{"id":"a4e4a0b0-3973-4222-9c15-58db7d8a4520","arxiv_id":"1908.07729","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A penalized atomic norm minimization estimator is formulated and solved through its dual to simultaneously estimate sparse OFDM channels and locate impulsive noise.","lead":"This paper presents a method for estimating OFDM wireless channels when some received measurements are corrupted by sudden high-power impulsive noise, by solving one convex optimization problem that finds both the channel and the corrupted locations. It matters because a single optimization step could replace the ad hoc blanking devices used in current receivers and avoid the grid mismatch of compressed-sensing channel estimation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Even-P index set in SDP (8) is dimensionally inconsistent for every experimental configuration; the dual equivalence fails for P=64 and P=128.","rationale":"The paper proposes a plausible extension of atomic norm minimization to impulsive-noise-robust OFDM channel estimation, and the dual-based support localization idea is standard. However, the central claim depends on (8) being exactly the dual of (6). The even-P index set breaks this for all experiments in the paper: a P−1-element J cannot index a P-dimensional observation, so one pilot measurement is dropped from the dual polynomial and from the Toeplitz constraints. This is not a cosmetic notation issue; it means the SDP being solved is not the stated optimization problem. The reader's weakest assumption identified exactly this point, and I agree with that assessment. The error is concrete and likely fixable by using a proper P-element index set and rerunning the experiments, but as written the central claim is not substantiated. No additional manufactured concerns are needed; the dimensional inconsistency is sufficient to sustain the reader's REJECT verdict.","tokens_in":7403,"tokens_out":34343,"duration_ms":293372,"concrete_test":"Analytical check: re-derive (8) from (7) for P=4 (even). Write q=(q(-1),q(0),q(1),q(2)) and J={-1,0,1}; show that Q(f) in (9) does not depend on q(2), while the SDP still optimizes over q(2). Then compare the feasible sets: the correct Lagrangian dual of (6) (via the SDP characterization of the atomic norm in [8]) requires the trace constraints tr[Θ_k Q0]=δ_k for all k=−(P−1),...,P−1, i.e., seven constraints for P=4, not the three used in (8). If the two SDPs differ, (8) is not equivalent to (7). Computational confirmation: run (8) for P=64 with J={-31,...,31} and with the corrected J={-32,...,31} (or {0,...,63}); check whether the phase-transition grid in Fig. 2 changes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section IV, for even P the paper sets m=P/2−1 and J={-m,...,m}, so |J|=P−1. Yet the observation vector y, the primal variables h,w,z, and the declared dual variable q are all P-dimensional (Section III: y,h,w,z∈C^{P×1}; q∈C^{P×1} in (8)). The trace constraints tr[Θ_k Q0]=δ_k run only over k∈J, and the dual polynomial in (9) sums only over k=−m,...,m, so for P=64 and P=128—the only sizes tested—one coordinate of the measurement vector does not enter the atomic-norm constraint or the polynomial whose unit-modulus points define the channel estimate. There is no bijection between {0,...,P−1} and a P−1-element J, so the assertion 'without loss of generality' is false. Consequently, the SDP (8) is not the exact dual of (7) in the tested regime, and the claimed simultaneous recovery of frequencies from |Q(f)|=1 and impulse support from |q|=λ is not established for the experiments. For odd P the symmetric index set has the right cardinality, so this is a fixable indexing/constraint error rather than a dead end; but as written the central claim fails precisely where it is tested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a penalized atomic norm minimization (PANM) estimator for pilot-aided OFDM channel estimation when the received pilots are corrupted by both Gaussian and impulsive noise. The authors formulate a convex primal problem (6), derive its dual (7), and then convert the dual into a semidefinite program (8) using positive trigonometric polynomial theory. They propose to estimate the continuous multipath delays from the points where the dual polynomial Q(f) satisfies |Q(f)|=1 and to localize impulsive noise from the entries of the dual solution with |q|=lambda. The numerical section reports simultaneous frequency and impulse localization, phase-transition diagrams for P=64 and P=128, and an MSE comparison against penalized l1 minimization.","tokens_in":7622,"tokens_out":17333,"duration_ms":344886,"significance":"If the SDP equivalence were correct, the paper would offer a gridless approach to jointly estimate a sparse continuous-domain channel and an impulsive-noise support, avoiding basis mismatch and bypassing the need for detailed impulsive-noise power modeling. The experimental design is reasonable in spirit: ground-truth MSE is used, phase transitions are averaged over 100 trials, and an external PLM benchmark is included. The main claimed contribution, however, rests entirely on the exactness of the SDP reformulation in Eq. (8), and that reformulation contains a load-bearing indexing error. As written, the estimator tested in Section V is not the exact dual of the stated PANM. No recovery theorem is provided; the contribution is algorithmic and experimental, and it is incremental relative to Refs. [8] and [9].","major_comments":[{"comment":"The SDP (8) is not equivalent to the dual problem (7) as written. After the 'without loss of generality' reindexing J={-m,...,m}, the trace constraints tr[Theta_k Q0]=delta_k are imposed only for k in J, but in the PTP/Gram-matrix parametrization the index k is a Toeplitz lag, not a vector-entry index. For a P-by-P matrix the lags range over {-(P-1),...,P-1}; constraining only k with |k|<=m leaves, for example, the lags k=+(m+1),...,+(P-1) and their negatives unconstrained, and no argument shows these are redundant. For even P the mismatch is worse: with m=P/2-1 the set J has P-1 elements, while y and q are P-dimensional, so the constraint system and the dual polynomial in (9) omit one measurement coordinate. Since all experiments in Section V use P=64 and P=128, the estimator that was implemented and tested is not the dual of the stated PANM. The reindexing should be applied only to the atomic vectors, while the trace constraints should run over the lags 0,...,P-1, or equivalently over the full difference set of the reindexed vector positions. This is a fixable formulation error, but it is load-bearing because it invalidates the empirical claims as stated.","section":"IV, Eqs. (8) and (9)"},{"comment":"There is a dimensional inconsistency in the noise constraint. The paper denotes the noise variance by sigma_np^2 in Section II and uses that same symbol in the SNR definition SNR=10 log10(1/sigma_np^2) in Section V; Eq. (6) uses the right-hand side sigma_np^2. The dual objective in Eq. (7), however, contains sigma_np ||q||_2. If the primal constraint is ||y-h~-z~||_2 <= sigma_np^2, the dual should contain sigma_np^2 ||q||_2; if the intended primal constraint is ||y-h~-z~||_2 <= sigma_np, then Eq. (6) should be changed. This discrepancy changes the numerical estimator and all simulation results that depend on the noise-level parameter, so it must be corrected and the experiments rerun.","section":"III Eq. (6) and IV Eq. (7)"},{"comment":"The paper does not specify how the frequencies are extracted from the dual polynomial Q(f) in the noisy experiments. The stated rule 'find values of f for which |Q(f)|=1' is an exact condition; under Gaussian noise the estimated dual polynomial does not necessarily attain the value 1, so a thresholding or peak-picking rule is needed. Without specifying that rule, the simulation results in Figs. 1-3 cannot be reproduced, and the reported phase-transition success rates are unverifiable. The authors should state the exact detection criterion, for example local maxima of |Q(f)| above a threshold or interpolation to |Q(f)|=1, and describe how the number of detected sources is determined.","section":"V, Experiments; IV recovery procedure"}],"minor_comments":[{"comment":"The subscript notation y_n(n) appears to be a typo; the received signal should be denoted simply y(n) or consistent with the later vector notation.","section":"II, Eq. (2)"},{"comment":"The symbol J is overloaded: it is defined as {0,...,P-1} in Section II and then redefined as {-m,...,m} in Section IV. The paper should either use distinct names or explicitly give the bijection between the two index sets, especially because the reindexing is central to the SDP formulation.","section":"II and IV, notation"},{"comment":"The paper asserts that Slater's condition holds for the primal-dual pair, but gives no justification; since sigma_np>0 in all experiments the condition is likely satisfied, yet a brief verification would make the duality argument rigorous.","section":"IV, Eq. (7)"},{"comment":"There are small typos: 'forall i={1,...,s}' should be 'forall k in {1,...,s}', and 'The gain of each frequencies can be calculated by (IV)' should refer to the least-squares procedure described at the end of Section IV rather than to an equation.","section":"V, first experiment text"},{"comment":"The regularization parameter lambda=0.1 is fixed without sensitivity analysis. The phase-transition and MSE results can depend strongly on lambda, so the paper should include at least a brief study of the influence of lambda or a justification for the chosen value.","section":"V, parameter choice"},{"comment":"The statement that solving (8) with CVX has complexity O(P^3) is not supported; an interior-point solver for a P-by-P variable SDP with P constraints typically has complexity far exceeding a single O(P^3) bound. The complexity claim should be corrected or removed.","section":"VI, complexity claim"}],"recommendation":"major_revision","confidential_remarks":"The indexing error in Section IV is serious enough that the current numerical results do not validate the claimed estimator, but it is a correctable formulation error rather than a dead end. If the authors reformulate the SDP with the correct lag-index set, re-run the experiments, and fix the noise-level inconsistency, the paper could become publishable as an algorithmic contribution. The lack of any recovery guarantee is a limitation that should be stated clearly, but it need not be fatal if the empirical claims are made reproducible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read on arXiv:1908.07729. The paper proposes a convex SDP that jointly estimates an OFDM channel on a continuous dictionary and localizes impulsive noise, by adding an l1 penalty to the atomic norm and working with the dual. That combination is a natural but real extension: prior ANM for OFDM didn't handle outliers, and prior l1 outlier methods were grid-based. The dual-polynomial recovery and the phase transition experiments are the right tools. If the estimator worked as stated, it would be a genuinely useful receiver-side tool.\n\nBut there's a load-bearing indexing bug. Section IV redefines J={-m,...,m} with m=P/2-1 for even P, which gives |J|=P-1. All experiments use P=64 and 128, so every run solves an SDP with a measurement vector that has one coordinate not entering the atomic-norm constraint or the dual polynomial. The 'without loss of generality' claim is false for even P; you can't index P measurements with a symmetric set of P-1 integers. The consequence is that the extra coordinate is effectively an unconstrained dual variable, bounded only by λ, so the optimizer can always declare it an impulsive-noise location. That's not the estimator the authors describe, and it undermines both the channel recovery and the noise localization claims as tested. This is fixable (use J={-P/2+1,...,P/2} or just {0,...,P-1}), so it's a major-revision problem, not a dead end.\n\nOther soft spots are minor by comparison: λ=0.1 is fixed with no sensitivity study; no error bars on the MSE curve; no code; the run-time claim O(P^3) is stated without substantiation. The citation pattern looks fine. No circular reasoning.\n\nBottom line: the paper is a reasonable idea with a specific, fixable correctness flaw in its central SDP for every experimental configuration. If the authors correct the indexing and add a sensitivity analysis, it's worth a second look. As is, I would not rely on the empirical claims. For peer review, I'd send it out rather than desk reject—the flaw is clear enough that a good referee can give the authors a concrete path to fix it.","headline":"Useful extension of atomic norm to impulsive noise, but the advertised SDP is dimensionally inconsistent for every P tested (even P), so the empirical claims rest on a fixable indexing error.","tokens_in":8205,"tokens_out":7221,"would_cite":false,"duration_ms":564788,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A12","90C22"],"pacs":[],"model":"deepseek-v4-flash","headline":"One convex estimator recovers both sparse OFDM channel delays and impulsive-noise locations from a single dual problem.","keywords":["OFDM","impulsive noise","atomic norm minimization","semidefinite programming","dual polynomial","continuous sparse recovery","pilot-aided channel estimation","super-resolution"],"falsifier":"Solve the SDP (8) for $P = 64$ with a single true frequency $f = 0.37$ and no impulsive noise, then evaluate the dual polynomial: the true $f$ should be a contact point $|Q(f)| = 1$. If the even-$P$ reindexing $J = \\{-m, \\dots, m\\}$ is not equivalent to the original $0, \\dots, P-1$ indexing, the contact point will be displaced or missing.","tokens_in":7161,"feed_emoji":"📶","tokens_out":5835,"duration_ms":57200,"temperature":0.7,"pith_summary":"The paper proposes a single convex program that cleans a pilot-aided OFDM measurement corrupted by both Gaussian noise and sporadic high-amplitude impulses. The estimator minimizes a penalized atomic norm that promotes a sparse channel together with an $\\ell^1$ penalty on impulsive entries, then solves the dual problem as a semidefinite program using positive trigonometric polynomial theory. From the dual solution, the channel's continuous delay frequencies are read off where the dual polynomial has magnitude one, and the impulsive-noise positions are read off where the dual vector has magnitude $\\lambda$. The payoff is a tractable SDP that avoids the grid quantization of standard compressed-sensing channel estimators and does not require modeling the impulsive-noise power.","feed_headline":"One convex program recovers OFDM delays and impulse noise","feed_subtitle":"Solving the dual of a penalized atomic-norm problem yields continuous delay estimates and impulsive-noise locations at once.","key_machinery":"The load-bearing object is the dual polynomial $Q(f)$ built from the SDP dual variable $\\hat{q}$, together with the semidefinite constraint that controls its sup norm. Positive trigonometric polynomial theory converts the infinite-dimensional constraint $\\|q\\|_A^* \\le 1$ into a linear matrix inequality involving a Hermitian Toeplitz matrix $Q_0$ with trace constraints $\\operatorname{tr}[\\Theta_k Q_0] = \\delta_k$, yielding the SDP (8). The dual variable then acts as a certificate: contact points $|Q(f)| = 1$ locate the continuous frequencies, and saturation of the $\\ell^\\infty$ bound $|\\hat{q}| = \\lambda$ locates impulsive noise.","core_discovery":"The paper's central claim is that the dual of the penalized atomic norm minimization, reformulated as the semidefinite program (8), simultaneously recovers the sparse time-dispersive OFDM channel and the support of impulsive noise. Under the observation model $y = h + w + z$, where $h$ is a sum of sinusoidal atoms at continuous frequencies $f_k$ and $z$ is sparse on the pilot index set, the dual optimum $\\hat{q}$ defines a trigonometric polynomial $Q(f) = \\sum_{k=-m}^{m} \\hat{q}(k) e^{j2\\pi f k}$. The channel frequencies are estimated as the points where $|Q(f)| = 1$, the impulsive-noise support as the indices where $|\\hat{q}| = \\lambda$, and the path gains then follow from a least-squares fit. The authors support this claim with simulations showing simultaneous localization of both entities and with phase-transition plots for $P = 64$ and $P = 128$ pilots.","pith_inferences":["The dual-polynomial contact rule suggests a direct statistical support test: count exceedances of $|Q(f)| = 1$ or $|\\hat{q}| = \\lambda$ and compare against a threshold calibrated by the Gaussian variance, though the paper does not derive false-alarm rates.","Because the method estimates frequencies on $[0,1)$ without a grid, the same dual construction could carry to two-dimensional delay-Doppler estimation if the atomic set is replaced by time-frequency atoms; this is not developed in the paper.","The phase-transition plots define an empirical recoverable region in the $(s,r)$ plane, but the paper gives no proof of a phase transition; proving one under a minimum-separation condition would turn the empirical region into a guarantee.","The even-$P$ reindexing deserves direct numerical scrutiny: for $P = 64$ or $P = 128$ the SDP (8) may be solving a symmetric-index observation model rather than the original pilot model, and the paper offers no construction proving they coincide."],"forward_implications":["Channel delays are estimated on a continuous dictionary, so the basis-mismatch error of grid-based compressed sensing is removed.","One SDP simultaneously gives the channel frequency support and the impulsive-noise support, so no separate blanking or nonlinearity block is needed before demodulation.","The estimator's success depends on the number of scatters $s$, impulsive entries $r$, and pilots $P$, rather than on the statistical power of the impulse noise.","Once the frequencies are known, the path gains are obtained by least squares, and the whole channel estimate is available at polynomial-time cost $O(P^3)$."],"supporting_citations":[{"why":"Supplies the atomic-norm semidefinite characterization and the Vandermonde decomposition used to localize frequencies off the grid.","marker":"[8]"},{"why":"Provides the positive trigonometric polynomial theory that turns the dual's infinite-dimensional norm constraint into linear matrix inequalities.","marker":"[11]"},{"why":"Establishes atomic-norm OFDM channel estimation on a continuous dictionary, the setting this paper extends to impulsive noise.","marker":"[7]"},{"why":"Provides the penalized $\\ell^1$ minimization with outliers used as the simulation baseline in the MSE comparison.","marker":"[16]"},{"why":"Supplies the pilot-allocation scheme and sparse channel model that determine the number $P$ and the index set $J$.","marker":"[3]"},{"why":"Represents the conventional impulsive-noise mitigation approach whose need for noise-power modeling the proposed method avoids.","marker":"[2]"}],"fun_headline_variants":["Dual atomic-norm program finds OFDM delays and impulse noise together","One convex solve recovers both OFDM channel and impulse noise support","Penalized atomic norm dual: joint channel and noise recovery in OFDM","Simultaneous delay and impulse-noise recovery via atomic norm dual","SDP dual yields continuous delays and sparse noise locations in one run"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For even numbers of pilots, the paper assumes without proof that the pilot indexes $0, \\dots, P-1$ can be re-centered as $-m, \\dots, m$ with no loss or reordering of the measurements; if that reindexing is not equivalent, the SDP solved is not the stated estimation problem.","fun_headline_variants_meta":{"raw":{"variants":["Dual atomic-norm program finds OFDM delays and impulse noise together","One convex solve recovers both OFDM channel and impulse noise support","Penalized atomic norm dual: joint channel and noise recovery in OFDM","Simultaneous delay and impulse-noise recovery via atomic norm dual","SDP dual yields continuous delays and sparse noise locations in one run"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000806,"raw_usage":{"total_tokens":3520,"prompt_tokens":909,"completion_tokens":2611,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":2518}},"tokens_in":525,"tokens_out":2611,"duration_ms":18897,"temperature":1.0,"reasoning_tokens":2518,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:59:09.727536+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve the SDP (8) for $P = 64$ with a single true frequency $f = 0.37$ and no impulsive noise, then evaluate the dual polynomial: the true $f$ should be a contact point $|Q(f)| = 1$. If the even-$P$ reindexing $J = \\{-m, \\dots, m\\}$ is not equivalent to the original $0, \\dots, P-1$ indexing, the contact point will be displaced or missing.","supporting_citations":[{"cited_title":"Dumitrescu, Positive trigonometric polynomials and signal processing applications","cited_arxiv_id":null,"evidence_quote":"Provides the positive trigonometric polynomial theory that turns the dual's infinite-dimensional norm constraint into linear matrix inequalities."},{"cited_title":"Estimation of sparse time dispersive channels in pilot aided OFDM using atomic norm,","cited_arxiv_id":null,"evidence_quote":"Establishes atomic-norm OFDM channel estimation on a continuous dictionary, the setting this paper extends to impulsive noise."},{"cited_title":"Signal recovery from incomplete measurements in the presence of outliers,","cited_arxiv_id":null,"evidence_quote":"Provides the penalized $\\ell^1$ minimization with outliers used as the simulation baseline in the MSE comparison."},{"cited_title":"Pilot-based channel estimation for OFDM systems by tracking the delay-subspace,","cited_arxiv_id":null,"evidence_quote":"Supplies the pilot-allocation scheme and sparse channel model that determine the number $P$ and the index set $J$."},{"cited_title":"Analysis and comparison of several simple impulsive noise mitigation schemes for OFDM receivers,","cited_arxiv_id":null,"evidence_quote":"Represents the conventional impulsive-noise mitigation approach whose need for noise-power modeling the proposed method avoids."}],"review_version":1}