{"id":"445c1833-84db-496b-bf01-7adc446ba223","arxiv_id":"2505.16230","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A beyond-diagonal intelligent reflecting surface is optimized to maximize the minimum user rate subject to a posterior Cramér-Rao bound in an uplink integrated sensing and communication system.","lead":"This paper designs a reconfigurable surface whose elements are interconnected to help a base station simultaneously sense a moving target and serve multiple users. It derives a statistical accuracy bound and an optimization algorithm that balances sensing and communication, and compares this with a time-sharing scheme.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equations (9) and (24) define G and U as unweighted integrals, silently replacing the Gaussian-mixture prior with a uniform one; this removes the claimed prior-aware behavior from both the rate bound and the PCRB.","rationale":"The paper's central contribution is a BD-IRS reflection design that exploits prior target-location information under a PCRB constraint; its analytical claim is that Eq. (24) is an explicit function of Φ with the prior included. That claim requires all Eθ[·] computations to use pΘ(θ). The printed definitions of G and U are the most load-bearing weak point because both the communication lower bound and the sensing PCRB depend on them. The fixed-r assumption is acknowledged and less damaging: a PCRB for θ alone is still a valid bound for angle MSE, and the paper explicitly lists joint angle-and-range estimation as future work, so this is a scope limitation rather than a contradiction. The missing prior density is a genuine internal inconsistency within the mathematical development. The rest of the derivation, including the SINR bound, the PDD reformulation, and the SOCP subproblems, appears internally consistent, and the claimed polynomial complexity is plausible. The numerical comparisons to diagonal IRS are suggestive, but they reproduce the claimed prior-aware concentration only if the implementation follows the intended weighted integrals rather than the printed unweighted ones. The absent code makes it impossible to tell which version was simulated, so this should be an explicit condition for acceptance, supporting a conditional rather than unconditional verdict.","tokens_in":26641,"tokens_out":6925,"duration_ms":60824,"concrete_test":"Using the exact Gaussian-mixture parameters from Section VII, recompute G = ∫_0^π g(θ)gH(θ)pΘ(θ)dθ and U = ∫_0^π ˙g(θ)˙gH(θ)pΘ(θ)dθ, substitute these into Eqs. (8) and (24), and rerun the proposed PDD algorithm for (P1) and (P5). Compare the resulting Fig. 7 beam pattern and Fig. 6 rate/PCRB curves with the unweighted-integral version. If the weighted version changes the optimized Φ or the PCRB by more than numerical noise, the printed equations omit the prior term on which the paper's central claims rely. Also re-derive FO independently from the complex Gaussian log-likelihood in Eq. (13) with the Gaussian-mixture prior and verify that the resulting PCRB matches Eq. (24) only when U is weighted by pΘ(θ).","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Under the stated prior pΘ(θ), Section II defines G ≜ Eθ[g(θ)gH(θ)] = ∫_0^π g(θ)gH(θ)dθ and Section III defines U ≜ Eθ[˙g(θ)˙gH(θ)] = ∫_0^π ˙g(θ)˙gH(θ)dθ. Both integrals are unweighted, so as printed they correspond to a uniform prior, not to the Gaussian mixture used in Remark 1 and Section VII. Concretely: (i) the Jensen lower bound in Eq. (8) is computed with the wrong interference covariance, so the optimized rate is not the actual expected-rate lower bound under the stated prior; (ii) the PCRB in Eq. (24) depends on the prior only through the scalar FP, so the designed Φ does not preferentially illuminate high-probability angles; the claim in Section VII-C that the design concentrates power toward high-probability angles cannot follow from the printed U unless the implementation silently uses pΘ(θ) weighting. This is an internal inconsistency in the central sensing and communication metrics, and it is more immediately load-bearing than the acknowledged fixed-r simplification: if the unweighted formulas were actually used, the numerical comparisons and the prior-information contribution would not test the claimed mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper considers a BD-IRS aided uplink ISAC system in which a multi-antenna BS estimates the azimuth angle of an active target using its uplink probing signals and a known prior PDF, while simultaneously serving multiple single-antenna communication users. The authors derive a posterior Cramér-Rao bound (PCRB) for the angle estimate, formulate a max-min expected rate optimization subject to a PCRB constraint and the lossless/reciprocal BD-IRS constraints, and propose a penalty dual decomposition (PDD) algorithm. They also propose a TDMA variant with closed-form time allocation. Numerical results claim that BD-IRS outperforms diagonal IRS and that the optimized reflection matrix concentrates sensing power toward high-probability target angles.","tokens_in":26882,"tokens_out":6132,"duration_ms":53240,"significance":"The paper is one of the first to optimize BD-IRS for Bayesian sensing with prior location information, and it provides a tractable PCRB expression and a polynomial-complexity algorithm. The analytical derivations are self-contained and parameter-free, and the SDMA/TDMA comparison is a useful design insight. If the PCRB and rate expressions are corrected, the work would be a solid contribution to BD-IRS ISAC.","major_comments":[{"comment":"The definitions G ≜ Eθ[g(θ)gH(θ)] = ∫_0^π g(θ)gH(θ)dθ and U ≜ Eθ[˙g(θ)˙gH(θ)] = ∫_0^π ˙g(θ)˙gH(θ)dθ are internally inconsistent with the stated prior: an expectation under pΘ(θ) must be ∫ f(θ)pΘ(θ)dθ. As printed, the integrals are unweighted, so the Jensen lower bound in Eq. (8) and the observation Fisher information in Eqs. (17)-(24) do not depend on the prior distribution except through the scalar FP. Consequently, the optimized Φ in Problems (P1)-(P3) is not prior-aware, and the claim in Section VII-C that the design concentrates power toward high-probability angles cannot follow from the printed equations. The authors should correct the integrals to include pΘ(θ), or state and justify a uniform prior; the numerical results with the Gaussian mixture would need to be recomputed accordingly.","section":"§II, Eq. (8); §III, Eq. (17)"},{"comment":"The stated goal is sensing the target's location, but the PCRB is derived for the azimuth angle θ only, assuming the target-to-IRS distance r is known exactly. If r is unknown or estimated with error, the PCRB for θ alone is not a lower bound on the location MSE, and the optimized Φ may be mismatched for joint range-angle estimation. The manuscript acknowledges this in a footnote but does not analyze the sensitivity of the PCRB constraint or the design to range errors; this should be discussed as a limitation or addressed by extending to joint estimation.","section":"§II, Eq. (2); §III, Eq. (24)"}],"minor_comments":[{"comment":"The target angle distribution is plotted on the same dBm axis as the effective sensing power, but the PDF is not expressed in dBm; please clarify the normalization or use a separate axis.","section":"§VII-C, Fig. 7"},{"comment":"The channel model in Eq. (2) includes only the x-dimension phase and does not explicitly show the z-dimension of the planar array; this is consistent with footnote 3 but should be stated in the main text before Eq. (2) for clarity.","section":"§II, Eq. (2)"},{"comment":"The optimal time allocation q* is written as a max with 0; when the problem is feasible, the term is at most 1, but it would be helpful to explicitly state that q* is capped at 1 in the feasible case.","section":"§VI, Eq. (92)"},{"comment":"The complexity formula in Table I has a formatting issue with the parentheses around L_P L_O L_I L_P4; please correct the typesetting.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle to acceptance is the internal inconsistency in the definitions of G and U: the unweighted integrals do not match the stated Gaussian mixture prior, and the numerical claims depend on prior-aware behavior that the printed equations do not implement. The paper does not provide a reproducibility statement or code, which would be helpful given the complexity of the algorithm."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent, incremental ISAC paper, and the main thing to know is that the authors define the two key expectations without the prior density. In Eq. (8) they write G = E_theta[g g^H] = int_0^pi g(theta) g(theta)^H dtheta, and in Eq. (24) U = E_theta[gdot gdot^H] = int_0^pi gdot(theta) gdot(theta)^H dtheta. Under the Gaussian-mixture prior from Remark 1, those should be weighted by p_Theta(theta). Taken literally, the rate lower bound and the PCRB are computed under a uniform prior, not the stated mixture. That is not cosmetic: the whole claimed contribution, that BD-IRS can concentrate reflection power on high-probability target angles, depends on the prior entering G and U. If the unweighted integrals were actually used, Figure 7's concentration toward the mixture peaks would not follow from the printed math. I suspect the implementation used the weighted integrals and the display equations are sloppy, but the paper needs to say so and the numerical section needs to confirm it.\n\nWhat is genuinely new: the combination of BD-IRS with PCRB-based sensing and max-min expected-rate optimization under a target location prior. I have not seen that exact combination. The PCRB derivation is standard but carefully done, the PDD reformulation with the SOCP subproblem looks plausible, and the complexity table is honest. The TDMA variant with closed-form time allocation is a reasonable addition, and the SDMA/TDMA comparison in Fig. 8 is useful.\n\nSofter spots, in order: first, the missing p_Theta in G/U; second, the fixed target-to-IRS distance r. The paper acknowledges r is assumed known and leaves joint range-angle estimation to future work, so I read that as a known simplification rather than a hidden flaw. It limits the sensing claim to angle-only estimation with perfectly known range, which is exactly what they state. Minor: Eq. (92) for the TDMA time fraction is typeset ambiguously; the derivation is easy to redo but the display should be cleaned up. No code is released, but for this subfield that is normal, and the numerical claims are plausible once the prior is restored.\n\nWho benefits: researchers working on IRS-aided ISAC, especially the BD-IRS subgroup. It will not shift paradigms, but it is a useful data point and a decent optimization framework. If I were an editor I would send it to review with a request to fix the integral definitions and rerun or recheck the simulations; the core idea survives.\n\nRecommendation: engage with it, but insist on the p_Theta fix before acceptance.","headline":"Solid BD-IRS ISAC extension with a clean PDD/SOCP framework, but the printed definitions of G and U drop the prior density and that must be fixed before the prior-aware results can be trusted.","tokens_in":27408,"tokens_out":4563,"would_cite":true,"duration_ms":38573,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A12","90C26"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper derives a closed-form posterior CRB for target angle as an explicit function of the BD-IRS reflection matrix, and optimizes that matrix to maximize the minimum user rate under a sensing-accuracy constraint.","keywords":["beyond diagonal IRS","integrated sensing and communication","posterior Cramer-Rao bound","device-based sensing","reflection matrix optimization","penalty dual decomposition","uplink ISAC","TDMA"],"falsifier":"Run the same uplink ISAC setup with $r$ drawn, say, uniformly from 5 m to 15 m and estimate both $\\theta$ and $r$ jointly from the received signals, then compare the empirical MSE of $\\theta$ against $\\mathrm{PCRB}_\\theta$ computed with $r$ fixed at its true value; if the empirical MSE exceeds the claimed PCRB bound while that bound is used as the constraint in the optimization, the paper's central claim that the PCRB captures sensing performance fails.","tokens_in":26425,"feed_emoji":"📡","tokens_out":4165,"duration_ms":34756,"temperature":0.7,"pith_summary":"This paper studies an uplink integrated sensing and communication (ISAC) system in which a base station both receives data from multiple users and estimates the azimuth angle of an active target, with the direct path blocked so a beyond-diagonal intelligent reflecting surface (BD-IRS) provides the only link. The authors derive a closed-form posterior Cramer-Rao bound (PCRB) for the target angle as an explicit function of the BD-IRS reflection matrix, and they formulate the reflection design as a max-min user-rate problem subject to a PCRB constraint. They propose a penalty dual decomposition (PDD) algorithm that finds a high-quality suboptimal reflection matrix in polynomial time, and a TDMA alternative with closed-form time allocation that eliminates sensing-communication mutual interference. If the claimed PCRB expression and algorithm performance hold, BD-IRS provides a new degree of freedom for interference management in uplink ISAC with prior location information.","feed_headline":"BD-IRS design boosts uplink ISAC rates without losing sensing","feed_subtitle":"Optimizing the whole reflection matrix, not just phase shifts, improves the tradeoff between sensing accuracy and user rates.","key_machinery":"The load-bearing identity is the closed-form PCRB expression in Eq. (24), which turns the sensing accuracy into a quadratic-in-$\\Phi$ form with the inverse of the interference-plus-noise covariance $\\Sigma_0(\\Phi)$, so that both sensing and communication become functions of the same reflection matrix. The optimization machinery is the penalty dual decomposition (PDD) algorithm: it introduces auxiliary unitary matrices $\\Psi_g$, penalizes the equality constraints $\\Phi_g = \\Psi_g$, and alternates between a convex second-order cone program for $(\\alpha, \\Phi)$ and closed-form updates for the auxiliary variables, using the duplication matrices $D_g$ and half-vectorization $\\mathrm{vech}(\\Phi_g)$ to reduce the dimension of the reflection optimization.","core_discovery":"The paper's central claim is that the posterior Cramer-Rao bound for the azimuth angle of an active target, computed with a known prior distribution, can be written in closed form as $\\mathrm{PCRB}_\\theta(\\Phi) = \\frac{1}{2P_0 L \\sum_{\\zeta=1}^{R} \\kappa_\\zeta (R\\Phi u_\\zeta)^H \\Sigma_0^{-1}(\\Phi) (R\\Phi u_\\zeta) + F_P}$, where $\\Sigma_0(\\Phi)$ is the interference-plus-noise covariance and $F_P$ is the prior Fisher information. The paper then establishes that maximizing the minimum expected user rate subject to this PCRB threshold, under the lossless and reciprocal constraints of the BD-IRS reflection matrix, is solvable by a PDD-based algorithm that alternates between convex subproblems. The resulting reflection design concentrates sensing energy on high-probability target angles while suppressing interference toward communication users, and numerical results show that fully- and group-connected BD-IRS architectures outperform conventional diagonal IRS and two benchmark reflection schemes.","pith_inferences":["The derivation treats only azimuth $\\theta$ with known distance $r$; a natural extension is joint angle-range estimation, where the PCRB matrix would couple $r$ and $\\theta$ and the optimized reflection could differ substantially.","The PCRB expression suggests a design principle: the optimal reflection should align the columns $R\\Phi u_\\zeta$ with the dominant eigenvectors of the inverse interference-plus-noise covariance, a form of whitened matched filtering across the surface that could be used to build cheaper greedy or codebook-based designs.","The rate metric in the paper is a lower bound via Jensen's inequality; although the paper's numerics show the bound is close to the actual expected rate in the tested regime, under heavy interference the gap could widen, so a design maximizing the bound may not maximize the true average rate.","The group-connected tradeoff in Table II hints at a Pareto frontier between circuit complexity and ISAC performance, so future hardware-aware designs could choose the group size based on both performance and implementation cost."],"forward_implications":["If the PCRB expression and algorithm are correct, the designed BD-IRS can meet a sensing-accuracy constraint while improving the minimum user expected rate, with the rate increasing as the PCRB threshold or surface size grows.","The fully-connected and group-connected BD-IRS architectures outperform conventional single-connected (diagonal) IRS, and the gain widens as the number of elements grows; the paper shows BD-IRS with fewer elements can match the performance of a diagonal IRS with more elements.","TDMA with optimized time allocation can beat simultaneous sensing and communication when a user lies near high-probability target locations, because removing interference compensates for the time-sharing loss.","The proposed design concentrates effective sensing power toward high-probability angles while suppressing power toward communication users, a location-dependent interference management that isotropic or random reflection benchmarks cannot achieve."],"supporting_citations":[{"why":"Supplies the BD-IRS architecture with lossless and reciprocal constraints $\\Phi^H\\Phi = I$ and $\\Phi = \\Phi^T$, which are the structural constraints in the optimization.","marker":"[20]"},{"why":"Provides the penalty dual decomposition framework used to solve the nonconvex reflection optimization problem.","marker":"[53]"},{"why":"Supplies the PCRB-with-prior-information modeling and the Gaussian mixture prior example used for the target angle distribution.","marker":"[42]"},{"why":"Serves as the diagonal-IRS uplink ISAC comparison and the source of the concept of PCRB/BCRB for device-based sensing with prior information.","marker":"[17]"},{"why":"Gives the posterior Fisher information decomposition $F(\\Phi) = F_O(\\Phi) + F_P$ used in the PCRB derivation.","marker":"[60]"},{"why":"Establishes the single-, group-, and fully-connected BD-IRS architecture and its performance advantages, motivating the reflection structure studied here.","marker":"[24]"}],"fun_headline_variants":["BD-IRS: Full reflection matrix beats diagonal in ISAC","Optimizing entire IRS matrix improves ISAC rates and sensing","BD-IRS improves ISAC: higher rates, same sensing accuracy","PDD-based BD-IRS design outperforms diagonal IRS in ISAC","Beyond-diagonal IRS yields better uplink ISAC performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The derivation assumes the target-to-BD-IRS distance $r$ is exactly known and only the azimuth angle $\\theta$ is random, so the PCRB does not account for errors in estimating range; if range is uncertain, the computed bound may not be a valid lower bound on location MSE and the optimized reflection may be mismatched.","fun_headline_variants_meta":{"raw":{"variants":["BD-IRS: Full reflection matrix beats diagonal in ISAC","Optimizing entire IRS matrix improves ISAC rates and sensing","BD-IRS improves ISAC: higher rates, same sensing accuracy","PDD-based BD-IRS design outperforms diagonal IRS in ISAC","Beyond-diagonal IRS yields better uplink ISAC performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000313,"raw_usage":{"total_tokens":1830,"prompt_tokens":1051,"completion_tokens":779,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":692}},"tokens_in":667,"tokens_out":779,"duration_ms":6022,"temperature":1.0,"reasoning_tokens":692,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:06:08.547298+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same uplink ISAC setup with $r$ drawn, say, uniformly from 5 m to 15 m and estimate both $\\theta$ and $r$ jointly from the received signals, then compare the empirical MSE of $\\theta$ against $\\mathrm{PCRB}_\\theta$ computed with $r$ fixed at its true value; if the empirical MSE exceeds the claimed PCRB bound while that bound is used as the constraint in the optimization, the paper's central claim that the PCRB captures sensing performance fails.","supporting_citations":[{"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 framework used to solve the nonconvex reflection optimization problem."},{"cited_title":"MIMO integrated sensing and communi cation exploiting prior information,","cited_arxiv_id":null,"evidence_quote":"Supplies the PCRB-with-prior-information modeling and the Gaussian mixture prior example used for the target angle distribution."},{"cited_title":"RIS-assisted joint sensing and commun ications via fractionally constrained fractional programming,","cited_arxiv_id":null,"evidence_quote":"Serves as the diagonal-IRS uplink ISAC comparison and the source of the concept of PCRB/BCRB for device-based sensing with prior information."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the posterior Fisher information decomposition $F(\\Phi) = F_O(\\Phi) + F_P$ used in the PCRB derivation."},{"cited_title":"Beyond diagonal reconfig urable intelli- gent surfaces: From transmitting and reflecting modes to sin gle-, group- , and fully-connected architectures,","cited_arxiv_id":null,"evidence_quote":"Establishes the single-, group-, and fully-connected BD-IRS architecture and its performance advantages, motivating the reflection structure studied here."}],"review_version":1}