m*!

EXPONENTIAL SUMS 213

ON T H E ZEROS OF E X P O N E N T I A L
SUMS A N D I N T E G R A L S *
BY R. E. LANGER

1. Introduction. It is of frequent occurrence in problems of

both pure and applied mathematics that certain values sought
may be specified and must be determined as the roots of a tran-
scendental equation. In particular, the equation may be of the
class in which the unknown is involved only through the
medium of exponential or trigonometric functions, with coef-
ficients which are power functions or essentially such. As simple
examples one has such equations as
tan z - z = 0, e' + z2eZz + (1 - z) = 0,

sin z + z3 tan 5s = 0, eiz — 2s5 = 0.

A determination of the distribution of roots of such an equation

may, therefore, be of interest and importance. Inasmuch as
this problem and certain of its generalizations have been the
subjects of recent and successful investigations I have presumed
that an outline of methods and attainments associated with it
might not be unsuitable for presentation at this meeting of the
Society.
The relation between the trigonometric functions of z and the
function eiz evidently makes possible the expression of each of
the functions involved in the equations above in the form
n

(1) $(a) = 2>/(*)e°''.

This type of function we shall refer to as an exponential sum.

It is to be the subject of Part I of the discussion to follow. If
the values A3-(z) and Cj are constants, and the Cj are real, the
function (1) can be expressed with the use of an appropriate
constant ZQ and a suitable step-function <£(/) in the form

* An address delivered at the meeting of the Society at Columbia, Mis-

souri, November 29, 1930, by invitation of the program committee.
214 R. E. LANGER [April,

(2) $(z - zo) = z J <j>{t)etzdt.

The integral of the type here involved with less specialized func-
tion </>(/) represents, therefore, a generalization of certain sums
of type (1). Such integrals include a number of standard func-
tions of importance, among which the Bessel functions may be
specifically mentioned. They will be considered briefly in part
II öf this discussion.
Returning to the consideration of the sum (1) we shall under-
stand that the quantities Cj, to be called the exponents, are con-
stants (real or complex). The functions Aj{z)1 to be called the
coefficients, will be characterized as the discussion progresses by
various sets of hypotheses. It will be my general plan to pass in
the considerations from forms of less to forms of greater gener-
ality, and to indicate at each stage the results and their rela-
tion with those previously established.

PART I. T H E EXPONENTIAL SUM

2. Constant Coefficients and Real Commensurable Exponents.

When the exponents Cj in a sum of the form (1) are real, it is a
matter merely of the arrangement of terms to suppose that these
exponents occur in the order of increasing algebraic magnitude.
With this arrangement it is then permissible without loss of
generality to assume that the first exponent c0 is zero, inasmuch
as the removal of any exponential factor can in no way affect
either the number or distribution of the zeros of the function.
These adjustments will be assumed to have been made. Theo-
retically the simplest type of sum is one, then, in which the
problem of the distribution of zeros is essentially an algebraic
one, and this occurs in particular when the coefficients are con-
stants (,4,-(2;)=a,-), and the exponents are commensurable, that
is, Cj = ap3', i = l, 2, • • • , n, with some real constant a and in-
tegral values pj. The sum is then of the form
n
(3) *(s) = DaX*»)1", p» = 0,

which is a polynomial of degree pn in the quantity eaz. If this

polynomial admits as zeros the values §i, £2, • • • , ?Pn, the func-
193*1 EXPONENTIAL SUMS 215

tion (3) vanishes for such values of z, and for only such, as
satisfy a relation
eaz = h.
The zeros of $(z) are, therefore, given by the formula
1
fA\ (o • . i y \ (7 = 1, 2, • • • , pn),
(4) z = —{2f»Tt + logt,-), o
a (w = 0, ± 1, ± 2, • • • ).
They are thus seen to be denumerably infinite in number and to
be distributed in the complex z plane at regular intervals of
length 27r/a along pn lines (not necessarily distinct) which are
normal to the axis of reals.
If the explicit solution of the polynomial equation involved is
feasible, the determination of the zeros of $(z) is completed by
the formula (4). In the alternative case the following further
considerations will not be futile. It is clear from the result (4),
that with any specifically given function $(z), the choice of a
constant K is possible so that the zeros of 3>(z) all lie within the
rectilinear strip of the z plane given by the relation
(5) | x | < K, (z = x + iy).
Any line y=yi, parallel to the axis of reals, cuts this strip, and
on such a line it is found by direct computation that

(6) ƒ{—*(*)} = X>/(?i)(*"0*',

Uo ; -•„!
the coefficients b3- being real constants with values depending
upon yi. The expression (6) is a polynomial in the positive
quantity eax. As such it can vanish, by Descartes's rule, at
most as many times as there are changes of sign in the sequence
of values b3: This cannot exceed (n — 1), since the number of
terms in (6) is n.
Let R designate the rectangular region cut from the strip (5)
by two lines y=yi, and y =y2} chosen so that on neither of them
there lies a zero of $(z). (There will evidently be such lines
through any interval of the y axis, however small.) Then as z
traces either of these lines the value arg $(2) can change (in-
crease or decrease) by at most rnr since l{ ( l / a 0 ) * ( s ) } vanishes
no oftener than (n — 1) times. As z traces the side x= — K of
216 R. E. LANGER [April,

the rectangle R the value of the sum $(0) remains arbitrarily

near that of its first term a0 (provided K has been chosen suf-
ficiently large), and the change in arg $(0) accordingly remains
arbitrarily near to 0. Lastly as z traces the side x = K the
change in arg $(0) is an increase in value arbitrarily near to
cn(y2~yi), since $(0) on this side of the rectangle is approxi-
mated with an arbitrary degree of accuracy by its last term
aneCn\ From this evaluation of the extreme possible changes in
arg $(0) as z traces the boundary of the rectangle it must be
concluded that the number n{R) of zeros of $(0) within the
rectangle is subject to the bounds

(7) - n + —(;y2 — y\) ^ n(R) ^ n + —(y2 - yi).

Z7T Z7T

This evidently limits, both above and below, the density of zeros
in any portion of the strip (5). In particular it shows that no sum
of the form (3) with (n+1) terms can have a zero of multiplicity
greater than n.
T H E O R E M 1. If in the exponential sum (1) the coefficients are
constants and the exponents are real and commensurable, then the
distribution of zeros is given explicitly by the formula (4). In this
distribution the number of zeros which lie between two lines y=yu
and y=y2, is restricted by the relation (7).
The hypothesis of commensurability of the exponents is evi-
dently a restrictive one. Yet many functions $(0) of importance
and interest are included by it. It will be observed, for instance,
that the hypothesis is vacuous whenever the sum involves only
two terms, and the results obtained apply, therefore, to all such
sums. Similarly the familiar trigonometric sums
n
X X COSJ0,
(8) n
^aj sinj0

are included, as is in fact any linear combinaton of trigono-

metric functions of integral multiples of 0 (iz naturally replacing
0). While the problem is in this case theoretically simple, the
interrelation of algebraic and transcendental features may be
193*'} EXPONENTIAL SUMS 217

of distinct interest. The following theorem, due to Pólya,* is at

once illustrative of this point and useful for later reference.
THEOREM 2. If the coefficients a3> are real and the zeros of the
polynomial

(9) P(Ö = Z a&

3-0

all lie within the unit circle about £ = 0, then the zeros of the corre-
sponding trigonometric sums (8) are all real and simple. Each of
these sums has precisely 2n zeros on the interval 0^z<2w and the
zeros of either sum alternate with those of the other. (By a theorem
of Kakeya the hypothesis is fulfilled if 0^<Zo<#i< • • • <an.)
The proof is exceedingly simple. Under the hypothesis made
ar
£ -P(£) increases by 2nw as £ traces the unit circle, that is,
arg P(eiz) increases by that amount as z traces the axis of reals
from 0 to 27T. But when z is real the sums (8) are respectively
the real and imaginary parts of P(eiz). Hence these sums vanish
in alternation, each 2n times on the interval 0^2<27r. T h a t
the zeros are thus all accounted for follows from the fact that
each of the sums (8) may be expressed as a polynomial of
degree 2n in eiz.
3. Constant Coefficients and General Real Exponents. When
the exponents are not commensurable the determination of the
distribution of zeros of $(0) is not in general of an algebraic
character. Independent discussions of the problem in this case
have been given by C. E. Wilder 2 and by J. D. Tamarkin. 3 In-
asmuch as the subject has to some extent been anticipated by
the discussion of the preceding section a slight amplification of
the reasoning there used will suffice here for the derivation of
such results as have been obtained.
The sum $(z) for the case in hand is expressed by the formula
n
(10) $(*) = 2>y^«, co = 0.
y=o
Given such a function it is easily seen that with K chosen as a
constant sufficiently large the zeros of $(z) all lie within a strip
of the z plane determined by a relation (5). This may be de-

* See list of references at the end of this paper.

218 R. E. LANGER [April,

duced from the dominance of the first term of the sum (10) over
all remaining terms for values of z satisfying the condition
x< — K, and the similar dominance of the last term of the sum
when x>K. The change in arg #(2) as z traces any portion of
the lines x= —K or x = K may be computed precisely as in the
case of the preceding section and the establishment of the relation
(7) depends only on the consideration of the quantity

(il) /{-#(*)} = 2X(yV'*,

on a line for which y is constant. Since the expression (11) is
now not a polynomial Descartes's rule may not be invoked. It
is, however, easily proved that in this case also the expressions
can vanish at most as many times as there are changes in sign
in the sequence of coefficients (see Pólya and Szegö, Aufgaben
und Lehrsdtze aus der Analysis, vol. 2, p. 49) and hence the re-
lation (7) is valid as before for any rectangle R cut from the
strip (5) by a pair of lines y—yi and y=y2 on neither of which
I $(z) I is zero. The existence of such lines through any interval
of the y axis follows, of course, from the regularity of the func-
tion $(2).
It is a fact of considerable importance in many considerations
that in the case of an exponential sum the value |4?(z) | is uni-
formly bounded from zero when the variable z is uniformly
bounded from the zeros of $(s). This means that given a suf-
ficiently small positive 8 there exists a constant H depending
only on ô and such that
(12) J *(*) I > H, for \ z - zm\ >ô,

where zm designates the set of zeros of $(z). The fact is obvious

when consideration is restricted to any finite portion of the z
plane. In the case of commensurable exponents it then follows
for the entire strip (5) and hence for the entire z plane because
of the periodicity of the function $(s). The fact may also be
established (see Wilder 2 or Tamarkin 3 ), though less simply, in
the general case.
T H E O R E M 3. If in the exponential sum (10) the coefficients are
constants and the exponents are real, then the zeros of the sum all
lie within a strip (5), and in any portion of this strip the number
*93*J EXPONENTIAL SUMS 219

of zeros is limited by the relation (7). When z is uniformly bounded

from the zeros of $(z),then \$(z) \ is uniformly bounded from zero.
It may not be devoid of interest to remark that the distance
between distinct zeros of <£(s) admits of no positive lower bound
in the general case of incommensurable exponents. This is shown
by the simple and explicit example $(z) = (1— eaz)(l— e^z)f in
which a/(3 is irrational. The formula (4) shows that when the
exponents are commensurable a minimum distance between
zeros does exist.
4. Coefficients Asymptotically Constant. In many considera-
tions associated with the distribution of zeros of a function of
type (1) the interest centers directly only upon the character of
this distribution in the region of the z plane remote from the
origin, that is, in the region |s|>ikf, where M is a constant
arbitrarily large. When such is the case it is natural that the
hypothesis of the constancy of the coefficients may be replaced
without essential loss by a less stringent one to the effect that
the coefficients be sufficiently like constants when \z | is large.
To facilitate the formulation of such hypotheses we shall agree
upon the use of symbolism as follows. In a region R of the z
plane including the point z = oo a function will be called an
epsilon function provided that it is analytic in every finite portion
of R and that it approaches zero uniformly in R as \z |—»oo. To
indicate such a function we shall use the generic symbol e(z).
Let it be supposed now either that the coefficients A 3-(z) in the
sum (1) are single valued and of the form

(13) Ai(z) = *, + €(*),

in the region \z \ > M, or that they are multiple valued but such
that in the region \z \ > M, — w <arg z S T, their various
branches are each of the form (13). In the latter case the several
determinations of the coefficients correspond to distinct branches
of the sum $(2), and the considerations which follow apply di-
rectly to any one of these branches. The discussion may be
completed by a similar consideration of each individual branch.
The form assumed for the sum (1) is, therefore,
n

(14) * ( * ) = £ { « ƒ + «(*)}«"*, a0an?éO.

j-0
220 R. E. LANGER (April,

As already indicated it will be made a part of the hypothesis

that neither of the two constants a0, an is zero. It will be shown
that the zeros of the sum (14) are asymptotically represented
by those of the related sum
n
(15) $i(s) = 2>/*"',

which is of the type considered in §§2 and 3.

It may be shown to begin with, and by precisely the reasoning
employed in the analogous consideration of §3, that there can be
no zeros of the sum (14) outside a strip of the form (5) when
\z | > M. Within this strip, on the other hand, each exponential
is bounded and hence the function (14) is evidently expressible
in the form
*(s) = $!(*) + €(*), | x\ < K, | z\ > M.

Now in virtue of the distribution of zeros of $1(2:), as de-

termined in the preceding section, these zeros may be enclosed
in groups of n or less by a succession of contours yr such that
no point of such a contour lies within a distance less than 5 from
a zero of $1(2), and each of the regions enclosed is of major
diameter no. For values of z not within such a contour, \$i(z) \
is bounded uniformly from zero, as stated in Theorem 3, and
hence for such values of 0,
$(2) = $!(*) {1 + e(z)} .

From this relation it follows that when \z \ >M, <&(z) has within
each contour yr precisely as many zeros as $1(2), and that all
zeros of $(2) are thereby accounted for.
T H E O R E M 4. If the function $(z) (or a determination of it) is
of the form (14), then in the region \z\>M the distribution of zeros
of $(z) (or of the branch of $(z) in question) may be described as
in Theorem 3. The zeros are asymptotically represented by those
of the related sum (15).
5. Coefficients which are Asymptotically Power Functions. It
will be supposed now that in the form (1) the coefficients A 3-(z),
or chosen branches of them, are of the form
(16) A,-(*) = z"{a3- + e(z)}, a0an 5* 0,
m*l EXPONENTIAL SUMS 221

for z in the region \z\ >M, — 7r<arg Z^T. The values v,- are
taken to be any real constants positive, negative or zero.
The sums with coefficients of type (16) evidently include as
a sub-class those with coefficients which are polynomials in z.
A discussion of the latter was given first by Pólya 6 and
Schwengler. 7 Their method, which is largely geometric, is in
many respects convenient and illuminating and will be utilized
in the following discussion whenever it is of advantage.
Without further convention the values v3 (except for j = 0,
and j = n) are not fixed beyond ambiguity by the form (16), in-
asmuch as an increase in v3 may be compensated for by the
choice of a3 as zero. It will be assumed, therefore, for the sake
of definiteness that whenever the form of the coefficient Aj{z)
permits, the value v3 will be chosen so that the corresponding
constant a3 is not zero. The terms of the sum for which this is
possible will be called ordinary terms. In the exceptional non-
ordinary terms the value of v3 must remain unspecified. Under
the form (16) the constant a3- for every such term will be zero.
Thus, by way of example, if A3-(z)=log z, then the form (16)
implies a3=0, with v3>0.
As a preliminary to the discussion of the general case under
(16) it will be convenient to consider that in which all terms are
ordinary terms, and in which the constants v3 are proportional
to the exponents c3.
6. The Values v3 and c3 Proportional. If the real constant j3
is defined by the relation

vj = Pc h (j = 1, 2, • • • , n),

the formula
(17) f = z+ p\ogz

defines a single-valued analytic map of the portion of the z plane

to which z was restricted above upon a complex f plane, the
point z = oo corresponding to f = oo. It follows that any epsilon
function of z is also such a function of f, and it is evident, there-
fore, that under the relation (17)
n

*(«) = 2 1 «f + «(f) 1 eCit> «oa» T* 0.

j-0
222 R. E. LANGER [April,

This, however, as a function of f is precisely of the form (14),

and in the f plane the zeros are accordingly distributed as de-
scribed in Theorem 4. It is necessary only to translate this
result into terms of the variable z and this can be readily done.
Thus if f = ^+irj the components of the relation (17) are
£ = x + p log | z | ,
rj = y + p arg z.
The rectilinear strip |£ | <K corresponds, therefore, to the curvi-
linear strip bounded by the logarithmic curves
(18) x + p\og\z\ = ± K.
This strip which is asymptotically of constant width 2K, and
whose bounding curves approach parallelism with the axis of
imaginaries, contains all the zeros of $(z) in the region \z \ > M.
Since for z in the strip (18) the value arg z approaches either 7r/2
or — 7r/2 according as the upper or lower half plane is considered,
it follows that any line 77=^1 on which rj is constant maps
asymptotically upon the line y=r]ii.l3ir/2, on which y is con-
stant. Moreover, the distance between two such lines both in
the upper or both in the lower half plane is asymptotically pre-
served.
T H E O R E M 5. If in the exponential sum (1) the coefficients are
of the form (16) with values Vj proportional to the exponents Cj,
and all terms are ordinary terms, then the zeros of the sum are asymp-
totically located within a logarithmic curvilinear strip bounded
by curves of the form (18), and the number of zeros lying between
any two lines parallel to the axis of reals is asymptotically subject
to the relations (7).
7. General Real Values vj. With each ordinary term of the sum
n

(19) Hz) = 5 > ' { a , + €(*)}«•'',

there can be associated in the z plane the corresponding point

Pj with Cartesian coordinates (c/, Vj). These points determine
the broken (or straight) line L, which (i) joins P 0 with Pn, (ii)
has vertices only at points of the set P,-, (iii) is convex upward
(or straight), and (iv) is such that no points of the set lie above
it. Let Lr designate the rth. segment of this broken line L, the
m1'] EXPONENTIAL SUMS 223

order being from left to right, and let the scheme of single sub-
script notation employed in (19) and in designating the set of
points Pj be replaced by one in which the points of the set
lying on the segment Lr are denoted in order by Prh, (A==l,
2, • • • , fir).
If the sum (19) contains exceptional terms it must be made a
matter of hypothesis that for every such term a choice of the
value Vj is possible under the form (16) such that the correspond-
ing point (e,-, Vj) lies below the broken line L. All points of the
set located below the line L may then be designated in any
order by P0h, (& = 1, 2, • • • , #0)> similar double subscripts being
assigned to the corresponding terms in (19).
Consider now any intermediate segment Lr of the line L, and
let its slope be mn the slope of the preceding segment being
ntr-i. Then as the real parameter k varies over the range
mr-\ — € ^ k > mr — e,
with e positive and sufficiently small but otherwise arbitrary,
the curve
(20) x = - k log | z |
sweeps out the region of the z plane bounded by the curves
x = - (mr-i - e) log | z\ ,
x — — (mr ~ e) log | z | .
The figure is symmetrical in the upper and lower half planes.
On any curve (20) it is obvious that
I zvl8eclaZ I = IsI Vla hcis
~
and hence the relative magnitude of the terms in (19) for z on
this curve depends upon the corresponding values (vu — kcu).
These, however, admit of ready geometrical interpretation. They
are, namely, the y intercepts of the lines having the slope k
and passing through the respective points Pu. In virtue of this
fact and the restriction upon the range of k it is readily seen
that for every l^r the inequality
(vu — kCu) < (vri ~ kCri)

obtains. It follows, therefore, that the form of expression

224 R. E. LANGER [April,

is valid for z on any curve (20), and hence that the sum (19) can
be writ! m in the form
nr
(22) $(*) = J2zvrh{arh + e(z)}ecrk^

for all z of the region bounded by the curves (21). Upon removal
of a factor zv^eCr^y however, the sum (22) takes the form con-
sidered in §6, since the constants {vrn — Vr\) are proportional to
the exponents (crh — cri), the constant of proportionality being
mr, the slope of the segment Lr. It may be concluded, therefore,
by Theorem 5, that the zeros of $(z) in the region (21) are
asymptotically confined to the logarithmic strip bounded by
the curves
(23) x + mr log | z | = ± K,
and that the number of zeros in this strip and between two lines
y=yi and y=y% is restricted by the relation obtained from (7)
by replacing n and cr by nr and (cr,nr — crl).
The geometric picture is completed if it is observed that in the
upper half of the plane the central line of the strip (23), that is,
the curve
(24) x -\- mr log | z I = 0 ,
passes through the point (0, 1) with slope —l/mr, namely in
the direction perpendicular to that of the associated segment L r .
There is a strip of the form (23) associated with each segment of
the broken line L (obvious simple modifications of the discussion
being necessary in the case of an end segment), and the zeros of
<&(z) in the remote part of the z plane are all contained within
these strips.
T H E O R E M 6. If $(z) is an exponential sum with coefficients of
the form (16) (the Vjfor exceptional terms satisfying the hypothesis
of the text), then the zeros of the sum are asymptotically confined to
a finite number of logarithmic strips (23), the number of zeros in
any strip between two lines parallel to the axis of reals being
asymptotically subject to a relation similar to (7).
8. Collinear Complex Exponents. The hypotheses which have
been made in the various cases of the sum (1) hitherto con-
sidered have been concerned primarily with the structure of the
193* -I EXPONENTIAL SUMS 225

coefficient functions Aj(z). The exponents have been assumed

throughout to be real. It will be shown now that this assump-
tion is dispensable, and that the distribution of zeros of the sum
having complex exponents is also determinable. Preliminary to
the general discussion the considerations of this section will be
given to the case in which the values Cj viewed as points in the
complex plane are collinear.
Let / designate the line on which the points Cj are located, and
let the subscripts be assigned to these points in the order of their
location on /. Then if 6 designates the inclination angle of the
line I with respect to the axis of reals (the positive sense of I
being from c0 to cn), it follows that the relations
Ci = co + y^e*', (j = 0, 1, 2, • • • , n),

are satisfied by a set of real values 0 = Y 0 < 7 I < • • • <7n. In

terms of the variable f defined by the relation
(25) f = ze»,
the sum (1), therefore, takes the form
n

(26) * = e°" 2X(r)eV,

where B3-(Ç)=Aj(z) under the substitution (25). The sum in
(26) is one in which the exponents 7 / are real, and the coef-
ficients 5/(f) have the essential structural characteristics of
Aj(z). The distribution of zeros of the sum 3> in the £* plane is,
therefore, as described in the theorems of the preceding sections,
that is, the zeros are asymptotically confined to one or more
strips which are parallel or approach parallelism with the axis
arg f = ± 7 T / 2 . Since the transformation (25) is a mere rotation
of the plane, it may be concluded that in the z plane the zeros
of the sum are confined for | s | > A f to a similar set of strips
parallel or approaching parallelism with a direction arg z = — 6
± 7 T / 2 . These directions, it will be noted, are those perpendicu-
lar to the line in the complex plane which contains the points
Cj, the complex conjugates of the exponents Cj.
T H E O R E M 7. If the exponents ct- in the exponential sum (1) are
collinear complex constants, the distribution of zeros of $(z) is
obtainable from the theorem previously enunciated by substituting
226 R. E. LANGER (April,

in the role of the axis of reals the line containing the points c3- con-
jugate to the exponents c3\
9. General Complex Constants. In the most general type of
exponential sum to which consideration will here be given the
exponents may be any set of complex constants. This case is
now mature for discussion. While treatments of it by purely
analytic methods are available2,3'4«5 and effective, a geometric
method 6 will again be resorted to as convenient for the purpose
at hand. Let it be supposed that the values dj (conjugate to the
exponents) have been plotted in the complex plane, and let the
points of the set so represented be surrounded by the polygon
P which (i), is convex, (ii), has vertices only at points of the set,
and (iii), includes all points of the set either in its interior or on
its perimeter.
Let the sides of the polygon P taken in counterclockwise
succession with any one as the initial one be designated by lr,
r = l, 2, • • • , g, and with this done let the points of the set Cj
which lie on the side lr be redesignated by crh, h = l, 2, • • • , hr,
the succession being again in the counterclockwise sense around
the polygon P. The points Cj lying within the polygon may in
conclusion be denoted (in any order) by coh, h = 1, 2, • • • , h0.
If cor represents the inclination angle of the outer normal to
the side lr, the relation
COr_i + € g T < C0r + €,

with € sufficiently small and positive but otherwise arbitrary,

restricts the real parameter r. As r varies over the range per-
mitted, the line

(27) arg z = r

sweeps out in the z plane a sector described by the relation

(28) cor_i + € ^ arg z < œr + e.

Such a sector is thus associated with each side of the polygon,

and the set of these sectors completely fills out the complex
plane. It is proposed to consider the sum $(z) within the typical
sector (28). The extension of the results to the entire plane will
then be immediate.
WJ EXPONENTIAL SUMS 227

When z lies on a line (27) it is readily computed that

| Zv*kec*kz\ = gl'l^rt-'H-^sinH-i'rtlogUI/I.IJ^

where c8k = a8k+i^ak* It follows from this that the relative

magnitudes of the terms in the sum (1) are ultimately deter-
mined by the relative magnitudes of the associated quantities
asjc cos r + &/<• sin r.
These, however, are represented geometrically by the respective
projection of the vectors c8k upon the line (27), and from this
interpretation it is quickly seen, in virtue of the restriction
upon the range of r, t h a t
aSk cos r + Psk sin r < ar\ cos r + /3ri sin r, for any s 7^ r.
It follows that for z on any line (27), that is, anywhere in the
sector (28), the form of expression
zv*ew = zVrleCrlze(z)

is valid for s^r and any k. The sum $(z) with coefficients of
the form (16) may, therefore, be expressed as a sum with a re-
duced number of terms, that is,
hr
(29) $(s) = 2> r *(*)ew,

and in this expression which is valid for all z in the section (28)
the coefficients are still of the form (16). The exponents in the
sum (29) are, however, collinear, their conjugates crh lying on
the side lr. It may, therefore, be concluded from Theorem 7
that the zeros of $(2) within the sector (28) are asymptotically
confined to one or more strips which are parallel or approach
parallelism with the normal to the associated side of the polygon,
A similar conclusion may obviously be drawn in connection
with the values of z in any other sector.
T H E O R E M 8. If in the sum (1) the exponents are any complex
constants, the zeros of $(2) are confined for \z\>M to a finite
number of strips each of asymptotically constant width. These
strips are associated in groups with the exterior normals to the
sides of the polygon described in the text, and approach parallelism
with the respective normals. Within each group of strips the dis-
228 R. E. LANGER [April,

tribution of zeros may be described as in the previously stated

theorems, the role of the axis of reals being transferred to the re-
spective side of the polygon.
In closing here this brief and necessarily incomplete discussion
of the distribution of the zeros of an exponential sum, I wish at
least to call attention to the related work which has been done
by Ritt 8 on the factorization of such sums and on algebraic
combinations of exponentials.
PART II. T H E EXPONENTIAL INTEGRAL

10. Discussion of the Integral. If in the exponential sum with

constant coefficients and real exponents as given in form (9), the
origin is assumed to be a zero of the sum (and this can be made
so by a simple translation of the plane), then S/=o a3- = 0. With
this adjustment the sum <ï>(s) can be simply represented in the
form of an integral as follows. Let the coefficient function <£(/)
be constant in successive intervals as defined by the relation
r

3-0

for CrSt<cr+i, (r = 0, 1, 2,- • • , n—l). Then it is found directly

that

$00 = z I cj>(t)eztdt.

The exponential integral of the general type

(30) *aib(z) = f t(t)eztdt

Ja
thus includes among its specializations (except for a factor z)
the exponential sums of the types considered in §2 and §3, and
hence, of course, also all trigonometric functions which are so
expressible. Besides these a number of other functions, in-
cluding Bessel functions, are amenable to representation in this
form.
In turning to a consideration of the integral (30) a restriction
of the discussion is necessitated by the brevity of available time.
It is proposed, therefore, to summarize only certain investiga-
I93I-] EXPONENTIAL SUMS 229

tions which deal with the integral with finite limits of integra-
tion and which impose none but qualitative hypotheses on the
coefficient function cj>(t) involved. Many investigators have con-
sidered either integrals of specialized forms included under the
type (30), or the integral with infinite limits. It is not due to a
want of interest in the results obtained that these subjects are
omitted from the present discussion.
Various normalizations of the integral (30) may be attained
by simple substitutions on the variables involved. Appropriate
normalizations of this kind make for simplicity in the enuncia-
tion of hypotheses and results, and for this reason the following
ones will be assumed as indicated from point to point in the en-
suing discussion. To begin with, the substitution of at+f3 and
z/a, respectively, for t and z affects a change in the limits of
integration, and by suitable choice of the constants a, /3, any
desired finite limits may be attained. It will be found of par-
ticular convenience to normalize the interval thus in some cases
to (0, 1) and in others to ( — 1, 1). Further, the substitution of
z-\-c for z affects the introduction of a factor ect into the coef-
ficient function ^(2), a n d it follows that whenever ^(<Z)T^0 and
^ ( & ) T ^ 0 , a normalization may be made such that \p(a) =\//(b) = l.
Lastly it will be observed that the trigonometric integrals

(a) Vs(z) = I ƒ(/) sin zt dt,

Jo

(b) V0(z) = ! ƒ(/) cos zt dt

Jo
are both expressible in the form ML-u^s), the coefficients being
given respectively by the formulas [l/(2i)]{/(0—ƒ( — $ } , and
(l/2){/«)+/(-fl}.
11. The Integral with an Infinity of Real Zeros. The distri-
bution of zeros of an exponential sum (9) as determined in § 3
features conspicuously the existence of a bound upon the nu-
merical values of x for which a zero of the sum may occur.
This fact, it will be recalled, followed from the ultimate domi-
nance of the one or the other extreme term of the sum when \x \
becomes sufficiently large. The integral (30) is in one sense the
analogue of an exponential sum with infinitely many exponents
lying on the range between a and b. Either of these extreme
230 R. E. LANGER [April,

values might in such event be a limit point for the infinite set,
and if such were the case it would conceivably involve an indefi-
nite delay in the accession to dominance of the respective ex-
treme term of the sum. The absence of the corresponding bound
upon the x coordinate of the zeros might then be reasonably ex-
pected. T h a t the situation thus conceived does as a matter of
fact exist is shown by the following example constructed by
Titchmarsh. 9
Let the constants 0 and v be chosen subject to the conditions
0 < l / 2 , *>>2, and let a = v/6. The function

( ( - a)ke~vk, for 6k - a~k < t ^ 6k (jfe = 1, 2, 3, • • • ),

yp{t) = \
\ 0 , elsewhere on (0, 1),
is integrable (absolutely) in the sense of Lebesgue. But for any
positive integer n it is found respectively that

f yp{t)e-«ntdt\ ^ f liKfll dt

= £ tr'k < e-*n+1 JTe-*1'= e->n+1 Ç | *(*) | dt;

k=n+l 3=0 J0

that

I r1
^ e _ a » (() »- i _ a -«+ i ) I | ^
Jo
| dt

Jo
and that

t(t)e-«ntdt
ƒ'n
'e -oT
= ( - ! ) » ( * - l)e~
Jen
Hence it follows that

| e2"nM>oi(- a-) - (e- 1 ) ( - 1)»|

< |^ w (2-,) + ^(2-i/ô)+«j r |^ | dtf
J0
193*1 EXPONENTIAL SUMS 231

and since the quantity on the right approaches zero with 1/n it
must be concluded that ^o.iO3) takes when z = —an (and n is
sufficiently large) the sign of ( —l) n . The existence of at least
one zero in each corresponding interval — an+l ^ z ^ — an is thus
established.
12. Hypotheses of Specific Type. While the example of the
preceding section shows that the distribution of zeros of an ex-
ponential integral may differ sharply from that of an exponen-
tial sum, yet in a very wide class of cases they show considerable
similarity. The coefficient function \p(t) is of course the deter-
minative factor of the integral, and hence is always the direct
subject of the hypothesis. The substance of the present section
is essentially an abstract of results which are due to Pólya 1 and
is characterized by the fact that the hypotheses are on the
whole of rather specific character. The results are accordingly
precise to a considerable degree of detail.
If the function \p(t) is a step-function with discontinuities
finite in number and occurring only at rational points of the
interval, the function will be called exceptional. With such a
function as coefficient the integral may be directly evaluated
into the form of an exponential sum of the elementary type
considered in §2. Hence no essential loss is involved in the re-
striction of the theorems to unexceptional functions, and this
will be tacitly assumed in the absence of statements to the con-
trary.
THEOREM 9. If the function \//(t) is positive and non-decreasing
on the interval (0, 1), then the zeros of the integral ^o,i(z) are all
located in the open half-plane x<0.
By way of proof let the value z=iy be first considered. For
any value t\ on the interval of integration it is seen that

iy I ^e^dtl ^ $(h) - ^(0) - I eiy'#

Jo I 1 J0
But if 0 is defined by the relation

I eiytd\// = e~iB j eiytd\l/,

J0 I «^o
then as h~>1 the relation above becomes
232 R. E. LANGER [April,

| iy&oi(iy) | è lim I {1 — cos (yt — 6)}d\[/.

Since d\p is never negative, the conclusion l^oifeO | > 0 will be

expected, and is in fact correct. Its derivation is, however, im-
peded somewhat by the necessity of recognizing and disposing
of the potential effect of the functional dependence of 6 upon
the variable t\. In conclusion the substitution of z = x+iy in the
discussion given merely replaces \f/(t) by \f/(t)ext> and since this
satisfies the hypotheses made for any fixed non-negative x, the
assertion of the theorem is established.
The result obtained admits of ready extension into the follow-
ing form.
T H E O R E M 10. If the coefficient \f/(t) is positive and continuous,
and if \p'(t) exists (except possibly at a finite number of points) and

a Ik - — - ^ 0,

then the zeros of the integral ^o.i(^) &*& confined to the open strip
of the complex plane given by the relation
a < x < p.
(The trivial case \{/(t) =ect+d is excepted.)
The transformations of the integral into the forms

^o.i(z) = { e^^e^-^'dt = ez \ e~ai^(\ - t)e^~z)tdt

*J o *o
are easily verified. In these forms the coefficients efit\[/(t)t and
e~ott\f/(l — t), satisfy the hypothesis of Theorem 9, and hence the
zeros are simultaneously restricted by the relations
R(z - p) < 0, R(a - z) < 0.
From this the theorem follows.
T H E O R E M 11. If the function f(t) is positive and non-decreasing
on the interval (0, 1), then the zeros of the trigonometric integrals
^s(z) and ^c(z) defined in (31) are all real and simple. The zeros
of tyc(z) occur singly in each of the intervals
[ ( W - 1 / 2 ) T T , (W + 1 / 2 ) X ] , m^O,
w-l EXPONENTIAL SUMS 233

while those of ^«(2) occur (except for z = 0) singly in each of the

intervals [tnw, (m+l)w], w ^ O , — 1 .
The facts of this theorem may be established by suitable pas-
sages from finite to corresponding infinite sums as is indicated in
the following. The trigonometric sum

*.(*) = — £ ƒ (-) *»"* sin (-z)

n j=0 \n/ \n /

has, for each value of n, coefficients which are positive and in-
creasing. Its zeros, therefore, by Theorem 2, are real, and alter-
nate with those of the corresponding sum of cosines. From this
it may be concluded, with suitable reasoning as n-^<x>, that the
zeros of the integrals ^fs(z) and ^c(z) are real, and that a multiple
zero of the one would necessarily be also a zero of the other.
Since in this latter event the integral

¥,(») + *¥.(*) = f f(t)eiztdt

Jo
would have a real zero, contrary to Theorem 9, the simplicity
of the zeros must be granted.
The more precise location of the zeros as stated in the theorem
is derived by Pólya from the expansion
i>wftw
(a» i « - no cm - ± <- rj-+-i-n,
z sinz M->OO ( z J== i JTT Lz—jir z+jirj)
which may be established by methods familiar in the theory of
functions of a complex variable. Each coefficient on the right of
this expansion is positive, for

( - l ^ t f . O V ) = ( - 1)' + 1 f f(t) smjirtdt,

Jo
while the integral represents a sum of areas, alternating in sign,
non-decreasing in numerical magnitude, with the final one posi-
tive. Because of this the rational function within the brace in
(32) is found by elementary means to have one zero in each of
the intervals (— 00, —mr), (rnr, 00), [mw, (w+l)7r], m=—ny
• • • , — 1, 1, 2, • • • , n — 1. A suitable passage to the limit as
234 R. E. LANGER [April,

n—>oo establishes the theorem. An analogous discussion for

SPcOs) is, of course, possible.
If the hypothesis of Theorem 11 is extended to include the
assumptions that limt^of(t) = 0, and that f(t) is nowhere con-
cave, a refinement of the theorem is possible which determines
the specific half of each of the intervals given in which the zero
is located. The method of proof is similar to that outlined above,
a remark which applies also to the following theorem.
T H E O R E M 12. If the function f (t) is positive, decreasing, and
nowhere convex downward, and if f'(f) exists and is not an excep-
tional function, then the zeros of ^c(z) are all real and there is just
one in each of the intervals \mir, (W + 1 ) 7 T ] , W T ^ 0 , — 1 .
As examples of integrals covered by the theorems of this sec-
tion one has the familiar functions

r1 2
Jo(z) = I cos ztdt,
Jo TT(I - t*yi*
f1 2/
=
Ji(z) I sin ztdt.
Jo *r(l - Z2)1'2
13. Hypotheses of Intermediate Type. Less specific than those
of the preceding section are the hypotheses and results to which
the attention will now be turned. The material has been taken
essentially from a memoir by Miss M. L. Cartwright 10 under date
of the present year. The two theorems to be given first, how-
ever, are attributed by her to Hardy.
As a preliminary it is perhaps desirable to note at this point
the following relations which have been proved by Titchmarsh, 9
namely that with any coefficient function \{/(t) which is inte-
grate,

(33) l^2)l{^otri-N»),forany0>0.

These assertions may be established by a direct analysis of the

integral involved.
T H E O R E M 13. If the function \[/(t) is itself an integral, if it is
continuous at t = l and t= — 1, and if \(/(l) = ^ ( —1) = 1, then the
W-I EXPONENTIAL SUMS 235

zeros, zm, of the integral SP-i.iOs) are given by the asymptotic for-
mula

nnri.
The proof is immediate. An integration by parts yields the
equality

g ^ - i fl (2) = ez - e~z — Ifj\t)e zt

\p'(t)e dt.

Of this the right hand member is of the form

* { 1 + e(a)} - r - { l + €(«)},
because of the first of the relations (33). T h a t the zeros are
located as stated in the theorem may be concluded then from
the deductions of §2 and §4.
T H E O R E M 14. If the function \f/(t) is of bounded variation, is
continuous at t = l and t=— 1, and yp(l) — yp{ — 1) = 1, then the
zeros of ^-it\(z) all lie within a strip of the complex plane given by
a relation \x \ <K.
The proof of this theorem has been constructed on the follow-
ing lines. When x>0, an integration yields the evaluation

eztd\p — I eztdyp,
-I Ji-$
in which the integrals on the right are numerically less, respec-
tively, than Hexa~8), and exrj(ô), with H some constant and
7)(ô) the maximum of \yp(t2)-yp(h) | for 1 - ô<h<t2^l. Hence
| 2^_i,i0) | ^ ex{ 1 — e~2x — He~hx — rj(d)} .

However, the right member of this inequality is assuredly posi-

tive if S is chosen sufficiently small (because of the continuity of
x[/(t) at t = 1 this makes rj(ô) small), and if K is then chosen suffi-
ciently large and x >K. It evidently follows that for such values
of z the integral can have no zero, and a similar discussion estab-
lishes the fact for x < — K.
The hypotheses of Theorem 14 are seen from the conclu-
sions thus drawn to be sufficient to assure for the exponential
integral a distribution of zeros resembling in at least one salient
236 R. E. LANGER [April,

feature the distribution associated with an exponential sum. By

the following theorem this similarity is shown to extend beyond
the mere restriction of the zeros to a strip of the plane even to
the density of zeros within that strip.
T H E O R E M 15. Under the hypotheses of the preceding theorem
the number, n(r), of zeros of the integral ty-ifi(z) which are in abso-
lute value less than r satisfies the relation

2r
n(r) = — + 0(1).

It is readily found that this relation is included in (7) which

restricts the zeros of exponential sums.
The proofs of Theorem IS, and of the following Theorem 16,
are less simple than any which have hitherto been outlined.
They are based upon facts and relations familiar in the theory of
entire functions and involve considerable detail. The author of
the theorem 10 has deduced also a formula for the value of the
quantity
f'n(s)
I ds.
JQ S

T H E O R E M 16. If the function ypit) is continuous, with modulus

of continuity co(ô) on the interval (—1, 1), andif^/(\) = ^ ( — 1) = 1,
then the zeros of the integral ty-i,i(z) all lie within a region

| x\ < K\ Z\ co
T^T
with a suitably chosen constant K> and

n(r) = — + O(rco(—YV

It will be recalled that the modulus of continuity co(S) is de-

fined as the maximum value of \i/(h) —^(h) | for t2, h on ( — 1, 1)
and | / 2 ~ ^ i | < ô . As ô—»0, co(ô) =0(1), and in fact if the function
satisfies a L i p s c h i t z c o n d i t i o n |^fe) —tih) | <A \h — t\ | a ,
I93I-] EXPONENTIAL SUMS 237

0 <a ^ 1, then w(5) = 0(ô a ). As in the case of the preceding theo-

rem an evaluation of the quantity

Jo s
is also deducible. Moreover, it can be shown that the require-
ment of continuity over the entire interval ( — 1, 1) can, with
suitable modification in the results, be replaced by hypotheses
demanding continuity only near the end points.
14. The Integral with Most General Coefficient. The exponen-
tial integral of greatest generality results from the admission of
any function which is absolutely integrable to the role of the
coefficient ip(f). As will be indicated below even this minimum
of hypotheses is not unproductive of significant results.
The function St'o.iO2) is analytic over the entire plane, and as
such its zeros may be enumerated in the order of increasing
absolute magnitude, thus, zm = rmeidm. The following theorems
involve a qualitative characterization of the density with which
these zeros are distributed in the complex plane.
T H E O R E M 17. If the coefficient function \p(t) is absolutely in-
tegrable, the zeros zm of the integral ^o,i(z) are distributed so that
the series

1+
m—li II Z
« m II «

converges for every € > 0 , and diverges for e = 0.

As an entire function the integral ^o.iO3) is of order at most

unity, as may be seen from the evident relation

|*o.i(*)| S e |f| f \m\dt.

Jo
Inasmuch as this order at least equals the exponent of conver-
gence, the statement of the theorem in so far as it concerns the
convergence for € > 0 is established. The divergence for e = 0 is
not so immediately verified. Independent proofs of it have, how-
ever, been given by both Pólya 1 and Titchmarsh. 9 The method
employed by the latter may be indicated briefly as follows.
238 R. E. LANGER [April,

Convergence of the series for e = 0 would imply the absolute

convergence of the canonical product

POO -*oi(o)fi f 1 - - ) -
In such event, however, a product of this type is known to
satisfy the relation

with any positive constant rç. Since ^0,1(2) is expressible in the

form ^0,1(2) = e^a+i^)zP{z)i its order being unity, it would follow
that
| * o . i ( ± x)\ = 0(*<±«+'>l*'),
and this contradicts the second of t h e relations (33).
T H E O R E M 18. If the function \f/(t) is absolutely integrable, then
the zeros of the integral ^0,1(2) satisfy the relations
r
(a) n{r) ~ — >
n
A cos 6m
(b) 2^f 1 r converges absolutely,

(c) 23 -j p converges conditionally.

m=l I 2m I

These results have all been obtained by Titchmarsh 9 by the

use of relations in the theory of entire functions. The result
(a) evidently implies the assertions of Theorem 17, and is in
fact more precise. The result (b) in conjunction with (a) indi-
cates that in general | c o s 0 w | is small when m is large.
The example of §11 would show, however, that a relation
limm^oo cos 0W = O is not in general valid. The result (c), which
is obvious in the case of a real coefficient function ^(/), since
the zeros present themselves then in conjugate pairs, is not
easily established in the general case. The author of the result 9
has, however, succeeded in explicitly summing both the series
(b) and (c) in terms of the values taken by S^o.iO2) and its de-
rivative at 2 = 0.
I93I-] EXPONENTIAL SUMS 239

LIST OF REFERENCES
1. Pólya, G., Über die Nullstellen gewisser ganzer Funktionen, Mathema-
tische Zeitschrift, vol. 2 (1918), pp. 352-383.
2. Wilder, C. E., Expansion problems of ordinary linear differential equa-
tions, etc., Transactions of this Society, vol. 18 (1917), pp. 415-442.
3. Tamarkin, J. D., Some general problems in the theory of ordinary linear
differential equations, 1917', Petrograd; and Mathematische Zeitschrift, vol. 27
(1927), p. 24.
4. Birkhoff, G. D., Boundary value and expansion problems, etc., Transac-
tions of this Society, vol. 9 (1908), p. 373.
5. Langer, R. E., The asymptotic location of the roots of a certain transcen-
dental equation, Transactions of this Society, vol. 31 (1929), p.837.
6. Pólya, G., Geometrisches über die Verteilung der Nullstellen, etc., Sitzungs-
berichte der Bayerischen Akademie, München, 1920, p. 285.
7. Schwengler, E., Geometrisches über die Verteilung der Nullstellent etc.,
Dissertation, Zurich, 1925.
8. Ritt, J. F., A factorization theory, etc., Transactions of this Society, vol.
29 (1927), p. 584.
On the zeros of exponential polynomials, and algebraic combinations of ex-
ponentials, Transactions of this Society, vol. 31 (1929).
9. Titchmarsh, E. C , The zeros of certain integral functions, Proceedings of
the London Mathematical Society, vol. 25 (1926), p. 283.
10. Cartwright, M. L,, The zeros of certain integral functions, Quarterly
Journal of Mathematics, vol. 1 (1930), p. 38.
T H E UNIVERSITY OF WISCONSIN