{"id":"ab5ea9e4-1a1e-473a-a76a-3dd57b995a4c","arxiv_id":"2506.05944","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A GaBP-based receiver detects QPSK data and computes AirComp sums at the same time, matching prior state-of-the-art performance without large matrix inversions.","lead":"This paper proposes a receiver design that lets a base station simultaneously decode data from many users and compute aggregate functions like sums over the air using a message-passing algorithm. The design scales to hundreds of antennas and users, and it generalizes to multiple computation streams.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The GaBP loop that replaces the AirComp matrix inversion solves a noiseless, dense linear system b=Au; its convergence to A^{-1}b is unverified and the Gaussian prior on u is ad hoc.","rationale":"I read the paper's central claim as a complexity/performance trade-off: a GaBP receiver that simultaneously detects data and computes f(s) with a matrix-inversion-free combiner at O(NK)+O(N^2). The diagonal approximation to R_wtilde (eq. 30) is a concern, but it has asymptotic support from channel hardening and small sigma_s^2, and the paper's simulations are in that regime. The combiner GaBP loop is less secure: it is a noiseless linear solve on a dense loopy graph, and the only stabilizing ingredient is an ad hoc Gaussian prior whose parameters are not derived from the statistics of u. This is the point where the argument is least supported by derivation or external evidence. I do not claim the loop is broken; the simulations suggest it worked in the tested cases. But the paper provides no convergence analysis, no complexity accounting for forming A (O(N^2K)), and no reported beta_u, so the central claim is conditional on the loop behaving well. This does not change the reader's overall CONDITIONAL verdict; it sharpens the condition that must be verified.","tokens_in":20983,"tokens_out":16443,"duration_ms":180857,"concrete_test":"Run Algorithm 2's combining stage in isolation for N=K=100 i.i.d. Rayleigh channels with E_D=1 and sigma_w^2 set for SNR_S in {-20, 0, 20} dB, forming A with Omega from eq. (32) at the final data iteration. Compare the GaBP output u_hat to the direct solve A^{-1}b using relative error ||u_hat - A^{-1}b||_2 / ||A^{-1}b||_2, sweeping beta_u in {0.3, 0.5, 0.8} and i_max in {10, 30, 100}. Also report the beta_u used to generate Fig. 3. If any configuration gives relative error above 1%, the claim that the GaBP loop replaces the matrix inversion is not established outside that configuration.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The framework's headline contribution is that the combiner u can be obtained by a second GaBP loop over b=Au (Section IV-B, eqs. (44)-(58), Algorithm 2 steps 7-12) instead of the O(N^3) inversion of A. This is load-bearing for both the complexity and the 'matrix-inversion-free' claims. Unlike the data-detection loop, this loop has no observation-noise term: the conditional variance in eq. (51) is just residual interference, so the likelihood messages can become overconfident as residuals shrink. GaBP on a dense loopy factor graph is not guaranteed to converge to the linear solve; convergence conditions such as those in [48] are not checked for A. The Gaussian prior used in eq. (55) is initialized with sigma_u^2=1 and mu_u updated by eq. (58), but the entries of u=A^{-1}b are not i.i.d. N(0,1); for i.i.d. Gaussian H the per-entry variance is typically O(1/N), so the prior is not matched to the quantity being estimated. The EM update (58) averages N^2 extrinsic estimates with no derivation that this estimates the true mean. If the loop's fixed point is not A^{-1}b, the NMSE agreement in Fig. 3 is only a statement about the particular simulated parameters, and the promised O(N^2) combiner is not a reliable replacement for matrix inversion. The paper does not report beta_u or give code, so this cannot be checked from the manuscript alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a Gaussian belief propagation (GaBP) framework for integrated communication and computing (ICC) receivers in a SIMO uplink. The receiver jointly estimates the data vector d and evaluates a nomographic function f(s)=φ(Σ_k ψ_k(s_k)) over-the-air. A benchmark algorithm first detects both d and the per-user computing signals s, then applies an MMSE combiner that requires an N×N matrix inversion. The proposed single-stream method treats Hs+w as an approximately diagonal effective noise, detects d by GaBP with soft interference cancellation, and replaces the combiner matrix inversion with a second GaBP loop that solves b=Au. The framework is then extended to multiple independent computing streams and to a multi-access setup where users transmit data, computing, or both. Simulations for systems with up to 200 antennas and 300 users show that the proposed algorithms match the benchmark BER/NMSE and approach matched-filter bounds in underloaded and fully loaded cases.","tokens_in":21296,"tokens_out":14149,"duration_ms":138832,"significance":"Assuming the two central approximations hold, the framework is a valuable step toward scalable ICC receivers: the MMSE combiner derivation in Appendix VII-A is explicit and self-contained, the GaBP updates are element-wise, and the multi-stream/multi-access extension is clean and general. The simulations cover a wider range of loading conditions than many AirComp papers and demonstrate that the proposed method approaches the matched-filter bound in the under/fully loaded cases. The main reservations are that the diagonal effective-noise approximation and the GaBP solution of the noiseless linear system are not supported by any formal analysis, and the prior variance and EM update in the combining loop appear to be mismatched to the problem; these points need to be addressed before the generality and complexity claims can be accepted.","major_comments":[{"comment":"The approximation R_tilde_w ≈ σ_s^2 diag(ξ) + σ_w^2 I is load-bearing for the soft-IC variances in Eq. (35), but its stated justification is incomplete. For i.i.d. CN(0,1) channels and E_S = E_D/K, the off-diagonal entries of σ_s^2 H H^H have zero mean and standard deviation σ_s^2 sqrt(K) = E_D/sqrt(K), while the diagonal entries have mean σ_s^2 K = E_D. Hence at high SNR the relative off-diagonal magnitude is 1/sqrt(K), which is 0.11 for K=75 and 0.07 for K=200; the approximation is an uncontrolled O(1/sqrt(K)) neglect of correlated interference. The paper should either bound the resulting error in the variances of Eq. (35) or explicitly restrict the claim to regimes where this error is negligible.","section":"Section IV-A, Eq. (30)"},{"comment":"The claim that the combining vector can be obtained by a second GaBP loop over b = A u lacks a convergence guarantee. Unlike the data-detection loop, this system has no observation noise, so the likelihood variance in Eq. (51) is purely residual interference left by imperfect estimates; as the estimates improve, these variances shrink, and the messages become overconfident. The standard GaBP convergence conditions of reference [48] are not checked for the dense matrix A, and the paper gives no argument that the loop's fixed point equals A^{-1} b. Without such support, the NMSE agreement in Fig. 3 is evidence only for the simulated parameter sets, not for the general equivalence of the GaBP combiner to the exact MMSE combiner.","section":"Section IV-B, Eqs. (46)-(51), Algorithm 2 steps 7-12"},{"comment":"The Gaussian prior used in the combining loop is initialized with σ_u^2 = 1 and μ_u = 0, but for i.i.d. Gaussian H and C ≈ σ_s^2 I, the entries of u = A^{-1} b have typical per-entry variance of order K/[N(N-K)] for N > K (about 0.03 for N=100, K=75), not 1. The prior in Eq. (55) therefore biases the solution, and the EM update in Eq. (58), which averages all N^2 extrinsic estimates into a single mean μ_u, has no derivation: the entries of u are distinct random variables, not i.i.d. samples of a common parameter. This step should be either derived, replaced by a non-informative initialization, or reported as a tuned heuristic.","section":"Section IV-B, Eqs. (55)-(58)"},{"comment":"The claimed O(N^2) complexity for the combining stage is not supported by the equations as written. For each of the N^2 pairs (n,n'), Eqs. (49) and (51) require a sum over q≠n' of N terms, and Eqs. (53) and (54) require a sum over q≠n of N terms; a direct implementation is therefore O(N^3) per iteration. The O(N^2) complexity is achievable only by maintaining row-wise totals such as Σ_q a_{n,q} \\hat u^{(i-1)}_{n,q} and subtracting the self term for each message; this schedule should be stated explicitly, together with the number of combining iterations, before the complexity claim can be assessed.","section":"Section IV-C, Algorithm 2 combining stage"}],"minor_comments":[{"comment":"The sentence refers to 'equation (65)', which does not exist; it should refer to Eq. (46).","section":"Section IV-B, before Eq. (47)"},{"comment":"The combiner-loop variables \\hat u^{(0)}_{n,n'} and \\hat σ^{2(0)}_{u:n,n'} are not initialized, so the pseudocode is not fully reproducible.","section":"Algorithm 2, initialization"},{"comment":"Step 6 says to compute A and b from Eq. (44), but the data-error covariance Ω entering A is not defined in that section; it should be specified (as in Section III-B) in terms of the data-detection variances.","section":"Algorithm 2, step 6"},{"comment":"The damping factor β_u is listed as an input to Algorithms 2 and 3, but its value is not reported in the simulations, so the numerical results cannot be reproduced without code.","section":"Section IV-C, simulation parameters"},{"comment":"Footnote 9 notes that joint estimation of d and s is overloaded even for K<N, which conflicts with the 'underloaded' label used in the figures and text; the terminology should be reconciled.","section":"Footnote 9 and Figs. 2, 3, 6"},{"comment":"The abstract's claim of 'up to 200 antennas and 300 UEs/EDs' is not supported by the proposed-method simulations: Fig. 3 reaches N=200 with K=200, and Fig. 4 reaches K=120; the 300-user point appears only in the benchmark results of Fig. 2.","section":"Abstract and simulation scope"}],"recommendation":"major_revision","confidential_remarks":"The paper compares only against the authors' own prior conference work, which is labeled 'SotA' in the figures; an independent baseline or a clear statement that [1] is the only applicable reference would strengthen the novelty claim. The two load-bearing approximations—diagonal effective noise and GaBP-based solution of the noiseless linear system—need either formal justification or a substantial expansion of the numerical evidence before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a reasonable next step from the authors' ICNC work: instead of estimating individual computing signals, they estimate the AirComp aggregate directly, and they replace the O(N^3) combiner inversion with a second GaBP loop over the linear system b=Au. The multi-stream and multi-access extensions are straightforward but useful. The simulations show that the low-complexity receiver matches the direct-inversion benchmark in BER and NMSE up to 200 antennas and 300 users, which is credible evidence the approach works in the tested regimes.\n\nThe derivations of the MMSE combiners and the GaBP updates are consistent, and the appendix fills in the algebra. I also think the diagonal approximation of the effective noise covariance (eq. 30) is less fragile than it first looks: with sigma_s^2 = 1/K, the off-diagonal entries of sigma_s^2 H H^H are O(1/sqrt(K)) and the diagonal entries are O(1), so the approximation gets better exactly in the massive regime the paper targets. The authors could strengthen that point with a concentration argument.\n\nThe weak spot is the GaBP loop for the combiner. It solves a noiseless linear system, and there is no check of the convergence conditions from [48] on A, nor a derivation that the fixed point is A^{-1}b. The prior sigma_u^2=1 is ad hoc and the EM update averages N^2 extrinsic estimates without proof. These are real gaps, but they are gaps in proof, not contradictions, and the NMSE agreement with the benchmark in Figs. 3-4 suggests the loop does converge in the parameter ranges shown. The authors should report beta_u, provide code or at least convergence diagnostics, and ideally prove convergence for positive definite A or state conditions under which the loop is guaranteed to converge. The complexity claim also needs a small correction: forming A costs O(N^2 K), so the total complexity is not simply O(NK)+O(N^2).\n\nWho gets value: researchers in AirComp or message-passing receivers will find the framework useful as a design template. I would not build on it without the convergence question settled, and I'd want code to reproduce the combiner loop. But the paper is coherent, honest about its limitations, and the contribution is real. It deserves a serious referee; I'd recommend major revision with the convergence and reproducibility items addressed.","headline":"Solid engineering extension of GaBP to joint data/AirComp detection, but the matrix-inversion-free combiner loop lacks convergence analysis and should be validated with code and more scenarios before I'd trust it generally.","tokens_in":21863,"tokens_out":4388,"would_cite":false,"duration_ms":44516,"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 establishes a two-loop Gaussian belief propagation receiver that detects data symbols and computes AirComp functions, obtaining the combiner from a linear system solve rather than a matrix inversion.","keywords":["integrated communication and computing","over-the-air computation","Gaussian belief propagation","massive MIMO","nomographic function","multi-stream AirComp","multi-access receiver","low-complexity detect-and-combine"],"falsifier":"Take the $N=100$, $K=75$ fully loaded setup at a fixed $\\mathrm{SNR}_S=20$ dB and replace the i.i.d. Rayleigh channel by one with exponential correlation matrix whose off-diagonal coefficient $\\rho$ is swept from 0 to 0.9; if the NMSE of $\\hat{f}(\\mathbf{s})$ rises significantly with $\\rho$, the diagonal approximation in equation (30) is the limiting assumption.","tokens_in":20781,"feed_emoji":"📡","tokens_out":10135,"duration_ms":91676,"temperature":0.7,"pith_summary":"This paper proposes a receiver design in which one base station simultaneously detects independent data symbols and computes over-the-air functions of the form $f(\\mathbf{s})=\\phi(\\sum_k \\psi_k(s_k))$, with each user able to transmit data, computing signals, or both. The receiver estimates the data vector $\\mathbf{d}$ with a Gaussian belief propagation (GaBP) loop, then computes the AirComp target $\\hat{f}(\\mathbf{s})$ by applying a second GaBP loop to obtain the combining vector $\\mathbf{u}$ from the linear system $\\mathbf{b}=\\mathbf{A}\\mathbf{u}$, avoiding the cubic-complexity matrix inversion used by earlier combiners. The central claim is that, despite the reduced complexity $O(NK)+O(N^2)$, the method matches the benchmark's bit error rate and normalized mean square error in underloaded and fully loaded setups with up to 200 antennas and 300 users. A sympathetic reader would care because it offers a systematic, low-complexity way to integrate communication and computation at the physical layer, generalizing to arbitrary nomographic functions and multiple computation streams.","feed_headline":"Two belief-propagation loops unite data detection and AirComp","feed_subtitle":"The combiner comes from solving a linear system by message passing, matching benchmark BER and NMSE at lower complexity.","key_machinery":"The load-bearing object is the MMSE AirComp combiner $\\mathbf{u}=(\\mathbf{H}(\\sigma_s^2\\mathbf{I}_K+\\boldsymbol{\\Omega})\\mathbf{H}^{\\mathsf H}+\\sigma_w^2\\mathbf{I}_N)^{-1}\\mathbf{H}\\sigma_s^2\\mathbf{1}_K$. Instead of inverting the $N\\times N$ matrix, the paper recasts it as the linear system $\\mathbf{b}=\\mathbf{A}\\mathbf{u}$ and applies Gaussian belief propagation (GaBP), a message-passing scheme for inference in Gaussian models whose soft interference cancellation and belief generation operate element-wise on scalars. Gaussian denoisers for the combiner use prior variance $\\sigma_u^2\\approx 1$ with mean updated by EM, while the data-detection loop uses the same GaBP machinery under a diagonal approximation of the effective noise covariance. The two loops together replace one cubic matrix inversion with $O(NK)+O(N^2)$ scalar updates, and this is what allows the scheme to scale to 200 antennas and 300 users.","core_discovery":"The paper demonstrates that the receive combiner for over-the-air computing can be treated as a linear estimation problem inside the same message-passing framework as data detection. After GaBP yields an estimate of $\\mathbf{d}$, the receiver forms $\\mathbf{A}=\\mathbf{H}(\\sigma_s^2\\mathbf{I}_K+\\boldsymbol{\\Omega})\\mathbf{H}^{\\mathsf H}+\\sigma_w^2\\mathbf{I}_N$ and $\\mathbf{b}=\\mathbf{H}\\sigma_s^2\\mathbf{1}_K$, and obtains the MMSE combining vector by running GaBP on $\\mathbf{b}=\\mathbf{A}\\mathbf{u}$ with a Gaussian prior on $\\mathbf{u}$ and an EM update for its mean. Because the effective noise covariance $R_{\\tilde{\\mathbf w}}=\\sigma_s^2\\mathbf{H}\\mathbf{H}^{\\mathsf H}+\\sigma_w^2\\mathbf{I}_N$ is approximated as diagonal, $\\sigma_s^2\\mathrm{diag}(\\xi)+\\sigma_w^2\\mathbf{I}_N$, all message updates reduce to element-wise scalar operations. The paper shows by simulation that this two-loop GaBP receiver reproduces the BER and NMSE of the benchmark joint estimator in underloaded and fully loaded settings, and extends to multiple computation streams through sparse index vectors $\\mathbf{p}_m$ and to multi-access operation by zeroing the data estimates of computing-only users.","pith_inferences":["Because the diagonal effective-noise approximation is what keeps the updates scalar, a natural test is to sweep the spatial correlation of the channel: if NMSE rises sharply as $\\mathbf{H}\\mathbf{H}^{\\mathsf H}$ gains off-diagonal weight, that confirms the approximation as the bottleneck.","The combining loop solves a Hermitian positive-definite linear system, so preconditioning or low-rank approximations of $\\mathbf{A}$ could plausibly push the cost below $O(N^2)$ without changing the message-passing structure.","The paper states that correlated computation streams would require dedicated conditional-probability message passing; a decorrelating preprocessing step might preserve scalar updates and extend the framework to correlated streams at modest cost.","The claim that overloaded AirComp is combiner-limited could be checked by replacing the MMSE combiner with a regularized variant and measuring whether the NMSE floor in the $K>N$ regime drops."],"forward_implications":["Data detection and over-the-air computing can share one receiver with no cubic matrix inversion: the communication loop costs $O(NK)$ and the combining loop costs $O(N^2)$.","The combiner design is agnostic to the target function, so any nomographic function supported by suitable pre- and post-processing can be computed with the same receiver structure.","Multiple target functions can be computed from the same set of computing symbols by defining sparse index vectors $\\mathbf{p}_m$ and running the combining loop once per stream.","Multi-access operation requires only zeroing data estimates for computing-only users, so communication-only, computing-only, and hybrid users coexist with minimal structural change.","In overloaded regimes, AirComp NMSE degrades, and the paper attributes the degradation to the combiner rather than to data detection."],"supporting_citations":[{"why":"Supplies the AirComp system model, nomographic function formulation, and transmit power allocation scaling used throughout.","marker":"[31]"},{"why":"Supplies the low-complexity AirComp combiner design that the proposed GaBP-based combiner replaces.","marker":"[36]"},{"why":"The preliminary GaBP joint data and computing detector used as the performance benchmark.","marker":"[1]"},{"why":"Defines Gaussian belief propagation theory underlying the scalar message-passing rules.","marker":"[43]"},{"why":"Supplies the adaptively scaled belief denoiser used for the QPSK data symbols.","marker":"[45]"},{"why":"Supplies the Gaussian-prior denoiser and bilinear GaBP update structure used for the combining vector.","marker":"[47]"},{"why":"Lists nomographic function examples and pre-/post-processing forms that the framework supports.","marker":"[46]"}],"fun_headline_variants":["GaBP receiver fuses data and AirComp in two loops","Message passing unifies data detection and over-the-air computing","Two-loop GaBP handles data and AirComp jointly","Receiver framework merges detection and computing via GaBP","Scalable GaBP design for simultaneous data and AirComp"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the interference plus noise seen by the data detector is well approximated by independent per-antenna noise, which the paper justifies by channel hardening and small $\\sigma_s^2$; when $\\mathbf{H}\\mathbf{H}^{\\mathsf H}$ has strong off-diagonal entries, the soft interference-cancellation variances are miscalibrated and both data detection and the computed function degrade.","fun_headline_variants_meta":{"raw":{"variants":["GaBP receiver fuses data and AirComp in two loops","Message passing unifies data detection and over-the-air computing","Two-loop GaBP handles data and AirComp jointly","Receiver framework merges detection and computing via GaBP","Scalable GaBP design for simultaneous data and AirComp"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000608,"raw_usage":{"total_tokens":2912,"prompt_tokens":1104,"completion_tokens":1808,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":720,"completion_tokens_details":{"reasoning_tokens":1728}},"tokens_in":720,"tokens_out":1808,"duration_ms":11799,"temperature":1.0,"reasoning_tokens":1728,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:13:22.603398+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the $N=100$, $K=75$ fully loaded setup at a fixed $\\mathrm{SNR}_S=20$ dB and replace the i.i.d. Rayleigh channel by one with exponential correlation matrix whose off-diagonal coefficient $\\rho$ is swept from 0 to 0.9; if the NMSE of $\\hat{f}(\\mathbf{s})$ rises significantly with $\\rho$, the diagonal approximation in equation (30) is the limiting assumption.","supporting_citations":[{"cited_title":"Over-the-Air Computation Systems: Optimization, Analysis and Scaling Laws,","cited_arxiv_id":null,"evidence_quote":"Supplies the AirComp system model, nomographic function formulation, and transmit power allocation scaling used throughout."},{"cited_title":"Low-Complexity and High- Performance Combiners for Over the Air Computing,","cited_arxiv_id":null,"evidence_quote":"Supplies the low-complexity AirComp combiner design that the proposed GaBP-based combiner replaces."},{"cited_title":"From theory to reality: A design framework for integrated communication and computing receivers,","cited_arxiv_id":null,"evidence_quote":"The preliminary GaBP joint data and computing detector used as the performance benchmark."},{"cited_title":"Gaussian Belief Propagation: Theory and Aplication,","cited_arxiv_id":null,"evidence_quote":"Defines Gaussian belief propagation theory underlying the scalar message-passing rules."},{"cited_title":"Design of Adaptively Scaled Belief in Multi-Dimensional Signal Detection for Higher-Order Modulation,","cited_arxiv_id":null,"evidence_quote":"Supplies the adaptively scaled belief denoiser used for the QPSK data symbols."},{"cited_title":"Bilinear gaussian belief propagation for massive MIMO detection with non-orthogonal pilots,","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian-prior denoiser and bilinear GaBP update structure used for the combining vector."}],"review_version":1}