Pith. sign in

REVIEW 4 major objections 5 minor 45 references

Is Noise Conditioning Necessary? A Unified Theory of Unconditional Graph Diffusion Models

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Graph diffusion models can be built without timestep inputs on large graphs.

desk verdict A useful empirical study and a plausible conjecture, but the headline O(T/M) bound rests on a geometric-series error that should be fixed before the theory is cited. read the letter →

arxiv 2505.22935 v2 pith:MP3MTBQZ submitted 2025-05-28 cs.LG

Jipeng Li , Yanning Shen This is my paper
classification cs.LG
keywords graphdiffusionmodelsnoiseconditioningtimestep-freedenoisingBernoulliedgeflipsposteriorconcentrationerrorpropagationgenerationcoupledstructure-attribute
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 tries to establish that the timestep input used by graph diffusion models is not actually needed: on large graphs, the noisy adjacency structure itself tells the denoiser how much corruption is present. For Bernoulli edge-flip noise it proves the flip-rate posterior concentrates with variance $O(M^{-1})$, where $M$ is the number of potential edges, so the noise level is inferable from the graph alone. It then proves that dropping the explicit timestep changes a single denoising step by $O(M^{-1})$ and that $T$ reverse steps accumulate error at most $O(T/M)$. Empirically, unconditional GDSS and DiGress variants match or beat their conditioned counterparts on molecular and social-network benchmarks while using 4–6% fewer parameters and 8–10% less compute. If the argument holds, simpler graph diffusion models can be built without noise-conditioning machinery.

What carries the argument

The argument rides on three named bounds plus their coupled counterparts. Edge-Flip Posterior Concentration (EFPC) shows that the posterior of the Bernoulli flip rate given the noisy graph has variance $O(M^{-1})$, with leading constant $\beta_t(1-\beta_t)/M$; this is the statistical fact that makes the noise level identifiable. Edge-Target Deviation Bound (ETDB) converts that concentration into a bound on how much the optimal denoising target moves when the explicit timestep is marginalized out, using the Lipschitz regularity of the conditional target. Multi-Step Denoising Error Propagation (MDEP) then proves, by the recursion $B_i \le \delta_i + L_{\max} B_{i+1}$, that the per-step errors add only linearly rather than exponentially, provided the denoiser is near-non-expansive. For scale-free degree distributions the EFPC rate is modified to $\tilde{O}(M^{-(\alpha-2)/(\alpha-1)})$; for coupled Gaussian structure-attribute noise, the same three-step structure is restated with total dimensionality $D$ in place of $M$.

What would settle it

Measure the Lipschitz constant $L$ of a trained unconditional denoiser and the per-step deviation $\delta$ on graphs of increasing $M$, then compare observed cumulative error to the predicted $O(T/M)$. If the empirical exponent against $M$ deviates from $-1$, or if cumulative error grows with $T$ faster than linearly for fixed $M$, the central bound is false for that model. In particular, with $L_{\max}=1+\eta$ for fixed $\eta>0$, the factor $(L_{\max}^T-1)/(L_{\max}-1)$ is not $O(T)$ at large $T$, so a single numerical check at $T=1000$ would settle the prefactor claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that a graph denoiser does not need the timestep $t$ supplied as an input: for corruption processes based on Bernoulli edge flips, the noisy adjacency matrix $\tilde A_t$ already pins down the noise level. Concretely, it proves that the posterior variance of the flip rate $\beta_t$ is $O(M^{-1})$, that replacing the conditional target $\mu^{\rm cond}_t$ with the unconditional target $\bar\mu_t$ costs only $O(M^{-1})$ in squared Frobenius norm, and that over $T$ reverse steps the accumulated reconstruction error is $O(T/M)$, where $M=\binom{n}{2}$. It then extends the same three-part structure to a coupled Gaussian structure-attribute noise model, replacing $M$ by the total dimensionality $D=M+n\,d_f$, and reports empirical scaling exponents on synthetic graphs that match these rates. On QM9 and soc-Epinions1, the paper finds that unconditional GDSS and DiGress variants match or beat conditioned baselines while saving parameters and time.

Load-bearing premise

The whole argument leans on the trained denoiser already making errors that shrink like one over the number of possible edges, and on that denoiser being nearly non-expansive so that small errors do not amplify—if either fails, the linear error bound collapses.

Editorial extensions

If this is right

  • Practitioners can remove timestep embeddings from GDSS-style and DiGress-style graph diffusion models without sacrificing generation quality on graphs large enough to make the posterior concentrate.
  • Parameter savings of 4–6% and per-epoch time savings of 8–10% follow directly, since the time-embedding layers and projections are deleted.
  • The $O(T/M)$ bound means the cost of going unconditional vanishes as graph size grows, so the benefit should be largest for molecular and social graphs with many potential edges.
  • Small or sparse graphs are the predicted failure regime: the paper's experiments show explicit conditioning still helps there, consistent with weak posterior concentration at small $M$.
  • The same design principle is claimed to transfer to Poisson, Beta, multinomial, and coupled structure-attribute corruptions, because the target Lipschitz constants for those families are $O(1)$.

Reading between the lines

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

  • This suggests a crossover experiment the paper does not run: plotting the quality gap between conditioned and unconditional models against graph size should show a threshold $M$ below which conditioning wins and above which it does not; a practitioner could measure that threshold for a new dataset before choosing an architecture.
  • The architecture sensitivity the paper reports hints that timestep embeddings may function partly as an optimization curriculum rather than as information; an ablation that adds timestep-free warm-starting to a model like DiGress would test whether the gap is about representation or about training stability.
  • The linear-in-$T$ factor of the bound is the part most sensitive to the Lipschitz assumption; for very long sampling chains, the theory predicts that unconditional models need either larger $M$ or explicit conditioning, which is a testable scaling law on fixed datasets with more reverse steps.
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 / 5 minor

Summary. This paper argues that explicit noise-level/timestep conditioning can be omitted in graph diffusion models. It presents a theoretical framework (EFPC, ETDB, MDEP, and coupled extensions) claiming posterior concentration O(M^-1) for the flip rate, single-step target deviation O(M^-1), and multi-step cumulative reconstruction error O(T/M), where M is the number of potential edges. It validates the predicted scaling laws on synthetic graphs and compares t-aware versus t-free variants of GDSS and DiGress on QM9 and soc-Epinions1, reporting comparable or better quality with parameter and runtime savings.

Significance. The question is timely, and the empirical demonstration that t-free GDSS/DiGress variants can match or beat t-aware baselines is valuable; if the O(T/M) bound held under realistic assumptions, it would constitute a substantive theoretical justification for simpler graph diffusion architectures. The paper merits credit for multi-seed benchmarking on two architectures and datasets, for reporting parameter/computation savings, and for clean-looking scaling-law experiments with high R^2. However, the synthetic validations largely recompute the paper's own Bayesian formulas, and the main multi-step bound is not established as stated; the theoretical contribution therefore needs substantial revision before the central claim is supported.

major comments (4)
  1. [Theorem 4.3 and Appendix D] The claim that the geometric prefactor (Lmax^T - 1)/(Lmax - 1) is O(T) is false for fixed eta > 0. With Lmax = 1 + eta, eta < 0.2, and T = 1000 as used by DiGress, the prefactor is on the order of e^{200}/eta, not 1000. The approximation (1+eta)^T \approx 1 + T eta requires eta*T << 1. As written, the central O(T/M) bound and the conclusion that large graphs make t-free models safe do not follow from Assumptions A2 and A4. The theorem must either state an O(M^{-1} (1+eta)^T/eta) bound or impose a condition such as eta = O(1/T). The same issue affects Theorem 5.3 (JMEP), whose proof repeats the same geometric-series argument.
  2. [Assumption A4, Appendix A] Assumption A4 assumes that the trained denoiser already achieves per-component MSE O(M^{-1}). This is close to the conclusion that ETDB is supposed to provide, but ETDB bounds the gap between ideal conditional and unconditional Bayesian targets, not the error of a learned denoiser. Without a separate capacity, optimization, or generalization argument linking the learned f_theta to the ideal target, the multi-step theorem is conditional on the very property it aims to guarantee. The authors should either prove A4 from more basic assumptions or explicitly restate it as a limitation rather than an assumption that is naturally satisfied.
  3. [Section 6.1 and Appendix G] The synthetic validation is not independent evidence for the theory. The EFPC, ETDB, and MDEP quantities are computed using the same Beta-Bernoulli posterior formulas and Bayesian target expressions derived in Appendices B-D. High R^2 values therefore confirm internal consistency of the formulas, not the validity of the theorem's assumptions (Lipschitz denoiser, O(M^{-1}) learned error, or the O(T) geometric prefactor). The real-world t-free versus t-aware comparisons are independent and informative, but they do not test the Lmax or A4 conditions that the theory requires.
  4. [Section 4.1 and Appendix F] The scale-free rate Var(beta | A_t) = \tilde{O}(M^{-(alpha-2)/(alpha-1)}) is stated as part of the framework, but Appendix F describes the derivation as a 'conceptual outline' and 'Hypothesis' and does not give a rigorous Fisher-information computation. This rate should be labeled as a conjecture or removed from the main theorem statements until a complete proof is supplied.
minor comments (5)
  1. [Appendix D] The appendix contains editorial artifacts that must be removed, including 'ensure this label is correct' and 'ensure this label is correct for your main paper'; Appendix B similarly contains the phrase 'a specific form provided in the user's context is...'. These passages are not appropriate in a submitted manuscript.
  2. [Theorem 5.3] The norm expression \|A*_0 - \hat A_0\|_E + \|X*_0 - \hat X_0\|_F is ambiguous because the two summands use different norms; the proof bounds each separately, but the theorem statement should define the combined norm precisely.
  3. [Abstract and Section 6.4] The abstract says parameter reduction is 4-6% and computation time reduction is 8-10%, while Section 6.4 reports up to 19.9% parameter reduction for GDSS on soc-Epinions1. These numbers should be reconciled or the range clarified.
  4. [Table 2] The QM9 uniqueness values for DiGress variants are extremely low (around 4.7-5.1%), which is concerning for a claim of comparable generative quality; the authors should discuss this mode-collapse-like behavior and its interaction with the validity metrics.
  5. [Appendix H] The main paper's Table 2 appears to report the permissive evaluation protocol, but the table caption does not state this; please indicate which protocol is used wherever numbers are cited from the strict versus permissive evaluations.

Circularity Check

2 steps flagged · score 6.0 of 10

MDEP's O(T/M) scaling is imported via Assumption A4 rather than derived from EFPC/ETDB, and the synthetic 'validation' recomputes the same Bayesian formulas used in the proofs.

  1. fitted input called prediction [Section 4.3, Theorem 4.3; Appendix A, Assumption A4]
    "Assumption A4. (Model Capacity and Optimization Quality) ... the Graph Diffusion Model is assumed to achieve a per-component mean-squared error that scales as O(M^{-1}) ... when predicting the clean graph signal (or the noise) at each reverse step. ... Under Assumption A2(i) ... and Assumption A4 (Model Capacity, implying maximum single-step target deviation δmax = O(M^{-1}) from ETDB), the Frobenius norm ... ∥A∗0 − ˆA0∥F ≤ (L_T_max−1)/(L_max−1) δmax = O(T M^{-1})."

    The 'prediction' O(T/M) is already contained in A4: the theorem assumes exactly the per-step O(M^-1) error that the framework claims to establish as a consequence of noise-level inferability. ETDB bounds the deviation between ideal conditional and unconditional Bayesian targets, not the trained denoiser's prediction error; A4 supplies the learned-denoiser error separately. Therefore MDEP is a Lipschitz recursion applied to an assumed scaling, not a derivation from EFPC/ETDB. The central claim that unconditional GDMs are safe on large graphs thus rests on assuming the key magnitude it purports to prove. The paper's own limitation acknowledges this: 'The guarantees require ... near-optimal single-step error O(M^{-1}).'

  2. other [Appendix G, Posterior and Target Computations; Section 6.1, Table 1]
    "For EFPC, the posterior distribution of the flip rate β given the observed number of flips and total potential edges was modeled as a Beta distribution (Beta(flips+α0, |E|−flips+β0) with priors α0 = 1.0, β0 = 1.0), and its mean and variance were computed analytically. For ETDB (Bernoulli noise), the conditional target Rcond(At, βtrue, p0) and unconditional target Runcond(At, flips, |E|, p0) were computed based on Bayesian optimal estimation. ..."

    The synthetic experiments 'validating' EFPC, ETDB, and MDEP compute the same closed-form conjugate-Beta posterior and the same conditional/unconditional regression targets that Appendix B and C derive and bound. The empirical exponents in Table 1 are therefore algebraic consequences of the proof formulas, not independent measurements of a learned denoiser. In particular, the MDEP scaling is obtained by summing the analytic per-step deviation from the same posterior-mean substitution, so the observed O(T/M) behavior is built in by construction. These experiments confirm internal consistency but cannot validate the assumptions (A2, A4) on which the theoretical transfer to real models depends.

full rationale

The paper's real-world DiGress/GDSS comparisons are independent evidence that t-free variants can match t-aware ones, and EFPC itself is a standard Bernstein-von Mises concentration result. The circularity enters when the ideal-target bounds are transferred to learned denoisers: Theorem 4.3 (MDEP) explicitly assumes A4, which already postulates the O(M^-1) per-step error that the theoretical framework is supposed to deliver. The cumulative O(T/M) bound is then a restatement of that assumption under a Lipschitz recursion, not a derivation from noise-level inferability. The synthetic validation amplifies the issue by recomputing the same Bayesian formulas used in the proofs, so its scaling laws are identities rather than falsifiable tests. No load-bearing self-citation chain was found. Overall score 6 reflects that the central theoretical conclusion partially reduces to its own input assumption, while the empirical real-world results retain independent value.

Assumptions & free parameters 2 free parameters · 7 assumptions · 0 invented entities

The theory rests on standard Bayesian asymptotics plus strong modeling assumptions. A4 is the most consequential because it assumes the very O(M^-1) error scale the paper wants to conclude for learned denoisers. The synthetic scaling validation recomputes the same Bayesian formulas used in the proofs, so it is not independent evidence. No new physical or conceptual entities are postulated.

free parameters (2)
  • eta in Lmax=1+eta = eta<0.2, but needs eta=O(1/T) for the O(T) prefactor
    Introduced in Assumption A2 to make MDEP error accumulation linear; the stated range is too loose for large T.
  • Beta prior hyperparameters alpha0, beta0 = 1.0, 1.0
    Chosen for the synthetic posterior computations in Appendix G; affects constants but not the O(M^-1) rate.
assumptions (7)
  • domain assumption A1 Degree Condition: bounded maximum degree or power-law degree distribution with exponent alpha>2.
    Used to control degree heterogeneity in Section 4 and Appendix F; the power-law case is not rigorously derived.
  • ad hoc to paper A2 Global Lipschitz regularity: denoiser and ideal conditional target are Lmax=1+eta Lipschitz with eta<0.2.
    Introduced in Section 4 assumptions; the eta<0.2 range is insufficient for the claimed O(T) prefactor when T is large.
  • standard math A3 Prior regularity: prior over noise parameters is smooth and bounded away from the boundary.
    Standard condition for the Bernstein-von Mises theorem used in Appendix B.
  • ad hoc to paper A4 Model capacity: trained GDM achieves per-component MSE O(M^-1).
    Stated in Appendix A; this imports the target error scale as an assumption rather than deriving it.
  • domain assumption Bernoulli edge flips are independent across potential edges.
    Core of the forward corruption model in Section 3; extensions relax to local dependence.
  • domain assumption Coupled Gaussian covariance is non-singular for gamma<1.
    Needed in Appendix E for Fisher information and Bernstein-von Mises application; gamma=1 is excluded.
  • standard math Bernstein-von Mises theorem and Chernoff, Hoeffding, and Janson concentration inequalities.
    Background probabilistic tools invoked in Appendices B through E.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Is Noise Conditioning Necessary? A Unified Theory of Unconditional Graph Diffusion Models." pith.science (2026). https://pith.science/paper/MP3MTBQZ

@misc{pith2026250522935,
  author       = {Pith},
  title        = {Pith review of: Is Noise Conditioning Necessary? A Unified Theory of Unconditional Graph Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MP3MTBQZ}},
  note         = {Machine review of arXiv:2505.22935}
}
read the original abstract

Explicit noise-level conditioning is widely regarded as essential for the effective operation of Graph Diffusion Models (GDMs). In this work, we challenge this assumption by investigating whether denoisers can implicitly infer noise levels directly from corrupted graph structures, potentially eliminating the need for explicit noise conditioning. To this end, we develop a theoretical framework centered on Bernoulli edge-flip corruptions and extend it to encompass more complex scenarios involving coupled structure-attribute noise. Extensive empirical evaluations on both synthetic and real-world graph datasets, using models such as GDSS and DiGress, provide strong support for our theoretical findings. Notably, unconditional GDMs achieve performance comparable or superior to their conditioned counterparts, while also offering reductions in parameters (4-6%) and computation time (8-10%). Our results suggest that the high-dimensional nature of graph data itself often encodes sufficient information for the denoising process, opening avenues for simpler, more efficient GDM architectures.

Figures

Figures reproduced from arXiv: 2505.22935 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. , we varied γ from 0 (independent noise) to 0.99 (highly coupled shared noise component). 0.0 0.2 0.4 0.6 0.8 1.0 75000 80000 85000 90000 95000 100000 105000 Reconstruction error (a) Reconstruction error vs. 0.0 0.2 0.4 0.6 0.8 1.0 0.965 0.970 0.975 0.980 0.985 0.990 0.995 1.000 Classification accuracy (b) Accuracy vs [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. EFPC verification. Posterior variance Var(β | At) versus potential edge count |E| on SBM graphs with β = 0.2. The log–log fit has slope −1.02 ± 0.02 (R2 = 0.999), matching the theoretical −1. 10 5 10 6 potential edge count |E| 10 7 10 6 10 5 d e viatio n / |E| (a) deviation scaling slope=-1.12 deviation 0 1 2 3 4 5 potential edge count |E| 1e6 0 100000 200000 300000 400000 ||Runcond||2 (b) norm growth R 2=1.000 10 5… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: ETDB verification. Deviation between the conditional target R(At | t) and unconditional R(At), normalized by |E|, on the same SBM graphs. The log–log slope is −1.06 ± 0.03 (R2 = 0.998). H Experimental Setup for Real-World Datasets This appendix provides a detailed desc…
Figure 5
Figure 5. Figure 5: JMEP verification. Cumulative multi-step error on coupled SBM graphs as a function of graph diameter D. The log–log slope is −1.04 ± 0.05, confirming the O(D−1 ) bound. General Setup Hardware and Software. All models were trained and evaluated primarily on NVIDIA L4 GP…
Figure 6
Figure 6. Figure 6: GDSS on QM9: Training-time evolution of molecular metrics. Bands denote 95% CI over [PITH_FULL_IMAGE:figures/full_fig_p033_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 31 canonical work pages

  1. [1]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. InAdvances in Neural Information Processing Systems (NeurIPS) 34, 2021

    work page 2021

  2. [2]

    Weiss, Mohammad Norouzi, and William Chan

    Nanxin Chen, Yu Zhang, Heiga Zen, Ron J. Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. InICASSP 2021 - IEEE International Conference on Acoustics, Speech and Signal Processing, 2021

    work page 2021

  3. [3]

    Diffwave: A versatile diffusion model for audio synthesis

    Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. InInternational Conference on Learning Representations (ICLR), 2021

    work page 2021

  4. [4]

    Diffusion probabilistic models for 3d point cloud generation

    Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2837–2846, 2021

    work page 2021

  5. [5]

    Geodiff: A geometric diffusion model for molecular conformation generation

    Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: A geometric diffusion model for molecular conformation generation. InInternational Conference on Learning Representations (ICLR), 2022

    work page 2022

  6. [6]

    Difusco: Graph-based diffusion solvers for combinatorial optimization

    Zhiqing Sun and Yiming Yang. Difusco: Graph-based diffusion solvers for combinatorial optimization. InAdvances in Neural Information Processing Systems (NeurIPS) 36, 2023

    work page 2023

  7. [7]

    DiGress: Discrete diffusion for graph generation.Advances in Neural Information Processing Systems (NeurIPS), 2022

    Guillaume Vignac, Louis-Philippe Morency, et al. DiGress: Discrete diffusion for graph generation.Advances in Neural Information Processing Systems (NeurIPS), 2022

    work page 2022

  8. [8]

    Denoising diffusion probabilistic models.Advances in Neural Information Processing Systems (NeurIPS), 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in Neural Information Processing Systems (NeurIPS), 33:6840–6851, 2020

    2020

Show all 45 references
  1. [9]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations (ICLR), 2021

    2021

  2. [10]

    Is noise conditioning necessary for denoising generative models?, 2025

    Qiao Sun, Zhicheng Jiang, Hanhong Zhao, and Kaiming He. Is noise conditioning necessary for denoising generative models?, 2025. URLhttps://arxiv.org/abs/2502.13129

    2025

  3. [11]

    Margossian, Ruben Ohana, and Bruno Régaldo-Saint Blan- card

    David Heurtel-Depeiges, Charles C. Margossian, Ruben Ohana, and Bruno Régaldo-Saint Blan- card. Listening to the noise: Blind denoising with gibbs diffusion. InProceedings of the 41st International Conference on Machine Learning (ICML). PMLR, 2024

    2024

  4. [12]

    Dif- fusion models with learned adaptive noise

    Subham Sekhar Sahoo, Aaron Gokaslan, Christopher De Sa, and V olodymyr Kuleshov. Dif- fusion models with learned adaptive noise. InAdvances in Neural Information Processing Systems (NeurIPS) 2024. Curran Associates, Inc., 2024

    2024

  5. [13]

    Non gaussian denoising diffusion models

    Eliya Nachmani, Robin San Roman, and Lior Wolf. Non gaussian denoising diffusion models. arXiv preprint arXiv:2106.07582, 2021

    2021 arXiv

  6. [14]

    Diffusion models for graphs benefit from discrete state spaces.arXiv preprint arXiv:2210.01549, 2022

    Kilian Konstantin Haefeli, Karolis Martinkus, Nathanaël Perraudin, and Roger Wattenhofer. Diffusion models for graphs benefit from discrete state spaces.arXiv preprint arXiv:2210.01549, 2022. 10

    2022 arXiv

  7. [15]

    Tseng, Nathaniel Diamant, Tommaso Biancalani, and Gabriele Scalia

    Alex M. Tseng, Nathaniel Diamant, Tommaso Biancalani, and Gabriele Scalia. Graphguide: interpretable and controllable conditional graph generation with discrete bernoulli diffusion. arXiv preprint arXiv:2302.03790, 2023

    2023 arXiv

  8. [16]

    Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg

    Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Struc- tured denoising diffusion models in discrete state-spaces. InAdvances in Neural Information Processing Systems (NeurIPS) 34, 2021

    2021

  9. [17]

    Discrete-state continuous-time diffusion for graph generation, 2024

    Zhe Xu, Ruizhong Qiu, Yuzhong Chen, Huiyuan Chen, Xiran Fan, Menghai Pan, Zhichen Zeng, Mahashweta Das, and Hanghang Tong. Discrete-state continuous-time diffusion for graph generation, 2024. URLhttps://arxiv.org/abs/2405.11416

    2024 arXiv

  10. [18]

    Score-based generative modeling of graphs via the system of stochastic differential equations

    Jaehyeong Jo, Seul Lee, and Sung Ju Hwang. Score-based generative modeling of graphs via the system of stochastic differential equations. InInternational conference on machine learning, pages 10362–10383. PMLR, 2022

    2022

  11. [19]

    Deep unsuper- vised learning using nonequilibrium thermodynamics.International Conference on Machine Learning (ICML), pages 2256–2265, 2015

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsuper- vised learning using nonequilibrium thermodynamics.International Conference on Machine Learning (ICML), pages 2256–2265, 2015

    2015

  12. [20]

    Equivariant diffusion for molecule generation in 3d, 2022

    Emiel Hoogeboom, Victor Garcia Satorras, Clément Vignac, and Max Welling. Equivariant diffusion for molecule generation in 3d, 2022. URL https://arxiv.org/abs/2203.17003

    2022 arXiv

  13. [21]

    Noise2self: Blind denoising by self-supervision

    Joshua Batson and Loic Royer. Noise2self: Blind denoising by self-supervision. InInternational conference on machine learning, pages 524–533. PMLR, 2019

    2019

  14. [22]

    Permutation invariant graph generation via score-based generative modeling

    Chenhao Niu, Yang Song, Jiaming Song, Shengjia Zhao, Aditya Grover, and Stefano Ermon. Permutation invariant graph generation via score-based generative modeling. InProceedings of the 23rd International Conference on Artificial Intelligence and Statistics (AISTATS), pages 4474...

    2020

  15. [23]

    Hyperbolic graph diffusion model

    Lingfeng Wen, Xuan Tang, Mingjie Ouyang, Xiangxiang Shen, Jian Yang, Daxin Zhu, Mingsong Chen, and Xian Wei. Hyperbolic graph diffusion model. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 15823–15831, 2024

    2024

  16. [24]

    Advancing graph generation through beta diffusion, 2024

    Xinyang Liu, Yilin He, Bo Chen, and Mingyuan Zhou. Advancing graph generation through beta diffusion, 2024. URLhttps://arxiv.org/abs/2406.09357

    2024 arXiv

  17. [25]

    Fisher information and stochastic complexity.IEEE transactions on information theory, 42(1):40–47, 1996

    Jorma J Rissanen. Fisher information and stochastic complexity.IEEE transactions on information theory, 42(1):40–47, 1996

    1996

  18. [26]

    A. W. van der Vaart.Asymptotic Statistics. Cambridge University Press, 1998

    1998

  19. [27]

    Tom A. B. Snijders, Gert G. van de Bunt, and Christian E. G. Steglich. Introduction to stochastic actor-based models for network dynamics.Social Networks, 32(1):44–60, 2010. doi: 10.1016/j.socnet.2009.02.004

    2010 doi

  20. [28]

    Higher-order interactions shape col- lective dynamics differently in hypergraphs and simplicial complexes.Nature Communications, 14:1605, 2023

    Yuanzhao Zhang, Maxime Lucas, and Federico Battiston. Higher-order interactions shape col- lective dynamics differently in hypergraphs and simplicial complexes.Nature Communications, 14:1605, 2023. doi: 10.1038/s41467-023-37214-w

    2023 doi

  21. [29]

    Dynamic networks and behavior: Separating selection from influence.Sociological methodology, 40(1):329–393, 2010

    Christian Steglich, Tom AB Snijders, and Michael Pearson. Dynamic networks and behavior: Separating selection from influence.Sociological methodology, 40(1):329–393, 2010

    2010

  22. [30]

    Complex contagion process in spreading of online innovation.Journal of the Royal Society Interface, 11(101):20140694,

    Márton Karsai, Gerardo Iñiguez, Kimmo Kaski, and János Kertész. Complex contagion process in spreading of online innovation.Journal of the Royal Society Interface, 11(101):20140694,

  23. [31]

    Gnnguard: Defending graph neural networks against adversarial attacks.arXiv preprint, 2020

    Xiang Zhang and Marinka Zitnik. Gnnguard: Defending graph neural networks against adversarial attacks.arXiv preprint, 2020

    2020

  24. [32]

    Provably robust explainable graph neural networks against graph perturbation attacks

    Jiate Li, Meng Pang, Yun Dong, Jinyuan Jia, and Binghui Wang. Provably robust explainable graph neural networks against graph perturbation attacks. InInternational Conference on Learning Representations (ICLR), 2025. 11

    2025

  25. [33]

    Ad- versarial examples on graph data: Deep insights into attack and defense

    Huijun Wu, Chen Wang, Yuriy Tyshetskiy, Andrew Docherty, Kai Lu, and Liming Zhu. Ad- versarial examples on graph data: Deep insights into attack and defense. InProceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI), 2019. URL https://arxiv.or...

    2019 arXiv

  26. [34]

    Shuwen Chai and Miklós Z. Rácz. Efficient graph matching for correlated stochastic block models.arXiv preprint, 2024

    2024

  27. [35]

    Duranthon and L

    O. Duranthon and L. Zdeborová. Optimal inference in contextual stochastic block models, 2024. URLhttps://arxiv.org/abs/2306.07948

    2024 arXiv

  28. [36]

    Aegraph: Node attribute-enhanced graph encoder method.Expert Systems with Applications, 236:121382, 2024

    Kang Sun, Liqing Qiu, and Wenxiu Zhao. Aegraph: Node attribute-enhanced graph encoder method.Expert Systems with Applications, 236:121382, 2024

    2024

  29. [37]

    On the evolution of random graphs.Publ

    Paul Erd6s and Alfréd Rényi. On the evolution of random graphs.Publ. Math. Inst. Hungar. Acad. Sci, 5:17–61, 1960

    1960

  30. [38]

    Stochastic blockmodels: First steps.Social networks, 5(2):109–137, 1983

    Paul W Holland, Kathryn Blackmond Laskey, and Samuel Leinhardt. Stochastic blockmodels: First steps.Social networks, 5(2):109–137, 1983

    1983

  31. [39]

    Quantum chemistry structures and properties of 134 kilo molecules.Scientific data, 1(1):1–7, 2014

    Raghunathan Ramakrishnan, Pavlo O Dral, Matthias Rupp, and O Anatole V on Lilienfeld. Quantum chemistry structures and properties of 134 kilo molecules.Scientific data, 1(1):1–7, 2014

    2014

  32. [40]

    Graph evolution: Densification and shrinking diameters.ACM transactions on Knowledge Discovery from Data (TKDD), 1(1):2–es, 2007

    Jure Leskovec, Jon Kleinberg, and Christos Faloutsos. Graph evolution: Densification and shrinking diameters.ACM transactions on Knowledge Discovery from Data (TKDD), 1(1):2–es, 2007

    2007

  33. [41]

    Large deviations for sums of partly dependent random variables.Random Structures & Algorithms, 24(3):234–248, 2004

    Svante Janson. Large deviations for sums of partly dependent random variables.Random Structures & Algorithms, 24(3):234–248, 2004. A Detailed Assumptions and Rationale for Theoretical Framework This appendix provides a detailed statement and justification for the assumptions u...

    2004

  34. [43]

    Concentration of the Sufficient Statistic:We demonstrate that the total number of observed edge flips, X, which is a sufficient statistic for βt, concentrates sharply around its expected value

  35. [44]

    This is first done by assuming a Beta conjugate prior to derive an exact analytical form for the variance

    Posterior Variance Analysis using Conjugate Priors and Laplace’s Method:We analyze the posterior variance Var[β|X] . This is first done by assuming a Beta conjugate prior to derive an exact analytical form for the variance. We then generalize this to show that the O(M −1) scal...

  36. [45]

    sharp rate

    Asymptotic Normality and Refined Rate via Bernstein–von Mises Theorem:Finally, we employ the Bernstein–von Mises theorem to formally establish the asymptotic normality of the posterior distribution. This allows for a precise determination of the leading constant in the O(M −1)...

    2022

  37. [2014]

    doi: 10.1098/rsif.2014.0694

    2014

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.