-'

536 CH,o,p, 9 SOLUTToN on DtpnengruTrAL EeuATroNS

9.9 Finite-Difference Method

Methods involving difference quotient approxinations
for derrvatives can be used for
solving certain second-order boundary valLre ploblens.
Consider the linear equation
(l) x" : p(t)x'(t) + q(t)x(t) + r(t)
over fa, ôl with x(a) : a and x(b):
B. Forn a parrition of
a : tO < /1 <'.. < t1y : b, where h : (b _ a)lN andti : [a,bl jhforT _
Lrsing rhe points
cL + 0, -,
central-difference fbrmulas discLrssed in chapter 6
:"^'I. 113
tne oeftvatlves
are r-rsed to approximare

(2) x'(1.i):'t(ri'-r)-x(ri-r) -rO(h2)

' 2lt
and

(3) x"ç111 : x(ti-t) - 2xQi) -f x(ti- t)

, O(h27.
To start the derivation, we replace each term x(r7) on
the right side of (2) and (3)
with 17, and the resulting equations ale sr:bstitr-rted into (l)
to obtain the relation

!fI-2x1i'_.ti 1 t o(tt')
^.1 : /r.,,r_ r.,.1 "\
(4) -;r- P(tit l'+
\
r oth21\
_/l
/

-t c1lt.1)x1 i r(t1),
Next, we drop the two terms o(h2) i, (4) and intr.odr,rce the notation p7 : pG),
q.i : q (rr), and r j : r(ri)i this prodLrc s the cliffèrence eqnation

(5) ti-t-2rit\L-t I Ln.v,..

It2 t'.t
--n,r.lri-\'i. .r(li^j tt-i.
2h
which is used to compute nr-tmerical approximations to
the differential equation (1).
This is carried oLrt by multiplying each side of (5) by h2
and then collectirg terms
' involvin g x j
-t ' x i , and x 1a1 ana àr.ang ing them in a system of linear equations:
/ _ lt \
(6) -,)r,-r F(2. tr2t1i)t1r(!r,i-')",- t:-h2,j.
lî,,,
Sec. 9.9 FrnrrE-DTpIERENCE METHoD 537

for j : 1,2,., N- 1, where x0 : aand,rry : p. Thesystemin (6)has thefamiliar

tridiagonal folm, which is r.nore visible wlien displayed with rnatlix notation:

h,
/ + n-ql
^
!)

;pt - | x

Jn^-
2 rz
|
2 * h2qz Tpr- t o xz

,)
^zïn-qj
-h
2 vJ tPt-t xj
h.^
o -+DN-)
-h
- ) 2 + h2qN-z l UN-) -
t
r xN-2
-11
ïPN-t - t /+
^
t')
n-qN_l .ïN- I

-h2'r + eo
t)
-n-12
_hrrj
-h2, N -z
-h2ry-1 * e1y

where
/h \ /-h \
eo:lit, t t1" and ur-lipr-t+tlB
\Z ,/ \ L /
When computations wrth step size h are used, the numerical approximation to the
solution is a set of discrete points {(/;, x;)}; if the analytic solLrtion x(ri) is known, we
can compare x. j and x (t j ) ,

Example 9.18. Solve the boundary valr"re ploblen-r

..2t2
x"(t) : (t) +
4 px - U Ux(r)
I

with .r(0) : 1.25 and x(4) : -0,95 over the interval [0, 4l
The functions p, q,and r are p(t) : 2t l0 ) r2), q(t) : -21(l * t2), and r(t) : l,
respectively. The linite-diffel'ence nrethod is r.rsed to constrLrcl numelical solutions {-v;} us-
ing the system of eqr-rations (6). Sample values of the zrpploximations {xi t}, {x i,z}, {xi t},
and {x7,+) con'esponcLing to the step sizes ht : 0.2, hy : Q 1, hj :0.05, and hq :0.025
are given in Table 9.17. Figuu-e 9.26 shows the graph of the polygonal path formed from
{(t1, x1,t)} fol the case hr : 0 2. There are 4l terms in the sequence generated with
hz:0.1 , and the sequence {xJ.2} only inclirdes evely other tet'm from these computations;
they correspond to the 21 valr,res of {r;} given in Table 9.17. Similally, the seqLrences {xl.:}
and [x7.+] ale a poltion of the valires genelated with step sizes hr : 0.05 and h4: 0.025,
respectively, and they correspond to the 2l values of {ti} in Tabte 0.t2.
Next we compal'e numelical solutions in Table 9.17 with the analytic solirtion: x(t1 :
L25 + 0 486089652t - 2,2512 +2t arÇTan(r) - j lnç I l- t') + lt2 tnçt + 12 ). The numerical
538 CHp.p. 9 SolurroN oF DTFFERENTTAL EouArroNS

)t )
Table 9.17 Nr"rrnelical Approximations fbr v"ll) : -- r'(r) - .-
| +f' -r(l) + l
I +r-
rj, r
x j.2 ri3 j'l x(t 1)
h-0.2 /r-01 h:005 h :' 0.025 exzrct

00 r 250000 I.250000 1.250000 r.250000 r 250000

02 l 3r4s03 | 316646 | 3r1t74 t.311306 I 3 17350
04 | 320607 | 325045 1 326141 |.326414 1 326s05
0.6 1.272'7 55 | 219533 r.28t206 |.28t623 t,281'762
08 ) t77399 I l 86438 | | 88670 1 t89227 |
189412
l0 | 042t06 1 053226 | 0-5_s973 l 056658 r 056886
12 0.874878 0 887823 0 89 r023 0.89 r 82 l 0 892086
1.4 0 6837 12 0 698181 0 7017-58 0.702650 0 702947
l6 0 416312 0 49202'7 0 495900 0 49686s 0 49'7r8'7
1.8 0 260264 0,276'749 0 280828 0 281846 0 282184
20 0 042399 0 059343 0 063537 0 064s83 0 06493 1

2,2 -0 r70616 -0 r 53592

-0 149378
-0 148321 -0 1419't'7
2,4 -0 312551 -0.355841 -0.35 t702 -0.350669 -0 3-50325
26 -0 5s7565 -0 s4t546 -0 -537_580 -0 536590 -0 53626 I

28 -0 1201 t4 -0 705 r B8 -0 101492 -0 700570 -0 100262

30 -0 854988 -0,84 1 55 1
-0.838223 -0 837393 -0 8371 l6
32 -0 L)51250 0 945100 -0 942839 -0 942125 -0 94 r 888
1t1
-I 022221 - I 0t 2958 - | 010662 - r .01 0090 -r 009899
36 - | 04s4s'7 - I 038880 - | ,031250 - r 036844 -1 036'709
3.8 -r 02272'7 - | 019238 -1 018373 -r0l8l58 -l 018086
4.0 -0 950000 -0 950000 -0 950000 -0.950000 -0 950000

solutiorls can be shown to have error of order O0t2). Hence reducing the step size by a
factor of j resLrlts iu the errol being reduced by aboLrt ]. A calefirl scl'utiny of Table 9.18
will reveal that this is happening. For instance, ar ti : 1.0 the erlors incr"rn'ed with step
sizes/r1,h2,h3, and/r4are e1t:0 014780. e1.z:0 003660. ei.,:0 0009 13, ande14:
0 000228, respectively. Their successive latios e i.2le i.r - 0 003660 10.014180 - 0 2416.
e1tle1.z: 0 000913/0 003660 :0.2495, and e 1.+le i: : 0 000228/0 000913 :0.2491
are approaching ].
Finally, we show how Richalclson's improvemenf scheme can be used to extrapolate
the seemingly inaccurate seqLrences {x7. 1}, {x7 z}, {xi 3},.and {xr.+} and obtain six digits
of precision, Elirninate the ellol terms O(h2) ancl O((hl2)25 in the apploxirr-rations {x; r)
and{.xi.z} bygenerating-l-heexl-rapolaLedseqLrence{iir}: {(4x12-x1t)13} Similarly,
the ellor terms O((h/2)r.) ancl Ottlt1412 l for {r;,u } ancl {-r, j) are eliminatecl by generat-
ing {27 z) : {(4,r7..t - x1ù13|. It has been shown that the seconcl level of Richardson's
improvement schene applies to the sequences {zi 1} and {i; z}, so the third improvement
is {(162, 2- z j,t)115).Let Lrs illustrate the situation by finding the extrapolared values that
Sec. 9.9 FINITE-DIFFERENCE METHoD s39

0.0

-05

-1.0
Figure 9.26 The graph of the nnmerical approximation for
x(t1 : u(t) + w(t), which is the solution to

,,2t
x'(l): 2

r+t. - rx(r)
ta t" I I

(using /z :0.2).
-l'(r)

colrespond to /; : I 0. The lirst extrapolated valr-re is

4x;t - x; t 4(1.053226\ - 1.042106

3 - 3
-tns60'tr-...
- '.vJv/JL-4j,t.

The second extlapolated valLre is

4xj,3 - xj,z 4(1.055973) - 1.053226

Finally, the third extrapolation involves the terms z j.t and z j.2:
-:-:l'05688e
162;) - z.i | 16(1.056889) - 1.056932
: : I 056886.
15 15

This last computation contains six decirnal places of accuracy. The vah,res at the other
points are given in Table 9,19. r
Program 9,12 will call Program 9.11 to solve the tridiagonal system (6). Pro-
gram 9.12 requires that the coeflicient functions p(r), cl (r), and r(r) (boLrndary value
problem ( 1)) be saved in M-files p ,m, q.m, and r . m, respectively.

Program 9.11 (Tridiagonal Systems). To solve the tridiagonal system C.X : B,

where C is a tridiagonal matnx.

function X=trisys (A, D,C, B)

540 CHap. 9 SoLUrroN oF DTFFERENTTAL ÊeuArroNS

Table 9'18 Ert'ors in Numerical Approximations

Using rhe Finite-Differ.ence Method

x(tj)-xj,4
: j.4 e

lt1 :9 2 h2:01 h3 :
00 0.000000
9.95 114 :0.025
0 000000 0 000000
02 0 002841 0 000000
0.000704 0 000176
04 0 005898 0.00 | 460
0.000044
0 000364 0 00009 I
06 0 009007 0 002229 0 000556 0.000r39
0.8 0.012013 0.002914 0.000742
10 0 014780 0 0001 85
0 003660 0 000913
1.2 0 017208 0.000228
0.004263 0 001063
1.4 0.019235 0 000265
0.004766 0.001 189
].O 0 02081s 0.000297
0.005160 0.001281
l8 0.021920 0 005435
0 000322
0 0013s6 0 000338
2.0 0.022533 0 00-5588 0.001 394
2.2 0.022639 0.000348
0 0056t5 0.00 t40l
2.4 0.022232 0 000350
0 005516 0 00t31'7
z,o 0 02t304 0.000344
0 00s2Bs 0 001319
28 0.0i9852 0.000329
0.004926 0 001230
30 0 017872 0 000308
0 00443s 0 001 107
J.Z 0.015362 0.000277
0.0038 t2 0.0009s l
3.4 0.0t2322 0 000237
0 0030s9 0 000763
36 0 008749 0 000191
0 002t1) 0 000s4r
38 0 00464r 0 000135
0 001 152 0 000287
40 0 000000 0 000072
0 000000 0 000000 0 000000
Sec, 9.9 FrulrE-DTFpnRENCE METHOD 54:

Table 9'19 Extrapolation of the NLrmericar Approximarions

{.r;, ,}, {x 1 ù, {.xr,:r} obtained
with the Fi nite-Difference Methocl

4x j,2-r j.l l6z. j.2-z j.l x(t 1)

3 3 Exact
ti solntion
0.0 1 250000 | 250000 I 250000 1.250000
0.2 1.317360 r 3 t73-s l 1 311350 1.31'73s0
0.4 I 326524 | 326506 | 326s04 I 326505
0.6 1.281792 t 28tl64 ).281162 1,281762
08 I 189451 I t894t4 r t894r2 1 189412
10 r 056932 I 056889 1.056886 1.056886
T,2 0.892138 0.892090 0 892086 0.892086
1.4 0.703003 0'702951 0 70294'7 0.702948
l.o 0.497246 0.497191 0 497 t87 0 491187
1.8 0.282244 0 282188 0.282184 0.282184
2.0 0.064991 0 064935 0 064931 0 064931
2.2 -0 147918 t4'7913
2.4
-0. -0 147917
-0.t4797'r
-0 3-50268 -0 350322 -0 350325 -0 350325
26 -0 536201
28
-0.53 625 8
-0 53626r -0 536261
-0.7002)3 -0 700259
-0.100263 -0.'/00262
3.0 -0.837012 -0.837113 -0.837116 -0.837 1 16
3.2
-0 941850 -0 941885
-0.941888 -0 941888
34 - I .009870 -r 009898
- l 009899 - I 009899
3.6 -l 036688
-t 036107
- L036708 - 1.036708
3.8
- 1 01 807-5
- r.0 t8085
-r 018086
-l 018086
40 -0.950000 -0 9-50000
-0.950000 -0 950000