Abstract. The deterministic numerical valuation of American options under Heston's stochastic volatility model is considered. The prices are given by a linear complementarity problem with a twodimensional parabolic partial differential operator. A new truncation of the domain is described for small asset values while for large asset values and variance a standard truncation is used. The finite difference discretization is constructed by numerically solving quadratic optimization problem aiming to minimize the truncation error at each grid point. A Lagrange approach is used to treat the linear complementarity problems. Numerical examples demonstrate the accuracy and effectiveness of the proposed approach.Key words. American option pricing, stochastic volatility model, linear complementarity problem, finite difference method, quadratic programming, multigrid method, Lagrange method, penalty method AMS subject classifications. 35K85, 65M06, 65M55, 65Y20, 91B281. Introduction. The seminal papers [2] and [28] by Black & Scholes, and Merton, respectively, laid the foundations of the modern theory of pricing financial options. Since then vast body of scientific work has been devoted to the development of methods for pricing options. Particularly American options are challenging to evaluate due to their early exercise possibility and various approaches to approximate their price have been proposed. The paper [4] by Brennan and Schwartz was one of first ones to formulate a linear complementarity problem (LCP) and then to solve it using a finite difference discretization. Empirical evidence has shown that the assumption on log-normal behavior of the value of the underlying asset is not realistic for many asset categories including shares of companies. To alleviate this, many generalizations have been introduced for the log-normal model. We consider stochastic volatility models which assume that the volatility of the value of the asset follows a stochastic process; see [11] and references therein. Particularly, our model problem for American options is based on Heston's model [14], but the techniques considered in this paper can be generalized also for other stochastic volatility models like the Hull-White model [17] and the Stein-Stein model [33].Based on Heston's stochastic volatility model, a LCP with a parabolic partial differential operator can derived for the price of American options with the value of underlying asset and its variance being the spatial variables. Near the axes, the first-order derivatives dominate the second-order derivatives in the operator. It also includes the second-order cross derivative. Due to these properties, it is not easy to construct an accurate and stable discretization. In financial literature, most often finite differences are used for the discretization while sometimes also finite elements are used; see [1], [34], [38], for example. It is desirable to use discretizations leading to matrices with the M-matrix property [39]; see also [41]. This property guarantees the stability of the ...