{"id":"fcad2135-b709-444d-885d-a00a4774f8ab","arxiv_id":"2607.27778","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Joint radio-fronthaul-cloud optimization cuts total power in cell-free massive MIMO ISAC networks by 50-59% versus transmit-power-only optimization for distributed multi-target sensing.","lead":"A 6G network design study shows that jointly optimizing radio, fronthaul, and cloud resources can cut total power consumption of cell-free massive MIMO systems with sensing by more than half compared to transmit-power-only optimization. The authors model distributed multi-target detection and cross-layer power consumption, then solve the mixed-integer optimization.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 50% savings may be an artifact: the relaxed fronthaul constraint in P1 does not match the exact P0 constraint, so the optimized W can underestimate true fronthaul/cloud power.","rationale":"The reader's weakest_assumption is the neglected TX-RX communication interference in the sensing model. That is a genuine, self-acknowledged limitation (Footnote 1) and could erode detection performance in practice, but it is a modeling simplification external to the optimization claim. A more load-bearing concern is internal: whether the algorithm whose outputs generate the headline numbers actually solves or even satisfies the stated problem P0. The manuscript itself contains the raw material for this concern: (50j) uses unrelaxed z_l and drops the z_l/z_l multipliers present in the exact Rfront_tot definition, and the reader's rationale also noted that '(50j) appears inconsistent.' Because the numerical percentages are the central claim, an infeasible fronthaul constraint would directly invalidate them. I therefore identify this as the primary check to run, while agreeing with the reader's overall CONDITIONAL stance. Since the proposed test has not been run, I do not move the verdict: if the test reveals violations, the paper would need major revision; if all operating points satisfy (45l), the concern is resolved and the conditional verdict can stand. I mark agreement as 'partial' because the reader also saw the (50j) issue in the rationale, but their stated weakest assumption was the sensing-interference simplification.","tokens_in":22458,"tokens_out":12352,"duration_ms":129958,"concrete_test":"Use the exact Fig. 2 scenario (L=25, K=8, FIS, tau_s=20). After running Algorithm 1, take the returned binary variables and compute Rfront_tot from the unrelaxed definitions in Section IV-A for all APs; check whether (45l) holds with W=ceil(W_tilde). Repeat for every plotted operating point in Figs. 2-4. If any final solution violates (45l), re-solve P1 with an exact convexification of the bilinear terms (e.g., McCormick envelopes or fixing RX-mode indicators in an outer loop) and recompute the total-power savings; if the >50% and 13-15% margins disappear, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claim depends on Algorithm 1 actually solving a faithful relaxation of P0. In Section V-A, constraint (50j) is presented as the convexified version of P0's fronthaul constraint (45l), where Rfront_tot is defined in Section IV-A as sum_l(z_l Rtx_l + z_l Rrx_l). But (50j) does not implement that sum: the communication and local-test-statistic terms appear as tau_d sum eta_tilde + tau_s sum zeta_tilde + sum xi_tilde without the z_l/z_l indicators required by (29) and the Rrx_l expression; the FIS/PIS terms use the unrelaxed RX-mode variable z_l inside a problem whose optimization variables are z_tilde_l and z_hat_l; and those terms are products of RX and TX counts, which are neither convex nor part of the stated P1 variable set unless previous-iterate values are silently frozen. This is not a cosmetic typo: the claimed savings come largely from switching off RX-APs, LCs, and GPPs, so underestimating the fronthaul load in the relaxation can make the optimizer choose a smaller W and more inactive APs than the exact P0 permits. The P2 recovery uses the same flawed expression and only enforces a derived bound Zmax, so it does not guarantee the final integer solution satisfies (45l). Consequently, the 50% and 13-15% savings reported in Section VI may describe an infeasible solution of P0 rather than the E2E framework.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an end-to-end resource orchestration framework for cell-free massive MIMO ISAC with distributed sensing. Each AP can be a TX-AP, RX-AP, or idle; RX-APs compute local MAPRT test statistics under FIS and PIS scenarios and forward them to the cloud for weighted aggregation. The authors model fronthaul data rates, radio/cloud GOPS, and total power consumption (40)-(44), formulate the joint MINLP P0 in (45), and solve it via a relaxed penalized problem P1 (50) with discrete recovery P2/P3 (51)-(54). Numerical results claim more than 50% power savings over transmit-power-only benchmarks and about 13-15% over radio-optimization benchmarks, with detection probabilities reported in Table I.","tokens_in":22841,"tokens_out":10221,"duration_ms":110765,"significance":"The paper addresses a timely and important problem: quantifying the energy cost of adding sensing to cell-free massive MIMO and showing that radio-only or transmit-power-only optimization misses large savings in fronthaul and cloud domains. Its strengths are the cross-layer power model, the explicit FIS/PIS detector distinction, the broad MINLP formulation, and a numerical comparison with four benchmarks. However, the central numerical claim is not yet supported because the convex relaxation used in Algorithm 1 does not faithfully implement the exact fronthaul constraint of P0, and the P1 subproblem is not well-posed as written. These issues are fixable but require reworking the optimization and re-running the results.","major_comments":[{"comment":"The relaxed constraint (50j) does not implement P0's exact fronthaul constraint (45l). In Section IV-A, Rfront_tot = sum_l(z_l Rtx_l + z_l Rrx_l), with Rtx_l in (29) and Rrx_l containing 2 sum_s xi_{s,l} + I_FIS 2 tau_s sum_l z_l + (1-I_FIS)(sum_l z_l)^2. In (50j), the communication terms tau_d sum eta_tilde + tau_s sum zeta_tilde and the local-statistic term sum xi_tilde appear without the z_l/z_l indicators, and the FIS/PIS terms are products of RX and TX counts involving z_l, which is not in the P1 variable set. Unless z_l is a frozen previous iterate, this product is nonconvex. P2's (51b) uses the same mismatch, and the Zmax bound in (52) is derived from it. Consequently the recovery step does not certify (45l), and the Section VI power savings may describe infeasible solutions of P0. Please replace by a valid SCA of the exact constraint and add a final feasibility check.","section":"V-A, Eq. (50j), (51b), (52)"},{"comment":"The MSE penalty in (48) is written as squared differences between binary variables z_l, z_l, eta, zeta, xi and relaxed surrogates, but P1's optimization variables listed in (50a) are only the relaxed continuous variables and slack. As written, P1 contains binary variables in the objective and in (50j) that are not optimized, so the first step of Algorithm 1 is not a well-defined convex problem. If the binary variables are meant to be previous-iterate values, this must be stated explicitly; if they are meant to be replaced by their relaxed counterparts, the expressions must be rewritten accordingly.","section":"V-A, Eq. (48) and P1 (50)"},{"comment":"The sensing SINR in (15) ignores communication interference. RX-APs are located in a network where TX-APs simultaneously transmit data to UEs and sensing signals; those communication waveforms can propagate into RX-APs through direct/reflected paths and appear in (8) as an unmodeled term. The detection probabilities in Table I and the optimized power/resource results are therefore optimistic if this leakage is non-negligible. Since this is a stated assumption rather than a negligible effect, the authors should either add a communication-interference term to the SINR and evaluate its impact in at least one numerical scenario, or explicitly restrict the claimed power savings and detection performance to the interference-free case.","section":"II-B, Footnote 1, Eq. (8)-(15)"},{"comment":"The PIS detector is not derived in this manuscript. After stating the test statistic in (25), the paper gives update equations (26)-(28) and then refers to [16, Algorithm 1] 'due to space limitations'. Because the PIS detector is one of the two named contributions and feeds the detection results in Table I, a journal paper should either include the derivation in an appendix or formally state that it is identical to the cited work. Also, the false-alarm threshold lambda_d is said to be 'selected empirically'; the procedure for setting it and matching the 0.03 false-alarm probability is not described, which makes the PD numbers hard to reproduce.","section":"III, Eqs. (25)-(28)"}],"minor_comments":[{"comment":"In (45f) the quantifier should be for all s,l rather than k,l; in (45g) the term zeta_{k,l} should be zeta_{s,l}.","section":"V, Eqs. (45f), (45g)"},{"comment":"The notation z_l and \\tilde{z}_l / \\hat{z}_l is visually similar; in (48)-(50) the binary and relaxed variables are not always clearly distinguished, which contributes to the ambiguity in P1.","section":"V-A, notation"},{"comment":"The weighting exponent v is set to 0.25 based on [17] with no sensitivity analysis. Since v controls the aggregation weights and hence detection probability, a short sensitivity study would strengthen the results.","section":"II-B, Eq. (20)"},{"comment":"The transmit-power-only benchmark uses a heuristic energy-unaware association. Savings relative to this benchmark should be interpreted as an upper bound; it would be helpful to include a stronger power-only baseline with optimized association.","section":"V-B, benchmarks"},{"comment":"The caption of Fig. 3c says 'FIS with L=25, K=8' while the text describes a plot of total power versus SE threshold; please check the caption and the text.","section":"VI, Fig. 3c"}],"recommendation":"major_revision","confidential_remarks":"The paper has a clear scope and the main issue is fixable: the relaxation must be made to match P0's exact fronthaul constraint, with a post-hoc feasibility check. I do not see a circularity problem; the power-savings numbers are optimization outputs, not fitted quantities. The main risk is that the reported savings come from infeasible solutions. If the authors reformulate P1/P2 correctly and verify feasibility of the final integer solutions, the paper could become a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a serious attempt at something useful: an end-to-end power optimization for cell-free massive MIMO ISAC that actually accounts for fronthaul and cloud processing, not just transmit power. The problem is well formulated, the power model is detailed, and the FIS/PIS distinction gives a nice view of a real implementation trade-off. It is not a theoretical breakthrough, but it is a legitimate extension of the authors' established line, and the numerical exploration of AP density, SE thresholds, and SINR thresholds is useful.\n\nThe main problem is the fronthaul constraint in the relaxed problem. The original P0 constraint (45l) involves \\sum_l (z_l Rtx_l + z_l Rrx_l), with Rrx_l depending on the number of TX APs and RX APs. In P1, constraint (50j) writes that as a product of sums, but it uses the unrelaxed binary z_l alongside the relaxed \\tilde{z}_l. The paper never states that z_l is fixed from a previous iteration, and if it is not fixed, the constraint is bilinear and nonconvex. The recovery step P2/P3 only enforces a bound on the number of RX APs (Zmax), not the full per-AP fronthaul expression, so the final binary solution is not guaranteed to satisfy (45l). Since the claimed savings come largely from switching off RX-APs, LCs, and GPPs, an underestimate of fronthaul load in the relaxation can easily produce W too small and an infeasible solution. This is not cosmetic; it directly concerns the 50% and 13-15% savings claims.\n\nI also note the sensing model ignores communication interference at the RX-APs (footnote 1). The authors are explicit about it, so it is an acknowledged limitation, but it does mean the detection probabilities and, by extension, the required powers could be optimistic. Table I gives detection probabilities without indication of Monte Carlo runs or confidence intervals, which makes it hard to judge whether the 'competitive detection performance' claim is solid.\n\nThe benchmark for transmit-power-only also uses a heuristic energy-unaware association, which flatters the proposed method, but that is a common and acceptable choice if clearly disclosed.\n\nThe stress-test note I saw is on target. The paper deserves a serious referee and the authors know their material, but the optimization pipeline needs to be made faithful and the numerical claims verified. If the authors can fix the relaxation, re-run the numbers, and report feasibility of the final solution, this could be a useful contribution.\n\nRecommendation: send to peer review, but prepare for major revision.","headline":"The joint radio-fronthaul-cloud formulation is the real deal, but the headline power savings are built on a fronthaul relaxation that does not faithfully represent the original constraint, so the numbers are not yet trustworthy.","tokens_in":23330,"tokens_out":7555,"would_cite":false,"duration_ms":71248,"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's central claim is that adding multi-target sensing to cell-free massive MIMO (a network of many distributed access points cooperating to serve users) need not nearly double network energy if radio, fronthaul, and cloud resources","keywords":["integrated sensing and communication","cell-free massive MIMO","distributed sensing","power minimization","resource allocation","multi-target detection","network orchestration","energy efficiency"],"falsifier":"Restore the communication-interference terms in the sensing SINR expression (Eq. 15), or simulate the same system with actual communication waveforms present at the RX-APs, and re-run the proposed optimization; if the detection probability at the 0.03 false-alarm rate drops below the reported ~0.9 (FIS, R=1) or the power savings narrow materially, the central claim depends on the neglected interference.","tokens_in":22383,"feed_emoji":"⚡","tokens_out":8184,"duration_ms":73486,"temperature":0.7,"pith_summary":"The paper sets out to show that the extra power consumed by adding multi-target sensing to a cell-free massive MIMO network is mostly a coordination problem, not an unavoidable cost. Its proposed end-to-end framework models the total network power—radio hardware, transmit power, fronthaul signaling, and cloud processing—as one closed-form function of AP modes, associations, and active cloud resources, and minimizes it jointly. The authors report that this joint optimization consumes roughly 13–15% less power than optimizing radio resources alone and more than 50% less than transmit-power-only schemes, while keeping detection probability above 0.9 at a 0.03 false-alarm rate. A sympathetic reader would take this as evidence that 6G networks can incorporate sensing without a proportional energy penalty if the cloud, fronthaul, and radio decisions are made together.","feed_headline":"Joint orchestration halves energy cost of adding sensing to 6G","feed_subtitle":"Coordinating transmit power, AP sleep modes, and cloud resources saves 13–15% over radio-only tuning while keeping detection.","key_machinery":"The load-bearing object is the joint end-to-end optimization problem P0, built on a closed-form power model (Eq. 44) that converts AP operation modes, UE/SSA associations, RX-AP assignments, transmit power coefficients, and the integer number of active cloud line cards/processors into a single power objective. On the sensing side, the framework uses distributed maximum a posteriori ratio test detectors—fully informed (FIS, waveform known at the RX-AP) and partially informed (PIS, statistical knowledge only)—whose local statistics are fused at the cloud with SIR-based weights. The optimization method makes the problem tractable by rewriting communication SINR constraints in second-order cone","core_discovery":"On the paper's own terms, the discovery is that the total network power consumption of a distributed multi-target sensing-plus-communication system can be written as a single linear-ish expression in the discrete network decisions—which APs transmit, which receive, which sleep, how UEs and sensing areas are associated, and how many line cards/processors are active—and that minimizing this expression jointly across domains yields large savings that transmit-power-only or radio-only optimization miss. The authors formulate the mixed-integer non-convex problem P0 (Eq. 45), convexify it with second-order-cone reformulation and successive convex approximation, and solve it with a two-stage penalt","pith_inferences":["If communication interference is no longer assumed negligible at sensing receivers, the optimal operating point would likely separate communication and sensing transmissions in time/frequency or allocate more antennas; the paper's 13–15% and 50%+ savings figures would then shrink, but the qualitative case for joint orchestration would probably survive.","The same power-model-plus-binary-optimization template could be re-derived for other functional splits or centralized processing; the constants change, but the insight that hardware idle power and fronthaul loads dominate transmit power would likely transfer.","Since the largest baseline waste comes from energy-unaware AP selection, a lightweight power-aware heuristic (e.g., favoring APs with low fixed power or reusing already-active infrastructure) might capture a good fraction of the E2E gain at a fraction of the solving cost—an avenue worth testing.","The SIR-based fusion weights are static in the paper; an online variant that adapts weights to measured interference or to the FIS/PIS mode could improve detection robustness in time-varying environments."],"forward_implications":["If the E2E claim holds, a network operator can add sensing capability while keeping the added energy burden to roughly half of what transmit-power-only planning would require.","The bulk of the savings comes from switching off unneeded APs and scaling down fronthaul/cloud hardware—transmit power itself is a minor part of the total, so optimizing it alone leaves most of the opportunity untouched.","Full coordination, where fronthaul and processing resources are dynamically pooled across APs, consistently beats local static allocation, pointing toward virtualized architectures with flexible resource sharing.","The FIS/PIS comparison shows a direct, tunable trade-off: full waveform knowledge yields better detection and lower power per target, while partial information cuts fronthaul load at the cost of needing more RX-APs for the same detection probability.","Increasing the sensing SINR threshold (7 to 10 dB) affects total power far less than structural choices like the number of RX-APs per area, so the framework's main levers are topology and mode selection."],"fun_headline_variants":["Joint cloud-fronthaul-radio tuning cuts 6G ISAC power by 50%+ vs power-only","Cross-layer resource allocation saves 50%+ power for 6G sensing and comms","Coordinating AP sleep and cloud resources cuts energy 50%+ vs transmit-only","6G ISAC: Joint cloud-fronthaul-radio optimization saves >50% energy","Sensing and comms: Joint resource allocation achieves >50% power savings"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claimed power savings and detection results assume that communication signals leaking into the sensing receivers are negligible (Footnote 1); if they are not, the sensing SINR, the computed detection probabilities, and therefore the reported savings are optimistic.","fun_headline_variants_meta":{"raw":{"variants":["Joint cloud-fronthaul-radio tuning cuts 6G ISAC power by 50%+ vs power-only","Cross-layer resource allocation saves 50%+ power for 6G sensing and comms","Coordinating AP sleep and cloud resources cuts energy 50%+ vs transmit-only","6G ISAC: Joint cloud-fronthaul-radio optimization saves >50% energy","Sensing and comms: Joint resource allocation achieves >50% power savings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001257,"raw_usage":{"total_tokens":5023,"prompt_tokens":813,"completion_tokens":4210,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":4092}},"tokens_in":557,"tokens_out":4210,"duration_ms":29137,"temperature":1.0,"reasoning_tokens":4092,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T01:32:18.861447+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Restore the communication-interference terms in the sensing SINR expression (Eq. 15), or simulate the same system with actual communication waveforms present at the RX-APs, and re-run the proposed optimization; if the detection probability at the 0.03 false-alarm rate drops below the reported ~0.9 (FIS, R=1) or the power savings narrow materially, the central claim depends on the neglected interference.","supporting_citations":[],"review_version":1}