{"id":"c4571e7c-d90c-4e96-b8e9-edb1ca043c95","arxiv_id":"2505.17877","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A proposed unified lower bound for active noise cancellation error combines a mutual-information term and a spectral-support term, but the information-theoretic term is derived incorrectly and is false.","lead":"This paper claims to prove a theoretical lower bound on how well any active noise cancellation system can suppress noise, combining a mutual-information term and a spectral-support term. The support term is reasonable, but the information-theoretic term rests on a derivation error and does not hold even for simple Gaussian noise, so the central claim is not established.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 1's linear information bound is algebraically invalid: the correct rate-distortion exponent gives exp(-2I), and the claimed 1 - I/H bound is violated by a simple Gaussian example.","rationale":"The paper's central claim is a unified lower bound on NMSE for any ANC algorithm. The information-theoretic component, Lemma 1, is the load-bearing part: without it, Theorem 2 reduces to the support-based bound alone, which the authors themselves note is insufficient to characterize algorithmic limits. The derivation of Lemma 1 contains a decisive algebraic error. Starting from the Shannon lower bound, the correct relation is D >= (1/2pi e) exp(2(H(d) - I(y;d))) = sigma_d^2 exp(-2I) for a Gaussian source. The proof instead writes the exponent as -I/H and then takes a first-order Taylor expansion of exp(-I/H) to obtain 1 - I/H. This is not a valid lower bound: exp(-2I) is not bounded below by 1 - I/H in the relevant regime. The Gaussian example in the concrete test demonstrates the failure numerically and exactly. The example is not exotic; it is the canonical additive-noise setting with a linear optimal estimator, and it directly violates the claimed inequality. The proof also contains a secondary inconsistency in Eq. (8), which adds 10log10(||P||^2) after normalizing by sigma_d^2, though the paper defines NMSE as E[|e|^2]/E[|d|^2]. The empirical results cannot rescue the theory because the bound is computed from the same algorithm outputs via KDE-based mutual information estimates, and no code or error bars are provided; even if the plots show the bound below observed NMSE, a false bound can still appear to hold on finite samples. The support-based Lemma 2 is correct and is a genuine contribution, but it is not enough to carry the paper's claims. Therefore the verdict should remain REJECT.","tokens_in":10773,"tokens_out":4490,"duration_ms":36982,"concrete_test":"Evaluate the Gaussian counterexample exactly: set d ~ N(0,1), y = d + n, n ~ N(0,1) independent. Compute H(d) = 0.5 log(2pi e) = 1.4189 nats, I(d;y) = 0.5 log(2) = 0.3466 nats, and the MMSE estimate E[d|y] = y/2, which yields NMSE = 0.5. Plug these into Eq. (2): the claimed lower bound is 1 - 0.3466/1.4189 = 0.7557, and the inequality 0.5 >= 0.7557 fails. This single check settles that Lemma 1 is false, independent of any empirical validation.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim collapses at Lemma 1 (Eq. 2). From the Shannon lower bound and I(y;d) >= R(D), the correct implication is D >= (1/2pi e) exp(2(H(d)-I(y;d))) = sigma_d^2 exp(-2I) for a Gaussian source. The proof then replaces exp(-2I) with 1 - I/H, which is not a valid lower bound: for d ~ N(0,1), y = d + n with n ~ N(0,1) independent, I = 0.3466 nats, H = 1.4189 nats, so the claimed bound is 0.7557, while the true minimum NMSE is 0.5 (the MMSE estimator gives E[d|y] = y/2). Since 0.5 < 0.7557, Eq. (2) is false. The Taylor step is also misstated: exp(-x) >= 1 - x would give D >= sigma_d^2(1 - 2I), not sigma_d^2(1 - I/H), and 2I is not dimensionless in the normalized ratio. Additionally, Eq. (8) appends +10log10(||P||^2) after already normalizing by sigma_d^2, which is inconsistent with the stated NMSE definition. The support-based Lemma 2 is sound, but Theorem 2's max inherits a false first argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified theoretical lower bound on the normalized mean-squared error (NMSE) of active noise cancellation systems. The bound in Theorem 2 is the maximum of two terms: an information-theoretic term sigma_d^2 (1 - I(y;d)/H(d)) from Lemma 1, and a spectral-support term from Lemma 2 measuring the disturbance power in frequencies where the secondary path has no gain. The authors validate the bound on the NOISEX dataset with three deep-learning ANC baselines over several reverberation times. The support-based Lemma 2 is a correct elementary observation, but the central information-theoretic Lemma 1 is false, and the dB conversion in Eq. (8) is inconsistent. Because Theorem 2 takes the maximum of these two terms, the paper's main claim is not established.","tokens_in":11199,"tokens_out":7128,"duration_ms":71504,"significance":"A valid, algorithm-independent lower bound on ANC performance would be a useful benchmark for the deep-learning ANC literature, and the empirical comparison across reverberation times is a sensible design. The support-based bound (Lemma 2) is a sound and simple frequency-domain constraint, and the authors are right that it is model-independent in its idealized form. However, the information-theoretic component, which is the paper's main novelty, is invalid: Lemma 1 is contradicted by a Gaussian counterexample, and the proof rests on an incorrect Taylor truncation. The paper also ships no code or reproducibility artifacts, and the experimental validation is weakened by the data-dependent estimation of I(y;d) from the very signals used to measure NMSE. The contribution therefore does not meet the standard for a theoretical result in information theory.","major_comments":[{"comment":"Lemma 1 is false. Take d ~ N(0,1) and y = d + n with n ~ N(0,1) independent. The minimum mean-squared error estimator is E[d|y] = y/2, giving NMSE = 0.5. For this Gaussian pair, H(d) = (1/2) log(2*pi*e) approx 1.419 nats and I(y;d) = (1/2) log 2 approx 0.347 nats, so the claimed right-hand side sigma_d^2 (1 - I/H) approx 0.756. Since 0.5 < 0.756, Eq. (2) is violated. The proof's transition from Eq. (5) to Eq. (6) is invalid: the Shannon lower bound yields D >= sigma_d^2 exp(-2I), and replacing exp(-2I) by 1 - I/H is neither a valid Taylor truncation nor the correct normalized exponent; in fact exp(-x) >= 1 - x would give a different linear term. Theorem 2 inherits this false first argument through its maximum.","section":"Section 4, Lemma 1 and Eq. (2)"},{"comment":"The conversion to dB double-counts the primary path norm. With NMSE = E[|e|^2]/E[|d|^2], Eq. (2) immediately gives NMSE >= 1 - I(y;d)/H(d) in linear units; there is no additional 10 log10(||P||_2^2) term in the decibel expression. Equation (8) appends +10 log10(||P||_2^2) after the variance ratio has already been normalized by sigma_d^2 = (1/2*pi) integral |P|^2 S_xx d*omega. The experiments in Section 8 report boundaries computed from Eq. (8), so the information-theoretic curves are offset by an unjustified constant that is not part of the stated NMSE definition.","section":"Section 4, Eq. (8)"},{"comment":"The information-theoretic bound is not a model-independent ceiling. Lemma 1 bounds sigma_e^2 using I(y;d), which is a functional of the algorithm's output y(n); in the experiments I(y;d) is estimated from the same d(n) and y(n) that define the measured NMSE (Sections 7 and A.1). The manuscript itself states in Section 4 that the bound is 'inherently dependent on the specific algorithm.' A lower bound that is a function of the algorithm's own statistics cannot support the claimed 'theoretical ceiling on the NMSE attainable by any ANC algorithm' (Abstract and Theorem 2) unless the mutual information is extremized over the admissible class of y(n) or replaced by a source/channel quantity independent of the implementation. The current derivation gives no such extremization.","section":"Sections 4, 7, and A.1"},{"comment":"Lemma 2 is sound as stated for exact spectral supports, but the numerical implementation does not evaluate this quantity. Appendix A.2 defines support by a magnitude threshold (e.g., 45 dB below the peak) and treats all bins below the threshold as having zero gain; with finite-length room impulse responses the true support is generically the full frequency band, so supp(P)\\supp(S) is empty and Lemma 2 is vacuous. The reported support-based boundary is therefore an artifact of the threshold choice, and no sensitivity analysis over the threshold is provided. Consequently the experimental 'support-based bound' in Figures 2 and 4 is not the model-independent bound of Lemma 2.","section":"Section 5, Lemma 2 and Appendix A.2"}],"minor_comments":[{"comment":"The sentence 'Recall, that for gaussian noise, the entropy is maximazied...' contains typos and an imprecise statement; the equality sigma_d^2 = (1/(2*pi*e)) e^{2H(d)} holds only for Gaussian sources, not for general noise.","section":"Section 4, proof of Lemma 1"},{"comment":"The phrase 'for negative exponent it is upper bound' is unclear and, as written, incorrect; the Taylor step needs to be re-derived with the exponent I/H rather than 2I.","section":"Section 4, after Eq. (5)"},{"comment":"The proof would benefit from explicitly stating that supp(P)\\supp(S) is interpreted as the set of frequencies where S has exactly zero gain; otherwise for typical finite-length impulse responses the bound is vacuous.","section":"Section 5, Lemma 2"},{"comment":"The limitation paragraph admits that insufficient sampling 'distorts the bound calculations,' but no convergence analysis or bias correction for the KDE/MI estimates is provided; this should be quantified if the empirical validation is to be meaningful.","section":"Section 9"},{"comment":"'Ressources' should be 'Resources'; the figure captions should state whether error bars or repeated trials are included.","section":"Section 7"}],"recommendation":"reject","confidential_remarks":"The central result is false, so rejection is appropriate. A repair would require replacing Lemma 1 with a correct rate-distortion bound and re-deriving the unified statement; this is beyond the scope of a revision. The manuscript does not provide code or data, and the empirical section does not include error bars or a sensitivity analysis for the threshold used in the support-based bound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the thesis is false. The paper's central Lemma 1 is invalid, and Theorem 2 inherits the failure. The Gaussian counterexample in the stress-test note is decisive: for d~N(0,1), y=d+n, n~N(0,1), the true minimum NMSE is 0.5 while the claimed bound is 0.756. A universal lower bound that is violated by the simplest Gaussian case is not a bound.\n\nThe framing deserves some credit. Pairing an information-theoretic ceiling with a spectral-support ceiling for ANC is a reasonable idea, and the support-based half (Lemma 2) is derived correctly and is genuinely model-independent. The empirical section also does more than the field often does: three baselines, three noise types, five reverberation times. The related-work coverage is adequate, and the self-citation is to a co-author's recent ANC paper, which is legitimate.\n\nThe problems are load-bearing. Lemma 1 starts from the Shannon lower bound, which correctly gives D >= sigma_d^2 exp(-2I), then replaces exp(-2I) with 1 - I/H. That replacement is algebraically invalid: the Taylor inequality runs the wrong way (exp(-x) >= 1 - x), and the exponent loses the factor of 2 and gains an H-normalization. The paper's own Eq. (8) additionally appends 10 log10(||P||^2) after normalization by sigma_d^2, which is inconsistent with the stated NMSE definition. And there is a circularity: I(y; d) is estimated from the same d(n) and y(n) used to compute the residual error, so the information-theoretic 'bound' is a data-dependent restatement of the algorithm's own statistics, not an independent theoretical ceiling. None of that is fixed by the empirical plots.\n\nThe support-based part could have been a short note on its own; as packaged, it does not rescue the claims.\n\nWho gets value: someone wanting a quick map of recent deep-learning ANC methods and the standard simulator setup would get something from the survey and the experimental description. As a theory paper it fails on its own terms.\n\nRecommendation: desk-reject. The leading theorem is false and easily falsified; referee time would just reproduce the counterexample. If the authors strip out Lemma 1 and publish the support-based bound with the multi-reverberation-time evaluation as an empirical note, that could be worth a look.","headline":"The central Lemma 1 is algebraically invalid and is violated by a simple Gaussian example, so the paper's unified lower bound does not hold.","tokens_in":11590,"tokens_out":2772,"would_cite":false,"duration_ms":27024,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A17","94A12"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that any active noise cancellation algorithm has its normalized residual error bounded below by the maximum of an information-theoretic term and a spectral-support term.","keywords":["active noise cancellation","mutual information lower bound","normalized mean squared error","rate-distortion theory","spectral support","reverberation","deep learning ANC"],"falsifier":"For a scalar Gaussian channel with d ~ N(0,1) and y = d + n where n ~ N(0,1) is independent, compute the mutual information I(y;d) = 0.5 ln 2 and the entropy H(d) = 0.5 ln(2πe). The claimed bound gives NMSE ≥ 1 - I/H ≈ 0.756, whereas the true minimum mean-squared error estimate of d from y, the conditional mean, has NMSE = 0.5. A direct calculation or a small simulation that estimates NMSE from samples will show the bound is violated, contradicting Lemma 1.","tokens_in":10598,"feed_emoji":"🎧","tokens_out":6258,"duration_ms":45856,"temperature":0.7,"pith_summary":"The paper sets out to find a theoretical floor for the performance of active noise cancellation (ANC) systems. It proposes a unified lower bound on the normalized mean squared error (NMSE) that combines two independent constraints: how much information the anti-noise signal carries about the disturbance, and how much disturbance power lies in frequency bands the cancellation path cannot touch. If the bound is correct, it gives every ANC algorithm—classical or deep learning—a concrete benchmark to measure against, and tells engineers whether to improve the algorithm or the physical hardware. The authors test the bound on the NOISEX dataset with three deep ANC models over several reverberation times and find their results lie above the claimed floor.","feed_headline":"New bound sets the floor for noise-cancelling error","feed_subtitle":"A unified mutual-information and spectral-support limit tells when an ANC algorithm has hit its theoretical ceiling.","key_machinery":"The engine of the paper is Lemma 1, an information-theoretic inequality: σ_e² ≥ σ_d²(1 - I(y;d)/H(d)). It is obtained by starting from the Shannon lower bound on the rate-distortion function, expressing distortion in terms of the entropy H(d) and mutual information I(y;d), then linearizing the resulting exponential expression with a first-order Taylor expansion. The second component is the spectral-support ratio, the fraction of disturbance energy in the frequency set supp(P) \\ supp(S), computed directly from path transfer functions. The unified bound is the maximum of these two separate floors.","core_discovery":"The central claim is Theorem 2: for any ANC system with a primary path P(z) and secondary path S(z), the NMSE in decibels must satisfy NMSEdB ≥ max{10 log10(1 - I(y;d)/H(d)), 10 log10(∫_{supp(P)\\supp(S)} S_dd($e^{{jω}}$) dω / ∫ S_dd($e^{{jω}}$) dω)}, where I(y;d) is the mutual information rate between the synthesized signal and the disturbance and H(d) is the disturbance's differential entropy rate. The first term, derived in Lemma 1, states that the residual error power is at least σ_d²(1 - I(y;d)/H(d)); the second term is the portion of disturbance power concentrated in frequencies the secondary path cannot reproduce. Taking the maximum of the two yields a single, model-independent ceiling that the paper argues applies to every cancellation algorithm, including deep networks trained end-to-end.","pith_inferences":["The critical step is the Taylor linearization from an exponential bound to 1 - I/H; because exp(-x) ≥ 1 - x, the linear form is a weaker bound than the exponential, but the paper's specific replacement of exp(-2I) with exp(-I/H) is where a reader should check the direction of the inequality.","A simple scalar Gaussian test—d ~ N(0,1), y = d + n with n ~ N(0,1)—gives I/H ≈ 0.244 and a claimed floor of about 0.756, while the true minimum mean-squared error is 0.5, indicating the linear bound may not hold as stated for Gaussian sources.","The paper's numerical estimation of mutual information relies on kernel density estimates from finite samples, which the authors acknowledge as a limitation; the bound's practical reliability therefore depends on sample size and bandwidth choices.","The support-based term is robust and likely to survive scrutiny, but the full unified bound inherits whatever flaws the information-theoretic term possesses."],"forward_implications":["If the bound holds, any ANC model whose NMSE sits far above the ceiling still has headroom for algorithmic improvement, while a model near the bound is limited by information or hardware constraints.","The identity of the dominant term directs engineering effort: a large information term signals the algorithm is not extracting enough about the disturbance, while a large support term signals the actuator or secondary path must be redesigned.","The support-based term is entirely model-independent, so it can be computed before training any neural controller and used as a sanity-check lower bound.","The information-theoretic term grows with reverberation time, providing a quantitative explanation for why ANC is harder in more reverberant rooms.","The bound offers a unified reporting standard for the ANC literature, allowing different methods to be compared against a common theoretical reference rather than only against each other."],"supporting_citations":[{"why":"Provides the NOISEX dataset (babble, engine, factory noises) used to empirically evaluate the bound against deep ANC baselines.","marker":"[42]"},{"why":"Supplies the room impulse response generator used to simulate primary and secondary acoustic paths with varying reverberation times.","marker":"[43]"},{"why":"DeepANC is one of the three deep learning ANC baselines whose measured NMSE is compared against the proposed bound.","marker":"[12]"},{"why":"DeepASC, the most recent deep ANC model, is reproduced and used as a baseline to test the bound's tightness.","marker":"[14]"},{"why":"The Attentive Recurrent Network (ARN) baseline, as cited in the experiments, provides another test point for the unified bound.","marker":"[33]"}],"fun_headline_variants":["ANC's theoretical floor: when error can't drop","Max-bound reveals noise-cancelling limits","Optimal ANC? New bound says when","Mutual info + spectral support = ANC ceiling","Noise cancellation hits a hard floor"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes that the first-order Taylor approximation of an exponential rate-distortion bound, turned into the linear inequality σ_e² ≥ σ_d²(1 - I/H), is a valid lower bound on the residual error; if that linearization direction is wrong, the information-theoretic half of the unified bound collapses.","fun_headline_variants_meta":{"raw":{"variants":["ANC's theoretical floor: when error can't drop","Max-bound reveals noise-cancelling limits","Optimal ANC? New bound says when","Mutual info + spectral support = ANC ceiling","Noise cancellation hits a hard floor"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000866,"raw_usage":{"total_tokens":3734,"prompt_tokens":903,"completion_tokens":2831,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":2763}},"tokens_in":519,"tokens_out":2831,"duration_ms":16297,"temperature":1.0,"reasoning_tokens":2763,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:39:35.023296+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a scalar Gaussian channel with d ~ N(0,1) and y = d + n where n ~ N(0,1) is independent, compute the mutual information I(y;d) = 0.5 ln 2 and the entropy H(d) = 0.5 ln(2πe). The claimed bound gives NMSE ≥ 1 - I/H ≈ 0.756, whereas the true minimum mean-squared error estimate of d from y, the conditional mean, has NMSE = 0.5. A direct calculation or a small simulation that estimates NMSE from samples will show the bound is violated, contradicting Lemma 1.","supporting_citations":[{"cited_title":"Assessment for automatic speech recognition: Ii","cited_arxiv_id":null,"evidence_quote":"Provides the NOISEX dataset (babble, engine, factory noises) used to empirically evaluate the bound against deep ANC baselines."},{"cited_title":"Image method for efficiently simulating small-room acoustics","cited_arxiv_id":null,"evidence_quote":"Supplies the room impulse response generator used to simulate primary and secondary acoustic paths with varying reverberation times."},{"cited_title":"Deep anc: A deep learning approach to active noise control","cited_arxiv_id":null,"evidence_quote":"DeepANC is one of the three deep learning ANC baselines whose measured NMSE is compared against the proposed bound."},{"cited_title":"Deep mcanc: A deep learning approach to multi-channel active noise control","cited_arxiv_id":null,"evidence_quote":"The Attentive Recurrent Network (ARN) baseline, as cited in the experiments, provides another test point for the unified bound."}],"review_version":1}