A Simultaneous Multi-resolution Discretization of PDEs with Random Inputs

1D Elliptic PDE
Consider the linear PDE d d a(x, y ) u(x, y ) dx dx u(0, y ) = u(1, y ) = 0, where y is a uniform random variable y U(1, 1 ). The scaler controls the variability in the diusion coecient a(x, y ), hence the solution u(x, y ), and is taken to be = 0.02 here. We consider the following model for a(x, y ), a(x, y ) = 1 + 4y (x 2 x).

= 1,

x (0, 1),

1D Elliptic PDE
The solution of this problem exhibits rather sharp gradients close to x = 0.5 and y = 1 .

u(x, y )

y x

1D Elliptic PDE
The standard low-order Gauss Legendre quadrature rule does not allocate a sucient number of abscissas close to y = 1 thus will not lead to accurate estimation of the solution statistics.
10
2

10

10

N=3 N=5 N=10 N=25 N=50 N=100

10

rel. error (%)

10

10

10

10

10

12

10

14

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Figure: Relative error in Variance of the solution. N is the number of Gauss Legendre abscissas.