Quadrature based finite element approximations to time dependent parabolic equations with nonsmooth initial data (Q1282124)

The authors consider semidiscrete and fully discrete finite element approximations of the parabolic problem \[ \begin{alignedat}{2} u_t + A(t)u &=0\quad &&\text{in }\Omega \times (0,T], \\ u &= 0 \quad&&\text{on } \partial \Omega \times (0,T],\\ u(\cdot ,0) &= u_0\quad&&\text{in }\Omega,\end{alignedat} \] in which \(\Omega\) is a convex bounded domain in \(\mathbb R^2\) with smooth boundary \(\partial \Omega\). The operator \(A\) is uniformly elliptic, selfadjoint, and of second order, with smooth coefficients. In the case of fully discrete approximations, the backward Euler method is used to discretize in time. The problem is posed in weak form, and all integrals are approximated by a quadrature scheme. The main goal of the work is to study the effect of numerical quadrature on approximate solutions, for the case in which the data is rough, in the sense that \(u_0\) is required to belong only to the Sobolev space \(H^1_0(\Omega)\). The quadrature scheme approximates the \(L^2\) inner product on a triangle by \[ (\phi,\psi)_h := (1/3)(\text{area of triangle)(sum of products of \(\phi\) and \(\psi\) at the vertices)}, \] and a similar scheme is applied to the term involving \(A\). The work differs from that of \textit{C. M. Chen} and \textit{V. Thomée} [J. Aust. Math. Soc., Ser. B 26, 329-354 (1985; Zbl 0576.65110)] in that these authors considered only the case \(A = -\Delta\) (the Laplacian), and furthermore they applied quadrature only to the term involving \(u_t\). The authors establish a set of stability estimates, in which norms of the solution and its derivatives depend on appropriate norms of \(u_0\). They then, in a lengthy analysis, derive error estimates in \(L^2\), \(H^1\), and \(L^\infty\) norms, which are respectively of order \(h^2t^{-1/2}\), \(ht^{-1}\), and \(h^{2-\epsilon}t^{-1}\), for the semidiscrete problem. For the fully discrete scheme a further lengthy analysis yields estimates of the error in various norms; for example, the error in the \(H^1\) norm is of order \(t_n^{-1}( h + k (1+ (\log k^{-1})^{1/2}))\).