Does the Euler-Lagrange condition for variance-optimal Malliavin Monte-Carlo weights force the minimizer to be a function of the payoff variable?

Malliavin integration by parts represents the sensitivities ('Greeks') of option prices as Monte-Carlo expectations of the payoff times a random weight, and the weight is highly non-unique. Fournié, Lasry, Lebuchoux and Lions (2001) posed the variational problem of finding the variance-minimizing weight within the class of Malliavin-constructible (Skorohod-integral) weights and derived an Euler-Lagrange optimality condition which formally implies that the Malliavin derivative of the minimizer is collinear with that of the payoff variable: $D_t\bar{\pi}=A(\omega)D_tF$ for all $t$. Their Appendix B proves, only in particular cases (for example when the conditioning variable is $W_1$), that such collinearity of Malliavin derivatives implies functional dependence, i.e. that the random variable is a deterministic measurable function of the payoff variable. The open problem is to settle the general implication: prove that, under general nondegeneracy and smoothness conditions on $F$ (including the multidimensional case), the variance-minimizing Malliavin weight $\bar{\pi}$, whose derivative process is collinear with $D F$, is $\sigma(F)$-measurable, or construct an explicit counterexample showing that a variance-minimizing Malliavin weight need not be a function of the payoff variable.

Background

Monte-Carlo pricing and the Greeks. In mathematical finance, the price of a European-style claim on a diffusion asset process is an expectation $\mathbf{E}[\Phi(F)]$, where $F$ is a functional of the driving Brownian motion $W$ on a time interval $[0,T]$ (for example the terminal asset value $S_T$, or the time average that enters an Asian option) and $\Phi$ is the payoff. The price sensitivities to model inputs, called the 'Greeks' (for instance the delta, the derivative of the price with respect to the initial asset value), must be computed numerically and are the basis of hedging. A finite-difference Monte-Carlo estimate of a sensitivity, obtained by perturbing the input and re-simulating, has high variance and behaves badly for discontinuous payoffs such as digital options.

Malliavin calculus and integration by parts. Malliavin calculus, also called the stochastic calculus of variations, is a differential calculus on the Wiener space of Brownian paths. Its two basic operators are the Malliavin derivative $D_tF$, which for a random variable $F$ measurable with respect to $W$ is a process measuring the sensitivity of $F$ to an infinitesimal perturbation of the Brownian path at time $t$, and the Skorohod integral $\delta(u)$, the adjoint of the derivative, which extends the classical Ito integral to integrands $u$ that may anticipate the future of the Brownian motion. The integration-by-parts formula transfers a derivative of the payoff onto a random weight, so that a sensitivity can be written as a single expectation $\mathbf{E}[\Phi(F)\pi]$ with an explicitly constructible weight $\pi=\delta(u)$; the choice of the auxiliary process $u$ is highly non-unique. Fournié, Lasry, Lebuchoux, Lions and Touzi (1999) used this to compute Greeks for path-dependent and discontinuous payoffs with far lower variance than finite differences.

Weights, variance, and the variational problem. All admissible weights give the same expectation but have different variances, so Monte-Carlo efficiency is governed by weight selection. Since the variance of $\pi$ satisfies $\operatorname{Var}(\pi)\ge\operatorname{Var}(\mathbf{E}[\pi\mid\sigma(F)])$, the smallest variance over all admissible weights is attained by a weight that is a deterministic function of $F$ (the conditional score, the likelihood-ratio type weight, when the law of the perturbed functional is known); but that density is typically unknown and this optimum is not computable. Part II of the series, Fournié, Lasry, Lebuchoux and Lions (2001), therefore posed the variational problem within the class of Malliavin-constructible weights $\pi=\delta(u)$: characterize the variance-minimizing weight and approach it by iterative (gradient-type) schemes. Optimizing the estimator variance over this admissible Malliavin class yields an Euler-Lagrange first-order optimality condition, from which the authors derive formally the pointwise collinearity relation

$D_t\bar{\pi}=A(\omega)\,D_tF,\qquad 0\le t\le T,$

for a minimizer $\bar{\pi}$, with a random coefficient $A$ that does not depend on $t$.

Functional dependence and the $C(F,G)$ criterion. A random variable is called $\sigma(F)$-measurable if it is a deterministic measurable function of $F$, written $G=f(F)$; such weights depend on the simulated path only through the payoff variable and are the Malliavin analogue of likelihood-ratio weights. The standard structural assumption on $F$ is nondegeneracy: the Malliavin covariance matrix $\int_0^T D_tF\,D_tF^\top\,dt$ is invertible almost surely and its inverse is integrable, $\mathbf{E}\big[(\det\int_0^T D_tF\,D_tF^\top\,dt)^{-p}\big]<\infty$ for all $p$, which implies that $F$ has a density and is the usual hypothesis under which the integration-by-parts machinery applies. Appendix B of the 2001 paper introduces a Malliavin-derivative-based test of functional dependence, the quantity $C(F,G)$ defined as the essential supremum over outcomes of the squared cosine of the angle between the derivative processes $DF$ and $DG$ in $L^2([0,T])$ (the ratio is taken to be $1$ where one of the two norms vanishes identically). One always has $C(F,G)=1$ when $F$ is a function of $G$, and $C$ is invariant under Lipschitz reparametrizations; the converse implication, collinearity implying functional dependence, is established only in particular cases, for example when the conditioning variable is $W_1$ (the Brownian motion at the terminal time). Whether the Euler-Lagrange condition forces the minimizer $\bar{\pi}$ to be $\sigma(F)$-measurable in general, and hence whether variance-optimal Malliavin weights must be functions of the payoff variable, is the unresolved question.

Neighboring later work. Benhamou (2003) re-expressed the Malliavin weights of the 1999 paper in terms of their Skorohod integrands, showed they must satisfy necessary and sufficient conditions given as conditional expectations, and derived the weight with the smallest total variance, identifying it with the likelihood-ratio weight; this settles optimal-weight selection inside parametric, density-based settings but not the general structural question. Kohatsu-Higa and Pettersson (2002) built a general error-analysis and variance-reduction framework for simulating densities on Wiener space. Bouchard, Ekeland and Touzi (2004) solved a companion variance-reduction problem, the optimal choice of localizing functions in Malliavin representations of conditional expectations of multidimensional Markov diffusions, proving existence and uniqueness of an optimal localizer in a Sobolev class with a PDE characterization. None of these establishes, or refutes, the general functional-dependence implication above.

Problem Statement

Let $W$ be an $N$-dimensional Brownian motion on $[0,T]$, and let $F$ be an $m$-dimensional Malliavin-differentiable functional of $W$ satisfying the usual $C^\infty$ (infinitely Malliavin-differentiable) smoothness and nondegeneracy assumptions, imposed on $F$ alone: the Malliavin covariance matrix $\int_0^T D_tF\,D_tF^\top\,dt$ is invertible almost surely and its inverse is integrable, $\mathbf{E}\big[(\det\int_0^T D_tF\,D_tF^\top\,dt)^{-p}\big]<\infty$ for all $p$ (so $F$ admits a density). In the Monte-Carlo sensitivity framework of Fournié, Lasry, Lebuchoux and Lions, a sensitivity $\mathbf{E}[\Phi'(F)G]$ is represented as $\mathbf{E}[\Phi(F)\pi]$ over the class of admissible Malliavin weights $\pi=\delta(u)$ (Skorohod integrals of Malliavin-differentiable processes $u$), and the estimator variance is minimized over this class. The Euler-Lagrange optimality condition derived in the source paper formally implies, for a minimizer $\bar{\pi}$, the pointwise almost-sure collinearity relation

$D_t\bar{\pi}=A(\omega)\,D_tF,\qquad 0\le t\le T,$

for some random variable $A$ not depending on $t$.

The open problem is to settle, in full generality, whether this collinearity relation implies that $\bar{\pi}$ is $\sigma(F)$-measurable, that is, $\bar{\pi}=f(F)$ for a deterministic measurable function $f$. In the notation of the source paper's Appendix B this is the statement that the Euler-Lagrange minimizer satisfies the collinearity criterion $C(\bar{\pi},F)=1$ — with $C$ defined as the essential supremum of the squared cosine of the angle between $D\bar{\pi}$ and $DF$ in $L^2([0,T])$ — and the open question is whether that criterion forces $\bar{\pi}$ to be a deterministic measurable function of $F$.

The question is whether the collinearity relation forces $\bar{\pi}$ to be $\sigma(F)$-measurable under general nondegeneracy and smoothness conditions on $F$, including the multidimensional case $m>1$ where collinearity becomes a span condition on the derivative matrix $DF$; or whether a nondegenerate $C^\infty$ payoff $F$ admits an Euler-Lagrange variance-minimizing Malliavin weight $\bar{\pi}$ over the admissible Skorokhod class that is not $\sigma(F)$-measurable.

References

1. E. Fournié, J.-M. Lasry, J. Lebuchoux, P.-L. Lions, Applications of Malliavin calculus to Monte-Carlo methods in finance. II, Finance and Stochastics 5 (2001) 201-236, DOI 10.1007/PL00013529, https://doi.org/10.1007/PL00013529
2. E. Fournié, J.-M. Lasry, J. Lebuchoux, P.-L. Lions, N. Touzi, Applications of Malliavin calculus to Monte Carlo methods in finance, Finance and Stochastics 3 (1999) 391-412, DOI 10.1007/s007800050068, https://doi.org/10.1007/s007800050068
3. E. Benhamou, Optimal Malliavin Weighting Function for the Computation of the Greeks, Mathematical Finance 13 (2003) 37-53, DOI 10.1111/1467-9965.t01-1-00004, https://doi.org/10.1111/1467-9965.t01-1-00004
4. A. Kohatsu-Higa, R. Pettersson, Variance Reduction Methods for Simulation of Densities on Wiener Space, SIAM Journal on Numerical Analysis 40 (2002) 431-450, DOI 10.1137/S0036142901385507, https://doi.org/10.1137/S0036142901385507
5. H.-P. Bermin, A. Kohatsu-Higa, M. Montero, Local Vega Index and Variance Reduction Methods, Mathematical Finance 13 (2003) 85-97, DOI 10.1111/1467-9965.00007, https://doi.org/10.1111/1467-9965.00007
6. E. Gobet, A. Kohatsu-Higa, Computation of Greeks for Barrier and Lookback Options Using Malliavin Calculus, Electronic Communications in Probability 8 (2003) 51-62, DOI 10.1214/ECP.V8-1069, https://doi.org/10.1214/ECP.V8-1069
7. G. Bernis, E. Gobet, A. Kohatsu-Higa, Monte Carlo Evaluation of Greeks for Multidimensional Barrier and Lookback Options, Mathematical Finance 13 (2003) 99-113, DOI 10.1111/1467-9965.00008, https://doi.org/10.1111/1467-9965.00008
8. B. Bouchard, I. Ekeland, N. Touzi, On the Malliavin approach to Monte Carlo approximation of conditional expectations, Finance and Stochastics 8 (2004) 45-71, DOI 10.1007/s00780-003-0109-0, https://doi.org/10.1007/s00780-003-0109-0
9. V. Bally, L. Caramellino, A. Zanette, Pricing and hedging American options by Monte Carlo methods using a Malliavin calculus approach, Monte Carlo Methods and Applications 11 (2005) 97-133, DOI 10.1515/156939605777585944, https://doi.org/10.1515/156939605777585944
10. Y. El-Khatib, N. Privault, Computations of Greeks in a market with jumps via the Malliavin calculus, Finance and Stochastics 8 (2004) 161-179, DOI 10.1007/s00780-003-0111-6, https://doi.org/10.1007/s00780-003-0111-6
11. G. Di Nunno, T. Meyer-Brandis, B. Øksendal, F. Proske, Malliavin calculus and anticipative Itô formulae for Lévy processes, Infinite Dimensional Analysis, Quantum Probability and Related Topics 8 (2005) 235-258, DOI 10.1142/S0219025705001950, https://doi.org/10.1142/S0219025705001950
12. M. Mrad, N. Touzi, A. Zeghal, Monte Carlo Estimation of a Joint Density Using Malliavin Calculus, and Application to American Options, Computational Economics 27 (2006) 497-531, DOI 10.1007/S10614-005-9005-3, https://doi.org/10.1007/S10614-005-9005-3
13. F. J. Mhlanga, Malliavin Calculus and the Optimal Weighting Function in a Pure Jump Lévy Setting, International Journal of Engineering Research in Africa 55 (2021) 66-81, DOI 10.4028/www.scientific.net/JERA.55.66, https://doi.org/10.4028/www.scientific.net/JERA.55.66