{"id":"e3a1da83-22ad-4711-8dd0-6d9eefc4f947","arxiv_id":"1908.02862","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"An explicit algorithm solves convolution Volterra equations with impulsive inputs by replacing the kernel with piecewise constant steps, giving exact solutions and proven error bounds.","lead":"This paper gives an explicit step-by-step algorithm for solving a common type of integral equation where the input can be a sharp impulse. The method replaces the kernel by a step function, solves that version exactly, and proves error bounds that let the user control accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2.7 is internally inconsistent: its inner sum stops at r=n-1, but the shifted triangle L^rγ_n contributes for r≥n when t≥n, so the claimed explicit solution is false for simple two-step kernels.","rationale":"The reader correctly notes the k<1 condition, but the more serious problem is internal. After Theorem 2.6, the expansion of A^nγ_n has infinitely many r-terms; (2.26) drops r≥n. Since shifted γ_n has support [r,r+n), such terms are exactly the ones active when t≥n, so (2.26) is not the solution h but a truncated series. The counterexample with a two-step kernel is unambiguous: every coefficient is nonnegative, so the omitted a1^2/2 term cannot cancel. This is not a disagreement with consensus; it is an algebraic inconsistency in the central theorem. The error bounds in the appendix appear to be correct for the true h, but they cannot rescue an algorithm that computes a different function. A corrected theorem would sum r over all nonnegative integers (for fixed t, only r<t survive), after which the method is plausible; but as written, the paper's main claim fails. Hence the verdict should move from CONDITIONAL to REJECT, with the path to revision being a corrected index range and re-verification.","tokens_in":12672,"tokens_out":12250,"duration_ms":130563,"concrete_test":"Re-derive (2.26) from Theorem 2.6 by expanding A^n and keeping all r. Then, for g=a0 rect(t)+a1 rect(t-1) with a0=a1=0.1, compute both sides at t=5/2: evaluate the direct Volterra Neumann sum Σ_{n=1}^{100} g^{*n}(5/2) by exact piecewise-polynomial convolution, and evaluate the right side of (2.26) truncated to n≤100. If the difference is at least 0.005 plus higher-order terms and does not vanish as the truncation grows, the theorem's sum range is the cause.","verdict_should_be":"REJECT","load_bearing_attack":"The decisive defect is in Theorem 2.7, eq. (2.26). Theorem 2.6 correctly gives g^{*n}=A^nγ_n=Σ_{r=0}^{∞}β_r^{(n)}L^rγ_n. Since L^rγ_n(t)=γ_n(t-r) and γ_n is supported on (0,n), a term with r≥n is nonzero whenever n≤t<r+n. The paper truncates this sum at r=n-1 without proof. The remark after (4.7) argues that γ_n(t)=0 for t≥n implies only r≤n-1 is needed; this reverses the support condition, which is r<t<r+n, i.e. r<t. Concrete failure: take g=a0 rect + a1 L rect with a0,a1>0 and a0+a1<1. For n=2, A^2=a0^2 I+2a0a1 L + a1^2 L^2, so β_2^{(2)}=a1^2. At t=5/2, L^2γ_2(t)=γ_2(1/2)=1/2, so the omitted term a1^2/2 belongs to g^{*2}(5/2); direct convolution gives g^{*2}(5/2)=a0a1+a1^2/2, while (2.26) includes only a0a1. Thus (2.26) is false, and the explicit algorithm built on it is not a solution method for the stated Volterra equation. The same truncation appears in corollary (4.5).","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an explicit numerical algorithm for the Volterra integral equation of the second kind, y(t)=f(t)+∫_0^t g(t−τ)y(τ)dτ, with emphasis on the case f=δ, where the fundamental solution is E=δ+h and h satisfies h=g+h∗g. The main idea is to approximate a general kernel g by a piecewise constant step kernel, write the step kernel as g=A rect with A=Σ a_j L^j, and express the n-fold convolution as g^{∗n}=A^n γ_n. The paper claims an explicit closed form for h, proves a scaling lemma to handle arbitrary step size δ, and gives L1 and L∞ stability bounds (1.6) and (1.7) showing that approximating g by a step kernel produces a controlled error. Applications to Hawkes processes and expressions for y=f+h∗f are also discussed.","tokens_in":13008,"tokens_out":13081,"duration_ms":150229,"significance":"If the explicit formula for h were correct, the paper would supply a practical, non-Monte-Carlo method for evaluating Hawkes-type intensity solutions with delta inputs, and the stability estimates (1.6) and (1.7) are useful and broadly standard. The paper also provides a self-contained derivation of the identity h=Σ g^{∗n} and a scaling argument for step size reduction. However, the central explicit formula in Theorem 2.7 is false as stated, and the error propagates into the numerical algorithm and the claimed finite termination of the series. The underlying idea is repairable, but the manuscript in its current form does not deliver the promised exact algorithm.","major_comments":[{"comment":"The truncation of the inner sum at r=n−1 is invalid. By Theorem 2.6, g^{∗n}=Σ_{r=0}^∞ β_r^{(n)} L^r γ_n, and L^rγ_n(t)=γ_n(t−r). The support condition is r<t<r+n, so terms with r≥n are nonzero whenever t>n. The remark after (4.7), which argues that γ_n(t)=0 for t≥n implies only r≤n−1 is needed, reverses the support condition. A concrete failure is obtained by taking g=a0 rect + a1 L rect with a0,a1>0 and a0+a1<1. For n=2, A^2=a0^2 I+2a0a1 L+a1^2 L^2. At t=5/2, direct convolution gives g^{∗2}(5/2)=a0a1+a1^2/2, whereas the right-hand side of (2.26) gives only a0a1 because the r=2 term a1^2γ_2(1/2) is omitted. Thus (2.26) is false, and the algorithm built on it is not an exact solution method.","section":"§2, Theorem 2.7 and eq. (2.26)"},{"comment":"The recurrence for β_r^{(n+1)} sums l only from 0 to n+1, but A=Σ_{j=0}^∞ a_j L^j has infinitely many nonzero coefficients in general. The coefficient of L^r in A^{n+1} is Σ_{l=0}^{r} a_l β_{r−l}^{(n)} (with β_{negative}=0), not a sum truncated at n+1. For example, β_3^{(2)} from A^2 includes the term a_0a_3, which is omitted by the stated upper limit. If the corrected full series for h is used, this recurrence must be corrected accordingly.","section":"§2, Lemma 2.5 and eq. (2.23)"},{"comment":"The claim that for 0≤t≤T only finitely many terms of the series are nonzero is false. For fixed t>0, γ_n(t)>0 for all n>t, so the outer series over n has infinitely many nonzero terms. The correct statement is that the tail can be made small using the bound Σ_{n>N} ‖g^{∗n}‖_1 ≤ k^{N+1}/(1−k), which introduces a tolerance rather than exact finite termination. Section 7's assertion of 'no approximation' therefore needs to be replaced by an explicit tail-estimate argument.","section":"§7, Numerical Implementation"}],"minor_comments":[{"comment":"The outer summation in (2.26) starts at n=0, but h is defined as Σ_{n=1}^∞ g^{∗n}; the n=0 term is undefined and should be removed or the index changed to n≥1.","section":"§2, Theorem 2.7"},{"comment":"In the displayed inequality leading to (8), the term arising from g2∗(h1−h2) is printed with ‖g1−g2‖_t in the second summand; it should be ‖h1−h2‖_t. As printed, the displayed estimate does not close the argument.","section":"§10, sup-norm proof"},{"comment":"Applying (1.7) to the estimate (6.1) yields ‖h−hδ‖∞ ≤ kθ(1+θ)δ/(1−k)^2, not kθ(1+θ)δ/(1−k); the displayed bound misses one factor of (1−k)^{-1}.","section":"§6, Example 2, eq. (6.2)"},{"comment":"The paper is titled as a numerical algorithm but contains no numerical experiments or implementation; Section 8 states that implementation is future work. A small computed example illustrating the corrected formula would substantially strengthen the manuscript.","section":"§8, Future Outlook"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the stress-test concern about Theorem 2.7 is correct and is the central issue of the manuscript. The false truncation is repairable by replacing the truncated sums with full sums over r and adding an explicit tail bound, so I recommend major revision rather than rejection. The paper is also in a preliminary state: no numerical experiments, no code, and the arXiv source contains heavy OCR corruption that makes verification difficult."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou can skip the heavy math unless you work on Hawkes processes or Volterra numerics, but if you do, the punchline is that the central explicit formula (Theorem 2.7) is wrong. The paper truncates the series for g^{*n} = A^n γ_n = ∑_{r≥0} β_r^{(n)} L^r γ_n at r = n−1. That is not justified. The support condition is that γ_n(s) is zero outside (0,n), so L^r γ_n(t) = γ_n(t−r) is nonzero whenever 0 < t−r < n, i.e. r ∈ (t−n, t). That includes r ≥ n as soon as t > n. The remark after (4.7) has the implication reversed: γ_n(t)=0 for t≥n does not imply L^rγ_n(t)=0 for r≥n.\n\nConcrete failure: take g = a0 rect + a1 L rect with a0, a1 > 0 and a0+a1 < 1. For n=2, the convolution g^{*2} includes a1^2 L^2 γ_2. At t=5/2, that term contributes a1^2/2. Direct convolution gives g^{*2}(5/2)=a0a1 + a1^2/2, while the truncated formula (2.26) omits the a1^2/2. So the algorithm is not a correct solution method for even a two-step kernel.\n\nWhat is actually good: the Neumann-series splitting E=δ+h, the operator A = ∑ a_j L^j, and the recurrence (2.22)-(2.23) are correct and neat. Theorem 2.6 is fine. The scaling lemma (3.2) and the L1/sup error bounds in the appendix are standard and correctly derived. The aim—an explicit, implementable series with controlled accuracy—is sensible.\n\nThe bigger soft spot beyond the bug is that there is no implementation or numerical experiment. That is unusual for a paper calling itself an explicit numerical algorithm, and a simple sanity check would have exposed the error. The contraction assumption ||g||_1 < 1 is clearly stated and fine for the intended Hawkes application.\n\nThe reader's report gave soundness 7 and a conditional recommendation; I disagree. The error is load-bearing. Theorem 2.7 and Corollary (4.5) are false as written. The paper might be repairable by keeping the full sum over r and truncating based on t, but until that is done and demonstrated, the contribution is not usable. I would not send this to peer review in its current form; a desk reject with a clear explanation is the right call. If the authors fix the series and add one numerical test, it becomes a reasonable short computational note.","headline":"The paper's central explicit solution formula truncates the convolution series incorrectly, making the main algorithmic claim false, despite some correct and neat algebraic machinery.","tokens_in":13515,"tokens_out":5562,"would_cite":false,"duration_ms":57580,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["45D05","65R20","45E10"],"pacs":[],"model":"deepseek-v4-flash","headline":"For a step-function kernel, a Volterra integral equation of the second kind has an exact solution given by an explicit series, and approximating a general kernel by steps gives a certified error bound controlled by the kernel's $L^1$ norm.","keywords":["Volterra integral equation","second kind","convolution kernel","piecewise constant approximation","explicit algorithm","Hawkes process","Neumann series","error bounds"],"falsifier":"Take the Rayleigh kernel $g(t)=\\frac{k t}{\\sigma^2}e^{-t^2/(2\\sigma^2)}$ with, say, $k=0.99$ and $\\sigma=1$, approximate it by step functions of width $\\delta$ using the paper's sampled coefficients, and compare $\\|h-h_\\delta\\|_1$ computed from Theorem 2.7 (via the scaling identity) against a high-accuracy direct solve of the Volterra equation; inequality (1.6) predicts the error stays below $(1-k)^{-2}\\|g-g_\\delta\\|_1$, so any violation at some $\\delta$ refutes the bound. A sharper check: for the step kernel with $g(t)=1$ on $[0,1)$ and $0$ elsewhere, $k=1$, the series in Theorem 2.7 diverges term by term, so the theorem's conclusion should fail exactly at the boundary.","tokens_in":12467,"feed_emoji":"📐","tokens_out":14099,"duration_ms":126068,"temperature":0.7,"pith_summary":"The paper considers the Volterra equation $y(t)=f(t)+\\int_0^t g(t-\\tau)y(\\tau)\\,d\\tau$, the linear integral equation that underlies self-exciting point-process models. Its aim is to replace numerical quadrature or iterative solution by an explicit algorithm: approximate the memory kernel $g$ by a piecewise-constant step function $g_\\delta$, for which the fundamental solution $h$ of $h=g+h*g$ is shown to have an exact closed form. The main results are the explicit double-series formula for $h$ in Theorem 2.7, a scaling identity that reuses one unit-step computation at any step size $\\delta$, and error bounds showing that $\\|h-h_\\delta\\|$ is at most $(1-k)^{-2}\\|g-g_\\delta\\|$ in $L^1$ and supremum norm when $k=\\|g\\|_1<1$. A sympathetic reader cares because the method turns a class of integral equations into a finite, explicit computation whose accuracy can be certified before the computation starts, at the price of choosing a step size.","feed_headline":"Step-function kernels make Volterra equations exactly solvable","feed_subtitle":"No quadrature, no iteration: a step-function kernel yields an exact series, and the step-size error is certified.","key_machinery":"The machinery is a small algebra of operators on $L^1_+(\\mathbb{R})$: the unit shift $L$ with $(Lf)(t)=f(t-1)$, the difference $\\Delta=I-L$, the integration operator $J$, and the scaling operator $S_\\delta$ with $(S_\\delta f)(t)=\\frac{1}{\\delta}f(t/\\delta)$. The step kernel is written $g=A\\,\\mathrm{rect}$ with the operator series $A=\\sum a_j L^j$, and the $n$-fold convolution becomes $g^{*n}=A^n\\gamma_n$, because the Laplace transform of $\\mathrm{rect}$ factors as $(\\widehat{J})(\\widehat{\\Delta})$; $\\gamma_n$ is the explicit piecewise-polynomial function from (2.15), and $\\beta_r^{(n)}$ are the coefficients of $A(z)^n$. The scaling identity $L_\\delta=S_\\delta L S_\\delta^{-1}$ transports the whole construction from unit steps to steps of size $\\delta$, while the error bounds (1.6)-(1.7) are derived in the appendix from the Neumann series and the convolution norm inequality, using the assumption that both kernels have $L^1$ norm $k<1$.","core_discovery":"For a kernel of the form $g(t)=\\sum_{j\\ge 0} a_j\\,\\mathrm{rect}(t-j)$ with $\\mathrm{rect}(t)=1$ on $[0,1)$ and $0$ elsewhere, the paper proves that the solution of $h(t)=g(t)+\\int_0^t g(t-\\tau)h(\\tau)\\,d\\tau$ is $$h(t)=\\sum_{n\\ge 1}\\sum_{r=0}^{n-1}\\$beta_r^{{(n)}}$\\,\\gamma_n(t-r),$$ where $\\gamma_n$ is the $n$-fold convolution of $\\mathrm{rect}$ with itself, given explicitly by $\\gamma_n(t)=\\sum_{r=0}^n (-1)^r\\binom{n}{r}\\frac{(t-r)_+^{n-1}}{(n-1)!}$, and $\\beta_r^{(n)}$ are the coefficients of $(\\sum_j a_j z^j)^n$ obtained from the recurrence (2.22)-(2.23). The paper then shows that if $g_\\delta$ is the width-$\\delta$ step approximation built from samples $g(j\\delta)$, its solution obeys $h_\\delta(t)=\\frac{1}{\\delta}h(t/\\delta)$, so a single unit-step computation serves every $\\delta$. If $\\|g\\|_1\\le k<1$ and $\\|g_\\delta\\|_1\\le k<1$, the error is bounded by $(1-k)^{-2}\\|g-g_\\delta\\|$ in both $L^1$ and supremum norms, and the same factor controls the error in the convolutions $h*f$ that give the solution $y$ for general inputs $f$.","pith_inferences":["The algebra of the coefficients $\\beta_r^{(n)}$ never uses positivity of the $a_j$, so the same explicit series and error bounds should hold for sign-changing kernels with $\\|g\\|_1<1$; the paper's nonnegative examples, like the Rayleigh kernel, are a special case.","The functions $\\gamma_n$ are the probability densities of sums of $n$ independent uniform random variables, so the series for $h$ can be read as an infinite mixture of such densities; this points toward probabilistic interpretations and possibly renewal-theoretic proofs of the convergence.","The constant $(1-k)^{-2}$ is derived by a crude contraction argument; a natural testable conjecture is that for monotone nonnegative kernels the factor improves to $(1-k)^{-1}$, which a comparison against direct numerical solutions can decide.","The condition $k<1$ is exactly the subcritical regime of the self-exciting point process that motivates the equation, so the algorithm covers the stable branch of the application space; extending to $k\\ge1$ would require a different representation and is left open by the paper."],"forward_implications":["For inputs $f$ that are a finite Dirac comb plus a locally integrable part, the solution is assembled directly as $y(t)=f(t)+\\sum_i w_i h(t_i)+(h*f_1)(t)$ using the explicit $h$, so no quadrature near the impulses is needed.","For a kernel with Lipschitz constant $L_g$, choosing $\\delta$ makes the supremum error at most $\\frac{L_g\\,\\delta}{(1-k)^2}$, giving an a priori recipe: pick $\\delta$ to meet a tolerance, then compute the series once.","Because $\\gamma_n(t)=0$ for $t\\ge n$, only finitely many terms of the double series contribute on $[0,T]$, so the algorithm terminates after a finite number of arithmetic operations.","The scaling lemma means the unit-step solution $h$ is a reusable table: recomputing for a different step size $\\delta$ costs only a rescaling of the input and output."],"supporting_citations":[],"fun_headline_variants":["Step kernels give exact Volterra solutions via explicit series","Explicit series solves Volterra exactly for step-function kernels","Volterra made exact: step kernels yield closed-form series","Certified error: step-kernel Volterra solver with exact series"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method depends on the kernel's $L^1$ norm being strictly smaller than 1, since that is what makes the Neumann series converge and what keeps the factor $1/(1-k)^2$ in the error bounds finite.","fun_headline_variants_meta":{"raw":{"variants":["Step kernels give exact Volterra solutions via explicit series","Explicit series solves Volterra exactly for step-function kernels","Volterra made exact: step kernels yield closed-form series","Certified error: step-kernel Volterra solver with exact series"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000352,"raw_usage":{"total_tokens":2002,"prompt_tokens":1111,"completion_tokens":891,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":727,"completion_tokens_details":{"reasoning_tokens":823}},"tokens_in":727,"tokens_out":891,"duration_ms":8668,"temperature":1.0,"reasoning_tokens":823,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:31:04.129456+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the Rayleigh kernel $g(t)=\\frac{k t}{\\sigma^2}e^{-t^2/(2\\sigma^2)}$ with, say, $k=0.99$ and $\\sigma=1$, approximate it by step functions of width $\\delta$ using the paper's sampled coefficients, and compare $\\|h-h_\\delta\\|_1$ computed from Theorem 2.7 (via the scaling identity) against a high-accuracy direct solve of the Volterra equation; inequality (1.6) predicts the error stays below $(1-k)^{-2}\\|g-g_\\delta\\|_1$, so any violation at some $\\delta$ refutes the bound. A sharper check: for the step kernel with $g(t)=1$ on $[0,1)$ and $0$ elsewhere, $k=1$, the series in Theorem 2.7 diverges term by term, so the theorem's conclusion should fail exactly at the boundary.","supporting_citations":[],"review_version":1}