{"id":"ee9961ba-9ebd-460f-a9a1-8daf717cb928","arxiv_id":"2505.20034","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A GNN built on a wave-equation recurrence is presented, but the claimed unconditional stability is disproved by the paper's own characteristic equation.","lead":"The paper models graph neural network message passing with the wave equation and proposes GWN layers, claiming an explicit discretization that is stable for any time step. The central stability proof is mathematically wrong, so the proposed efficiency guarantee does not hold.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2 is false: for any graph with an edge, the normalized adjacency has eigenvalue 1, so the recurrence has a mode with spectral radius (2+τ²+√(τ⁴+4τ²))/2 > 1, contradicting constant stability.","rationale":"The reader's weakest_assumption exactly identifies the fatal gap: in the real-root case λ′ > 2, the larger root exceeds 1. I verified this independently via the block-determinant reduction for the first-order form, and confirmed that the problematic case always occurs because D^{−1/2}AD^{−1/2} has eigenvalue 1 on any nonempty graph. The paper's own Figure 7 and the case analysis in Appendix A.3 evade this by only inspecting λ′ = 2 from below and λ′ ≤ −2, never the interval (2, 2+τ²]. Since Theorem 2 is the theoretical foundation for the claimed 'constantly stable' explicit scheme and the associated efficiency gains, its falsehood is a load-bearing defect that invalidates the central contribution. The empirical improvements over baselines on heterophilic datasets may be genuine, but they do not rescue the theoretical claim, and the GWN-fa stability theorem is inapplicable to the implemented time-dependent attention model. I therefore agree with the reader's REJECT verdict; no adjustment is needed.","tokens_in":28703,"tokens_out":5479,"duration_ms":55424,"concrete_test":"Run the scalar mode test on a two-node graph with adjacency [[0,1],[1,0]], so L_a = [[0,1],[1,0]] has eigenvalue μ = 1. Take initial vectors aligned with the eigenvector [1,1]^T and set τ = 1. The recurrence reduces to x_{n+1} = 3x_n − x_{n−1}, whose characteristic roots are (3±√5)/2 ≈ 2.618 and 0.382. Iterating from x_0 = 1, x_1 = 3 shows the norm growing like 2.618^n rather than staying bounded, directly disproving Theorem 2. Equivalently, compute the eigenvalues of C = [[2I + τ²L_a, −I],[I,0]] for this graph and verify that its spectral radius exceeds 1.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim of the paper is Theorem 2, which asserts that the explicit scheme X^{(n+1)} = (2I + τ²L_a)X^{(n)} − X^{(n−1)} with L_a = D^{−1/2}AD^{−1/2} is constantly stable for every τ > 0. The proof in Appendix A.3 converts the second-order recurrence to first-order form with matrix C and studies eigenvalues λ satisfying λ² − λ′λ + 1 = 0, where λ′ = 2 + τ²μ and μ ∈ [−1,1] is an eigenvalue of L_a. For μ > 0, λ′ > 2 whenever τ > 0, and the larger real root λ₁ = (λ′ + √(λ′² − 4))/2 is strictly greater than 1 (at μ = 1, λ₁ = (2 + τ² + √(τ⁴ + 4τ²))/2 > 1). The proof only analyses λ′ = 2 and λ′ ≤ −2 for the larger root, and never treats λ′ > 2. Since D^{−1/2}AD^{−1/2} always has eigenvalue 1 for any graph with at least one edge (with eigenvector D^{1/2}1), the defective case λ′ > 2 always occurs, so ρ(C) > 1. Thus the scheme is not stable for any positive time step; it diverges exponentially for generic initial data. The same flaw invalidates Theorem 3: L_{a,·} = εI ± D^{−1/2}AD^{−1/2} has eigenvalues up to ε+1, giving λ′ = 2 + τ²(ε+1) > 2 for ε > 0. Moreover, GWN-fa's implemented recurrence (Eq. 18) includes a time-dependent attention matrix and an ε^{(n)}X^{(0)} term, so Theorem 3 does not even apply to the deployed model. The claimed 'constantly stable' property is the basis for the efficiency advantage over GRAND and heat-equation methods; without it, the headline contribution is unsupported, even though the empirical results may still reflect a useful heuristic.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes modeling message passing in GNNs as wave propagation. It derives a graph wave equation, discretizes it with the forward Euler method to obtain an explicit scheme, and claims that the scheme is \"constantly stable\" for any positive time step (Theorems 2 and 3). Two implementations, GWN-sym and GWN-fa, are introduced and evaluated on node classification benchmarks, with additional experiments on over-smoothing and heterophily.","tokens_in":29136,"tokens_out":5566,"duration_ms":58557,"significance":"If the central stability claim were correct, it would be a significant contribution: explicit schemes for second-order graph PDEs that are unconditionally stable would enable larger time steps and more efficient training than conditionally stable heat-equation methods such as GRAND. The paper also provides a useful connection between wave equations and spectral GNNs, and the experimental section is extensive, including code release and large-scale benchmarks. However, the main theorem is false, and the proof omits a growing eigenmode that is present for every graph with at least one edge. Since the efficiency and robustness advantages are explicitly built on this theorem, the theoretical foundation of the paper is unsupported, even though the empirical results may reflect a useful heuristic.","major_comments":[{"comment":"The stability proof for the explicit scheme X^{(n+1)} = (2I + τ²L_a)X^{(n)} − X^{(n−1)} with L_a = D^{−1/2}AD^{−1/2} is incorrect. The proof reduces the second-order recurrence to the characteristic equation λ² − λ′λ + 1 = 0 with λ′ = 2 + τ²μ, μ ∈ [−1,1], and then analyzes only the smaller real root λ₂ for λ′ > 2, together with the complex-root case. For every graph with at least one edge, L_a has an eigenvalue μ = 1 (with eigenvector D^{1/2}1), so λ′ = 2 + τ² > 2 for any τ > 0. In that case the larger real root λ₁ = (2 + τ² + √(τ⁴ + 4τ²))/2 is strictly greater than 1, giving ρ(C) > 1. Thus the scheme is not stable for any positive time step; it diverges exponentially for generic initial data, contradicting the theorem's statement of constant stability.","section":"Appendix A.3, Theorem 2"},{"comment":"The same omission invalidates Theorem 3 for the frequency-adaptive Laplacian. For L_{a,·} = εI ± D^{−1/2}AD^{−1/2} with ε ∈ (0,1), the eigenvalue bound is λ′ ∈ [2 + τ²(ε−1), 2 + τ²(ε+1)]. Since the normalized adjacency has eigenvalue 1, λ′ reaches 2 + τ²(ε+1) > 2 for τ > 0, and the characteristic equation again has a real root λ₁ > 1, so ρ(C) > 1. Additionally, the GWN-fa recurrence actually implemented, Eq. (18), includes the time-dependent attention matrix α^(n) and the ε^(n)X^(0) term; the constant-coefficient stability analysis of Theorem 3 does not apply to this deployed model.","section":"Appendix A.5, Theorem 3"},{"comment":"The graph wave equation is written as ∂²X/∂t² = a²LX with L = D − A. Since D − A is positive semidefinite, the continuous PDE itself has exponentially growing modes for every positive eigenvalue of L. A standard, well-posed wave equation on a graph requires the Laplacian to have nonpositive spectrum (or the sign in front of L to be negative). This sign choice is not merely a convention: it is exactly what makes both the continuous problem and the discrete scheme unstable, and it is consistent with the failure of Theorem 2.","section":"Proposition 1 / Eq. (11)"},{"comment":"The empirical stability analysis does not support the claimed constant stability. For example, Table 8 reports Texas 48/32/20 accuracy varying from 89.85 (sym-0.2) to 82.75 (sym-5.0), and Cornell 60/20/20 accuracy varying from 89.57 (sym-2.0) to 84.26 (sym-1.0). These are substantial swings across the tested time-step values. The observed behavior is better explained by the bounded terminal time T, early stopping, and the specific hyperparameter tuning than by an unconditional stability property of the scheme.","section":"Section 4.4 and Tables 7–8"}],"minor_comments":[{"comment":"The word \"Caes\" appears twice and should be \"Case\".","section":"Appendix A.3 and A.5"},{"comment":"The figure is referenced to justify assertions about the real roots λ₁ and λ₂, but the text never explains which root is plotted in each panel. The caption should state both roots and explicitly show why λ₁ > 1 for λ′ > 2, which is exactly the case the proof omits.","section":"Figure 7"},{"comment":"The initial-value notation is inconsistent: Eq. (13) defines φ₀(X) and φ₁(X), while the text after Eq. (27) refers to \"φ₁(X) and φ₂(X)\".","section":"Eqs. (13) and (27)"},{"comment":"The word \"Datesets\" in table headers is a typo; also references [34] and [35] are identical and should be merged or corrected.","section":"Tables 2 and 3"},{"comment":"The objects L_{a,l} and L_{a,h} are called \"Laplacians\", but they are affine combinations of the identity and the normalized adjacency matrix, not Laplacian matrices in the standard graph-theoretic sense. Consider renaming them \"filters\" or \"operators\" to avoid confusion.","section":"Eq. (17)"}],"recommendation":"reject","confidential_remarks":"The false stability theorem is the central contribution of the paper, and the error is not a local proof gap: the proposed scheme is unconditionally unstable for every graph with at least one edge. Because the efficiency and robustness claims in Sections 3.5, 4.4, and the abstract all rest on Theorem 2, the theoretical core is invalid. The experimental results may still indicate a useful heuristic, but the manuscript would need a substantially revised theoretical framing (e.g., an explicit conditional stability condition or a different operator with nonpositive spectrum) before it could be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: the paper's central theorem is false, and the \"constantly stable\" claim that motivates the whole efficiency story collapses. The proof of Theorem 2 in Appendix A.3 only ever bounds the smaller root |λ2| and ignores λ1. For any positive eigenvalue μ of D^{−1/2}AD^{−1/2}—and every graph with at least one edge has one—λ′ = 2 + τ²μ > 2 for any τ>0, and then λ1 = (λ′ + √(λ′²−4))/2 > 1. So the spectral radius of the iteration matrix C exceeds 1, and the explicit scheme is not unconditionally stable. The paper's own Eq. (36) contains exactly this case; the proof skips it. The same flaw invalidates Theorem 3, which also treats a constant Laplacian while the deployed GWN-fa uses time-dependent attention in Eq. (18) and an εX(0) term.\n\nThat said, the empirical work is real. The authors ship code, use held-out splits, report means and stds over 10 runs, and compare against a wide set of baselines. GWN-fa's results on heterophilic datasets are genuinely strong, and the over-smoothing experiments are informative even if the stability analysis is wrong. The frequency-adaptive attention layer is a new design, even though second-order ODE and hyperbolic PDE message passing already exist in GraphCON and PDE-GCN.\n\nBeyond the theorem, the operator in Eq. (15) is the normalized adjacency, not the normalized Laplacian, so the paper's framing is misleading. And the stability proof never actually applies to the model that runs, because the Laplacian changes each layer via attention.\n\nFor a referee: yes, this deserves serious refereeing. The empirical contribution is substantive and the flaw is instructive. But as it stands the central claim is unsupported. The authors would need to remove or thoroughly rewrite the stability theorems and reposition GWN as a heuristic with good empirical performance. Without that, the paper should not be published with the constant-stability claim intact.\n\nI'd bring it to reading group precisely to discuss how a stability proof can look plausible and fail. I wouldn't cite it in my own work until the claim is corrected.","headline":"The paper's central stability theorem is false, but the empirical GWN-fa design and careful benchmarking are worth a second look.","tokens_in":29759,"tokens_out":4713,"would_cite":false,"duration_ms":49665,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that modeling GNN message passing as wave propagation—not heat diffusion—yields a two-step explicit scheme stable at any time step, enabling faster training and stronger accuracy on homophilic and heterophilic graphs.","keywords":["graph neural networks","wave equation","message passing","spectral graph convolution","PDE-based learning","stability analysis","heterophily","over-smoothing"],"falsifier":"Compute the spectral radius of $C = \\begin{pmatrix} 2I + \\tau^2 D^{-1/2}AD^{-1/2} & -I \\\\ I & 0 \\end{pmatrix}$ for any connected graph (a 10-cycle suffices) at $\\tau = 1$: the eigenvalue $\\mu = 1$ of $D^{-1/2}AD^{-1/2}$ yields an eigenvalue $(2 + \\tau^2 + \\sqrt{(2+\\tau^2)^2 - 4})/2 > 1$ of $C$, so the iteration grows by that factor per step; running the recurrence from a generic initial condition confirms the norm explosion, which would refute \"constantly stable for any $\\tau$\".","tokens_in":28440,"feed_emoji":"🌊","tokens_out":12057,"duration_ms":111866,"temperature":0.7,"pith_summary":"This paper argues that message passing in graph neural networks should be modeled as wave propagation rather than heat diffusion, because graph signals are superpositions of eigenvector waves. It derives a graph wave equation from the physics wave equation, discretizes it with the forward Euler method into a two-step explicit recurrence, and claims this recurrence is constantly stable—stable for every positive time step $\\tau$. If the claim holds, wave-based GNNs can take far larger time steps than heat-equation models such as GRAND, cutting training cost while keeping or improving accuracy. The paper reports that the resulting networks, GWN-sym and GWN-fa, beat prior methods on homophilic and heterophilic benchmarks, stay accurate as layer count grows, and mitigate over-smoothing.","feed_headline":"Treating graph message passing as wave motion beats heat diffusion","feed_subtitle":"A two-step wave recurrence tolerates large time steps, so training runs faster without losing accuracy.","key_machinery":"The central object is the graph wave equation $\\frac{\\partial^2 X}{\\partial t^2} = L_a X$, in which the graph Laplacian replaces the spatial derivative operator of the physics wave equation, with $L_a = a^2(D - A)$ in discrete form (Proposition 1). The argument is carried by the explicit two-step scheme $X^{(n+1)} = (2I + \\tau^2 L_a)X^{(n)} - X^{(n-1)}$, recast as a first-order iteration $U^{(n+1)} = C U^{(n)}$ with iteration matrix $C = \\begin{pmatrix} 2I + \\tau^2 L_a & -I \\\\ I & 0 \\end{pmatrix}$; Theorem 2 claims that the spectral radius of $C$ never exceeds 1 for $L_a = D^{-1/2}AD^{-1/2}$. That spectral-radius bound is what licenses arbitrarily large time steps and, with them, the paper's efficiency and robustness claims.","core_discovery":"The central claim, stated on the paper's own terms, is that a second-order wave equation on the graph is the faithful description of message passing, and that its explicit forward-Euler solution $X^{(n+1)} = (2I + \\tau^2 L_a) X^{(n)} - X^{(n-1)}$ is stable for every $\\tau \\in \\mathbb{R}_+$ whenever $L_a = D^{-1/2}AD^{-1/2}$ has eigenvalues in $[-1,1]$ (Theorem 2). The paper connects the wave equation to spectral GNNs by letting the Laplacian play the role of the spatial operator and the wave speed act as either a constant or learnable parameters, yielding two architectures: GWN-sym, built on the symmetric normalized adjacency, and GWN-fa, built on a frequency-adaptive Laplacian with signed attention weights that act as low-pass and high-pass filters. The payoff is that stability no longer constrains the time step, so larger $\\tau$ means faster training without an accuracy penalty, and the depth of the network can grow without over-smoothing.","pith_inferences":["Editorial extension: the explicit scheme is the classical leapfrog (Verlet) integrator for second-order dynamics, so the layer stack conserves a discrete energy; tracking that energy during training could serve as an early warning for when a chosen $\\tau$ leaves the empirically stable range.","Editorial extension: the initial velocity term $\\varphi_1(X)$ in Eq. (13) injects momentum; on temporal graphs, feeding the embedding change from the previous snapshot as $\\varphi_1$ would yield a natural wave-based model for evolving networks, which the paper does not explore.","Editorial extension (testable): the paper's accuracy-versus-$\\tau$ curves are flat on large datasets, while smaller heterophilic datasets show a drop at the largest tested step ($\\tau = 5$); mapping where that drop begins, relative to the top eigenvalue of $D^{-1/2}AD^{-1/2}$, would pin down the practical operating regime for large-$\\tau$ training."],"forward_implications":["If Theorem 2 holds, the time step $\\tau$ can be made as large as the solver needs, so the number of iterations—and with it training time—can be cut far below what heat-diffusion models require, which are stable only for tiny steps such as $\\tau = 0.005$.","Deeper networks become safe: more terminal time $T$ no longer threatens numerical blow-up or over-smoothing, matching the paper's layer-depth experiments on Cora, CiteSeer, and PubMed.","Because the equation is defined through the Laplacian, any spectral GNN expressed as a Laplacian filter can be lifted into wave form; the paper demonstrates this by turning GCN into GWN-sym and FAGCN into GWN-fa, each improving on its base model.","Signed attention in GWN-fa provides simultaneous low-pass and high-pass behavior, which the paper's heterophilic benchmarks show as large gains over diffusion-based and standard GNN baselines."],"supporting_citations":[{"why":"Supplies the heat-diffusion paradigm the paper replaces, the forward Euler discretization it borrows, and the conditional-stability baseline (stable only for small time steps) that the constant-stability claim is defined against.","marker":"[6]"},{"why":"Provides the graph signal processing view that graph signals are linear combinations of eigenvector waves, which motivates treating message passing as wave propagation.","marker":"[39]"},{"why":"GCN is the base model of GWN-sym; its symmetric normalized adjacency defines the Laplacian $L_a = D^{-1/2}AD^{-1/2}$ used in Eq. (16).","marker":"[20]"},{"why":"FAGCN is the base model of GWN-fa; its low-pass and high-pass filter design is extended into the time-dependent frequency-adaptive Laplacian of Eq. (17).","marker":"[3]"},{"why":"GraphCON supplies the second-order ODE baseline and the layer-depth evaluation protocol used to show that GWN mitigates over-smoothing.","marker":"[36]"}],"fun_headline_variants":["Wave equation replaces heat diffusion for stable, fast GNN message passing","Graph wave networks: second-order dynamics beat first-order heat flow","Wave-based message passing stabilizes GNNs and speeds training","From heat to waves: a stable GNN layer that trains faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a two-step recurrence built from any normalized-adjacency eigenvalue between $-1$ and $1$ stays bounded; a graph whose normalized adjacency has an eigenvalue near $+1$ instead produces a growing recurrence at every positive time step, so that premise is what the claim stands on.","fun_headline_variants_meta":{"raw":{"variants":["Wave equation replaces heat diffusion for stable, fast GNN message passing","Graph wave networks: second-order dynamics beat first-order heat flow","Wave-based message passing stabilizes GNNs and speeds training","From heat to waves: a stable GNN layer that trains faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1626,"prompt_tokens":1078,"completion_tokens":548,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":694,"completion_tokens_details":{"reasoning_tokens":475}},"tokens_in":694,"tokens_out":548,"duration_ms":5367,"temperature":1.0,"reasoning_tokens":475,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:01:17.136034+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the spectral radius of $C = \\begin{pmatrix} 2I + \\tau^2 D^{-1/2}AD^{-1/2} & -I \\\\ I & 0 \\end{pmatrix}$ for any connected graph (a 10-cycle suffices) at $\\tau = 1$: the eigenvalue $\\mu = 1$ of $D^{-1/2}AD^{-1/2}$ yields an eigenvalue $(2 + \\tau^2 + \\sqrt{(2+\\tau^2)^2 - 4})/2 > 1$ of $C$, so the iteration grows by that factor per step; running the recurrence from a generic initial condition confirms the norm explosion, which would refute \"constantly stable for any $\\tau$\".","supporting_citations":[{"cited_title":"Gorinova, Michael M","cited_arxiv_id":null,"evidence_quote":"Supplies the heat-diffusion paradigm the paper replaces, the forward Euler discretization it borrows, and the conditional-stability baseline (stable only for small time steps) that the constant-stability claim is defined against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FAGCN is the base model of GWN-fa; its low-pass and high-pass filter design is extended into the time-dependent frequency-adaptive Laplacian of Eq. (17)."},{"cited_title":"Konstantin Rusch, Ben Chamberlain, James Rowbottom, Siddhartha Mishra, and Michael M","cited_arxiv_id":null,"evidence_quote":"GraphCON supplies the second-order ODE baseline and the layer-depth evaluation protocol used to show that GWN mitigates over-smoothing."}],"review_version":1}