{"id":"fe98b9df-1d3a-4fe4-af81-56495887f084","arxiv_id":"2607.20227","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A joint chirp-parameter and sparsification-pattern optimization enables a banded-Cholesky MMSE receiver for AFDM with near-full-MMSE BER at O(N d^2) complexity.","lead":"This paper designs a low-complexity MMSE receiver for AFDM, a 6G candidate waveform, by sparsifying the channel matrix and jointly optimizing the chirp parameter. It reports an order-of-magnitude faster chirp search and about 1 dB BER gain at 10^-3 over baseline AFDM configurations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FCDA's surrogate metric P_e(c1) rests on a false diagonal-invariance assertion (Eq. 19) and is validated by only one simulation; if P_e misranks BER, the order-of-magnitude c1-search claim collapses.","rationale":"The paper addresses a real and practically relevant problem: joint chirp-parameter selection and low-complexity MMSE reception for AFDM. The complexity reduction claimed for the banded receiver is plausible and the algorithmic structure is coherent. I read the strongest claim as a bundle: FCDA gives an order-of-magnitude faster c1 search, and HS-JCPS with LC-BMMSE gives about 1 dB gain at BER 10^-3. The weakest load-bearing element is the P_e(c1) surrogate metric used for the FCDA search-time claim. The derivation breaks at Eq. (19): the diagonal part of H_t is not invariant under FΛ_{c1}(·)Λ_{c1}^H F^H. The correct diagonal is a constant plus the transformed off-diagonal contribution, so P_e measures only the latter and does not account for the nonlinear way that BER depends on the constant offset c. This makes P_e a heuristic rather than a derived proxy. The paper's only validation is Fig. 5, which is a single operating point. A rank-correlation audit across varied N, SNR, and channel realizations would settle whether the heuristic is reliable. I am not recommending rejection: the HS-JCPS algorithm itself uses PA(c1,w), not P_e, so the 1 dB empirical claim may hold even if FCDA is flawed. But the reader's conditional verdict is appropriate, and my concern reinforces that conditionality rather than escalating it. No code or data are provided, so the audit cannot be done from the manuscript alone and should be requested.","tokens_in":21888,"tokens_out":16946,"duration_ms":157724,"concrete_test":"Run a rank-correlation audit of P_e(c1) versus exact full-ECM MMSE BER across settings not used in Fig. 5: N ∈ {64, 256, 512}, Eb/N0 ∈ {10, 15, 20} dB, and at least 50 independent TR38.811/NTN-TDL-A channel realizations, including a 6-tap configuration. For each setting, compute Eq. (28) and the Eq. (14) BER for all candidate c1 = k/(2N), k = 1,...,N; measure Spearman rank correlation and record whether argmin P_e = argmin BER. Also numerically verify the corrected diagonal identity [FΛ_{c1} H_t Λ_{c1}^H F^H]_{k,k} = tr(H_t)/N + [L]_{k,k}. If the rank correlation is below 0.8 or argmin mismatch occurs in more than 5% of settings, the FCDA claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central FCDA claim—that P_e(c1) from Eq. (28) can replace exact BER for c1 selection—is not supported by the derivation. In Eq. (19), the diagonal part of H_t is asserted to be invariant under the similarity transformation FΛ_{c1}(·)Λ_{c1}^H F^H. That is false: F diag(H_t) F^H is circulant with every diagonal entry equal to tr(H_t)/N, so the k-th diagonal of the transformed matrix is c + [L]_{k,k}, not [H_t]_{k,k} + [L]_{k,k}. Consequently, SINR_k = (Es/σ_n^2)/(c + [L]_{k,k}) − 1, and the BER in Eq. (14) is a nonlinear Q-function of SINR_k. Minimizing Σ_k |[L]_{k,k}|^2 is therefore not equivalent, in general, to minimizing BER; it is a heuristic. The only evidence that P_e ranks BER is Fig. 5, which shows one channel realization, one SNR setting, and one N. If the ranking degrades at other operating points, the FCDA algorithm may select a c1 with substantially worse BER, undermining both the order-of-magnitude search-time claim and the paper's 'near-optimal c1' conclusion. The HS-JCPS 1 dB result is less directly affected because Algorithm 2 evaluates PA(c1,w) rather than P_e, but the FCDA contribution and the abstract's joint claim remain exposed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates joint chirp-parameter (c1) selection and low-complexity MMSE detection for AFDM over doubly selective channels. It proposes: (i) an FCDA algorithm that selects c1 by minimizing a simplified metric P_e(c1) derived from the regularized inverse Gram matrix; (ii) an LC-BMMSE receiver that sparsifies the effective channel matrix path-wise, reorders the regularized matrix with RCM, and solves the banded system via banded Cholesky factorization; and (iii) an HS-JCPS algorithm that jointly optimizes c1 and per-path pruning widths under a fixed complexity budget. The reported results include an order-of-magnitude reduction in c1 search time and roughly 1 dB BER gain at 10^-3.","tokens_in":22307,"tokens_out":12668,"duration_ms":117699,"significance":"If correct, the FCDA metric would be a useful tool for AFDM parameter selection, and the LC-BMMSE receiver offers a convincing complexity reduction from O(N^3) to O(N d~^2 + N d~ + Nd), supported by Table I and the BER-versus-d~ tradeoff. The paper also provides useful waveform comparisons (AFDM vs OCDM/OFDM/OTFS) under equal guard overhead and with perfect CSI, which is a fair baseline. These are genuine strengths. The main reservation concerns the derivation of the FCDA surrogate metric, which contains an algebraic error; the LC-BMMSE and HS-JCPS parts are less affected because HS-JCPS evaluates the actual cost P_A(c1,w) rather than P_e(c1).","major_comments":[{"comment":"The derivation of P_e(c1) relies on the claim that the diagonal part of H_t is invariant under the similarity transformation U = F Lambda_c1 (.) Lambda_c1^H F^H. This is false. For a diagonal matrix D, diag(U D U^H) = (tr(D)/N) 1_N, not diag(D). Therefore Eq. (19) should be [U H_t U^H]_{k,k} = c + [L]_{k,k}, with c = tr(H_t)/N and L defined by (20). Consequently, P_e in (21)-(28) omits the constant c, and since the BER in (14) is a nonlinear Q-function of SINR_k, minimizing sum_k |[L]_{k,k}|^2 is not equivalent to minimizing BER. Remark 1 overstates the metric; Fig. 5 validates it for one channel, one SNR, and one N. This error directly affects the FCDA near-optimal c1 claim and the order-of-magnitude search-time claim. Please correct the derivation (e.g., include c) or reframe FCDA as a heuristic with additional validation across SNR, N, and channel realizations.","section":"Section III-B, Eq. (19)-(28)"},{"comment":"The proof of the lower bound is not justified as written. In (51), the term (sigma_n^2 + sigma_Delta^2)/E_s [D(Psi~^{-1})^H]_{k,k} appears with a minus sign. Proposition 2 asserts that this term, together with the preceding nonnegative term, is positive and smaller than C(sigma_Delta). But D = Psi~^{-1} H~_eff^H Delta H_eff is not Hermitian, and Delta H_eff is an arbitrary pruning perturbation; the diagonal entries of D(Psi~^{-1})^H need not be real, let alone nonnegative. The inequality (54) follows from (51) only if these subtracted terms are known to be nonnegative. As it stands, Lemma 1 is not established. The HS-JCPS algorithm does not depend on this bound, but the paper explicitly lists the lower bound as a contribution and uses it in Fig. 9. Please prove the sign property or state clearly that the bound is heuristic.","section":"Section IV / Appendix A, Lemma 1 and Proposition 2"}],"minor_comments":[{"comment":"The expression d = w1 - xi1 + wP + xiP appears to be off by one if d denotes the number of retained rows in the cyclic interval; it should be w1 + wP + xiP - xi1 + 1.","section":"Section II-D"},{"comment":"The condition (p+xi_i-w_i)_N <= q <= (p+xi_i+w_i)_N is ambiguous when the cyclic interval wraps around the index boundary. Please define the retained set using modular arithmetic or interval union notation.","section":"Eq. (9)"},{"comment":"The mask S uses negative powers of Pi_u (u = -d+1, ..., -1) but Pi_u was only defined for nonnegative u. Define Pi_{-u} as Pi_{N-u} or use a different notation.","section":"Eq. (29)"},{"comment":"The right y-axis appears to be labeled '10^3' with no clear unit or scaling. Please clarify whether P_e(c1) is scaled and how the two curves are related.","section":"Fig. 5"},{"comment":"The complexity comparison for FCDA does not mention the one-time cost of forming H_t = (H^H H + sigma_n^2/E_s I)^{-1}. State explicitly that this cost is amortized over all c1 candidates; otherwise the comparison is incomplete.","section":"Section III-B, complexity"}],"recommendation":"major_revision","confidential_remarks":"The key issue is Eq. (19): the diagonal-invariance assertion is algebraically wrong and undermines the FCDA derivation. The LC-BMMSE receiver and HS-JCPS algorithm are largely independent and have merit; the paper is likely salvageable after a substantive revision of the FCDA claim or its justification. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading for the receiver design, but do not take the FCDA metric derivation at face value. The LC-BMMSE receiver exploits the cyclic-banded structure of the sparsified AFDM channel matrix and uses RCM reordering plus banded Cholesky to drop MMSE complexity from O(N^3) to O(N  d~^2). That part is standard but competently done, and the complexity accounting in Table I holds up. The joint optimization of c1 and per-path pruning widths under a fixed complexity budget is also genuinely new relative to prior work, and the simulations support the claimed ~1 dB gain and order-of-magnitude search-time reduction. Those are real contributions.\n\nThe soft spot is Eq. (19). The authors assert that the diagonal part of H_t is invariant under the similarity transformation FΛc1(·)Λc1^H F^H. That is false. Fourier-transforming a diagonal matrix yields a constant diagonal, so the derivation of the surrogate metric P_e(c1) does not go through. This means P_e is a heuristic, not a derived simplification. The only evidence that it ranks BER correctly is Fig. 5, one simulation setting. That is thin, especially because the FCDA 'order-of-magnitude faster c1 search' claim depends on P_e being reliable. The authors should either fix the derivation or explicitly reframe P_e as a heuristic and add more validation across channel realizations and SNR points.\n\nThere are also two smaller issues. The definition of the sparsified regularized matrix \\tilde\\Psi is inconsistent between Eq. (30) and Appendix A, which is confusing. And the SINR lower bound in Lemma 1 / Proposition 2 has a proof gap: the inequality leading to the bound is asserted rather than proved. These are fixable.\n\nNo code or data is provided, which limits reproducibility, but the simulations are described in enough detail that the main empirical claims could be checked independently.\n\nWho is this for? Researchers working on AFDM, OTFS, or low-complexity receivers for doubly selective channels. They will get useful ideas from the receiver architecture and the joint optimization formulation, but they should treat P_e as an empirical tool, not a theorem.\n\nRecommendation: send to peer review. The paper deserves referee time, but the referee should push for a corrected derivation or a clear heuristic framing of P_e, a resolved \\tilde\\Psi definition, and a completed proof of the bound. If those are fixed, the paper would be a solid contribution to the AFDM literature.","headline":"Useful engineering contribution with a solid low-complexity receiver design, but the theoretical foundation of the proposed fast chirp-parameter search metric has a load-bearing algebraic error that needs fixing.","tokens_in":22801,"tokens_out":1737,"would_cite":false,"duration_ms":19005,"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":"AFDM's chirp parameter and receiver sparsification should be chosen jointly; a hierarchical search does this at near-linear MMSE cost, with a 1 dB BER gain over separate design.","keywords":["AFDM","chirp parameter selection","MMSE receiver","banded matrix factorization","low-complexity detection","doubly selective channel","structured sparsification","6G waveform"],"falsifier":"Enumerate all candidate c1 values for several different delay-Doppler channel realizations, compute both Pe(c1) and the true MMSE BER curve, and check whether the argmin of Pe equals the argmin of BER in every case; one counterexample where the two argmins differ — especially in a channel with strongly varying diagonal channel gain — would falsify the FCDA selection claim.","tokens_in":21754,"feed_emoji":"📶","tokens_out":6780,"duration_ms":64511,"temperature":0.7,"pith_summary":"Affine frequency division multiplexing (AFDM) is a 6G candidate waveform that can resist fast-varying, doubly selective channels, but its bit-error rate depends on chirp parameters and on how much of the channel matrix the receiver keeps. This paper tries to prove that those two choices — chirp selection and receiver sparsification — can be optimized together rather than treated separately. The authors derive a cheap closed-form metric for ranking chirp parameters, design a banded MMSE receiver that prunes the channel matrix path by path, and wrap both in a hierarchical search that respects a fixed complexity budget. If the paper is right, AFDM can run a near-optimal MMSE receiver at a small fraction of the usual cubic cost, gaining about 1 dB at a bit-error rate of 10^-3 over non-joint designs.","feed_headline":"AFDM detection cost drops from N-cubed to near-linear with a 1 dB gain","feed_subtitle":"Joint chirp-parameter and sparsification search keeps near-optimal BER while making the MMSE receiver cheap enough for large blocks.","key_machinery":"The engine is the identity Pe(c1)=1/N Σ_{u=1}^{N-1}|x_u|^2, which converts BER ranking of chirp parameters into an energy computation: x_u is the DFT-domain aggregate of the chirp-modulated circulant-diagonal vectors of the regularized inverse Gram matrix. On the receiver side, the load-bearing structure is a cyclic-banded binary mask S that prunes each multipath's coupling neighborhood; a sparse-matrix reordering step (RCM) turns the masked Gram matrix into a fixed banded matrix, and banded Cholesky factorization solves the MMSE system without a full inversion. The pruning widths w, the band width d, and c1 are linked: d determines the complexity, and the search must balance how much channe","core_discovery":"The paper sets out to show that the full MMSE receiver for AFDM can be replaced by a structured approximation with almost no performance loss. On the paper's own terms: the chirp parameter c1 affects post-MMSE SINR only through the diagonal entries of a transformed coupling matrix; the paper aggregates these into Pe(c1), a Parseval-energy metric, and observes that the c1 minimizing Pe also minimizes true BER. The receiver then sparsifies the effective channel matrix into a cyclic banded matrix, reorders it into a fixed banded form, and solves the MMSE system with banded Cholesky factorization. A hierarchical search, HS-JCPS, picks c1 and per-path pruning widths under a fixed complexity budge","pith_inferences":["A direct corollary the authors do not develop: the same hierarchical-search structure could be rerun per block or per slot to track a slowly drifting channel, since the reordering matrix and band pattern can be precomputed offline and only the local search cost is paid online.","The surrogate-metric idea should transfer to MIMO-AFDM and to index-modulated AFDM, where the per-path coupling masks become per-path spatial masks; the complexity of the local search would grow, but the Parseval-based ranking would still apply to each spatial stream.","The derivation of Pe(c1) relies on Eq. (19) treating the diagonal of Ht as invariant under FΛc1(·)Λc1^H F^H; as written, F diag(Ht) F^H has a constant diagonal equal to the average trace, not the original one. A corrected derivation would make Pe(c1) a slightly different quantity; whether its argmin still tracks BER is a test the paper leaves open."],"forward_implications":["Per-block MMSE detection in AFDM can be implemented with on the order of N d~^2 + 2N d~ + N d complex multiplications instead of roughly (4/3)N^3, making large-N blocks practical under high mobility.","Because the optimal c1 depends on the complexity budget (the band width d~), parameter selection and receiver design cannot be decoupled; the paper shows the best c1 shifts as d~ grows.","The c1 search becomes cheap enough to run adaptively: FCDA cuts the search time by about an order of magnitude relative to direct BER evaluation, which matters for fast-varying channels where c1 may need updating per block.","Sparsification need not cost diversity: the retained path responses can still have full column rank, so the pruned receiver keeps the channel's path diversity while dropping only residual interference energy.","Under equal guard overhead and perfect CSI, the joint design outperforms OFDM and OCDM and matches OTFS at the same complexity, with more headroom when the complexity budget is loosened."],"fun_headline_variants":["AFDM receiver: near-linear MMSE with near-optimal BER","Chirp selection and banded MMSE: AFDM complexity slashed","AFDM gets 10x faster chirp search, low-complexity receiver","Joint chirp and sparsification search keeps AFDM BER near-optimal","AFDM: cubic to near-linear receiver complexity without BER loss"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole fast-selection pipeline rests on the surrogate metric Pe(c1) ranking chirp parameters in the same order as true bit-error rate; the paper validates this ranking on a single simulation setting, and the derivation of Pe rests on an invariance of the diagonal of Ht under the chirp transform that is stated without proof.","fun_headline_variants_meta":{"raw":{"variants":["AFDM receiver: near-linear MMSE with near-optimal BER","Chirp selection and banded MMSE: AFDM complexity slashed","AFDM gets 10x faster chirp search, low-complexity receiver","Joint chirp and sparsification search keeps AFDM BER near-optimal","AFDM: cubic to near-linear receiver complexity without BER loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000694,"raw_usage":{"total_tokens":3034,"prompt_tokens":862,"completion_tokens":2172,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":2075}},"tokens_in":606,"tokens_out":2172,"duration_ms":16282,"temperature":1.0,"reasoning_tokens":2075,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T10:25:12.445129+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Enumerate all candidate c1 values for several different delay-Doppler channel realizations, compute both Pe(c1) and the true MMSE BER curve, and check whether the argmin of Pe equals the argmin of BER in every case; one counterexample where the two argmins differ — especially in a channel with strongly varying diagonal channel gain — would falsify the FCDA selection claim.","supporting_citations":[],"review_version":1}