{"id":"f4001ef3-7b24-4724-a0d2-ae377f2d3fb8","arxiv_id":"2508.15257","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Optimizing SIM phase shifts before the digital beamformer, with an inner iterative projected-gradient loop, achieves up to 115.53% higher achievable sum rate than benchmark AO schemes in simulation.","lead":"This letter tests two tweaks to the standard alternating-optimization (AO) recipe for beamforming in SIM-aided multiuser systems: update the metasurface phase shifts before the digital precoder, and run the projected-gradient (PG) inner loop to convergence. In simulations, the tweaks are reported to raise achievable sum rate by up to 115.53% and to avoid the saturation seen in prior AO schemes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains rest on a single random initialization/channel draw with no error bars, so the ordering and 'no saturation' claims are not yet established.","rationale":"In good faith, the paper is a plausible engineering letter: the algorithm is clearly described, the AO ordering effect is a reasonable observation for non-convex problems, and the simulation figures are internally consistent. The reader's weakest-assumption analysis correctly identifies the lack of statistical replication and single-initialization dependence as the principal threat to the empirical central claim. I do not see a fatal logical error in the design, and the 'no saturation' statement is explicitly hedged to 'the considered setting.' The most load-bearing concern is therefore not the model choice per se but the absence of evidence that the observed ranking is robust across initial points and channel realizations. The additional gradient-formula index mismatch in Eq. (9)/(11) is a concrete internal inconsistency that should be fixed, but it is likely a typo given the derivation in the appendix; if corrected, it probably would not change the qualitative conclusions. Thus the appropriate verdict remains CONDITIONAL, requiring the authors to provide error bars/averaging, initialization sensitivity, and a corrected gradient formula before the headline percentages are accepted as general findings.","tokens_in":9481,"tokens_out":9764,"duration_ms":121323,"concrete_test":"Run the identical SIMwDB setup of Fig. 2(a) for at least 30 random seeds (independent UE drops plus random θ(0), W(0)), and report mean/median with min–max or 10–90% bands for all four variants at each L and at N=100. If the iterative PG θ-first curve is not above the others for nearly all seeds, or if its advantage shrinks dramatically, the claimed design guideline and the 115.53% gain are not robust. Separately, re-derive Eq. (9) from the log objective to check that the second denominator should be j≠k; if it should be, the published formula needs correction and the simulations should be regenerated with the corrected gradient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claims—θ-first ordering, iterative-PG gains up to 115.53%, and saturation avoidance—are purely empirical, but Section IV does not state how many random UE drops/channel realizations or initial points are used. Fig. 1 is explicitly 'starting from the same randomly generated initial point' (singular), and Figs. 2–3 show no error bars or trial counts. Because (P1) is non-convex and the paper itself notes AO is sensitive to initialization, a single trajectory does not establish that θ-first is generally better or that the proposed method avoids saturation; the observed ranking and percentage gains may be specific to one favorable initial point/realization. This is the load-bearing gap for the headline claim. Additionally, the denominator in the second term of Eq. (9)/(11) appears to sum over j≠1 instead of j≠k; if this is not merely a typo, the PG update as printed is not the gradient of the stated objective, which would further undermine reproducibility.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers sum-rate maximization for a stacked intelligent metasurface (SIM) aided multiuser MISO downlink, optimizing digital beamforming and SIM phase shifts alternately. The proposed method differs from prior AO schemes in two ways: it optimizes the SIM phase shifts before the digital beamformer, and it uses an iterative projected-gradient (PG) procedure for the phase-shift subproblem instead of a single PG step. The main theoretical contribution is a closed-form expression for the gradient of the sum rate with respect to the phase-shift vector (Theorem 1). The performance claims—that the theta-first ordering and iterative PG yield higher achievable sum rate and avoid saturation as the number of SIM layers grows—are supported only by simulation results for a Rayleigh-Sommerfeld propagation model and correlated Rayleigh user channels. The headline gain is stated as up to 115.53% over benchmark schemes.","tokens_in":9715,"tokens_out":4788,"duration_ms":59078,"significance":"If the empirical claims are robust, the paper offers a simple and practically relevant design guideline for SIM-aided systems: optimize phase shifts first and iterate PG to convergence. The derivation of the gradient is direct and appears correct apart from a likely typographical error in the denominators of Eqs. (9) and (11), and the simulation setup is clearly described, which aids reproducibility. However, the central claims are empirical ordering statements over non-convex problems, and the current evidence is not statistically grounded: Fig. 1 uses a single initial point, Figs. 2–3 show no error bars or trial counts, and the 'no saturation' conclusion is extrapolated from L up to 10. The comparison between iterative PG and single PG also conflates algorithmic refinement with additional computational effort. With additional Monte Carlo results and a corrected gradient expression, the contribution would be a valid and useful engineering insight.","major_comments":[{"comment":"The central claims are empirical: the theta-first ordering, the advantage of iterative PG, and the percentage gains in Fig. 3 all depend on random initializations and channel realizations. Fig. 1 is explicitly described as starting 'from the same randomly generated initial point' (singular), and Figs. 2–3 do not report the number of Monte Carlo runs, the number of initial points, or error bars. Since (P1) is non-convex and the paper itself notes in Remark 2 that AO performance is sensitive to the initial point, the observed ordering and the 115.53% gain could be specific to a favorable draw. Please provide statistics over many random channel realizations and initial points (e.g., mean ± standard deviation or box plots) and state the trial count.","section":"Section IV, Figs. 1–3"},{"comment":"The denominator of the second term in the gradient expression is written as Σ_{j≠1} |(θ^l)^T e^l_{k,j}|^2 + σ^2 in both Eq. (9) and Eq. (11). By the derivation in the Appendix, this term should be the interference-plus-noise power for user k, i.e., Σ_{j≠k} |(θ^l)^T e^l_{k,j}|^2 + σ^2. As printed, the PG update is not the gradient of the stated sum-rate objective. If this is a typographical error, please correct it and confirm that the simulations used the correct formula; if it is not a typo, the algorithm description is internally inconsistent and the simulation results may not correspond to the claimed objective.","section":"Eqs. (9) and (11)"},{"comment":"The conclusion that 'saturation does not occur with the proposed method' is extrapolated from simulations with L only up to 10. A monotonically increasing curve over L=1,...,10 does not establish absence of saturation at larger L; many functions increase for small arguments and flatten later. To support this headline claim, either provide a scaling argument or simulate larger L (with computational cost reported) and show that the ASR continues to grow, for example by plotting the slope or relative increment versus L. As it stands, the evidence is too limited for a 'no saturation' claim.","section":"Section IV, Fig. 2 and Conclusion"},{"comment":"The comparison between 'iterative PG' and 'single PG' does not equalize computational effort. In the proposed method, PG is repeated until convergence (Lines 12–13 of Algorithm 1) before updating W, whereas the single-PG benchmarks perform only one PG step per outer iteration. The observed gain may therefore be due to more inner iterations rather than to the algorithmic 'refinement' per se. To support the design guideline that iterative PG is inherently better, compare the variants under a matched computational budget (e.g., equal total number of PG iterations, equal runtime, or ASR versus per-iteration cost) and report the additional complexity.","section":"Section IV, Algorithm 1"}],"minor_comments":[{"comment":"The notation 'Compute ∇θ(m)R(θ(m))' is awkward; it should be 'Compute ∇θ R(θ)|_{θ=θ(m)}' or simply 'Compute ∇θR(θ(m))'.","section":"Algorithm 1, line 4"},{"comment":"The caption '... (a). SIMwDB (b). SIMwoDB' is missing a separator verb; it should read '(a) SIMwDB, (b) SIMwoDB' for clarity.","section":"Fig. 2 caption"},{"comment":"When reporting percentage gains, please report the absolute ASR values and, if possible, confidence intervals; a percentage computed from a single point is sensitive to small variations.","section":"Section IV"},{"comment":"The paper states that the SCA method of [20] achieves the same performance as WMMSE. Since the proposed algorithm relies on this SCA solver, a brief description of its computational complexity or a reference to a version with more details would improve reproducibility.","section":"Section IV"},{"comment":"Minor typos and notation inconsistencies exist (e.g., 'P A' in Section I, and the use of 'θ(m)' as both a subscript and an argument in Algorithm 1). A careful proofread is recommended.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a communications letters venue, and the topic is timely. However, the claimed novelty—theta-first ordering and iterative PG—is modest, and the empirical evidence is not yet statistically convincing. If the authors add proper Monte Carlo statistics, correct the gradient formula, and address the computational-fairness issue, the paper could become acceptable. I would not recommend reject at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis letter is a useful empirical follow-up in the SIM-aided MU-MISO line: the authors show that optimizing the SIM phase shifts before the digital beamformer, and running the projected-gradient inner loop to convergence, gives substantially higher sum rate than the usual W-first, single-PG AO, and avoids the saturation with layer count reported in earlier SIM papers. The contribution is a design guideline, not a new technique — the authors admit as much — and the four-way comparison is clear, with a standard simulation setup for the subfield.\n\nWhat the paper does well: the ordering and inner-loop questions are explicitly isolated, the benchmarks are sensible, and the paper is honest that performance depends on the initial point. The closed-form gradient in Theorem 1 is a direct calculus derivation, and the appendix makes the chain of dependencies readable. The self-citations supply the SCA solver and prior SIM scaling results, which is fine.\n\nThe soft spots are mostly about statistical support. Fig. 1 shows a single random initial point; Figs. 2 and 3 have no error bars or trial counts. Given that the problem is non-convex and the authors themselves emphasize sensitivity to initialization, one trajectory does not establish that θ-first is generally better, nor that saturation is avoided. The 115.53% headline is measured at one configuration (N=100, L=10). The 'no saturation' conclusion is only checked up to L=10; phrasing it as 'no saturation in the considered range' would be more accurate. There's also a likely typo: the denominator in the second sum of eq. (9)/(11) runs over j≠1 instead of j≠k. If that's not a typo, the printed gradient is not the gradient of the stated objective; either way, it should be corrected.\n\nThese are addressable, not fatal. The paper would be much stronger with multi-initialization statistics, error bars, and a corrected formula. As it stands, it's a plausible engineering claim that merits a serious referee, though I wouldn't cite it in my own work until the empirical basis is firmer. If you work on SIM optimization or AO for phase-shift design, it's worth a look.","headline":"Useful design-guideline letter on SIM AO ordering, but the headline gains rest on a single simulation trajectory without error bars.","tokens_in":10215,"tokens_out":3449,"would_cite":false,"duration_ms":38852,"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":"Optimizing SIM phase shifts before the digital beamformer, with an iterated projected-gradient loop, raises the achievable sum rate by up to 115.53% over benchmark schemes and lets the rate keep growing with SIM layer count instead of satur","keywords":["stacked intelligent metasurface","alternating optimization","projected gradient","sum rate maximization","multiuser MISO","wave-domain beamforming","achievable sum rate","digital beamforming"],"falsifier":"Repeat the comparison of the four algorithm variants (Figs. 1–3) under a different inter-layer propagation law—for example, full-wave electromagnetic simulation of the metasurface stack or measurements from a SIM prototype—and across many random initial points and channel realizations. If the θ-first iterative-PG scheme no longer consistently outperforms the W-first single-PG scheme, or if its sum rate saturates as L grows under the alternative model, the paper's central claims are refuted.","tokens_in":9375,"feed_emoji":"📶","tokens_out":10644,"duration_ms":99379,"temperature":0.7,"pith_summary":"The paper studies how to maximize the total downlink data rate in a multiuser system whose base station transmits through a stacked intelligent metasurface (SIM)—several layers of tunable meta-atoms that shape the signal in the electromagnetic domain before it reaches single-antenna users. The standard solution method is alternating optimization (AO): update the digital beamformer, then update the SIM phase shifts, and repeat. This letter argues that the recipe has been applied badly in two small-looking ways, and that these details matter enormously. It shows numerically that updating the SIM phase shifts first, before the digital beamformer, and solving that update with an iterative projected-gradient loop instead of a single step, yields up to 115.53% higher achievable sum rate than the common baselines, and—unlike them—keeps growing with the number of SIM layers instead of saturating. If correct, the result means the saturation observed in prior SIM studies can be avoided by more careful algorithm design, at least within the simulated setting.","feed_headline":"Putting SIM phase shifts first lifts sum rates 115%","feed_subtitle":"A reordered alternating optimization—phase shifts first, iterated to convergence—keeps SIM gains from saturating.","key_machinery":"The load-bearing mechanism is the ordering and refinement inside an alternating optimization loop. Algorithm 1 cycles between an iterative projected-gradient ascent on the SIM phase-shift vector θ—using the closed-form complex gradient of the sum rate from Theorem 1, with backtracking line search and a stopping tolerance—and an SCA-based update of the digital beamforming matrix W. The phase-shift gradient is derived from the Rayleigh–Sommerfeld diffraction model of inter-layer propagation, which makes the entire wave-domain channel a structured function of θ, and the paper's two design rules—optimize θ first, and iterate the PG loop to convergence before returning to W—are what convert that","core_discovery":"The paper's discovery is that the performance of alternating optimization for SIM-aided systems hinges on two implementation details prior work treated as interchangeable. First, the update order is not neutral: optimizing the SIM phase-shift vector θ before the digital beamformer W converges to a significantly higher sum rate than the reverse order, because the SIM's wave-domain beamforming is the dominant source of interference suppression and benefits from early optimization. Second, the phase-shift subproblem, usually handled by a single projected-gradient step, is far better solved by an iterative projected-gradient loop that runs to near-convergence inside each AO cycle before switchin","pith_inferences":["An implicit corollary the paper does not develop: because the ordering advantage comes from θ being the higher-leverage variable, the θ-first principle plausibly transfers to other SIM optimization problems (energy efficiency, cell-free networks, joint sensing and communication), all of which currently use AO with single-step phase-shift updates.","The saturation-avoidance claim is demonstrated starting from a single randomly generated initial point; a natural extension is to test whether θ-first iterative PG preserves its advantage across many random initializations, since AO for non-convex problems is initial-point-sensitive and the paper explicitly identifies that sensitivity.","If the results generalize beyond the simulated Rayleigh–Sommerfeld/sinc-correlated model, comparative studies of SIM should include a well-tuned AO baseline; otherwise 'saturation' and other reported SIM limits may be algorithmic artifacts rather than physical ones.","A testable engineering prediction follows: the gap between θ-first iterative PG and the baselines should widen as the number of users K grows, because more users mean more interference for the SIM's wave-domain suppression to handle."],"forward_implications":["Any existing SIM-aided AO design can adopt the two guidelines as a drop-in change—move the phase-shift update to the front of each AO cycle and iterate the PG loop to convergence—without changing the channel model or the objective.","Within the considered setting, the previously reported sum-rate saturation with increasing SIM layer count (at fixed thickness) is avoided, so a correctly tuned AO can keep extracting gains from added layers.","The reported gains widen with the number of meta-atoms per layer (roughly 55–78% at N=49 up to 115.53% at N=100), so the benefit of the refined AO matters most in the large-SIM regime.","The same algorithm and ordering apply to SIM without digital beamforming (power-allocation variant), with comparable gains (up to 73.53% at L=10), making the design principle architecture-independent within the SIM family.","The ordering advantage is attributed to SIM's interference-suppression capability, which implies the phase-shift subproblem should be solved more thoroughly than the beamforming subproblem in any AO-based SIM design."],"supporting_citations":[{"why":"Supplies the Rayleigh–Sommerfeld diffraction law in eq. (1) that models every inter-layer channel matrix; the simulated SIM propagation used for all numerical results stands on it.","marker":"[18]"},{"why":"Establishes the sinc-correlated spatial correlation model for user channels and serves as a leading baseline (W-first, single PG) whose saturation behavior the proposed method is designed to beat.","marker":"[4]"},{"why":"Provides the low-complexity successive convex approximation (SCA) algorithm used for the digital beamforming and power-allocation subproblem in Algorithm 1.","marker":"[20]"},{"why":"Supplies the WMMSE method for the beamforming subproblem, cited to justify that the adopted SCA update matches WMMSE sum-rate performance at lower complexity.","marker":"[19]"},{"why":"A benchmark AO scheme for the same MU-MISO wave-domain beamforming setting (W-first, single PG) whose converged sum rate the proposed method is compared against.","marker":"[6]"},{"why":"A single-PG (θ-first) baseline the paper compares against, used to isolate the effect of replacing one PG step with an iterative PG loop.","marker":"[8]"}],"fun_headline_variants":["Put phase shifts first for 115% SIM rate boost","Iterate PG phase shifts to beat SIM saturation","Reorder AO: Phase shifts before beamforming wins","115% more sum rate: SIM optimization order matters","Iterative PG outperforms single-step in SIM AO"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The ranking of the four schemes—including the 115.53% gain and the absence of saturation—is established in simulation under a single propagation model (Rayleigh–Sommerfeld diffraction between layers and sinc-correlated Rayleigh fading to users), started from one randomly generated initial point; if real SIM propagation or the user-channel distribution differs from that simulator, the ordering of the methods could change.","fun_headline_variants_meta":{"raw":{"variants":["Put phase shifts first for 115% SIM rate boost","Iterate PG phase shifts to beat SIM saturation","Reorder AO: Phase shifts before beamforming wins","115% more sum rate: SIM optimization order matters","Iterative PG outperforms single-step in SIM AO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000144,"raw_usage":{"total_tokens":1040,"prompt_tokens":798,"completion_tokens":242,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":167}},"tokens_in":542,"tokens_out":242,"duration_ms":3135,"temperature":1.0,"reasoning_tokens":167,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:59:53.266334+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the comparison of the four algorithm variants (Figs. 1–3) under a different inter-layer propagation law—for example, full-wave electromagnetic simulation of the metasurface stack or measurements from a SIM prototype—and across many random initial points and channel realizations. If the θ-first iterative-PG scheme no longer consistently outperforms the W-first single-PG scheme, or if its sum rate saturates as L grows under the alternative model, the paper's central claims are refuted.","supporting_citations":[{"cited_title":"All-optical machine learning using diffractive deep neu ral networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the Rayleigh–Sommerfeld diffraction law in eq. (1) that models every inter-layer channel matrix; the simulated SIM propagation used for all numerical results stands on it."},{"cited_title":"On the Joint Beamforming Design for Large-scale Downlink RIS-assisted Multiuser MIMO Systems","cited_arxiv_id":"2412.08320","evidence_quote":"Provides the low-complexity successive convex approximation (SCA) algorithm used for the digital beamforming and power-allocation subproblem in Algorithm 1."},{"cited_title":"An iteratively weighted MMSE approach to distributed sum-utility maximization for a MIMO interfering broadcast channel,","cited_arxiv_id":null,"evidence_quote":"Supplies the WMMSE method for the beamforming subproblem, cited to justify that the adopted SCA update matches WMMSE sum-rate performance at lower complexity."},{"cited_title":"Stacked intelligent metasurfaces for multiuser downlin k beamforming in the wave domain,","cited_arxiv_id":null,"evidence_quote":"A benchmark AO scheme for the same MU-MISO wave-domain beamforming setting (W-first, single PG) whose converged sum rate the proposed method is compared against."},{"cited_title":"Achievable rate optimization for large stacked intelligent metasurfaces based on statistical CSI ,","cited_arxiv_id":null,"evidence_quote":"A single-PG (θ-first) baseline the paper compares against, used to isolate the effect of replacing one PG step with an iterative PG loop."}],"review_version":1}