{"id":"c0dc48fc-1bd3-4831-a0ad-102288c90d60","arxiv_id":"2505.05971","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"For parameter estimation over blocked direct links, optimizing a fully connected beyond-diagonal RIS yields provably higher Fisher information than diagonal RIS designs while limiting eavesdropper information.","lead":"A wireless link whose direct path is blocked can estimate parameters more accurately using a beyond-diagonal reflective surface than a conventional diagonal one, even when an eavesdropper's information is capped. The paper gives closed-form and iterative designs for the surface and reports large Fisher information gains in simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'significant enhancement' claim rests on one channel realization and a heuristic diagonal baseline; a Monte Carlo comparison is needed to confirm the gap is robust.","rationale":"The reader already identified the single-realization numerics and heuristic diagonal baseline as part of the weakest assumption, and I agree that this is the most load-bearing concern for the empirical claim. My focus is on the lack of statistical evidence and the possibly suboptimal diagonal baseline, rather than on the physical realizability of arbitrary unitary BD-RIS matrices, which is a standard modeling premise in the BD-RIS literature and applies to the field as a whole. The mathematical core, especially Lemma 1, is correct: the unitary set contains the diagonal set, so BD-RIS cannot be worse than diagonal RIS in the no-eavesdropper case, and the proposed optimal form Omega = V_E V_M^H is consistent with the von Neumann trace inequality. The concern does not invalidate the paper, but it supports keeping the reader's CONDITIONAL verdict: the central claim of a significant and general enhancement needs a Monte Carlo test with a stronger diagonal baseline before it can be accepted as fully established.","tokens_in":8092,"tokens_out":13493,"duration_ms":150124,"concrete_test":"Repeat the Section V experiments over, say, 1000 independent channel realizations using the same distributions and parameters (r=36, k=10 and r=64, k=15). For each realization, compute the ratio of Bob's trace-FIM under the BD-RIS design to that under the best diagonal-RIS design. For the diagonal baseline, replace the relaxed-and-clipped SDP with a proper phase-only optimization over unit-modulus diagonal matrices, for example manifold optimization or SDP with many rank-one randomizations and no magnitude relaxation, and record feasibility of the Eve constraint after optimization. Report the median ratio, the 10th and 90th percentiles, and the fraction of realizations where the ratio exceeds 1.1 and 2. If the median ratio remains large, the claimed significant enhancement is robust; if it drops near 1, the single-realization figure overstates the gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that BD-RIS significantly improves parameter estimation over diagonal RIS is supported by Lemma 1 for the no-eavesdropper non-reciprocal case, but the magnitude of the improvement is established only numerically. In Section V, all channel matrices are generated from a single realization in MATLAB, explicitly stated in the text, and the diagonal-RIS baseline is obtained with SDP plus Gaussian randomization and, in the eavesdropping case, a relaxation to diagonal magnitudes no greater than one followed by clipping. This baseline is heuristic: Gaussian randomization does not guarantee optimality, and clipping can produce a solution that is not feasible for the Eve constraint because it increases Eve's Fisher information. Consequently, the reported gap may be partly an artifact of a weak baseline or of the particular channel draw. The no-eavesdropper theory guarantees only that the BD-RIS optimum is at least as large as the diagonal optimum, since the unitary set contains the diagonal set; it does not establish that the advantage is significant. A second, related premise is that fully connected BD-RIS can realize arbitrary unitary or symmetric-unitary response matrices (Section III, constraints (7b) and (11b)). This is a modeling assumption inherited from the BD-RIS literature; if practical devices cannot realize those matrices, the derived limits are upper bounds rather than achievable gains. The paper does not discuss this limitation or report any statistical variability, so the empirical claim of significant enhancement is not yet robustly supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the design of a beyond-diagonal reconfigurable intelligent surface (BD-RIS) for parameter estimation in a system where direct links between the transmitter (Alice) and both the intended receiver (Bob) and an eavesdropper (Eve) are blocked. The performance metric is the trace of the Fisher information matrix (FIM) at Bob, optionally subject to a limit on Eve's average Fisher information. For the non-reciprocal BD-RIS without an eavesdropper, the authors prove Lemma 1, providing the closed-form optimal response Ω = V_E V_M^H and the maximum trace equal to the sum of sorted eigenvalue products. For the reciprocal (symmetric-unitary) case, they propose a manifold alternating-optimization algorithm (Algorithm 1). For secure estimation, they formulate a constrained problem and solve it via penalty dual decomposition (PDD). Numerical results compare the proposed BD-RIS designs with a conventional diagonal RIS and report substantial improvements in Fisher information and Cramér–Rao bound.","tokens_in":8324,"tokens_out":7172,"duration_ms":72045,"significance":"If the results are solid, the paper provides a valuable extension of RIS-assisted parameter estimation to the increasingly important BD-RIS hardware model. The main theoretical contribution is Lemma 1, which is cleanly stated and correctly proved via von Neumann's trace inequality; it gives an ultimate bound for the non-reciprocal no-eavesdropper case. The algorithms for the reciprocal and secure cases follow established techniques (manifold optimization and PDD) and appear plausible, though no convergence proofs are given. The central weakness is the numerical validation: the claimed significant gain over diagonal RIS rests on a single channel realization and a heuristic diagonal baseline, so the magnitude of the advantage is not established robustly. The modeling assumption that arbitrary unitary response matrices are realizable is also load-bearing and not explicitly discussed.","major_comments":[{"comment":"The numerical comparison is based on a single channel realization in MATLAB, which is insufficient to support the paper's central claim that BD-RIS 'notably enhances' estimation performance. Since the channel matrices are random, the reported gaps may be specific to that draw. Please provide Monte Carlo averages over many channel realizations, with confidence intervals or box plots, and ensure the conclusions hold statistically.","section":"Section V, first paragraph"},{"comment":"The diagonal-RIS baseline is obtained by SDP followed by Gaussian randomization, and in the eavesdropping case by relaxing the unit-modulus constraint and then clipping entries. Because the diagonal RIS is the reference point for the claimed improvement, the baseline's suboptimality could artificially inflate the BD-RIS gain. Please validate the baseline, e.g., for small r compare against a global optimization method, report the number of randomizations used, and verify that the clipped solutions still satisfy the Eve information constraint.","section":"Section V, conventional RIS baseline"},{"comment":"The assumption that a fully connected BD-RIS can realize any unitary (or symmetric-unitary) response matrix is load-bearing for the interpretation of the results as 'ultimate limits'. If practical devices restrict the achievable response matrices, the derived optima are upper bounds rather than achievable gains. Please add a discussion of the hardware feasibility of arbitrary unitary responses, citing the relevant BD-RIS literature, or explicitly state that the results are valid under the ideal unitary model.","section":"Section III, constraints (7b) and (11b)"},{"comment":"The alternating-optimization algorithm for the reciprocal case lacks a convergence proof or even a monotonicity argument. The paper cites similar manifold methods in [13] and [10], but the specific update with adaptive step size is not shown to converge to a stationary point. This affects the credibility of the reciprocal-BD-RIS curves in Figs. 1–4. Please state the convergence properties or characterize the algorithm's behavior (e.g., monotone increase of the cost function).","section":"Section III, Algorithm 1"}],"minor_comments":[{"comment":"The phrase 'in ∆ and Σ' should read 'in Δ_E and Δ_M' to refer to the eigenvalue matrices of E_b and M, respectively.","section":"Eq. (10)"},{"comment":"In the term '2αiℜ(e^{φ i}bH ui)', the exponential should be e^{jφ_i} (with the imaginary unit j) to be dimensionally consistent and to match the subsequent phase optimization.","section":"Eq. (23)"},{"comment":"The convergence threshold ǫ and the eavesdropping information limit ε share the same symbol; using different symbols would prevent confusion for the reader.","section":"Algorithm 1"},{"comment":"The augmented Lagrangian in (18) is not fully standard in the sign of the penalty term; adding a brief explanation or a pointer to [9, Alg. 1] would help readers verify the subsequent update equations.","section":"Section IV, Eq. (18)"},{"comment":"The axis label '10 104' in Fig. 1 appears to be a formatting artifact; please check that the tick labels render correctly.","section":"Figure 1 and 2 captions"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competent but incremental extension of [2] to BD-RIS. Lemma 1 is a nice closed-form result, and the algorithmic framework is plausible. The main concern is that the experimental section does not provide the statistical evidence needed for the claimed 'significant enhancement' over diagonal RIS; the single-realization comparison and heuristic baseline are substantial weaknesses. I would recommend major revision, with the requirement of additional Monte Carlo simulations and a stronger baseline study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper applies BD-RIS to a problem previously studied with diagonal RIS only: secure parameter estimation via trace-FIM maximization. That is a legitimate and natural extension, and the paper's main theoretical result, Lemma 1, is correct and clean. The closed-form unitary response Ω = V_E V_M^H follows directly from von Neumann's trace inequality, giving the exact optimum for the no-eavesdropper non-reciprocal case. The reciprocal variant via manifold optimization and the eavesdropping case via PDD are plausible extensions with standard machinery. I verified the algebra in the Ψ-update; the eigendecomposition and bisection step are fine. So as a letter, the contribution is real.\n\nWhere it gets shaky is the numerical claim of significant enhancement over diagonal RIS. The figures rest on a single channel realization, and the diagonal baseline is heuristic: SDP relaxation plus Gaussian randomization, and in the Eve-constrained case, clipping of diagonal entries to unit modulus. Clipping can break the Eve information constraint, making the comparison unfair. The reader is right to flag this. The theoretical ordering only guarantees BD-RIS is never worse than diagonal RIS; it says nothing about the magnitude of the gap. A Monte Carlo average would be needed to claim 'significant' in a robust way. The other soft spot is the modeling premise that a fully connected BD-RIS can realize arbitrary unitary or symmetric-unitary responses; if practical devices cannot, the derived limits are upper bounds. The paper says nothing about this, but it is an inherited assumption from the BD-RIS literature, not an oversight unique to this work.\n\nOverall the core is sound. The no-eavesdropper result deserves to be published; the secure-case algorithm is reasonable and cites the right literature. The numerical section needs strengthening before the 'significant enhancement' language is justified. If I were refereeing, I'd ask for Monte Carlo results and a feasible baseline; with that, the letter would be solid. This is a good candidate for a serious referee, not a desk reject.","headline":"A correct and clean extension of secure parameter estimation to BD-RIS, whose numerical claims of significant gain rest on a single realization and a heuristic diagonal baseline.","tokens_in":8862,"tokens_out":3993,"would_cite":true,"duration_ms":38314,"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":"A beyond-diagonal RIS whose response is an arbitrary unitary matrix can substantially improve secure parameter estimation, with the no-eavesdropper optimum given in closed form by an eigenvector alignment.","keywords":["beyond-diagonal RIS","reconfigurable intelligent surface","parameter estimation","Fisher information","Cramér–Rao lower bound","physical layer security","eavesdropping","unitary response matrix"],"falsifier":"Repeat the numerical experiments over many independent random channel realizations, for example 1000 draws at $r=36,\\,k=10$, and compare the average and median trace Fisher information of the BD-RIS designs against the paper's diagonal-RIS baseline at the same eavesdropping threshold; if BD-RIS does not consistently achieve higher average Fisher information than the diagonal baseline, the central numerical claim fails.","tokens_in":7889,"feed_emoji":"📡","tokens_out":12616,"duration_ms":111400,"temperature":0.7,"pith_summary":"This paper asks whether a beyond-diagonal reconfigurable intelligent surface (BD-RIS), whose response is a full unitary matrix rather than diagonal phase shifts, can improve parameter estimation when the transmitter-to-receiver and transmitter-to-eavesdropper direct links are blocked. The paper's answer is yes: it derives the optimal unitary response in closed form for the no-eavesdropper case, gives algorithms for the symmetric (reciprocal) and eavesdropper-constrained cases, and shows numerically that BD-RIS substantially outperforms conventional diagonal RIS, especially when the eavesdropper's allowed Fisher information is small. If the result holds, BD-RIS gives wireless systems a concrete way to shape the propagation environment for both estimation accuracy and information-leakage control.","feed_headline":"Beyond-diagonal RIS beats diagonal for secure estimation","feed_subtitle":"Unitary surface response maximizes receiver Fisher information and caps the eavesdropper's leakage.","key_machinery":"The central object is the $r\\times r$ BD-RIS response matrix $\\boldsymbol{\\Omega}$, constrained to be unitary ($\\boldsymbol{\\Omega}^H\\boldsymbol{\\Omega}=\\mathbf{I}_r$) in the general non-reciprocal case and additionally symmetric ($\\boldsymbol{\\Omega}=\\boldsymbol{\\Omega}^T$) in the reciprocal case. The objective is the trace of the Fisher information matrix, written as $\\operatorname{tr}(\\boldsymbol{\\Omega}^H\\mathbf{E}_b\\boldsymbol{\\Omega}\\mathbf{M})$ with $\\mathbf{E}_b=\\mathbf{H}_{rb}^H\\boldsymbol{\\Sigma}_b^{-1}\\mathbf{H}_{rb}$ and $\\mathbf{M}=\\mathbf{H}_{ar}\\mathbf{P}\\mathbf{P}^H\\mathbf{H}_{ar}^H$. The argument is carried by von Neumann's trace inequality, which upper-bounds this trace by the sum of descending-ordered eigenvalue products and is achieved by aligning the eigenbases of $\\mathbf{E}_b$ and $\\mathbf{M}$. For the eavesdropping case, the constraint $\\operatorname{tr}(\\boldsymbol{\\Omega}^H\\mathbf{E}_e\\boldsymbol{\\Omega}\\mathbf{M})\\le\\epsilon$ is handled by vectorization, which turns the inner update into a quadratically constrained least-squares problem whose KKT solution is a bisection search over the Lagrange multiplier $\\mu$.","core_discovery":"The paper's central claim is that replacing the diagonal phase-shift matrix of a conventional RIS with a fully connected beyond-diagonal RIS, whose response matrix $\\boldsymbol{\\Omega}\\in\\mathbb{C}^{r\\times r}$ is an arbitrary unitary matrix (or a symmetric unitary matrix in the reciprocal case), can substantially improve estimation quality at the intended receiver while limiting the information leaked to an eavesdropper. In the no-eavesdropper non-reciprocal case, the paper shows that the trace of the Fisher information matrix, $\\operatorname{tr}(\\boldsymbol{\\Omega}^H \\mathbf{E}_b \\boldsymbol{\\Omega} \\mathbf{M})$, is maximized by $\\boldsymbol{\\Omega}=\\mathbf{V}_E\\mathbf{V}_M^H$, where $\\mathbf{V}_E$ and $\\mathbf{V}_M$ hold the eigenvectors of $\\mathbf{E}_b$ and $\\mathbf{M}$ sorted by descending eigenvalue, and the resulting maximum is the sum of the ordered eigenvalue products. For the reciprocal case the paper supplies a manifold alternating-optimization algorithm initialized at the closest symmetric unitary matrix to that unconstrained optimum, and for the secure case it supplies a penalty-dual-decomposition algorithm with a semi-closed-form update obtained by diagonalizing $\\mathbf{M}^T\\otimes\\mathbf{E}_e$. The numerical section reports that BD-RIS achieves substantially higher trace Fisher information, equivalently a lower Cramér–Rao bound and lower MSE of the maximum-likelihood estimator, than the diagonal-RIS baseline, with the reciprocal design nearly matching the non-reciprocal one.","pith_inferences":["Because the numerical evidence uses a single channel realization, the size of the gain over diagonal RIS should be checked across many random realizations; the closed-form result suggests the gain depends on how mismatched the eigenbases of $\\mathbf{E}_b$ and $\\mathbf{M}$ are.","The eigen-alignment view implies that when $\\mathbf{E}_b$ and $\\mathbf{M}$ are nearly diagonal or share eigenvectors, the BD-RIS advantage over a diagonal RIS should shrink, which is a testable prediction the paper does not make.","The eavesdropper-constrained problem could be extended to compute the full trade-off frontier between Bob's trace Fisher information and the eavesdropper threshold $\\epsilon$; the semi-closed-form structure suggests the frontier is piecewise-parametric in $\\mu$.","A natural follow-up is to jointly optimize the power allocation $\\mathbf{P}$ with the BD-RIS response, since the objective is linear in $\\mathbf{M}$ and the closed form for $\\boldsymbol{\\Omega}$ may simplify that joint optimization."],"forward_implications":["Without an eavesdropper and with a non-reciprocal BD-RIS, the maximum average Fisher information is exactly $\\sum_{i=1}^r \\delta_{E,i}\\delta_{M,i}$, so any suboptimal design can be measured against a closed-form ceiling.","Because the model is linear and Gaussian, the Cramér–Rao bound is achievable by the maximum-likelihood estimator, so the reported trace-Fisher-information gains translate directly into mean-squared-error reductions.","The negligible gap between reciprocal and non-reciprocal designs means reciprocal hardware, which is symmetric and easier to build, can capture most of the BD-RIS benefit.","Under tight eavesdropping limits, conventional diagonal RIS may become infeasible, whereas BD-RIS keeps the problem feasible and maintains high estimation quality at the intended receiver."],"supporting_citations":[{"why":"Supplies the system model, the trace-Fisher-information metric, and the diagonal-RIS baseline whose performance BD-RIS is compared against.","marker":"[2]"},{"why":"Provides the scattering-parameter modeling that justifies treating fully connected BD-RIS responses as arbitrary unitary matrices.","marker":"[8]"},{"why":"Gives the penalty-dual-decomposition template and the unitary projection lemma used in the eavesdropper-constrained algorithm.","marker":"[9]"},{"why":"Provides the theorem used to find the closest symmetric unitary matrix for initialization and for the reciprocal constrained update.","marker":"[10]"},{"why":"Establishes the capacity-style ultimate-limit analogy and supports the eigen-alignment result for unitary response matrices in related MIMO settings.","marker":"[12]"},{"why":"Supplies the manifold optimization procedure used for the reciprocal BD-RIS problem.","marker":"[13]"},{"why":"Provides the penalty-dual-decomposition convergence analysis used for the outer-layer update.","marker":"[14]"},{"why":"Supplies the vectorization/Kronecker identity used to cast the eavesdropper-constraint update as quadratic programming.","marker":"[15]"}],"fun_headline_variants":["Unitary BD-RIS boosts estimation, limits eavesdropper","Beyond-diagonal RIS ups Fisher info, cuts leakage","BD-RIS secures estimation via unitary surface","Non-diagonal RIS wins for secure estimation","Fully connected RIS sharpens estimation, blocks eavesdropping"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a fully connected BD-RIS can realize any $r\\times r$ unitary (orthonormal-column) response matrix, symmetric in the reciprocal case; if practical hardware cannot realize such matrices, or if the single simulated channel realization is unrepresentative, the derived limits and the large reported gains over diagonal RIS may not hold.","fun_headline_variants_meta":{"raw":{"variants":["Unitary BD-RIS boosts estimation, limits eavesdropper","Beyond-diagonal RIS ups Fisher info, cuts leakage","BD-RIS secures estimation via unitary surface","Non-diagonal RIS wins for secure estimation","Fully connected RIS sharpens estimation, blocks eavesdropping"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000461,"raw_usage":{"total_tokens":2341,"prompt_tokens":1016,"completion_tokens":1325,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":1248}},"tokens_in":632,"tokens_out":1325,"duration_ms":10045,"temperature":1.0,"reasoning_tokens":1248,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:51:12.035573+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the numerical experiments over many independent random channel realizations, for example 1000 draws at $r=36,\\,k=10$, and compare the average and median trace Fisher information of the BD-RIS designs against the paper's diagonal-RIS baseline at the same eavesdropping threshold; if BD-RIS does not consistently achieve higher average Fisher information than the diagonal baseline, the central numerical claim fails.","supporting_citations":[{"cited_title":"Joint RIS pha se profile design and power allocation for parameter estimation in pre sence of eavesdropping,","cited_arxiv_id":null,"evidence_quote":"Supplies the system model, the trace-Fisher-information metric, and the diagonal-RIS baseline whose performance BD-RIS is compared against."},{"cited_title":"Opti mizing power consumption, energy efficiency, and sum-rate using be yond diagonal RIS—a unified approach,","cited_arxiv_id":null,"evidence_quote":"Gives the penalty-dual-decomposition template and the unitary projection lemma used in the eavesdropper-constrained algorithm."},{"cited_title":"Wideband channel capacity maximization with beyond diagonal RIS reflection matrices,","cited_arxiv_id":null,"evidence_quote":"Provides the theorem used to find the closest symmetric unitary matrix for initialization and for the reciprocal constrained update."},{"cited_title":"Capacity Maximization for MIMO Channels Assisted by Beyond-Diagonal RIS","cited_arxiv_id":"2411.18298","evidence_quote":"Establishes the capacity-style ultimate-limit analogy and supports the eigen-alignment result for unitary response matrices in related MIMO settings."},{"cited_title":"MIMO capacity maximization with beyond-diagonal RIS,","cited_arxiv_id":null,"evidence_quote":"Supplies the manifold optimization procedure used for the reciprocal BD-RIS problem."},{"cited_title":"Penalty dual decomposition method f or non- smooth nonconvex optimization-Part I: Algorithms and conv ergence analysis,","cited_arxiv_id":null,"evidence_quote":"Provides the penalty-dual-decomposition convergence analysis used for the outer-layer update."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the vectorization/Kronecker identity used to cast the eavesdropper-constraint update as quadratic programming."}],"review_version":1}