{"id":"98d90162-ef93-4417-9d78-8e97b751278d","arxiv_id":"2501.14675","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A digital over-the-air computation scheme leaves channel randomness in the received constellation so that different function values map to distinct received points almost surely, without explicit overlap-avoiding constellation optimization.","lead":"This paper proposes a way for many wireless devices to transmit at the same time and have a base station compute functions such as sums directly from the mixed signals, using a channel-aware digital constellation. The goal is to remove the need for complex constellation design and to avoid the excessive transmit power that earlier digital over-the-air computation methods require.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1 guarantees collision-free almost surely, but not a positive minimum distance; dE can be arbitrarily small, so AR = dR/dE is unbounded, and the unproven parameter adjustment in Sec. III-B leaves the central reliability claim unsupported.","rationale":"The paper's own text concedes the issue: Section III-B says 'while the constellation points ... do not overlap, a small dE can occur, leading to a large AR. This may exceed the limit ARM...' The proof of Proposition 1 uses a measure-zero hyperplane argument, which is correct for continuous i.i.d. fading. But the proposition is then used to justify reliability and fixed-power transmission; that inference fails because the minimum distance has no positive lower bound. The amplification factor AR in Eq. (14) is dR/dE when dE < dR; although multiplying signal and noise by AR preserves the symbol SNR, a finite receiver gain ARM turns small dE events into outages, and no outage or AR analysis is reported. The proposed fix ('CP can adjust a1, a2, Q1, Q2, θ') is an assertion, not a derivation; it also requires reconfiguring the common transmit constellation inside the coherence block, which is nontrivial under the protocol of Fig. 1. The reader's weakest_assumption correctly identified practical fragility, but the more precise issue is not finite precision alone: even with perfect continuous channel knowledge, the near-collision probability is positive and the paper does not control it. I therefore partially agree with the reader; the verdict remains CONDITIONAL, with the condition sharpened to require either a proof that the parameter adjustment always yields AR ≤ ARM with bounded loss, or an AR distribution and outage analysis.","tokens_in":16656,"tokens_out":7410,"duration_ms":73095,"concrete_test":"Run a Monte Carlo test on the Section V setup (K = 4, Q = 8, sum function, Rayleigh fading with Table I parameters). For 10^6 channel realizations, compute dE via Eq. (12) and AR via Eq. (14) with dR chosen as the receiver decision threshold (e.g., 3σ). Record the empirical complementary CDF of AR and the frequency of AR > ARM for a practical ARM (e.g., 30 dB). Then attempt to implement the Section III-B adjustment: for each realization, search over a1, a2, Q1, Q2, θ (with Q1Q2 ≥ Q) for parameters satisfying AR ≤ ARM while preserving f(i) ≠ f(j) ⇒ si ≠ sj; if the search fails or the outage rate stays non-negligible, the unproven adjustment does not rescue the scheme.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the no-overlap guarantee of Proposition 1 (Appendix A). That proof only establishes that the event of an exact collision has probability zero. It does not establish any positive lower bound on dE = min_{i≠j, f(i)≠f(j)} |Σ_k sqrt(Pt) |h_k| (x_{k,q_i} − x_{k,q_j})|. For any ε > 0, P(dE < ε) > 0 under Rayleigh fading, because the finite minimum of continuous random variables has a density with mass arbitrarily close to zero. Consequently, in Eq. (14), AR = dR/dE is unbounded over the channel ensemble. The paper explicitly acknowledges this in Section III-B, noting that a small dE may exceed the receiver limit ARM, and states that the CP can adjust a1, a2, Q1, Q2, θ so that AR ≤ ARM. However, no existence result, algorithm, or performance analysis is provided for that adjustment, and a change of Xnorm would need to be communicated to all nodes within the coherence block. The numerical sections report NMSE but do not report the distribution of AR or the outage probability P(AR > ARM). Without such analysis, the claim that channel-aware constellations 'inherently avoid overlap' and enable reliable fixed-power OTA computation is not established; the no-overlap property is too weak a guarantee. Finite-precision channel estimates are a secondary but related issue: they make exact collisions possible and make the a.s. argument inapplicable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a digital over-the-air (OTA) computation scheme in which the computation point builds the demodulation constellation from the channel coefficients (or their magnitudes) of the participating nodes. Because the constellation is channel-aware, the authors claim that overlap between constellation points representing different function values is inherently avoided, that nodes can transmit at fixed power instead of inverting channel gains, and that both symmetric and asymmetric functions can be computed. The scheme is extended to cellular and cell-free massive MIMO, and NMSE simulations are reported for several functions, node counts, and quantization levels. The paper's central formal claims are Proposition 1 (the minimum distance between constellation points corresponding to different function values is positive) and Proposition 2 (a one-to-one mapping from transmit vectors to combined constellation points), with proofs in Appendices A and B.","tokens_in":16997,"tokens_out":6038,"duration_ms":58344,"significance":"If the central reliability claim were established, this would be a useful contribution to digital OTA computation: it would remove the combinatorial constellation-design step of earlier schemes, avoid high transmit power under poor channels, and broaden the class of computable functions. The paper has clear strengths: Proposition 1 is a correct measure-theoretic statement for continuous i.i.d. Rayleigh channels, the fixed-power Type II transmission is practically attractive, the numerical study covers several functions and both cellular and cell-free architectures, and the simulations appear to be honest evaluations rather than curve fitting to a target. The significance is nonetheless conditional because the paper does not establish a positive lower bound on the minimum receive distance, does not analyze the outage probability of the receiver amplification factor, and contains an inconsistency in its treatment of channel estimation error.","major_comments":[{"comment":"Proposition 1 only proves that the event of an exact collision between two combined constellation points has probability zero for continuous i.i.d. fading. It does not establish any positive lower bound on dE = min([D_{i,j}]). For any epsilon > 0, P(dE < epsilon) > 0 under Rayleigh fading, because the finite minimum of continuous random variables has a density with mass arbitrarily close to zero. Consequently, the amplification factor AR = dR/dE defined in Eq. (14) is unbounded over the channel ensemble. The paper itself acknowledges this in Section III-B by stating that the CP may adjust a1, a2, Q1, Q2, and theta so that AR <= ARM, but no existence result, algorithm, or performance analysis is provided for that adjustment. The numerical sections do not report the distribution of AR or the outage probability P(AR > ARM), so the NMSE results do not validate the claimed reliable fixed-power operation under a receiver amplification limit.","section":"Section III-A, Eq. (12)-(14), and Appendix A"},{"comment":"The treatment of channel estimation error is internally inconsistent. For Type I, Eq. (19) gives the noiseless received term as AR sqrt(Pt) sum_k (h_k hhat_k^*/|hhat_k|) x_k, but the combined constellation used for demodulation is defined as s_tilde_m = AR sqrt(Pt) sum_k |hhat_k| x_{k,q_k,m}. These two expressions are not equal unless h_k = hhat_k. For Type II, the received signal contains AR sqrt(Pt) sum_k h_k x_k, while the demodulation constellation uses hhat_k in place of h_k. The paper provides no analysis of this mismatch, and the almost-sure no-overlap result of Proposition 1, which was proved for the true channels, does not apply to the estimated-channel constellation actually used at the receiver. This is a load-bearing issue because the massive MIMO simulations in Section IV are based on MMSE estimates, so the numerical results may not reflect the proposed demodulation rule.","section":"Section III-C, Eqs. (19)-(20)"},{"comment":"The same estimation-error mismatch appears in the multi-antenna extension. For Type I, the combined constellation is defined with |sum_{nA} hhat_{k,nA}|, whereas the actual noiseless received contribution of node k is (sum_{nA} h_{k,nA}) (sum_i hhat_{k,i}^*)/|sum_i hhat_{k,i}|, which is not equal to |sum_i hhat_{k,i}| in general. For Type II, the constellation uses sum_{nA} hhat_{k,nA} while the received signal contains sum_{nA} h_{k,nA}. The paper should either correct these definitions or explicitly adopt a small-error approximation with a quantified bound on the resulting demodulation error. As written, the claimed reliability of the massive MIMO extensions is not supported.","section":"Section IV-B, Eqs. (28)-(29)"}],"minor_comments":[{"comment":"The NMSE definition appears to have a typo: the denominator should likely be N_s * |f_max - f_min|^2 rather than N_s |f_max - f_min|^2 without the multiplication dot, since the sum is over N_s trials.","section":"Section V, Eq. (36)"},{"comment":"The caption states \"0 <= x_tilde_k <= 7, ˘x_k in {0,...,7}\" but the quantization mapping from the input value x_tilde_k to the quantized scalar ˘x_k is never defined. Please clarify the quantization rule used in the simulations.","section":"Fig. 4 caption and Section II"},{"comment":"The notation for the decoded function in Eq. (20) uses D_tilde on an argument containing the true channels, while the demodulation constellation is defined with estimated channels only. Please state explicitly what demodulation rule is implemented at the receiver when only hhat_k is available, and align the notation with that rule.","section":"Section III-C"}],"recommendation":"major_revision","confidential_remarks":"The paper contains an appealing idea and a valid measure-theoretic observation, but the current draft does not support the central reliability claim: the no-overlap property is too weak, the amplification factor is unbounded and not analyzed, and the channel-estimation-error section contains a substantive model inconsistency. I would like to see a corrected Section III-C, a precise treatment of the AR constraint (including an outage analysis or a positive-distance guarantee under a suitable model), and numerical reporting of AR or of the outage probability before reconsidering the manuscript. The topic fits the journal's scope, and the fixed-power scheme is worth pursuing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a real incremental step in digital OTA computation. The core move—letting continuous channel randomness break the symmetry that causes overlapping combined constellation points—is simple and different from ChannelComp and SumComp. Proposition 1 is correct under i.i.d. continuous fading: exact collisions are measure-zero. Proposition 2's injectivity claim is also fine for the same reason. The fixed-power transmit coefficient (Type II) is worth having, and the cell-free LPCV with weighted voting is a sensible extension. The simulations are honest evaluations, not fitted to a target, and the self-citations are background rather than load-bearing.\n\nThe soft spots are real and they sit on the central claim. Proposition 1 only guarantees no exact overlap. It does not give a lower bound on dE, the minimum distance between constellation points that represent different function values. Under Rayleigh fading, dE can be arbitrarily small with positive probability, so the receiver amplification AR = dR/dE is unbounded. The paper acknowledges this in Sec. III-B and says the CP can adjust a1, a2, Q1, Q2, and theta so that AR <= ARM, but it provides no existence result, no algorithm, and no analysis of the channel realizations where this fails. That is not a minor omission—it is the difference between \"almost surely distinguishable in the noiseless limit\" and \"reliable at finite SNR.\" The simulations report NMSE but never report the distribution of AR, the outage P(AR > ARM), or NMSE conditioned on near-collisions. Finite-precision channel estimates are secondary but related: with quantized estimates the almost-sure argument does not even apply.\n\nOther weaknesses are more ordinary: the Q^K combined constellation size grows exponentially, error bars are missing, and the analog OTA baseline is underspecified (power control, antenna count, and whether it uses the same cell-free setup). None of these are fatal on their own.\n\nWho is this for: researchers working on digital OTA computation and cell-free massive MIMO aggregation. They will get a clear, if narrow, idea to build on. The paper deserves a serious referee; the almost-sure argument is sound and the system-level contribution is concrete, but the reliability and power-saving claims need substantial revision before publication. A conditional accept with a request for AR/outage analysis and a finite-precision discussion would be appropriate.","headline":"Channel-aware constellations solve the overlap problem almost surely, but the paper never shows the minimum distance is large enough, so the reliability claim rests on an unanalyzed parameter adjustment.","tokens_in":17475,"tokens_out":2146,"would_cite":true,"duration_ms":20100,"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":"A channel-aware constellation built from estimated fading makes digital over-the-air computation overlap-free while allowing fixed-power transmission.","keywords":["digital over-the-air computation","channel-aware constellation","demodulation mapper","massive MIMO","cell-free massive MIMO","function computation","NMSE","fading channels"],"falsifier":"Run Monte Carlo trials with channel estimates quantized to $B$ bits and measure, across many channel realizations, how often two distinct function values produce combined constellation points closer than the receiver resolution $d_R$, together with the resulting NMSE as $B$ varies. If any finite $B$ yields a nonzero collision rate, or if NMSE rises steeply as $B$ shrinks, the almost-surely-overlap-free claim does not survive imperfect channel knowledge.","tokens_in":16459,"feed_emoji":"📡","tokens_out":8608,"duration_ms":71063,"temperature":0.7,"pith_summary":"This paper tries to show that digital over-the-air (OTA) computation can be made simpler and more power-efficient by building the receiver's demodulation constellation from the estimated channel conditions of the participating nodes, rather than designing a fixed constellation that provably keeps all function values apart. The central claim is that continuous random fading makes overlap between constellation points representing distinct function values an event of probability zero, so the costly overlap-avoiding constellation search can be dropped and nodes can transmit at fixed power $\\sqrt{P_t}$ instead of inverting their channels. A one-to-one mapping from transmit vectors to combined constellation points then lets the same receiver compute symmetric and asymmetric functions alike. The paper backs these claims with two propositions and with NMSE simulations in cellular and cell-free massive MIMO settings, comparing fixed-power and channel-conjugate transmit coefficients under different processing and voting strategies.","feed_headline":"Random fading makes digital over-the-air constellations overlap-free","feed_subtitle":"Fixed-power nodes and a channel-aware constellation keep function values distinct without overlap design.","key_machinery":"The load-bearing object is the channel-aware combined constellation at the computation point, whose points are channel-weighted sums of the nodes' modulated values. For Type I transmit coefficients (phase-only correction, fixed magnitude) the point is $\\tilde{s}_m = \\sum_k \\sqrt{P_t}\\,|h_k|\\,x_{k,q_{k,m}}$; for Type II (no channel correction) it is $\\breve{s}_m = \\sum_k \\sqrt{P_t}\\,h_k\\,x_{k,q_{k,m}}$. The randomness of the continuous channel coefficients makes distinct function values land on distinct points almost surely, so the receiver needs no combinatorial constellation search to avoid overlap, and the transmitter needs no channel inversion. This same object carries the one-to-one mapping that supports asymmetric functions and determines the minimum distance $d_E$ that, together with the receiver resolution $d_R$, sets the amplification factor $A_R = d_R/d_E$ when $d_E < d_R$.","core_discovery":"The paper's central claim is that a digital OTA computation system can avoid explicit overlap-avoiding constellation design by making the demodulation constellation at the computation point depend on the estimated channels of the participating nodes. With transmit coefficient Type I, the combined constellation point is $\\tilde{s}_m = \\sum_{k=1}^K \\sqrt{P_t}\\,|h_k|\\,x_{k,q_{k,m}}$; with Type II it is $\\breve{s}_m = \\sum_{k=1}^K \\sqrt{P_t}\\,h_k\\,x_{k,q_{k,m}}$. Because each $h_k$ (or $|h_k|$) is a continuous random variable, two distinct transmit vectors produce the same combined point only when the channel vector lies on a lower-dimensional hyperplane, an event of probability zero; this is Proposition 1. Proposition 2 turns the same reasoning into a one-to-one mapping from transmit vectors to combined constellation points, which is what allows asymmetric functions to be decoded. Since the channel is absorbed into the constellation rather than inverted, each node can transmit at fixed power $\\sqrt{P_t}$, eliminating the excessive transmit power that arises when power scales inversely with channel gain. The mechanism is then extended to cellular and cell-free massive MIMO, with simulations reporting NMSE for sum, product, maximum, and sum-of-squares functions under fully centralized processing and local processing with centralized voting.","pith_inferences":["An implication the authors leave implicit is that the practical outage behaviour is governed by the distribution of the minimum distance $d_E$ between distinct function values: designers could choose constellation parameters $a_1, a_2, Q_1, Q_2, \\theta$ from a target probability that $d_E$ falls below the receiver resolution $d_R$, rather than relying on the almost-sure statement alone.","The mechanism only requires continuous random channel coefficients, so the same fixed-power, channel-aware demodulation idea should transfer to other superposition-based aggregation tasks such as federated-learning gradient aggregation, where reducing device transmit power is a major concern.","A testable extension is to replace the Rayleigh fading model with measured channel statistics: the no-overlap guarantee should degrade gracefully as the empirical channel distribution becomes discrete or quantized, and the degradation rate can be checked by counting near-collisions in field data."],"forward_implications":["Nodes can transmit at fixed power $\\sqrt{P_t}$ instead of scaling power inversely with channel quality, removing the excessive transmit-power problem under poor fading.","The computation point no longer needs to solve a combinatorial constellation-design problem to guarantee that distinct function values map to distinct received points, cutting computational complexity.","Because every transmit vector maps one-to-one to a combined constellation point, the receiver can compute asymmetric functions as well as symmetric ones, broadening the range of usable target functions.","The scheme extends to cellular and cell-free massive MIMO, and the cell-free case offers a tunable trade-off: fully centralized processing gives the lowest NMSE at high fronthaul cost, while local processing with centralized voting and channel-based weights reduces fronthaul with only a modest NMSE penalty in most simulated settings.","Numerical results indicate that the fixed-power Type II transmit coefficient can outperform the channel-conjugate Type I as the cell radius grows, because the larger separations between some constellation points help at low signal-to-noise ratio."],"supporting_citations":[{"why":"Supplies the prior digital OTA computation method whose overlap-avoiding constellation design this work removes.","marker":"[13]"},{"why":"Represents the follow-up digital coding scheme whose computational complexity motivates the channel-aware simplification.","marker":"[14]"},{"why":"Provides the cell-free massive MIMO OTA model and channel-estimation setup that the proposed constellation design extends.","marker":"[16]"},{"why":"Supplies the MMSE channel-estimation machinery used to form the estimated channels feeding the constellation.","marker":"[17]"},{"why":"States the computation-over-multiple-access-channels principle that makes superposition-based function evaluation possible.","marker":"[5]"}],"fun_headline_variants":["Channel-aware constellations avoid overlap without explicit design in OTA computation","Random fading replaces explicit constellation overlap avoidance in OTA computation","Fixed-power nodes and channel-aware mapping make OTA computation overlap-free","Channel-aware constellations turn fading into a feature for digital OTA computation","No explicit overlap design: channel-aware constellations for reliable OTA computation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scheme's no-overlap guarantee assumes the receiver knows each channel coefficient (or its magnitude) exactly and that the coefficients are continuous random variables, so collisions occur only on a set of channel outcomes with probability zero; with finite-precision, correlated channel estimates, those collisions become possible and near-collisions can force large noise-amplifying receiver gains.","fun_headline_variants_meta":{"raw":{"variants":["Channel-aware constellations avoid overlap without explicit design in OTA computation","Random fading replaces explicit constellation overlap avoidance in OTA computation","Fixed-power nodes and channel-aware mapping make OTA computation overlap-free","Channel-aware constellations turn fading into a feature for digital OTA computation","No explicit overlap design: channel-aware constellations for reliable OTA computation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000639,"raw_usage":{"total_tokens":2980,"prompt_tokens":1022,"completion_tokens":1958,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":1870}},"tokens_in":638,"tokens_out":1958,"duration_ms":13320,"temperature":1.0,"reasoning_tokens":1870,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:54:59.016367+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Monte Carlo trials with channel estimates quantized to $B$ bits and measure, across many channel realizations, how often two distinct function values produce combined constellation points closer than the receiver resolution $d_R$, together with the resulting NMSE as $B$ varies. If any finite $B$ yields a nonzero collision rate, or if NMSE rises steeply as $B$ shrinks, the almost-surely-overlap-free claim does not survive imperfect channel knowledge.","supporting_citations":[{"cited_title":"ChannelComp: A General Method for Computation by Communications,","cited_arxiv_id":null,"evidence_quote":"Supplies the prior digital OTA computation method whose overlap-avoiding constellation design this work removes."},{"cited_title":"SumComp: Coding for Digital Over-the-Air Computation via the Ring of Integers,","cited_arxiv_id":null,"evidence_quote":"Represents the follow-up digital coding scheme whose computational complexity motivates the channel-aware simplification."},{"cited_title":"Making Cell-Free Massive MIMO Competitive With MMSE Processing and Centralized Implementation,","cited_arxiv_id":null,"evidence_quote":"Supplies the MMSE channel-estimation machinery used to form the estimated channels feeding the constellation."},{"cited_title":"Computation over multiple-access channels,","cited_arxiv_id":null,"evidence_quote":"States the computation-over-multiple-access-channels principle that makes superposition-based function evaluation possible."}],"review_version":1}