{"id":"2ee5ca0c-3a6e-42ec-833d-8db6c3018d0b","arxiv_id":"2507.19309","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A rotation-first, statistics-based optimizer for 6DMA antenna panels achieves nearly the same sum rate as alternating optimization with far lower complexity.","lead":"This paper proposes a cheaper way to aim and place movable antenna panels on a 6G base station, using long-term channel statistics instead of fast instantaneous estimates. It first picks each panel's rotation, then finds positions that physically fit, and reports near the performance of much more expensive alternating optimization.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline performance claim is not yet established: Fig. 2 appears to compare the proposed method's Jensen-approximated rate (19) with MC-AO's Monte Carlo rate (16), and the paper never states that all curves are evaluated on the true rate.","rationale":"The reader's rotation-restriction concern in Eq. (22) is real: P2-A restricts q_b to d_ins n(u_b) and labels this a relaxation, and P2-B's heuristic has no guarantee. However, the most load-bearing issue for the central claim is empirical: whether the simulation actually demonstrates comparable true performance. The paper's MC-AO benchmark is defined by Monte Carlo evaluation of Eq. (16), while the proposed optimization uses approximation (19); if Fig. 2 plots the approximation for the proposed method, the comparison cannot support 'comparable communication performance.' The concrete Monte Carlo re-evaluation directly settles this and also tests the Jensen bound's effect. I do not propose changing the CONDITIONAL verdict; the test would turn the condition into a checkable requirement.","tokens_in":9781,"tokens_out":8341,"duration_ms":88383,"concrete_test":"Reproduce Fig. 2 using the paper's parameters but report true average sum rate R_k(z)=E[log2(1+h_k^H B_k^{-1}h_k)] (Eq. 16) for every benchmark, evaluated with W=10^4 Monte Carlo channel realizations at the final optimized configurations. Also evaluate the approximation (19) for the proposed configuration and compare it with its Monte Carlo value. If the proposed method's true-rate curve is significantly below MC-AO's or if the approximation deviates from the true rate by more than the gap reported in Fig. 2, then the headline performance claim is not supported; if true rates match, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed sequential scheme achieves communication performance comparable to AO. The optimization objective (P1) uses rk in Eq. (19), obtained by two Jensen steps from the true rate in Eq. (16): first an upper bound, then a lower bound, so the gap to the true rate is unquantified. In Section IV, MC-AO is explicitly described as 'computing the expectation over rk by the Monte Carlo method' with W=10^3, but the paper never specifies whether the proposed scheme's curve in Fig. 2 is the true Monte Carlo rate or the approximation (19). If it is the approximation, the comparison is not apples-to-apples and 'comparable performance' may be an artifact of the unquantified Jensen gap. This matters before the rotation-restriction issue in Eq. (22): even if the restriction is harmless, a performance claim needs a valid true-rate comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers a 6DMA-equipped base station serving K users and optimizes the positions and rotations of B antenna surfaces to maximize the average sum log-rate under a unified blockage/overlap-avoidance constraint. Because Monte Carlo alternating optimization (MC-AO) is expensive, the authors propose a sequential method: first optimize rotations with all surface centers restricted to an inscribed sphere inside the feasible volume (P2-A, Eq. (22)), then construct feasible positions for the optimized rotations via a geometry-based greedy algorithm (P2-B, Eqs. (30)-(42)). The objective in (P1) replaces the true average rate in Eq. (16) by the double-Jensen surrogate in Eq. (19). Section IV reports one simulation comparing the proposed scheme with MC-AO, a fixed-position BS, and a position-adjustable antenna BS, and concludes that the proposed scheme is comparable to MC-AO with much lower complexity.","tokens_in":10051,"tokens_out":4303,"duration_ms":45373,"significance":"If the performance and complexity claims hold, the paper would make 6DMA configuration more practical by replacing Monte Carlo evaluation with a closed-form statistical-channel surrogate and a sequential rotation/position design. Strengths of the manuscript include a parameter-free analytical surrogate (no fitted constants), a greedy initialization strategy for rotations, a geometric position-construction idea, and comparisons against several benchmarks. However, the central claim of 'comparable performance with lower complexity' is not yet established: the surrogate's error relative to the true rate is unquantified, the rotation-only problem is a restriction mislabeled as a relaxation, the position-search heuristic lacks a feasibility proof, and the complexity reduction is not quantified. These are load-bearing gaps, so the manuscript requires major revision.","major_comments":[{"comment":"The manuscript never states whether the proposed scheme's curve in Fig. 2 is the true Monte Carlo rate from Eq. (16) or the double-Jensen surrogate from Eq. (19). The MC-AO benchmark explicitly uses W=10^3 Monte Carlo samples to evaluate Eq. (16), while the proposed algorithm optimizes the surrogate in Eq. (19). If the proposed curve is the surrogate, the comparison with MC-AO is not apples-to-apples, and the 'comparable performance' conclusion may be an artifact of the unquantified gap between (16) and (19). Please either evaluate the optimized configuration using Monte Carlo samples of the true rate and plot that curve, or provide a quantitative bound or empirical check showing that the surrogate error is small for the reported configurations.","section":"§II-C, Eq. (19) and §IV, Fig. 2"},{"comment":"P2-A is called a relaxation of (P1), but setting q_b = d_ins n(u_b) is a restriction: it removes the position variables entirely and confines every surface center to the inscribed sphere of the feasible volume. No argument, proof, or ablation is given that rotations optimizing the restricted problem remain near-optimal after positions are restored in P2-B. If the optimal rotations for the original problem require surface centers that are not on the inscribed sphere, the proposed rotation-first method can be substantially suboptimal. Please provide a dominance or approximation relation between P2-A and P1, or add an ablation that optimizes rotations with free positions (e.g., a coarse position grid) and compares against the inscribed-sphere restriction.","section":"§III-A, Eq. (22)"},{"comment":"The geometry-based position search is a heuristic, and the paper does not prove that the returned q satisfies the feasibility constraints (30). In particular, the shift rule in Eq. (40) moves already-positioned surfaces outward along their projected normals; the manuscript does not show that this operation preserves the mutual constraints among those already-positioned surfaces. Moreover, the first surface is assigned 'arbitrarily, for example, the coordinate origin,' which may not lie inside V6DMA in general. Please state and prove the invariant maintained by Steps 1-3, or empirically report the fraction of runs for which all returned positions satisfy (30), and confirm that the simulation's returned positions indeed satisfy the constraints.","section":"§III-B, Eqs. (30)-(42)"},{"comment":"The abstract and introduction claim that the proposed scheme 'significantly reduces the computational complexity' of AO, but Section IV provides no runtime, number of objective evaluations, floating-point-operation count, or iteration-count comparison. Since complexity reduction is a central advertised contribution, please quantify it, for example by reporting the average number of objective/gradient evaluations per configuration for the proposed method versus MC-AO, or by reporting wall-clock times.","section":"§IV (complexity claim)"}],"minor_comments":[{"comment":"Equation (8) appears ill-defined: the antenna-gain matrix G(z, f_{k,l}) depends on the path index l, so it cannot factor out of the concatenated matrix [a(z,f_{k,1}), ..., a(z,f_{k,L_k})] as written. Please write the l-th column as G(z,f_{k,l})a(z,f_{k,l}) and define \\tilde A column-wise.","section":"§II-A, Eq. (8)"},{"comment":"The objective in (P1) uses log(r_k) without specifying the base, while Eq. (16) uses log2 and Fig. 2 is labeled 'log-rate.' Please state the base consistently; if natural log is used in (P1), the figure's numerical values will differ from log2 rates by a factor of ln(2).","section":"§II-C and §IV, Fig. 2"},{"comment":"The simulation reports a single curve per scheme with no confidence intervals, error bars, or description of how many random channel/geometry realizations were averaged. Adding such details would make the 'comparable performance' claim more convincing.","section":"§IV, Fig. 2"},{"comment":"The gradient is computed by finite differences with epsilon = 2^{-16}; at this scale, cancellation error may be significant for the double-Jensen surrogate. Please comment on the sensitivity of the results to the choice of epsilon or use a complex-step/analytic gradient where possible.","section":"§III-A, Eqs. (25)-(27)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things. First: this is a genuinely useful direction for 6DMA optimization. The authors replace Monte Carlo alternating optimization with a sequential method: optimize all surface rotations assuming positions lie on the inscribed sphere of the feasible region, then run a geometric CER algorithm to find actual positions that respect blockage and overlap constraints. That decomposition is new relative to their earlier AO work, and using statistical channel information is the right practical framing for mechanically tuned antennas. Second: the headline claim—comparable performance to MC-AO at much lower complexity—is not yet established. The comparison in Fig. 2 appears to pit MC-AO's true Monte Carlo rate (16) against the proposed method's double-Jensen approximation (19), and the paper never states which curve is plotted for the proposed scheme. Two Jensen steps (upper then lower) leave an unquantified gap, so 'comparable' may be partly an artifact.\n\nThe paper does good things. The system model is careful, the unified halfspace overlap constraint is sensible, and the CER algorithm is a concrete, well-specified heuristic that does find feasible placements in the simulation. The manuscript is honest in its limitations section insofar as it admits the method is heuristic, though it does not flag the relaxation/restriction problem or the surrogate-rate comparison.\n\nSoft spots, in proportion. The biggest is Eq. (22): the paper calls P2-A a relaxation, but setting q_b = d_ins n(u_b) restricts every surface to the inscribed sphere of V6DMA. That shrinks the feasible set, and no bound or ablation shows the lost position freedom is harmless. Second, the CER algorithm has no proof of success—no guarantee that the greedy shifting yields a feasible configuration, or that the returned positions stay inside V6DMA. Third, the complexity reduction is asserted but never measured; there is no runtime or per-iteration cost comparison. Fourth, the entire performance evaluation is a single simulation with one user/scatterer geometry. None of these are fatal individually, but together they mean the central claim is currently supported by a single, potentially apples-to-oranges curve.\n\nThe math and derivations are otherwise sound; the Jensen bounds are correctly applied, and I found no circular reasoning or fitted constants. The self-citation pattern is normal for a group pushing this line of work.\n\nWho is this for? People working on 6DMA or movable-antenna systems will want to read it, and the sequential idea is worth citing even in its current form. It deserves a serious referee: the idea is strong enough that the gaps should be addressed rather than the paper desk-rejected. I would send it out with a request for major revision: quantify the Jensen gap or plot true-rate curves for both schemes, justify or ablate the sphere restriction, and report actual complexity numbers.\n\nFor you: worth a skim if you follow the area, but I would not change any of my current work based on the performance claim yet.","headline":"A plausible low-complexity 6DMA scheme that reduces the problem to a rotation-then-position heuristic, but the performance claim rests on an unquantified Jensen approximation and a restriction mislabeled as a relaxation.","tokens_in":10479,"tokens_out":1450,"would_cite":true,"duration_ms":17133,"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":"The paper proposes and simulates a rotation-first, statistics-based sequential optimizer that configures 6DMA surfaces at a fraction of the computational cost of alternating optimization while keeping comparable average sum log-rate.","keywords":["six-dimensional movable antenna","6DMA","statistical channel information","sequential optimization","alternating optimization","sum log-rate maximization","antenna placement constraints","gradient ascent"],"falsifier":"Take a user-scatterer geometry whose optimal 6DMA configuration is known, by exhaustive search over a coarse grid, to place surface centers near the corners of the feasible volume rather than on the inscribed sphere; run the proposed rotation-first method and the Monte Carlo alternating-optimization benchmark on that geometry and compare average sum log-rate. If the proposed method lags by more than the slack of the Jensen bound, the inscribed-sphere restriction is the cause.","tokens_in":9569,"feed_emoji":"📡","tokens_out":7167,"duration_ms":69790,"temperature":0.7,"pith_summary":"This paper is trying to show that configuring a six-dimensional movable antenna (6DMA) base station need not alternate between optimizing surface positions and rotations. Its proposed scheme first fixes every surface's rotation while placing the surface center on the largest inscribed sphere of the allowed region, then finds feasible positions that realize those rotations without overlap or mutual blockage. Because the average user rates are replaced by a closed-form lower bound built from channel covariance matrices, the method avoids Monte Carlo averaging over channel realizations. If the paper is right, a low-complexity, statistics-driven procedure can deliver the same average sum log-rate as costly alternating optimization while beating fixed-position and position-only antenna arrays.","feed_headline":"Rotation-first 6DMA tuning matches Monte Carlo at far lower cost","feed_subtitle":"A rotation-first scheme cuts complexity while keeping 6DMA sum-rate close to alternating optimization.","key_machinery":"The enabling move is the substitution $z_b(u_b) = [d_{\\rm ins} n(u_b)^\\top, u_b^\\top]^\\top$, which ties each surface center to the inscribed sphere of radius $d_{\\rm ins}$ inside the feasible volume, so rotation variables alone drive the first-stage objective. The objective is the covariance-based lower bound $\\underline{r}_k$ from (19), whose finite-difference gradient is ascended with a backtracking line search; rotations are initialized by a greedy search over a uniformly generated candidate set of rotations. Positions are then found by representing each surface by a circular extended region of radius $d/2$ and sequentially placing tangent planes so that all extended regions lie in the correct half-spaces, guaranteeing the optimized rotations are realizable without surface overlap or mutual blockage.","core_discovery":"The central claim is that the joint 6DMA position-rotation problem can be solved sequentially: optimize rotations first with positions pinned to the inscribed sphere, then repair positions through a geometry-based feasibility algorithm. The rate objective is the analytic Jensen lower bound $\\underline{r}_k(\\Sigma(z)) = \\log_2\\left(1+\\operatorname{tr}\\left(\\mathbb{E}[B_k]^{-1}\\Sigma_k(z)\\right)\\right)$, so gradients can be computed by finite differences and no Monte Carlo expectation is needed. The paper reports simulations, with eight four-antenna surfaces, five users, and three dominant scatterers, in which the sequential scheme achieves average sum log-rate close to the Monte Carlo alternating-optimization benchmark at lower complexity, and superior to a fixed three-sector array and a position-adjustable array.","pith_inferences":["A natural testable extension is to stress the two-stage ordering in near-field or corner-heavy user geometries, where optimal surface centers may lie far from the inscribed sphere; if the gap to full joint optimization grows there, the position-repair stage would need to feed back into the rotation choice.","The same rotation-first separation could apply to other mechanically steerable surfaces, such as intelligent reflecting surfaces or UAV-mounted arrays, where the slow variable is orientation and position is mostly a feasibility constraint.","The finite-difference gradient over the rotation space is simple but scales with the number of surfaces; an analytic gradient of the Jensen bound would remove the per-surface function-evaluation overhead and is a straightforward extension of this scheme."],"forward_implications":["6DMA reconfiguration can be driven by statistical channel information that changes slowly, so mechanical adjustment does not need instantaneous channel knowledge.","The explicit Jensen bound removes the Monte Carlo expectation, cutting the dominant computational cost of the alternating-optimization benchmark while keeping rates comparable.","The geometry-based position repair yields feasible placements inside the 6DMA region, so the optimized rotations are practically implementable rather than abstract.","In the simulated settings, the scheme outperforms fixed-position and position-adjustable baselines, which suggests 6DMA surfaces' rotation freedom, not just position freedom, is what buys the rate gain."],"supporting_citations":[{"why":"Supplies the 6DMA channel model, the alternating-optimization formulation, and the Monte Carlo benchmark that the proposed scheme is compared against.","marker":"[6]"},{"why":"Extends 6DMA design to discrete position and rotation choices and motivates the joint position-rotation optimization problem addressed here.","marker":"[7]"},{"why":"Supplies the backtracking line-search rule used to set step sizes in the rotation-gradient ascent.","marker":"[20]"},{"why":"Defines the antenna gain pattern used in the simulations that produce the reported rate comparisons.","marker":"[21]"},{"why":"Supplies the particle-swarm optimizer used by the position-adjustable antenna benchmark.","marker":"[22]"}],"fun_headline_variants":["Rotate-then-place 6DMA tuning cuts complexity, holds rate","6DMA sequential design: rotations first, then positions","Low-cost 6DMA: optimize rotations before antenna positions","Rotation-first 6DMA achieves AO rate at lower complexity","Sequential 6DMA: fix rotations, then repair positions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that restricting every surface center to the inscribed sphere during rotation optimization does not exclude the configurations needed for near-optimal rates once positions are later repaired.","fun_headline_variants_meta":{"raw":{"variants":["Rotate-then-place 6DMA tuning cuts complexity, holds rate","6DMA sequential design: rotations first, then positions","Low-cost 6DMA: optimize rotations before antenna positions","Rotation-first 6DMA achieves AO rate at lower complexity","Sequential 6DMA: fix rotations, then repair positions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000615,"raw_usage":{"total_tokens":2821,"prompt_tokens":870,"completion_tokens":1951,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":1862}},"tokens_in":486,"tokens_out":1951,"duration_ms":15206,"temperature":1.0,"reasoning_tokens":1862,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:55:37.429783+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a user-scatterer geometry whose optimal 6DMA configuration is known, by exhaustive search over a coarse grid, to place surface centers near the corners of the feasible volume rather than on the inscribed sphere; run the proposed rotation-first method and the Monte Carlo alternating-optimization benchmark on that geometry and compare average sum log-rate. If the proposed method lags by more than the slack of the Jensen bound, the inscribed-sphere restriction is the cause.","supporting_citations":[{"cited_title":"6D movable antenna en- hanced wireless network via discrete position and rotation optimization,","cited_arxiv_id":null,"evidence_quote":"Extends 6DMA design to discrete position and rotation choices and motivates the joint position-rotation optimization problem addressed here."},{"cited_title":"Particle swarm optimization,","cited_arxiv_id":null,"evidence_quote":"Supplies the particle-swarm optimizer used by the position-adjustable antenna benchmark."}],"review_version":1}