Variance-optimal Malliavin weights and localization for Monte Carlo Greeks in general diffusion models

The Malliavin integration-by-parts method of Fournié–Lasry–Lebuchoux–Lions–Touzi (1999) expresses the Greeks (price sensitivities) of general diffusion models as Monte Carlo expectations of the payoff multiplied by a computable random weight. The weight - parameterized by an auxiliary function $a\in\Gamma_m$, more generally by a Skorohod-integral direction, and combined for nonsmooth payoffs with a localization of bandwidth $\delta$ - is highly non-unique, and different admissible choices yield unbiased estimators with widely different variances. The question the 1999 paper deferred and its 2001 sequel sharpened is: characterize the variance-minimizing admissible weight and localization in full generality (in particular, whether the variance-minimizing weight over the Skorohod-integral-representable weights is necessarily a deterministic function of the payoff variables, i.e. the likelihood-ratio score weight, which is itself uncomputable when the payoff law is unknown); construct computable, payoff-independent weight-selection rules, together with computable localization-selection rules, with theoretical variance guarantees for general multidimensional and path-dependent models, including higher-order Greeks; and derive the associated complexity-accuracy trade-offs. Later literature resolved special cases (European Greeks in Black-Scholes-type models; score-dominance when densities exist, proved by Benhamou (2003); model-specific variational optima; asymptotic optimality rates), but no general characterization or guaranteed computable procedure exists, leaving a precise open core.

Background

The Greeks are the sensitivities of an option price - the expectation $u(x)=\mathbf{E}^x[\phi(X(\cdot))]$ of a payoff functional $\phi$ of a diffusion process $X$ - with respect to inputs such as the initial state (Delta), volatility (Vega), or second-order combinations (Gamma). In Monte Carlo pricing these sensitivities must themselves be estimated by simulation. Finite-difference bump estimators suffer a bias-variance trade-off in the perturbation size and degrade sharply for discontinuous payoffs such as digitals (binary options paying a fixed amount on one side of a strike level); pathwise differentiation requires smooth payoffs; and the likelihood-ratio (score) method requires knowing the joint density of the state variables, which is unavailable in general models.

Malliavin calculus - the stochastic calculus of variations on Wiener space - provides an alternative. Its integration-by-parts formula moves the derivative off the (possibly discontinuous) payoff onto a random weight built from the Malliavin derivative and the Skorohod integral (the divergence operator, an extension of the Ito integral to non-adapted integrands). Fournié, Lasry, Lebuchoux, Lions and Touzi (1999) showed that for an $n$-dimensional diffusion $X$ with smooth, uniformly elliptic coefficients (i.e. $\xi^{*}\sigma^{*}(x)\sigma(x)\xi\ge\varepsilon|\xi|^{2}$ for all $x,\xi$), with first-variation (Jacobian) process $Y(t)=\partial X(t)/\partial x$, and for any payoff $\phi\in L^{2}$ depending on the path at $m$ monitoring dates, each Greek admits the representation, for the gradient,

$$\nabla u(x)=\mathbf{E}^{x}\Big[\phi\big(X(t_1),\dots,X(t_m)\big)\int_0^T a(t)\big[\sigma^{-1}(X(t))Y(t)\big]^{*}\,dW(t)\Big],$$

for any auxiliary function $a$ in $\Gamma_m=\{a\in L^{2}([0,T]):\int_0^{t_i}a(t)\,dt=1,\ i=1,\dots,m\}$. The weight is an Ito integral computable from simulated paths, is independent of the payoff, and gives the standard $n^{-1/2}$ Monte Carlo convergence even for discontinuous payoffs; in the Black-Scholes model with the simple choice $a(t)=1/T$ the Delta weight reduces to $W_T/(x\sigma T)$.

The weight, however, is far from unique: every admissible $a\in\Gamma_m$ - and, more generally, every Skorohod-integral weight satisfying the integration-by-parts constraint - yields an unbiased estimator, but with different variance. Because the global weights are polynomials in $W_T$ that are large exactly where nonsmooth payoffs are singular, the 1999 paper also introduced a localization device: decomposing, for example, the call payoff $(S_T-K)_+$ into a smoothed part and a compactly supported remainder controlled by a bandwidth $\delta>0$, and applying integration-by-parts only to the remainder, which markedly reduces variance in the reported experiments. The 1999 paper deferred the optimality question - the optimal choices of $a\in\Gamma_m$ and of the localization parameter $\delta$, indicating that optimality of a particular weight would be proved in a forthcoming paper (realized, for the European formulas, by the 2001 sequel). Work on this deferred question concerns: the systematic design of variance-minimizing weights for general multidimensional and path-dependent models; high-dimensional scalability and complexity-accuracy analysis; and stability/variance analysis of higher-order Greek estimators obtained by iterated integration by parts, beyond formal expressions such as the Gamma weight $\frac{1}{x^{2}\sigma T}\big(\frac{W_T^{2}}{\sigma T}-W_T-\frac{1}{\sigma}\big)$.

The authors' sequel, Fournié, Lasry, Lebuchoux and Lions (2001), states that it answers the question of the optimal weight functional in the sense of minimal variance for the European formulas of Part I. It introduces the admissible class $\mathscr{W}=\{\pi\in L^{1}(\Omega):\mathbf{E}[\pi\mid\sigma(F)]=\pi_0\}$, where $F$ collects the payoff variables and $\pi_0=\partial_\lambda\log f_\lambda(F)|_{\lambda=0}$ is the score (likelihood-ratio) weight of the perturbed law - exactly the class of weights that are unbiased for all payoffs simultaneously - derives Euler-Lagrange first-order optimality conditions for minimizing the estimator variance over Skorohod-integral weights, and settles special cases. Two facts frame the frontier. First, when the density of $F$ is known, the score $\pi_0$ is $\sigma(F)$-measurable (a deterministic function of the payoff variables) and dominates every other admissible weight: replacing a weight by its conditional expectation given $F$ preserves unbiasedness and can only reduce variance. This projection principle was proved by Benhamou (2003), who derived necessary and sufficient conditional-expectation conditions on the Skorohod integrand and identified the likelihood-ratio weight as the weighting function with the smallest total variance; Chen and Glasserman (2007) reached the same correspondence from the likelihood-ratio side, deriving such weights without Malliavin calculus; and Detemple, Garcia and Rindisbacher (2005) established the complementary asymptotic fact that only Malliavin-based estimators attain the optimal convergence rate for discontinuous payoffs. Second, precisely this optimal weight is uncomputable in general because the law of $F$ is unknown - the very reason Malliavin weights are used. Over the full class $\mathscr{W}$ the projection principle therefore already identifies the variance-minimizing weight: every admissible weight can be replaced by its conditional expectation given $F$, namely $\pi_0$ itself, without increasing the variance, so whenever $\pi_0$ is admissible the minimum over $\mathscr{W}$ is attained at the $\sigma(F)$-measurable score. The unresolved structural question concerns the narrower class of Skorohod-integral-representable weights - the class over which the 2001 sequel's Euler-Lagrange conditions are derived - whose closure under conditioning on $F$ is not established, and in which membership of $\pi_0$ is not assumed: whether the variance-minimizing weight in that class is necessarily $\sigma(F)$-measurable under general nondegeneracy and smoothness conditions remains open, and no practical optimization algorithm (convergence, discretization, regularization) for variance-optimal weights has been developed.

Problem Statement

For a general multidimensional diffusion with smooth, uniformly elliptic coefficients, and a payoff functional $\phi\in L^{2}$ (possibly discontinuous and path-dependent, observed at finitely or infinitely many dates), the Malliavin integration-by-parts representation expresses each Greek as $\mathbf{E}[\phi\,\pi]$ with a highly non-unique admissible weight $\pi$ - in the 1999 parameterization $\pi=\int_0^T a(t)[\sigma^{-1}(X(t))Y(t)]^{*}\,dW(t)$ with $a\in\Gamma_m$ and $Y$ the first-variation (Jacobian) process of the diffusion; in the 2001 formulation $\pi$ ranges over Skorohod-integral weights with $\mathbf{E}[\pi\mid\sigma(F)]=\pi_0$, where $F$ denotes the payoff variables, $\sigma(F)$ the sigma-field they generate, and $\pi_0$ the score (log-derivative of the perturbed law of $F$) - and, for nonsmooth payoffs, combined with a localized payoff decomposition of bandwidth $\delta$. (Notation: the scalar $\delta>0$ is the localization bandwidth, while $\delta(u)$ below always denotes the Skorohod divergence integral of the process $u$.) Complete the variance theory of these estimators in full generality:

(i) Characterize the variance-minimizing admissible weight and localization in general. The admissible localization class is the localization family of the cited Fournié et al. (1999, 2001) papers: the exact decompositions $\phi=\phi_\delta+r_\delta$ of the payoff into a smooth part $\phi_\delta$ and a compactly supported remainder $r_\delta=\phi-\phi_\delta$ supported in a $\delta$-neighbourhood of the singular set of $\phi$, parameterized by the bandwidth $\delta>0$ and the localization function; the weight and the localization are optimized jointly over this class. Over the full conditional-expectation class $\mathscr{W}$, the projection principle recalled in the Background identifies the score $\pi_0$ as the variance-minimizing weight, so the open structural question concerns the subclass of Skorohod-integral-representable admissible weights (the class over which the 2001 sequel's Euler-Lagrange conditions are derived), whose closure under conditioning on $F$ is not established; whether $\pi_0$ belongs to this subclass is itself part of what this part must decide. In particular, prove or disprove that the minimizer of the estimator variance over this Skorohod-representable admissible class is necessarily $\sigma(F)$-measurable (a deterministic function of the payoff variables, coinciding with the score weight when the law of $F$ is known) under general nondegeneracy and smoothness conditions, beyond the special cases resolved in the 2001 sequel.

(ii) Construct computable selection rules for the weight (the function $a$, or the process $u$ defining the Skorohod integral $\delta(u)$) that are payoff-independent, together with computable selection rules for the localization bandwidth $\delta$ - the latter may depend on the payoff's singularity structure (e.g. a localization function centered at the strike level) and may be chosen adaptively during the simulation - none of them requiring the law of $F$, for general multidimensional and path-dependent models, with theoretical guarantees - provable near-optimality or explicit variance bounds relative to the optimal weight - and extend the analysis to higher-order Greeks obtained by iterated integration by parts (stability and variance of the iterated-weight estimators).

(iii) Derive the corresponding complexity-versus-accuracy trade-offs for large-scale Monte Carlo implementations: the scaling of estimator variance and of the cost of weight evaluation with the state-space dimension and with the time-discretization mesh, including the control of the $1/\sqrt{\Delta t}$-type blow-up of discretized weights.

The common objective is a complete, constructive theory of minimal-variance Malliavin Monte Carlo estimation of Greeks: items (i)-(iii) are respectively the characterization, the computable construction, and the complexity analysis of one and the same variance-minimization objective. Parts (i)-(iii) are jointly posed and logically independent: a resolution of any one part leaves the others open. Within (ii), the payoff-independent weight-selection rule, the localization-selection rule, and their variance guarantees are required, and the extension to higher-order Greeks obtained by iterated integration by parts is likewise required.

References

1. E. Fournié, J.-M. Lasry, J. Lebuchoux, P.-L. Lions, N. Touzi (1999), Applications of Malliavin calculus to Monte Carlo methods in finance, Finance and Stochastics 3(4):391-412, DOI 10.1007/s007800050068, https://doi.org/10.1007/s007800050068
2. E. Fournié, J.-M. Lasry, J. Lebuchoux, P.-L. Lions (2001), Applications of Malliavin calculus to Monte-Carlo methods in finance. II, Finance and Stochastics 5(2):201-236, DOI 10.1007/pl00013529, https://doi.org/10.1007/pl00013529
3. E. Benhamou (2002), Smart Monte Carlo: various tricks using Malliavin calculus, Quantitative Finance 2(5):329-336, DOI 10.1088/1469-7688/2/5/301, https://doi.org/10.1088/1469-7688/2/5/301
4. E. Benhamou (2003), Optimal Malliavin Weighting Function for the Computation of the Greeks, Mathematical Finance 13(1):37-53, DOI 10.1111/1467-9965.t01-1-00004, https://doi.org/10.1111/1467-9965.t01-1-00004
5. N. Chen, P. Glasserman (2007), Malliavin Greeks without Malliavin calculus, Stochastic Processes and their Applications 117(11):1689-1723, DOI 10.1016/j.spa.2007.03.012, https://doi.org/10.1016/j.spa.2007.03.012
6. C. P. Fries (2007), Sensitivities (Partial Derivatives) of Monte Carlo Prices, chapter 17 of C. P. Fries, Mathematical Finance: Theory, Modeling, Implementation, Wiley, pp. 243-259, DOI 10.1002/9780470179789.ch17, https://doi.org/10.1002/9780470179789.ch17
7. A. Kohatsu-Higa, R. Pettersson (2002), Variance Reduction Methods for Simulation of Densities on Wiener Space, SIAM Journal on Numerical Analysis 40(2):431-450, DOI 10.1137/S0036142901385507, https://doi.org/10.1137/S0036142901385507
8. H.-P. Bermin, A. Kohatsu-Higa, M. Montero (2003), Local Vega Index and Variance Reduction Methods, Mathematical Finance 13(1):85-97, DOI 10.1111/1467-9965.00007, https://doi.org/10.1111/1467-9965.00007
9. J. Detemple, R. Garcia, M. Rindisbacher (2005), Asymptotic Properties of Monte Carlo Estimators of Derivatives, Management Science 51(11):1657-1675, DOI 10.1287/mnsc.1050.0398, https://doi.org/10.1287/mnsc.1050.0398
10. D. David, N. Privault (2009), Numerical computation of Theta in a jump-diffusion model by integration by parts, Quantitative Finance 9(6):727-735, DOI 10.1080/14697680902814191, https://doi.org/10.1080/14697680902814191
11. Y. Yolcu-Okur, T. Sayer, B. Yilmaz, B. A. Inkaya (2018), Computation of the Delta of European options under stochastic volatility models, Computational Management Science 15(2):213-237, DOI 10.1007/s10287-018-0316-y, https://doi.org/10.1007/s10287-018-0316-y
12. B. Yilmaz (2018), Computation of option greeks under hybrid stochastic volatility models via Malliavin calculus, Modern Stochastics: Theory and Applications 5(2):145-165, DOI 10.15559/18-VMSTA100, https://doi.org/10.15559/18-VMSTA100
13. F. J. Mhlanga (2021), Malliavin Calculus and the Optimal Weighting Function in a Pure Jump Lévy Setting, International Journal of Engineering Research in Africa 55:66-81, DOI 10.4028/www.scientific.net/JERA.55.66, https://doi.org/10.4028/www.scientific.net/JERA.55.66
14. M. Kharrat (2021), Pricing American Put Options Using Malliavin Calculus with Optimal Localization Function, 'Prof. Marin Drinov' Publishing House of the Bulgarian Academy of Sciences, DOI 10.7546/crabs.2021.10.04, https://doi.org/10.7546/crabs.2021.10.04
15. N. Cufaro Petroni, P. Sabino (2011), Multidimensional Quasi-Monte Carlo Malliavin Greeks, arXiv:1103.5722, https://arxiv.org/abs/1103.5722
16. S. Sojudi, M. Tahmasebi (2026), Conditional Expectation expression in mean-field SDEs and its applications, arXiv:2602.14479, https://arxiv.org/abs/2602.14479
17. V. Krishnamurthy, L. Snow (2026), Efficient Counterfactual Estimation of Conditional Greeks via Malliavin-based Weak Derivatives, arXiv:2602.02811, https://arxiv.org/abs/2602.02811
18. A. Kohatsu-Higa, M. Montero (2001), An application of Malliavin Calculus to Finance, arXiv:cond-mat/0111563, https://arxiv.org/abs/cond-mat/0111563