208 SOLVING ALGEBRAIC EQUATIONS BY AN AUTOMATIC COMPUTER

1. Iterate the function 2y2 — 1, starting with the number y whose arc sin is
required.
2. Record the signs of the iterates in order.
3. Accumulate the signs; that is, record the "partial products" of the signs
in order.
4. Write descending powers of 2 between the signs accumulated.
5. Multiply the series obtained by ir/2.

Example : Compute arc sin V-75

1. VT75, .5, -.5, ■•5,

2. + +
3. + + +
4. +1/2 +1/4 -1/8 + 1/16 -1/32 ••• =2/3
5. ir/2-(2/3) = ir/3
arc sin V-75 = 7r/3.

S. Comparison with Other Methods. The usefulness of the method described

above as compared with other methods depends, of course, on the function to be
evaluated and on the features of the machine to be used.
The fact that each iteration yields exactly one binary bit may be an advantage
or a disadvantage ; a method where error decreases faster than 2~" will converge
with fewer iterations than this one. On the other hand, one iteration of this
method may consist of fewer commands than an iteration of another method.
The logarithm program for the CRC-102A described above has 4 commands per
iteration as compared with 14 commands per iteration in another program for
logarithm on the same machine. The program for arcsin by the method described
above has 9 commands per iteration as compared to 22 for another arcsin pro-
gram. The fact that each iteration yields exactly one binary bit also simplifies
error analysis, for the number of digits of accuracy is exactly one less than the
number of digits computed.
D. R. Morrison
SANDIA Corp.
Albuquerque, New Mexico

A Method for Solving Algebraic Equations

using an Automatic Computer
Introduction. Many methods have been developed for solving algebraic equa-
tions and several of these have been used with automatic computers [1, 2]. Those
methods which are most suitable for use with automatic computers are ones
which apply to a wide class of equations and which are relatively rapid when the
degree of the equation is large. The method described here has been constructed
with these considerations in mind and has been programmed for the Illiac
computer at the University of Illinois.
 SOLVING ALGEBRAIC EQUATIONS BY AN AUTOMATIC COMPUTER 209

A process is to be constructed to find n solutions to the general algebraic

equations of nth degree

(1) fix) = aox" + aix""1 + • • • + an = 0

where the coefficients ao, at, ■■■, a„ are complex numbers and ao ^ 0. Each root
is found by an iterative procedure. Successive iterations toward a particular root
are obtained by finding the nearer root of a quadratic whose curve passes through
the last three points. The quadratic will in general have complex coefficients and
complex roots. This solution is accomplished by a variation of the standard
quadratic formula. Although the method derived here is rather complicated, no
evaluation of derivatives of fix) and only one evaluation of the polynomial fix)
is required per iteration. If the degree of the equation is large a greater amount
of time is spent evaluating the function than is spent in the remainder of the
process. Thus, the time spent per iteration is less with this process than with
iterative schemes which require the calculation of derivatives, whenever the degree
of the equation is large.
The Lagrange interpolation formula will yield a quadratic

(2) Liifix)) = box2+ bix + 02

whose curve passes through the last three points (x,-, fix/)), (x,-_i, /(x,_i)),
(x,_2, /(x,_2)) where the coefficients bo, 6i, b2 satisfy the system of equations

box2 + biXi + b2 = fix/),

(3) box2i-i + &ix,_i + b2 = /(xi_i),
box2i-i + &ix,_2 + b2 = /(x,_2).

A somewhat more convenient representation for this quadratic is obtained by

introducing the new quantities h = x — *,-, hi = x¿ — Xj_i, Ä,_i = x,_i — x,_2,
X = h/hi, X; = hi/hi-i, and 5¿ = 1 + X¿.The Lagrange interpolation formula may
now be written as the following quadratic in X.

(4) Liifix)) = X25,-1[/(xi_2)X<2

- fixi-OXJi + fixdXi-]
+ Xôr'lfixi^Xi2 - /(xt_i)5,-2+ fix/)i\i + «<)] + /<*<)•
A single iterative step is obtained by letting xi+i be a value of x which makes
Li(f(x)) vanish. We may solve the quadratic equation in X obtained by setting
expression (4) equal to zero. We then obtain X = Xl+i = (xj+i — x¿)/(x, — x,_i)
by using the inverse of the standard quadratic formula

\J) A,+ l — ,
gi ± Vg<2- 4/(x,)5,Xt[/(xi_2)Xi - f(xi-i)Si + /(*<)]

where g¿ = /(x,_2)X<2 — /(x,_i)5i2 + /(x,)(X< + 5¿). From Xi+i we may obtain

hi+i = Xi+ihi and x,+i = x,- + Ât+i. This x,+i represents a zero of the quadratic
described above. The sign in the denominator of (5) is always taken so as to make
the denominator have the greater magnitude. This makes Xt+i and A,-+i each be
210 SOLVING ALGEBRAIC EQUATIONS BY AN AUTOMATIC COMPUTER

that one of the two possible choices having the smaller magnitude so that x¿+i
is the root which is closer to Xj.
A convenient starting method for this process uses artificial starting values
at Xo = —1, Xi = 1, and x2 = 0 :

an — a„_i + an_2 is used for f(xo),

(6) a„ + a„_i + a„_2 is used for /(xi),
an is used for f(x2).

thus making X2 = —\ and hi = —1. This choice of starting values makes

Liifix)) = a„ + a„_ix + a„_2x2

which approximates to f(x) in the neighborhood of x = 0. The advantage of

this starting process is that it requires no special evaluations of the polynomial
f(x) and is therefore rapid.
Convergence of the Process. A final value of the root x% is taken when
|x< — 3Ci_ij/|at,-| becomes less than some preassigned number. Such a convergence
criterion is consistent with the use of floating point arithmetic in the calculation.
As a result of this criterion we see that convergence occurs if x,_i = x,. This
means that before convergence no two consecutive iterative results will be equal.

Furthermore, if x¿ = x,_2 we have 5¿ = —■-^— = 0 so by (5) X,+i = 0 and

X,_l Xj_2
X;+i = Xi also giving convergence unless x; = 0. Thus in normal operation of the
process we see that x„ x,_i, and x,_2 are distinct.
As each root is found it may be divided into the polynomial f(x) thus reducing
the degree of the polynomial by one. The algorithm for this reduction is the
commonly used one
(7) a/ = ra,_i + a,, (i = 0, 1, 2, • • •)

where a/ is the new coefficient to replace a¿ and r is the root which has just been
found. We make a_i = 0. Errors introduced by this process will be reduced if the
roots are eliminated in order of increasing magnitude. By always starting at the
point x = 0 one will tend to find roots in roughly this order.
No general proof of convergence in the large has been obtained for this
process, but convergence can be shown to occur whenever the process leads one
sufficiently close to a single or double root.
In order to facilitate the study of convergence let us assume that x,+i = 0.
This loses no generality since a simple shift of origin is always possible within the
system. At point Xj+i we also have L;(x¿+i) = 0 so that

(8) Li(xi+i) = èox21+i + ¿>ix,+i + b2 = b2 = 0.

Now each of the functions appearing on the right hand sides of equations (3)
may be expanded about x,+i = 0 as

(9) /(«^-¿«»^(OVi!
t-=o
 SOLVING ALGEBRAIC EQUATIONS BY AN AUTOMATIC COMPUTER 211

and b2 may be written as

(10) è2= zZbikf(k)(0)/k\ = 0

where bu is obtained by solving the system of equations

bokXi2 + bikXi + b2k = x/

(11) bokx2i-i + &iiX,_i + b2k = x*i_i
bokx2i-i + bikXi-i + bik = x*<_2.

This system of equations may be solved by elimination provided Xi, x¿_i, and
x,_2 are distinct and we obtain Ô20= 1, Ô21= b22 = 0 for the first three solutions.
When k > 3 the solution becomes somewhat more difficult but may be carried
out by eliminating bu between the first two equations to give
k-1
bokXiXi-i — biz = x<Xi_i ¿Z xty¡:í~p.
p=0

A similar elimination may be made between another pair and the result combined
with the above equations to give

(12) o2k = x¿x,_iXt_2 ¿_, x,x,_iX¿_2.

p+q+a—k— 3

The sum is to be taken over all non-negative integral p, q, and s satisfying

p-\-q + s = k — 3. All of these results may be inserted in (8) and (10) giving

(13) /(0) = -XiXi-iXi-A If"i0) + £ ¿Z xtxUxSi-2fwiO)/k\

\ k=i p+q+t=k— 3

Up to this point no approximations have been made and no limits have been
taken. Equation (13) expresses the same relationship contained in (5). We now
assume that the points x,-, X;_i, x,_2 lie in the neighborhood of a root r. Thus if we
let e,-+i = Xi+i — r, e,- = Xi — r, «t_i = x,_i — r and éí_2 = x,_2 — r the magni-
tudes of the last three quantities are all assumed to be less than some upper
bound em

(14) Iti I < em, Ie,_i I < em, \ e¿_21< em.

We also now make the tentative assumption

(15) |í;+i| < em,

and we shall seek to justify this assumption later. If (14) and (15) are inserted
in (13) and the functions are expanded about r we obtain

(16) ei+if'ir) + e2i+ihf"

(r) + 4+11/'"M ■
= - (e; - ti+i)(e¿_i - ei+i)(Éi_2 - ei+i)if'"ir) + 0(e*m).
212 SOLVING ALGEBRAIC EQUATIONS BY AN AUTOMATIC COMPUTER

If r is a single root we see that ei+i = 0(«m) so that we may write

/'" (f)
(17) «¿+i= -e,t,_ie,--2 ■+ 0(4,).
6f'(r)
A solution e,+i to this equation does exist if em is sufficiently small and will satisfy
(15). This solution will also satisfy L¿(/(xi+i)) = 0, and hence we are justified
in assumption (15) and hence (17) for at least one of the two x,+i for which
Li(f(xi+i)) = 0 holds. We now wish to show that (17) holds for that xi+i which
is actually chosen by the process described in connection with equation (5). If
both x.+i satisfy (17), the proof need not be given. If, however, one does and one
does not, we must make some further analysis. It was pointed out that the process
chooses the point x¿+i which is nearer to x,-. The point given by equation (17)
must satisfy

(18) |xi+i — Xi\ = |eí+i — €¿| < 2em.

This must therefore also hold for the x¿+i which is chosen by the process in (5)
and hence |í,+i| < 3em for this case. But |«,+i| < 3em is adequate to give (17) for
sufficiently small em and we may therefore assume (17) for the e,+i obtained in
the process.
A general limiting formula for the e,- in the neighborhood of a root may be
obtained from equation (17). If logarithms are taken on both sides we obtain

(19) log u+i = log e,-+ log ei_i + log 6,_2+ log ( - Jjrr!) + 0(0-

Neglecting the terms 0(em) we may solve (19) as a difference equation using
standard techniques and obtain

/ f'"(r)\
(20) log tj = cmx* + c2m2' + C3m3' — 5 log I — ,,,, : )■
\ 6/'(r)/

Where the constants Ci, c2, and c3 are determined by the starting values and the
three orders of convergence mi, m2, and m3 are roots of the characteristic equation

(21) m3 = m2 + m + 1.

The roots are

í»i = 1.84,
mi, mi = —.420 ± .606?.

Since the last two roots have magnitude less than 1 their effect will die out and
the order of the process is given by m\. After these approximations become valid
we have from (20)

(22) ej+i = £./ «

 SOLVING ALGEBRAIC EQUATIONS BY AN AUTOMATIC COMPUTER 213

where
K_( f'"(r)Y

In the case of a double root a similar argument exists. Equation (17) is then
replaced by

(23) é-+i = —€,e,_ie»_2 + 0(4,)

3fir)
and the characteristic equation becomes

(24) 2m3 = m2 + m + 1.

It has roots
mi = 1.23,
m2, mi = —.367 ± .520t,

and again the order of convergence is given by mi. We therefore have in the limit

(25) ei+i = Ke/-2*

with
/ fir) y
K-\~3rV)) •
Convergence of the Generalized Process. One might imagine a generalized
process in which an a degree Lagrange interpolation polynomial Li(x, a) is used
rather than the quadratic of equation (2). This presumes that some new method
for obtaining the nearest root of this polynomial is to be used. Since the direct
method corresponding to equation (5) would no longer be practical, one would
probably use some iterative method for solution of the equation Li(x,+i, a) = 0.
We now wish to investigate the convergence rate for such a process.
A general set of equations corresponding to equations (11) may be formed.
They are
a

(26) £ bikxlZÍ= xU, s = 0, 1, • ■•, a.

When k > a the quantity bak may be obtained by elimination. (This direct method
for obtaining bak was pointed out to the author by Mr. W. Scott Bartky.) Let us
eliminate ba-i,k between the first equation and each succeeding equation giving

a-2

(27) ¿Z bjkXiXi-e(x?~J~l - x"r/-1) - bak(xi - x,_.) = x¿x¿_,(x¿-1 — x^),

J-0

s = 1, 2, • • •, a.
214 SOLVING ALGEBRAIC EQUATIONS BY AN AUTOMATIC COMPUTER

Each equation may be divided by (x¿ — x¿_„) giving

a—1 a—j—i k—1

(28) ¿_ bjkXiXi-, ¿_ XiXi_s ' bak — XiXis ¿_, x¡x¡ , s = 1, 2, • ■•, a.
)=>0 1=0 ¡=0

We next eliminate bx-i,k • ■-, bok in a similar manner until the result

(29) bak = (-\)aXiXi_i ■■• ¡t,_ £ *f4L, • • • x\ta,

Po + pi +•••+/»« = *- (a + 1)

is obtained. In this expression the summation is made over all terms for which
a
the exponents po, pi, ■■• ,pa are non-negative integers and zZ Pi — h — (a + 1).
J-0
We also see directly from (26) that bak = 0 if k < a, except that bao = 1.
We may therefore obtain a generalization of equation (17)

(30) ei+i= (-1)«+W, ••• «_ (a {T+"!2,

+ 1)!/'(/)
, + °('-+2)
and obtain for the characteristic equation corresponding to (21)

(31) ma+l = ma + «■-' + • • • +1

This equation has one root «i which lies between 1 and 2 on the real axis and
which approaches 2 with increasing a. The remaining roots lie within the unit
circle and therefore represent perturbations which die out. The order of con-
vergence of the process to single roots is therefore given by mi. Since this can never
reach 2 we conclude that there is little to be gained in speed of convergence by
letting a exceed 2.
One should not ignore the possibility of letting a = 1. In this case the formula
corresponding to (5) is greatly simplified since a linear equation rather than a
quadratic now must be solved. This choice, however, suffers from a disadvantage
if all the coefficients of the original equation are real. If one starts from a real point
xo then all successive iterative results x, will also be real and hence only real roots
will be found.
Tests of the Method. The process with a = 2 as outlined in the preceding
sections was altered slightly in practice. Whenever the new value of the function
/(xi+i) is calculated the quantity |/(xi+i) |/ |/(x¿) | is formed. If this latter quan-
tity exceeds 10 the quantity X¿+i is halved and hi+iXi+i, and /(x,+i) are recomputed
accordingly. With this revision the process has produced convergence in all the
cases tested.
Another alteration was made to handle the case in which the denominator
of (5) is zero. This occurs whenever /(x¿) = /(x,_i) = /(x,_2) and in such cases
the arbitrary value X,+i = 1 is chosen since (5) may no longer be used.
The process (a = 2) was tested for equations of varying degree. Fourteen
equations were solved starting with degree 10 and progressing in steps of ten
through degree 140. Each equation was formed by choosing random points as
 SOLVING ALGEBRAIC EQUATIONS BY AN AUTOMATIC COMPUTER 215

roots within the square having vertices rfc 1 ± t. Polynomials were then formed
from these roots. The solutions to these polynomials were then compared with the
original random numbers which were used to generate the polynomial. Results
are summarized in the table.
Time Taken by
Accuracy of Last Illiac for Corn-
Degree of Accuracy of Least Root to be plete Solution
Equation Accurate Root Found in Minutes
10 10-7 10-9 1
20 10"8 10-8 2
30 IO-5 10-8 5
40 IO-4 lu"9 6
50 IO-4 10-6 10
60 IO-5 IO-7 12
70 10-" IO-5 17
80 lu"4 IO-5 20
90 lO-1 10-6 20
100 10-6 33
110 — — 42
120 — — 43
130 — 10-8 48
140 — — 60

Dashes in the table indicate that some roots were too inaccurate to be identi-
fied. In all equations some roots appeared which were correct to 10-8 or better.
The solutions to the 100th degree equation, which had some unidentifiable roots,
were used to generate a polynomial. All coefficients of this polynomial agreed with
the coefficients of the original polynomial to at least 6 decimal places. This result
indicates that the obtaining of accurate values of the roots of the equations of
higher degree was precluded by the limited accuracy of the coefficients and
independent of the method of solution.
The equation x128—1=0 was solved as an example of a special type of
equation whose solution could be easily checked. The maximum error occurring
in any root was of order 10-7 and the time for solution was 70 minutes.
No equation whose solution has been attempted has failed to yield convergence
although as indicated in the table, the solutions of equations of large degree may
be greatly in error. We conclude that convergence in the large does occur in most
practical cases in spite of the fact that convergence has only been proved for
single and double roots when the process has brought one to the neighborhood
of a root.
David E. Müller
University of Illinois
Urbana, Illinois
1. R. A. Brooker, "The solution of algebraic equations on the EDSAC," Cambridge Phil.
Soc, Proc, v. 48, 1952, p. 255-270.
2. Hans J. Maehly, "Zur Iterativen Auflösung Algebraischer Gleichungen," Z. Angew. Math.
Physik, v. 5, 1954, p. 260-263.