Pith. sign in

REVIEW 3 major objections 6 minor 48 references

Ontology Neural Network and ORTSF: A Framework for Topological Reasoning and Delay-Robust Control

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that semantic reasoning and delay-robust control can both be guaranteed by a single topological framework, with persistent-homology distances bounded by curvature losses and phase margins preserved up to a computable…

desk verdict The delay-robustness theorem is invalid on its own terms, and the rest of the paper is mostly known results wrapped in fitted constants. read the letter →

arxiv 2506.19277 v3 pith:GBAXPMJN submitted 2025-06-24 cs.RO cs.SYeess.SY

Jaehong Oh This is my paper
classification cs.ROcs.SYeess.SY MSC 55N3168T0793D09
keywords OntologyNeuralNetworkORTSFForman-Riccicurvaturepersistenthomologysemanticscenegraphsdelay-robustcontrolphasemargincognitiverobotics
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

The paper is trying to close the gap between high-level relational semantics and low-level control in robotics. It proposes ONN, which treats scene meaning as a dynamically evolving graph, and ORTSF, which turns reasoning traces into delay-compensated control commands. The central claim is that both processes carry formal guarantees: the persistent homology distance between scene graphs at different times is bounded by a curvature-related loss, and the closed-loop phase margin survives delays up to an explicit limit. If correct, this would give robots a single mathematically grounded pipeline that can explain its reasoning and still stay stable under latency.

What carries the argument

The argument is carried by a few interlocking objects. The ONN works in the graph 1-cochain space $X=\mathbb{R}^m$ with edge Laplacian $L_1=B_1B_1^T$, and its update is the projection-consensus operator $T=P_C\circ(I-\eta(\nabla L+L_1))$, where $P_C$ projects onto cycle constraints $Cx=\tau$. Forman-Ricci curvature gives each edge a discrete bending value that enters the filtration $f_t(e_{ij})=\alpha\|S_i(t)-S_j(t)\|^2+\beta|\mathrm{Ric}_F(e_{ij})|$; persistent homology converts that filtration into persistence diagrams compared by bottleneck distance. On the control side, ORTSF is the composition $F_{\text{ORTSF}}=T_{\text{control}}\circ T_{\text{delay}}\circ T_{\text{predict}}$, with the delay margin derived from a small-gain condition. The seven theorems tie these together through averaged-operator convergence, gauge-anchored connection Laplacian uniqueness, dynamic tracking bounds, exact penalty and hierarchical optimization, and the contextual topology stability bound.

What would settle it

For the paper's own numerical example ($\gamma=2.5$, $K_c=1.2$, $\|G\|_{H^\infty}=0.8$), compute the $H^\infty$ norm of $L(j\omega)=C(j\omega)G(j\omega)e^{-j\omega\Delta t}$; because $|e^{-j\omega\Delta t}|=1$, the norm is independent of $\Delta t$, so the small-gain condition either holds for every delay or fails before the claimed limit. Then check the Nyquist contour at $\Delta t=0.954$ s: if it encircles $-1$, the stated delay margin is false; if the true margin is the 55 ms reported in the validation text rather than the 0.954 s from Eq. (89), the claimed factor-2.6 improvement over baselines must be revised.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that relational semantic reasoning can be encoded as a constrained topological optimization with provable convergence and stability, and the same reasoning trace can be passed through a delay-aware operator with a proven phase-margin bound. The ONN defines a projection-consensus iteration over graph 1-cochains with cycle constraints $Cx=\tau$ and connection-Laplacian consistency, and its loss integrates consensus energy, connection consistency, Forman-Ricci curvature regularization, and context preservation. The paper claims that the persistent homology distance between time-separated scene graphs satisfies $d_{PH} \le L_c\kappa \sqrt{L_{\text{ricci-internal}}} + \eta(L_{\text{context}})$, with an integrated multi-scale bound in Theorem IV.15. Empirically, on TUM RGB-D sequences the PH distance decays as $d_{PH}(k)=0.127k^{-0.51}+0.003$ with $R^2=0.94$. For control, ORTSF composes predictor, delay compensator, and controller so that the effective phase margin stays above a safe threshold for delays satisfying $\Delta t < \ln(\gamma)/(K_c\|G\|_{H^\infty})$, and the simulations report a maintained phase margin of $28^\circ \pm 2^\circ$ against a $30^\circ$ design target.

Load-bearing premise

The delay-robustness claim rests on the unproven step in Theorem IV.8 that the small-gain condition implies the delay bound $\Delta t<\ln(\gamma)/(K_c\|G\|_{H^\infty})$, a step that is not derivable because the $H^\infty$ norm of $C(s)G(s)e^{-s\Delta t}$ is independent of $\Delta t$.

Editorial extensions

If this is right

  • Scene meaning, as measured by persistence diagrams, drifts by at most a computable amount whenever curvature variance and context loss stay bounded.
  • A robot running ONN+ORTSF can operate with a designed phase margin (reported $30^\circ\pm 2^\circ$) for delays up to the computed limit, while direct compensation and Smith predictors degrade earlier.
  • The unified loss objective lets one training procedure enforce cycle constraints, connection consistency, and temporal continuity simultaneously, so semantic reasoning and control no longer need separate pipelines.
  • The $O(k^{-1/2})$ convergence rate for PH distance gives a predictable schedule for when relational structure stabilizes, with the empirical fit (exponent $-0.51$) matching the theoretical rate.
  • The quantified robustness envelope for control, topological consistency, and contextual adaptation can be checked online to define safe operating boundaries before deployment.

Reading between the lines

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

  • If the PH stability bound holds, the same loss could be used as an online semantic-integrity monitor: raise an alert whenever $d_{PH}$ exceeds a threshold, independent of the control path.
  • If the delay bound can be proven rigorously, it would generalize beyond ONN: any linear time-invariant compensator fed by a Lipschitz reasoning trace would inherit the phase-margin guarantee, making the semantic-to-control bridge reusable.
  • The paper reports two values for the delay margin (0.954 s from the formula and 55 ms from validation); a direct experimental sweep of time delay would settle which one is the effective operating limit.
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

3 major / 6 minor

Summary. The manuscript proposes an architecture combining an Ontology Neural Network (ONN) and an Ontological Real-Time Semantic Fabric (ORTSF) to unify topological semantic reasoning with delay-robust control. The ONN models relational semantics through graph cochains, Forman-Ricci curvature, and persistent homology, with a loss function comprising consensus, connection, and contextual terms. The ORTSF is a delay-compensation layer that converts reasoning traces into control commands. The paper claims seven core theorems providing convergence, uniqueness, tracking bounds, delay margins, exact penalty properties, hierarchical optimization, and contextual topology stability. Empirical sections report a persistent homology decay d_PH(k) = 0.127 k^{-0.51} + 0.003 (R^2=0.94) on TUM RGB-D data and a delay margin of 52 ms. The central claims are that ONN+ORTSF provably preserves topological structure and provably preserves phase margin under bounded delays.

Significance. If the theorems were sound, the framework would constitute a rare unified bridge between semantic/topological reasoning and delay-tolerant control, with explicit quantitative guarantees. Some components are standard and correctly used in isolation: the projection-consensus convergence argument in Theorem IV.2 is a textbook application of averaged operator theory, and the bottleneck stability bound in the appendix correctly invokes the Cohen-Steiner-Edelsbrunner-Harer theorem. However, the load-bearing claims in the manuscript are not sound. The delay-margin theorem is mathematically invalid because the small-gain argument ignores that the H-infinity norm of a delayed transfer function is independent of the delay. The empirical validation contradicts the paper's own theoretical formula by a factor of 17. The claimed O(k^{-1/2}) convergence rate is contradicted by the paper's own appendix, which derives O(k^{-1/4}). The purported independent verification of topological stability is circular because the bound on persistent homology distance is expressed directly in terms of the loss terms being minimized. These failures remove the paper's central contributions as stated.

major comments (3)
  1. [Theorem IV.8 (Section IV.E), Eq. (29)] The small-gain stability proof is invalid. The proof sketch writes \|L(jω)\| = \|C(jω)G(jω)e^{-jωΔt}\| and then requires K_c\|G\|_{H∞} e^{-ω_c Δt} < 1, treating the exponential as a magnitude decay. In fact |e^{-jωΔt}| = 1 for every ω, so \|C(s)G(s)e^{-sΔt}\|_{H∞} = \|C(s)G(s)\|_{H∞}, independent of Δt. A small-gain condition of the form \|L\|_∞ < 1 cannot produce a delay-dependent bound such as Δt < ln(γ)/(K_c\|G\|_{H∞}). Any valid delay margin must come from phase-margin or Nyquist reasoning, which is not supplied. This invalidates the restated guarantees in Proposition VI.6, Eq. (71), Eq. (83), and Eqs. (88)-(91). The internal inconsistency is concrete: Eq. (89) with γ=2.5, K_c=1.2, and \|G\|_{H∞}=0.8 gives Δt_max = ln(2.5)/(1.2·0.8) = 0.954 s, while Eq. (90) claims the theory predicts 55 ms. A 17-fold discrepancy is not a rounding artifact, and it indicates that the central delay-robustness claim of the paper is unsupported.
  2. [Corollary IV.16 (Eq. 41) vs. Appendix "Convergence Rate Theorem"] The convergence rate claims are internally contradictory. The main text states in Corollary IV.16 and Section VII.C (Eq. 85) that E[d_PH(G_C(k), G^*_C)] = O(k^{-1/2}), and Section VII.C uses the empirical exponent -0.51 as confirmation. However, the appendix's complete proof of the Convergence Rate Theorem derives that the loss components converge as O(k^{-1}) (convex) and O(k^{-1/2}) (non-convex), then computes the PH distance bound as O(k^{-1/4}) and only then asserts, without derivation, that "empirical observations show O(k^{-1/2}) due to beneficial coupling." The manuscript therefore contains two mutually exclusive claims about the theoretical rate, one of which is labeled 'proven' in the main text while the appendix explicitly says the theoretical rate is O(k^{-1/4}). This is a load-bearing inconsistency because the paper's empirical validation is presented as matching the proven bound.
  3. [Section V.H and Eq. (98) (also Theorem IV.14, Eq. 37; Theorem IV.15, Eq. 40)] The claimed 'independent verification' of topological stability is circular. Section V.H states that the persistent homology distance is computed after convergence and 'provides independent verification that the contextual constraints preserve topological structure without circular dependency on the loss function.' But the bounds used in the paper, e.g., Theorem IV.14, Theorem IV.15, and the Discussion's Eq. (98), give d_PH(G_C(t), G_C(t+δ)) ≤ C_1 sqrt(L_ricci-internal) + C_2 L_context, where L_ricci-internal and L_context are exactly the loss terms being minimized. Moreover, the constants C_1 and C_2 are said to be 'empirically in the range [0.5, 2]' for typical graphs, so the resulting 'verification' reports that a quantity bounded by a fitted function of the minimized objective is small. This does not independently confirm topological stability; it only evaluates the fitted upper bound.
minor comments (6)
  1. [References] Reference [2] lists "N. Cadena" as the first author of SemanticFusion; the paper is by McCormac et al. The citation is inaccurate.
  2. [Appendix] The appendix contains the heading "Deleted Content: Rigorous Mathematical Foundations," which appears to be leftover from an editing process and should be removed or completed.
  3. [Figure 5] The stability-region caption uses the condition ρ_T + L_Q g(Δt) < 1, but ρ_T and L_Q are not defined anywhere in the text, making the figure's claimed visualization untraceable.
  4. [Eq. (27)] The summation index N in ∑_{k=0}^{N-1} e_k is never defined; the total observation time T appears in the bound but its relation to N is not specified.
  5. [Algorithm 2, Step 2] The displayed compensator C_delay(s) = (1+τ_D s)/(1+τ_D s/K_c) with τ_D = Δt is inconsistent with the lead-lag form in Eq. (69) and is likely a typesetting corruption.
  6. [Table I] The entry for Smith predictor read "Classical (¡ 15 ms)" using a nonstandard inverted punctuation; the intended comparison value should be typeset normally.

Circularity Check

3 steps flagged · score 7.0 of 10

Topological stability 'guarantee' is a definitional/fitted restatement of the minimized ONN losses, and the 'independent verification' is not independent; the delay-margin theorem is a separate invalid derivation.

  1. self definitional [Appendix, 'Relational Consistency Theorem (Complete Proof)', Eqs. (266)-(270); also Theorem IV.15 Eq. (40)]
    "Define the threshold function η(ϵ) such that: η(ϵ) = ϵ2 / [λ4(Σ3 k=0 αkC1,kκ)2 + λ4λcontext(Σ3 k=0 αkC2,k)2] ... < ϵ by construction of η(ϵ)."

    The theorem's conclusion is 'If total loss satisfies Ltotal < η(ϵ), then dPH < ϵ.' The proof constructs η from the exact expression to be bounded, so the implication holds by definition rather than by an independent derivation. Moreover, the constants C1,k and C2,k are not derived; Section VIII states they 'are empirically in the range [0.5, 2]'. Thus the advertised topological stability guarantee reduces to a fitted, definitional restatement of the loss terms rather than a first-principles prediction.

  2. fitted input called prediction [Section V.H 'Topological Stability Analysis (Independent Verification)', Def. V.4; Section VIII.A Eq. (98)]
    "StabilityPH(t, t+δ) = dB(Dk(ft), Dk(ft+δ)) ... This metric is computed after optimization convergence and provides independent verification that the contextual constraints preserve topological structure without circular dependency on the loss function. ... dPH(GC(t), GC(t + δ)) ≤ C1√Lricci-internal + C2Lcontext, where C1, C2 > 0 depend on graph density and label diversity and are empirically in the range [0.5, 2]."

    The 'independent verification' metric is bounded by Lricci-internal and Lcontext, which are exactly the loss components minimized in the ONN objective (Eqs. 49-53). A small measured PH distance is therefore the expected output of minimizing those same losses, not an independent confirmation. The constants C1, C2 are empirically fitted post hoc, so the bound is a fitted input presented as a validation of the theory.

1 more flagged steps
  1. fitted input called prediction [Section VII.C Eqs. (85)-(86) vs Appendix, 'Convergence Rate Theorem (Complete Proof)', Eqs. (287)-(304)]
    "dPH(k) = 0.127k−0.51 + 0.003, R2 = 0.94 (86) ... The empirical exponent −0.51 closely matches the theoretical rate −0.5 ... However, empirical observations show O(k−1/2) due to beneficial coupling between loss components that accelerates the topological convergence beyond the theoretical worst-case bound."

    The main text presents O(k^{-1/2}) as a proven convergence rate and uses the fitted power law dPH(k) = 0.127k^{-0.51}+0.003 as confirmation. But the appendix's own proof derives O(k^{-1/4}) and then restores O(k^{-1/2}) by appealing to 'empirical observations'. The theoretical rate is therefore not an independent prediction; it is read off the same fitted curve whose constants and exponent were estimated from the data, and the confirmed rate is weaker than claimed.

full rationale

The paper contains no significant self-citation chain: its references to Bottleneck Stability, Krasnosel'skii-Mann iteration, and gauge anchoring are external mathematical facts. However, the central topological-stability claim is circular in two substantive ways. First, the 'Relational Consistency Theorem' constructs the threshold function η(ϵ) precisely so that Ltotal < η(ϵ) implies dPH < ϵ; that is a tautology, not a guarantee. Second, the 'independent verification' in Section V.H is not independent: the bound on dPH is expressed through Lricci-internal and Lcontext, the very loss terms being minimized, with empirically fitted constants C1 and C2. The empirical convergence exponent is also used to confirm a main-text rate that the appendix itself weakens from O(k^{-1/2}) to O(k^{-1/4}). Separately, and not a circularity but a validity defect, the delay-margin proof in Theorem IV.8 is unsupported: |e^{-jωΔt}| = 1, so the H∞ small-gain condition cannot yield a delay-dependent bound, and the paper's own numbers contradict themselves (Eq. 89 gives 0.954 s while Eq. 90 reports 55 ms). Together these issues mean the claimed first-principles predictions are in large part fitted or definitional restatements of the framework's own loss and design choices.

Assumptions & free parameters 4 free parameters · 7 assumptions · 0 invented entities

The central claims rest on a large set of assumptions about convexity, bounded dynamics, gauge anchoring, nominal stability, and sub-Gaussian noise, many of which are not verified. The PH bound also relies on fitted constants. No new physical entities are introduced.

free parameters (4)
  • alpha, beta (filtration weights) = not specified; chosen by hand
    Weights in the filtration function ft(eij) = alpha||Si-Sj||^2 + beta|RicF(eij)| that define what 'topological coherence' means; the theorems depend on these choices.
  • C1, C2 (PH bound constants) = empirically [0.5, 2] for graphs of 10-50 nodes
    Constants in Eq. (98) are said to be 'empirically in the range [0.5,2]', making the theoretical bound numerically plausible rather than derived.
  • PH decay fit coefficients = 0.127 and -0.51
    Eq. (86) fits dPH(k) = 0.127 k^{-0.51} + 0.003 with R^2=0.94; this is a fitted curve presented as validation of the O(k^{-1/2}) claim.
  • Kc compensator gain = 1.2 in the example
    The delay margin Delta tmax = ln(gamma)/(Kc||G||_Hinf) depends directly on Kc, which is chosen (1.2) to obtain a favorable delay bound.
assumptions (7)
  • domain assumption Task energy L is L-Lipschitz differentiable and mu-strongly convex
    Required for Theorem 1's linear convergence rate; not justified for semantic losses in general.
  • domain assumption Graph G is connected and gauge anchoring matrix A has full row rank
    Needed for Theorem 2's uniqueness result on the connection Laplacian.
  • domain assumption Transition maps are uniformly Lipschitz (L_phi < inf) and dynamics are bounded (||cdot|| <= M)
    Assumed in Theorem 3 to obtain bounded tracking error; no mechanism is given to ensure these bounds.
  • domain assumption Nominal system C(s)G(s) is stable with gain margin gamma > 1 and ||C(jw)|| <= Kc
    Used in Theorem 4's small gain argument; the step from small gain to the delay bound is not shown.
  • standard math Bottleneck stability theorem of Cohen-Steiner, Edelsbrunner, and Harer
    Invoked in Patch-A to bound PH distance by the L-infinity change in the filtration function.
  • ad hoc to paper Sub-Gaussian filtration perturbations with parameter sigma^2
    Assumed in the appendix to obtain an exponential concentration bound; no empirical support or physical justification is given.
  • domain assumption The filtration function (6) with chosen alpha, beta encodes 'relational semantics' such that PH distance is a meaningful semantic metric
    The paper equates semantic meaning with edge weights derived from semantic tensors and curvature, which is a modeling choice that is never validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ontology Neural Network and ORTSF: A Framework for Topological Reasoning and Delay-Robust Control." pith.science (2026). https://pith.science/paper/GBAXPMJN

@misc{pith2026250619277,
  author       = {Pith},
  title        = {Pith review of: Ontology Neural Network and ORTSF: A Framework for Topological Reasoning and Delay-Robust Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GBAXPMJN}},
  note         = {Machine review of arXiv:2506.19277}
}
read the original abstract

The advancement of autonomous robotic systems has led to impressive capabilities in perception, localization, mapping, and control. Yet, a fundamental gap remains: existing frameworks excel at geometric reasoning and dynamic stability but fall short in representing and preserving relational semantics, contextual reasoning, and cognitive transparency essential for collaboration in dynamic, human-centric environments. This paper introduces a unified architecture comprising the Ontology Neural Network (ONN) and the Ontological Real-Time Semantic Fabric (ORTSF) to address this gap. The ONN formalizes relational semantic reasoning as a dynamic topological process. By embedding Forman-Ricci curvature, persistent homology, and semantic tensor structures within a unified loss formulation, ONN ensures that relational integrity and topological coherence are preserved as scenes evolve over time. The ORTSF transforms reasoning traces into actionable control commands while compensating for system delays. It integrates predictive and delay-aware operators that ensure phase margin preservation and continuity of control signals, even under significant latency conditions. Empirical studies demonstrate the ONN + ORTSF framework's ability to unify semantic cognition and robust control, providing a mathematically principled and practically viable solution for cognitive robotics.

Figures

Figures reproduced from arXiv: 2506.19277 by the authors.

Figure 1
Figure 1. ONN + Contextual Topology Framework Overview. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Cycle Constraint Projection and Convergence Anal [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Piecewise-Static Cochain Bundle Evolution. The time [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (7 more)
Figure 3
Figure 3. Figure 3: Connection-Consistency Layer (CCL) Resolution. A [PITH_FULL_IMAGE:figures/full_fig_p006_3.png]
Figure 5
Figure 5. Figure 5: Delay-Stability Condition Analysis. The coordinate [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Persistent Homology distance decay validation. The [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Phase margin preservation vs delay. ORTSF (blue) [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Initial scene graph topology (iteration 0): sparse [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Converged scene graph topology (iteration 500): dense [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 1
Figure 1. Figure 1: PH distance decay (final value [PITH_FULL_IMAGE:figures/full_fig_p015_1.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 33 canonical work pages

  1. [2]

    Semanticfu- sion: Dense 3d semantic mapping with convolutional neural networks,

    N. Cadena, J. McCormac, A. Davison, and S. Leutenegger, “Semanticfu- sion: Dense 3d semantic mapping with convolutional neural networks,” in IEEE International Conference on Robotics and Automation (ICRA) , 2017, pp. 4628–4635

    work page 2017

  2. [1]

    Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras,

    R. Mur-Artal and J. D. Tardos, “Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras,” IEEE Transactions on Robotics, vol. 33, no. 5, pp. 1255–1262, 2017

    work page 2017

  3. [3]

    Ricci curvature regularization for graph neural networks,

    T. Sandhu, S. Georgiou, E. Tzortzis, and S. Theodoridis, “Ricci curvature regularization for graph neural networks,” in Proceedings of the 28th ACM International Conference on Information and Knowledge Man- agement (CIKM), 2019, pp. 1743–1752

    work page 2019

  4. [4]

    Topology and data,

    G. Carlsson, “Topology and data,” Bulletin of the American Mathemat- ical Society, vol. 46, no. 2, pp. 255–308, 2009

    2009

  5. [5]

    Con- strained model predictive control: Stability and optimality,

    D. Q. Mayne, J. B. Rawlings, C. V . Rao, and P. O. M. Scokaert, “Con- strained model predictive control: Stability and optimality,” Automatica, vol. 36, no. 6, pp. 789–814, 2000

    work page 2000

  6. [6]

    Robosherlock: Unstructured informa- tion processing for robot perception,

    F. Balint-Benczedi and M. Beetz, “Robosherlock: Unstructured informa- tion processing for robot perception,” in IEEE International Conference on Robotics and Automation (ICRA) , 2015, pp. 1549–1556

    work page 2015

  7. [7]

    Bochner’s method for cell complexes and combinatorial ricci curvature,

    R. Forman, “Bochner’s method for cell complexes and combinatorial ricci curvature,” Discrete & Computational Geometry , vol. 29, no. 3, pp. 323–374, 2003

    work page 2003

  8. [8]

    Ricci curvature of the internet topology,

    C.-C. Ni, Y .-Y . Lin, J. Gao, X. D. Gu, and E. Saucan, “Ricci curvature of the internet topology,” INFOCOM 2015, pp. 2758–2766, 2015

    work page 2015

Show all 48 references
  1. [9]

    Edelsbrunner and J

    H. Edelsbrunner and J. Harer, Computational Topology: An Introduction. American Mathematical Society, 2010

    2010

  2. [10]

    H. H. Bauschke and P. L. Combettes, Convex Analysis and Monotone Operator Theory in Hilbert Spaces , 2nd ed. Springer, 2017

    2017

  3. [11]

    R. T. Rockafellar, Convex Analysis. Princeton University Press, 1970

    1970

  4. [12]

    Absil, R

    P.-A. Absil, R. Mahony, and R. Sepulchre, Optimization Algorithms on Matrix Manifolds. Princeton University Press, 2008

    2008

  5. [13]

    Godsil and G

    C. Godsil and G. Royle, Algebraic Graph Theory. Springer, 2001, vol. 207

    2001

  6. [14]

    Stability of persis- tence diagrams,

    D. Cohen-Steiner, H. Edelsbrunner, and J. Harer, “Stability of persis- tence diagrams,” Discrete & Computational Geometry , vol. 37, no. 1, pp. 103–120, 2007

    2007

  7. [15]

    The structure and stability of persistence modules,

    F. Chazal, V . de Silva, M. Glisse, and S. Oudot, “The structure and stability of persistence modules,” 2016

    2016

  8. [16]

    Appli- cations of persistent homology to machine learning,

    M. Vejdemo-Johansson, W. Chacholski, and A. Zomorodian, “Appli- cations of persistent homology to machine learning,” arXiv preprint arXiv:1311.5668, 2013

    2013 arXiv

  9. [17]

    Ricci curvature of markov chains on metric spaces,

    Y . Ollivier, “Ricci curvature of markov chains on metric spaces,”Journal of Functional Analysis , vol. 256, no. 3, pp. 810–864, 2009

    2009

  10. [18]

    Closer control of loops with dead time,

    O. J. M. Smith, “Closer control of loops with dead time,” ISA Journal, vol. 6, no. 2, pp. 28–33, 1959

    1959

  11. [19]

    Geometric deep learning: going beyond euclidean data,

    M. M. Bronstein, J. Bruna, Y . LeCun, A. Szlam, and P. Vandergheynst, “Geometric deep learning: going beyond euclidean data,” IEEE Signal Processing Magazine, vol. 34, no. 4, pp. 18–42, 2017

    2017

  12. [20]

    Semi-supervised classification with graph convolutional networks,

    T. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” International Conference on Learning Repre- sentations (ICLR), 2017, arXiv:1609.02907

    2017 arXiv

  13. [21]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in Neural Information Processing Systems (NeurIPS) , vol. 30, 2017

    2017

  14. [22]

    Curvature graph network,

    Z. Zhang, L. Zhao, J. Zhu, and S. Ji, “Curvature graph network,” NeurIPS, vol. 33, pp. 4863–4875, 2020

    2020

  15. [23]

    Topological methods in machine learning: Theory and applications,

    C. Chen and X. Gu, “Topological methods in machine learning: Theory and applications,” IEEE Signal Processing Magazine, vol. 32, no. 1, pp. 134–142, 2015

    2015

  16. [24]

    Distributed optimization and statistical learning via the alternating direction method of multipliers,

    S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed optimization and statistical learning via the alternating direction method of multipliers,” Foundations and Trends in Machine Learning , vol. 3, no. 1, pp. 1–122, 2011

    2011

  17. [25]

    The entropy formula for the ricci flow and its geometric applications,

    G. Perelman, “The entropy formula for the ricci flow and its geometric applications,” arXiv preprint math/0211159 , 2002

    2002 arXiv

  18. [26]

    Emergence of scaling in random net- works,

    A.-L. Barab ´asi and R. Albert, “Emergence of scaling in random net- works,” Science, vol. 286, no. 5439, pp. 509–512, 1999

    1999

  19. [27]

    The structure and function of complex networks,

    M. E. J. Newman, “The structure and function of complex networks,” SIAM Review, vol. 45, no. 2, pp. 167–256, 2003

    2003

  20. [28]

    The gudhi library: Simplicial complexes and persistent homology,

    J.-D. Boissonnat and C. Maria, “The gudhi library: Simplicial complexes and persistent homology,” in International Congress on Mathematical Software, 2014, pp. 167–174

    2014

  21. [29]

    Mesbahi and M

    M. Mesbahi and M. Egerstedt, Graph Theoretic Methods in Multiagent Networks. Princeton University Press, 2010

    2010

  22. [30]

    Learning deep generative models of graphs,

    Y . Li, O. Vinyals, C. Dyer, R. Pascanu, and P. Battaglia, “Learning deep generative models of graphs,” in International Conference on Machine Learning (ICML), 2018

    2018

  23. [31]

    Inductive representation learning on large graphs,

    W. L. Hamilton, R. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” in NeurIPS, vol. 30, 2017

    2017

  24. [32]

    Wasserstein generative adver- sarial networks,

    M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein generative adver- sarial networks,” in ICML, 2017

    2017

  25. [33]

    Measuring and regularizing networks in function space,

    T. Chen, E. Fertig, and C. Guestrin, “Measuring and regularizing networks in function space,” in ICLR, 2019

    2019

  26. [34]

    J. R. Munkres, Elements of Algebraic Topology. Addison-Wesley, 1984

    1984

  27. [35]

    Attention models in graph neural networks: A review,

    J. Lee, I. Lee, and J. Kang, “Attention models in graph neural networks: A review,” Applied Sciences, vol. 9, no. 22, p. 4708, 2019

    2019

  28. [36]

    How powerful are graph neural networks?

    K. Xu, W. Hu, J. Leskovec, and S. Jegelka, “How powerful are graph neural networks?” International Conference on Learning Representa- tions (ICLR), 2019, arXiv:1810.00826

    2019 arXiv

  29. [37]

    Geometric deep learning: Grids, groups, graphs, geodesics, and gauges,

    M. M. Bronstein, J. Bruna, T. Cohen, and P. Veliˇckovi´c, “Geometric deep learning: Grids, groups, graphs, geodesics, and gauges,” arXiv preprint arXiv:2104.13478, 2021

    2021 arXiv

  30. [38]

    Graph neural networks: A review of methods and applications,

    J. Zhou, G. Cui, S. Hu, Z. Zhang, C. Yang, Z. Liu, L. Wang, C. Li, and M. Sun, “Graph neural networks: A review of methods and applications,” in AI Open, vol. 1, 2020, pp. 57–81

    2020

  31. [39]

    Gated graph sequence neural networks,

    Y . Li, D. Tarlow, M. Brockschmidt, and R. Zemel, “Gated graph sequence neural networks,” International Conference on Learning Rep- resentations (ICLR), 2016, arXiv:1511.05493

    2016 arXiv

  32. [40]

    Barcodes: The persistent topology of data,

    R. Ghrist, “Barcodes: The persistent topology of data,” Bulletin of the American Mathematical Society , vol. 45, no. 1, pp. 61–75, 2008

    2008

  33. [41]

    A review of delay compensation techniques for con- trol systems,

    R. C. Quiroga, “A review of delay compensation techniques for con- trol systems,” Journal of Control, Automation and Electrical Systems , vol. 31, pp. 1039–1055, 2020

    2020

  34. [42]

    Graph signal processing for graph neural networks,

    F. Gama, E. Isufi, and A. Ribeiro, “Graph signal processing for graph neural networks,” inICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2020, pp. 8509– 8513

    2020

  35. [43]

    An introduction to topological data analysis: Fundamental and practical aspects for data scientists,

    F. Chazal and B. Michel, “An introduction to topological data analysis: Fundamental and practical aspects for data scientists,” arXiv preprint arXiv:1710.04019, 2017

    2017 arXiv

  36. [44]

    Graph neural tangent kernel: Fusing graph neural networks and graph kernels,

    S. S. Du, W. Hu, M. Belkin, J. D. Lee, L. Wang, and X. Zhu, “Graph neural tangent kernel: Fusing graph neural networks and graph kernels,” NeurIPS, vol. 32, 2019

    2019

  37. [45]

    Hatcher, Algebraic Topology

    A. Hatcher, Algebraic Topology. Cambridge University Press, 2002

    2002

  38. [46]

    Topics in optimal transportation,

    C. Villani, “Topics in optimal transportation,” Graduate Studies in Mathematics, vol. 58, 2003

    2003

  39. [47]

    Ollivier’s ricci curvature, local clustering and curvature-dimension inequalities on graphs,

    J. Jost and S. Liu, “Ollivier’s ricci curvature, local clustering and curvature-dimension inequalities on graphs,” Discrete & Computational Geometry, vol. 51, no. 2, pp. 300–322, 2014

    2014

  40. [48]

    J. M. Lee, Introduction to Smooth Manifolds , 2nd ed. Springer, 2013

    2013

Pith tools

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