ANNALES SOCIETATIS MATHEMATICAE POLONAE Series I : COMMENTATIONES MATHEMATICAE X I V (1970) ROCZNIKI POLSKIEGO TO W AR ZYSTW A MATEMATYCZNEGO
Séria I : PRACE MATEMATYCZNE X I V (1970)
A. P l i s (Krakow)
Loss of uniqueness property in difference approximation of a Dirichlet problem
In this note we give an example of a Dirichlet problem for a (smooth) elliptic differential equation possessing a nniqne solution and admitting a difference approximation without uniqueness property. The differential equation is of the form
(
1) a ( x , у)ихх+Ъ(х, y)uxy+ c { x , y)uyy+ d { x , y)u =
0and satisfies the assumption of elipticity
(
2) 4a(x, y)c(x, у) — Ъ{х, у)* >
0as well as uniqueness conditions
(3) d ( x , y ) ^ 0, a ( x , y ) > 0.
Inequalities (2), (3) are classical conditions implying the uniqueness of solutions for any Dirichlet problem for (1).
Eeplacing derivatives in (
1) by suitable difference quotients gives a difference equation possessing non-vanishing solution of homogeneous Dirichlet problem. This example shows that the translation of uniqueness theorems from differential to difference equations need not be straight
forward.
We shall start with construction of a solution of the difference equation, then we define the difference equation and from it we obtain the differential equation.
N o t a t i o n s . We shall consider a set Z of nodal points Z =
i , j integers,
0< i < 3, \j\ < n}, where integer n >
1will be determined
later (to be sufficiently large). On the set of interior nodal points
8 = {(i, j ): i
= 1or
2, j integer |j| < n} the difference equation will
88 A. Plis
be satisfied. Difference quotients corresponding to the derivatives uxx, uxy, uyy will be denoted by ul{, u\{, u%
= ui+1,i — 2iUij' + ,
— \{ux+l,j+l — u% ~1,i+1— ui+1,j~1 + ,
= ui’j+1- 2 u iJ + uiJ- 1, for (i,j)€S.
Now we define a solution of difference equation. We define uli
= 0for (i, j ) e Z \ S ,
u 1* = + |j| < n
n U21 = \ + r,
W11 = r for |jj < n, j Ф
1, where r = (w-j-l)~3.
Obviously
(4) 0 < uv < 1 for ( i , j ) e S
and therefore (5)
W e have
Nix|, N
2 21 < 2 for
(6) i
till = %r for 0 < 1 j| < n,
(7) 10 2 2
0 > ul
°2= --- \-2r > --- ,
n n
(8) < = h
(9) < = l - 1
--- Ы + (j2—n2—2)r > 2r for \ j \ < n . j Ф 1 , n
(10) 21 1
Щ] — и — 9 — — — (n2-\-l)r > 2r for large n.
n
Now we define coefficients of the difference equation. We put a
1 0—
1, Ъ
1 0= — 9^ i — %nul 22 , c
1 0= n, d
1 0—
0. It is easy to see that for {i, j) = ( 1 , 0 ) the difference equation
(
1 1) + + + =
0is satisfied and in virtue of (5) and (7) we have the inequality
(
1 2) âaijcij— (bi])2 >
0,
Difference approximation o f \DiricMet problem
89
for large n, as well as the inequalities
(13) 0,
(14) aij > 0.
For i = 1, 0 < \j\ < n we put
a 11 = r , b1} = 0 , clj =
1, d 11 — — (ги]^-\- u^2) fu^.
Conditions (
1 1), (
1 2) and (14) are evident and (13) results from (5), (
6) and (4).
For i — 2, \j\ < n we put
a 21 —
1, b2j =
0, c 21 — r, d2] = — (%]{-{-ruli) I u2\
Similarly as in the previous case conditions (11), (
1 2) and (14) are evident. Condition (13) results from (9) or (10), (5) and (4).
From now on we consider integer n to he sufficiently large for (12),.
to be satisfied, and fixed.
We have constructed the coefficients al\ b%i, cli, dlJ satisfying (12), (13) and (14) for (i , j ) e S . Equation (11) has the non-zero solution ulf defined on Z and vanishing on Z \ 8 .
Now we define coefficients of (1) on nodal points. We put a { i , j ) = aij', b( i , j ) = Ьц , c { i , j ) = cij, d { i , j ) = dij
for i = 1 or 2, j being integer and |j| < n. Inequalities (
2) and (3) result from (12), (13) and (14). It is easy to see that the definition of functions a(æ, y), b(æ, y), c(æ, y) and d ( x ,y ) can be extended in a smooth manner on the whole rectangle R = {(x , y):
0< x < 3, \y\ ^ n} in such a way that inequalities (2) and (3) are satisfied on R.
If we suitably replace derivatives in (1) by difference quotients u\{) Uui u 22 and function u ( x , y ) as well as coefficients of (
1) by their values at the nodal points (x, y) eR, x, y being integers, we obtain back the difference equation (
1 1), and therefore equation (
1) has the claimed properties.
INSTITUTE OF MATHEMATICS OF THE POLISH ACADEMY OF SCIENCES INSTYTUT M ATEM ATYCZNY POLSKIEJ A K A D E M II N A U K