{"id":"93cb7793-eb68-4ab7-9d31-bed8f5ebfac8","arxiv_id":"2412.12724","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The first-order difference of the modulo residual is sparse, enabling LASSO-based unfolding, and a 1-bit fold indicator reduces recovery to a closed-form least-squares solution.","lead":"This paper introduces LASSO-B2R2, a faster algorithm to unfold modulo-sampled signals by exploiting the sparsity of the first-order difference of the folding residual. It also shows that adding a single side-information bit to indicate fold events reduces recovery to a simple pseudo-inverse, demonstrated on a hardware prototype.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's sparsity bound depends on uncontrolled truncation/periodization and the same leaky 99%-energy choice of N; the proof gap is load-bearing but likely repairable.","rationale":"The reader's weakest assumption identifies exactly the same soft spot: the proof of Theorem 1 rests on an unquantified spectral-leakage approximation, and the 99% energy rule for choosing N inherits the same truncation problem. My read agrees with that assessment. I do not escalate to REJECT because the theorem's conclusion may be true despite the flawed proof: classical zero-counting bounds for entire functions of exponential type could plausibly deliver a bound of the same form, and the numerical evidence is consistent with the claimed behavior. However, the central theoretical contribution is not established as written, and the algorithm's operating regime depends on a heuristic that can fail for adversarial bandlimited signals. Since the reader already assigned CONDITIONAL with medium risk, my stress-test does not change the verdict; it sharpens what would be needed to upgrade to acceptance: a direct proof of the bound and a principled way to choose N, or a clear demonstration that the truncation error is negligible for the intended signal class.","tokens_in":16783,"tokens_out":15315,"duration_ms":163408,"concrete_test":"Settle the concern by (a) independently re-deriving Theorem 1 without periodization: bound the number of real zeros of f(t) - (2m+1)lambda on [0,T] directly from the exponential-type zero-counting theorem (e.g., Boas), and compare the resulting bound with (9). If the direct bound matches (9) with no extra terms, the leakage approximation is removable and the theorem stands; if it contains an additive term depending on ||f||_{L_inf(R \\ [0,T])} or on f(0), f(T), Theorem 1 is not proved. (b) Separately, run LASSO-B2R2 on a bandlimited signal constructed to have a sub-threshold-energy tail bump of amplitude > lambda just outside the 99%-energy window; an NMSE collapse would show the time-limitedness assumption is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-A proves the sparsity bound L <= min(4K + 4K floor((||f||_inf - lambda)/(2 lambda)), N) by (i) truncating the finite-energy bandlimited signal f to [0,T], (ii) asserting the truncated signal lies in B(omega_m + Delta_omega) and then 'approximating' the bandwidth as [-omega_m, omega_m], and (iii) periodizing to a trigonometric polynomial of order K. Steps (i)-(ii) are uncontrolled: a nonzero finite-energy bandlimited function cannot be time-limited, so the Fourier series in (10) has infinitely many terms; dropping terms with |k|>K changes the function and can change the number of level crossings, and no error estimate is supplied. The same truncation is used in Section II-B to claim that the energy condition (4) implies |f(t)| < lambda outside [0,T]. For bandlimited L2 signals this does not follow from tail energy alone; a narrow excursion of amplitude > lambda and energy below (lambda - epsilon)^2 can sit outside [0,T] while staying within the bandwidth. If such a tail bump exists, z(n) is not N-limited, and the measurement model V zhat = Fhat in (7) is corrupted by folds outside the window. Numerical experiments do not close this gap because the random signal model makes the leakage/truncation error small by construction. The bound is plausible (a zero-counting argument for exponential-type functions may yield a similar inequality), but as written the proof is an approximation argument, not a theorem about the actual residual.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LASSO-B2R2, a compressed-sensing-based recovery algorithm for modulo (unlimited) sampling. The central idea is that the first-order difference of the residual signal, z-hat(n), is sparse; the authors derive an upper bound on its sparsity (Theorem 1), cast residual recovery as a LASSO problem solved by ISTA, and then introduce a 1-bit folding-indicator mechanism that reduces recovery to a pseudo-inverse computation. The paper also reports simulation comparisons against HOD, prediction-based, and B2R2 algorithms, and presents a hardware prototype that captures the 1-bit folding information with minimal additional circuitry. The claimed advantages are higher speed and robustness, and an improved performance at low modulo thresholds when the 1-bit side information is used.","tokens_in":17128,"tokens_out":4055,"duration_ms":37077,"significance":"If the sparsity bound and the recovery guarantees are made rigorous, the LASSO-B2R2 formulation is a valuable computational alternative to the sample-by-sample B2R2 method, and the 1-bit bits-distribution mechanism is a practical and economical hardware enhancement. The paper's strengths include a concrete algorithmic contribution, extensive numerical comparisons, and a working hardware prototype that demonstrates the 1-bit folding channel with only an OR gate added to an existing modulo-ADC. The main weaknesses are that the proof of the sparsity bound relies on an unquantified truncation/periodization approximation, and the hardware validation shows a substantially larger MSE than the corresponding simulations; both issues need to be addressed before the central claims are fully supported.","major_comments":[{"comment":"The proof of Theorem 1 truncates the finite-energy bandlimited signal f(t) to duration T, asserts that the truncated signal is approximately bandlimited to [-omega_m, omega_m], and then periodizes it to a trigonometric polynomial of order K. A nonzero finite-energy bandlimited signal cannot be time-limited, so the truncated signal has unbounded spectral support and the Fourier series in (10) is not finite. Dropping the terms with |k|>K changes the function and can change the number of level crossings, but no error estimate is supplied for this approximation. This is load-bearing because the sparsity of z-hat, and hence the justification for the LASSO formulation in (13), rests on this bound. Please provide a rigorous error estimate or replace the approximation argument with a direct bound on the number of level crossings of the original bandlimited signal, for example using properties of exponential-type functions.","section":"Section III-A, Theorem 1"},{"comment":"The claim that choosing T via the energy condition (4) guarantees |f(t)| < lambda for all t outside [0,T] is not justified for general bandlimited L2 signals. Tail energy below (lambda - epsilon)^2 does not preclude a narrow excursion of amplitude larger than lambda outside [0,T]; such an excursion could have arbitrarily small energy while still being present in the bandlimited signal. If such a tail bump exists, z(n) is not N-limited, and the measurement model V z-hat = F-hat in (7) is corrupted by folds outside the window. The numerical experiments use random signals for which this effect is negligible, but the theoretical statement needs additional assumptions, such as a bound on the signal amplitude or a decay rate, or an explicit quantification of the resulting model error.","section":"Section II-B, time-limitedness property"},{"comment":"The simulated LASSO-B2R2 with 1-bit reports MSE of -9.0683 dB (Table II), while the hardware result in Table IV reports MSE of 4.8 dB, a degradation of roughly 14 dB. Similarly, B2R2 and LASSO-B2R2 show hardware MSE of 7.8149 dB, far from the simulated values. The paper does not discuss this discrepancy. Since the hardware section is presented as demonstrating practical feasibility, please analyze the sources of degradation (e.g., comparator thresholds, timing, board non-idealities, noise) and report the actual folding-detection error rate, or temper the claim to indicate that the prototype demonstrates the 1-bit channel concept but does not yet achieve the simulated recovery performance.","section":"Section IV-B and IV-C, hardware results"},{"comment":"The pseudo-inverse recovery in (18) is evaluated in simulations using the true support T of the sparse vector z-hat. In the hardware, the support is estimated from the 1-bit folding stream, which is subject to detection errors. The comparison in Table II and Fig. 8 is therefore an oracle upper bound, not a performance evaluation of the full pipeline with estimated support. Please either simulate recovery using the support obtained from the actual 1-bit stream (with realistic errors) or explicitly state that the simulated results assume perfect folding detection, and assess how support-estimation errors affect the MSE.","section":"Section IV-A and IV-B, oracle support in simulations"}],"minor_comments":[{"comment":"The sentence 'Modern data acquisition systems relay on Analog-to-Digital Converters' contains a typo: 'relay' should be 'rely'.","section":"Section I, first paragraph"},{"comment":"The typo 'LASSSO' appears twice in (17); it should be 'LASSO'.","section":"Section III-B, equation (17)"},{"comment":"The text says 'Table II tabulates the MSE and the computational time required by each algorithm' when describing hardware results; the correct reference is Table IV.","section":"Section IV-C, hardware prototype"},{"comment":"The rounding operation 'z <- ceil(floor(z/lambda)/2)' is not explained; please clarify how this maps the soft-thresholded estimate to values in 2*lambda*Z.","section":"Algorithm 1, line 15"},{"comment":"The statement that 'the spark of V is at least M+1' is correct for a partial DFT matrix, but the phrase 'at least' combined with the subsequent condition L < (M+1)/2 is slightly imprecise; the relevant property is that the spark equals M+1 when any M columns are linearly independent.","section":"Section III-D"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of the journal and the core idea is promising, but the theoretical sparsity bound is not yet rigorous and the hardware validation is weaker than the simulations suggest. Both issues are fixable within the manuscript's scope, so I recommend a major revision rather than rejection. The authors should also ensure that the simulation methodology for the 1-bit case is clearly labeled as an oracle bound, and that the hardware discrepancy is quantified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful part of this paper is the observation that the first-order difference of the residual in modulo sampling is sparse, and that a 1-bit folding indicator can reduce recovery to a pseudo-inverse. The LASSO formulation is a natural application of compressed sensing, but the sparsity insight is new for this problem and the bits-distribution mechanism is a genuine hardware-friendly addition. The simulations showing speed gains over B2R2 are credible, and the hardware prototype with an OR gate is a nice practical touch.\n\nThe soft spot is Theorem 1. The proof truncates a bandlimited signal to [0,T], then periodizes and truncates the Fourier series to order K, explicitly approximating away the spectral leakage. That is not a theorem about the actual residual. A finite-energy bandlimited signal is not time-limited, so the truncated signal has unbounded bandwidth; dropping terms with |k|>K changes the signal and can change the level-crossing count. No error estimate is given. The 99% energy rule for choosing N has a related problem: small tail energy does not preclude a narrow excursion above lambda outside [0,T], so the claim that z(n) is N-limited is not guaranteed. The numerical experiments use random signals whose tails decay fast, so they don't stress this gap. The sparsity bound may well be true with a different proof, but as written it's an approximation argument.\n\nThe hardware demo also deserves a closer look. The 1-bit experiment reports MSE 4.8 dB, versus -9.07 dB in simulation, and the comparator-based fold detection is not validated against ground truth. That gap is never explained. Still, this is a preliminary prototype, so I wouldn't call it fatal.\n\nOverall, the algorithmic contribution is plausible and useful, and the proof gap is likely repairable. The paper deserves a serious referee, but it needs a rigorous statement of the sparsity bound (or an honest heuristic label) and a more careful hardware evaluation before acceptance. I'd send it to review, and I'd cite the 1-bit mechanism if I worked in the area.","headline":"Useful sparsity insight and a neat 1-bit side-information mechanism, but the main theorem's proof rests on an unquantified truncation approximation and the hardware demo is much weaker than the simulations.","tokens_in":17627,"tokens_out":3210,"would_cite":true,"duration_ms":29191,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A12","94A20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that modulo-ADC recovery can be made fast and robust by treating the first-order difference of the residual as a sparse signal, and that spending one ADC bit on folding indicators reduces recovery to a pseudo-inverse.","keywords":["modulo sampling","unlimited sampling","compressed sensing","LASSO","residual recovery","analog-to-digital converter","high dynamic range","1-bit side information"],"falsifier":"Construct a bandlimited signal whose energy is concentrated near the edges of the chosen $T$-duration window so that truncation creates substantial spectral leakage, count the level crossings of the true residual difference, and check whether the count exceeds $4K + 4K \\lfloor (\\|f\\|_\\infty - \\lambda)/(2\\lambda)\\rfloor$; exceeding it would falsify Theorem 1. Alternatively, simulate LASSO-B2R2 on such a signal and test whether recovery fails despite satisfying the spark condition, which would indicate that the spectral-leakage approximation is load-bearing.","tokens_in":16602,"feed_emoji":"⚡","tokens_out":4960,"duration_ms":38355,"temperature":0.7,"pith_summary":"This paper claims that modulo sampling, which folds a high-dynamic-range signal into an ADC's range, can be unfolded by exploiting sparsity in the first-order difference of the residual (the difference between folded and original samples). The authors prove an upper bound on how many nonzero entries this difference can have, and use that to cast recovery as a LASSO problem solved by ISTA from out-of-band Fourier measurements. They then show that if the hardware spends one bit marking when folding occurs, recovery reduces to a simple pseudo-inverse computation, which is far faster and works at lower modulo thresholds. Numerical simulations and a hardware prototype with an OR gate support the claims. A sympathetic reader would care because this promises ADC designs that handle high dynamic range without clipping while recovering near the Nyquist rate.","feed_headline":"Sparse residual differences make modulo-ADC recovery a LASSO problem","feed_subtitle":"First-order difference of the residual is provably sparse; with 1-bit folding info, recovery collapses to a pseudo-inverse.","key_machinery":"The central object is the first-order difference of the residual, $\\hat{z}(n) = \\Delta z(n)$, where $z(n) = f_\\lambda(n) - f(n)$ is a piecewise-constant signal taking values in $2\\lambda\\mathbb{Z}$. Nonzero entries of $\\hat{z}$ mark each $2\\lambda$ level jump, i.e., each modulo folding event. Its sparsity is bounded by a level-crossing count for trigonometric polynomials of order $K = \\lfloor N/(2\\,\\mathrm{OF})\\rfloor$, and that sparsity makes the partial-DFT equation $\\hat{\\mathbf{F}}_\\lambda = \\mathbf{V}\\hat{\\mathbf{z}}$ solvable by $\\ell^1$ minimization. When a 1-bit folding indicator reveals the support, the same equation collapses to the pseudo-inverse (18).","core_discovery":"The first-order difference of the residual, $\\hat{z}(n)$, is sparse: Theorem 1 bounds its support by $L \\le \\min(4K + 4K \\lfloor (\\|f\\|_\\infty - \\lambda)/(2\\lambda)\\rfloor, N)$ with $K = \\lfloor N/(2\\,\\mathrm{OF})\\rfloor$. Because $\\hat{z}$ lives in the out-of-band Fourier coefficients of the folded samples, recovery of $\\hat{z}$ from the partial DFT matrix $\\mathbf{V}$ is a sparse linear inverse problem, solved here by the LASSO in (13) via ISTA. With a 1-bit folding-indicator stream $b(n)$, the support of $\\hat{z}$ is known in advance, and the estimate becomes the oracle least-squares pseudo-inverse $\\hat{\\mathbf{z}}_T = \\mathbf{V}_T^\\dagger \\hat{\\mathbf{F}}_\\lambda^q$ in (18). The paper also demonstrates that the extra bit can be obtained from the existing modulo hardware by adding a single OR gate, and reports numerical and hardware results showing lower MSE and much lower runtime than prior B2R2 recovery.","pith_inferences":["The same sparse-residual viewpoint could extend to other signal models, such as finite-rate-of-innovation or graph signals, where the residual's difference may be sparse under a different bound.","The 1-bit side-information idea could be refined: instead of one bit per sample, flag only genuine folding events, potentially saving bits when folding is rare in a long record.","The spectral-leakage approximation in Theorem 1 is unquantified; a tighter analysis might change the sparsity bound and the operating region where LASSO-B2R2 is guaranteed to succeed.","Combining the spark condition with the bound on $L$ yields a concrete recipe for choosing $\\mathrm{OF}$ and $\\lambda$; stress-testing it on adversarial signals with rapid repeated foldings would show whether the rule is tight or pessimistic."],"forward_implications":["LASSO-B2R2 recovers modulo samples robustly at lower oversampling than higher-order-difference and prediction-based methods, with runtime much lower than B2R2.","The sparsity bound connects oversampling factor, modulo threshold $\\lambda$, and the number of out-of-band measurements $M$ through the spark condition $L < (M+1)/2$, giving a design rule for choosing sampling parameters.","Bits distribution: using 1 of $B$ bits to mark folding events reduces recovery to a closed-form least-squares estimate, removing iterative solving entirely.","Hardware: the existing modulo-ADC prototype needs only an OR gate plus comparators to supply the 1-bit folding stream, making the faster algorithm practical.","At low $\\lambda$, LASSO-B2R2 alone requires a higher sampling rate, but the 1-bit variant overcomes this limitation while keeping performance comparable to B2R2."],"supporting_citations":[{"why":"Introduces modulo sampling and the higher-order-difference recovery baseline that LASSO-B2R2 must beat.","marker":"[19]"},{"why":"Establishes prediction-based perfect recovery above the Nyquist rate and supplies the time-limitedness property used in the formulation.","marker":"[20]"},{"why":"Proposes B2R2 residual recovery, the constrained per-sample optimization that this work replaces with a one-shot sparse problem.","marker":"[21]"},{"why":"Extends B2R2 and provides the BL-ness and time-limitedness properties that justify the partial-Fourier measurement model.","marker":"[22]"},{"why":"Describes the hardware prototype that the authors augment with an OR gate to obtain the 1-bit folding information.","marker":"[36]"},{"why":"Gives the trigonometric-polynomial level-crossing bound used in the proof of Theorem 1.","marker":"[42]"},{"why":"Supplies the ISTA proximal algorithm and its convergence initialization used to solve the LASSO problem.","marker":"[43]"},{"why":"Provides the restricted-isometry and stable-recovery guarantee cited for partial DFT matrices.","marker":"[45]"}],"fun_headline_variants":["Sparse residual differences turn modulo-ADC recovery into a LASSO problem","1-bit folding info turns modulo recovery into a pseudo-inverse","LASSO solves sparse residual recovery in modulo sampling","Hardware demo: 1-bit modulo recovery via pseudo-inverse","Sparsity in residual unlocks fast modulo-ADC recovery"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The sparsity bound in Theorem 1 treats the truncated, periodized signal as if it were exactly bandlimited to $[-\\omega_m,\\omega_m]$, ignoring the spectral leakage introduced by truncation, and no error estimate is given; the practical choice of $N$ also relies on a 99%-energy heuristic.","fun_headline_variants_meta":{"raw":{"variants":["Sparse residual differences turn modulo-ADC recovery into a LASSO problem","1-bit folding info turns modulo recovery into a pseudo-inverse","LASSO solves sparse residual recovery in modulo sampling","Hardware demo: 1-bit modulo recovery via pseudo-inverse","Sparsity in residual unlocks fast modulo-ADC recovery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000905,"raw_usage":{"total_tokens":3968,"prompt_tokens":1093,"completion_tokens":2875,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":709,"completion_tokens_details":{"reasoning_tokens":2789}},"tokens_in":709,"tokens_out":2875,"duration_ms":19186,"temperature":1.0,"reasoning_tokens":2789,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:47:57.869441+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a bandlimited signal whose energy is concentrated near the edges of the chosen $T$-duration window so that truncation creates substantial spectral leakage, count the level crossings of the true residual difference, and check whether the count exceeds $4K + 4K \\lfloor (\\|f\\|_\\infty - \\lambda)/(2\\lambda)\\rfloor$; exceeding it would falsify Theorem 1. Alternatively, simulate LASSO-B2R2 on such a signal and test whether recovery fails despite satisfying the spark condition, which would indicate that the spectral-leakage approximation is load-bearing.","supporting_citations":[{"cited_title":"On Unlimited Sampling and Reconstruction,","cited_arxiv_id":null,"evidence_quote":"Introduces modulo sampling and the higher-order-difference recovery baseline that LASSO-B2R2 must beat."},{"cited_title":"Residual Recovery Algorithm for Modulo Sampling,","cited_arxiv_id":null,"evidence_quote":"Proposes B2R2 residual recovery, the constrained per-sample optimization that this work replaces with a one-shot sparse problem."},{"cited_title":"Unlimited sampling beyond mod- ulo,","cited_arxiv_id":null,"evidence_quote":"Extends B2R2 and provides the BL-ness and time-limitedness properties that justify the partial-Fourier measurement model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the trigonometric-polynomial level-crossing bound used in the proof of Theorem 1."},{"cited_title":"Proximal algorithms,","cited_arxiv_id":null,"evidence_quote":"Supplies the ISTA proximal algorithm and its convergence initialization used to solve the LASSO problem."},{"cited_title":"Stable signal recovery from incomplete and inaccurate measurements,","cited_arxiv_id":null,"evidence_quote":"Provides the restricted-isometry and stable-recovery guarantee cited for partial DFT matrices."}],"review_version":1}