadratic-Newton estimator is derived from the eigenvalue condition itself, not from post-hoc performance tuning. External benchmarks (MobileNetV2, ResNet-50) provide independent comparison, confirming the central pipeline does not collapse to self-definition or fitted-input prediction.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 1 invented entities

The central claim rests on the transferability of Ising-model phase-transition concepts to CNN feature manifolds and on the validity of the spectral-topological correspondence for guiding graph design; both are modeling assumptions rather than derived results.

free parameters (2)
  • embedding dimension (32 for ImageNet-10, 64 for ImageNet-100)
    Chosen to achieve the stated compression while preserving accuracy; no first-principles derivation is given.
  • number of graphs in the ensemble (three)
    Selected to produce the reported soft-ensemble accuracy; appears tuned for the final metric.
axioms (2)
  • domain assumption Random-Bond Ising Model at Nishimori temperature is appropriate for classification on natural-image feature manifolds
    Invoked when the authors state that operating at the temperature where the smallest Bethe-Hessian eigenvalue vanishes yields good classifiers.
  • domain assumption Trapping sets in the Tanner graph correspond to poles of the Ihara-Bass zeta function and can be systematically suppressed
    Used to justify the topology-guided graph construction that is claimed to improve accuracy by more than a factor of four.
invented entities (1)
  • Quasi-cyclic multi-edge-type LDPC graph for feature embedding no independent evidence
    purpose: To place high-dimensional CNN features onto a sparse graph suitable for the random-bond Ising model
    New construction introduced for this pipeline; no independent evidence outside the reported experiments is provided.

pith-pipeline@v0.9.0 · 5908 in / 1861 out tokens · 64163 ms · 2026-05-18T21:51:15.756848+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

47 extracted references · 47 canonical work pages

  1. [1]

    In particular, modeling feature vectors as spins (nodes) on a sparse graph under a Gibbs distribu­ tion (a Random-Bond Ising Model) enables powerful clustering and classification

    INTRODUCTION Graph-based spectral methods have recently shown great promise for machine learning tasks. In particular, modeling feature vectors as spins (nodes) on a sparse graph under a Gibbs distribu­ tion (a Random-Bond Ising Model) enables powerful clustering and classification. For example, Nishimori and Bethe free-energy ideas have been used to tune...

    work page

  2. [2]

    Assign a binary spin variable 𝑠𝑖 ∈ {−1, +1} to each vertex 𝑖 ∈ 𝑉 and denote the spin configuration bys = ( 𝑠1,

    RANDOM BOND ISING MODELS Let 𝒢 = ( 𝑉, 𝐸) be an undirected graph with|𝑉 | = 𝑛. Assign a binary spin variable 𝑠𝑖 ∈ {−1, +1} to each vertex 𝑖 ∈ 𝑉 and denote the spin configuration bys = ( 𝑠1, . . . , 𝑠𝑛)⊤. For a symmetric coupling matrix𝐽 = (𝐽𝑖𝑗)𝑛 𝑖,𝑗=1, the Hamiltonian of the Random-Bond Ising Model is ℋ𝐽(s) = − ∑︁ 𝑖<𝑗 𝐽𝑖𝑗𝑠𝑖𝑠𝑗 = − 1 2 s⊤𝐽s. At inverse tempe...

    work page 2025

  3. [3]

    A Quasi­ Cyclic LDPC (QC-LDPC) codes provide a structured and hardware-friendly subclass defined by a quasi-cyclic parity-check matrix𝐻 [8]

    MULTI-EDGE TYPE QUASI-CYCLIC LDPC GRAPHS Low-density parity-check (LDPC) codes naturally define sparse bipartite graphs. A Quasi­ Cyclic LDPC (QC-LDPC) codes provide a structured and hardware-friendly subclass defined by a quasi-cyclic parity-check matrix𝐻 [8]. An (𝑁, 𝐾) QC-LDPC code consists of 𝑁 total codeword bits, with𝐾 information bits and𝑁 − 𝐾 parit...

    work page 2025

  4. [4]

    twisted” or “curved

    TOPOLOGICAL SIGNATURES OF TRAPPING SETS Trapping sets (TS) in Tanner graphs correspond to local topological defects that disrupt the decoding dynamics of LDPC codes, [16]. A trapping set𝑇 𝑆(𝑎, 𝑏), formed by cycles (block-cycle for QC-LDPC) or cycle (block-cycle) overlap, consists of𝑎 variable nodes and 𝑏 odd-degree check nodes, where the configuration pre...

    work page 2025

  5. [5]

    # neg.𝜆(𝐻1)

    EXPERIMENTAL EVALUATION All experiments use the same MobileNetV2 pre-trained backbone, Fig. 5. The feature ex­ tractor (CNN layers) is frozen, continuous features are binarized using sign, only the graph embedding (replace of feed-forward MLP) and final classifier are trained. We use following datasets: ImageNet10: 10-class subset (13000 images, 10 000 tr...

    work page 2025

  6. [6]

    A natural next step is to make all three components differen­ tiable and train them jointly

    FUTURE DIRECTIONS The present pipeline keeps the CNN backbone frozen, employs a similarity kernel and uses a MET-QC-LDPC adjacency. A natural next step is to make all three components differen­ tiable and train them jointly. One can back-propagate through the feature extractor𝑓𝜃, learn a Mahalanobis (or other) metric for the similarity kernel, and adjust ...

    work page

  7. [7]

    Trapping sets create Ihara–Bass zeta poles that appear as isolated eigenvalues of the non-backtracking operator, encoding Z/2-torsion, Betti numbers and bordism obstructions

    CONCLUSIONS We introduced a physics-inspired embedding that maps CNN features onto spins of a Random-Bond Ising Model defined on MET-QC-LDPC graphs. Trapping sets create Ihara–Bass zeta poles that appear as isolated eigenvalues of the non-backtracking operator, encoding Z/2-torsion, Betti numbers and bordism obstructions. A Nishimori-temperature estimator...

    work page 2025

  8. [8]

    Dall’Amico et al., J

    L. Dall’Amico et al., J. Stat. Mech. 2021, 093405 (2021). https://doi.org/10.1088/1742-5468/ac21d3

    work page doi:10.1088/1742-5468/ac21d3 2021

  9. [9]

    V. S. Usatyuk, D. A. Sapozhnikov, S. I. Egorov, Moscow Univ. Phys. Bull.79, S647–S665 (2025). https://doi.org/10.3103/S0027134924702102

    work page doi:10.3103/s0027134924702102 2025

  10. [10]

    Nishimori, Prog

    H. Nishimori, Prog. Theor. Phys.66, 1169–1181 (1981). https://doi.org/10.1143/PTP.66.1169 x–25 The 9th International Conference in Deep Learning in Computational Physics

    work page doi:10.1143/ptp.66.1169 1981

  11. [11]

    Hasenbusch et al., Phys

    M. Hasenbusch et al., Phys. Rev. E 77, 051115 (2008). https://doi.org/10.1103/PhysRevE.77.051115

    work page doi:10.1103/physreve.77.051115 2008

  12. [12]

    Fortunato, D

    S. Fortunato, D. Hric, Phys. Rep.659, 1–44 (2016). https://doi.org/10.1016/j.physrep.2016.09.002

    work page doi:10.1016/j.physrep.2016.09.002 2016

  13. [13]

    Biazzo, A

    I. Biazzo, A. Ramezanpour, Phys. Rev. E 89, 062137 (2014). https://doi.org/10.1103/PhysRevE.89.062137

    work page doi:10.1103/physreve.89.062137 2014

  14. [14]

    Saade et al., NeurIPS27 (2014)

    A. Saade et al., NeurIPS27 (2014). https://dl.acm.org/doi/10.5555/2968826.2968872

    work page doi:10.5555/2968826.2968872 2014

  15. [15]

    R. G. Gallager, IEEE Trans. Inf. Theory 8, 21–28 (1962). https://doi.org/10.1109/TIT.1962.1057683

    work page doi:10.1109/tit.1962.1057683 1962

  16. [16]

    A class of group-structured LDPC codes,

    R. M. Tanner et al., "A class of group-structured LDPC codes," ISCTA, 365–370 (2001). [Online]. Available: https://www.researchgate.net/publication/2370505_A_class_ of_group-structured_LDPC_codes

    work page arXiv 2001

  17. [17]

    M. P. C. Fossorier, IEEE Trans. Inf. Theory 50, 1788–1793 (2004). https://doi.org/10.1109/TIT.2004.831841

    work page doi:10.1109/tit.2004.831841 2004

  18. [18]

    R. M. Tanner et al., IEEE Trans. Inf. Theory 50, 2966–2984 (2004). https://doi.org/10.1109/TIT.2004.838370

    work page doi:10.1109/tit.2004.838370 2004

  19. [19]

    Multi-Edge Type LDPC Codes

    T. J. Richardson and R. L. Urbanke, "Multi-Edge Type LDPC Codes" IEEE ISIT talk (2002). [Online]. Available: http://wiiau4.free.fr/pdf/Multi-Edge%20Type%20LDPC%20Codes.pdf

    work page 2002

  20. [20]

    Divsalar et al., IEEE ISIT, 1622–1626 (2005)

    D. Divsalar et al., IEEE ISIT, 1622–1626 (2005). http://dx.doi.org/10.1109/ISIT.2005.1523619

    work page doi:10.1109/isit.2005.1523619 2005

  21. [21]

    Stoica et al., ACM SIGCOMM31, 149–160 (2001)

    I. Stoica et al., ACM SIGCOMM31, 149–160 (2001). https://doi.org/10.1145/964723.383071

    work page doi:10.1145/964723.383071 2001

  22. [22]

    Khalitov et al., Neural Netw

    R. Khalitov et al., Neural Netw. 152, 160–168 (2022). https://doi.org/10.1016/j.neunet.2022.04.014

    work page doi:10.1016/j.neunet.2022.04.014 2022

  23. [23]

    Vasi´ c et al., Allerton, 1–7 (2009)

    B. Vasi´ c et al., Allerton, 1–7 (2009). https://doi.org/10.1109/ALLERTON.2009.5394825

    work page doi:10.1109/allerton.2009.5394825 2009

  24. [24]

    Tian et al., IEEE ICC5, 3125–3129 (2003)

    T. Tian et al., IEEE ICC5, 3125–3129 (2003). https://doi.org/10.1109/ICC.2003.1203996

    work page doi:10.1109/icc.2003.1203996 2003

  25. [25]

    Tian et al., Trans

    T. Tian et al., Trans. Commun. 52, 1242–1247 (2004). https://doi.org/10.1109/TCOMM.2004.833048

    work page doi:10.1109/tcomm.2004.833048 2004

  26. [26]

    Bass, Int

    H. Bass, Int. J. Math.3, 717–797 (1992). https://doi.org/10.1142/S0129167X92000357

    work page doi:10.1142/s0129167x92000357 1992

  27. [27]

    Atiyah, Proc

    M. Atiyah, Proc. Cambridge Philos. Soc. 57, 200–208 (1961). https://doi.org/10.1017/S0305004100035064

    work page doi:10.1017/s0305004100035064 1961

  28. [28]

    Naitzat, A

    G. Naitzat, A. Zhitnikov, L. Lim, JMLR, 21(184), 1-40 (2020). [Online]. Available: https://jmlr.org/papers/v21/20-345.html

    work page 2020

  29. [29]

    Hodge laplacians on graphs.SIAM Review, 62(3):685–715, 2020

    Lek-Heng L., SIAM Rev.62, 685–715 (2020). https://doi.org/10.1137/18M1223101 x–26 Moscow University Physics Bulletin80(8), x (2025)

    work page doi:10.1137/18m1223101 2020

  30. [30]

    S. P. Novikov, Math. USSR-Izv.4(3), 479–505 (1970). https://doi.org/10.1070/IM1970v004n03ABEH000916

    work page doi:10.1070/im1970v004n03abeh000916 1970

  31. [31]

    de Tiliere, Electron

    B. de Tiliere, Electron. J. Probab.26, 1–86 (2021). https://doi.org/10.1214/21-EJP601

    work page doi:10.1214/21-ejp601 2021

  32. [32]

    Bolte and J

    J. Bolte and J. Harrison, J. Phys. A: Math. Gen. 36, 2747–2769 (2003). https://doi.org/10.1088/0305-4470/36/11/307

    work page doi:10.1088/0305-4470/36/11/307 2003

  33. [33]

    M. F. Atiyah and I. M. Singer, Bull. Am. Math. Soc. 69, 422–433 (1963). https://doi.org/10.1090/S0002-9904-1963-10957-X

    work page doi:10.1090/s0002-9904-1963-10957-x 1963

  34. [34]

    G. G. Kasparov, Math. USSR-Izv.16, 513–572 (1981). https://doi.org/10.1070/IM1981v016n03ABEH001320

    work page doi:10.1070/im1981v016n03abeh001320 1981

  35. [35]

    A. Dua, D. J. Williamson, J. Haah, and M. Cheng, Phys. Review B,99(24), 245135, (2019). https://doi.org/10.1103/PhysRevB.99.245135

    work page doi:10.1103/physrevb.99.245135 2019

  36. [36]

    Dodziuk, V

    J. Dodziuk, V. Mathai, J. Funct. Anal. 154, 359–378 (1998). https://doi.org/10.1006/jfan.1997.3205

    work page doi:10.1006/jfan.1997.3205 1998

  37. [37]

    D. G. Glynn, Eur. J. Comb.31, 1887–1891 (2010). https://doi.org/10.1016/j.ejc.2010.01.010

    work page doi:10.1016/j.ejc.2010.01.010 2010

  38. [38]

    Smarandache, P

    R. Smarandache, P. O. Vontobel, IEEE Trans. Inf. Theory 58, 585–607 (2012). https://doi.org/10.1109/TIT.2011.2173244

    work page doi:10.1109/tit.2011.2173244 2012

  39. [39]

    Smarandache, IEEE ISIT, 2059–2063 (2013)

    R. Smarandache, IEEE ISIT, 2059–2063 (2013). https://doi.org/10.1109/ISIT.2013.6620588

    work page doi:10.1109/isit.2013.6620588 2059

  40. [40]

    P. O. Vontobel, IEEE Trans. Inf. Theory 59, 1866–1901 (2013). https://doi.org/10.1109/TIT.2012.2227109

    work page doi:10.1109/tit.2012.2227109 1901

  41. [41]

    IEEE Conference on Computer Vision and Pattern Recognition , pages=

    M. Sandler et al., IEEE CVPR, 4510–4520 (2018). https://doi.org/10.1109/CVPR.2018.00474

    work page doi:10.1109/cvpr.2018.00474 2018

  42. [42]

    RBIM Natural Image Clustering

    V. S. Usatyuk, D. A. Sapozhnikov, "RBIM Natural Image Clustering" [Online]. Available: https://github.com/Lcrypto/Classical-and-Quantum-Topology-ML-toric-spherical/ tree/main/Topology_Optim_RBIM_Trapping_Sets

    work page

  43. [43]

    C. Di, R. Urbanke and T. Richardson, IEEE ISIT, pp. 50- (2001), https://10.1109/ISIT.2001.935913

    work page doi:10.1109/isit.2001.935913 2001

  44. [44]

    Shokrollahi, CSG,123, (2001), https://doi.org/10.1007/978-1-4613-0165-3_9

    M.A. Shokrollahi, CSG,123, (2001), https://doi.org/10.1007/978-1-4613-0165-3_9

    work page doi:10.1007/978-1-4613-0165-3_9 2001

  45. [45]

    Baldassia, et al., (PNAS), 113(48), E7655-E7662 (2016), https://www.pnas.org/doi/full/10.1073/pnas.1608103113

    C. Baldassia, et al., (PNAS), 113(48), E7655-E7662 (2016), https://www.pnas.org/doi/full/10.1073/pnas.1608103113

    work page doi:10.1073/pnas.1608103113 2016

  46. [46]

    He et al., Curr Biol

    Y. He et al., Curr Biol. , 34(20):4623-4638 (2024). 10.1016/j.cub.2024.08.037

    work page doi:10.1016/j.cub.2024.08.037 2024

  47. [47]

    E. I. Knudsen, Trends Neurosci.,41(11), 789–805 (2018), doi: 10.1016/j.tins.2018.06.006 x–27

    work page doi:10.1016/j.tins.2018.06.006 2018