{"id":"232ce585-c107-4b0e-8db7-e9b28244c908","arxiv_id":"2507.01876","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A sparse multi-dimensional graph neural network reduces computational complexity of joint power control and precoding in cell-free massive MIMO by about 48% with only a 2.11% spectral efficiency loss.","lead":"Cell-free massive MIMO networks need fast power control and precoding; this paper sparsifies a graph neural network's access point to user connections to cut computation by about half while losing only a few percent of spectral efficiency. It is a practical engineering trade-off study for deploying GNN-based resource allocation in large wireless networks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (11) violates the per-AP power constraint, so the reported SE and complexity trade-off may rely on infeasible transmit powers; this must be corrected and rerun.","rationale":"The reader's weakest-assumption analysis focused on threshold selection using the same evaluation data, which is a valid generalization concern. However, the more fundamental issue is that the output layer in Eq. (11), the only mechanism that enforces the per-AP power constraint, does not actually enforce it unless the unconstrained network outputs satisfy D_j ≥ K. Since the optimization objective rewards larger SINR, there is an incentive for the network to produce small norms and thereby violate the power budget, making the reported SE values feasible only outside the stated system model. This concern is more load-bearing than the threshold protocol because it affects the validity of every reported result, not just generalization to unseen thresholds. The threshold-selection issue remains relevant and should be addressed in revisions, but the first priority is to correct or clarify the output normalization and rerun the simulations. If the corrected rerun preserves the reported trade-off, the paper could be acceptable after adding implementation details; if not, the central claim fails. I therefore recommend keeping the reader's CONDITIONAL verdict, with the condition now including a verification of the power constraint rather than only threshold-selection and variance reporting.","tokens_in":7171,"tokens_out":5016,"duration_ms":69583,"concrete_test":"Re-run the joint power-control and precoding experiments with the corrected feasible normalization F_{j,k} = X^{(L)}_{j,k} · sqrt(P_max / Σ_{k'} ||X^{(L)}_{j,k'}||²), or with any other output mapping that provably satisfies the per-AP constraint (2). Then recompute the mean SE and wall-clock complexity reported in Sec. IV-C and Fig. 5. If SP-MDGNN's gap to MDGNN changes materially from 2.11% SE loss / 48% complexity reduction, or if the original outputs violate P_max on the test set, the central claim does not hold.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing assumption is hidden in the output normalization, Eq. (11). As written, F_{j,k} = X^{(L)}_{j,k} / ||X^{(L)}_{j,k}|| · sqrt(P_max / Σ_{k=1}^{K} ||X^{(L)}_{j,k}||²). Let D_j denote the denominator sum for a fixed AP j. Then Σ_k ||F_{j,k}||² = K · P_max / D_j. The per-AP power constraint (2) requires this to be ≤ P_max, i.e. D_j ≥ K. Nothing in the network architecture, the loss, or the training procedure enforces D_j ≥ K. If the network outputs small-norm feature vectors, the resulting F can exceed the power budget; in the limit D_j → 0 the scale factor diverges, giving unbounded transmit power and hence inflated SINR and SE. Consequently, the reported 2.11% SE loss and 48% complexity reduction in Sec. IV-C and Fig. 5 are not computed under the feasible power constraint stated in problem (6). Even if the implementation actually uses a different normalization, the paper as written is internally inconsistent, and the central empirical claim depends on a feasible output mapping.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SP-MDGNN, a sparse multi-dimensional graph neural network for joint power control and precoding in cell-free massive MIMO systems. The idea is to prune low-impact AP-UE connections using a learnable adjacency tensor with a threshold, thereby reducing inference complexity while approximately preserving spectral efficiency. The paper reports that, compared with the non-sparse MDGNN baseline, SP-MDGNN reduces computational complexity by 48% with only a 2.11% SE loss for the joint task, and it also compares against WMMSE and an attention-based MDGNN variant. The central claim is an empirical performance-complexity trade-off that favors the proposed sparse architecture.","tokens_in":7359,"tokens_out":6701,"duration_ms":68894,"significance":"If the reported trade-off is reliable, the paper would be a useful practical contribution: it addresses a real bottleneck in deploying learned resource allocation in CF mMIMO by reducing inference cost without major SE loss. The sparsification approach is plausible and the paper identifies a relevant baseline comparison. However, the central quantitative claims are not established as written: the output normalization in Eq. (11) can violate the per-AP power constraint, the threshold selection may use the test data, the complexity metric is wall-clock time without implementation detail, and no statistical uncertainty is reported. The idea is worth pursuing, but the manuscript needs substantive revision before the claims can be accepted.","major_comments":[{"comment":"The normalization in Eq. (11) does not enforce the per-AP power constraint (2). For a fixed AP j, let D_j = Σ_k ||X^(L)_{j,k}||². Then each F_{j,k} has squared norm P_max / D_j, so Σ_k ||F_{j,k}||² = K · P_max / D_j, which equals P_max only if D_j = K. Nothing in the architecture, the loss, or the training procedure enforces D_j ≥ K; if the network outputs small-norm feature vectors, the transmit power exceeds P_max, and in the limit D_j → 0 the scale factor diverges. Consequently, the SINR and SE values reported in Sec. IV-C and Fig. 5 may be computed under infeasible transmit powers. In addition, the paper states that for precoding tasks the output X^(L) 'directly forms the precoding vectors' without any normalization, leaving the per-AP constraint unaddressed for those tasks. Please correct the output mapping so that (2) is guaranteed by construction and rerun all affected experiments.","section":"Sec. III-C, Eq. (11)"},{"comment":"The sparsity thresholds τ_pc = 0.63 and τ_prec = 0.62 are selected by maximizing the harmonic-mean score H_s on evaluations whose dataset is not specified. The manuscript only defines a training set of 10,000 samples and a testing set of 2,000 samples. If the same testing data are used both for threshold selection and for the final SE and complexity numbers in Sec. IV-C, then the reported performance is not an independent evaluation and the claimed trade-off may not generalize. Please use a separate validation set for model selection and report final results on a held-out test set, or explicitly justify the protocol if the same data were used.","section":"Sec. IV-B"},{"comment":"The complexity reduction is measured solely as wall-clock computation time 'using the same hardware setup' with no further details: no hardware specifications, software/library, number of samples, batch size, or averaging across runs are given. Wall-clock time is implementation-dependent and does not by itself establish an algorithmic complexity reduction. Please report a hardware-independent metric (e.g., FLOPs, number of retained multiplications, or parameter counts), or provide a complete benchmarking setup with statistical variation, so that the claimed 48% reduction can be assessed.","section":"Sec. IV-C, Fig. 5"},{"comment":"The binary mask M = 1[A_{j,k,n} > τ] is non-differentiable, yet the paper does not explain how gradients are propagated through this hard threshold during training (e.g., straight-through estimator, Gumbel-sigmoid relaxation, or a two-phase training procedure). Additionally, Eq. (10) uses A⊙X^(l-1) with the adjacency tensor A, while Eq. (9) defines H_sparse = H⊙M and states that H_sparse is the GNN input; the relationship between A, M, and the input features needs to be made precise for the method to be reproducible.","section":"Sec. III-A, Eqs. (8)-(10)"},{"comment":"All quantitative results (SE losses of 1.3%, 1.44%, and 2.11%; complexity reductions of 55%, 49%, and 48%) appear to come from single simulation runs; no error bars, confidence intervals, or multiple-seed statistics are reported. Given the stochastic nature of GNN training and the small performance differences claimed, these point estimates are insufficient to support the paper's central trade-off claims.","section":"Sec. IV"}],"minor_comments":[{"comment":"The symbol H_{j,k,n} is used both for the channel matrix element and for the small-scale fading coefficient; use lowercase h_{j,k,n} for the small-scale fading coefficient to avoid confusion.","section":"Sec. II, Eq. (1)"},{"comment":"The sum over i = 1 to 4 in the layer propagation is unexplained; the manuscript should state what the four aggregation terms represent.","section":"Sec. III-B, Eq. (10)"},{"comment":"The definitions of Sparsity (S) and Performance Retention (P) are not given as equations; without formal definitions, the harmonic-mean analysis is not reproducible.","section":"Sec. IV-B"},{"comment":"The attention-based MDGNN (A-MDGNN) is introduced as a baseline, but no description or equations for the attention mechanism are provided in the method section; please add a reference or a concise description.","section":"Sec. IV-C"},{"comment":"The power-control simulation uses 'M = 8 antennas' while the system model in Sec. II uses N antennas; the notation should be made consistent across the paper.","section":"Sec. IV-A, Sec. IV-C"},{"comment":"The conclusion states that robustness is 'validated across a myriad of network conditions,' but the experiments cover only two specific settings (9 APs/8 UEs for power control and 4 UEs/16 antennas for precoding); please either add the supporting experiments or temper the claim.","section":"Sec. V"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the power-normalization error in Eq. (11), which may invalidate the reported SE numbers; the authors should be asked to correct the output mapping and rerun the experiments before acceptance. The threshold-selection and wall-clock-complexity issues also need to be addressed in a revision. The paper is within the scope of a correspondence and the sparsification idea is plausible, but the current evidence is not yet convincing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this one because the title is exactly the kind of thing that gets cited. The short version: the sparse-GNN idea is sensible and the paper is not a crank document, but the main quantitative claim is undermined by a load-bearing bug in the output normalization.\n\nWhat's new: combining a learnable sigmoid-threshold mask with MDGNN for joint power control and precoding in cell-free massive MIMO, and reporting the resulting complexity-performance trade-off. The harmonic-mean threshold selection in Sec. IV-B is a reasonable heuristic, and the WMMSE comparison grounds the results externally. The sparsity-as-regularization observation (faster convergence) is plausible and worth noting.\n\nThe soft spots are real, and the biggest is Eq. (11). As written, F_{j,k} = X/||X|| · sqrt(Pmax / Σ||X||²). Squaring and summing over k gives K·Pmax / Σ||X||². The per-AP constraint (2) requires this to be ≤ Pmax, i.e. Σ||X||² ≥ K. Nothing in the architecture, loss, or training enforces that. If the network outputs small-norm features, the scale factor blows up and the reported SINR/SE are computed at infeasible transmit powers. The 2.11% SE loss and 48% complexity reduction in Sec. IV-C therefore may not be feasible. This is an internal contradiction with the problem statement, and it must be corrected and the simulations rerun.\n\nOther issues are milder but still matter. Thresholds τ_pc=0.63 and τ_prec=0.62 are chosen by maximizing H_s on the same 2000 test samples used for the final SE numbers (Sec. IV-B), so the final evaluation is not independent. Complexity is reported only as wall-clock time without implementation details or flop counts, so the 48% reduction could be an artifact of sparse-layer implementation. There are no error bars or multiple seeds, and single-run differences of 1-2% could be noise. The binary mask training (Eq. (7)-(9)) is not explained—thresholding is non-differentiable, and the paper never says whether they used a straight-through estimator or some other trick.\n\nNone of the secondary issues are fatal on their own; the power-constraint bug is the one that blocks acceptance. If the authors fix the normalization, move threshold selection to a validation set, report variance, and clarify the mask training, this would be a credible incremental engineering result for a wireless-communications venue. As is, I would send it to peer review only with a strong request for major revision, not desk-reject it. The topic is relevant and the approach is not wrong-headed; the evidence just doesn't yet support the headline numbers.\n\nFor you: probably not worth citing until revised. If you want a classroom example of a power-constraint bug in a GNN paper, this is a clean one.\n\nRecommendation: send to peer review as a borderline major-revision; don't take the numbers at face value.","headline":"Sparse GNN pruning for CF mMIMO is a reasonable extension, but the power-normalization in Eq. (11) violates the stated constraint, so the headline 48%-for-2% trade-off is not trustworthy as written.","tokens_in":7918,"tokens_out":5783,"would_cite":false,"duration_ms":62529,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Sparsifying the access-point-to-user connections in a multi-dimensional graph neural network cuts computational complexity by 48% for joint power control and precoding in cell-free massive MIMO while losing only 2.11% of spectral…","keywords":["cell-free massive MIMO","graph neural networks","power control","precoding","sparsification","spectral efficiency","computational complexity","WMMSE"],"falsifier":"Run the same joint power-control and precoding experiment but choose the thresholds on a validation set that is completely disjoint from the test set, then measure mean SE and computation time on the held-out test data. If the 48% complexity reduction or the 2.11% SE loss changes noticeably, or if the harmonic-mean score peaks at a different threshold, the reported optimal balance is an artifact of selection on the test set.","tokens_in":6945,"feed_emoji":"📡","tokens_out":9859,"duration_ms":93354,"temperature":0.7,"pith_summary":"This paper argues that in cell-free massive MIMO, the connections between access points and user terminals are not all equally important, and that pruning the weak ones in a graph neural network can dramatically cut computational cost with a negligible loss in spectral efficiency. It proposes SP-MDGNN, which sparsifies the multi-dimensional graph neural network (MDGNN) by thresholding a learned adjacency tensor, and reports that for joint power control and precoding this lowers computation time by 48% while reducing mean SE by only 2.11% relative to the dense MDGNN. For the separate tasks the trade-off is even gentler: 55% complexity reduction at 1.3% SE loss for power control, and 49% at 1.44% for precoding. The paper also demonstrates that a harmonic-mean score of sparsity and performance retention picks a clear optimal threshold, and that the classical WMMSE algorithm remains more accurate but far more expensive.","feed_headline":"Sparse graph network cuts cell-free MIMO compute by 48%","feed_subtitle":"Pruning weak AP-UE links keeps 97.89% of spectral efficiency while halving computation.","key_machinery":"The core mechanism is the sparse adjacency mask. A trainable tensor $W$ is mapped through a sigmoid to link weights $A_{j,k,n}=\\sigma(W_{j,k,n})$, and a binary mask $M_{j,k,n}$ is set to 1 when $A_{j,k,n}>\\tau$ and 0 otherwise; the sparse channel matrix is $H_{\\text{sparse}}=H\\odot M$. This mask is folded into each GNN layer update $X^{(l)}=\\phi\\left(\\sum_{i=1}^4 P_i^{(l)}(A\\odot X^{(l-1)})\\right)$, so pruned links consume no aggregation cost. The threshold $\\tau$ is not global but chosen per task by maximizing the harmonic-mean score $H_s=2SP/(S+P)$, where $S$ is normalized sparsity and $P$ is performance retention relative to the dense baseline; this yields $\\tau_{\\mathrm{pc}}=0.63$ for power control and $\\tau_{\\mathrm{prec}}=0.62$ for precoding. The mask is the single element that turns a dense MDGNN into a sparse one and is also what lets the paper quantify the trade-off with just one parameter.","core_discovery":"The central claim is that a sparse version of MDGNN can replace the dense MDGNN for joint power control and precoding in cell-free massive MIMO with only a small spectral-efficiency penalty. The sparsification is data-driven: each AP-UE link gets a learnable weight passed through a sigmoid, and links below a threshold are pruned before the GNN layers aggregate, so they do not contribute to message-passing cost. On the joint task, SP-MDGNN retains 97.89% of the non-sparse MDGNN's mean SE while reducing computation time by 48%; on power control alone the numbers are 98.7% SE retention with 55% less computation, and on precoding alone 98.56% with 49% less. The paper concludes that this is the optimal performance-complexity balance among the compared approaches, including WMMSE and an attention-augmented MDGNN.","pith_inferences":["The threshold-selection procedure likely overestimates the generality of the trade-off: a fair evaluation would require thresholds tuned on a separate validation set, and the reported 48% and 2.11% figures could shift for data outside the current test set.","Wall-clock computation time on one hardware setup is the only complexity metric used; an energy-per-inference or asymptotic-flop comparison might give a different picture of the trade-off, and the numbers may not transfer to other hardware or to imperfect CSI.","The harmonic-mean score is a reasonable starting point for tuning sparsity, but it weights sparsity and performance equally; a deployment that values latency or energy differently could choose a different threshold and get a different 'optimal' balance.","The sparsification concept naturally extends to time-varying channels: one could test whether the thresholds need to be re-learned per coherence block or can be adapted incrementally, which would matter for real-time resource allocation."],"forward_implications":["In a cell-free massive MIMO system, replacing the dense MDGNN with SP-MDGNN at the reported thresholds roughly halves per-inference computation time while keeping mean spectral efficiency within about 2% of the dense baseline.","The sparsification recipe transfers to power control and precoding separately, cutting complexity by 55% and 49% with SE losses of only 1.3% and 1.44%, respectively.","Because the pruned network updates only the parameters of surviving connections, training converges faster and needs fewer cycles, which makes the method more practical for large-scale deployments.","Against the classical WMMSE algorithm, SP-MDGNN trades a small amount of spectral efficiency for a much lower computational footprint, positioning GNN-based resource management as a deployable alternative for large antenna counts."],"supporting_citations":[{"why":"Defines the MDGNN architecture that SP-MDGNN sparsifies and serves as the non-sparse performance baseline.","marker":"[12]"},{"why":"Presents the WMMSE-based processing method used as the classical optimization benchmark in the SE and complexity comparisons.","marker":"[4]"},{"why":"Introduces sparse graph neural networks with l0-norm regularization for resource allocation, the prior work whose different sparsification approach SP-MDGNN builds on and contrasts with.","marker":"[14]"},{"why":"Shows GNN-based power control in partially connected cell-free massive MIMO, providing the graph-modeling context for the power control pathway.","marker":"[9]"},{"why":"Documents the computational bottleneck of large-scale cell-free massive MIMO, motivating the complexity reduction that SP-MDGNN targets.","marker":"[13]"}],"fun_headline_variants":["Sparse GNN prunes AP-UE links, halves cell-free MIMO compute","Cell-free MIMO compute down 48% with sparse graph network","Pruned graph neural net retains 97.9% SE in cell-free MIMO","Sparse MDGNN: 48% less compute, near-ideal spectral efficiency","Cell-free MIMO: sparse graph cuts computation while keeping SE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The thresholds that decide which links to prune are selected by maximizing the harmonic-mean score on the same evaluation data that is later used to compute the reported spectral efficiency and complexity numbers, so the claimed balance has not been shown to hold on data that was not used for model selection.","fun_headline_variants_meta":{"raw":{"variants":["Sparse GNN prunes AP-UE links, halves cell-free MIMO compute","Cell-free MIMO compute down 48% with sparse graph network","Pruned graph neural net retains 97.9% SE in cell-free MIMO","Sparse MDGNN: 48% less compute, near-ideal spectral efficiency","Cell-free MIMO: sparse graph cuts computation while keeping SE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000934,"raw_usage":{"total_tokens":3979,"prompt_tokens":913,"completion_tokens":3066,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":2964}},"tokens_in":529,"tokens_out":3066,"duration_ms":25152,"temperature":1.0,"reasoning_tokens":2964,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:40:37.116721+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same joint power-control and precoding experiment but choose the thresholds on a validation set that is completely disjoint from the test set, then measure mean SE and computation time on the held-out test data. If the 48% complexity reduction or the 2.11% SE loss changes noticeably, or if the harmonic-mean score peaks at a different threshold, the reported optimal balance is an artifact of selection on the test set.","supporting_citations":[{"cited_title":"Multidimensional graph neural networks for wireless communications,","cited_arxiv_id":null,"evidence_quote":"Defines the MDGNN architecture that SP-MDGNN sparsifies and serves as the non-sparse performance baseline."},{"cited_title":"WMMSE- based processing in cell-free massive MIMO systems,","cited_arxiv_id":null,"evidence_quote":"Presents the WMMSE-based processing method used as the classical optimization benchmark in the SE and complexity comparisons."},{"cited_title":"Sparse graph neural networks for two-timescale wireless resource allocation,","cited_arxiv_id":null,"evidence_quote":"Introduces sparse graph neural networks with l0-norm regularization for resource allocation, the prior work whose different sparsification approach SP-MDGNN builds on and contrasts with."},{"cited_title":"Graph neural network aided power control in partially connected cell-free massive MIMO,","cited_arxiv_id":null,"evidence_quote":"Shows GNN-based power control in partially connected cell-free massive MIMO, providing the graph-modeling context for the power control pathway."},{"cited_title":"Double-layer power control for mobile cell-free XL-MIMO with multi-agent reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Documents the computational bottleneck of large-scale cell-free massive MIMO, motivating the complexity reduction that SP-MDGNN targets."}],"review_version":1}