{"id":"c7025bb8-65a7-44e2-b546-05f2c6a10584","arxiv_id":"2504.20777","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A delay-domain sparse precoder, learned by a DNN with variational Bayesian training, cuts DMRS pilot overhead up to 8x and improves BER in simulated MIMO-OFDM downlinks.","lead":"This paper proposes a precoder design that makes the effective channel sparse in the delay domain, allowing demodulation pilots to be shared across up to eight streams in MIMO-OFDM. In simulations with 4K-QAM, it reports up to 8x lower pilot overhead, about 5 dB better error rate, and much lower compute.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim of 8x pilot reduction depends on exact delay-domain support of D=72 taps; any fractional-delay leakage or learned window Dv exceeding 56 breaks the anti-aliasing reconstruction in Eq. (10), so the overhead gain is not robust.","rationale":"The paper's advertised contribution—up to 8x DMRS reduction with better BER—stands or falls on the claim that the effective channel has delay-domain support Deff=127, which fits within the K/Ar=128 decimation window. This claim requires two exact-sparsity conditions: the propagation channel must be exactly contained in the first D=72 DFT delay taps, and the precoder must be exactly a Dv=56-tap window. Both are idealized: continuous multipath delays cause spectral leakage that spreads energy beyond any fixed tap window, and the adaptive-window component (Eq. (55)-(57)) might select Dv>56, enlarging Deff beyond the alias-free region. The paper demonstrates only a single channel realization (Fig. 6) and does not quantify the out-of-window energy in its simulations. This is precisely the reader's weakest assumption, and I agree it is the most load-bearing concern. A concrete fractional-delay experiment would settle whether the issue lands. Since the reader already marked the paper CONDITIONAL with this as the main caveat, my stress-test does not change the verdict—it reinforces it. The other issues (no code, no error bars, omitted DNN baselines) are real but secondary relative to the enabling sparsity assumption.","tokens_in":18040,"tokens_out":8365,"duration_ms":83091,"concrete_test":"Generate the same 802.11ax model B channel but add a small fractional delay offset (e.g., 0.3 sample durations) to all taps, while keeping the identical DNN architecture, pre-training, and Ar=8 DMRS allocation. Measure the effective-channel estimation NMSE (Fig. 13 equivalent) and the end-to-end BER at SNR=38 dB, L=Nt/2. If NMSE degrades by more than 3 dB relative to the aligned case, or if the BER at 38 dB worsens by more than 2x, the exact-support assumption is load-bearing and the 8x reduction claim is not robust. Alternatively, directly compute the energy fraction outside the first 72 delay taps for the paper's simulated channel; if it exceeds about 1%, the premise of Eq. (10) is already violated in the reported setup.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's pilot-overhead reduction is enabled by the assertion that the propagation channel has exactly D=72 non-zero taps in the K-point DFT delay domain and that the learned precoder is exactly a Dv=56-tap low-pass filter, giving an effective delay spread Deff = D + Dv - 1 = 127 < K/Ar = 128. This exact-support condition is required for the FDM reconstruction in Eq. (10) to be alias-free: the anti-aliasing filter selects the first 127 taps and discards the other K/Ar windows. In a physically realistic ray-tracing or 802.11ax channel with arbitrary (non-sample-spaced) delays, the K-point IDFT of a sum of complex exponentials has sinc-like leakage across all taps, so the channel is not exactly supported on the first 72 taps. Any energy outside the first 127 taps folds into the selected window under Ar=8 decimation and cannot be separated by Eq. (10), setting an error floor independent of SNR. The paper provides only Fig. 6 as an illustration and no measurement of out-of-window energy, and it does not analyze robustness to channel delay profile or to the learned window length Dv (Eq. (55)-(57)) deviating from 56. If either condition is violated, the claimed 8x reduction is not achievable and the end-to-end BER gains over baselines may shrink or vanish.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a delay-domain sparse precoder for MIMO-OFDM downlink transmission, combined with an end-to-end DNN transceiver and a variational Bayesian training framework. The central idea is to design the precoder as a short delay-domain tap vector so that the effective channel has small delay spread, allowing demodulation reference signals to be sent on frequency-division-multiplexed subcarrier subsets with up to an 8x reduction in pilot overhead. Three modules are presented: propagation channel estimation from uplink SRS, sparse precoder design obtained by unrolling a block-coordinate-ascent solution to an EVM-minimization problem, and effective channel (CSIR) estimation from DMRS. The training loss is derived from an evidence-lower-bound objective with Markov priors for delay-domain support. Simulations over a modeled 802.11ax channel report NMSE close to genie-aided MMSE and end-to-end BER gains of about 5 dB over the baselines. The main load-bearing assumptions are that the propagation channel has exactly D=72 nonzero delay-domain taps and that the learned precoder window has Dv=56 taps, so the effective delay spread Deff=127 fits within the K/Ar=128 FDM pilot window.","tokens_in":18376,"tokens_out":4697,"duration_ms":54498,"significance":"If the delay-support and pilot-decimation assumptions hold, the paper would make a useful contribution: it combines a sparse-precoder design principle with an interpretable, model-assisted DNN data path, and the optimization problem P(A) and its algorithmic unrolling are clean and standard. The VBI training framework is extensive and ties each module to a well-defined probabilistic model, which is a genuine strength relative to black-box end-to-end designs. However, the central 8x pilot-reduction claim is not an emergent result; it is enforced by choosing Dv=56 and Deff=127, and it is demonstrated only for an exact-support channel model. The absence of code, error bars, and sensitivity analysis means the empirical gains cannot yet be independently verified. The contribution is therefore promising but currently conditional on assumptions that are not defended in the manuscript.","major_comments":[{"comment":"The anti-aliasing FDM reconstruction in Eq. (10) assumes that the delay-domain channel is exactly zero outside the first D=72 taps. For the ray-tracing channel model in Eq. (1) with non-sample-spaced path delays, the K-point DFT of a sum of complex exponentials has sinc-like leakage across all delay taps, so energy outside the first 127 taps folds into the selected window when Ar=8 and cannot be separated by Eq. (10). The paper provides only the single illustration in Fig. 6 and does not report the out-of-window energy for the simulated channel delay profile or any variation of the delay profile. Since the 8x reduction in Contribution 1 depends on this exact-support condition, a robustness analysis over delay profiles with leakage, or a derivation of the required guard, is needed.","section":"Sec. II-A, Eq. (10); Sec. III-C, Eq. (44); Fig. 6"},{"comment":"The claimed 8x reduction appears to be set by hand rather than derived. With D=72 and Dv=56, Eq. (44) gives Deff=127, which is just below K/Ar=128; the text also states 'We set Deff = K/8', i.e., 128, which is inconsistent with Eq. (44) for these parameters. In addition, the adaptive window component in Eqs. (55)-(57) permits Dv as large as Deff, which would increase Deff and reduce Ar below 8; the paper does not report the learned Dv distribution or test sensitivity to Dv different from 56. The paper should state explicitly how Dv and Deff are selected and show that the 8x reduction is robust to the learned window length.","section":"Sec. V.C.2, Eq. (44), Eqs. (55)-(57)"},{"comment":"All performance claims rely on single-curve simulations with no error bars, confidence intervals, seeds, or repeated trials, and no code is released. With 10^4 test channel realizations and no indication of variance across training runs, the reported 2.5-5 dB BER gains and NMSE improvements cannot be statistically assessed or reproduced. For a journal submission, the authors should provide either code, error bars over multiple training seeds, or a statement of the simulation setup that makes the empirical claims reproducible.","section":"Sec. VI, Figs. 12-15"}],"minor_comments":[{"comment":"The condition 'K≥DA' should be written more precisely as K ≥ D A_u, and the definition of A_u should be repeated in the sentence following Eq. (10) for readability.","section":"Sec. II-A"},{"comment":"The notation 'ELO' is used for the evidence lower bound; the standard abbreviation is 'ELBO'. Please correct this throughout Section V.","section":"Sec. V.C.1"},{"comment":"The caption says 'for Nr = 8,N r = 8,L = 4'; the first entry should be Nt, not Nr.","section":"Table II caption"},{"comment":"The sentence 'where the architecture of gshp is illustrated in Fig. 11a' appears to refer to the wrong panel; the support supplementary network is shown in Fig. 11(b), and the text should be corrected accordingly.","section":"Sec. V.A.2, Fig. 11"}],"recommendation":"major_revision","confidential_remarks":"I would not accept the paper in its current form because the headline 8x pilot reduction is a direct consequence of the chosen Dv and Deff and is not shown to be robust to delay-domain leakage. The request for robustness analysis, explicit Dv/Deff selection, and reproducible simulation results is, in my view, a necessary condition, not a stylistic preference."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this is a solid engineering paper whose headline number (8x pilot reduction) is real only inside a specific tap-spaced channel model. The core idea is good, but the advertised gain is a design choice plus a support assumption, not a derived or robustly tested fact.\n\nWhat's new: the authors notice that the spatial-delay precoders from [29,30] already produce sparse delay-domain effective channels, and they turn that into a DMRS reduction scheme. The delay-domain sparse precoder formulation P(A) with EVM minimization is clean, and the BCD updates are standard but sensible. The DNN unrolling of those updates is a reasonable model-assisted design, and the VBI training with Markov priors for clustered sparsity is elaborate but grounded in the problem. The simulation results are consistent: the proposed DNN beats LS, OMP, VAMP, LASSO, and the sparse WMMSE baseline by meaningful margins, and the complexity table is useful. Credit where due: the paper is honest that [29,30] had the precoder but missed the pilot-reduction significance.\n\nSoft spots, in order of size. First, the 8x claim is hand-set: Deff=K/8 and Dv=56 are chosen so Deff=127 fits under the K/Ar=128 decimation window. The anti-aliasing reconstruction in Eq. (10) is exact only if the propagation channel is supported on exactly the first D=72 taps and the learned precoder uses exactly 56 taps. Real channels with fractional delays produce sinc leakage across all 1024 taps; that energy aliases into the selected window under decimation by 8 and sets an error floor. Fig. 6 shows one illustrative case; there is no sweep over delay profiles or over Dv. This is the load-bearing assumption and it needs robustness analysis. Second, there is no code or data, and no error bars; single-run BER curves in a DNN paper make it hard to judge whether the 5 dB gains are stable. Third, the BER comparison omits black-box end-to-end DNN baselines, so the claim that model-assisted unrolling is what buys the gain is not fully tested. The 'perfect reconstruction' wording in Section III-C overstates the case even under the model. The citation pattern is fine; [29,30] is properly credited.\n\nOverall: the central argument is plausible and the paper deserves serious refereeing, but the 8x reduction should be reported as conditional on the tap-spaced support assumption. I'd send it to review with a request for code, error bars, and a fractional-delay delay-profile robustness study.","headline":"A solid engineering paper with a genuinely useful idea, but the headline 8x pilot reduction is conditional on an exact tap-spaced delay-support assumption that needs more robustness analysis.","tokens_in":18889,"tokens_out":3014,"would_cite":false,"duration_ms":32806,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A sparse delay-domain precoder cuts MIMO-OFDM pilots by up to 8x.","keywords":["precoder design","pilot reduction","MIMO-OFDM","model-assisted deep learning","end-to-end learning","channel estimation","delay-domain sparsity","variational Bayesian inference"],"falsifier":"Simulate the same 802.11ax MIMO channel but increase the maximum delay spread beyond 72 taps (e.g., insert an 80th tap) or add a fractional delay; if the anti-aliasing reconstruction in Eq. (10) shows NMSE degradation beyond a few dB or the BER at SNR=38 dB with 4096-QAM no longer beats VAMP by the reported margin, the 8x pilot-reduction claim fails.","tokens_in":17862,"feed_emoji":"📡","tokens_out":10573,"duration_ms":95801,"temperature":0.7,"pith_summary":"This paper tries to establish that a MIMO-OFDM transmitter can deliberately shape its precoder so that the channel the receiver must estimate is sparse in the delay domain, allowing demodulation reference signals to use only a fraction of the subcarriers instead of one full OFDM symbol per spatial stream. The authors claim this reduces pilot overhead by up to 8x for 1024- and 4096-QAM downlinks, and that a DNN built by unrolling the resulting precoding algorithm, trained end-to-end with a variational Bayesian loss, beats iterative sparse-recovery baselines in bit error rate while using far less computation. If correct, this would let high-order modulation be used in TDD systems without spending most of the frame on pilots, which is precisely the bottleneck for future WiFi throughput targets.","feed_headline":"Delay-sparse precoder cuts MIMO-OFDM pilots up to 8x","feed_subtitle":"A joint DNN design matches genie-aided channel estimation and beats iterative sparse-recovery baselines by 5 dB in BER.","key_machinery":"The central object is the delay-domain-sparse precoder $V=(\\mathbf{F}_v\\otimes I_{N_t})\\tilde{W}$, where $\\mathbf{F}_v$ is the $K\\times D_v$ partial DFT matrix keeping the first $D_v$ columns. This low-pass filter in the delay domain makes the effective channel support exactly $D_{\\mathrm{eff}}=D+D_v-1$ taps, turning the receiver's channel-estimation problem into a sparse anti-aliasing reconstruction for which a uniform FDM pilot scheme can host $A_r=\\lfloor K/D_{\\mathrm{eff}}\\rfloor$ orthogonal streams in one OFDM symbol. The mechanism that makes it trainable is algorithmic unrolling of the block-coordinate-ascent iterations (42)-(43), with a learned Lagrange multiplier, an adaptive window that chooses $D_v$ dynamically, and a variational Bayesian objective that couples the modules through Markov priors on channel support.","core_discovery":"The central claim is that a precoder restricted to $D_v=56$ delay-domain taps behaves as a low-pass filter, so convolving the propagation channel (which occupies the first $D=72$ taps) with it creates an effective channel of support $D_{\\mathrm{eff}}=D+D_v-1=127$ taps. With $K=1024$ subcarriers this support fits inside the $K/A_r=128$-sample window of a uniform FDM pilot pattern, so $A_r=8$ orthogonal pilot streams can share one OFDM symbol and the receiver reconstructs the effective channel with an anti-aliasing projection. The paper unrolls the resulting EVM-minimization block-coordinate-ascent solver into a DNN, adds NN-based channel estimation modules at the transmitter and receiver, and trains the entire datapath with a variational Bayesian objective whose Markov priors encode clustered delay-domain sparsity. The claimed result is effective channel estimation close to genie-aided MMSE and an end-to-end BER gain of about 5 dB over S-WMMSE combined with VAMP recovery.","pith_inferences":["The paper fixes $D_v=56$ and reports one delay profile; a natural extension is to sweep $D_v$ against pilot overhead and BER to trace the tradeoff curve, which the adaptive window component could in principle navigate automatically.","Because the sparse effective channel is deliberately engineered at the transmitter, the same trick could be applied to other pilot-limited links, such as positioning reference signals or OTFS, where the receiver-estimated quantity can be made sparse by design.","The Bayesian training framework's Markov priors on delay-domain support are not tied to OFDM; they could serve any datapath where clustered sparsity of an intermediate variable is the inductive bias, including massive MIMO at higher carrier frequencies."],"forward_implications":["With $K=1024$ and $D=72$, the design needs $M=\\lceil L D_{\\mathrm{eff}}/K\\rceil$ DMRS OFDM symbols instead of $L$; for $L=8$ and $D_{\\mathrm{eff}}=127$ this is one symbol instead of eight, an 8x pilot reduction.","The effective channel estimation NMSE is shown to be within a few dB of genie-aided MMSE and more than 7 dB better than VAMP, with BER gains of 5 dB over S-WMMSE at a BER of $10^{-3}$.","The end-to-end BER objective gives about 2.5 dB over the EVM-optimized precoder and another 2.5 dB over the optimization-based design, so jointly training the three modules is the main source of the 5 dB total gain.","Running the DNN costs roughly 0.4 GFLOPs and about 17 ms on CPU, which is an order of magnitude faster than the iterative baselines it beats."],"supporting_citations":[{"why":"Defines the uniformly sampled FDM pilot scheme whose anti-aliasing reconstruction the proposed design exploits.","marker":"[7]"},{"why":"Provides the VAMP sparse-recovery baseline that must be beaten in channel estimation and BER comparisons.","marker":"[12]"},{"why":"Gives the S-WMMSE delay-domain precoder baseline against which the sparse precoder's BER gain is measured.","marker":"[30]"},{"why":"Contributes the learned-Lagrange-network idea that makes the unrolled precoder solver trainable.","marker":"[38]"},{"why":"Establishes the algorithmic-unrolling principle used to turn the block-coordinate-ascent solver into a DNN.","marker":"[41]"},{"why":"Defines the ray-tracing clustered channel model that yields the $D=72$ delay-domain sparsity.","marker":"[42]"},{"why":"Provides the NN-based QAM de-mapper that converts soft estimates into bit probabilities for the cross-entropy loss.","marker":"[43]"}],"fun_headline_variants":["Sparse delay-domain precoder cuts MIMO-OFDM pilots 8x","Bayesian deep MIMO-OFDM: sparse precoder, 8x fewer pilots","Sparse precoder + Bayesian DNN: 8x pilot cut, 5 dB BER gain","Interpretable deep datapath: 8x pilot reduction in MIMO-OFDM","Delay-domain sparse precoder achieves 8x pilot savings"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pilot-reduction gain rests on the assumption that the physical channel occupies only the first $D=72$ delay taps, so the sparse precoder's effective support $D+D_v-1=127$ stays inside the $K/A_r=128$-sample decimation window and the anti-aliasing filter can reconstruct it losslessly.","fun_headline_variants_meta":{"raw":{"variants":["Sparse delay-domain precoder cuts MIMO-OFDM pilots 8x","Bayesian deep MIMO-OFDM: sparse precoder, 8x fewer pilots","Sparse precoder + Bayesian DNN: 8x pilot cut, 5 dB BER gain","Interpretable deep datapath: 8x pilot reduction in MIMO-OFDM","Delay-domain sparse precoder achieves 8x pilot savings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000376,"raw_usage":{"total_tokens":1987,"prompt_tokens":915,"completion_tokens":1072,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":963}},"tokens_in":531,"tokens_out":1072,"duration_ms":10229,"temperature":1.0,"reasoning_tokens":963,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:20:23.725470+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the same 802.11ax MIMO channel but increase the maximum delay spread beyond 72 taps (e.g., insert an 80th tap) or add a fractional delay; if the anti-aliasing reconstruction in Eq. (10) shows NMSE degradation beyond a few dB or the BER at SNR=38 dB with 4096-QAM no longer beats VAMP by the reported margin, the 8x pilot-reduction claim fails.","supporting_citations":[{"cited_title":"Channel estimation for wideband mmWave MIMO OFDM system exploiting block sparsity,","cited_arxiv_id":null,"evidence_quote":"Provides the VAMP sparse-recovery baseline that must be beaten in channel estimation and BER comparisons."},{"cited_title":"Cross-subcarrier precoder design for massive MIMO-OFDM downlink,","cited_arxiv_id":null,"evidence_quote":"Gives the S-WMMSE delay-domain precoder baseline against which the sparse precoder's BER gain is measured."},{"cited_title":"Robust WMMSE precoder with deep learning design for massive MIMO,","cited_arxiv_id":null,"evidence_quote":"Contributes the learned-Lagrange-network idea that makes the unrolled precoder solver trainable."},{"cited_title":"Simplified spatial correlation models for clustered MIMO channels with different array configura- tions,","cited_arxiv_id":null,"evidence_quote":"Defines the ray-tracing clustered channel model that yields the $D=72$ delay-domain sparsity."}],"review_version":1}