{"id":"e10ef584-2c59-4003-aa83-0154e7d372ab","arxiv_id":"2505.15403","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A combined-impacts RIS beam model, calibrated against measured patterns, reproduces beam patterns with error 0.94 compared to 4.15 and 8.75 for simpler models, and gives sub-meter simulated localization errors.","lead":"This paper proposes three beam models for reconfigurable intelligent surfaces that include real hardware imperfections such as mutual coupling, imperfect phase settings, and non-uniform element patterns, and calibrates them against measured beam patterns. The most detailed model matches measurements closely and yields sub-meter localization errors in simulation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The positioning claim is inferred from a lower bound: Section IV-C plots only the squared-bias term of the mismatched lower bound as 'ALB', and Section V-C treats this bound as the achievable position error, so the 'high likelihood below 1 m' conclusion does not follow.","rationale":"The reader's weakest assumption (the 16x16 measurement versus 1x16 simulation geometry mismatch) is a valid and concrete gap: the paper does not state how the planar prototype patterns are reduced to the linear-array steering model in Eq. (3), and if the reduction is invalid, the calibration errors lose their meaning. However, the single most load-bearing concern is the interpretation of the ALB. Even if the geometry issue were fully resolved by an explicit column-array reduction, the conclusion that the CI beam 'leads to a high likelihood of achieving a positioning error below 1 m' still would not follow from a lower bound. A lower bound below 1 m is necessary but not sufficient for achievability. The authors explicitly discard the MCRB term in Eq. (22) for tractability, and then plot the residual bias term as the 'positioning error'. This is a logical error in the central claim, not merely a missing detail. The in-sample evaluation and the questionable closed-form update in Eq. (19) are secondary: they affect the optimality or generalizability of the calibration, but the ALB issue directly invalidates the headline performance statement. The verdict remains CONDITIONAL because the paper can be repaired by recomputing the full bound or presenting empirical estimator errors and by softening the claim accordingly; the geometry mismatch should also be resolved, but it is not the most decisive concern.","tokens_in":9143,"tokens_out":14825,"duration_ms":135917,"concrete_test":"Compute the full mismatched lower bound in Eq. (22), including the MCRB term A^{-1} B A^{-1}, at the SNR implied by Table I (noise figure 10 dB and a specified transmit power), or alternatively simulate a practical estimator (e.g., maximum likelihood or least squares) on received signals generated with the calibrated CI beam model. Then plot the CDF of the resulting RMSE alongside the ALB curve in Fig. 4. If the full bound or empirical RMSE exceeds 1 m for a substantial fraction of UE positions where the ALB is below 1 m, the paper's central positioning claim is not supported by its evidence and must be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central performance claim, that the calibrated CI beam yields a high likelihood of positioning error below 1 m, is derived from the 'absolute lower bound' (ALB) defined in Section IV-C. In Eq. (22), the authors write the misspecified lower bound as MLB(η0) = A^{-1} B A^{-1} + Bias(η0), then state 'we focus on the biased term' and define ALB = Bias(s0). This is only the squared-bias component of the MLB; the positive-semidefinite MCRB term is omitted. A lower bound on the error cannot establish that an estimator achieves an error below a threshold: even if the bound is below 1 m, the actual RMSE may be arbitrarily larger when the variance term is non-negligible. Section V-C then plots the CDF of this ALB and describes it as 'positioning error', equating a necessary condition with the claimed outcome. The issue is compounded by Eq. (24), where the pseudo-true parameter is sought by gradient descent initialized at the true parameter āeta; if the mismatch objective is non-convex, this can underestimate the bias. Independent of the beam-model fitting quality, the localization benefit claimed in the abstract and results is therefore not supported by the presented evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses beam calibration for RIS-aided ISAC localization. It proposes three beam models of increasing complexity: a mutual-coupling model (MCM, Eqs. (4)-(5)), a non-ideal-codebook model (NC, Eq. (6)), and a combined-impacts model (CI, Eq. (7)). Calibration algorithms are developed in Section IV, using least squares and alternating optimization, together with a mismatch analysis based on a misspecified Cramer-Rao bound (Eq. (22)) that is specialized to a bias-only 'absolute lower bound' (ALB). The evaluation uses measured beam patterns from a 16x16 RIS prototype (Section V-A) and a localization simulation with a 1x16 RIS (Table I). The paper reports calibration errors of 8.75, 4.15, and 0.94 for MCM, NC, and CI, respectively, and uses CDFs of the ALB to claim that the CI-calibrated beam yields a high likelihood of positioning error below 1 m.","tokens_in":9455,"tokens_out":9085,"duration_ms":80474,"significance":"If the claims were fully substantiated, the work would be a useful step toward practical RIS calibration for ISAC: it uses real measured beam patterns, compares models of increasing physical fidelity, and attempts to connect beam-model mismatch to localization bounds. The comparison of beam-reconstruction accuracy on real prototype data is informative, and the CI model clearly gives the best in-sample fit. However, the localization claim currently rests on a bias-only lower bound, the calibration estimators contain identifiable technical flaws, and the evaluation is partly in-sample. The significance as stated is therefore not yet established, but the underlying problem and measurement effort are valuable.","major_comments":[{"comment":"The calibration algorithm for the MCM model is underdetermined and does not identify the mutual-coupling matrix. Eq. (11) minimizes over all matrices M_g in C^{T x N} using the single vector equation M_g w_g = \\bar{b}_g; with T equations and T*N unknowns, the least-squares solution is highly non-unique (for T=181 and N=16 or N=256, T*N is much larger than T). The subsequent step (12) assumes that an arbitrary minimizer \\hat{M}_g equals A^T C_g for some Toeplitz C_g, but nothing in (11) enforces that structure. Consequently the coefficients \\hat{c}_{g,1}, \\hat{c}_{g,2} from (14) are not identifiable from the stated formulation. The model should instead be estimated directly as b_g = A^T Toeplitz(c_g) w_g, which is linear in the coupling coefficients.","section":"IV-B1, Eqs. (10)-(14)"},{"comment":"The proposed closed-form update in Eq. (19) is not the solution to the stated unit-norm least-squares problem. For min_w ||\\bar{b}_g - \\tilde{\\Gamma} A^T w||^2 subject to ||w||=1, the normal equations give (A \\tilde{\\Gamma}^2 A^T) w = A \\tilde{\\Gamma} \\bar{b}_g for real diagonal \\tilde{\\Gamma}, not the expression (A A^T)^{-1} A \\tilde{\\Gamma}^{-1} \\bar{b}_g in (19). Moreover, the normalized unconstrained least-squares solution is not in general the solution to the unit-norm constrained problem. Since the CI calibration depends on these updates, the reconstructed CI beam patterns and all downstream ALB results are based on an incorrect optimizer.","section":"IV-B2, Eqs. (18)-(19)"},{"comment":"The 'ALB' is only the squared-bias term of the misspecified lower bound, and the CDF of this ALB is presented as positioning error. A lower bound cannot establish that the actual positioning error is below a threshold; the omitted MCRB(\\eta_0) term is positive semidefinite and may dominate. Furthermore, Bias(s_0) in Eq. (22) is an outer-product matrix, and the scalar quantity plotted in Figs. 3 and 4 is never defined. The conclusion in Section V-C that the CI model 'exhibits a high likelihood of achieving a positioning error below 1 m' is therefore not supported. In addition, Eq. (24) obtains the pseudo-true parameter by gradient descent initialized at the true value \\bar{\\eta}; for a non-convex mismatch objective this can converge to a local minimum and underestimate the bias.","section":"IV-C and V-C, Eq. (22), Figs. 3-4"},{"comment":"The ground-truth measurements are for a 16x16 planar RIS prototype, while the localization simulation uses N = 1x16 with the steering vector [a(\\phi)]_n = e^{j\\pi n cos(\\phi)}. The paper does not state how the 16x16 measured beam patterns are reduced to the 1x16 linear array. If the measured patterns do not correspond to the simulated array geometry, the calibration errors in Section V-C and the ALB results do not describe the simulated system. A precise mapping, or an adaptation of the beam models to the planar prototype, is required before the numerical conclusions can be accepted.","section":"V-A and V-B, Table I"},{"comment":"The evaluation is in-sample. The calibration error in Eq. (9) is the objective minimized in Eq. (16) and in the corresponding MCM/NC fits, and the same measured patterns are then used as ground truth to generate the received signal matrix Y in the localization simulation (Section V-B). The reported errors 8.75, 4.15, and 0.94 therefore measure training fit, not predictive accuracy, and the CDF in Fig. 4 reflects the fitted model on the training data. A cross-validation split or a separate measurement set is needed before the paper can claim that the CI model accurately reconstructs measured beam patterns in general.","section":"IV-A, IV-B and V-B"}],"minor_comments":[{"comment":"The y-axis label 'Pr(possition error< )' contains a typo and an incomplete expression, and the x-axis should be defined consistently with the ALB quantity (squared or not).","section":"Fig. 4"},{"comment":"The sampled angles are denoted \\theta_s in Eq. (9) after the text uses \\phi for the observation angle; the notation should be unified.","section":"Eq. (9)"},{"comment":"The text says the NC-model parameter \\tilde{W} can be estimated by least squares, but the unit-norm constraint ||\\tilde{w}_g||=1 in Eq. (6) is not imposed in that least-squares estimate; the role of the constraint should be clarified.","section":"IV-B, after Eq. (6)"},{"comment":"Eq. (22) mixes scalar, vector, and matrix quantities; the dimensions of A_{\\eta_0}, B_{\\eta_0}, and Bias(\\eta_0) should be defined explicitly.","section":"Eq. (22)"},{"comment":"Reference [24] lists volume 8 for IEEE Transactions on Wireless Communications in 2023; the volume number appears inconsistent with the journal's 2023 volume numbering.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and addresses a timely problem with real measurement data. The technical issues identified above are substantive but, in my view, fixable within a revision: the calibration estimators can be reformulated, the ALB claim can be reinterpreted as a bias-only analysis, and the geometry mapping between the measured prototype and the simulated array can be clarified. I do not see grounds for rejection at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nBottom line: this paper has a genuinely useful experimental comparison of three RIS beam models, but the localization claim in the abstract and Section V is built on a lower-bound misstatement, and the link between the measurements and the simulation geometry is missing.\n\nWhat is new and good: the combined-impacts (CI) model of Eq. (7), which multiplies a non-ideal codebook by a diagonal angle-dependent correction, fits the authors' 30 GHz 16x16 RIS prototype data much better than the mutual-coupling-only (MCM) or codebook-only (NC) models: calibration error 0.94 versus 4.15 and 8.75. That is a real result, and it is the kind of empirical grounding the field needs. The measured patterns as ground truth, the per-codeword mutual coupling, and the 2-bit phase shift details are all reasonable choices.\n\nThe soft spots are serious. First, the 'absolute lower bound' (ALB) in Eq. (22) is only the bias term of the mismatched lower bound; the MCRB variance term is dropped, and then the CDF of this ALB is plotted in Fig. 4 and described as 'positioning error'. A bound on the squared bias of the pseudo-true parameter does not bound the actual RMSE, so the 'high likelihood of positioning error below 1 m' does not follow. Second, the evaluation is in-sample: the calibration error (9) is exactly the objective minimized, so the model comparison is a goodness-of-fit, not a prediction. There is no held-out data or cross-validation. Third, the measurements are from a 16x16 planar RIS, but the localization simulation uses a 1x16 linear array (Table I) with no explanation of how the patterns are mapped. This disconnects the simulation from the measured data. Fourth, the MCM estimation in (11)-(14) is underdetermined: a T-by-N matrix is estimated from T equations, and the later vectorization assumes the least-squares result equals A^T C_g without identifiability guarantees. Finally, the closed-form solution in (19) appears to minimize a slightly different objective than (18), since it uses Γ^{-1} rather than Γ^H; this should be reconciled.\n\nNone of these are unfixable. The model comparison and the calibration algorithm formulation are a solid starting point, and the prototype data are valuable. But the paper as written overclaims positioning performance.\n\nFor whom: researchers working on RIS or millimeter-wave array calibration who want a real-data benchmark. A serious referee should engage; the paper needs major revision: fix the bound interpretation, add a held-out evaluation, resolve the array-size mismatch, and address the MCM identifiability.\n\nI would bring this to a reading group to discuss what counts as a valid performance bound, but I would not cite the localization claim in its current form.","headline":"Useful real-data model comparison, but the localization claim is built on a bias-only bound and an unresolved array-geometry mismatch.","tokens_in":9988,"tokens_out":4316,"would_cite":false,"duration_ms":34356,"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":"Calibrating RIS beams against measured patterns cuts reconstruction error to 0.94.","keywords":["RIS beam calibration","ISAC","reconfigurable intelligent surface","mutual coupling","non-ideal codebook","localization","mismatched lower bound","beam pattern modeling"],"falsifier":"Keep the original −40° to 40° azimuth data for fitting, but measure the same 16×16 prototype over additional azimuth and elevation cuts, then test whether the fitted $\\gamma(\\phi)$ and $\\tilde{W}$ predict those withheld cuts; if the reconstruction error climbs well above 0.94, the reported calibration is a fit to a single angular slice rather than a general beam model.","tokens_in":8951,"feed_emoji":"📡","tokens_out":6354,"duration_ms":48807,"temperature":0.7,"pith_summary":"This paper sets out to show that the idealized steering-vector beam model used in RIS-aided localization is too simplistic, and that a beam model incorporating mutual coupling, non-ideal codebook weights, and an angle-dependent correction coefficient can reconstruct real measured RIS beam patterns closely enough to make sub-meter positioning plausible. Using beam patterns measured from a 16×16, 2-bit RIS prototype as ground truth, the authors fit three progressively richer models: a mutual-coupling-matrix model, a non-ideal-codebook model, and a combined-impacts model that includes both plus an angle-dependent element-pattern factor. In the reported experiments, the combined model reproduces the measured patterns with calibration error 0.94, compared with 4.15 and 8.75 for the simpler models. When the calibrated beams drive a simulated 2D localization scenario, the combined model yields a high probability of positioning error below 1 m, whereas the other two models do not. The paper concludes that realistic beam modeling and calibration are prerequisites for high-accuracy RIS-aided ISAC localization.","feed_headline":"Calibrated RIS beams shrink localization error below 1 m","feed_subtitle":"Folding three hardware effects into one beam model makes sub-meter RIS-aided positioning likely.","key_machinery":"The load-bearing object is the combined-impacts (CI) beam model $b_{\\mathrm{CI}}(\\phi)=\\gamma(\\phi)\\,\\tilde{W}^{\\top}a(\\phi)$, with $a(\\phi)$ the steering vector, $\\tilde{W}$ the perturbed codebook matrix capturing RF-chain inhomogeneity and mutual coupling, and $\\gamma(\\phi)$ a diagonal correction for element pattern and measurement-gain imbalance. It carries the argument by being the only fitted model whose reconstructed beams stay close to the measured patterns, which is what pushes the absolute lower bound on positioning error below 1 m. The calibration procedure is an alternating minimization: a closed-form unit-norm least-squares update for each column of $\\tilde{W}$, then an independent linear update for each diagonal entry of $\\Gamma(\\phi)$, iterated to convergence.","core_discovery":"On the paper's own terms, the central discovery is that the dominant source of RIS beam error in practice is not mutual coupling alone or codebook perturbation alone, but their combination with element-pattern and measurement-gain variations, captured by the beam model $b_{\\mathrm{CI}}(\\phi) = \\gamma(\\phi)\\,\\tilde{W}^{\\top}a(\\phi)$, where $a(\\phi)$ is the steering vector, $\\tilde{W}$ is a codebook matrix perturbed to absorb RF and coupling impairments, and $\\gamma(\\phi)$ absorbs element pattern and measurement uncertainties. Fitting this model to measured beam patterns with alternating least squares between $\\tilde{W}$ and $\\gamma$ produces a calibrated beam that tracks ground truth with calibration error 0.94, versus 4.15 for the non-ideal-codebook model and 8.75 for the mutual-coupling-matrix model. A mismatched lower bound analysis of a 2D LOS-plus-RIS localization scenario then shows that the combined-impacts calibrated beam makes positioning errors below 1 m highly likely, while the simpler models settle at comparable, larger errors. The paper reads this as evidence that realistic beam modeling and calibration are prerequisites for high-accuracy RIS-aided ISAC localization.","pith_inferences":["An extension the paper leaves implicit: if the CI parameterization is as general as the measured data suggest, it should fit other RIS geometries (different element counts, 1-bit phase resolution, larger apertures) with calibration error staying near 1; that is a direct transfer test.","The paper's ALB maps show error spikes in the gaps between the 11 discrete scanning angles, which implies that codebook design, not just calibration, is the next lever; denser or interpolated codewords could remove those angular dead zones.","Because the beam modeling is 2D azimuth-only, the angle-dependent factor $\\gamma(\\phi)$ may actually absorb elevation effects that the model does not name; a 3D extension would show whether the correction separates cleanly into azimuth and elevation components."],"forward_implications":["RIS-aided ISAC systems that calibrate with the combined-impacts model can expect positioning errors below 1 m to become the typical case rather than an outlier.","Systems designed from the ideal steering-vector model in equation (2) will carry a localization bias that no amount of signal averaging can remove, because the bias term in the ALB is independent of SNR.","The calibration error metric of equation (9) gives a practical acceptance threshold: a fitted beam with error near 1 is sufficient for sub-meter localization, while errors of 4 to 9 are not.","The alternating least-squares calibration procedure can be run directly from anechoic-chamber beam measurements, with no second reference array or special hardware."],"supporting_citations":[{"why":"Supplies the measured ground-truth beam patterns from the 16×16 RIS prototype used to fit and evaluate all three beam models.","marker":"[25]"},{"why":"Supplies the Toeplitz mutual-coupling-matrix model that the MCM beam model is built on.","marker":"[20]"},{"why":"Documents over-the-air array calibration in beam-steering mode and motivates the non-ideal codebook model for RF-chain inhomogeneities.","marker":"[11]"},{"why":"Provides the misspecified Cramér–Rao bound used to derive the absolute lower bound (ALB) for localization under model mismatch.","marker":"[23]"},{"why":"Supplies the simplification that the pseudo-true parameter minimizes the residual norm between true and modeled noise-free signals, used to compute the ALB bias term.","marker":"[24]"},{"why":"Provides the SISO RIS-enabled downlink localization and synchronization system model on which the signal model in (1) is based.","marker":"[17]"},{"why":"Quantifies how hardware impairments create model mismatch in OFDM-based localization, motivating the need for beam calibration.","marker":"[8]"},{"why":"Frames the calibration problem in RIS-aided ISAC systems and positions this work's contribution.","marker":"[5]"}],"fun_headline_variants":["Combined beam model cuts RIS calibration error by 77%","Fusing three real-world beam errors enables sub-meter ISAC location","One model that adds coupling, codebook, and noise beats separate fixes","Sub-meter ISAC localization achieved by merging three beam errors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The measured ground-truth beam patterns come from a 16×16 planar RIS prototype, but the localization simulation uses a 1×16 linear array, and the paper does not state how the planar measurements are mapped onto the linear steering model in equation (3).","fun_headline_variants_meta":{"raw":{"variants":["Combined beam model cuts RIS calibration error by 77%","Fusing three real-world beam errors enables sub-meter ISAC location","One model that adds coupling, codebook, and noise beats separate fixes","Sub-meter ISAC localization achieved by merging three beam errors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000837,"raw_usage":{"total_tokens":3672,"prompt_tokens":986,"completion_tokens":2686,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":2613}},"tokens_in":602,"tokens_out":2686,"duration_ms":17575,"temperature":1.0,"reasoning_tokens":2613,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:18:36.090163+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Keep the original −40° to 40° azimuth data for fitting, but measure the same 16×16 prototype over additional azimuth and elevation cuts, then test whether the fitted $\\gamma(\\phi)$ and $\\tilde{W}$ predict those withheld cuts; if the reconstruction error climbs well above 0.94, the reported calibration is a fit to a single angular slice rather than a general beam model.","supporting_citations":[{"cited_title":"Design of low-cost single-layer 2-bit reflective pro- grammable metasurface based on folded ground,","cited_arxiv_id":null,"evidence_quote":"Supplies the measured ground-truth beam patterns from the 16×16 RIS prototype used to fit and evaluate all three beam models."},{"cited_title":"Online mutual coupling calibration using a signal source at unknown location,","cited_arxiv_id":null,"evidence_quote":"Supplies the Toeplitz mutual-coupling-matrix model that the MCM beam model is built on."},{"cited_title":"Over-the-air array calibration of mmwave phased array in beam-steering mode based on measured complex signals,","cited_arxiv_id":null,"evidence_quote":"Documents over-the-air array calibration in beam-steering mode and motivates the non-ideal codebook model for RF-chain inhomogeneities."},{"cited_title":"Performance bounds for parameter estimation under misspecified models: Fundamental findings and applications,","cited_arxiv_id":null,"evidence_quote":"Provides the misspecified Cramér–Rao bound used to derive the absolute lower bound (ALB) for localization under model mismatch."},{"cited_title":"RIS-aided near-field localization under phase- dependent amplitude variations,","cited_arxiv_id":null,"evidence_quote":"Supplies the simplification that the pseudo-true parameter minimizes the residual norm between true and modeled noise-free signals, used to compute the ALB bias term."},{"cited_title":"Siso RIS-enabled joint 3D downlink localiza- tion and synchronization,","cited_arxiv_id":null,"evidence_quote":"Provides the SISO RIS-enabled downlink localization and synchronization system model on which the signal model in (1) is based."},{"cited_title":"Modeling and analysis of ofdm-based 5G/6G local- ization under hardware impairments,","cited_arxiv_id":null,"evidence_quote":"Quantifies how hardware impairments create model mismatch in OFDM-based localization, motivating the need for beam calibration."},{"cited_title":"Calibration in RIS-aided Integrated Sensing, Localization and Communication Systems","cited_arxiv_id":"2409.16931","evidence_quote":"Frames the calibration problem in RIS-aided ISAC systems and positions this work's contribution."}],"review_version":1}