{"id":"c780e340-eaf0-4f63-b4da-b8860b1954f4","arxiv_id":"2505.14033","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A partition-wise graph filtering method, CPF, unifies graph-wise and node-wise filtering and achieves state-of-the-art node classification on 13 benchmark graphs and anomaly detection on 3 datasets.","lead":"This paper introduces CPF, a graph neural network that applies different filters to groups of nodes rather than one filter for the whole graph or one filter per node. It claims better accuracy on both homophilic and heterophilic graphs by using graph coarsening to form the groups.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.2's proof applies the RSA inequality to signals outside the subspace R, so at r=0.5 the structural-equivalence claim is unsupported for arbitrary features.","rationale":"The paper has real strengths: the method is clearly specified, code is available at the stated repository, and the empirical comparisons are extensive and consistently favorable. Proposition 3.3 is a straightforward but correct unification of graph-wise and node-wise filtering as extreme cases, and the ablation studies support the practical value of both filtering stages. The load-bearing weakness is in the theoretical analysis, not the experiments. The reader identified the RSA/epsilon assumption and the unproven norm bound in Appendix B as the weakest point; my analysis agrees but locates the defect more precisely: the proof of Theorem 3.2 applies an inequality that is restricted to a subspace R to arbitrary signals Delta^{k-1} x. This is not merely a missing epsilon measurement; it is a quantitative failure of the argument as stated. Even if the claim happens to hold in practice for some datasets, the theorem does not prove it for arbitrary graph signals. Empirically, r=0.5 is fixed across all datasets and epsilon is never reported, so the condition required by the theorem is unchecked. Thus the theoretical unification claim should be regarded as conditional pending either a corrected proof with explicit assumptions on x, or a numerical demonstration that the relevant epsilon is small for the actual coarsenings and feature subspaces used. Because the reader's verdict was already CONDITIONAL with requests to fix or qualify the theoretical statements, my concern does not change the verdict; it strengthens the reason for that condition. I found no basis to reject the empirical claims, and no evidence of bad faith or sloppy experimentation.","tokens_in":34657,"tokens_out":7707,"duration_ms":85851,"concrete_test":"Run a numerical check on Cora and ogbn-Arxiv using the released code with the LV coarsening at r=0.5. Extract the subspace R actually used by the coarsening routine (or, if unspecified, the span of the top-k eigenvectors used by the RSA optimizer). Sample 1000 random signals x from a standard Gaussian in R^n and another 1000 from R. For k=2,3,4, compute the left-hand side of Eq. (5), ||Delta^k x - Pi Delta Pi Delta^{k-1} x||_L, and compare it to the claimed right-hand side, epsilon ||x||_L (||Delta||_L + ||Pi Delta||_L), with epsilon computed as in Definition 2.1. Also compute ||Delta^k x - (Pi Delta Pi)^k x||_L to test Eq. (6). If any x not in R violates the inequality, Theorem 3.2 as stated is false; if violations occur only outside R, the theorem must be restricted to x in R, which would not cover typical node features.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical justification for structure-aware partition-wise filtering is Theorem 3.2, whose proof (Appendix B) is invalid as written. The RSA constant in Definition 2.1 bounds reconstruction error only for signals in a subspace R: epsilon_{L,C,R} = sup_{x in R, ||x||_L=1} ||x - C^+ C x||_L. The proof starts from the cited inequality (17), ||Delta x - Pi Delta Pi x||_L <= epsilon ||x||_L (||Delta||_L + ||Pi Delta||_L), which is only guaranteed for x in R. It then replaces x with Delta^{k-1} x and explicitly states 'x is assumed to be in R^n' (the whole space). Unless R = R^n, Delta^{k-1} x need not lie in R, so the bound in Eq. (5) does not follow. For node feature vectors that are not contained in the low-dimensional spectral subspace R -- precisely the high-frequency content relevant to heterophilic graphs -- the R-perpendicular component is uncontrolled, and the claimed approximation ||Delta^k x - (Pi Delta Pi)^k x||_L -> 0 in Eq. (6) is unsupported. The induction step (24) repeats the same substitution, again relying on the unproved inequality for signals outside R. Since r=0.5 is fixed for all 13 datasets and epsilon is never measured, the structural equivalence between original-graph propagation and coarsened-graph propagation -- the core motivation for sharing filters within partitions -- rests on an unverified condition. The empirical method may still work, but the paper's theoretical unification claim is not established.","agreement_with_reader":"partial"},"referee_report":null,"author_rebuttal":null,"desk_editor":null,"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-07T15:41:37.651996+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}