{"id":"9b359d3b-ff67-4e9d-92cf-cf80bfc5306d","arxiv_id":"2412.11264","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The iVi scheme simulates the integrated square-root process via an Inverse Gaussian law, preserving non-negativity and first moments, and shows high accuracy in Heston tests with as few as one time step.","lead":"This paper introduces a simulation scheme for the square-root (CIR) process that first simulates the integrated process as an Inverse Gaussian random variable. This yields a simple, nonnegative, accurate method for Heston option pricing with very few time steps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Accuracy hinges on an unproved right-endpoint approximation that makes the simulated Z conditionally degenerate given U; the paper defers convergence analysis to a companion paper, so the central claim is not yet self-contained.","rationale":"The reader's conditional verdict is sound. The iVi scheme is genuinely simple and nonnegative, Theorem 1.3 and Proposition 1.4 are correct as far as I checked, and the numerical experiments in Sections 2 and 3 are encouraging across the tested parameter sets. The weakest link is the discretization step in Section 1.1, which the paper itself does not analyze: the scheme's joint law for (U,Z) is degenerate conditional on U, and this is exactly what the Heston update (3.2) uses through the rho Z term. The paper defers convergence to Abi Jaber and Attal (2025), so the present manuscript does not, on its own, establish the central accuracy claim. That gap, rather than any internal inconsistency, supports keeping the verdict conditional. The proposed characteristic-function test would directly quantify the one-step joint-law error, and a clean match would weaken the concern substantially.","tokens_in":16750,"tokens_out":22879,"duration_ms":209426,"concrete_test":"Use the exact Heston characteristic function (B.1)-(B.2) as reference. For a single step of length delta = T = 1 with Case 1 parameters, compute the scheme's one-step joint characteristic function of (log S_{t+delta}-log S_t, U_hat) implied by (3.2) and the IG law in (0.4)-(0.6): Phi_hat(u,w) = E[exp(u log(S_{t+delta}/S_t) + w U_hat)] = E[exp(((w - u/2 + (1-rho^2)u^2/2) + u*rho/sigma) U_hat - u*rho*alpha/sigma)] with U_hat ~ IG(alpha, alpha^2/sigma^2). Compare Phi_hat(i y, -s) pointwise with the exact characteristic function exp(phi(delta) + psi(delta) v) over a grid of y in [-20,20] and s in [0,5]. If the log-characteristic functions differ by more than 5% at any grid point, the right-endpoint/point-mass approximation is not benign for Heston pricing; if they match to O(delta^2), the missing error analysis is not a practical obstacle.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy claim depends on the unproved replacement of the exact representation of U in (1.2) by the right-endpoint approximation leading to (1.4). Concretely, Algorithm 1 sets Z_hat_i,i+1 = (U_hat_i,i+1 - alpha_i)/sigma_i, so conditional on U_hat the increment Z_hat is a point mass. For the true process, Z_s,t = W_tilde_{U_s,t} by the Dubins-Schwarz theorem, and conditional on U_s,t the increment Z_s,t is not degenerate. The omitted dispersion of Z|U is exactly the part of the joint law that the Heston update (3.2) uses when rho is nonzero. Proposition 1.4 checks only first conditional moments, and Proposition 1.5 compares only the marginal Laplace transform of U, not the joint (U,Z) law used in (3.2). The paper provides no bound on the local weak error of this step, and the convergence guarantee is outsourced to Abi Jaber and Attal (2025), so the current manuscript does not itself establish the central accuracy claim. If the joint-law error is large in some regime, e.g. |rho| near 1 or short maturities, the reported option-price accuracy need not persist.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a new simulation scheme, called iVi, for the square-root process and the Heston model. The scheme first samples the integrated variance increment U from an Inverse Gaussian distribution whose parameters are obtained by a right-endpoint Euler discretization of the integrated process dynamics, then defines the Brownian-integral increment Z deterministically from U, and finally updates V through the exact drift relation. The main theoretical results are Theorem 1.3 (nonnegativity of the simulated V), Proposition 1.4 (exact first conditional moments for V, U, and Z), Proposition 1.5 and Remark 1.6 (the Inverse Gaussian law arises from an implicit Euler discretization of the Riccati equations), and Remark 1.7 (recovery of known Inverse Gaussian limiting laws in high-mean-reversion and long-maturity regimes). Numerical experiments compare the scheme to the QE scheme of Andersen and the Alfonsi scheme, for variance swaps, volatility swaps, Laplace transforms of integrated variance, and Heston call option prices and implied volatilities, across three calibrated parameter sets. The paper reports high accuracy with as few as one to five time steps. Section 4 states extensions to Volterra/rough Heston settings and refers to a companion paper for convergence proofs.","tokens_in":17002,"tokens_out":12400,"duration_ms":113792,"significance":"If the numerical findings are representative, the iVi scheme is a practically attractive addition to the simulation toolbox: it is parameter-free, simple to implement, preserves nonnegativity, matches first conditional moments exactly, and is competitive with, or better than, established schemes in the tested regimes, especially for short-maturity high-mean-reversion calibrations. The nonnegativity and moment-matching proofs are clean and easily verifiable, and the numerical study is extensive and reproducible in structure. The main limitation is that the accuracy that makes the scheme useful is not established theoretically in this manuscript; the right-endpoint discretization is heuristic, and convergence is deferred to a companion paper. The paper is therefore best read as a numerical-method proposal with strong empirical support rather than as a fully analyzed scheme.","major_comments":[{"comment":"The defining approximation of the scheme is the replacement of the exact integrated dynamics (1.2) by the right-endpoint rule (1.3), which collapses the path of Z_{t_i,·} to its endpoint. There is no local weak-error estimate in the manuscript for this step, and Section 4 explicitly outsources convergence to Abi Jaber and Attal (2025). Because the accuracy claims in Sections 2 and 3 rest on this approximation, the manuscript does not, on its own, prove the scheme's accuracy. Please add a local error bound for the right-endpoint replacement (or state precisely what remains to be proved), or downgrade the claim to an empirical/numerical one.","section":"Section 1.1, Eqs. (1.2)-(1.4)"},{"comment":"The Heston update (3.2) uses the joint law of (U_{i,i+1}, Z_{i,i+1}) through the term ρ Z_{i,i+1}. Proposition 1.4 only matches first conditional moments, and Proposition 1.5 compares only the marginal Laplace transform of U with the z-argument set to zero. Algorithm 1 makes Ẑ conditionally degenerate given Û, whereas the true Z_{i,i+1} conditional on the single value U_{i,i+1}=u is generally not degenerate, because the path of V inside the interval is not determined by the one integrated value. The numerical tests include ρ = -0.68, -0.70, and -0.9, but the paper gives no bound on the joint weak error, so the reported price accuracy is not guaranteed to persist for, e.g., |ρ| close to 1 or very short maturities. Please analyze the joint-law error or add a systematic stress test over ρ and maturity.","section":"Section 3, Eq. (3.2), Propositions 1.4-1.5"},{"comment":"The abstract states that the scheme 'yields the exact limiting Inverse Gaussian distributions' in two regimes. The manuscript does not prove these limiting results; it cites Mechkov (2015), Abi Jaber and De Carvalho (2024), McCrickerd (2019), and Forde and Jacquier (2011). Since the one-step distribution is Inverse Gaussian by construction, the matching is a consequence of known limit theorems. Please either give self-contained proofs of the limits (including the parameters of the limiting laws, which are not stated for the long-maturity case) or reformulate the abstract and Remark 1.7 as 'recovers/reproduces' the known limiting distributions.","section":"Abstract and Remark 1.7"}],"minor_comments":[{"comment":"For α_i=0, the convention IG(0,0)=0 is stated, but the update (0.5) still divides by σ_i; specify the convention for the boundary case, for example when c=0.","section":"Section 1.1 and Algorithm 1"},{"comment":"The proposition says that ψ̂ is 'a root' of the quadratic, but only later in Remark 1.6 is it stated that the relevant root has non-positive real part; make this explicit in the proposition.","section":"Proposition 1.5 and Eq. (1.9)"},{"comment":"The statement that φ̂ corresponds to a trapezoidal discretization is argued only asymptotically; a one-line bound on the remainder would make the remark precise.","section":"Remark 1.6"},{"comment":"The independence of N_i from the random variables used in Algorithm 2 is stated in prose; it should be part of the algorithm specification for reproducibility.","section":"Section 3, Eq. (3.2)"},{"comment":"The reference to Duffie, Filipović, and Schachermayer (2003) appears twice in the same sentence; moreover, the branch of the square root in D(u,w) in Eq. (B.2) is not discussed, so the Albrecher et al. (2007) branch choice should be stated for numerical evaluation.","section":"Introduction and Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The numerical claims are plausible and the scheme is attractive, but the missing local error analysis for the right-endpoint step and the joint-law issue in the Heston update are load-bearing gaps. The author could resolve them by importing or proving the companion-paper convergence results, or by explicitly limiting the paper's claims to empirical validation. I do not see a fatal mathematical error, but the manuscript as it stands is not self-contained on its central accuracy claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely useful simulation scheme, and the core idea is new as far as I can tell. The iVi scheme samples the integrated variance first from an Inverse Gaussian distribution, then reconstructs the instantaneous variance and the Brownian-integral term by algebraic relations. That flips the usual order (V first, then U) and avoids endpoint conditioning entirely. The derivation connects a right-endpoint Euler discretization of the integrated dynamics to a first-passage-time problem for a drifted Brownian motion, which is elegant and gives a transparent reason for the IG choice. Nonnegativity and first conditional moment matching are proved cleanly (Theorem 1.3, Proposition 1.4); those proofs are correct. The numerical work is also solid: three parameter sets, all violating Feller, comparisons against QE (Andersen) and Alfonsi, reference values from Fourier, and the scheme is competitive or better for short maturities and high vol-of-vol. Figure 1 with one time-step per slice is striking.\n\nThe soft spots are real but not fatal. First, the central accuracy claim is not fully self-contained. The right-endpoint step is a heuristic, and no error analysis appears in this paper. The stress-test point is correct that the scheme's Z_hat is a point mass conditional on U_hat, while the true Z is not, and that omitted dispersion feeds directly into the Heston update through rho. The numerical tests include rho = -0.9 and short maturities, so the evidence covers a lot, but it is evidence rather than a guarantee. Second, the abstract's \"exact limiting Inverse Gaussian\" claim is asserted with references to Mechkov and to Forde-Jacquier, but the parameter matching is not proved here. Third, no code is shipped; Algorithm 2 gives the IG sampler, so reproducibility is okay but not turnkey.\n\nI would send this to peer review. The idea is new, simple, and likely to be used; the included proofs are correct; the missing convergence analysis is explicitly deferred to a companion paper (Abi Jaber and Attal 2025), which a referee can check. The referee should push for a precise statement and proof of the limiting result, and for a discussion of the Z|U degeneracy and its effect on the joint law, but those are addressable revisions, not reasons to desk-reject.","headline":"Simulating the integrated square-root process first via an Inverse Gaussian first-passage time is a fresh, simple idea with strong numerics; the paper is worth refereeing, but the 'exact limiting' and convergence claims need to be substantiated.","tokens_in":17489,"tokens_out":5545,"would_cite":true,"duration_ms":45252,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H10","60H35","65C05","91G20","91G60"],"pacs":[],"model":"deepseek-v4-flash","headline":"The iVi scheme simulates the integrated square-root process first, draws each increment from an Inverse Gaussian law, and then reconstructs the variance process, giving a simple, accurate, nonnegative Heston simulator.","keywords":["square-root process","Heston model","Inverse Gaussian distribution","integrated variance","simulation scheme","nonnegative preservation","Riccati equation","time-changed Brownian motion"],"falsifier":"Run the iVi scheme on a parameter set outside the tested cases, for example $b>0$ with large $c$ and small $a$, and compare the empirical Laplace transform of $U_{0,T}$, $\\mathbb{E}[e^{-uU_{0,T}}]$, with the exact value from the Heston characteristic function (B.1) over several maturities; if the relative error does not decay as the number of time steps grows to 1000, the right-endpoint approximation in (1.3) is the source and the Inverse Gaussian proxy is not valid in that regime.","tokens_in":16500,"feed_emoji":"📈","tokens_out":12161,"duration_ms":99098,"temperature":0.7,"pith_summary":"This paper argues that the difficulty of simulating square-root dynamics is removable: instead of discretizing the process $V$ itself, one should first simulate its time integral $U_{t_i,t_{i+1}}$ between grid points, draw each increment from an Inverse Gaussian distribution, then recover the Brownian increment and the next value of $V$. The resulting iVi scheme is simple—one Inverse Gaussian draw and a few arithmetic operations per step—and provably keeps simulated variance nonnegative. It matches the first conditional moments of $V$, $U$, and the Brownian-integrated increment $Z$ exactly, and it reproduces the Inverse Gaussian limiting laws that appear in fast-mean-reversion/high-volatility-of-volatility regimes and for long maturities. Numerical tests on market-calibrated Heston parameters show accurate prices for variance swaps, volatility swaps, and implied-volatility slices with very few time steps, often one.","feed_headline":"Simulate integrated variance first for accurate Heston pricing","feed_subtitle":"The iVi scheme keeps variance nonnegative, matches first moments, and prices Heston options accurately with few steps.","key_machinery":"The load-bearing object is the integrated process $U_{s,t}=\\int_s^t V_r\\,dr$ together with the time-changed Brownian representation $Z_{s,t}=W_{U_{s,t}}$ provided by the Dambis–Dubins–Schwarz theorem. This identity closes the right-endpoint discretization of the $U$-dynamics into the fixed-point equation $\\widehat{U}_{i,i+1}=\\alpha_i+\\sigma_i W_{\\widehat{U}_{i,i+1}}$, whose solution is a first-passage time of a drifted Brownian motion; the law of that passage time is Inverse Gaussian with parameters $(\\alpha_i,\\alpha_i^2/\\sigma_i^2)$. The same Inverse Gaussian law re-emerges when the Riccati equation for the conditional characteristic function of $U$ is discretized by the same right-endpoint rule, which is the paper's mathematical justification for the choice of sampling distribution.","core_discovery":"The central discovery is that the integrated square-root process $U_{s,t}=\\int_s^t V_r\\,dr$ carries enough structure to be simulated directly, and that its conditional sampling law can be read off from an Euler-style right-endpoint discretization of the integrated dynamics. Writing $Z_{s,t}=\\int_s^t\\sqrt{V_r}\\,dW_r$, the variation-of-constants form of the SDE gives $U_{t_i,t_{i+1}}\\approx \\alpha_i+\\sigma_i Z_{t_i,t_{i+1}}$, and because $Z$ is a Brownian motion run at the clock $U$, this is the fixed-point equation $\\widehat{U}=\\alpha_i+\\sigma_i W_{\\widehat{U}}$. A natural solution is the first passage time of a drifted Brownian motion to level $\\alpha_i$, whose law is Inverse Gaussian with mean $\\alpha_i$ and shape $(\\alpha_i/\\sigma_i)^2$. The scheme draws $\\widehat{U}$ from that law, sets $\\widehat{Z}=(\\widehat{U}-\\alpha_i)/\\sigma_i$, and updates $\\widehat{V}$ from the original SDE; the paper proves that $\\widehat{V}$ stays nonnegative and that first conditional moments are exact, and reports convergence for the scheme in a companion work.","pith_inferences":["Because the scheme samples $U$ and $Z$ jointly from a first-passage law, it is naturally suited to payoffs on the whole integrated-variance trajectory, such as Asian-style vol products, rather than only terminal values.","The same 'simulate the integrated quantity first' idea may extend to other affine processes whose integrated component admits a tractable hitting-time representation; the paper notes that multivariate Wishart-type extensions are not straightforward.","A natural testable extension is to derive an explicit error bound for the right-endpoint approximation as a function of step size and parameters; such a bound would determine the regimes where the one-step version is reliable.","For local-stochastic-volatility or rough-Heston settings where the Heston component is simulated alongside an outer model, the scheme's low cost per path and few-step accuracy make it a candidate building block."],"forward_implications":["The simulated variance process $\\widehat{V}$ never goes negative for any real $b$ and any nonnegative $a$, so the scheme needs no reflection or absorption fix.","First conditional moments of $V$, $U$, and $Z$ are matched exactly at every step, which makes variance-swap prices nearly unbiased even with one time step.","On calibrated fast-mean-reversion/high-volatility-of-volatility parameters, the one-step scheme already reproduces the correct Inverse Gaussian limiting law of $U$, and the reported implied-volatility slices stay accurate.","For Heston option pricing, each step adds only one independent standard Gaussian, and the experiments show accurate ITM/ATM/OTM option prices with 1–5 time steps.","The scheme has no switching thresholds or stored grids, so it is easy to code and appeared stable in low-sample-path tests."],"supporting_citations":[{"why":"supplies the Dambis–Dubins–Schwarz theorem that turns $Z$ into a Brownian motion run at the clock $U$, closing the fixed-point equation (1.4).","marker":"Revuz and Yor (2013)"},{"why":"provides the acceptance-rejection sampler (Algorithm 2) used to draw the Inverse Gaussian increments.","marker":"Michael, Schucany, and Haas (1976)"},{"why":"defines the model and the affine characteristic function used for reference option prices and for the Riccati equations in Appendix B.","marker":"Heston (1993)"},{"why":"gives the affine-process framework under which the exponential-affine characteristic function (B.1) and Riccati equations hold.","marker":"Duffie, Filipović, and Schachermayer (2003)"},{"why":"identifies the fast-reversion limit in which the integrated square-root process converges to an Inverse Gaussian law, the regime where the one-step scheme is exact.","marker":"Mechkov (2015)"},{"why":"establishes the large-maturity Inverse Gaussian limit of the integrated variance, another regime where one step is sufficient.","marker":"Forde and Jacquier (2011)"},{"why":"provides the Quadratic-Exponential benchmark against which the iVi scheme's accuracy is compared.","marker":"Andersen (2008)"},{"why":"provides a second-order benchmark scheme used in the numerical comparisons.","marker":"Alfonsi (2010)"}],"fun_headline_variants":["Simulate integrated variance first for Heston pricing","Inverse Gaussian trick speeds square-root simulation","Few-step scheme for accurate Heston option pricing","Exact moments from integrated variance simulation","New scheme makes square-root simulation simple"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme hinges on replacing an integral in the dynamics of $U$ by its right-endpoint value; if that approximation is poor in some parameter regime, the Inverse Gaussian law used at each step becomes a biased proxy for the true conditional distribution, and the error will feed into the Heston update.","fun_headline_variants_meta":{"raw":{"variants":["Simulate integrated variance first for Heston pricing","Inverse Gaussian trick speeds square-root simulation","Few-step scheme for accurate Heston option pricing","Exact moments from integrated variance simulation","New scheme makes square-root simulation simple"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1281,"prompt_tokens":911,"completion_tokens":370,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":304}},"tokens_in":527,"tokens_out":370,"duration_ms":4146,"temperature":1.0,"reasoning_tokens":304,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:07:19.057940+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the iVi scheme on a parameter set outside the tested cases, for example $b>0$ with large $c$ and small $a$, and compare the empirical Laplace transform of $U_{0,T}$, $\\mathbb{E}[e^{-uU_{0,T}}]$, with the exact value from the Heston characteristic function (B.1) over several maturities; if the relative error does not decay as the number of time steps grows to 1000, the right-endpoint approximation in (1.3) is the source and the Inverse Gaussian proxy is not valid in that regime.","supporting_citations":[{"cited_title":"Continuous martingales and B rownian motion , volume 293","cited_arxiv_id":null,"evidence_quote":"supplies the Dambis–Dubins–Schwarz theorem that turns $Z$ into a Brownian motion run at the clock $U$, closing the fixed-point equation (1.4)."},{"cited_title":"Generating random variates using transformations with multiple roots","cited_arxiv_id":null,"evidence_quote":"provides the acceptance-rejection sampler (Algorithm 2) used to draw the Inverse Gaussian increments."},{"cited_title":"Affine processes and applications in finance","cited_arxiv_id":null,"evidence_quote":"gives the affine-process framework under which the exponential-affine characteristic function (B.1) and Riccati equations hold."},{"cited_title":"Fast-reversion limit of the H eston model","cited_arxiv_id":null,"evidence_quote":"identifies the fast-reversion limit in which the integrated square-root process converges to an Inverse Gaussian law, the regime where the one-step scheme is exact."},{"cited_title":"The large-maturity smile for the H eston model","cited_arxiv_id":null,"evidence_quote":"establishes the large-maturity Inverse Gaussian limit of the integrated variance, another regime where one step is sufficient."},{"cited_title":"Simple and efficient simulation of the H eston stochastic volatility model","cited_arxiv_id":null,"evidence_quote":"provides the Quadratic-Exponential benchmark against which the iVi scheme's accuracy is compared."},{"cited_title":"High order discretization schemes for the CIR process: application to affine term structure and H eston models","cited_arxiv_id":null,"evidence_quote":"provides a second-order benchmark scheme used in the numerical comparisons."}],"review_version":1}