Chow 1962
Chow 1962
Chow 1962
1. Introduction
In earlier papers by MILNES and POTTS [1, 2] and CHOW and MILNES [3--7] a
method has been introduced for the numerical solution of partial differential
equations and associated boundary value problems by the contraction of the bound-
aries of the region over which the problem is defined. Although the technique has
been shown to apply to equations of hyperbolic and mixed types, special attention
has been given to LAPLACE'S equation which is of elliptic type. The simplicity
of this equation has permitted the illustration of the principles involved in the
construction of suitable approximating relations for the contraction process and
for the discussion of such topics as stability, convergence, accuracy and the
type of boundary conditions to which the method is applicable, without giving
rise to the complications necessarily involved when these same questions have
been discussed in a general way, for wider classes of partial differential equations.
In previous work, the contraction method has been applied almost without
exception to regions which have been circular, since the symmetry of such a
figure implies considerable simplification of the mathematical analysis and aids
in obtaining a more ready understanding of the fundamental principles under-
lying the method. In this paper, the contraction process is now generalized
to regions, the shape of which is subject only to quite mild constraints. The
applicability of the technique to problems of practical interest is thereby greatly
improved. This generalization is, however, presently restricted to LAPLACE'S
equation in the plane with particular emphasis upon the Dirichlet problem.
2. Proportional Coordinates
Consider a closed, bounded and simply connection region ~ in the plane,
Fig. t, which is referred originally to a polar coordinate r, {9 system. The boundary
of ~ will be assumed to be a Jordan curve defined by a relation of the form:
where f (0) nowhere vanishes and possesses second derivatives at all but a finite
number of points.
It will be further assumed that the pole can be chosen at such a point O,
interior to ~ , that the function l(O) is single-valued; that is, the radius vector
from 0 to the boundary ~go of ~ is assumed to intersect the boundary only once
for any value of 0 and to be nowhere tangent to the boundary.
Numer. Math. Bd. 4 15
210 TsE-Sux CHow and HAROLD W. ~{ILNE$:
= o, R = r/l (O), (2 3)
Let :
Ri --~ ( , (i = O, t, 2 . . . . ), (2.4)
where 0 < Q < I is a constant, then the corresponding sequence of contours:
{g~} so defined are similar figures in perspective from 0.
Fig. t
Fig. 2
of the grid system are taken to be (Fig. 2) the points of intersection (Qi, 0~)
of the radial lines (2.5) with the contours ~ i ' (2.4) and the value of the solution
u = u (R, zg) of LAPLACE'Sequation at these points will be denoted as u (Qi,tg,) -----ui, ~-
The angular indices are assumed to be reduced modulo N.
Solution of Laplace's Equation by Boundary Contraction 2t
3. Contraction
The contraction process has been described in detail in reference [2j but the
fundamental principles m a y be briefly reviewed. Consider the case of the Dirichlet
problem in which the boundary values are prescribed at the boundary of ~ .
This implies that the numerical solution is known initially at the nodal points
of g0. Suppose now an approximating scheme can be found which relates only
the unknown values on 51 to those on g0. The approximating relations m a y
be either implicit in character or explicit, more generally the former, but the
number of unknowns to be determined in finding the values on ~1 is, in any
event, limited by the number of grid points on 51 and does not involve any
other unknown values on the other contours. Thus, to determine the values
of the solution on ~1, it is required to solve at most N simultaneous equations.
After the solution has been determined on 51, the original set of values prescribed
on g0 is discarded and is replaced b y the newly computed data on c~1, giving
rise to a new boundary value problem on this contracted contour. The values
on ~ are now computed from those on ~1 and the process can be repeated as
m a n y times as desired. At any stage, the number of unknowns to be ascertained
is limited b y the number of grid points on a single contour rather than by the
total number of points on all contours. I t m a y be noted initially, simply on
an intuitive basis, t h a t the proportional coordinate system lends itself well to
this process since the calculations necessary to determine the values on c~2 from
~71 are likely to be similar to those required to evaluate those on ~1 from g0-
It m a y be mentioned that startling advantages are to be derived from the
contraction method in actual machine calculations. An improvement over the
classical approach is obtained, since the implicit relations of the usual method
cannot be solved for any group of values short of the total set, so that J •
simultaneous equations must be resolved at one time, where J denotes the total
number of contours of interest; the contraction method, on the other hand,
requires the solution of at most N simultaneous equations at any one time
and, as a consequence, the amount of calculation is tremendously reduced. With
suitable programming precautions accuracy ranging as high as six and seven
significant digits has been obtained in certain instances. For systems involving
several thousand nodal points, the machine time using the IBM 704 can be
conservatively estimated from five to ten minutes. Actual calculations that
have been performed for regions of different shapes and for a variety of boundary
conditions are discussed in w 7 of this paper.
4. Orthogonal Functions
Consider the complex variable
where the proportional coordinate system of the previous paragraph has now
been applied in an obvious way to the complex plane. Then:
V ~ u = V 2 u ( R , *9) = O, (4.4)
T h e functions {~0• (t, *9)} are seen to be linearly independent for otherwise there
exists a finite linear combination of t h e m such t h a t
K K
~,ak[/(O)]kcoskO+~,bk[/(O)JksinkO~O (0~0< 2n). (4.6)
k~0 k=l
T h e corresponding function:
K K
u(R,O)=~akRk[l(O)]kcoskO+ Y, bkRk[l(*9)~ksink*9, (4.7)
k=0 k=l
K K
u(r',O)=~,akr'kcoskO+~,b~r'ksinkO~O, (0~0< 2~). (4.8)
k=0 k=l
[lF II = F)*.
Solution of Laplace's Equation by Boundary Contraction 2t 3
The functions y~• tg) defined only on f0 b y (4.10), are readily extended to
the functions ~v• vq) on any bounded region of the R,,9 plane, except at
R = 0, b y the following recursive relationships
k k-1 0) '
i=o s=l
(k=t,2,3 .... ).
It will be noted t h a t the functions ~• va) are neither orthogonal or normal
for values of R ~ I . Since the recursive process is a finite one for a n y value
of k, it follows t h a t the functions ~• 0) satisfy (4.4), i.e., V ~ • v~)=0.
Consider the subspace H 2 of Hilbert space defined as the closure, according
to (4.9), of finite linear combinations of the functions ~• 0), ( k = 0 , t, 2 . . . . ).
Since the functions ~0• , 0) are linearly independent they form a basis for
H 2 and, moreover, so do the ~p• ~), ( k = 0 , 1, 2 . . . . ), as each ~0• v~) can
be expressed as a finite linear combination of the ip• va) and conversely.
Indeed, it is easily seen t h a t any function in the closure of the space generated
by the q~• (1, tg) is likewise in the closure of the space generated b y the ~• (I, va),
and conversely.
Now, let
u (1, O) = g (O), (4.t 2)
be a continuous function defined on % , then a function [8] u = u (r, O ) = u (R, ~)
exists, harmonic in ~ and tending to g(4) on % as a limit from the interior.
I t is known [9] that under the conditions t h a t have been imposed upon go
214 TsE-SuN CHOW a n d HAROLD W . MILNES:
It is easily seen that this series is convergent for values: 0 < R < = I , since the
sequence of harmonic functions
+K
uK = ~. a~v~(R, O), (4.t6)
--K
for each fixed value of K assumes its maximum absolute value for R : 1 and in
the limit as K-+oo, uK(l, 0)-+g(0) by (4.t3).
5. Numerical Solution
A numerical solution is inherently a finite process: -- data can be prescribed
only at a finite set of nodal points on c~o and only a finite number of arithmetical
operations can be performed in calculating a solution at some discrete set of
mesh points interior to ~ . It is therefore necessary to modify the analytic
solution (4.15) somewhat to conform to these conditions. The transition is made
by replacing the functions 9~:k (R, 0) with the vectors:
coskoo ]
[RE/(o)?coskOl /
(R)=1 cos /'
[R~ [! (0~_1)] ~ cos k 0~_~3
(5.~)
] Rk[[(O~176 ][
R k [[ (0i)]k sin k Oi /
q--k (R) = R ~ [[ (0~)]~ sin k 0~ ,
- R* I/(0N-i)] k sin k 0 N - i 3
and then b y replacing the product (4.9) with
N--1
--~ N Z •(0n} F2(On}' (5.2)
n=0
Solution of Laplace's Equation by Boundary Contraction 215
Based on the inner product (5.2) the orthonormalized vectors ~ : k (t) are calculated
from the ~o~:k(l) according to the G r a m - S c h m i d t process in a m a n n e r similar
to that outlined b y (4.t0) for the functions ~• ~9). In like manner, the new
vectors "-r
~• R ) are obtained from the -> ~/,• in accordance with (4At).
Both (4A0) and (4.1 t) express t h e ~ • (t) and the ~• (R) as linear combinations
of the ~• and the ~:~k(R), and the previously orthogonalized vectors; how-
ever, for facilitation of the ensuing discussion it is convenient to express the
~v•
-* and the +~v• as linear combinations of the ~~v• ) and the -+ ~• R ):
~)k(1)
-+ = bk,o ~~00 (~) -~ bk,1 ">
~01(~) -t'- b k,-1 "-->
~0-1(1) "~- "'" -t- bk,/~ ~~
~k (R) = bk, -~o (R) + bk,1 -~l (R) + bk, -1 -~-1 (R) + . . . + bk,k ~k (R) ,
(5.4)
(k=o, • . .
"3 2
I t will be shown later t h a t in the final formulas for c o m p u t a t i o n the coefficients
bk, z can be eliminated.
The numerical solution can then be expressed in vector notation, as:
(N--l}
2
k (N--l}
2
where:
ui'~ ] I g{}o) ]
(5.6)
I_% N - l ,
and, in accordance with (5.2) :
Consider the N x N matrices ~b(R), ~r/(R) of which the columns are the vectors
~• (R) and ~+k(R) respectively:
q~(R) = [~0 (R), ~1 (R), ~_1 (R) . . . . . ~ _ w - , (R)], (5.8)
2
and
~'(R) = [~o (R), ~1 (R), ~-1 (R) . . . . . ~_ N-1 (R)]. (5.9)
2
If B is the m a t r i x :
-boo 0 0 ... 0 I
I
Ib~o b~l 0 ...0 I
:o ..... I
/ b N-1 o b g - 1 b N-1 1 "'" b N-1 N--l|
L-T, - - V -'1 ~- ' - ~ ' g-J
then it follows from (5.4) t h a t :
T(R) = r B'. (5.tt)
2t6 TsE-SuN CHOW a n d HAROLD W . MILNES:
But by (5.7):
~ = 7-"(t) ~ = ~ ' 0 ) ~o, (5.t7)
and therefore by (5.15), (5.17) and (5.t4)
6. Stability and A c c u r a c y
The stability of the numerical solution (5.18) m a y be d e m o n s t r a t e d most
readily b y consideration of the eigenvalues of the m a t r i x M. F r o m (5.24) it
follows t h a t the eigenvectors of M are just the vectors ~ : k (t) and the correspond-
ing eigenvalues are ~lkl. Let v be a vector in the space generated b y the 9~k(l)
and let :
~ = ~oG0) + ~G(~) + ~-~-~0) + . . . + ~ u - ~ _ u--~ 0), (6.~)
2 2
then a norm relative to this basis m a y be defined in the usual w a y as:
U0 = C0~0(]) + C1~1(1) + C-1~-1(1) + C2~2(1) + ' ' " + C_ N~=I ~ _ N--I ( l ) , (6.3)
2 2
where the c~k are scalars, then from (5.23):
-7 = c 0-7
u, % 0 ) + Cl~ i -~
~01(1)+ c 4 ~ i -~
~_,(t)+
N-1 . (6.4)
+~2~'G(l)+...+c_N-~e ~ ~_~-1(t)
fl $
and:
-->
1I-,11=(4 + e2;+ c--1 ~)2j_~_ C2~47 ~_ ... _t_ C2 N--1 ~ (N-1)I
2 (6.5)
__<( 4 + c ~ + G + 4 + . . . + c 2 N~I)~= [lu011.
This relation GII
< II~0I1 holds independently of N, ~ and 1' so t h a t the process
is seen to be stable.
The intrinsic error of the solution (5.26) is v e r y small and results entirely
from the error due to the truncation of the analytic solution so as to satisfy
the requirements of the numerical process. Consider to begin with the basis
218 TsE-SuN CHOW and HAROLD V~7. MILNES:
functions (4.3) and assume that the boundary data g(tg) prescribed on (~0 is
equal to the value of one of these functions for R = t. Then the analytic solution
on ~. should be, according to the choice of sine or cosine functions:
u (e i, t~) = 9k (e j, t}) = e ki [! (t})] kj cos k ~, (k = 0, t, 2 , . . . ) ,
(6.6)
u(gi,~9) = 9_~(Qi, t}) = Q~i[f(tg)]~isink~9, (k = 1,2, 3 . . . . ).
But in this case the boundary d a t a is given b y the vector ~:Lk(t), which for
k--< N - - t is an eigenvector of the matrix M with the corresponding eigenvalue
-- 2
and this is identical with the analytic solution at the grid points of ~., provided
I kl __< (N- 1)/2.
I t follows from the above discussion and from the finiteness of the ortho-
gonalization process that if the boundary data can be decomposed in the form
g(tg):ao~Vo(l,tg)d-a~vl(t,tg)+a_l~v_l(t,tg)+...+a N-1 ~p_ N-1 (t, tg) (6.8)
2
t h e n the numerical solution, apart from computational error, is exact. If, how-
ever, ak ~ 0 for k > ( N - - ! ) / 2 then an intrinsic error arises from the truncation of
the series (4.t5). The restriction I kl ___(N--t)/2 is quite realistic since it is
difficult to fit the boundary data with fewer than 2 k mesh points if it contains
a component r ~ El (6})~kc~ k 0.
Computational error can arise from three causes: the finite word length of
the machine, the accumulation of round-off errors and the loss of significance
due to the order of arithmetical operations. These factors depend in a large
measure upon the machine being employed and upon the manner in which the
calculations are performed. Using the IBM 704 computer with the floating
decimal arithmetic and compiling the actual machine program in a straight-
forward manner with the F O R T R A N compiler, the authors have obtained solu-
tions of six or seven significant figures for moderate values of N. The results
of actual calculations are reported in the next paragraph and only some general
comment will be made here.
The principal source of error is in the orthogonalization process. Here care
must be taken to retain as much significance as possible, since an error committed
in calculating ~• (R, ~9) affects the accuracy in the calculation of ~3• (R, tg) with
Ill >1 k I. Regarding the choice of N, a balance is struck between too large a
value of N which generates an accumulation of computational error and too
small a value of N which produces an inherent truncation error. It m a y be
noted from (5.25) and the definition of W({}) that as ~-+1 the matrix hv(Q}-->~(1),
so that M tends to the identity matrix. Consequently, for values of ~ close
to unity, M is strongly diagonal; the off-diagonal elements m a y become so
small t h a t they lose significance due to the finite word length of the machine.
On the other hand, too small a value of ~ makes the off diagonal elements large
enough so that the inaccuracies in the data, which have previously been generated
due to round-off, are magnified and accumulated. However, with smaller e a
Solution of Laplace's Equation by Boundary Contraction 219
smaller accumulation of round-off error will result because fewer steps are required
to contract to a given R. Examination of numerous actual calculations indicates
t h a t empirically the o p t i m u m balance in both these cases seems to be a choice
of Q and N so t h a t :
2~ = A ~ . (6.9)
7. N u m e r i c a l Calculations
Numerous calculations have been made for regions of different shapes and
for various values of Q and N. As a basis for evaluation of the method, the d a t a
on c~0 given to the program was selected so as to correspond to the f u n d a m e n t a l
harmonic functions given b y the real p a r t of z~, i.e., u = r ~cos k O = R k [] (v~)] k cos kva,
for various values of k. The exact solutions at the mesh points of t h e grid were
c o m p u t e d and compared with the values of the solutions obtained b y the
220 TsE-SuN CHow a n d HAROLD W. MILNES:
TTT ~??
sss 0 0 0 O 0 0 0 0 0 0 0
0 b-oO 0 0 Oh
b-.(1 ~ Ox ~ (Dh
O1 O0 ' , ~ ',.00X ~
sss
XXX XX x XXX XXX XXX XXX
0
t'l
0
o ooo dd d
O1
ddd d d d ~ d d d ddd
I I i ~ I I
00
6 6 7 ?
I::, Ii ii o o
x x x xx x x x X X XX XXX XXX
..4 0
.-,
0
O'x t,..
0
t'q
0,1
b. 0
9 0
sss
XXX XX x x X X X X X X X X X X X
'..0 eq ,~"', 0 0 '~- ~'-~ 0 00x ",r
b.. q ",,O O0 ,~-,, 0 0
o'h t',q ~ Oh'.O
0O t'~ O ~t, ,b-. ,..~ '..O
,-t', Oh i"... O'x ' . ~ 1--. b-. e q ~'t rq
o0 9 61 aO '4~, q "~"~ 0
0 Ox e,1 1".,,. -,~., ..~ 00
o ooo dd d o d dd ddo o odo doo
0
~ I I I I I I I
O0 ~q
II II o o o o o o
x x x xx x x X X X X X x x x x x x
0 '-C, ',r t'~ ~ ',.~
1>,,. ',,,Q o ~ e,1 0", ~'-.
".~ ' , ~ r r'q 0 R ~ ~22
00 ,'~ 0 ~t', O0 '@ oO
0o 0 t",! O0 '4~ 61 ~-,-', 0
Oh t"q ~ ~ .-., aO
ood dd d ,5 6 o 6do 6d6 ddo
I I I I I I I I
~--a iI-~ ~--, u,,h
o0 0o ',,C, o0'.0
eq ~ eq
II It II II II II II II II li II II
Solution of Laplace's E q u a t i o n b y B o u n d a r y Contraction 22!
oo ~o~ s*s
X X X x XX XXX XXX XXX XXX
~ O~O0 0 ~ ~
O o O O O ~ ssg ~ O 9 O 9 1 6 9
X X X X X X XXX XXX X X X X X X
O Cq O Cq ~b-.. b...
0 ,~', o,~ ,~', em oO
~ " 0"~ 0 Gx 0 r
t-~ b-. i'-,- Ox Ox em
~ ~ 0 ~ ~ o ~
o ddd d,Sd ddd ddd odd odo
e,1
II II I I
d o d d d d
s II II ~%% ~ sO~ II II Ts II II
x x x x x x x xx xxx X X X X X X
0 ~ ~ e,1 0 b-.. O0
0 ~ 0 ~ 0 ~
.~'- 0", 0 ~"~ 0 Cq oO
01 .~.... I~... 0", 0", rm oo
ooo 00% O O O O O
XXX XXX X X X X X X X X X X X X
O ~ O O O O 01 b-- 0", O t',~ 0,1
O ~ O O ~ O 0",o90 ~". 0 ~0
o:: o:: o
a0 ',o
~u
II II II tl II II II II II II II II
222 TsE-SuN CHOW a n d HAROLD W'. MILNES:
. . sss as oso o o o
~ ~ 0 ~ 0 ~ O "~" ~.. ~- ~- ~
d o o d d d d d d d d d ddd ddd
I II li I I I
el
0 0 0 0 0
X X X X X X X XX XXX x x x x x x
~'~ ~ " ~1 r~'~ ~'~, (~
~ 0 ~
~ ~ 0 ~
II II ~ II II o o o o o o
X X X X X X X XX XXX XXX XXX
~ 0 ~ ~
0
~ " O0 ~', O0 ',~ O'x .,-, ~ ~g2 ~ ~ 0 ~
0 ~ o ~ o
oo ~O 00 ~,,O
II II II II II II II II I1 II II II
Solution of L a p l a c e ' s Equation by Boundary Contraction 223
$ $ ?s
XXX XX x x x X X X X XXX XXX
.q-
00 e,1
9 o',,
'0
it-,,, 0", eq
e,1
~ood oo 6 ~o d d odo s ooo
I n t I
do . . . . . . do . . . . . .
o So Soo t~ ~u SSS SS8
~ X X X XX x ~J ~ X x x x x x ~ X X X X X X
0
d 0", 61 O0
0",
ood Od d 0 dd oc5o doo o6o
I I I I I I
$$$ $$$ $$$ $$$
X X X X X X X X X XX X XXX XXX
0 ~o0 O~ 0o c"q
XX X X X x x XX XXX X X X X X X
G-, O0 u-', ~- C,1 ,r~
O 0 0 ~
~ ~ 0 ~
O ~ ~
0 0 a', 0 ".~ ',-
1>.0",0,1 O0 ~
0 '~ 0'~ ",O oO 0
G', ',-'
o o d od 6 o dd o o o d o o o d o
I I I II t I I
q~ o:: o::
~0
~..~0 ~ 0
nl IF II Ul II II Ii II II fl II It
224 7 u~z
!7
. . . . . . . . / a~e
~-~--t--~ I/
uo, o 7
~ L ~ !I ~<'~
ao,N-2
Fig. 3
~2
~aJ-2
Fig. 4
Fig. 5
TsE-Suz~ CHow and HAROLD W. MILNES: Solution of Laplace's Equation 225
~
digits, R > ( t - - e ) . For k----(N--l)/2
the accuracy dropped away to a b o u t
three decimal digits a n d for values of
.uI,o ~o,e
k > (N - - t)/2 the solution became m e a n i n g -
less as the contraction progressed inwards. R=7
A s u m m a r y of some of these results d
is given in the following tables. Here a
general outline of the region is indicated
a n d the exact a n d n u m e r i c a l solutions
are compared at certain points on the
c o n t o u r : ~., with a value of 7' chosen so Fig. 6
that R ~.5.
Acknowledgements. The authors wish to acknowledge their indebtedness to
Mrs. J. KRULL who was responsible for the numerical calculations that were per-
formed in connection with this work. Her capable handling of this aspect of the
endeavor has contributed to it very significantly.
References
[1] MILNES, H. W., and R. B. POTTS: Boundary Contraction Solution of Laplace's
Differential Equation. J. of the Association for Computing Machinery 6,
No. 2, 226--235 (1959).
[2] -- -- Numerical Solution of Partial Differential Equations b y Boundary Con-
traction. Quarterly of Applied Mathematics 18, No. 1, t - - t 3 (1960).
E3] CHOW, T. S , and H \V. MILNES: Boundary Contraction Solution of Laplace's
Differential Equation II. J. of the Association for Computing Machinery
7, No. t, 37--45 (1960).
[4] -- -- Numerical Solution of the Neumann and Mixed Boundary Value Problems
by Boundary Contraction. J. of the Association for Computing Machinery
8, No. 3, 336--358 (1961).
L@ - - Numerical Solution of a Class of Hyperbolic-Parabolic Partial Differential
Equations by Boundary Contraction. J. of the Society for Industrial and
Applied Mathematics 10, No. 1, 124-- t48 (1962).
[6] -- -- Boundary Contraction Method for Numerical Solution of Partial Dif-
ferential Equations: Part I. Quarterly of Applied Mathematics (to appear).
[7] -- -- Boundary Contraction Method for Numerical Solution of Partial Dif-
ferential Equations: Part II. Quarterly of Applied Mathematics (to appear).
E8] SmERNBERG,W. J., and T. L. SMITH: The Theory of Potential and Spherical
Harmonics, p. 291. Toronto: The University of Toronto Press 1952.
[9] W A L S H , J. L.." The Approximation of Harmonic Functions by Harmonic Poly-
nomials and by Harmonic Rational Functions. Bull. of the American Mathe-
matical Society 35, 499--544 (1929).
Boeing Airplane Company
Renton, Washington
and
Land-Air, Inc.
Pacific Missile Range
Box 48
Point Mugu, California
(Received July 25, 1961)
Numer. Math. Bd, 4 t6