Scribd
Upload
554 views

An Improved Pegasus Method For Root Finding : Abstract

This document presents an improved method for finding the root of a non-linear equation based on the Pegasus root-finding procedure. The improved method eliminates one or two slower substeps of the Pegasus method by interchanging points when a root has been crossed. This results in higher asymptotic convergence rates of either 3 function evaluations per cycle when the root is crossed after the first step, or 5 function evaluations per cycle when two steps are needed. The improved method is more computationally efficient than both the Pegasus method and Snyder's method. Numerical tests confirm the predicted convergence behavior.

Uploaded by

Anonymous F1UlyJbqQ
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
554 views

An Improved Pegasus Method For Root Finding : Abstract

This document presents an improved method for finding the root of a non-linear equation based on the Pegasus root-finding procedure. The improved method eliminates one or two slower substeps of the Pegasus method by interchanging points when a root has been crossed. This results in higher asymptotic convergence rates of either 3 function evaluations per cycle when the root is crossed after the first step, or 5 function evaluations per cycle when two steps are needed. The improved method is more computationally efficient than both the Pegasus method and Snyder's method. Numerical tests confirm the predicted convergence behavior.

Uploaded by

Anonymous F1UlyJbqQ
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

BIT 13 (1973), 423-427

AN IMPROVED PEGASUS METHOD


FOR ROOT FINDING*
RICHARD F. KING

Abstract.
A method of Regula Falsi type for finding a simple root of a non-linear equation
is presented. I t is similar to the Pegasus procedure, but exhibits a higher asymp.
tobic convergence efficiency. This is accomplished by eliminating one or two of
the slower Pegasus substeps.
CR Category: 5.1.
Key words and phrases: ]~egula Falsi, nonlinear equations, root finding, Pegasus
method, order of convergence, Snyder's method, Illinois method.

1. Introduction.
The classical Regula Falsi method finds a simple root of the non-linear
equation
(1) f(x) = 0

by repeated linear interpolation between the two current bracketing


estimates. There is a distinct shortcoming, however: one endpoint is
retained step after step whenever a concave or convex region of f has
been reached.
Recently two modifications aimed at overcoming this difficulty have
been analyzed by M. Dowell and P. J a r r a t t ([2], [3]). In these methods,
the ensured convergence of Regula Falsi is maintained while the asymp-
totic convergence rate is increased from linear to superlinear.
The first such modification, called the Snyder method here 1, halves
the offending (large) endpoint function value in each successive step
until the root has been crossed and a better endpoint found.
The second and faster modification, called the Pegasus method [3],
reduces this endpoint value by a fortuitous rational combination of
other current function values (see (5) below). I n the improved method

*Work performed under the auspices of the United States Atomic Energy Commission.
1 Referred to in [2]-[4] as the Illinois method. We cite reference [1] as justification for
the present terminology.
Received May 2, 1973. Revised July 27, 1973.
424 RICHARD F. KING

of this paper, as many as two of the slower Pegasus substeps are elimi-
nated b y timely interchange of points.

2. An improved Pegasus method.


Assume that a simple root 0 of (1) has been bracketed b y previous
estimates x 0 and xl. That is,

(2) fofx < O,


for f~ =f(x~), with f assumed to be continuous. The proposed method,
which is closely related to the Pegasus scheme b u t is asymptotically
faster, goes as follows:
(3) Perform a secant-type step to get x2,

fl
x2 = x 1 - (x o - xl) f o - f l '

and then calculate f2.


(4) If fl.f~ < 0 (so that fof2 > 0), then interchange (x0,f0) and (xx,fx).
(5) With the latest (xx,fl) and (x~,f2): if flf2 > O, then replace (Xo,fo)
b y (Xo,fofl/[fl+f2]) and (xl,fx) b y (x2,f2), use the expression in (3)
to get a new x2, and calculate f2.
(6) With the latest (xx,fl) and (x2,f2): if fxf~ < 0, then replace (Xo,fo)
b y (xx,fl) and (xl,fx) b y (x2,f2), and go to (3). Otherwise, go to (5).
This method differs from Pegasus in that never are two or more suc-
cessive secant-type steps allowed. In effect, it recognizes and takes ad-
vantage of the symmetry in (x0,fo) and (xl,fx) of the secant formula in (3)
[see also the asymptotic error formula (8) below]. In particular, if after
the very first step it happens that flf2 < 0, then the interchange of
(Xo,fo) and (xl,fl) prescribed b y (4) constitutes a more judicious labeling
of initial conditions. Later this interchange not only eliminates one of
Pegasus' secant steps, b u t m a y also (depending upon the particular
function f ) asymptotically preclude the need for a second successive step
of type (5). We shall see in the next section how this comes about.

3. Convergence rates and efficiency.


In determining asymptotic convergence rates, we follow much of the
notation and analysis of [3]. Define the error of iterate x r to be er = x r - 0,
and assume that fr can be expanded about the (simple) root 0:
AN IMPROVED P E G A S U S I%IETHOD F O R R O O T F I N D I N G 4~5

OO

(7) fr = ~ c~e/, where ci = f(~)(O)/j!.


j=l
W e presume t h a t x 0 a n d x 1 are near 0.
a. F o r t h e first (secant-type) step of a Pegasus cycle, we substitute
x r = e r + 0 a n d f ~ from (7) into the formula for x2 in (3). This leads to the
a s y m p t o t i c error formula

C2
(8) e~ ~ ~e0el-De0el(e0+el), where D = (c2/ci)~-c3/cl.

I n [2] a n d [3] this unmodified secant step is called one of t y p e U. W i t h -


o u t loss of generality we can presume t h a t flf2 > 0, so now the modified
formula of (5) is used. The error for this (type-M1) step turns out to be

\C1/

Thus e3 asymptotically has the same sign as e2, so a step of t y p e M 1


m u s t be followed b y a n o t h e r modified step (M2), whose error formula is

C2
(I0) e4 ~ -- s2e3 .
Cl

T h a t is, an M 2 step has the same leading error t e r m as a secant step.


A n d finally a U step completes the Pegasus cycle.
I n [3] it is shown t h a t this cycle, UM1M2U (or UUM1M~), gives
Pegasus a convergence rate of 7.275 at an expense of four function
evaluations per cycle. Thus its c o m p u t a t i o n a l efficiency is

~ . 2 7 5 -- 1.642,

compared with 1.449 for Snyder's m e t h o d (see [2]) a n d 1.000 for Regula
Falsi itself.
b. :For the proposed scheme (3)-(6), an unmodified first step U with
f~f~ < 0 is followed b y an interchange of (x0,f0) a n d (xl,fl) in step (4).
T h e n t h e n e x t step, (5), is a modified step t h a t we shall call of t y p e
MI*. I t has the error formula

\~31/

T h a t is to say, the error (I1) is t h a t of (9) w i t h e0 a n d ex interchanged.


If f~f~ > 0 after the first step, t h e n a Pegasus sequence is followed
through t h e n e x t U-type step. B u t now flf~ < 0, so an interchange is
426 R I C H A R D F. K I N G

prescribed and an MI* step performed. In either case, eventually each


cycle starts with U followed b y MI*.
B u t for a step MI*, the sign of ea differs from that of ~2 when D > 0.
Thus the root 0 has been crossed and one cycle of the sequence

(12) UMx* , U M I * . . . .

has been completed, with just two function evaluations. If D < 0, on the
other hand, then MI* must be followed b y another modified step M2*,
with the same leading error term (10) as M~. Furthermore this M2* step
completes the cycle. In this case, the procedure requires three function
evaluations for a sequence of cycles

(13) UMI*M2* , U M I * M 2 * ,. . . .

We exclude the special case D = 0.


To determine the convergence rate of (12), we observe that for one
full step
c2 D
E2 N __EOEI ~ E3 ~ D ~0E1E2 ~'~
~--- - - ~2 2
C1 C2 /C 1

while for the next step

c2 ~)
C1 C2/C 1 ?42 ,*~ Cl

Consequently the error formula for the full cycle is

(14) ~5 ~ \cl) o3

Thus for D > 0 the process has convergence of order 3 and an efficiency of
~f3-- 1.732.
A similar analysis for (13) gives
(15) ~7 ~= D ~
for the three-step cycle. Consequently for D < 0 this Pegasus modification
has a convergence rate of 5 and a computational efficiency of

~f5 -- 1.710.
In either case, the proposed method's efficiency is greater than the
1.642 of Pegasus.
A number of the functions tested in [2] and [3] have been tried with
AN IMPROVED PEGASUS METHOD FOR ROOT FINDING 427

the proposed algorithm. In all cases, the predicted errors (11) were in
good agreement with the numerical results. Furthermore, the asymptotic
sequences (12) and (13) were seen to occur.
Both the author and the referee have observed that the idea in (4) of
interchanging estimates to speed convergence can also be applied to
method "A" of Anderson and BjSrek [4]. The result is a method with
exactly the same convergence rates as the Improved Pegasus method.

4. Acknowledgements.
Brian Smith provided continuing help and encouragement during this
study, and P. Jarratt kindly made an early ts~peseript of [3] available
to the author.

REFERENCES
1. J. N. Snyder, Inverse inte~po~atlon, a real root o f f ( x ) = O , University of Illinois Digital
Computer Laboratory, ILLIAC I Library R o u t i n e H1-71 (1953), 4 pages.
2. M. Dowell a n d P. J a r r a t t , A modified Regula Falsi method for computing the root of an
equation, B I T 11 (1971), 168-174.
3. M. Dowell a n d P. J a r r a t t , The "Pegasus" method for computing the root of an equation,
B I T 12 (1972), 503-508.
4. N. Anderson a n d A. Bj6rck, A new high order method of Regula Falsi type for computing
a root of an equation, B I T 13 (1973), 253-264.

ARGONNE NATIONAL LABORATORY


9700 S O U T H CASS A V E N U E
ARGONNE, ILLINOIS 60439
USA

You might also like