{"id":"4cae69ba-1cf5-46e7-9969-d49f291e4ffc","arxiv_id":"2607.23277","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A GLRT-based detector for RIS-aided spectrum sensing under correlated noise estimates the unknown PU-RIS channel and transmit power, then uses the optimally tuned RIS to outperform energy detection.","lead":"This paper designs a smarter radio-listening method: a programmable mirror-like surface (RIS) is tuned to reflect a primary user's signal toward a secondary listener, and a statistical test (GLRT) then decides whether that signal is present in correlated noise. The method estimates the unknown reflection channel and transmit power on the fly, and simulations show it beats a simple energy counter, especially with correlated noise and few samples.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Phase-1 channel identifiability and training protocol are not established; default simulation parameters violate Lbar<=N and H1-only estimation, so the claimed GLRT gains are unsupported until these are fixed.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the estimation phase is only meaningful under H1 and with full column rank of the per-group sensing matrix. I agree with that assessment. Appendix A and B derivations are formally correct under those assumptions, but the manuscript's numerical section violates them: the default L=512, G=16, N=8 gives Lbar=32>N, and Fig. 3's caption (L=16) conflicts with the default. No training protocol for the H0 case is provided. These are fixable, but they touch the core empirical demonstration, so the verdict should remain CONDITIONAL rather than ACCEPT or REJECT. The proposed computational check would determine whether the claimed superior detection probability survives when the stated constraints are respected and when the PU is silent during Phase 1.","tokens_in":9020,"tokens_out":10913,"duration_ms":99114,"concrete_test":"Reproduce Fig. 3 using the stated default parameters (N=8, L=512, G=16, hence Lbar=32) with the exact group-wise model and the ML estimator (9)-(12). If the matrix pair (A_g^H R A_g, A_g^H A_g) is singular or the GLRT statistic (20) is undefined, the reported ROC is an artifact of the L=16 caption. Then, with a valid configuration (e.g., G=64 so Lbar=8), repeat the ROC after replacing Phase-1 data with H0-only noise; if the optimized RIS gives no improvement over random phases, the training protocol is circular.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the GLRT with jointly estimated PU-RIS channel and optimized RIS beats ED. This rests on the Phase-1 estimates of h_g and sigma_x^2 being valid, which requires (i) A_g = W G_g Phi_g to have full column rank (Lbar <= N) for the generalized eigenvalue problem in (9) to be well-posed, and (ii) Phase-1 observations to be generated under H1. The paper states an L<=N constraint in Section II but not the per-group Lbar<=N requirement, and its default parameters (N=8, L=512, G=16) give Lbar=32>N, making (9) singular; Fig. 3's caption says L=16, contradicting the default. The detection model in (18b) uses the same full A* with L=512>N, so the GLRT statistic (20) inherits a singular pencil unless the group-wise Lbar<=N formulation is applied consistently. Moreover, the two-phase protocol never specifies how the SU knows the PU is active during estimation: under H0, the ML estimates are noise-dominated and the 'optimal' RIS phase matrix (16) is uninformative. Since these conditions are prerequisites for the RIS configuration used in detection, the ROC advantage in Figs. 3-4 is currently unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies RIS-aided spectrum sensing under correlated Gaussian noise, proposing a generalized likelihood ratio test (GLRT) that estimates the unknown PU-RIS channel and PU transmit power, then optimizes the RIS phase shifts to maximize the estimated effective channel gain. A group-wise estimation scheme is introduced to handle large RISs, and an energy detector (ED) with optimized RIS is also analyzed. Numerical ROC results claim that the proposed GLRT outperforms ED, especially at low SNR and under correlated noise. The appendices provide algebraic derivations of the ML estimates and the GLRT statistic.","tokens_in":9354,"tokens_out":10330,"duration_ms":95148,"significance":"If the technical gaps are repaired, the paper would be a useful contribution: it extends RIS-aided sensing to unknown PU-RIS channels, considers correlated noise, and provides closed-form GLRT and ED statistics together with a grouped processing strategy for large RISs. The algebraic derivations in Appendices A and B are a strength, and the GLRT statistic is self-contained rather than a fitted quantity. However, the manuscript currently contains load-bearing errors and missing assumptions: the rank-one claim in Section IV is false for N>1, the identifiability condition for the per-group ML estimator is not respected by the default simulation parameters, and the two-phase framework presupposes PU activity during parameter estimation without specifying a training protocol. These issues must be addressed before the claimed detection improvements can be considered supported.","major_comments":[{"comment":"The statement that \"\\tilde G^H \\tilde G and \\hat h\\hat h^H are rank-one matrices\" is false for N>1. With the paper's N=8, \\tilde G is 8×L, so \\tilde G^H \\tilde G has rank at most 8, not 1. The correct maximized value of J is λ_max(\\tilde G^H \\tilde G)λ_h, where λ_max is the largest eigenvalue. The phase solution in (16) can still achieve this bound if λ_{\\tilde G} is redefined as the largest eigenvalue, but as written Eq. (17) and the subsequent ED moment expressions in Section V.B inherit the error. Please correct the rank argument and propagate the fix.","section":"Section IV, Eq. (17)"},{"comment":"The ML estimator in Eq. (9) is a generalized eigenvalue problem that is well-posed only if A_g = W G_g Φ_g has full column rank, i.e., \\bar L = L/G ≤ N. The paper only imposes L ≤ N in Section II, not the per-group condition. The stated default parameters N=8, L=512, G=16 give \\bar L = 32 > N, so A_g^H A_g is singular and the \"dominant generalized eigenvector\" is undefined. Figure 2 uses G=8 (\\bar L=64=N), while Figure 3's caption states L=16; the simulation settings are inconsistent. All experiments supporting the claimed ROC gains must use \\bar L ≤ N and report G and L explicitly for each figure.","section":"Section III, Eq. (9) and Section VI defaults"},{"comment":"Phase 1 'channel estimation and RIS configuration' uses the signal model in Eq. (2b), which assumes the PU is active (H1). The paper never states how the SU knows the PU is active during this phase, nor what happens if Phase 1 is run while the PU is silent. Under H0, the estimates from (9)–(12) are noise-dominated, \\hat h is essentially arbitrary, and the optimized phase matrix in (16) is not aligned to the true channel. The detection gains reported in Figs. 3–4 are therefore unsupported unless the authors explicitly define a training protocol with a known PU-active interval or otherwise justify the H1 assumption during estimation.","section":"Section II, two-phase framework"},{"comment":"The detection model in Eq. (18b) uses the full matrix A⋆ = √ν W G Φ⋆ with L columns, whereas estimation is performed group-wise with \\bar L ≤ N. For L > N, which is the default scenario (L=512, N=8), C = A⋆^H A⋆ is singular, and the generalized eigenvalue problem in Eq. (20) is not well-defined without restricting to the range of A⋆^H. Appendix B's maximization of Eq. (39) subject to h^H C h = 1 is ill-posed for singular C. The manuscript should either define the effective rank and domain for the GLRT statistic or perform Phase 2 detection group-wise and combine the per-group statistics.","section":"Section V, Eq. (20)"}],"minor_comments":[{"comment":"The parameter settings are inconsistent: the default text gives N=8, L=512, G=16; Fig. 2 says G=8; Fig. 3 caption says L=16. Please unify the notation and state exactly which (N,L,G,\\bar L) pair is used in each curve.","section":"Section VI, Figs. 2 and 3"},{"comment":"Eq. (28) contains a stray factor 'G' in γ_ED = G μ_0 + σ_0 Q^{-1}(P_FA). The ED statistic in Eq. (23) does not involve the number of groups, and the threshold should be μ_0 + σ_0 Q^{-1}(P_FA).","section":"Section V.B, Eq. (28)"},{"comment":"When only one RIS group is active, the phase-shift matrix Φ_g has zeros outside the group indices and is not unitary on the full L×L space. The unitary constraint (15b) should be stated for the active \\bar L×\\bar L block; otherwise the model is ambiguous.","section":"Section II, group-wise activation"},{"comment":"There are minor typos and grammar issues, e.g., 'V on Neumann' (Von Neumann), 'SU have control' (has), and 'SG g=1 Ig' (union symbol). A careful proofreading pass is needed.","section":"Throughout"},{"comment":"The text mentions a tradeoff between detection performance and computational complexity, but no complexity analysis is given. A short discussion of the cost of the generalized eigen-decomposition and the group-wise estimation would improve the paper.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is combining GLRT with a grouped beyond-diagonal RIS for spectrum sensing under correlated noise and unknown PU-RIS channel. That is a real gap: prior work either assumes the channel is known or only handles white noise. The derivations in Appendices A and B are algebraically correct for the models they address, and the group-wise ML formulation is a sensible way to deal with large RISs. The ED analysis is standard and clean. Credit where it is due: the paper is not sloppy in its math, and the statistical approach is coherent.\n\nThe soft spots are real, but mostly fixable. First, the identifiability constraint. The paper motivates grouping by saying L <= N is impractical, but it never states the per-group condition Lbar <= N that its own generalized eigenvalue problem requires. The default simulation parameters (L=512, G=16, N=8) give Lbar=32 > N, which makes A_g^H A_g singular. Fig. 3's caption says L=16, and Fig. 2 uses G=8 so Lbar=N=64. The text and figures do not agree. If the authors actually ran with Lbar > N, both the ML estimator in (9) and the GLRT statistic in (20) are not well-defined because the matrix pencil is singular. This is a load-bearing inconsistency, not a cosmetic one.\n\nSecond, the rank-one claim in Section IV is simply wrong for N > 1. \\tilde{G}^H \\tilde{G} is not rank-one in general; the maximum effective gain is sigma_max^2(\\tilde{G}) ||h||^2, not the product of two single eigenvalues. This affects the ED analysis and the phase-optimization derivation, and should be corrected.\n\nThird, the two-phase protocol assumes the PU is transmitting during Phase 1. If the PU is silent, the ML estimates are noise-dominated and the \"optimal\" RIS matrix is uninformative. The paper never specifies how the SU knows the PU is active, nor what fallback exists if Phase 1 sees only noise. That is an acknowledged circularity in the framework, and it needs an explicit statement or a workaround.\n\nThe core statistical ideas are sound, and the GLRT derivation is correct under the proper rank condition. The paper deserves a serious referee, but the authors must fix the parameter constraints and reconcile the simulation setup before the claimed detection gains are credible. I would send it to review with a clear request for these revisions, not desk-reject it.","headline":"The GLRT math is largely sound, but the paper's own parameter choices violate the identifiability condition its estimator needs, so the ROC claims are not yet supported.","tokens_in":9808,"tokens_out":3974,"would_cite":false,"duration_ms":37520,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A13","62F03","62H15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A GLRT that jointly estimates the unknown primary-to-RIS channel and transmit power, then configures a grouped RIS to maximize the estimated gain, is claimed to outperform energy detection under correlated noise and few observations.","keywords":["spectrum sensing","reconfigurable intelligent surface","generalized likelihood ratio test","energy detection","correlated noise","maximum likelihood estimation","beyond-diagonal RIS","cognitive radio"],"falsifier":"Reproduce the detection experiment under the paper's stated default parameters N=8, L=512, G=16, giving 32 RIS elements per group against 8 antennas. The matrix A_g^H A_g is singular, so the generalized-eigenvalue MLE in (9) is not well-defined; if the detector still shows a large ROC gain over energy detection, the gain cannot come from the paper's estimation theory. Conversely, run the full two-phase procedure with only noise present during training and check that the resulting phase yields no better detection than random phases.","tokens_in":8879,"feed_emoji":"📡","tokens_out":11712,"duration_ms":88795,"temperature":0.7,"pith_summary":"The paper studies spectrum sensing for cognitive radio with a reconfigurable intelligent surface (RIS) helping the secondary user hear the primary user. Its key move is to remove the usual assumption that the channel between the primary user and the RIS is known. It splits the RIS into groups, estimates each group's channel and the primary transmit power by maximum likelihood from whitened received data, then chooses the RIS phase matrix in closed form to maximize the gain of the estimated channel. With that optimized surface, it builds a GLRT decision statistic from the dominant generalized eigenvalue of a data matrix pencil. In simulated ROC curves, this GLRT consistently beats an energy detector at the same false-alarm rate, and the gap is largest under correlated noise and few observations, exactly the regime where energy detection struggles.","feed_headline":"Smart-surface sensor that estimates channel beats energy detection","feed_subtitle":"New detector learns the hidden channel, steers the surface, and beats energy detection under correlated noise.","key_machinery":"The central mechanism is the grouped beyond-diagonal RIS with sequential activation. Partitioning the L elements into G blocks makes the estimation problem low-dimensional enough for ML: each block's effective matrix A_g = W G_g Φ_g is N × L_bar, so the unknown channel is identifiable only when each group has at most N elements. The mathematical load is carried by two closed forms: the ML estimates as a generalized eigenvector/eigenvalue problem, derived via matrix determinant and Woodbury identities, and the phase-optimization solution Φ* = U_G~ U_h^H obtained from the von Neumann trace inequality. These reduce the whole detection problem to a scalar generalized-eigenvalue test statistic wi","core_discovery":"The paper claims that with grouped, sequential activation of RIS elements, the unknown primary-to-RIS channel and transmit power can be estimated on the fly: per group, the channel estimate is the dominant generalized eigenvector of the pencil formed by the whitened sample covariance and the effective sensing matrix, and the power estimate is the corresponding excess eigenvalue. These estimates feed a closed-form RIS phase design Φ* = U_G~ U_h^H that maximizes the norm of the estimated reflected channel through the von Neumann trace inequality. The resulting GLRT statistic is T(λ̂_max − 1 − ln λ̂_max), where λ̂_max is the dominant generalized eigenvalue under the optimally configured RIS. Th","pith_inferences":["If the estimation phase runs while the primary user is silent, the estimates are noise-driven and the phase solution aligns to noise; a robust design would need a coarse pre-detection trigger or a known pilot, which the paper does not describe.","The identifiability condition L_bar ≤ N is load-bearing; the paper's default simulation parameters (N=8, L=512, G=16) appear to violate it with 32 elements per group, so the claimed ROC gains may not be supported by the MLE theory unless the figure used a compliant grouping.","A natural next experiment is to inject an explicit channel-estimation error model and measure how the optimal-phase GLRT degrades; since the phase is a rotation that aligns to the estimate, small angular errors in the channel estimate could disproportionately hurt the beamforming gain.","The grouped-estimation plus closed-form-phase front-end is agnostic to the downstream test, so it could equally serve maximum-eigenvalue or other covariance-based detectors, making the grouping and phase design the more portable contribution than the GLRT itself."],"forward_implications":["If the claim holds, RIS-aided spectrum sensing no longer requires perfect knowledge of the primary-to-RIS channel, removing the main practical obstacle in previous RIS-sensing designs.","The closed-form phase design lets the hardware switch from estimation mode to sensing mode without iterative optimization, keeping sensing latency short.","Working on whitened data makes the detector specifically effective in correlated-noise environments, where energy detectors need very long observation windows to stabilize their threshold.","The groupwise estimation trick caps the required training dimension: sensing with large RISs is feasible as long as each group stays small enough, but the per-group element count must respect the identifiability limit."],"fun_headline_variants":["RIS-aided GLRT learns channel to beat energy detection","On-the-fly channel estimation powers smarter spectrum sensing","GLRT with adaptive RIS outperforms energy detection in noise","Estimating hidden channel boosts cognitive radio detection","RIS steering from learned channel improves sensing odds"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise, stated by the paper in Section II as needing L ≤ N, is that each group's effective matrix is identifiable: per group, the number of active RIS elements must not exceed the number of receive antennas, and the training data must come from an active primary user. If either requirement fails — e.g., the default N=8, L=512, G=16 gives 32 elements per group — the estimated channel and the RIS phase built from it are not tied to the true channel.","fun_headline_variants_meta":{"raw":{"variants":["RIS-aided GLRT learns channel to beat energy detection","On-the-fly channel estimation powers smarter spectrum sensing","GLRT with adaptive RIS outperforms energy detection in noise","Estimating hidden channel boosts cognitive radio detection","RIS steering from learned channel improves sensing odds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1278,"prompt_tokens":701,"completion_tokens":577,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":503}},"tokens_in":445,"tokens_out":577,"duration_ms":4806,"temperature":1.0,"reasoning_tokens":503,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T23:52:10.490733+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the detection experiment under the paper's stated default parameters N=8, L=512, G=16, giving 32 RIS elements per group against 8 antennas. The matrix A_g^H A_g is singular, so the generalized-eigenvalue MLE in (9) is not well-defined; if the detector still shows a large ROC gain over energy detection, the gain cannot come from the paper's estimation theory. Conversely, run the full two-phase procedure with only noise present during training and check that the resulting phase yields no better detection than random phases.","supporting_citations":[],"review_version":1}