SOLUTION OF THE EQUATION ze* = a

BY E. M. WRIGHT
Communicated by Richard Bellman, December 15, 1958
The roots of the equation zez = a (a7*0) play a role in the iteration
of the exponential function [2 ; 3 ; 11 ] and in the solution and applica-
tion of certain difference-differential equations [l; 9; 10; 12]. For this
reason, several authors [4; 5; 7; 8; 9; 12] have found various proper-
ties of some or all of the roots. Here we "solve" the equation in the
following sense. We list the roots Zny where n takes all integral values,
and define Zn precisely for each n. We give a rapidly convergent series
for Zn for all n such that \n\ >n0(a); the first few terms provide a
very good approximation to Z n . In general, n0 is fairly small. Finally
we show how to calculate each of the remaining Zn ( — ^ o ^ ^ ^ ^ o )
numerically by giving a variety of methods to find a first approxima-
tion to Z n and showing how to improve this to any required degree of
accuracy.
We cut the complex s-plane along the negative half of the real axis
and take | arg z\ Sir in the cut-plane. If we put w = z+log z, we have
dw/dz = (z + l)/z and there is a branch-point at z— •—1. The cuts in
the w-plane are the two semi-infinite lines on which w = u ± m, u ^ — 1.
It can be proved that there is a one-to-one correspondence between
the points of the s-plane and those of the w-plane, excluding the cuts
in each case, so that the function z(w) is uniquely defined in the cut
w-plane.
We write A = \ a\, take log A real and log a = log A+ia, where
— x < a ^ m. All the roots of our equation are given by Zn
= z(log a-\-2nwi), where n takes all integral values. Zn is thus precisely
defined except when <x = w and log A<—\, (i.e. when a is real and
— e~l ^a < 0 ) . In this one case, log a and log a — liri lie one on each of
the two cuts in the w-plane; s(log a) has two real values, one less than
— 1 and one between —1 and 0, while z (log a — 2iri) has the same two
values. If — 6 ~ 1 < a < 0 , we define Z_i and Z 0 to be these two real
values, distinguishing them arbitrarily by Z _ i < — K Z o < 0 . If
a— —e~x, the equation (1) has a double root at z— —\ and we put
Z_i = Zo= — 1. In addition, when a is real and positive, Z 0 is real.
There are no other real roots for any a.
For every nonreal root Z», we write Zw = X n + i F n . It is easily
proved t h a t F 0 lies between 0 and a, t h a t
(2n - 1)TT + a < Yn < 2nr + a (n è 1)
and t h a t
89
90 E. M. WRIGHT [March

Ifiit + a < Y„ < (2» + 1)T + a (M g - 1).

We define the sequence of polynomials Pm(t) by

Pi{t) = /, Pm+i{i) = Pm(t) +m f PmW)d<T.

Jo
In particular,
1 3 1
P 2 = / + _ i\ p3 = t + — fi + — fi,

11 1 25 25 1
P 4 = t + 3t* + — t3 + — t\ P 6 = t + 5t* + — ts + — t' + — <6.
6 4 6 12 5
For every sufficiently large positive n, we write H = 2nir+a — w/2,
(8 = log (i4/fl) and

(i) u = Zc-iyiw^ff-2'-1.
y-o
We can show then that
F n = H + 77, Xw = (ff + ij) tan 97
or, if we wish to calculate Xn only without first calculating t\, we
may use the series

(2) Xn = 0 + £ (-l)iPvQ3i)H-*>:

To obtain these expansions we take i/7 as a first approximation to

Z n and note that f3—w(Zn) —w(iH). Hence, by the Taylor's series for
z(w)> we have
00

(3) Zn = iH + £ ^[dmz/dwm]l=iH.
W—1

Some manipulation enables us to deduce (1) and (2). We can show

that the series in (1) and (2) are convergent and the results valid if
2ff|j8| < ( f f - l) 2 ,

(log A)2 < l H TTJ + 2 (1 + log A) log H + 1

are both satisfied, which they clearly are for large enough n.
To calculate 77 from a reasonable number of terms of (1), we must
19591 SOLUTION OF THE EQUATION » * - « 91

have jS/fiT fairly small. We observe that the series in (1) and (2) have
real terms, a matter of importance for numerical calculation.
For n negative, we write H= —2mr—Tr/2 —cuX), /3 = log (A/H) and
define rj by (1). We have then
Yn= - H -rj, Xn = (H + V) tan rj
and (2) is still true.
There will remain a few values of n for which the series (1) and (2)
diverge or converge too slowly to provide a convenient means of cal-
culating Zn. For such an n, we have to calculate z(w), where w = log A
+ (2mr+a)i. Now
dw z+ 1
dz z
dz z
dw z+ 1
2
dz dz d / z \ z
2
dw ~ dw dz \z + 1/ (s + l ) 8
Hence, if ôz, 8w denote corresponding small changes in z and w, we
have
(4) ôz = z(z + l)~l»w + 0{z(z + l)-*(Ôwy}.
Thus, if we have a first approximation Zo to z, we calculate z£/o = z£>(2o)
and take 8w = w — w0. We then apply the correction ôz — Zoizo + l^ôw
to Zo to obtain Zi (say). If we write w = u+iv and z = x+iy, we may
calculate X= {(xo + l ) 2 + ^ o } _ 1 and use the correction in the real form
ox = {1 — X (XQ + l)}ôu ~ ^oXôz;,
ôy = yokôu + {1 - X(#o + l)}ôv.
Next we calculate Wi=w(zi) and, if this still differs appreciably from
wf we repeat the process. It is usually possible to use the same coeffi-
cients of du, ôv in (5) at each step. Provided z0 is not near — 1, the
process converges fairly rapidly by (4).
But 0o is near —1 if and only if w is near —1 ±wi. Let us suppose,
for example, that w lies near — 1 +7ri, so that z must be near — 1 . We
can show that
00

(6) z = - 1 + X) ^cow,
W l = l

where
92 E. M. WRIGHT [March

ci = - 3c2 = 36c3 = 270^4 = 4320^5 = - 1701(k6 = 1,

139 1 _ 571
Cl C C
~~ ~~ 5443200 ' * ~~ ~ 204120 ' ' ~" ~ 2351462400 '
^ m—1

Cm = — Cm-l(w + 1)~X X) C/i^-M-1 0» à 3)

2 /t=2

and co = i2 1/2 (w + l — iri)112. If ze> lies on the lower edge of the cut in
the w-plane ending at — 1 +wif we take co real and positive; if w does
not lie on this cut, we take £(a>) > 0 . The radius of convergence of the
series in (6) is 27r*. If w lies near — 1 — 7ri, the same series gives us
z(w), but oy — il112^-^!-^^)112 and tf(co) < 0 , unless o> is real. Thus if
(say) log a is near — 1 +iri, (6) enables us to calculate Zo and Z_i.
If w lies between the cuts, i.e. if u < — 1, we have (see [ó], for exam-
pie)
00

(7) z = X ) (-l) wr " 1 w ,fl " 1 (w!)" 1 ^ mtt '.

m=l

For u^—2, the first few terms give the value of s with sufficient
accuracy. This gives us Zo when \a\ Se~2.
Even if the series (1) does not converge sufficiently rapidly to be
useful to calculate Z», the first one or two terms may provide a suffi-
cient approximation to enable us to apply our correction procedure.
(A similar remark applies to (7) and even to (6).)
If \w\ > 4 and w does not lie between the cuts in the w-plane, a
useful value for Zo is w — log w, where log w has its principal value. The
next approximation Z\ will be accurate to at least one decimal place
and further approximations converge rapidly. For \w\ ^ 4 , we have
constructed a table of w(z), which gives a satisfactory value of ZQ by
inspection, except near z — — 1.
Alternatively drawing can be used to obtain the first approxima-
tion. Given u, v, we have to solve
(8) x + log r = u, y + 6 = v,
where r2 = x2+y2, tan d=y/x. To solve these equations graphically,
we use (i) a sheet of paper, the (x, y) plane, carrying circles r ~k and
radii 0 = h for various values of k and h, and (ii) a sheet of tracing
paper, the {X, Y) plane, on which the lines X— —log k and Y= — h
are drawn. We place the origin of the (X, F) plane at the point (uy v)
on the (x, y) plane, make the corresponding axes parallel and then
plot on a second sheet of tracing paper (the second (#, y) plane) placed
over the first the intersections of X = —log k with r — k and that of
1959] SOLUTION OF THE EQUATION ze*~a 93

F = — h with 8 = h. Through these two sets of points can be drawn

the two curves (8) and their intersection in the (x, y) plane gives the
required solution.
REFERENCES
1. R. Bellman, Ann. of Math. vol. 50 (1949) pp. 347-355.
2. G. Eisenstein, J. Reine Angew. Math. vol. 28 (1844) pp. 49-52.
3. L. Euler, Opera Omnia (i) vol. 15, Leipzig and Berne, 1927, pp. 268-297.
4. N. D. Hayes, J. London Math. Soc. vol. 25 (1950) pp. 226-232.
5. , Quart. J. Math. Oxford Ser. (2) vol. 3 (1952) pp. 81-90.
6. A. Hurwitzand R. Courant, Funktionentheorie, 3d éd., Berlin, 1929, pp. 141-142.
7. E. M. Lémeray, Nouv. Ann. de Math. (3) vol. 15 (1896) pp. 548-556 and vol. 16
(1896) pp. 54-61.
8. , Proc. Edinburgh Math. Soc. vol. 16 (1897) pp. 13-35.
9. O. Polossuchin, Thesis, Zurich, 1910.
10. F. Schürer, Ber. Verh. Sachs. Akad. Wiss. Leipzig Math.-phys. Kl. vol. 64
(1912) pp. 167-236 and vol. 65 (1913) pp. 239-246.
11. E. M. Wright, Quart. J. Math. Oxford Ser. (2) vol. 18 (1947) pp. 228-235.
12. , J. Reine Angew. Math. vol. 194 (1955) pp. 66-87, esp. pp. 72-74.
UNIVERSITY OF ABERDEEN, SCOTLAND