{"id":"52b68944-f58c-402b-b832-bc23087c7a86","arxiv_id":"2412.06178","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Deep unfolding networks, PGD-Net for hybrid beamforming and GNN-aided AO-Net for power control, are applied to multi-port matching networks, approaching WMMSE and MO-AltMin with lower complexity.","lead":"This paper models 6G antenna systems as multi-port electrical circuits and uses deep unfolding to design beamforming and power control. It reports near-classical performance with lower run time, but a key constraint on the analog beamformer is not actually enforced as written.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (46) clamps real and imaginary parts separately to [−1,1], so |[FA]_{a,b}| can reach √2; Algorithm 3 has no normalization, so the PGD-Net output need not satisfy constant-modulus constraint (34).","rationale":"The reader's weakest assumption is the same load-bearing issue I find. PGD-Net is advertised as a low-complexity approximation of the feasible MO-AltMin solver, but Eq. (46) only enforces a box constraint on the real and imaginary parts separately; it is not a constant-modulus projection. Algorithm 3's reconstruction step does not normalize, and Algorithm 2's V^{-1} is also unspecified, so the produced FA can have entries with magnitude up to √2 and is not in A. The comparison in Fig. 8 is therefore not between feasible designs unless an undocumented projection is inserted. I did not find a stronger objection: the DMA insertion-loss study and the AO-Net power-control experiment are independent enough that the main problem is confined to the beamforming claim, though that is central to the abstract. A normalization could repair the issue, so the paper is a candidate for revision rather than hopeless, but as written the central claim is unsupported.","tokens_in":19016,"tokens_out":5669,"duration_ms":58075,"concrete_test":"Re-implement Algorithm 3 as written on the same simulation setup (M=120, NRF=6, K=6) and, for a held-out channel set, record the per-entry magnitudes |[F_A]_{a,b}| of the matrix produced in step 11 before the power scaling in (50). If the maximum magnitude exceeds 1+ε (the theoretical bound is √2), the output violates (34) as claimed. Then run a second variant where Step 11 is replaced by the explicit constant-modulus projection F_A = exp(j∠(Re(n_L)+jIm(n_L))). Compare the resulting SE curves against MO-Altmin in Fig. 8: if the normalized variant is needed to match MO-Altmin, the manuscript must document it; if the normalized variant loses SE relative to the reported curves, the reported PGD-Net gain came from infeasible, non-constant-modulus solutions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is that PGD-Net approximates MO-AltMin with low complexity for hybrid beamforming in multi-port matching networks. This claim depends on PGD-Net producing an analog precoder FA that lies in the feasible set A={FA: [FA]_{a,b}=e^{jϱ_{a,b}}} (Eq. 34), because MO-AltMin is being used as the feasible benchmark. The description of the projection step does not establish this. The activation ψ(t,x) in Eq. (46) is ψ(t,x)=−1+(1/|t|)(ReLU(x+t)−ReLU(x−t)), i.e., a clamped/saturated linear map whose output lies in [−1,1] for each real coordinate. Since n stacks Re(vec(FA)) and Im(vec(FA)), the activation bounds each component independently, not the complex magnitude. A pair (Re, Im) ∈ [−1,1]^2 yields magnitudes anywhere in [0,√2], including 0, so ψ does not implement the constant-modulus projection required by Eq. (42). Algorithm 3 step 11 says only 'Reconstruct the complex RF beamforming matrix FA from n_L' and step 12 applies (49)–(50); no normalization or unit-modulus projection is described. Algorithm 2 step 19 likewise reconstructs FA via V^{-1}(n_ℓ) without specifying such a projection. If V^{-1} is just the inverse of the real/imag stacking map, the output is infeasible and the SE comparison is not on equal footing with MO-AltMin. If a normalization is silently applied, it is absent from the algorithm and the training loss (47) is computed on unnormalized n_ℓ, so the unfolded iterations do not correspond to PGD on the stated feasible set. Either way the central comparison is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript models a multi-user MIMO downlink (and later uplink) as a multi-port matching network with impedance matrices, covering DMA, full-digital, and hybrid architectures. It first analyzes the impact of mutual coupling and insertion losses, then proposes two deep-unfolding designs: a PGD-Net that unfolds projected gradient descent to approximate a fully digital precoder under a constant-modulus analog constraint, and a GNN-aided AO-Net that unfolds WMMSE-type iterations for power control. Numerical results are reported for spectral efficiency, energy efficiency, complexity, and runtime, and the paper claims near-MO-AltMin beamforming performance and faster-than-WMMSE power control.","tokens_in":19466,"tokens_out":17380,"duration_ms":156489,"significance":"If the claims were fully established, the paper would make a useful contribution: it applies deep unfolding to circuit-theoretic MIMO/DMA models, emphasizes that insertion loss and mutual coupling matter for performance analysis, and uses unsupervised training that avoids label generation. I do not regard the unsupervised training as circular, since the networks minimize the same objectives (LS approximation and negative sum rate) that define the tasks. However, the constant-modulus constraint is not actually enforced by the proposed PGD-Net activation, and several dimensional and index inconsistencies appear in the system model and algorithms. These issues currently prevent the numerical comparisons from validating the headline claims, so a careful revision is needed.","major_comments":[{"comment":"The activation ψ(t,x)=−1+(1/|t|)(ReLU(x+t)−ReLU(x−t)) clamps each real coordinate of n_ℓ to [−1,1]. Since n stacks the real and imaginary parts of vec(FA), this bounds each component separately and allows complex entries with modulus anywhere in [0,√2]; for example Re=Im=1 gives modulus √2 and Re=Im=0 gives modulus 0. Thus ψ does not project onto the constant-modulus set A in (34). Algorithm 3 step 11 reconstructs FA from n_L without any normalization or unit-modulus projection, and Algorithm 2 step 19 does the same during training while the loss (47) is computed on unnormalized n. The SE comparison with MO-AltMin, which enforces (34), is therefore not on equal footing, and the claim that PGD-Net approximates the feasible MO-AltMin solution is unsupported as written. The authors should add an explicit unit-modulus projection to the forward pass and to the reconstruction, or show that the unnormalized output nevertheless satisfies (34).","section":"Section III-B, Eq. (46), Algorithms 2-3"},{"comment":"The dimensions of the beamforming matrix and the impedance matrices are inconsistent. With jt∈C^{N_RF} in (1), Ztt∈C^{N_RF×N_RF}, Zrs∈C^{K×M}, and Zst∈C^{M×N_RF}, the equivalent channel H in (7) is K×N_RF, so the beamforming matrix in y=HFs should be N_RF×K, not the M×K stated before (6). The transmitted power in (11) is only defined if F is N_RF×K, yet hybrid beamforming later defines F=FAFD with FA∈C^{M×N_RF} and FD∈C^{N_RF×K}, making F M×K; using the same Ztt in (31c) is then dimensionally invalid. The DMA precoding in (26)-(28) implicitly uses F_dma∈C^{N_RF×K}, confirming the inconsistency. The uplink constraint (53c) is also unclear, since Zrr was defined as K×K while g_k∈C^{M×1}. The authors should state the port count consistently for the DMA, FD, and hybrid cases and rewrite the affected definitions and constraints.","section":"Section II/III-B/IV, Eqs. (5)-(11), (31c), (53c)"},{"comment":"The DMA ZF precoder F_dma=ξ_dma H†_dma is normalized with ξ_dma based on the supplied power P_s^max in (26)-(27), but the resulting SINR in (29) is written with P_t^max in the numerator. Since the SINR expression should be ξ_dma^2 σ_x^2/σ_n^2, the numerator should be P_s^max unless a separate argument equates the two power limits. The same ambiguity appears in (32)-(33), where F_opt is scaled by P_s^max while the hybrid constraint (31c) is on P_t. This needs to be corrected and stated explicitly.","section":"Section III-A, Eqs. (26)-(29)"},{"comment":"The WMMSE-derived updates contain reproducibility errors. Algorithm 4 steps 8-10 refer to Eqs. (56), (57), and (58) for u, w, and q, but those variables are defined in (55), (56), and (57), respectively. More substantively, the denominator in (57) reads ∑_m ar h_{m,k} u_{ℓ,m} u_{ℓ,m} w_{ℓ,m}, with u_{ℓ,m} appearing twice and no square on ar h_{m,k}; the standard WMMSE form is ∑_m |ar h_{m,k}|^2 u_{ℓ,m} w_{ℓ,m}. The subscript in q(k−1) in (56) is also ill-defined. These issues must be fixed before the AO-Net results can be independently reproduced.","section":"Section IV, Eqs. (55)-(57), Algorithm 4"}],"minor_comments":[{"comment":"The text says 'we set the frequency f to 10 Hz', while Table I states 10 GHz; the waveguide dimensions and all numerical results are consistent only with 10 GHz, so this should be corrected.","section":"Section V-A, Table I"},{"comment":"The symbol 'P maxss' in (27) is garbled and should presumably be P_s^max; please normalize the notation throughout Section III-A.","section":"Eq. (27)"},{"comment":"The captions refer to 'PCG-Net' although the proposed method is called PGD-Net; the acronym should be made consistent.","section":"Figures 3 and 4 captions"},{"comment":"The MO-Altmin baseline is cited to [33], but the described algorithm appears to match [22]; please verify the citation.","section":"Section V-C"},{"comment":"Table II reports point values without standard deviations, number of test samples, or repeated trials; given the small SE gap between AO-Net (6 layers) and WMMSE, error bars would support the comparison.","section":"Table II"},{"comment":"The paper does not state the training-set size, the number of test channels, or the random seeds used for channel generation, which limits the reproducibility of the numerical results.","section":"Section V-A"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the proposed framework is potentially salvageable. The constant-modulus handling, dimensional inconsistencies, and the AO-Net update errors are load-bearing but fixable by revising the forward models, adding an explicit projection, and rerunning the comparisons. I therefore recommend major revision rather than rejection. It would also be worth asking the authors to check the MO-Altmin citation and to provide more details on the test setup."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a modest extension of known deep-unfolding techniques to a circuit-theoretic MIMO model. The DMA analysis (insertion loss, mutual coupling, energy efficiency) is reasonable, and the GNN-aided AO-Net for power control looks sound. But the PGD-Net beamforming design has a genuine feasibility problem: the activation in Eq. (46) clamps the real and imaginary parts of each analog-precoder entry separately to [−1,1], which does not enforce the unit-modulus constraint in Eq. (34). Algorithm 3 step 11 says only \"reconstruct\" FA from n_L, with no normalization or unit-circle projection described. So the output can have entries with magnitude anywhere from 0 to √2. That makes the SE comparison with MO-AltMin, which does enforce constant modulus, not on equal footing, and it undercuts the paper's main beamforming claim. The stress-test note is correct on this; I don't see a way to read Eq. (46) as a unit-modulus projection.\n\nThe genuinely new part is narrow: refs [19] and [20] already provide the DMA multi-port channel model and insertion-loss treatment, and refs [21] and [25] already establish unfolded PGD and GNN-unfolded WMMSE. The contribution here is combining those pieces for a matching-network system, which is a valid but incremental step. The NAG-based insertion-loss optimization is simple and appears effective. The AO-Net section is a straightforward extension of [25] with graph layers, and its runtime gains over WMMSE are believable.\n\nSoft spots in proportion: the constant-modulus issue is load-bearing and needs to be fixed before the beamforming results can be accepted. The minor inconsistencies (10 Hz vs 10 GHz, P_t^max vs P_s^max in Eq. (29), the duplicated u_{ℓ,m} in Eq. (57)) are mostly typographical but reduce confidence. The complexity analysis is plausible, though the asymptotic claims are a bit hand-wavy.\n\nIf the beamforming section were revised to use a proper unit-modulus projection (or to explicitly model amplitude-and-phase analog hardware, which would contradict the stated system model), the paper would be a competent incremental contribution. As it stands, the central beamforming comparison is unsupported, but the DMA analysis and AO-Net results are probably publishable on their own.\n\nThe paper deserves a serious referee: the flaw is substantial but fixable, and the DMA/power-control content has enough value that an editor should not desk-reject. A good reviewer would ask for major revisions, not dismissal.","headline":"A competent but incremental deep-unfolding paper whose beamforming section has a real constant-modulus feasibility flaw; the DMA and power-control parts are more defensible.","tokens_in":19997,"tokens_out":3078,"would_cite":false,"duration_ms":33010,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that deep unfolding can accelerate multi-port MIMO beamforming and power control: PGD-Net matches alternating-minimization hybrid precoding at about 30% lower complexity, and GNN-aided AO-Net approaches WMMSE sum rate…","keywords":["multi-port matching networks","deep unfolding","hybrid beamforming","power control","dynamic metasurface antenna","mutual coupling","graph neural network","projected gradient descent"],"falsifier":"Record $F_A$ output by the trained PGD-Net for a channel matrix in the test set and compute $\\max_{a,b} \\big||F_A|_{a,b}| - 1\\big|$; if this exceeds the numerical tolerance used for MO-AltMin's constant-modulus constraint, then the two methods are not compared under the same feasible set.","tokens_in":18828,"feed_emoji":"📡","tokens_out":7855,"duration_ms":73241,"temperature":0.7,"pith_summary":"This paper tries to show that circuit-level effects—impedance matching, mutual coupling, and insertion loss—should be built into MIMO signal-processing models, and that deep unfolding can turn two classical optimizers into fast learned solvers for those models. For hybrid beamforming, it unfolds projected gradient descent into a network (PGD-Net) that solves the matrix-approximation problem $\\min \\|F_{\\mathrm{opt}} - F_A F_D\\|_F^2$; numerical results report spectral efficiency close to the conventional MO-AltMin baseline with roughly 29.68% fewer complex multiplications. For power control, it unfolds the weighted MMSE algorithm into AO-Net, with graph-neural-network layers replacing the per-iteration weights, and reports nearly the same sum rate as WMMSE in about a sixth of the runtime. The same multi-port model is used to show that DMA spectral efficiency saturates as elements per waveguide grow and that ignoring mutual coupling or insertion loss misestimates performance. If these claims hold, near-field and metasurface MIMO systems can use physics-aware channel models without paying the full computational cost of model-based optimization.","feed_headline":"Unfolded beamforming matches optimal MIMO at ~30% lower cost","feed_subtitle":"Projected-gradient steps become network layers, cutting beamforming operations by ~30% and power-control runtime sixfold.","key_machinery":"The load-bearing objects are the multi-port admittance matrix $Z$ that couples transmitter, DMA element, and user ports, and the deep-unfolded layers derived from two classical iterations. In beamforming, the working identity is the real-valued splitting $n = [\\operatorname{Re}(\\mathrm{vec}(F_A)); \\operatorname{Im}(\\mathrm{vec}(F_A))]$, which turns the hybrid precoding objective into the least-squares form $\\|m - Dn\\|^2$; PGD-Net then iterates $n_\\ell = F_\\ell(n_{\\ell-1} - \\mu_\\ell \\bar{m} + \\mu_\\ell \\bar{D} n_{\\ell-1})$ with a learned projection $F_\\ell$ and the activation $\\psi(t,x)$. In power control, the mechanism is the WMMSE fixed-point update on $u_{\\ell,k}$, $w_{\\ell,k}$, and $q_{\\ell,k}$, with graph-neural-network functions $a_\\ell = \\Theta(H;\\vartheta_{\\ell,1})$ and $b_\\ell = \\Theta(H;\\vartheta_{\\ell,2})$ replacing hand-tuned MMSE weights, making the network unsupervised and size-adaptive. The DMA analysis additionally uses the supplied-power formula $P_s = \\frac{\\sigma_x^2}{2}\\operatorname{Tr}(\\operatorname{Re}(F^H Z_q F))$ with $Z_q = (I - \\Lambda^H \\Lambda)^{-1} Z_p$ to carry insertion-loss effects into the optimization.","core_discovery":"On the paper's own terms, the central discovery is that the optimization problems governing a multi-port matching network—hybrid beamforming under unit-modulus analog constraints and power allocation under per-user power limits—can be unrolled into trainable networks that retain the structure of the underlying algorithms. The equivalent channel takes the circuit form $H = \\tilde{Z}_r (Z_{rs} (Z_s + Z_{ss})^{-1} Z_{st} - Z_{rt})$, and the beamforming design minimizes $\\|F_{\\mathrm{opt}} - F_A F_D\\|_F^2$ with $F_A$ phase-only; PGD-Net replaces each projected-gradient step by a layer whose parameters $\\{\\vartheta_{\\ell,1}, \\vartheta_{\\ell,2}\\}$ are learned without labels. Power control is formulated as sum-rate maximization with constraints on transmit power and receiver regularization; AO-Net unrolls WMMSE updates for $u_{\\ell,k}$, $w_{\\ell,k}$, and $q_{\\ell,k}$, using a GNN $\\Theta(H;\\vartheta)$ to produce the per-layer weights. The numerical section reports that PGD-Net's spectral efficiency matches MO-AltMin while cutting complex multiplications by about 29.68%, and that a six-layer AO-Net reaches 33.127 bps/Hz against WMMSE's 33.298 bps/Hz with 28 ms versus 61 ms runtime. It further reports that DMA systems have an optimal number of elements per waveguide and that insertion-loss-aware optimization tracks the lossless ideal closely.","pith_inferences":["A direct testable extension is to normalize $F_A$ after reconstruction (or use a true unit-modulus projection) and re-run the spectral-efficiency comparison; if the gap to MO-AltMin widens, part of PGD-Net's reported efficiency comes from relaxing the constant-modulus constraint.","The same unfolding recipe—unroll an optimization iteration and replace its per-step hyperparameters with a learned graph network—should transfer to other coupled-antenna architectures such as holographic MIMO surfaces or RIS-aided systems, where the impedance matrix changes structure but the update form remains.","Because AO-Net uses GNN layers rather than per-array dense layers, it may generalize across numbers of ports and users without retraining; the paper hints at this via GNN's permutation-equivariance but does not demonstrate cross-size transfer.","A natural objective-level extension is to train PGD-Net and AO-Net with a loss that includes supplied power or energy efficiency, so the unfolding directly optimizes the energy-efficiency metric shown in the simulations rather than spectral efficiency alone."],"forward_implications":["A deployed PGD-Net would deliver hybrid beamformers for large arrays in significantly fewer complex multiplications than MO-AltMin, with the paper reporting about 29.68% lower complexity and nearly identical spectral efficiency.","AO-Net power control can run in milliseconds, making per-channel power allocation practical under fast fading where WMMSE's 61 ms runtime is too slow.","DMA designers should treat spectral efficiency per waveguide as saturating; the paper's results put a cost-performance limit on elements per waveguide, so adding elements beyond the threshold wastes hardware.","Circuit-level modeling matters: ignoring mutual coupling or insertion loss in the multi-port network changes the predicted spectral efficiency, so physics-aware channel models are needed for DMA and near-field MIMO.","Energy-efficiency comparisons favor DMA and hybrid architectures over full digital, since fewer RF chains and no phase-shifter array reduce total power for comparable spectral efficiency."],"supporting_citations":[{"why":"Provides the alternating-minimization hybrid precoding framework and the MO-AltMin and SDR-AltMin baselines that PGD-Net is compared against.","marker":"[22]"},{"why":"Gives the weighted MMSE algorithm whose iterations AO-Net unfolds, and whose sum-rate and runtime are the main power-control baseline.","marker":"[27]"},{"why":"Supplies the electromagnetic multi-port channel model for dynamic metasurface antennas that the paper extends with insertion-loss and coupling analysis.","marker":"[19]"},{"why":"The prior DMA insertion-loss and coupling study that motivates Algorithm 1's gradient-based load-admittance optimization.","marker":"[20]"},{"why":"Establishes the GNN-unfolded WMMSE architecture that AO-Net adapts for multi-port matching networks.","marker":"[25]"},{"why":"Provides the near-field MO-AltMin variant used as a fully-connected hybrid beamforming baseline in the numerical comparisons.","marker":"[33]"},{"why":"Supplies the CNN 'learning to optimize' power-control baseline whose poor performance in multi-port settings motivates the GNN-aided approach.","marker":"[28]"},{"why":"Defines the deep-unfolding methodology that both PGD-Net and AO-Net are built on.","marker":"[13]"}],"fun_headline_variants":["Deep-unfolded beamforming matches optimal MIMO at ~30% lower complexity","AO-Net power control matches WMMSE at 2x faster runtime","PGD-Net: Unfolded beamforming cuts multiplications by ~30%","Unrolled optimization matches MIMO performance with ~30% fewer ops"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that PGD-Net respects the unit-modulus phase-shifter constraint rests on the activation in equation (46) bounding the stacked real and imaginary parts of the analog beamformer to $[-1,1]$, which actually allows entry magnitudes up to $\\sqrt{2}$ and is never followed by normalization in Algorithm 3.","fun_headline_variants_meta":{"raw":{"variants":["Deep-unfolded beamforming matches optimal MIMO at ~30% lower complexity","AO-Net power control matches WMMSE at 2x faster runtime","PGD-Net: Unfolded beamforming cuts multiplications by ~30%","Unrolled optimization matches MIMO performance with ~30% fewer ops"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001174,"raw_usage":{"total_tokens":4948,"prompt_tokens":1133,"completion_tokens":3815,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":749,"completion_tokens_details":{"reasoning_tokens":3731}},"tokens_in":749,"tokens_out":3815,"duration_ms":29756,"temperature":1.0,"reasoning_tokens":3731,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:55:48.585281+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record $F_A$ output by the trained PGD-Net for a channel matrix in the test set and compute $\\max_{a,b} \\big||F_A|_{a,b}| - 1\\big|$; if this exceeds the numerical tolerance used for MO-AltMin's constant-modulus constraint, then the two methods are not compared under the same feasible set.","supporting_citations":[{"cited_title":"Alternating minimization algorithms for hybrid precoding in millimeter wave MIMO systems,","cited_arxiv_id":null,"evidence_quote":"Provides the alternating-minimization hybrid precoding framework and the MO-AltMin and SDR-AltMin baselines that PGD-Net is compared against."},{"cited_title":"Electromagnetic based communication model for dynamic metasurface antennas,","cited_arxiv_id":null,"evidence_quote":"Supplies the electromagnetic multi-port channel model for dynamic metasurface antennas that the paper extends with insertion-loss and coupling analysis."},{"cited_title":"Performance evaluation of dynamic metasurface antennas: Impact of insertion losses and coupling,","cited_arxiv_id":null,"evidence_quote":"The prior DMA insertion-loss and coupling study that motivates Algorithm 1's gradient-based load-admittance optimization."},{"cited_title":"Unfolding WMMSE using graph neural networks for efficient power allocation,","cited_arxiv_id":null,"evidence_quote":"Establishes the GNN-unfolded WMMSE architecture that AO-Net adapts for multi-port matching networks."},{"cited_title":"Beam focusing for near-field multiuser MIMO communications,","cited_arxiv_id":null,"evidence_quote":"Provides the near-field MO-AltMin variant used as a fully-connected hybrid beamforming baseline in the numerical comparisons."},{"cited_title":"Learning to optimize: Training deep neural networks for interference management,","cited_arxiv_id":null,"evidence_quote":"Supplies the CNN 'learning to optimize' power-control baseline whose poor performance in multi-port settings motivates the GNN-aided approach."}],"review_version":1}