{"id":"3d26bb9e-245f-4ba8-9a2d-8850d696b523","arxiv_id":"2505.08356","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An adaptive-radius Lorentzian mapping (ARLCH) for DMA beamforming reduces transmit power by over 20% versus existing Lorentzian holography schemes in simulated multi-user MISO networks.","lead":"This paper proposes a new way to tune beamforming in dynamic metasurface antennas (DMAs), called ARLCH, and reports about 20% lower transmit power than existing tuning methods in simulated multi-user networks. It also provides a unified framework, GMLCH, for comparing three standard Lorentzian-constrained tuning schemes in the same optimization setup.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ARLCH cannot reduce Euclidean projection error below LCEH, so its stated objective (25) is not the mechanism for the reported power gains; the central claim's rationale is unsupported.","rationale":"The reader's weakest assumption was that reducing the Euclidean projection error in (25) is only empirically, not analytically, linked to lower transmit power. My stress-test finds a stronger, structural problem: ARLCH's final weights are on the same unit Lorentzian circle Q as LCEH, and LCEH is by definition the elementwise Euclidean projection onto Q. Therefore ARLCH cannot improve the Euclidean projection error relative to LCEH; it can only worsen it. The optimization in (25) minimizes distance to a scaled, pre-normalization circle, and the algorithm then discards the scale, so the stated objective does not describe the final output. This means the paper's explanation for ARLCH's advantage is geometrically impossible as stated. The reported transmit-power gains may still be reproducible, but they would have to come from an unexamined and unstated mechanism, and the paper's central narrative would need substantial revision. This does not by itself disprove the empirical result, but it raises the correctness risk and demands a major revision with additional evidence, hence CONDITIONAL rather than ACCEPT or UNCHANGED.","tokens_in":22926,"tokens_out":19266,"duration_ms":193442,"concrete_test":"For the Monte-Carlo realizations of Section VI-B (e.g., Fig. 11), record the SDR ideal vector q* from Step 6, the LCEH-mapped vector q_LCEH, and the ARLCH-mapped vector q_ARLCH at the final AO iteration for the same q*. Compute and report the squared Euclidean projection errors ||q*−q_LCEH||^2 and ||q*−q_ARLCH||^2 averaged over realizations and per realization. Since LCEH is the elementwise projection onto the unit Lorentzian circle, the test should show ||q*−q_ARLCH||^2 ≥ ||q*−q_LCEH||^2 for every realization; if instead ARLCH is sometimes closer, the LCEH implementation is not performing the projection (tie-breaking issue). In either case, the paper must either demonstrate that a larger projection error is causally linked to lower transmit power, or revise the claimed mechanism.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"ARLCH's final output lies on the unit Lorentzian circle Q (Algorithm 2, Step 11, via (26)). For a fixed ideal vector q*, the Euclidean projection onto Q^N is obtained elementwise, and LCEH (GMLCH with center (0,0.5), Section IV-A2) is exactly that projection: it minimizes |q*_n − q_n| for each n independently. Hence no mapping whose output is in Q^N can have a smaller Euclidean distance to q* than LCEH. ARLCH nevertheless solves (25), which minimizes ||q* − D(j+e^{jΦ})/2|| over a free diameter D before normalization; a large D can make the scaled point close to q*, but the algorithm then discards D and keeps only the phases (Step 11). The final normalized point is therefore not the minimizer of (25) and is generally farther from q* than LCEH's projection. The paper's claim that ARLCH 'minimizes the discrepancy between ideal and physically feasible weights' is consequently false: LCEH already achieves that minimum. Any transmit-power improvement of ARLCH over LCEH must arise from an unstated mechanism, such as the power objective being non-monotonic in Euclidean projection error, which the paper neither derives nor tests. This leaves the central 20% power-reduction claim without its stated theoretical support.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers a downlink multiuser MISO system served by a DMA-aided base station, where the design variables are the digital precoding vectors and the Lorentzian-constrained analog DMA weights. The authors formulate a total-transmit-power minimization problem with per-user SINR constraints, relax it into two SDP subproblems for the digital precoders and for the DMA weights, and combine them in an alternating optimization loop (Algorithm 1). A projection step then maps the SDR-derived ideal weights onto the Lorentzian circle. The paper generalizes existing Lorentzian mappings into a framework called GMLCH, parameterized by a mapping center, which specializes to LCPH, LCEH, and LCUSH. It further proposes ARLCH, which adapts the Lorentzian-circle diameter during the projection step and is claimed to provide an additional degree of freedom, lower transmit power, and better scalability as the number of users grows. Numerical results report transmit-power reductions of about 20% relative to LCUSH/ADMM-SCA benchmarks, with gains over LCEH of 16.7% at K=6 and 29.1% at K=8.","tokens_in":23241,"tokens_out":9871,"duration_ms":108628,"significance":"The unified GMLCH treatment is a useful organizing contribution: it provides one SDP-based alternating-optimization platform in which LCPH, LCEH, and LCUSH can be compared under identical conditions, and the comparison of mapping centers in a SINR-constrained multiuser setting appears not to have been done systematically before. The numerical finding that LCEH outperforms the more commonly used LCUSH is potentially valuable for practitioners. Lemma 1 is algebraically correct for the scalar subproblem it solves, and the SDP derivations in Section III follow standard lines. However, the central theoretical justification for ARLCH is not supported: the final ARLCH weights do not minimize the stated Euclidean projection error (25), and in fact cannot have smaller Euclidean distance to the ideal weights than LCEH. The claimed power gains may well be real, but they are currently validated only by Monte Carlo point estimates, with no confidence intervals and no derivation linking the projection objective to the power objective.","major_comments":[{"comment":"The final output of ARLCH does not minimize the objective in (25), and the paper's claim that ARLCH minimizes the discrepancy between ideal and physically feasible weights is false. Since LCEH (Section IV-A2, center (0,0.5)) is the elementwise Euclidean projection onto the unit Lorentzian circle Q, any output qhat in Q^N satisfies ||q* - qhat|| >= ||q* - M(q*;0,0.5)||. ARLCH solves (25) by allowing a free diameter D, but Step 11 discards D_f and outputs qhat_f=(j+e^{j Phi_f})/2, which lies in Q^N. Therefore ARLCH cannot reduce the Euclidean projection error below LCEH. The paper must either provide a different, correct mechanism for the reported power gains (for example, showing that the mapping minimizing the total power (7) is not the Euclidean projection) or substantially revise the claims made around (25), the Abstract, and the Conclusion. Relatedly, the statement that ARLCH works in a 'greater optimization space' is not supported, because the final weights are normalized back onto the same unitary Lorentzian circle.","section":"Section IV-B, Eq. (25), Algorithm 2 Step 11"},{"comment":"The one-dimensional optimization problems defining the mapped phases are degenerate. For any line through the mapping center, every intersection point of that line with the Lorentzian circle gives zero value of the objective in (22) and in (29), and a line through the center intersects the circle in two points. The text does not specify whether the near or far intersection point is selected. This ambiguity affects every GMLCH variant and the ARLCH phase update, and it is not resolved by the figures, which appear to implicitly select the near point. The authors should state the selection rule explicitly and discuss whether the far point can be selected and what effect that would have on SINR feasibility.","section":"Section IV-A, Eq. (22); Section IV-B, Eq. (29)"},{"comment":"The rank-one recovery from the SDP solution of (16) is not shown to be tight for K>1, and no convergence proof is given for the alternating procedure in Algorithm 1. The text asserts that the alternating execution 'effectively drives the convergence' and that the DMA weights are optimized 'toward high-quality suboptimal solutions,' but these claims are unsupported. Since the headline 20% power-reduction result is measured from this algorithm, the paper should provide at least a monotone-convergence argument (Step 11 already enforces non-increasing P_Tx) and some quantification of the SDR rank-one gap for the system sizes used in the Monte Carlo experiments.","section":"Section III-B and Algorithm 1"},{"comment":"The optimal diameter in (27) is not constrained to be positive, although D is described as the physical diameter of the Lorentzian circle. For ideal weights with Re(qhat^H q*) < 0, the expression gives D* < 0, and Algorithm 2 does not specify any clamping, absolute value, or feasibility handling. This is not a remote case: SDR ideal weights can have arbitrary phases, so negative values can occur. The authors should either prove D* >= 0 under the SDR solution or add a D >= 0 constraint and re-derive the closed form, and they should state how Algorithm 2 behaves when the unconstrained optimum is negative.","section":"Appendix, Lemma 1, Eq. (27)"},{"comment":"The reported gains, including the abstract's 'over 20%' reduction and the 16.7% and 29.1% reductions relative to LCEH at K=6 and K=8, are point estimates over 1000 realizations without confidence intervals or significance tests. The spread shown in Fig. 10 indicates that the per-realization ratio varies considerably, so mean-only reporting is insufficient. Please provide quantiles or standard errors, state whether the percentages are ratios of means or means of ratios, and report the fraction of realizations in which ARLCH actually outperforms LCEH.","section":"Section VI-B, Figs. 10 and 11"}],"minor_comments":[{"comment":"The line equation appears to contain a typo: 'sin(x_c)' should presumably be 'sin(phi*_n)' in the slope term, since the line is meant to pass through (cos(phi*_n), sin(phi*_n)) and (x_c, y_c).","section":"Section IV-A, Eq. (20)"},{"comment":"The notation switches between 'd' and 'D' without definition; please define d = D* before Eq. (29) and use the same symbol throughout the phase-update subproblem.","section":"Section IV-B, around Eq. (29)"},{"comment":"The reported wall-clock runtimes (LCUSH 117.28 s, baseline 116.07 s, ARLCH 119.25 s) are for a single instance; please state the hardware/software platform and either report variance over multiple runs or move the timing discussion to an appendix.","section":"Section VI-A"},{"comment":"The x-axis labels '6/6, 6/4, 6/2, 6/1' are ambiguous; they should be clarified as pairs (N_e, d_x) or replaced with explicit spacing values.","section":"Fig. 12"}],"recommendation":"major_revision","confidential_remarks":"The unified GMLCH comparison could be a solid contribution, and the empirical power gains reported for ARLCH may be genuine. However, the paper's central justification for ARLCH contains a correctness error: the final ARLCH weights cannot improve the Euclidean projection error relative to LCEH, so the stated objective (25) is not the mechanism behind the gains. This is fixable within the manuscript's scope by reframing ARLCH as a heuristic mapping tuned to the power objective and by adding appropriate analysis and statistics, but it currently undermines the novelty claim. I would also ask the authors, during revision, to resolve the near/far intersection ambiguity and the sign issue in Lemma 1, since these affect the reproducibility of all the mapping methods."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper's main new algorithmic claim is unsupported: ARLCH is not minimizing Euclidean projection error, because LCEH already does that. Second, the multi-user comparison of Lorentzian mapping centers is new and useful, and the simulations appear honestly run.\n\nWhat the paper does well: it embeds GMLCH (from Bowen et al.) into an SDP-based alternating optimization for SINR-constrained power minimization, giving a clean way to benchmark LCPH, LCEH, and LCUSH in the same framework. The result that LCEH beats LCUSH, with the gap growing from 13.9% to 17% as K goes from 1 to 8, is a real empirical contribution and directly relevant to anyone designing DMA beamformers. Lemma 1 is a correct algebraic projection. The Monte Carlo setup is standard and the comparisons against the ADMM-SCA baseline are fair.\n\nThe soft spot is load-bearing. The stress-test note is right: LCEH with center (0,0.5) is exactly the elementwise Euclidean projection onto the Lorentzian circle Q. ARLCH solves min(Φ,D) ||q* - D(j+e^{jΦ})/2||, but then discards D and outputs (j+e^{jΦ})/2, which lies on Q. That output cannot have smaller Euclidean distance to q* than LCEH's projection. So the paper's claim that ARLCH 'minimizes the discrepancy' between ideal and feasible weights is false. The reported power gains must come from some other mechanism — perhaps the phase profile ARLCH picks happens to help after digital precoder re-optimization — but the paper neither derives nor tests that. The 20% headline therefore has no stated theoretical support, and the 'optimal diameter' language is misleading: D* is optimal only for the scaled subproblem, not for the final beamforming objective.\n\nOther issues are minor or fixable: mapping definition (22) is ambiguous about which intersection point to pick, (20) has a typo, SDR rank-one recovery is asserted without proof, and Algorithm 1 has no convergence guarantee. No code or data is released, so the exact numbers are not independently checkable.\n\nBottom line: the LCEH-vs-LCUSH comparison deserves to be in the literature, and ARLCH may be a genuinely useful heuristic — but its stated rationale is wrong, and the paper needs major revision before the ARLCH claims can be trusted. This is exactly the kind of paper that should go to peer review, not be desk-rejected, because the empirical question is interesting and the flaw is addressable.","headline":"ARLCH's stated rationale is mathematically wrong — LCEH already achieves the projection minimum — but the empirical comparison of mapping centers is solid and worth referee time.","tokens_in":23750,"tokens_out":2824,"would_cite":true,"duration_ms":31010,"reading_group":"yes","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 claims that the largest beamforming loss in DMA holographic systems comes from projecting ideal weights onto the Lorentzian circle, and that adaptively relaxing that circle's radius cuts transmit power by over 20% in multiuser…","keywords":["dynamic metasurface antennas","holographic beamforming","Lorentzian-constrained holography","transmit power minimization","semidefinite relaxation","alternating optimization","multi-user MISO","near-field beamforming"],"falsifier":"For many random user placements, compute both the projection error that ARLCH minimizes and the actual transmitted power after the digital precoder is re-optimized; a single realization in which ARLCH's smaller projection error leads to higher transmit power than LCEH would falsify the proxy assumption that carries the claimed 20% gain.","tokens_in":22720,"feed_emoji":"📡","tokens_out":9590,"duration_ms":88603,"temperature":0.7,"pith_summary":"Dynamic metasurface antennas (DMAs) form beams with elements whose amplitude and phase cannot be set independently; they must lie on the Lorentzian circle $q=(j+e^{j\\Phi})/2$. This paper tries to establish that the way idealized beamforming weights are projected onto that circle is a major, under-appreciated source of power inefficiency, and that letting the circle's diameter adapt during the projection recovers much of that loss. It first unifies the three existing projection schemes, LCPH, LCEH, and LCUSH, into one framework called GMLCH, then introduces ARLCH, which optimizes the circle's diameter together with the element phases by alternating a closed-form diameter update with per-element phase searches. In simulated multiuser downlink networks with SINR guarantees, ARLCH reduces transmit power by over 20% relative to conventional fixed-circle benchmarks, and the gain grows with the number of users: 16.7% over LCEH at six users and 29.1% at eight users. If this holds, DMA arrays can approach fully digital power efficiency without adding radio-frequency chains.","feed_headline":"Adaptive projection radius cuts beamforming power by over 20%","feed_subtitle":"By letting the holographic weight radius adapt, the method beats fixed mapping schemes and scales to denser networks.","key_machinery":"The central object is the Lorentzian circle $Q=\\{(j+e^{j\\Phi})/2 : \\Phi\\in[0,2\\pi]\\}$, the set of physically realizable DMA weights, where amplitude and phase are coupled through a sinusoidal amplitude profile. GMLCH is the projection operator that maps ideal unit-modulus weights onto this circle by intersecting lines drawn from a chosen center $(x_c,y_c)$ with the circle, so LCPH, LCEH, and LCUSH become three special cases of one one-dimensional minimization. ARLCH replaces the fixed circle with a diameter $D$ that is itself optimized: for fixed phases, Lemma 1 gives the closed-form optimal diameter $D^\\star=\\mathrm{Re}(\\hat{q}^H q^\\star)/(\\hat{q}^H \\hat{q})$, then per-element phase searches find new intersection points, and the result is normalized back to the unit Lorentzian circle. This adaptive diameter is the mechanism that lets the mapping center move during optimization and reduces the projection mismatch that costs transmit power.","core_discovery":"On the paper's own terms, the central discovery is that the Lorentzian constraint should be treated as an optimizable projection target rather than a fixed hardware boundary. The authors formulate the DMA weight design as a relaxed SDP, obtain ideal unconstrained weights $q^\\star$, and then show that every standard Lorentzian mapping, LCPH, LCEH, and LCUSH, is a special case of GMLCH, a line-intersection projection parameterized by a mapping center $(x_c,y_c)$. Their proposed ARLCH goes further: it relaxes the Lorentzian circle's diameter $D$, solves $\\min_{\\Phi,D}\\|q^\\star - D(j+e^{j\\Phi})/2\\|^2$ by alternating the closed-form optimal diameter $D^\\star=\\mathrm{Re}(\\hat{q}^H q^\\star)/(\\hat{q}^H \\hat{q})$ from Lemma 1 with per-element phase searches, and normalizes the result back onto the unit circle. The paper claims this extra degree of freedom lowers transmit power under per-user SINR constraints, with the advantage over fixed mappings growing as user count increases; it also reports that among fixed mappings, LCEH consistently beats LCUSH, the most common choice in prior work.","pith_inferences":["The authors leave implicit that ARLCH's projection step is portable: any DMA pipeline that projects ideal weights onto the Lorentzian circle, whether manifold-based or codebook-based, could adopt the adaptive diameter without changing its outer optimizer.","A natural untested extension is per-element or per-microstrip diameter adaptation, since Lemma 1 optimizes a single global $D$; directionally clustered users might benefit from a spatially varying radius.","The paper does not prove that smaller projection error implies lower transmit power; a formal bound connecting the ARLCH objective (25) to the power objective (7) would turn the empirical 20% gain into a guarantee."],"forward_implications":["Among fixed mappings, LCEH should replace LCUSH as the default Lorentzian projection in DMA beamforming: the paper reports 13.9% lower power than LCUSH at one user rising to 17% at eight users.","ARLCH's advantage grows with network density: 16.7% lower power than LCEH at six users and 29.1% at eight users, so the method is most useful in crowded cells.","The gains persist at subwavelength antenna spacings, about 15% at $d_x=\\lambda$ rising to nearly 20% at $d_x=\\lambda/6$, meaning the method suits dense aperture designs.","The unified GMLCH framework reproduces the ADMM-SCA baseline when configured as LCUSH, giving future work a single SDP-based platform to test new projection schemes against existing ones."],"supporting_citations":[{"why":"Defines the Lorentzian resonance response $q=(j+e^{j\\Phi})/2$ that is the paper's core constraint.","marker":"[10]"},{"why":"Introduces the GMLCH mapping from unit-modulus weights to the Lorentzian circle, which the paper generalizes and benchmarks.","marker":"[26]"},{"why":"Supplies the ADMM-SCA baseline method that the paper re-implements and compares against for LCUSH.","marker":"[25]"},{"why":"Provides the DMA block-diagonal weight model, near-field spherical-wave channel, and the vec identity used in the DMA-weight SDP.","marker":"[18]"},{"why":"Establishes rank-one optimality of the SDR solution for the digital precoder subproblem.","marker":"[31]"},{"why":"Supplies the rank-one approximation and interior-point complexity results used to recover weights from the SDP.","marker":"[33]"},{"why":"Introduces DMA-assisted MIMO with Lorentzian-constrained holographic beamforming as the baseline approach the paper extends.","marker":"[11]"}],"fun_headline_variants":["Adaptive holographic weights slash beamforming power 20%+","ARLCH: smarter Lorentzian mapping cuts network power 20%","Beamforming power down 20% via adaptive Lorentzian radius","Dynamic metasurfaces: adaptive radius beats fixed holography","Lorentzian radius adaptation trims transmit power 20%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that making the analog weights closely match the ideal unconstrained weights is a reliable stand-in for reducing transmitted power; if that proxy fails for some channel layouts, the claimed savings disappear.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive holographic weights slash beamforming power 20%+","ARLCH: smarter Lorentzian mapping cuts network power 20%","Beamforming power down 20% via adaptive Lorentzian radius","Dynamic metasurfaces: adaptive radius beats fixed holography","Lorentzian radius adaptation trims transmit power 20%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1449,"prompt_tokens":1068,"completion_tokens":381,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":684,"completion_tokens_details":{"reasoning_tokens":290}},"tokens_in":684,"tokens_out":381,"duration_ms":3925,"temperature":1.0,"reasoning_tokens":290,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:57:43.009075+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For many random user placements, compute both the projection error that ARLCH minimizes and the actual transmitted power after the digital precoder is re-optimized; a single realization in which ARLCH's smaller projection error leads to higher transmit power than LCEH would falsify the proxy assumption that carries the claimed 20% gain.","supporting_citations":[{"cited_title":"Analysis of a waveguide-fed metasurface antenna,","cited_arxiv_id":null,"evidence_quote":"Defines the Lorentzian resonance response $q=(j+e^{j\\Phi})/2$ that is the paper's core constraint."},{"cited_title":"Near- field beamforming optimization for holographic XL-MIMO multiuser systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the ADMM-SCA baseline method that the paper re-implements and compares against for LCUSH."}],"review_version":1}