The Gibbs effect in difference schemes (Q610541)

Problems for one-dimensional linear parabolic equations are considered well studied. There is a family of schemes with weights, which contains the classical half-sum scheme. It is unconditionally stable and asymptotically conditionally stable and has precision \(O(\tau^2+h^2)\), where \(\tau\) is a time stepsize and \(h\) is a space stepsize. This scheme is extensively used in applied computations. Other schemes are used rarely at present. It was mentioned in earlier work that the half-sum scheme is not monotone. If the initial data are discontinuous, then, at the first time step, the computed profile of solution becomes nonmonotone, while the exact solution is monotone. However, this phenomenon has not been studied in detail. It was believed that this effect fades out in a small number of time steps and does not affect convergence at any finite moment of time. This paper studies the convergence of various difference schemes for parabolic problems with discontinuous initial data. It is shown that the half-sum scheme does not converge in the \(C\) norm, and it converges in the \(L_2\) norm with accuracy \(\sim\tau^{0.5}\), i.e., very slowly. More over, near the initial position of the discontinuity, the numerical solution strongly oscillates, and the oscillation amplitude does not decrease as \(\tau\to 0\) and \(h\to 0\). This resembles the well-known Gibbs effect, which arises in the Fourier approximations of discontinuous functions. Such a behavior seriously complicates computations and substantially narrows the domain of application of the scheme. The half-sum scheme gives good results only when the initial and boundary data, as well as the sources, are continuous and their derivatives are not very large. We explain the reason for this phenomenon. The parabolic problems are stiff. The half-sum scheme is only \(A\)-stable, and it belongs to the boundary of the range of \(A\)-stability. The best damping of the Gibbs effect is provided by \(L\)-stable schemes, such as the \(L_1\)-stable purely implicit scheme with precision \(O(\tau)\) and the \(L_2\)-stable scheme of \textit{H. H. Rosenbrock} [Comput. J. 5, 329--330 (1963; Zbl 0112.07805)], which has precision \(O(\tau^2)\). The latter scheme gives best results in numerical calculations.