{"id":"d01e07cb-1871-4495-929e-ebaccfa09d3e","arxiv_id":"2607.07510","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":5,"one_line_summary":"GNN-parametrized continuous normalizing flows for graph signals are permutation equivariant and satisfy Wasserstein stability bounds under relative graph perturbations, motivating a Lipschitz-regularized training strategy.","lead":"This paper proves that graph-based continuous flow models for signal generation preserve permutation equivariance and remain stable when the underlying graph is perturbed. The results matter because graph-structured data often has noisy or estimated topologies, and knowing how errors propagate through generative dynamics helps build more robust models.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Proof of Theorem 1 applies the one-sided Lipschitz condition (13)—stated for the nominal graph S—to the perturbed graph S̃ without justification; the gap is fixable but currently unaddressed.","rationale":"The reader's identified concern—that the integral Lipschitz constant C and filter bound B from [30, Thm. 4] are assumed but not enforced during training—is valid and important for practical applicability. However, it is a limitation the paper openly acknowledges (§V: 'our proposed regularization strategy solely penalizes M... though alternative approaches could be chosen to jointly regularize both the spatial and integral Lipschitz constants').\n\nMy concern is distinct: it is about proof correctness, not assumption satisfaction. The one-sided Lipschitz condition (13) and the state stability condition (Definition 4) are both stated for the nominal graph S but applied to the perturbed graph S̃ in the proofs of Theorems 1 and 2. This is a genuine gap in the proof as written.\n\nHowever, the gap is benign. The fix—decomposing the perturbed-graph vector field difference into nominal-graph plus perturbation terms—introduces only additional O(Γε) contributions that change the constant prefactor but not the bound's structure. The exponential growth factor Ω_t, the trajectory supremum C_g, and the linear-in-ε dependence are all preserved.\n\nGiven that (a) the proof gap is easily fixable and doesn't change the qualitative result, and (b) the reader's concern about C and B is acknowledged by the authors and is a known limitation rather than a flaw, the CONDITIONAL verdict with MODERATE confidence is appropriate. The theory is sound under its stated (or slightly strengthened) assumptions; the gap between theory and practice remains the main open issue, which the paper honestly reports.","tokens_in":24449,"tokens_out":9254,"duration_ms":623833,"concrete_test":"Re-derive the Grönwall bound in Appendix B using the three-way decomposition of u_t^θ(x̃_t; S̃) - u_t^θ(P₀x_t; S̃) into nominal-graph and perturbation components: apply (13) only to u_t^θ(·; S) and (12) to the two cross terms. If the resulting bound matches (14) up to a constant factor (expected ~3×), the proof gap is confirmed benign and the theorem stands with a minor constant adjustment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Appendix B, the proof of Theorem 1 decomposes d/dt||z_t||² into two terms. The first term bounds ⟨x̃_t - P₀x_t, u_t^θ(x̃_t; S̃) - u_t^θ(P₀x_t; S̃)⟩ ≤ m_t||x̃_t - P₀x_t||² by invoking the one-sided Lipschitz condition (13). However, (13) is stated as ⟨u_t^θ(x; S) - u_t^θ(y; S), x - y⟩ ≤ m_t||x - y||² for the nominal graph S only. Here it is applied to u_t^θ(·; S̃) on the perturbed graph S̃, which is not covered by the assumption as written. The same pattern recurs in Theorem 2 (Appendix G), where the state stability condition (Definition 4) is defined for S but applied to S̃ in the recurrence ||T_k^θ(x̃_k; S̃) - T_k^θ(P₀x_k; S̃)|| ≤ α||x̃_k - P₀x_k||.\n\nThe gap is fixable: decompose u_t^θ(x; S̃) - u_t^θ(y; S̃) = [u_t^θ(x; S̃) - u_t^θ(x; S)] + [u_t^θ(x; S) - u_t^θ(y; S)] + [u_t^θ(y; S) - u_t^θ(y; S̃)], apply (13) to the middle term (nominal graph) and (12) to the cross terms. This introduces two additional O(Γε) contributions, changing the constant prefactor from 1 to approximately 3, but preserving the bound's structure: Ω_t · C_g · (Γε + O(ε²)). The qualitative result is unchanged, but the proof as written contains an unjustified step that should be made explicit, either by assuming (13) holds uniformly over an ε-neighborhood of S or by performing the decomposition above.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"The paper studies permutation equivariance and stability of continuous normalized flow (CNF) models parameterized by graph neural networks (GNNs) for graph signal generation. The central contribution is a Wasserstein stability bound (Theorem 1, Corollary 2) showing that under relative graph perturbations of magnitude ε, the 2-Wasserstein distance between generated distributions on nominal and perturbed graphs is bounded by Ω_t · C'_g · (Γε + O(ε²)), where Γ captures base GNN stability, Ω_t is an exponential growth factor from the one-sided Lipschitz constant, and C'_g is a trajectory-dependent supremum. The framework extends to discrete Euler and Heun samplers (Theorem 2, Corollaries 3–4). Motivated by the bounds, the authors propose a regularized flow matching (RFM) objective penalizing the spatial Lipschitz constant M, and validate on synthetic SBM graphs and fMRI connectome data.","tokens_in":24800,"tokens_out":1561,"duration_ms":307071,"significance":"The paper provides a clean, self-contained theoretical treatment of a timely problem: how structural perturbations propagate through generative flow dynamics on graphs. The proofs use standard tools appropriately (Grönwall's inequality, triangle inequality, induction). The derivation is from first principles, building on the external GNN stability result of Gama et al. [30] without circularity. The practical regularization strategy is directly motivated by the theory, and the experiments honestly report that the bounds are loose but that bound-informed regularization still improves robustness. The extension to discrete samplers with explicit Euler and Heun bounds, including the convergence to the continuous bound as h→0, is a solid contribution.","major_comments":[{"comment":"Appendix B, proof of Theorem 1: The one-sided Lipschitz condition (13) is stated for the nominal graph S, but in the proof it is applied to u_t^θ(·; S̃) on the perturbed graph S̃. Specifically, the bound ⟨x̃_t - P₀x_t, u_t^θ(x̃_t; S̃) - u_t^θ(P₀x_t; S̃)⟩ ≤ m_t‖x̃_t - P₀x_t‖² invokes (13) with the perturbed GSO S̃, which is not covered by the assumption as written. The same pattern recurs in Appendix G (proof of Theorem 2), where the state stability condition (Definition 4) is defined for S but applied to S̃ in the recurrence ‖T_k^θ(x̃_k; S̃) - T_k^θ(P₀x_k; S̃)‖ ≤ α‖x̃_k - P₀x_k‖. This is a load-bearing gap: the proof structure can be repaired (e.g., by decomposing u_t^θ(x; S̃) - u_t^θ(y; S̃) into nominal-graph and cross-graph terms, applying (13) to the former and (12) to the latter, introducing additional O(Γε) contributions that change the constant prefactor but preserve the bound's O(","section":null},{"comment":"§II-B and §V: The base GNN stability bound (4) from [30, Thm. 4] requires the filter frequency response h(λ) to be integral Lipschitz with constant C and uniformly bounded by B. These conditions constrain the learned filter coefficients but are not enforced during training—the paper only regularizes the spatial Lipschitz constant M (Proposition 3, Eq. 25), leaving C and B uncontrolled. The paper acknowledges this in §VII ('exploring...jointly regularizing the spatial and frequency-response Lipschitz constants') but does not state as a formal assumption that the trained filters satisfy integral Lipschitzness, nor does it verify this empirically. If the learned filters violate these conditions, the base bound (4) does not hold and the entire stability edifice (Theorems 1–2, Corollaries 1–4) is unsupported. The authors should either (a) add an explicit assumption that trained filters are验证d","section":null},{"comment":"§V, Eqs. (22)–(24) and Proposition 3: The bound on M in (24) is stated as M ≤ ∏_ℓ max_i ‖∑_p Θ_ℓp λ_i^p‖_2, but the proof in Appendix J derives M ≤ ∏_ℓ M_ℓ where M_ℓ := max_i ‖∑_p Θ_ℓp λ_i^p‖_2. The product of maxima is not the same as the maximum of products. Please clarify whether (24) is the tight bound from the proof or a further relaxation, and if the latter, note this explicitly.","section":null}],"minor_comments":[{"comment":"§III, Eq. (12): The bound ‖g(x,t)‖ appears on the right-hand side, but the subsequent definition of C_g(x₀) in Theorem 1 evaluates ‖g(Φ_τ^θ(P₀^⊤x₀; S), τ)‖ along the trajectory. Please clarify that (12) is being applied with x = Φ_τ^θ(x₀; S̃) (the perturbed trajectory) while C_g is defined along the unperturbed trajectory, and justify that the latter bounds the former.","section":null},{"comment":"§VI-A: The architecture description states F₀ = 65 (input feature dimension from 1 node feature + 64 time embedding dimensions), but earlier in §III it is assumed F₀ = F_L = 1. Please reconcile this discrepancy or note that the theoretical analysis assumes d=1 for simplicity while experiments use d=65.","section":null},{"comment":"Figures 1–2: The y-axis labels on the center panels ('K(x₀; S)') are not defined in the text. Presumably this refers to the empirical stability metric ‖Ψ_k^θ(x₀; S̃) - Ψ_k^θ(x₀; S)‖, but this should be stated explicitly.","section":null},{"comment":"§VI-A, Synthetic perturbation: The perturbation model generates a random diagonal matrix E with entries drawn from [(1-ε)ε, ε]. This differs from the relative perturbation model in (3) where E is a general symmetric matrix. Please clarify the relationship or justify the restriction to diagonal E.","section":null},{"comment":"§II-B: The eigenvector misalignment parameter δ is stated as δ ≤ 8 but the source [30, Thm. 1] should be checked for the exact conditions under which this bound holds.","section":null},{"comment":"Typos: 'errrors' (§II-D), 'F(6)' should likely be 'g (6)' in Appendix H (Euler sampler graph stability verification).","section":null}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about applying the one-sided Lipschitz condition to the perturbed graph is valid and is the primary reason for the major_revision recommendation. The fix is straightforward (decompose into nominal + cross terms), but it touches the core proofs of both main theorems and should be made explicit before acceptance. The secondary concern about unenforced integral Lipschitz conditions is also legitimate but somewhat standard in this literature—the authors inherit assumptions from [30] and the gap between theory and practice is acknowledged. I would not block on this alone, but it deserves an explicit assumption statement."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive reading of our manuscript. The report identifies three substantive issues, all of which are valid and warrant revision. We address each below.","responses":[{"response":"The referee is correct. Condition (13) is stated for the nominal GSO S, but in the proof of Theorem 1 (Appendix B) it is applied to u_t^θ(·; S̃), i.e., the vector field evaluated on the perturbed graph. The same issue arises in Appendix G for Definition 4. As the referee notes, the proof structure can be repaired by decomposing u_t^θ(x̃_t; S̃) - u_t^θ(P₀x_t; S̃) into a nominal-graph term (to which (13) applies directly) and a cross-graph perturbation term (bounded via (12)). This introduces an additional O(Γε) contribution that changes the constant prefactor but preserves the O(Γε + O(ε²)) structure of the bound. We will revise the assumption statements in (13) and Definition 4 to explicitly require that the one-sided Lipschitz and state stability conditions hold for all GSOs in the relative perturbation neighborhood of S (i.e., for any S̃ satisfying (3) with ||E||₂ ≤ ε). This is a natural and mild strengthening: the conditions depend on the filter coefficients Θ_{ℓp}, which are shared across S and S̃, and on the spectrum of the GSO, which changes continuously under relative perturbations. We will also add a remark explaining why this strengthened assumption is satisfied under the same integral Lipschitz and boundedness conditions on the filter frequency response that underpin the base GNN stability bound (4). The revised proofs in Appendices B and G will explicitly show the decomposition the referee suggests.","revision_made":"yes","referee_comment":"Appendix B, proof of Theorem 1: The one-sided Lipschitz condition (13) is stated for the nominal graph S, but in the proof it is applied to u_t^θ(·; S̃) on the perturbed graph S̃. The same pattern recurs in Appendix G (proof of Theorem 2), where Definition 4 is defined for S but applied to S̃. This is a load-bearing gap."},{"response":"The referee is correct that the integral Lipschitz condition (constant C) and the uniform boundedness condition (constant B) on the filter frequency response are necessary assumptions for the base GNN stability bound (4) and hence for our Theorems 1–2 and Corollaries 1–4. In the current manuscript, these are invoked as assumptions inherited from [30] but are not stated as explicit assumptions on the trained filters, nor are they verified empirically. This is a gap. We will revise the manuscript as follows. First, we will add an explicit assumption (in §II-B or §III-B) stating that the trained filter coefficients satisfy the integral Lipschitz and boundedness conditions, making clear that our stability guarantees are conditional on this assumption. Second, we will add empirical verification in §VI: after training, we will compute the bounds on C and B from (22)–(23) for the learned filters and report whether they are finite and well-behaved. We expect this to be the case given the shallow architecture (L=2, P=4) and the regularized training, but the referee is right that this must be verified rather than assumed. Third, we will strengthen the discussion in §V and §VII to acknowledge that jointly regularizing M, C, and B is the principled approach, and that our current focus on M alone is a practical simplification whose validity depends on the trained filters happening to satisfy the C and B conditions. We note that our framework is general (as stated after (4): 'our analytical framework is general and naturally extends to any base GNN satisfying a stability condition like (4)'), so the theoretical results hold for any GNN whose filters satisfy these conditions—whether by architecture design, by regularization, or by verification.","revision_made":"yes","referee_comment":"§II-B and §V: The base GNN stability bound (4) from [30, Thm. 4] requires the filter frequency response h(λ) to be integral Lipschitz with constant C and uniformly bounded by B. These conditions constrain the learned filter coefficients but are not enforced during training—the paper only regularizes the spatial Lipschitz constant M, leaving C and B uncontrolled. If the learned filters violate these conditions, the base bound (4) does not hold and the entire stability edifice is unsupported."},{"response":"The referee is correct that there is a notational discrepancy between (24) and the proof in Appendix J. The proof derives M ≤ ∏_{ℓ=1}^{L} M_ℓ, where M_ℓ := max_i ‖∑_p Θ_{ℓp} λ_i^p‖_2. This is indeed a product of per-layer maxima, not a maximum of products. Equation (24) as written, M ≤ ∏_ℓ max_i ‖∑_p Θ_{ℓp} λ_i^p‖_2, is in fact the same expression—the product symbol ∏_ℓ ranges over layers and the max_i is taken within each layer. So (24) is the tight bound from the proof, not a further relaxation. However, we agree the notation is ambiguous as written because the scope of the product and maximum operators is not clearly delineated. We will revise (24) to make the notation unambiguous, writing M ≤ ∏_{ℓ=1}^{L} M_ℓ with M_ℓ := max_{i∈{1,...,N}} ‖∑_{p=0}^{P-1} Θ_{ℓp} λ_i^p‖_2, and we will add a sentence clarifying that this is a product of per-layer maxima (each maximum taken over graph frequencies), which is the direct result of the recursive application of the layer-wise Lipschitz bound in Appendix J.","revision_made":"partial","referee_comment":"§V, Eqs. (22)–(24) and Proposition 3: The bound on M in (24) is stated as M ≤ ∏_ℓ max_i ‖∑_p Θ_ℓp λ_i^p‖_2, but the proof in Appendix J derives M ≤ ∏_ℓ M_ℓ where M_ℓ := max_i ‖∑_p Θ_ℓp λ_i^p‖_2. The product of maxima is not the same as the maximum of products. Please clarify whether (24) is the tight bound from the proof or a further relaxation."}],"tokens_in":24531,"tokens_out":1446,"duration_ms":209728,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"Bottom line: this paper derives Wasserstein stability bounds for GNN-parametrized continuous normalizing flows under relative graph perturbations, extends them to discrete Euler/Heun samplers, and uses the bounds to motivate a Lipschitz regularization strategy. The theory is mostly sound and the contribution is genuinely new — nobody has formally characterized how graph structural errors propagate through generative flow dynamics before. It deserves a serious referee, but there are two issues the authors need to address before publication, one of which is a real proof gap the stress-test correctly identifies, and the other is a training-implementation mismatch the reader correctly flags but which is less severe than it sounds at first reading. Let me explain both in proportion to their actual weight. The main new results are Theorem 1 (continuous Wasserstein bound via Grönwall's inequality on the ODE error dynamics) and Theorem 2 (discrete bound via geometric series accumulation). The equivariance proofs (Propositions 1–2) are straightforward but necessary — they establish the prerequisite that makes stability modulo permutation well-defined. The proofs in Appendices B and G are clean and use standard techniques correctly. The Wasserstein lifting in Appendix C is a standard coupling argument done right. The regularization strategy (Proposition 3) is a reasonable practical contribution: the bound on the spatial Lipschitz constant M is graph-spectrum-aware, and the experiments show RFM improves stability without degrading generation quality. Now the two issues. First, the stress-test concern is legitimate. In Appendix B, the proof of Theorem 1 applies the one-sided Lipschitz condition (13) — stated for the nominal graph S — to the perturbed graph S̃ without justification. The same pattern recurs in Theorem 2's proof (Appendix G), where the state stability condition (Definition 4) is defined for S but applied to S̃. This is a real gap. The fix is straightforward: decompose u_t^θ(x; S̃) − u_t^θ(y; S̃) into nominal-graph and cross-graph terms, apply (13) to the nominal part and (12) to the cross terms. This introduces two additional O(Γε) contributions, changing a constant prefactor from 1 to roughly 3, but the bound's structure is preserved. The qualitative result is unchanged. The authors should either assume (13) holds uniformly over an ε-neighborhood of S or perform this decomposition explicitly. Second, the reader's concern about the integral Lipschitz condition (C and B from [30]) being unenforced during training is valid but somewhat overstated. The paper only regularizes M, leaving C and B uncontrolled — true. But the paper is transparent about this: Section V explicitly states the focus on M as the dominating term and acknowledges joint regularization of C and M as future work. The gap between assumed filter regularity and what is enforced is a limitation, not a flaw in the derivation. The bounds are conditional guarantees: if the filters satisfy integral Lipschitzness, the bounds hold. The practical question of whether trained filters actually satisfy this is empirical and somewhat orthogonal to the theoretical contribution. The bounds are admittedly loose — up to 12 orders of magnitude off in the fMRI case — but the paper is honest about this and the qualitative insight (controlling M improves robustness) is empirically supported. No code or data is released, which limits reproducibility but does not undermine the theoretical claims. This paper is for the graph signal processing community, particularly those working on generative models. Readers interested in stability theory for graph-based learning will get the most value. The contribution is real and the execution is competent. Recommend accepting for peer review with the expectation that the authors fix the proof gap in Theorems 1 and 2 and add a sentence acknowledging the C/B enforcement question more prominently.","headline":"Wasserstein stability bounds for GNN-parametrized continuous normalizing flows on graphs — sound theory with a fixable proof gap and a training-implementation mismatch worth flagging.","tokens_in":25554,"tokens_out":868,"would_cite":true,"duration_ms":153655,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Graph signal generators stay stable under topology noise","keywords":["graph neural networks","continuous normalizing flows","Wasserstein stability","permutation equivariance","flow matching","graph signal processing","relative graph perturbations","Lipschitz regularization"],"falsifier":"Train a GNN-parametrized flow where the learned filter coefficients produce a frequency response h(λ) that is not integral Lipschitz, then measure the empirical Wasserstein distance between generated distributions on nominal and perturbed graphs; if the distance grows superlinearly in ε or violates the predicted Ω_t scaling, the bound's dependence on the inherited GNN stability assumption is exposed.","tokens_in":24523,"feed_emoji":"","tokens_out":705,"duration_ms":100939,"temperature":0.7,"pith_summary":"When a generative flow model uses a graph neural network as its vector field, the resulting probability distribution over graph signals degrades gracefully under structural perturbations of the graph. Specifically, for a relative perturbation of magnitude ε, the 2-Wasserstein distance between generated distributions on the nominal and perturbed graphs is bounded by a product of three factors: Γ (inherited GNN stability), Ω_t (exponential error growth governed by the one-sided Lipschitz constant of the flow dynamics), and C'_g (a trajectory-dependent supremum). This bound holds for both the continuous-time ODE and its discrete Euler or Heun approximations. The key structural insight is that the spatial Lipschitz constant M of the vector field drives the exponential growth factor Ω_t, making it the dominant controllable source of instability. Penalizing M during training—via a spectral bound on the GNN filter responses—yields models that are more robust to graph estimation errors without degrading generation quality.","feed_headline":"Flow models on graphs stay stable under topology noise","feed_subtitle":"Wasserstein bounds show how graph perturbations propagate through generative dynamics, and penalizing one spectral constant during training.","key_machinery":"The Wasserstein stability bound (Theorem 1, Corollary 2) for the continuous flow, its discrete counterparts (Theorem 2, Corollaries 3–4) for Euler and Heun samplers, and the spectral Lipschitz bound (Proposition 3) that enables graph-aware regularization of the vector field during flow matching training.","core_discovery":"The paper proves that a continuous normalized flow parameterized by a GNN inherits permutation equivariance for both the continuous ODE and discrete samplers (Euler, Heun), and derives explicit Wasserstein stability bounds showing that distributional drift under relative graph perturbations factors into (i) base GNN stability Γ, (ii) an exponential growth term Ω_t governed by the one-sided Lipschitz constant m_t of the vector field, and (iii) a trajectory-dependent constant C'_g. The exponential growth factor is the mechanism through which small per-step vector field errors accumulate over the integration horizon, and because Ω_t is monotone in the Lipschitz constant M, regularizing M via a谱","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Wasserstein bounds for stable graph signal generation","Provable Wasserstein bounds for graph signal flow models","Graph flow models inherit GNN topology stability bounds","Stabilizing graph flow models against structural perturbations","Regularizing Lipschitz constants for stable graph flow models"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The entire stability framework inherits a bound from prior work requiring that the GNN's graph filter frequency response be integral Lipschitz (constant C) and uniformly bounded (constant B), but the training procedure only regularizes the spatial Lipschitz constant M, leaving C and B uncontrolled—if the learned filters violate integral Lipschitzness, the base stability bound does not hold.","fun_headline_variants_meta":{"raw":{"variants":["Wasserstein bounds for stable graph signal generation","Provable Wasserstein bounds for graph signal flow models","Graph flow models inherit GNN topology stability bounds","Stabilizing graph flow models against structural perturbations","Regularizing Lipschitz constants for stable graph flow models","Bounding Wasserstein drift in graph signal flow models"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":1566,"prompt_tokens":515,"completion_tokens":1051,"prompt_tokens_details":null},"tokens_in":515,"tokens_out":1051,"duration_ms":43255,"temperature":1.0,"reasoning_tokens":1007,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T08:27:15.109022+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Train a GNN-parametrized flow where the learned filter coefficients produce a frequency response h(λ) that is not integral Lipschitz, then measure the empirical Wasserstein distance between generated distributions on nominal and perturbed graphs; if the distance grows superlinearly in ε or violates the predicted Ω_t scaling, the bound's dependence on the inherited GNN stability assumption is exposed.","supporting_citations":[],"review_version":1}