{"id":"9cd8ddcf-f7a0-47a6-a053-197267d5a24b","arxiv_id":"2507.14982","paper_version":3,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For an ISAC base station serving K users and estimating L parameters, at most K + sqrt(L(L+1)/2) beamformers are needed if users cancel sensing interference, and at most sqrt(K^2 + L(L+1)/2) if they cannot.","lead":"This paper proves simple formulas for the maximum number of beamforming signals a base station needs when it must both talk to K users and estimate L sensing parameters at the same time. The formulas show that combining both tasks needs fewer beams than doing each task separately, which can simplify future 6G integrated sensing and communications hardware.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The rank-reduction scaling step in Appendix B is under-proved: the paper does not show the max-magnitude solution component can be taken as an eigenvalue of M'_s, but the missing sign argument is supplied by the positive SINR targets and is repairable.","rationale":"The central claim is the pair of upper bounds in Theorems 1 and 2. Their proofs are driven by iterative rank reduction, so Lemmas 1 and 3 are the load-bearing components. The single most delicate point is the scaling in Appendix B: the paper chooses δ with |δ| equal to the largest absolute value among the a'_k's and the eigenvalues of M'_s, and needs I - M_s to be both PSD and singular. As written this does not follow if the maximum is attained by an a'_k, because then the corresponding d_k is zero and no eigenvalue of M'_s equals δ. I checked whether this case can occur. It cannot for any nonzero solution when γ' > 0 and σ² > 0: scaling so that a_k = 1 turns the k-th SINR equation into a contradiction, because the signal-to-target ratio equals the interference sum plus σ², while the equation forces it to be no larger than the interference sum. Hence the maximum component is necessarily an eigenvalue, and the scaling step is valid. Lemma 3 is cleaner because it has only M variables and no SINR unknowns. The d-quadratic extension in Theorem 3 inherits the same reasoning, and the examples are consistent with the bounds rather than fitted to them. The reader identified exactly the right place to look, and the concern is real as a gap in exposition, but it is repairable without changing any theorem statement or bound. I therefore find no fatal defect, and the reader's ACCEPT verdict remains appropriate.","tokens_in":40601,"tokens_out":37686,"duration_ms":436309,"concrete_test":"Independently re-derive the scaling step in Appendix B. Take any nonzero solution of the homogeneous system (109)-(110) and test whether the maximum absolute value can be attained by an a'_k component. For a candidate solution with |a'_k| = δ, divide by δ so that a_k = 1, then use the k-th SINR equation to conclude |h_k^H vhat_k|² / γ'_k = |Σ_{n≠k} a_n h_k^H vhat_n|² ≤ Σ_{n≠k} |h_k^H vhat_n|², contradicting γ'_k = |h_k^H vhat_k|² / (Σ_{n≠k} |h_k^H vhat_n|² + σ²). If this contradiction holds for every nonzero solution, Lemma 1 is sound; if a counterexample is found, the rank-reduction engine of Theorem 1 would need revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorems 1 and 2 stand or fall on Lemmas 1 and 3. In Appendix B, after obtaining a nonzero solution (a'_k, M'_s) of (109)-(110), the proof scales by δ with |δ| equal to the largest absolute value among the a'_k's and the eigenvalues of M'_s, and asserts that I - M_s is both PSD and singular. This is not immediate: if the maximum magnitude is attained by an a'_k rather than an eigenvalue of M'_s, scaling by δ makes d_k = 0 and does not make I - M_s singular. The paper does not explicitly rule out this case. However, such a case is impossible for any nonzero solution whenever the SINR targets are positive and σ² > 0. After scaling so that a_k = 1, the k-th SINR equation in (110) gives |h_k^H vhat_k|² / γ'_k = Σ_{n≠k} a_n |h_k^H vhat_n|². Taking absolute values and using |a_n| ≤ 1 yields |h_k^H vhat_k|² / γ'_k ≤ Σ_{n≠k} |h_k^H vhat_n|², which contradicts γ'_k = |h_k^H vhat_k|² / (Σ_{n≠k} |h_k^H vhat_n|² + σ²) because the denominator exceeds the interference sum by σ² > 0. Hence the maximum component is necessarily an eigenvalue, and the scaling step is valid. Lemma 3 has no a_k variables, so it does not suffer this subtlety. Because the missing argument is short and does not change the bound, I do not treat this as a fatal flaw.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a downlink ISAC system in which a base station uses linear beamforming to serve K single-antenna communication users while estimating L real-valued sensing parameters, and asks how many beamformers are needed to attain the same joint performance as a full set of N_T + K beamformers. The central results are universal upper bounds on this minimum number: with interference cancellation at the users, at most K + floor(sqrt(L(L+1)/2)) beamformers suffice; without interference cancellation, at most floor(sqrt(K^2 + L(L+1)/2)) suffice. These bounds are extended to a class of d-quadratic sensing metrics, yielding K + floor(sqrt(d)) and floor(sqrt(K^2 + d)), respectively. The paper also derives applications to channel-matrix estimation, estimation of LoS target parameters, radar SNR/SCNR detection, and beam-pattern matching, with several tightness results, most notably the exact characterization for one LoS target without interference cancellation.","tokens_in":40942,"tokens_out":20884,"duration_ms":248987,"significance":"If the proofs are completed as claimed, this is a substantial contribution. The bounds are parameter-free, channel-independent, and expressed directly in terms of the number of users K, the number of sensing parameters L, or the number of quadratic terms d, which makes them directly useful for system design and for avoiding over-parameterized SDR-based optimization. The paper improves on prior results such as the 2N_tr bound of [7] and generalizes earlier special-case results of [8], [9], [14]. The derivation is self-contained, builds on standard rank-reduction ideas of Pataki and Huang-Palomar, and the authors explicitly disclose the relation to their conference version [1] and to the concurrent work [18]. The numerical examples are honest about the gap between worst-case bounds and typical behavior, and several tightness cases are identified. The main reservations concern proof completeness in three load-bearing places rather than the validity of the overall approach.","major_comments":[{"comment":"The proof that I_Ns - M_s is singular is under-specified. The text chooses delta with |delta| equal to the largest magnitude among the a'_k and the eigenvalues of M'_s, then asserts that if I_Ns - M_s were nonsingular, delta would equal some a'_i and the corresponding v'_i would be zero, violating the SINR constraint. As written, this does not explain why the case of a largest-magnitude eigenvalue is handled, nor why delta = a'_i is impossible. The missing argument is: after scaling so that a_i = 1, the i-th equation of (110) gives |h_i^H vhat_i|^2 / gamma'_i = sum_{n != i} a_n |h_i^H vhat_n|^2, while the definition of gamma'_i gives |h_i^H vhat_i|^2 / gamma'_i = sum_{n != i} |h_i^H vhat_n|^2 + sigma^2, a contradiction because |a_n| <= 1 and sigma^2 > 0. Please add this inequality (or an equivalent argument) and specify the sign convention for delta. This is load-bearing because Lemma 1 is the engine of Theorem 1.","section":"Appendix B, proof of Lemma 1, scaling step after Eqs. (112)-(113)"},{"comment":"The proof invokes Lemma 1 for d-quadratic constraints, but Lemma 1 is stated and proved only for the BFIM constraint J_V = J, and its proof uses strong duality of P_IC_N together with the dual stationarity conditions (115)-(117) to establish power preservation via (118). For the d-quadratic problem (57a)-(57c), the analogous strong-duality result, the analogous dual problem, and the power-preservation identity are not stated or proved. The extension is plausible and likely follows by repeating the SDR-tightness argument of Lemma 4, since the constraints tr(Q_i V V^H) = c_i are linear in R = V V^H, but as the text stands Theorem 3 rests on an unproved generalization of the central rank-reduction lemma. Please add a formal lemma covering d linear quadratic constraints, or expand the proof of Theorem 3 to include the strong-duality and scaling steps for both the IC and NIC cases.","section":"Section IV-B, proof of Theorem 3"},{"comment":"The proof of Lemma 2 uses a perturbation argument in which the dual problem D_epsilon is solved for every epsilon > 0, and then a limit point V'_c of the sequence {V'_{c,epsilon}} is taken as epsilon -> 0. The text asserts that 'since epsilon -> 0, V'_c must achieve the same optimal value' as the unperturbed problem (132), but no continuity or compactness argument is supplied to justify convergence of the optimal values or feasibility and optimality of the limit point. Since Lemma 2 provides the orthogonality condition (33) on which Theorem 2 depends, this is a load-bearing step. Please add a short argument showing that the optimal value of R_epsilon converges to that of R_0 and that the limit point is optimal for (127), or replace the perturbation step with a direct KKT-based proof.","section":"Appendix D, proof of Lemma 2, limiting argument near Eqs. (139)-(143)"}],"minor_comments":[{"comment":"The identity matrices in (114b) and (114c) should be I_{N_T}, not I_N, since the matrices tilde{G}_{i,j} and the channel vectors h_k act on the N_T-dimensional transmit space; the dual problem is independent of the number of beamformers N.","section":"Appendix B, dual problem (114)"},{"comment":"The displayed constraint uses vSINR_IC, but the theorem is for the no-interference-cancellation scenario and should use vSINR_NIC.","section":"Theorem 2, Eq. (39)"},{"comment":"There is a factor-of-two inconsistency with Eq. (66): (66) defines J_V with 2Upsilon/sigma^2 in the off-diagonal blocks, but Appendix F defines B_1, B_2, and bar{J} with Upsilon/sigma^2, and the final trace expression misses the factor 2 inside the inverse. Please align Appendix F with Eq. (67a), which appears to have the correct factor.","section":"Appendix F"},{"comment":"The label 'Hypoenuse Bound' in the figure legend should read 'Hypotenuse Bound'.","section":"Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The technical core appears sound and the results are valuable, but the paper currently leaves several load-bearing proof steps implicit: the scaling argument in Lemma 1, the generalization of the rank-reduction lemmas to d-quadratic metrics in Theorem 3, and the limiting argument in Lemma 2. All three appear repairable without changing the stated bounds, which is why I recommend major revision rather than rejection. The authors should also correct the small notational inconsistencies listed in the minor comments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThis paper gives clean worst-case upper bounds on the number of beamformers needed for integrated sensing and communication: K + sqrt(L(L+1)/2) with interference cancellation, and sqrt(K^2 + L(L+1)/2) without. Those are genuinely new, especially the hypotenuse bound, which is tighter than the sum bound and leads to the nice corollary that communication beamformers alone suffice once K >= L(L+1)/4. The d-quadratic extension covers radar SNR/SCNR and beam-pattern metrics and is a useful unification.\n\nThe proof strategy is sound. The key is a rank-reduction procedure: for the no-cancellation case, Lemma 2 first shows the sensing beams can be taken orthogonal to all communication channels, which makes the SINR constraints invariant to the transformation, and then a counting argument gives the bound. The appendices are detailed and the strong-duality groundwork is handled properly. I also appreciate that the numerical section says plainly that these are worst-case bounds and that typical numbers are often much smaller.\n\nThe soft spot is in Appendix B, Lemma 1. The paper chooses a scaling delta whose magnitude is the largest among the solution components, and then asserts I - M_s is PSD and singular. To get singularity, it needs delta to be one of the eigenvalues of M'_s, not one of the a'_k. The proof sketches a contradiction for the a'_k case, but it doesn't fully spell out the sign choice when the largest magnitude is attained by a negative component. The stress-test note gives a correct fix: if scaling makes d_k = 0 for some user, that beamformer vanishes and the positive SINR target is violated, so the maximum must be an eigenvalue. With that repair, the lemma goes through. This is a minor presentation gap, not a structural flaw.\n\nA smaller issue: the d-quadratic generalization is largely mechanical once the BCRB case is done. That's fine, but the novelty there is modest.\n\nOverall, the counting question is basic and the answer is clean. I'd bring it to a reading group focused on ISAC or MIMO beamforming, and I'd cite it. The paper deserves a serious referee; the math is self-contained and the claims are explicit about what is tight and what is not.","headline":"Solid answer to a basic ISAC counting question; the main theorems hold up after a minor fix in one lemma's scaling step.","tokens_in":41478,"tokens_out":3642,"would_cite":true,"duration_ms":43832,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C22","94A12"],"pacs":[],"model":"deepseek-v4-flash","headline":"At most $K+\\lfloor\\sqrt{L(L+1)/2}\\rfloor$ beamformers suffice for simultaneous sensing and communication; without interference cancellation the cap is $\\lfloor\\sqrt{K^2+L(L+1)/2}\\rfloor$.","keywords":["integrated sensing and communications","beamforming","minimum number of beamformers","Cramér-Rao bound","Bayesian Fisher information matrix","rank reduction","semidefinite relaxation","MIMO radar"],"falsifier":"A counterexample would be a feasible ISAC instance with $N_s^2 > L(L+1)/2$ in which every nonzero solution of the homogeneous system (109)-(110) fails, after scaling, to make $I_{N_s}-M_s$ positive semidefinite and singular with $1-a_k \\ge 0$ for all $k$; searching random channel and target realizations for such a case would settle whether the sum bound is universal.","tokens_in":40388,"feed_emoji":"📡","tokens_out":12616,"duration_ms":125608,"temperature":0.7,"pith_summary":"This paper asks how many transmit beamformers (spatial weight vectors) a base station needs to simultaneously serve $K$ communication users and estimate $L$ environmental parameters. It proves a pair of universal upper bounds: with interference cancellation at the users, $K+\\lfloor\\sqrt{L(L+1)/2}\\rfloor$ beamformers always suffice; without cancellation, $\\lfloor\\sqrt{K^2+L(L+1)/2}\\rfloor$ suffice. The bounds hold for every channel realization, every number of antennas, and every feasible combination of sensing accuracy and user SINR. If the bounds are right, massive-antenna ISAC transmitters can be designed and optimized with far fewer beamforming chains than antennas, and the extra sensing beams are often unnecessary.","feed_headline":"Sensing plus communication: at most K + sqrt(L(L+1)/2) beams","feed_subtitle":"Even without canceling sensing interference, sqrt(K^2 + L(L+1)/2) beamformers match the full antenna set.","key_machinery":"The engine is a rank-reduction argument that turns performance preservation into a counting problem. Starting from an optimal full beamformer set, the paper writes the requirement that a new set keep the same BFIM, SINRs, and power as quadratic equations, then substitutes $a_k = 1-|d_k|^2$ and $M_s = I_{N_s}-U_s U_s^H$ to obtain a homogeneous linear system. This system has $K+L(L+1)/2$ equations and $K+N_s^2$ unknowns, so whenever $N_s^2 > L(L+1)/2$ a nonzero solution exists; scaling it by its largest component makes $I_{N_s}-M_s$ positive semidefinite and singular, yielding a shorter beamformer matrix with identical performance. In the no-cancellation case a separate lemma forces the optimal sensing beams to be orthogonal to every user channel, which lets the communication beams absorb sensing beams and shifts the counting threshold to $N^2 > K^2+L(L+1)/2$. Iterating until the counting condition fails leaves at most the advertised number of beamformers.","core_discovery":"The paper establishes that the minimum number of downlink beamformers for ISAC is governed by the number of distinct quadratic terms in the performance metrics rather than by the number of antennas. For Bayesian Cramér-Rao bound (BCRB) estimation of $L$ real parameters, any pair consisting of a Bayesian Fisher information matrix (BFIM) and user SINRs that is achievable with the full $N_T+K$ beamformers is also achievable with at most $K+\\lfloor\\sqrt{L(L+1)/2}\\rfloor$ beamformers when users cancel sensing interference, and with at most $\\lfloor\\sqrt{K^2+L(L+1)/2}\\rfloor$ when they cannot. Because the second bound is less than the sum $K+\\sqrt{L(L+1)/2}$, joint operation can require strictly fewer beamformers than the two tasks taken separately. For any sensing metric depending on $d$ quadratic trace terms, the same argument gives $K+\\lfloor\\sqrt{d}\\rfloor$ and $\\lfloor\\sqrt{K^2+d}\\rfloor$; for $N_{\\mathrm{tr}}$ line-of-sight targets this scales roughly as $K+1.871N_{\\mathrm{tr}}$ with cancellation, and without cancellation the single-target worst case is exactly 2 beamformers for $K=0$ or 1 and $K$ beamformers for $K \\ge 2$.","pith_inferences":["The same counting logic suggests that any sensing metric whose Fisher information or objective is a function of $m$ independent quadratic forms will inherit bounds of order $\\sqrt{m}$ or $\\sqrt{K^2+m}$, even when the underlying unknowns are not angles or path losses.","Because the bounds are worst-case over channels, typical instances need far fewer beams (the paper's own simulations show 2-3 beams in many cases); a practical design rule is to start from the upper bound and then prune using the same reduction algorithm.","The paper flags that its tighter AoA-only bound relies on zero-mean path-loss priors; if that caveat bites, active or multi-stage sensing that updates priors to nonzero means should revert to the larger $O(N_{\\mathrm{tr}})$ scaling, which is a testable prediction for multi-stage ISAC.","The result gives a direct argument for reduced-RF or hybrid architectures: since the needed number of beamformer ports is bounded independently of $N_T$, the beamformer count can be fixed before the antenna-array size is chosen."],"forward_implications":["Any feasible ISAC performance pair can be realized with at most the stated number of beamformers, so optimizing directly in beam space loses nothing compared with a full-antenna semidefinite relaxation.","Without interference cancellation, if $K \\ge L(L+1)/4$ then communication beamformers alone achieve any feasible BFIM and SINR targets.","For single-target line-of-sight sensing without cancellation, the worst-case minimum is exactly 2 beamformers for $K=0$ or 1 and exactly $K$ beamformers for $K \\ge 2$.","For radar SNR or SCNR, which are $d$-quadratic metrics with $d=1$ or 2, no extra sensing beamformers are needed when at least one user is served without cancellation.","For beam-pattern probing over $N_g$ grid points, the bounds give at most $K+\\lfloor\\sqrt{N_g+1}\\rfloor$ beamformers with cancellation and $\\lfloor\\sqrt{K^2+N_g+1}\\rfloor$ without."],"supporting_citations":[{"why":"Supplies the rank-constrained separable SDP theorem whose rank-sum inequality the proof refines into a rank-one communication-beam solution.","marker":"[19]"},{"why":"Provides the underlying SDP rank-reduction existence argument that the iterative beamformer-removal procedure adapts.","marker":"[20]"},{"why":"Gives the classical $2N_{\\mathrm{tr}}$ bound for MIMO radar CRB estimation that the new bounds improve.","marker":"[7]"},{"why":"Prior one-target one-user result without cancellation that the paper generalizes to arbitrary $K$.","marker":"[8]"},{"why":"Shows communication beamformers alone suffice for radar SNR without cancellation; recovered as the $d=1$ hypotenuse case.","marker":"[9]"},{"why":"Shows when sensing symbols are unnecessary for one user and provides a tightness benchmark for the sum bound.","marker":"[14]"},{"why":"Supplies the optimal transmit-beamforming formulation and SDR tightness argument used to establish strong duality.","marker":"[12]"},{"why":"Provides the CRB-based joint beamforming optimization whose linear BFIM structure the proof exploits.","marker":"[5]"}],"fun_headline_variants":["ISAC beams: at most K + sqrt(L(L+1)/2) for full accuracy","No cancel? Still sqrt(K^2 + L(L+1)/2) beams suffice","ISAC joint beam count can beat sum of single-task needs","For one target, comm beams alone cover ISAC when K ≥ 2"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result rests on the scaling step: whenever the homogeneous linear system has more unknowns than equations, a nonzero solution can always be scaled so that deleting one beamformer does not change the BFIM, SINRs, or power, and this must hold for every feasible configuration.","fun_headline_variants_meta":{"raw":{"variants":["ISAC beams: at most K + sqrt(L(L+1)/2) for full accuracy","No cancel? Still sqrt(K^2 + L(L+1)/2) beams suffice","ISAC joint beam count can beat sum of single-task needs","For one target, comm beams alone cover ISAC when K ≥ 2"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000303,"raw_usage":{"total_tokens":1891,"prompt_tokens":1238,"completion_tokens":653,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":854,"completion_tokens_details":{"reasoning_tokens":566}},"tokens_in":854,"tokens_out":653,"duration_ms":7612,"temperature":1.0,"reasoning_tokens":566,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:44:26.424319+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A counterexample would be a feasible ISAC instance with $N_s^2 > L(L+1)/2$ in which every nonzero solution of the homogeneous system (109)-(110) fails, after scaling, to make $I_{N_s}-M_s$ positive semidefinite and singular with $1-a_k \\ge 0$ for all $k$; searching random channel and target realizations for such a case would settle whether the sum bound is universal.","supporting_citations":[{"cited_title":"Rank-constrained separable semidefinite programming with applications to optimal beamforming,","cited_arxiv_id":null,"evidence_quote":"Supplies the rank-constrained separable SDP theorem whose rank-sum inequality the proof refines into a rank-one communication-beam solution."},{"cited_title":"On the rank of extreme matrices in semidefinite programs and the multiplicity of optimal eigenvalues,","cited_arxiv_id":null,"evidence_quote":"Provides the underlying SDP rank-reduction existence argument that the iterative beamformer-removal procedure adapts."},{"cited_title":"Range com- pression and waveform optimization for MIMO radar: A Cram ´er–Rao bound based study,","cited_arxiv_id":null,"evidence_quote":"Gives the classical $2N_{\\mathrm{tr}}$ bound for MIMO radar CRB estimation that the new bounds improve."},{"cited_title":"Integrated sensing and communication exploiting prior information: How many sensing beams are needed?","cited_arxiv_id":null,"evidence_quote":"Prior one-target one-user result without cancellation that the paper generalizes to arbitrary $K$."}],"review_version":1}