{"id":"5d0318a3-0775-45cb-a180-fdd7ae05b18d","arxiv_id":"2501.01506","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper derives MSE guarantees for dithered modulo ADCs, claiming OF>3 and b>3 suffice for 1/OF^3 scaling, but the central proof omits a noise-leakage term.","lead":"An oversampled modulo ADC with a 1-bit folding marker is claimed to suppress quantization noise at rate 1/OF^3, far better than a conventional ADC's 1/OF. The paper derives this in a dithered-quantization framework and adds an OMP-based recovery for systems without the marker, but the central proof contains an unhandled noise-leakage term.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof drops the V_S^† V_{S^c} Δε_{S^c} leakage term; b>3 alone does not guarantee the rounding step, so the central O(1/OF^3) guarantee is unproven.","rationale":"Agree with the reader's identification: the weakest point is the unproven orthogonality in Appendix B. The correct LS error is Δε_S + V_S^† V_{S^c} Δε_{S^c}; no argument in the paper bounds the second term. The central theorem, Theorem 1, depends on exact rounding, so this is load-bearing. The simulations matching Eq. (11) are not enough: the theorem claims a deterministic guarantee for all bandlimited signals, and a single counterexample to the proof would invalidate it. The OMP result (Theorem 3) inherits the flaw and also relies on an unverified RIP assumption. The case studies are extensions of the same argument. No issue found with Lemma 1's rank argument; the concern is specifically the noise leakage. Therefore the reader's REJECT verdict stands; no adjustment needed.","tokens_in":22284,"tokens_out":6649,"duration_ms":57744,"concrete_test":"Using the exact matrices from Simulation-1 (or any admissible f, OF, b), compute V, S, and the operator norm ‖V_S^† V_{S^c}‖∞; if ‖V_S^† V_{S^c}‖∞ ≥ (2^b−2)/6, then the worst-case leakage 6λ/(2^b)‖V_S^† V_{S^c}‖∞ can reach or exceed λ′, so the rounding step in Theorem 1 is not guaranteed for b>3. Alternatively, for a fixed noise realization, compute the full error vector V_S^† V Δε and test whether any entry exceeds λ′; if yes, the simulated MSE should deviate from Eq. (11). Both checks settle whether the dropped cross term breaks the claimed sufficient condition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Appendix B, the proof of Theorem 1 writes ∥V†_S y − Δz_S∥∞ = ∥V†_S V_S(Δz_S + Δε_S) − Δz_S∥∞. This implicitly replaces the true observation y = V_S Δz_S + V Δε by V_S(Δz_S + Δε_S). Since Δε is nonzero on all N samples, the least-squares estimate is Δz_S + V†_S V Δε = Δz_S + Δε_S + V†_S V_{S^c} Δε_{S^c}. The cross term V†_S V_{S^c} Δε_{S^c} is not zero: V has only the K out-of-band Fourier rows, so columns in S are not orthogonal to columns in S^c. No bound on this term is provided. The subsequent rounding argument (inequality (24)) only controls 6λ/(2^b) = 6λ′/(2^b−2), the ℓ∞ norm of Δε, not the leakage through V†_S V_{S^c}. Consequently b > 3 is not shown to be sufficient for perfect unfolding, Eq. (11) does not follow, and Theorem 3 and Corollary 1 inherit the gap. The numerical matches in Figs. 4–7 are for particular signals and do not establish the universal guarantee.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies modulo analog-to-digital converters under dithered quantization. For a modulo ADC with an extra 1-bit folding-information signal, it claims that OF > 3 and b > 3 are sufficient for exact unfolding and derives the mean squared error MSE = c^2 / (OF (2^b - 2)^2) * (||f||_inf / (OF - 2))^2, which is O(1/OF^3). For the system without side information, it proposes an OMP-based support recovery algorithm and claims the same MSE under OF > 3 and b > 3 + log_2(zeta), conditional on a restricted isometry property (RIP) of the sensing matrix. The results are extended to the simultaneous acquisition of strong and weak signals in disjoint frequency bands, and several numerical experiments are reported.","tokens_in":22513,"tokens_out":10271,"duration_ms":103265,"significance":"If the O(1/OF^3) quantization-noise scaling were rigorously established, this would be an important contribution to modulo sampling: it would show that, for the same amplitude-quantization bit budget, a modulo ADC strictly outperforms a conventional oversampled ADC. The paper is clearly organized, the comparison with conventional ADCs is thoughtfully set up, and the numerical study covers the main claimed scenarios, including dither-free operation and weak/strong components. However, the central guarantees rest on proof steps that are not justified as written, so the stated contribution is not currently established.","major_comments":[{"comment":"The proof of Theorem 1 is invalid because it replaces the true observation model y = V_S Delta z_S + V Delta epsilon with V_S(Delta z_S + Delta epsilon_S). From the definition of y in Eq. (6), Delta epsilon is present at every time sample, so the least-squares estimate is Delta z_S + V_S^dagger V Delta epsilon = Delta z_S + Delta epsilon_S + V_S^dagger V_{S^c} Delta epsilon_{S^c}. The cross term V_S^dagger V_{S^c} Delta epsilon_{S^c} is not zero and is never bounded; the argument around inequality (24) controls only the l_infinity norm of Delta epsilon. Consequently the rounding step is not guaranteed by b > 3, Eq. (11) does not follow, and Theorem 3 and Corollary 1 inherit the gap.","section":"Section III-A, Appendix B, Eq. (7)"},{"comment":"The proof of Lemma 1 asserts that the columns of V_S are linearly independent because they are columns of the Fourier basis, and therefore |S| <= K suffices for full column rank. This statement is false in general for a partial Fourier matrix: after restriction to a subset of rows, distinct Fourier columns can become linearly dependent. The dimensional inequality in Eq. (22)-(23) does not establish full column rank, so OF > 3 alone does not justify the unique-recovery step in Eq. (7).","section":"Appendix A, Lemma 1"},{"comment":"The OMP-based guarantee is conditional on an RIP hypothesis on the matrix V that is not verified for the specific deterministic OOB-frequency submatrix used in the paper. The text acknowledges this limitation, but the subsequent appeal to random partial DFT matrices does not apply because V is not random. Thus the paper does not establish the claimed sufficient conditions OF > 3 and b > 3 + log_2(zeta) for the no-side-information system; the bit condition in Eq. (16) depends on an unknown RIP constant delta_{L0+1}.","section":"Section III-B, Theorem 3"},{"comment":"The numerical experiments are consistent with the formulas, but they use particular signals and averaging over random coefficients and time offsets; they do not probe the leakage term V_S^dagger V_{S^c} Delta epsilon_{S^c} or the possible rank deficiency of V_S. The simulations therefore do not fill the gap in the theoretical proof.","section":"Section V and overall"}],"minor_comments":[{"comment":"The abstract uses delta in the condition b > 3 + log_2(delta), while the body and Theorem 3 use zeta for the same penalty term; the notation should be unified.","section":"Abstract and Section III-B"},{"comment":"The text writes 'Reimann-Lebesgue lemma'; this should be 'Riemann-Lebesgue lemma'.","section":"Section II-A"},{"comment":"The phrase 'sufficient condition of OF >= 3' conflicts with the OF > 3 condition used in Theorem 1 and Lemma 1; the threshold should be stated consistently.","section":"Section III-A"}],"recommendation":"reject","confidential_remarks":"I concur with the reader's assessment: the central proof gap in Appendix B is load-bearing, and the rank assertion in Appendix A is also unjustified. The paper extends the authors' ISIT conference work and leans on their own sparsity bound [20], but the new results are not supported by the current proofs. In my view this is a reject rather than a minor revision, because the claimed universal sufficient conditions would require substantially different arguments or additional hypotheses to be made correct."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this if you work on unlimited sampling or high-DR ADCs. The genuinely new content is the OMP-based recovery for the modulo ADC without the 1-bit folding marker (Section II-C, Theorem 3) and the weak/strong simultaneous acquisition analysis (Section IV). The 1-bit side-information MSE analysis is a journal expansion of the authors' ISIT 2024 paper, and they disclose that. The numerical section is more honest than most: it tests the violating regime, includes dither-free runs, and compares against conventional ADC baselines. Good credit where due.\n\nThe soft spot is load-bearing. In Appendix B the proof of Theorem 1 writes the least-squares error as V_S^† V_S(Δz_S+Δε_S) - Δz_S, which assumes y has noise only on the folding support. But y also contains Δε_{S^c} on all non-folding samples. The actual error is Δε_S + V_S^† V_{S^c}Δε_{S^c}. The cross term is not zero for a partial Fourier matrix and no bound on it is provided. Full column rank of V_S does not imply orthogonality between the selected and unselected columns. So b>3 does not follow as a sufficient condition for exact rounding, and Eq. (11) is not established. The stress-test note is correct.\n\nTheorem 3 inherits that gap, and it also assumes an RIP condition on the deterministic matrix V that the paper does not prove. The authors acknowledge the RIP constant cannot be evaluated; the cited random partial Fourier results do not apply to their fixed V. The OMP stopping analysis is conditional on an unverified assumption. This is not a manufactured flaw: it is the step that carries the O(1/OF^3) claim.\n\nThe numerical MSE curves match the predicted formula for the tested signals, and the ISIT 2024 result is probably fine. But simulations over a few random signals cannot repair a missing universal bound. If the leakage term can be bounded—or an assumption added that controls it—the main theorem can be salvaged. As written, the central guarantee is unproven.\n\nWho is this for: researchers in modulo sampling and quantization-aware ADC theory. It deserves a serious referee rather than a desk reject, because the framework is clean and the gap is checkable. I would send it out with the expectation of heavy revision. My own verdict is skeptical: I would not cite the journal version until Theorem 1 is fixed.","headline":"The O(1/OF^3) guarantee rests on a dropped cross-term in the least-squares step; the paper is a useful extension but Theorem 1 as written is not proven.","tokens_in":23153,"tokens_out":3753,"would_cite":false,"duration_ms":38048,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A20","94A12"],"pacs":[],"model":"deepseek-v4-flash","headline":"A modulo ADC that folds the input before quantization can, with oversampling factor above 3 and more than 3 amplitude bits, achieve mean-squared error $O(\\mathrm{OF}^{-3})$, strictly better than a conventional ADC's $O(\\mathrm{OF}^{-1})$…","keywords":["modulo ADC","unlimited sampling","quantization noise","dithered quantization","oversampling","orthogonal matching pursuit","bandlimited signals","weak and strong signals"],"falsifier":"Run the modulo ADC recovery with b=4 and OF=4 on a bandlimited signal with many closely spaced folds, and compare the simulated MSE to the Theorem 1 curve; then compute the leakage norm $\\|V_S^\\dagger V_{S^c}\\Delta\\epsilon_{S^c}\\|_\\infty/\\lambda'$ for the same signal. If the leakage ratio is not negligible, or if the simulated MSE stays above the predicted $O(\\mathrm{OF}^{-3})$ curve by more than the filtered quantization-noise power, the exact-rounding step fails and the sufficient condition in Theorem 1 does not hold.","tokens_in":21937,"feed_emoji":"📉","tokens_out":11803,"duration_ms":109431,"temperature":0.7,"pith_summary":"The paper sets out to prove that a modulo ADC—one that folds a large-amplitude signal through a modulo operator before quantizing—does more than avoid clipping: it actively suppresses quantization noise. With an extra 1-bit signal that marks where folds occur, the authors show that oversampling factor $\\mathrm{OF}>3$ and quantizer resolution $b>3$ guarantee exact unfolding after a rounding step, so the only remaining error is lowpass-filtered quantization noise. The resulting MSE is $$\\mathrm{MSE} = \\frac{$c^{2}$}{\\mathrm{OF}(2^b-2)^2}\\left(\\frac{\\|f\\|_\\infty}{\\mathrm{OF}-2}\\right)^2,$$ scaling as $O(\\mathrm{OF}^{-3})$, while a conventional ADC with the same amplitude-quantization bits scales only as $O(\\mathrm{OF}^{-1})$. The paper further claims the same MSE for an OMP-based recovery that does not need the folding flag, at the price of the stronger bit condition $b>3+\\log_2(\\zeta)$. A sympathetic reader should care because, if true, this gives a concrete engineering recipe: oversample moderately and fold the input rather than buying more bits, to capture high-dynamic-range signals without clipping.","feed_headline":"Modulo ADC turns oversampling into cubed noise gains","feed_subtitle":"With OF > 3 and more than 3 amplitude bits, folding beats a conventional ADC that spends the same bits on amplitude.","key_machinery":"The argument has three load-bearing parts. First, the dithered quantization framework: a triangle dither with the right amplitude makes the b-bit quantization noise white with variance $\\lambda^2/2^{2b}$, so the lowpass-filtered quantization power is known exactly. Second, sparsity of the first-order difference of the modulo residue: the folding positions form a sparse set $S$, and the out-of-band DFT of the folded samples yields the sensing equation $y = V_S \\Delta z_S$ plus filtered quantization noise; Lemma 1 shows $V_S$ is full column rank once $\\mathrm{OF}>3$ and $\\lambda' \\ge \\|f\\|_\\infty/(\\mathrm{OF}-2)$. Third, exact rounding: the least-squares estimate $\\Delta\\hat{z}_S = V_S^\\dagger y$ is claimed to stay within $\\lambda'$ of the true multiples of $2\\lambda'$, so rounding reproduces $z[n]$ exactly when $b>3$; for the system without the flag, OMP with a bounded-noise stopping rule first recovers $S$, using the fact that the smallest nonzero entry of $\\Delta z$ has magnitude $\\lambda'$ and an RIP condition on $V$.","core_discovery":"On the paper's own terms, the central discovery is that quantization noise in a modulo ADC can decay as the inverse cube of the oversampling factor rather than the inverse first power. Theorem 1 states that with modulo threshold $\\lambda' = c\\|f\\|_\\infty/(\\mathrm{OF}-2)$ for $1 \\le c \\le \\mathrm{OF}-2$, with $\\mathrm{OF}>3$, $b>3$, and negligible truncation, the reconstruction error of the modulo ADC with 1-bit folding information is exactly the filtered in-band quantization noise, $\\mathrm{MSE} = \\frac{c^2}{\\mathrm{OF}(2^b-2)^2}\\left(\\frac{\\|f\\|_\\infty}{\\mathrm{OF}-2}\\right)^2$. A conventional ADC using the same bits for amplitude quantization has $\\mathrm{MSE} = \\|f\\|_\\infty^2/(\\mathrm{OF}(2^b-2)^2)$, which is strictly larger for every $\\mathrm{OF}>3$ because of the extra $(\\mathrm{OF}-2)^{-2}$ factor. Theorem 3 extends the same MSE expression to the modulo ADC without folding information, assuming OMP recovers the folding support under condition $b>3+\\log_2(\\zeta)$. Corollary 1 carries the result to simultaneous weak/strong band acquisition, giving a per-component NMSE bound that also decays as $O(\\mathrm{OF}^{-3})$.","pith_inferences":["If the $O(\\mathrm{OF}^{-3})$ law survives hardware non-idealities, the practical design rule is to oversample rather than add bits: increasing OF by a factor of about 1.6 buys the same noise suppression as one extra bit, since each octave of OF gives roughly 9 dB and each bit gives 6 dB.","The leakage condition in Appendix B is directly testable: computing $V_S^\\dagger V_{S^c}\\Delta\\epsilon_{S^c}$ for partial Fourier matrices with adjacent folding indices would show whether the paper's sufficient conditions are loose or whether the guarantee needs an additional assumption on the folding pattern.","A coherence-based alternative to the RIP condition in Theorem 3, computable for this deterministic Fourier sensing matrix, would make the no-flag guarantee practically checkable, since the paper itself notes that verifying RIP is NP-hard.","The same machinery of sparse first-order difference, out-of-band Fourier sensing, and exact rounding is likely to transfer to other signal classes studied under unlimited sampling, such as finite-rate-of-innovation or sparse multi-band signals, though the paper does not prove that."],"forward_implications":["If the theorem is right, oversampling is three times more powerful in a modulo ADC than in a conventional one: doubling OF reduces quantization noise by about 9 dB instead of 3 dB.","A modulo ADC without the folding flag still reaches the same $O(\\mathrm{OF}^{-3})$ scaling, so the flag is not needed for the asymptotic gain—it only tightens the sufficient bit condition.","In a two-band signal with a strong and a weak component, the per-component NMSE bound grows with the component bandwidth and shrinks with its energy, so a weak narrowband signal is not buried by the strong signal's quantization noise.","The paper's simulations show the predicted curves hold with and without the dither signal, and that the advantage disappears when $b=2$ or $b=3$, consistent with the sufficient conditions.","If the scaling law holds, every finite-energy bandlimited signal becomes digitizable by a modulo ADC with better quantization-noise scaling than a conventional b-bit ADC at the same amplitude-quantization bit budget."],"supporting_citations":[{"why":"It supplies the sparsity bound on the first-order difference of the modulo residue that Lemma 1 uses to reduce OF>3 to full column rank of the sensing matrix.","marker":"[20]"},{"why":"It provides the modulo-ADC design and the 1-bit folding-information recovery scheme whose quantized version is analyzed in Theorem 1.","marker":"[21]"},{"why":"It gives the dithered-quantization result that the quantization noise is white with variance $\\lambda^2/2^{2b}$, the basis of every MSE formula in the paper.","marker":"[40]"},{"why":"It states the Schuchman dither conditions used to justify the first- and second-order statistics of the quantization noise.","marker":"[39]"},{"why":"It supplies the OMP exact-support-recovery theorem restated as Theorem 2, on which the no-flag guarantee in Theorem 3 rests.","marker":"[49]"},{"why":"It provides the bounded-noise OMP stopping rule used by the proposed no-flag support-recovery algorithm.","marker":"[46]"},{"why":"It supplies the bandlimited-signal amplitude bound used to convert Theorem 1 into the weak/strong NMSE bound in Corollary 1.","marker":"[55]"},{"why":"It defines the unlimited-sampling baseline and the higher-order-differences recovery whose noise-dependent sampling-rate penalty the paper compares against.","marker":"[11]"}],"fun_headline_variants":["Modulo ADC: cubic noise drop with oversampling","Oversampling cubed: how modulo ADCs beat conventional","Noise floor drops as OF^-3 with modulo sampling","Folding info yields OF^-3 MSE vs OF^-1 for conventional","Modulo ADC: noise scales as OF^-3, not OF^-1"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes that, when the folding positions are known, the noise at the other sample positions does not contaminate the least-squares estimate of the difference signal; if that contamination is large, rounding to the nearest multiple of the folding threshold is not guaranteed even when b>3.","fun_headline_variants_meta":{"raw":{"variants":["Modulo ADC: cubic noise drop with oversampling","Oversampling cubed: how modulo ADCs beat conventional","Noise floor drops as OF^-3 with modulo sampling","Folding info yields OF^-3 MSE vs OF^-1 for conventional","Modulo ADC: noise scales as OF^-3, not OF^-1"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000802,"raw_usage":{"total_tokens":3650,"prompt_tokens":1194,"completion_tokens":2456,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":810,"completion_tokens_details":{"reasoning_tokens":2369}},"tokens_in":810,"tokens_out":2456,"duration_ms":19372,"temperature":1.0,"reasoning_tokens":2369,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:29:08.901845+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the modulo ADC recovery with b=4 and OF=4 on a bandlimited signal with many closely spaced folds, and compare the simulated MSE to the Theorem 1 curve; then compute the leakage norm $\\|V_S^\\dagger V_{S^c}\\Delta\\epsilon_{S^c}\\|_\\infty/\\lambda'$ for the same signal. If the leakage ratio is not negligible, or if the simulated MSE stays above the predicted $O(\\mathrm{OF}^{-3})$ curve by more than the filtered quantization-noise power, the exact-rounding step fails and the sufficient condition in Theorem 1 does not hold.","supporting_citations":[{"cited_title":"Lasso-based fast residual recovery for modulo sampling,","cited_arxiv_id":null,"evidence_quote":"It supplies the sparsity bound on the first-order difference of the modulo residue that Lemma 1 uses to reduce OF>3 to full column rank of the sensing matrix."},{"cited_title":"Compressed Sensing Based Residual Recovery Algorithms and Hardware for Modulo Sampling","cited_arxiv_id":"2412.12724","evidence_quote":"It provides the modulo-ADC design and the 1-bit folding-information recovery scheme whose quantized version is analyzed in Theorem 1."},{"cited_title":"Dithered quantizers,","cited_arxiv_id":null,"evidence_quote":"It gives the dithered-quantization result that the quantization noise is white with variance $\\lambda^2/2^{2b}$, the basis of every MSE formula in the paper."},{"cited_title":"Dither signals and their effect on quantization noise,","cited_arxiv_id":null,"evidence_quote":"It states the Schuchman dither conditions used to justify the first- and second-order statistics of the quantization noise."},{"cited_title":"The exact support recovery of sparse signals with noise via orthogonal matching pursuit,","cited_arxiv_id":null,"evidence_quote":"It supplies the OMP exact-support-recovery theorem restated as Theorem 2, on which the no-flag guarantee in Theorem 3 rests."},{"cited_title":"Limits on bandlimited signals,","cited_arxiv_id":null,"evidence_quote":"It supplies the bandlimited-signal amplitude bound used to convert Theorem 1 into the weak/strong NMSE bound in Corollary 1."},{"cited_title":"On unlimited sampling and reconstruction,","cited_arxiv_id":null,"evidence_quote":"It defines the unlimited-sampling baseline and the higher-order-differences recovery whose noise-dependent sampling-rate penalty the paper compares against."}],"review_version":1}