{"id":"889d27b4-9e77-4642-87c3-e7f92ba4aa4f","arxiv_id":"2501.11453","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Integrate-and-fire with reset-to-mod is shown to be the integral version of send-on-delta sampling, with error bounds, sparsity, and sparse regularization expressed in the Alexiewicz norm.","lead":"This paper argues that integrate-and-fire (IF) and send-on-delta (SOD) are two views of the same event-based sampling process, related as integration and differentiation, and that the Alexiewicz norm provides the right mathematical setting for error analysis. It also claims maximum sparsity and sparse regularization properties, though several key proofs are deferred to the authors' own earlier and forthcoming papers.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4's uniqueness claim is false: for f(t)=t on [0,1] every u in (1/2,1] solves u = ||SOD_u(f)||_1, and the minimizer of (21) depends on lambda, so the sparse-regularization result collapses.","rationale":"The paper's central identity SOD_theta(integral f) = IF^M_theta(f) and the quantization view of IF/mod are plausible and are partly supported by the definitions in Section 2. However, the strongest advertised consequences include the maximal-sparsity and sparse-regularization interpretations, and the latter rests on Theorem 4. That theorem is asserted without a proof in this manuscript, and its uniqueness assumption is demonstrably false for the simple linear signal f(t)=t. The fixed-point equation has a continuum of solutions, and the proposed regularizer's minimizer depends on lambda, directly contradicting the theorem's statement. This is not a matter of outside-consensus disagreement or a missing reference; it is an internal mathematical failure in a result that the paper explicitly presents as a theorem. The reader's verdict of REJECT is therefore appropriate: even if Theorems 1 and 2 are later verified in the cited under-review work, the current manuscript contains a false headline theorem and cannot be accepted as is.","tokens_in":7296,"tokens_out":9193,"duration_ms":104550,"concrete_test":"Take f(t)=t on [0,1] and compute SOD_u, ||SOD_u(f)||_1, chi_u, alpha(u), and beta(u) explicitly for u in (0,1]. Tabulate the fixed-point residual r(u)=|u - ||SOD_u(f)||_1| and the objective F_lambda(u)=u + lambda * floor(1/u) for lambda=1 and lambda=0.01. If r(u) vanishes on a continuum of u in (1/2,1] and the argmin of F_lambda shifts with lambda, Theorem 4's uniqueness and lambda-independence claims are disproved.","verdict_should_be":"REJECT","load_bearing_attack":"Section 5's Theorem 4 is load-bearing: it is the paper's advertised link between SOD/IF sampling and sparse regularization. Its proof is explicitly deferred ('an extended proof is postponed to an upcoming paper'), and the asserted mechanism is incorrect. Let f(t)=t on [0,1] with f(0)=0. For threshold u in (0,1], SOD_u emits amplitude-u spikes at times u,2u,...,(floor(1/u)-1)u, so ||SOD_u(f)||_1 = u * floor(1/u). Hence the fixed-point equation u* = ||SOD_{u*}(f)||_1 reduces to floor(1/u*)=1, which is satisfied by every u* in (1/2,1]. Uniqueness fails. The induced step function chi_u satisfies ||f - chi_u||_infinity = u and (1/u)||chi_u||_TV = floor(1/u), so the objective in (21) equals u + lambda * floor(1/u). Its minimizer depends on lambda: for lambda=1, u=0.51 gives about 1.51 while u=0.1 gives about 10.1; for lambda=0.01, u=0.51 gives about 0.52 while u=0.1 gives about 0.2. Thus no single theta = u* can solve (21) for arbitrary lambda, contrary to the theorem. The 'Vice versa' remark does not repair the stated claim. Because the paper advertises the sparse-regularization characterization as a consequence, this is a correctness-level objection, not merely a missing proof.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified mathematical treatment of integrate-and-fire (IF) with reset-to-mod and send-on-delta (SOD) threshold-based sampling. Its central claims are the identity SOD_ϑ(∫f)=IF^M_ϑ(f) (Eq. 7), the quantization-operator characterization IF^M_ϑ=A^{-1}∘q_ϑ∘A (Theorem 1), a quasi-isometry relation in the Alexiewicz norm (Eq. 12), a maximal sparsity property (Theorem 2), and a sparse-regularization interpretation (Theorem 4). The paper also gives reconstruction formulas for SOD and IF/mod. The main structural theorems are not proven in the manuscript: Theorem 1 and Theorem 2 cite an under-review preprint [9], and Theorem 4 postpones its proof to an upcoming paper.","tokens_in":7764,"tokens_out":4360,"duration_ms":42901,"significance":"If the main claims were correct, the paper would provide a useful conceptual bridge between neuromorphic integrate-and-fire encoding and event-based send-on-delta sampling, offering unified error bounds, sparsity guarantees, and a regularization-based interpretation. The identity in Eq. (7) and the quasi-isometry derivation are interesting and plausible. However, the sparse-regularization characterization in Theorem 4 is demonstrably false, and the other central theorems are deferred to references rather than proved here. Because the advertised regularization result collapses, the paper does not deliver its stated contribution.","major_comments":[{"comment":"The fixed-point condition u* = ||SOD_{u*}(f)||_1 is not uniquely solvable. For f(t)=t on [0,1], the SOD spike train has ||SOD_u(f)||_1 = u floor(1/u), so the fixed-point equation reduces to floor(1/u)=1, which holds for every u in (1/2,1]. Thus uniqueness fails. Moreover, the objective in (21) for this f equals u + λ floor(1/u) (since α(u)=u and (1/u)||χ_u||_TV = floor(1/u)), and its minimizer depends on λ: for λ=1, u=0.51 yields about 1.51 while u=0.1 yields about 10.1, whereas for λ=0.01, u=0.51 yields about 0.52 while u=0.1 yields about 0.2. Hence no single value ϑ=u* can solve (21) for arbitrary λ, contrary to the theorem. The 'Vice versa' remark does not repair the stated claim.","section":"Section 5, Theorem 4"},{"comment":"The proofs of Theorem 1 (quantization operator) and Theorem 2 (maximal sparsity) are not contained in this manuscript; they are cited to reference [9], which is under review. Theorem 4's proof is likewise postponed to an upcoming paper. Since these theorems carry the paper's central structural claims, the manuscript is not self-contained, and the results cannot be verified from the presented material. This is a load-bearing completeness issue independent of the counterexample to Theorem 4.","section":"Sections 3.1 and 3.3, Theorems 1 and 2"},{"comment":"The proof sketch asserts α(u)=||f−χ_u||_∞ = u for continuous f and convexity of β(u)=(1/u)||χ_u||_TV, but neither is established, and the claimed consequence does not follow. The counterexample in the previous comment shows that even when α(u)=u holds, the conclusion of Theorem 4 fails. The sketch is therefore insufficient to support the sparse-regularization characterization.","section":"Section 5, proof sketch"}],"minor_comments":[{"comment":"The sentence 'While the mathematical principles are outlined in this article, their exploitation, e.g. for a novel design approach for adapting analog-to-spike converters for signals with jumps' is incomplete; it lacks a main clause.","section":"Section 1, Introduction"},{"comment":"The caption reads 'q(x) := q_ϑ(x) for ϑ = 1, i.e., q_ϑ(x) = ϑ q(x/ϑ)', which is confusing because q is defined through q_ϑ; it would be clearer to define q as the integer truncation and q_ϑ as the scaled version.","section":"Section 2, Figure 2 caption"},{"comment":"The reconstruction formula G_k(t) uses s_{k+1} within the interval [t_k, t_{k+1}); please clarify the treatment of the final interval and the case |s_{k+1}| > ϑ, which appears not to be covered by the displayed formula.","section":"Section 4, Eq. (17)"},{"comment":"The numerical illustration of Eq. (21) uses discrete acceleration data, for which the paper states α(u)=u is only approximately valid; the figure should be described as a heuristic illustration rather than a verification of Theorem 4, since the theorem assumes continuous f.","section":"Section 6, Figure 3 discussion"}],"recommendation":"reject","confidential_remarks":"The paper's main advertised result, the sparse-regularization characterization, is false as stated (f(t)=t gives a continuum of fixed points and a λ-dependent minimizer). Combined with the heavy reliance on under-review and forthcoming references for Theorems 1, 2, and 4, the manuscript is not acceptable in its current form. The identity in Eq. (7) and the conceptual framing are potentially salvageable in a future, fully self-contained revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one genuinely new thing here is Eq. (7): SOD applied to the integral of f equals IF/mod applied to f. That identity is proven in the text via the recursion, and it is a useful observation. The generalized SOD for discontinuous signals is also new and makes the framework apply to signals with jumps. The Alexiewicz-norm setting gives clean error bounds and the reconstruction formulas are explicit. I read the paper as a fairly honest synthesis of the authors' prior work, and the explanatory parts are clear.\n\nThe soft spots are real. The paper advertises maximal sparsity and sparse regularization as consequences, but those are not established in this manuscript. Theorem 1 and Theorem 2 are cited to an under-review paper by the same authors, and Theorem 4's proof is postponed to an upcoming paper. That alone would be a reason to be cautious. Worse, Theorem 4 as stated is false. For f(t)=t on [0,1], every threshold u in (1/2,1] satisfies u = ||SOD_u(f)||_1, so the fixed point is not unique. And the minimizer of the regularization objective (21) depends on lambda: with lambda=1, u=0.51 is better than u=0.1, while with lambda=0.01, u=0.1 is better. No single theta can solve (21) for arbitrary lambda, contrary to the theorem. This is a correctness-level objection, not a stylistic one.\n\nThe central identity and the quasi-isometry bounds are still worth knowing. The paper would be useful to researchers in neuromorphic computing and event-based sampling who want the SOD/IF connection and the Alexiewicz-norm interpretation. But the regularization claim should be withdrawn or corrected before this is citable as a research result.\n\nI would not reject the paper out of hand, but I would send it to a referee who can check the math. The identity is solid enough to deserve referee time, even if the current version has a load-bearing flaw.","headline":"A genuine new identity between SOD and IF/mod, but the sparse-regularization theorem is false as stated and the main proofs are deferred to the authors' own under-review papers.","tokens_in":8178,"tokens_out":2212,"would_cite":false,"duration_ms":25468,"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":"Integrate-and-fire with reset-to-mod is exactly send-on-delta sampling applied to the integral of the signal, making threshold-based spiking a quantization operator in the Alexiewicz norm.","keywords":["integrate-and-fire","send-on-delta","threshold-based sampling","Alexiewicz norm","sparse regularization","neuromorphic computing","spike train quantization","event-based signal processing"],"falsifier":"For the ramp $f(t)=t$ on $[0,1]$, compute $\\mathrm{SOD}_u(f)$ for thresholds $1/2 < u \\le 1$; the output is a single spike of amplitude $u$, so $\\|\\mathrm{SOD}_u(f)\\|_1 = u$ and every such $u$ satisfies the fixed-point equation $u = \\|\\mathrm{SOD}_u(f)\\|_1$. This contradicts the uniqueness of $u^*$ required by Theorem 4, so the regularization claim as stated cannot hold for this simple continuous signal.","tokens_in":7106,"feed_emoji":"⚡","tokens_out":16278,"duration_ms":141078,"temperature":0.7,"pith_summary":"This paper establishes that the integrate-and-fire (IF) neuron model, when its membrane potential is reset by a modulo operation, is the integral version of send-on-delta (SOD) threshold sampling: the identity $\\mathrm{SOD}_\\vartheta(\\int f) = \\mathrm{IF}^M_\\vartheta(f)$ holds for bounded integrable signals with finitely many Dirac impulses. The same identity is reframed as $\\mathrm{IF}^M_\\vartheta = A^{-1} \\circ q_\\vartheta \\circ A$, which turns spiking into a quantization of the signal's running integral into packets of size $\\vartheta$. This makes the Alexiewicz norm $\\|f\\|_A = \\sup_T |\\int_{t_a}^T f|$ the natural metric for spiking error, with the bound $\\|\\mathrm{IF}^M_\\vartheta(f) - f\\|_A < \\vartheta$. The paper then derives quasi-isometry bounds, a maximal-sparsity property, and a sparse-regularization reading of threshold-based sampling, extending these guarantees to signals with jumps.","feed_headline":"Integrate-and-fire is the integral twin of send-on-delta","feed_subtitle":"The identity yields a neuron encoder with uniform error bounds and optimally sparse spike trains.","key_machinery":"The key object is the reset-to-mod integrate-and-fire operator $\\mathrm{IF}^M_\\vartheta$, in which each spike discharges the membrane potential by the largest multiple of the threshold $\\vartheta$ that leaves the potential subthreshold, effectively a modulo operation. The argument runs through the operator identity $\\mathrm{IF}^M_\\vartheta = A^{-1} \\circ q_\\vartheta \\circ A$, with $A(f)(t) = \\int_0^t f(\\tau)\\,d\\tau$ the accumulating integral and $q_\\vartheta$ the threshold quantization by truncation to multiples of $\\vartheta$; this integrate-quantize-differentiate structure is what ties threshold-based spiking to the Alexiewicz norm $\\|f\\|_A = \\sup_T |\\int_{t_a}^T f|$. The sparsity and regularization results then follow from this quantization view, because the spike train is chosen as the minimal $\\ell^1$ grid point within an Alexiewicz ball around the input.","core_discovery":"The central discovery is that reset-to-mod integrate-and-fire is not merely analogous to send-on-delta sampling but is the same operation in the integral domain: for any signal $f$ that is integrable and bounded except for finitely many superimposed Dirac impulses, $\\mathrm{SOD}_\\vartheta(\\int f) = \\mathrm{IF}^M_\\vartheta(f)$. The decomposition $\\mathrm{IF}^M_\\vartheta = A^{-1} \\circ q_\\vartheta \\circ A$, where $A$ is the accumulating integral operator and $q_\\vartheta$ is pointwise truncation to multiples of $\\vartheta$, shows that IF/mod behaves as a quantizer in the Alexiewicz norm with error strictly below the threshold. From this the paper derives a global quasi-isometry between the signal space and the spike-train space, a maximal-sparsity property identifying the IF spike train as the minimal-$\\ell^1$ spike train inside an open Alexiewicz ball around the signal, and a characterization of SOD (and, analogously, IF/mod) as the solver of a total-variation sparse-regularization problem.","pith_inferences":["Editorial extension: the identity suggests a direct hardware recipe - an analog integrator feeding a SOD comparator realizes an IF/mod neuron, and an SOD sampler yields an IF neuron if an integrator is placed in front; this could guide analog-to-spike converter designs for jump-heavy signals, which the paper only sketches as future work.","Editorial extension: if the Alexiewicz-norm view is right, then reconstruction quality and error bounds for neuromorphic sensing should be evaluated on cumulative sums of the signal rather than pointwise values, which may explain why spiking codes tolerate small timing jitter.","Editorial extension: the same quantization argument could be carried over to leaky integrate-and-fire and to other reset rules through a weighted or discounted integral operator, yielding a family of Alexiewicz-type norms; the paper does not develop this, but it follows directly from the operator decomposition."],"forward_implications":["IF/mod spike trains reconstruct the input with an Alexiewicz-norm error below $\\vartheta$, and the piecewise reconstruction (19) stays within $2\\vartheta$, giving a uniform error guarantee that does not require the signal to be continuous.","The quasi-isometry (12) preserves signal distances up to an additive $2\\vartheta$, so spike-based representations become asymptotically isometric as the threshold goes to zero.","The maximal-sparsity property makes the IF/mod spike train the unique minimal-$\\ell^1$ spike train among all spike trains inside the open Alexiewicz ball of radius $\\vartheta$, so event-based encoding is optimally sparse in this metric.","SOD and IF/mod solve sparse-regularization problems, linking the threshold $\\vartheta$ to a regularization parameter $\\lambda$ and giving a principled way to choose the threshold.","The generalized SOD applies to discontinuous signals with jumps and Dirac impulses, so threshold-based sampling can handle mixed analog signals in communication and measurement applications."],"supporting_citations":[{"why":"Defines the Alexiewicz norm that serves as the metric for the quantization error bounds and quasi-isometry.","marker":"[1]"},{"why":"Introduces the send-on-delta concept that IF/mod is shown to generalize through identity (7).","marker":"[4]"},{"why":"Establishes the quasi-isometry of threshold-based sampling that the paper extends to IF/mod.","marker":"[6]"},{"why":"Provides the sharp global quasi-isometry bounds used in Section 3.2.","marker":"[7]"},{"why":"Establishes leaky integrate-and-fire as a spike-train quantization operator on Dirac-superimposed signals, the setting used here.","marker":"[8]"},{"why":"Contains the proofs of the quantization-operator decomposition (Theorem 1) and the maximal-sparsity property (Theorem 2).","marker":"[9]"},{"why":"Introduces the reset-to-mod variant and the Alexiewicz-topology perspective on spiking networks on which this paper builds.","marker":"[10]"}],"fun_headline_variants":["Integrate-and-fire equals send-on-delta in integral domain","IF and SOD are the same operation under integration","Reset-to-mod IF is integral-domain SOD","Neuron spiking is integral send-on-delta sampling","Integrate-and-fire is the integral form of SOD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The sparse-regularization theorem depends on a proof deferred to a later paper and on the assumption that the reconstruction error $\\|f - \\chi_u\\|_\\infty$ equals $u$ and that the threshold equation $u^* = \\|\\mathrm{SOD}_{u^*}(f)\\|_1$ has a unique solution; that uniqueness already fails for the ramp $f(t)=t$ on $[0,1]$.","fun_headline_variants_meta":{"raw":{"variants":["Integrate-and-fire equals send-on-delta in integral domain","IF and SOD are the same operation under integration","Reset-to-mod IF is integral-domain SOD","Neuron spiking is integral send-on-delta sampling","Integrate-and-fire is the integral form of SOD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000163,"raw_usage":{"total_tokens":1219,"prompt_tokens":895,"completion_tokens":324,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":248}},"tokens_in":511,"tokens_out":324,"duration_ms":3549,"temperature":1.0,"reasoning_tokens":248,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:15:28.961565+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the ramp $f(t)=t$ on $[0,1]$, compute $\\mathrm{SOD}_u(f)$ for thresholds $1/2 < u \\le 1$; the output is a single spike of amplitude $u$, so $\\|\\mathrm{SOD}_u(f)\\|_1 = u$ and every such $u$ satisfies the fixed-point equation $u = \\|\\mathrm{SOD}_u(f)\\|_1$. This contradicts the uniqueness of $u^*$ required by Theorem 4, so the regularization claim as stated cannot hold for this simple continuous signal.","supporting_citations":[{"cited_title":"Linear functionals on Denjoy-integrable functions","cited_arxiv_id":null,"evidence_quote":"Defines the Alexiewicz norm that serves as the metric for the quantization error bounds and quasi-isometry."},{"cited_title":"Send-on-delta concept: An event-based data reporting strategy","cited_arxiv_id":null,"evidence_quote":"Introduces the send-on-delta concept that IF/mod is shown to generalize through identity (7)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the quasi-isometry of threshold-based sampling that the paper extends to IF/mod."},{"cited_title":"Moser and Michael Lunglmayr","cited_arxiv_id":null,"evidence_quote":"Provides the sharp global quasi-isometry bounds used in Section 3.2."},{"cited_title":"On Leaky-Integrate-and Fire as Spike-Train-Quantization Operator on Dirac-Superimposed Continuous-Time Signals","cited_arxiv_id":"2402.07954","evidence_quote":"Establishes leaky integrate-and-fire as a spike-train quantization operator on Dirac-superimposed signals, the setting used here."},{"cited_title":"On the Sampling Sparsity of Neuromorphic Analog-to-Spike Conversion based on Leaky Integrate-and-Fire","cited_arxiv_id":"2410.17441","evidence_quote":"Contains the proofs of the quantization-operator decomposition (Theorem 1) and the maximal-sparsity property (Theorem 2)."},{"cited_title":"Moser and Michael Lunglmayr","cited_arxiv_id":null,"evidence_quote":"Introduces the reset-to-mod variant and the Alexiewicz-topology perspective on spiking networks on which this paper builds."}],"review_version":1}