{"id":"6122bc71-384c-4ab2-903c-32f96003f82d","arxiv_id":"2507.12881","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A robust beamforming design for secure near-field ISAC systems, using S-Procedure and sequential rank-one constraint relaxation, maximizes the minimum sensing beampattern gain under channel uncertainty.","lead":"The authors design a beamforming scheme for a near-field integrated sensing and communication system that must serve users, detect targets, and protect against eavesdroppers. The scheme is robust to imperfect channel knowledge and, in simulations, outperforms existing baseline methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SROCR stopping criterion can declare convergence when v=1 is reached via the δ offset, so the returned W_k may not be rank-one; the claimed extraction of beamforming vectors is then invalid.","rationale":"The S-Procedure transformations in Sec. III-A are individually correct: each quadratic uncertainty set is a norm ball with nonempty interior, so the single-constraint S-lemma applies losslessly, and the LMIs (8) and (10) match the standard forms. The complexity expression is also dimensionally consistent. The load-bearing weakness I find is in the SROCR termination (Alg. 1, Sec. III-B). The update v = min(1, λmax/tr + δ) with δ(0)=0.1 means v hits 1 while the eigenvalue ratio is only 0.9. If Step 6's second condition |t(i)-t(i-1)| ≤ ε is met in that same iteration, the algorithm outputs a PSD W_k with rank > 1, contradicting Remark 1. Without rank-one W_k, the 'beamforming vectors' w_k do not exist, so the proposed scheme is not a feasible beamforming design and the numerical gains in Figs. 2-4 are not achieved by any actual transmission. The reader's concern about bounded CSI errors is a modeling assumption shared with most robust beamforming papers; the rank-one termination issue is an internal correctness risk that can be settled by a direct numerical check. Since the issue is fixable (e.g., continue until λmax/tr ≈ 1 or check rank explicitly), a conditional accept with a required revision is appropriate.","tokens_in":9952,"tokens_out":19229,"duration_ms":194803,"concrete_test":"Run Algorithm 1 with the default parameters of Sec. IV (N=64, K=4, L=2, M=2, η=0.1) over 50 independent channel realizations. At the first iteration satisfying the Step-6 conditions, record r_k = λmax(W_k)/tr(W_k) for each k. If any r_k is materially below 1 (e.g., < 0.99, or below 1−δ^(0)), then the stopping rule does not certify rank-one, so the output cannot be used to form w_k. As a second check, inspect whether the iteration at which v first reaches 1 has |t^(i)-t^(i-1)| ≤ ε; if this occurs before rank-one is attained, the flaw is realized.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 (Sec. III-B) updates the relaxation parameter as v^(i)_k = min(1, λmax(W^(i)_k)/tr(W^(i)_k) + δ^(i)_k) and terminates when v^(i)_k = 1 for all k and |t^(i)-t^(i-1)| ≤ ε (Step 6). Because δ^(0) = 0.1 and δ may remain positive, v=1 is reached as soon as the eigenvalue ratio exceeds 0.9; the matrix W_k need not be rank-one. Remark 1 states that 'when v^(i)_k = 1, we have rank(W_k) = 1', which is inconsistent with the update rule. If the objective t has already converged at that iteration, the algorithm outputs a PSD W_k of rank > 1. Such a W_k cannot be factored as w_k w_k^H, so no beamforming vector is produced, and the performance comparisons in Sec. IV do not correspond to a valid transmission scheme. The cited convergence guarantee to a KKT point of the rank-one constrained problem (P2) therefore does not follow from the stated stopping rule.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The letter studies robust transmit beamforming for a near-field secure integrated sensing and communication (ISAC) system with multiple CUs, sensing targets, and eavesdroppers under norm-bounded CSI errors. The authors formulate a max-min optimization that maximizes the worst-case sensing beampattern gain subject to worst-case SINR lower bounds for CUs, SINR upper bounds for eavesdroppers, and a total power constraint. They apply the S-Procedure to convert the semi-infinite CSI-uncertainty constraints into LMIs, then use the sequential rank-one constraint relaxation (SROCR) algorithm to handle the rank-one beamforming constraints. Numerical results are presented to show convergence and performance gains over SDR, information-only, and far-field baselines.","tokens_in":10186,"tokens_out":9366,"duration_ms":110729,"significance":"If the proposed algorithm as stated produced valid rank-one beamformers and the numerical comparisons are reproducible, the work would be a useful extension of robust secure beamforming to near-field ISAC. The S-Procedure transformations in Section III-A are standard and appear algebraically correct, and the paper provides a complexity estimate and explicitly discusses limitations of near-field CSI acquisition. However, the rank-one termination flaw in Algorithm 1 and the under-specified numerical baselines currently prevent the central claims from being accepted as stated. The contribution is potentially fixable, but the validation path requires substantive changes to the algorithm and the experiments.","major_comments":[{"comment":"The stopping rule in Algorithm 1 (Step 6) can terminate with a non-rank-one W_k. In Step 4, v^(i)_k = min(1, lambda_max(W^(i)_k)/tr(W^(i)_k) + delta^(i)_k), and delta^(0)=0.1 is halved only when (P3) is infeasible. If (P3) remains feasible, delta stays 0.1, so v=1 is reached when lambda_max/tr >= 0.9, not only when rank(W_k)=1. Since the loop stops immediately after v=1, the returned W_k need not be rank-one, and Remark 1's assertion that 'when v^(i)_k = 1, we have rank(W_k)=1' is inconsistent with the update. Such a W_k cannot be factored as w_k w_k^H, so no valid beamforming vector is delivered and the objective t in Figs. 2-4 does not correspond to a feasible solution of (P1). The algorithm should either continue until lambda_max/tr(W_k) >= 1-epsilon, solve one final SDP with v=1, or verify rank-one status and apply a valid extraction procedure before reporting results.","section":"Sec. III-B, Algorithm 1, Remark 1"},{"comment":"The claimed convergence to a KKT point of the rank-one constrained problem (P2) relies on Theorem 1 of [16], but the authors do not verify that the hypotheses of that theorem hold for Algorithm 1 as written. In particular, the update of delta^(i)_k does not drive delta to zero on the feasible path, and the criterion v=1 can be satisfied without the rank-one relaxation being tight. Without a delta-to-zero mechanism or an explicit check that the rank-one constraint is active, the convergence guarantee is not established.","section":"Sec. III-B, Remark 1"},{"comment":"The numerical evaluation does not specify whether the plotted points are single random deployments or averages, no error bars or confidence intervals are given, and the SDR baseline is not fully defined (in particular, no rank-one recovery or Gaussian randomization step is described). The far-field and LoS/NLoS curves are also not accompanied by the exact channel model used. Since the abstract's claim of outperformance is empirical, these omissions prevent the reader from assessing the statistical significance and reproducibility of the comparison.","section":"Sec. IV, Figs. 2-4"}],"minor_comments":[{"comment":"The conclusion contains a typo: 'SROCR agorithm' should be 'SROCR algorithm'.","section":"Sec. V"},{"comment":"The sentence introducing Fig. 4 repeats 'versus eta_C,k and eta_C,k'; the second quantity should presumably be eta_E,l.","section":"Sec. IV, Fig. 4"},{"comment":"Remark 1's statement that 'the trace of W_k approaches one' is not implied by any normalization in (P3); the SROCR constraint compares lambda_max to tr(W_k) but does not fix tr(W_k). Please rephrase.","section":"Sec. III-B, Remark 1"},{"comment":"Footnote 2 says the design can be extended to LoS/NLoS scenarios, but Fig. 3(c) already reports a LoS/NLoS curve; the NLoS channel model used in that figure should be given explicitly.","section":"Sec. II-A and Fig. 3(c)"},{"comment":"The phrase 'near-field unlink/downlink ISAC performance' appears to contain a typo ('unlink' should be 'uplink').","section":"Sec. I"}],"recommendation":"major_revision","confidential_remarks":"The main technical issue is fixable by revising the stopping criterion and re-running the simulations; the manuscript is within scope for a communications letter. I would not reject on the basis of disagreement with the near-field channel model, since the paper acknowledges its limitations. Please ensure the authors address the rank-one extraction issue and the baseline specification before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper takes a real problem — robust secure beamforming for near-field ISAC with multiple users, targets, and eavesdroppers under norm-bounded CSI errors — and applies standard tools (S-Procedure to LMI, SROCR for rank-one constraints) in a way that is mostly algebraically correct. The formulation looks new relative to the cited work, and the complexity analysis is a useful addition. If the algorithm is repaired, this is a reasonable letter contribution.\n\nThe soft spot that matters is in Algorithm 1. The update rule sets v_k = min(1, λmax(W_k)/tr(W_k) + δ_k) with δ_k initialized at 0.1 and kept positive until the SDP is infeasible. Stopping triggers when v_k = 1 for all k and t has converged. With δ positive, v crosses 1 as soon as the eigenvalue ratio exceeds 0.9. The returned W_k need not be rank-one. Remark 1 says v=1 implies rank one, which is only true after solving the next subproblem with that v. As written, the algorithm can terminate without producing any w_k, and the curves in Section IV may be computing a value that no actual beamforming vector achieves. This is not a fatal flaw in the overall approach — one extra SROCR iteration with v=1 would force rank-one — but as published, the performance claims need rechecking.\n\nOther issues are minor. The numerics lack error bars or any indication of how many channel realizations were used; the SDR baseline is not fully specified; and the bounded-error model, while standard, is doing a lot of work. The paper acknowledges near-field CSI acquisition remains challenging, and Fig. 4(c) honestly shows the design is very sensitive to target channel error.\n\nBottom line: this deserves a serious referee, but I would not accept until the stopping criterion is fixed and the simulations rerun. The new formulation is worth engaging with.","headline":"A new application of S-Procedure and SROCR to robust near-field secure ISAC, but the SROCR stopping rule can declare convergence without a rank-one solution, so the headline numbers need verification.","tokens_in":10664,"tokens_out":3370,"would_cite":false,"duration_ms":41192,"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 proposes a robust beamforming design for near-field secure ISAC that maximizes the minimum sensing beampattern gain while enforcing worst-case SINR constraints under bounded CSI errors, and reports that it outperforms…","keywords":["near-field ISAC","secure beamforming","robust beamforming","S-Procedure","SROCR","sensing beampattern gain","worst-case SINR","CSI uncertainty"],"falsifier":"Run the proposed design with a fixed scenario, such as $N=64$, $K=4$, $L=2$, $M=2$, and $P_0=30$ dBm, then draw many error realizations uniformly from the declared norm balls of radii $\\epsilon_{C,k}$, $\\epsilon_{E,l}$, and $\\epsilon_{T,m}$; if any realization drives a CU SINR below $\\bar{\\gamma}_{C,k}$ or an eavesdropper SINR above $\\bar{\\gamma}_{E,l,k}$, the S-Procedure reformulation is not actually worst-case, and the robustness claim fails.","tokens_in":9797,"feed_emoji":"📡","tokens_out":6281,"duration_ms":65277,"temperature":0.7,"pith_summary":"This letter addresses a practical gap: near-field ISAC systems that must simultaneously communicate with users, sense targets, and keep eavesdroppers from intercepting, while the base station only holds imperfect channel estimates. The paper formulates the design as maximizing the minimum sensing beampattern gain subject to worst-case SINR constraints for every user and every eavesdropper under bounded channel-error norms, together with a transmit-power budget. Because the resulting problem is non-convex with semi-infinite constraints, the authors use the S-Procedure to turn each uncertainty constraint into a linear matrix inequality, then apply a sequential rank-one constraint relaxation (SROCR) method to enforce the rank-one beamforming structure. They report that this SROCR-based design converges within a few iterations and achieves a higher worst-case sensing beampattern gain than semidefinite relaxation and information-only baselines while preserving security constraints. If correct, this gives a tractable robust beamforming recipe for near-field ISAC under imperfect CSI.","feed_headline":"Robust beamforming beats SDR for near-field secure ISAC","feed_subtitle":"Worst-case SINR is kept for users and eavesdroppers while sensing gain is maximized under bounded CSI errors.","key_machinery":"The two load-bearing tools are the S-Procedure and the SROCR algorithm. The S-Procedure converts each semi-infinite constraint, an inequality that must hold for every channel error inside a norm ball, into an equivalent linear matrix inequality by introducing a nonnegative slack variable, provided the error set is convex with nonempty interior. SROCR then handles the rank-one constraints by replacing $\\operatorname{rank}(W_k)=1$ with the equivalent condition $\\lambda_{\\max}(W_k)=\\operatorname{tr}(W_k)$, relaxing it to $u^{(i-1)H}_{k,\\max} W_k u^{(i-1)}_{k,\\max} \\geq v^{(i-1)}_k \\operatorname{tr}(W_k)$ using the leading eigenvector from the previous iteration, and increasing $v$ toward one until a rank-one solution is reached. The near-field steering vector $\\alpha(r,\\theta)$ with spherical-wavefront phases carries the geometric model that distinguishes this problem from far-field beamforming.","core_discovery":"The central claim is that robust secure beamforming for near-field ISAC can be formulated as a rank-one constrained semidefinite program with linear matrix inequalities obtained from the S-Procedure, and that this program can be solved by iterating a convex relaxation that gradually forces each communication beamforming matrix $W_k$ to satisfy $\\lambda_{\\max}(W_k)=\\operatorname{tr}(W_k)$, i.e., to have rank one. The paper shows numerically that this SROCR-based solution converges to a local Karush-Kuhn-Tucker point in a few iterations, and that for the tested configurations it yields a larger minimum sensing beampattern gain than the SDR baseline, which drops rank-one constraints and may fail to return a feasible beamformer, and larger than an information-only transmission without dedicated sensing signals. Robustness is demonstrated by degrading gracefully as the normalized CSI error bound grows, with the worst-case sensing gain eventually collapsing as the target-channel error grows.","pith_inferences":["Because the S-Procedure step requires only convex error sets with nonempty interior, the same LMI structure could be reused for other convex uncertainty sets, such as ellipsoidal or elementwise bounds, without re-deriving the algorithm; only the LMI entries would change.","The reported sensitivity to target-channel error suggests that in practice more estimation effort should go into target position accuracy than into user or eavesdropper channel accuracy.","If the bounded-error assumption were replaced by stochastic errors, the S-Procedure step would need to give way to chance-constrained or distributionally robust reformulations, but the SROCR outer loop over rank-one constraints would likely carry over unchanged.","The same SROCR machinery should transfer to related near-field secure ISAC objectives, such as maximizing secrecy rate or minimizing power subject to sensing and secrecy constraints, since only the objective and the LMI definitions would change."],"forward_implications":["For a fixed scenario, the proposed design satisfies every CU worst-case SINR constraint and every eavesdropper leakage cap under any error realization inside the declared norm balls, while maximizing the minimum target beampattern gain.","Compared with SDR, the SROCR method returns an actual rank-one beamforming matrix rather than a possibly high-rank relaxation, so the solution is directly implementable.","Jointly allocating power to dedicated sensing signals and communication beams improves worst-case sensing gain over information-only transmission, so the sensing covariance $R_0$ is not wasted.","The design degrades gracefully with increasing CSI error at users and eavesdroppers, up to normalized error $0.9$ in the simulations, but is most sensitive to target-channel error, where the beampattern gain collapses as that error grows.","The per-iteration complexity is polynomial in the array size $N$ and the numbers of users, targets, and eavesdroppers, so the approach remains feasible for typical system sizes."],"supporting_citations":[{"why":"Supplies the S-Procedure equivalence theorem used to convert each semi-infinite worst-case constraint into an LMI.","marker":"[15]"},{"why":"Supplies the SROCR algorithm and the convergence guarantee to a local KKT point for the rank-one constrained problem.","marker":"[16]"},{"why":"Provides the bounded CSI error model with deterministic norm constraints and the SDP complexity expression adopted in the paper.","marker":"[11]"},{"why":"Supplies the near-field beam training method that motivates the paper's CSI acquisition discussion and the proposed two-stage extension.","marker":"[13]"},{"why":"Defines the near-field region and Fresnel and Rayleigh distances used in the spherical-wavefront channel model.","marker":"[4]"}],"fun_headline_variants":["Robust beamforming secures near-field ISAC against eavesdroppers","Near-field ISAC beamforming: robust to channel errors, beats SDR","Robust ISAC beamforming: max sensing, min eavesdropper SINR","Robust beamforming for secure near-field ISAC under uncertainty"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every channel error lies inside a known norm ball of a given radius, and the near-field channel is dominated by the line-of-sight path; if real errors fall outside that ball or the LoS model is wrong, the worst-case guarantees do not hold.","fun_headline_variants_meta":{"raw":{"variants":["Robust beamforming secures near-field ISAC against eavesdroppers","Near-field ISAC beamforming: robust to channel errors, beats SDR","Robust ISAC beamforming: max sensing, min eavesdropper SINR","Robust beamforming for secure near-field ISAC under uncertainty"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000916,"raw_usage":{"total_tokens":3920,"prompt_tokens":923,"completion_tokens":2997,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":2917}},"tokens_in":539,"tokens_out":2997,"duration_ms":22807,"temperature":1.0,"reasoning_tokens":2917,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:36:12.790481+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the proposed design with a fixed scenario, such as $N=64$, $K=4$, $L=2$, $M=2$, and $P_0=30$ dBm, then draw many error realizations uniformly from the declared norm balls of radii $\\epsilon_{C,k}$, $\\epsilon_{E,l}$, and $\\epsilon_{T,m}$; if any realization drives a CU SINR below $\\bar{\\gamma}_{C,k}$ or an eavesdropper SINR above $\\bar{\\gamma}_{E,l,k}$, the S-Procedure reformulation is not actually worst-case, and the robustness claim fails.","supporting_citations":[{"cited_title":"Multivariate nonn egative quadratic mappings,","cited_arxiv_id":null,"evidence_quote":"Supplies the S-Procedure equivalence theorem used to convert each semi-infinite worst-case constraint into an LMI."},{"cited_title":"A sequential constr aint relaxation algorithm for rank-one constrained problems,","cited_arxiv_id":null,"evidence_quote":"Supplies the SROCR algorithm and the convergence guarantee to a local KKT point for the rank-one constrained problem."},{"cited_title":"A fr amework of robust transmission design for IRS-aided MISO communica tions with imperfect cascaded channels,","cited_arxiv_id":null,"evidence_quote":"Provides the bounded CSI error model with deterministic norm constraints and the SDP complexity expression adopted in the paper."},{"cited_title":"Fast near-field beam trainin g for extremely large-scale array,","cited_arxiv_id":null,"evidence_quote":"Supplies the near-field beam training method that motivates the paper's CSI acquisition discussion and the proposed two-stage extension."},{"cited_title":"Near-Field Communications: A Comprehensive Survey","cited_arxiv_id":"2401.05900","evidence_quote":"Defines the near-field region and Fresnel and Rayleigh distances used in the spherical-wavefront channel model."}],"review_version":1}