{"id":"992f154a-de9d-4343-b981-8a20fec7adaa","arxiv_id":"2506.20158","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A closed-loop estimator for rotatable-antenna systems, alternating MUSIC/LS estimation with gradient-ascent antenna reorientation, achieves lower NMSE than fixed, random, or isotropic antennas in LoS simulations.","lead":"This paper proposes a two-step channel estimation scheme for wireless base stations whose antennas can be mechanically rotated. The scheme alternates between estimating the direction and strength of incoming signals, then re-aiming the antennas to improve the next measurement.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The alternating estimation-orientation loop never fuses measurements across blocks, and the per-antenna orientation subproblems are identical, so the claimed 'diverse angular perspectives' gain is not implemented; the reported NMSE advantage appears to be a beamforming/received-SNR effect.","rationale":"The reader's weakest assumption was the LoS single-path model and the unsupported multipath extension. That concern is valid but secondary. The load-bearing issue I identify is more structural: the proposed algorithm, as written, does not implement the 'diverse angular perspectives' mechanism that motivates the paper. Because the per-antenna orientation subproblems are identical, all RAs converge to the same orientation, so the array manifold is unchanged up to a scalar gain. Because the estimator in each block uses only that block's data, there is no multi-perspective fusion across the M blocks. The observed NMSE improvement over benchmarks is therefore attributable to increased received SNR from steering, not to an improved estimation method that exploits additional spatial degrees of freedom. This does not make the empirical result false, but it means the central claim as stated in the abstract is unsupported by the actual equations. The authors should either demonstrate a mechanism that actually combines observations from different orientations (e.g., a joint estimator over all blocks) or reframe the contribution as 'RA orientation adjustment increases the effective SNR for a standard MUSIC+LS estimator.' Since this is a substantial interpretational gap but not a disproof of the simulated performance, the conditional verdict remains appropriate. The reader already noted a possible beamforming effect in their rationale, so there is partial agreement, but their formal weakest assumption was the LoS limitation rather than this implementation-level concern.","tokens_in":8635,"tokens_out":8220,"duration_ms":99425,"concrete_test":"Rerun the simulation while forcing all RAs to share a single common orientation equal to the optimized orientation produced by the proposed algorithm (i.e., replace the per-antenna optimization in Section III-B with a one-parameter common-orientation search). If the NMSE is unchanged from the reported proposed scheme, then the per-antenna orientation DoFs are unused and the alternating loop is equivalent to a fixed UPA with a common gain. As a second check, plot NMSE versus received SNR, e.g., ||h_k(Theta)||^2/sigma^2, for the proposed and 'without orientation adjustment' schemes; if the curves collapse, the claimed diversity gain is purely a received-SNR effect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central narrative is that adaptively rotating RAs during training 'enriches environmental observations from diverse angular perspectives' (Abstract, Section I). The implemented algorithm does not do this. In Section III-B, problem (18) is decomposed into N independent subproblems (19). Since neither the objective (19a) nor the constraints (19b)-(19c) depend on the antenna index n, every subproblem is identical, so the optimizer returns the same orientation f* for all N RAs. Consequently, the gain vector g_k(Theta) in (6) is a common scalar times the all-ones vector, and the effective array response b(theta,phi;Theta) = g(Theta) ⊙ a(theta,phi) in (12) is a scalar multiple of the fixed-array response a(theta,phi). Thus MUSIC's spectrum shape is unchanged; only the received SNR is scaled. Furthermore, the CSI estimation in (10)-(15) uses only the current block's observations; no estimator fuses measurements from different orientation blocks. The loop only re-aims the common boresight, increasing the received pilot power for the next single-block estimate. Since SNR is defined as transmit SNR, rho = 10log10(pbar/sigma^2) (Section IV), the comparison against 'without orientation adjustment' and 'random orientation' benchmarks does not control for received power. The reported NMSE gain can therefore be fully explained by a beamforming gain rather than by the claimed multi-perspective, diversity-based estimation mechanism. This directly undermines the abstract's claim that the scheme works by enriching observations from diverse angular perspectives, even though the empirical NMSE curves may reproduce.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers an uplink multiuser system in which the base station is equipped with rotatable antennas (RAs) whose boresights can be adjusted. The proposed channel estimation scheme alternates, over M blocks, between a CSI estimation sub-block (MUSIC for angle-of-arrival, least squares for path gains) and an orientation-adjustment sub-block (projected gradient ascent to maximize the sum channel gain). Simulations with K=3 LoS users show that the proposed scheme achieves lower NMSE than three benchmarks: random orientation, no orientation adjustment, and an isotropic-antenna system. The letter is clearly written, and the constituent estimation and optimization tools are standard, but the central narrative that adaptive rotations provide 'diverse angular perspectives' is not actually realized by the algorithm as specified, and the performance comparison does not control for received power.","tokens_in":8995,"tokens_out":6992,"duration_ms":76881,"significance":"This letter addresses a timely topic—channel estimation for rotatable antenna architectures—and proposes a two-stage training loop that combines standard spectral estimation with a cleanly formulated orientation optimization. The signal model, MUSIC spectrum in (12), least-squares path-gain estimator in (15), and projected-gradient-ascent update in (21) are internally consistent and clearly presented. If the claimed diversity benefit were genuinely achieved, the work would be a useful step for RA-enabled systems. However, the current algorithm reduces to a common orientation for all antennas, and no cross-block fusion is used, so the claimed 'multi-perspective' mechanism is not implemented. The reported NMSE gains are therefore plausibly explained by received-power/beamforming effects rather than by a new estimation principle. With a revised algorithm or a reframed contribution, the idea could still be of interest, but the present version requires substantial revision.","major_comments":[{"comment":"The orientation optimization in (18) is decomposed into N subproblems (19), but since neither the objective (19a) nor the constraints (19b)-(19c) depend on the antenna index n, every subproblem is identical. Consequently, all RAs are driven to the same orientation f*, so the gain vector g_k(Theta) in (6) is a common scalar times the all-ones vector, and the effective array response b(theta,phi;Theta) in (12) is a scalar multiple of the fixed-array response a(theta,phi). This means the array does not observe the environment from diverse angular perspectives; the MUSIC spectrum is reshaped only by a direction-dependent scalar factor, and the central claim in the Abstract and Section I of 'enriching environmental observations from diverse angular perspectives' is not realized. The reported NMSE improvement in Figs. 2-4 is thus likely a received-SNR/beamforming effect. The authors should either modify the design so that different antennas can take different orientations (e.g., by adding diversity or orthogonality constraints), or explicitly reframe the paper's contribution as adaptive beam pointing for improved received SNR rather than multi-perspective estimation.","section":"Section III-B (Eqs. (18)-(19))"},{"comment":"The CSI estimation in each block uses only the current block's observations; the received-signal model in (13)-(14) and the least-squares estimator in (15) do not fuse measurements from different orientation states across the M blocks. Thus the sequential adjustment of orientations over time does not combine information from multiple measurement perspectives. This directly contradicts the abstract's and Section I's claim that sequentially adjusting orientations enriches the channel observations. If the authors intend to use the multi-block structure to improve estimation, they need to specify how measurements are aggregated (e.g., joint MUSIC over all blocks or an iterative refinement); if not, the paper should not claim diversity gains from the temporal adaptation.","section":"Section III-A (Eqs. (10)-(15))"},{"comment":"The SNR is defined as transmit SNR, rho = 10 log10(pbar/sigma^2), and the benchmarks (without orientation adjustment, random orientation, isotropic antenna) do not have the same directional gain as the proposed scheme. Since the proposed scheme steers all antennas toward the users, it enjoys a receive array gain that the benchmarks lack, so the NMSE comparison conflates estimation quality with receive power. A fairer evaluation would either equalize the received SNR across schemes (e.g., by comparing against a fixed-orientation system with the same directional gain but no adaptation) or report NMSE as a function of received SNR. Without such a control, the plots in Figs. 3-4 do not demonstrate that the adaptive orientation loop improves estimation per se rather than merely boosting SNR.","section":"Section IV (simulation setup and Fig. 3)"},{"comment":"The channel model assumes one LoS path per user, and the algorithm estimates only top-K MUSIC peaks and optimizes the orientation objective (19) using a single direction per user. The claimed extension to multipath environments in footnote 1 is not derived: with multiple paths, the top-K peaks of the MUSIC spectrum do not map cleanly to the K users, and the optimization in (18)-(19) would need a multi-path per-user formulation. This limitation should be stated explicitly in the abstract and conclusion, or the scheme must be extended and validated for scattering environments. As written, the general claim in the title and introduction overstates the applicability of the results.","section":"Section II-A and footnote 1"}],"minor_comments":[{"comment":"'light-of-sight' should be 'line-of-sight'.","section":"Section II-A"},{"comment":"The definition of b(theta,phi;Theta) = g(Theta) * a(theta,phi) is ambiguous because the gain vector g(Theta) is written without an explicit dependence on the candidate direction (theta,phi); clarify that the gain pattern is evaluated at the candidate AoA.","section":"Eq. (12)"},{"comment":"The sample covariance estimator would benefit from explicitly stating that the expectation is replaced by the sample average over T_m^E snapshots; the current phrasing 'by exploiting T_m^E time slots' is unclear.","section":"Eq. (10)"},{"comment":"The stated gradient-ascent complexity O(3N/epsilon^2) relies on a given solution accuracy epsilon, but the number of iterations is not specified and the projected gradient method for the nonconvex problem (19) does not have known O(1/epsilon^2) iteration complexity; please provide a justification or state the iteration count.","section":"Section III-C"},{"comment":"The x-axis label 'Directions (degree)' is awkward; 'Angle (degrees)' would be clearer.","section":"Fig. 2 caption"},{"comment":"The phrase 'the proposed RA-enabled channel estimation method with preset orientations' is confusing because the whole scheme is about adapting orientations; consider rewording to avoid implying orientations are preset.","section":"Section IV, paragraph after Eq. (NMSE)"},{"comment":"The conclusion is very brief and mostly restates the abstract; a sentence or two about limitations (LoS, known K, single-path model) and possible extensions would strengthen the paper.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"This paper's basic components are standard, and the proposed training protocol is clearly described. However, the central claim that adaptive rotation provides diverse angular perspectives is not implemented by the actual algorithm: the per-antenna orientation subproblems are identical, and no cross-block fusion is performed. The reported performance gains are therefore explainable by receive array gain rather than by a new estimation mechanism. I recommend requiring the authors to either modify the algorithm to truly exploit diversity (e.g., different orientations per antenna or joint multi-block processing) or substantially reframe the contribution and adjust the simulation comparisons accordingly. The issue is verifiable from Eqs. (18)-(19) and (10)-(15) and is within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read Xiong et al. in one sitting. The stress-test note is right, and it matters more than the reader's take suggests. What's genuinely new is the closed-loop structure: alternate MUSIC/LS estimation of AoAs and path gains with a projected-gradient update of RA orientations, repeated over M training blocks. That loop has not appeared in the RA literature they cite, and the component math is standard and internally consistent. The steering vector, directional gain model, MUSIC spectrum in (12), LS gain estimate in (15), gradient ascent in (20)-(21), and the complexity summary all check out. Simulations are described well enough to reproduce and do not fit any constants, so there is no circularity. The citation pattern is fine: the self-citations are the prior RA modeling works this builds on.\n\nThe soft spot is not just the LoS assumption, though that is real and footnote 1's multipath extension is hand-waving. The bigger issue is in Section III-B. Problem (18) decomposes into N subproblems (19), and neither the objective nor the constraints depend on n. Every antenna therefore gets the same orientation f*. Since all RAs share the same gain pattern, g_k(Theta) in (6) is a scalar times the all-ones vector, and b(theta, phi; Theta) in (12) is a scalar multiple of the fixed-array response a(theta, phi). The MUSIC spectrum shape is unchanged; only received SNR is scaled. Moreover, each block's CSI estimate uses only that block's observations; no measurement fusion happens across M blocks. So the abstract's 'diverse angular perspectives' is not implemented. The reported NMSE gains are a beamforming/received-power effect, not a multi-perspective estimation gain, and since SNR is defined as transmit SNR, the benchmarks do not control for received power.\n\nThat does not make the paper worthless. A short letter saying 'steer all RA boresights toward estimated user directions during training, then re-estimate' is a plausible engineering contribution for RA hardware. But the claimed mechanism is wrong. The contribution should be reframed, and a stronger version would intentionally give different antennas different orientations, fuse measurements across blocks, include an oracle or Cramer-Rao baseline, and test in multipath.\n\nI would send it to review because the math is checkable and the flaw is fixable, but I would expect the accepted version to be about beam-steering-aided training, not diverse perspectives.","headline":"The closed-loop RA channel-estimation loop is real and checkable, but its advertised diversity mechanism is not implemented: all RAs get the same orientation, so the NMSE gain is beamforming, not multi-perspective estimation.","tokens_in":9522,"tokens_out":3440,"would_cite":false,"duration_ms":40462,"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":"This paper claims that a rotatable-antenna base station can improve its channel estimation accuracy by alternating between estimating angles of arrival and path gains and re-aiming antenna boresights to maximize effective channel gain.","keywords":["rotatable antenna","channel estimation","angle-of-arrival estimation","MUSIC","path gain","orientation adjustment","directional antenna gain","normalized mean square error"],"falsifier":"A simulation or over-the-air measurement in a scattering environment where each user has two or more paths with comparable strengths: if the proposed scheme's normalized mean square error is no better than a fixed-orientation array, the central claim fails. Concretely, generate K users each with a two-ray channel, run the proposed alternating estimation, and compare NMSE to the 'without orientation adjustment' benchmark.","tokens_in":8349,"feed_emoji":"📶","tokens_out":3013,"duration_ms":29656,"temperature":0.7,"pith_summary":"This letter proposes a channel estimation scheme for a base station whose antennas can rotate their boresight directions. Over each training period the scheme alternates between estimating angles of arrival and path gains with MUSIC and least squares, and rotating the antennas to maximize the expected received signal strength. The paper argues that actively steering the antennas during training yields more accurate channel estimates than fixed, random, or isotropic antenna configurations.","feed_headline":"Active antenna rotation cuts channel estimation error","feed_subtitle":"Re-aiming boresights during training lowers CSI error below fixed, random, or isotropic settings.","key_machinery":"The key machinery is the alternating training block: (1) a MUSIC pseudo-spectrum whose steering vector is modified by the orientation-dependent gain pattern, $\\mathbf{g}(\\Theta)\\odot \\mathbf{a}(\\vartheta,\\varphi)$, used to read off the top-$K$ angles of arrival; (2) a least-squares estimate of the path gains from the same measurements; and (3) a projected gradient ascent step on the unit-sphere constraint $\\mathbf{f}_n^T\\mathbf{e} \\ge \\cos\\theta_{\\max}$ that rotates each antenna to maximize the sum channel gain $\\sum_{k=1}^K \\mu_k(\\mathbf{f}_n^T\\bar{q}_k)^{2p}$. The gradient $2p\\sum_{k=1}^K \\mu_k(\\mathbf{f}_n^T\\bar{q}_k)^{2p-1}\\bar{q}_k^T$ drives the update.","core_discovery":"The central claim is that a rotatable-antenna array can improve its own channel estimation accuracy by using the current estimate to reposition antenna boresights before the next measurement block. In the system model each user produces a single line-of-sight path, and the channel vector is the Hadamard product of an orientation-dependent gain pattern, a steering vector, and a scalar path gain. The paper shows numerically that alternating MUSIC-based angle-of-arrival estimation with projected gradient ascent for orientation adjustment reduces normalized mean square error relative to benchmarks across SNR and array size.","pith_inferences":["Editorial inference: The same alternating principle could be applied to other reconfigurable-aperture architectures such as movable or fluid antennas, where the position (instead of the orientation) is updated based on the current channel estimate.","Editorial inference: The paper's reliance on a single line-of-sight path per user suggests the performance gain may shrink in rich scattering environments; extending the top-$K$ MUSIC selection to a per-path association would be the natural next step.","Editorial inference: A testable extension is to replace the fixed number of training blocks $M$ with an adaptive stopping rule that halts orientation updates once the estimated channel gain stops improving, which could reduce pilot overhead in slowly varying channels.","Editorial inference: The cosine pattern model with exponent $p$ is idealized; real antenna patterns have sidelobes and beamwidths that could affect the MUSIC spectrum, so a robustness study with measured patterns would clarify the practical gains."],"forward_implications":["If the central claim is correct, a rotatable-antenna base station can estimate channel state information more accurately without extra spectrum or pilot power, by exploiting the already-available orientation degree of freedom.","The proposed alternating procedure implies that channel estimation and beamforming can be coupled in a single training period, reducing the need for a separate offline calibration stage.","Because the orientation optimization depends only on the estimated user directions and path gains, the scheme is directly compatible with standard uplink pilot designs.","The numerical results suggest that the accuracy gap grows with the number of antennas, meaning the benefit of active orientation adjustment is larger for larger arrays.","The method's complexity is moderate and scales polynomially with array size and training block count, making it feasible for practical deployment scenarios."],"supporting_citations":[{"why":"Supplies the rotatable antenna system model and the independent boresight adjustment mechanism that the channel estimation scheme builds on.","marker":"[8]"},{"why":"Provides the MUSIC algorithm used for angle-of-arrival estimation in the first procedure.","marker":"[12]"},{"why":"Provides the cosine power antenna gain pattern model that couples orientation to effective channel gain.","marker":"[11]"},{"why":"Supports the assumption that the number of users $K$ is known or correctly estimated, which the MUSIC peak selection relies on.","marker":"[13]"},{"why":"Establishes the rotatable antenna system model and quantifies its performance gains, forming the basis for the channel vector in (7).","marker":"[7]"}],"fun_headline_variants":["Rotatable antennas adapt to sharpen channel estimates","Adaptive rotation improves channel estimation","Rotate antennas mid-training for better CSI","RA channel estimation: adjust boresights on the fly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The channel is assumed to have exactly one resolvable line-of-sight path per user and a known number of users K; if a user's signal arrives over multiple paths, the top-K MUSIC peaks do not correspond cleanly to users and the orientation objective is not defined per user.","fun_headline_variants_meta":{"raw":{"variants":["Rotatable antennas adapt to sharpen channel estimates","Adaptive rotation improves channel estimation","Rotate antennas mid-training for better CSI","RA channel estimation: adjust boresights on the fly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000663,"raw_usage":{"total_tokens":2985,"prompt_tokens":856,"completion_tokens":2129,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":472,"completion_tokens_details":{"reasoning_tokens":2073}},"tokens_in":472,"tokens_out":2129,"duration_ms":15935,"temperature":1.0,"reasoning_tokens":2073,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:20:55.252558+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A simulation or over-the-air measurement in a scattering environment where each user has two or more paths with comparable strengths: if the proposed scheme's normalized mean square error is no better than a fixed-orientation array, the central claim fails. Concretely, generate K users each with a two-ray channel, run the proposed alternating estimation, and compare NMSE to the 'without orientation adjustment' benchmark.","supporting_citations":[{"cited_title":"Statistical analysis of the performance of information theoretic criteria in the detection of the number of signals in array processing,","cited_arxiv_id":null,"evidence_quote":"Supports the assumption that the number of users $K$ is known or correctly estimated, which the MUSIC peak selection relies on."}],"review_version":2}