0% found this document useful (0 votes)
30 views8 pages

Hwang 1973

This document summarizes the behavior of the recurrence relation xn+1 = xn(3N - xn3/2N) for finding the square root of N. It finds that: 1) If 0 < x0 < (3N)1/2, the sequence converges quadratically to N1/2. 2) If x0 > (5N)1/2, the sequence oscillates infinitely. 3) There is a increasing sequence βr converging to (5N)1/2 such that if βr < x0 < βr+1, the sequence converges to (-1)rN1/2.

Uploaded by

Florin
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
30 views8 pages

Hwang 1973

This document summarizes the behavior of the recurrence relation xn+1 = xn(3N - xn3/2N) for finding the square root of N. It finds that: 1) If 0 < x0 < (3N)1/2, the sequence converges quadratically to N1/2. 2) If x0 > (5N)1/2, the sequence oscillates infinitely. 3) There is a increasing sequence βr converging to (5N)1/2 such that if βr < x0 < βr+1, the sequence converges to (-1)rN1/2.

Uploaded by

Florin
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 8

JOURNAL OF APPROXIMATION THEORY 9, 299-306 (1973)

A Recurrence Relation for the Square Root

W. G. HWANG

Center for Applied Mathematics, Cornell University, Ithaca, New York 14850

AND

JOHN TODD*

Department of Mathematics, California Institute of Technology, Pasadena, California 91109


Received April 19, 1971

The behavior of the sequence x,+, = x,(3N - x,,3/2N is studied for N > 0 and
varying real x0. When 0 < x0 < (3N)l” the sequence converges quadratically
to wjz. When x0 > (5N)1/2 the sequence oscillates infinitely. There is an increasing
sequence & with /I-r = (3N)rle which converges to (5N)‘le and is such that when
& < x0 < jr+, the sequence {x,.1 converges to (--1)‘N”‘. For x0 = 0, ,!L1 ,
B0 9e-e the sequence converges to 0. For x0 = (5N)“2 the sequence oscillates:
x, = (-1)“(5N)l/“. The behavior for negative x, is obtained by symmetry.

1. INTRODUCTION

The recurrence relation


X n+1 = .mJ = &z(3N - &2)/2x (1)
which converges quadratically to N1/2-we shall assume N > O-was
popular in the days of computers without division instructions since it did
not involve division by a variable quantity as does the Newton-Raphson
relation
Yn+1 = dJJn> = CL + w7Jw, (2)
although (1) is slightly less rapidly convergent than (2). In fact if
E, = x, - WI2 we have
En+1 = - l ,“(x, + 2N1/3/2N - - •,~(3/(2Nl/~)),
* The preparation of this paper was supported in part by the National Science Founda-
tion. We are grateful to Dr. Oved Shisha for his comments on a preliminary version of
this paper.
299
Copyright 0 1973 by Academic Press,Inc.
AU rights of reproduction in any form reserved.
300 HWANG AND TODD

while if v?, = y,, - Nil2 we have

%+1 = %“/(2Ym) - %2(l/(2N1’2)).


It is well known that the behavior of the iterates of polynomial or rational
functions can be quite complicated. (See, e.g., Chaundy and Phillips [2] for
the behavior in the case x,+~ = axn2 + bx, + c and, for more general
results, De Bruijn [l] and Monte1 [4]. Stein and Ulam [6] and Metropolis
ef al. [3] discuss higher dimensional nonlinear transformations, partly
experimentally.)
The behavior in the case of (2) is simple: If 0 < x0 < N1/2 then x1 > N1/2
and {x,}: decreases to N1/2. If x,, = N112 then x, = N112. If N1/2 < x0
then (x,$’ decreases to N1i2. If x0 is negative we get convergence to -N1j2,
g being odd.
The behavior in the case of (1) summarized in the abstract, is moderately
complicated. In our discussion of (l), sincef(x) is odd, we may restrict our
attention to the case when x,, is positive.

2. THE CASE WHEN 0 < x0 d (3i19’/~

If x,, = 0 we have X, = 0 and x, -+ 0. If x0 = (3N)l/” then xl = x2 =


.., = 0 and we have x, -+ 0. If x,, = N112 then x, = N112. We need to deal
with 0 < x0 < (3N)‘/” only.
Note that if the sequence x, + 1 then we must have
I = 1(3N - 12)/2N
so that
I=0 or 1 = *NW*. (3)
We have
xn+1 - N1f2 = [xn(3N - x,3 - 2N312]/2N
= -(x, - N”2)2(xn + 2N1i2)/2N. (4)
This shows that convergence will be quadratic. Moreover, if x, is positive,
&a+1 - N1j2 is negative, i.e., x,+~ < N1i2. Again

&+1 - x, = x,(N - x,,~)/~N, (5)


which is positive if 0 < x, < N1i2, negative if x, > N1i2. If 0 < x0 < (3N)li2
then x, > 0 from (1) and from (4) we conclude that xl < N1/2. It follows
from (5) that x1 < x2 < a**. Since we have already noted that x,,+~ < N1/2
it follows that x, + N1j2.
SQUARE ROOT 301

These cases are illustrated in Fig. 1. If x0 = A, then x1 = B, x2 = C


while if x,, = D, then x1 = B, x2 = C.

y=x(JN-x’)/2N

FIGURE 1

3. THE CASE WHEN y = (5N)1/2 < x0

From (I), if x, = &y, then x~+~ = Ffr. Hence, when x0 = y, we have


x, = (- 1)“~ and the sequence oscillates finitely.
Take x0 > y. Then from (1)

x1 + x, = x,(5N - xo2)/2N < 0 (7)


so that x1 < -y. Also, since x1 is negative and xl2 > 5N,
x2 + x1 = x,(5N - x~~)~Z'I> 0. (8)

Subtracting (7) from (8) we find


x2 - x, > 0. (9)
Hence, the subsequence {x2,} is increasing; similarly, the subsequence {x~++~}
is decreasing. We shall show that neither can have a finite limit, and so the
sequence oscillates infinitely.
302 HWANG AND TODD

From (1) we find

x w-2 -- x,(3N - x,3(12N3 - xn2(3N - xn2)3/16N4. (IO)

If the subsequence {x2,} has a finite limit 1 then we must have

16N*Z = Z(3N - 12)(12N3 - i2(9N2 - 6N12 + Z*)),

which gives
19 - 9NP + 27N21s - 39N313 + 20N41 = 0.

The same is true for the subsequence (x~+~}.


This I quation has five real roots

1 = 0, 1 = &jjW 3 I = j-(5N)'12 (11)


and fnllr ml nplex ones which are the roots of

l* - 3N12 + 4N2 = 0.

Since, from (9), the absolute values of all the x, exceed x0 > y, none
of the limits in (11) is possible. Hence, the sequence oscillates infinitely.

4. THE CASE WHEN (3N)lf2 -=cx9 -c (5N)l12

This is where the complications are. We shall show that there is an increas-
ing sequence of numbers flV , r = - 1, 0, I,... satisfying (3N)ll” ,< & < (5N)1/2
and such that 8,. -+ (5N)lj2 which have the following properties:

if x,, = tBTthen x, -+ 0 and in fact x,+~ = x,+~ = ... = 0. (12)

if f17 < x,, < /37+1then x, + (: 1)’ N1j2. (13)

We begin by discussing the equation

h(x) = 3x3 - 3x - 28 = 0, (14)

which we can also write as

x(3x2 - 5) + 2(x - 0) = 0, (15)


SQUARE ROOT 303

when 1 < 8 < (5/3)1/2 + 1.2910. The graph of h(x) has the following form:

FIGURE 2

It is clear from the graph that for any such 0 the equation (14) has exactly
one real root x(0) which satisfies 8 < x(0) < (5/3)1/2.
We define a sequence 01, , n = - 1, 0, I,..., as follows: CL-~= 1 and for
n = -1, 0, 1, 2 ,..., CX,+~is the unique real root of

3x3 - 3x - 2ci, = 0. (16)


Clearly 01, is an increasing sequence bounded above by (5/3)1/2. Its limit 1,
therefore, satisfies
313 - 31- 21 = 0

and is, therefore, (5/3)l12. Convergence is ultimately geometric with a common


ratio of l/6 for we have

We can find the early al’s by use of a computer, or from the tables of
Salzer et al. [5]:

a-1 = 1, a,, = 1.2600, a1 = 1.2826, 01~= 1.2896,...


lim 01, = 1.2910.
304 HWANG AND TODD

We now define

fin = 43N)1/2, n = - 1, 0, l,... .

This means that Eq. (16), which defines 01,+~,

3U3n+1 -3301n+l - 201, = 0, (17)

can be rewritten as
%+I - 3l3m.,N = 2Wn, (18)
i.e.,
A = -.f@n+1). (19)
We shall establish the results (12) and (13) for r = - 1, 0, I,... by induc-
tion; however, reference to the diagrams will be helpful. For instance, if
Bo -=cxo = E -=cA , then -PO < x1 = f(xo) = F < -/I, and 0 < x2 =
f(x,) = G < B-I .
(a) If x0 = /3-I = (3N)li2, then, as already noted in Section 2, we
have x1 = x2 = .** = 0. Assume that (12) holds. Take x0 = /3T+1. Then
x1 =f(xo) =f@,+r) = -j$. by (19). The induction hypothesis applies and
gwes x~+~ = xv+4 = .. - = 0.
(b) If P-1 < xo < Bo 3 then 0 > x1 >f(flo) = -Be1 . Hence, since
f(x) is odd, we have from Section 2 that x, + -N1i2. Now assume (13)
holds. Take x0 such that j?r+l < x0 < /3,.+, . Then

f(Br+2) < Xl =f(xo) <f@r+A

i.e., by (19),

-ST+1 < x1 < --pT or BY < -x1 < Pr+1-


The induction hypothesis applied to x1 gives the limit (-1)’ N1f2 and the
fact that f is odd shows that for -x, the limit is
(-1) x (-l)r NV = (-l)r+l NW
as required.

5. ANOTHER RECURRENCE RELATION

The behavior of the recurrence relation

h+l = k3/[3~2 - ~1, (20)


SQUARE ROOT 305

which also converges quadratically to N112 can be discussed similarly, or it


can be read off from our results by observing that putting x, = N/yn in (1)
gives (20).

6. THE ORIGIN OF THE RELATIONS

All three relations are obtainable from Newton’s formula,

x n+1 = & - ~fhw’(x?z)l>

for suitable $ Specifically:

f(x) = I - Nx-2 gives (l),


f(x) = N - x2 gives (2)
f(x) = x3 - Nx gives (20).

7. OPTIMAL STARTING APPROXIMATIONS

There has recently been considerable activity in the discussion of optimal


starting values for square root algorithms. In the present context this means
determining a polynomial S(N) of assigned degree such that if we take
x,, = S(N) then the algorithm for N1i2 is optimal in an appropriate sense.
In particular Wilson [7] discusses this problem for the algorithm

W n+l = ~~(312 - (N/2) ~3, (1’)


which converges to N-‘j2, from which N1/2 can be obtained with one multi-
plication. Wilson notes that the optimal polynomials for (1) are got by
reversing those he obtained for (1’). Wilson does not discuss the regions of
convergence.

REFERENCES

1. N. G. DE BRUIJN, “Asymptotic Methods in Analysis,” North-Holland, Amsterdam,


1958.
2. T. W. CHAUNDY AND ERIC PHILLIPS, The convergence of sequences defined by
quadratic recurrence-formulae, Quart. J. Math. Oxford Ser. 7 (1936), 74-80.
3. N. METROPOLIS, M. L. STEIN, AND P. R. STEIN, Stable states of a non-linear transforma-
tion, Numer. Math. 10 (1967), 1-19.
4. P. MONTEL, “Lecons sur les recurrences et leurs applications,” Gauthier-Villars, Paris,
1957.
306 HWANG AND TODD

5. H. E. SALZER, C. H. RICHARDS, AND I. ARSHAM, “Tables for the Solution of Cubic


Equations,” McGraw-Hill, New York, 1958.
6. P. R. STEIN AND S. M. ULAM, Non-linear transformation studies on electronic computers,
Dissertationes Math. = Rozprawy Mat. 39 (1964), l-65.
7. M. W. Wnso~, Optimal starting approximations for generating square root for slow
or no divide, Comm. ACM 13 (1970), 559-560.

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy