{"id":"4c239d9e-0a7a-464d-8033-153a2fc12d93","arxiv_id":"1908.04650","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A manifold-optimized DFRC waveform design reduces range sidelobes while improving communication sum-rate compared to the authors' prior closed-form design.","lead":"This paper proposes a waveform design for systems that do radar and communication at once, reducing troublesome radar echoes while keeping data rates up. The authors use a manifold-optimization algorithm to balance communication quality, radar sidelobes, and waveform similarity under antenna power limits.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The objective in (12) and the problem solved in (15) are not equivalent: (13) equates ρ2-weighted ISL with a ρ2-weighted similarity term, and the printed gradient (17) drops the ISL weight. The numerical gains cannot be attributed to Algorithm 1 as written.","rationale":"The reader's conditional verdict is appropriate, but the weakest point is not only the unproven convergence of RCG; before convergence is even relevant, the paper's optimization problem is internally inconsistent. Eq. (12) assigns ρ2 to PISL and ρ3 to PSIM, while Eq. (15), derived through the false identity (13), assigns ρ2 to PSIM and ρ3 to PISL. Since the numerical section quotes only one weight triple, the reported 12/17 dB sidelobe improvements cannot be connected to a definite objective. The gradient (17) also omits the ρ3 factor on the ISL term; if implemented literally, Algorithm 1 does not descend on the sidelobe component of F. This is a concrete, checkable defect rather than a disagreement with consensus. The authors may have intended a corrected version, and the general manifold-optimization idea is plausible, but the manuscript as written does not establish the central claim. I keep the reader's CONDITIONAL verdict unchanged, with the explicit conditions that the equations and weight notation be corrected, the gradient be verified, and code or iteration traces be released. The paper's own Remark admits that strict convergence analysis remains open, which is an additional but secondary concern.","tokens_in":6548,"tokens_out":11194,"duration_ms":107098,"concrete_test":"Independently re-derive the Euclidean gradient of F in (15) and compare it with (17); then implement Algorithm 1 twice on the Section V setup (N=16, K=4, L=100, P=8, Rayleigh-faded H, QPSK S, X0 from (9), weights ρ1=ρ3=0.15, ρ2=0.7): once with the printed gradient (17) and once with the corrected ρ3-weighted ISL gradient. If the corrected version reproduces the reported 12/17 dB range-sidelobe reductions while the printed version does not, the printed equations are typos; if both behave identically, the printed gradient is sufficient as written; otherwise the central numerical claim is unsupported and the manuscript must be revised before its conclusions can be accepted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing fault is that Algorithm 1 is not actually minimizing the problem the paper claims. Eq. (13) asserts ρ1 PMUI + ρ2 PISL = ||AX−B||^2_F with A=[√ρ1 H; √ρ2 I_N] and B=[√ρ1 S; √ρ2 X0], but direct expansion gives ρ1||HX−S||^2_F + ρ2||X−X0||^2_F = ρ1 PMUI + ρ2 PSIM, not ρ2 PISL. Consequently problem (15), which is derived via (13), is not a recasting of (12): in (12) the ISL weight is ρ2, while in (15) the ISL term carries ρ3, and the similarity weight is swapped correspondingly. The Section V statement that the weights are ρ1=ρ3=0.15 and ρ2=0.7 is therefore ambiguous; depending on which equation is normative, the ISL weight is either 0.7 or 0.15, so the claimed 12–17 dB sidelobe reductions are not tied to a single well-defined optimization problem. In addition, the Euclidean gradient printed in (17) omits the ρ3 prefactor on the quartic ISL term; the correct gradient for F in (15) includes 2ρ3 Σ_{p≠0} (XJ_pX^H XJ_p^H + XJ_p^H X^H XJ_p). If Algorithm 1 literally uses (17), the descent direction is independent of the sidelobe term and the reported sidelobe suppression has no explanation in the printed algorithm. The Remark's admission that strict convergence analysis of RCG remains open does not mitigate this, because the defect is in the objective itself, not in the stopping criterion or local-convergence behavior.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a waveform design for MIMO dual-functional radar-communication (DFRC) systems that jointly minimizes multiuser interference (MUI), integrated sidelobe level (ISL), and a similarity penalty with respect to a reference waveform, subject to a per-antenna power constraint. The optimization problem is non-convex, and the authors develop a Riemannian conjugate gradient (RCG) method on the complex oblique manifold. Numerical simulations for N=16, K=4, L=100 report 12 dB and 17 dB range-sidelobe reductions in omni-directional and directional designs, respectively, together with improved sum-rate over the closed-form baseline of [4]. The core idea is plausible and relevant, but the printed formulation contains algebraic inconsistencies that affect the exact algorithm as described.","tokens_in":6929,"tokens_out":6491,"duration_ms":56288,"significance":"If the formulation and algorithm were corrected, the contribution would be significant for DFRC waveform design: it is one of the first attempts to control range sidelobes explicitly in a joint radar-communication waveform, and the manifold-based approach offers a computationally efficient first-order descent method with per-antenna power control. The numerical study is reasonably comprehensive for the system size considered. However, the load-bearing derivations in Section IV are not correct as printed, so the reported gains cannot be attributed to Algorithm 1 as written. The strengths are the clear system model, the explicit integrated-sidelobe metric, and the use of the oblique manifold to handle the per-antenna constraint; the weakness is the inconsistency among equations (12), (13), (15), and (17).","major_comments":[{"comment":"The identity ρ1PMUI + ρ2PISL = ‖AX−B‖_F^2 with A=[√ρ1 H; √ρ2 I_N] and B=[√ρ1 S; √ρ2 X0] is false; direct expansion gives ρ1‖HX−S‖_F^2 + ρ2‖X−X0‖_F^2 = ρ1PMUI + ρ2PSIM. Consequently, problem (15), which is derived from this identity, is not a recasting of (12): the weights of the ISL and similarity terms are swapped between the two formulations. Since Section V sets ρ1=ρ3=0.15 and ρ2=0.7, the reported 12–17 dB sidelobe reductions are not tied to a single well-defined objective; the paper must be corrected so that (12), (13), and (15) refer to the same weighted problem.","section":"Section IV, Eq. (13)"},{"comment":"The Euclidean gradient printed for the objective in (15) omits the factor ρ3 multiplying the quartic ISL term. The correct gradient is 2 A^H(AX−B) + 2ρ3 Σ_{p≠0}(XJp X^H X Jp^H + X Jp^H X^H X Jp). As printed, the gradient has no dependence on the ISL weight, so Algorithm 1, if implemented literally with (17), would not account for the sidelobe term; the documented sidelobe suppression therefore has no explanation in the printed algorithm. The authors should either state that the numerical results were obtained with a corrected gradient (and provide it) or revise the algorithm description.","section":"Section IV, Eq. (17)"},{"comment":"The algorithm input specifies ρ_i for all i, but the construction of A and B in (14) uses only ρ1 and ρ2, and the gradient (17) does not use ρ3 at all; the role of ρ3 in the iteration is therefore undefined. The Remark's admission that strict convergence analysis remains open is a useful limitation, but it does not resolve the inconsistency between the stated problem (12) and the problem actually implemented by the algorithm.","section":"Algorithm 1 and Remark, Section IV"}],"minor_comments":[{"comment":"The term PXk(F(Xk−1)) is not well-defined because F(Xk−1) is a scalar; the standard Polak-Ribière formula on a manifold involves the projected previous gradient (or descent direction), for instance PXk(grad F(Xk−1)) or PXk(Πk−1), depending on convention.","section":"Section IV, Eq. (20)"},{"comment":"The sentence 'by solving (12) under only a small weighting factor ρ3 = 0.15' appears to treat ρ3 as the ISL weight, whereas in (12) ρ3 multiplies the similarity term PSIM; this sentence should be reworded after the weight notation is fixed.","section":"Section V"},{"comment":"The list of inputs should be aligned with the corrected notation, making explicit which weights enter the quadratic part and which enter the quartic part of the objective.","section":"Algorithm 1 input"},{"comment":"The number of Monte-Carlo trials used for the sum-rate curves is not reported, and no error bars are shown; adding this information would improve reproducibility.","section":"Section V, numerical results"},{"comment":"Minor typographical issues: 'interets' in the Introduction should be 'interests', and the phrase 'remains open problem' in the Remark should be 'remains an open problem'.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The algebraic inconsistency in Section IV is severe enough that the paper, as submitted, does not describe a well-defined algorithm for the stated problem. The authors should be asked to provide the exact objective minimized by their implementation (including the role of ρ3), correct equations (13), (15), and (17) accordingly, and rerun or confirm the simulations with the corrected algorithm. The use of the authors' own prior work [4] as the baseline is acceptable but should be acknowledged as an internal comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThe paper proposes adding range-sidelobe energy to the authors' earlier closed-form DFRC design and solving the new nonconvex problem with a Riemannian conjugate gradient method on the complex oblique manifold. That is a sensible and relevant extension: sidelobe control is a real missing piece in DFRC waveform design, and the manifold geometry is the right tool for the per-antenna power constraint. The numerical section shows big sidelobe reductions and sum-rate gains over the closed-form baseline.\n\nThe trouble is that the printed math does not support the algorithm as stated. Equation (13) claims ρ1PMUI + ρ2PISL = ‖AX−B‖²_F with A and B as in (14). Direct expansion gives ρ1‖HX−S‖²_F + ρ2‖X−X0‖²_F, i.e. ρ1PMUI + ρ2PSIM, not ρ2PISL. So problem (15), which is derived from (13), is not a recast of (12): the ISL weight in (12) is ρ2, while in (15) the ISL term carries ρ3, and the similarity weight is swapped. Section V then says ρ1=ρ3=0.15 and ρ2=0.7, so the reported 12–17 dB sidelobe reductions are not tied to one well-defined objective. Add to this that the gradient printed in (17) omits the ρ3 prefactor on the quartic ISL term. If Algorithm 1 literally uses (17), the descent direction does not reflect the sidelobe weight, and the numerical sidelobe suppression has no explanation in the printed algorithm. The authors also acknowledge that strict convergence for RCG remains open, so the near-optimality claim rests only on empirical behavior.\n\nLet me be careful: these look like fixable typos, not a fundamentally wrong approach. The intended objective—weighted MUI plus ISL plus similarity—is coherent, and the manifold machinery is standard. But as printed, a reader cannot reproduce the results. The comparison baseline is the authors' own earlier design, which is acceptable for a follow-up but means the gains are not independently benchmarked.\n\nWho is this for? Researchers working on dual-functional radar-communication waveform design. The idea is worth engaging with, but only after the authors correct the equations and, ideally, release code and data. I'd send it to peer review: a serious referee can verify the corrections and the numerics. Desk rejection would throw out a useful contribution to a busy subfield on the basis of repairable errors.","headline":"A useful sidelobe-control extension of the authors' prior DFRC design, but the printed equations do not actually implement the stated objective and need correction before the results can be trusted.","tokens_in":7456,"tokens_out":5151,"would_cite":false,"duration_ms":47505,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A MIMO DFRC waveform can be jointly optimized for radar and communication, cutting range sidelobes by up to 17 dB while improving achievable sum-rate.","keywords":["dual-function radar-communication","waveform design","range sidelobe reduction","integrated sidelobe level","manifold optimization","oblique manifold","multi-user interference","MIMO radar"],"falsifier":"Re-run the algorithm from many random initializations under the paper's settings ($N=16$, $K=4$, $L=100$, $P=8$, $\\rho_1=\\rho_3=0.15$, $\\rho_2=0.7$) and record the converged objective and range-sidelobe level; if the outcomes spread widely or the gradient norm does not fall below $10^{-6}$ within the reported tens of iterations, the claimed 12/17 dB reductions are initialization-dependent rather than a stable property of the design. A second check is to vary the weights and channel realizations and see whether convergence and the sidelobe gain survive.","tokens_in":6334,"feed_emoji":"📡","tokens_out":13891,"duration_ms":116592,"temperature":0.7,"pith_summary":"This paper is trying to establish that a MIMO dual-function radar-communication (DFRC) waveform can control its range sidelobes, the time-domain clutter that appears in nearby range bins, without giving up much communication performance. The authors add the integrated sidelobe level (ISL) to a weighted objective that penalizes multi-user interference and deviation from a reference radar waveform, and they solve the resulting nonconvex problem with a first-order Riemannian conjugate gradient method on the complex oblique manifold of matrices satisfying per-antenna power constraints. With the weights $\\rho_1=\\rho_3=0.15$ and $\\rho_2=0.7$, the reported simulations show a 12 dB sidelobe reduction for an omni-directional design and a 17 dB reduction for a directional design, together with an improved achievable sum-rate over the closed-form baseline that only matches the radar beampattern.","feed_headline":"Sidelobe-aware DFRC waveforms cut range sidelobes by up to 17 dB","feed_subtitle":"Per-antenna power stays fixed, so the optimized signal is directly usable by a dual-function radar-communication base station.","key_machinery":"The carrying object is the complex oblique manifold $\\mathcal{M}=\\{X\\in\\mathbb{C}^{N\\times L}:\\operatorname{diag}(XX^H)=\\frac{LP_T}{N}\\mathbf{1}_N\\}$, the set of waveforms that respect the per-antenna power budget. On this manifold the paper runs a Riemannian conjugate gradient (RCG) algorithm: the Euclidean gradient is projected onto the tangent space, combined with the previous descent direction using a Polak-Ribiere coefficient, and brought back to the manifold by the retraction $R_X(Z)=\\beta\\,\\operatorname{ddiag}((X+Z)(X+Z)^H)^{-1/2}(X+Z)$ with $\\beta=\\sqrt{LP_T/N}$. This machinery keeps every iterate feasible while descending, and the per-iteration complexity is dominated by the gradient computation, stated as $O(N^2PL-N^2P^2/2)$.","core_discovery":"The central claim is that range-sidelobe suppression can be folded directly into the DFRC waveform optimization without requiring the waveform covariance to equal a fixed radar covariance. The paper minimizes $F(X)=\\rho_1\\|HX-S\\|_F^2+\\rho_2\\sum_{p=-P,p\\neq 0}^{P}\\|XJ_pX^H\\|_F^2+\\rho_3\\|X-X_0\\|_F^2$ subject to $\\operatorname{diag}(XX^H)=\\frac{LP_T}{N}\\mathbf{1}_N$, where the first term is multi-user interference, the second is the integrated sidelobe level, and the third keeps the waveform close to a radar-friendly reference. Because the feasible set is a complex oblique manifold, the algorithm projects the Euclidean gradient onto the tangent space, forms a conjugate descent direction via the Polak-Ribiere coefficient, and retracts back to the manifold. The reported result is that this relaxed, weighted design beats the closed-form beampattern-constrained design in both communication sum-rate and range sidelobe level, at the cost of a small mismatch in the spatial beampattern.","pith_inferences":["Beyond the paper, the same manifold formulation is a natural template for adding other radar constraints, such as Doppler sidelobe levels or peak-to-average-power ratio, by changing the objective or the retraction while keeping the per-antenna power constraint built in.","Beyond the paper, initializing at the closed-form solution (10) instead of a random point may improve convergence stability and give a more direct comparison, since that solution is the reference the relaxation is built around.","Beyond the paper, the reported 12/17 dB gains are tied to one weight setting and simulation choice; a deployment-oriented test would vary antenna counts, user numbers, block lengths, and weights to see whether the sidelobe reductions persist and whether a principled weight-selection rule emerges."],"forward_implications":["Should the claim hold, DFRC base stations can produce a single waveform that serves downlink users and suppresses range sidelobes at the same time, with the reported 12 dB omni-directional and 17 dB directional sidelobe reductions.","The weighted objective relaxes the rigid covariance equality of the closed-form design, so the radar beampattern is only approximately matched; the authors argue this mismatch is acceptable given the improved sum-rate.","The per-antenna power constraint is maintained exactly at every iteration, so the resulting waveform is directly transmittable with the intended power distribution.","Because the per-iteration cost is a matrix gradient computation, the design is computationally light enough to permit re-optimization as the communication channel changes, assuming the observed convergence behavior persists."],"supporting_citations":[{"why":"Supplies the closed-form beampattern-constrained design (9) that is the benchmark and the source of the reference waveform $X_0$ and the sum-rate formula.","marker":"[4]"},{"why":"Supplies the complex oblique manifold and RCG machinery, including tangent-space projection and retraction, and is the reference for the paper's admitted open convergence question.","marker":"[8]"},{"why":"Supplies the MIMO radar beampattern definition and the directional covariance construction used for the reference beampatterns.","marker":"[5]"},{"why":"Justifies minimizing multi-user interference by linking it to a lower bound on achievable sum-rate, grounding the communication metric.","marker":"[6]"},{"why":"Gives the target-echo and clutter model with the temporal shift matrix $J_p$ that defines the integrated sidelobe level.","marker":"[7]"},{"why":"Provides the classical conjugate gradient and Armijo line-search background used in the descent-direction and step-size updates.","marker":"[9]"}],"fun_headline_variants":["DFRC waveform optimization trims range sidelobes by 17 dB","Manifold-based DFRC design reduces sidelobes, increases rate","Weighted DFRC optimization improves link rate and cuts sidelobes","Range sidelobes reduced in DFRC via manifold optimization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the iterative algorithm actually finding a low-enough local minimum of the nonconvex objective in tens of iterations; the paper states that strict convergence analysis remains open, so a stall in a poor local valley would undo the claimed sidelobe and rate gains.","fun_headline_variants_meta":{"raw":{"variants":["DFRC waveform optimization trims range sidelobes by 17 dB","Manifold-based DFRC design reduces sidelobes, increases rate","Weighted DFRC optimization improves link rate and cuts sidelobes","Range sidelobes reduced in DFRC via manifold optimization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001305,"raw_usage":{"total_tokens":5293,"prompt_tokens":885,"completion_tokens":4408,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":4330}},"tokens_in":501,"tokens_out":4408,"duration_ms":30851,"temperature":1.0,"reasoning_tokens":4330,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:36:10.800011+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the algorithm from many random initializations under the paper's settings ($N=16$, $K=4$, $L=100$, $P=8$, $\\rho_1=\\rho_3=0.15$, $\\rho_2=0.7$) and record the converged objective and range-sidelobe level; if the outcomes spread widely or the gradient norm does not fall below $10^{-6}$ within the reported tens of iterations, the claimed 12/17 dB reductions are initialization-dependent rather than a stable property of the design. A second check is to vary the weights and channel realizations and see whether convergence and the sidelobe gain survive.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the closed-form beampattern-constrained design (9) that is the benchmark and the source of the reference waveform $X_0$ and the sum-rate formula."},{"cited_title":"Hua and S","cited_arxiv_id":null,"evidence_quote":"Supplies the complex oblique manifold and RCG machinery, including tangent-space projection and retraction, and is the reference for the paper's admitted open convergence question."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MIMO radar beampattern definition and the directional covariance construction used for the reference beampatterns."},{"cited_title":"Li and P","cited_arxiv_id":null,"evidence_quote":"Justifies minimizing multi-user interference by linking it to a lower bound on achievable sum-rate, grounding the communication metric."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the target-echo and clutter model with the temporal shift matrix $J_p$ that defines the integrated sidelobe level."},{"cited_title":"Absil, R","cited_arxiv_id":null,"evidence_quote":"Provides the classical conjugate gradient and Armijo line-search background used in the descent-direction and step-size updates."}],"review_version":1}