{"id":"d12d783f-6847-4864-bef0-5e3078a397b0","arxiv_id":"2606.17543","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"DNN-based optimization of movable antenna positions from partial CSI achieves near-optimal performance in single-user and outperforms conventional methods in multi-user scenarios.","lead":"The paper proposes a deep neural network framework to optimize the positions of movable antennas in multi-user MISO systems using only partial channel power measurements, avoiding the need for full channel state information. This approach could significantly reduce channel estimation overhead in wireless communications.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"DNN capture of nonlinear mapping from partial power gains to optimal MA positions, trained only on simulated data, lacks verification for practical channels.","rationale":"The reader's weakest assumption directly identifies the same load-bearing empirical-mapping step; the abstract supplies no counter-evidence that would resolve it, so the UNVERDICTED status is unchanged.","tokens_in":1724,"tokens_out":325,"duration_ms":32143,"concrete_test":"Retrain both the single-user and multi-user DNNs on a mismatched channel model (e.g., replace the original model with Rician fading having K=5 while keeping the same partial-measurement subset and architecture); recompute the reported sum-rate gaps to the alternating optimizer on 500 new test realizations—if the gap shrinks by more than 15 % or becomes negative, the simulation-to-practice transfer is unreliable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that a DNN (supervised for single-user, unsupervised sum-rate maximizing for multi-user) can learn the mapping from a subset of location power gains to good MA positions. The abstract states this mapping is too complex to characterize analytically for practical models, so performance rests entirely on empirical learning from simulated channels. No analytical bound or identifiability argument is supplied, and the multi-user case adds the further requirement that the attention architecture can implicitly resolve interference without global labels. If the learned function fails to generalize when the true propagation deviates from the training distribution, both the near-optimal single-user claim and the outperformance versus full-CSI alternating optimization become unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes a DNN framework to optimize positions of multiple transmit movable antennas (MAs) in a multi-user MISO system using only partial channel power gain measurements from a subset of locations, thereby avoiding full CSI acquisition. In the single-user case, a supervised DNN is trained to learn the nonlinear mapping from these partial gains to near-optimal MA positions. In the multi-user case, an unsupervised attention-based DNN is trained by directly maximizing the sum-rate objective. Simulation results are reported to show near-optimal single-user performance and outperformance versus conventional full-CSI alternating optimization in the multi-user setting.","tokens_in":1869,"tokens_out":548,"duration_ms":20679,"significance":"If the empirical mappings generalize beyond the simulated training distributions, the work would meaningfully reduce the prohibitive channel estimation overhead that currently limits practical deployment of movable-antenna systems. The unsupervised multi-user formulation is a notable technical contribution because it sidesteps the need for globally optimal position labels.","major_comments":[{"comment":"The central claim that a DNN can reliably capture the mapping from partial power gains to optimal MA positions rests entirely on empirical performance for the specific channel models used in training; no analytical identifiability result, approximation bound, or robustness analysis against model mismatch is supplied (single-user analysis and multi-user training sections).","section":"single-user analysis and multi-user training sections"},{"comment":"The multi-user claim that the attention-based unsupervised network surpasses CSI-based alternating optimization is load-bearing for the paper's contribution, yet the manuscript provides no details on Monte-Carlo trial count, error bars, or statistical testing of the reported sum-rate gains (simulation results section).","section":"simulation results section"},{"comment":"The unsupervised loss directly maximizes the instantaneous sum-rate expression, but the paper does not address how the attention mechanism implicitly resolves inter-user interference when only partial measurements are available, nor does it compare against stronger baselines that also operate on partial CSI.","section":"multi-user framework"}],"minor_comments":[{"comment":"Notation for the partial measurement set and the attention-module input dimension should be introduced earlier and used consistently.","section":"system model"},{"comment":"Figure captions for the simulation plots should explicitly state the number of independent channel realizations and the exact channel model parameters.","section":"simulation results"}],"recommendation":"major_revision","confidential_remarks":"The work is squarely within the scope of eess.SP. The citation list appears balanced; no obvious self-citation inflation is visible from the abstract alone."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback. We address each major comment below and indicate planned revisions to strengthen the manuscript.","responses":[{"response":"We acknowledge that the proposed DNN approach is fundamentally empirical, as deriving closed-form identifiability results or approximation bounds for general channel models is intractable due to the highly nonlinear mapping involved. In the revision, we will add a dedicated limitations subsection discussing this aspect and include new simulation results evaluating robustness under channel model mismatch (e.g., different path-loss exponents and Rician factors).","revision_made":"partial","referee_comment":"[single-user analysis and multi-user training sections] The central claim that a DNN can reliably capture the mapping from partial power gains to optimal MA positions rests entirely on empirical performance for the specific channel models used in training; no analytical identifiability result, approximation bound, or robustness analysis against model mismatch is supplied (single-user analysis and multi-user training sections)."},{"response":"We agree that additional statistical details are needed. The revised manuscript will explicitly state that results are averaged over 1000 independent Monte-Carlo trials, include error bars showing one standard deviation, and note that the observed sum-rate improvements are consistent across trials.","revision_made":"yes","referee_comment":"[simulation results section] The multi-user claim that the attention-based unsupervised network surpasses CSI-based alternating optimization is load-bearing for the paper's contribution, yet the manuscript provides no details on Monte-Carlo trial count, error bars, or statistical testing of the reported sum-rate gains (simulation results section)."},{"response":"The attention layers are intended to learn cross-user feature correlations from the partial power gains that implicitly account for interference; we will expand the multi-user framework section with a qualitative explanation of this mechanism supported by attention weight visualizations. We will also add comparisons against partial-CSI baselines such as position selection via greedy search on measured gains and a simple DNN regressor trained on partial data.","revision_made":"partial","referee_comment":"[multi-user framework] The unsupervised loss directly maximizes the instantaneous sum-rate expression, but the paper does not address how the attention mechanism implicitly resolves inter-user interference when only partial measurements are available, nor does it compare against stronger baselines that also operate on partial CSI."}],"tokens_in":1453,"tokens_out":500,"duration_ms":22158,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper shows a DNN can optimize movable antenna positions using only partial channel power gains from a few locations, trained with supervised learning for one user and unsupervised sum-rate maximization for multiple users.\n\nThe new part is applying this to the multi-user case with an attention architecture that handles interference without needing optimal position labels. That unsupervised training directly on the objective is a practical choice when labels are unavailable.\n\nIt does a good job identifying the channel estimation burden for MAs and offering a learning-based workaround that runs in real time after training. The simulations indicate it gets near optimal for single user and outperforms alternating optimization in multi-user settings.\n\nSoft spots center on the simulation-only validation. Training and testing use the same channel models, so there's no check against distribution shift or real measurements. The paper acknowledges the mapping is too nonlinear to derive analytically, which means the claims depend entirely on empirical performance in the chosen scenarios. Without seeing variance across runs or how sensitive the results are to model parameters, it's hard to gauge robustness. The multi-user gains versus full-CSI methods could also depend on implementation details of the baselines.\n\nThis is for researchers in wireless systems design, particularly those exploring movable antennas or machine learning for physical layer optimization. A reader focused on practical MA deployment would find the partial CSI angle relevant.\n\nThe paper shows clear thinking on the problem setup and literature, so it deserves peer review. I would send it out for referees to check the simulation details and generalization potential.","headline":"The paper shows a DNN can optimize movable antenna positions using only partial channel power gains from a few locations, trained with supervised learning for one user and unsupervised sum-rate maximization for multiple users.","tokens_in":2316,"tokens_out":382,"would_cite":false,"duration_ms":28935,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A deep neural network learns to set optimal movable-antenna positions from partial power measurements alone.","keywords":["movable antennas","deep neural network","position optimization","partial CSI","multi-user MISO","sum-rate maximization","unsupervised learning"],"falsifier":"Compare the sum-rate achieved by the DNN positions against the sum-rate of positions found by exhaustive search over the full movement region with complete CSI; a large persistent gap falsifies the claim.","tokens_in":2651,"feed_emoji":"📡","tokens_out":618,"duration_ms":20856,"temperature":0.7,"pith_summary":"The paper establishes that a DNN can capture the nonlinear relationship between optimal movable-antenna locations and channel power gains measured at only a few points inside the movement region. Supervised training works for the single-user case; an unsupervised attention-based network directly maximizes multi-user sum-rate when no globally optimal labels exist. Both versions run in real time and avoid estimating channels at every possible antenna coordinate. The result lowers the overhead that has blocked practical use of movable antennas in multi-user MISO systems.","feed_headline":"DNN sets movable-antenna positions from partial power measurements","feed_subtitle":"Single-user performance nears the optimum; multi-user rates beat full-CSI alternating optimization while using far fewer channel samples.","key_machinery":"Attention-based deep neural network that maps a vector of partial power measurements to a set of movable-antenna coordinates.","core_discovery":"A DNN trained on partial channel power gains predicts the positions of multiple transmit movable antennas that maximize sum-rate; in single-user settings the network is trained by supervised learning to match known optima, while in multi-user settings an unsupervised attention architecture directly optimizes the rate expression without requiring globally optimal position labels.","pith_inferences":["The overhead of channel estimation in movable-antenna systems drops from scaling with the number of candidate positions to scaling only with the number of measurement points.","Similar partial-observation learning could apply to other high-dimensional antenna reconfiguration problems where exhaustive search is infeasible.","Transfer from simulation to real channels remains untested and would be the next concrete experiment.","If the learned mapping proves robust, it suggests many wireless resource-allocation tasks with prohibitive measurement costs can be solved by neural networks trained on subsets of observations."],"forward_implications":["Single-user systems reach performance within a small gap of the full-CSI optimum.","Multi-user systems exceed the rates of conventional alternating-optimization methods that rely on full CSI.","Position decisions are obtained from a small number of power measurements without estimating channels everywhere.","The same network can be deployed in real time once trained offline."],"fun_headline_variants":["DNN learns MA positions from partial power measurements","Partial CSI trains DNN for movable antenna optimization","Attention DNN maximizes rates with incomplete channel data","Movable antennas positioned without full CSI via DNN","Unsupervised DNN optimizes multi-user MA placement"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The relationship between optimal antenna positions and power gains at a small subset of locations can be learned by a neural network trained only on simulated data.","fun_headline_variants_meta":{"raw":{"variants":["DNN learns MA positions from partial power measurements","Partial CSI trains DNN for movable antenna optimization","Attention DNN maximizes rates with incomplete channel data","Movable antennas positioned without full CSI via DNN","Unsupervised DNN optimizes multi-user MA placement"]},"model":"grok-4.3","cost_usd":0.003397,"raw_usage":{"total_tokens":1814,"prompt_tokens":695,"num_sources_used":0,"completion_tokens":66,"cost_in_usd_ticks":33974500,"prompt_tokens_details":{"text_tokens":695,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1053,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":695,"tokens_out":66,"duration_ms":9204,"temperature":1.0,"reasoning_tokens":1053,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T23:25:38.279025+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Compare the sum-rate achieved by the DNN positions against the sum-rate of positions found by exhaustive search over the full movement region with complete CSI; a large persistent gap falsifies the claim.","supporting_citations":[],"review_version":1}