0; ie, the error tends to zero as ht -» 0 like Ch? for some constant
C for tx € io, b], We shall now state a resujt which tells us about how the
error €, tends to zero,
THEOREM 2.3 Iff is sufficiently differentiable, €, will satisfy
n= WP B (ta) +0 (hE) (2.11)
where 8 (1) is the solution of the ini value problem
8) = Ale, yO) BO
Let us denote »,, (i) the aie en of y (tn) calculated from (2.6)
with step 4. Then (2.11) can be written as
Yn (It) = y(t) AP 8 (tn) +0 (he) (2.13)
In conclusion, Taylor’s series method has advantages; it is easily derived in
* any order, and values of y (1) for 1 not on the grid are easily obtained, How-
ever, the method suffers from the time consumed in calculating the higher
derivatives.
+ PHN), 8 (to) = 0 (2.12)
2.3 RUNGE-KUTTA METHODS
We first explain the principle involved in the Runge-Kutta methods. By
the Mean-Value Theorem any solution of
L(ts §))s to) = Yor 1 E [oy 6]
satisfies 9" (tuts) = y(ta)-thy’ (En)
Vita) hf (ny ¥ (End)
tt nh OK On
h f
oui = self (tot Fe sot E flim on) @u4)
Alternatively, again using Euler's method, we proceed as follows:
a” (ot b )= $b" (Fy (nea)
7 LLC MAL Unb Seen 32 NUMERICAL SOLUTIONS.
and thus we have the approximation
yaa = Yet LLM Altes In FAO Dd (15)
Either (2.14) or (2.15) can be regarded as
Yati = Yah (average slope) (2.16)
This is the underlying idea of the Runge-Kutta approach. In general, we
find the slope at f, and at several other points: average these slopes, multiply
by A, and add the result to yn, Thus the Runge-Kutta method with v slopes
can be written as
Ki = hf lteter, Yok i ayKi)s er = OE = Vy Queers 9 (2.17)
or K, =hf(ta; ya)
Ky = hf (tat eah, yebon Ki)
Ky = hflturbes, yottay Ki tas: Ka)
Kg = hf (tut cqhy Yn aay Ki t@ag Kat aus Ks)
and Ines = Yat a wi Ky
where the parameters c2. C3, +1, Co, 42) +++) @etv_y) aNd wy are arbitrary,
From (2.16), we may interpret the increment function as the linear
combination of the slopes at f, and at several other points between ta and
ta4,..To obtain specific values for the parameters, we expand yn41 in powers
of f such that it agrees with the Taylor series expansion of the solution of
the differential equation to a specified number of terms,
Let us study this approach with just two slopes.
2.3.1 Second Order Methods
Define
Ky = hf(te yn)
Ky = hf(tetberh, yo an Ky)
and Yat, = Yow, Ki pw, Kz (2.18)
where the parameters c>, a1, ; and ware chosen to make y,+, closer to
Hines).
Now Taylor's series gives
2
I oes) = ¥ Ce) bhy" (10+ y(t
(2.19)
where
Sit {fe
Lut Uf firth AL LAG) SINGLESTEP METHODS 33
The values of y' (tn), y" (tn), «+ ate obtained by substituting ¢ = t,. We
expand X, and K, about the point (tn, yn).
Ki=hfe
Ky = hf (tee h, yw ay h fn)
= AE flte, Yn)-Hlez bh fb an hi fr fr
Fy (GIA Sich Den aa fF IE ooo]
= hifurt Heafitayfefe)+
FPG hic Rehr oh Pio
Substituting the values of K, and K in (2.18), we get
Yarn = Yok (re bwadh fob A? (wea fie waar fif)-+
F Wulf Restarhfirb a f3fylt (220)
Comparing (2.19) with (2.20) and matching coefficients of powers of h, we
obtain three equations for the parameters
Wy ti = 1
2 Wy = 1/2
ayy W2 = 1/2
From these equations, we see that if cz is chosen arbitrarily (nonzero), then
1 1
Ay = 62 55 1-5 (2.21)
Using (2.21) in (2.20), we get
2 13
Yast = InN GAGA + fut aft fehodt- (2.22)
Subtracting (2.22) from (2.19), we obtain the local truncation error
Ta = ¥ (tats) —Ynt
=P [(F-B)ter enter in) +t (seths )]e
2
= Hr2—3ea +302 frp] ten
We observe that no choice of the parameter c, will make the leading term
of 7» vanish for all f(t, y). The local truncation error depends not only on
derivatives of the solution y(r) but also on the function f. This is typical of
all the Runge-Kutta methods; in most other methods the truncation error
depends only on certain derivatives of y(t). Generally, cz would be chosen
between 0 and I. From the definition of the Runge-Kutta equations, we see 34 NUMERICAL SOLUTIONS
that any Runge-Kutta formula must reduce to a quadrature formula of the
same order or greater for,f(t, 3’) independent of y,, {wy} and fey) being, the
weights and abscissas of such formula. In the present case, cp = 1/2 and
¢z = 1 give the mid-point and the trapezoidal rule of integration for f(t. y)
independent of y. An alternative way of choosing the arbitrary parameters
isto produce zeros among iiy’s, where possible, to simplify the final formula,
The choice of c; = 1/2, for example, makes w, = 0. Sometimes the free
parameters are chosen either to have as large as possible the interval of
absolute stability or to minimize the sum of the absolute values of the
coefficients in the term Ty. Such a choice is called optimal. In the latter case
we define
atty rey.
aren < rp hia O12
We find
\fIYou might also like
