On the Stability of the Iterated Crank-Nicholson Method in Numerical Relativity

Saul A. Teukolsky
Newman Laboratory, Cornell University, Ithaca, NY 14853
(February 3, 2008)

are similar.) A simple first-order accurate differencing

The iterated Crank-Nicholson method has become a pop- scheme is FTCS (Forward Time Centered Space):
ular algorithm in numerical relativity. We show that one
should carry out exactly two iterations and no more. While un+1 − unj
j unj+1 − unj−1
the limit of an infinite number of iterations is the standard = . (2)
Crank-Nicholson method, it can in fact be worse to do more ∆t 2∆x
than two iterations, and it never helps. We explain how this Here n labels the time levels and j the spatial grid points.
paradoxical result arises.
It is a standard textbook result that this scheme is
arXiv:gr-qc/9909026v1 7 Sep 1999

04.25.Dm, 04.20.-q, 04.70.-s unconditionally unstable. One sees this with a von Neu-
mann stability analysis: Put

unj = ξ n eikj∆x (3)

I. INTRODUCTION
and find that the amplification factor ξ is
There is currently a large worldwide effort underway ξ = 1 + iα sin k∆x, (4)
attempting to solve Einstein’s equations numerically for
astrophysically interesting scenarios. The problem is ex- where α = ∆t/∆x. Since |ξ|2 > 1 for any choice of α,
tremely challenging technically. Among the difficulties the method is unconditionally unstable.
is that of finding a finite-difference scheme that allows a Backwards differencing gives a stable scheme:
stable time evolution of the system. It is well-known that
implicit differencing schemes tend to be stable. However, un+1
j − unj un+1 n+1
j+1 − uj−1
the difficulty of solving the resulting implicit algebraic = , (5)
∆t 2∆x
equations, especially in three spatial dimensions, has led
most researchers to stay with explicit methods and their for which
potential instabilities. 1
Several years ago, Choptuik proposed solving the im- ξ= . (6)
plicit Crank-Nicholson scheme by iteration. This would 1 + iα sin k∆x
effectively turn it into an explicit scheme, but hopefully Now |ξ|2 < 1 for any choice of α, and so the method is
by iterating until some convergence criterion was met one unconditionally stable.
would preserve the good stability properties of Crank- The Crank-Nicholson scheme is a second-order accu-
Nicholson. The iterated Crank-Nicholson scheme has rate method obtained by averaging equations (2) and (5).
subsequently become one of the standard methods used Now one finds
in numerical relativity.
In this note, we point out that when using iterated 1 + 12 iα sin k∆x
Crank-Nicholson, one should do exactly two iterations ξ= . (7)
1 − 12 iα sin k∆x
and no more. While the limit of an infinite number of
iterations is the implicit Crank-Nicholson method, it can Since |ξ|2 = 1, the method is stable. It is the presence
in fact be worse to do more than two iterations, and it of the quantities un+1 on the right hand side of equation
never helps. (5) that makes the method implicit.
The first iteration of iterated Crank-Nicholson starts
by calculating an intermediate variable (1) ũ using equa-
II. ITERATED CRANK-NICHOLSON tion (2):
(1) n+1
To understand this paradoxical result, consider differ- ũj − unj unj+1 − unj−1
encing the simple advective equation = . (8)
∆t 2∆x
∂u ∂u Then another intermediate variable (1) ū is formed by av-
= . (1)
∂t ∂x eraging:
(Many equations in numerical relativity are general- (1) n+1/2
ūj = 12 ((1) ũn+1
j + unj ). (9)
izations of this form, and the differencing techniques

1
Finally the timestep is completed by using equation (2) smallest number of iterations for which the method is
again with ū on the right-hand side: stable is two, there is no point in carrying out more iter-
ations than this.
(1) n+1/2 n+1/2
un+1
j − unj ūj+1 − (1) ūj−1 Note that there was nothing special about using the
= . (10) advective equation (1) for this analysis. Similar behavior
∆t 2∆x
is found for the wave equation, written in first-order form
(Iterated Crank-Nicholson can alternatively be imple-
mented by averaging the right-hand side of equation (1). ∂u
= v, (18)
For linear equations, this is completely equivalent.) ∂t
Iterated Crank-Nicholson with two iterations is carried ∂v ∂2u
out in the same way. After steps (8) and (9), we calculate = , (19)
∂t ∂x2
(2) n+1 (1) n+1/2 n+1/2 with the standard centered difference formula for the sec-
ũj − unj ūj+1 − (1) ūj−1
= , (11) ond derivative term. One recovers the usual Courant
∆t 2∆x
(2) n+1/2
condition (without the factor of 2) for the stable cases.
ūj = 21 ((2) ũn+1
j + unj ). (12)

Then the final step is computed analogously to equation III. ACKNOWLEDGEMENTS

(10):
(2) n+1/2 n+1/2 This work was supported in part by NSF grant PHY
un+1
j − unj ūj+1 − (2) ūj−1
= . (13) 99-00672 and NASA Grant NAG5-7264 to Cornell Uni-
∆t 2∆x versity.
Any number of iterations can be carried out in the same
way.
Now consider the stability of these iterated schemes.
If we define β = (α/2) sin k∆x, and call the FTCS
scheme (2) the zeroth-order method, then direct calcula-
tion shows that the amplification factors are
(0)
ξ = 1 + 2iβ, (14)
(1) 2
ξ = 1 + 2iβ − 2β , (15)
(2) 2 3
ξ = 1 + 2iβ − 2β − 2iβ , (16)
(3) 2 3 4
ξ = 1 + 2iβ − 2β − 2iβ + 2β , (17)

and so on. As one would expect, these are exactly the

same values one gets by expanding equation (7) in powers
of β and truncating at the appropriate point.
To check stability, compute |ξ|2 for each of these ex-
pressions. You find an alternating pattern. Levels 0 and
1 are unstable; levels 2 and 3 are stable provided β 2 ≤ 1;
levels 4 and 5 are unstable; levels 6 and 7 are stable
provided β 2 ≤ 1; and so on. Since the stability require-
ment must hold for all wave numbers k, it translates into
α2 /4 ≤ 1, or ∆t ≤ 2∆x. This is just the Courant condi-
tion (the 2 occurs because of the 2 in eqn. [2]).
Now we see the resolution of the paradox: While the
magnitude of the amplification factor for iterated Crank-
Nicholson does approach 1 as the number of iterations
becomes infinite, the convergence is not monotonic. The
magnitude oscillates above and below 1 with ever de-
creasing oscillations. All the cases above 1 are unstable,
although the instability might be very slowly growing for
a large number of iterations.
The accuracy of the scheme is determined by the trun-
cation error. This remains second order in ∆t and ∆x
from the first iteration on. Doing more iterations changes
the stability behavior, but not the accuracy. Since the