{"id":"8102537d-70bf-4ce8-aa62-8545d4835b6a","arxiv_id":"2502.05819","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A stacked intelligent metasurface at the base station can approximate zero-forcing multiuser beamfocusing in the near field using passive wave propagation.","lead":"This paper shows that a stack of programmable radio-wave surfaces at a base station can shape signals for multiple users in the near field, replacing heavy digital hardware with passive analog computation. Simulations indicate the stack can match ideal beamforming and beat far-field setups by 7.6% in sum rate.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The downlink signal model uses Q = H^H G while Eq. (6) defines H as the uplink channel; physical reciprocity requires Q = H^T G, so the optimized SIM and all numerical results are computed for a non-physical channel.","rationale":"The reader's weakest assumption concerned the physical fidelity of the G factorization (no inter-layer coupling). That is an important external validation risk, but it is not the first thing that breaks. Before asking whether a real 12-layer SIM realizes the single-pass chain, one should ask whether the paper's own simulation represents the downlink of the described passive reciprocal system. It does not: H is explicitly defined as the channel from each UE to the last metasurface layer (Eq. 6), yet the downlink received signal is written with H^H (Eqs. 7-8). Reciprocity for a passive linear structure gives the transpose, not the conjugate transpose, in complex baseband. The difference is not a harmless convention: it changes the sign of every inter-element phase. A transmit field that maximizes |h_k^H g_k| is not the field that focuses on UE k; the correct focusing field is conjugate to the channel, which corresponds to |h_k^T g_k|. Consequently, the optimization P1 is fitting the SIM to the wrong target, and Figs. 2-5 do not support the central claim as stated. This concern is concrete and checkable with the data already in the paper, and it is more load-bearing than the inter-layer coupling question, which would only matter after the signal model is corrected. A corrected resubmission could potentially re-establish the claim, but the present version's numerical evidence is invalid, so the verdict should move from CONDITIONAL to REJECT.","tokens_in":10501,"tokens_out":14154,"duration_ms":146594,"concrete_test":"Re-run the entire optimization and evaluation with Q = H^T G instead of H^H G: replace Eq. (7), the NMSE objective (11a), the gradient expressions (12)-(14), and the Fig. 5 energy calculation with the non-conjugated channel. If the SIM no longer approaches ZF or the beamfocusing maps lose their foci, the central claim fails. A cheaper diagnostic: take the optimized G reported for one configuration, compute the two NMSEs ||H^H G - I||_F^2 and ||H^T G - I||_F^2; the former is expected to be ~0.002 and the latter should be large if the error is real. Since the paper gives all simulation parameters, this check requires no new physics or full-wave simulation.","verdict_should_be":"REJECT","load_bearing_attack":"The paper defines h_{m,k} in Eq. (6) as the channel from UE k to meta-atom m, with propagation phase e^{-j2πd/λ}. For the passive reciprocal SIM described, the downlink channel from meta-atom m to UE k is the same scalar, so the end-to-end downlink matrix is Q = H^T G, not Q = H^H G as used in Eq. (7), the objective (11a), the gradient (12)-(14), and the Fig. 5 field maps. Using H^H reverses the sign of every propagation phase. The intended focusing condition h_k^H g_k = 1 would correspond to a physical received signal h_k^T g_k = Σ_m e^{-2j(2π/λ)d_{m,k}} g_{k,m}, a sum of rapidly varying phases that generically has small magnitude; conversely, true focusing requires the transmitted field to be the conjugate h_k^*, i.e., the inner product h_k^T g_k. Thus the optimized meta-atom coefficients and the reported sum-rate and beamfocusing curves are computed for a non-physical, non-reciprocal channel. This is an internal inconsistency, not an external validation gap: even if the multiplicative factorization of G were exact, the simulated system would not be the physical SIM described.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a stacked intelligent metasurface (SIM) integrated at a base station to perform near-field multiuser beamfocusing in the wave domain, with the goal of replacing fully digital zero-forcing precoding. The SIM is modeled as a cascade of programmable metasurface layers with multiplicative propagation matrices, and the authors formulate a non-convex optimization problem to minimize the normalized mean-square error between the SIM-based end-to-end channel and a ZF target. They propose a gradient-descent algorithm with a coupled phase-amplitude model and report sum-rate and beamfocusing simulations showing that the SIM approaches ZF performance, outperforms codebook and random baselines, and achieves a 7.6% near-field gain over far-field operation. The central claims are that passive multilayer metasurfaces can substitute digital baseband beamforming and that near-field spherical wavefronts provide higher spatial gain.","tokens_in":10750,"tokens_out":9769,"duration_ms":99043,"significance":"If the proposed system model and results were correct, the paper would offer an interesting step toward replacing power-hungry digital beamforming with passive wave-domain computation for near-field multiuser communications. The paper includes a specific channel model, a concrete optimization algorithm, and extensive numerical comparisons against codebook and random baselines, which are useful for benchmarking. However, the manuscript contains a fundamental channel-modeling error that invalidates the reported results: the downlink end-to-end channel is defined with a conjugate transpose of the uplink channel instead of the transpose required by physical reciprocity. As a result, the optimized SIM coefficients and all numerical curves correspond to a non-physical, non-reciprocal system. The paper also provides no validation of the idealized multilayer propagation model, and the near-field advantage is based on a single unquantified percentage. These issues mean the main contribution, as presented, is not reliable.","major_comments":[{"comment":"The end-to-end downlink channel is defined as Q = H^H G, where H in Eq. (6) is the uplink channel from the k-th UE to the m-th meta-atom, with entries h_{m,k} = e^{-j2πd_{m,k}/λ}. By reciprocity, the downlink channel from each meta-atom to each UE is the same scalar h_{m,k}, so the physical end-to-end channel from the BS antennas to the UEs is Q_phys = H^T G, not H^H G. Using H^H reverses the sign of every propagation phase. This error propagates into the SINR in Eq. (9), the optimization objective in Eq. (11a), the gradient in Eqs. (12)-(14), and the beamfocusing maps in Fig. 5. For a single UE, the optimized G ≈ h/||h||^2 (since W_ZF = h/||h||^2) would yield a physical received signal h^T h / ||h||^2, which is a sum of random unit phasors and tends to 1/√M in magnitude for random phases, not a focused beam. Thus all simulation results describe a non-physical, non-reciprocal system. This is a load-bearing error: the central claim that the SIM approaches ZF performance is not supported for the actual physical channel. The authors must re-derive the model and all subsequent optimization with H^T in place of H^H and rerun the simulations.","section":"§II-C, Eqs. (6)-(9)"},{"comment":"The objective P1 explicitly minimizes the NMSE between H^H G and H^H W_ZF, i.e., it fits the SIM response to the ZF target. Therefore, the observation in Figs. 2-4 that the SIM 'approaches ZF performance' is largely a statement about the convergence of the gradient descent to its own fitting target, not an independent discovery. The meaningful comparisons are against the codebook and random baselines; the ZF-approximation claim should be framed as an algorithmic fitting result, and the paper should rely on the NMSE values (Fig. 4(b)) to quantify the fitting quality rather than treating the sum-rate proximity as an unexpected outcome.","section":"§III-A, Eq. (11a)"},{"comment":"The near-field vs. far-field gain is reported as a single 7.6% improvement with no confidence intervals, despite the statement that results are averaged over 100 experiments. It is unclear whether this difference is statistically significant across the random UE placements. Additionally, the path-loss model β = (λ/4π)^2 d^{-α} is introduced in Section IV but does not appear in the signal model of Eqs. (6)-(9), where the channel coefficients have unit magnitude. The authors should clarify exactly where β is applied (e.g., in the SNR, in the effective channel gain) and confirm that the reported sum-rate values include this path loss; otherwise the absolute rates and the near-field/far-field comparison are ambiguous.","section":"§IV, Fig. 2 and signal model"},{"comment":"The single-pass factorization G = Φ_L W_L ... Φ_1 W_1 assumes no multiple reflections between metasurface layers, no mutual coupling between meta-atoms, and that tuning one meta-atom does not alter the propagation matrices of other layers. These are strong idealizations for a 12-layer stack with meta-atoms spaced at one wavelength. The paper does not state these assumptions or provide any validation (full-wave simulation, measurement, or external reference) for the factorization in a multilayer configuration. The authors should explicitly acknowledge these idealizations and, if possible, provide evidence that the model is representative of physical multilayer metasurfaces, since the entire wave-domain beamforming result depends on this model.","section":"§II-A, Eq. (5)"}],"minor_comments":[{"comment":"There is a placeholder citation '[?]' in 'the authors of [ ?], [24]'; this must be replaced with the correct reference.","section":"§I, paragraph 4"},{"comment":"Eq. (5) states G ∈ C^{M×K}, but W_1 ∈ C^{M×S}; the product is M×S. Since the paper assumes S=K in the simulations and in the system setup, this is consistent only if S=K is stated explicitly. Please clarify the general dimensions or restrict the claim to S=K.","section":"§II-A, Eq. (5)"},{"comment":"The complexity expression is garbled: 'Also 4[...], O(...), ...' is not a clear list. Please rewrite the complexity breakdown as separate terms for forward propagation, gradient computation, regularization, and parameter updates.","section":"§III-C, complexity analysis"},{"comment":"There are typos in the figures: 'Syetem sum rate' in Fig. 2 and 'y-aixs'/'x-aixs' in Fig. 5 should be corrected.","section":"§IV, figures"},{"comment":"The proposed algorithm uses a codebook-based initial solution, but the construction of the codebook and how the initial phase values are selected are not described. Please specify this step, as it affects reproducibility.","section":"§III-B, initialization"}],"recommendation":"reject","confidential_remarks":"The transposition error identified in the major comments is the decisive issue: the manuscript's model is internally inconsistent with a reciprocal passive SIM, and the numerical results are for a different system. This is not a minor fix; it requires redoing the optimization and all simulations. In addition, the authors' extensive prior work on stacked intelligent metasurfaces makes the incremental novelty modest even if the model were corrected. The paper could be resubmitted as a new contribution after a thorough reworking, but in its current form it should not proceed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a coherent, well-written extension of SIM beamforming to near-field spherical wavefronts with a coupled amplitude-phase meta-atom model, but it has a load-bearing channel orientation error. The paper defines h_{m,k} as the propagation phase from UE k to meta-atom m (Eq. 6). For a passive reciprocal SIM, the same phase applies from meta-atom m to UE k, so the downlink end-to-end channel is Q = H^T G, not Q = H^H G as used in Eq. (7), the objective (11a), the gradients (12)-(14), and the field maps in Fig. 5. Using H^H conjugates every propagation phase, so the optimization is effectively focusing the phase-conjugated field. The physical received signal h_k^T g_k would be a sum of rapidly varying phases and would generically be small; the focusing condition is h_k^T g_k = 1, which requires the SIM to generate the conjugate of h_k. Thus the numerical results describe a non-reciprocal, non-physical channel, not the SIM in Fig. 1.\n\nWhat the paper does well: it introduces a near-field spherical-wavefront model and a practical coupled amplitude-phase response, and it gives a detailed gradient descent algorithm with normalization and codebook initialization. The extension is natural and the complexity analysis is useful. The 'approaches ZF' claim, however, is partly circular: the objective (11a) minimizes the NMSE between H^H G and H^H W_ZF, so reaching ZF is a statement about convergence to the fitting target, not an independent result. The near-field superiority is a single 7.6% point with no error bars. The multilayer model also assumes no inter-layer coupling or mutual coupling between meta-atoms, with no full-wave validation.\n\nIf the channel orientation is fixed (H^T G) and the simulations re-run, the results may or may not survive; the architecture is still worth studying. As is, I would not cite it and would not trust the quantitative claims. I think a serious editor should still send it to review, because the flaw is subtle and the paper is from a group that could fix it, but the referee should be pointed at the reciprocity issue immediately.","headline":"The paper's simulations use H^H where physical reciprocity requires H^T for the downlink, so the optimized SIM is focusing a conjugated field and the reported gains likely do not carry over to the hardware described.","tokens_in":11327,"tokens_out":5490,"would_cite":false,"duration_ms":52699,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A05","94A12"],"pacs":[],"model":"deepseek-v4-flash","headline":"A stack of programmable metasurface layers can shape near-field beams in the wave domain, matching zero-forcing digital precoding without its hardware cost.","keywords":["stacked intelligent metasurfaces","near-field communications","beamfocusing","wave-domain computing","multiuser MISO","zero-forcing precoding","reconfigurable intelligent surfaces","terahertz communications"],"falsifier":"A full-wave electromagnetic simulation or a prototype measurement of a multi-layer stack would settle it: if the measured end-to-end transmission matrix for random phase settings deviates from the product $G=\\Phi_L W_L\\cdots\\Phi_1 W_1$ by more than the noise floor, or if the beamfocusing maps fail to match the predicted focusing, the cascade model would be falsified.","tokens_in":10274,"feed_emoji":"📡","tokens_out":12283,"duration_ms":104063,"temperature":0.7,"pith_summary":"The paper tries to establish that a stacked intelligent metasurface (SIM) — several programmable layers of meta-atoms that reshape radio waves as they pass through — can do the beamforming that a base station normally does in digital baseband, but in the wave domain and at the speed of light. The setting is a near-field multiuser downlink, where wavefronts are spherical and a surface can focus energy at points in space rather than only in directions. The authors model the end-to-end channel through the stack, optimize the meta-atom responses with a gradient-descent algorithm, and report that the SIM's sum rate approaches fully digital zero-forcing precoding. If the result holds, near-field and terahertz systems could use many fewer RF chains and lower-resolution converters, shifting signal processing from chips to passive surfaces.","feed_headline":"Metasurface stack approaches zero-forcing near-field beamforming","feed_subtitle":"Near-field tests show the stack approaches digital precoding with fewer RF chains and low-resolution converters.","key_machinery":"The central mechanism is the cascaded propagation model of the SIM. Each layer is a diagonal diffraction-coefficient matrix $\\Phi_l=\\operatorname{diag}(\\varphi^l_1,\\dots,\\varphi^l_M)$, each gap contributes a free-space propagation matrix $W_l$ whose entries are scalar diffractive couplings, and the whole stack maps $S$ transmit antennas to the final layer through $G=\\Phi_L W_L\\cdots\\Phi_1 W_1$. The optimization drives the product $H^H G$ toward the target $H^H W_{\\mathrm{ZF}}$ by minimizing the normalized mean-square error, using gradient descent on the coupled phase-amplitude response of each meta-atom. This cascade identity is what lets a passive stack perform matrix-style computation on the wavefront instead of in a digital processor.","core_discovery":"On the paper's own terms, the central claim is that tuning the phase and amplitude of each meta-atom in a multilayer SIM can make the end-to-end channel $Q = H^H G$, with $G=\\Phi_L W_L\\cdots\\Phi_1 W_1$, approximate the zero-forcing channel $H^H W_{\\mathrm{ZF}}$ closely enough that inter-user interference nearly vanishes. The paper's simulations show the near-field model adding 7.6% sum rate over a far-field model at 150 m, a 239% sum-rate gain over the codebook baseline at $L=11$ layers, and the fitting NMSE falling from 0.74 to 0.002 as the layer count grows from 1 to 12. It also shows a single layer focusing on only two of four users, while four layers produce beamfocusing maps comparable to zero-forcing. The claim is that computation happens in the wave domain: the stack is not a passive reflector but an analog computer that customizes the propagation itself.","pith_inferences":["Going beyond the paper, the same wave-domain computing picture implies that a SIM could act as an analog front-end preprocessor for uplink reception and integrated sensing, separating users or targets before any digital conversion.","The paper leaves the energy budget implicit: more layers suppress interference but add insertion loss and tuning circuitry, so a net power-efficiency comparison against digital baseband is a natural next step.","The optimization assumes the SIM knows the near-field channel matrix; estimating $H$ through the stack is not modeled, so the reported gains are an upper bound on a real end-to-end protocol.","A direct experimental check of the cascade model—full-wave simulation or a prototype transmission-matrix measurement—would determine whether the beamfocusing result transfers from the scalar diffraction model to physical hardware."],"forward_implications":["If the claim is correct, a SIM with enough layers can replace fully digital zero-forcing precoding in near-field multiuser systems, eliminating most RF chains and high-resolution DACs.","Near-field operation adds a measurable spatial gain over far-field modeling: the paper reports 7.6% higher sum rate at 150 m, because spherical wavefronts give the channel higher rank.","More metasurface layers translate directly into better interference suppression: fitting NMSE falls from 0.74 at one layer to 0.002 at twelve, and the beamfocusing maps approach ZF with just four layers.","The optimization algorithm has complexity $O(4I_{\\mathrm{GDA}}[(2L-2)M^3+MLK^2])$, so the hardware savings come at the cost of offline optimization that grows with stack size, number of meta-atoms, and number of users."],"supporting_citations":[{"why":"Supplies the multilayer cascade model $G=\\Phi_L W_L\\cdots\\Phi_1 W_1$ and the free-space diffraction formula used to build each propagation matrix.","marker":"[15]"},{"why":"Provides the equivalent-circuit model of meta-atom impedance and the coupled phase-amplitude response used in the optimization.","marker":"[28]"},{"why":"Provides the codebook-based initialization and the codebook baseline that the proposed scheme is compared against.","marker":"[12]"},{"why":"Establishes the wave-domain multiuser beamforming formulation and supplies the codebook and random benchmark configurations.","marker":"[24]"},{"why":"Supplies the near-field spherical-wavefront channel modeling used for the link between the SIM and the users.","marker":"[7]"},{"why":"Motivates multilayer wave-domain computing by showing diffractive deep neural networks perform parallel matrix operations optically.","marker":"[16]"}],"fun_headline_variants":["Stacked metasurfaces compute beamfocusing in the wave domain","Near-field beamfocusing with stacked intelligent metasurfaces","Metasurface stack does wave-domain computing to cut RF chains","Stacked metasurfaces approach zero-forcing near-field beamforming","Wave-domain analog computer: stacked metasurfaces for multiuser beamfocusing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result depends on the assumption that a real multilayer metasurface behaves exactly as a chain of independent diagonal phase masks with free-space diffraction between layers, with no multiple reflections, no mutual coupling between meta-atoms, and no tuning-dependent change to the propagation matrices.","fun_headline_variants_meta":{"raw":{"variants":["Stacked metasurfaces compute beamfocusing in the wave domain","Near-field beamfocusing with stacked intelligent metasurfaces","Metasurface stack does wave-domain computing to cut RF chains","Stacked metasurfaces approach zero-forcing near-field beamforming","Wave-domain analog computer: stacked metasurfaces for multiuser beamfocusing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000775,"raw_usage":{"total_tokens":3443,"prompt_tokens":975,"completion_tokens":2468,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":2377}},"tokens_in":591,"tokens_out":2468,"duration_ms":17696,"temperature":1.0,"reasoning_tokens":2377,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T17:50:14.777038+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A full-wave electromagnetic simulation or a prototype measurement of a multi-layer stack would settle it: if the measured end-to-end transmission matrix for random phase settings deviates from the product $G=\\Phi_L W_L\\cdots\\Phi_1 W_1$ by more than the noise floor, or if the beamfocusing maps fail to match the predicted focusing, the cascade model would be falsified.","supporting_citations":[{"cited_title":"Stacked intelligent metasurfaces fo r efficient holographic MIMO communications in 6G,","cited_arxiv_id":null,"evidence_quote":"Supplies the multilayer cascade model $G=\\Phi_L W_L\\cdots\\Phi_1 W_1$ and the free-space diffraction formula used to build each propagation matrix."},{"cited_title":"Intellig ent reflecting surface: Practical phase shift model and beam forming optimization,","cited_arxiv_id":null,"evidence_quote":"Provides the equivalent-circuit model of meta-atom impedance and the coupled phase-amplitude response used in the optimization."},{"cited_title":"Codebook-based solutions for reconfigurable intell igent surfaces and their open challenges,","cited_arxiv_id":null,"evidence_quote":"Provides the codebook-based initialization and the codebook baseline that the proposed scheme is compared against."},{"cited_title":"Near-Field Communications: A Tutorial Review","cited_arxiv_id":"2305.17751","evidence_quote":"Supplies the near-field spherical-wavefront channel modeling used for the link between the SIM and the users."},{"cited_title":"All-optical machine learning using di ffractive deep neural networks,","cited_arxiv_id":null,"evidence_quote":"Motivates multilayer wave-domain computing by showing diffractive deep neural networks perform parallel matrix operations optically."}],"review_version":1}