{"id":"8d1bb233-dc91-4f69-b6c0-759901dba059","arxiv_id":"2504.15737","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"For SIM-enhanced multiuser MISO downlinks, a hybrid digital and wave precoding framework with a realistic power model shows that 2 to 5 SIM layers balances energy and spectral efficiency, and can improve energy efficiency by 80% over digital-only precoding.","lead":"This paper designs a hybrid precoding framework for a base station with a stacked intelligent metasurface (SIM), maximizing energy efficiency in a multi-user downlink system. It proposes alternating optimization with SDP and gradient-based beamforming, and suggests 2 to 5 SIM layers as a good compromise between spectral and energy efficiency.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The BS precoding subproblem in Eq. (29a) does not match the stated EE objective in (20a): the square root is applied to each rate term rather than to the sum, so the algorithm may not maximize the defined EE.","rationale":"The reader's weakest assumption is the unmeasured power constants P_meta and P_controller, which affect the quantitative 2–5 layer recommendation. That is a valid concern, and the authors' own Footnote 5 acknowledges the lack of experimental data. However, I identified a more fundamental, internal inconsistency: the convex subproblem in Eq. (29a) optimizes a different objective than the EE defined in Eq. (15) and Eq. (20a). If the implementation follows the displayed equation, the proposed algorithm does not actually maximize the claimed metric; if it follows the intended expression, the manuscript has a serious typo that must be corrected and the results re-verified. Either way, the central claim about maximizing EE and the layer guidance derived from Figs. 8–9 is currently unsupported. The power-constant issue would only shift the layer count within a correct algorithm, whereas the objective mismatch undermines the algorithmic contribution itself. I therefore disagree with the reader's choice of weakest assumption, although the final verdict remains CONDITIONAL pending a fix and rerun. The paper otherwise applies standard tools (AO, SCA, SDP, PGA) and the channel model and power model are clearly described; these are positive features, but they do not resolve the objective inconsistency.","tokens_in":22871,"tokens_out":13922,"duration_ms":120214,"concrete_test":"Re-derive the SCA step for the objective 2t(Σ log2(1+γ_k))^{1/2} using a valid concave lower bound (e.g., first-order Taylor of the square-root function) and replace (29a) with the corrected surrogate. Rerun the BS precoding optimization for the settings of Figs. 3 and 8, keeping all other parameters unchanged, and compare the converged EE and the optimal number of SIM layers. If the EE at any operating point changes by more than a few percent, or if the 2–5 layer range shifts, the reported results depend on the erroneous objective.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-C introduces the SCA subproblem for BS precoding. The quadratic-transform objective in (20a) is 2t (Σ_k log2(1+γ_k))^{1/2} - t^2 P_total. The displayed convex subproblem in (29a) instead uses 2t Σ_k (log2(1+γ_k))^{1/2} - t^2 P_total. These are not equivalent; for nonnegative rates, √(Σ R_k) ≤ Σ √R_k, so (29a) is an upper bound, not a lower-bound surrogate. Maximizing (29a) therefore does not maximize the EE in (15). Since the SCA step is the core of the BS digital precoding update, the converged precoders and all subsequent SIM-layer sweeps (Figs. 8–9) are computed for a different objective unless the implementation silently corrects the formula. The paper provides no proof that the SCA updates monotonically improve the true EE, nor a rank-one verification for the SDP relaxation. Footnote 5 admits the power constants are unmeasured, but even if those were measured, the algorithm itself is currently not shown to optimize the claimed metric. The 2–5 layer guidance and the 80% EE improvement are therefore not justified as stated.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies the energy efficiency (EE) of a stacked-intelligent-metasurface (SIM)-assisted multi-user MISO downlink with hybrid digital precoding at the base station and wave-based beamforming across multiple SIM layers. The authors present a power consumption model for the SIM-assisted BS (Eqs. (6)-(12)), formulate the joint digital precoding and SIM phase-shift design as a non-convex EE maximization problem (P0 in Eq. (16)), and solve it via an alternating optimization framework: a quadratic transformation for the fractional objective, an SCA-based update for the BS precoding (Section III-C), and both SDP-based and projected-gradient-ascent (PGA) algorithms for the SIM phases (Section III-D). Extensive simulations evaluate EE and spectral efficiency versus the number of antennas, meta-atoms, layers, and power parameters. The central conclusions are that the proposed hybrid precoding improves EE by up to 80% over digital precoding without SIM beamforming, and that 2 to 5 SIM layers provide a good compromise between EE and SE.","tokens_in":23121,"tokens_out":21971,"duration_ms":184740,"significance":"If the claims hold, the paper provides a useful design rule for SIM hardware sizing and a joint precoding algorithm with an explicit power model, going beyond the SE-only studies in the SIM literature. The numerical study is broad (Figs. 2-10), the complexity analysis is included, and the authors are candid about the lack of experimental grounding for the SIM power constants (footnote 5). The layer-count guidance (2-5 layers) is a falsifiable, practically relevant claim that would interest system designers. However, the significance is contingent on (i) correcting the mismatch between the stated EE objective and the actual SCA subproblem solved at the base station, and (ii) demonstrating the robustness of the layer-count recommendation to the unmeasured power constants.","major_comments":[{"comment":"The SCA subproblem in Eq. (29a) is inconsistent with the objective defined in Eq. (17) and used in Eq. (20a). The quadratic-transform objective is f = 2t (sum_k log2(1+gamma_k))^{1/2} - t^2 P_total, with the square root applied to the sum of rates. In (29a), the objective is instead written as 2t sum_k (log2(1+gamma_k))^{1/2} - t^2 P_total, with the square root applied to each rate before summation. Since sqrt(sum_k R_k) <= sum_k sqrt(R_k) for nonnegative R_k, the surrogate in (29a) is an upper bound on the true objective, not a lower-bound surrogate of the kind required for a monotone SCA/MM procedure. Consequently, the paper provides no proof that the iterates of (29a) monotonically increase the EE defined in (15) or converge to a stationary point of P0; the BS precoding block is the first block of the AO loop, so Figures 2-9 are generated by an algorithm that optimizes a different objective than the one stated. The claims that the proposed framework maximizes EE and achieves an 80% EE improvement are therefore not justified as written. The authors should either correct (29a) to keep the outer square root (the resulting objective in gamma remains concave), provide a valid lower-bound surrogate with an ascent proof, or re-derive the SCA around the true objective and re-run the simulations.","section":"Section III-C, Eq. (29a) vs. Eqs. (17) and (20a)"},{"comment":"The headline design rule of the paper, that 2 to 5 SIM layers yield satisfactory EE and SE, is a numerical outcome of the power model in Eqs. (11)-(12) with P_meta = 10 dBm and P_controller = 25 dBm. Footnote 5 states that no large-scale experimental data quantify SIM controller power consumption, and P_meta is likewise a chosen value. Because the marginal cost of adding a layer is N*P_meta, the location of the EE peak in Fig. 8 is directly controlled by these unmeasured constants. Fig. 10 sweeps P_meta for a fixed M=4, but no sensitivity of the optimal number of layers to P_meta or P_controller is reported, and the abstract and conclusion state the 2-5 layer guidance without this caveat. The claim in footnote 5 that P_controller 'does not affect the overall trends and conclusions' is also not generally true: as an additive constant in the denominator of (15), P_controller changes the ratio R(M)/P(M) across M and can shift the argmax, so this claim should either be proven or removed. I recommend re-scoping the conclusions or adding a systematic sensitivity analysis (e.g., a two-parameter sweep over P_meta and P_controller) to establish the robustness of the 2-5 layer recommendation.","section":"Table I, Eqs. (11)-(12), footnote 5, Figs. 8-10"}],"minor_comments":[{"comment":"In Eq. (44), chi_k is defined with a summation over all j from 1 to K, whereas Eq. (52) defines zeta_k (of which chi_k is the reciprocal, per the proof in (51)-(53)) with a summation over j != k. The summation in (44) should exclude j = k; as displayed, the two expressions are inconsistent and the quotient-rule derivation does not follow.","section":"Eq. (44) vs. Eq. (52)"},{"comment":"The gradient formula in Eq. (45) and its derivation in Eq. (54) are mutually inconsistent: (45) uses 2 Im[(e^{jφ} h^H b (q)^H W1 p_k)(h^H G W1 p_k)], while (54) yields Im[(e^{jφ} h^H b (q)^H W1 p_j)^H (h^H G W1 p_j)] without the factor of 2. The correct derivative of |f(φ)|^2 with respect to φ for f = e^{jφ} a + c is 2 Im[A^* f] with A = e^{jφ} a; the displayed (45) lacks the conjugation and (54) lacks the factor of 2. Since (48) updates the phases using this gradient, the formulas should be corrected and the PGA implementation checked against them.","section":"Theorem 1, Eqs. (45) and (54)"},{"comment":"The outer stopping criterion reads |Opt(r) − Opt(r)|/Opt(r) ≤ ε, which compares Opt(r) with itself; it should presumably read |Opt(r) − Opt(r−1)|/Opt(r) ≤ ε.","section":"Algorithm 1, line 21"},{"comment":"Step 10 says 'Compute {p_k} by solving the problem in (18)', but the BS precoding subproblem solved at that point is (29), not (18).","section":"Algorithm 1, step 10"},{"comment":"The sentence 'the problem in (42) is not convex' appears before problem (42) is defined and evidently refers to problem (35); also, 'the coupling of variables coupling' contains a duplicated word. In the same paragraph, 'the problem P_SDP in (42)' should refer to the convex approximation \\hat{P}_SDP.","section":"Section III-D1, text before Eq. (42)"},{"comment":"The smoothing parameter ε in the indicator approximation (27) is not reported in Table I or in the simulation setup. Since Remark 3 states the activation threshold is x ≥ 10^{-4}, the authors should specify the ε used so that the simulation is reproducible.","section":"Remark 3 and Eq. (27)"},{"comment":"The rank-one relaxation via (40)-(41) is claimed to 'ultimately arriv[e] at a rank-one solution' as ε_m increases to 1, but no numerical evidence for the achieved rank of V_m (for example, the λ_max/Tr(V_m) ratio or the final ε_m) is reported; without such evidence, the SDP results cannot be verified as solutions of the rank-constrained problem.","section":"Eqs. (35)-(42), rank-one relaxation"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and the contributions are potentially useful, but the two major comments are load-bearing: the displayed SCA subproblem optimizes a different objective than the stated EE, and the central layer-count recommendation is tied to unmeasured power constants. I would ask the editor to require that the simulations be re-run with a corrected objective or that a rigorous convergence proof for the surrogate be provided, since the numerical values in Figs. 2-9 depend on the actual objective being optimized. The novelty relative to Ref. [38], which already addresses EE in SIM-based broadcast MIMO, is moderate; the additional elements here are the joint digital/wave precoding with user fairness, antenna selection, and the layer-count study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent engineering paper with a real optimization gap. The SIM-layer guidance and the 80% EE gain are conditional on unmeasured power constants and, more importantly, on an SCA subproblem whose displayed objective does not match the stated EE.\n\nWhat is new: the paper adds user fairness (QoS constraints) and RF-chain/antenna selection to EE maximization in SIM-assisted multiuser MISO, and it gives two SIM beamforming algorithms, SDP and PGA, with complexity analysis. The gradient derivation in Theorem 1 is standard but appears correct, and the simulation campaign is broad: antenna count, meta-atoms, Pmax, QoS, and layer sweeps. Against the prior SIM-EE work ([38]), this is a real extension.\n\nWhere it gets soft. First, the objective mismatch. Eq (20a) uses 2t (Σ_k log2(1+γ_k))^{1/2}, but the convex subproblem (29a) maximizes 2t Σ_k (log2(1+γ_k))^{1/2}. These are not equivalent; since √ΣR ≤ Σ√R, (29a) maximizes an upper bound, not a lower-bound surrogate, and no proof is given that iterates improve the original EE. If the implementation silently keeps the square root outside the sum, the paper needs to say so; as written, the algorithmic claims and everything downstream (Figs. 8–9, the 80% number) are not justified. This is the kind of thing a referee should catch, and it should be fixed before publication.\n\nSecond, the layer-count guidance rests on hand-picked P_meta and P_controller values. Footnote 5 admits the controller number is unmeasured, and there is no sensitivity sweep for the optimal layer count as P_meta varies. Fig. 10 shows meta-atom power matters a lot, which strengthens the worry. The 2–5 layer rule might survive a sensitivity check, but the paper doesn't show it.\n\nThird, smaller stuff: no code/data, no error bars on the improvement percentages, and the SDP rank-one recovery is described only as increasing ε_m from 0 to 1—the actual extraction of ϕ_m from V_m is not specified.\n\nVerdict: conditional, leaning positive. The core modeling and solution framework are sensible, and the objective mismatch is probably a bracket typo rather than a flaw in the overall approach, but the paper as written does not establish its headline claims. Who it's for: researchers working on SIM and holographic MIMO EE who want a benchmark framework and a starting point for layer sizing. It deserves a serious referee; the fixes are within reach.","headline":"Competent SIM energy-efficiency study with a genuine SCA objective mismatch and an unmeasured-power dependence behind the 2–5 layer rule; deserves review but not acceptance as is.","tokens_in":23679,"tokens_out":2946,"would_cite":false,"duration_ms":28218,"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 hybrid digital and wave-based precoding maximizes energy efficiency in SIM-assisted multiuser MISO downlinks, and that 2-5 SIM layers give near-optimal performance for both energy and spectral efficiency.","keywords":["stacked intelligent metasurface","energy efficiency","hybrid precoding","wave-based beamforming","multi-user MISO","alternating optimization","semidefinite programming","projected gradient ascent"],"falsifier":"Measure the actual power consumption of a fabricated SIM's meta-atoms and controller, then recompute the energy-efficiency versus layer-count curves in the paper with those measured constants; if the measured $P_{\\rm meta}$ and $P_{\\rm controller}$ depart substantially from 10 dBm and 25 dBm, the 2-5 layer recommendation should move accordingly.","tokens_in":22645,"feed_emoji":"🔋","tokens_out":7771,"duration_ms":68673,"temperature":0.7,"pith_summary":"This paper sets out to establish that a stacked intelligent metasurface (SIM) mounted at a multi-antenna base station can be jointly optimized with digital precoding to maximize the downlink energy efficiency of a multiuser MISO system, while respecting per-user rate and per-antenna power limits. The problem is a non-convex fractional program, and the paper solves it with an alternating optimization loop: a quadratic transform handles the ratio objective, successive convex approximation updates the digital precoders, and either semidefinite programming or projected gradient ascent tunes the SIM phase shifts. The quantitative claim is that this hybrid design achieves up to 80% higher energy efficiency than digital precoding without SIM beamforming. The practical sizing claim is that although the layer count that maximizes spectral efficiency differs from the one that maximizes energy efficiency, 2 to 5 SIM layers give satisfactory performance on both metrics.","feed_headline":"2-5 SIM layers deliver the best energy efficiency","feed_subtitle":"Joint digital and wave beamforming beats digital-only by up to 80% in simulations.","key_machinery":"The load-bearing object is the wave-based beamforming matrix $G = \\Phi_M W_M \\cdots \\Phi_1$, built from $M$ stacked metasurface layers, each a diagonal phase-shift matrix $\\Phi_m$, separated by diffraction-based transmission matrices $W_m$. This matrix is what turns the SIM into an analog beamformer that can orthogonalize user channels before the digital precoder acts. The argument is carried by the alternating optimization loop: the fractional energy-efficiency objective is reformulated with a quadratic transform, the digital precoders are updated by successive convex approximation with an approximate indicator function for RF-chain activation, and the SIM phase shifts are updated either by semidefinite programming or by projected gradient ascent using the closed-form gradient of the sum rate with respect to each phase shift.","core_discovery":"The central claim is that the best energy-efficient transmitter for a SIM-assisted base station is a hybrid of digital precoding and wave-based beamforming, not either mechanism alone. The SIM's stacked phase-shift layers reshape the effective channel, creating extra optimization space, while the digital precoder controls multiuser interference and selects which RF chains and antennas stay active. Under the assumed power model, energy efficiency is not monotone in hardware scale: adding layers or meta-atoms first improves the achievable sum rate and then hurts efficiency because every added meta-atom contributes a term $N M P_{\\rm meta}$ to the power budget. The numerical conclusion is that 2-5 layers is a robust compromise, with the proposed hybrid design beating digital-only precoding by up to 80% in energy efficiency.","pith_inferences":["If hardware measurements show per-meta-atom power far below 10 dBm, the optimal layer count will likely move above 5; if the controller or meta-atom power is higher than assumed, fewer layers will look better.","The same alternating structure should transfer to near-field, wideband, and cell-free massive MIMO settings, where the channel model changes but the tradeoff between wave-domain gains and fixed SIM power remains.","A prototype measurement campaign that records actual power draw per meta-atom and per controller, together with achieved rates, is the direct test of whether 2-5 layers is the right sizing rule.","The layer-count result also suggests that SIM hardware designers should aim to make per-meta-atom power scale down with array size, since the fixed $N M P_{\\rm meta}$ term is precisely what caps the useful number of layers."],"forward_implications":["A base station that adds a SIM and runs the proposed hybrid design can expect up to an 80% energy-efficiency gain over the same station using only digital precoding, under the assumed channel and power model.","Energy efficiency peaks at an intermediate number of layers and then falls, so designers should size the SIM stack rather than simply adding layers; the paper's recommended range is 2-5 layers.","Using the SDP-based SIM update buys about a 7% energy-efficiency gain over the projected-gradient update, at substantially higher computational cost, so the choice between the two is a quantified accuracy-complexity tradeoff.","More meta-atoms per layer can substitute for more base station antennas, giving a hardware design knob that trades array size against SIM size.","Joint RF-chain and antenna selection keeps energy efficiency rising as antennas are added, whereas without selection it would peak near L=8 and decline."],"supporting_citations":[{"why":"Defines the wave-based beamforming matrix and the diffraction-based inter-layer transmission model that the system model is built on.","marker":"[39]"},{"why":"Introduces the stacked intelligent metasurface MIMO transceiver concept that this paper applies to a multiuser downlink with hybrid precoding.","marker":"[23]"},{"why":"Provides the closest prior energy-efficiency design for SIM-based broadcast MIMO, giving the baseline that this work extends with fairness constraints and a two-algorithm approach.","marker":"[38]"},{"why":"Supplies the BS power-consumption model with active RF chains, antenna selection, and PA efficiency used in the energy-efficiency objective.","marker":"[45]"},{"why":"Supplies the quadratic transform that converts the fractional energy-efficiency objective into the tractable form the alternating optimization solves.","marker":"[46]"},{"why":"Supplies the SINR constraint transformation and successive convex approximation used in the digital precoding update.","marker":"[47]"},{"why":"Cited as the basis for treating SIM as a semi-passive low-power structure in the power-consumption model.","marker":"[20]"}],"fun_headline_variants":["Hybrid beamforming boosts SIM energy efficiency up to 80%","Optimal SIM layers: 2 to 5 for best energy efficiency","Hybrid precoding beats digital-only by 80% in SIM efficiency","2-5 SIM layers: the sweet spot for energy-efficient hybrid precoding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the SIM's fixed power costs are 10 dBm per meta-atom and 25 dBm for the controller; the paper itself notes that no large-scale experimental data yet pins these values down, and changing them shifts the optimal layer count.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid beamforming boosts SIM energy efficiency up to 80%","Optimal SIM layers: 2 to 5 for best energy efficiency","Hybrid precoding beats digital-only by 80% in SIM efficiency","2-5 SIM layers: the sweet spot for energy-efficient hybrid precoding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001051,"raw_usage":{"total_tokens":4416,"prompt_tokens":947,"completion_tokens":3469,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":3390}},"tokens_in":563,"tokens_out":3469,"duration_ms":20933,"temperature":1.0,"reasoning_tokens":3390,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:18:47.001047+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the actual power consumption of a fabricated SIM's meta-atoms and controller, then recompute the energy-efficiency versus layer-count curves in the paper with those measured constants; if the measured $P_{\\rm meta}$ and $P_{\\rm controller}$ depart substantially from 10 dBm and 25 dBm, the 2-5 layer recommendation should move accordingly.","supporting_citations":[{"cited_title":"Sta cked intelligent metasurfaces for multiuser downlink beamform ing in the wave domain,","cited_arxiv_id":null,"evidence_quote":"Defines the wave-based beamforming matrix and the diffraction-based inter-layer transmission model that the system model is built on."},{"cited_title":"Stacked intelligent metasurfaces for efficie nt holo- graphic MIMO communications in 6G,","cited_arxiv_id":null,"evidence_quote":"Introduces the stacked intelligent metasurface MIMO transceiver concept that this paper applies to a multiuser downlink with hybrid precoding."},{"cited_title":"Energy-efficient multicell multigroup multica sting with joint beamforming and antenna selection,","cited_arxiv_id":null,"evidence_quote":"Supplies the BS power-consumption model with active RF chains, antenna selection, and PA efficiency used in the energy-efficiency objective."},{"cited_title":"Fractional programming for communic ation systems—part i: Power control and beamforming,","cited_arxiv_id":null,"evidence_quote":"Supplies the quadratic transform that converts the fractional energy-efficiency objective into the tractable form the alternating optimization solves."},{"cited_title":"A programmable diffractive deep 14 neural network based on a digital-coding metasurface array ,","cited_arxiv_id":null,"evidence_quote":"Cited as the basis for treating SIM as a semi-passive low-power structure in the power-consumption model."}],"review_version":1}