mathematics of computation

volume 61,number 204

october 1993,pages 733-744

ON ENCLOSINGSIMPLE ROOTS OF NONLINEAREQUATIONS

G. ALEFELD, F. A. POTRA, AND YIXUN SHI

Abstract. In this paper we present two efficient algorithms for enclosing a

simple root of the nonlinear equation f(x) = 0 in the interval [a, b]. They
improve recent methods of Alefeld and Potra by achieving higher efficiency
indices and avoiding the solution of a quadratic equation per iteration. The
efficiency indices of our methods are 1.5537... and 1.618... , respectively.
We show that our second method is an optimal algorithm in some sense. Our
numerical experiments show that the two methods of the present paper compare
well with the above methods of Alefeld and Potra as well as efficient solvers of
Dekker, Brent, and Le. The second method in this paper has the best behavior
of all, especially when the termination tolerance is small.

1. Introduction
In a recent paper, Alefeld and Potra [2] proposed three efficient methods
for enclosing a simple zero x* of a continuous function f(x) in the interval
[a, b] provided that f(a)f(b) < 0. Starting with the initial enclosing interval
[a\ ,b\] = [a, b], the methods produce a sequence of intervals {[a„, bn]}™=l
such that
(1) x, £[an+i, bn+l]ç[a„, bn] ç ••■ç [ai, bi] = [a, b],

(2) lim (bn -an) = 0.

n—Kx

The asymptotic efficiency indices of each of those three methods, in the sense of
Ostrowski[10], are s/1 = 1.4142..., \ß = 1.5874... , and y/(3 + ^/U)/2 =
1.4892... , respectively. The numerical experiments in that paper show that
the practical behavior of those methods is comparable to that of the efficient
equation solvers of Dekker [6] and Brent [5], although they perform slightly
worse on some problems.
Although there are many enclosing methods for solving the equation
(3) f(x) = 0,
where f(x) is continuous on [a, b] and has a simple root x» in [a, b], most
of them do not have nice asymptotic convergence properties of the diameters
{(b„ -an)}^=l. For example, in case of Dekker's method, the diameters b„ - a„
may remain greater than a relative large positive quantity until the last iteration
Received by the editor June 26, 1991 and, in revised form, June 24, 1992.
1991 Mathematics Subject Classification. Primary 65H05.
©1993 American Mathematical Society
0025-5718/93 $1.00+ $.25 per page

733

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 734 G. ALEFELD,F. A. POTRA, AND YIXUN SHI

when a " ¿-step" is taken. In case of Le's Algorithm LZ4 of [8], the conver-
gence properties of {(bn - a„)}%L, have not been proved except that the total
number of function evaluations will be bounded by four times that needed by
the bisection method, which is also the upper bound of the number of function
evaluations required by our second method in this paper. For other examples,
like Brent's method, the Illinois method, the Anderson-Björck method, the Reg-
ula Falsi method, Snyder's method, the Pegasus method, and so on, only the
convergence rate of {|x„ -x»|}~ ,, where x„ is the current estimate of x», has
been studied and not the convergence rate of the diameters (b„ - an).
In case f(x) is convex on [a, b], the classical Newton-Fourier method [10,
p. 248], J. W. Schmidt's method [12], and the methods of Alefeld and Potra
[1] produce a sequence of enclosing intervals whose diameters are superlinearly
convergent to zero. The highest asymptotic efficiency index of those methods,
1.5537... , is attained by a method of J. W. Schmidt [12] and a slight modifi-
cation of this method due to Alefeld and Potra [1].
In the paper of Alefeld and Potra [2] three iterative methods are proposed that
produce enclosing intervals satisfying (1) and (2) without any convexity assump-
tions on /. Surprisingly enough, under appropriate smoothness assumptions,
one of the methods of [2] has the efficiency index 1.5874... , which is higher
than the efficiency index of the above-mentioned method of J. W. Schmidt [12].
In the present paper two new algorithms for enclosing zeros of nonconvex
functions are presented. Our first method requires at most 3, while our second
method requires at most 4 function evaluations per step. Both methods reduce
the length of the enclosing interval by at least one half at each step, so that in
the worst case scenario our methods require 3 times, respectively 4 times, more
function evaluations than the bisection method. As the bisection method, or
the methods of Brent [5], Dekker [6], or Le [8, 9], our methods are applicable
to rather general problems involving discontinuous functions and derivatives,
multiple zeros, etc. (see Theorem 3.1). However, in case of simple zeros of
C3-functions we can prove that, asymptotically, our first method requires only
2, and our second method only 3 function evaluations per step. Moreover, in
this case the sequence of diameters {(bn - an)}%>=lconverges to zero with R-
order at least 1 + \[2 = 2.414... for our first method, and 7?-order at least
2 + \/5 = 4.236... for our second method. Hence the corresponding efficiency
indicesare \/l + V2 = 1.5537... and y/ÏTTs = (1 + v%)/2 = 1.618... ,
respectively. As far as we know, the latter is the highest efficiency index for
iterative methods that produce monotone enclosing intervals for simple zeros
of sufficiently smooth functions.
This paper improves the results of [2] in two ways. First, by making better use
of available information, we obtain a higher efficiency index. Second, our new
algorithms do not use the exact solution of a quadratic equation at each step.
Instead, we use 2 or 3 Newton steps to get a convenient approximation. This
modification saves the work of computing the square root, makes the subroutine
program much simpler, and preserves the good convergence properties. For
convenience of comparison, we list the three algorithms of [2] in the Appendix
of this paper.
In our numerical experiments we compared our methods with the methods in
[2], with the methods of Dekker [6] and Brent [5] which are used in many stan-

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 ON ENCLOSINGSIMPLEROOTS 735

dard software packages, and also with the Algorithm LZ4 of Le [8]. The results
are presented in §5. The numerical results show that the two methods of the
present paper compare well with the other six methods. The second method in
this paper has the best behavior of all, especially when the termination tolerance
is small.
In §6, we show that in a certain sense our second method is an optimal
procedure.

2. Preliminary subroutines and lemmas

In this section we present some notations and results to be used later. We
assume throughout that f(x) is continuous on [a, b] and that f(a)f(b) < 0.
We consider a point c £ [a, b].
Subroutine bracket(a, b, c,a, b, d).
If f(c) = 0, then print c and stop;
If f(a)f(c) < 0, then ä = a, b = c, d = b;
If f(b)f(c) < 0, then a = c, b = b, d = a. a

After calling the above subroutine, we will have a new interval [ä, b] c [a, b]
with f(a)f(b) < 0. Furthermore, we will have a point d &[ä, b] such that if
d < â then f(a)f(d) > 0 ; otherwise f(d)f(b) > 0.
Subroutine Newton-Quadratic(a, b, d, r, k).
Set A = f[a,b,d], B = f[a,b];
If /I = 0, then r = a-B~x-f(a);
If A ' f(a) > 0, then r0 = a, else r0 = b;
For i = 1, 2, ... , k do:
(A) r-r P{ri~x) -r P^~l)
W r¡ - ri-i - snz—T - r*-i _
P'in.t) '-' B + A(2ri_l-a-b)'
r = rk. D

The above subroutine has a, b, d, and k as inputs and r as output. It is

assumed that d <¿[a, b] and that f(d)f(a) > 0 if d < a and f(d)f(b) > 0
if d > b. Furthermore, k is a positive integer and r is an approximation of
the unique zero z of the quadratic polynomial
P(x) = P(a,b, d)(x) = f(a) + f[a, b](x -a)+f[a,b, d](x - a)(x - b)
in [a, b], where f[a, b] = (f(b) - f(a))/(b - a), and f[a, b, d] =
(f[b, d] - f[a, b])l(d - a) ; note that, P(a) = f(a) and P(b) = f(b). Hence,
P(a)P(b) < 0.
Lemma 2.1. (i) Under the above assumptions, r £ (a, b).
(ii) Furthermore, if {a, b, d) ç [e, f], and if f(x) is twice continuously
differentiablein [e, f] with f'(x) ^ 0 for all x £[e, f] and
Ô = min |/'(x)| - (b - a) max |/"(x)| > 0,
e<x<f e<x<f
then
(5) \r-z\<XL(b-af, whereX=maXe^f/"{x)l, L = 2k-\.
2d

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 736 G. ALEFELD,F. A. POTRA, AND YIXUN SHI

Proof, (i) follows from the monotone convergence of Newton's method on

quadratics, while (ii) follows by remarking that
maxg<x</|/,/(s)|

and \P'(x)\ > ô > 0 for all x £ [a, b]. Indeed, we have that

(6) \rk-z\ = |r*_i - z|2—J^L < ^_, _ z\2k < XL\r0- zf,

where L= 1+2 + --- + 2*"1 =2k - 1. D

The next lemma can be proved in a straightforward manner; it will be needed
in §6.
Lemma 2.2. Let In = (n + Vl + n2)xl(n+X)for n = 1, 2, 3, ... ; then I2 > In
for all n/2.

3. Algorithms
In this section we present two algorithms for enclosing a simple zero x* of
a continuous function f(x) in [a, b] where f(a)f(b) < 0. These two algo-
rithms are improvements of the methods in [2]. The first algorithm requires at
most 3, and asymptotically 2, function evaluations per iteration, and the sec-
ond algorithm requires at most 4, and asymptotically 3, function evaluations
per iteration. Under certain assumptions the first algorithm has an asymptotic
efficiency index Vl-h/2 = 1.5537... and the second algorithm has an asymp-
totic index (1 + •v/5)/2= v2 + \/5 = 1.6180... . In the following algorithms,
p < 1 is a positive parameter which is usually chosen as p = 0.5.
Algorithm 1.
1.1 set ax =a, b\=b, C\ = ax - f[ait b\]-xf(a\);
1.2 call bracket(a\ ,b\,c\,a2,b2, d2) ;
For » = 2,3,..., do:
1.3 call Newton-Quadratic(an , b„ , d„ , c„ , 2) ;
1.4 call bracket(an ,_bn, c„, an, bn, d„) ;
1.5 if \f(ä~n)\ < \f(bn)\ > then set un — a„ , else set un = bn\
1.6 set cn = un - 2f[äj,, b~„]-xf(un) ;
1.7 if \c„ - u„\ > 0.5(b„ - an), then c„ = 0.5(b„ + an), else cn = cn ;
1.8 call bracket(a„, b„ , cn , än , b„ , d„) ;
1.9 if b„ - än <p(b„ - an), then an+\ = ân , b„+i = bn , dn+l = dn , else call
bracket(an , b„ , 0.5(â„ + b„), an+\, bn+\, dn+x). D

Algorithm 2.
2.1-2.2: same as 1.1-1.2;
For n = 2,3, ... , do :
2.3 call Newton-Quadratic(an , bn , dn , c„ , 2) ;
2.4 call bracket(an , bn , cn, an , bn , dn) ;
2.5 call Newton-Quadratic(än ,b„,d„,c„,3);
2.6 call bracket(an , bn, cn,an, b„, dn);
2.7-2.11: same as 1.5-1.9. D

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 ON ENCLOSINGSIMPLEROOTS 737

The following theorem is a basic property of the above two algorithms, whose
proof is straightforward and hence will be omitted.
Theorem 3.1. Let f be a real function defined on [a, b] suchthat f(a)f(b) < 0,
and consider one of the Algorithms 1 or 2. Then either a zero of f is found in
a finite number of steps, or an infinite sequence of intervals [an, bn] is produced
such that
f(an)f(bn)<\,
an < an+\ < b„+i < b„,
b„+\ -an+\ < \(b„ -an),
lim an = x* = lim bn ,
n—»oo n—»oo

f(x* - 0)/(x, + 0) < o. a

Corollary 3.2. Under the hypothesis of Theorem 3.1, assume that f is continu-
ous at x». Then x* is a zero of f. D

4. Convergence results
In §3 it is easy to see that the intervals {[a„, b„]}^=l produced by either
Algorithm 1 or Algorithm 2 satisfy that bn+\ - an+i < p\(bn - an) for n>2,
where p\ = max{/j, 0.5} . Since p\ < 1, this shows at least linear convergence.
In what follows we show that under certain smoothness assumptions, Algorithm
1 and Algorithm 2 produce intervals whose diameters {(b„ - an)}™=yconverge
to zero with 7?-orders at least 1 + \/2 = 2.414... and 2 + \/5 = 4.236... ,
respectively.
First, we have the following two lemmas.
Lemma 4.1 (Alefeld and Potra [2]). Assume that f is continuously differentiable
in [a, b] and f(a)f(b) < 0, and x» is a simple zero of f(x) in [a, b]. Suppose
that Algorithm 1 (or Algorithm 2) does not terminate after a finite number of
iterations. Then there is an «3 such that for all n > «3, the quantities cn and
u„ instep 1.6 (or in step 2.8) satisfy that
(7) f(Cn)f(Un) < 0.
Lemma 4.2. Under the assumptions of Lemma 4.1, also assume that f(x) is
three times continuously differentiable on [a, b] ; then
(i) for Algorithm 1, there are an r\ > 0 and an n\ such that for all n > n\
(8) \f{cn)\<rx{bn-a„)2(bn-x-an-ù,
where c„ is defined in step 1.3 ;
(ii) for Algorithm 2, there are an r2>0 and an n2 such that for all n > n2
(9) \f(cn)\<r2(bn-an)\bn-x-an-x),
where c„ is defined in step 2.5.
Proof. By Theorem 3.1, b„ - a„ -> 0 and x» £ (an , bn). Since x» is a simple
root, f'(x*) ^ 0. Therefore, when n is big enough, f'(x) ^ 0 for all x €
[a„ , b„]. For simplicity, we assume that f'(x) ¿ 0 for all x £ [a, b]. Also, it
is easy to see that in both algorithms we have that
bn-an< p(bn-{ - a„-i) < (bn-X - an-i).

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 738 G. ALEFELD,F. A. POTRA, AND YIXUN SHI

Since Xo —mina<x<¿, |/'(x)| > 0 and bn-a„ -> 0, then for any fixed 0 < ô < X0
there is an n\ such that for all n > n\ we have that b„ - a„ < 1 and

(10) S„ = X0- max \f"(x)\(bn - a„) >S > 0.

a<x<b

(i) For Algorithm 1, when n > n\, suppose zn is the unique zero of
P(a„ , b„ , d„)(x) in [a„ , bn]. Then using the error formula for Lagrange inter-
polation, we see that
1/(2,1)1<Xi\zn -a„\\z„ -b„\\z„ -d„\
< 0.25A,(ô„ - an)2(bn-x - «„_,), where A, = I max \f"(x)\.
"• a<x<b

By Lemma 2.1 and (10),

(12) \ 26n )
<X2(bn-an)2(bn-X-an„{), where¿2=(maXa^l/WI) .

Combining (11) and (12), we have that

\f(c„)\ < \f(z„)\ + ( max |/'(x)| ) \c„ - zn\ < n(b„ - an)2(b„-X - <%,_,),
\a<x<b J

where rx = 0.25Ai + X2maxa<^.<¿|/'(x)|.

(ii) For Algorithm 2, when n > n\, we have that

(13) \f(c„)\ < r{(b„ - an)2(bn-Y - an-i),

where cn is given by 2.3. Suppose z„ is the unique zero of P(an , bn , c„)(x) =

P(ä„ , b„ , dn)(x) in [ä„ , b„] ; then as in Alefeld and Potra [2], we deduce that
there is an n2 (we can choose n2 > «1) such that for all n > n2

(14) \f(¿n)\<X3(bn-an)2\f(c„)\, where X3= 2 (o.25^j .

Finally, similar to (12), by Lemma 2.1 and (10),

\d„-z„\ <X4(bn-ä„)s <X4(b„-a„)4(bn^ -a„_i),

(15) „.u„. , /maxû<;(<è|/"(x)|V
/maxa<
where A4= I-=
20
Combining (13), (14), and (15), when n > n2 > n\ , we get

(16) \f(cn)\ < \f(z„)\ + max |/'(x)| \c„ - z„\ < r2(bn - an)4(bn-l - aH-i)
a<x<b

with r2 = Xerx + X4maxa<x<ft |/'(x)|. D

The following two theorems show the asymptotic convergence properties of
Algorithm 1 and Algorithm 2, respectively.

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 ON ENCLOSINGSIMPLEROOTS 739

Theorem 4.3. Under the assumptions of Lemma 4.2, the sequence of diameters
{(bn-a,,)}^ produced by Algorithm 1 converges to zero, and there is an L\ > 0
such that
(17) bn+i -an+i <L{(bn-an)2(bn-x -a«-i), V« = 2, 3, ... .
Moreover, there is an N\ such that for all n> N\ we have
an+i = an and bn+x = b„.
Hence, when n > N\, Algorithm 1 requires only two function evaluations per
iteration.
Proof. As in the proof of Lemma 4.2, we assume without loss of generality that
f'(x) t¿ 0 for all x £ [a, b]. Take Nx such that TV-> max{«i, «3} . Then by
Lemma 4.1, (7) holds for all n > N\. For steps 1.6-1.8 of Algorithm 1 and the
fact that un,cn£ [a~n, bn] we deduce that
(18) b„-â„ < \c„-u„\, V«>/Vi.
From step 1.6 we also see that

(19) \c„- u„\= \2f[ä„, b„rlf(u„)\ < hf(un)\.

¿0

Finally, since cn £ {än, bn) , we have that \f(u„)\ < \f(cn)\ ■ Combining that
with (18) and (19), we have

(20) bn-â„<hf(c„)\, V»>/V,.

Ao
Now by Lemma 4.2, \f(cn)\ < rx(b„ - a„)2(bn-\ - an-\); hence
2
(21) b„-ân < -r-r,(¿?„ - a„)2(è„_i -a„_i), V« > Nx.

Since {(bn - an)}™=l converges to zero, if /Vi is large enough, then

b„ - ân < p(bn - an), V«>/Vi.
This shows that for all n > Ni we will have an+\ — ân and bn+\ = b„ . By
taking

L, > max (fr, {bn%~han+X\ J. n = 2,3,...,Nl,

[X0 (bn -a„Y(bn-\ -fl„_i)J
and using (21), we obtain (17). D

Corollary 4.4. Under the assumptions of Theorem 4.3, {sn)n<'=\= {(bn-an)}^Lx

converges to zero with an R-order at least 1 + V2 = 2.414... . Since, asymp-
totically, Algorithm 1 requires only two function evaluations per iteration, its
efficiencyindex is Vi-m/2 = 1.5537... .
Proof. By Theorem 4.3, {e«}^! converges to zero and e„+i < Lie2e„-i, for
« = 2,3,..., and the result follows by invoking Theorem 2.1 of [11]. D
Theorem 4.5. Under the assumptions of Lemma 4.2, the sequence of diameters
{(¿>„-a„)}£L. produced by Algorithm 2 converges to zero, andthereisan L2 > 0
such that
(22) 6„+i -an+x < L2(b„ - an)4(bn-i -an.x), V« = 2,3,....

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 740 G. ALEFELD,F. A. POTRA, AND YIXUN SHI

Moreover, there is an N2 such that for all n > N2 we have

an+x = an and bn+x - b„.

Hence, when n > N2, Algorithm 2 requires only three function evaluations per
iteration.
Proof. The proof is almost the same as that of Theorem 4.3. We assume that
f'(x) ^ 0 for all x £ [a, b]. Take N2 such that /V2> max{«2, "3} • Then,
when n > N2, as in the proof of Theorem 4.3, we have that

(23) bn-â„<hf(cn)\.
^0

Now by Lemma 4.2, \f(c„)\ < r2(b„ - a„)4(b„-i - an-\). Therefore,

2
(24) 6„-â„ < T-r2(è„-an)4(6„_i-a„_i), Vn > N2.
Arj

The rest of the proof is similar to the corresponding part of the proof of Theo-
rem 4.3 and is omitted. D
Corollary 4.6. Under the assumptions of Theorem 4.4, {£„}£!, = {(bn-an)}%L ¡
converges to zero with an R-order at least 2 + V5 = 4.236.... Since asymp-
totically, Algorithm 2 requires only three function evaluations per iteration, its
efficiencyindex is v2 + -\/5=1.618.... D

5. Numerical experiments
In this section we present some numerical experiments. We compared our
methods with the methods in [2], with the methods of Dekker [6] and Brent [5],
and also with the Algorithm LZ4 of Le [8]. In our experiments, the parameter
p in all the methods of this paper and [2] was chosen as 0.5. For Dekker's
method we translated the ALGOL 60 routine Zeroin presented in [6] into For-
tran; for Brent's method we simply used the Fortran routine Zero presented in
the Appendix of [5], while for the Algorithm LZ4 of Le we used his Fortran
code. The machine used was Encore-Multimax, and double precision was used.
The test problems are listed in Table 5.1. The termination criterion was the one
used by Brent [5], i.e.
(25) b-a<2-tole(a,b),
where [a, b] is the current enclosing interval, and

tole(a, b) = 2 • \u\ • macheps + toi.

Here, u £ {a, b} is such that \f(u)\ = min{\f(a)\, \f(b)\}, macheps is the

relative machine precision, which in our case is 2.2204460492504 x 10-16, and
toi is a user-given nonnegative number.
Owing to the above termination criterion, a natural modification of the sub-
routine bracket was employed in our implementations of all the methods in this
paper and in [2]. The modified subroutine is the following:

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 ON ENCLOSINGSIMPLEROOTS 741

Subroutine bracket(a ,b,c,a~,b) (or bracket(a, b, c,â~, b, d)).

Set ô = X • tole(a, b) for some user-given fixed A e (0, 1) (in our experi-
ments we took X = 0.7) ;
if b - a < 4ô , then set c - (a + b)/2, goto 10;
if c < a + 2ô , then set c = a + 23 , goto 10;
if c > b - 2ö , then set c = b - 26 , goto 10;
10 if f(c) = 0, then print c and terminate;
if f(a)f(c) < 0, then a = a, b = c, (d = b);
if f(b)f(c)<0, then a = c, b = b, (d = a);
calculate tolera, b) ;
if b - a < 2 • tole(a~, b), then terminate. D
In our experiments we tested all the problems listed in Table 5.1 with different
user-given to¿ (toi = 10-7, 10-10, 10-15, and 0). The total number of func-
tion evaluations in solving all the problems (145 cases) are listed in Table 5.2,
where BR and DE stand for Brent's method and Dekker's method, respectively,
and "unsolved" means a problem is not solved within 1000 iterations. From
there we see that our two methods compare well with the other six methods.
The second method in this paper has the best behavior of all, especially when
the termination tolerance is small. This reconfirms the fact that the efficiency

Table 5.1. Test problems

function f(x) [a,b] parameter
sin* —x/2 l*/2,n]
-2£ï,(2'-5)2/(*-'2)3 [a„, b„]
an=n2+ 10~9
b„ = {n+ l)2 - HT9 n ■ 1(1)19
[-9,31] a = -40, b = -1
a = -100, b = -2
a = -200, b = -3
x —a 0,5] a = 0.2, 1, n = 4(2)12
0.95, 4.05] a= 1, « = 8(2)14
sin* - 0.5 0, 1.5]
2xe-" - 2e- + 1 [0, 1] « = 1(1)5, 15, 20
[1 + (1 -n)2];c-(l -nx) [0. 1] 1, 2, 5, 10, 15, 20
x2 -(1 -x)" [0, 1] « = 1, 2, 5, 10, 15, 20
[1 + (1 -n)4]x-(l - nx)4 [0, 1] « = 1, 2,4, 5, 8, 15, 20
e~nx(x- \) + x" [0, 1] « = 1, 5, 10, 15, 20
(nx - !)/((« - 1)*) [0.01, 1] 2, 5, 15, 20

0 ifx = 0
[-1,4]
xe~ otherwise

Í ftft
ft( + sinx- 1) if;c>0
[-104, tt/2] n= 1(1)40
I* —n
20 otherwise

e- 1.859 if x >
« = 20(1)40
e(n+lW2xl03 _ , g59 ¡f x g [0i 2XHCÍJ [-104, 10"4] « = 100(100)1000
-0.859 ifx < 0

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 742 G. ALEFELD,F. A. POTRA, AND YIXUN SHI

Table 5.2. Total number of function evaluations in solving all

the problems listed in Table 5.1
toi Bl B2 B3 Alg.l Alg.2 BR DE LE
HT7 3139 2895 2580 2800 2604 2693 2658 2643
10-io 3447 2995 2773 2990 2708 2794 2819 2808
1 unsolved
10" 3672 3017 2948 3134 2746 2860 2955 2971
1 unsolved
3714 3041 3007 3137 2793 2873 2936 3025
4 unsolved 3025

Table 5.3. Total number of function evaluations in solving

x" = 0 in [-1, 10] for « = 3,5, 7,9, 19,25
toi Bl B2 B3 Alg.l Alg.2 BR DE LE BIS
10- 402 510 384 355 349 434 1340 185 174
10-i0 561 718 529 521 461 611 1987 237 234
10- 785 1034 721 757 746 867 2 unsolved 377 325
2219 2959 1793 2208 1830 2624 6 unsolved 1680 921

index is an asymptotic notion. In order to give an interesting example where

methods having higher efficiency index are outperformed by methods with lower
efficiency indices, we compare those methods with the bisection method, solving
the problem
x" = 0, « = 3,5,7,9,19,25
with the initial interval [a, b] = [-1, 10]. The results are listed in Table 5.3,
where BIS stands for the bisection method.
6. Discussion
We notice that our Algorithm 2 is an optimal procedure in the following
sense. It is clear that Algorithm 2 improves our Algorithm 1 by repeating 2.3-
2.4 in 2.5-2.6. If we repeat 2.3-2.4 a total of k times, then we get an algorithm
of the form
Algorithm 3
3.1-3.2: same as 2.1-2.2;
for n —2, 3, ... , do
3.3 call Newton-Quadratic (an , b„, dn, c„, 2) ;
3.4 call bracket(a„ ,bn,c„, anX),bnl], d„X));

3.2k + 1 call Newton-Quadratic^"1', bnk ^ , af~l) ,cn,k+l);

3.2k + 2 call bracket(a„k~l), bnk~l),c„,ä„,b„, d„) ;
3.2k + 3 - 3.2k + 1 : same as 2.7-2.11. D

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 ON ENCLOSINGSIMPLE ROOTS 743

It is clear that Algorithms 1 and 2 are special cases of Algorithm 3. Fur-

thermore, similar to Theorem 4.3 and Theorem 4.5, we see that for Algorithm
3,
(b„+\ -an+i) <Lk(bn-an)2k(bn-x -aH-\), n—2, 3,...,
for some Lk > 0. Hence, Algorithm 3 has an i?-order at least x = fc-hVl + k2,
which is the positive root of the equation t2-2kt-1=0. Since asymptotically,
Algorithm 3 requires k+1 function evaluations per iteration, its efficiency index
is Ik = (k + VtTTc2)1!^^ . By Lemma 2.2, Ik < I2, for all k ¿ 2. Therefore,
Algorithm 2 is the optimal choice.
7. Appendix
In what follows we list the three algorithms proposed in Alefeld and Potra
[2], assuming that f(x) is continuous on [a, b] and f(a)f(b) < 0. For con-
venience, we use the names Bl, B2, and B3 for the first, the second, and the
third method in [2], respectively.
Algorithm Bl
set ax = a, bx = b, for n = 1, 2,... do :
Bl.l c„ = an - f[an , bn]-xf(a„) ;
B1.2-B1.5: same as 1.4-1.7 in Algorithm 1 of this paper;
B1.6 call bracket(a~n,bn,cn,ân,bn);
B1.7 if b„-ân< p(bn - a„), then set an+l = an, bn+l = bn ;
else call bracket(â„ , b„ , 0.5(â„ + bn), an+x, bn+\). D
Algorithm B2
set ax = a, b\ = b, for n = 1, 2, ... do :
B2.1 c„ = a„ - f[an, b„]~xf(a„) ;^
B2.2 call bracket(an ,b„,c„,än, b„) ;
B2.3 c„ - the unique zero of P(a„ , b„ , c„)(x) in [ä„ , b„] ;
B2.4 call bracket(än , bn, c„, a~n,bn) ;
B2.5-B2.9: same as B1.3-B1.7. □
Algorithm B3
set a\ = a, b\ = b, for n = 1, 2, ... do :
B3.1 c„ = 0.5(a„ + bn) ;
B3.2-B3.6: same as B2.2-B2.6;
B3.7 call bracket(a~n, bn,c„, an+\, bn+i). D

Acknowledgment
We thank Dr. D. Le for kindly sending us the Fortran code of his Algorithm
LZ4, and an anonymous referee whose comments led to an improved presenta-
tion of our results.

Bibliography

1. G. Alefeld and F. A. Potra, On two higher order enclosing methods of J. W. Schmidt,

Z. Angew.Math. Mech. 68 (1988), 331-337.
2. _, Some efficient methods for enclosing simple zeroes of nonlinear equations, BIT 32
(1992), 334-344.

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use

 744 G. ALEFELD,F. A. POTRA, AND YIXUN SHI

3. Ned Anderson and Ake Björck, A new high order method of regula falsi type for computing
a root of an equation, BIT 13 (1973), 253-264.
4. Ken Atkinson, An introduction to numerical analysis, Wiley, New York, 1989.
5. R. P. Brent, Algorithms for minimization without derivatives, Prentice-Hall, Englewood
Cliffs,NJ, 1972.
6. T. J. Dekker, Finding a zero by means of successive linear interpolation, Constructive As-
pects of the Fundamental Theorem of Algebra (B. Dejon and P. Henrici, eds.), Wiley-
Interscience, London, 1969, pp. 37-48.
7. R. F. King, Methods without secant steps for finding a bracketed root, Computing 17 ( 1976),
49-57.
8. D. Le, An efficient derivative-free method for solving nonlinear equations, ACM Trans. Math.
Software 11 (1985), 250-262.
9. _, Three new rapidly convergent algorithms for finding a zero of a function, SIAM J. Sei.
Statist. Comput. 16 (1985), 193-208.
10. A. M. Ostrowski, Solution of equations in Banach spaces, Academic Press, New York, 1973.
11. F. A. Potra, On Q-order and R-order of convergence,J. Optim. Theory Appl. 63 (1989),
415-431.
12. J. W. Schmidt, Eingrenzung der Lösungen nichtlinearer Gleichungen mit höherer Konver-
genzgeschwindigkeit,Computing 8 (1971), 208-215.

Institut für Angewandte Mathematik, Universität Karlsruhe, Kaiserstrasse 12, W-

7500 Karlsruhe, Germany
Department of Mathematics, University of Iowa, Iowa City, Iowa 52242
Department of Mathematics and Computer Science, Bloomsburg University, Blooms-
burg, Pennsylvania 17815

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use