J. Soc. INDUST. APPL. MATH.

Vol. 3, No. 1, March, 1955

Printed in U.S.A.

--

ON THE NUMERICAL INTEGRATION OF Ou

IMPLICIT

METttODS*

JIM DOUGLAS, JR.

02u

Ou
BY
Ot

Oy

Introduction. The purpose of this report is to investigate how solutions

of boundary value problems for the heat flow equation in two space dimensions can be obtained by solving certain difference equations in a rectangular lattice and taking the limit as the mesh size of the lattice tends to zero.
All problems treated are of the form

(I)

02u

02u

Ox

Oy

u(x, y, O)
u(O, y, t)
u(x, O, t)

Ou
Ot

(0 < x,y < 1, t> 0),

f(x, y),
gl(y, t),
h(x, t),

u(1, y, t)
u(x, 1, t)

(t > 0),
(t > 0).

g(y, t)
h(x, t)

Interest in the derivation of a numerical solution arises from two sources.

First, the formal solution as given by either a quadrature using an influence
function or a series resulting from the separation of variables technique
(if possible) is usually difficult to evaluate. Second, numerical procedures
applicable to the heat flow equation often extend in modified form to nonlinear parabolic equations.
Quite a number of papers have been devoted to the study of numerical
methods for the solution of the related one-dimensional equation

Ou

(II)

Ou
Ot

Ox

under various conditions. Most of the work has been centered around the
explicit difference equation

(III) w(x,

+ At)

w(x, t) -t" At [w(x Ax, t)

(Ax)

2w(x, t)

w(x -b Az, t)].

It has been known [2] for some time that the solution of (III) agreeing with
the initial condition of u will not approximate closely the solution of (II)
except for certain special initial values [9] unless
At
gv)

x*

* Presented to the American Mathematical Society, August 31, 1954.

] Received by the editors October 18, 1954.
42

43

NUMERICAL INTEGRATION BY IMPLICIT METHODS

Under the assumption that },* __< 1/2, various authors [5, 7] have proved
that the solution of (III) converges to that of (II) in any finite region.
Moreover, the dependence of the rate of convergence on the smoothness
of the initial condition has been studied [7] in the special case of vanishing
boundary values.
More general differential equations of the form
Ou
Ot

(V)

ao(x, t)

Ou

+ al(x, t) Ou -I- as(x, t, u)

have also been studied [6] for the initial value problem on the infinite interval
< x < Again, explicit difference formulae were used"

(VI) w(x, --[- At) "r c,(x, t)w(x + rAx, t) -t- (At)as(x, t, W(X, t) ).
For convergence, it was necessary to restrict to be small.

The restriction of small },* leads to the taking of a very large number of
time steps to complete the solution of the problem. Various implicit methods
have been proposed [3, 4, 8, 10] to alleviate this difficulty. Two of these
difference equations, implicit in form are"

u(x
(VII)

t
[u(x /x t+
(5x)
2u(x, +/t) + u(x

+ t)

u(x, +/t)

2(hx)

[u(x -/x,

2u(x, -{- At)

(VIII)

u(x, t)

+ x, + t)]

u(, t),

+ u(x W Ax,

At
[u(x -/x, t)
+ 2(5X)

-t- At)]
2u(x, t)

+ u(x + Ax, t)].

It is easy to show that either of these two difference equations is stable for
any value of h*; however, to the authors knowledge no convergence proof
has been published for either method, but the methods of this report can
be used to give proofs. Thus, the number of time steps necessary to complete the problem can be comparatively small. There is, of course, the
problem of solving the linear equations, but this is not troublesome if the
method outlined in [1] is used.
In the two-dimensional case, the analogue of (III) is

u(x, y,

(IX)

+ at)

u(x, y, t)

+ h*[u(x- 5x, y, t)- u(x

-t- u(x, y + Ay, t) + u(x, y

/x, y, t)
5y, t)- 4u(x, y, t)]

44

JIM

DOUGLAS

JR.

where

(x)

x*

(Ay)

(Ax)

"

Again it is necessary to limit the size of ),*. Equation (IX) is stable only
the results in the explicit case above have obvious analogues in
if h* -<_
the two-dimensional problem. Thus, the complete solution involves a vast
amount of calculation.
This report is devoted to the theoretical study of various implicit difference schemes, each stable for all },* for the numerical integration of (I).
An accompanying report of Peaceman and Rachford [11] treats the practical aspects of the calculation in considerable detail and gives application
of one method to the Dirichlet problem by treating the time step as an
iteration parameter; machine time and storage requirements are included
in [11].
Chapter I contains a discussion of an implicit method that has no analogue in the one space dimensional case. Chapters II and III concern the
direct analogues of equations (VII) and (VIII). Each of the three procedures leads to a set of N linear equations at each time step, where N is
the number of internal points on the side of the mesh. The great difference
between the schemes is that the first method gives equations each one of
which involves three unknowns and the other two give equations involving
five unknowns. Moreover, the equations in the first procedure break into
N groups of N equations each, and each group may be solved directly by
the method mentioned previously [1], whereas the equations of the second
and third schemes must be solved by iteration.
Convergence theorems are given for all three methods. The result is that
the three are essentially equivalent, as the truncation error is shown to be
uniformly O(At), assuming that the initial and boundary conditions are
such that (I) possesses a six-times differentiable solution. No effort is made
in this report to study the behavior of the rate of convergence with changing smoothness restrictions on the initial and boundary data.
In the appendix a lemma on Fourier coefficients is proved.

-;

CHAPTER

1. Approximation method, Consider the heat flow equation

(1)

O2u

Ou_ Ou

Ox

Oy

in the semi-infinite cylinder 0 _-< x

condition

=<

Ot

1, 0

=< y <-

1,

>

0 with the initial

45

NUMERICAL INTEGRATION BY IMPLICIT METHODS

u(x, y, O)

(la)

F(x, y)

and boundary conditions

fl(y, t)
f(y, t)

y, t)

y, t)

(lb)

u(x, O, t)

(t > 0),

f3(x, t)
f,(x, t)

u(x, 1, t)
We shall assume that the initial and boundary conditions are such that a

=<

y -< 1, 0 -<
solution u(x, y, t) exists in the closed region 0 -<_ x _-< 1, 0
T and that u is six times continuously differentiable in the closed region.

=<

Let

where Ax

(3)

1IN. Let

/y

gi,,n

[Ag,j

(gi+l,,,

2gi,,,,

(g,j+,,

2g,,,

gi_,,,)/(Ax)

g,_l,,)/(Ay) 2.

Replace the differential system (1) by the following finite difference approximation:

(4)

+ yt Wi,,2n-I
/%,,+ + W,,n+
n Wi,j,2n

W,,

(n > 0),
(n >-- 0),
0 or N).

(w, j,2n-l wi, ,:,)/At

(W,,+ W,,:+l)//t
u,.,
(i 0 or N, j

2. Difference equation for truncation error. Let us develop the difference

equation for the truncation error.
ti,j,2n

(5)

02Uii ,
Ox

02Ui,j ,2n+i
Ox

OaU (x)

1 03u /t
20t Ox

+ Ou (x),

04u/Ox is evaluated at ((i + O)Ax, jAx, 2nAt),

Ou/OtOx at (i/x, jAx, (2n )At), I1 < 1/2. Similarly,

where

< 1,

and

46

JIM DOUGLAS, JR.

and
Ui,j,2n+l

(7)

Ui,j,2n

At

Ou (At)2.
+ ,.1,
z -0-[r

Ou,i,+
Ot

As u is four times boundedly differentiable, if (2n -t- 1)At -< T,

A ua,2,,

(8)

+ vA ui,,,,+
g,,,=.

Ui,j,2n+l

Ui,j,2n

At

O(,t

gi,y,2n,

+ (/x)).

Similarly,
At

(9)

O(t + x)).
Let the truncation error be denoted by v,,.. Then,
(10)
u,,. w,,..
vi,,.
g,,.+

Combining (4), (8), and (9),

vij,o

O,
v,,.

v,,.+

X.

+ g,,.

(n

>= 0),

(11)

Av""+ + uAv"+
v,,.

v,,..+

v,,.+

At

(i

+ g,,.+

0orN, j

(n >- 0),
0orN).

This is the basic equation for the truncation error; it is our purpose to give
bounds on the size of vi,,. in terms of the length of the time step. Since
the difference equation (11) is linear, a superposition principle may be applied to simplify the analysis. Let*

(2)
where

v.

n--iv(k)
kO

is defined as below"

(n -< k),
(13)

* Yn

)i,i,n.

NUMERICAL INTEGRATION BY IMPLICIT METHODS

(k)
_2

()

y(k)
2

2+

(2n > k)

(14)

+ a v2+

v:.+

(12)

Direct substitution shows that

At
is a solution of (11).

3. Homogeneous system. Each of the functions v.__ satisfies the

homogeneous equation associated with (11) (or the analogue with x and
y interchanged) except that the initial condition is not identically zero.
Thus, the solution of the homogeneous equation with arbitrary initial
condition will enable us to solve (11). Let
0 or N),
(if i 0 or N, j
f, 0
f,
v,,0

v,,.+- v,,.

(n > 0),

v,,.+- v,,+
At

(n

(5)

x2yij2n+2

Y52yij2n+l

v,,.

Let
(16)

-- v

()
As

--4(Ax) sin
(18)
Thus,

a sin ax sin fy.

--4 sin

A sin ax

a:.

4(Ay)

-4(zXx) sin aAx

-if-- a,+
a2+1
a2n

(19)

(Ax)

Ay

sin

4(Ay)_. sin
At

-I- 4 (A)

(Ax)

1 "l- 4

(Ax)

an+l

(a.,+

a,)/ht

a,+

(a.+.

a,+l)/At

aAx
2

sin

fAy
2

At

At

sin ax,

sin

sin
a2n+2
a2n+l

0orN.

ao sin ax sin y,

f,.

and assume

0orN, j

ifi

0),

sin

A___y
2
2

48

JIM

JR.

DOUGLAS

Moreover,
(20)

Now, for arbitrary a and

the ratio a2,+l/a2,, or a2,,+/a.,,+l may be

larger than one; however, the ratio a,+/a2,, is always less than one;
i.e., stable. To prevent the intermediate ratio from becoming too large
when At, Ax, and Ay tend to zero, set

(21)

At

(Ax)

At

(5y)

X,

constant.

Then, the intermediate ratios are limited by X (assuming X larger than one,
as it will be for the desired large time step), as the numerator is no greater
than the larger of one and X 1 in magnitude and the denominator is greater
than one.
The solution with (16) as initial condition is

w_

ao

x
-t-Xsin--/

+Xsin _y]
2 /

-t- X sin

W X sin: -5-/

(22)

X sin:

sin ax sin

fax
2

-t- X sin thx

2

sin ax sin fly.

For arbitrary f.j vanishing on the boundary of the unit square,

(23)

f.

,= c sin rpx sin rqy,

cq

4
N2

N--1

where

(24)

N--1

Ef

i,i

d--

sin

qj
rpi sin-

y,

NUMERICAL INTEGRTI01 BY IMPLICIT METHODS

Hence, the general solution of (15) is

)2n

k sin

1+ sin

q\
2N

sinrpx sin rqy,

Cpq

q-

(25)

sin

/ 1 + sin

)2n+l

sin
sin

sin px sin -qy.

4. A convergence theorem. Equation (25) may be applied to obtain a

result guaranteeing the convergence of the solution of (4) to that of (1)
uniformly. Let
,(n)

(26)
Then,*

Vn+l

N-1

v.k)
n+l

q=*

_(2k)

(n)
P,q=I Gp,q

- -

sin rpx sin qy.

rp
X sin

sin

7v

n--k

X sin

rq

sin rpx sin qy

(27)

sin -px sin -qy.

Thus,
* Only vn+t will be discussed; the argument for v is similar.

50

JIM

DOUGLAS

JR.

)2n+l

(27)

To investigate the behavior of V2n.t_ aS N increases, it is necessary to

study the properties of the Fourier coefficients and also of the ratio
1
1

X sin rp
2N

+ X sin

2N

as functions of the indices p and q. The following lemma on the Fourier

coefficients is proved in the appendixLEMMA. Let f(x, y) be twice boundedly differentiable for 0
x
1,
0 <- y <- 1. Let

<=

c,q

lf( )

.=1

sin

.pi

--ff sin

.qj

(p, q

<=

1,

1).

Then, there exists a constant A > 0 depending only on the magnitudes of

f(x, y) and its first two derivatives such that

Ic,ql <- A

<__N 1-.
In particular, if f(x, y) contains (At) as a factor, then A does as well.
Note that there is no assumption that f(x, y) vanishes on the boundary;
hence, it applies to the coefficients C (k)
pq uniformly in k as u is six times con(k)
contains a factor (At) This follows
tinuously differentiable. Moreover, Cpq

ifl

<_ p,q

from the following lemma.

51

NUMERICAL INTEGRATION BY IMPLICIT METHODS

LEMMA. /f
Proof" Let g

>

g(x), V(0)

p(X)Y

AY

V(1)

0, and

> O,

then

< 0; for if not, t the first positive mximum,

--pv
g-- A2v > O,

0. First,

which contradicts the assumption that v were positive. At a negative

minimum,

p-(g-

p-g.

v)

Thus, the result follows for positive g. Similarly, for g 0, v 0 and (29)
is stisfied. The proof is completed by separating g into its positive nd
negative prts.
We see from (13) that

(30)

-.+,

v.+,(0)

g.,

.+,

where the second difference is taken in the x or y direction as n is odd or

even.

Hence,

As g.
(32)

0(t),
()
11"+

o((ht))

Now, applying the lemma on Fourier coefficients,

Cq

(33)

c,

(0 < p, q <

(At)

0((t) )

<

p or q $ N).

A bound for the stability ratio may be obtained as follows. Let

1
1

()
Then,
(35)
if X

(36)

sin z <

1.

h sin z
h sin z

+ 2 Zk=l (-- 1)k(k sin

Moreover, there exists

[ol <

Xsin

asz/2 < sinz < zfor0 < z < r/2.

z0

Z)

<

2X sin z

such that if 0

z < 1 X-,
Z

<z<

+ 2X
zo,

sin z,

52

JIM

For Zo

DOUGLAS

JR.

<= z <- -/2, there exists > 0 such that

(37)
A posteriori,

(38)

[pl_-<l--

Thus, if

rain

()

(i,),r/16),
(0 < x < /2)

For z

p/2N,

IPi -< 1--y.

(40)

Hence,

=< N1-)

(1 =p<

(41)

(N1-

=< p =< N- 1).

Let us consider the first sum in (28).

sin px sin -qy

k=,.O

k=O porq)N -1.

p,q-=l

<A(At)N ( 1

1)1(

+ A (At)
<

A_, (At).fN

Z I><Z

p orq>N I-

p,q=l

NOW,
(43)

__1

,.q- Pq

Thus, as At

(1

0(N-2),

)(log N) :,

k-----0

1--

:p orq>N 1-a

N-,

n--1

1
orq>N I-

< N

- -

NUMERICAL INTEGRATION BY IMPLICIT METHODS

(44)

2 ,=
2 c(2

=o

X sin

n--1 N--1

1 -k- X sin

n--k

Xsin-

1 -b X sin

53

n--k

sin px sin rqy

0 (N-** -b N-2(log N) -b N-+*).

The second sum in (28) may be approximated in the same manner with
the same result. Thus,

(45)

0(N-2+2

v.+

+ N-

N-2 (log N) ) (0 __< e __< 1).

The best result is obtained by choosing e so that the first two exponents
are equal. Thus e
1/2, and

<6>

O(y-/)

v.

O((At) /)

uniformly in 0 -<_ x, y =< 1, 0 -<_ _-< T. Collecting, we have the

THEOREM 1. If, in the region 0 <= x <- 1, 0 <- y <= 1, 0 <= <- T, there
exists a solution u(x, y, t) of the boundary value woblem (1) having continuous
derivatives of the sixth order, then the solution of the finite difference scheme
X constant),
(4) converges uniformly to u(x, y, t) as At --+ O, (At/(Ax)
with the error being O((At) m)
O(Ax4/3) at worst.
5. Refinement. To obtain the above result, the sums in (28) were broken
into two parts; a better bound can be calculated by breaking the sums
into many parts. Divide the set of index pairs (p, q) into the following
subsets:

I:0
(47)

<

p, q -<_ N -

(1 < p, q < N 1-i ,j= 2,

m),
(1 _-< p,q__< N- 1).

>N
I.+, p or q > N -

I:

p or q

If e0
1 and e,,+
forj
1, 2,...,m
n--1
k-O

p,q) el

0 are included, the general description for

1. Let

X sin

+ X sin

rp

I- holds

n--k

(48)
X sin

rq
sin -px sin -qy

-I-

X sin

From the lemma on Fourier coefficients and (41),

54

ffIM

DOUGLAS

JR.

(49)

(p, q)
<1

Thus,

_, (_

(50)

< A-I(At)N._ (,q)i

1<2 v=-

(51)

(52)

(p,q)(Ii

1
N2

< 3 (1

q= Pq

N-2.i(N2-2ti

1)
.) (log N)2,

.) (._

N2-2i-1).

Set
(j

0,.-.,m+ 1).

i-2+(2j/m+l)(N2j/m+l

N(2(i-1)/m+))t
1)N2/n+l.

(53)

m+l

Then,

[Si[__ AI-I(AO2N2-(2(-1)/m+1)

2(log N)

(54)

A_(At)=

--{-

{3(lOg(mN)2N2+(/+)+
1)

jN-s/+

+ (N

2/m+l

Summing,
m+l

]SI] _-< AV-(At)

j=l

(55)

< An-(At) (m
&S

3(log N)2N
1)(1 N-:("+))

N2+(21m+l)

55

NUMERICAL INTEGRATION BY IMPLICIT METHODS

m+l

jx <

(56)

jxJ=

(0

<= x < 1),

+1

+1

-y_yy y-- 1 1

(y 1).

The choice of m can be made anywhere between one and N. Take

(57)
Then, N+
1=1

(58)

log N.

-t- 1

e and

I.3"1 < A,-I(At) {3(1-

}N

e-2) -2 + e} N-2

lvAXv-1 {3(1

0(N-2).

The second sum in (28) can also be approximated as above. Thus,

(59)

O(N-)

v,

O(at)

T.
x,y
1,0
TEOnEM 2. Under the same assumptions as in Theorem 1, the truncation
error of method (4) is O(At) for any X
4ht/(hx)
uniformly in 0

CHAPTER II

6. Second approximation method. Again we shall treat the boundary

value problem (1). Consider as a second implicit scheme for integrating (1)
the backwards difference equation:
w,,+x

A.

(60)
|
wi,.

wi,i,.

(n >_-- 0),
(on boundary).

u,.

Note that all of the Laplacian operator is approximated on the advanced

time line. As a consequence, the linear equations are considerably more
complex than in the first method.

7. Difference equation for truncation error. The solution u of (1) satisfies

a non-homogeneous difference equation analogous to (60). Itcan be obtained as follows.

xA Ui,j,nT1

yA2Ui,j,n+l

0
(al)
[

Ot

+0
20t

,,+

()

, ,

56

JIM

DOUGLAS

JR.

04u/Ox Ou/Oy

and 02u/Ot are evaluated at ((i -f- 01)Ax, jAx,

(n "4- 1)At, (iAx, (j "4- 2)Ax, (n "4- 1)At), and (iAx, jAx, (n -4- 03)At), re-

where

0il <

spectively, with

(62)
A Ui,i,n+l

(63)

1. Setting

0/0

(64)

4At

(5y)

yA Ui,j,n+l

, ,

assuming O/Ox
0 -< __< T. Set

4At

(Ax)

and

v,.,

X,

i,j,n+l

Ui,i,n

At

02/0
Ui,j,n

+ gi.j.,

go be bounded in 0

=<

x,

-< 1,

Wi,j,n.

Subtracting (60) from (63),

0
vi,j,n+

(65)

Vi,j,n_l

At

(on boundary).

A different method of attack based on [12] will be used on the truncation

error equation (65) from the harmonic analysis procedure for equation (11).
8. Bound for solution of (65). The following lemma will be useful in relating the magnitude of v+ to that of v..
LEMMA. If Vi. "4- vA vi, p (x, y)v, gi, 0 < x, y < 1, p > O,
and v
0 on the boundary, then
max

lv(x, Y)

<=

max

p(x, y)
Prtoor" Case 1. g <- O. First, v >= 0. Otherwise, there would exist (x, y)
at which v would be at a negative minimum. Then,
A v(x, y)
--p(x, y)v(x, y) g(x, y)
AV(x, y) <= O,
as the second differences are positive at a minimum. But the left hand side
must be positive; thus, a contradiction arises from the assumption of negative v. Therefore, v -> 0. Consider then a point (x, y) at which v has a
maximum"

Av(x, y) >= g(x, y),

as the second differences are negative. As p(x, y) > 0,
-p(x, y)v(x, y)

g(x, y)

v(x, y)

NUMERICAL INTEGRATION BY IMPLICIT METHODS

g(x, y)
p(x, y)"

v(x, y) <-

Thus, the conclusion holds for g __< 0.

Case 2. g >- O. In like manner, v -< 0 and -v(x, y) <__ g(x, y)/p(x, y) at
minimum.
Case 3. g
and v + -< 0,

g+ g-, where g+g- O, g+ >- O, g- <-_ O. Then, v

v- >= 0. It is easily seen that
v(x, Y)

max

--< max

Rewriting (65),

(66)
As At

h v+

>

v+x

0 and v vanishes on the boundary,

(67)

max

(68)
As iv01

[v.+

max

v. + At max

O(At), there exists A such that

Now, as g.

(69)

+ hv+

max

Iv.+,

max

iv.

+ A(St)

v+

v-,

+ g.

[.

0,
max

Iv. <= nA (St)

At(At),

and, consequently, v, tends to zero uniformly in the closed region 0 =< x,

y -< 1, 0 -< -< T as At tends to zero. Collecting, we have
THEOREM 3. If there exists a solution u of (1) in 0 <-_ x, y <-_ 1, 0 <- <= T
having bounded ux uuuuu and u then the solution of (60) converges uniformly to u in this region with the error being O(tAt).
The convergence for this method then is of the same order as that of
method I; it should be noted that the method of proof in this case allows
the use of a weaker hypothesis. However, when both methods apply, the
first method involves only O(N) operations per time step and the second
requires O(N 3) by standard iteration techniques or O(N log N) using an
iteration technique based on the first method itself [11].
The numerical stability of method II follows from the lemma as applied
to the homogeneous form of (66).
CHAPTER III

9. Third approximation method. As a third implicit method, let us consider a direct extension of the procedure (VIII) of the introduction.

58

-JIM

DOUGLAS

Fi,j

Wi,y,o

A Wi,j,n-4-1

(70)

A2Wi,j,n-{-1

JR.

xA Wi,j,n

A2Wi,,n

(on boundary).
u,,.
The inclusion of the Laplacian difference operator at the old time level
will reduce the elemental error due to truncation in the time direction without introducing any appreciable addition to the amount of computation
for the completion of one time step.
10. Difference equation for truncation error. Assuming that the solution
u is six times boundedly differentiable in 0 =< x, y =< 1, 0 -< =< T, it is
easy to see that
w,,.

j,+l

(71)

-4- A u i,j,+l A-

u,.,. -t-

A u,,
2 ti,j,n+

gi,j,n

(72)

+ (t) )

o((x3

At

Ui,j,n

At

o(t),

).

If we again let
we obtain the basic truncation error equation by subtracting
v,i,o

:A2v,,,,+ -[- ,A2v,,,+l -4- A2v,,,, 4- ,A2v,,,,

(74)

v,.,

vi,j,nq-1

(70) from (71)-

Vi,],n

gi,Y,n

(on boundary).

The appearance of the second differences on the known time level prevents the useful application of the method of analysis of Chapter II. We
must return to the harmonic analysis procedure of Chapter I.

11. Homogeneous system. As a preliminary to the solution of (74) by

means of superposition, we shall study the corresponding homogeneous

equation. Let

59

NUMERICAL INTEGRATION BY IMPLICIT METHODS

(75)

A2Yi,j,n+l

+ xA Vi,j,n +

A2Vi,j,n+l

v..,

0 on boundary),

(fi.

f.

vi..0

yA

Vi,,n

Yi,i,,,+I

vi,i,n

At

(on boundary).

As f can be expanded in a finite Fourier series, it is sufficient to consider

the special case:

(76)

Assume that

a0

sin ax sin

y.

- --

a= sin axi sin y.

Substituting into (75) and using (62), we obtain
(77)

vi,,

--a+ h sin
Hence,

a X sin

-4- sin

am

+ sin
cAx
sin2
+ sin

sin2

am/,

(78)

-+-

2(a,+- a).

-4- sin

v)

As h > 0, the stability ratio is less than one in magnitude, i.e., stable.
The solution of (75) under the initial condition (76) is

sin2 + sin

(79)

vn

a
1

+X(sinx

As

(80)

_1

Cvq
p,q=i

v)]

sinxsing.

N--1
E,q=l
Cpq sin -pxi sin rqy,

fie

(81) v,,.

+ sin

sin

i+(sin2

+sin

A +sin2rA )J
x

sin px sin qy.

12. Convergence eorem. To apply the results of the last section to (74),
v as a sum of functions satisfng homogeneous
equations. Let

it is necessary to express

(82)
where

=0

v(k)

60

JIM

DOUGLAS

JR.

(83)

To show that v as defined by (82) satisfies the difference equation (74),

let us substitute directly"
At
n--1

()
Yn+i
k=l

(84)

v,,+

=o

2 Yn+l

At

by (83). Thus, (74) is satisfied. Next, let us obtain a bound

lemma of section 8 may again be used.

(85)
Expand

max
-,+

l)

(k)
j

g,,
()
the
for+

O((at)).

at max

+,

=0

in its Fourier series"

(86)
Then,

.q_ c sin px sin rqy.

-1- sin
(87)

rqAxl]n--k--1
2 -_.

sin -px sin rqy.

+
Adding,
1

n--1

X sin pAx

Cpq
kO p,ql

1 "4-

sin

61

NUMERICAL INTEGRATION BY IMPLICIT METHODS

Under the assumption that u is six times continuously differentiable, g

is twice boundedly differentiable; thus, the lemma of section 4 on Fourier
coefficients is again applicable.

((p--1q--1 -t-- N-2) (At) 2) (1--<p,q=<

((At) )
(p or q > N-).

c(q)

(89)
Also,
1

(90)
1

(
+(

sin
sin

pAx
2

rpAx
2

qAx)
-t-sin-,rqAx)
2,
"t-

sin

)
4N

<

(1 <

P, q <

(p or q

8N,

N 1-)

> N 1-)

Hence,

Iv"<= A (At)

N.q=l (1___ _i_,l)=o1( )

_t..

1--

+ A(At) Z1

porq>N 1-

(91)
For

< A*(Nv

(92)

1
k=0

+ N:+)(At): A*h(N- + N-+)

o(g-)

O((At)/a).

It is obvious that the refinement procedure is identical here to that of

Chapter I; hence, we may state the final result
THEOREM 4. If there exists a solution u of (1) such that u is six times conT, then the solu1, 0
tinuously differentiable in the region 0 x, y
tion of (70) converges uniformly to u in this region .with the truncation error
being no worse than O(At).
Note that, while much more work is involved in the solution by this
method, no improvement in the convergence over that of the first method
was obtained. The additional work is O(N/log N).
APPENDIX

13. A lemma on Fourier coefficients. Consider a function f(x, y) which

is twice boundedly differentiable on the closed unit square. Its Fourier sine
coefficients are given by

(93)

Cq

] fo f(x, y)

sin rpx sin rqy dx dy

(p, q

1, 2,...).

62

JIM

DOUGLAS

JR.

It is easy to see by integration by parts first with respect to x and then with
respect to y that

The Fourier sine coefficients off(x, y) on the finite group

are given by

(95)

qi

vpi

fsinsin

(p,q

,,.-

[Ji=o X
N- 1).

1,

As the sum in (95) is an approximate for the integrM of (93),

be little different from c for small p, q. Let

(96)

p, q

c should

-.

Then,

L L f(x, y)sin zp sin zqy d dy

.+:,

as the

interand

-m
(98)

is

-+ +-++/(,

) sin

sin q

O(N- in the borderin strip left out and its area

(x, y) sin rpx sin qy dx dy

-m -m f

Le u onider ech of hese erms eprgely.

s vpx

sin zpx
=0

(99)

+ 0(-+),

sin

x
xi/

+ cos

(2n)!

(2n+ 1)I

x----

- -

63

NUMERICAL INTEGRATION BY IMPLICIT METHODS

Hence,
sin rpx sin rqy

=o

sin

rpi

=o

(2m)[
*

(oo)

( )m I +

(2m

vpi

cos

rqj

x-

0?T

=o

+ cos

+ 1)[

(+)

cos

(-)()

+sinNcos

rqjE(--1)Orp)
( )"]
(2n)

(--1)0rq)2" y

=o

sin

-N

sin

(m)

Thus,

+1/2f+1/2 ( J)sin-pxsinrqydxdy
f
(o)

.
X

sin

=o

(2m+ 1)

sin

=o

(2n

+ 1)!

(2 +

fsinsin

(2+

For p, q as required by (96),

(o2)

(e

+ 1),.

Therefore,

(o3)

+"[+" ( J)sinvpxsinpydxdy
1

Nf sin

rpi

sin

rqj

W 0(-).

64

JIM DOUGLAS, JR.

(104)

Hence,
(x, y)

J-I/r Ji-I/2v

f,

(105)

X
=0

cos

f
sin

sin rpx sin rqy dx dy

--0

(2n

(-1) (rq)
2
(2m+l) (2 + Wf

Collecting,

f+/v f(x,
-t r

(106)

Nfii
sin rpx sin rqy dx dy

(107)

N--1

N
for 0

(108)

< <

sin

1) terms in the sum of (97),

f(x, y)

cos

sin

--

y) sin rpx sin rqy dx dy

1

As there are (N

+ 1)!(2n + 3) (2N) +

rpi

fii sin

sin

rpi

qj

sin

+ O(N--).

rqj

+ 0(N-)

1; .i.e.,
c,q

C,q

+ O(N-").

Combining (108) and (94), we have proved the following

LEMMA. If f(x, y) is twice boundedly differentiable in 0 <- x, y <- 1 and
c,q is defined by (95), then, for p, q <__ N i-,

/
the constant depending on f(x, y) and its first two derivatives.

NUMERICAL INTEGRATION BY IMPLICIT METHODS

1.

2.

3.

4.

5.
6.

7.
8.
9.

10.
11.

12.

65

BIBLIOGRAPHY
G.
D.
W.
Bruce, H.,
Peaceman, H. H. Rachford, and J. D. Rice, Calculations of
unsteady-state gas flow through porous media, Journal of Petroleum Technology 198(1953) pp. 79-91.
Courant, R., K. Friedrichs, and H. Lewy, ber die Partiellen Differenzengleichungender Mathematischen Physik, Math. Ann. 100(1928) pp. 32-74.
Crank, J., and P. Nicholson, A practical method for numerical integration of solutions of partial differential equations of the heat-conduction type, Proc.
Camb. Phil. Soc. 43(1947) pp. 50-67.
Hartree, S., and J. R. Womersley, A method for the numerical or mechanical solution of certain types of partial differential equations, Proc. Royal Soc. London, Series A, 161(1937) pp. 353-366.
Hildebrand, F. B., On the convergence of numerical solutions of the heat-flow equation, J. Math. and Physics 31(1952) pp. 35-41.
John, F., On integration of parabolic equations by difference methods, Comm. on
Pure and App. Math. 5(1952) pp. 155-211.
Juncosa, M. L. and D. M. Young, On the order of convergence of solutions of a difference equation to a solution of the diffusion equation, J. Soc. Indust. Appl.
Math. 1(1953) pp. 111-135.
Laasonen, P., ber eine methode zur Losung der Wrmeleitungsgleichung, Acta
Math. 81 (1949) pp. 309-317.
Leutert, W., On the convergence of approximate solutions of the heat equation to
the exact solution, Proc. Amer. Math. Soc. 2(1951) pp. 433-439.
OBrien, G. G., M. A. Hyman, and S. Kaplan, A study of the numerical solution
of partial diffeiential equations, J. Math. and Physics 29(1951) pp. 223-251.
Peaceman, D. W., and H. H. Rachford, Jr., The numerical solution of parabolic
and elliptic differential equations, this ournal, 3(1955), pp. 28-41.
Rothe, E., Zweidimensionale parabolische Randweraufgaben als Grenzfall eindimensionaler Randwertaufgaben, Math. Ann. 102(1930) pp. 650-670.

HUMBLE OIL AND REFINING COMPANY

HOUSTON, TEXAS

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.