{"id":"1912b028-21f5-48f0-927f-a26a160c674f","arxiv_id":"2507.10074","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A superimposed-pilot MIMO-OFDM receiver with variational and deep-learning channel estimators improves throughput and block error rate over orthogonal-pilot baselines in simulations and over-the-air tests.","lead":"This paper builds a wireless receiver that overlays probe signals on data symbols to use the spectrum more efficiently, then cleans up the resulting interference with iterative decoding and learned channel estimators. Simulation and over-the-air tests show higher throughput than standard pilot designs, at the cost of a more complex receiver.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"High-mobility throughput claims rest on a single simulated channel model (CDL-C in Sionna); the OTA experiment is static and matched, so the headline gains are not experimentally established.","rationale":"The paper's internal derivations are plausible: the VMP posterior mean in (22) and VMP-L in (27) are consistent up to an irrelevant scaling of the covariance, and the OTA results at v=0 reproduce the low-mobility throughput trade-off (7.7% OTA vs. 7.5% simulated). The concern is not a mathematical error but external validity of the central quantitative claim. The strongest claim is explicitly about high mobility and robustness to statistical mismatch, yet the only evidence for those regimes is a single-channel-model simulation. This does not justify rejection because the simulation is carefully specified and the mechanism (pilot density over all REs) is physically plausible; it justifies the reader's CONDITIONAL verdict until a high-mobility OTA/emulator test or an iterative-OP control is provided. Releasing code or per-point confidence intervals would also help, but the decisive missing evidence is the fast-fading experimental comparison.","tokens_in":17582,"tokens_out":13385,"duration_ms":163656,"concrete_test":"Use the existing OTA prototype plus a programmable channel emulator (or a moving UE) to reproduce the 2x2 high-mobility setting at 324 km/h, first with CDL-C and then with a different profile such as CDL-A or a measured urban channel. Measure BLER and throughput for SIP with I=2 and OP 2P/4P under identical SNR points. Separately, add an 'OP receiver, 2P/4P, I=2' baseline that uses the same soft-remapping feedback and refined LMMSE/VMP-L estimation as the proposed receiver. If the observed throughput gains fall materially below 19.7%/39.8%, or the DL mismatch gain below 2 dB outside CDL-C, the central claim must be qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline high-mobility gains (19.7% over OP 2P and 39.8% over OP 4P in Section IV-B; >2 dB DL robustness in Section IV-C) are produced only by the Sionna simulator with a CDL-C channel at 3.5 GHz. The OTA validation in Section V is explicitly static (v = 0 km/h), line-of-sight, and uses channel statistics and DL training data collected from the same OTA environment; it therefore exercises only the matched, low-mobility regime. The mismatch study in Section IV-C varies only the training speed (72 vs. 324 km/h) while fixing the CDL-C delay profile, so it tests one Doppler mismatch, not the broader distribution shift that deployment would encounter. Because the claimed quantitative benefits are specific to fast fading, the absence of any high-mobility experimental or emulated validation leaves the central claim dependent on CDL-C being representative of real fast-fading channels. A secondary confound is that the OP baselines are not given the same iterative JCDD feedback loop, so part of the simulated gain could be receiver-processing gain rather than SIP-specific pilot overhead recovery.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an iterative joint channel estimation, detection, and decoding (JCDD) receiver for superimposed-pilot (SIP) MIMO-OFDM, augmenting a prior interference-cancellation design with LMMSE-based refinement and two adaptive estimators: a variational message-passing (VMP) method with a low-complexity variant (VMP-L), and a CNN-plus-attention deep learning estimator. Simulations under CDL-C channels in 2x2 and 64x4 MIMO at 15 and 324 km/h compare the receiver against orthogonal-pilot (OP) baselines, reporting throughput gains up to 19.7% over OP 2P and 39.8% over OP 4P in the 2x2 high-mobility case, and more than 2 dB robustness gains under mismatched channel statistics. A 2x2 MIMO over-the-air (OTA) prototype at 2 GHz demonstrates a 7.7% throughput gain over an OP 1P baseline in a static line-of-sight environment. The paper includes complexity analysis and releases simulation details via the Sionna package.","tokens_in":17845,"tokens_out":5521,"duration_ms":67004,"significance":"If the claims are fully supported, the paper would make a useful contribution by showing that SIP with an iterative receiver and adaptive estimation can recover pilot overhead while maintaining high-mobility and multi-stream performance, which matters for spectral-efficiency-oriented 6G designs. The inclusion of an OTA prototype and explicit complexity comparisons is a strength, as is the use of standard channel models and the Sionna simulator. However, the current evidence does not yet establish the headline high-mobility and generalization claims: the OTA test is static and matched, the mismatch study covers only one type of Doppler mismatch, and no comparison is made with the prior SIP receiver [24] that the paper directly extends. These gaps are central to the claimed contributions rather than cosmetic.","major_comments":[{"comment":"The throughput gains of 19.7% and 39.8% are computed against OP receivers that use LMMSE-based estimation and detection without the iterative JCDD feedback loop described in Section II-B. Because the proposed receiver changes both the pilot scheme and the receiver architecture, these percentages do not isolate the SIP-specific benefit of recovering pilot overhead. Please add an OP baseline that is run with the same I=2 JCDD iterations, and also compare against the prior SIP IC-based JCDD receiver in [24], which is the design this paper modifies. The exact SNR operating point at which the percentage gains are evaluated should also be stated explicitly.","section":"Section IV-B, Figs. 6-7"},{"comment":"The OTA experiment is static (v=0 km/h), line-of-sight, and uses channel statistics and DL training data collected from the same OTA environment, so it validates only the matched, low-mobility regime. The high-mobility gains in Section IV-B rest entirely on Sionna simulations with the CDL-C delay profile. The mismatch study in Section IV-C varies only the training speed from 72 to 324 km/h while keeping the same delay profile and the same frequency/spatial statistics. This is a narrow distribution shift and does not substantiate the abstract's broad claim of robustness under mismatched channel conditions. Please either provide high-mobility emulator or OTA validation, or substantially temper the generalization claims in the abstract and conclusion.","section":"Sections V-C and IV-C"},{"comment":"The quantitative robustness claims, such as 'a performance gain of more than 2 dB' for the DL estimator, are presented without confidence intervals, standard errors, or the number of Monte Carlo frames used for each BLER point. In Fig. 8(b), the authors themselves describe the VMP-L gain over LMMSE as marginal, and the DL and VMP-L curves may overlap within statistical uncertainty. Please provide error bars or confidence intervals and state the number of independent trials per point, so that the relative merits of VMP-L and DL can be assessed.","section":"Section IV-C, Fig. 8"},{"comment":"The despreading operation in the DL estimator assumes that the residual interference terms in (25) are zero-mean, but after only I=2 JCDD iterations the soft-symbol and channel-estimation errors can be biased, especially under high mobility. The averaging over L1=6 and L2=2 neighboring REs may therefore introduce systematic bias rather than simply suppressing noise. Please provide a justification for the zero-mean assumption or quantify the bias it introduces, for example by comparing despreading with and without an unbiasedness correction.","section":"Section III-C, Eq. (25)"}],"minor_comments":[{"comment":"Please define Omega explicitly for each OP pilot pattern (1P, 2P, 4P) and state the SNR at which the throughput gains in Section IV-B are computed, since the throughput curves cross at different SNR values.","section":"Eq. (31) and Section IV-B"},{"comment":"The 1P pilot pattern is described only by a diagram; please clarify whether it uses one OFDM symbol with all subcarriers or a frequency-domain comb, as this directly affects the reported overhead and the comparison with SIP.","section":"Fig. 5"},{"comment":"In the DL complexity expression, the constant factor '10' in '10 F^2 C^2' is unexplained. A brief derivation or a reference to the network layer count would help the reader reproduce the complexity comparison.","section":"Table II and Section III-C"},{"comment":"After Eq. (12), the superscript (t) is dropped for notational convenience; please remind the reader that all variables in Section III-A still depend on the time index, as this omission initially makes the equations ambiguous.","section":"Section III-A"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and contains a useful prototype and a fair amount of careful derivation. The main risk is overclaiming: the central quantitative advances are not yet isolated from the iterative-receiver baseline, and the OTA evidence is limited to a static, matched environment. A revision that adds the missing [24] comparison, an OP-with-feedback baseline, and error bars would materially strengthen the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a well-executed engineering paper that delivers what it promises. The new content is the specific combination: an IC-based JCDD loop starting from LMMSE channel estimates, two VMP-based estimators that avoid time-domain correlation, an attention-augmented CNN with despreading, and a real over-the-air prototype. The OTA experiment, even at v=0 over a short LOS link, is more than most papers in this subfield bother to do. The Section III derivations look internally consistent, and the complexity table is honest, including the admission that full VMP becomes intractable in the 64x4 case and that VMP-L's gain is marginal there.\n\nThe weaknesses are real but not fatal. First, there is no comparison against the closest prior SIP receiver, Qian et al. [24]. Without that baseline, the reported gains over OP receivers could partly be generic iterative-receiver processing gain rather than something specific to SIP and these estimators. That is a concrete missing experiment, and the paper would be much cleaner with it. Second, there are no error bars anywhere; the Sionna curves are deterministic-looking, but one would like to know the spread across channel realizations or training runs. Third, the high-mobility headline numbers (19.7% and 39.8% in 2x2; 19.4% and 40.5% in 64x4) come only from simulated CDL-C. The mismatch study shifts training speed from 72 to 324 km/h, but keeps the same delay profile, so it is a narrow distribution shift. The OTA test is static, matched, and LOS, so it does not exercise the fast-fading claim at all. These limitations matter because the design is explicitly motivated by high mobility.\n\nI do not think the paper is overclaiming in a deceptive way. It states the OTA conditions clearly and acknowledges VMP-L's limited benefit. The main issue is that the quantitative gains are less firmly established than the narrative suggests. A revision that adds the [24] comparison, error bars, and ideally a channel-emulator test at high speed would substantially raise confidence.\n\nThis deserves a serious referee. It is an incremental but useful contribution with genuine experimental effort. I would send it to review, and I would ask the authors to address the missing baseline and the experimental scope before publication.","headline":"Solid engineering contribution with real OTA validation, but the quantitative high-mobility claims rest on simulations alone and the closest prior SIP baseline is missing.","tokens_in":18411,"tokens_out":1555,"would_cite":false,"duration_ms":20024,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A12","94A40"],"pacs":[],"model":"deepseek-v4-flash","headline":"An iterative receiver makes superimposed pilots beat orthogonal pilots in MIMO-OFDM throughput, with gains up to about 40 percent in fast-fading simulation and a working over-the-air prototype.","keywords":["superimposed pilots","iterative receiver","joint channel estimation and decoding","variational message passing","deep learning channel estimation","MIMO-OFDM","high mobility","over-the-air prototype"],"falsifier":"Run the proposed receiver at a user speed of 324 km/h over an outdoor measurement or ray-traced channel whose Doppler spread and delay profile differ from CDL-C, and check whether the >2 dB mismatched-statistics gain and the roughly 20-40% throughput advantage over the 2P and 4P orthogonal-pilot baselines survive; if they vanish or invert, the central claim fails.","tokens_in":1885,"feed_emoji":"📡","tokens_out":2444,"duration_ms":66501,"temperature":0.7,"pith_summary":"This paper tries to establish that superimposed pilots—where reference symbols are added onto every data resource element—can outperform dedicated orthogonal pilots in both throughput and reliability, provided the receiver couples channel estimation, detection, and decoding into a few iterative rounds. The payoff would be higher spectral efficiency in MIMO-OFDM uplinks, especially at high user speeds and with many spatial streams, where orthogonal pilot patterns either fail to track the channel or consume too many resources. The paper reports that two iterations of its joint receiver recover the pilot overhead: a 19.7% throughput gain over a two-symbol orthogonal pilot pattern and a 39.8% gain over a four-symbol pattern in a 2x2 high-mobility simulation, with similar gains in a 64x4 setting. It also claims three refined channel estimators—two variational message-passing variants and one convolutional-neural-network estimator—that keep the receiver robust when the channel statistics used in training do not match the test channel.","feed_headline":"Iterative receiver lifts MIMO throughput by up to 40 percent","feed_subtitle":"Superimposed pilots plus two decoding iterations beat dedicated pilot patterns in fast-fading 2x2 and 64x4 links.","key_machinery":"The engine is the iterative IC-based JCDD receiver: estimate the channel, subtract pilot contamination and soft-data interference, detect with LMMSE, decode, and feed extrinsic LLRs back as soft symbol estimates for the next estimation round. The initial estimate uses separable time-frequency LMMSE interpolation, whose dependence on time-domain correlation is the weak point that VMP, VMP-L, and the DL estimator are designed to remove. VMP builds a factor graph of the equivalent MIMO system and updates a Gaussian posterior over the channel without time-domain correlation; VMP-L decouples variable nodes per transmit antenna, cutting complexity from O(T(NrNtK)^3) to O(T Nt (Nr+1)$K^{3}$); the DL estimator applies a despreading operation, a residual CNN denoiser, an attention block that ingests soft-symbol confidence scores, and a DNN interpolator to produce full-grid channel estimates.","core_discovery":"The central discovery is that a joint channel-estimation, detection, and decoding (JCDD) receiver operating on superimposed pilots can exploit soft-decoded data feedback to cancel mutual pilot-data interference so effectively that the full-grid pilot occupation becomes a net win rather than a loss. The paper builds on a prior interference-cancellation JCDD design, replacing its least-squares estimator with LMMSE, and then adds two families of refined estimators: variational message passing (VMP) and its low-complexity variant VMP-L, which perform inference without time-domain correlation statistics, and a deep-learning (DL) estimator that combines despreading, a residual CNN denoiser, an attention mechanism fed by soft-decision confidence, and a neural interpolator. Under matched statistics, the receiver with I=2 iterations approaches the ideal SIP-CSI performance upper bound; under mismatched time-varying statistics, the DL estimator maintains more than 2 dB gain over LMMSE in the 2x2 case and about 1.5 dB over VMP-L in the 64x4 case. An over-the-air static line-of-sight experiment with a 2x2 software-defined-radio prototype confirms the receiver's practical viability, showing a 7.7% throughput gain over an orthogonal-pilot baseline despite a roughly 1 dB BLER penalty.","pith_inferences":["The same iterative data-aided cancellation could transfer to OTFS or wideband systems where orthogonal pilots are even costlier; the despreading-plus-attention recipe is generic across time-frequency lattices.","A fair comparison at constant overhead—for instance SIP with reduced pilot power versus orthogonal pilots with fewer pilot symbols—would isolate whether the gain comes from the receiver iterations or from the extra data resource elements.","The attention block's use of soft-decision confidence suggests a testable extension: use the confidence map to early-exit the iterations, cutting latency when the feedback is already reliable.","Because the over-the-air test is static and line-of-sight, the high-mobility claims hinge on simulation realism; a ray-traced or measured high-mobility channel test would be the natural next validation."],"forward_implications":["Two JCDD iterations are enough to approach the SIP-CSI bound, so the iterative overhead is modest and the configuration is practical.","SIP with this receiver removes the dependence of pilot density on the number of transmit antennas, which is the reason the 64x4 massive-MIMO case improves over orthogonal pilots by roughly 19-40% in high mobility.","Under mismatched channel statistics, VMP and VMP-L avoid the degradation of LMMSE by ignoring time-domain correlation, while the DL estimator stays robust with over 2 dB gain over LMMSE.","The over-the-air prototype shows the receiver works outside simulation, achieving a 7.7% throughput gain over an orthogonal-pilot baseline even though the experiment is static.","The complexity analysis indicates the DL estimator scales linearly with resource blocks, making it the most deployable option among the three refined estimators."],"supporting_citations":[{"why":"Supplies the base IC-based JCDD receiver design that the paper modifies with LMMSE estimation and refined extrinsic information.","marker":"[24]"},{"why":"Provides the variational message-passing formulation for joint channel estimation and decoding that the VMP and VMP-L estimators build on.","marker":"[16]"},{"why":"Defines the code-domain multiplexing despreading operation that the DL estimator uses before CNN-based denoising.","marker":"[25]"},{"why":"Supplies the soft-input soft-output LMMSE detection and extrinsic LLR computation used in the iterative receiver.","marker":"[26]"},{"why":"Provides the deep-residual-learning CNN structure for OFDM channel estimation that the DL denoiser network adapts.","marker":"[27]"},{"why":"Motivates the attention mechanism integrated into the DL channel estimator for robust feature weighting.","marker":"[29]"},{"why":"Defines the CDL-C power delay profile used to generate the simulated high- and low-mobility channels.","marker":"[32]"},{"why":"Supplies the link-level simulation library used for the numerical results.","marker":"[33]"}],"fun_headline_variants":["Iterative receiver with DL estimator boosts MIMO throughput","Superimposed pilots plus deep learning beat conventional schemes","AI-aided receiver outdoes orthogonal pilots in fading MIMO","Joint estimation and decoding lift MIMO capacity with pilots","Deep-learning receiver wins over-the-air with superimposed pilots"],"cache_read_input_tokens":20480,"weakest_assumption_plain":"The high-mobility throughput gains are demonstrated only in simulated CDL-C channels with matched statistics, while the over-the-air validation is static, line-of-sight, and uses matched training data, so the fast-fading benefit rests on the assumption that the simulated channel captures the relevant real-world dynamics.","fun_headline_variants_meta":{"raw":{"variants":["Iterative receiver with DL estimator boosts MIMO throughput","Superimposed pilots plus deep learning beat conventional schemes","AI-aided receiver outdoes orthogonal pilots in fading MIMO","Joint estimation and decoding lift MIMO capacity with pilots","Deep-learning receiver wins over-the-air with superimposed pilots"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1381,"prompt_tokens":1032,"completion_tokens":349,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":271}},"tokens_in":648,"tokens_out":349,"duration_ms":4843,"temperature":1.0,"reasoning_tokens":271,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:40:32.933610+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the proposed receiver at a user speed of 324 km/h over an outdoor measurement or ray-traced channel whose Doppler spread and delay profile differ from CDL-C, and check whether the >2 dB mismatched-statistics gain and the roughly 20-40% throughput advantage over the 2P and 4P orthogonal-pilot baselines survive; if they vanish or invert, the central claim fails.","supporting_citations":[{"cited_title":"Enhancing wideband mul- tiuser MIMO uplink using superimposed pilots: Joint receiver design,","cited_arxiv_id":null,"evidence_quote":"Supplies the base IC-based JCDD receiver design that the paper modifies with LMMSE estimation and refined extrinsic information."},{"cited_title":"Variational message-passing for joint channel estimation and decoding in MIMO-OFDM,","cited_arxiv_id":null,"evidence_quote":"Provides the variational message-passing formulation for joint channel estimation and decoding that the VMP and VMP-L estimators build on."},{"cited_title":"NR; Physical channels and modulation,","cited_arxiv_id":null,"evidence_quote":"Defines the code-domain multiplexing despreading operation that the DL estimator uses before CNN-based denoising."},{"cited_title":"ASIC implementation of soft- input soft-output MIMO detection using MMSE parallel interference cancellation,","cited_arxiv_id":null,"evidence_quote":"Supplies the soft-input soft-output LMMSE detection and extrinsic LLR computation used in the iterative receiver."},{"cited_title":"Deep residual learning meets OFDM channel estimation,","cited_arxiv_id":null,"evidence_quote":"Provides the deep-residual-learning CNN structure for OFDM channel estimation that the DL denoiser network adapts."},{"cited_title":"Deep residual learning with attention mechanism for OFDM channel estimation,","cited_arxiv_id":null,"evidence_quote":"Motivates the attention mechanism integrated into the DL channel estimator for robust feature weighting."},{"cited_title":"Study on channel model for frequencies from 0.5 to 100 GHz,","cited_arxiv_id":null,"evidence_quote":"Defines the CDL-C power delay profile used to generate the simulated high- and low-mobility channels."}],"review_version":1}