{"id":"aaf7f352-88b0-422d-a74a-8913a38c479d","arxiv_id":"2412.08679","paper_version":1,"verdict":"UNVERDICTED","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A COST Action working group whitepaper reviews 6G signal processing for communications, localization, and ISAC, and compiles the group's own contributions and research directions.","lead":"This whitepaper from the COST INTERACT Working Group 2 surveys signal processing techniques for 6G communications, localization, and integrated sensing and communication. A generalist reader will find a structured map of where 6G physical-layer research stands and which open problems the group plans to tackle.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The generalized Bussgang claim in §1.5.6 is correct under standard finite-variance and independence assumptions, but the phrase 'regardless of distribution' overstates; it fails for dependent or heavy-tailed noise.","rationale":"The reader identified the same weakest assumption: the derivation relies on zero-mean, independent s and n with finite second moments. My independent check of §1.5.6 confirms that Eqs. (15)–(20) are algebraically correct under those assumptions, and the central mathematical claim is not fraudulent or internally inconsistent. However, the wording 'regardless of the distribution of signal and noise' is too broad. The formulas break down when s and n are correlated or when second moments do not exist. The whitepaper itself considers α-stable interference in §1.4.3, so the heavy-tailed case is not merely hypothetical in the same document. Still, this is a qualification issue, not a fundamental flaw: the decomposition by orthogonal projection is always available, and the paper honestly notes that αs is distribution-dependent. Since the paper is a whitepaper with no single testable central hypothesis, the UNVERDICTED status remains appropriate. The concrete counterexample with n = s provides a fast, unambiguous check of the missing assumption, and would settle whether the claim needs to be restated with the usual hypotheses.","tokens_in":45799,"tokens_out":6105,"duration_ms":62144,"concrete_test":"Construct a counterexample with s ~ N(0,1) and n = s (so s and n are dependent). For a memoryless nonlinearity, e.g., soft limiter f(x) = tanh(x), compute αs = E[ys]/E[s^2] and γs = E[y^2]/E[s^2] numerically. Compare with Eq. (19) using γx = E[y^2]/E[x^2] and σ_n^2/σ_s^2 = 1; the formula will fail because E[x^2] = 4σ_s^2 rather than 2σ_s^2. Alternatively, take n as α-stable with α = 1.5 and show σ_n^2 is undefined, so Eq. (19) cannot be evaluated. Either test settles whether the 'regardless of distribution' claim requires explicit uncorrelatedness and finite-second-moment assumptions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assertion in §1.5.6 is that 'a version of the Bussgang decomposition applies in general, regardless of the distribution of signal and noise,' supported by Eqs. (14)–(20). The derivation requires E[x^2] = σ_s^2 + σ_n^2, which presupposes s and n are uncorrelated (in the intended setting, independent) and have finite second moments. Without these, Eq. (19) does not follow. Example: if n = s, then x = 2s and γs = E[y^2]/σ_s^2 = 4γx, whereas Eq. (19) with σ_n^2/σ_s^2 = 1 would give γs = 2γx. If n is α-stable with α < 2, σ_n^2 is undefined and the formula is not meaningful. Moreover, the classical Bussgang constant-gain property (αx depending only on the nonlinearity and input variance) does not generalize; the authors concede that αs depends on the distributions of s and n. What remains is the orthogonal projection (linear minimum mean-square error decomposition) y = αs s + δ with uncorrelated δ, not the substantive content of the Bussgang theorem. This does not invalidate the whitepaper's practical use cases (OFDM with approximately Gaussian signals), but the unqualified 'regardless of distribution' is an overstatement that should be qualified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This document is a whitepaper from COST Action INTERACT Working Group 2. It surveys signal processing for communications (waveforms, channel coding, massive MIMO, massive access, RF impairments, O-RAN, security, underwater communications), signal processing for localization (AoA, fingerprinting, SLAM, machine learning, RSSI/UWB/LoRa, RIS, testbeds), and integrated sensing and communication (terminology, resource allocation, waveform design, channel measurements, parameter estimation, WiFi/cellular sensing). It reports a large number of the group's contributions, with heavy citation to internal WG2 papers, and contains one self-contained derivation: a generalized Bussgang decomposition for nonlinear systems with noisy inputs in Section 1.5.6.","tokens_in":46033,"tokens_out":11692,"duration_ms":118074,"significance":"As a survey and roadmap, the whitepaper is useful: it consolidates a broad range of 6G physical-layer topics and documents the activity of a large European working group. Its strengths are its breadth, its connection of each topic to concrete WG2 publications, and the explicit enumeration of open problems. The Bussgang derivation in Section 1.5.6 is internally consistent under the standard assumptions of zero-mean, finite-variance, independent signal and noise, and it gives a useful reminder that the SDNR identity itself does not require Gaussianity. However, the manuscript does not contain a single falsifiable central claim or reproducible experiments; most programmatic statements are qualitative, and many quantitative claims are borrowed from cited papers without providing enough detail for independent verification. As a position or survey document it could be valuable, but its technical authority rests on the cited primary literature rather than on the content of this paper.","major_comments":[{"comment":"The statement near Eqs. (14)-(20) that \"a version of the Bussgang decomposition applies in general, regardless of the distribution of signal and noise\" is an overstatement. The derivation of γs = γx(1 + σ_n^2/σ_s^2) and the SDNR formula in Eq. (19) presupposes that s and n are zero-mean, finite-variance, and uncorrelated (in the intended setting, independent), because it uses E[x^2] = σ_s^2 + σ_n^2. If n = s, then σ_n^2 = σ_s^2 but E[x^2] = 4σ_s^2, so γs = 4γx rather than the 2γx that Eq. (19) would give; if n is α-stable with α < 2, σ_n^2 is undefined and the formula is not meaningful. Moreover, for arbitrary distributions the decomposition y = αs s + δs with δs uncorrelated with s is a linear orthogonal projection (linear MMSE), not the Bussgang constant-gain property, and the paper itself concedes that αs depends on the distributions of s and n. Please state the required assumptions explicitly before Eq. (14) and replace the unqualified \"regardless of distribution\" claim with a qualification such as \"for zero-mean, finite-variance, uncorrelated signal and noise; in the general case the decomposition reduces to the linear MMSE/projection property.\"","section":"1.5.6"}],"minor_comments":[{"comment":"The title contains a typo: \"Intergrated\" should be \"Integrated.\"","section":"Title"},{"comment":"The notation is inconsistent: the text before Eq. (11) refers to the input standard deviation as σ, whereas Eq. (11) uses σx, and the noisy-case definitions of σs and σn do not state the independence or finite-variance assumptions. Please align the notation and place the assumptions with the definitions.","section":"1.5.6"},{"comment":"The density PN in Eq. (23) is not defined; please define it explicitly (presumably the noise PDF) before using it in the mutual-information expression.","section":"1.6.2"},{"comment":"There are several typos: \"see water\" should be \"sea water,\" \"Baltic See\" should be \"Baltic Sea,\" and the citation \"[P.521]\" does not match the \"ITU-R Recommendation P.527-6\" mentioned in the same sentence.","section":"1.8"},{"comment":"The maximum-likelihood expression in Eq. (2) is garbled; the minimization should be over b ∈ {0,1}^N and the norm should be written explicitly, for example as || y - Σ_{i=1}^N h_i b_i c_i ||^2.","section":"1.4.2"},{"comment":"The phrase \"Voltera series\" should be \"Volterra series.\"","section":"1.5.2"},{"comment":"The caveat in the last paragraph, that the distortion-plus-noise term is not generally Gaussian even in the Gaussian-input case, limits the usefulness of the SDNR formula for BER analysis; move this caveat immediately after Eq. (19) so that it appears together with the SDNR claim.","section":"1.5.6"},{"comment":"In Figures 9 and 10, the legend labels \"mq 3\" and \"mq 4\" are unclear; please use m_q = 3, m_q = 4 or define the notation in the caption.","section":"1.6.2"}],"recommendation":"major_revision","confidential_remarks":"This is a working-group whitepaper rather than a conventional research article. Its main value is documentary, and the heavy self-citation (BS22, BV22, BS24, Baj23, PR23b, etc.) is natural in that context but means the claims are largely the group's own reports rather than independently verified results. The Bussgang overstatement in Section 1.5.6 is the one substantive technical issue; it can be fixed locally. If the journal publishes position papers or comprehensive surveys, the manuscript is in scope after revision; otherwise it may be better placed in a venue for workshop or working-group reports."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a whitepaper from COST Action INTERACT WG2, not a research preprint. The useful thing is the breadth: it surveys waveforms, coding, MIMO, localization, and ISAC, with pointers to the group's own publications. Someone entering 6G signal processing would get a reasonable orientation. I would not look to it for new results.\n\nThe one piece of claimed novelty, the Bussgang decomposition in §1.5.6, is a small and mostly correct observation: for any signal s and noise n with finite variances, the output of a memoryless nonlinearity can be written y = αs·s + δ with δ uncorrelated with s, where αs is the linear MMSE coefficient. That is just orthogonal projection, and it does hold regardless of distribution as long as the moments exist. The stronger statements in that section—the factor (1 + σn²/σs²) in γs and the SDNR formula—require s and n to be independent (or at least uncorrelated) with finite second moments. The phrase “regardless of the distribution of signal and noise” overstates it; dependent or infinite-variance noise breaks the specific formulas. The authors are transparent that αs depends on the distributions, so the gap is an unqualified sentence rather than a hidden error. Small revision, not rejection.\n\nThe rest is a summary of the group's prior work, with heavy self-citation. That is normal for a whitepaper. I did not see circularity where the citations are used to derive new claims. The survey is competently written and the citations look appropriate. There are no reproducibility artifacts because there are no new experiments or code.\n\nThe soft spot is generic to the genre: the proposed directions are programmatic, with little critical comparison to alternatives. It is a position piece, so I do not penalize heavily, but it limits the counterfactual impact.\n\nSend it to peer review as a survey, but ask the authors to qualify the Bussgang passage. The audience is graduate students and researchers wanting a map of 6G signal processing, or people tracking COST INTERACT output. A legitimate survey, not a must-read.","headline":"A competent, broad 6G signal-processing survey from a COST working group; the only new math is a Bussgang note that is correct but oversold.","tokens_in":46580,"tokens_out":3160,"would_cite":false,"duration_ms":34910,"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":"A Bussgang decomposition applies to any signal-noise distribution.","keywords":["6G","physical layer","integrated sensing and communication","localization","Bussgang decomposition","waveform design","massive MIMO","physical layer security"],"falsifier":"Feed a memoryless nonlinearity a sum of a Bernoulli-distributed signal and $\\alpha$-stable noise with infinite variance, compute $\\alpha_s = \\mathbb{E}[ys]/\\mathbb{E}[s^2]$, and test whether $y - \\alpha_s s$ is uncorrelated with $s$; a nonzero correlation, or a divergence in $\\gamma_s$, would falsify the claimed general decomposition. Even within finite variance, a simulation where Eq. (19) fails to predict the measured signal-to-distortion-plus-noise ratio would settle the matter.","tokens_in":1623,"feed_emoji":"📡","tokens_out":7157,"duration_ms":153834,"temperature":0.7,"pith_summary":"This whitepaper argues that meeting 6G targets (1 Tbps data rates, 1 ms latency, 1000 km/h speeds, and integrated sensing) requires rethinking the physical layer. Its most concrete technical claim is that the Bussgang decomposition works for any signal and noise distribution, not just Gaussian. The paper derives that an output of a memoryless nonlinearity can be written as a scaled input plus uncorrelated distortion, with coefficients based on conditional moments, and it gives a signal-to-distortion-plus-noise ratio formula valid without Gaussianity. It also surveys and proposes work on waveforms, coding, massive MIMO, massive access, fronthaul compression, security, localization, and ISAC. A sympathetic reader should care because if the Bussgang generalization holds, nonlinear amplifiers can be modeled for non-Gaussian and non-OFDM waveforms that 6G may use.","feed_headline":"Bussgang decomposition shown to hold for any signal and noise","feed_subtitle":"A 6G whitepaper extends the classic nonlinear-amplifier model to non-Gaussian inputs, easing waveform and ISAC design.","key_machinery":"The carrying object is the Bussgang decomposition: for a memoryless nonlinearity $y = f(x)$, the output is split as $y = \\alpha_x x + \\delta_x$, where $\\delta_x$ is uncorrelated with $x$ and $\\alpha_x = \\mathbb{E}[yx]/\\mathbb{E}[x^2]$. The generalized version uses conditional moments $\\mu_y(s) = \\mathbb{E}[y|s]$ and $\\mu_{y^2}(s) = \\mathbb{E}[y^2|s]$, defining $\\alpha_s$ and $\\gamma_s$ through integrals over the signal distribution $p_s(s)$. This yields the distribution-free relation $\\gamma_s = \\gamma_x(1 + \\sigma_n^2/\\sigma_s^2)$ that anchors the SDNR formula.","core_discovery":"Section 1.5.6 claims that 'a version of the Bussgang decomposition applies in general, regardless of the distribution of signal and noise.' Writing the input to a memoryless nonlinearity as $x = s + n$ with zero-mean, independent, finite-variance signal and noise, the paper defines $\\alpha_s = \\mathbb{E}[ys]/\\mathbb{E}[s^2]$ and shows that $y - \\alpha_s s$ is uncorrelated with $s$. It also defines $\\gamma_s = \\mathbb{E}[y^2]/\\mathbb{E}[s^2]$ and shows that $\\gamma_s = \\gamma_x (1 + \\sigma_n^2/\\sigma_s^2)$ holds for arbitrary distributions, while the classical simplification $\\alpha_s = \\alpha_x$ requires Gaussianity. The resulting signal-to-distortion-plus-noise ratio in Eq. (19) is claimed to hold generally, with Eq. (20) as the Gaussian special case. The paper notes that the distortion-plus-noise term is not usually Gaussian even when the input is.","pith_inferences":["A direct test would feed a nonlinear amplifier with a non-Gaussian signal plus heavy-tailed interference and check whether $y - \\alpha_s s$ remains uncorrelated with $s$; residual correlation would pinpoint where independence or finite-variance premises break.","Quantization is itself a memoryless nonlinearity, so the generalized decomposition could be applied to optimize fronthaul quantization intervals without assuming Gaussian signals.","If $\\alpha_s$ must be computed from actual distributions, data-driven estimates of $\\alpha_s$ may become useful for neural-network receivers and ISAC distortion models."],"forward_implications":["Non-Gaussian waveforms such as OTFS, chirp, and ISAC signals can be analyzed with a linear-plus-uncorrelated-distortion model for nonlinear amplifiers.","The SDNR formula gives a way to choose transmit power, backoff, and power allocation under arbitrary signal and noise statistics, assisting distortion-aware precoding and energy-efficient designs.","In the Gaussian case, results reduce to the classical Bussgang formulas, so existing receiver designs remain valid as a special case."],"supporting_citations":[{"why":"This is the original Bussgang theorem that supplies the Gaussian-case cross-moment relation behind the classical decomposition.","marker":"[Bus52]"},{"why":"This source provides the signal-to-distortion-plus-noise ratio formula that the paper shows does not depend on the Gaussian assumption.","marker":"[Zil10]"}],"fun_headline_variants":["Bussgang decomposition goes beyond Gaussian inputs","6G whitepaper generalizes Bussgang decomposition","Non-Gaussian Bussgang: new signal processing tool","Bussgang holds for any signal, noise distribution"],"cache_read_input_tokens":48768,"weakest_assumption_plain":"The generalization assumes the wanted signal and noise are independent zero-mean random variables with finite second moments; if the noise is heavy-tailed, correlated with the signal, or has infinite variance, the derived relations for $\\alpha_s$ and $\\gamma_s$ no longer follow.","fun_headline_variants_meta":{"raw":{"variants":["Bussgang decomposition goes beyond Gaussian inputs","6G whitepaper generalizes Bussgang decomposition","Non-Gaussian Bussgang: new signal processing tool","Bussgang holds for any signal, noise distribution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1432,"prompt_tokens":969,"completion_tokens":463,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":400}},"tokens_in":585,"tokens_out":463,"duration_ms":4670,"temperature":1.0,"reasoning_tokens":400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:43:29.563122+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed a memoryless nonlinearity a sum of a Bernoulli-distributed signal and $\\alpha$-stable noise with infinite variance, compute $\\alpha_s = \\mathbb{E}[ys]/\\mathbb{E}[s^2]$, and test whether $y - \\alpha_s s$ is uncorrelated with $s$; a nonzero correlation, or a divergence in $\\gamma_s$, would falsify the claimed general decomposition. Even within finite variance, a simulation where Eq. (19) fails to predict the measured signal-to-distortion-plus-noise ratio would settle the matter.","supporting_citations":[],"review_version":1}