Pure Sciences
205
ferential equations, however, it is sufficient for the method to be A (a) -stable, for some a (0, π/2), i.e., the region oif absolute stability should include the sector K : Re K <0, arg (-k )< a) in the complex -plane. This is so be- cause parabolic equations tend to have spectrum inside a narrow sector about the positive axis (the spectrum is contained in the numerical range, which is contained in a parobola).
Methods which are A-stable are necessarily implicit. Dahlquist (6) has shown that linear multistep methods cannot be A-stable if their order is greater than two and that in fact the trapezoid rule is the 2nd order method with the smallest error constant.
Axelsson (2) shows that certain m-stage fully implicit Runge-Kutta methods based on Radau quadrauture are A-stable. These methods, unfortunately, re- quire that we solve a system of m implicit equations simultaneously. This is often not an easy task, particularly if the equation is not linear. In this paper, however, we will consider only Rung-Kutta methods which are diagonally im- plicit, that is, one is required only to solve a single implicit equation at each of the m successive stages of the method.
We also mention that Rosenbrock has also designed semi-implicit Rung- Kutta methods which make use of f' = f/ y (see Axeisson (2) and Rosen- brock (12). In Wilson (3), a class of specialized methods for linear second order systems arising in structural dynamics is discussed. These methods turn out to be nearly absolutely stable for all eigenfrequencies.
=
In Section 2, one step diagnoally implicit Runge-Kutta mehods are present- ed. Applying such a method to the test equation y' y amounts to approxi- mating e K by a rational function R(k) with real poles in contrast to fully im- plicit Runge-Kutta methods which usually have complex poles. Miller (11) de- signed two such methods. His 2nd order method is a 2-stage method and his 3rd order method is a 3-stage method. Both of these methods are also strongly sta- ble at infinite, i.e., lim IR (K )1 = 0. We will derive a 3rd order 2-stage me- thod which is not strongly stable t infinity. We will also derive three 4th order methods which are strongly stable at infinity, althogh some of the intermediate stages are not.
In Section 3, we discuss multistep diagonally implicit Rung-Kutta methods. These methods will be A-stable provided that the spectral radius of a certain matrix G (z) is less than one for Re z < 0. We develop two 4th order methods; one is a 3-step stiffly A-stable method and the other is a 2-step method which is only A-stable.
ص ٣٥٩