A. C. DOWNING, JR.

, Editor

Finding Zeros of a Polynomial mentioned furnish only one zero at a time; as soon as a
zero is found, it is divided out, and the method is started
By the Q-D Algorithm afresh with a new random start.
Other methods (such as the ones ascribed to Bernoulli
P. ~-IENRICI and Graeffe) do not require any first approximations, b u t
University of California,* Los Angeles, and Eidgenhssisehe then the convergence is usually slower and again only one
Technisehe Hochschule, Zurich, Switzerland (or two) zeros are found at a time.
I n this note we relate some experimental results cover-
BRUCE O. WATKINS
ing a technique for finding zeros based on the Quotient-
Utah State University, ~ Logan, Utah
Difference (Q-D) algorithm. This method was introduced
b y H. Rutishauser [5], and its theoretical aspects h a v e
since been discussed in a number of papers [6, 7, 8, 9]. Un-
A method which finds simultaneously all the zeros of a like any of the methods mentioned above, the Q-D al-
polynomial, developed by H. Rutishauser, has been tested on a gorithm has the advantage of furnishing simultaneous
number of polynomials with real coefficients. This slowly con- approximations to M1 zeros of a given polynomial, using
verging method (the Quotient-Difference (Q-D) algorithm) no information other t h a n the coefficients of the given
provides starting values for a Newton or a Bairstow algorithm polynomial. Theoretically the method could be used to
for more rapid convergence. determine the zeros of a polynomial with arbitrary ac-
Necessary and sufficient conditions for the existence of the curacy; however, since the method converges at least as
Q-D scheme are not completely known; however, failure may slowly as the Bernoulli method, and since it is subject to
occur when zeros have equal, or nearly equal magnitudes. propagation of rounding error, we use the Q-D method
Success was achieved, in most of the cases tried, with the only to determine first approximations to the zeros, which
failures usually traceable to the equal magnitude difficulty. then are determined more accurately b y the Newton or
In some cases, computer roundoff may result in errors which the Newton-Bairstow procedure.
spoil the scheme. Even if the Q-D algorithm does not give all T h e present paper is concerned solely with the com-
the zeros, it will usually find a majority of them. putational aspects of the algorithm. For the theory we
infer to the references given previously.
1. I n t r o d u c t i o n 2. T h e A l g o r i t h m
M a n y welLknown methods for determining the zeros Let
of a polynomial (such as the Muller method [1], Laguerre's p(z) = aoz~ + alz ~-1 + . . . -4- aN (1)
method [2], the Newton method or its variant for pairs
of complex conjugate zeros, the Bairstow method [3]) be a polynomial whose coefficients (real or complex) are
converge v e r y rapidly if good first approximations to the all different from zero. The Quotient-Difference Algorithm
zeros are known. If no such approximations are known, consists in constructing a two-dimensional array of num-
the convergence m a y be much slower or, in certain cases, bers, whose elements are arranged as indicated below ~
m a y not take place at all [4]. Furthermore, the methods in the special case N = 4 :
~o+ q~> q~ q2>
The preparation of this paper was sponsored in part by the
Office of Naval Research and the U.S. Army Research Office
(Durham). Reproduction in whole or in part is permitted for any q~'> qi: ql :> q~')
purpose of the United States Government. Part of the computing
time was supplied by the Western Data Processing Center, UCLA,
Los Angeles. Zeros were checked by a double precision program
developed by General Motors Corp., Allison Div., Indianapolis, etc.
Indiana.
* Numerical Analysis Research. 1 Our notation is adapted to the situation on hand; it does not
Department of Electrical Engineering. agree with the standard notation in [3] or [5].

570 Communications of the ACM Volume 8 / Number 9 / September, 1965

 The entries in this scheme are connected b y the relations (i) For every index k such that ] zk_~ ] > ] z~ ] > ] z~+~ I,
q<k) q(k)
,+1 = + (e(,k ) - e ~ k-~)) (k = 1,2, . . . , N ) , (2) lira q~k) = zk. (7)
k~av
~(k+l)
e(k)
,~+~ = e(nk) ~n+l
(k) (k = 1 , 2 , . . . , N - - l ) , (3) Thus, if zk is the only zero of modulus I zk 1, it can be found
qn+l
as the limit of the k-th q-column. Thus in particular, if
where n = 0 , 1, 2, . . . the moduli of all zeros are distinct (which happens, for
Since each of these relations involves four adjacent instance, if all zeros are real, positive, and simple), then
entries of the scheme, they can be used to generate the the limits of all q-colunms exist and are just the zeros, in
scheme row after row as soon as the first two rows and descending order.
the first and last column of the scheme are known. The
first two rows are determined from the coefficients of the (ii) For every k such that I zk ] > I zk+a I,
polynomial, as follows: lim e~k) = 0.
k~av
q0(1) _ a~ (qgk) = 0) (lc = 2, 3, . . . , N ) , (4) Thus, if one or several e-colulnns do not converge to zero,
a0
this is an indication t h a t several zeros have the same
e0(k) _ ak+l (k = 1 , 2 , . - . , N - - l ) . (5) modulus. Most frequently this occurs when a real poly-
ak nomial has a pair of complex conjugate zeros. Such zeros,
(Here we use the assumption t h a t the coefficients of p(z) too, can be recovered from the scheme, as follows:
do not vanish.) The first and the last columns of the method (iii) For every ]c such that
consist entirely of zeros:
IZk--ll > I Z k l - ~ - I Z k + i ] ~> [ Z k + 2 l,

e~°) = e~N) = 0 (n = 0, 1, 2, . . - ) . (6) the limits

Thus for instance for the polynomial lira rtq~(k) + ~~( k + i ) 1j = Ak (8a)
n~ao
p(z) = 128z 4 -- 256z 3 + 160z~ -- 32z + 1, lim q(.~1 ~/,
- (k+l) = Bk (8b)
n --~¢¢
the first six rows of the scheme are as follows:
exist and have the values zk + zk+~ and zkzk+l , respectively;
n rows
e0 q~X) e(1) q(2) e(2) q(a) e(a) q(4) e(4) thus z 2 -- Akz + B~ is the factor of p(z) with the zeros zk
2.000000 0.000000 0.000000 0.000000 and Zk+l.
0 --0.625000 --0.200000 --0.031250 0 Relations similar to (8) also exist when there are more
1.375000 0.425000 0.168750 0.031250
0 --0.193182 --0.079412 --0.005787 0 than two zeros of equal modulus, but they are more com-
1.181818 0.538770 0.242375 0.037037 plicated [3].
0 --0.088068 --0.035725 --0,000884 0
As computational checks, two relations m a y be used:
The q k converge to the zeros of the polynomial as n in-
creases. To five significant figures the zeros for this case (1) + (2) + . . . + G ~ ) = _ a~v-1
are 0.96194, 0.69134, 0.30866, 0.03806, and are obtained ahr (9a)
with n about 15. (n = 0,1,2...)

3. Convergence Properties q(1) ^(2) q(n~)N ~- ( _ _ i ) N ao

n+1 (//n+2 """
Clearly, the method as described above m a y fail to exist, a. (9b)
since one of the divisors ~/.+i
~(k) in (3) m a y turn out to be (n= 0, 1, 2 . . - ) .
zero. Necessary and sufficient conditions for the existence
of the scheme can be given [3, 5], but they are too involved 4. E x p e r i m e n t a l Results
to be of much use in practice. Suffice it to say t h a t the As indicated above, the convergence of the Q-D scheme
method, to the extent to which it is necessary to compute is only linear and therefore slow. Furthermore, since the
it, existed in the great majority of cases which have been scheme is not self-checking, it is subject to propagation of
tried. rounding error. For these reasons it would be unwise to
If the scheme exists, it has some remarkable convergence rely solely on the Q-D scheme for the determination of
properties. We number the zeros of p in descending mag- the zeros. Following a suggestion [3], we have written a
nitude. routine which uses the scheme to produce good starting
Iz~t _>_ tz2! > . . . . >_ [z~l values for the Newton or Bairstow methods, which are
applied to the complete (undisturbed) polynomial.
and set I z0 I = ~o, I z~+1 I = 0 for convenience. The fol- I t is of interest to compare this procedure with the v e r y
lowing convergence theorems then hold. elaborate programs described in [4]. There the zeros (or

V o l u m e 8 / N u m b e r 9 / S e p t e m b e r , 1965 C o m m u n i c a t i o n s o f t h e ACIVI 571

 Number
Problem Degree Coe ff~cienP Zeros or Factors 2 of Zeros
Found 3

TABLE I. POLYNOMIAL COEFFICIENTS, ZEROS AND Q-D PROGRAM ~ESIILTS

1 3 I, --i, 2, 5 --1.1325, 1.0662 2: il.8106 3

2 3 1, 0, 3, 1 --0.32219, 1.6109 ± il.7544 3
3 3* i, 1.0004, --1.0002, --1.0006 --1.00000, 1.00010, 1.00050 3
4 3* 1, 3.00006, 3.000120, 1.00006 --1.00000, --1.00001, 1.00005 3
5 3 1, 0, --3, 1 --1.8794, 1.5321, 0.34730 3
6 3* 1, 8, 21, 18 --3, --3, --2 3
7 4 128, --256, 160, --32, 1 0.96194, 0.69134, 0.30866, 0.03806 4
8 4 1, --8, 39, --62, 50 3 ±i4,12:il 4
9 4* 1, 7, 13, --3, --18 1, --2, --3, --3 4
10 4* 1, --6, 9, 4, --12 2, 2, 3, --1 4
11 4* 1, --5, 6, 4, --8 2, 2, 2, --1 4
12 4* 1, --8, 24, --32, 16 2, 2, 2, 2 4
13 4 1, --3, --14, 52, --48 4, --4.1273, 1.5637 2:i0.68002 4
14 4* 1, --3, --12, 52, --48 2, 2, 3, --4 4
15 4* 1, 3.05, 4.101, 3.10205, 1.05105 z~ z W 1.001, z W 1, z + 1.05 4
16 4 1, --3, 0, 52, --48 -3.2635, 0.95756, 2.6530 2:i2.8848 4
17 4 1, --3, --14, 0, --48 5.7098, -3.1068, 0.19852 ± il.6329 4
18 5 1, --6, 14, --16, --7, --30 3.682, - 0 . 5 ± i0.868, 1.65 2:i2.23 5
19 6 1, 0, 0, 0, 0, 1 1, - 1 , 0.33333 2: i0.86667, -0.33333 ± i0.86667 6
20 6* 1, --9, 45, 85, 34, 74, --100 --1,2,3±i4,1 ±il 6
21 6* 1, --7, 25, 25, --246, +422, --300 2, --3, 3 2: i4, 1 2: i l 6
22 6* 1, --13, 85, --305, 594, --622, 300 2, 3 , 3 2 : i 4 , 1 2 : i l 6
23 6* 1, --4, --5, 190, --666, 944, --600 2, - - 6 , 3 ± i 4 , 1 2 : i l 6
24 6* 1, --11, 65, --195, 314, --274, 100 2, 1 , 3 2 : i 4 , 1 2 : i l 6
25 7* 1, 4.870, --0.670, --0.15430003, --0.4265, --1.02113, z 2 + z + 1.01, z ~ -- z + 1.10, z ~- 1.0, z -- 1.13, 7
-- 2.48608, -- 6.2771496 z+5
26 7* 1, 507.9, --2079.899 9865.5087, --8504.8717, z ~ - 3z + 16, z 2 - z Jr 1.001, z + 1, z - - 1.1, 7
--2072.1159, 9864.8527, --9020.211 z ~ 512
27 10" 1, --2.5, --460.8, --9133.4, --50761.8, --88653.098, z ~ - 2z + 2 , z ~ + 2z-~ 2.2, z 2 + 20z + 200, z + 1, 10
-- 53510.4, -- 37313, -- 197170, -- 364800, -- 198000 z + 1.5, z + 5 , z - - 3
28 12" 1, 1288, 271787, 1337920, -1.5021189E8, 1.2664523E9, z ~ + 3z + 4, z 2 - 4z + 8, z 2 - 10z ~- 64,z-~ 1, 12
-5.2810183E9, -1.9582773E10, 9.621956E9, z + 2, z - - 16, z + 32, z + 256, z + 1024
1.9383320E10, - 3.1504466Ell, -- 8.6167780Ell,
-- 5.4975580Ell
29 13" 1, 128901, 273087, 1612424.9, -1.488606E8, z ~ + 3z-4- 4, z ~ - 4z + 8, z 2 - 10z + 64, z + 1, 13
1.1147382E9, --4.0019013E9, --2.49166E10, z A- 2, z -- 16, z -4- 32, z + 256, z + 1024,
--1.0156643E10, 2.9101496E10, --2.954675Ell, z + 1.01
- 1.1798729E12, -- 1.4200503E12, -- 5.5525336Ell
30 15" 1, 2, 334, --592, 36352, 60716, --1486310, -2191720, z z + 4z + 8, z ~ - 4z + 9, z -~ 1, z -- 2, z + 3, 15
23210431, 30731586, --169919420, --134375270, z + 4, z -- 5, z-4- 6, z-4- 7, z - b 10, z -- 11,
1.634846E9, 2.1045721E9, -5.7149106E9, z -- 12, z + 13
-- 6.2270208E9
31 18 1, 5, 10, 9, 50, 40, 30, 105, 98, 15, 81, 72, 60, 48, 36, --2.9028 2: il.8927, 0.73184 ± il.2325, 16
25, 130, 100, 520 0.50380 2: il.6390, --1.2458 2: i0.31540,
0.89442 2: i0.88178, --1.0386 2: i0.67193,
0.05192 2: i1.1350, --0.51543 2: i0.96378,
1.0210 2:i0.35144
32 19 1, 5, 10, 9, 50, 40, 30, 105, 98, 15, 81, 72, 60, 48, 36, --1.0507, --2.9028 ± il.8927, 1.0770 2: i0.32915, 15
25, 130, 100, 520, 672 --0.41890 2: il.0476, 0.50824 2: il.6396,
0.95083 2: i0.85507, 0.76888 ± il.2373,
--1.1915 ± i0.41270, 0.16098 ± i1.1971,
--0.92737 ± i0.72873

T A B L E 2. RANDOMPOLYNOMIAL COEFFICIENTS OF 5 DIGITS, ZEROS AND Q-I) PROGRAM RESULTS

1 6 53.081, --1429.7, 0.07721, --2.0143, 56.131, --561.12, 0.54839 ± i0.54964, 0.04197, --0.56935 ± i0.56935, 6
23.451 26.934
2 6 --2.9041, --6.9933, --8438.8, 23.804, --47.947, 0.15886, --0.018289, --0.068874 ± i0.15235, 6
36.539, 0.68540 --1.2054 ± i53.892
3 7 --48.773, 3.8121, --74.419, --116.96, 4.2614, --0.00035978 ± i0.020882, --0.79461 2: i2.8190, 7
--9915.8, --7.1330, --4.3250 2.2513 2: il.8147, --2.8345
(eontlnue~)

572 C o m m u n i c a t i o n s o f t h e ACM V o l u m e 8 / N u m b e r 9 / S e p t e m b e r , 1965

 Numb~
Problem Degree Coe~cients I Zeros or Factors 2 ~ Zeros
Found s

4 7 3.8121, --74.419, --116.96, 4.2614, --9915.8, 0.038108 :::l:i0.066060, 0.076218, 1.7322 ~ i4.5447, 7
--0.07133, --0.04325, --4.3936 --5.1782, 21.236
5 8 1, 2.9742, --6.7676, 6.2608, --54.215, --97.167, --4.90436, 2.77582, --1.18832, --0.690732, 8
7.0080, 7.7130, --15.806 0.198214 ~ i2.37181, 0.318483 ~ i0.385059
6 8 4.3936, --8.3619, 7.7182, --5545.9, --288.08, 3803.4, 1.2106, 0.35283 ~ i0.91569, --0.98342 d:::i0.54629, 8
105.40, --54.646, 8191.9 11.420, --4.7333 ::l=i9.3942
7 8 7.7182, --5545.9, --288.08, 3803.4, 105.40, --54.646, 0.14221, 1.2085, --0.14332, 0.35080 ::t::i0.92018, 8
8191.9, 8.7029, --167.00 --0.97998 ~/0.54791, 718.60
8 9 1, --993.56, --21.125, 122.20, 2.6288, --7.7351, 0.014036 ~ i0.11897, 0.99662 ~ i0.56000, 9
7.2159, --2184.9, 61.422, --31.350 --0.0080556 A:: i1.1198, --1.0132 ::t: i0.56143,
993.58
9 9 8.7029, --167.00, 463.33, 1126.1, 76.241, --7.0508, --0.31157, 0.029598 ± i0.20480, 1.4716, 9
4085.4, -- 1036.1, -- 99.729, -- 54.649
-- --0.39604 =l::il.3425, --1.9311, 5.6018, 15.091
10 9 463.33, 1126.1, 76.241, --7.0508, --4085.4, --1036.1, 0.16456 ± i0.18925, --0.36540, 9
99.729, -- 54.649, -- 2.9041, -- 6.9933
-- --0.10821 ~ i0.25013, 0.13911,
--0.40776 ~ il.4637, --2.7533
11 15 1, 39.247, --20.573, --8.3243, 22.834, --0.78440, --0.0082883, --1.2152, 1.1806 =t::i0.53560, 15
--4.2754, 504.15, --21.134, 72.874, 2.9240, --94.501, 0.36708 ~ il.2673, --0.70578 ~ i0.96858,
5.5945, 4.0532, 2549.3, 21.129 0.022019 ± il.4604, 1.3624 =t= i0.76961,
--1.3588 ::t::i0.74633, --39.759
12 18 1, --2.6027, 693.57, --6.5842, 1.6240, --6113.5, 0.74344 ::t=/0.66073, --0.98881, 0.97248 ~ i0.24402, 12
21.652, 4.5896, --0.71410, --7.8138, 8.3005, --0.58013 ::t=i0.82796, --0.11371 ::t=i0.99865,
37.504, --64.678, 4.9990, --6.8560, 45.808, 337.62, 0.35904 ::i=i0.92990, --0.88517 ::t=i0.45229, 2.0683,
14.910, -- 6093.8
-- --1.0373 ~ il.7847, 1.3029 ± i26.304
13 36 --9265.3, 6468.0, --42.015, 70.311, 3072.4, 2.9530, 0.81276 ~ i0.061480, 0.58720 =t= i0.48197, 22
5.6163, 870.73, --7.9141, --74.110, --22.964, 9.2252, 0.22083 ± /0.700000, --0.88460 ~ /0.30152,
--2.4987, --39.063, 6.5810, --6.8461, --7.8867, 0.63308 ::t::i0.69314, --0.20948 ~ /0.89996,
--32.151, --34.637, 67.916, --390.57, 60.247, 1.0530 ± i0.087703, --0.83729 ± i0.39630,
265.74, -- 453.86, -- 7015.6, -- 309.67, -- 2.0574, --0.072582 ± /0.99088, 0.57888 ::i=i0.77201,
--85.581, --99.394, --20.775, 49.225, 3924.5, --0.583111 ~ /0.78522, 0.15853 ~ il.0003,
--0.083830, 73.941, 0.049060, 88.312, --993.56 0.82646 ± i0.57735, --0.98285 ± i0.11248,
0.39262 ~ /0.92500, 0.95385 ::t: i0.36071,
--0.74665 ::l=i0.60412, --0.37723 ~ 0.89807
* Polynomials made up from selected zeros er factors. Otherwise polynemials made u p from selected coefficients.
1 Coefficients are listed in order of descending powers of the polynomial.
Zeros and coefficients are at least six significant figures for * polynomials. Otherwise these values are to five significant figures. Factors are i ndicated in the form (z+l), etc.,
otherwise the values given are zeros.
s See discussion on equal zeros.

pairs of zeros) are found b y the N e w t o n - B a i r s t o w tech- ure to pass the test if the count was less t h a n three; other-
nique using unsophisticated first guesses, and each zero wise the count was reduced b y one on failure to pass the
(or each quadratic factor) is divided out as soon as found. test. T h e n u m b e r of iterations, n, was an i n p u t value.
As a final step, when all the zeros have been found, the Experience showed t h a t an E P S value of a b o u t 0.01 and
original polynomial is used to produce i m p r o v e d zeros an n value of a b o u t 150 gave satisfactory results for m o s t
b y iteration. I t would seem that, considered as a m e t h o d polynomials.
for finding first approximations, the Q-D scheme is m o r e Besides t r y i n g the m e t h o d on a n u m b e r of fairly simple
economical. polynomials (Table 1) a n u m b e r of other polynomials were
Several methods for checking the convergence of the investigated (Table 2). These later polynomials were m a d e
e(k)-columns in the Q-D m e t h o d were tried. T w o success- up of coefficients formed b y using r a n d o m n u m b e r s of five
ful test expressions, either of which m a y be used, follow: digits f r o m [10]. T h e decimal was placed b y spinning a
pointer with six possible stopping positions, two positions
(2 l e£~) I/(I + I + I q.-1 + [)) being used to m o v e the decimal to the right one place,
-- E P S l e~k) ]/I e(k)
25 I -- E P S . two positions being used to m o v e the decimal to the right
two places, one position three places, and one position
E P S m u s t be chosen sufficiently small so t h a t the test four places. T h e sign was established b y a coin flip. I t is
expression used will remain positive until the kth e-column of interest to note t h a t the m a g n i t u d e of m a n y of the
has effectively converged. To further ensure such con- zeros of these polynomials tend to cluster a b o u t the value
vergence, a count of five or more passages of the test set one. F o r example, in one 18th degree polynomial, 13 zeros
an indicator I(k). This count was set back to zero on fail- h a v e m a g n i t u d e s differing f r o m one b y less t h a n 1.2 per

Volume 8 / Number 9 / September, 1965 C o m m u n i c a t i o n s of the ACM 573

cent. Similar clustering was observed on most of the other showing up as slightly unequal values. A fairly simple
polynomials. program for equM real zeros of orders three and four can
Two principal difficulties were encountered in using the be devised from reference [3]. However, in actual practice,
Q-D ~lgorithm. The first consisted of obtaining a value of exactly equal zeros are highly unlikely.
zero in relation (2) due to computer roundoff. The second Another recently developed program proceeds as
consisted in the difficulty of separating roots of almost follows:
equal magnitude. Due to the clustering effect previously
discussed, the second difficulty was more serious. The (A) Q-D method (single precision) to find starting
technique used to overcome both difficulties was to make values for the zeros.
a shift of the variable along the real axis. Due to the strong (B) Newton-Bairstow method (double precision) to find
influence of coefficient variation on the zeros, as pointed more accurate values.
out by Wilkinson [4], this shift was made in double pre- (C) In (B) reduce p(z) b y dividing out the zeros Zk as
cision arithmetic. A similar shift was used in case some they are located, after a test of the remainder of the poly-
coefficient was zero. nomial evaluated at z = zk •
The technique used to obtain the results indicated in (D) Check zero count, and if count not equal to the
Table 2 was as follows: degree of polynomial, shift the variable in the reduced
polynomial on the real axis (double precision).
(A) Q-D method (single precision) to find starting (E) R e t u r n to (A) with reduced polynomial.
values for the zeros. F r o m 100 to 500 iterations were This program found all the zeros in all problems shown
used, depending on the polynomial degree. in Table 2, thus demonstrating that the difficulty in the
(B) Newton-Bairstow method (single precision) to find previously discussed program lay in the fact that the zeros
more accurate values (always using the original coeffi- were found in single precision arithmetic.
cients).
(C) Check root count, and if root count not equal to 5. C o n c l u s i o n s
the degree of the polynomial: (1) Reduce p(z) b y dividing The Q-D algorithm appears to be a useful and powerful
out the zeros that have been located. (2) Shift the variable tool in obtaining the zeros of a polynomial with real co-
in the reduced polynomial on the real axis (see [3, p. 173]). efficients since (1) all computations are with real numbers;
(D) Return to (A) with a reduced polynomial to find (2) a complete set of starting values for rapidly converging
new starting values. methods are obtained at one time, rather than one value
per iteration. The method of shifting the variable along
As can be seen from the tabular results on roots found, the real axis solves the inherent problem of separating
this method was quite successful for polynomials of ninth zeros of almost equal magnitude.
degree or lower. Shifting the variable greatly assisted in
separating zeros close in magnitude, which otherwise could RECEIVED MARCH, 1965; REVISEDJUNE, 1965
not have been located. The failures are traceable, we
believe, to the fundamental problems of finding zeros in REFERENCES
single precision arithmetic, as discussed b y Wilkinson [4]. 1. MULLER,D. A method for solving algebraic equations using
The largest magnitude zeros are produced first, and an automatic computer. Math. Tab. Wash. 10 (1956), 208--215.
thence in order of descending magnitude. Hence, by vary- 2. M,tEHLr, HAmS J. Zur iterativen auflosung algebraischer
ing the amount of the axis shift mentioned in C(2) above, gleichungen. Z. Angew. Math. und Physilc. 5 (1954), 260--263.
different zeros will be output, and if two or three different 3. HENRICI,P. Elements of Numerical Analysis. John Wiley and
shift values are used, all zeros m a y be obtained in even the Sons, New York, 1964.
most difficult cases, as the greatest computational errors 4. WIL~SON, J. H. The evaluation of the zeros of ill-condi-
occur in the later stages of the process outlined above. tioned polynomials III. Numerische Mathemematik I (1959),
150-180.
This feature might be automated, but this was not done 5. RUTISH&USER, H. Der quotienten-differenzen-algorithmus.
in this study. Z. Angew. Math. Physik. 5 (1954), 233-251.
In Table 1, zeros of exactly equal magnitude were 6. - -. Anwendungen des quotienten-differenzen-algorith-
obtained in an earlier program which did not involve the mus. Z. Angew. Math. Physik. 5 (1954), 496-507.
axis shift feature. The later program used for Table 2 7. Bestimmung der eigenvektoren und eigenwerte einer
deliberately suppressed zeros previously found, since in matrix. Z. Angew. Math. Physik. 6 (1955), 387-401.
such cases it could be assumed that the Newton process 8. Eine formel von Wronski and ihre bedautung ffir den
was incorrectly converging to a previously found zero. quotienten - differenzen - algorithmus. Z. Angew. Math.
Physik. 7 (1956), 164-165.
(This could be avoided if a reduced polynomial were used 9. HENRICI, P. The quotient-difference algorithm. Nat. Bur.
in the Newton-Bairstow scheme.) However, it was gener- Standards Applied Mathematics Series 49, U. S. Govern-
ally true that exactly equal zeros did not appear as equal ment Printing Off., Washington, D. C., 1958, 23--46.
in the results due to computer rounding. Hence, equal 10. The RAND Corp. A Million Random Digits. Free Press, Chi-
real roots, at least, will usually be found in any case, cago, 1955.

574 Communications of the ACM Volume 8 / Number 9 / September, 1965