{"id":"552a692d-d1bb-401a-8914-714b3029ac56","arxiv_id":"2505.07717","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A constrained deep-unrolling channel estimator with a U-Net proximal operator and monotonic descent outperforms prior estimators in simulated wideband XL-MIMO ULA and UPA channels.","lead":"The paper presents a neural network that estimates wireless channels for extremely large antenna arrays by unrolling a standard iterative optimization algorithm and adding a constraint that each layer must improve the estimate. It reports large gains over existing methods in simulated wideband XL-MIMO settings with near-field, beam split, and spatial non-stationarity effects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The convergence guarantee in Theorem 2 is not established: the proof's step from Eq. (32) to Eq. (33) is invalid, so the claimed theoretical support is unsupported as written.","rationale":"The reader's formal weakest_assumption was the synthetic training/test distribution and contiguous visibility regions. That is a legitimate generalization caveat, but the paper's headline claim is explicitly about simulation results, so distribution shift is less central to whether the paper's internal argument is sound. The proof flaw in Theorem 2 is internal and directly affects a stated contribution: the constrained unrolled PGD network is claimed to have theoretical convergence guarantees, and Fig. 9 uses the monotonic-descent constraint to motivate improved intermediate-layer behavior. The reader did note in the rationale that 'the theoretical convergence proof contains a questionable inequality,' so there is partial agreement, but the reader did not make this the weakest assumption. Because the empirical claim may still be supported by the simulations, I would not change the conditional verdict; I would, however, require the authors to repair or clearly weaken the convergence claim before the theoretical contribution is accepted.","tokens_in":18945,"tokens_out":10050,"duration_ms":107640,"concrete_test":"Check whether Eq. (33) can be derived from Eq. (32) under Assumptions 1–5. A direct test: substitute the admissible abstract sequence Z_t = 0 into Eq. (32). Since 0 ≤ (1−δ)(1−ε)·0 + (1−δ)ζ + δC, the recurrence holds, but the expression in Eq. (33) equals −ζ − δC/(1−δ), so the claimed limit inferior fails. If the authors believe the cited result [52] supplies additional conditions that rule out such a sequence, those conditions must be stated and verified for the trained network; otherwise Theorem 2 must be revised or removed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim may be correct, but one of the paper's advertised contributions—the theoretical convergence guarantee—does not follow from the proof. In Appendix A, the proof obtains the recurrence Z_t ≤ (1−δ)(1−ε)Z_{t−1} + (1−δ)ζ + δC, labeled Eq. (32), with (1−δ)(1−ε) < 1. This inequality alone only bounds the limsup of Z_t. The next line claims, citing [52], that lim inf_t (ε Z_t − ζ − δC/(1−δ)) = 0, labeled Eq. (33). No derivation is shown, and the recurrence does not imply it: as an abstract sequence of nonnegative numbers, Z_t = 0 satisfies Eq. (32), since 0 ≤ (1−δ)ζ + δC, but makes the expression in Eq. (33) equal to −ζ − δC/(1−δ), not 0. Thus the asserted limit inferior is not a consequence of the stated bound. The subsequent use of Fatou's lemma and Eqs. (36)–(39) therefore does not establish Theorem 2. In addition, Eq. (30) conditions on event A_t while using an unconditional expectation of Z_{t−1}, and the bound C on ||h_t−h||^2 is not shown to follow from Assumption 1, which bounds the loss and constraint functions rather than the network outputs. Because the monotonic-descent constraint and its theoretical analysis are central to the paper's contribution, this gap matters even if the simulation comparisons are accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses channel estimation in wideband XL-MIMO systems under near-field propagation, beam split, and spatial non-stationarity. It formulates estimation as a maximum a posteriori problem, unrolls proximal gradient descent with learnable step sizes and a U-Net proximal operator, and imposes a monotonic descent constraint on layer-wise error. A primal-dual training algorithm is proposed, with Theorem 1 bounding the duality gap and Theorem 2 claiming convergence of the expected layer-wise distance to the true channel. Simulations under ULA and UPA configurations report that the proposed PGD-Net outperforms LMMSE, OMP, ISTA-Net+, AMP-SBL, and a same-architecture CNN (D2-CNN).","tokens_in":19262,"tokens_out":2619,"duration_ms":27469,"significance":"If the results hold, the paper makes a useful contribution to model-based deep learning for XL-MIMO channel estimation. The experimental design is internally consistent and stronger than most papers in this area because it includes a same-architecture CNN baseline (D2-CNN), which helps isolate the benefit of unrolling rather than the U-Net backbone. The channel model jointly capturing near-field, beam split, and spatial non-stationarity for both ULA and UPA is also a useful asset. However, the advertised theoretical convergence guarantee is not established by the proof as written: the key step in Appendix A is invalid, and a central inequality conditions on an event while using an unconditional expectation. This weakens the claim that the monotonic descent constraint is backed by theory. The empirical comparison remains plausible, but the theoretical gap is load-bearing because convergence analysis is listed as one of the paper's contributions. No code or data is provided, which limits reproducibility.","major_comments":[{"comment":"The proof of Theorem 2 does not establish the claimed limit inferior. Inequality (32) only gives a one-sided upper recurrence on Z_t, which implies at most a limsup bound. The assertion in Eq. (33) that lim inf_t (epsilon Z_t - zeta(J,delta) - delta C/(1-delta)) = 0 is not a consequence of (32); for example, the sequence Z_t = 0 satisfies (32) but makes the expression equal to -zeta - delta C/(1-delta). Without a derived lower bound or an additional argument, the application of Fatou's lemma and the subsequent display (39) do not follow. This gap invalidates Theorem 2 as stated, and the convergence analysis is a central advertised contribution of the paper.","section":"Appendix A, Eqs. (32)-(33)"},{"comment":"Inequality (30) conditions on the event A_t that the constraint in (26) is satisfied, but its right-hand side uses the unconditional expectation E[||h_{t-1}-h||^2]. Conditioning on A_t changes the distribution of h_{t-1}, so the inequality needs a conditional expectation or a separate argument controlling the conditional law. Additionally, Eq. (31) bounds the second term by a constant C said to follow from Assumption 1, but Assumption 1 bounds the loss and constraint functions, not the network outputs ||h_t||; boundedness of the layer outputs must be assumed or proved separately. These issues affect the derivation of recurrence (32) and therefore the proof of Theorem 2.","section":"Appendix A, Eq. (30)"},{"comment":"The empirical claim of general superiority is made for channels drawn from a single synthetic distribution in which each visibility region is a contiguous block selected uniformly at random and all training, validation, and test channels share the same SNR range and geometry statistics. If deployed channels have non-contiguous visibility regions, different path counts, or different SNR distributions, the reported gains are not guaranteed to transfer. The paper should either qualify the abstract claim as applying to this synthetic distribution or test robustness to at least one mismatched geometry, e.g., non-contiguous visibility masks or out-of-range SNR. This is not a fatal issue for the in-distribution comparison, but it is load-bearing for the broader claim of outperforming prior methods in wideband XL-MIMO.","section":"Section V-A and Section V-B"}],"minor_comments":[{"comment":"The legend and captions spell the baseline as 'ISAT-Net+' instead of 'ISTA-Net+'; this appears in both ULA and UPA panels and should be corrected.","section":"Figures 6 and 7"},{"comment":"The paper states that 'for each SNR level, 5e7 data pairs are generated' but does not specify the number of SNR levels or whether the four plotted SNR values (-5, 0, 5, 10 dB) are the only ones used. The total training data size should be stated unambiguously.","section":"Section V-A"},{"comment":"Several training hyperparameters are not reported: the number of epochs, total number of training batches, the primal and dual step sizes mu_w and mu_lambda in Algorithm 1, the contraction factor epsilon in Eq. (17), and the noise variance schedule sigma_t^2 in Eq. (19). These are needed to reproduce the results and to interpret the convergence experiments.","section":"Section V-A"},{"comment":"No code or data is made available, which is unusual for a heavily empirical paper of this type. A public implementation would materially strengthen confidence in the reported NMSE numbers.","section":"Section V-A and Table I"},{"comment":"The shaded confidence intervals are not defined. The paper should specify whether they are standard deviations or confidence bands over test realizations, and how many realizations were used.","section":"Figure 9"}],"recommendation":"major_revision","confidential_remarks":"The invalid step from Eq. (32) to Eq. (33) in Appendix A is the main technical problem. If the authors cannot repair the proof, they should consider removing or substantially downgrading Theorem 2 and reframing the monotonic descent constraint as a heuristic training regularizer whose benefit is shown only empirically. The empirical comparison itself is credible and would justify publication once the theoretical claims are corrected or removed. I would also gently encourage the editor to ask for code or at least detailed hyperparameter reporting, since the paper's central claim depends on reproducible training."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The empirical claim is plausible: constrained unrolled PGD with a U-Net proximal operator beats dictionary-based and deep baselines on simulated wideband XL-MIMO channels. The comparison is fair, includes a same-architecture CNN baseline, and the gains over AMP-SBL and ISTA-Net+ are around 3 dB in the UPA configuration. If reproducible, that is a useful incremental contribution for 6G receiver design. What is actually new: first application of constrained unrolling with monotone descent to wideband XL-MIMO with both ULA and UPA, jointly modeling near-field, beam split, and spatial non-stationarity. The delta over earlier unrolled PGD and Hadou et al. is modest but real. The soft spots are in the theory. The convergence proof in Appendix A has a genuine gap: recurrence (32) only bounds the limsup, but (33) claims the liminf of epsilon*Z_t - zeta - delta*C/(1-delta) is zero without derivation. That does not follow; a zero sequence satisfies the recurrence but makes the expression negative. So Theorem 2 is unsupported as written. That matters because the monotonic constraint and convergence analysis are advertised contributions. Lesser issues: no code or data released, the noise schedule is unreported, and the synthetic visibility regions are contiguous blocks, so real-world transfer is untested. The empirical central claim still holds up as plausible. I would send this to peer review. A serious referee should check the proof, and the authors should either fix it or downgrade the claim. I would also ask for code or at least full hyperparameters. My own verdict would be conditional. Not a theory reference, but if the numbers reproduce, a reasonable application reference.","headline":"Plausible empirical gains for wideband XL-MIMO channel estimation, but the advertised convergence theorem does not follow from the proof; still worth a serious referee.","tokens_in":743,"tokens_out":1084,"would_cite":false,"duration_ms":36808,"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 claims that a constrained unrolled proximal gradient descent network with a learned U-Net proximal mapping estimates wideband XL-MIMO channels more accurately than dictionary-based and deep baselines, reporting a 2.85 dB NMSE…","keywords":["XL-MIMO","channel estimation","deep unrolling","proximal gradient descent","near-field propagation","beam split","spatial non-stationarity","monotonic descent constraint"],"falsifier":"Running the same planar-array experiment at 0 dB signal-to-noise ratio with visibility regions drawn as multiple disjoint blocks instead of one contiguous rectangle would settle it: if the normalized mean-squared-error advantage over the AMP-SBL baseline falls below the reported 2.85 dB, the learned proximal mapping is fitted to the simulator's block geometry rather than to spatial non-stationarity in general.","tokens_in":18736,"feed_emoji":"📡","tokens_out":11529,"duration_ms":99063,"temperature":0.7,"pith_summary":"The paper tries to establish that channel estimation for wideband XL-MIMO systems—where near-field spherical wavefronts, frequency-dependent beam split, and spatially non-stationary visibility regions break conventional sparse representations—can be solved by unrolling proximal gradient descent into a learned network. The network replaces the handcrafted channel prior with a trainable U-Net (a multi-scale encoder-decoder network) proximal mapping and gives every layer its own learned step size, so no explicit dictionary or regularization function is needed. A monotonic descent constraint forces each layer's estimate to move closer to the true channel, and a primal-dual training method with decaying noise injection stabilizes convergence. If the claim holds, the approach supplies one estimation pipeline for both linear and planar arrays that beats LMMSE, OMP, ISTA-Net+, AMP-SBL, and D2-CNN in the simulated wideband XL-MIMO scenarios.","feed_headline":"Deep unrolling beats XL-MIMO channel estimation baselines","feed_subtitle":"No handcrafted dictionaries; the network learns the channel prior, with gains growing as SNR and antenna count rise.","key_machinery":"The load-bearing object is the unrolled proximal gradient descent network. Each layer first takes a gradient step on the data-fidelity term with a learnable step size $\\alpha_t$, then applies a trainable proximal mapping implemented as a U-Net—a multi-scale encoder-decoder with residual blocks and skip connections—so the unknown channel prior is learned rather than specified. The convergence mechanism is the monotonic descent constraint $\\|h_t-h\\|_2-(1-\\epsilon)\\|h_{t-1}-h\\|_2 \\leq 0$, enforced through a primal-dual update of a Lagrangian with one dual variable per layer, together with noise injection whose variance decays across layers. Theorem 2 uses this constraint to show that the expected distance to the true channel eventually enters a region whose size is controlled by the sample complexity, the failure probability, and the contraction factor $\\epsilon$.","core_discovery":"The central discovery is that the maximum a posteriori (MAP) formulation of wideband XL-MIMO channel estimation can be solved without constructing polar-domain dictionaries or explicit regularizers: a five-layer unrolled proximal gradient descent network, whose proximal operator is a U-Net with residual blocks and whose step sizes are learned per layer, implicitly learns the channel prior from data. Adding the monotonic descent constraint $\\|h_t-h\\|_2-(1-\\epsilon)\\|h_{t-1}-h\\|_2 \\leq 0$ and training with a primal-dual Lagrangian update yields layer-wise estimates that descend toward the true channel, with a convergence bound that shrinks as the number of training samples grows. In simulation, the network reports NMSE around $-12.61$ dB at 0 dB SNR under the UPA configuration, about 2.85 dB better than AMP-SBL and 3.39 dB better than ISTA-Net+, with larger margins over LMMSE and OMP and at higher SNR. The paper further reports that the method improves with antenna dimensionality and with pilot overhead, and that the constrained variant has more stable intermediate-layer convergence than the unconstrained one.","pith_inferences":["Beyond the paper, the same unrolled PGD skeleton could be applied to other ill-posed XL-MIMO inverse problems, such as beam training, localization, or integrated sensing and communication, since the proximal network removes the need to design a new dictionary for each channel structure.","The contiguous-block visibility model is the most restrictive simulation choice; a direct extension would test disjoint or statistically measured visibility regions to see whether the learned prior captures antenna-domain structure or only block geometry.","The convergence bound in Theorem 2 gives a sample-complexity handle: one could measure how the NMSE gap over AMP-SBL shrinks as the training set is reduced, isolating the value of the monotonic constraint from sheer data volume.","Another testable extension is cross-band transfer: train at 100 GHz and evaluate at 28 GHz or 300 GHz without retraining, which would reveal whether the learned proximal mapping is tied to the simulated carrier frequency or captures transferable near-field structure."],"forward_implications":["Under the UPA configuration at 0 dB SNR, the method reports NMSE around $-12.61$ dB, about 2.85 dB better than AMP-SBL and 3.39 dB better than ISTA-Net+, with the gap widening at 10 dB SNR.","The learned proximal mapping improves monotonically with pilot overhead from 256 to 1024 symbols under UPA, reaching $-16.63$ dB at 512 pilots.","A network trained only on three-path channels keeps most of its accuracy for two-path channels and degrades gracefully at five paths; joint training on two to five paths improves the five-path case by about 0.48 dB.","The monotonic descent constraint makes intermediate-layer NMSE decrease layer by layer, whereas the unconstrained network shows wider variation across layers.","Learned estimators exploit higher antenna dimensionality: the proposed method improves from about $-19.21$ dB at 10 dB SNR in the ULA to $-20.04$ dB in the UPA, while LMMSE degrades as the array grows."],"supporting_citations":[{"why":"Defines the proximal gradient descent algorithm whose iterations are unrolled into network layers.","marker":"[49]"},{"why":"Provides the constrained learning theory used to bound the primal-dual training gap in Theorem 1.","marker":"[51]"},{"why":"Supplies the stochastic-descent convergence argument that the proof of Theorem 2 builds on.","marker":"[52]"},{"why":"Provides the trainable proximal-gradient channel-estimation framework from which the proposed unrolled PGD network is developed.","marker":"[23]"},{"why":"Defines the enhanced residual-block architecture used in the U-Net proximal mapper.","marker":"[50]"},{"why":"Provides the AMP-SBL deep-unrolling baseline used as a comparison target in the wideband near-field experiments.","marker":"[45]"},{"why":"Provides the ISTA-Net+ optimization-inspired deep baseline used as a comparison target.","marker":"[53]"},{"why":"Provides the polar-domain dictionary on which the OMP baseline's sparse representation is built.","marker":"[25]"},{"why":"Gives the polar-domain sparse representation used alongside [25] for the OMP baseline.","marker":"[27]"}],"fun_headline_variants":["Unrolled network learns channel prior, no dictionaries needed","XL-MIMO estimation: deep unrolling without handcrafted priors","Unrolled proximal gradient learns XL-MIMO channel prior","Deep unrolling with learned step sizes beats AMP-SBL and ISTA","Constrained unrolled network improves wideband XL-MIMO estimation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that real channels resemble the synthetic training distribution, in particular that each propagation path is visible over one contiguous block of the array and that path counts, distances, and signal-to-noise ratios stay inside the ranges used for training.","fun_headline_variants_meta":{"raw":{"variants":["Unrolled network learns channel prior, no dictionaries needed","XL-MIMO estimation: deep unrolling without handcrafted priors","Unrolled proximal gradient learns XL-MIMO channel prior","Deep unrolling with learned step sizes beats AMP-SBL and ISTA","Constrained unrolled network improves wideband XL-MIMO estimation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000792,"raw_usage":{"total_tokens":3515,"prompt_tokens":995,"completion_tokens":2520,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":2447}},"tokens_in":611,"tokens_out":2520,"duration_ms":16237,"temperature":1.0,"reasoning_tokens":2447,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:09:41.771822+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Running the same planar-array experiment at 0 dB signal-to-noise ratio with visibility regions drawn as multiple disjoint blocks instead of one contiguous rectangle would settle it: if the normalized mean-squared-error advantage over the AMP-SBL baseline falls below the reported 2.85 dB, the learned proximal mapping is fitted to the simulator's block geometry rather than to spatial non-stationarity in general.","supporting_citations":[{"cited_title":"Proximal algorithms,","cited_arxiv_id":null,"evidence_quote":"Defines the proximal gradient descent algorithm whose iterations are unrolled into network layers."},{"cited_title":"Constrained learning with non-convex losses,","cited_arxiv_id":null,"evidence_quote":"Provides the constrained learning theory used to bound the primal-dual training gap in Theorem 1."},{"cited_title":"Robust st ochastically- descending unrolled networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the stochastic-descent convergence argument that the proof of Theorem 2 builds on."},{"cited_title":"Trainable proximal gradi ent descent- based channel estimation for mmWave massive MIMO systems,","cited_arxiv_id":null,"evidence_quote":"Provides the trainable proximal-gradient channel-estimation framework from which the proposed unrolled PGD network is developed."},{"cited_title":"Enhanced dee p residual networks for single image super-resolution,","cited_arxiv_id":null,"evidence_quote":"Defines the enhanced residual-block architecture used in the U-Net proximal mapper."},{"cited_title":"Deep unfolding based chann el estimation for wideband terahertz near-field massive MIMO systems,","cited_arxiv_id":null,"evidence_quote":"Provides the AMP-SBL deep-unrolling baseline used as a comparison target in the wideband near-field experiments."},{"cited_title":"ISTA-Net: Interpretable optim ization-inspired deep network for image compressive sensing,","cited_arxiv_id":null,"evidence_quote":"Provides the ISTA-Net+ optimization-inspired deep baseline used as a comparison target."},{"cited_title":"Channel estimation for extremely lar ge-scale MIMO: Far-field or near-field?","cited_arxiv_id":null,"evidence_quote":"Provides the polar-domain dictionary on which the OMP baseline's sparse representation is built."},{"cited_title":"Multiple access for near-field communi cations: SDMA or LDMA?","cited_arxiv_id":null,"evidence_quote":"Gives the polar-domain sparse representation used alongside [25] for the OMP baseline."}],"review_version":1}