{"id":"a16f7768-562d-419e-bf1e-30759967a42a","arxiv_id":"2506.17010","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Gaussian Belief Propagation detector for AFBM is shown, via simulation, to outperform AFDM by about 2 dB at BER 10^-3 in doubly-dispersive channels at low per-iteration cost.","lead":"This paper designs a low-complexity receiver for Affine Filter Bank Modulation (AFBM), a new waveform for high-mobility wireless and radar-communication systems. It reports that the receiver, built on Gaussian Belief Propagation, beats the established AFDM waveform in bit-error-rate simulations while keeping the waveform's low out-of-band emissions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The receiver's likelihood in Eqs. (20)-(23) treats G^H n as scalar white noise; G is an overlapping Toeplitz filter bank, so the reported 2 dB GaBP gain over AFDM may hinge on an unverified noise-whitening assumption.","rationale":"I read the manuscript as a detector-design paper: given a known doubly-dispersive channel and the AFBM transmit matrix, the proposed GaBP receiver should estimate QPSK symbols with element-wise operations and outperform LMMSE/AFDM. For that claim to hold, the likelihood model in Eq. (20) must be correct, which requires the residual after matched filtering to be Gaussian with known variance and independent across the \\bar n index. The derivation starts from a valid-looking model, but the noise term is G^H n, and G is a non-orthogonal filter-bank matrix. Rather than accusing the authors, I note that the paper gives no whitening step, no estimate of G^H G, and no statement that C_f also equalizes noise; 'complex orthogonality' logically constrains the signal term only. This is the same weakness the reader flagged. It is load-bearing because the headline 2 dB gain is a small BER difference at 10^-3; any mismatch in the noise statistics can shift curves by that amount or more, and without code the direction of the shift cannot be assessed. The proposed test would settle it by comparing the physical noise insertion with the assumed scalar-sigma^2 insertion. If the covariance is near-identity, the concern is retired; if not, the paper should be revised to include a whitening stage or to justify the approximation. I therefore recommend keeping the reader's CONDITIONAL verdict: the contribution is interesting and the complexity argument is structurally sound, but the noise model needs verification before the 2 dB claim is accepted.","tokens_in":8453,"tokens_out":11799,"duration_ms":122718,"concrete_test":"Reproduce the Fig. 1 PHYDYAS (O=4) and Hermite (O=1.5) setups exactly (including channel parameters, \\xi, and \\beta_x) under two noise injections: (A) physical, n \\sim \\mathcal{CN}(0,\\sigma_n^2 I) added before G^H, i.e. \\bar r = \\bar H x + G^H n; (B) idealized, w \\sim \\mathcal{CN}(0,\\sigma_n^2 I) added directly to \\bar r as Eqs. (20)-(23) assume. Also compute \\|G^H G - \\alpha I\\|_F / \\|G^H G\\|_F for each filter. If the norm ratio is small and BER(A) agrees with BER(B) to within about 0.5 dB at 10^-3, the concern is resolved. If BER(A) is materially worse, the GaBP receiver needs an explicit noise-whitening stage and the 2 dB claim as stated is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the linear model in Eq. (16), \\bar r = \\bar H x + \\bar w, with \\bar H = G^H H G(I_K\\otimes Q_P C_f)\\Xi. From Eq. (10), \\bar r = G^H r, so the effective noise is \\bar w = G^H n. Equations (20), (21) and (23) then treat this noise as white with scalar variance \\sigma_n^2 and combine observations across the \\bar n index as independent. But G in Eq. (9) is a block Toeplitz matrix whose columns are the prototype filter shifted by N/2 samples; for the Hermite O=1.5 and PHYDYAS O=4 filters used in Fig. 1, adjacent and nearby columns overlap, so G^H G has off-diagonal bands. No whitening filter is applied, and no proof that G^H G is a scaled identity is given; 'transceiver complex orthogonality' in [17] constrains the signal path, not the covariance of the matched-filtered noise. If the coloring is material, Eq. (21) assigns the wrong residual variance and Eq. (22) treats dependent observations as independent, so the GaBP fixed point is not the true posterior estimate. The paper withholds normalized delays, Doppler shifts, \\xi, \\beta_x, and code, so the reported 2 dB advantage at 10^-3 cannot currently be separated from this model mismatch.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Gaussian belief propagation (GaBP) receiver for affine filter bank modulation (AFBM), a waveform recently introduced for doubly-dispersive (DD) channels. The receiver models the input–output relationship as r̄ = H̄x + w̄ and performs detection through soft interference cancellation, belief generation, and soft replica generation with damping, using only element-wise scalar operations. The authors claim a per-iteration complexity of O(N̄M̄) versus O(M̄^3) for LMMSE, and simulation results show about 2 dB gain over AFDM at a BER of 10^-3, together with low out-of-band emissions.","tokens_in":8715,"tokens_out":4558,"duration_ms":45060,"significance":"If the claims hold, the contribution is significant for ISAC and high-mobility communications: it provides a low-complexity detection scheme for a waveform with good spectral containment, avoiding costly matrix inversions. The GaBP framework is standard and the complexity analysis is straightforward. The reported 2 dB gain, however, rests on simulation curves and on a noise model that is not fully justified; these issues need to be resolved before the performance claim can be accepted.","major_comments":[{"comment":"The model sets r̄ = H̄x + w̄ with w̄ = G^H n, but Eq. (21) treats w̄ as white noise with scalar variance σ_n^2. The matrix G in Eq. (9) is an overlapping block Toeplitz filter matrix, so G^H G generally has nonzero off-diagonal entries; the paper provides neither a whitening step nor a proof that G^H G is a scaled identity. This mismatch can bias the GaBP variance updates in Eq. (21) and the consensus estimate in Eq. (26), so the reported BER gain may not hold outside the exact simulated setup. Please either whiten the observation or incorporate the true noise covariance into the variance updates.","section":"Section III, Eq. (16) and Eq. (21)"},{"comment":"The damping update is written as x̂^(i) = β_x x̂^(i) + (1−β_x) x̂^(i−1), which is self-referential: the right-hand side uses the quantity being defined. The equation should use the undamped estimate from Eq. (24), e.g., a separate notation such as x̃^(i), to be algebraically consistent. As written, the algorithm is not executable as stated.","section":"Section III, Eq. (25a)"},{"comment":"The iteration indices are inconsistent. In Eq. (18), the soft replica at iteration i is denoted x̂^(i)_{n,m}; Eq. (19) uses x̂^(i−1); the soft IC equation below Eq. (19) uses x̂^(i) in the interference term; and the text says that replicas from a previous iteration are used. This prevents exact reproduction of the algorithm. Please unify the superscripts and define the order of updates clearly.","section":"Section III, Soft IC and belief generation"},{"comment":"The central claim of a 2 dB gain at BER 10^-3 is based solely on simulation curves with no confidence intervals, number of trials, or statistical significance statement. Given the noise-model concern in the first comment, the reader cannot assess whether the observed gain is robust or an artifact of the specific channel realization and filter choices. Please include error bars or at least state the number of independent Monte Carlo runs.","section":"Section IV, Figs. 1–2"}],"minor_comments":[{"comment":"The matrix Ξ̄ uses a block 0_{L/2} in the middle, but the text refers to 0_L as a full zero matrix of size L; the notation should be made consistent, e.g., 0_{L/2} for the zero block.","section":"Section II, Eq. (3)"},{"comment":"The second tanh term contains a typo: the subscript is \\bar{x}^{(i)}_{\\bar n,\\bar k} instead of \\bar{x}^{(i)}_{\\bar n,\\bar m}. Please correct.","section":"Section III, Eq. (24)"},{"comment":"The caption contains a typo: 'Abmibuity' should be 'Ambiguity'.","section":"Section IV, Fig. 3"},{"comment":"The condition that P must be smaller than N is stated in the text, but in the simulations P=256 and N=256, and P=128,192,256 are used with N=256; please clarify the exact constraint and its relationship to the simulation settings.","section":"Section II, Eq. (9)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for the journal, but the load-bearing noise-whitening assumption and the self-referential damping equation need to be addressed before the central performance claim can be credited. The authors might consider releasing code or providing more detailed simulation configuration to improve reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it is the first receiver designed specifically for AFBM, and that is a useful step: AFBM is a new ISAC waveform from the same group, and nobody had shown how to detect it. Second, the central claim, that GaBP detection beats AFDM by about 2 dB at a BER of 10^-3, is supported only by simulations that leave out several parameters and ignore a potential noise-whitening problem. So my honest take is: plausible, worth reviewing, not yet solid.\n\nWhat is genuinely good: the paper takes the AFBM signal model, derives an element-wise GaBP detector with per-iteration complexity O(NK * KL/2) instead of O((KL/2)^3) for LMMSE, and shows BER curves plus OOBE. The equations are standard GaBP/SGA and are mostly consistent. The complexity advantage is real if the algorithm converges, and the OOBE advantage of the filter bank over AFDM is unsurprising but well illustrated.\n\nNow the soft spots, in rough order of severity. The biggest one is the noise model. In Eq. (16) the receiver uses r_bar = G^H r, so the effective noise is w_bar = G^H n. G is a block Toeplitz filter matrix with overlapping columns (for Hermite O=1.5 and PHYDYAS O=4). G^H n is generally colored, and no whitening step or proof that G^H G is a scaled identity is given. The GaBP variance updates in Eqs. (20)-(23) treat each observation as if it had independent white noise with variance sigma_n^2. If the coloring is significant, the likelihood is mismatched and the reported BER doesn't mean what it appears to mean. This issue deserves a whitening filter or at least a numerical check of diag versus off-diagonal entries of G^H G. It is not necessarily fatal; many receivers ignore noise coloring when it is mild. But the paper does not show that it is mild.\n\nSecond, the paper withholds several numbers needed to reproduce the simulations: normalized delays and Doppler shifts, the guard width xi, the damping factor beta_x, and all Monte Carlo details. The authors should give these values. There is also a typo in Eq. (25a): the same variable appears on both sides; presumably the RHS should be the undamped estimate from Eq. (24). The iteration indices in the soft IC description are a bit sloppy, for instance around Eq. (19). These are minor but should be fixed.\n\nThird, the 2 dB gain over AFDM is claimed for a particular channel setup, but without the channel parameters it cannot be checked. The GaBP for AFDM may also not be optimally tuned, which could bias the comparison. Not necessarily a flaw, but a fair referee should ask.\n\nWho is this for? People working on AFDM, AFBM, and filter-bank ISAC waveforms. If that is you, it is worth reading. If you are outside the area, the specific contribution is narrow: a new application of known message-passing to a new matrix model, not a new principle.\n\nMy recommendation: send it to peer review. A good referee can push on the noise whitening and the missing simulation details. If those are resolved, it becomes a solid contribution to the waveform-receiver literature.","headline":"First AFBM receiver is a useful step, but the 2 dB gain over AFDM rests on an unexamined white-noise assumption that a referee should push on.","tokens_in":9319,"tokens_out":6524,"would_cite":true,"duration_ms":66382,"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":"This paper claims that a Gaussian Belief Propagation receiver can decode AFBM, a filter-bank ISAC waveform, with element-wise scalar operations and beat AFDM by about 2 dB at a bit error rate of 0.001.","keywords":["affine filter bank modulation","Gaussian belief propagation","doubly-dispersive channels","integrated sensing and communications","low-complexity detection","out-of-band emissions","affine frequency division multiplexing","soft interference cancellation"],"falsifier":"Compute the empirical covariance of $\\bar{\\mathbf{w}}=\\mathbf{G}^{H}\\mathbf{n}$ for the two prototype filters used in the simulations and compare its off-diagonal entries with $\\sigma_{n}^{2}$; if they are not negligible, rerun the BER curves with a whitened or colored-noise-aware GaBP detector and check whether the 2 dB gain at a bit error rate of $10^{-3}$ survives.","tokens_in":8201,"feed_emoji":"📡","tokens_out":11070,"duration_ms":96068,"temperature":0.7,"pith_summary":"Affine Filter Bank Modulation (AFBM) is a recently proposed waveform for integrated sensing and communications that keeps the low out-of-band emissions of filter-bank systems while remaining usable in doubly-dispersive channels, but no receiver had been designed for it. This paper claims that a Gaussian Belief Propagation (GaBP) detector, built from element-wise scalar operations, can decode AFBM without matrix inversion. In simulations over a three-path doubly-dispersive channel it reports a bit error rate about 2 dB better than AFDM at $10^{-3}$, with per-iteration complexity $\\mathcal{O}(\\bar{N}\\bar{M})$ versus $\\mathcal{O}(\\bar{M}^{3})$ for LMMSE. The same simulations show that AFBM keeps very low out-of-band emissions relative to AFDM, so the spectral advantage does not have to be paid for in receiver complexity.","feed_headline":"Element-wise GaBP detection lifts AFBM 2 dB over AFDM","feed_subtitle":"The ISAC waveform is decoded with scalar operations only, replacing matrix inversion with cheap iterative updates.","key_machinery":"The load-bearing object is the filtered time-domain channel matrix $\\bar{\\mathbf{H}}=\\mathbf{G}^{H}\\mathbf{H}\\mathbf{G}(\\mathbf{I}_{K}\\otimes\\mathbf{Q}_{P}\\mathbf{C}_{f})\\boldsymbol{\\Xi}$, which collapses the AFBM transmit filter bank, doubly-dispersive channel, and receive demodulation into one linear model with known coefficients. On this model the paper runs Gaussian Belief Propagation: each receive element sends a Gaussian message whose mean and variance come from subtracting the interference of all other symbols (soft interference cancellation), the messages are combined into an extrinsic belief, a Bayes-optimal QPSK denoiser with damping produces soft replicas, and the consensus update fuses the resulting estimates. This mechanism carries the argument because it converts detection into repeated element-wise scalar updates, avoiding the matrix inversion required by LMMSE, while the scalar Gaussian approximation keeps the messages Gaussian.","core_discovery":"The paper's central claim is that the AFBM transceiver, whose received signal after demodulation can be written as $\\bar{\\mathbf{r}}=\\bar{\\mathbf{H}}\\mathbf{x}+\\bar{\\mathbf{w}}$ with $\\bar{\\mathbf{H}}=\\mathbf{G}^{H}\\mathbf{H}\\mathbf{G}(\\mathbf{I}_{K}\\otimes\\mathbf{Q}_{P}\\mathbf{C}_{f})\\boldsymbol{\\Xi}$, can be inverted by a GaBP message-passing receiver that never forms or inverts the full Gram matrix. The receiver alternates soft interference cancellation, belief generation under a scalar Gaussian approximation, Bayes-optimal soft replica generation for QPSK, and a consensus update that merges the per-row estimates. The paper reports that this scheme matches or beats LMMSE detection: in the simulated doubly-dispersive channel the GaBP-based AFBM receiver outperforms AFDM by roughly 2 dB at a bit error rate of $10^{-3}$ across the chirp sizes tested, with per-iteration cost $\\mathcal{O}(\\bar{N}\\bar{M})$ rather than $\\mathcal{O}(\\bar{M}^{3})$. It also verifies that AFBM retains very low out-of-band emissions relative to AFDM while its ambiguity function stays similar, so the sensing side of ISAC is not sacrificed.","pith_inferences":["Editorial: The 2 dB comparison is at one bit error rate operating point in one channel profile; the ordering at higher SNR or with more resolvable paths is untested, and the colored-noise question below should be checked before generalizing.","Editorial: The same GaBP structure should extend to higher-order QAM via the adaptive belief scaling cited in the paper, and to joint channel-and-data estimation if the channel matrix is estimated iteratively; the paper only demonstrates QPSK with a known channel.","Editorial: Because the filtered channel matrix inherits the Toeplitz and Kronecker structure of the AFBM filter bank, fast matrix-vector products may lower the per-iteration cost below the reported product of dimensions in practice; the paper does not exploit this.","Editorial: If the colored-noise concern is real, a whitening step or a colored-noise-aware variance model could be inserted into the GaBP updates without changing the architecture; the paper does not test either."],"forward_implications":["AFBM detection scales to larger frames: each GaBP iteration costs a product of the frame dimensions in element-wise operations instead of a cubic matrix inversion, so the practical gap over LMMSE widens as the frame grows.","The reported 2 dB gain over AFDM at a bit error rate of 0.001, if it persists across channel profiles, makes AFBM a stronger candidate for high-mobility ISAC links where spectral containment matters.","Because the receiver works on the collapsed filtered channel matrix, it can be reused with any prototype filter and any chirp size that satisfies the orthogonality condition, as the simulations with varying chirp size indicate.","The complexity comparison is per iteration; with damping and consensus the number of iterations is a modest multiplier, so the end-to-end cost remains far below LMMSE for the simulated sizes."],"supporting_citations":[{"why":"Defines the AFBM waveform and its complex-orthogonality compensation, the scheme whose receiver the paper proposes.","marker":"[17]"},{"why":"Introduces AFDM, the baseline waveform whose BER the proposed AFBM-GaBP receiver is compared against.","marker":"[9]"},{"why":"Provides the doubly-dispersive channel model and the chirp-parameter orthogonality condition used to choose simulation parameters.","marker":"[8]"},{"why":"Supplies the pruned DFT-spread filter-bank principles behind AFBM's low out-of-band emissions.","marker":"[14]"},{"why":"Supplies the generalized DFT-precoded filter bank structure that AFBM adapts with affine precoding.","marker":"[16]"},{"why":"Gives convergence conditions for Gaussian belief propagation, supporting the damping update used to avoid local minima.","marker":"[20]"},{"why":"Provides Bayes-optimal denoisers for higher-order modulations, backing the claim that QPSK is without loss of generality.","marker":"[19]"},{"why":"Explains the oversampling tradeoff that justifies AFBM's double-rate transmission structure.","marker":"[18]"}],"fun_headline_variants":["Scalar-only GaBP receiver lifts AFBM 2 dB over AFDM","AFBM decoded with element-wise GaBP beats AFDM in DD channels","Low-complexity GaBP detector gives AFBM edge over AFDM","Element-wise beliefs make AFBM outperform AFDM for ISAC","GaBP receiver: AFBM tops AFDM with scalar operations only"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The detector assumes the filtered noise $\\bar{\\mathbf{w}}=\\mathbf{G}^{H}\\mathbf{n}$ is white with a single scalar variance, even though the overlapping Toeplitz filter columns generally color it; the paper gives no whitening step or proof that the coloring is negligible.","fun_headline_variants_meta":{"raw":{"variants":["Scalar-only GaBP receiver lifts AFBM 2 dB over AFDM","AFBM decoded with element-wise GaBP beats AFDM in DD channels","Low-complexity GaBP detector gives AFBM edge over AFDM","Element-wise beliefs make AFBM outperform AFDM for ISAC","GaBP receiver: AFBM tops AFDM with scalar operations only"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000145,"raw_usage":{"total_tokens":1159,"prompt_tokens":908,"completion_tokens":251,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":157}},"tokens_in":524,"tokens_out":251,"duration_ms":3259,"temperature":1.0,"reasoning_tokens":157,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:15:10.391306+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the empirical covariance of $\\bar{\\mathbf{w}}=\\mathbf{G}^{H}\\mathbf{n}$ for the two prototype filters used in the simulations and compare its off-diagonal entries with $\\sigma_{n}^{2}$; if they are not negligible, rerun the BER curves with a whitened or colored-noise-aware GaBP detector and check whether the 2 dB gain at a bit error rate of $10^{-3}$ survives.","supporting_citations":[],"review_version":2}