{"id":"4344d760-91d4-46e1-900a-46e7c27d9b71","arxiv_id":"2501.11229","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A SIC-aided diffusion model for joint channel estimation and data detection is claimed to outperform existing baselines in low-rank (overloaded) MIMO channels.","lead":"This paper proposes a wireless receiver that combines a diffusion model with successive interference cancellation to jointly estimate the channel and detect users' data when there are more users than antennas. If it works, it could make overloaded multi-user MIMO systems, where prior methods struggle, more reliable.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 3's MAP objective and gradient updates (Eqs. 2-5) are internally inconsistent: Eq. (4) is not the derivative of Eq. (3), so the algorithm whose simulations are reported is not uniquely or correctly specified.","rationale":"I read the paper as making an empirical claim, so the only way to evaluate it from the manuscript is through the algorithm in Section 3. The reader's weakest assumption identifies exactly the MAP objective and gradient updates. My independent check of the calculus supports that concern: Eq. (4) sums j from i to k and uses residual Y - H(j)X(j), whereas the derivative of Eq. (3) would involve Y - sum_{l<j}H(l)X(l) and only j>i. Thus the published equations do not describe a valid MAP ascent for the stated factorization. The covariance definition is also inconsistent with the stage residual. Because the paper gives no code, no pseudocode, and no training details, the simulations cannot be attributed to a uniquely defined method. This is an internal mathematical inconsistency, not a disagreement with consensus, and it directly undermines the central claim as written. I credit the simulation figures as reported evidence, but without a coherent algorithm specification they do not settle the claim. The reader's conditional verdict is reasonable; I would sharpen it to 'unverified' until Eqs. (4)-(5) are corrected or the implementation is made available. If the corrected gradients reproduce the reported gains, the central claim would be supported and acceptance would be appropriate.","tokens_in":6118,"tokens_out":9052,"duration_ms":89080,"concrete_test":"Independently differentiate Eq. (3): for fixed j, compute the gradient of log N(sum_{l=1}^{j-1} H(l)X(l), Sigma_j) with respect to H(i), and compare with the RHS of Eq. (4). The expressions match only if the residual is Y - sum_{l<j}H(l)X(l) and j>i; Eq. (4) uses Y - H(j)X(j) and starts at j=i. If they disagree, obtain the authors' code or a corrected derivation and rerun Figure 1b (Nu=64, Nr=32, P=30, D=50). The central claim stands only if the corrected algorithm reproduces the reported NMSE/SER gains.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim is empirical: the proposed SIC-aided diffusion method outperforms LMMSE, L-DAMP, and Joint Langevin. The load-bearing component is the algorithm definition in Section 3. Equations (2)-(3) define a MAP objective in which each stage's Gaussian likelihood has mean equal to the cumulative reconstructed signal from previously decoded users, sum_{l<i} H(l)X(l), and covariance Sigma_i that treats later users as noise. If this is the objective, then differentiating with respect to H(i) gives contributions only from stages j>i, with residual Y - sum_{l<j} H(l)X(l). Equation (4), however, sums j from i to k and uses residual Y - H(j)X(j). That is the gradient of a different objective, sum_j log N(H(j)X(j), Sigma_j), not of Eq. (3). Equation (5) has the same structure. Also, Sigma_i is defined with later users as noise but omits the current user's signal, which is inconsistent with the stage-i residual under Eq. (3)'s mean. No pseudocode, network architecture, noise schedule, or code is provided, so the experiments cannot be mapped to a unique implementable algorithm. The reported curves may be correct for some implementation, but the manuscript does not define that implementation; the derivation gap is internal inconsistency, not a disagreement with external consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a successive interference cancellation (SIC)-aided score-based diffusion algorithm for joint channel estimation and data detection in uplink MIMO systems where the number of users exceeds the number of access-point antennas. The channel matrix is split into full-rank submatrices, a score network is trained on the partial-channel priors, and the algorithm alternates between estimating channel submatrices, detecting symbol blocks, and updating a channel-gain-based SIC decoding order. Simulations on QuaDRiGa 3GPP channels compare normalized mean squared error and symbol error rate against LMMSE, L-DAMP, and Joint Langevin baselines in both full-rank and low-rank scenarios, reporting improvements across SNR.","tokens_in":6473,"tokens_out":7292,"duration_ms":69868,"significance":"If the reported results are correct, the paper addresses a genuine and understudied problem: standard massive MIMO receivers degrade when the number of users exceeds the number of antennas, and the proposed decomposition into approximately full-rank channel submatrices is a plausible way to adapt diffusion-based priors to this regime. The experimental setup uses a realistic 3GPP channel model and compares against several relevant baselines. However, the central algorithm is not well defined in the manuscript: the stated MAP objective is not a valid factorization of the joint posterior, the gradient updates do not follow from that objective, and the implementation details needed to reproduce the experiments are largely absent. The potential significance is therefore clear, but the current text does not establish a correct or reproducible method.","major_comments":[{"comment":"The MAP objective in Eqs. (2)-(3) is not a correct factorization of the joint posterior p(XD,H|Y,XP). For an SIC decoder, after decoding stages 1 through i-1, the residual is R_i = Y - sum_{l<i} H(l)X(l), and the stage-i likelihood should be a density for this residual, such as N(H(i)X(i), Sigma_i) with later users treated as noise. Equation (3) instead places the cumulative previous-signal sum H(i-1)X(i-1)+...+H(1)X(1) in the mean of a Gaussian for Y, which is inconsistent with the observation model Y=HX+Z and with the stated SIC residual. As a result, the objective being maximized is not the MAP objective for the system in Eq. (1).","section":"Section 3, Eqs. (2)-(3)"},{"comment":"Equations (4)-(5) are not the derivatives of the objective in Eq. (3). Differentiating the first term of Eq. (3) with respect to H(i) yields contributions only from stages j>i, since H(i) does not appear in the stage-i mean, and the residual that appears should be Y - sum_{l<j} H(l)X(l), not Y - H(j)X(j). The sums in Eqs. (4)-(5) start at j=i and use Y - H(j)X(j), so they solve a different optimization problem. Consequently, the iterative update rule that is claimed to maximize Eq. (3) is not correctly specified, and the algorithm that generated the simulation results is not uniquely defined by the manuscript.","section":"Section 3, Eqs. (4)-(5)"},{"comment":"The manuscript does not contain a complete algorithm specification. There is no pseudocode, no score-network architecture or training loss, no noise schedule, no Langevin step size or iteration count, and no convergence criterion for the SIC loop. Section 4 also does not state whether the 10,000 QuaDRiGa channel instances were split into training and test sets. Because the central claim is an empirical performance comparison, these omissions prevent reproduction and make it impossible to map the curves in Fig. 1 to a unique implementation.","section":"Sections 3 and 4"}],"minor_comments":[{"comment":"The covariance Sigma_i is written as \"Sigma_i = Z + PU_{i+1} H(i)X(i)\", which is not a valid covariance expression and P U is never defined. It should presumably be something like sigma_0^2 I + sum_{l>i} H(l)X(l)X(l)^H H(l)^H if later users are treated as noise.","section":"Section 3, Eq. (3)"},{"comment":"The notation for the SIC ordering is confusing: pi(i) is called the decoding order for user i, while pi^{-1}(i) is called the user index decoded at the ith stage, and later H(i) is defined using pi^{-1}(iNr:...). Please define the ordering and the grouping consistently.","section":"Section 3"},{"comment":"The first paragraph lists baselines [18, 13, 19], but the experimental description also cites [20]; please reconcile the reference list and clearly identify which baseline is used in each experiment.","section":"Section 4"},{"comment":"The figure shows no markers, error bars, or confidence intervals on the NMSE and SER curves. Please specify the number of Monte Carlo runs and report the statistical spread of the results.","section":"Figure 1"},{"comment":"The sentence \"we observe that as XD increases, the SER efficacy increases\" should be rephrased, since the SER should decrease as the number of data symbols grows under the stated comparison; please clarify the intended claim.","section":"Section 4"},{"comment":"The abstract claims the method \"outperforms the baseline methods\" generally, but only three specific baselines are evaluated; please qualify the claim as applying to the evaluated baselines.","section":"Abstract and Section 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is not ready for publication in its present form. The derivation in Section 3 is internally inconsistent, and the experimental section omits essential implementation details such as pseudocode, network architecture, hyperparameters, and train/test splitting. If the authors can provide a corrected derivation, a complete algorithm specification, and properly reported experiments, a revised version could merit reconsideration. As it stands, I do not recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the paper has a genuinely new angle—splitting the channel into Nr-by-Nr submatrices and running a diffusion prior on each with a dynamic SIC order—and the reported gains over LMMSE, L-DAMP, and Joint Langevin are consistent across their plots. But the central derivation doesn't hold together. Equations (2)–(3) are not a valid factorization of the joint posterior; Y appears in every stage term, so L is a product of likelihoods rather than a posterior. More importantly, the gradient updates in Eqs. (4)–(5) are not the gradients of Eq. (3). The stated objective puts H(i) only in the covariance (as noise) for the i-th term, so its gradient w.r.t. H(i) would not produce a residual Y − H(i)X(i). The summations over j ≥ i suggest the authors actually optimized a different objective, one where each stage's contribution is the mean of the observation with later users treated as noise. That's a plausible SIC objective, but it's not what is written. As a result, the manuscript does not uniquely define the algorithm whose simulations are reported.\n\nCredit where due: the problem is well-motivated, the submatrix splitting to restore full-rankness for the diffusion prior is clever, and the dynamic channel-gain-based decoding order is a sensible addition. The simulation setup (QuaDRiGa, 3GPP channels, Nu=64, Nr=32) is standard, and the baselines are appropriate. The performance curves are consistent and show the method working in both full- and low-rank settings.\n\nThe soft spots beyond the derivation: implementation details are sparse—no score network architecture, no Langevin step sizes, no annealing schedule, no SIC convergence criteria, no code. That makes replication impossible. The plots have no error bars, so we can't tell whether the gains are significant. Some references are arXiv preprints rather than peer-reviewed venues, but the key baseline [13] is an IEEE conference paper.\n\nFor whom: researchers working on diffusion-based receivers for overloaded MIMO will want to understand the idea, but they can't trust the equations as written. A serious referee could ask for a corrected derivation and code/implementation details. If the corrected derivation matches the simulations, the paper would be a solid contribution. As is, it's not ready.\n\nMy recommendation: send it to peer review, but make clear that the derivation must be fixed and the implementation fully specified. If the authors cannot reconcile Eqs. (2)–(5), it should be rejected.","headline":"Promising SIC-plus-diffusion idea for overloaded MIMO that is undermined by an internally inconsistent MAP derivation in Section 3.","tokens_in":6962,"tokens_out":3541,"would_cite":false,"duration_ms":33087,"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 SIC-aided diffusion receiver jointly estimates channels and detects data in low-rank MIMO systems where users outnumber antennas.","keywords":["successive interference cancellation","score-based diffusion models","joint channel estimation","data detection","low-rank MIMO","massive MIMO","Langevin dynamics","MAP estimation"],"falsifier":"Re-derive the gradient of the MAP objective in Eq. (2) with respect to a fixed $H^{(i)}$ for a two-stage system and compare term by term with Eq. (4); if any stage whose $H^{(j)}$ and $X^{(j)}$ do not appear in that objective's likelihood factor still contributes to the gradient sum, then the algorithm is maximizing a different objective than the one stated.","tokens_in":5950,"feed_emoji":"📶","tokens_out":8625,"duration_ms":74277,"temperature":0.7,"pith_summary":"The paper aims to establish that joint channel estimation and data detection in MIMO uplinks remains feasible when the number of users meets or exceeds the number of receive antennas, a low-rank regime where prior diffusion-based receivers fail. Its proposed receiver splits the channel into near-square submatrices, learns the prior of each submatrix with a score-based diffusion (Langevin) model, and cancels interference between submatrices through successive interference cancellation with a decoding order updated from current channel gains. Simulations with a 3GPP channel model and a 64-user, 32-antenna configuration show lower normalized mean squared error and symbol error rate than learned D-AMP, joint Langevin diffusion, and linear MMSE baselines across all tested SNRs, with the largest advantage concentrated in the low-rank scenarios.","feed_headline":"SIC-aided diffusion beats baselines on low-rank MIMO","feed_subtitle":"A joint channel-estimation and data-detection method for overloaded access points where users outnumber antennas.","key_machinery":"The load-bearing object is the SIC-aided score-based diffusion update over partial channel submatrices. The full channel matrix $H$ is split into $k = \\lceil N_u/N_r \\rceil$ blocks $H^{(i)}$ of size $N_r \\times N_r$ each, and a diffusion network $s_\\theta(H^{(i)}, \\sigma_{i,H})$ is trained to estimate the score (gradient of the log-prior) of each block. The MAP objective is written as a sum over SIC stages, with later-stage users folded into a Gaussian noise covariance $\\Sigma_i$, and the algorithm alternates between reordering users by current channel gain and updating each $H^{(i)}$ and data block $X_D^{(i)}$ with the gradients in Eqs. (4)-(5). The split is what makes the per-stage subproblem nearly full-rank, so the same diffusion machinery that works for massive MIMO can be applied in the overloaded regime.","core_discovery":"The central claim is that the failure of prior diffusion-based receivers in low-rank channels is not fundamental. By estimating the score of each $N_r \\times N_r$ channel submatrix separately rather than the full channel, and by cancelling the contributions of already-decoded submatrices through successive interference cancellation, the remaining per-stage estimation problem is close to full rank and can be handled by the same Langevin-diffusion machinery that already works in massive MIMO. The paper derives a MAP objective that factorizes over SIC stages and an iterative update that alternates channel-submatrix estimation, data-symbol detection, and a decoding order based on current channel gains. The reported simulations show this receiver matching or beating learned D-AMP, joint Langevin diffusion, and LMMSE baselines in NMSE and SER for both rank-sufficient and rank-deficient channels, with the clearest advantage at low SNR and in the low-rank case.","pith_inferences":["The submatrix-splitting trick is independent of the generative model, so the same SIC ordering and likelihood factorization could be combined with other learned priors, such as GANs or normalizing flows, rather than only score-based diffusion; the paper does not test this.","The channel-gain-based ordering is a heuristic; a decoding order optimized for post-cancellation signal-to-interference-plus-noise ratio would likely improve the low-SNR gains further, but the paper does not explore it.","The reported gains come from simulations with one antenna ratio (64 users, 32 antennas) and perfect SIC assumptions, so how the gains degrade with imperfect cancellation or with more extreme ratios, such as eight users per antenna, remains an open question.","Because diffusion sampling is iterative, the improvement over L-MMSE and L-DAMP may come with a substantially higher computational cost; the paper does not report runtime or complexity comparisons, so practical deployment would need to trade the error gains against latency."],"forward_implications":["In overloaded uplink systems where users outnumber receive antennas, a receiver can estimate the channel and detect data with lower NMSE and SER than learned denoising, joint Langevin diffusion, or LMMSE baselines at every tested SNR.","The largest gains appear in low-rank channels and at low SNR, exactly where the baselines degrade most, so the method widens the operating range of diffusion-based receivers.","The dynamic SIC order, re-derived from current channel strengths, gives a practical rule for deciding which partial channels to estimate first as the channel estimates improve.","For a fixed pilot length, increasing the number of data symbols changes the SER, so the method remains usable as block length grows without requiring more pilots."],"supporting_citations":[{"why":"It supplies the baseline joint Langevin diffusion method for massive MIMO and the annealing trick that approximates the discrete symbol prior, both of which the proposed algorithm extends to partial channel submatrices.","marker":"[13]"},{"why":"It provides the low-complexity diffusion-based channel estimator architecture and the simulation configuration that the paper's experiments follow.","marker":"[12]"},{"why":"It defines the learned D-AMP baseline that the proposed method is compared against for both channel estimation and symbol error rate.","marker":"[18]"},{"why":"It defines the linear MMSE semi-blind channel estimation baseline used as a comparison in the NMSE experiments.","marker":"[19]"},{"why":"It provides the QuaDRiGa 3D channel model used to generate the 3GPP low-rank channel instances for the simulations.","marker":"[21]"}],"fun_headline_variants":["Diffusion model handles overloaded low-rank MIMO","SIC makes diffusion work in low-rank MIMO","Joint channel and data recovery via SIC-diffusion","Low-rank MIMO cracked by SIC-aided diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole iterative loop depends on treating users decoded later as Gaussian noise with a known covariance, and on the gradient formulas in Eqs. (4)-(5) really describing the stated MAP objective; if either is off, the loop may not solve the problem it claims to solve.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model handles overloaded low-rank MIMO","SIC makes diffusion work in low-rank MIMO","Joint channel and data recovery via SIC-diffusion","Low-rank MIMO cracked by SIC-aided diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1326,"prompt_tokens":896,"completion_tokens":430,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":365}},"tokens_in":512,"tokens_out":430,"duration_ms":4435,"temperature":1.0,"reasoning_tokens":365,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:30:42.471205+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-derive the gradient of the MAP objective in Eq. (2) with respect to a fixed $H^{(i)}$ for a two-stage system and compare term by term with Eq. (4); if any stage whose $H^{(j)}$ and $X^{(j)}$ do not appear in that objective's likelihood factor still contributes to the gradient sum, then the algorithm is maximizing a different objective than the one stated.","supporting_citations":[{"cited_title":"Reduc- ing pilot overhead: channel estimation with symbol rep- etition in mimo-fbmc systems,","cited_arxiv_id":null,"evidence_quote":"It supplies the baseline joint Langevin diffusion method for massive MIMO and the annealing trick that approximates the discrete symbol prior, both of which the proposed algorithm extends to partial channel submatrices."},{"cited_title":"Joint channel es- timation and symbol detection in mimo-ofdm systems: A deep learning approach using bi-lstm,","cited_arxiv_id":null,"evidence_quote":"It provides the low-complexity diffusion-based channel estimator architecture and the simulation configuration that the paper's experiments follow."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It defines the learned D-AMP baseline that the proposed method is compared against for both channel estimation and symbol error rate."},{"cited_title":"Vlsi designs for joint channel estimation and data de- tection in large simo wireless systems,","cited_arxiv_id":null,"evidence_quote":"It defines the linear MMSE semi-blind channel estimation baseline used as a comparison in the NMSE experiments."},{"cited_title":"Cddm: Chan- nel denoising diffusion models for wireless communi- cations,","cited_arxiv_id":null,"evidence_quote":"It provides the QuaDRiGa 3D channel model used to generate the 3GPP low-rank channel instances for the simulations."}],"review_version":1}