{"id":"3e394ee7-6838-4530-9bdd-2055d56f3741","arxiv_id":"2505.21343","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GNN-augmented AMP and expectation propagation detectors approach ML BER in simulated compressed-sensing space-frequency index modulated multi-user MIMO uplinks, while tolerating varying numbers of active users.","lead":"Compressed sensing, index modulation, and graph neural network detectors are combined for a multi-user MIMO uplink, and simulations report that the learned detectors approach maximum-likelihood bit error rates at lower complexity. The result matters because a detector trained once and usable across different user counts would simplify practical large-scale MIMO receivers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'train once for any number of users' claim is unsupported: the promised permutation-equivariance proof is absent, and the Fig. 11 experiment trains on a mixture of all user counts, so its evaluation is interpolation rather than generalization.","rationale":"The reader's conditional verdict is already sensitive to the train-once issue; I focus on it rather than the perfect-CSI caveat, because the CSI assumption is explicitly scoped in footnotes 3 and 4 and is a limitation rather than an internal gap. The train-once claim is central to the paper's advertised advantage over DNN detectors, and the manuscript itself asserts a proof that it does not provide. The Section IV training description makes the evaluation circular: training on all Ka values and then reporting success over those same Ka values cannot establish generalization to unseen user counts. Permutation equivariance is a known property of sum-aggregation GNNs and may help, but the paper does not connect that property to cross-size transfer, and the EP/AMP observation modules depend on problem size through matrix inverses. I would keep the CONDITIONAL verdict: the performance comparisons in Figs. 9-10 are plausible if the implementation matches the text, but the headline 'train once for any number of users' requires either a proof or an out-of-sample experiment.","tokens_in":26646,"tokens_out":9752,"duration_ms":116263,"concrete_test":"Retrain GNN-AMP and GEPNet from scratch using only Ka=1 samples (equivalently, only Umax=4) and evaluate at Ka=8 and Ka=16 with Umax=16. If the BER at the 10^-2 threshold degrades substantially relative to the mixture-trained curves in Fig. 11, the train-once/any-number-of-users claim is falsified. If the curves are unchanged, the claim survives this test, and the missing permutation proof should still be supplied.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the Abstract and Section I contribution 4 includes a single GNN training that works for any number of users. Contribution 4 says this is 'verified by mathematical proof that the proposed detectors share GNN's permutation equivariance property,' but no proof appears in Sections II-IV. The only supporting experiment, Section IV and Fig. 11, states that all GNN-aided methods 'are trained by a scheme that applies a mixture of a single active user to the maximum number of users to see the full range.' Training on Ka=1,...,Umax and testing on the same Ka range is interpolation, not cross-user generalization. Permutation equivariance of the graph processor is not by itself sufficient: it concerns relabeling nodes within one graph, whereas going from Umax=4 to Umax=16 changes the graph size and the dimensions of the EP/AMP linear modules. Unless the promised proof also covers addition or removal of users, or unless an unseen user count is tested, the train-once guarantee does not follow. The reported 'robustness to changes in the number of users' is therefore based on an omitted proof and a circular evaluation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes CS-SFIM for the large-scale multiuser MIMO uplink, in which each user's information bits separately select active transmit antennas and active subcarriers and are also mapped to APM symbols. The BS detection problem is represented as a factor graph, and the authors develop AMP and EP detectors and then two GNN-aided variants, GNN-AMP and GEPNet, together with a GNN-MMSE baseline. The paper claims that the GNN-aided detectors approach ML performance at reduced complexity and that a single GNN training suffices for systems with a variable number of users, verified by an alleged permutation-equivariance proof. The evidence is BER simulations over Rayleigh channels for U=4 and U=16 users, plus complexity comparisons and a robustness experiment across active user counts.","tokens_in":26923,"tokens_out":9039,"duration_ms":97913,"significance":"If the performance claims hold, the paper makes a useful contribution to index-modulation-based massive MIMO uplink detection: it extends compressed-sensing-aided multidimensional IM to a multiuser setting, formulates the factor graph for the resulting sparse signal, and demonstrates that unrolling AMP/EP with GNN modules can give large BER gains over the classical counterparts. The ML benchmark is an independent upper bound rather than a training target for the GNN, so the overall comparison is not circular. The complexity table and the evaluation across different numbers of active users are informative. However, the headline 'train once for any number of users' claim is supported neither by the promised proof nor by the presented experiment, and the EP/GEPNet derivation contains a mean/variance labeling inconsistency; these issues must be resolved before the central claims can be accepted.","major_comments":[{"comment":"Equations (46) and (47) assign the two parameters of the cavity distribution in (45) in an internally inconsistent way. For the quotient N(mu_i,Sigma_ii)/N(eta_i/V_i,1/V_i), the standard Gaussian identity gives variance v_o,i = Sigma_ii/(1 - Sigma_ii*V_i) and mean m_o,i = v_o,i*(mu_i/Sigma_ii - eta_i). The text instead labels the variance expression as m_o,i in (46) and the mean expression as v_o,i in (47). These labels feed Algorithm 2 (line 4), Algorithm 4 (line 4), and the GEPNet node attribute in (69), so an implementation following the text would use the wrong cavity statistics. The authors should either swap the labels/expressions or explicitly define a nonstandard convention and use it consistently; as written, the GEPNet derivation is inconsistent and the near-ML GEPNet results cannot be checked against the equations.","section":"Section III-A2, Eqs. (46)-(47)"},{"comment":"Contribution 4 in Section I claims that a single GNN training works for any number of users and states that this is verified by a mathematical proof of permutation equivariance, but no such proof appears in Sections II-IV. The only supporting experiment, Fig. 11, explicitly trains all GNN-aided methods on a mixture of one active user up to the maximum number of users and then evaluates the same range of Ka; this is interpolation over the training distribution, not a test of generalization to an unseen number of users. In addition, permutation equivariance of the GNN message-passing module concerns relabeling nodes inside a fixed graph, whereas changing Umax changes the graph size, the dimensions of Hg, and the EP/AMP linear modules in (42)-(43) and Algorithm 1. To support the train-once claim, the authors should either provide a proof that covers insertion/removal of users, not merely node permutations, or test on unseen Umax/Ka settings, for example training at Umax=4 and testing at Umax=8 and 16.","section":"Section I, contribution 4; Section IV, Fig. 11"},{"comment":"Section IV reports BER curves for the GNN and DNN detectors but does not state the training setup: number of training samples, SNR values or distribution used for training, optimizer, learning rate, number of epochs, train/validation split, or the DNN architecture and its label generation beyond 'trained by the data of ML detection'. Because the central claims are established purely by simulation, these details are needed to assess sensitivity and reproducibility. Please add a training-configuration table and, if possible, a statement on code/data availability.","section":"Section IV"}],"minor_comments":[{"comment":"The rate expression contains a stray ']' and unbalanced parentheses ('K log2 C(Nt, Na) +] + NaK log2 L'); it should be rewritten.","section":"Section II-A3, Eq. (2)"},{"comment":"The ML objective uses uppercase Yg while the received signal is defined as yg in Eq. (7); use consistent notation.","section":"Section III-A1, Eq. (15)"},{"comment":"The model introduces a per-user noise w_i^u and then defines w_i = sum_u w_i^u in (5), which inflates the noise variance by U unless the individual variances are scaled; this is inconsistent with the conventional per-antenna noise model later used in (6). Please clarify the noise model.","section":"Section II-A, Eqs. (3)-(5)"},{"comment":"The scheme description lists 64 RAs, but the accompanying text says 'employ a 16x16 MIMO size' and Fig. 12 is described in terms of U=4; please resolve this inconsistency.","section":"Section IV, Scheme 5"},{"comment":"The paper uses 'AMP-GNN' and 'GNN-AMP' interchangeably (e.g., contribution 3, Section III-C3, Fig. 7, and Table III); pick one consistent name.","section":"Notation throughout"},{"comment":"The claimed DNN complexity O(Nh1 Nh2) is missing the dependence on the input/output dimensions and the number of layers; specify the complexity in terms of the architecture actually used.","section":"Section III-D and Table III"}],"recommendation":"major_revision","confidential_remarks":"The main technical blocker is the missing equivariance proof and the interpolation-only experiment for the train-once claim; if the authors reframe the claim as robustness across the trained user range or provide genuine cross-user evaluation, the paper could be suitable. I would not reject on the basis of the ML-style training approach; the issue is internal consistency of the equations and the gap between claim and evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid incremental paper, not a breakthrough. The new CS-SFIM variant (separate spatial and frequency index activation) is genuinely new relative to [27], and applying GNN-AMP and GEPNet to it gives consistent gains over EP/AMP/DNN in simulation. But the headline 'train once for any user count' claim is not supported as written: the promised permutation-equivariance proof does not appear, and the Fig. 11 experiment trains on a mixture of all Ka values, so it only demonstrates interpolation within the trained range, not generalization to unseen user counts.\n\nWhat's good: The system model is carefully laid out, the factor graph and detector derivations are mostly clear, and the simulation set covers both 4- and 16-user cases and variable active users. The ML benchmark is independent, so there's no circular fitting. The complexity tables, while approximate, give a useful comparison. For the MIMO detection and index modulation community, the paper provides a workable recipe and confirms that GNN-aided MP detectors transfer to a new sparse index-modulation uplink.\n\nWhere it's soft: The missing proof is the main issue. Permutation equivariance of a graph processor concerns node relabeling within a fixed graph; it doesn't automatically cover changing graph size or the dimensions of the EP/AMP linear modules. Unless the authors can show the entire detector is equivariant across user counts, or test on an unseen Ka, the abstract's claim should be downgraded to 'trained on a mixture of user counts.' Second, equations (46) and (47) swap the mean and variance of the EP cavity distribution; as written, m_o gets the variance expression and v_o gets the mean. This looks like a typo rather than a fundamental error, but it needs fixing because GEPNet uses those as node attributes. Third, the text around Fig. 9(a) has an off-by-one labeling error: the paragraph describing 'Scheme 1d)-1f)' actually refers to schemes 1e)-1g). Finally, there's no code or training hyperparameters, so the results can't be independently reproduced from the manuscript.\n\nOverall: the central BER results are plausible and consistent with prior work. The paper deserves a serious referee but needs major revision to either supply the missing generalization argument or soften the claim, fix the equation typo, and add reproducibility details.","headline":"Solid incremental application of GNN-aided MP detectors to a new CS-SFIM variant; the 'train once for any user count' claim is unsupported as written.","tokens_in":27441,"tokens_out":3234,"would_cite":true,"duration_ms":33973,"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":"Graph neural networks fused with message-passing detectors (GNN-AMP and GEPNet) approach the maximum-likelihood performance on a compressed sensing-aided space-frequency index modulation uplink, while a single GNN training suffices for…","keywords":["Index Modulation","Multi-User MIMO","Graph Neural Network","Message Passing","Approximate Message Passing","Expectation Propagation","Compressed Sensing","Space-Frequency Index Modulation"],"falsifier":"Measure BER of GNN-AMP and GEPNet with imperfect channel estimates (e.g., 10% normalized MSE) or with a number of active users outside the training range, and compare against ML; a gap that grows by more than a few dB, or a broken equivariance under user permutation, would refute the near-ML and train-once claims.","tokens_in":26446,"feed_emoji":"📡","tokens_out":6741,"duration_ms":68627,"temperature":0.7,"pith_summary":"This paper claims that adding a graph neural network module to two message-passing detectors, AMP and EP, yields near-optimal detection for a compressed sensing-aided space-frequency index modulation (CS-SFIM) uplink with many users. The proposed GNN-AMP and GEPNet detectors are shown to outperform conventional MP and GNN detectors and to come within a few tenths of a dB of the maximum-likelihood (ML) bound, while avoiding ML's exponential complexity. The paper further claims that a single GNN training suffices across different numbers of active users, because the underlying detector is permutation-equivariant. If true, this would give large-scale multi-user MIMO a low-complexity, near-optimal receiver that does not need retraining as user load changes.","feed_headline":"GNN-aided detectors approach ML performance on MU-MIMO uplink","feed_subtitle":"Graph neural networks wrapped around AMP and EP cut complexity while keeping near-optimal BER across user counts.","key_machinery":"The central object is the joint factor graph of the space-frequency domain, where variable nodes are the elements of the equivalent sparse transmit vector x and factor nodes are the received observations across receive antennas and subcarriers. The load-bearing mechanism is the GNN module that replaces the Gaussian approximation of the cavity distribution in AMP and EP: node attributes are the MP-computed mean and variance, edge attributes are the channel cross-correlations and noise variance, and an MLP-plus-GRU message-passing scheme iteratively refines the posterior. A claimed permutation equivariance of this message-passing graph is what licenses the single-training/many-users property.","core_discovery":"The paper designs a CS-SFIM transceiver in which each user maps bits to subcarrier activation indices and antenna activation indices separately, forming a sparse space-frequency matrix per subcarrier group. Treating the entire multi-user SF matrix as a joint factor graph, the authors build AMP and EP detectors, then interleave a graph neural network into each iteration: the GNN takes the cavity parameters (mean, variance) of the MP algorithm as node attributes and refines the posterior distribution that the MP module uses. The resulting GEPNet detector reaches a BER of $10^{-4}$ at an SNR within about 0.3-0.35 dB of ML in the 4-user configurations, and both GNN-AMP and GEPNet remain the best-performing practical detectors in the 16-user configurations. The train-once claim is supported by the permutation equivariance of the GNN message-passing architecture, which lets the same trained weights act on any subset of users.","pith_inferences":["The permutation-equivariance argument, made rigorous, would imply the learned message-passing weights depend on channel statistics rather than on user identity; a direct stress test is to apply the detector to a user count outside the training range.","Treating channel estimation error as part of the edge attributes and training on noisy channel realizations could extend the near-ML gains to imperfect CSI; the paper does not explore this.","Because the GNN module refines posterior distributions rather than hard decisions, it could be coupled with channel coding for joint detection and decoding, as the conclusion itself suggests.","The complexity numbers in the paper count multiplications only; a fairer systems comparison would include memory, latency, and training cost for the GNN modules."],"forward_implications":["GEPNet achieves BER within about 0.3 dB of the maximum-likelihood detector in the 4-user configurations, while GNN-AMP remains within about 0.3-0.35 dB with lower complexity.","Both GNN-aided detectors outperform the conventional EP, AMP, GNN-MMSE, and DNN detectors by several dB in all four tested configurations, with the gap widening as user count grows.","A single GNN training, performed on a mixture of active-user counts, suffices for any number of active users from 1 to the maximum supported, without retraining per user count.","Increasing the number of receive antennas improves all GNN-aided detectors, and the near-ML performance is maintained in the large-scale 16-user, 64-128 receive-antenna uplink.","The detector family tolerates user activity variation, degrading gracefully as the number of active users grows."],"supporting_citations":[{"why":"Proposed CS-SFIM for the large-scale multi-user MIMO uplink; the scheme this paper extends to GNN-aided detection.","marker":"[27]"},{"why":"Introduces the approximate message passing (AMP) algorithm that the paper's AMP and GNN-AMP detectors are built on.","marker":"[28]"},{"why":"Provides the orthogonal AMP (OAMP) detector that integrates AMP with MMSE filtering, a baseline and component for GNN-AMP.","marker":"[33]"},{"why":"Establishes expectation propagation (EP) detection for MIMO systems, the basis of the GEPNet detector.","marker":"[34]"},{"why":"Supplies the EP detector formulation with cavity distributions and moment matching used in GEPNet.","marker":"[35]"},{"why":"Presents the pairwise Markov random field and GNN framework for massive MIMO detection that the paper adapts.","marker":"[39]"},{"why":"Proposes an AMP-based GNN model for MU-OFDM with near-ML performance, direct predecessor of GNN-AMP.","marker":"[41]"},{"why":"Introduces the EP-based GNN (GEPNet) detector for MU-MIMO that the paper extends to CS-SFIM.","marker":"[42]"},{"why":"Develops GNN-aided MU-MIMO detectors and shows the performance-complexity trade-offs the paper compares against.","marker":"[43]"}],"fun_headline_variants":["GNN-AMP and GEPNet cut complexity while nearing ML BER","Train-once GNNs serve any user count in MU-MIMO","Graph neural nets aid MU-MIMO detection near ML performance","GNN-Aided detectors hit near-ML BER in MU-MIMO uplink","GNNs wrapped in AMP/EP near ML for multi-user MIMO"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results assume perfect channel knowledge and perfect timing at the base station, and the claim that one training works for any number of users rests on a symmetry property that the paper states but does not fully prove.","fun_headline_variants_meta":{"raw":{"variants":["GNN-AMP and GEPNet cut complexity while nearing ML BER","Train-once GNNs serve any user count in MU-MIMO","Graph neural nets aid MU-MIMO detection near ML performance","GNN-Aided detectors hit near-ML BER in MU-MIMO uplink","GNNs wrapped in AMP/EP near ML for multi-user MIMO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000921,"raw_usage":{"total_tokens":4019,"prompt_tokens":1081,"completion_tokens":2938,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":697,"completion_tokens_details":{"reasoning_tokens":2842}},"tokens_in":697,"tokens_out":2938,"duration_ms":22684,"temperature":1.0,"reasoning_tokens":2842,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:29:13.739334+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure BER of GNN-AMP and GEPNet with imperfect channel estimates (e.g., 10% normalized MSE) or with a number of active users outside the training range, and compare against ML; a gap that grows by more than a few dB, or a broken equivariance under user permutation, would refute the near-ML and train-once claims.","supporting_citations":[{"cited_title":"Two-Dimensional Index Modulation for the Large-Scale Multi-User MIMO Uplink,","cited_arxiv_id":null,"evidence_quote":"Proposed CS-SFIM for the large-scale multi-user MIMO uplink; the scheme this paper extends to GNN-aided detection."},{"cited_title":"Message Passing Algo- rithms for Compressed Sensing,","cited_arxiv_id":null,"evidence_quote":"Introduces the approximate message passing (AMP) algorithm that the paper's AMP and GNN-AMP detectors are built on."},{"cited_title":"Orthogonal AMP,","cited_arxiv_id":null,"evidence_quote":"Provides the orthogonal AMP (OAMP) detector that integrates AMP with MMSE filtering, a baseline and component for GNN-AMP."},{"cited_title":"Expectation Propagation Detection for High-Order High-Dimensional MIMO Systems,","cited_arxiv_id":null,"evidence_quote":"Establishes expectation propagation (EP) detection for MIMO systems, the basis of the GEPNet detector."},{"cited_title":"Expecta- tion Propagation Detector for Extra-Large Scale Massive MIMO,","cited_arxiv_id":null,"evidence_quote":"Supplies the EP detector formulation with cavity distributions and moment matching used in GEPNet."},{"cited_title":"GNN-Enhanced Approximate Message Passing for Massive/Ultra-Massive MIMO Detection,","cited_arxiv_id":null,"evidence_quote":"Proposes an AMP-based GNN model for MU-OFDM with near-ML performance, direct predecessor of GNN-AMP."},{"cited_title":"Graph Neural Network Aided Expectation Propagation Detector for MU-MIMO Systems,","cited_arxiv_id":null,"evidence_quote":"Introduces the EP-based GNN (GEPNet) detector for MU-MIMO that the paper extends to CS-SFIM."},{"cited_title":"Graph Neural Network Aided MU-MIMO Detectors,","cited_arxiv_id":null,"evidence_quote":"Develops GNN-aided MU-MIMO detectors and shows the performance-complexity trade-offs the paper compares against."}],"review_version":1}