{"id":"90cfb99e-fe84-4f08-b2c4-2b4c98a1ea2d","arxiv_id":"2607.15602","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Searching over fractional Fourier orders before graph-signal sampling reduces reconstruction error on the tested signals, but the optimal order is chosen using the true signal itself.","lead":"This paper proposes sampling and reconstructing graph signals in a fractional Fourier domain, tuning a fractional order parameter to pick the best spectral basis for each signal. The idea is that a better-chosen basis concentrates signal energy and could improve recovery from only a fraction of node samples.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fractional-order selection uses ground-truth x (Eq. 32, Algorithm 1), so the reported GFRFT gains are in-sample oracle bounds, not evidence that a practical reconstruction method outperforms GFT.","rationale":"The reader's weakest_assumption identifies the same load-bearing flaw: Eq. (32) and Algorithm 1 select the fractional order by minimizing reconstruction error against the true signal x. This is an oracle selection rule. The paper's headline claim—'GFRFT domain sampling and reconstruction generally achieve better recovery performance than GFT domain methods'—is meant as a practical statement about graph-signal reconstruction, where unobserved nodes are exactly those whose values are being estimated. The reported improvements are in-sample, oracle-type lower bounds, not achievable by a procedure that only has access to y. The theory in Theorem 1 is internally consistent but conditional; it does not propose or validate a data-driven way to choose a. The experiments therefore do not support the central claim as stated. I agree with the reader's rejection, while acknowledging the algebraic framework and conditional theory could be rehabilitated by an operational order-selection rule and out-of-sample validation.","tokens_in":17278,"tokens_out":4641,"duration_ms":60054,"concrete_test":"Re-run the PeMS-BAY experiment with a* selected solely from the observed samples y: for each candidate a, reconstruct from the K sampled nodes, then choose a* by leave-one-out cross-validation over those K samples (or by holding out a random subset of sampled nodes for order selection). Evaluate the final MSE on the unobserved nodes and compare it against a=1 and against the oracle a* from Eq. (32). If the data-driven a* yields ΔMSE ≈ 0 or negative while the oracle ΔMSE is positive, the reported improvements are selection artifacts and the central practical claim fails.","verdict_should_be":"REJECT","load_bearing_attack":"The central empirical claim—that GFRFT-domain sampling and reconstruction 'generally achieve better recovery performance' than GFT—rests on selecting the fractional order by Eq. (32): a* = argmin_a ||x - x_hat_a||^2. Algorithm 1 requires the true signal x as input and returns the order whose reconstruction is closest to it. The reported ΔMSE = MSE_{a=1} - MSE_{a=a*} is therefore the minimum over 51 candidate orders of the same in-sample error later reported as the gain. Because the candidate family is flexible enough to overfit the reconstruction error, this procedure will typically beat the fixed a=1 baseline even when no rule based only on the observed samples y could identify a better domain. The paper explicitly acknowledges that ground truth is used ('In supervised experiments, the ground truth graph signal is available'), but graph-signal reconstruction from partial observations is exactly the setting where x is unknown on unobserved nodes. Theorem 1 gives a conditional sufficient condition for an oracle-aided advantage; it does not supply an operational selection rule from y. The real-data experiment uses one time step and one graph with no error bars. Thus the experiments establish only that the best fractional order in hindsight can reduce error, not that the proposed method achieves such reductions in practice.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a graph signal sampling and reconstruction framework in the graph fractional Fourier transform (GFRFT) domain. The fractional order a is treated as an adjustable spectral-domain parameter, and the authors formulate a unified sampling-correction-reconstruction pipeline under subspace, smoothness, and stochastic priors, with both unconstrained and predefined reconstruction modes. They provide theoretical analysis showing that a suitable fractional subspace can reduce projection residual, while residual leakage and noise amplification must be controlled; Theorem 1 gives a sufficient condition for the fractional-domain reconstruction to beat the GFT-domain reconstruction. Experiments on simulated and real traffic signals compare a fixed order a=1 with an 'optimal' order a* selected by minimizing the reconstruction error against ground truth.","tokens_in":17588,"tokens_out":3356,"duration_ms":41434,"significance":"The theoretical framework is systematic and the algebraic derivations in the appendices are internally consistent; Theorem 1 provides a valid conditional sufficient condition. The idea of adapting the spectral representation to the signal via a fractional transform is of potential interest. However, the key empirical claim that GFRFT-domain methods 'generally achieve better recovery performance' is not supported because the optimal order is selected using ground truth, turning all reported improvements into in-sample oracle gains. Without an operational selection rule that uses only the observed samples, the practical significance of the framework remains unestablished.","major_comments":[{"comment":"The optimal order a* is defined as the minimizer of ||x - x_hat_a||^2 using the true signal x, and Algorithm 1 selects a* by evaluating reconstruction error against ground truth. The same MSE is then reported in Tables II–V as the improvement over a=1. This makes the measured improvements in-sample oracle results; they are a consequence of the selection rule, not evidence that the method can outperform GFT in a practical reconstruction task where x is unknown on unobserved nodes. The abstract's claim of 'generally better recovery performance' is therefore not supported.","section":"§III-F, Eq. (32), Algorithm 1"},{"comment":"Theorem 1 gives a sufficient condition (51) involving ||r_{K,a}||, tau_a, and nu_a. These quantities depend on the unknown signal x and on the chosen fractional subspace, and the paper provides no way to estimate them or to choose a* from the observed samples y_a alone. The theorem thus characterizes when a hindsight oracle would benefit, but it does not supply an operational order-selection rule. To connect theory to practice, the authors need either a data-driven estimator of the condition or a practical criterion based only on y.","section":"§IV-B, Theorem 1"},{"comment":"The real-data experiment uses one traffic signal at a single time step, with no error bars, repeated trials, or statistical significance tests. Many reported improvements are below 0.01 dB (e.g., SS PD G2: 0.0010 dB; ST PD G2: 0.0008 dB) and are likely within sampling and measurement variability. This does not establish that GFRFT-domain methods 'generally' outperform GFT on real data. Please provide confidence intervals, multiple time steps, or a significance analysis, and also show the a=1 baseline in the sensitivity analysis of Table V.","section":"§V-C, Table IV"}],"minor_comments":[{"comment":"The symbol U_{K,a} is used both for the K-dimensional subspace and for an orthonormal basis of that subspace. This is confusing; please use distinct notation, e.g., V_{K,a} for the basis.","section":"§IV-A, Definition 1"},{"comment":"Typographical error: 'aaainduced' should be 'induced'.","section":"Appendix A, Prop. 2 proof"},{"comment":"The tables do not report standard deviations or the number of random trials. Fixed random seeds ensure reproducibility, but variability across seeds or sampling sets is not quantified.","section":"§V, Tables II–V"}],"recommendation":"major_revision","confidential_remarks":"The central empirical result is oracle-based, so the paper's main claim is not yet supported. If the authors can add a practical order-selection rule and redo the experiments, the contribution could be significant. Otherwise, the manuscript may need to be reframed as an analysis of oracle performance bounds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — quick take on arXiv:2607.15602. The paper builds a GFRFT-domain version of generalized sampling for graph signals, covering subspace, smoothness, and stochastic priors, each in unconstrained and predefined recovery modes. The formulas in Propositions 1–3 and the predefined variants are new to this setting; they are not verbatim in Wei and Yan. The appendices check out: the least-squares derivations are consistent, and Theorem 1 is a valid sufficient condition for the fractional domain to beat the GFT domain in terms of expected MSE. That theorem is honestly stated as a conditional result, and the decomposition into projection residual, leakage, and noise amplification is genuinely useful.\n\nThe soft spot is large and central. The order selection in Eq. (32) is a* = argmin_a ||x − x̂_a||², and Algorithm 1 takes the true graph signal x as required input. So the reported improvement ΔMSE is the minimum over 51 candidate orders of the very error that is later reported as the gain. That is an in-sample oracle bound. It does not demonstrate that a practical reconstruction method using only the sampled y can find a better domain than a=1. The paper itself says 'In supervised experiments, the ground truth graph signal is available' — which is honest but does not rescue the claim. Theorem 1 gives a condition on the true signal, not a way to pick a* from y. Without an operational selection rule, the experiments establish only that hindsight helps.\n\nThe real-data section is also thin: one graph, one time step, no error bars, and the baseline comparisons are limited. Some GFT-domain methods get a*=1, which is fine, but the ST UNC gains of ~0.9 dB on real data are exactly the cases where a* = 0.04, far from 1, so the sensitivity to the oracle is especially visible.\n\nWould I send this to review? Yes — the theoretical framework is clean, the conditional theorem is real, and the flaw, while damaging to the empirical claim, is fixable in principle by designing a selection rule from y (e.g., prediction error on held-out samples, or an order selector based on a criterion like energy concentration estimated from the observations). A serious referee should ask for exactly that. I wouldn't cite it for the 'GFRFT generally outperforms GFT' claim, but the conditional analysis might be citable if you work on fractional graph transforms.","headline":"Mathematically clean GFRFT-domain sampling framework whose headline 'optimal order' experiment is an oracle selection on the true signal; the conditional theory is valid but the practical claim is unproven.","tokens_in":18050,"tokens_out":2243,"would_cite":false,"duration_ms":27417,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Sweeping the fractional order lowers graph-signal reconstruction error","keywords":["graph signal processing","graph fractional Fourier transform","graph signal sampling","graph signal reconstruction","fractional order selection","energy concentration","projection residual","generalized sampling"],"falsifier":"Take a graph-signal dataset, split into training/validation/test sets, choose the fractional order using only validation data, and compare test-set reconstruction error against the GFT case a = 1. If the validation-selected order does not beat a = 1 consistently, the central claim is falsified. Alternatively, use a pure graph-Fourier eigenvector as the signal: for that signal the GFT is already maximally concentrating, so ρ_K(a) ≤ ρ_K(1) for all a, and any reported improvement would contradict Theorem 1.","tokens_in":17153,"feed_emoji":"🎛️","tokens_out":6434,"duration_ms":67957,"temperature":0.7,"pith_summary":"Graph signals are often sampled and reconstructed in the graph Fourier transform (GFT) domain, but the GFT basis is fixed by the graph and many real signals are not concentrated in it. This paper proposes instead to work in the graph fractional Fourier transform (GFRFT) domain, where a fractional order tunes the spectral representation continuously. The paper argues that choosing the order that best concentrates a signal's energy reduces the low-dimensional projection residual, and it proves a sufficient condition under which the fractional-domain reconstruction has strictly smaller expected error than any GFT-domain reconstruction. Experiments on simulated and real traffic-sensor signals show that the optimized order usually lowers reconstruction error relative to the GFT case. The contribution matters because it turns spectral-domain choice into a tunable parameter inside a unified sampling-correction-reconstruction framework.","feed_headline":"Tuning one spectral knob lowers graph signal recovery error","feed_subtitle":"Searching the fractional Fourier order per signal beats the fixed graph Fourier domain in most tested cases.","key_machinery":"The central object is the GFRFT matrix F_a = V Λ_F^a V^{-1}, fractional powers of the GFT eigenvalues. The scalar a is an adjustable spectral-domain parameter: a = 1 is the GFT; other values rotate the spectral basis continuously. For fixed K, each a defines a K-dimensional fractional spectral subspace with projector P_{K,a}; the energy concentration ratio ρ_K(a) = ||P_{K,a}x||^2/||x||^2 captures how much signal energy is kept. Theorem 1 carries the argument: it decomposes expected reconstruction error into projection residual ||r_{K,a}||^2, residual-leakage coefficient τ_a = ||A^†_a S^*_a (I - P_{K,a})||_2, and noise amplification ν_a = σ^2||A^†_a||_F^2, and states an explicit condition und","core_discovery":"The paper claims that the graph fractional Fourier transform (GFRFT) with a carefully chosen fractional order gives a better low-dimensional spectral representation for sampling and reconstruction than the fixed GFT. There often exists an order a such that the K-dimensional fractional subspace captures more energy and leaves a smaller projection residual r_{K,a} than the GFT subspace. Theorem 1 makes this precise: if (1 + τ_a^2)||r_{K,a}||^2 + ν_a < ||r_{K,1}||^2, then the expected error of the fractional-domain least-squares reconstruction is strictly below any reconstruction confined to the K-dimensional GFT subspace. τ_a is residual leakage through sampling inversion; ν_a is noise amplifi","pith_inferences":["The experiments select the optimal order by comparing reconstructions against the ground-truth signal (Eq. 32 and Algorithm 1), so the reported improvements are oracle gains; a practical version would need a validation-based or unsupervised order-selection rule, and the observed margins may shrink.","The fractional order can be viewed as a cheap way to adapt the spectral dictionary to each signal; the same tuning idea could be transferred to graph filtering, denoising, or classification tasks.","Theorem 1 suggests a practical diagnostic: before reconstructing, compute ρ_K(a), τ_a, and ν_a on a candidate grid and flag orders that satisfy the inequality; this could predict when GFRFT will help without knowing the test signal.","The large improvement of the stochastic unconstrained method at a* ≈ 0.04 on traffic data hints that real graph signals have low-rank covariance structure poorly aligned with the Laplacian eigenvectors; testing on more datasets would show whether such extreme orders recur."],"forward_implications":["For signals whose energy is spread across the GFT spectrum, searching over fractional orders can reduce reconstruction error at the same sampling rate.","The optimal fractional order is not fixed: it depends on the signal, the sampling count, the spectral response, and the prior, so the method is genuinely adaptive.","The gain is conditional: Theorem 1 shows that an order which improves energy concentration can still hurt if it increases residual leakage or noise amplification after sampling inversion.","The same unified sampling-correction-reconstruction chain covers subspace, smoothness, and stochastic priors, so the order-selection rule applies across different recovery models.","On real traffic data, the stochastic unconstrained variant shows the largest gains, suggesting that mismatches between the signal covariance and the GFT spectrum are common in practice."],"fun_headline_variants":["One spectral knob: lower graph recovery error","Fractional order tuning cuts graph signal error","Spectral knob: choose the right order, less error","Tune fractional order for lower graph sampling error"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The experiments choose the optimal fractional order using the true complete signal to compute reconstruction error, so the reported 'optimal' gains are in-sample oracle results; when the true signal is unknown, the selection rule in Eq. (32) cannot be applied, and the practical advantage over the GFT is not established.","fun_headline_variants_meta":{"raw":{"variants":["One spectral knob: lower graph recovery error","Fractional order tuning cuts graph signal error","Spectral knob: choose the right order, less error","Tune fractional order for lower graph sampling error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000334,"raw_usage":{"total_tokens":1681,"prompt_tokens":725,"completion_tokens":956,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":897}},"tokens_in":469,"tokens_out":956,"duration_ms":8597,"temperature":1.0,"reasoning_tokens":897,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T22:48:52.179202+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a graph-signal dataset, split into training/validation/test sets, choose the fractional order using only validation data, and compare test-set reconstruction error against the GFT case a = 1. If the validation-selected order does not beat a = 1 consistently, the central claim is falsified. Alternatively, use a pure graph-Fourier eigenvector as the signal: for that signal the GFT is already maximally concentrating, so ρ_K(a) ≤ ρ_K(1) for all a, and any reported improvement would contradict Theorem 1.","supporting_citations":[],"review_version":1}