{"id":"e340c377-84dd-448b-89c2-25df4dd21e45","arxiv_id":"2502.01438","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A staged graph neural network jointly learns pinching-antenna positions and transmit beamforming, and the authors report it outperforms a heuristic baseline in simulated pinching-antenna systems.","lead":"This paper proposes GPASS, a graph neural network that places pinching antennas and computes transmit beamforming in pinching-antenna wireless systems. The authors report that it beats a simple closest-user zero-forcing baseline in simulated spectral efficiency and runs in a few milliseconds.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The nested PA-permutation equivariance in (8) is likely false: constraint (5c) makes the feasible set label-order-dependent, so PBF-sub-GNN's core design premise is invalid.","rationale":"The reader's weakest_assumption focused on Proposition 1 being stated without proof. My concern is more fundamental: the property that Proposition 1 is meant to realize, namely nested permutation equivariance over pinching antennas, does not actually hold for problem P. Constraint (5c) makes the labels m order-dependent, so the feasible set is not invariant under arbitrary PA permutations. Adding s as an input feature does not alter the optimization problem or the constraints, so it cannot create an equivariance that the problem lacks. This means PBF-sub-GNN is built on a false symmetry premise, and the architecture's claimed benefit of leveraging permutation properties is not justified. I partial-agree with the reader because the reader identified the proof gap but not the incorrectness of the underlying equivariance claim. The baseline comparison issue is also real, but the equivariance flaw is more central to the paper's novel contribution. If the concrete check confirms the infeasibility under PA permutation, the paper needs a major redesign or a substantially revised claim; hence a REJECT verdict for the current version is appropriate. The numerical SE claim is not necessarily false, but the theoretical foundation of the proposed GPASS architecture is.","tokens_in":8618,"tokens_out":15309,"duration_ms":151866,"concrete_test":"Analytical check: For N=1, M=2, Δmin>0, take any feasible pair (x1,x2) with x2−x1≥Δmin. The PA transposition maps it to (x2,x1), and the transformed Δ'=x1−x2≤−Δmin, violating (5c). Thus the feasible set is not closed under the PA permutation used in (8). To make this fully computational, solve problem P for a single user and compare the optimal labeled positions with the permuted-label problem where Δ is imposed on the permuted order; if the transformed optimal A depends on the order s, then eq. (8) fails.","verdict_should_be":"REJECT","load_bearing_attack":"Property (8) is the load-bearing premise of PBF-sub-GNN. It asserts that the pinching-beamforming policy is equivariant to arbitrary permutations of pinching antennas within each waveguide. This is not a property of problem P. Constraint (5c) (equivalently (6b)) defines Δp_{m,n}=xp_{m,n}-xp_{m-1,n}≥Δmin, imposing a monotone spatial order on the label m. For a feasible labeled configuration, e.g., N=1, M=2, x1=0, x2=2Δmin, the transposition x'1=x2, x'2=x1 gives Δ'=-2Δmin, violating (5c). The feasible set is therefore not invariant under PA-label permutations, so the mapping cannot satisfy the Ω_P equivariance in (8) for nontrivial permutations. Adding the index s as an input does not change problem P, which is independent of s; \\hat F1(Φ,s) is the same policy as \\tilde F1(Φ), which explicitly depends on PA order. Consequently, the 2D-PE update rule and Proposition 1 are applied to a symmetry that the problem does not possess. The omission of the proof is not the central issue: the stated equivariance itself is unsupported and likely false. This undermines the theoretical justification of the architecture, even though the raw SE numbers could in principle still be reproduced.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GPASS, a two-stage graph neural network for jointly learning pinching beamforming (antenna positions) and transmit beamforming in pinching-antenna systems. The first sub-GNN (PBF-sub-GNN) learns the antenna positions with an edge-update architecture that is claimed to be equivariant to nested permutations of pinching antennas, waveguides, and users. The second sub-GNN (TBF-sub-GNN) learns only the uplink and downlink power allocations and recovers the transmit beamforming matrix using the optimal solution structure in [13]. The network is trained unsupervised with the negative spectral efficiency as the loss. Numerical results compare the SE of GPASS against a heuristic baseline for M=1 and M=3 and report an average inference time of 6 ms.","tokens_in":8927,"tokens_out":7336,"duration_ms":59209,"significance":"If the claims were correct, this would be the first deep-learning approach for pinching-antenna systems and could provide a low-complexity real-time beamforming solution. The staged architecture and the use of the optimal beamforming structure are sensible engineering choices, and the unsupervised training setup is standard in learning-to-optimize works. However, the theoretical foundation of the PBF-sub-GNN rests on a permutation equivariance property that is likely false, and the numerical comparison is undermined by the baseline being defined only for M=1. The paper also overclaims an FNN comparison that is not present in the results. These issues make the central contribution not defensible in its current form.","major_comments":[{"comment":"The nested permutation equivariance property (8) is invalid because the feasible set of problem P is not invariant under permutations of pinching-antenna labels within a waveguide. Constraint (6b), equivalently (5c), defines Δp_{m,n} = xp_{m,n} - xp_{m-1,n} ≥ Δmin, which imposes a monotone order on the labels. For example, with N=1, M=2, let xp_1=0 and xp_2=2Δmin; the feasible point has Δp_2=2Δmin, but swapping the labels gives Δp'_2=-2Δmin, violating (6b). Thus the mapping from user positions to optimal A cannot satisfy Ω_P equivariance. Adding the index s to the input does not change this, because the underlying optimization problem is unchanged and still depends on the label-ordered constraints. Since the PBF-sub-GNN design and Proposition 1 are built on property (8), the theoretical justification of the proposed architecture collapses.","section":"III.A, Eq. (8) and constraint (6b)"},{"comment":"The proof of Proposition 1 is omitted with the statement 'Due to limited space, the proof is not provided.' This is not acceptable for a proposition that is load-bearing for the claimed equivariance. Moreover, given the issue raised in the previous comment, even a correct proof of the form in (11) being nested-permutation-equivariant would not establish the equivariance of the beamforming policy, because the policy is not equivariant due to the order-dependent constraints. The proposition as stated is therefore not a sufficient foundation for the architecture.","section":"III.A, Proposition 1"},{"comment":"The baseline method is defined only for M=1, where the x-axis position of each pinching antenna is set to that of the closest user and transmit beamforming is zero-forcing. However, Fig. 3(b) reports results for M=3, and the text states that the performance gain is larger when M=3. Without a baseline at M=3, the reported gains are not evidence that GPASS outperforms a reasonable alternative at the same number of antennas; the gain may simply be due to the larger number of pinching antennas. A fair baseline for M=3 must be provided, or the claim of higher SE must be restricted to M=1.","section":"IV, baseline comparison"},{"comment":"The conclusion states that the permutation properties were incorporated 'for better learning performance than FNNs,' but no FNN comparison appears anywhere in Section IV. This claim is unsupported and should be removed or substantiated with the corresponding experiments.","section":"V, conclusion"},{"comment":"The numerical results do not include error bars, confidence intervals, or multiple training runs. Since the training is stochastic and the SE differences between GPASS and the baseline appear modest in some regimes, the reader cannot assess whether the reported improvements are statistically significant. The authors should report the mean and standard deviation over several random seeds and training runs.","section":"IV, numerical results"}],"minor_comments":[{"comment":"The transformation uses the expression xp_{m,n} = xp_{1,n} + Σ_{i=2}^m Δp_{m,n}; the index of Δp in the summation should be i, not m. This typo makes the relationship confusing.","section":"II, Eq. (6a)"},{"comment":"The activation function in (9) uses max(·, 0) + Δmin for Δp_{m,n}, but it is not clear whether this applies for m=1 as well, and the definition of Δp_{1,n} is ambiguous because xp_{0,n} is not defined. Please clarify the indexing of Δp.","section":"III.A, Eq. (9)"},{"comment":"The pooling function is denoted as ' summation P(·)' but P is also used elsewhere for the power constraint; using a distinct symbol for the pooling operation would avoid confusion.","section":"III.A, Eq. (10)"},{"comment":"The caption of Fig. 3 does not specify the baseline configuration for panel (b) with M=3; it should state whether the baseline curves correspond to M=1 or M=3.","section":"IV, Fig. 3"},{"comment":"The paper sets N=K in all simulations, which is restrictive. A brief discussion of how the results would generalize to N≠K would be helpful, especially since the GNN is claimed to be size-generalizable.","section":"IV, simulation setup"}],"recommendation":"reject","confidential_remarks":"The central theoretical claim of the paper—the nested permutation equivariance in Eq. (8)—appears to be false because the feasible set is not invariant under pinching-antenna label permutations. This is not a missing-proof issue but a fundamental flaw in the problem formulation. The numerical comparison is also weakened by the baseline being defined only for M=1 while the main gains are claimed for M=3, and the conclusion contains an unsupported FNN comparison. These issues are beyond a routine revision and affect the validity of the paper's main contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read on arXiv:2502.01438. The paper adapts an established edge-update GNN framework to jointly learn pinching-antenna positions and transmit beamforming in PASS. The new elements are the staged PBF/TBF decomposition and the claim that the policy satisfies a nested permutation equivariance that the GNN can exploit. The system model is standard, the writing is clear, and using a GNN for this problem is a reasonable idea. I give credit for the unsupervised training formulation and for leveraging the classical optimal TBF structure from [13].\n\nThe soft spot is not minor. Property (8) is load-bearing and very likely false. Constraint (5c) enforces a spatial order on the label m: Δ^p_{m,n} = x^p_{m,n} - x^p_{m-1,n} ≥ Δ_min. Permuting pinching-antenna labels on a waveguide does not preserve feasibility; for M=2, swapping x1 and x2 gives a negative Δ. The feasible set is not invariant under PA-label permutations, so \\hat F1(Φ,s) cannot satisfy the claimed nested equivariance. Adding the index s as an input does not change the underlying problem, which is independent of s. The omitted proof of Proposition 1 is a symptom, not the disease—the proposition is applied to a symmetry the problem does not have. This undermines the theoretical justification of PBF-sub-GNN. The raw SE numbers could in principle still be reproduced, but the architecture's claimed advantage is unsupported.\n\nThe numerical evaluation has its own gaps: the baseline is defined only for M=1 while the largest reported gains are for M=3, no error bars, no code, and the conclusion overclaims an FNN comparison that never appears in the results. These are fixable, but the equivariance issue is not a simple fix; it requires rethinking what symmetry actually holds.\n\nWho is this for? Readers working on GNNs for wireless policy learning will find it a useful cautionary example of checking whether the symmetry you encode is truly present. The PASS community may be interested in the problem setup, but should not rely on the theoretical claims. It deserves a serious referee—not because it is likely to be accepted, but because the flaw is subtle and worth documenting. My recommendation: send it to review, but expect the authors to either prove a restricted form of equivariance or drop the claim entirely.","headline":"The GPASS paper applies a known GNN template to a new problem, but its core permutation-equivariance claim is false under the stated spacing constraint, so the theoretical basis collapses.","tokens_in":9423,"tokens_out":4734,"would_cite":false,"duration_ms":41283,"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":"This paper claims that joint pinching-antenna placement and transmit beamforming can be learned end-to-end by two cascaded graph neural networks, and that the learned policy beats a heuristic baseline in spectral efficiency while running…","keywords":["pinching-antenna systems","beamforming","graph neural networks","permutation equivariance","spectral efficiency","deep learning","transmit beamforming","real-time inference"],"falsifier":"For a small system such as $N=2$, $M=2$, $K=2$, enumerate all user, waveguide, and antenna permutations, feed them to the trained GPASS, and check whether the output positions and power allocations permute exactly as equations (7) and (8) require; any violation would falsify the equivariance claim. Comparing the GPASS result against brute-force optimal spectral efficiency for that tiny system would also show whether the heuristic baseline comparison understates the achievable gain.","tokens_in":1588,"feed_emoji":"📡","tokens_out":3886,"duration_ms":68570,"temperature":0.7,"pith_summary":"The paper proposes GPASS, the first deep-learning architecture for beamforming in pinching-antenna systems, where antennas are placed on dielectric waveguides and their positions can be tuned. It claims that a staged graph neural network—first learning antenna positions, then transmit beamforming—can exploit the permutation symmetries of the beamforming policy and the known structure of optimal transmit beamforming to reach spectral efficiencies at least as high as a heuristic baseline while running in milliseconds. A sympathetic reader would care because pinching-antenna optimization is non-convex and coupled, and learned policies could offer real-time operation and generalization across numbers of users, waveguides, and pinching antennas.","feed_headline":"Deep learning places pinching antennas and beamforms in 6 ms","feed_subtitle":"A staged graph network sets antenna positions and beams, beating a heuristic baseline in spectral efficiency.","key_machinery":"The load-bearing object is a two-stage edge-update graph neural network whose layers respect two-dimensional permutation equivariance. In the pinching-beamforming stage, edge representations are updated by functions of the form in (11) that combine a pinching antenna's own state, summed contributions from antennas on the same waveguide, and summed contributions from all other waveguides; this form is what preserves equivariance to nested permutations of pinching antennas and waveguides under user permutations. In the transmit-beamforming stage, a permutation-equivariant edge-update GNN learns power allocations, and the beamforming matrix is reconstructed using the optimal solution structure of the multiuser beamforming problem. This decomposition reduces the learning task to small power-allocation vectors instead of raw beamforming matrices.","core_discovery":"The central claim is that joint pinching beamforming and transmit beamforming in a pinching-antenna system can be learned end-to-end by two cascaded graph neural networks. The first sub-GNN, the PBF-sub-GNN, outputs pinching-antenna positions, represented as first positions and spacings, from user locations and antenna indices; the second, the TBF-sub-GNN, outputs uplink and downlink power allocations, from which the transmit beamforming matrix is recovered via the known optimal structure $\\mathbf{W}^\\star = \\tilde{\\mathbf{H}}(\\Lambda \\tilde{\\mathbf{H}}^H \\tilde{\\mathbf{H}} + \\sigma_0^2 \\mathbf{I}_K)^{-1} \\mathbf{P}^{1/2}$. The architecture is built so that its update equations are equivariant to permutations of users, waveguides, and pinching antennas, matching the permutation properties of the optimal policy. Numerical results show spectral efficiency close to or above a zero-forcing-style baseline at transmit SNRs from 50 to 90 dB, with an average inference time of 6 ms for 8 users, 8 waveguides, and 3 pinching antennas per waveguide.","pith_inferences":["A testable extension is to compare GPASS against an iterative optimization baseline at small sizes; the current comparison is only against a heuristic placement-plus-zero-forcing baseline, so the gap to the true optimum remains unknown.","If Proposition 1's nested equivariance truly holds, the same two-recursion update form could be applied to other nested-permutation problems, such as multi-cell precoding with antenna-array subgroups, which the paper does not explore.","The 6 ms inference time is reported without hardware details; reporting per-antenna or per-parameter timing would let readers judge real-time claims across platforms.","One could probe the equivariance claim empirically by feeding randomized user permutations to the trained network and checking whether outputs permute accordingly; failure would point to Proposition 1 or its implementation."],"forward_implications":["If GPASS works as claimed, pinching-antenna beamforming can be run in real time: 6 ms inference for an 8-user, 8-waveguide, 3-antenna-per-waveguide setup.","The staged design means antenna-placement decisions and transmit beamforming no longer need to be solved as one coupled non-convex optimization; each stage learns a smaller mapping.","Because both sub-GNNs are permutation-equivariant, the learned policy should transfer to different numbers of users, waveguides, and pinching antennas without retraining from scratch.","The transmit-beamforming stage's use of the optimal solution structure means the GNN only needs to learn $K$ uplink and $K$ downlink power scalars, not the full beamforming matrix.","Unsupervised training on user positions alone avoids the need for labeled optimal solutions during training."],"supporting_citations":[{"why":"Supplies the edge-update GNN framework and the two-dimensional permutation-equivalence property that both sub-GNNs are designed to satisfy.","marker":"[10]"},{"why":"Provides the optimal multiuser transmit beamforming structure used to recover the beamforming matrix from learned power allocations.","marker":"[13]"},{"why":"Motivates the interference-aware recursive update equation in the pinching-beamforming sub-GNN and its size-generalizable form.","marker":"[9]"},{"why":"Establishes the edge-update empowered GNN architecture for radio resource management that the paper adapts to pinching-antenna systems.","marker":"[7]"}],"fun_headline_variants":["GNNs jointly learn pinching antenna positions and beamforming","Graph networks set antenna positions and beams in 6 ms","Permutation-equivariant GNNs beat heuristic in pinching-antenna beamforming","Two-stage GNN learns antenna placement and beamforming for pinching arrays","Deep learning positions pinching antennas and beamforms jointly"],"cache_read_input_tokens":11520,"weakest_assumption_plain":"The architecture's permutation equivariance rests on Proposition 1, whose proof is omitted; if the update rule in (11) does not actually preserve nested permutations of antennas and waveguides, the claimed learning advantage loses its foundation.","fun_headline_variants_meta":{"raw":{"variants":["GNNs jointly learn pinching antenna positions and beamforming","Graph networks set antenna positions and beams in 6 ms","Permutation-equivariant GNNs beat heuristic in pinching-antenna beamforming","Two-stage GNN learns antenna placement and beamforming for pinching arrays","Deep learning positions pinching antennas and beamforms jointly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00053,"raw_usage":{"total_tokens":2527,"prompt_tokens":894,"completion_tokens":1633,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":1543}},"tokens_in":510,"tokens_out":1633,"duration_ms":11966,"temperature":1.0,"reasoning_tokens":1543,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T15:17:21.711776+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a small system such as $N=2$, $M=2$, $K=2$, enumerate all user, waveguide, and antenna permutations, feed them to the trained GPASS, and check whether the output positions and power allocations permute exactly as equations (7) and (8) require; any violation would falsify the equivariance claim. Comparing the GPASS result against brute-force optimal spectral efficiency for that tiny system would also show whether the heuristic baseline comparison understates the achievable gain.","supporting_citations":[{"cited_title":"Multidimensional graph neural networks for wireless communications,","cited_arxiv_id":null,"evidence_quote":"Supplies the edge-update GNN framework and the two-dimensional permutation-equivalence property that both sub-GNNs are designed to satisfy."},{"cited_title":"Recursive GNNs for learning precoding policies with size-generalizability,","cited_arxiv_id":null,"evidence_quote":"Motivates the interference-aware recursive update equation in the pinching-beamforming sub-GNN and its size-generalizable form."},{"cited_title":"ENGNN: A general edge- update empowered GNN architecture for radio resource management in wireless networks,","cited_arxiv_id":null,"evidence_quote":"Establishes the edge-update empowered GNN architecture for radio resource management that the paper adapts to pinching-antenna systems."}],"review_version":1}