{"id":"b7085fe5-f2ba-4168-9e81-4df6914aa4bd","arxiv_id":"2506.18863","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A variational Bayesian joint estimation and detection algorithm works with low-resolution quantized fronthaul and outperforms LMMSE and prior nonlinear methods in simulated cell-free massive MIMO.","lead":"This paper proposes a variational Bayesian method for joint channel estimation and data detection in cell-free massive MIMO with low-bit quantized fronthaul links. It reports roughly 2 dB SER gains over LMMSE with perfect fronthaul, and shows the quantize-then-estimate design beats estimate-then-quantize.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (19) adds cross-AP residual terms that do not follow from the factorized likelihood in Eq. (16), so the implemented channel update is not the claimed variational Bayes solution.","rationale":"The reader's rejection rationale correctly identifies the inconsistency in Eq. (19), and I agree that this is the most load-bearing issue: the update does not follow from the stated mean-field factorization and likelihood, so the paper's central algorithmic claim is unsupported as written. I mark agreement as partial because the reader's formal 'weakest_assumption' field names the known-covariance assumption rather than the Eq. (19) derivation error; the covariance issue is a legitimate secondary concern, but the algebraic discrepancy is more decisive. The proposed concrete test isolates the error in a minimal setting and then checks whether the numerical headline survives correction. No additional objection is needed: the paper's own conclusion admits that learning Sigma within VB is future work, and the Gamma hyperparameter tuning in Fig. 7 is scenario-dependent, but these would not change the verdict once the update derivation is invalid. The verdict should remain REJECT until the channel updates are re-derived and the simulations are rerun with a corrected update.","tokens_in":24526,"tokens_out":6263,"duration_ms":73916,"concrete_test":"Independently re-derive q(h_i,l) for the PFL model from Eq. (16) with L=2, M=1, K=1, T_p=1, fixed x_1=1, and no data phase. The resulting Gaussian mean should depend only on r_1,1, the observation at the AP whose channel is being updated. Substitute the same setting into Eq. (19) and check whether the r_2,1 term appears; if it does, the published update is not the CA-VI update. Then, if the simulation code is available, rerun Fig. 2 with Eq. (19) replaced by the correctly derived local-block update and compare the SER at 10^{-3}; a material change in the VB(Q-E, 3-bit) curve would show that the headline gain depends on the erroneous cross-AP term.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing defect is in the derivation of the channel update. In the PFL model, p(rt|H,gamma_p;xt) = CN(rt; Hxt, gamma_p^{-1} I_ML) factorizes over AP blocks as prod_l CN(r_l,t; H_l xt, gamma_p^{-1} I_M). When forming q(h_i,l) from Eq. (16), h_i,l appears only in the l-th block, because H_l is the only block containing h_i,l. The expected log-likelihood can therefore produce the local residual r_l,t - sum_{j != i} <h_j,l> x_j,t, but never the cross-AP residual r_l',t - <H_l'> xt for l' != l. Yet Eq. (17) and Eq. (19) include exactly those cross-AP terms, for both the pilot and data phases, and Algorithm 1 step 14 feeds these updates into the PFL and Q-E algorithms. The same cross-AP terms reappear in the E-Q data-phase update in Eq. (47). Unless a different, unstated model is intended in which h_i,l affects observations at every AP, these terms are not part of the CA-VI fixed point; their inclusion turns the method into a heuristic that is not the variational Bayesian solution described in the paper. Since the central claim is that this VB estimator outperforms LMMSE(PFL), the numerical gains in Figs. 2-10 do not support the stated derivation. The known-covariance caveat is real but secondary; the Eq. (19) inconsistency breaks the derivation even when Sigma is known perfectly.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies uplink joint channel estimation and data detection (JED) in a cell-free massive MIMO network in which access points (APs) forward low-resolution quantized signals to a central processing unit (CPU). It proposes a variational Bayesian (VB) inference framework with a mean-field factorization, deriving coordinate-ascent variational inference (CAVI) updates for the channel coefficients, data symbols, and precision parameters under three scenarios: perfect fronthaul links (PFL), quantization-and-estimation (Q-E), and estimation-and-quantization (E-Q). The paper reports that the proposed VB methods outperform LMMSE with perfect fronthaul, with about 2 dB SER gain at 1e-3 for 3-bit quantization, and also compares against GAMP-based and VB-EM-based nonlinear benchmarks. Complexity and fronthaul signaling overhead are analyzed.","tokens_in":24843,"tokens_out":8169,"duration_ms":80769,"significance":"If the derived updates were the true CAVI fixed points and the numerical results were reproducible, the paper would make a useful contribution by demonstrating that nonlinear VB-based JED can overcome low-resolution fronthaul limitations and outperform linear processing with perfect fronthaul, while providing a unified treatment of PFL, Q-E, and E-Q. The manuscript includes extensive simulations, comparisons with state-of-the-art nonlinear methods, and complexity/overhead analysis. However, the central derivation contains an error that invalidates the claim that the implemented algorithm is the variational Bayesian solution of the stated probabilistic model; the reported numerical gains therefore cannot be attributed to the method as described.","major_comments":[{"comment":"The update for q(h_{i,ℓ}) in Eq. (17) and the mean update in Eq. (19) include cross-AP residual terms Σ_{ℓ'≠ℓ}(r_{ℓ',t} − ⟨H_{ℓ'}⟩x_t) that do not follow from the factorized likelihood. Since p(r_t|H, γ_p; x_t) = CN(r_t; Hx_t, γ_p^{-1} I_{ML}) factorizes over AP blocks, the channel h_{i,ℓ} appears only in the ℓ-th block, and the expectation in Eq. (6) can only produce the local residual r_{ℓ,t} − Σ_{j≠i}⟨h_{j,ℓ}⟩x_{j,t}. The cross-AP residuals are constant with respect to h_{i,ℓ} and cancel in the normalization of q(h_{i,ℓ}). Their inclusion in Eq. (19), in the data-phase terms of Eq. (17), in Eq. (47) for the E-Q scenario, and in Algorithm 1 step 14 means that the implemented update is not the CAVI fixed point of the stated model. The paper provides no alternative model or justification for these terms. This is load-bearing because the claimed SER gains of VB(Q-E) and VB(E-Q) over LMMSE(PFL) are obtained with this heuristic update, not with the variational Bayesian solution described in Sections II-C and III-A.","section":"III-A1, Eq. (17)-(19)"},{"comment":"The comparison against VB-EM in Fig. 7 tunes the Gamma hyperparameters a_p, b_p, a_{d,t}, b_{d,t} via bisection over the same SNR range on which performance is reported, whereas the VB-EM baseline uses the limiting case a_p = b_p = a_{d,t} = b_{d,t} = 0. This gives the proposed method an additional fitted degree of freedom and makes the claimed improvement over VB-EM not an apples-to-apples comparison. The paper should either fix the hyperparameters a priori, report sensitivity to their choice, or clearly state that the reported curves use hyperparameters tuned to the evaluation setting.","section":"IV-A, Fig. 7"}],"minor_comments":[{"comment":"The complexity of VB(PFL) and VB(Q-E) is stated as O(Itr [M^3 L^3 K + ...]); since the matrix inversion in Eq. (18) is performed per (i,ℓ) pair on an M×M matrix, the leading term should be O(K L M^3), which overstates the cost by a factor of L^2.","section":"IV-C"},{"comment":"The residual terms ζ_p and ζ_d are defined in Eqs. (10)-(13) using estimated values Ĥ and X̂_d that are not yet available when the probabilistic model is specified; the subsequent Gaussian assumption is a modeling choice that should be clearly stated as part of the variational approximation rather than presented as a consequence of the definitions.","section":"II-C"},{"comment":"The hard-decision step 'Compute ˆx_{i,t} = arg max_{a∈S} q_i(a)' is written outside the iterative loop but does not specify whether the variational distribution from the final iteration is used; this should be stated explicitly, and the same applies to the final channel estimate in each algorithm.","section":"Algorithms 1 and 2"}],"recommendation":"reject","confidential_remarks":"The derivation error in Eq. (19) is fundamental: the cross-AP residual terms cannot arise from the factorized likelihood, so the implemented algorithm is not the variational Bayesian solution claimed in the paper. Correcting the update would change the algorithm and likely the simulation results, so the current numerical evidence cannot support the paper's central claims. I recommend rejection, although a corrected derivation with re-run experiments could form the basis of a future submission. The acknowledged overlap with [35] is not by itself a concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a plausible extension of variational Bayesian JED to cell-free massive MIMO with quantized fronthaul, and the E-Q variant is genuinely new. But the central derivation has a load-bearing problem: Eq. (19) includes cross-AP residual terms that do not follow from the factorized likelihood. As written, the algorithm is not the variational Bayes solution the paper claims.\n\nIn the PFL model, p(rt|H,γp;xt) = ∏_ℓ CN(r_ℓ,t; H_ℓ x_t, γp^{-1} I_M). The channel h_i,ℓ appears only in the ℓ-th AP block, so the mean-field update for q(h_i,ℓ) can only involve the local residual r_ℓ,t − Σ_{j≠i}⟨h_j,ℓ⟩x_j,t. Equation (17) and (19) add Σ_{ℓ'≠ℓ}(r_ℓ',t − ⟨H_ℓ'⟩x_t), which is not a function of h_i,ℓ under the stated model and cannot appear in a CAVI step. The same spurious term reappears in Algorithm 1 step 14 and in the E-Q data-phase update, Eq. (47). This is not a minor typo: it changes the fixed point and invalidates the convergence guarantee. Even with perfect Σ, the derivation fails.\n\nWhat is good: the E-Q scenario (local AP channel estimation followed by quantized forwarding to the CPU) is a new combination; the Gamma-prior on precisions cleanly contains VB-EM as a special case; the benchmark set is appropriate, including GAMP and VB-EM under correlated channels; and the complexity/overhead comparison is useful. The numerical SER/NMSE results are consistent with the claims, but they are evaluating the heuristic version, not the derived VB estimator.\n\nSecondary issues: the Gamma hyperparameters in Fig. 7 are tuned on the same evaluation scenario by bisection, there are no error bars, and no code is provided. So I cannot tell whether the simulations implement Eq. (19) literally or a corrected local-residual update. That matters for interpreting every figure.\n\nRecommendation: I would not accept this in current form. The repair is conceptually straightforward—delete the cross-AP terms and re-derive the CAVI updates, or explicitly present the method as a heuristic with cross-AP information exchange and drop the claim that it is the VB solution. Either way the simulations need to be repeated or justified under the corrected algorithm. The E-Q contribution is worth preserving, so I would send it to a serious referee rather than desk reject, but the letter should make clear that the derivation must be fixed before the results can be trusted.","headline":"Useful extension of VB-JED to quantized cell-free MIMO with a genuinely new E-Q variant, but Eq. (19) adds cross-AP residual terms that do not follow from the factorized likelihood, so the simulated algorithm is not the claimed variational Bayes solution.","tokens_in":25451,"tokens_out":8825,"would_cite":false,"duration_ms":96419,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A variational Bayesian receiver lets 3-bit fronthaul links beat an unquantized linear baseline in cell-free massive MIMO.","keywords":["cell-free massive MIMO","variational Bayesian inference","joint channel estimation and data detection","low-resolution quantization","fronthaul links","quantization-and-estimation","estimation-and-quantization","symbol error rate"],"falsifier":"Reproduce the paper's main simulation (L=8, M=4, K=16, Tp=32, Td=128, QPSK) and check whether VB(Q-E, 3 bits) really holds about a 2 dB SER advantage over LMMSE(PFL) at 1e-3; if the gap vanishes or reverses, the central claim fails. A second probe targets the weakest assumption: rerun with mismatched or estimated covariance matrices and see whether the advantage disappears.","tokens_in":24317,"feed_emoji":"📡","tokens_out":9476,"duration_ms":88901,"temperature":0.7,"pith_summary":"This paper sets out to show that the fronthaul bottleneck in cell-free massive MIMO does not force a system onto linear receivers. It proposes a variational Bayesian receiver that performs joint channel estimation and data detection at the central processor from low-resolution quantized signals sent by the access points, under two fronthaul strategies: quantization-and-estimation and estimation-and-quantization. The central claim is that both VB(Q-E) and VB(E-Q) outperform the linear LMMSE receiver even when that receiver has perfect, unquantized fronthaul, with about a 2 dB symbol-error-rate gain at 1e-3 for 3-bit quantization. If true, cell-free networks with limited fronthaul bandwidth can keep the benefits of centralized nonlinear processing rather than accepting a linear-processing penalty.","feed_headline":"Quantized 3-bit link beats unquantized linear receiver","feed_subtitle":"Variational Bayesian processing turns 3-bit AP signals into a 2 dB SER gain over LMMSE with perfect fronthaul.","key_machinery":"The load-bearing object is the mean-field variational family $q(R_p,R_d,X_d,H,\\gamma_p,\\gamma_d)$ and the coordinate-ascent updates that cycle through its factors. The critical update is the one for $r_t$: conditioned on the quantized bin, the variational posterior of the unquantized signal is a truncated complex Gaussian whose mean and variance are computed from Gaussian pdf/cdf functions. This step lets the method treat quantization as an exact likelihood rather than a Bussgang linearization. The second piece is the Gamma prior on the precision parameters $\\gamma_p$ and $\\gamma_{d,t}$; their VB updates use Lemma 1 to evaluate expectations of squared residuals, and setting the Gamma parameters to zero recovers the EM-based precision estimate of [35].","core_discovery":"The paper's central discovery is that the nonlinear quantization operation can be absorbed into a fully variational Bayesian joint estimation and detection loop without linearizing the quantizer. The method approximates the intractable posterior over channel, data symbols, and precision parameters by a mean-field distribution, then cycles through coordinate-ascent updates. At each access point the received signal is quantized to a few bits; at the CPU, each unquantized receive sample is given a truncated Gaussian variational distribution whose support is the quantization bin, so the quantizer enters as an exact likelihood rather than an approximation. The precision of the residual inter-user interference, which contains noise plus estimation and detection errors, is treated as a Gamma random variable and updated inside the loop, and the authors show this contains the earlier VB-EM scheme as a special case. In simulations, VB(Q-E, 3 bits) and VB(E-Q, 3 bits) each gain about 2 dB over LMMSE with perfect fronthaul at symbol error rate 1e-3, VB with perfect fronthaul gains about 4 dB, and VB(Q-E) slightly beats VB(E-Q), which the authors attribute to local channel-estimation errors in the E-Q approach. Under correlated channels, the proposed VB methods converge while GAMP-based benchmarks can diverge.","pith_inferences":["If the channel covariance matrices must be estimated rather than given, the 2 dB advantage could shrink or vanish; a natural test is to feed the VB receiver estimated or mismatched covariances and measure the SER gap to LMMSE(PFL).","The truncated-Gaussian quantization update is a modular ingredient: the same treatment could be dropped into other Bayesian receivers, such as expectation propagation or bilinear inference, to replace Bussgang linearization, and that substitution can be tested directly.","The reported saturation of SER at high SNR for 1- and 2-bit quantizers implies a design rule the paper does not state: choose quantizer resolution according to the operating SNR, since quantization noise dominates above a few dB.","The E-Q comparison assumes the local channel-estimation error variance is known; in deployment that variance would need tracking, and a mismatch could change the ordering between Q-E and E-Q."],"forward_implications":["Low-resolution fronthaul no longer forces a linear-processing penalty: a 3-bit quantized link can outperform an unquantized link running LMMSE, which changes the cost–bandwidth trade-off in cell-free network design.","Because VB(E-Q) needs fewer fronthaul bits than VB(Q-E) yet suffers only a small SER loss, it is the preferred scheme when both fronthaul bandwidth and computational complexity are tight.","The fully VB precision updates make the method stable under spatially correlated channels, where AMP/GAMP-type receivers may diverge, so the receiver is usable in realistic correlated propagation.","Joint estimation and detection benefits compound with data-block length: longer data phases improve SER for the VB methods while LMMSE plateaus, so iterative refinement is most valuable in long coherence blocks."],"supporting_citations":[{"why":"Supplies the VB-EM baseline, Lemma 1 for expectations of quadratic forms, and the truncated-Gaussian mean/variance formulas reused in the proposed updates.","marker":"[35]"},{"why":"Supplies the GAMP(Q-E) benchmark and the quantized-Gaussian moment computation used for the E-Q channel update.","marker":"[34]"},{"why":"Defines the Q-E and E-Q fronthaul scenarios and the earlier finding that E-Q performs slightly worse than Q-E, which the paper reproduces.","marker":"[30]"},{"why":"Provides the Bussgang-based bilinear-inference JED receiver for cell-free massive MIMO with low-resolution ADCs against which the paper positions itself.","marker":"[32]"},{"why":"Supplies the variational Bayesian perspective on massive MIMO detection, including the LMMSE complexity baseline and the motivation for VB over AMP.","marker":"[14]"},{"why":"Supplies the coordinate-ascent variational inference convergence guarantee to a local optimum used to justify the algorithm's stability.","marker":"[25]"}],"fun_headline_variants":["VB with 3-bit fronthaul beats perfect-fronthaul LMMSE","3-bit VB wins over unquantized LMMSE","Variational Bayes tames 3-bit fronthaul","Low-bit fronthaul? VB still wins by 2 dB","3-bit VB outperforms full-precision LMMSE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the statistics of every user–AP channel (the covariance matrices) are known exactly in advance; if they are wrong or must be learned, the reported gains may shrink.","fun_headline_variants_meta":{"raw":{"variants":["VB with 3-bit fronthaul beats perfect-fronthaul LMMSE","3-bit VB wins over unquantized LMMSE","Variational Bayes tames 3-bit fronthaul","Low-bit fronthaul? VB still wins by 2 dB","3-bit VB outperforms full-precision LMMSE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000903,"raw_usage":{"total_tokens":4000,"prompt_tokens":1174,"completion_tokens":2826,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":790,"completion_tokens_details":{"reasoning_tokens":2737}},"tokens_in":790,"tokens_out":2826,"duration_ms":18434,"temperature":1.0,"reasoning_tokens":2737,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:42:13.586440+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the paper's main simulation (L=8, M=4, K=16, Tp=32, Td=128, QPSK) and check whether VB(Q-E, 3 bits) really holds about a 2 dB SER advantage over LMMSE(PFL) at 1e-3; if the gap vanishes or reverses, the central claim fails. A second probe targets the weakest assumption: rerun with mismatched or estimated covariance matrices and see whether the advantage disappears.","supporting_citations":[{"cited_title":"V a riational Bayes for joint channel estimation and data detection in few -bit massive MIMO systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the VB-EM baseline, Lemma 1 for expectations of quadratic forms, and the truncated-Gaussian mean/variance formulas reused in the proposed updates."},{"cited_title":"B ayes-optimal joint channel-and-data estimation for massive MIMO with lo w-precision ADCs,","cited_arxiv_id":null,"evidence_quote":"Supplies the GAMP(Q-E) benchmark and the quantized-Gaussian moment computation used for the E-Q channel update."},{"cited_title":"Uplink spectral and energy efficiency of c ell- free massive MIMO with optimal uniform quantization,","cited_arxiv_id":null,"evidence_quote":"Defines the Q-E and E-Q fronthaul scenarios and the earlier finding that E-Q performs slightly worse than Q-E, which the paper reproduces."},{"cited_title":"Bayesian receiver design via bilinear inference for cell- free massive MIMO with low-resolution ADCs,","cited_arxiv_id":null,"evidence_quote":"Provides the Bussgang-based bilinear-inference JED receiver for cell-free massive MIMO with low-resolution ADCs against which the paper positions itself."},{"cited_title":"A Variational Bayesian Perspective on Massive MIMO Detection","cited_arxiv_id":"2205.11649","evidence_quote":"Supplies the variational Bayesian perspective on massive MIMO detection, including the LMMSE complexity baseline and the motivation for VB over AMP."}],"review_version":2}