{"id":"8aedf1a6-8da1-4f6e-8dd4-da28eff28500","arxiv_id":"2502.03162","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A successive convex approximation plus shifted generalized power iteration algorithm is proposed to maximize the weighted sum of communication rate and negative Cramér-Rao bound in ISAC beamforming.","lead":"This paper proposes a low-complexity beamforming algorithm for integrated sensing and communications that jointly optimizes communication sum rate and sensing accuracy. It combines successive convex approximation with a shifted generalized power iteration to avoid expensive semidefinite relaxation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 2's Taylor expansion omits tr(Z0^-1), making surrogate (12) an upper bound on -tr(F^-1); the SCA-SGPI monotone-convergence claim is unsupported.","rationale":"The reader's verdict identifies the same load-bearing weakness: surrogate (12) is not a minorant. My independent check confirms the technical mechanism. The first-order expansion in Lemma 2 is missing the tr(Z0^-1) term, so equality at the current iterate fails; even correcting the constant, the negative of a lower bound on tr(F^-1) is an upper bound on the sensing term, so the surrogate cannot be a lower bound for maximization. This invalidates the standard SCA monotone-ascent argument and means the convergence and tradeoff claims rest on numerical evidence alone. The numerical experiments are suggestive but do not establish the algorithmic guarantee. Because the flaw is in the theoretical framing rather than necessarily in the numerical outcomes, I do not call for rejection; the appropriate disposition is to keep the conditional verdict and require a fixed Lemma 2 and a convergence analysis, or a clear statement that SCA-SGPI is a heuristic, before the claims are accepted.","tokens_in":8646,"tokens_out":11615,"duration_ms":108550,"concrete_test":"Independently re-derive Eq. (9) from the first-order Taylor expansion of tr(Z^-1) and evaluate it at Z=Z0; if the RHS is 0 instead of tr(Z0^-1), the lemma is mis-stated. Then, at a random feasible W[t] with auxiliary variables from (11) and Phi=F[t]^-2, compare the true objective (7) with the surrogate (12) at W[t]; if the surrogate exceeds the true objective, the minorant property fails and the SCA convergence claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is that Algorithm 1 is an SCA method whose surrogate problem (12) minorizes the true objective (7). This fails for the sensing term. Lemma 2 (Eq. 9) states tr(Z^-1) >= -tr(Z0^-1 Z Z0^-1) + tr(Z0^-1), with equality at Z=Z0. The first-order Taylor expansion of the convex function tr(Z^-1) at Z0 is 2tr(Z0^-1) - tr(Z0^-2 Z); therefore Eq. (9) is missing one tr(Z0^-1) term and its RHS is 0 at Z=Z0, not tr(Z0^-1). Moreover, the inequality direction is wrong for the maximization objective: because tr(F^-1) is convex, its tangent is a lower bound, so negating it gives an upper bound on -tr(F^-1). The sensing part of (12) is therefore an upper bound, not a minorant, and at the current iterate the surrogate objective exceeds the true objective by tr(F[t]^-1). Standard SCA requires a global lower-bound surrogate tangent at W[t]; neither condition holds. The paper explicitly omits the convergence proof, so the observed monotonicity in Fig. 1 is the only support, and it cannot be attributed to the stated SCA mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies joint beamforming for a monostatic ISAC system with Nt transmit and Nr receive antennas, K single-antenna users, and one point target. It formulates the problem of maximizing a weighted sum of the multiuser sum rate and the trace of the inverse Fisher information matrix, subject to a total transmit power constraint. The authors propose an SCA-SGPI algorithm in which SCA builds a quadratic surrogate for the objective and a shifted generalized power iteration solves the resulting subproblems. They provide a complexity analysis, a runtime comparison with an SCA-SDR benchmark, and simulations showing fast convergence and a slightly better communications-sensing tradeoff than the benchmark.","tokens_in":8999,"tokens_out":16055,"duration_ms":146182,"significance":"The practical goal is worthwhile: CRLB-based ISAC beamforming is typically solved by SDR with high complexity, and a low-complexity iterative method with closed-form updates could be valuable. The paper gives a clean system model, explicit surrogate equations, a simple projection update, and numerical evidence of large runtime reductions (0.0362 s versus 9.0803 s in one setting). However, the central theoretical justification is not sound: Lemma 2 is incorrect, and the sensing surrogate in (12) is not a valid minorant of the objective. As a result, the claimed SCA monotone-convergence mechanism is unsupported; the contribution is currently an empirically motivated heuristic rather than a proven optimization algorithm. If the surrogate issue is repaired or the claims are appropriately weakened, the complexity and simulation results would be of interest to the ISAC community.","major_comments":[{"comment":"Lemma 2 is false as stated. Evaluating (9) at Z = Z0 yields tr(Z0^-1) >= 0, whereas equality is claimed; the right-hand side should contain an additional tr(Z0^-1) term (the correct tangent lower bound is tr(Z^-1) >= 2tr(Z0^-1) - tr(Z0^-1 Z Z0^-1)). This is load-bearing because the sensing surrogate in (12) is built directly from this inequality and therefore does not equal the true sensing term at the current iterate W[t].","section":"Section III-A, Lemma 2 (Eq. (9))"},{"comment":"Even after correcting Lemma 2, the sensing term in (12) is an upper bound on -tr(F^-1), not a lower bound. Since tr(F^-1) is convex in F and F depends affinely on WW^H, the first-order expansion of tr(F^-1) is a global lower bound for tr(F^-1); negating it yields an upper bound on -tr(F^-1). A standard SCA for maximization needs a surrogate that globally minorizes the objective and is tight at the current iterate. The surrogate (12) fails both conditions for the sensing term, so the SCA monotone-ascent argument for Algorithm 1 does not apply.","section":"Section III-A, Eq. (12)"},{"comment":"The paper explicitly omits the convergence proof (\"We skip the detailed convergence analysis...\") and instead relies on Fig. 1. Because the surrogate in (12) is not a minorant, the observed monotone increase in Fig. 1(a) cannot be attributed to the claimed SCA mechanism. The algorithm may be a useful heuristic, but the central theoretical claim needs to be either repaired (for instance, by constructing a valid minorant via a proximal or regularization term) or replaced by a different convergence analysis that does not rely on SCA minorization.","section":"Section III-B, Algorithm 1"}],"minor_comments":[{"comment":"There are several typos: \"Cram ´er\" in the title and abstract, \"it’s\" in Lemma 3, \"the the objective\" in the caption of Fig. 1, \"significantly reduction\" in the Introduction, and \"SCA-SPGI\" for \"SCA-SGPI\" in the Conclusion.","section":"General"},{"comment":"The inner-loop iteration uses W[n] without an outer-iteration index; writing W[t,n] would clarify that the inner iterations are performed at a fixed outer iteration t.","section":"Section III-B, Eqs. (15)-(16)"},{"comment":"The runtime of SCA-SGPI is non-monotonic in K (0.0270 s for K=2, 0.0185 s for K=4, 0.0137 s for K=6), which may be due to varying numbers of outer iterations; a brief explanation would improve reproducibility.","section":"Section IV, Table I"},{"comment":"The transmit power constraint is written as an equality, tr(WW^H) = Pt. Since both the sum rate and the sensing CRLB improve with transmit power, the equality is natural, but the paper should state that using full power is always optimal for this objective.","section":"Section II-B, Eq. (7)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a communications/signal processing journal, and the runtime improvements over SDR-based benchmarks are striking. The main blocker is the incorrect surrogate derivation: Lemma 2 is wrong, and even its corrected version would not provide a minorant for the sensing term. I recommend asking the authors to rework the SCA justification or to reframe the algorithm as a heuristic with an appropriate convergence study. The self-citation [23] is relevant, but the authors should verify Lemma 1's provenance and ensure the cited results are publicly available. The numerical comparison is limited to a single SDR-based benchmark; adding a simple baseline (e.g., maximum-ratio or zero-forcing beamforming) would strengthen the tradeoff claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core SCA convergence argument in this paper does not hold as written. Lemma 2 states a lower bound for tr(Z^-1) that misses one tr(Z0^-1) term; the correct first-order Taylor expansion is 2tr(Z0^-1) - tr(Z0^-2 Z). The paper's expression -tr(Z0^-1 Z Z0^-1) + tr(Z0^-1) is not tangent and equality at Z=Z0 fails. When negated for the sensing term, the surrogate in (12) becomes an upper bound on -tr(F^-1), not a minorant. At the current iterate, the surrogate sensing term is 0 while the true term is -tr(F[t]^-1), so the surrogate objective exceeds the true objective. Standard SCA monotone ascent therefore does not apply, and the paper explicitly skips the convergence proof. The only support is Fig. 1's empirical monotonicity, which is unexplained.\n\nWhat the paper does well: the SCA-SGPI combination for trace-CRLB and sum-rate beamforming is new, and the complexity reduction is substantial — from SDR's O(K^3.5 Nt^6.5) to O(I1 I2 K Nt^2). The reported runtimes (0.0362 s vs 9.0803 s) support the practical claim. The SGPI inner loop is a sensible way to solve the quadratic subproblem, and the overall aim of avoiding standard optimization toolboxes is worthwhile.\n\nMinor issues: the benchmark implementation is underdescribed, the tradeoff curves have no error bars, and the runtime comparison is a single table without variance. The self-citation to [23] for SGPI is fine.\n\nThe central flaw is load-bearing but fixable: replacing the sensing surrogate with the correct tangent (tr(FΦ) - 2tr(F0^-1)) would restore both tightness and the minorant property, likely preserving the algorithmic structure. Until that is done, the convergence claims and the objective-value interpretation in Fig. 1 are not established. A serious referee should see this — the idea is useful and the repair looks tractable — but the revision should be major.","headline":"The SCA-SGPI algorithm is a promising practical recipe, but the central convergence argument is broken by an incorrect Taylor expansion in Lemma 2, so the current version needs a major fix before the theory can be trusted.","tokens_in":9470,"tokens_out":6378,"would_cite":false,"duration_ms":51879,"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":"The paper proposes a low-complexity SCA-SGPI algorithm for joint beamforming that maximizes a weighted sum of communications sum rate and sensing CRLB under a power constraint, matching or slightly beating SDR-based optimization at a…","keywords":["integrated sensing and communications","Cramér-Rao lower bound","beamforming optimization","successive convex approximation","generalized power iteration","sum rate optimization","MIMO radar","low-complexity algorithm"],"falsifier":"Compute both sides of Lemma 2's inequality (9) at $Z = Z_0$: the left side equals $\\mathrm{tr}(Z_0^{-1})$ while the right side is zero, so the claimed equality is false. A direct algorithmic check is to record the true objective (7) after every outer iteration on random channels; any decrease would disprove the claimed monotone convergence and confirm that the missing minorant property matters.","tokens_in":8438,"feed_emoji":"📡","tokens_out":12971,"duration_ms":104657,"temperature":0.7,"pith_summary":"Integrated sensing and communications (ISAC) beamforming must balance how much data the base station delivers to users against how accurately it can estimate a target, and the sensing side of that balance is usually measured by the Cramér-Rao lower bound (CRLB). Optimizing the CRLB with semidefinite relaxation is accurate but expensive, scaling steeply in the number of users and antennas. This paper proposes to solve the weighted sum-rate-plus-CRLB problem with a two-step iteration: successive convex approximation builds a quadratic surrogate for the nonconvex objective, and a shifted generalized power iteration solves each quadratic subproblem directly. The resulting SCA-SGPI algorithm is claimed to match or slightly beat the SCA-SDR benchmark on the communications-sensing tradeoff while cutting the per-iteration complexity from $O(I_1 K^{3.5} N_t^{6.5})$ to $O(I_1 I_2 K N_t^2)$, and the reported runtime drops from 9.08 seconds to 0.036 seconds in the simulated setup. The practical payoff, if the claims hold, is ISAC beamforming that can be recomputed nearly in real time.","feed_headline":"ISAC beamforming gets a 250x faster CRLB optimizer","feed_subtitle":"Power-iteration solver cuts per-design cost from K^3.5 to linear in users, with the same tradeoff curve.","key_machinery":"The load-bearing object is the quadratic surrogate (13) together with the shifted generalized power iteration (16). The SCA layer converts the nonconvex objective into $\\mathrm{tr}(\\mathbf{W}\\mathbf{W}^H \\mathbf{A}) + 2\\delta\\,\\Re\\bigl\\{\\mathrm{tr}(\\mathbf{W}\\boldsymbol{\\Sigma}_1^H \\mathbf{H}^H)\\bigr\\}$, where $\\mathbf{A}$ is made positive semidefinite by a shift parameter $\\lambda$; this keeps the subproblem a quadratically constrained quadratic program. The SGPI layer is a fixed-point method that repeatedly maps $\\mathbf{H}\\boldsymbol{\\Sigma}_1 + \\mathbf{A}\\mathbf{W}^{(n)}$ onto the feasible sphere $\\mathcal{S}$, costing only $O(K N_t^2)$ per update and avoiding interior-point solvers and semidefinite relaxation. The two layers together are what let the algorithm escape the $K^{3.5} N_t^{6.5}$ scaling of SDR-based CRLB optimization.","core_discovery":"The paper's central claim is that the nonconvex beamforming problem (7), which maximizes $\\delta\\sum_k R_k - \\mathrm{tr}(F^{-1})$ under a fixed transmit power, can be solved efficiently by alternating two cheap steps. The outer SCA step replaces each user's log-rate with its first-order minorant (Lemma 1) and replaces the sensing term with a linearized trace expression built from Lemma 2, turning the objective into a quadratic form in the beamforming matrix $W$. The inner SGPI step then solves that quadratic subproblem by the fixed-point update $\\mathbf{W}^{(n+1)} = \\Pi_{\\mathcal{S}}\\big(\\mathbf{H}\\boldsymbol{\\Sigma}_1 + \\mathbf{A}\\mathbf{W}^{(n)}\\big)$, a projection onto the power-constraint sphere. The authors assert that this yields a locally optimal solution with per-iteration cost $O(K N_t^2)$, and their simulations show a slightly better communications-sensing tradeoff than the SCA-SDR benchmark with two to three orders of magnitude less runtime.","pith_inferences":["The paper leaves implicit that, once the Lemma 2 constant error is repaired, the same surrogate-plus-power-iteration recipe should transfer to other CRLB-type sensing metrics such as the determinant or the largest eigenvalue of $F^{-1}$, since those also admit tractable surrogates.","A testable extension is to apply SCA-SGPI to wideband or multi-carrier ISAC; the linear-in-$K$ scaling suggests the speed advantage over SDR should persist.","Recording the true objective (7) at every outer iteration on random channels would separate the algorithm's numerical behavior from the unproven minorant property, since any decrease would localize the gap to the sensing surrogate."],"forward_implications":["In the simulated $N_t=16$, $N_r=20$, $K=4$ setup, the algorithm converges in roughly ten outer and five inner iterations, so the complexity reduction translates into a runtime of 0.0362 s versus 9.0803 s for the SCA-SDR benchmark.","Because complexity grows linearly in the number of users $K$, the method remains tractable for many-user ISAC systems where SDR-based solvers scale as $K^{3.5}$.","The sweep over $\\delta$ in $[10^{-7}, 10]$ shows SCA-SGPI slightly dominating the SCA-SDR tradeoff curve, indicating the speed gain does not come at the cost of communications or sensing performance.","The same SCA-plus-power-iteration structure can be extended to multi-target ISAC, which the authors identify as the immediate future direction."],"supporting_citations":[{"why":"Supplies the monostatic ISAC signal model, the FIM block expressions, and the L-time-slot assumption used in (6).","marker":"[11]"},{"why":"The SCA-SDR benchmark whose tradeoff and runtime are compared in the numerical results.","marker":"[15]"},{"why":"Establishes that global optimality for this problem requires exponential complexity, justifying a local fast algorithm.","marker":"[20]"},{"why":"Justifies using the trace of F^-1 rather than a single diagonal entry as the sensing accuracy metric.","marker":"[22]"},{"why":"Provides the first-order log-rate lower bound used in Lemma 1 to construct the communications surrogate.","marker":"[23]"},{"why":"Provides the first-order tr(Z^-1) lower bound used in Lemma 2 to linearize the sensing term.","marker":"[24]"},{"why":"Gives the O(K^3.5 N_t^6.5) SDR complexity figure that motivates the low-complexity design.","marker":"[26]"}],"fun_headline_variants":["ISAC beamforming: SCA-SGPI cuts CRLB optimization cost 250x","SCA-SGPI: 250x faster CRLB optimization for ISAC beamforming","Power iteration speeds up ISAC CRLB optimization by orders of magnitude","Low-complexity ISAC beamforming via SCA-SGPI achieves same tradeoff","SCA-SGPI: Efficient CRLB optimization for ISAC systems"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The algorithm's convergence relies on each surrogate being a lower bound of the true objective, yet the paper's Lemma 2 misses a constant term that breaks equality at the current point, so the lower-bound property is not established.","fun_headline_variants_meta":{"raw":{"variants":["ISAC beamforming: SCA-SGPI cuts CRLB optimization cost 250x","SCA-SGPI: 250x faster CRLB optimization for ISAC beamforming","Power iteration speeds up ISAC CRLB optimization by orders of magnitude","Low-complexity ISAC beamforming via SCA-SGPI achieves same tradeoff","SCA-SGPI: Efficient CRLB optimization for ISAC systems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000273,"raw_usage":{"total_tokens":1638,"prompt_tokens":952,"completion_tokens":686,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":579}},"tokens_in":568,"tokens_out":686,"duration_ms":5784,"temperature":1.0,"reasoning_tokens":579,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T05:41:16.114332+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute both sides of Lemma 2's inequality (9) at $Z = Z_0$: the left side equals $\\mathrm{tr}(Z_0^{-1})$ while the right side is zero, so the claimed equality is false. A direct algorithmic check is to record the true objective (7) after every outer iteration on random channels; any decrease would disprove the claimed monotone convergence and confirm that the missing minorant property matters.","supporting_citations":[{"cited_title":"Cr am´ er-Rao bound optimization for joint radar-communication beamfor ming,","cited_arxiv_id":null,"evidence_quote":"Supplies the monostatic ISAC signal model, the FIM block expressions, and the L-time-slot assumption used in (6)."},{"cited_title":"Transmitter side beyond-diagonal R IS for mmwave integrated sensing and communications,","cited_arxiv_id":null,"evidence_quote":"The SCA-SDR benchmark whose tradeoff and runtime are compared in the numerical results."},{"cited_title":"Globally optimal b eamforming design for integrated sensing and communication systems,","cited_arxiv_id":null,"evidence_quote":"Establishes that global optimality for this problem requires exponential complexity, justifying a local fast algorithm."},{"cited_title":"R ange compression and waveform optimization for mimo radar: A cra m ´Er–rao bound based study,","cited_arxiv_id":null,"evidence_quote":"Justifies using the trace of F^-1 rather than a single diagonal entry as the sensing accuracy metric."},{"cited_title":"Majorization-minim ization algorithms in signal processing, communications, and mach ine learning,","cited_arxiv_id":null,"evidence_quote":"Provides the first-order tr(Z^-1) lower bound used in Lemma 2 to linearize the sensing term."},{"cited_title":"Partially-conne cted hybrid beamforming design for integrated sensing and commu nication systems,","cited_arxiv_id":null,"evidence_quote":"Gives the O(K^3.5 N_t^6.5) SDR complexity figure that motivates the low-complexity design."}],"review_version":1}