{"id":"e07fb9c3-16ec-4106-95c5-c1c47b5eb62e","arxiv_id":"2504.19885","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An inverse-Gaussian implicit scheme for integrated Volterra square-root processes is proved weakly convergent and shown numerically accurate with very few time steps, including for hyper-rough fractional kernels.","lead":"This paper introduces a fast new way to simulate integrated Volterra square-root processes, which underpin rough and hyper-rough Heston models in finance. The method uses an Inverse Gaussian distribution at each step, keeps the simulated integrated variance non-decreasing, and needs very few time steps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'any L1 kernel' claim is false: nonnegativity preservation is not automatic, and a simple indicator kernel makes α_i negative.","rationale":"The reader's conditional verdict is the right one. The paper's main convergence theorem, Theorem 2.2, is a conditional statement and its proof structure is coherent apart from the one-word sign typo in Lemma 5.4. The decisive weakness is the gap between the advertised scope and the assumptions actually used: well-definedness of the scheme depends on the integrated kernel preserving nonnegativity, which Proposition 1.4 establishes only for completely monotone kernels. The indicator-kernel counterexample shows that this is not a harmless technical condition implied by L1 membership, and that the Introduction's claim of handling any K in L1 is literally false. All reported numerics use fractional or shifted-fractional kernels, which are completely monotone, so the practical conclusions of the paper stand; however, the paper should present Assumption 2.1 as an explicit restriction on the kernel class and revise the universal-L1 wording in the abstract and introduction. This is precisely the same load-bearing concern the reader identified, so my stress-test does not change the verdict.","tokens_in":29694,"tokens_out":18404,"duration_ms":190812,"concrete_test":"Run the scheme with K(t)=1_{[0,1]}(t), b=0, c=1, g0≡1, n=5, T=1.25. From (0.5)-(0.8), α0=1/4, α1=U0, α2=U1, α3=U2, and α4=1/4+U3−U0. Simulate 10^6 paths and record the frequency of α4<0. If that frequency is positive, as it must be because all involved IG laws have positive density on (0,∞), the well-definedness claim for general L1 kernels fails. A purely analytic version is to evaluate P(U0>2, U1<1/8, U2<1/8, U3<1/8)>0 by iterated conditioning; this event forces α4<0.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition is well-definedness of Algorithm 1, which requires α_i ≥ 0 so that the Inverse Gaussian parameters in (0.7) are valid. The proof rests on Assumption 2.1 / Theorem 1.3: the integrated kernel \\bar K in (1.8) must be non-increasing and preserve nonnegativity. Proposition 1.4 verifies this only for completely monotone kernels, yet the Introduction claims the scheme works 'for any kernel K in L1([0,T])'. This is not merely unproved; it is false. Take K(t)=1_{[0,1]}(t), which is nonnegative, non-increasing, and in L1. With b=0, c=1, g0≡1, and Δ=T/n=1/4, (0.5) gives k0=k1=k2=k3=1/4 and k4=0. The recursion yields α0=1/4, α1=U0, α2=U1, α3=U2, and α4=1/4+U3−U0. Since the IG law has positive density on (0,∞), the event {U0>2, U1<1/8, U2<1/8, U3<1/8} has positive probability, and on this event α4<0. Thus Algorithm 1 can be undefined even for a simple nonnegative non-increasing L1 kernel, so the advertised L1 generality must be qualified. Separately, Lemma 5.4 states that \\bar K^n is 'non-decreasing' under Assumption 2.1, which says 'non-increasing'; the intended correction is clear, but as written the tightness proof reads inconsistently.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces an Inverse-Gaussian-based implicit Euler-type scheme (the iVi scheme) for the integrated Volterra square-root process U_{0,t} and for pricing in Volterra Heston models. The increment bU_{i,i+1} is sampled from an Inverse Gaussian law whose parameters are formed from integrated kernel quantities, and bZ_{i,i+1} is set by the implicit relation, preserving the non-decreasing property of U. The main theoretical results are: Theorem 1.3, showing well-definedness and nonnegativity under an assumption on the integrated kernel; and Theorem 2.2, proving J1-tightness of the piecewise-constant scheme and characterizing every accumulation point as a solution of the limiting stochastic Volterra equation, with weak convergence following under weak uniqueness (Corollary 2.3). The numerical section reports fast convergence for fractional kernels, including negative Hurst indices, with reference values from Fourier inversion and Monte Carlo noise floors.","tokens_in":30022,"tokens_out":16102,"duration_ms":153580,"significance":"If the claims are read in their formal scope, this is a substantial contribution: a simple, hyperparameter-free scheme with a genuine convergence proof under an explicit kernel condition, together with a self-contained stability theorem for stochastic Volterra equations with measure-valued kernels. The martingale property of the discretized Z and the identification of its limiting quadratic variation are proved in detail, and the numerical study is careful, using Fourier-inversion reference values and reporting Monte Carlo noise levels. However, the advertised L1-kernel generality is not supported by the formal assumptions and is in fact false as stated; the scope must be narrowed or the scheme must be modified before the paper can be accepted.","major_comments":[{"comment":"The paper advertises in the abstract and introduction that the scheme simulates U in (0.3) for any kernel K in L1([0,T]). The formal well-definedness result, Theorem 1.3, requires the integrated kernel \\bar K to be non-increasing and to preserve nonnegativity (Assumption 2.1), and this property is only verified in Proposition 1.4 for kernels with the completely monotone-type representation (1.9). The advertised L1 claim is false. A concrete counterexample is obtained with T=5/4, n=5, \\Delta=1/4, K(t)=1_{[0,1]}(t), b=0, c=1, g0\\equiv 1. Then (0.5) gives k0=k1=k2=k3=1/4 and k4=0, and writing X_i=\\hat U_{i,i+1} the recursion (0.6)-(0.8) yields \\alpha_0=1/4, \\alpha_1=X_0, \\alpha_2=X_1, \\alpha_3=X_2 and \\alpha_4=\\Delta+X_3-X_0. On the positive-probability event \\{X_0>3,\\ X_1<0.1,\\ X_2<0.1,\\ X_3<0.1\\} we have \\alpha_4<0, so the Inverse Gaussian parameters in (0.7) are invalid at the last step of Algorithm 1. The scheme is therefore undefined for a nonnegative, non-increasing L1 kernel. This does not contradict Theorem 1.3 as stated, but it contradicts the paper's advertised scope. The authors should either restrict all claims to kernels satisfying Assumption 2.1 (and state the verified examples), or modify the scheme, for example by replacing \\alpha_i with \\alpha_i^+, and prove well-definedness for that modified scheme.","section":"Introduction, Theorem 1.3, Assumption 2.1"},{"comment":"Proposition 1.4 is the bridge that turns Theorem 1.3 into a usable statement for fractional and exponential kernels, so its proof needs to be correct as written. Two issues need attention. First, the displayed formula for \\nu(dx) has a sign error: when K(t)=\\int e^{-xt}\\mu(dx), Fubini gives \\bar K(t)=\\int e^{-xt}(1-e^{-xT/n})/x\\,\\mu(dx), so the exponent in the paper's formula should be e^{-xT/n}, not e^{xT/n}. Second, the statement allows \\mu to be a measure of locally bounded variation, which may be signed; it is not automatic that Alfonsi (2025, Theorem 2.11) applies to such \\nu. Please clarify whether \\mu is assumed nonnegative (as in the completely monotone case of Example 1.5) or supply the additional argument needed for signed \\mu. This is important because the numerical examples rely on this proposition for the fractional and shifted-fractional kernels.","section":"Proposition 1.4"}],"minor_comments":[{"comment":"In the proof of Lemma 5.4, the sentence 'In virtue of Assumption 2.1, \\bar K^n is non-decreasing and therefore (k_i^n) is a non-decreasing sequence' should read 'non-increasing' in both places. The subsequent inequality uses k_{j-l}^n-k_{i-l}^n \\ge 0, which is exactly what the non-increasing case gives; as printed the sign is inconsistent.","section":"Section 5.2, Lemma 5.4 proof"},{"comment":"The discretized Riccati equation in Remark 1.7 contains a typo: the term 'bk0\\psi2(\\Delta_i)' should be 'bk0\\psi(\\Delta_i)', and the quadratic term should be written as k0 c^2/2 \\psi(\\Delta_i)^2. As printed, the expression is dimensionally inconsistent and does not match the quadratic equation (1.14).","section":"Remark 1.7"},{"comment":"The paper repeatedly uses the phrase 'any kernel K in L1([0,T])' in the abstract and introduction. Even after the main scope is corrected in Theorem 1.3, the abstract and the 'Relation to the Literature' paragraph should be reworded so that the formal assumption (integrated kernel non-increasing and nonnegativity-preserving) is aligned with the claimed scope.","section":"General presentation"}],"recommendation":"major_revision","confidential_remarks":"The core convergence argument under Assumption 2.1 appears sound and the numerical study is convincing for the completely monotone kernels covered by Proposition 1.4. The decisive issue is the mismatch between the advertised L1 generality and the actual assumptions, which is not merely a wording problem: the indicator-kernel example shows the scheme can be undefined for an admissible L1 kernel. This is fixable by narrowing the claims or by adding a clipping/well-definedness repair to the algorithm, so I recommend major revision rather than rejection. The paper also leans heavily on the first author's prior work for weak uniqueness and the nonnegativity-preservation toolkit; this is legitimate, but the new contributions should be stated more carefully against that background."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Off the record: this paper is worth your time. It extends Abi Jaber's IG-implicit scheme from the exponential kernel to general Volterra kernels, including singular L1 kernels, by discretizing through integrated kernel quantities. The scheme is simple, keeps the integrated process non-decreasing without clipping, and comes with an honest weak convergence proof via a measure-kernel reformulation. Sections 4 and 5 contain real mathematics: moment estimates, tightness, and a stable limit characterization. The numerics are also done responsibly, with Fourier-inversion reference values and Monte Carlo noise levels marked. For rough and hyper-rough Heston simulation, this is a useful contribution.\n\nThe main soft spot is the scope claim. The abstract and introduction say the scheme handles 'any L1 kernel', but the rigorous condition is Assumption 2.1: the integrated kernel \\bar K must be non-increasing and preserve nonnegativity. That is not automatic. The stress-test counterexample is correct: for K=1_{[0,1]}, g0=1, b=0, c=1, T=1, n=4, one gets α0=1/4, α1=U0, α2=U1, α3=U2, α4=1/4+U3-U0, and the event {U0>2, U1,U2,U3<1/8} has positive probability, making α4 negative and the IG parameters invalid. So the 'any L1' claim is false as stated. The fix is straightforward: state theorems under Assumption 2.1, give the completely monotone case as the main example, and soften the abstract and introduction. There is also a likely typo in Lemma 5.4, where \\bar K^n is called non-decreasing; the proof needs non-increasing, matching Assumption 2.1.\n\nMinor caveats: no code is released, and the observed faster convergence as H → -1/2 is empirical, not proven. Neither undermines the core contribution. The self-citation to Abi Jaber (2021) for weak uniqueness is appropriate, not a red flag.\n\nBottom line: this paper deserves a serious referee. It needs revision to qualify the kernel assumptions and fix the typo, but the core scheme and convergence result are solid and useful for the rough-volatility community.","headline":"A genuinely useful IG-based simulation scheme for integrated Volterra square-root processes with a real convergence proof, but the advertised 'any L1 kernel' scope is false and needs qualification before it is publishable.","tokens_in":30546,"tokens_out":4721,"would_cite":true,"duration_ms":45487,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H20","60H35","91G20","60G44"],"pacs":[],"model":"deepseek-v4-flash","headline":"An inverse-Gaussian one-step discretization of integrated Volterra square-root processes is proven to converge weakly, with non-decreasing simulated paths and fast numerics even for negative Hurst indices.","keywords":["integrated Volterra square-root process","Volterra Heston model","Inverse Gaussian distribution","weak convergence","Skorokhod J1 topology","rough volatility","nonnegativity-preserving kernels","stochastic Volterra equations"],"falsifier":"Choose a nonnegative $L^1$ kernel with a singularity that is not completely monotone, for example a truncated or shifted version of $t^{-1/2}$ whose integrated average $\\bar K$ has an increasing segment, and run Algorithm 1 on a coarse grid with constant $g_0$, computing the $\\alpha_i$ recursively; if any $\\alpha_i<0$ appears, the well-definedness claim (Theorem 1.3) fails as stated, while finding no such kernel would support widening the assumption to all $L^1$ kernels.","tokens_in":29487,"feed_emoji":"📈","tokens_out":11676,"duration_ms":95927,"temperature":0.7,"pith_summary":"This paper introduces a simulation scheme, iVi (integrated Volterra implicit), for the integrated Volterra square-root process $U_t=\\int_0^t V_s\\,ds$, the object that drives Volterra Heston models. The scheme discretizes the dynamics of $U$ directly, using one Inverse Gaussian draw per time step, and is designed for $L^1$ kernels with singularities such as the fractional kernel $t^{H-1/2}$ with $H\\in(-1/2,1/2]$. The authors prove that the piecewise constant processes produced by the scheme are tight in the Skorokhod $J_1$ topology and that every accumulation point solves the original integrated Volterra equation with the correct martingale structure; under weak uniqueness this gives weak convergence. They also prove the scheme is well defined, keeps the simulated integrated process non-decreasing, and report that numerically a handful of time steps already matches reference option prices, with convergence accelerating as the Hurst index approaches $-1/2$.","feed_headline":"Inverse-Gaussian scheme converges for Volterra square-root processes","feed_subtitle":"Simulates singular rough-Heston kernels without clipping; weak convergence proved, few steps needed.","key_machinery":"The load-bearing reformulation writes the discrete scheme as a stochastic Volterra equation with a measure kernel, $U^n_t=\\int_0^{\\lfloor nt/T\\rfloor T/n}g_0(s)\\,ds+\\int_{[0,t]}(bU^n_{t-s}+cZ^n_{t-s})K^n(ds)$ with $K^n=\\sum_{i=0}^{n-1}k_i^n\\delta_{t_i^n}$, which turns the non-Markovian recursion into a convolution equation whose stability can be studied by $L^1$ arguments. The Inverse Gaussian law itself comes from the Dambis-Dubins-Schwarz representation: writing $Z$ as a time-changed Brownian motion, the implicit step $(1-bk_0)\\hat U=\\alpha_i+ck_0\\tilde W_{\\hat U}$ identifies $\\hat U_{i,i+1}$ as a first passage time, whose distribution is Inverse Gaussian. Well-definedness is carried by the requirement that the integrated kernel $\\bar K(t)=\\int_0^{T/n}K(t+s)\\,ds$ be non-increasing and preserve nonnegativity, a condition verified for completely monotone kernels.","core_discovery":"At each step $i$, Algorithm 1 computes $\\alpha_i=\\int_{t_i}^{t_{i+1}}g_0(s)\\,ds+\\sum_{j<i}k_{i-j}(b\\hat U_{j,j+1}+c\\hat Z_{j,j+1})$ from integrated kernel weights $k_\\ell=\\int_0^{T/n}K(\\ell T/n+s)\\,ds$, then samples $\\hat U_{i,i+1}$ from an Inverse Gaussian law with mean $\\alpha_i/(1-bk_0)$ and shape $(\\alpha_i/(c k_0))^2$, and sets $\\hat Z_{i,i+1}=((1-bk_0)\\hat U_{i,i+1}-\\alpha_i)/(c k_0)$. The paper's central result is that the piecewise constant processes $U^n_t=\\sum_{i<nt/T}\\hat U_{i,i+1}$ and $Z^n_t=\\sum_{i<nt/T}\\hat Z_{i,i+1}$ are $J_1$-tight and any accumulation point $(U,Z)$ satisfies $U_t=\\int_0^t g_0(s)\\,ds+\\int_0^tK(t-s)(bU_s+cZ_s)\\,ds$ with $Z$ a continuous square-integrable martingale whose quadratic variation is $U$. When the law of the limit is unique, the scheme converges weakly (Theorem 2.2 and Corollary 2.3). The proof also shows $\\alpha_i\\ge 0$ and that $\\hat U$ is non-decreasing, so no clipping is needed.","pith_inferences":["Extending the nonnegativity-preservation assumption from $\\bar K$ to broader classes than completely monotone kernels would widen the proven scope; Proposition 1.4 already suggests Bernstein representations with signed measures of locally bounded variation may suffice.","The single-step accuracy as $H\\to -1/2$ suggests a testable closed-form approximation: compare the one-step iVi distribution with the explicit Inverse Gaussian limit law established for the continuous process, and measure the gap as a function of $H+1/2$.","Eliminating the drift term $bU$ via the resolvent kernel $R_b$, as the paper sketches, should remove an additional discretization bias; a numerical comparison of the original and resolvent-based versions would quantify that bias.","The finite-$n$ martingale property of $\\hat Z$ suggests the scheme could be used as a control variate or as the basis of a multilevel estimator, since the bias in $\\langle Z^n\\rangle$ versus $U^n$ is explicit."],"forward_implications":["The scheme applies to $L^1$ kernels with singularities, including the fractional kernel with negative Hurst index, without reconstructing the instantaneous variance $V$.","Under weak uniqueness of solutions, the discretized processes converge weakly to the integrated Volterra square-root process, so option prices computed from the scheme inherit the model's limit behavior.","The simulated integrated process is non-decreasing by construction, eliminating the need for clipping or projection steps that other schemes require.","For the fractional kernel, numerical convergence improves as $H$ decreases toward $-1/2$, the opposite of existing schemes, and one time step already approximates the limiting Inverse Gaussian regime.","Because the scheme needs very few steps, it mitigates the quadratic cost in the number of steps inherent to non-Markovian simulation."],"supporting_citations":[{"why":"Supplies the nonnegativity-preserving kernel tools and the results used to prove $\\alpha_i\\ge 0$ for the integrated kernel.","marker":"Alfonsi (2025)"},{"why":"Establishes the integrated Volterra formulation for $L^1$ kernels and the weak uniqueness invoked in Corollary 2.3.","marker":"Abi Jaber (2021)"},{"why":"Introduces the original iVi scheme for the standard square-root process that this scheme generalizes.","marker":"Abi Jaber (2024)"},{"why":"Provides the acceptance-rejection algorithm used to sample Inverse Gaussian increments.","marker":"Michael, Schucany, and Haas (1976)"},{"why":"Establishes the limiting Inverse Gaussian law as $H\\to -1/2$, used to explain the scheme's fast convergence in the hyper-rough regime.","marker":"Abi Jaber, Attal, and Rosenbaum (2025)"},{"why":"Provides the affine Volterra framework and the characteristic function used for reference option prices.","marker":"Abi Jaber, Larsson, and Pulido (2019)"},{"why":"Serves as the explicit-scheme benchmark against which the iVi scheme is compared.","marker":"Richard, Tan, and Yang (2023)"}],"fun_headline_variants":["Inverse Gaussian scheme tames rough Volterra kernels","Volterra Heston: no clipping, weak convergence in few steps","New iVi scheme: singular kernels, few steps, proven convergence","Rough kernels no match: iVi converges as H approaches -1/2","Integrated Volterra sim: Inverse Gaussian does it fast"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the kernel's average over each time step, $\\bar K(t)=\\int_0^{T/n}K(t+s)\\,ds$, is a non-increasing and nonnegative function; this is verified for completely monotone kernels, but not for every $L^1$ kernel with a singularity that the paper claims to handle, and if it fails the $\\alpha_i$ can become negative and the Inverse Gaussian sampling breaks down.","fun_headline_variants_meta":{"raw":{"variants":["Inverse Gaussian scheme tames rough Volterra kernels","Volterra Heston: no clipping, weak convergence in few steps","New iVi scheme: singular kernels, few steps, proven convergence","Rough kernels no match: iVi converges as H approaches -1/2","Integrated Volterra sim: Inverse Gaussian does it fast"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000784,"raw_usage":{"total_tokens":3491,"prompt_tokens":1003,"completion_tokens":2488,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":2400}},"tokens_in":619,"tokens_out":2488,"duration_ms":19893,"temperature":1.0,"reasoning_tokens":2400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:40:22.070273+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a nonnegative $L^1$ kernel with a singularity that is not completely monotone, for example a truncated or shifted version of $t^{-1/2}$ whose integrated average $\\bar K$ has an increasing segment, and run Algorithm 1 on a coarse grid with constant $g_0$, computing the $\\alpha_i$ recursively; if any $\\alpha_i<0$ appears, the well-definedness claim (Theorem 1.3) fails as stated, while finding no such kernel would support widening the assumption to all $L^1$ kernels.","supporting_citations":[{"cited_title":"Nonnegativity preserving convolution kernels","cited_arxiv_id":null,"evidence_quote":"Supplies the nonnegativity-preserving kernel tools and the results used to prove $\\alpha_i\\ge 0$ for the integrated kernel."},{"cited_title":"Weak existence and uniqueness for affine stochastic V olterra equations with L1 -kernels","cited_arxiv_id":null,"evidence_quote":"Establishes the integrated Volterra formulation for $L^1$ kernels and the weak uniqueness invoked in Corollary 2.3."},{"cited_title":"On the discrete-time simulation of the rough H eston model","cited_arxiv_id":null,"evidence_quote":"Serves as the explicit-scheme benchmark against which the iVi scheme is compared."}],"review_version":1}