Numerical solution of ordinary and
partial dierential Equations
Module 16: Predictor-Corrector methods
Dr.rer.nat. Narni Nageswara Rao
£
August 2011
1 General Multistep Methods
Consider the linear multistep methods
yj+1 =
kX
i=0
aiyj i + h
kX
i=0
bifj i + hb 1fj+1j = k; k + 1; ¡¡¡ (1)
Which are k+1 step methods, k ! 0. For k = 0, we recover one-step methods.
The coecients ai, bi are real and fully identify the scheme, thus are such
that ak T= 0, or bk T= 0. If b 1 T= 0 the scheme is implicit, otherwise it is
explicit.
2 Predictor-Corrector Methods
When solving nonlinear IVP of the form
y
H = f(t; y); y(t0) = y0
at each time step implicit schemes require solving nonlinear algebraic equa-
tions. For instance, if the Crack-Nicolson method is used, we get the nonlin-
ear equation
yj+1 = yj +
h
2
[fj + fj+1] = (yj+1)
£nnrao maths@yahoo.co.in
1
that can be cast in the form (yj+1) = 0, where (yj+1) = yj+1   (yj+1).
To solve this equation the Newton method would give
y
(n+1)
j+1 = y
(n)
j+1   (y
(n)
j+1)
H(y
(n)
j+1)
for n = 0; 1; ¡¡¡ until convergence and require an initial datum y
(0)
j+1 su-
ciently close to yj+1. Alternatively one can resort to
xed point iterations
y
(n+1)
j+1 = (y
(n)
j+1) (2)
for n = 0; 1; ¡¡¡ until convergence. In such a case, the global convergence con-
dition for the
xed-point method (see theorem in
rst year B-Tech content)
sets a constraint on the diccretization step size of the form
h 
1
jb 1jL
(3)
where L is the Lipschitz constant of f with respect to y. In practice, ex-
cept for the case of shift problems, this restriction on h is not signi
cant
since considerations of accuracy put a much more restrictive constraint on h.
However, each iteration of (2) requires one evaluation of the function f and
the computational cost can be reduced by providing a good initial guess y
(0)
j+1.
This can be done by taking one step of an explicit multistep method and then
iterating on (2) for a
xed numbe m of interations. By doing so, the implicit
multistep method that is employed in the
xed-point scheme Corrects the
value of yj+1 predicted by the explicit multistep method. A procedure of
this sort is called a predictor-corrector method. or PC method. There are
many ways in which a predictor-corrector method can be implemented.
2.1 Basic Version
The value y
(0)
j+1 is computed by an explicit ~k + 1-step method called the
Predictor (here identi
ed by the coecients f~ai; ~big).
[P ] y
(0)
j+1 =
~kX
i=0
~aiy
(1)
j i + h
~kX
i=0
~bif
(0)
j i
where f
(0)
k = f(tk; y
(0)
k ) and y
(1)
k are the solutions computed by the PC method
at the previous steps or are the initial conditions. Then, we evaluate the
function f at the new point (tj+; y
(0)
j+1)(evaluation step)
[E] f
(0)
j+1 = f(tj+; y
(0)
j+1)
2
and

Ma2002 1.16 rm

  • 1.
    Numerical solution ofordinary and partial dierential Equations Module 16: Predictor-Corrector methods Dr.rer.nat. Narni Nageswara Rao £ August 2011 1 General Multistep Methods Consider the linear multistep methods yj+1 = kX i=0 aiyj i + h kX i=0 bifj i + hb 1fj+1j = k; k + 1; ¡¡¡ (1) Which are k+1 step methods, k ! 0. For k = 0, we recover one-step methods. The coecients ai, bi are real and fully identify the scheme, thus are such that ak T= 0, or bk T= 0. If b 1 T= 0 the scheme is implicit, otherwise it is explicit. 2 Predictor-Corrector Methods When solving nonlinear IVP of the form y H = f(t; y); y(t0) = y0 at each time step implicit schemes require solving nonlinear algebraic equa- tions. For instance, if the Crack-Nicolson method is used, we get the nonlin- ear equation yj+1 = yj + h 2 [fj + fj+1] = (yj+1) £nnrao maths@yahoo.co.in 1
  • 2.
    that can becast in the form (yj+1) = 0, where (yj+1) = yj+1   (yj+1). To solve this equation the Newton method would give y (n+1) j+1 = y (n) j+1   (y (n) j+1) H(y (n) j+1) for n = 0; 1; ¡¡¡ until convergence and require an initial datum y (0) j+1 su- ciently close to yj+1. Alternatively one can resort to
  • 3.
    xed point iterations y (n+1) j+1= (y (n) j+1) (2) for n = 0; 1; ¡¡¡ until convergence. In such a case, the global convergence con- dition for the
  • 4.
  • 5.
    rst year B-Techcontent) sets a constraint on the diccretization step size of the form h 1 jb 1jL (3) where L is the Lipschitz constant of f with respect to y. In practice, ex- cept for the case of shift problems, this restriction on h is not signi
  • 6.
    cant since considerations ofaccuracy put a much more restrictive constraint on h. However, each iteration of (2) requires one evaluation of the function f and the computational cost can be reduced by providing a good initial guess y (0) j+1. This can be done by taking one step of an explicit multistep method and then iterating on (2) for a
  • 7.
    xed numbe mof interations. By doing so, the implicit multistep method that is employed in the
  • 8.
    xed-point scheme Correctsthe value of yj+1 predicted by the explicit multistep method. A procedure of this sort is called a predictor-corrector method. or PC method. There are many ways in which a predictor-corrector method can be implemented. 2.1 Basic Version The value y (0) j+1 is computed by an explicit ~k + 1-step method called the Predictor (here identi
  • 9.
    ed by thecoecients f~ai; ~big). [P ] y (0) j+1 = ~kX i=0 ~aiy (1) j i + h ~kX i=0 ~bif (0) j i where f (0) k = f(tk; y (0) k ) and y (1) k are the solutions computed by the PC method at the previous steps or are the initial conditions. Then, we evaluate the function f at the new point (tj+; y (0) j+1)(evaluation step) [E] f (0) j+1 = f(tj+; y (0) j+1) 2
  • 10.