{"id":"d8e710f3-1dbf-4bb0-acde-29ae36627686","arxiv_id":"1908.02884","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"BEACHES estimates mmWave massive MIMO channels by soft-thresholding the DFT beamspace channel, selecting the threshold automatically via Stein's unbiased risk estimate to minimize mean squared error in the large-antenna limit.","lead":"BEACHES proposes a low-complexity method to estimate millimeter-wave channels by transforming them into a sparse beamspace domain and denoising with an automatically tuned soft threshold. The threshold is chosen using Stein's unbiased risk estimate, and the method matches far more complex algorithms in simulated bit error rate while running orders of magnitude faster.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof of Theorem 2 invokes an i.i.d. CAMP theorem for correlated DFT beamspace entries, and even a repaired pointwise SURE convergence would not by itself justify the argmin claim; the advertised large-antenna optimality is unproven as written.","rationale":"The paper's central contribution is the provable large-antenna MSE optimality of the SURE-based threshold selection. That claim is not supported by the proof in Appendix B: the cited CAMP result requires independence and a fixed limiting distribution, neither of which is established for the DFT beamspace channel of (1). This is a genuine proof gap, and it is in the same area the reader flagged. I would not turn this into a rejection: the conclusion of Theorem 2 is plausibly true, because SURE is exactly unbiased for each fixed h and the summands are bounded, so a direct variance argument can likely repair pointwise convergence. The larger missing piece is that the paper never proves the uniform or argmin convergence needed to move from 'SURE_tau converges to MSE_tau for each tau' to 'the tau chosen by Algorithm 1 minimizes MSE in the limit.' That is why the result should remain CONDITIONAL: the method is well specified and simulation evidence is encouraging, but the advertised guarantee is not proven. The reader's weakest assumption captures the same proof gap, though I would broaden it to include the argmin step; hence partial agreement. A careful re-derivation plus the numerical convergence test would settle whether the concern is cosmetic or substantive.","tokens_in":9449,"tokens_out":33788,"duration_ms":380756,"concrete_test":"Test the convergence claim numerically: simulate (1) with B = 2^8, 2^10, 2^12, 2^14, L = 5, E0 = 1, random gains and angles; for fixed tau in {0.5, 1, 2}, compute SURE_tau from Eq. (6) on single noisy realizations and empirical MSE averaged over 10^4 noise draws, and check whether |SURE_tau - MSE_tau| tends to 0 as B grows. Independently, attempt to re-derive Eq. (7) from Theorem 1 without invoking [20] and verify the additional argmin/uniform-convergence step; if the re-derivation requires an assumption absent from the paper, the proof remains incomplete even if the numerics pass.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that BEACHES provably minimizes MSE in the large-antenna limit. This rests on Theorem 2 (Appendix B), whose proof applies [20, Thm. III.15 and III.16] through Eq. (14). That theorem is a state-evolution result for sequences with independent entries and a fixed limiting distribution H; no such H is defined or verified for the beamspace channel in (1). The entries of h_hat = F h are DFT coefficients of L complex sinusoids: they are correlated, and their magnitudes range from O(sqrt(B)) down to O(1/sqrt(B)), so the empirical measure does not converge to a fixed random variable H in the sense required by (14). In addition, the derivative term in SURE is discontinuous at |y_b| = tau, so it is not pseudo-Lipschitz, another unverified condition. Finally, Theorem 2 as stated gives pointwise convergence in tau; the algorithm selects tau by minimizing SURE, and the paper does not prove uniform convergence or argmin consistency. The consequence is that the main theoretical guarantee is not established as written, even though a direct variance argument may repair pointwise convergence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BEACHES, a beamspace channel estimation method for massive MIMO mmWave/THz uplink systems. The channel vector is modeled as a superposition of L complex sinusoids in the antenna domain; after taking a DFT, the paper denoises the beamspace vector by complex soft-thresholding with threshold τ. To avoid manual tuning, BEACHES computes Stein's unbiased risk estimate (SURE) for the soft-thresholding estimator and selects τ by minimizing the SURE expression over intervals defined by the sorted magnitudes of the noisy beamspace observations. Theorem 1 proves the standard SURE identity in the complex domain, and Theorem 2 claims that SURE converges to MSE as B→∞, so the chosen threshold minimizes MSE in the large-antenna limit. Simulations with a QuaDRiGa 60 GHz model for LoS and non-LoS channels show BER comparable to ANM and NOMP at orders-of-magnitude lower runtime.","tokens_in":9709,"tokens_out":6722,"duration_ms":69017,"significance":"If Theorem 2 were rigorously established, the paper would provide a practical, low-complexity denoiser with automatic parameter selection for a relevant wireless setting; the runtime comparison in Table I and the BER results in Fig. 2 are concrete and favorable. The derivation of Theorem 1 in Appendix A is a correct application of Stein's identity, and the algorithmic idea of minimizing SURE over sorted intervals is clean and computationally attractive. The weakness is that the central asymptotic optimality claim rests on an unproved application of a CAMP state-evolution theorem to correlated beamspace entries, so the paper's headline theoretical guarantee is not currently supported as written.","major_comments":[{"comment":"The proof of Theorem 2 applies [20, Thm. III.15 and III.16], a state-evolution result for complex approximate message passing, to the beamspace entries ```\\hat h_b```. Those theorems require the entries of the observed vector to be independent (or to satisfy a specific exchangeability/state-evolution condition) and the empirical distribution of the true entries to converge to a fixed random variable H. Under the channel model (1), ```\\hat h = F h``` is the DFT of a superposition of L complex sinusoids: the entries are correlated, and their magnitudes range from O(√B) down to O(1/√B) depending on whether a path is on the DFT grid. No limiting distribution H is defined for this model and none of the theorem's conditions are verified. Therefore Eq. (14) is not justified, and the convergence claim (7) is not established by the manuscript's argument.","section":"Appendix B, Eq. (14)"},{"comment":"The transition to Eq. (16) treats the sum of the derivative terms ```\\partial[\\eta_R]_b/\\partial[\\hat y_R]_b + \\partial[\\eta_I]_b/\\partial[\\hat y_I]_b``` as a pseudo-Lipschitz function and applies (14). These derivative terms are not pseudo-Lipschitz: they are discontinuous at |```\\hat y_b| = \\tau```, and SURE is undefined at those points, as the paper itself states after Eq. (6). The proof neither checks the pseudo-Lipschitz condition nor explains how the measure-zero discontinuity is handled in the limit interchange. This is a second unverified condition in the proof of Theorem 2.","section":"Appendix B, Eq. (16)"},{"comment":"Even if the pointwise limit ```\\lim_{B→∞} SURE_\\tau = MSE``` held for each fixed ```\\tau```, the paper's advertised conclusion that BEACHES 'provably minimizes the MSE in the large-antenna limit' would not follow. The algorithm selects ```\\tau^*``` by minimizing the empirical SURE function over the intervals defined by the sorted magnitudes ```|\\hat y_b|``` (Section III-C and Algorithm 1). The paper does not prove uniform convergence of ```SURE_\\tau``` to MSE or an argmin-consistency result, so convergence of the selected threshold to the MSE-minimizing threshold is not established. The central theoretical claim therefore needs an additional, nontrivial argument or a revised, weaker statement.","section":"Theorem 2 and Section III-C"}],"minor_comments":[{"comment":"The update ```V = \\sum_{k=1}^{B}(|\\hat y^s_k|)^{-1}``` is undefined if any sorted magnitude is zero. For Gaussian noise this has probability zero, but the pseudocode should state the convention (e.g., omit zero entries or replace them by a small positive value).","section":"Algorithm 1, line 4"},{"comment":"The pseudocode evaluates SURE_```\\tau``` at the interval boundaries ```\\hat y^s_{k-1}``` and ```\\hat y^s_k```, where the SURE expression is undefined, while the text says the values should be computed arbitrarily close to these boundaries. This inconsistency should be resolved either by defining the one-sided limits in Eq. (8) or by rewriting the loop to evaluate the limits explicitly.","section":"Algorithm 1, lines 5–7 and footnote 2"},{"comment":"The expression ```E_{\\hat y}[ B^{-1}\\|\\eta(\\hat y,\\tau)-\\hat y\\|^2 ]``` uses the symbol ```\\hat y``` for both the random vector and the fixed vector inside the norm; this is confusing and should be rewritten with a generic random vector such as ```\\tilde y```.","section":"Appendix B, Eq. (15)"},{"comment":"The phrase 'H is a random variable with the sparse distribution of the channel vector in the beamspace domain' is not a definition. The distribution of ```\\hat h_b``` depends on B through the DFT of L sinusoids; without an explicit asymptotic model for L and the angles ```\\Omega_\\ell```, the statement is not precise.","section":"Appendix B, after Eq. (14)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely of interest to the signal-processing community, but the advertised theoretical guarantee is currently unsupported. The issue appears fixable through a direct variance-based proof of pointwise convergence and an explicit uniform/argmin argument, or by softening the claims to 'SURE-based threshold selection' without large-antenna optimality. I recommend major revision rather than rejection because the algorithmic contribution and simulations are solid."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on BEACHES. The real contribution is the simple O(B log B) exact search over soft-threshold parameters: the authors derive a complex-domain SURE for the soft-thresholding operator, then sort the magnitudes of the beamspace observations and evaluate the SURE quadratic on each interval in linear time. That is a new, clean, and practical result. For a mmWave channel estimator, it gives a parameter-free denoiser that seems to match ANM and NOMP in the simulations at orders-of-magnitude lower complexity. That is worth something.\n\nThe theory is where I part ways. Theorem 1 (unbiased SURE in the complex domain) is a standard integration-by-parts argument and appears correct. Theorem 2, which claims SURE converges to MSE as B approaches infinity, is not proven as written. The proof invokes [20, Thm. III.15-16], a CAMP state-evolution result for sequences with i.i.d. entries, and applies it to the beamspace entries of a superposition of L sinusoids. Those entries are not i.i.d. (their magnitudes range from O(sqrt(B)) down to O(1/sqrt(B))), and no limiting distribution H is defined or verified. The derivative term in SURE is also discontinuous at |y_b| = tau, so it is not pseudo-Lipschitz; the paper says it is bounded, which is true away from the threshold but not in a neighborhood of it. Even if pointwise convergence were repaired, that alone would not justify the argmin claim: the algorithm selects tau by minimizing SURE, and the paper does not prove uniform convergence or argmin consistency. So the headline claim that BEACHES 'provably minimizes MSE in the large-antenna limit' is not established.\n\nThe saving grace is that the algorithm itself does not depend on the theorem. The efficient threshold search is an exact minimization of the SURE expression, and Theorem 1 already gives that SURE is unbiased for any fixed B. The practical value stands on those two legs plus the simulations. The simulations look plausible, though they lack error bars and code/data; that is a minor issue for a methods paper, not a fatal one. The boundary handling in Algorithm 1 is also under-specified (the footnote acknowledges the issue but the pseudocode does not implement the epsilon perturbation). Citations to the SURE and CAMP literature are appropriate.\n\nWho should read this? Anyone building a low-complexity mmWave channel estimator or looking for a baseline to compare against. It deserves a serious referee, but I would ask the authors to fix or soften the asymptotic claim, either by adding a proper state-evolution argument for the beamspace model or by stating SURE minimization as a design heuristic. This is not a takedown; it is a useful algorithm with an overreaching proof.","headline":"A genuinely useful O(B log B) SURE-tuned beamspace denoiser with a real proof gap: the advertised large-antenna optimality is not established as written.","tokens_in":10231,"tokens_out":2798,"would_cite":true,"duration_ms":32231,"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":"BEACHES shows that mmWave channel estimation can be reduced to soft-thresholding a DFT beamspace vector, with a SURE-based threshold that is provably MSE-optimal as the antenna count grows.","keywords":["beamspace channel estimation","massive MIMO","mmWave communications","terahertz communications","Stein's unbiased risk estimator","soft-thresholding","sparse channel estimation","low-complexity signal processing"],"falsifier":"Generate beamspace channels from the model with a small number of paths at finite array sizes such as $B=64$ and $B=256$, add noise at SNRs from -10 to 15 dB, and compare the $\\tau$ that minimizes SURE (6) with the $\\tau$ that minimizes the true MSE; a systematic gap between the two minimizers, or a loss of BER parity with ANM and NOMP when the beamspace entries are strongly correlated, would show that the large-antenna optimality claim does not transfer to finite arrays.","tokens_in":9265,"feed_emoji":"📡","tokens_out":7201,"duration_ms":69473,"temperature":0.7,"pith_summary":"This paper proposes BEACHES, a channel-estimation method for millimeter-wave (mmWave) and terahertz multi-antenna uplinks that treats the problem as denoising in the beamspace domain. Because high-frequency channels are a small superposition of directional paths, their DFT (beamspace) representation is approximately sparse, so a soft-thresholding step can suppress noise. The paper's central move is to set the threshold automatically with Stein's unbiased risk estimate (SURE), and it proves that in the limit of infinitely many antennas the SURE curve coincides with the true mean-square error, so the threshold that minimizes SURE is the MSE-optimal threshold. In simulations for line-of-sight and non-line-of-sight 60 GHz channels with 256 antennas, BEACHES matches the bit-error rate of atomic norm minimization and Newtonized orthogonal matching pursuit while running orders of magnitude faster. The upshot is that the expensive off-grid machinery may be unnecessary for practical mmWave channel denoising.","feed_headline":"Simple beamspace denoising rivals costlier mmWave estimators","feed_subtitle":"A SURE-tuned threshold matches atomic-norm and Newtonized-OMP bit error rates with orders-of-magnitude less compute.","key_machinery":"The load-bearing object is Stein's unbiased risk estimator for complex soft-thresholding, specialized to the beamspace channel model. For a weakly differentiable estimator $\\mu(\\hat{\\mathbf y})$ of a vector observed in complex Gaussian noise, SURE (5) is an unbiased estimate of MSE; for $\\mu=\\eta(\\cdot,\\tau)$ it reduces to expression (6), a piecewise quadratic function of $\\tau$ with jumps at the observed magnitudes $|\\hat y_b|$. BEACHES sorts the $|\\hat y_b|$ so the pieces are intervals between consecutive sorted magnitudes, evaluates the quadratic minimizer in each interval in constant time using running sums, and picks the $\\tau$ with smallest SURE. This reduces the parameter-selection problem to an $O(B\\log B)$ sweep, consisting of an FFT, a sort, and an inverse FFT, which is what makes the method cheap enough for practical hardware.","core_discovery":"BEACHES estimates the uplink channel vector by taking the DFT of the noisy antenna-domain observation, applying the complex soft-thresholding operator $\\eta(\\hat{\\mathbf y},\\tau)$ entry by entry, and transforming back. The parameter $\\tau$ is chosen by a sweep over the sorted magnitudes of the beamspace vector that minimizes the SURE expression (6), which depends only on the observations, the noise variance $E_0$, and $\\tau$. Theorem 2 states that $\\lim_{B\\to\\infty}\\mathrm{SURE}_\\tau = \\mathrm{MSE}$, so in the large-antenna regime minimizing SURE provably minimizes the channel-estimation MSE. The paper's simulations show that this simple procedure delivers uncoded bit-error rates on par with the off-grid methods ANM and NOMP, and 2-3 dB better than maximum-likelihood estimation at a bit-error rate of $10^{-2}$, while the runtime is lower by factors of about 120-1500 versus NOMP and 29,000-58,000 versus ANM.","pith_inferences":["The same SURE-sweep idea transfers to other element-wise denoisers, such as hard thresholding or nonnegative shrinkage, where closed-form SURE expressions exist; those variants would inherit the same $O(B\\log B)$ tuning cost.","One could test BEACHES as a front-end for hybrid or low-resolution-ADC architectures, where the paper's assumption of all-digital antennas and unquantized observations would need modification.","Because the algorithm only uses sorted magnitudes of the beamspace coefficients, it may be robust to phase errors or to channels with many weak paths where sparsity is less pronounced, though the paper does not explore this."],"forward_implications":["If the asymptotic SURE-MSE equivalence holds at operating array sizes, BEACHES removes the need for manual threshold tuning in beamspace denoising, since the data alone determine $\\tau$.","Off-the-grid methods such as ANM and NOMP are not needed for competitive bit-error rate in the simulated LoS and non-LoS 60 GHz scenarios, so mmWave channel estimation can be implemented with FFT, sorting, and scalar operations.","The computational savings grow with scenario difficulty: the paper reports speedups up to 1,500x over NOMP and 58,000x over ANM in the non-LoS case.","Because the SURE expression only needs the noise variance $E_0$ and the beamspace observation magnitudes, the method adapts automatically to changing propagation conditions and numbers of paths.","The same SURE-based thresholding logic can apply to any approximately sparse beamspace vector, making the approach a candidate for terahertz and other high-dimensional array systems."],"supporting_citations":[{"why":"Supplies the SURE principle and the soft-thresholding estimator that BEACHES uses as its denoiser.","marker":"[18]"},{"why":"Provides the complex-LASSO asymptotic convergence theorem invoked in the proof that SURE converges to MSE.","marker":"[20]"},{"why":"Provides the real-domain SURE convergence lemma that the complex-domain proof adapts.","marker":"[23]"},{"why":"Suggests the efficient interval-search strategy for minimizing a SURE-like expression, which BEACHES adapts.","marker":"[21]"},{"why":"Provides atomic norm minimization, the off-grid baseline whose bit-error rate BEACHES matches.","marker":"[14]"},{"why":"Provides Newtonized orthogonal matching pursuit, the second baseline and the main complexity comparison target.","marker":"[15]"},{"why":"Provides the 60 GHz QuaDRiGa channel model used to generate the LoS and non-LoS simulation channels.","marker":"[22]"}],"fun_headline_variants":["Beamspace soft-thresholding matches complex mmWave estimators","SURE-tuned threshold speeds mmWave channel estimation","Simple DFT denoising rivals costlier mmWave channel estimators","Low-complexity beamspace method matches ANM and NOMP","One transform and a threshold: fast, accurate mmWave estimation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof that SURE converges to the MSE treats the beamspace channel entries as independent and identically distributed draws from a fixed sparse distribution, even though the model represents them as a DFT of a few sinusoidal paths, which makes the entries correlated; the paper does not verify that the convergence theorem's conditions hold for this correlated model.","fun_headline_variants_meta":{"raw":{"variants":["Beamspace soft-thresholding matches complex mmWave estimators","SURE-tuned threshold speeds mmWave channel estimation","Simple DFT denoising rivals costlier mmWave channel estimators","Low-complexity beamspace method matches ANM and NOMP","One transform and a threshold: fast, accurate mmWave estimation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000307,"raw_usage":{"total_tokens":1758,"prompt_tokens":950,"completion_tokens":808,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":723}},"tokens_in":566,"tokens_out":808,"duration_ms":9214,"temperature":1.0,"reasoning_tokens":723,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:30:41.521402+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate beamspace channels from the model with a small number of paths at finite array sizes such as $B=64$ and $B=256$, add noise at SNRs from -10 to 15 dB, and compare the $\\tau$ that minimizes SURE (6) with the $\\tau$ that minimizes the true MSE; a systematic gap between the two minimizers, or a loss of BER parity with ANM and NOMP when the beamspace entries are strongly correlated, would show that the large-antenna optimality claim does not transfer to finite arrays.","supporting_citations":[{"cited_title":"Adapting to unknown smoothness via wavelet shrinkage,","cited_arxiv_id":null,"evidence_quote":"Supplies the SURE principle and the soft-thresholding estimator that BEACHES uses as its denoiser."},{"cited_title":"Asymptotic analysis of complex LASSO via complex approximate message passing (CAMP),","cited_arxiv_id":null,"evidence_quote":"Provides the complex-LASSO asymptotic convergence theorem invoked in the proof that SURE converges to MSE."},{"cited_title":"Consistent Parameter Estimation for LASSO and Approximate Message Passing","cited_arxiv_id":"1511.01017","evidence_quote":"Provides the real-domain SURE convergence lemma that the complex-domain proof adapts."},{"cited_title":"Parameterless Optimal Approximate Message Passing","cited_arxiv_id":"1311.0035","evidence_quote":"Suggests the efficient interval-search strategy for minimizing a SURE-like expression, which BEACHES adapts."},{"cited_title":"Atomic norm denoising with applications to line spectral estimation,","cited_arxiv_id":null,"evidence_quote":"Provides atomic norm minimization, the off-grid baseline whose bit-error rate BEACHES matches."},{"cited_title":"Newtonized orthog- onal matching pursuit: Frequency estimation over the continuum,","cited_arxiv_id":null,"evidence_quote":"Provides Newtonized orthogonal matching pursuit, the second baseline and the main complexity comparison target."},{"cited_title":"QuaDRiGa - Quasi Deterministic Radio Channel Generator User Manual and Documentation,","cited_arxiv_id":null,"evidence_quote":"Provides the 60 GHz QuaDRiGa channel model used to generate the LoS and non-LoS simulation channels."}],"review_version":1}