{"id":"07f2f0d9-d998-450e-ac1f-c3e4abacdac8","arxiv_id":"2412.13532","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A squint-aware analog precoder that tunes communication-sensing channel correlation via true-time-delay hardware and a complex-valued network improves dual-functional ISAC performance in sub-THz systems.","lead":"This paper designs analog precoders for sub-THz systems that serve both communication and radar sensing, using delay lines to align the two channels and a neural network to speed up the design. The authors show in simulation that their approach outperforms existing squint-aware baselines in both data rate and sensing accuracy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1's proof rests on an unproved and generally false monotonic link between reciprocal KL (Cor) and beamspace peak overlap S; if this bridge fails, the TTD objective lacks theoretical support.","rationale":"The reader's verdict is CONDITIONAL with medium correctness risk, and I agree. The paper's strongest claim is Proposition 1, and the proof's central bridge is the Cor-to-S monotonicity attributed to [28]. That bridge is neither stated nor implied by [28] and is generally false; hence the theoretical support for the TTD correlation-maximizing objective is missing. I do not reject the paper because the simulations and ablations provide independent empirical support for the proposed schemes in the tested settings, and the concern is fixable by either proving a restricted monotonicity statement under the LoS/sub-THz channel model or softening Proposition 1 to a design heuristic. Since the identified gap is exactly the one flagged by the reader, no verdict change is needed.","tokens_in":14102,"tokens_out":6014,"duration_ms":59352,"concrete_test":"Run a small numerical experiment (e.g., Nt=16 or 64, M=4, K=1) with random LoS angles. For a continuous sweep of the target/user angular separation across DFT bin boundaries, compute Cor (Eq. 16) and S (exact argmax peak overlap) and plot one against the other. If Cor and S are not monotonically related, the proof's step (2) is refuted. To settle Proposition 1 itself, take two channel realizations with Cor1>Cor2 but S1<S2 and compare their true rate-CRB Pareto frontiers, obtained by scalarized optimization over the full covariance (not the restricted subspace) at fixed total power. If the higher-Cor realization does not dominate the lower-Cor one, the claimed monotonicity of the Pareto boundary in Cor is false in that regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central design principle is that larger Cor(hc,G) in Eq. (16) moves the ISAC Pareto boundary outward. Appendix A establishes this through three steps: (1) assume the optimal transmit covariance lies in the subspace Um of [15]; (2) assert that 'a higher Cor gives rise to a higher S' over beamspace peak indices, citing [28]; (3) assert that larger S improves the entries xi_c and xi_s in (A.4), hence R* and 1/CRB*. Step (2) is the load-bearing bridge, and it is not a consequence of [28]: Hershey and Olsen bound or approximate KL divergence between Gaussian mixtures and state no monotonic relation to indicator-based peak overlap. The claim is false in general because KL is a functional of the whole normalized beamspace distributions, while S counts exact coincidences of argmax peaks. Two distributions can be made KL-close by sharing low-level sidelobes while their dominant peaks remain in different bins, so Cor can be large while S is zero; conversely, matching peaks with very different tail masses can keep S constant while KL varies. Step (3) is also not derived: xi_c,m,n1,n2 and xi_s,m,n1,n2 are inner products involving steering-vector overlaps and the subspace basis, not simple monotone functions of the indicator sum. The subspace restriction (A.1) is an additional imported assumption that is not re-derived for TTD-modified equivalent channels. If steps (2) or (3) fail, Proposition 1 is unsupported, and the objective of problem (17) and the loss (22) have no theoretical grounding, although the simulations can still be read as empirical evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a true-time-delay (TTD) based analog precoder design for sub-THz integrated sensing and communication (ISAC) systems. It introduces a communication-sensing (C-S) channel correlation measure Cor(hc,G) defined via the inverse KL divergence of beamspace distributions, and claims (Proposition 1) that higher Cor improves the rate-CRB Pareto boundary under the same transmit power. Based on this, the authors formulate an optimization benchmark (SA-Opt) that maximizes Cor through TTDs and then optimizes phase shifters and power allocation, and a low-complexity complex-valued neural network (CSP-Net) trained with an unsupervised loss that combines correlation, rate, and CRB. Simulation results show performance gains over controlled beam-squint and dedicated baselines.","tokens_in":14471,"tokens_out":3910,"duration_ms":35657,"significance":"If the main claim holds, the paper offers a practically meaningful design principle: TTDs can be used to actively tune the beamspace correlation between communication and sensing channels in sub-THz ISAC, improving joint performance over existing squint-mitigation approaches. The proposal of a lightweight complex-valued network with a tailored architecture is a useful contribution, and the empirical comparison against several external baselines (comm-dedicated, sensing-dedicated, CBS-ISAC, Opt w/o TTD) strengthens the plausibility of the proposed schemes. The paper also provides a complexity analysis showing a substantial reduction for CSP-Net relative to SA-Opt. However, the theoretical foundation is currently incomplete, as the proof of Proposition 1 relies on an unproved and generally questionable monotonicity bridge between a KL-based correlation measure and beamspace peak overlap, and the simulation results do not provide uncertainty quantification or a fully specified 'ideal' benchmark.","major_comments":[{"comment":"The step 'It can be proved that a higher Cor(hc,G) gives rise to a higher S(hc,G) [28]' is not established by the cited reference. Reference [28] (Hershey and Olsen, ICASSP 2007) addresses approximation of KL divergence between Gaussian mixture models and contains no statement relating KL divergence to indicator-based peak overlap S. Moreover, the claim is false in general: Cor is defined via the KL divergence between the full normalized beamspace distributions (Eq. 16), while S counts exact coincidences of argmax peaks. Two distributions can be KL-close while having disjoint dominant peaks if their sidelobes are similar, so Cor can be large while S is zero; conversely, matching peaks with different tail masses can keep S constant while Cor varies. Since the objective (17) and the loss (22) directly optimize Cor, this unproved bridge is load-bearing for the paper's central design principle.","section":"Appendix A, Proposition 1 proof"},{"comment":"The expressions for ξc,m,n1,n2 and ξs,m,n1,n2 are asserted without derivation, and the statement that both 'improve' as S increases is not formalized. These quantities are inner products involving steering-vector overlaps and the subspace basis; they are not simple monotone functions of the indicator sum S. Without a rigorous derivation or a counterexample-free argument, the claim that larger S increases R* and decreases CRB* remains unsupported. This is a second load-bearing gap in the proof of Proposition 1.","section":"Appendix A, Eq. (A.4)"},{"comment":"The proof assumes, following [15], that the optimal transmit covariance lies in the reduced subspace Um = [At,m*, Ȧt,θ,m*, Ȧt,φ,m*, at(θc,φc,fm)] of dimension Nr = 3K+1. This assumption is imported without re-derivation for the TTD-modified equivalent channels ẽc,m = FTD,m^H hc,m and ẽG,m = Gm FTD,m. The diagonal, frequency-dependent TTD matrix FTD,m alters the channel structure, so it is not immediate that the subspace optimality result of [15] carries over. This should be either proved or explicitly justified; otherwise the Pareto-boundary analysis in (A.2)-(A.3) is not applicable.","section":"Appendix A, Eq. (A.1)"},{"comment":"The paper repeatedly describes SA-Opt as 'near-optimal', but Fig. 5 shows an 'ideal case' curve whose definition is never stated in the text (e.g., is it infinite-resolution TTDs? an upper bound obtained by exhaustive search over unquantized delays? or a genie-aided scheme?). Without a precise definition and a quantitative measure of the gap, the near-optimal claim is not substantiated. Additionally, no error bars, confidence intervals, or statistical significance tests are reported for any of the simulation comparisons, so it is unclear whether the performance differences among schemes are meaningful given the random channel realizations.","section":"Section 6.3, Fig. 5"}],"minor_comments":[{"comment":"The correlation measure is defined as 1/KL(bhb_c, bhb_s), but KL divergence is asymmetric in general. The paper does not specify which argument is the reference distribution or whether a symmetrized version is intended. The lack of symmetry in Cor should be addressed explicitly.","section":"Eq. (16)"},{"comment":"The notation 'Cor(hc,Gs; T[qh,qv]=t)' is unclear: the subscripts on Gs are inconsistent with the rest of the paper (usually G), and the semicolon notation is not defined. It should be clarified that the correlation is computed on the equivalent channels after applying the candidate TTD value.","section":"Algorithm 1, line 5"},{"comment":"The closed-form solution for fPS is presented without showing the Cauchy-Schwarz relaxation steps or the conditions under which the relaxation is tight. Adding a short derivation would improve reproducibility and clarify the role of the sum over r of ẽG_m^H[r,:].","section":"Eq. (19)"},{"comment":"The quantities Cor*, R_max, and CRB_min in the loss function are not defined in the text. Cor* presumably denotes some normalization of the correlation value, but its definition is missing. Similarly, R_max and CRB_min likely refer to the dedicated-communication and dedicated-sensing points, but this should be stated explicitly. Also, the phase output φout = 2π/√2 |φ̂| is not constrained to lie in [0, 2π); the finite-resolution property of phase shifters (18a) should be handled explicitly.","section":"Eq. (22) and Section 4.2.3"},{"comment":"There are small presentation issues: 'sening' should be 'sensing' in Section 3.3; reference [25] is listed as arXiv:2405.14347 but the standard arXiv identifier for 'Deep complex networks' is 1705.09792 (please verify the correct version); and the word 'e ffective' appears with a misplaced space in Section 6.2 (and elsewhere). These do not affect the technical content.","section":"References and typos"}],"recommendation":"major_revision","confidential_remarks":"The paper's central design idea is appealing and the empirical results are suggestive, but the proof of Proposition 1 contains a load-bearing gap that is not merely a missing detail: the monotonic link between the KL-based correlation Cor and the beamspace peak overlap S is asserted with an inappropriate citation and is not generally true. The authors should either provide a rigorous proof (or a precise set of conditions) for this bridge, or reframe Proposition 1 as a heuristic design principle supported by simulations. If the theoretical claim is weakened, the contribution would still be meaningful as a demonstrated heuristic, but the current framing overstates the guarantee. The absence of a defined 'ideal case' and of uncertainty quantification further weakens the near-optimality and comparison claims. I recommend major revision rather than rejection, as the core engineering approach appears sound and the gaps are addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one for the engineering, not the theory. The authors propose using true-time-delay elements to reshape the beamspace distributions of communication and sensing channels so they overlap more, then train a complex-valued network (CSP-Net) to approximate the resulting precoder. That is a concrete, plausible idea, and the simulations against sensible baselines (CBS, dedicated schemes, no-TTD) show a few dB gains and an expanded rate-CRB region. The complex-valued architecture with attention is a reasonable tweak, and the complexity table is honest: SA-Opt is expensive, CSP-Net is cheap. What is genuinely new is actively tuning C-S channel correlation through TTDs, plus the specific network design. The ablation in Table 2 supports both design choices.\n\nNow the soft spots. Proposition 1 is the load-bearing wall, and it has a hole. The proof in Appendix A asserts that a higher Cor (reciprocal KL between beamspace distributions) gives a higher peak-overlap S, citing [28]. [28] is Hershey and Olsen on approximating KL divergence between Gaussian mixtures; it says nothing about peak coincidence. And the statement is false in general. KL is a functional of the whole normalized distributions; S is a count of exact argmax collisions. You can have distributions that are KL-close but with peaks in different bins (small KL, high Cor, S=0), or identical peaks with very different tail masses (same S, much larger KL). So the bridge is not merely unproved—it is not true without extra assumptions. The subsequent step, that larger S improves the xi_c and xi_s inner products, is also asserted rather than derived, and the subspace restriction (A.1) is imported from [15] without re-checking for TTD-modified channels. Consequently, the TTD objective in (17) and the loss in (22) have no theoretical grounding, and the 'near-optimal' benchmark label is not backed by any gap analysis.\n\nThe simulations may still be read as empirical evidence, and they might survive. But as written, the central claim is unsupported. Add to that no code/data and no error bars, and reproducibility is limited.\n\nWho benefits: researchers in sub-THz ISAC and hybrid precoding who want a concrete TTD heuristic and a network implementation to benchmark against. They should treat Proposition 1 as a conjecture, not a theorem.\n\nMy take: send it to peer review, but require the authors to either prove Proposition 1 under explicit conditions or rewrite the paper around the empirical claim. As is, the theory overclaims.","headline":"The paper offers a useful TTD beam-shaping heuristic and a complex-valued network, but the central monotonicity theorem is unproved and likely false as stated.","tokens_in":14997,"tokens_out":3355,"would_cite":false,"duration_ms":30726,"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":"By shaping beam squint with true-time delays, a sub-THz integrated sensing and communication transmitter can push both its data rate and its sensing accuracy outward on the achievable Pareto frontier.","keywords":["integrated sensing and communication","analog precoding","sub-terahertz","beam squint","true time delay","communication-sensing channel correlation","synesthesia of machine","complex-valued neural network"],"falsifier":"Fix one sub-THz channel realization with a user and several targets, hold phase shifters and transmit power fixed, and compute the rate-CRB Pareto frontier for the TTD setting that maximizes $\\mathrm{Cor}(h_c,\\mathbf{G})$ against a setting with lower correlation; if the lower-correlation setting achieves a strictly better rate at the same CRB or a lower CRB at the same rate, the monotonicity asserted in Proposition 1 is falsified.","tokens_in":13902,"feed_emoji":"📡","tokens_out":11300,"duration_ms":88638,"temperature":0.7,"pith_summary":"This paper argues that in sub-terahertz integrated sensing and communication, the beam squint that normally degrades array gain can be turned into a deliberate tuning knob. The central claim is that raising the correlation between the communication and sensing channel beamspace distributions, measured by the inverse Kullback-Leibler divergence $\\mathrm{Cor}(h_c,\\mathbf{G})$, moves the whole rate-versus-sensing-accuracy Pareto frontier outward: under equal transmit power, both the achievable data rate and the inverse Cramér-Rao bound increase. To exploit this, the paper proposes a squint-aware analog precoder benchmark that optimizes true-time delays to maximize this correlation, and then a lightweight complex-valued neural network that approximates the optimum with much lower complexity. A sympathetic reader would take away a design principle: analog front ends with frequency-dependent phase control can deliberately shape beam squint to serve both functions.","feed_headline":"Beam squint becomes a lever for better sub-THz ISAC","feed_subtitle":"Matching beam footprints lifts both rate and sensing accuracy; a light neural network finds the optimum fast.","key_machinery":"The load-bearing objects are the normalized beamspace power distributions $\\hat{\\mathbf{b}}_c$ and $\\hat{\\mathbf{b}}_s$, built by summing DFT-transformed per-subcarrier channels, and the correlation measure $\\mathrm{Cor}(h_c,\\mathbf{G})=1/\\mathrm{KL}(\\hat{\\mathbf{b}}_c,\\hat{\\mathbf{b}}_s)$. Proposition 1's monotonicity claim is what carries the argument: larger correlation improves both the rate and the inverse CRB along the Pareto frontier through greater overlap of the beamspace peaks of the user and the targets. True-time delays supply the physical knob, because their frequency-dependent phase shifts reshape the equivalent wideband channel and thereby change the correlation; the optimization benchmark searches over TTD values element-wise and then alternates between a closed-form phase-shifter update and a convex power allocation, while the proposed CSP-Net replaces that search with an unsupervised complex-valued convolutional architecture whose loss includes the normalized correlation.","core_discovery":"The paper's central discovery is Proposition 1, which states that under the same transmit power, every rate-CRB pair on the Pareto boundary of the achievable region improves as $\\mathrm{Cor}(h_c,\\mathbf{G})=1/\\mathrm{KL}(\\hat{\\mathbf{b}}_c,\\hat{\\mathbf{b}}_s)$ grows. The proof confines the transmit covariance to the reduced subspace spanned by the target steering vectors, their angle derivatives, and the user steering vector, then connects a higher correlation to a larger overlap of beamspace peaks, which raises both the communication and the sensing Fisher-information terms in the rate and inverse-CRB expressions. The paper further demonstrates in simulations that maximizing this correlation with true-time delays and alternating optimization yields near-optimal dual-functional performance, and that the learned network reproduces the behavior with far lower complexity.","pith_inferences":["Because the correlation criterion is defined on normalized beamspace distributions rather than on the specific TTD hardware, the same design heuristic may transfer to hybrid or fully digital arrays in lower bands, with the gain shrinking as beam squint weakens.","A natural testable extension is the multi-user or multi-target regime, where a single scalar correlation would need to be generalized to a set of beamspace coincidences, and it is not obvious that the monotone Pareto behavior survives.","The proof's restriction to the $3K+1$ subspace is the most fragile step; an independent numerical check across random channels could reveal whether the monotonicity holds for the true unconstrained optimum or only within the assumed subspace."],"forward_implications":["TTD tuning becomes a principled design degree of freedom: analog precoders in sub-THz ISAC can be designed by maximizing the correlation surrogate instead of solving the full nonconvex trade-off.","The SA-Opt benchmark approaches the separate communication-dedicated and sensing-dedicated limits, closing most of the rate-CRB gap within about 1.5--2 dB at high SNR in the paper's simulations.","CSP-Net reduces the design complexity to quadratic scaling in the number of antennas and subcarriers, with only a small performance loss, which matters for time-varying mobile channels.","Dual-functional gain depends on the spatial distribution of user and targets: it grows when user-target angular separation shrinks, and the squint-aware schemes degrade more gracefully than fixed beam-squint control as separation grows."],"supporting_citations":[{"why":"Supplies the reduced-subspace form of the optimal transmit covariance used in the proof of Proposition 1.","marker":"[15]"},{"why":"Cited as the basis for the claim that higher correlation implies higher beamspace peak overlap, a step central to the monotonicity result.","marker":"[28]"},{"why":"Provides the beamspace dictionary and sparsity argument used to define the correlation measure.","marker":"[23]"},{"why":"Introduces the delay-phase precoding architecture whose TTD hardware the paper exploits, and serves as the communication-dedicated baseline.","marker":"[12]"},{"why":"Demonstrates beam-squint control for sensing in ISAC, serving as both motivation and the sensing-dedicated baseline.","marker":"[13]"},{"why":"Supplies the convex reformulation used for the controlled-beam-squint CBS-ISAC benchmark.","marker":"[27]"}],"fun_headline_variants":["Beam squint turns into a feature for sub-THz ISAC","Squint-aware precoding lifts sub-THz ISAC performance","Exploiting beam squint for dual-functional sub-THz gains","Sub-THz ISAC: Squint as a resource, not a curse","Light neural network optimizes sub-THz analog precoding for ISAC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes without numerical verification that the optimal transmit covariance lies in the reduced subspace spanned by target steering vectors, their angle derivatives, and the user steering vector, and it takes as given a link between larger Kullback-Leibler-based correlation and larger beamspace peak overlap that the cited source does not actually prove.","fun_headline_variants_meta":{"raw":{"variants":["Beam squint turns into a feature for sub-THz ISAC","Squint-aware precoding lifts sub-THz ISAC performance","Exploiting beam squint for dual-functional sub-THz gains","Sub-THz ISAC: Squint as a resource, not a curse","Light neural network optimizes sub-THz analog precoding for ISAC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000344,"raw_usage":{"total_tokens":1871,"prompt_tokens":911,"completion_tokens":960,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":865}},"tokens_in":527,"tokens_out":960,"duration_ms":7251,"temperature":1.0,"reasoning_tokens":865,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:01:43.962497+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix one sub-THz channel realization with a user and several targets, hold phase shifters and transmit power fixed, and compute the rate-CRB Pareto frontier for the TTD setting that maximizes $\\mathrm{Cor}(h_c,\\mathbf{G})$ against a setting with lower correlation; if the lower-correlation setting achieves a strictly better rate at the same CRB or a lower CRB at the same rate, the monotonicity asserted in Proposition 1 is falsified.","supporting_citations":[{"cited_title":"On the performance gain of integrated sensing and communications: A subspace correlation per- spective,","cited_arxiv_id":null,"evidence_quote":"Supplies the reduced-subspace form of the optimal transmit covariance used in the proof of Proposition 1."},{"cited_title":"Approximating the Kullback Leibler di- vergence between gaussian mixture models,","cited_arxiv_id":null,"evidence_quote":"Cited as the basis for the claim that higher correlation implies higher beamspace peak overlap, a step central to the monotonicity result."},{"cited_title":"Mutual information maximizing wide- band multi-user (wMU) mmwave massive MIMO,","cited_arxiv_id":null,"evidence_quote":"Provides the beamspace dictionary and sparsity argument used to define the correlation measure."},{"cited_title":"Delay-phase precoding for wideband THz massive MIMO,","cited_arxiv_id":null,"evidence_quote":"Introduces the delay-phase precoding architecture whose TTD hardware the paper exploits, and serves as the communication-dedicated baseline."},{"cited_title":"Integrated sensing and communications with joint beam-squint and beam-split for mmWave /THz massive MIMO,","cited_arxiv_id":null,"evidence_quote":"Demonstrates beam-squint control for sensing in ISAC, serving as both motivation and the sensing-dedicated baseline."},{"cited_title":"Joint delay-phase precoding under true-time delay constraints in wideband sub-THz hybrid massive MIMO systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the convex reformulation used for the controlled-beam-squint CBS-ISAC benchmark."}],"review_version":1}