Advances in Applied Mathematics and Mechanics DOI: 10.4208/aamm.2014.5.

s6
Adv. Appl. Math. Mech., Vol. 6, No. 5, pp. 693-698 October 2014

A Remark on the Courant-Friedrichs-Lewy Condition

in Finite Difference Approach to PDE’s
Kosuke Abe1 , Nobuyuki Higashimori2 , Masayoshi Kubo3 ,
Hiroshi Fujiwara3 and Yuusuke Iso3, ∗
1 Nihon University, Tokyo, Japan
2 Hitotsubashi University, Kunitachi, Japan
3 Graduate School of Informatics, Kyoto University, Kyoto, Japan

Received 30 September 2013; Accepted (in revised version) 31 March 2014

Available online 23 July 2014

Abstract. The Courant-Friedrichs-Lewy condition (The CFL condition) is appeared in

the analysis of the finite difference method applied to linear hyperbolic partial differ-
ential equations. We give a remark on the CFL condition from a view point of sta-
bility, and we give some numerical experiments which show instability of numerical
solutions even under the CFL condition. We give a mathematical model for rounding
errors in order to explain the instability.
AMS subject classifications: 65M06, 65M12, 65G50
Key words: Numerical analysis, finite difference scheme, stability, CFL condition.

1 Introduction
The finite difference method (FDM) has been discussed as one of the mathematical tools
to deal with partial differential equations before the era of digital computers till now,
and R. Courant, K. O. Friedrichs and H. Lewy gave precise discussion about asymptotic
behaviour of FDM solutions in [1]; we can see ”we will find that for the case of the initial
value problem for hyperbolic equations, convergence is obtained only if the ratio of the
mesh widths in different directions satisfies certain inequalities which in turn depend on
the position of the characteristics relative to the mesh” in [1] (this sentence is quoted from
its English translation [2]). As is pointed out in [6], we note as follows; it is a necessary
condition of convergence for FDM solution that the region of dependence of the finite

∗ Corresponding author.
Email: [email protected] (K. Abe), [email protected] (N. Higashimori), kubo@i.
kyoto-u.ac.jp (M. Kubo), [email protected] (H. Fujiwara), [email protected] (Y. Iso)

http://www.global-sci.org/aamm 693 c 2014 Global Science Press

Downloaded from https:/www.cambridge.org/core. University of Wollongong, on 08 Jan 2017 at 12:27:14, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.
http://dx.doi.org/10.4208/aamm.2014.5.s6
 694 K. Abe, N. Nobuyuki, M. Kubo, H. Fujiwara and Y. Iso / Adv. Appl. Math. Mech., 6 (2014), pp. 693-698

difference schema contains that of the corresponding hyperbolic differential equation.

We call this condition the Courant-Friedrichs-Lewy condition (CFL condition).
Let us introduce the CFL condition for the case
∂ ∂
u(t,x) = c u(t,x), x ∈ R, t > 0, (1.1a)
∂t ∂x
u(0,x) = u0 ( x), x ∈ R, (1.1b)

where c is a positive constant. Let ∆t and ∆x be increments along the t-direction and
x-direction respectively, and we give a discretization by FDM

ukj +1 − ukj ukj+1 − ukj

=c , (1.2)
∆t ∆x

where ukj denotes the value corresponding to u(k∆t, j∆x). Introduction of λ := ∆t/∆x
leads us to
ukj +1 = cλukj+1 +(1 − cλ)ukj , (1.3)

and the inequality 1 − cλ ≥ 0 is the CFL condition for (1.2) and (1.3).
We sometimes encounter misunderstanding that the CFL condition is considered as a
stability condition of the schema (1.2) or (1.3). The authors are afraid that it is caused by
misunderstanding the proper meaning of the Lax equivalence theorem [5]. The analysis
of stability of numerical solutions is one of the most important ones in the theory of
FDM, and we may find its origin in the historic paper [7] of G. G. O’Brien et al. They
gave, following the fundamental study of von Neumann, definition of stability so as to
estimate numerical errors mainly coming from the rounding errors. Lax et al. [5] gave
definition of stability as a discrete analogue to well-posedness of differential equations,
and they declared that numerical errors were not taken into account in [5]. Lax et al., on
the other hand, referred O’Brien et al. [7] as analysis of influence of numerical errors, and
the authors are afraid that many researchers may have misunderstood the differences
between their analyses.
Proper understanding of the analysis by O’Brien et al. [7] and by Lax et al. [5] is very
much significant in order to understand real effects of rounding errors in computation by
FDM. The authors consider it to be equivalent to give an answer to the question whether
stability conditions guarantee stable computation less influenced by the rounding errors.
For clear discussion, we need a mathematical model of propagation of the rounding er-
rors in computation, but we should remark that modeling is dependent upon a system
of floating point numbers on computers. When we restrict ourselves on computation of
an evolutional finite difference schema, we notice that O’Brien et al. [7] adopted a naive
model in their analysis of propagation of the rounding errors, and the authors should
remark that it has close relation with stability analysis by Lax et al. [5]. Unfortunately it
is inadequate in explanation of severe effects of the rounding errors occurred on current
digital computers.

Downloaded from https:/www.cambridge.org/core. University of Wollongong, on 08 Jan 2017 at 12:27:14, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.
http://dx.doi.org/10.4208/aamm.2014.5.s6
 K. Abe, N. Nobuyuki, M. Kubo, H. Fujiwara and Y. Iso / Adv. Appl. Math. Mech., 6 (2014), pp. 693-698 695

One of the aims of the present paper is to reconfirm the proper meaning of the CFL
condition, and we have already mentioned it above. We remark again that it is a nec-
essary condition of convergence of the finite difference solutions to the Cauchy problem
of linear hyperbolic equations. It is not a sufficient condition for stability. The other of
the aims is to show behaviours of the rounding errors for a stable finite difference schema
through concrete computation. To this end, we shall give a mathematical model for prop-
agation of the rounding errors arisen in the IEEE754 [3] computer environment in Section
2. We shall show some numerical examples to illustrate instability of numerical solutions
of a stable schema in Section 3. They may not only diverge but may converge to zero,
and both of them are quite different from the exact solutions of the difference schema.
We note that they are unstable numerical computations of (1.2) or (1.3) for the case the
CFL condition 1 − cλ > 0 holds.

2 Mathematical modeling of the rounding errors

Influence of rounding errors is surely inevitable in computations which deal with real
numbers on digital computers, but it is sometimes left outside of discussion in theo-
retical numerical analysis. Nonetheless we often encounter problems of accumulation
of rounding errors in computation with digital computers and are obligated to come to
know the difference between theories of computation and realistic computations. In the
analysis of the rounding errors, we always focus on the fact that real numbers on digital
computers have only finite significant figures, but we are often liable to forget the fact
that basic arithmetics equipped on them admit also rounding errors.
Real numbers on computers are expressed as floating point numbers with finite fig-
ures of the binary codes, and each one, in general, contains rounding errors. We call, in
the present paper, the four rules of the basic arithmetics on computers the computational
arithmetics, and they are executed within some tolerance of rounding errors. We remark
rounding errors appeared in real number computation are strictly ruled by the IEEE754
standard [3], but we also note that it leaves some flexibility or freedom to users. Hence
the same real numbers may have different floating number expressions on different com-
puters. As is well-known, the computational arithmetics do not satisfy the associative
law.
In order to illustrate effects of rounding errors, let us consider addition a + b for two
real numbers a and b. Denoting computational real numbers with a bar ·, we write a and
b for the expression of a and b on computers. We also denote the computational addition
by ⊕. Hence a mathematical statement a + b is executed as a ⊕ b on computers, where all
the rounding errors adherent to a ⊕ b are ruled by the IEEE754 standard now.
For the analysis of the rounding errors, we need a suitable mathematical model cor-
responding to their propagation in the processes of computation. Denoting the total sum
of the rounding errors by δ, we know that

a+b+δ (2.1)

Downloaded from https:/www.cambridge.org/core. University of Wollongong, on 08 Jan 2017 at 12:27:14, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.
http://dx.doi.org/10.4208/aamm.2014.5.s6
 696 K. Abe, N. Nobuyuki, M. Kubo, H. Fujiwara and Y. Iso / Adv. Appl. Math. Mech., 6 (2014), pp. 693-698

is a popular model of the rounding errors to a ⊕ b, and it explains many types of compu-
tational phenomena. The authors strongly propose adopting, in place of (2.1),

(1 + ǫ)( a + b) (2.2)

as a mathematical model corresponding to a ⊕ b in analysis of an evolutional finite dif-

ference schema, where |ǫ| is small and is determined by the rule of rounding-off. For a
single calculation, there are no differences between (2.1) and (2.2), but they are completely
different to each other in analysis of iterative calculations.
The mathematical modeling (2.2) comes from strict understanding of the rule of
rounding errors of both real numbers and the computational arithmetics. We find analy-
sis based on the models of the type of (2.2) in the famous book of Wilkinson [8]. The mod-
els of the type of (2.1) are convenient to know rough estimates of the rounding errors in
finite numbers of calculations, and they often give reasonable estimates of accumulation
of the rounding errors in polynomial order. For recurrent schema or evolutional finite
difference schema, each computation should consist of finite number of calculations, but
the numbers of calculation become huge as change of parameters in the schema; the total
number of calculations for (1.2) or (1.3) becomes huge as a parameter ∆t to zero. Hence
the accumulation of the rounding errors sometimes does not only grow exponentially
but also tends to zero for the cases. We do not mention here detailed estimates of these
severe behaviours of the rounding errors, but we consider it possible to give their quan-
titative estimates by using the model of the type of (2.2). We remark that such estimates
can be shown by the analysis not only of a finite difference scheme but also of orders of
arithmetics appeared in the scheme, and they require delicate quantitative estimation of
rounding errors for case by case.
Some other researchers have tried to discuss the same problem with us using the
models of the type of (2.2), and we refer, for example, the paper of Jézéquel [4], which
deals with the case of the heat equation. The authors should emphasize that we discuss
the problem in connection with stability and convergence analysis of an evolutional finite
difference schema. We shall show two types of instability of numerical solutions under
the stability condition in the next section.
Returning to the historic work of O’Brien et al. [7], we notice that they adopted the
models of the type (2.1). Applying their approach to (1.3), we obtain

ũkj +1 = cλũkj+1 +(1 − cλ)ũkj + δjk ∆t, (2.3)

where δjk is the total sum of the rounding errors at k-th level. Since we can deal with the
rounding errors as an inhomogeneous term of the difference equation (1.3), we are suc-
cessful in Fourier analysis to conclude that 1 − cλ ≥ 0 is a sufficient condition for stability
in the sense of Lax. We should remark that rounding errors coming from the computa-
tional arithmetics are not taken into account in the formula (2.3) and that it is inadequate
for explanation of influence of the rounding errors.

Downloaded from https:/www.cambridge.org/core. University of Wollongong, on 08 Jan 2017 at 12:27:14, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.
http://dx.doi.org/10.4208/aamm.2014.5.s6
 K. Abe, N. Nobuyuki, M. Kubo, H. Fujiwara and Y. Iso / Adv. Appl. Math. Mech., 6 (2014), pp. 693-698 697

3 Numerical examples for instability

We show numerical examples for the equations (1.1a) and (1.1b) with c = 1 and u0 ( x) = 1.
Our finite difference schema is the formula (1.3) with a periodic boundary condition. The
numerical computation is done for the case ∆x = 1/20000 and ∆t = 1/30000, and it is
executed with the single precision of IEEE754. We note λ = 2/3 in this case and that the
stability condition in the sense of Lax holds. Furthermore we remark that λ = 2/3 also
satisfies the CFL condition.
The authors are afraid that we are liable to imagine divergence of numerical solutions
as their instability of numerical solutions, but we should notice that effects of the round-
ing errors are much complicated. The case (a) in Fig. 1 is the case that numerical solutions
show divergence, but the case (b) shows extinction of numerical solutions.
Since the IEEE754 standard leaves users’ flexibility in rounding-off within the single
precision environment, we computed our finite difference schema in two ways; in the
”round toward +∞” mode and in the ”round toward −∞” mode. The former case cor-
responds to ǫ > 0 in (2.2), and the latter does to ǫ < 0. We remark again that numerical
solutions may tend to zero as execution going and we should recognize that this case is
one of the types of numerical instability. We easily notice instability of computation when
numerical solutions show signs of divergence, but we are very much afraid that we may
lose sign of instability for the latter case of extinction.
We conclude that numerical solutions may be unstable even under the stability con-
dition in the sense of Lax, and the CFL condition is surely far from stability of numerical
solutions. Without precise analysis of the behaviour of the rounding errors, we might
misunderstand unstable numerical results as the aimed solution. The situation becomes
worse for accurate computation that the mesh sizes are very small. The stability condi-
tion in the theory of FDM is that for the exact solutions to the difference equations, and
they are different from numerical solutions obtained by computation using the differ-
ence equations. Furthermore, we should know that there are many types for numerical
instability caused by the rounding errors.

1
20000
0.8
15000 0.6
10000 0.4
5000 0.2
0
0

0
0
0.2 0.2
0.4
0.4
2500 3000 x 0.6 2500 3000
x 0.6
1500 2000 0.8 1500 2000
0.8
1 0 500 1000 1 0 500 1000
t t

(a) Within ”Round toward +∞” Mode (b) Within ”Round toward −∞” Mode

Figure 1: Numerical Results with the Single Precision of IEEE754, 0 ≤ t ≤ 3000.

Downloaded from https:/www.cambridge.org/core. University of Wollongong, on 08 Jan 2017 at 12:27:14, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.
http://dx.doi.org/10.4208/aamm.2014.5.s6
 698 K. Abe, N. Nobuyuki, M. Kubo, H. Fujiwara and Y. Iso / Adv. Appl. Math. Mech., 6 (2014), pp. 693-698

Acknowledgements
The authors would like to express their thanks to Dr. Jiwoon Kim, since they could not
notice the present problem without discussion with him. This research was partially
supported by Grant-in-Aid for Scientific Research (B) 22340018 from Japan Society for
the Promotion of Science.

References
[1] R. C OURANT, K. F RIEDRICHS AND H. L EWY, Über die partiellen Differenzengleichungen der
mathematischen physik, Math. Ann., 100 (1928), pp. 32–74.
[2] R. C OURANT, K. F RIEDRICHS AND H. L EWY, On the partial difference equations of mathemat-
ical physics, IBM J. Res. Develop., 11 (1967), pp. 215–234 (English translation of the original
work [1]).
[3] IEEE Standard for Binary Floating-Point Arithmetic, IEEE Std 754-1985, 1985.
[4] F. J ÉZ ÉQUEL, Round-off error propagation in the solution of the heat equation by finite differences,
J. Univ. Comput. Sci., 1 (1995), pp. 469–483.
[5] P. D. L AX AND R. D. R ICHTMYER, Survey of the stability of linear difference equations, Commun.
Pure Appl. Math., 9 (1956), pp. 267–293.
[6] P. D. L AX, Hyperbolic difference equations: a review of the courant-friedrichs-lewy paper in the light
of recent developments, IBM J. Res. Develop., 11 (1967), pp. 235–238.
[7] G. G. O’B RIEN , M. A. H YMAN AND S. K APLAN, A study of the numerical solution of partial
differential equations, J. Math. Phys., 29 (1951), pp. 223–251.
[8] J. H. W ILKINSON, Rounding Errors in Algebraic Processes, Her Majesty’s Stationery Office,
London, 1963.

Downloaded from https:/www.cambridge.org/core. University of Wollongong, on 08 Jan 2017 at 12:27:14, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.
http://dx.doi.org/10.4208/aamm.2014.5.s6