{"id":"304e631a-c629-4202-ae64-d45c0c382d87","arxiv_id":"2506.00655","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A two-phase over-the-air computation scheme lets the central processor of a cell-free massive MIMO system obtain the Gramian and matched-filter sufficient statistics directly from superimposed access point transmissions.","lead":"This paper proposes transmitting the two key statistics needed to decode uplink data, the Gramian and the matched filter, from many access points to a central processor over the air, letting the wireless channel add them up naturally. If it works, cell-free massive MIMO networks could avoid expensive wired fronthaul and scale the number of access points without proportionally increasing fronthaul resources.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Perfect synchronization is load-bearing for Eq. (14); the paper assumes it without modeling offsets, so the central OTA claim rests on an unverified idealization.","rationale":"The paper's central claim is that global sufficient statistics can be computed OTA with minimal performance loss. Equation (14) is the linchpin: it states the CPU observes a clean scaled sum of the local sufficient statistics plus noise. This requires perfect phase alignment of all AP transmissions. The paper explicitly assumes perfect synchronization in Footnote 1 and references synchronization literature instead of modeling errors. This is a standard idealization for a first framework paper, but it is load-bearing because any residual phase offset transforms the received signal into a weighted sum with random phases, invalidating the LMMSE/LS estimators (25)-(26), the SINR expression (41), and the closed-form MSEs (30)-(36). The paper does not provide any analysis or simulation of this degradation, unlike the imperfect UE-AP CSI case which is treated numerically in Fig. 10. Therefore, the central claim is conditional on an unverified assumption. The concrete test—adding per-AP phase offsets and measuring the resulting SER/rate gap—would quantify the sensitivity. If the gap is small for realistic offsets (a few degrees), the assumption is benign; if large, the paper's central claim needs qualification. The rest of the mathematical development (power control, LMMSE derivation, error floor analysis) is internally consistent and supported by simulations that match the derived expressions. No other concern I found rises to the same level: the imperfect AP-CPU CSI section sketches the error term but is secondary; the UatF bound is a valid lower bound for the given receiver; the novelty comparison to [12] cannot be resolved from this text alone.","tokens_in":19741,"tokens_out":31847,"duration_ms":281455,"concrete_test":"Modify the received-signal model in Eq. (14) to Z^{(i)} = sqrt(eta_c^{(i)}) sum_l e^{j phi_l} \\bar X_l^{(i)} + E, with phi_l i.i.d. uniform on [-Delta, Delta] degrees and independent across APs and coherence blocks, while keeping perfect CSI. Re-run the SER simulation of Fig. 3 and the rate CDF of Fig. 4 for Delta = 1, 5, 10 degrees. If the gap to the wired baseline grows by more than about 3 dB at a target SER or rate, perfect synchronization is indeed load-bearing and the paper should either model it or qualify the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (14) is the linchpin: the CPU observes Z^{(i)} = sqrt(eta_c^{(i)}) sum_l \\bar X_l^{(i)} + E, which requires every AP's transmission to arrive at the CPU perfectly phase-aligned. The paper states this as an assumption in Footnote 1 and cites [14], [15] rather than modeling synchronization errors. With a per-AP phase offset phi_l, the received signal becomes sqrt(eta_c) sum_l e^{j phi_l} \\bar X_l + E, which is not proportional to the desired sum. The LMMSE/LS estimators in (25)-(26), the SINR expression (41), and the closed-form MSEs in (30)-(36) all rely on the coherent model; under offsets, the deviation acts as an additional self-interference term whose statistics are not characterized. The paper provides no analysis or simulation of this degradation, in contrast to the imperfect UE-AP CSI case (Section III-E, Fig. 10). Because the central claim that sufficient statistics can be aggregated OTA depends entirely on this coherent sum, the assumption is load-bearing and unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an over-the-air (OTA) computation framework for the fronthaul of uplink cell-free massive MIMO systems. Each access point computes local sufficient statistics (the Gramian matrix and the matched-filter output), precodes them with a local zero-forcing precoder, and transmits them simultaneously to the central processing unit (CPU) using a common power-scaling factor, so that the CPU receives a noisy superposition proportional to the desired global sums. The manuscript provides a two-phase transmission protocol, LMMSE and LS estimators for the sufficient statistics, closed-form MSE expressions, an achievable-rate analysis based on the use-and-then-forget bound, an extension to imperfect CSI, and a comparison with a digital orthogonal fronthaul scheme. Numerical results show that the proposed OTA scheme closely matches the performance of a wired fronthaul in the considered scenarios.","tokens_in":19854,"tokens_out":23371,"duration_ms":213261,"significance":"If the claimed results hold, the framework is a notable step toward scalable wireless fronthaul for cell-free massive MIMO, because the number of fronthaul channel uses needed for the sufficient statistics is independent of the number of APs. The paper's strengths include closed-form performance expressions that are verified by simulation, a careful treatment of power scaling and error floors, and a quantitative comparison with digital fronthaul. The main risk is the reliance on perfect synchronization between the APs and the CPU, which is load-bearing for the coherent OTA sum.","major_comments":[{"comment":"The coherent OTA sum in Eq. (14) assumes that all AP-to-CPU transmissions arrive at the CPU perfectly phase-aligned. The paper adopts this as an assumption in Footnote 1 and cites [14], [15], but provides no model or robustness analysis for synchronization errors. With per-AP phase offsets, the received signal becomes a weighted sum with phase rotations that is not proportional to the desired sum, and the LMMSE/LS estimators in Eqs. (25)-(26), the SINR expression in Eq. (41), and the closed-form MSEs in Eqs. (30)-(36) no longer hold. Since the central claim that the global sufficient statistics can be computed OTA depends entirely on this coherent summation, the manuscript should either incorporate synchronization errors into the system model or provide a quantitative study of the degradation, along the lines of the imperfect-CSI treatment in Section III-E and Fig. 10.","section":"Section II-A, Eq. (14) and Footnote 1"},{"comment":"The definition of the common power-scaling factor is incomplete. The set V^(i) in Eq. (13) contains only APs that violate the average power constraint, and if no AP violates the constraint, V^(i) is empty and eta_c is undefined; the text should specify eta_c = 1 in that case. In addition, the condition N >= M stated in Section II-A does not guarantee that the expected transmit power in Eq. (10) is finite: for N = M, the expectation E{||G_l(G_l^H G_l)^{-1}||^2} diverges for Gaussian G_l, so the framework should assume N > M (or N >= M+1) for the average power constraint to be meaningful.","section":"Section II-A, Eqs. (10) and (13)"}],"minor_comments":[{"comment":"There is a duplicated word in the sentence 'as the the number of APs increases' in the related-work discussion.","section":"Section I-A"},{"comment":"The phrase 'over-the-air(OTA)' is missing a space before the parenthesis; it should read 'over-the-air (OTA)'.","section":"Abstract"},{"comment":"The pilot model leading to the channel-estimation error covariance in Eq. (51) is not stated. The sum over all UEs in the inverse suggests that all UEs share the same pilot sequence; if orthogonal pilots are used, the formula should be different. Please specify the pilot structure.","section":"Section III-E, Eq. (51)"},{"comment":"The statement that 'the CPU does not need to know the Gramian matrix A, but only its first and second order statistics' should be clarified, because the SINR expression in Eq. (41) involves the combining vector v_k, which in the LS example depends on the estimate of A. The required statistical knowledge at the CPU should be stated explicitly.","section":"Section III-B"},{"comment":"The ceiling notation in these equations is garbled in the displayed text (the 'V' symbols appear to represent ceiling brackets). Please typeset the ceiling operators correctly.","section":"Section IV, Eqs. (56), (61), (62)"}],"recommendation":"major_revision","confidential_remarks":"The technical core of the manuscript is sound under the stated ideal assumptions, and the closed-form results are a clear strength. The main concern is the perfect-synchronization assumption, which is load-bearing for the central OTA claim; I would like the authors to address it with either a model or an explicit robustness discussion before I can recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a solid, internally consistent paper on using over-the-air (OTA) computation to aggregate sufficient statistics—the Gramian and matched-filter outputs—from APs to a CPU in uplink cell-free massive MIMO. The two-phase transmission, the common power scaling that equalizes per-AP gains, and the closed-form MSE, achievable rate, and error-floor analysis are genuine contributions. The simulations match the derived expressions, and the comparison with a digital fronthaul is fair: OTA uses far fewer channel uses, at the cost of analog transmission without quantization. If you work on OTA computation or cell-free fronthaul, this is worth reading.\n\nThe derivations in Sections III-A and III-C hold up; I checked the MSE expressions and the error-floor argument. The observation that eta_c^{(2)} falls as UE power grows, causing an SER floor, is a nice, non-obvious consequence of the power constraint.\n\nThe soft spots are real but not disqualifying. The biggest is the perfect-synchronization assumption behind Eq. (14). The coherent OTA sum demands all AP transmissions arrive phase-aligned at the CPU. The paper explicitly states this in Footnote 1 and cites synchronization references, but it never models phase or timing offsets. With per-AP phase errors, the received signal is not the desired sum; the LMMSE/LS estimators and the SINR analysis all rest on the coherent model. That is load-bearing. I wouldn't desk-reject over it, but a referee should ask for a sensitivity analysis or at least a quantified discussion of how much misalignment the scheme tolerates.\n\nSecond, the novelty claim relative to [12] is not pinned down. The paper says [12] designs transmit coefficients and receive combining for OTA in cell-free massive MIMO, but it doesn't explain why computing the Gramian and matched filter isn't already covered there. A clean comparison is needed.\n\nThird, the achievable rate in Section III-B is a use-and-then-forget bound that assumes the CPU does not know the Gramian, only its statistics, while the detectors in Section II-D use the estimated Gramian. This is acknowledged implicitly but could be stated more carefully. It's a presentation issue, not a technical error.\n\nOverall: a solid engineering paper with honest derivations. It deserves serious peer review. The synchronization gap is the main thing to push on; the rest is addressable in revision.","headline":"Solid OTA-computation framework for cell-free massive MIMO with clean closed-form analysis; perfect synchronization is a genuine gap that needs sensitivity analysis, but the paper deserves peer review.","tokens_in":20470,"tokens_out":3603,"would_cite":true,"duration_ms":32809,"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 shows that in cell-free massive MIMO, the global sufficient statistics needed to decode uplink data can be computed over the air at the central processor with performance close to a wired fronthaul.","keywords":["cell-free massive MIMO","over-the-air computation","sufficient statistics","wireless fronthaul","uplink data detection","zero-forcing precoding","Gramian matrix","matched-filter output"],"falsifier":"Simulate or build a setup where APs transmit a known pilot with the proposed ZF precoder while their clocks have controlled phase offsets relative to the CPU; if the received signal deviates from the coherent sum predicted by Eq. (14) whenever offsets exceed a fraction of a symbol period, the claimed OTA summation and the derived MSE and rate expressions no longer hold.","tokens_in":19478,"feed_emoji":"📡","tokens_out":6848,"duration_ms":64801,"temperature":0.7,"pith_summary":"The paper seeks to establish that an uplink cell-free massive MIMO network can decode user data without sending each access point's raw signal over a wired fronthaul. Instead, the access points transmit their local sufficient statistics as analog signals, and the wireless channel itself sums them at the central processing unit. By using local zero-forcing precoders and a common power scaling factor, the CPU receives a coherent sum of all AP contributions plus noise, which is exactly what data detection needs. The authors derive closed-form mean-square errors, achievable rates, and symbol-error and bit-error expressions showing performance close to an ideal wired fronthaul while fronthaul resource use stays constant as the number of APs grows.","feed_headline":"Cell-free MIMO can decode uplink data without wired fronthaul","feed_subtitle":"Over-the-air summation at the CPU matches wired performance while fronthaul load stays flat as APs grow.","key_machinery":"The central objects are the local sufficient statistics at each access point: the Gramian $A_l = H_l^H H_l$ and the matched-filter output $t_l = H_l^H y_l$, whose sums $A$ and $t$ are sufficient for ML detection. The mechanism that carries the argument is the combination of local zero-forcing precoding on the fronthaul channel and a common average-power scaling factor $\\eta_c^{(i)}$; together they make the CPU's received matrix equal to the desired sum plus noise. The two phases use separate scalings because the matched-filter output's dynamic range depends on user transmit power, and this scaling choice is what creates the observed error floor in symbol-error rate.","core_discovery":"On its own terms, the paper's discovery is that the sufficient statistics for uplink maximum-likelihood detection decompose naturally as sums over access points: the Gramian $A = \\sum_l H_l^H H_l$ and the matched-filter output $t = \\sum_l H_l^H y_l$. With zero-forcing precoders $W_l = G_l(G_l^H G_l)^{-1}$ and a common power scaling $\\eta_c^{(i)}$, the CPU receives $Z^{(i)} = \\sqrt{\\eta_c^{(i)}} \\sum_l \\bar{X}_l^{(i)} + E^{(i)}$, a coherent sum of the local sufficient-statistic contributions plus noise. That makes the wireless fronthaul a physical summation channel rather than a data pipe. The paper then provides LMMSE and LS estimators for the summed sufficient statistics, closed-form NMSE expressions, a use-and-then-forget achievable rate, and numerical SER/BER results; the conclusion is that this over-the-air scheme achieves essentially the same performance as an ideal wired fronthaul while keeping fronthaul channel use independent of the number of APs.","pith_inferences":["A natural extension not developed in the paper is to replace the perfect phase-alignment assumption with per-AP phase correction based on a common pilot; then residual synchronization error, rather than the OTA summation itself, would set the achievable accuracy.","The paper notes that the sufficient-statistic structure is valid for any input distribution, so testing the same two-phase OTA scheme on continuous analog sensor data would be a direct follow-up that the paper does not simulate.","Because the error floor is caused by the common scaling factor being limited by the weakest AP, a hybrid fronthaul that offloads only the worst APs to digital or wired links should lift the floor without giving up the scalability of OTA for the rest of the network."],"forward_implications":["The number of fronthaul channel uses needed to deliver the sufficient statistics does not grow with the number of access points, unlike orthogonal digital fronthaul.","With perfect fronthaul CSI and ZF precoding, the CPU can use LMMSE or LS estimation of the summed Gramian and matched-filter output, with closed-form MSE expressions that the paper verifies numerically.","Per-user achievable rates and SER/BER performance closely match those of an ideal wired fronthaul, and coded transmission further narrows the gap.","An error floor appears in SER as user transmit power grows, because the common scaling factor $\\eta_c^{(2)}$ decreases to satisfy the AP power constraint.","Imperfect CSI at the APs and at the CPU can be incorporated through approximate sufficient statistics, with only modest degradation when pilot power is adequate."],"supporting_citations":[{"why":"Supplies the cell-free massive MIMO system model and the wired-fronthaul baseline the paper compares against.","marker":"[2]"},{"why":"Provides an earlier OTA precoding design for cell-free massive MIMO that the paper extends from downlink precoding to uplink sufficient-statistics aggregation.","marker":"[4]"},{"why":"Supports the claim that local processing at APs can keep the fronthaul load from scaling with the number of APs.","marker":"[13]"},{"why":"Cited as the synchronization study underlying the assumption that all AP and CPU transmissions arrive phase-aligned.","marker":"[14]"},{"why":"Cited together with [14] as the synchronization mechanism the paper relies on for perfect phase alignment.","marker":"[15]"},{"why":"Supplies the LMMSE and MVU/LS estimator theory and the error-covariance formulas used to derive the MSE expressions.","marker":"[19]"},{"why":"Provides the use-and-then-forget bound and massive MIMO rate framework used for the achievable-rate expressions.","marker":"[20]"},{"why":"Supplies the urban microcell path-loss model used in the numerical evaluations.","marker":"[23]"},{"why":"Supplies the LDPC code used for the coded BER results.","marker":"[24]"}],"fun_headline_variants":["Over-the-air fronthaul scales in cell-free MIMO uplink","Cell-free MIMO uplink decoded without wired fronthaul","OTA summation replaces wired fronthaul in cell-free MIMO","Sufficient statistics summed over the air for cell-free MIMO","Wireless fronthaul matches wired in cell-free MIMO uplink"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every access point and the CPU share perfect phase and timing synchronization, so all over-the-air contributions arrive aligned at the CPU; the paper assumes this in a footnote and only points to synchronization studies without modeling errors.","fun_headline_variants_meta":{"raw":{"variants":["Over-the-air fronthaul scales in cell-free MIMO uplink","Cell-free MIMO uplink decoded without wired fronthaul","OTA summation replaces wired fronthaul in cell-free MIMO","Sufficient statistics summed over the air for cell-free MIMO","Wireless fronthaul matches wired in cell-free MIMO uplink"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000574,"raw_usage":{"total_tokens":2736,"prompt_tokens":997,"completion_tokens":1739,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":1649}},"tokens_in":613,"tokens_out":1739,"duration_ms":11656,"temperature":1.0,"reasoning_tokens":1649,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:02:38.037547+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate or build a setup where APs transmit a known pilot with the proposed ZF precoder while their clocks have controlled phase offsets relative to the CPU; if the received signal deviates from the coherent sum predicted by Eq. (14) whenever offsets exceed a fraction of a symbol period, the claimed OTA summation and the derived MSE and rate expressions no longer hold.","supporting_citations":[{"cited_title":"Foundations of user- centric cell-free massive MIMO,","cited_arxiv_id":null,"evidence_quote":"Supplies the cell-free massive MIMO system model and the wired-fronthaul baseline the paper compares against."},{"cited_title":"Distributed precoding design via over-the-air signaling for cell-free massive MIMO,","cited_arxiv_id":null,"evidence_quote":"Provides an earlier OTA precoding design for cell-free massive MIMO that the paper extends from downlink precoding to uplink sufficient-statistics aggregation."},{"cited_title":"Distributed computation of a posteriori bit likelihood ratios in cell-free massive MIMO,","cited_arxiv_id":null,"evidence_quote":"Supports the claim that local processing at APs can keep the fronthaul load from scaling with the number of APs."},{"cited_title":"Massive synchrony in distributed antenna systems,","cited_arxiv_id":null,"evidence_quote":"Cited as the synchronization study underlying the assumption that all AP and CPU transmissions arrive phase-aligned."},{"cited_title":"Beamsync: Over-the-air synchronization for distributed massive MIMO systems,","cited_arxiv_id":null,"evidence_quote":"Cited together with [14] as the synchronization mechanism the paper relies on for perfect phase alignment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LMMSE and MVU/LS estimator theory and the error-covariance formulas used to derive the MSE expressions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LDPC code used for the coded BER results."}],"review_version":1}