Multilevel Monte Carlo front-tracking for random scalar conservation laws (Q285286): Difference between revisions

Many problems in physics and engineering are modeled by systems of hyperbolic conservation or balance laws. The Cauchy problem for such systems has the form \[ U_t + \sum_{j=1}^d {\partial \over \partial x_j} (F_j(U)) = 0, \;\;x = (x_1, \dots, x_d) \in \mathbb{R}^d, \;\;t >0, \] \[ U(x,0) = U_0(x), \;\;x \in \mathbb{R}^d. \] Here \(U: \mathbb{R}^d \to \mathbb{R}^d\) is the vector of unknowns and \(F_j: \mathbb{R}^d \to \mathbb{R}^d\) is the flux vector for the \(j\)-th direction with \(m\) being a positive integer. It is supposed that the flux functions are Lipschitz continuous with respect to the state variable, for which there exists a unique random entropy solution. A convergence analysis of the multilevel Monte Carlo front-tracking algorithm is presented. An improved complexity estimate in one space dimension is obtained. In Section, 2 some preliminary notions from probability theory and functional analysis are introduced. In Section 3, the concept of random entropy solutions is introduced and the well-posedness of the scalar hyperbolic conservation law, i.e. \(m=1\), with random interval data is developed. In Subsection 3.1, the Cauchy problem for the scalar conservation law by setting \(m=1\) is presented. In Subsection 3.2, the distributional solutions in the class of entropy solutions are presented. In Subsection 3.3, the spatially homogeneous random flux functions are considered. Here, the concept of random data for the scalar conservation law is developed. In Subsection 3.4, the random scalar conservation law is rewritten in the terms of random data, introduced in Definition 3.1. Theorem 3.2 states that there exists a random entropy solution of the scalar conservation law. In Section 4, the multilevel Monte Carlo front-tracking algorithm for numerical solution of hyperbolic conservation law is developed. In Subsection 4.1, the discretization of the scalar conservation law is given. In Theorem 4.1, the convergence of the Monte Carlo estimates \(\mathbb{E}_M[U(\cdot,t)]\) in \(L^2(\Omega;L^1(\mathbb{R}^d))\) to \({\mathcal M}^1(U(\cdot,t))\) is presented. In Subsection 4.2, the concept of the front-tracking in the one-dimensional case is described. Algorithm 1, which is a modification of Graham's scan, gives the scheme how to calculate all solutions of the Riemann problems. In Subsection 4.2.2, the front-tracking in dimension \(d \geq 2\) is considered. The approximate solutions of the scalar conservation law \(U^{\eta}(x,t)\) is given. In Theorem 4.3 many properties of the quantity \(U^{\eta}(x,t)\) are presented. In Theorem 4.4 the convergence results for the approximation of the random scalar conservation law are presented. In Subsection 4.3, the multiresolution decomposition of the random flux on the phase of the solution is developed. In Subsection 4.4, the convergence analysis of the difference \(\mathbb{E}[U(t)] - \mathbb{E}_L^{MLMC}[U^L(t)]\) of the statistical mean \(\mathbb{E}[U(t)]\) and the expectations of increments for each level are developed. The main result is presented in Theorem 4.5, where an upper bound of the \(L^2\)-norm \(|| \mathbb{E}[U(t)] - \mathbb{E}_L^{MLMC}[U^L(t)] ||^2_{L^2(\Omega;L^1(\mathbb{R}^d))}\) is shown. In Section 5, a performance of the multilevel Monte Carlo method is tested on several examples with random fluxes in one and two space dimensions.