{"id":"d58b3ae7-a0b3-4290-b9d5-68f5267790bf","arxiv_id":"2501.15207","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A joint phase-time array with one radio-frequency chain can serve near- and far-field users on different subbands, and a graph attention network can allocate resources nearly as well as iterative optimization at much lower complexity.","lead":"This paper simulates an antenna array that uses time delays and phase shifters to send different frequency bands to different users, including users close to the base station and users far away. It reports that this design gives about 8% higher user rates than a conventional phase-only array, and that a neural-network solver matches the iterative optimizer at much lower runtime.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The DL PS-normalization in Eq. (37) violates the unit-modulus constraint (9f), so the headline DL-based results are not reproducible as described; the reported JPTA gains rest on an undocumented implementation choice.","rationale":"I read the paper as making a system-design claim: the JPTA architecture, when optimized by the proposed AO or DL method, provides feasible frequency-dependent beamforming gains. The AO derivation is largely sound; Eq. (16) is a valid conjugate-phase target in the one-hot subband-allocation regime, and the alternating updates in Eqs. (19) and (21) respect the unit-modulus and delay constraints. The weakest point I find is in the DL solver, which is the method used for the headline reported comparisons. As written, Eq. (37) produces a PS vector with unit norm rather than unit-modulus entries, making the output infeasible. Because no code is provided, this cannot be dismissed as a simple typo; it is a concrete internal inconsistency in the description of the algorithm that produces the main numbers. The reader's identified weakest assumption - the LoS-only, Fresnel-truncated channel model - is legitimate but secondary in my reading: within the stated dominant-LoS scenario the model is standard, and the optimality of conjugate-phase beamforming for a single user per subband does not depend on the LoS structure as long as the channel is known. The lack of error bars on the 8% gains is also a real limitation but affects confidence rather than correctness. The Eq. (37) issue is more directly load-bearing because if the literal text is followed, the DL solution is infeasible and the claim that DL matches AO cannot be true; if the text is wrong, the paper needs a correction or code release to be reproducible. This reinforces the reader's CONDITIONAL verdict rather than changing it.","tokens_in":16284,"tokens_out":17975,"duration_ms":185242,"concrete_test":"Inspect the released code or request the authors' PS normalization layer. If Eq. (37) is implemented literally, rerun the Section V experiments with per-element unit-modulus normalization, i.e., set each entry of phi_i to (a + jb)/sqrt(a^2 + b^2), and recompute Table IV and Fig. 5. The concern is settled if (a) the DL-vs-AO logarithmic-rate gap remains at most about 0.01 and the JPTA-vs-PA rate gains remain about 8% with the corrected normalization, or (b) the gains collapse, in which case the headline comparisons are artifacts of an infeasible DL beamformer. Independently, on a single test sample, directly evaluate |[phi_i]_j| from the DL output and verify that every entry has magnitude 1 within numerical tolerance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing issue is that the DL branch, which produces the headline numerical comparisons, is not described feasibly. Eq. (37) forms the PS beamformer as phi_i = (phi_tilde_i^Re + j phi_tilde_i^Im) / ||phi_tilde_i^Re + j phi_tilde_i^Im||, normalizing the whole (N/N_T)-dimensional vector to unit Euclidean norm. This makes |[phi_i]_j| approximately 1/sqrt(N/N_T), not 1, so the DL output violates the constant-modulus PS constraint (9f) that the AO branch enforces exactly. For N=64 and N_T=16, the literal formula reduces the beamforming gain by a factor of N/N_T = 4 relative to a unit-modulus vector, i.e., about 6 dB. If the simulations actually used this normalization, the JPTA array gains reported in the paper would be near 7-10 rather than the 29-42 shown, and the DL/AO parity in Table IV could not hold. If, instead, the implementation used the intended per-element unit-modulus normalization, then Eq. (37) is incorrect and the paper is not reproducible from the text alone. Either way, the central quantitative claims - 8.21% and 8.07% JPTA-over-PA rate gains and DL matching AO - are currently supported only by an undocumented implementation choice, not by the described algorithm.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers a single-RF-chain joint phase-time array (JPTA) base station serving multiple users located in both the near-field and far-field regions over a wideband OFDM downlink. The authors formulate a network utility maximization problem over subband allocation, transmit power, PS-based beamformers, and TTD-based beamformers, with sum-rate and proportional fairness as special cases. They propose a 3-step alternating optimization (AO) algorithm in which subband allocation is handled by successive convex approximation with an exact penalty, the analog beamformer is approximated via a closed-form per-subband target followed by block coordinate descent, and power is allocated by water-filling. They also propose an unsupervised deep-learning approach combining a CNN feature extractor, a node-wise graph attention network, and a normalization module, and they report that the DL method matches the AO method at much lower complexity and that JPTA improves user rates by about 8% over conventional phased arrays.","tokens_in":16653,"tokens_out":7231,"duration_ms":72793,"significance":"If the numerical claims are reliable, the paper makes a useful contribution by extending JPTA beam-splitting to multi-user hybrid near/far-field scenarios with a single RF chain, and by demonstrating a scalable GAT-based alternative to iterative optimization. The problem decomposition is clean, the complexity analysis is provided, and the unsupervised learning formulation is interesting. However, the central quantitative claims currently rest on the DL branch, and the DL normalization as written in Eq. (37) violates the unit-modulus constraint, which puts the reported 8% rate gains and the DL/AO parity in doubt. Because the issue is localized and potentially fixable by correcting the normalization and re-running the DL experiments, the manuscript warrants major revision rather than rejection.","major_comments":[{"comment":"The normalization in Eq. (37) does not enforce the unit-modulus constraint (9f). The formula phi_i = (phi_tilde_i^Re + j phi_tilde_i^Im) / ||phi_tilde_i^Re + j phi_tilde_i^Im|| normalizes the entire (N/N_T)-dimensional vector to unit Euclidean norm, so each entry has magnitude approximately 1/sqrt(N/N_T), not 1. For N = 64 and N_T = 16 this is a factor-of-4 (6 dB) reduction in array gain relative to a unit-modulus vector. Since Section V explicitly states that 'in subsequent simulations, we shall only showcase the performance of the DL method,' the reported JPTA gains of 8.21% and 8.07% and the DL/AO parity in Table IV are not supported by the algorithm as written. Please either use a per-element unit-modulus normalization (for example, phi_{i,j} = exp(j * angle(...))) or, if the implementation used such a normalization, correct Eq. (37) so that the paper is reproducible from the text.","section":"Section IV-D, Eq. (37)"},{"comment":"The optimal analog beamformer w_m is asserted to be directly optimized as w_m = (1/sqrt(N)) sum_k b_{m,k} exp(j angle(h_{m,k})), but no derivation is given. Because the objective in (P2.1) applies a concave utility F to each user's sum of rates over subbands, the decoupling of subbands is not immediate. Please provide a proof, or at least a clear argument that for each subband m the only active term is the one for the user with b_{m,k}=1 and that maximizing |h^H_{m,k} w_m| subject to |[w_m]_n| = 1/sqrt(N) leads to this expression.","section":"Section III-B, Eq. (16)"},{"comment":"The convergence of the 3-step AO algorithm is not established. The algorithm combines SCA with a penalty factor updated as rho = 5*rho, an inner BCD loop for the PS/TTD matching problem, and water-filling, but no monotonic improvement or limit-point argument is provided. The paper calls the result 'near-optimal' without supporting analysis. Please either state explicitly that the algorithm is heuristic and validate it empirically (for example, by plotting objective versus outer iteration), or provide a convergence guarantee under the stated assumptions.","section":"Section III-D, Algorithm 1"},{"comment":"The paper does not specify how the subband allocation variables b_{m,k} are obtained at inference time. If the continuous Gumbel-softmax outputs are used directly in the rate expression (10), the binary constraint (9a) and the one-subband-per-user constraint (9b) are violated and the computed rates are not achievable. If instead an argmax or sampling step is applied at test time, that step and its effect on the reported rates should be described explicitly. This is necessary for the reproducibility of Table IV and Fig. 5.","section":"Section IV-D and Section V, Gumbel-softmax at inference"}],"minor_comments":[{"comment":"The equivalence in Eq. (11) should be stated more carefully; as written, it can be read as an equivalence for a single b_{m,k}, when in fact it is the conjunction of the sum constraint and the box constraint over all m,k that forces each b_{m,k} to be binary.","section":"Section II-B, Eq. (11)"},{"comment":"The caption 'Array gain for 2-user scenario at the distance of 1 meters achieved by different approaches' contains a typo ('1 meters' should be '1 meter') and is ambiguous about whether both users are at 1 m; please clarify.","section":"Section V-B, Fig. 4 caption"},{"comment":"The complexity expressions in Table IV are typeset in a way that splits formulas across lines and obscures the intended expressions; please format the table so that the AO and DL complexity formulas are readable and complete.","section":"Table IV"},{"comment":"There is a typo in the sentence 'The aim of this work is to explore the potential of JPTA architecture to generate frequency-dependent beamformers for hybird near-far field communications.' Please change 'hybird' to 'hybrid'.","section":"Section I, last paragraph"},{"comment":"The loss function includes the term -lambda_3 * sum b_{m,k} log b_{m,k}; the text says this drives b closer to 0 or 1, but this behavior should be stated explicitly in terms of the entropy interpretation to avoid confusion about the sign convention.","section":"Section IV-D, Eq. (38)"}],"recommendation":"major_revision","confidential_remarks":"The core JPTA architecture and the block-coordinate-descent matching step are taken from [13], which shares a co-author with this manuscript. The paper is not fully self-contained in this respect, and the authors should ensure that the overlap is clearly acknowledged and that the novel contributions (hybrid near/far-field extension, subband and power allocation, and the GAT-based solver) are distinct. More importantly, the DL normalization issue in Eq. (37) is a correctness problem for the headline results; the editor may wish to ask the authors to clarify whether the reported DL results were obtained with per-element unit-modulus normalization and, if so, to correct the manuscript and possibly rerun the experiments with the stated normalization."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe bottom line: this is a competent engineering extension of JPTA to hybrid near/far-field multi-user wideband systems, with a standard AO pipeline and a GAT-based learned solver. The new combination is real—no prior work in the cited literature jointly does subband allocation and JPTA beamforming for mixed-field users from a single RF chain. The 8% rate gains over PA are plausible for LoS channels, and the energy-efficiency discussion is sensible.\n\nWhat it does well: the problem formulation is clean, the AO decomposition follows standard SCA/BCD/water-filling, and the simulations cover array gains, rate CDFs, and energy efficiency. The graph attention construction is thoughtful, with type-specific weights and a symmetric adjacency matrix. The complexity analysis is honest about the offline training cost.\n\nThe soft spots are mostly minor, except one load-bearing issue. Eq. (37) normalizes the entire PS vector to unit Euclidean norm, so each entry has magnitude 1/sqrt(N/N_T). For N=64, N_T=16, that is 1/4, a 6 dB loss. If the DL output literally followed Eq. (37), the reported parity with AO in Table IV could not happen. If the implementation used per-element unit-modulus normalization, then the equation is wrong and the paper is not reproducible from the text. Either way, the DL-based headline results rest on an undocumented implementation choice. The authors must fix this and ideally release code.\n\nEq. (16) is asserted without proof. It is actually correct for fixed subband allocation because each subband serves one user and the utility is increasing, but the paper should state that reasoning. The AO convergence is empirical, not proven—acceptable for an engineering venue, but it should be labeled as such. The LoS-only channel model is a stated scope limitation; multipath would change the beamforming target, but that is not fatal.\n\nWho this is for: researchers working on TTD-based beamforming or learned resource allocation for 6G. The paper deserves a serious referee—the core idea is useful, the AO part is largely sound, and the DL flaw looks fixable. I would ask for a revised version with corrected normalization, code/data, and error bars. As written, I would not cite it.\n\nRecommendation: send to peer review, but flag the Eq. (37) problem prominently.","headline":"Useful JPTA extension to hybrid near/far-field with a real, fixable reproducibility flaw in the DL normalization that must be addressed before the headline numbers are trusted.","tokens_in":17148,"tokens_out":2649,"would_cite":false,"duration_ms":26860,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A12","94A05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A joint phase-time array with one RF chain can split its beam by frequency to serve near-field and far-field users simultaneously, achieving measurable rate gains over conventional phased arrays.","keywords":["joint phase-time arrays","frequency-dependent beamforming","true-time delay","near-field communications","far-field communications","graph attention network","subband allocation","hybrid beamforming"],"falsifier":"Run the same utility maximization on channels with strong multipath or with a user placed inside the Fresnel region where the quadratic distance approximation fails; if JPTA's rate gain over phase arrays falls well below the reported 8% or reverses, then the phase-aligned target in Eq. (16) is not adequate for those channels.","tokens_in":16098,"feed_emoji":"📡","tokens_out":7792,"duration_ms":66536,"temperature":0.7,"pith_summary":"This paper argues that a joint phase-time array (JPTA), which adds a bank of true-time delays to the phase shifters of a conventional array, can turn a wideband signal into a set of frequency-split beams: different subbands can simultaneously serve users in different directions and at different distances using a single radio-frequency chain. The target scenario is a base station with an extremely large array serving some users in the near-field, where the wavefront is spherical and the beam must be focused on a point, and others in the far-field, where the beam is steered by angle. The authors formulate the joint choice of subband assignment, transmit power, phase shifts, and time delays as a utility-maximization problem and solve it with a three-step alternating optimization and with an end-to-end graph attention network trained without labels. They report that JPTA raises average user rates by 8.21% and 8.07% over phase-array baselines in 2-user and 5-user mixed-region settings, while the learned solver matches the optimization-based one at orders-of-magnitude lower runtime.","feed_headline":"Frequency-splitting beams lift user rates 8% over phased arrays","feed_subtitle":"Time-delay beamforming reshapes each subband, lifting rates 8% over phased arrays in mixed near/far-field setups.","key_machinery":"The engine is the true-time delay unit, whose phase response $-2\\pi f_m \\tau_i$ makes the beam pattern frequency-dependent so that one analog front end realizes different beams on different subbands. The paper's key optimization target is the phase-aligned beamformer $w_m = \\frac{1}{\\sqrt{N}}\\sum_{k=1}^K b_{m,k}\\exp(j\\angle h_{m,k})$, which says what each subband's ideal beam should look like before the hardware is fit to it. The alternating optimization cycles through subband allocation via successive convex approximation, fitting the phase-shifter and delay hardware to $w_m$ via block coordinate descent, and power allocation via water-filling. The learning variant encodes the JPTA as a graph with subband, phase-shifter, and time-delay nodes, then uses a node-wise graph attention network with a Gumbel-softmax normalization to output allocations and beamformers directly from channel information.","core_discovery":"The paper's central claim is that JPTA can multiplex users in the hybrid near/far field by exploiting beam-splitting: because a true-time delay of $\\tau_i$ contributes phase $-2\\pi f_m \\tau_i$ that scales with subband frequency, the analog front end can point different subbands at different targets without a per-antenna RF chain. The design target at each subband is the phase-aligned beamformer $w_m = \\frac{1}{\\sqrt{N}}\\sum_{k=1}^K b_{m,k}\\exp(j\\angle h_{m,k})$, and the analog hardware is then fitted to approximate each $w_m$ by the product of the frequency-independent phase-shifter matrix $\\Phi$ and the frequency-dependent delay vector $T_m$. On this basis the paper formulates a network-utility maximization over subband allocation, power, and analog beamformers, solves it with a 3-step alternating optimization, and learns the same mapping with a graph attention network that treats subbands, phase-shifter groups, and TTDs as nodes. The numerical evidence is that JPTA outperforms phase-array baselines in user rate and fairness and lands between phase arrays and fully digital arrays in energy efficiency.","pith_inferences":["Editorial inference: under a measured multipath channel, the beam-splitting gain should shrink or vanish because the phase-aligned target is matched to a single line-of-sight path.","Editorial inference: the graph encoding of subbands, phase-shifter groups, and TTDs is a general pattern for coupling discrete resource decisions with continuous analog constraints, and could transfer to antenna selection or OFDMA scheduling with hardware-imposed phase ties.","Editorial inference: the same frequency-splitting hardware could be tested for integrated sensing and communication by assigning some subbands to data users and others to distance-focused sensing beams."],"forward_implications":["Under proportional-fairness (logarithmic) utility, JPTA's average user-rate gain over phase-array beamforming is 8.21% with two users and 8.07% with five users; under sum-rate utility the gains are 6.97% and 7.15%.","Using more TTDs improves array gain and rates: with one TTD per antenna ($N_T=64$) and a 5 ns delay range, JPTA closes 26.4% of the logarithmic-rate gap between phase arrays and fully digital beamforming, and even a 0.05 ns delay range still closes 17.3% of that gap.","The graph attention network reaches essentially the same logarithmic rate as the alternating optimization in the 5-user test (118.525 vs 118.531) while dropping average CPU runtime from 7.11 minutes to 0.11 seconds per sample.","JPTA's energy efficiency sits between phase-array and fully-digital designs, and increasing the number of TTDs reduces energy efficiency because the added delay hardware consumes more power than the spectral-efficiency gain it buys."],"supporting_citations":[{"why":"introduces the JPTA architecture and its beam-splitting capability, which the paper extends to hybrid near/far-field multi-user service","marker":"[13]"},{"why":"supplies the true-time-delay phase response that makes the frequency-dependent beamforming physically possible","marker":"[8]"},{"why":"provides the spherical-wave near-field distance model used in Eq. (1) for near-field channel responses","marker":"[32]"},{"why":"provides the graph attention mechanism that the learning-based solver is built on","marker":"[31]"},{"why":"supplies the Gumbel-softmax trick used to make discrete subband allocation differentiable in the neural network","marker":"[37]"},{"why":"supplies the logarithmic/proportional-fairness utility objective used to balance throughput against user fairness","marker":"[33]"}],"fun_headline_variants":["Joint phase-time arrays outperform phased arrays in wideband links","One RF chain, near and far users: phase-time beamforming","Frequency-splitting beams from joint phase-time arrays","Deep learning matches optimization for joint phase-time beamforming","JPTA: hybrid near/far beamforming with single RF chain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The channel model assumes a single dominant line-of-sight path for every user, with near-field distances truncated at the Fresnel quadratic term; the beam-splitting gain is built on phase alignment to that one path.","fun_headline_variants_meta":{"raw":{"variants":["Joint phase-time arrays outperform phased arrays in wideband links","One RF chain, near and far users: phase-time beamforming","Frequency-splitting beams from joint phase-time arrays","Deep learning matches optimization for joint phase-time beamforming","JPTA: hybrid near/far beamforming with single RF chain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00068,"raw_usage":{"total_tokens":3139,"prompt_tokens":1045,"completion_tokens":2094,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":2010}},"tokens_in":661,"tokens_out":2094,"duration_ms":13408,"temperature":1.0,"reasoning_tokens":2010,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:31:10.216147+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same utility maximization on channels with strong multipath or with a user placed inside the Fresnel region where the quadratic distance approximation fails; if JPTA's rate gain over phase arrays falls well below the reported 8% or reverses, then the phase-aligned target in Eq. (16) is not adequate for those channels.","supporting_citations":[{"cited_title":"Graph Attention Networks,","cited_arxiv_id":null,"evidence_quote":"provides the graph attention mechanism that the learning-based solver is built on"},{"cited_title":"Categorical Reparameterization with Gumbel-Softmax,","cited_arxiv_id":null,"evidence_quote":"supplies the Gumbel-softmax trick used to make discrete subband allocation differentiable in the neural network"},{"cited_title":"Charging and rate control for elastic traffic,","cited_arxiv_id":null,"evidence_quote":"supplies the logarithmic/proportional-fairness utility objective used to balance throughput against user fairness"}],"review_version":1}