S. Mandelbrojt (Auth.) - Dirichlet Series - Principles and Methods (1971, Springer Netherlands)

DIRICHLET SERIES
S. MANDELBROJT
DIRICHLET SERIES
Principles and Methods
D. REIDEL PUBLISHING COMPANY I DORDRECHT-HOLLAND

Library of Congress Catalog Card Number 75-170339
ISBN-13: 978-94-010-3136-3 e-ISBN-13: 978-94-010-3134-9

DOl: 10.1007/978-94-010-3134-9
All Rights Reserved

Copyright © 1972 by D. Reidel Publishing Company, Dordrecht, Holland
Softcover reprint of the hardcover 1st edition 1972
No part of this book may be reproduced in any form, by print, photoprint, microfilm,
or any other means, without written permission from the publisher
To the memory
o/Gladys
PREFACE
It is not our intention to present a treatise on Dirichlet series. This part

of harmonic analysis is so vast, so rich in publications and in 'theorems'
that it appears to us inconceivable and, to our mind, void of interest to
assemble anything but a restricted (but relatively complete) branch of the
theory.
We have not tried to give an account of the very important results of
G. P6lya which link his notion of maximum density to the analytic
continuation of the series, nor the researches to which the names of
A. Ostrowski and V. Bernstein are intimately attached. The excellent book
of the latter, which was published in the Collection Borel more than thirty
years ago, gives an account of them with all the clarity one can wish for.
Nevertheless, some scattered results proved by these authors have found
their place among the relevant results, partly by their statements, partly
as a working tool.
We have adopted a more personal point of view, in explaining the
methods and the principles (as the title of the book indicates) that originate
in our research work and provide a collection of results which we develop
here; we have also included others, due to present-day authors, which
enable us to form a coherent whole.
The reader will find chapters that are not connected with others. This
is the case in Chapter IV, which is really only concerned with Taylor
series, but is of a general character and, we believe, important enough to
draw attention to.
Chapter I is elementary - it only contains the classical results that are
indispensable for the techniques to be developed later.
Chapter X plays a special role: in the book itself (the first nine chapters)
we do not give any bibliographical indications; it is Chapter X that pro-
vides the history of the results demonstrated in the preceding chapters,
with a number in square brackets indicating the publication where the
result in question is to be found.
Certain statements, for example part of those in Chapter V - a chapter
devoted to theorems on the composition of singularities - reproduce in
VIII DIRICHLET SERIES
an improved form results contained in our Rice Institute Pamphlet, which

has been out of print for several years.
As to the guiding ideas or principles which have served almost through-
out the other chapters, we have to point first of all to what one could call
the 'principle of arithmetic isolation' of the exponents, which consists
in the following: the exponents whose fractional parts form a set that
can be isolated by certain neighbourhoods around the fractional parts
of the 'useful' exponents, neighbourhoods whose extent depends on the
problem in question, provide autonomous information that cannot be
destroyed by the other exponents; these exponents, which are isolated
(by their fractional parts) from the others, already determine in a rather
important manner the behaviour of the analytic continuation of the
series.
Another principle consists, as we have already seen in other branches
of analysis - by making the adherent series intervene - in inequalities
concerning the coefficients. These inequalities use, when the exponents
are given, the maximum of the function in a well-defined part of the plane;
this part of the plane being defined as to its form or its dimensions (or
both) by the arithmetic character of the exponents or by their growth.
I wish to thank Miss M. Caron and her collaborators at the College de
France for the perfect way in which the typescript has been prepared.
April 1969 s. MANDELBR01T

CONTENTS
PREFACE VII
CHAPTER 1/ SEQUENCES OF EXPONENTS AND ASSOCIATED

SEQUENCES. ELEMENTARY THEOREMS ON THE COEFFI-
CIENTS AND ON CONVERGENCE
1.1. Ascending Sequences of Positive Numbers 1
1.2. General Properties of Convergence 7
1.3. The Calculus of Coefficients and Some Important
Combinations of Coefficients 14
CHAPTER II / INEQUALITIES CONCERNING THE COEFFICIENTS 18

H.I. Inequalities Corresponding to the Arithmetic Character
of the Exponents 18
11.2. A General Inequality for the Coefficients Corresponding
to Finite Upper Density of the Exponents 27
CHAPTER III/ THEOREMS OF LIOUVILLE-WEIERSTRASS-PI CARD

TYPE OF ARITHMETICAL CHARACTER AND OF GENERAL
CHARACTER 31
III.I. Theorems of Arithmetical Type 31
IH.2. Liouville-Picard Theorems of General Type 35
CHAPTER IV I SINGULARITIES OF FUNCTIONS REPRESENTED

BY A TAYLOR SERIES 44
IV.1. Singularities of Taylor Series 44
CHAPTER V / COMPOSITION OF SINGULARITIES 54

V.I. Composition of Hadamard Type and Generalizations 54
V.2. Composition of Functions of Slow Growth 62
V.3. 'Fictitious Composition' of Singularities 75
V.4. Composition of Functions of Rapid Growth 79
V.5. Composition of 'Hurwitz Type' 89
x DIRICHLET SERIES
CHAPTER VI/SOME APPLICATIONS OF THE PRINCIPLES FOR

ANALYTIC CONTINUATION 99
VI. 1. Arithmetic Properties of the Exponents and the Analytic
Continuation 99
VL2. Analytic Continuation of General Dirichlet Series 102
VI.3. Entire Functions Represented by Dirichlet Series 106
VI.4. Some Applications of the Composition Theorems 108
CHAPTER VII / ON THE BEHAVIOUR OF THE REMAINDERS OF

A DIRICHLET SERIES IN THE DOMAIN OF EXISTENCE OF
THE FUNCTION. APPLICATIONS 116
VII. 1. Evaluation of the Remainders of (A, A) 116
VII.2. Applications of the Results of Chapter III and of VLl to
Series Meromorphic on an Interval of the Axis of Con-
vergence 124
CHAPTER VIII / APPLICATIONS TO THE GENERALIZED RIE-

MANN FUNCTIONAL EQUATION 129
VIlLI. Number of Independent Solutions. Links between Ex-
ponents 129
CHAPTER IX / INFLUENCE OF ARITHMETICAL PROPERTIES OF

THE EXPONENTS OF A DIRICHLET SERIES ON ITS ANALY-
TIC CONTINUATION 139
IX.l. Isolated Fractional Parts of the Exponents and the Possi-
bility of Analytic Continuation 139
IX.2. Relations Between Isolation of the Fractional Parts of
the Exponents and the Distribution of Singularities 149
IX.3. Variation of the Exponents and the Analytic Continuation 154
CHAPTER X / BIBLIOGRAPHICAL NOTES 156
NOTE 160
BIBLIOGRAPHY 164
CHAPTER I
SEQUENCES OF EXPONENTS AND ASSOCIATED

SEQUENCES, ELEMENTARY THEOREMS ON THE
COEFFICIENTS AND ON CONVERGENCE
1.1. Ascending sequences of positive numbers

Let A = {AnHn ~ 1) be a strictly increasing sequence of positive numbers:
0< Al < A2 , , • tending to infinity, The quantity
D = limsup~
An
is called the upper density l of A, and the quantity
h = lim inf (An +1 - An)
the step of A.
If h > 0, then D'h ~ 1, because for every e with 0 < e < h there is an ne
such that for n>ne we have: An-An.>(n-ne-l)(h-e), hence
D' ~ (h - e)-I,
Denote by N(x) the number of An less than x (x> 0) and let D(x) =
N(x)/x. Then D(x) is the density function of A, We also have
D' = lim sup D(x),
The function
15(x) = ! ("lC D(y) dy (x > 0)

x Jo
is called the function of mean density of 11, and the quantity
15' = lim sup 15(x)
is the upper mean density of J,

It is easy to see that 15' ~ D', But one can construct sequences for which
15' < D', Here is an example with D' =], 15' = e- I : let {vn} be an ascending
sequence of positive integers and set An=Vk+(n-Vk-I)/(Vk-Vk-l) for
2 DIRICHLET SERIES
Vk-l <n~vk (k~ 1, vo=O), Here D(x) < 1 (x>O) and from AVk=Vk+ 1 it
follows that
D' ~ I'1m Vk Vk
+ 1 = 1,
that is, D'=1. Also, for vp~x<vP+l (p~l)
(since AVp +1 =VP+l + (vP+l_Vp)-l) we have
D(x) ~ ~
x
xD(x) = k~-l i:k+l D(y) dy +i: D(y) dy,

From these relations for Vp ~ x < Vp + 1 and from D(x) < 1 (x> 0) it
follows that
And since log xlx~ lIe for x>O, we can write
If we choose,2 for example, Vn =E[exp (exp n)], then for Vp ~ x< vp+l:
that is,
lim sup 15(x) = 15' = e- 1 < 1 = D',
Observe that always D' ~ e15 (this remark is due to Deny and Dvoretzky),
In fact, since D(x) ~ nIx for x> An, we can write
15(x) ~~ [J:n D(y) dy + n log (xl An)],

and by choosing x=eAn we obtain
SEQUENCES OF EXPONENTS 3
that is,
15' = lim sup D(x) ~ lim sup D(eAn)
~ lim sup n/(eAn) = D'/e,
The preceding example shows that in this inequality e cannot be re-
placed by a smaller constant.
We also see that D' and D' are simultaneously equal to zero, finite,
or infinite,
If D' <co, then the products (n> 1)
Pn = n IA~ - A;I/A~
m;'n
converge; Pn>O, The convergence of the product is an immediate con-
sequence of the fact that An>kn with k>O, The sequence {An}, where
An =P; \ is said to be associated to A,
THEOREM 1.1.1. There is a positive function A(a), defined and finite for
a> 1, such that for h>O
lim sup An/An ~ A(a)D' - 2a log (hD') , D' = L(a; h, D').
When D' =0, the expression L(a; h, D') must be taken to be zero.
For A(a) we can choose the following function:
A(a) = a{4 + log [(a 2 - 1)/a3 ] + log [(a + 1)/(a - 1)]/a}.

By taking, for example, a=3/2 we have: 2a=3 and A(a) < 8.5.
Thus, in particular:
lim sup An/An ~ [8.5 - 3 log (hD·)]D·.
If m is a positive integer and p is fixed, we set fLn = Ap + n (n ~ 1) and

write:
P = Pm,p = min (fLn + 1 - fLn), (flo = 0),
O",n",N
where N=Np(afLm), a> 1 being fixed and Np(x) being the number of fLn
less than x, In other words, for X>Ap+l we have Np(x)=N(x)-p, and
for X~Ap+l we have NvCx) =0.
We denote by {Mn} the sequence associated to {fLp} and set
4 DIRICHLET SERIES
where in a simplified notation with /Lm = /L:
Am = n 11 -
q';;N
q¢m
/L:
/Lq
I
B = n 11 -/L21
-·
m q >N+ 1 /L~
Now
log Am = 2:
q.;;N log 11
q¢m
-I /L + t/Lq
2
_
/L)121
~ 2:
/c';;m-l
log [(
/L
~ k P) 2 - 1]
+ 2:
k';;N-m
log [ 1 - ( /L )
./L + kp
2] .
Since the function 1-[/L/(/L+xp)]2 increases for x> -/L/p, if we bear
in mind that - /L/ P < 1- m, we see that
log Am ~ f-m log [(/L/(/L + pX»2 - I] dx
+ LN-m log [I - (/L/(/L + pX»2] dx
= LN_-mm log [plxl(2/L + pX)/(/L + pX)2] dx
~ p-l j P(N-m) log [Itl/(/L + t)] dt

p(l-m)
~ (N - l)[log p - log (/L + (N - m)p)]
+ p-l jP(N-m) log (/tl!/L) dt.

p(1-m)
As Np=Nm(a/L)p~a/L, we have It I~a/L for It I~pN. Also
log [1 _(N ~ m)p] ~ (N ~ m)p,

which allows us to write (taking account of the inequality Np~alL):

[PN
log Am ;>-: 2p-l Jo log (t/alL) dt
+ (N - 1) log a - N(N - m)PIL- 1

[P'l)(GIl)
;>-: 2a1LP-1 Jo log T dT
+ (N - l)loga - a(N - m)
= [2 log {p~(alL)}
- (2 +a- log a)]N + ma - log a,
where ~(x) is the density function of {lLn}.

On the other hand, when we integrate by parts, we find
log Bm ;>-: fIX) log (I - 1L2 /X 2 ) dNm{x)

all
IX) ~(x)
= -log (1 - a- 2)NialL) - 21L2 f 2 2 dx
all X - IL
for from D' < 00 we have Nm(x) = O(x) (x ~ 00) (evidently the sequences
{lLn} and A have the same upper density D') and
lim log (1 - p,2/X2)Nm{x) = 0,
Consequently
IX) dx
log Bm ;>-: -log (I - a- 2)N - 2p,2~'{ap,) f 2 2
all X - P,
= -log (I - a- 2)N + p,~'{ap,) log [(a - 1)/(a + 1)],

where
~'(x) = sup ~(y),
y~X
log Qm ;>-: [2 log (p~{ap,» - (2 +a - log a)]N

- log (l - a- 2)N + 1L~'{ap,) log [(a - 1)/(a + 1)],
Now the following relations are obvious:
lim log (An +p/Mn)/lLn

n-toOO
= 0, p-+co
lim Pm,p = h; lim~'(x) = D'
6 DIRICHLET SERIES
°
and if D' > 0, no matter what e > is taken, we have n//Ln ~ (1 + e)D' for
sufficiently large n; hence pn//Ln(1+e)p~pD' e-1~rl; and since xlogx
decreases in ]0, e -1] :3
[pn//Ln(1 + e)e] log [pn//Ln(1 + e)e] ~ (hD'/e) log (hD'/e),

Therefore it is enough to let p tend to infinity to see that for every ~l> 0:
lim sup An/An ~ L(a; h, D') + 0,
The statement of proposition 1.1.1 follows from this, The case D' =0
is easier to derive, starting out from the inequality above for log Qm.
The next proposition gives another property of a sequence with finite
upper density, which is useful in what follows,
THEOREM 1.1.2, Suppose that D' < co and set for r> 0:
Ao(r) = n (1 + ~;),
This product converges uniformly on every compact set in [0, co[, and
r1m sup log Ao{r)

r = 7T ]j' ,
The uniform convergence of the product is a consequence of the fact

that An> kn with k > 0,
For r>O we have:
log Ao{r) = L log (1 + ~;) = f' log (1 + ~2) dN(x),

and integration by parts yields
log Ao(r) = 2r 21'"o D(x)+ dxr

X
2 2;
the terms that are not under the integral sign disappear on integrating
by parts, because on the one hand N(O) = 0, and on the other hand
lim N{x) log (1 +r2/x 2 ) =0, since D' < 00,
Integrating by parts a second time we obtain:
This enables us to write for x> 0:
log 11o(r)/r = -2r J: yD(y) d(y2 ~ r2)

-2r f' yD(y) d{y2 ~ r2)
~ 2 max D(y)x3 /r(x 2 + r2) + 7T sup D(y);
O~y:::;x y;::.x
and since lim SUPy;.t D(y) = D', we obtain the required inequality.
1.2. General properties of convergence

Let 11 be a sequence with strictly increasing positive terms tending to
infinity, and A = {an} a sequence of complex numbers.
The series
(Ll) 2: an e- ilnS ,
where s=a+it (a and t being real), is called a Dirichlet series. For given
sequences 11 and A = {an} we denote the series (Ll) by (A, 11).
If in a Taylor series (with ao=O)
(1.2) 2: anzn
we set z=e-s, then the series becomes a Dirichlet series
(1.3) 2: an e- ns
which we call a Taylor-D series. Here >'n=n (n~ 1).
Another example of a Dirichlet series is
(1.4) 2: ~ = 2: e-(lognls
which represents the Riemann Zeta-function '(s) , a function that plays
a principal role in the theory of the distribution of prime numbers.
Let us recall that a Taylor series has a radius of convergence R, which
may be zero, infinity, or a finite positive number. In the first case the series
converges for z=O only; if O<R<oo, the series converges for Izl <R,
and does not converge for any value z with Izl > R. The disk Izl < Rand
the circle Izl =R are, respectively, the disk and the circle of convergence
of the series.
8 DIRICHLET SERIES
In any case R is given by the formula
Moreover, it is known that the series converges uniformly on every com-

pact subset of the disk of convergence; in the disk it converges absolutely.
Since the transformation z=e- S gives u= -log 14 t= -arg z, we see
that when z varies on a curve L in the interior of the disk Izl < a, then s
varies on a curve L' = -log L situated in the half-plane u> -log a. (If
we associate with a point Zo of L the point So = Uo +ito with 0:::;; to < 211',
then the curve L' is well-defined, for we agree to let s vary continuously
when z varies on L.)
The series (1.3) then converges absolutely as long as
(1.5) u > lim sup log lanl = Uc

n
(if Uc< co), and does not converge for u < Uc (provided that - co < uc).
The series also converges uniformly on every compact set of the half-
plane (1.5). The series (1.3) does not converge anywhere if uc=co.
What we have just said concerning the series (1.3) cannot be trans-
ferred directly to every Dirichlet series simply by replacing n by "n.
From
the point of view of convergence, the essential difference lies in this: that
different half-planes of convergence do not coincide.
We begin with the following theorem:
THEOREM 1.2.1. If Lan e- ilnso (so=uo+it o) converges, then the series

(A, A) converges uniformly on every closed angle given by
larg (s - so) I : :; y < 11'/2.
We set S=So+S', s' =u' + it', a' >0, larg s'l <y, and we write
2: an e-
n.;m
ilnso = Am(so) = Am
We have
a' > 0, It'l/u':::;; tan y = M < co.

For s=so+s' and q>p~2 we have:

L
p"n"q
all e- hnS = L
p"n"q
an e-h"so e- hnS'
L (An
= p",n",q - All-I) e- AnS'
= L
p"n",q
[(All - S) -- (A n - l - S)] e- hns'
= L
p"n",q
[(An - S) - (A n- l - S)] e- hnS'
= L
p"lI .. q-1
(An - S)(e- hnS' - e- hn+1S,)
+ (Aq - S) e-hqs' - (A p - 1 - S) e-hps'.

Given e>O we choose p so that IAn-SI <e for n;:::p-l; then we have:
(1.6) L
p"'n"q
all e-hnS ~ L
p"n"q-l
IAII - SI·le- hns' - e- hn + 1s'l
+ IAq - SI e-hq'" + IA p - 1 - SI e-V'
~e L le-A"s' - e-hn+lS'1 + 2e.
p"n",q-l
Now
le-hns' - e-hn+lS'1 = IS'I\ (,'n+l e dul ~ Is'l

US ' [\'+1 eU'" du
Jh ll Jh ll
= 1;:1 (e- hn'" _ ehn+l"')
~ (M + l)(e-hn'" - e?!n+l"'),
and it follows from (1.6) that
ILp"'n"q
e- lI "sl ~ 2e + e(M + 1) L p",n",q-l
(e- 1I11 '" - elln+l"')
~ e(3 + M),
which proves the theorem.
Here is an immediate corollary to the preceding theorem.
THEOREM 1.2.2. If Lan e-ilnSo (so=uo+it o) converges, then the series

Lan e- llnS converges at every point s=u+it with U>Uo, and converges
uniformly on every compact set contained in the half-plane u> uo.
10 DIRICHLET SERIES
If we replace the series (A, A) by

(I.7) 2: lanl e- llnS ,
Theorem 1.2.2 permits us to assert
THEOREM 1.2.3. If the series (A, A) converges absolutely for S=So = CJo +ito,
it also converges absolutely for s=CJ+it with CJ> CJo.
Consequently,
THEOREM 1.2.4. If (A, A) converges at any point of the plane, two cases
are possible: (1) (A, A) converges for every value of s, (2) there exists a
number CJc (-OO<CJc<OO) such that (A, A) converges for every s with
CJ> CJc and does not converge for any s with CJ < CJc'
If (A, A) converges absolutely at any point of the plane, two cases are
possible: (1) (A, A) converges absolutely in the whole plane, (2) there
exists a number CJa (- <X) < CJa< (0) such that (A, A) converges absolutely
for CJ> CJa and (A, A) does not converge absolutely for any point of the half-
plane CJ<CJa.
If (A, A) does not converge at any point, we write CJc=OO.
If (A, A) does not converge absolutely at any point, we write CJa = 00.
If (A, A) converges at every point, we write CJc= - 00, and if it con-
verges absolutely at every point, we write CJa = -<X).
The quantities CJc and CJa are called, respectively, the abscissa of con-
vergence and abscissa of absolute convergence of the series (A, A), and the
straight lines CJ=CJc and CJ=CJa are, respectively, the axis of convergence
and the axis of absolute convergence of (A, A). Evidently CJa~CJc' For a
Taylor-D series ('\n=n) we have CJa=CJc, but this is not true in general.
It is known, for example, that the series
2: (-l)nn- S = 2: (-l)n e-(lognls
converges for CJ> 0 and does not converge for CJ < O. By contrast, the
series 2: n- S converges for CJ> 1 and does not converge for CJ~ 1. For the
series 2: (-l)nn- S we have CJc=O, CJa= 1.
Furthermore, a Dirichlet series may converge in the entire plane and
not converge absolutely anywhere. Here is an example.
We set an=(-I)nn-1, '\n=(log log n)1/2 (n~3). For k>O we have the
following relations:
lim an ellnk = 0
lanl ellnk < lan-ll elln - lk (n > nk)'
The first relation is evident, the second can be written in the form
-log (1 -~) > k [(log log n)1/2 - (log log {n - 1))1/2]
(n > nk)
{this follows from log ({n-l)/n)- -n- 1 and the relation
{log log n)1/2 - (log log {n - 1))1/2 - (2n log n(log log n)1/2)-1).
Hence the series L: an e-""k converges for every k, and this proves that
O'c= -00.
On the other hand, it is obvious that L: n - 1 e - (lOg log n)1/2 k diverges for
every k, hence O'a=OO.
Note that in our two examples
(An = log n,
we have D' = 00. But the following elementary proposition holds:
THEOREM 1.2.5. If D' < 00, then
O'a = O'c = lim sup 10gAlanl .

n
Suppose that n/An~L-1<00{n~1) and suppose, to begin with, that

lim sup (log lanl/An)=a<oo. If b>a, we have for n>nb: lanl ~e""b and
for O'>b:
Consequently L: lanl e-",,"<oo for O'>a. On the other hand, if

- 00 < a < 00and - 00 < 0' e""b and lanl e-"""~e",.(b-"» 1 (n> 1); hence the series (A, A)
does not converge for 0' < a. This proves the proposition.
In the proposition we have just proved the condition D' < 00 can be
replaced by the rather less restrictive condition: lim log n/ An = O. This
result (which is due to Valiron) will not be proved here.
But we prove the following theorem:
THEOREM 1.2.6. For every Dirichlet series

. logn
O'a - O'c ~ hm sup -A-'
n
12 DIRICHLET SERIES
°
We denote by b the value of the expression on the right. We have to
show that for every e >
(1.8) 21an l e-A,,<a.+b+B) < 00.

From the convergence of (A, A) for s=uc+e/2 it follows that
and
but for n> nB : log n/ "" ~ b +e/4, which proves that the general term of
(1.8) for n>ns is smaller than An-<b+B/2)J(b+B/4), and so the theorem is
proved.
Finally, here is a theorem which gives the abscissae of convergence
for an arbitrary sequence A:
THEOREM 1.2.7. We set
If a<oo, the series (A, A) converges for u>max (0, a), and if a>O,
then uc=a.
We set
b = lim sup (log 2:

m .. n
laml)/~n.
If b < 00, then the series (A, A) converges absolutely for u> max (0, b),
and if b>O, we have ua=b.
The part of the theorem concerning absolute convergence is an im-
mediate consequence of the first part if in (A, A) we replace an by lanl.
We set An= 2m"" am. For e>O and n>nB we have
if a> -00, and log IAnl/'\n<e/2 if a= -00. Hence
(1.9) IAnl < eh,,<a+S/2), or IAnl < eh.,<B/2) (if a = -(0).

We set a' = max (0, a). Then for q > p ~ 2 we have:

2: an e-An(a'+e) = 2: (An - A n - 1) e-An(a'+e)
p~n~q p~n~q
= 2: Aie-An(a'u) - e- An + 1(a'u)
p~n~q-l
From (1.9) it follows that for p > 1 +ne
I 2:
p~n~q
an e -An(a' + e) I ~ 2:
p~n~q-l
eAn(a' u/2)
+ eA.(a' + e/2) e -A.(a' + e)
~ 2:
p~n~q-l
e An (a'+B/2)(a' + e)
X f"n
"n+1
e-u(a'+e) du
~ (a' + e) f"· e- u(e/2) du

JA p
+ e- q(e/2) + e- Ap(ef2).
A
This last expression tends to zero as p tends to infinity. Hence the series
(A, A) converges for s=a' +e, consequently Gc<a'.
We are now going to show that if (A, A) converges for s=Go>O, then
a ~ Go. From this it follows that if a > 0, then also a ~ Gc. This fact, together
with the inequality Gc~a', which we have already proved, yields the
theorem.
We can write
2:
2~m~n-l
Am(Go)(eAm"o - eAm+l"o)
14 DIRICHLET SERIES
(AiO'o) = 2:n .. m an e- An <1o). If 2: an e- An <1o converges, then there exists a

constant M such that IAn(O'o)1 < M (n ~ 1), and
IAnl ~ M [ L
m"n .. l
(eAm +l<1o - eAm <1o) + eAn <1o]
which gives
a = lim sup log IAnl/An ~ 0'0'
The following theorem is an immediate corollary to Theorem 1.2.7.
THEOREM 1.2.8. If 2: an e- An <1o does not converge, then
O'c = 0'0 + lim sup (lOg In~ ame-Am<1°I/An)-

If2:la nl e- An <1 o=oo, then
O'a = 0'0 + lim sup (log L laml e-Am<1o/An).
nE:;m
We ought to remark that in Theorem 1.2.7 the condition a>O (as well
as the condition b > 0) is essential. In fact, suppose that the series (1.2)
represents a function fez) holomorphic in Izl < Rl with Rl> 1 and such
thatf(I):;IoO. The series
cp(z) = 1 ~ Z fez) = Lam)

2: Anzn (An = n"m
has radius of convergence I, for z= 1 is a singular point of cp(z), and
consequently
lim sup (log IAnl/n) = 0,
however, for the series (1.3) we have O'c ~ -log Rl < O.
1.3. The calculus of coefficients and some important combinations

of coefficients
First we are going to extend to Dirichlet series the classical Cauchy
formula for the coefficients of a Taylor series. We assume that O'a < 00,
although this condition is not essential. But we do not have to use the
most general cases when the formulae which we establish here subject to
this restrictive condition are still valid.
THEOREM 1.3.1. If U a < 00, and if f(s) denotes the sum of (A, A) for u> U a,
then
where to is arbitrary and Ul is an arbitrary real number with Ul < U a•

We have for S=Ul +it:
+ -1
T
iT (2:
to m;on+l
am e("n -lIm)S) dt.
Since for a real number k with k =1= 0:
lim -1
T
iT
to
elc«(fl + it) dt = 0,
and since the series under the integral signs converge uniformly with
respect to t, tEO] - 00, 00[, we see that
Consequently,
which corresponds to the statement of the theorem.

Strictly speaking, we see that in the theorem just proved we can re-
place U a by uu, the abscissa of uniform convergence of (A, A), that is,
16 DIRICHLET SERIES
the greatest lower bound of the quantities a' such that (A, A) converges
uniformly in every half-plane a ~ a' +e, where e > 0 is arbitrary. For in
the proof we have only used the uniform convergence of the series. Now
we need the following lemma:
I.3.2. For c>O, k a positive integer, and w real, we have:
1
21T
J(c + it)k dt
eW(c+tt)
{
=
wk-l
(k - I)!
(w > 0)
=0 (w :s; 0)
Suppose first that w > O. For T> 0, al < 0 we denote by 11( ± T, al) the
intervals (al:S; a:S; c, t = ±T) and by 12(T, al) and 13(T) the intervals
(a=al, It I :S;T) and (a=c, It I:S;T). Let R(T, al) be the rectangle formed
by these four intervals. By the residue theorem we have
But since for fixed al the integrals extended over h(T, al) and h( -T, al)
tend to zero, we see that
-
1 f
eW(c +tt)
211" (c + it)k
dt - -
1 f
e W("1 +It)
21T (al + it)k

dt =
wk - 1
(k - I)!
.
And since the second integral in this equality tends to zero as al tends
to - 00, the lemma is proved for w> O. For w < 0 the proof is similar. This
time R(T, al) must be replaced by the contour R'(T, (2) with a2> C, com-
posed of the intervals:
f{ (± T, (2) == (c :s; a :s; a2, t = ± T), f 2(T, (2)

== (a = a2, It I :s; T), f3(T).
The integral written above, but this time extended over the contour
R'(T, (2) (instead of R(T, al» is zero, because the integrand is holomor-
phic in the interior and on R'(T, al), and if we first let T tend to + 00
and then a2 to + 00, we obtain the required result.
The following theorem will be very useful later on.
THEOREM 1.3.3. Suppose that O'a < 00. If v > 0 and if k is a positive integer,
we have
= L (v - An)"-1am
(k - I)! fr(e + it) ev(c+tt) dt { An<V
27T (e + itY for
=0 for
where e> max (O'a, 0) and f(s) is the sum of (A, A).
For V>A1 and s=e+it we have
flC_s)_
eVS dt =
Sk
f" L,lIn<v an e(V-An)S dt
Sk
+ f"
L,lIn~v
an e(V-lIn)S dt.
sIt
From 1.3.2 and the uniform convergence of the series under the last
integral sign we see that this integral is zero. The same Lemma 1.3.2. allows
us to write:
(k - I)! ff(S) vs _ (k - I)! '" fe(V-"n)S

2 k e dt - 2 L. an k dt
7T S 7T lin < y S
= 2
i\n<V
an(v - An)"-l.
If v ~ A1> the integral is zero. In this theorem we can likewise replace

O'a by O'u'
We end this chapter with an elementary inequality, which generalizes
the Cauchy inequality for the Taylor coefficients.
THEOREM 1.3.4. Let 0'1> O'a, to be any real number and M=

SUPt~to If(O'l +it)l, wheref(s) is the sum of (A, A). Then for n~ 1:
(1.10) lanl ~ M eAn (11.
This follows immediately from 1.3.1.
NOTES
1 Unless the contrary is stated, the first index of a sequence ..::I ={.\n} is always 1 and
not 0; when the index or the variable in lim sup, lim inf, or lim tends to + 00, this is
not stated explicitly. Also lim SUPn~ Pn is written lim sup Pn.
<Xl
2 E[a] denotes the integral part of a.

3 We recall that hD';;;; 1.
CHAPTER II
INEQUALITIES CONCERNING THE COEFFICIENTS
II.1. Inequalities corresponding to the

arithmetic character of the exponents
Suppose that aa ~ 0, and let T( a) be an arbitrary function with T(a) > - 00

for a>aa. We write 'S'xfor the set of points s=a+it with a>aa, t>T(a).
If f(s) is the sum of (A, A) for a> aa, we set
M = sup If(s)l;
8e~T
obviously
for (1.10) holds by 1.3.4, for every al>O.

We are now going to show that if a member Ak of the sequence A 'is
distinguished' by an arithmetic property from the other members of the
sequence - this distinguishing feature could express itself, for example,
in the form of the difference Ak - Am for all positive integers m with m:l- k
- then it suffices that the analytic continuation f(s) of (A, A) should be
bounded by M in a domain 'S' whose properties depend on the arithmetical
character of Ak> to be able to state that lakl ~BM, where B is a numerical
constant independent of A and of k. This constant depends only on the
arithmetical character of k. No inequality of this form is to be valid if
we abandon this assumption that Ak is 'distinguished'. (In other words,
the inequality becomes false no matter what value is chosen for B.) If
Ak is an integer, then the value of min (Ak-Ak-h Ak+l-Ak) plays the
same role as the arithmetical character of the difference Ak - Am (m:l- k).
We begin with some definitions.
In the plane of the complex variable z=x+iy=rei8 (r= Izl) we use the
name starred curve, which we denote by a capital letter with an asterisk,
for example r*, for any closed Jordan curve that is symmetric with respect
to the real axis, passes through the point z= 1, and is given by a con-
tinuous function r = cp( 0) (0 E [0, 21/'], cp(O) = cp(21/'» cpC 0) > 0, where cp has a
INEQUALITIES CONCERNING THE COEFFICIENTS 19
continuous derivative in the neighbourhood of () = 0, subject to the con-

dition
lim dr.! = c
0 .... 0 d(} () ,
and where c is a finite quantity.

We denote by D(r*) the domain bounded by r*.
We say that a starred curve G*' shelters' the starred curve r* if for every
Z E r*, z# 1, we have z E D(G*) and if by writing for Z E G*
1. dr 1
i~ d(}'O = Clt
we have Cl> c.
Let cI>(z) be a function holomorphic on the closed disk Izl ::::; 1, real for
real z, with cI>(1)=I, cI>'(l) > -1, cI>(z)#O for Izl<l. Such a function is
said to be attached to a starred curve r* if for Izl < 1 we have zcI>(z) E D(r*).
Let A> 0; an integer m > 0 is called A-suitable for the function cI> if by writ-
ing
(II.1) (cI>(z)Y = dbA) + 2: d~) zm,
we have d~)#O. Two numbers A>O and A'>O are said to be distinct
modulo 1 relative to cI> if for no pair m ~ 0, m' ~ 0 of A- and A' -suitable
numbers for cI>, respectively, we can have the relation
" - '" = m' - m.

If ,,1 is any (open) domain containing the origin, we write 2,,1 for the
domain of the plane of s = a + it that consists of all the points for which
e- S ELl. Thus, 2 D(r*) is the set of all points of the s-plane such that
the points z=e- S form a domain whose boundary is r*.
We denote the half-plane a>a' by Pa ,.
Let ~ be a domain of the s-plane containing a half-plane P(J' whose
boundary is a curve defined by a continuous function a=a(t)
(t E ]-00, 00[, s=a+it).
If ff(s) is a function holomorphic in P(J" we say that it has a direct
analytic continuation into ~ starting from P(J' if CP(J' () ~#0 and if there
exists a holomorphic function in ~ that is equal to ff(s) in P(J" We denote
this analytic continuation into ~ (and only the direct one) by the same
letter ff.
20 DIRICHLET SERIES
We can now state the following general theorem; its character will be
easier to grasp by the special cases which we state below.
THEOREM 11.1.1. Let G* be a starred curve, r* a starred curve sheltered

by G* and q, the function attached to r*.
If (A, A) has an abscissa of absolute convergence aa < 00 and if its sum
f(s) has a direct analytic continuation starting from Pa" where a' > aa,
into .!l' D(G*), then there exists a constant H depending only on G* and on
q" such that for every Ak E A that is distinct modulo 1 relative to q, from
all the other An we have
(11.2) lakl ~ H sup If(s) I inf(ld~k)lvAk + m)-l.
SE.!l'DWO)
Here the d~) are defined in (11.1).

A particular choice of a curve r* on one hand and the function q,
attached to it on the other (a choice that permits us to make the proper-
ties demanded of the' distinguished' exponent Ak more precise) leads from
Theorem 11.1.1 to the following theorems, which can be regarded as par-
ticular cases.
THEOREM II.1.2. To every starred curve G* sheltering Izl = 1 there corres-

ponds a constant B depending only on G* such that if a a < 00 and if the func-
tion f(s) defined by (A, A) has a direct analytic continuation starting from
Pa" where a' ~ aa, into .!l'D(G*) the following inequality holds for every k:
(11.3) lakl ~ B sup If(s)IAk 1/2.
SE.!l'D(GO)
THEOREM 11.1.3. Suppose that a starred curve G* shelters the (starred)

curve defined by r=cos (8j3) (I 81 ~7T). Suppose also that aa < 00 and that
(s) defined by (A, A) has a direct analytic continuation starting from Pu"
where a' ~ aa, into .!l'D(G*).
Then there exists a constant Bl depending only on G* such that the
inequality
(11.4)
is valid for every Ak for which Ak - An is not an integer for any n i' k.
THEOREM 11.1.4. If Ak is an integer and if in the statement of Theorem

11.1.3 the condition concerning Ak is replaced by the following: every integral

An (n#k) satisfies one of the following two conditions An>2Ak' Ak> 2An,
while the other conditions of Theorem 11.1.3 are preserved, then the inequality
(11.4) also holds.
It is clear that if in 11.1.4 all the An are integers, we can write

F(z)= L: anz An , and with the assumption concerning Ak on the one hand
and the direct analytic continuation of F(z) into D(G*) on the other hand,
we can put the conclusion into the form
lakl ~ Bl sup IF(z)l.
2eD(GO)
THEOREM 11.1.5. The conclusions of Theorems 11.1.3 and 11.1.4 remain

valid if we replace G* by a starred curve Gt which shelters the starred curve
r* consisting of the arc r=cos (012), 101 ~7T/2, and of its symmetric image
in the y-axis; here the condition bearing on Ak in Theorem 11.1.3 is to be
replaced by the condition that Ak - An for n # k is not an even integer,' the
condition in Theorem 11.1.4 bearing on Ak has to be replaced by the follow-
ing condition: if Ak is an integer, then every integral An (n # k) for which
An - Ak, is even satisfies one of the two conditions An> 3Ab Ak> 3An.
To establish Theorem 11.1.1 we have, first of all, to prove the follow-
ing lemma:
11.1.6. Let r*, G*, and ~ be defined as in Theorem 11.1.1, let ea (a~O) be
the curve in the plane of z=rete defined by r=eae2 , 101 ~7T. For sufficiently
small a the image of ea under z~(z) is situated on D(G*).
(According to the notation introduced above, D(G*) denotes the com-
pact domain bounded by G*.)
We set ~1(Z)=Z~(z). Since ~1(1)=1, ~W)=l+~'(l»O, there exists
a disk with centre at Z= 1 in which ~1(Z} is a single-valued function; let
U1 be such a disk. There also exists a disk with centre at ,= I in which the
inverse function ~11m of ~1 (~-l(~(Z»=z) is single-valued; let U2 be
such a disk. We set ,=e+i1)=pety and denote by Qa the image of ea
under ~1' With every a> 0 we can associate an e=e(a) > 0 and a y=y(a) > 0,
sufficiently small for the following conditions to hold: on the one hand,
the part of Qa which is the image (under ~1) of the arc of ea corresponding
to 0 ~ e~ e, is a simple Jordan arc" which links the point' = 1 to a point
e
in the half-plane 1) > 0(, = + i1), without passing through the real axis;
22 DIRICHLET SERIES
on the other hand, the ray from the origin (g = 0) making the angle y
with the axis 7]=0 cuts ,I in a single point p. The curve composed of the
part of this ray between the origin and p, the arc ,I and the interval
0:::;; g:::;; 1 on the real axis then determines a simply-connected domain D.
In addition, we can choose e(a) and yea) so that e(a)=e(O) and y(a)=y(O)
(this is possible - it suffices to take yea) sufficiently small) and such that
the domain between the arcs Qo and Qa beginning at g= 1 and ending
on the ray (through the origin) making with 7]=0 the angle y(O)=y(a)
should be contained in U2 • We can also choose e(a)=e(O) small enough
for the corresponding arcs of Co (which is an arc of Izl = 1) and of Ca
to be contained in U1 • For a given a, having made the choice of the corres-
ponding quantities e and y we refer to the resulting domain D as the
'domain Da'. The image of Da under 4>i 1 is denoted by D;;l. The bound-
ary of D;;l is evidently a simple closed Jordan curve.
If a is sufficiently small, we can choose the domain Da in such a way
that when we write Ra for the part of Qa that is part of the boundary of
Da we have RacD(G*).
passing through ,=
This comes from the following arguments. If L is a curve in the '-plane
1 that can be represented in a neighbourhood of this
point by p = p(y) (polar coordinates: a single p for sufficiently small Iy\)
and such that the (finite) limit
lim dP.!
y-+O dy y
exists, we denote this limit by c(L). The fact that Ra c D(G*) for suffi-
ciently small a is due to the following inequality, which is valid for
sufficiently small a:
c(Ra) < c(G*).
For if we set
then we have:
lim A(eao2 +10) = 1 + 4>'(1) = A > 0
8-+0
lim dB(ea82+18) = OB] = OA] = A'(I),

0-+0 dO oy 2=1 OX 2=1
where A(x) is the restriction of A(z) to the real axis. Writing A'(I)=A l
we easily see that
20: Al 1
c(Ra) = A - A2 - Ji.'
and by definition: c(G*) > c(r*). Thus, it is sufficient that 0: > 0 should be
small enough for us to have c(Ra) < c(G*). And so with a suitable choice
of 0: we have ~l(Da)c D(G*). If 0: is sufficiently small, the image under
~l of that part of D(Ca) which is in the exterior of Da and its image in
the real axis is certainly contained in D(G*). But the same holds for the
image under ~l of the part of D(Ca) that is contained in Da (and its
image in the real axis). Therefore, there exists an 0:>0 such that the image
of D( Ca) under ~1 is in D( G*). And this demonstrates the lemma.
Now we proceed to the proof of Theorem 11.1.1. According to the
conditions of the theorem, the function ~(z) does not vanish in the disk
Izl < 1 (see the definition of a function attached to a curve r*); it is holo-
morphic on Izl ~ 1. It may have a finite number of zeros on Izi = 1, say,
zl> ... , Z/c'
We choose 0: > 0 sufficiently small so that:
(1) ~l[D(Ca)]cD(G*) (~l(Z)=Z~(z», which is possible by virtue of
the lemma we have just proved;
(2) ~(z) has no zeros on the disk Izi ~ ean2 other than the points Zj
(j=1,2, ... ,k).
We write L j (j= 1,2, ... , k) for the ray through the origin passing
through the point Zj' and S} for the segment of L j between Zj and Ca.
We denote by eta the curve composed of Ca and of the k segments Sf.
Finally, we denote by Lla the compact set bounded by eta.
We set
where log ~l(e-C1) is real for (7)0. The function pes) is holomorphic on
~Lla, and we have
sup IP(s) I ~ sup I/(s) I = M.

se..l!'LI" se..l!'D(G·)
In fact, we have Llac D(G*), and ~l(Z) has no zeros in Lla other than a
simple zero at the origin.
24 DIRICHLET SERIES
We set
<P(z) = do + L dmzm
(that is, dm=d~) according to (11.1».
We see that for sufficiently small r> 0
L lanl(r L Idmlrmt n < 00,

which enables us to write for sufficiently large a:
lJ'(s) = Lan e-"nS(L dme-ms)"n
= Lan L d:;n) e-(mHn)s = LIp e-/1. p S,
where the [.Lp are of the form m+An, Ip being the sum extended over all
the terms and~n) for which [.Lp=m+A n•
For every p~ 1 and sufficiently large al we can write:
Ip = lim ~ J: lJ'(al + it) eUp(fJ 1 +tt> dt.
Now let p be such that a7T 2[.Lp ~ 1, and let Ca,p denote the curve composed
of the arc a=[.Lpl-at 2 , O:::;t:::;1T, and the segments SJ,p, which are the
parts of the segments -log S1 (s E -log S1=e- s E S1) contained between
this arc and the line a=O (these segments are evidently parallel to the
real axis t=O), where the ordinates of these segments belong to ]0,1T].
In other words, the segments S"p which we consider are all of the form
[.Lp 1 - all :::; a :::; 0, o< t, :::; 1T
with <P (e- ltl ) = O.
If we denote by Ca,p,m the union of m curves of which one is the union
C~,p of Ca,p with its symmetric image in the line 1=1T, while the m-I
others are those that we obtain from C~,p by the successive translations
s+2mri (n= 1,2, ... , m-l), we see by an evident application of Cauchy's
theorem that
Ip = lim ~ f lJ'(s) ettpS ds,

m1Tl JCa,p,m
where the integral is taken (for every m) along the segments of Ca,Jl,m
that are parallel to the real axis in both senses. (Incidentally, lJ'(s) does
not have the same value when we approach such a segment with in-
creasing or with decreasing I, because <P(e- ItI)=O.)
Since IP(s) I ~ M on Ca,p,m, we can write:
I/pl ~ ~ [f e Up(U p -l-at 2) dt +2 f fq- up e- Up!1 da].
It follows that
I/pl ~ ~ [f e 1 - aU pt2 dt + k f,,2_ U

;1 e-UpfJ da].
But we can write for 0 < IX < 1 (and IX1T2p.p ~ 1):
where A is a finite constant. Starting from the obvious inequality I/pl ~M

for IX1T2p.p ~ 1 we also have:
All in all, no matter what p is, we have

I/pl ~ KaM p.;1/2.
But, if Ak is distinct modulo 1 relative to flJ from all the other An
(n#k), and if p.p=Ak+m with d~~k)#O, there exists no other couple
An, m' (n#k) such that p.p=An+m', dg:n)#O. Consequently, according to
the definition of p.p and Ip we have:
and the inequality we have demonstrated for Ip gives
laklldr\tk)1 ~ KaM(A" + m)-1!2.

The conclusion of the theorem is now an immediate consequence.
26 DIRICHLET SERIES
Next we show that Theorem 11.1.2 is a particular case of Theorem

11.1.1. The starred curve r* of Theorem 11.1.2 is the circle Izl = 1, the
function <1>(z) attached to this curve r* is the constant 1; and no matter
what A> 0 is, only the integer m = 0 is A-suitable for <1>. Hence An and Am
are distinct modulo 1 for all m and n relative to <1> (that is, they are simply
distinct, A being a strictly increasing sequence). So the inequality (11.2),
with the conditions of Theorem 11.1.2, becomes (11.3).
Now we make the passage from Theorem 11.1.1 to Theorems 11.1.3
and II.1.4. The (starred) curve in these statements which is sheltered by
G* is the curve r* corresponding to the statement of Theorem 11.1.1;
it is the exterior contour of the image of Izl = 1 under z<1>(z)=z(1 +z)/2.
As long as A> 0 is not an integer, every integer m;:;:: 0 is A-suitable for <1>.
And if A is an integer, then m is A-suitable for <1> whenever 0 ~ m ~ A. The
inequality (11.1) can also be written:
la,,1 ~ A sup
sE.!l'D(G*)
If(s) I inf(ld~Ak)1 VA"
m
+ m)-1
~ A sup
sE.!l'DW*)
If(s)I(ldE (AkI 2 )lvA" + E(A,,/2»-1
where, as before, E(a) denotes the integral part of a. A simple estimate of
dk~~/2) yields the required result.
In Theorem ILLS the starred curve r* is the exterior contour of the
image of the circle Izl = 1 under z<1>(z)=z(l +z2)/2. The passage from
Theorem 11.1.1 to Theorem 11.1.5 starts from this point and follows the
arguments used in the passage from Theorem 11.1.1 to Theorems 11.1.3
and II.1.4.
We remark that the conclusion of Theorem 11.1.1 is not valid if no
condition is imposed on the nature of A" (for a given k). Thus, the func-
tion f(s)=L:(-l)nne- ns is bounded in a domain .PD(G*), where G*
corresponds to the statement 11.1.3, but la,,1 =k is not bounded indepen-
dently of k. In fact, one can even take for a" the quantity (-l)"g(k),
where g(z) is an entire function of minimum type (of exponential type
zero), g(z) having positive Taylor coefficients; for g(n) = lanl tends to
infinity more rapidly than any power of n, and yet one knows that the
function F(z) = L: g(n)zn only has Z= I as singular point, that is, the func-
tion
f(s) = L: ( _l)ng(n) e- ns
is bounded in an .IE D(G*), where G* corresponds to the statement 11.1.3.
11.2. A general inequality for the coefficients

corresponding to finite upper density
of the exponents
In what follows we assume that the upper density D· of the sequence A
is finite. From 1.2.5 we know that aa=ac=lim sup (log lanl/An). We also
assume that the last quantity is not + 00, so that aa = ac< 00. As before,
we denote by f(s) the sum of (A, A) for a> ac.
We denote by C(so, R) the disk with centre at So and radius R. Let L
be a Jordan arc given by s=s(u), u E [a, ,8]. The union
<£(L; R) = U C(s, R)
seL
is called a channel of width 2R and with L as central line. The extreme

disks of the channel are C(a, R) and C(,8, R). Throughout we assume that
the extreme disk C(,8, R) lies in Pc. We say that (A, A) has a direct analy-
tical continuation into er(L; R) if there exists a holomorphic function in
<£(L; R) that is equal to (A, A) in C(,8, R). We frequently omit the adjec-
tive 'direct' when indicating this continuation. We also use the letter J,
or frequently (A, A), to denote the continuation.
We now establish the following general theorem, which plays a very
important role in what follows:
THEOREM II.2.l. Let D· < 00 and suppose that ac< 00. If f(s), the sum of
(A, A), has an analytic continuation in a channel cr.(L; 7TR) with R> 15",
such that C(so, 7TR) (so=ao+it o) is one of the disks:
C(so, 7T R) c er(L; 7T R),
then we have
lanl :;;; C(R)AnM(so) eAnrJO ,
where C(R) < 00 only depends 1 on R, where {An} is the sequence associated
with A, and where
M(so) = sup If(s)l.
ssC(so.nR)
We recall that iY is the mean upper density of A. We set
An(z) = n (1 -
m,on
~:)
m
(n ~ 1).
28 DIRICHLET SERIES
Since D' < 00, we see immediately that this product converges uniformly
on every compact set, so that An(z) is an entire function whose zeros are
only the points ± Am (mIn), and by 1.1.2 we have
(II,S) '
I1m log An{ir)./ D-'
sup "" 7T •
r
because it is evident that the upper density and the mean upper density
of the sequence obtained by removing from A a single term (or a finite
number of terms) are the same as for A itself,
It is clear that
max IAn(z) I = An{ir) ,
121=r
Hence the An(z) is an entire function of exponential type not exceed-

ing 7TlY,
We now set
An(z) = 2: a~n)z2fJ;
we see that (-I)fJa~n» 0 with a~n}= 1.
We introduce the operator fJn(f) defined by
(II,6) fJn(f) = 2: a~n>J(2fJ},
This operator represents a holomorphic function in the channel
<teL, 7T(R - j)'», because for every e > 0 with e < R - j)' the series (II,6)
converges uniformly (with respect to s) on every closed disk C(s,7Tr) with
s E L, 0 < r ~ R - j)' - e, It is easy to see first of all that we may argue under
the assumption that I is holomorphic on <t(L,7TR) (the closed channel),
For every point s' for which Is'-sl=7T(R-j)'-e), where sEL, we can
write (for every positive integer q):
J<q}(s') = il! 1m
27Ti j g-
d,
S')q+l
the integral being taken over the circle with centre at s' and of radius
7T(j)' +e), which gives
(II,7) If(q}(s') I ~ q~M(s)

"'" (IT(D' + eW
where M(s) is the maximum of Iii on C(s,7TR),
Since the type of An(z) does not exceed 7TIY, the series L: a~n)(2p)! Z-2P
has radius of convergence R:::;; 7T 15 (it converges at the exterior of the
circle). Hence the series (11.6) converges by (II.7) uniformly on the disk
C(s,7T(R-D·-e». The function Fn(s)=lYn[f(s)] is holomorphic in the
channel <f;(L, 7T(R-D' -e», and on L we have:
(II.8) IFn(s) I : :; L: (-1)Pa~n)(2p)! (7TR)-2 P. M(s) = An(R)M(s).
(It is enough to set in (11.7): D' +e=R.) But when we set (see 1.1.2)
Ao(r) = L: (-1)Papr2P ,
we have (-l)Pa~n):::;;(-l)Pap, and when we write
C(R) = L: (-1)Pa p(2p)! (7TR)-2P,
we have An(R):::;; C(R) < CI) for R> D'. Hence
(II.9) \Pn(s) I : :; C(R)M(s).
Note that by setting 11(S) = L: lanl e- ilnS we have another inequality
similar to (I1.7), where I is replaced by 11 and M(s) by M 1(s), Ml being
defined like M but with I replaced by 11> where s is taken on that part of
L for which C(s, 7T R) C P([c' Without loss of generality we may assume that
L contains part of the real axis. When we write F'n(s) = L: (-1)Pa~n~fi2p)(s),
we see that for sufficiently large a E L we have:
F~(a) = L: (-l)Pa~ny?P)(a)
= L(-l)Pa~n) LA~laml e-
P m
ilm ([;
and since all the terms of the double sum are positive, we can change the
order of summation also in the series
Fn(s) = L: a~ny(2p)(s) = L: a~n) L: A~ am e- ilms
= L: am e- ilms L: a~n)A~ = L: amAn(Am) e- ilmS,
as long as s E L, with a sufficiently large. But since An(Am)=O for m=j:n,
we finally obtain
Fn(s) = anAn{An) e-iln8 = A; Ian e- iln8,
where {An} is the sequence associated with A. The inequality (II.9) with
S=So gives us the conclusion of the theorem.
Here is an important immediate corollary to Theorem II.2.1.
30 DIRICHLET SERIES
11.2.2. Let D' < 00, Gc < 00. If (A, A) has an analytic continuation in a
channel [(L;rrR) with R>IY of which C(so,rrR) (so=Go+ito) is a disk,
then
·
11m log An
Go ;:: Gc - sup - - .
An
It is sufficient to take account of 1.2.5.
Taking account of 1.1.1 we obtain the following statement:
11.2.3. If the step h of A is positive and if the other conditions of 11.2.2

are satisfied (D' <00 holds automatically, because h>O), then we have
for every a> 1:
Go ;:: Gc + [2a log (hD') - A(a)]D' = Gc - L(a, h, D'),
where A(a) is defined by 1.1.1.
In particular, such a series CA, A) cannot be continued analytically in a
channel of width 2rrR, R> D', for which the abscissa of the centre of a disk
belonging to it is smaller than [3 log (hD') - 8.5]D· +Gc•
NOTE
1 C(1TR)=1TRL(1TR), where L is the Laplace transform of Ao (see 1.1.2).

CHAPTER III
THEOREMS OF LIOUVILLE-WEIERSTRASS-PICARD
TYPE OF ARITHMETICAL CHARACTER AND OF
GENERAL CHARACTER
111.1. Theorems of arithmetical type

We are going to prove the following theorem.
THEOREM IlL 1. 1. Suppose that Ak for a given k has the following property,'
the only rational numbers al> a2, ... , am (m> k), a, with aj ~ 0 (j = 1, 2, ... ,
m) satisfying the relation
are a1 = .. 'a"-l =ak+1'" =am=O, ak= 1, a=O.

Suppose that f(s) defined by (A, 11) with aa < 00 has a direct analytic
continuation starting from Pu' (a' > aa) into !l'D(G*), where the starred
curve G* shelters the curve defined by
r = cos «()J3) (I ()I ~ 77).
Then the set of values taken by f(s) in 2 D( G*) is everywhere dense in the
disk of radius lakl around the origin.
Let a be complex and e>O such that If(s)-al >e in 2D(G*). Now
limf(s) = 0 as a ~ 00, uniformly with respect to t (this is so because
AI> 0). It follows that lal ~ e. Next we set
1 1
F(s) = a - f(s) a
Then F(s) is a holomorphic function in 2 D(G*), and in this domain
IF(s)1 < e- 1 + lal- 1 ~ 2e- 1 •
If a is such that
2: lanl e-"nu < lal,

we can write
32 DIRICHLET SERIES
where {Ln} is a sequence of positive numbers increasing (strictly) to in-

finity.
Every Ln is of the form
where the Aj are positive integers.

If p is any positive integer, there exists a term Lq of {Ln} equal to PAk;
but according to the assumption made on Ak, no combination
AlAI + A2A2 + ... + AmAm + A,

where the Ai (j= 1, ... , m) are non-negative integers and where A is an
integer, can be equal to pAk' provided that the integers A and Aj (ji=k)
are not all zero and Ak=p.
In other words, no difference
Lq - Ln (n i= q)
is an integer.
Since for Lq=PAk we have Cq=a~/aV+I, Theorem II.I.3 shows that
la~1 ~ B1 1alV+ 1 sup
SE,:l'D(G')
1F(s) 1= 2B1 IalV+ 1e-l.
By extracting the p-th root and letting p tend to infinity we obtain

Let M={ftp} be an increasing partial sequence of A: MeA. We say
that M is a linearly isolated partial sequence of A if the difference between
any linear combination with positive integral coefficients
and a linear combination with positive integral coefficients
is not an integer unless the combinations are identical (that is, q=p,
Ani=fti> Aj=mj (j= 1,2, ... , p».
The following theorem is of 'Picard' type, with the conditions bearing
on the 'arithmetical' exponents.
THEOREMS OF LIOUVILLE-WEIERSTRASS-PICARD TYPE 33
THEOREM 111.1.2. Let {AV}} be a linearly isolated partial sequence of A.

We assume that f(s) defined by the series (A, A) with aa < 00 has an analytic
continuation starting from P (a' > aa) into If D(G*), where G* is defined
(I'
as in 111.1.1.
Then f(s) assumes in If D(G*) all the values of the disk qO,2: lav}D
(lsi < 2: lav}I), except possibly the value zero and one other value.
It is known that if a=l:-b, there exists in the complex plane a function
X(w) with the following properties:
(a) x(w) is regular at the origin and consequently is given in the neigh-
bourhood of the origin by a series 2: Anwn;
(f3) x(w) has an analytic continuation on every curve that does not pass
through a or b, with w=a and w=b as ramification points for x(w);
(y) Ix(w)1 < 1. Evidently
lim sup IAnl11n = [min (Ial, Ibl)]-l = c- 1.
Suppose then that, contrary to our assertion,J(s) does not take the two
values a and b such that a=l:-b, a=l:-O, b=l:-O, lal <A= 2: lav,l, Ibl <A, and
consider the function F(s)=xof(=x(f(s)), where X is the function we
have just defined.
If a' is such that 2: lanl e-An'" < c, we see that
converges and consequently that F(s) for sufficiently large a is given by a

Dirichlet series
a series having an abscissa of absolute convergence. In If D(G*) we have

IF(s)1 < 1. By the condition on {-'V}}, every Lk which is a linear combination
with positive integral coefficients of the Av, differs from any other L y
by a number that is not an integer. Let
34 DIRICHLET SERIES
The corresponding coefficient c" of F(s) is given by:
- ,n !An 'Vp
C" - amI • ••
amp
Vp ,
(m 1 + m2 + ... + mp -- 11) •
ml····mp.
Theorem II.l.3 allows us to write
(III.1)
where p, Vb ... , Vp and n are given and the inequality is valid for any
positive integers mb m2, ... , mp for which
ml + m2 + ... + mp = n.
We fix p and the integers Vll V2, ... , Vp. We take p positive constants
all ••• , a p with al + ... + a p = 1, and for every positive n we set
mj = ECna,) (1 ~ j ~ p - 1),
mp = n - (ml + m2 + ... + mp-l)'
By Stirling's formula we have
. ( n! ) lin _a _a
hm , , = <XlI ••• <Xp p.
n ml ..•. mp.
From (III.1) it follows that

(III •2) -a ••• <Xp-a p
<XlI IaV1 la 1... Iavp lap '"~ (1'1m sup IA n11In)-l
= C = min (lal, Ibl).
We see this by setting in (III.1) n=nj, where {nj} is a sequence for which
lim IAnjll/ni=lim sup IAnll/n, and by lettingj tend to infinity.
Since we may assume that all the an are non-zero, we may set
<Xj = lavj l/(lav1 1 + ... + lavpl) (1 ~ j ~ p).
We then have aJf= [(la vj l/Cla V1 1 + ... + lavpl)]af, and (III.2) gives us im-
mediately:
laV1 1 + ... + lavpl ~ min (lal, Ibl).
Since p is arbitrary, we arrive at the contradiction
Theorem III.l.2 is evidently more precise than Theorem 111.1.1 (not only
is Theorem III.1.2 of 'Picard' type, while Theorem 111.1.1 is of 'Weier-
strass' type, but also the radius of the 'disk of the distribution of values'
is smaller in the latter proposition). But we have thought it useful to
demonstrate both theorems, because although the two theorems are based
essentially on Theorem 11.1.3, the passage from this to Theorem 111.1.1
is elementary, whereas we had to use the modular function in order to
pass from Theorem 11.1.3 to 111.1.2.
Here is an immediate corollary to 111.1.2.
111.1.3. If the sequence {t\·n} is such that no combination

AlAI + A2A2 + ... + AnAm
where the Aj (j= 1, ... , n) are integers, is itself an integer, and if the condi-
tions oflll.l.2 on f(s) are satisfied, thenf(s) takes in 2 DCG*) all the values
of the disk qO,2: lanD except possibly the value zero and a single other
value.
We ought to emphasize the fact that in Theorems 111.1.1 and 111.1.2
a single coefficient, or a subset of coefficients whose exponents are dis-
tinguished by their arithmetical character from the other exponents, can
already furnish important information on the function. They give proper-
ties that cannot be annihilated by the coefficients corresponding to other
exponents. In Theorem III. 1.1 the object to be evaluated is the least upper
bound of I/(s) 1 in a domain whose form depends on the character of the
exponent, and in Theorem III. 1.2 it is the image of this domain under f
We shall see later that the distribution of singularities of 1 cim also be
determined to a large extent by the coefficients corresponding to exponents
that are' isolated' in a certain fashion.
llI.2. Liouville-Picard theorems of general type

We assume that the upper density D' on A is finite; we also assume now
that the series CA, A) converges in the entire plane: U c = -00. From the
fact that D' < 00 it follows that Uc=Ua = lim log lanl/An = - 00.
The function I(s) given by (A, A) is therefore entire. For every real U
we set:
M(u) = lub If(u + it)l.
t
36 DIRICHLET SERIES
It follows from 1.3.4 that IimM(O')=oo as 0' tends to -00, providedfis

not identically zero. In that case the quantity
. log log M(O')

(111.3) p = 11m sup ---"'---'''--~
a-+-oo -a
is called the (R)-order off(the Rilt order off). One must not confuse the
(R)-order off with the order of this entire function in the classical sense,
that is, the order defined by
~ l' log log ml(r)

0= In sup 1ogr '
where ml(r)=max If(s)I(lsl =r). Thus, the (R)-order of f(s)=e- s is 0, but

its order in the classical sense is 1.
The following proposition is useful:
THEOREM 111.2.1. If
(111.4) 'fAn
IimIn 1--> 0,
ogn
then a necessary and sufficient condition for the (R)-order off to be equal
to p is
. sup log lanl = -_.
11m 1
(Ill. 5)
An log An p
(The second term of this equality is - 00 if the (R)-order of f is 0, and

vice versa.)
Note that in the statement we do not demand that D' < 00; but the
condition (111.4) is satisfied when D' < 00. Since in this section we assume
the latter condition to be satisfied, the (R)-order off is here always given
by (111.5).
Suppose that p, the (R)-order of f(s) , is finite. For every 0' we have (see
(I.10»:
lanl ~ M(O') ell,,,.
It then follows from (1l1.3) that for every e > 0 and sufficiently large
-0' we have
log lanl ~ log M(O') + Ana < e-(p+e)" + Ana.
Now the minimum of the third term of this inequality is assumed for
a= -log [An/{P+ e)]/{p+ e), a quantity that tends to -00 as n tends to
+00. Hence for sufficiently large n we have:
log lanl ~ min (e-(P+e)<1+ Ana)
= An{l - log (An/(P + e»)/(p + e),
which gives
lim sup log lanl ~ ---,

An log An p +e
that is, since e> 0 is arbitrary,
lim sup log lanl ~ 1

An log An p
Evidently this inequality is also valid for p = 00. On the other hand, if this
inequality is also valid for p < 00, then to every e > 0 we can assign a quan-
tity B{e) such that
which, in turn, gives

(III. 6) M(a) ~ 2: lanl e-71n <1
~ B{e) 2: A;;71 n !(P+S) e-71 n t1
~ B{e) max [e-71 n log 71 n !(P+2S)-71 n<1] 2: e-a7lnIOg'\n,
with a=e/(p+e) (p+2e). But it follows from (I1I.4) that for a certain
positive constant b:
Thus, (I1I.6) allows us to write, by setting k=(p+2e)-1

M(a) ~ C(e) max e-k7lnlOgiln -71n <1
~ C{e) max e-kxIOgx-x<1 = C(e) e-(k+2<1) e- O +<1!k).
So we have for sufficiently large - a

log M{a) ~ e-(l!k+e)<1,
which proves that

.
11m log log M(a) ___
sup : : : p.
a-+-co -a
38 DIRICHLET SERIES
This completes the proof of 111.2.1.

We are now going to define the order of a function f in a horizontal
strip. Let S=S(to,a) denote the strip It-tol<a. Letf(s) be a function
holomorphic in the closed strip It-tol::;;.a and let us set for every a
Ms = max If(a
It-tol .. a
+ it)l.
The quantity
.
I1m log log M,la)
sup
(1-+-00
= Ps
-a
is called the order off in the strip S.

We can now prove the following theorem, which plays an important
role.
THEOREM III.2.2. Suppose that the step of A is positive and that the func-
tion f represented by (A, A) is entire. Then the order off in every strip
S=S(to,1Ta) with a> IY is equal to the (R)-order off(in the plane).
Let P be the (R)-order of f and let Ps be its order in a strip S(to,1Ta)
with IY < a. Clearly Ps::;;' p. So we have to show that Ps ~ p. Also, it is
sufficient to prove the theorem for p > 0, because as we shall see later,
for every f not identically 0 we have Ps ~ O.
We set
MI(a') = max If(s)l,
seC
where C=C(a'+ito,1Ta). It follows from Theorem 11.2.1 that for every

a and every n~ 1:
By (111.5), to every e> 0 there corresponds a sequence of positive inte-

gers {nj} such that
(III. 8) log lanil ~ - G+ e) Ani log An} (j ~ 1),
and according to Proposition 1.1.1 there exists a constant P such that
(III.9) log An ::;;. PAn (n ~ 1).

The inequalities (III.?), (III.8), and (111.9) give for j> jg and for every
a (wherejg is independent of a) and a constant C independent of a andj:
(111.10) log MI(a) ~ G+ e) An} log Anj - (P + a)An) -

- C
~ -G+ 2e) Ani log An} - aAn}.

If we set in (111.10), in particular, a= -(I/p+3e) log An}, we obtain for
j';~ je:
Now to every a there corresponds an /X with I/XI ~7Ta for which
If we denote by /Xj the value of a corresponding to
a = - [~ + 3e] log An},

then we have by virtue of (111.11) and (III.12):
log Ms [ - (~ + 3e) log An} + /X j ] ~ eAn} log An},
which allows us to write
.
I1m log log Ms(a)
sup --"''---''''------''..0..-"-
a-"-oo -a
and since e > 0 is arbitrary, the theorem is proved for p > O.

By (III.?) we see that if/is not identically 0, then we have lim MsCa)=oo
(a -+ - (0) and necessarily Ps ~ O.
The theorem is now proved in all cases.
40 DIRICHLET SERIES
Next we pass on to theorems generalizing the theorems of Liouville

and of Picard for general Dirichlet series (D' <(0). We begin with the
following definition. Let s(u) = u(u) +it(u) be a continuous complex func-
tion of the real variable u E ] - 00, +00 [ such that
lim u(u) = 00, lim u(u) = -00,
u=co 1.£=:::-00
with t(u) = to for u ~ uo, where to and Uo are constant. If R is a positive

quantity, the union of the disks C(s(u), R)(u E ]-00, +ooD is called a
curvilinear strip of width 2R, extending to - 00 (horizontal, to the right).
We denote such a strip by B(s(u), R).
The horizontal strips S which we have discussed above correspond
to the case when t(u)=to for every u.
The set of points s(u) is the central line of the strip B. Here is a general-
ization of Liouville's theorem.
THEOREM 111.2.3. Suppose that D' < 00 and that Uc < 00 for the series
(A, A). If f(s), represented by (A, A), has an analytic continuation across
a curvilinear strip B of width 27TR extending to - 00, with R> 15', and if
this continuation is bounded in B, then f is identically equal to zero (that is:
an=Ofor n~ I).
This theorem is an immediate corollary to 11.2.1, where we put so=s(u)
and then let u tend to - 00. The conclusion of 11.2.1 shows immediately
that an=O for n~ 1.
Next, here is a general theorem of 'Picard' type.
THEOREM 111.2.4. Suppose that D' < 00 and that f(s), represented by (A, A)
with Uc < 00, has an analytic continuation across a strip B(S(U),7TR) extend-
ing to - 00, with R> D' " suppose also that f is not identically zero.
If {un} is an arbitrary sequence with lim Un = - 00, then one of the state-
ments (1), (2) holds:
(1) f(s) assumes in the set of disks C(s(un), 7TR) all values except possibly
one.
(2) For every e with 0 < e < R we have
limf(s) = 00
uniformly (with respect to s) as long as s remains in the disks

C(s(un),7T(R - e»,
and n tends to infinity,
If s(u) is real for u<u' (s(u)=a(u», we can replace in this statement the
disks C(s(un), 11'R), C(s(un), 11'(R)-e» by the squares
I
10' - O'(un) < 11'R, I
It < 11'R;
10' - O'(un) I < 11'(R - e), It I < 7r{R - e).
We consider the family ff of holomorphic functions
fn(s) = f(s + s(un»

defined in the disk C=. C(O, 11'R). If (1) does not hold, the family is normal
in C. But no sequence extracted from ff can be bounded in C, because
then, by Theorem 11.2.1, we would havef(s)=.O; hence, if (1) does not
hold, then we can extract from every sequence fn(s) a sequence tending
to infinity uniformly for every closed disk in the interior of C. It then
follows that (2) holds. The last part of the theorem is proved similarly.
The following lemmas will be used later.
111.2.5. Let ff be a family of functions rp(s) holomorphic in a domain ~

and such that: 1<p(s)1 > 1 in ~. To every bounded domain ~1 for which
~1 c ~ there corresponds a constant a such that
where <P E 5'; S10 S2 E ~1.

For, the family of harmonic functions
'PS2(s) = l~: II:g~jl'

where <P is an arbitrary function of ff and where S2 is an arbitrary point
of ~h is normal in ~, because all the functions 'Psis) are positive in ~.
This proves that this family is bounded on ~h because 'P8,.cS~ = 1 for
every S2 E ~\. To obtain the lemma it suffices to make the same remark
with S2 replaced by S1.
111.2.6. Let a and b be two constants with 0 < a < b, and let F(s) be a func-
tion holomorphic in the horizontal strip S(to, b). If the order of the strip
S(to, a) is infinite, the function cannot tend to infinity uniformly in S(to, b)
as 0' tends to -00.
42 DIRICHLET SERIES
Suppose, contrary to our statement, that F(s) tends to infinity uniformly

in S(to, b) as a -+ -00. Consider the family tr of functions FaCs)=F(s+d),
where d is an arbitrary negative quantity and s varies in the square
AI: la-aol <b, It-tol <b. We have
lim FeJ(s) lim F(s + d) = 00
= a=-oo
a=-oo
uniformly in s for sEAl. And if ao is chosen negative and of sufficiently
large absolute value, then for every d~O, SEAl> we have: lEaCs)1 > 1.
Let n(l) be the integral part of -112a. Applying Lemma III.2.S to the
family FeJ(s) (d < 0), we see that there exists a constant a> 1 such that for
any sand s' with s, S' E.32 , where .32 is the closed square la-aol ~a,
It-tol ~a, we have
log lEaCs) I
(III.13) log lEaCs') I < a.
Now let al be a negative quantity of sufficiently large absolute value,

and let us write down the inequality (III.13) for d, s, and s' chosen in the
following manner:
(1) d=al, s is arbitrary in .32, s' =ao+a;
(2) d = al + 2ka (1 ~ k ~ neal»~, s = ao - a, s' = ao + a;

(3) d = 0, s = ao + a + 2an(al) + al> S' = ao - a.
Since
F<1l +2ka(aO + a) = F<1l +2(k+lla(aO - a),
we obtain for s E.32 by multiplying these inequalities:
log IF(al + s)1 = log IF<1l(s) I

~ an(<1 l l+llog IF(al + ao + a + 2n(al)a) I
~ a n(<1 l l+210g IF(ao - a)1
~ a2-<1l/2a log lE(ao - a)l·
It now suffices to recall the definition of the order in a horizontal strip
to see that for S=S(to, a)
Ps ~ log al2a < 00,
which contradicts our assumption. Thus, our lemma is proved.

By combining 111.2.2, 111.2.4, and 111.2.6 we obtain the following

theorem:
THEOREM 11I.2.7. Suppose that the step of A is positive and that f(s),
represented by (A, A), is an entire function of infinite (R)-order. Then f
assumes all values, except possibly one, infinitely often in every horizontal
strip S(to, 7Ta) with a> D'.
Note that Lemma 111.2.6 is a particular case of the theorem of Bieber-
bach-Valiron, which can be stated in the following form:
111.2.8. Let a and b be such that 0 < a < b. If the function F(s) is holomorphic
in the strip S(to, 7Ta) and if the order of F in S(to, 7Ta) exceeds 1/2a, then
the function assumes in every half-strip
It - tol < 7Tb,
all values, except possibly one.
For a proof of this theorem we refer the reader to Valiron [59].
By combining Theorem III.2.2 with 111.2.8 we obtain the following
theorem:
THEOREM 11I.2.9. If A has positive step, and if p is the order of the entire
function f(s) given by (A, A) (a c = -00), then f assumes in every horizontal
strip S(to,7Ta) with a>max (D', 1/2 p) all the values, except possibly one,
infinitely often.
CHAPTER IV
SINGULARITIES OF FUNCTIONS REPRESENTED

BY A TAYLOR SERIES
IV.t. Singularities of Taylor Series

This section is concerned with the case when An = n, and instead of writ-
ing the series in the form of a Taylor-D series: L: anr ns , we write it in
the usual form: L: anzn (z=x+ iy). There exists an enormous literature on
the analytic continuation of a Taylor series, the distribution of its singu-
larities, their nature, etc. Nevertheless we come back to this subject,
because we wish, first of all, to give an account of a general theorem which
plays for the argument of the singularities (on the circle of convergence)
a similar role to that of Hadamard's theorem for their modulus (the
formula of Cauchy-Hadamard providing the radius of convergence). Our
theorem is very general and has a considerable number of applications.
Since the Taylor-D series is a very important case of the general Dirich-
let series, we have thought it useful to devote the present chapter to this
subject.
Here is the theorem on the arguments of the singularities.
THEOREM IV. I. I. Let R (O<R<oo) be the radius of convergence of the

series
and let us set for h~O:
Then D(h) is a continuous function at h=O and has at h=O a right-hand

derivative
- l' D(h) - D(O)

D'+ (0) -1m
Iqo
h '
SINGULARITIES OF FUNCTIONS 45
for which
and if we write
D'+ (0) = cos rp,
then one of the points Rel'P or Re-I'P is a singularity of (IV.l), namely the
one that is nearest to z = R on the circle of convergence. 1
Consider the series
(IV. 2) go(z) = L: ~:.

If Zo is a singularity of F(z), then Zo 1 is a singularity of go(z), and vice
versa.
The radius of convergence of (IV.2) is R -1 = D(O).
Since the function go(z+h)/z is regular at infinity, we can write for
sufficiently large Izl
with
dn(h) = ~7T~f gO;z) (z + h)n dz
= ~}i f gO;z) (zn + (~) zn-1h + ... + hn) dz
= an + (~) an-1h + ... + aohn.

The radius of convergence of the series (IV.3) is equal to D(h). Now go(z)
and gh(Z) have the same singularities, except possibly the point Zo= -h,
which may be a singularity for go(z) without being one for gh(Z).
We denote by S+(h) and S-(h) the domains Iz+hl > D(h) and
Iz+hl < D(h), respectively, and by C(h) their common boundary. Since
gh(Z) has all its singularities in S-(h)=S-(h) u C(h), and since there is
at least one singularity of gh(Z) on C(h), we see that D(h) is positive
for sufficiently small h, that the intersection J(h) = S-(h) (\ S-(O) is
not empty, and that all the singularities of go(z) are on l(h). If Zo = R is
46 DIRICHLET SERIES
a singular point of F(z), then Z= D(O) is a singularity of go(z), and since

h ~O, we have:
D(h) = D(O) + h.
In this case the theorem is therefore trivial. Suppose then that z=R
is a regular point of F(z), so that Z= D(O) is a regular point for go(z).
Since I(z) is not empty, we have
D(O) - h ~ D(h) ~ D(O) + h.
In other words, D(h) is a continuous function at h=O. We denote by
Re1fP the singular point of F(z) on Izl =R whose argument has the smallest
absolute value (or one of these two points, if both Re1fP and Re- 1fP are
singular points for F(z)). The point q= D(O)e- 1fP is the singular point of
go(z) on Izl = D(O) with the smallest absolute value of the argument. We
denote by A(h) the arc of ceh) in the interior of ceO). Since gh(Z) has no
singularities on ceh) - A(h), there is at least one singularity of gh(Z) on
A(h). But go(z) has no singularity on the part of ceO) exterior to C(h).
Since the points - h +D(h)e19 of ceh) tend to the points of ceO), uniformly
in 8, as the points of A(h) tend to the part of the arc D(O)e1t/t where 1"'1 ~ rp,
and since the set of singularities is closed, we see immediately that the
points of intersection of ceO) with ceh) tend, respectively, to D(O)e1fP
and D(O)e- 1fP .
If Xh ± iYh are the points of intersection of the circles ceO) and ceh),
we have:
and by what we have just seen,

·
I1m I· D2(h) - D2(O)
Xh = 1m 2h
hlO hlO
-_ D(O) 1°1m D(h) -h D(O) -- D(O) cos rp,

hlO

We are now going to show that we can preserve the form of the terms
participating in the formation of dn(h), but neglect a certain number of
them, depending on nand h; this makes the formula more manageable
and more susceptible to yield interesting results. So we prove here a
theorem on the distribution of the singularities of F(z) on its circle of

convergence, which is of a rather curious character. Later we shall try
to clarify the difference between this proposition and the classical 'lacu-
nary' propositions.
We begin with some lemmas.
IV. 1.2. Let lim sup lanl 1/n = 1 and let a~O, A>O be such that
A
(IV.S) a > 1 + A- log (1 + A).
Let {Uj} be a sequence of positive numbers tending to zero and suppose
that to every j there corresponds a sequence {m~} of positive integers such
that
(IV.6) lim sup lamlll/m~

n-+ 00 n
> 1 + aUf + o(u j) (j ~ (0)
and that
am = 0, for m E ]mW - AUj), m~(1 + Uj)[, m =f. m~.
We set hj=uje-«, v~=E[mW+uj)] (E[a] denoting the integral part of

a); if dn(h) is defined by
(IV.7) dn(h) = aohn + G) a1hn- 1 + ... + an,

then
Remark. The condition (IV.S) holds, in particular, if a and A satisfy one

of the following conditions:
(1) A = 1, a > ! - log 2

(2) a = 0, A > 0 arbitrary.
To prove IV. 1.2 we first have to demonstrate the following two lemmas.
IV. 1.3. Let lim sup lanl 1/n = 1, :> > 1, and let {Pn} and {qn} be two sequences
of positive integers tending to infinity. If for h > 0:
(IV.9) lim sup (qn/Pn) ~ I - Oh,
48 DIRICHLET SERIES
then
(IV. 10) lim sup laohPn + (~n) a1hPn -1 + ... + (::) aqnhPn -qn 11/Pn
~ 1 + 0(1 - log o)h + o(h) (h ~ 0).
We set
on(h) = aoh Pn + (~n) a1hPn -1 + ... + (::) aqnhPn -qn.

We see, first of all, that for every a> 1
Also for every b > 0:
Pn(w) = 1 + (~n) W + ... + (::) wqn
= -1.
hz
f ( + -W)P
1
2
n
(1 + z + ... + zqn)-
dz
z
where the integral is taken over Izl = b; consequently for b> 1 :
which allows us to write, when (IV.9) holds:

lim sup IPn(w)1 1/P n ~ (b + Iwl)b- 6h •
It is sufficient to set w=hja, a> 1, b= 0 with at 1 to obtain (IV.10).
IV. 1.4. With the notation and conditions of Lemma IV. 1.3, let
c> 0(1 -log 0). If
lim sup I(~:) hPn -qn + ... + apn rlPn ~ 1 + ch + o(h).

and if dn(h) is defined by (IV.?), then:
lim sup Idpn(h)11/P n ~ 1 + ch + o(h) (h ~ 0).
This lemma is an immediate consequence of Lemma IV.1.3 (where qn

is replaced by qn -1).
We now prove Lemma IV. 1.2. Let e>O and for fixedj let:
Pn = E[mW + Uj)], qn = E[mW - ,\Uj)] (n ~ 1).
When j is sufficiently large (Uj sufficiently small), we have on the one
hand:
q I-'\u,
(lV.1I) lim sup -2! = 1 J = 1 + (1 + '\)Uj + o(Uj)
Pn + Uj
= 1 - (1 + '\)eah j + o(h j )
~ 1 - {(I + ''\)ea - e}h j ,

and on the other hand:
. sup I (Pm~ )am~hPn -mn Il/Pn

= hm j
~ (1 + Uj) (~
h) Uj/(l-U j )
(1 + aUj + o(Uj»)1/(1 +U j)
= 1 + eahj + o(hj) (j -+ (0).

Since it follows from (IV.11) that for sufficiently largej:
lim sup qn ~ 1 - Sh j (3 = (1 + '\)ea - e),

Pn
c = ea > 3(1 - log 3),
the conclusion of Lemma IV.1.2 follows from Lemma IV.1.4.
We now prove the following theorem.
THEOREM IV.1.5. Let

F(z) = 2: anzn
be a Taylor series with radius of convergence 1 and ,\ > 0, 0 ~ a> '\/(1 +,\)-
log (1 + ,\). Let {hj} be a positive sequence tending to zero and suppose that
to every j there corresponds a sequence of positive integers {m~} and a
sequence of positive numbers {~~} such that 2
(1) 9\(am eIJ1~) ~ 0 for mE](1 - ,\hj)m~, (1 + hj)m~[
(2) lim sup 19\(am;, eIP;')11/m{, ~ 1 + ahj + o(hj) (j -+ 0).
50 DIRICHLET SERIES
Then the series F(z) has at least one singularity at ei8 with
COROLLARY to IV.1.5. Let F, a, ,\ be defined as in IV.1.5, let {Uj} be a

sequence tending to zero, and suppose that to every j there corresponds a
sequence of positive integers {m~} such that
(IV.l2) lim sup lamd 1fm' ~ 1 + aUj + o(Uj) (j -+ (0)
am = 0 for m E ](1 - Mj)m~, (1 + hj)mM, m -# m~.
Under these conditions there exists a singularity of F on every arc of the

circle Izl = 1 with an angle 21arc cos elll, (Iarc cos elll ~7T/2).
To prove Theorem IV.1.5 it suffices to set
and to note that by Lemma IV.l.2:
lim sup Idv~(hjWfV'
~ .
hm sup I ((aohjn + (vt)t a1hvn
~ I I -1 + ... + av~) elfinf) IlfV~
~ 1 + ellhj + o(hj );
this proves that when we denote by eilP the singularity of F(z) on Izl = 1
nearest to z = 1 we have
cos cp = D'+ (0) = I'1m D(h)h - 1 ~ eIl .

h~O
This proves the theorem, at least for the singularities situated on an arc
of angle 21arc cos elll centered at z= 1. An arbitrary rotation proves the
rest.
Now we indicate the essential difference between Theorem IV.1.5 (and
its corollary) on the one hand, and the older theorems in the same range
of ideas on the other.
In the classical theorems one studies the behaviour of the coefficients
am when the index m varies in intervals around indices mn corresponding
to 'major' coefficients, that is, those for which lim lamnllfmn=1 (where,
for example, all the other am are zero, as this is the case in the Corollary
to Theorem IV.1.5), the length of these intervals being proportional to

mn with a fixed proportionality factor. In contrast, here we have just
introduced intervals around the indices m~ without assuming any se-
quence {am!} n
(for fixed j) to be 'major' (see (IV.l2), where the propor-
tionality factors of the length of these intervals decreases, and where the
'importance' of the sequence {am!} n
'increases', with decreasing Uj in
(IV.12), a being assumed to be non-negative). This proportionality fac-
tor (,.\ + 1)hj tends to zero as j tends to infinity, whereas the expression
lim sup laml,1 1 /ml,
n
(in (IV.12» can never be equal to 1 (if a < 0), but tends to 1 as j tends to
infinity.
We end this chapter with the following theorem, which is also essen-
tially based on Theorem IV. I. I.
THEOREM IV.I.6. If the radius of convergence of (IV.l) is I, if

lim la nJ 1 1/n! = 1'
and if there exists a constant 3> 0 such that no polynomial of the sequence
dniz) = aoZ n} + CJ)a1Znj-1 + ... + anj (j ~ 1)
vanishes in the disk Izl < 3, then F(z) has at least one singular point at ejt{J
with
. an -1
COS g; ~ hm Sup ~ - '-.
an}
We see from this theorem that if lim sup ~(anr1/an!)= 1, when the
other conditions are satisfied, then Z= I is a singular point of (IV. I).
In the notation of Theorem IV.l we have for h ~ 0,
·
(IV .13) Iog D(h) = 11m log Idn(h) I '- I· log Idnih)I
sup .,:;; 1m sup .
n nj
There exist two constants k>O and L>O such that lanl <Lkn (n~O);
consequently
Idn(z) I = laoz n + (~)alzn-l + ... + ani

~ L(lzln + G)k lzl n- 1 + ... + kn) = L(lzl + k)n.
52 DIRICHLET SERIES
Since dn,(z) does not vanish in Izl < 8, we can define arg dn,(z) in this
disk in a continuous manner, by defining arg dniO) equal to arg an" the
latter lying in [0,217'[. The functions log dn,(z) = log Idn,(z)1 +i arg dn,(z)
are then holomorphic in Izl < 8, and their real parts satisfy the inequality
~ log dn,(z) = log Idn,(z) I : : :; log L + nJ log (k + 8).
The real parts of the functions
CfJn (z) = log dn/(z)

/ nJ
are bounded. These functions therefore form a normal family in Izl < 8.
Consequently, we can extract from {nJ} a sequence {mp} having the follow-
ing two properties:
· am -1 = 1·Imsup~--
( 1) 1Im~-P- an,-1 = a;
amp an,
(2) the sequence CfJmp(z) = log dmp(z)/mp tends in Izl:::::; 8/2 uniformly to
a function holomorphic in this disk, because we have
·
1ImCfJmp log amp = 1·1m log lampl
(0) = 1·Im-- + i arg amp = 0.
mp mp
On the other hand, it is easy to show that
d~(z) = ndn_1(z),
and that in Izl < 8
1 d () 1 (2 d~(t) d 1 (2 dn- 1(t) d

og n Z = ogan + Jo dn(t) t = ogan + n Jo dn(t) t.
Hence, by (IV.13) we have for 0:::::;h:::::;8/2:
log D(h) ~ lim log Idmp(h)1

mp
· (~log amp
= 1~
mp
+~
'" ill
0 dmp(t)
d)
dmp - 1(t) t·
But since
we have
1 (h) 1· fit dmp-1(t) d

og D ~ 1m 9t Jo dmp(t) t.
Since D~(O) exists (Theorem IV.1.l) and from the fact that D(O)=
= lim sup lanl 11n = 1, we also have
D~(O) = (log D(h»~(It=O) = lim log f(h)

It.o
·
~ 11m -h
It.o
1 (1'1m 9t
p-+co
fit dmp-1(t)
0
dmp() d)
t
t
But by Theorem IV.1.l we know that (IV.l) has a singularity eiqJ with
cos p = D~ (0). This completes the proof of our theorem.
NOTES
1 We recall that a 'singularity of a Taylor series on its circle of convergence' is de-

fined as a point Zo with Izol=R such that for no e>O there exists a function holo-
morphic in the disk C(zo, e) that is equal to the series in C(O, R).
2 9l(a) denotes the real part of a.
CHAPTER V
COMPOSITION OF SINGULARITIES
V.l. Composition of Hadamard type and generalizations
Without doubt, one of the most important theorems in the theory of the
distribution of singularities of a Taylor series is that of Hadamard con-
cerning the mUltiplication of singularities, which he discovered in 1898.
Since later in this chapter we shall have occasion to state the theorem
in its general form, we give it here under a form (due to Hadamard)
which is, in fact, only valid when the singularities are isolated and the
functions themselves are uniform.l
If Eh i = 1, 2, 3, is the set oj singularities oj the Junctions represented by
2: anzn, 2: bnzn, and 2: anbnzn, respectively, then to every point ea oj Ea
there corresponds a point el oj El and a point e2 oj E2 such that ea = ele2'
Translated into Taylor-D series this statement becomes the follow-
ing.
If y is a singularity of the function represented by 2: anbne - ns, then there
is a singularity IX of the function represented by 2: an e- ns and a singularity
/3 of the function represented by 2: bn e- ns such that y=IX+/3.
We assume in these statements that the radii of convergence of the
series 2: anzn and 2: bnzn are finite and non-zero, or what comes to the
same, that the abscissae of convergence of the Taylor-D series are finite.
The statement on the Taylor-D series ceases to be true for general
Dirichlet series. In other words, the function represented by 2: anb n e- AnS
can have singularities other than those of the form IX +/3, where IX is a
singularity of the functionJ represented by 2: an e -An B, and /3 is a singularity
of the function cp represented by 2: bn e-An S • This remark is valid even if
each of the functions J and cp has only one singularity and the functions
themselves are uniform. For example, if an = bn= 1 (n ~ 1), and An = log n,
then bothJand cp are the Riemann zeta-function ~(s), which has no singu-
larity other than a simple pole at s= 1. Hence, if the theorem were true,
the function represented by the 'composite series' 2: anbn e- AnS could not
have singularities other than the point s=2. But of course, the composite
function is still ~(s).
COMPOSITION OF SINGULARITIES 55
The fact that the statement given for Taylor-D series cannot be 'trans-
ferred' directly to general Dirichlet series is a challenge to search for a
general theorem of a type depending on properties of A, which for A=N
(An=n) turns into that on the Taylor-D series.
We give here several theorems on the composition of singularities of
Dirichlet series, which all contain as a particular case Hadamard's
theorem corresponding to An = n, each of the theorems having its specific
character of generality.
Although the theorems we state are valid for multiform functions (after
the terminology has been made somewhat more precise), to simplify the
language we assume that they are uniform - at least 'uniform in a half-
plane' - an expression to be made explicit later. However, we shall see
that these restrictions can be removed.
So let us consider the series (A, A)=2:an e- llnS with Ua<OO, and let
Ul> - 00. Let LI be a domain contained in PfJl and containing
PfJ,(PfJ , c LI c PfJl ) for a certain u' > max (Ub ua), and such that the sum
f(s) of (A, A) has a uniform analytic continuation from PfJ , to LI.
Given Ub if there exists a domain, which we denote by LI(u!), containing
all the other LI having the properties listed, we say that the function J,
which is holomorphic in LI(u!) and is there the continuation of (A, A)
starting from PfJ " is uniform in Pal' The set consisting of all the points of
Pal that do not belong to LI(u!) is denoted by SfJl' We call this set the
singular set off relative to the half-plane Pal' The points on the line U=Ul
evidently belong to Sal' The theorems will not be interesting unless Sal
contains points not on the line U = Ul, but for the results we shall give later to
be always valid we have to define the singular set in the manner indicated.
If u~ < Ub the analytic continuations of (A, A) into LI(ul) on the one
hand, and into LI(uD on the other (if the two continuations are uniform)
take in LI(aD () LI(u!) the same value; it is therefore not inconvenient to
denote such a continuation (into a LI(u!» by the same letter f(s) as the
sum of (A, A) in PfJa . Note that LI(u!) c LI(uD. To say then that 'f(s)=
(A, A)= 2: an e- llns is uniform in the half-plane Pal' means that Llh)
exists, thatfis hoi om orphic in LI(u!), thatfis given by (A, A) in the part
of LI(u!) belonging to Pac' and that f has no analytic continuation in a
neighbourhood of a point Sal except along a curve intersecting the line
u=u!.
A general Dirichlet series does not necessarily have singularities on the
axis of convergence, in contrast to a standard fact for the Taylor-D series
56 DIRICHLET SERIES
(a Taylor series has at least one singular point on its circle of convergence;
hence a Taylor-D series has at least one singular point on every segment
of length 21T on its axis of convergence). For example, the function repre-
sented by the series 2: (-l)nn- S is entire, although O'e=O for this series.
Now let H be the greatest lower bound of the quantities 0" such that
f(s)=(A, A) is holomorphic in Pu" This H is called the abscissa of holo-
morphy of (A, A). Evidently H ~ 0'e ~ 0'a' The line 0' = H is the axis of
holomorphy of (A, A).
The quantity
H(O'l) = sup
ses a1
0', (s = 0' + it)
is called the abscissa of holomorphy of (A, A) in the half-plane PU1 ' We have
H(O'l)=H if 0'1 ~H, and H(O'l) = 0'1 if 0'1 > H.
Let E be a closed set contained in a half-plane 0' < 0'1.
The expression' the only possible singularities of (A, A) in PU1 are the
points of E' has the following meaning: if we denote by (P Ul - E) that part
of the difference PU1 - E (the complement of E n PU1 in puJ which forms
a domain (a connected set of interior points) containing a half-plane
PU2 with 0'2> O'e, thenf(s) is holomorphic in (PU1 - E) (and in PU2 we have:
f(s)=(A, A».
It is important to emphasize that O'~ < 0'1 does not imply SUl C Sa~.
This is the reason why in the definition of SUl we use the expression
'relative to Pu/ and not 'in Pu/.
in
Letf(s)=(A, A) be uniform PU1 ; for e>O we set:
If to every e> 0 there corresponds a constant K ( = K( e, 0'1» such that 2

on LI(O'b e)=(PU1-SU1(e»: If(s)1 <K, we say that f is bounded in Pup
except at the singularities. It is clear that a Taylor-D series, uniform in
0' > 0" is bounded except at the singularities. But this is not the case for
all the Dirichlet series. For example, the function ns) is not bounded,
except at the singularities in any half-plane PUl with 0'1 = 1 - 8 (8) 0).
Let us now suppose thatf(s) = (A, A), uniform in PU1 ' is such that there
exists a non-negative constant m with the property that to every e> 0
there corresponds a K(e, m) (=K(e, m, 0'1» such that If(s)1 <K(e, m)ltl m
holds inLl(O'b e) for sufficiently large Itl. Let p. be the greatest lower bound
of these quantities m. In other words, to every S> 0 and every e > 0 there
corresponds a constant
N(e, 8) (= N(e, 8, 0'1»
such that for sufficiently large It I

If(s) I ~ N(e,8)jtIItH,
in LI(a1o e) and no quantity smaller than f.L has this property. We then say
that f is of order f.L in Pal except at the singularities.
Let A(t) be a function that is positive for t;;:: 0, increasing, tending to
infinity. We say that f(s)=(A, 11) does not grow faster than A(t) in Pal
except at the singularities if to every e > 0 there corresponds a
K(e) = K(e, 0'1) such that in LI(a1o e) we have
If(s) I ~ K(e)A(ltj).
Let Q(0'10 to) be the quarter-plane given by a> 0'10 t> to. Suppose that
(A, 11) has the following properties. Let LI be a domain such that
Q( aI, to) :::> LI :::> Q(0'2, to), where 0'2 is a certain quantity 0'2> 0'1; f(s),
given by (A,11) in Pa2 , being holomorphic and uniform in LI. Suppose
also that there exists a domain D(a1' to) which contains all the domains
LI just described. Then we say thatf(s), given by (A, 11), is uniform in the
quarter plane Q(a1o to). We denote by SabtO the set of points s satisfying
the inequalities 0';;::0'10 t;;::t o and not belonging to D(a1o to). Then Sa,to is
called the singular set off relative to the quarter plane Q(a1o to). Next,
let Sal,tO(e) be the union of the disks C(s, e) with s E Sal,tO' If to every
e > 0 there corresponds a constant K(e) (=K(e, aI, to» such that in the
closed domain (Q( 0'10 to) - SalotO( e», which is the part of the closure of
Q(a1o to)-Sal.tO(e) connected with Q(a2' t 1) we have If(s) I < K(e) , then
we say that f is bounded in Q(0'10 to) except at the singularities. Iff is uni-
form and bounded in Pal except at the singularities, then it has the same
properties in Q(a1' to) for any to.
We introduce the quantities
H+(al) = lim arCto),

to ...... 00
It is evident that H+(a1) exists and that H+(al);;::al' We call H+(a1)

the ultimate (+ )-abscissa of holomorphy off in the half-plane Pal'
58 DIRICHLET SERIES
Since the first theorem of this chapter concerns functions f(s) = (A, A)
that are bounded, except at the singularities, in a half-plane or quarter-
plane, it is essential to prove that there exist Dirichlet series which are
not Taylor-D series (or series of the form L'. an e- kns , where k is a con-
stant), and which have these properties. Let (A, A) be an arbitrary series
with U a < 0, and let T(z) = L: cnz n be an entire function. Consider the func-
tion
F(s) =T [f(s) +e_1_] = T(L'. an e-

S - 1
AnS + L'. e- ns )
= L'. cmCL'. an e- An8 + L'. e-ns)m.
Since for u> 0 the series L'. Icml(L'. lanl e- An " +L'. e-M)m converges, we
see that F(s) is represented for u> 0 by a Dirichlet series
where the Vn are of the form Vn = L,..;p a)lj+{3, the (x,

and {3 being non-
negative integers. On the other hand, the functionf(s)+(e 8 -l)-1 has in
the half-plane u>ua only simple poles at the places 2hri (where k is any
integer). Besides, this function is bounded for u> U a + e outside the disks
C(2hri, e).
In the same domain F(s) is also bounded. But in each of these disks the
values taken by f(s) + (e S -1) -1 cover a neighbourhood of infinity, and by
Picard's theorem F(s) takes in these disks all values, except possibly one,
infinitely often. The points 2k7Ti are therefore (essential) singularities of
F. Thus, this function has the properties required for every U1 > ua".
Before we state our theorems, let us make a convention: if in a state-
ment several Dirichlet series occur, for examplef(s)=(A, A), <p(s) = (B, M),
we attach to the various elements we have to introduce the letter indi-
cating the relevant function. Thus, the abscissae of simple or absolute
convergence off will be denoted, respectively, by U e", u a ,,; the abscissa
of holomorphy by HI; the singular set of f with respect to P<11 by S<11,1'
etc.
If the analytic continuation of a series f(s) =(A, A) is uniform, we de-
note by S, its singular set (in the entire plane), that is, the complement
in the plane to the domain of existence of f (the total domain where
(A, A) can be continued analytically).
In all the subsequent cases we assume that the Dirichlet series occurring
in the conditions have an abscissa of absolute convergence without stat-

ing this every time.
The theorem we are about to state contains as a particular case Hada-
mard's theorem on Taylor series, although V.I.1 concerns much less
general series than the next theorem (V.1.2).
THEOREM V.I.I. Let f(s)=(A, A) be uniform and bounded in Q(Ub to),

except at the singularities, and let cp(s) = (B, A) be uniform and bounded in
P"2' except at the singularities. Let S""to,1 be the singular set off with re-
spect to Q(Ul, to) and let S"2,({! be the singular set of cp with respect to P"2'
We set cn=anb n (n?;: 1), C=:{cn}. The function F(s)=(C, A) has the
following properties: Ua,F'~ Ua,l + Ua, F is uniform in P"3' where
({!,
u3=Ht(U1)+U2, and the only possible singularities of (C, A) in P"3

are the points 3 of S""to'/ +S"2,({! and the limit points of this set.
Let c>ua,f, t1 > to, and let us write for z=x+iy
FT(z) = 1
T'
11
i C
c+t,i
+ T1 f(s)cp(z - s) ds,
where the integration is taken over the line u=c. Then FT(z) is holomor-
phic for x>c+u a ,({!' for when U=C, we have Z-SEP"a,({!' If z satisfies
this condition, we have
FT(z) = -.
1 [C +
f(s)
TI
2: bn e-II.(Z-s) ds
Tl, c + t1i
= L, bn e- lI • z -1.
TI
i C
c+t,i
+ Ti f(s) ell • s ds.
But according to Theorem 1.3.1 we have
(V.I) 1
T'
11
i
C
+ Ti f(s) e llns ds =
c+t,i
an + enCT),
with lim enCT) = 0 as (T -+ CI)). Thus,

F1{z) = 2: anbn e- ilnz + L en(T)-lIn Z•
The series (C, A) converges absolutely for X>Ua,({!+C, because
60 DIRICHLET SERIES
where
We also have L: Ibnl e-An(x-c)<oo for x-c>aa,rp.

Hence the series (C,11) converges absolutely for a>aa,t+aa,rp, and
aF~aa,,+aa,rp,
For x>aa,rp+c we have:
IFT{z) - F{z) I = IL: bn e-AnZ[an + en(T)] - L: anbn e-Anzi
and since by (V.1) leinl ~2.A(c) eAnC, we have for each p:
IFT(z) - F(z) I
~ 2: Ibnen{nl e- AnX + 2.A(c) 2: Ibnl e-An(X-C);
n~p n>p
but on the one hand, for fixed p, en{T) -+ 0 as T -+ 00, n = 1, 2, . , ., p,
and on the other hand, L:n>p Ibnl e-(A n-c) -+ 0 as p -+ 0, x- c> aa,rp. So
we see that lim FT{z)=F{z) for x>aa.rp+c.
Suppose now that c>max (aa", a1) and let us divide the strip al ~a~c
into squares whose sides are parallel to the axes and oflength e=(c-al)/q
(where q is an integer). We denote by A a square of this division having
the following property: the closure of the union of this square with all its
adjacent squares does not contain any point of SUl,to," Let '1){e) be the
set consisting of the union of all the squares A of this division. If e is
sufficiently small (q sufficiently large), then all the squares based on a = c
form part of'1){e). We assume e to be chosen this way. Let '1)*{e) be the
largest domain forming part of '1){e) and having a= c as part of its boun-
dary. Let '1)*{e, t1) be the part of '1)*{e) for which t> t 1 • Thenf(s) is holo-
morphic and bounded in '1)*{e, t1)' Let C(e) be the boundary of '1)*(e, t 1).
Let {am} be a sequence tending to +00, with am> t1 (m ~ 1); we denote by
'1)*{e, t1> m) the part of '1)*{e, t 1) for which t < am, and by C'{e, 11> m)
the part of its boundary that does not contain the points of a = c. It
follows from Cauchy's theorem that for x>aa,rp+c
(V.2)
(the integral being taken in a convenient sense) and consequently F(z)

for the same values of x is the limit of the integral (V.2) as m tends to
infinity.
Let 1m be the set of points forming the segments of C'(e, 11> m) for which
t=a m, and let L be the set of points forming the segments of C"(e, t1> m)
for which t= h. Since for given m the total length of each of these sets
of segments does not exceed c-a1> and since on these segments IJ(s) I
and Irp(z-s)1 (for x>aa,Q.l+c) are bounded, we see that when we denote
by C(e, t1. m) the set
C"(e, 11> m) - 1m - L
we have F(z)=lim F;m(z), where
F;m(z) = ~ r
lam Je(e,t1,m)
J(s)rp(z - s) ds.
Now letLl be a bounded domain of the plane z=x+iy containing points

with x> aa,Q.l + C such that for every z ELl
x > H/(al) + a2 - 3e
and that for every couple z, y with z ELl, y E SG1,to,f +SG2,f we have
Iz-yl > 6e. If 11 is sufficiently large, we have for every s E C(e, 11> m),
z ELl, fJ E SG2,Q.l
9\(z - s) = x - a > a2, Iz - s- fJl > e.
Indeed, if 11 is sufficiently large, then for every s E C(e, t1> m) we have
a<HtCa1)+3e; and for z ELl we have X-a>a2' Ifwe had Iz' -s' -WI <e
for a z' ELl, an s' E C(e, f1' m), and a fJ' E SG2."" because to every s' indi-
cated there corresponds an a' E SG1,t O.f such that Is' - a' I < 2V2 e, we would
have Iz' -a' -,8'1 ~ Iz' -s' -,8'1 + Is' -a'i < (2V2+ l)e, contrary to the as-
sumption that Iz' -a' -,8'1> 6e. With this choice of t 1 , for every
s E C(e, t1> m), z ELl, the variable u=z-s is situated in that part of the
half-plane a> a2 which is connected with the half-plane of absolute con-
vergence of rp and in which rp(s) is holomorphic, uniform and bounded.
Thus, the functions F;m(z) defined above form a bounded family of holo-
morphic functions in Ll. But in a closed part of Ll situated in the half-
plane x>aa,,,,+c the sequence F;m(z) converges uniformly to F(z)=
2: anbne-AnS, and by the Stieltjes-Vitali theorem, F;m(z) converges uniformly
to a function holomorphic on every compact set in Ll. This limit is the
analytic continuation of F(z). Thus, the 'composite' sum 2: an bn e- AnS
can be continued analytically into every domain Ll of a kind we have
defined. This proves the theorem.
62 DIRICHLET SERIES
Theorem V.1.1 contains as a particular case Hadamard's theorem,

because for the Taylor-D series 0"1 and 0"2 can be chosen arbitrarily, at
least if the functions are uniform. (For multiform functions a statement
valid for Dirichlet series of an order of generality comparable to that of
Theorem V.1.1 can also be obtained; later we shall have to say some-
thing about this.)
The fact that to is arbitrary in Theorem V.1.1 does not improve the
corresponding statement for the Taylor-D series, because f and cp are then
periodic functions of period 271'i. But the arbitrary choice of to plays an
important role in some applications of Theorem V.I.I to general Dirichlet
series. Note that for two Taylor-D series the sum Sql.f+ Sq2.<P is closed,
because the sets Sql.t, and Sq2,<P are themselves periodic of period 271'i.
V.2. Composition of functions of slow growth

Before dealing with Dirichlet series of a much more general character
we have to introduce some new definitions. Let M = {lLn} be a strictly in-
creasing sequence of positive numbers tending to infinity; as before, let
f(s) = (A, A) = L: an e- An8 , and let k be a positive integer. Let no be the
smallest integer n such that ILn> A10 and let us set
a~k) = 0 (0 < n < no, for no > 1)
al.k) = L (ILn -
Am<Jln
Am)kam (n ~ no).
The quantity al.k) is called the n-th coefficient off (or of (A, A» relative
to M, of order k.
THEOREM V.2.1. Let f(s)=(A, A) be uniform and of order v in Pq1 , except

at the singularities, and let cp(s) = (B, M) be uniform and of order IL in
Pq2 , except at the singularities.
Let Sql,t be the singular set off with respect to the half-plane Pq1 , and
Sq2,v+lL. Then the series L: a~k)bn e- lln8 ,
where al.k) is the n-th coefficient off relative to {lLn}, of order k, has an ab-
scissa of absolute convergence not exceeding
the function Fk(S) represented by this series is uniform in the half-plane

a> max (HI ( (1), 0) +a2, and the only possible singularities of Fk(S) in this
half-plane are the points belonging to Salol +Sa2,<P' the limit points of this
set, and the points of Sa2.!'
Let c> max (0, aa,!) and consider
(V.3)
_i~
F(z) - 2 .
TTl
C+1OO
c-loo
f(s)cp(z - s)
S
k+l ds,
the integration being taken over the line a = C.

F(z) is holomorphic if x>c+aa,<p (z=x+iy), for when a=c, the vari-
able z-s remains in the half-plane of absolute convergence of cpo For
such values of z we can write:
k ' fC+IOO 1:( )" b -Un(Z-Sl_S

_. d
2TTl. c-ioo J'S L. n e k+1
f
S
b e-UnZ -k! + lOO ds
C
= '" '" a e(Un-~mls_
~ n 2TTi c-ioo ~ m Sk+l
and 1.3.3 enables us to write (for these values of z)
F(z) = 2:
n~no
bn e- UnZ 2:
Am <Un
am(fLn - Am)k = 2: a~lbn e- unz ,
where no is the smallest integer n such that there exists a Am less than fLn.
The last series for F(z) converges for x>c+aa,<p, But (V.3) is still valid
when we replacef(s) by L lanl e- llns and cp(s) by L Ibnl e- UnS .
Consequently the series
L L
n~no i\m<Un
laml(fLn - Am)klbnl e- UnX
converges for x>c+aa,<p, which proves that the abscissa of absolute

convergence of 2: ditklbn e- UnS does not exceed max (aa,<p, aa,!+aa,<p) (c is
only subject to the inequality c> max (0, a a,!».
Evidently F(z) = Fk(Z).
Now we choose c> max (0, a1. aa.!); and as in the proof of Theorem
V.l.1 we divide the strip al~a~c into squares of sides parallel to
the axes and of length e=(C-al)/q; let 'l)l(e) be the set consisting of the
union of the squares of this division having the following property: the
closure of the union of such a square with all its adjacent squares does
not contain any point of Sal.! nor the point s = 0. If e is sufficiently small,
64 DIRICHLET SERIES
all the squares based on u=c form part of 'Ilt(e), the largest domain
contained in 'Il1(e) having u=c as part of its boundary. The functionJis
holomorphic in 'Ilt(e), and to every Sl > 0 there corresponds a constant
N1(Sl) (=N1(Sh e, Ul)) such that IJ(s) I< N1(Sl)ltI UHl on 'Ilt(e) for suffi-
ciently large It I.
Let 'Ilt(e, T) be the part of 'Ilt(e) for which It I<T, and let ct(e, T)
be the part of its boundary that does not contain the points of u = c.
For x>ua,rp+c we have, by Cauchy's theorem,
fc~
J(s)rp(z - s)
S
ds
k+l =
iC+lT J(s)rp(z -
c-iT
s)
S
ds
k+l
where the first integral is taken in a convenient sense over ct = ct(e, T),
and the second over the segment of the line u=c.
On all parts of ct(e, T) on which t = - T or t = T, the integrals tend to
oas T tends to +00. It follows that
(V.4) Fk(z) = k'i
-2 ds
'. • J(s)rp(z - s)"""k+I'
7Tl cl(e) S
where ct(e) is the part of the boundary of 'Ilt(e) that does not contain
the points of u=c.
Let L1 be a bounded domain of the plane of z=x+iy containing points
x>ua,.,+c and having the following properties: x>max (Hlul), 0)+
U2+2e for every zEL1, and Iz-yl>6e for every ZEL1 and yE
(SU2," + SUl,f) U SU2,rp. We see that
~(z - s) = x - u > U2,
that Iz-s-,81 >e for every z EL1, every s ECt(e) and every,8 E SU2,rp. This
can be shown in a manner similar to that we have used to demonstrate
an analogous fact in the proof of Theorem V.l.l. Consequently, if
s E ct(e), Z E L1, then the variable u = z - s is situated in that part of the
half-plane u> U2 which is connected with the half-plane of absolute con-
vergence where rp(s) is holomorphic and uniform and where to every
S2>0 there corresponds a constant N 2(S2) (=N2(S2, e, U2)) with the
property:
Irp(s) I :::; N 2 (3 2 ) IW+6 2
for sufficiently large Itl. Since L1 is bounded, we see that
Irp(z - s)1 < N 3 (S2)IW H2
for sufficiently large It I if S E et(e), Z E .d. Since v +fL < k, clearly 01 and O2
can be chosen so that
fL + v + 01 + O2 < k,
and the integral (V.4) converges uniformly in Z on every compact set in
.d. Hence Fk(z) is holomorphic in.d, and the theorem is proved. Theorem
V.2.1 also contains Hadamard's theorem, at least if we assume that the
Taylor series represents a uniform function. In fact, let:
(to simplify the calculations we assume that d1 =0; also do=bo=O as

everywhere in this text) and let zo=e- so be one of the singularities of
L: dnz n on its circle of convergence. The function
lh(s) = L: d n e- nso e- ns = L: An e- ns
has a singularity at the point s=O. We set
(1 - S-S)2 eS{l1(s) = /(s) = L: an e- ns •
Then
An = (n - I )a1 + (n - 2)a2 + ... + an -1
=
m<n
2 (n - m)a m (n;?; 2).
The singularities of/are those of {lb except the points 2k1Ti, which may
be singular for {l1 but not for f Since for the Taylor-D series f and cp
the quantities 0'1 and 0'2 can be taken negative, of arbitrarily large abso-
lute value, and since for such a choice of 0'1 and 0'2 we can always take
k to be equal to 1, we see by Theorem V.2.1 that the only singularities
of F(s) = L: d;,l)b n e- ns , where
a~1) = 2 amen -
m<n
m) = An (n > 1),
and a~1)=0, are the points of Sf+S", and of S",.

Let SljJl be the set of singularities of {l1; since S,it! contains Sf and the
point s=O, we see that SljJl + S",= (S, + S",) uS",. In other words, the only
possible singularities of
66 DIRICHLET SERIES
are the points of S"'l + S",. Clearly, every point of S"" the set of singulari-
ties of !fr, is obtained by adding So to a point of S"'l' In other words, the
only possible singularities of l/J = 2: dnb n e - ns are the points of S", + S",.
This is the most precise statement of Hadamard's theorem, at least for
uniform functions. For multiform functions we shall make the appro-
priate remarks later.
Coming back to Theorem V.2.l, we see that if (11 <0 and if s=O does
not belong to SU1.r, then for sufficiently small e>O the boundary Ct(e)
contains a square (of side 3e) containing in its interior the point s = 0,
the square being separated from the other points of Ct(e). Let c(e) be
the boundary of this square. The part of Fiz) obtained by integration
over c(e) gives (by the residue theorem) for sufficiently large x:
k'l
liz) = -2'.
7Tl c(e)
ds
J(s)g;(z - s) k+1
S
= Ok[J(S)g;(z - s)] (s = 0)
OSk
= (-I)k[J(O)g;(k)(z) - mf'(O)g;UC -1)(Z)
+ ... + (-1 Y'.f(k)(O)g;(z)].
On the other hand, we see by following the proof of Theorem V.2.1
that the only possible singularities of the function
Fk(z) - lk(z) = k'i

-2'.
1T1 C~(e)
ds l'
J(s)g;(z - s) k+
S
where C:(e)=Ct(e)-c(e), are the points of SU1.t + SU2.",· From the form
of Mz) it follows that the only possible singularities of this function are
those of g;(z). Theorem V.2.1 can therefore be made more precise in the
following fashion:
THEOREM V.2.2. Under the conditions oj Theorem V.2.l, if s=O is a regu-

lar point oj J(s) , then the Junction Fis) = 2: dnk)b n e- llnS is uniform in the
half-plane (1 > Hi (11) + (12, and the only possible singularities in this half-
plane oj the Junction
Fk(S) - lis),
where
Ms) = (-l)k(f(O)g;(kl(s) - mf'(0)g;(k-1)(s)
+ ... + (-l)kj<kl(O)g;(s)],
The fact that in Theorems V.2.l or V.2.2 the two sequences A={An}
and M = {fLn} may be chosen completely independently allows a large
number of interesting applications. Thus, since the distribution of singu-
larities of a Taylor series is much better known than that of Dirichlet
series, one would achieve considerable progress in reducing the problem
of distribution of singularities of a Dirichlet series to that concerning on
the one hand the Taylor series whose coefficients depend on the sequences
A and A (the coefficients and the exponents of a given series) and on the
other hand, to an investigation of the singularities of a Dirichlet series
whose coefficients are simple (and given in advance). For example, we
shall show that in fairly frequent cases an indication of the position of
the singularities of (A, A) can be obtained by starting out from a know-
ledge of the places of the singularities of L: dn e-ilnS, where the dn can be
expressed in a simple manner by the An, and of those of a Taylor series
whose coefficients are formed by means of the an and An.
Let A={An}, with AI> 1, let C={cn} be a complex sequence, and k a
positive integer.
We say that C has G={gn} as a generating sequence of order k with
respect to A if for every n ~ 1 :
C n= 2:
m<,1,n
(An - m)kgm.
In general, a sequence C does not have a generating sequence of a given

order with respect to a given sequence A. But it is easy to indicate con-
ditions on A for every sequence C to have a generating sequence of a
given order k.
V.2.3. Let en be the largest integer smaller than An (n ~ 1). If there exists a
sequence {nJ such that enj=nj and that the determinant LIn! (j~ 1) of order
nj whose term in the p-th row and q-th column is (Ap-q)k for 1 :(,p:(,nj,
q:(' ep, and is 0 for l:('p < nj, ep< q:(' enj , is different from 0, then every
sequence C has a generating sequence of order k with respect to A.
In fact, every system S} of equations (in the unknowns gn)
m<Al
2: (AI - m)kgm = Cl
m<nj
2: (Ani - m)kgm = Cn
68 DIRICHLET SERIES
has a system of solutions. If nh < nf2 (nil' ni2 E {nj}), then the quantities
... , g~l~ which are the first nil terms of the solution of the sys-
g~2), g~2),
tem (Sj2) also form a system of solutions of the system (Sil)' The set of
solutions of all the systems (Sf) (j~ 1) thus constitutes a sequence {gn}
(to every n there corresponds a single gn) satisfying all the systems of
equations
:z:
m<ill
(A1 - mtgm = C1
:z:
m<iln
(An - m)1cgm = cn'
V.2A. If n < An ~ n + 1 (n ~ 1), then every sequence C has a generating

sequence of every order k with respect to A.
Here nj = j, and it is clear that
the conclusion therefore follows from V.2.3.

Theorem V.2.1 also allows us to state the following theorem.
THEOREM V.2.5. Suppose that the function F(s) is represented by (A, A)

(A1> 1) and suppose that the function <pes), represented by a series (D, A),
where D={dn} (dn=l=O, n~ 1), is uniform of order fL, except at the singulari-
ties, in PU2' Suppose that {and;; 1} has a generating sequence {gn} of integral
order k > fL with respect to A, with
lim sup log Ignl = a < 00

n '
thefunctionf(s) represented by (G, N)= L gn e- ns being uniform.
Then F is uniform and the only possible singularities in the half-plane
u>max (a, O)+U2 are the points of SU2.<P+Sh and the points of SU2.<P'
This theorem is, in effect, a special case of Theorem V.2.1, where the
quantities gn play the role of the an in Theorem V.2.1, the dn play the
role of the bn in the same theorem, and U1 is an arbitrary quantity with
H,(ul)=a for Ul <a.
So we see, by Theorem V.2.5, that whenever A satisfies the conditions
of V.2.3 (or, in particular, the conditions of V.2A), then there is only
one Dirichlet series that is not a Taylor-D series whose singularities are
used to provide information on the singularities of the given series (A, A).
And, what is essential, the Dirichlet series in question is a series which
we can choose in advance, (D, A)= L: dn e- AnS • The other series with
which 'we compose' is a Taylor-D series.
It can be advantageous to choose dn = 1 (n ~ 1), in which case the
series (D, A) becomes L: e- An ", the analogue to the series L: e- ns for A=N.
One can also choose dn=l/A'(An), where
(V.S) A(z) = Il (1 - ~;) = Ao(iz),
and where Ao(z) has been defined in 1.1.2. In fact, it is known that if A
is a sufficiently regular sequence, then t~e series
(V.6)
presents certain interesting and important features. For example, its

analytic continuation can be expressed by an integral which represents
the extension to the complex plane of the Fourier transform of I/Ao(ix).
Incidentally, we find that JA'(An)J =2/AnAno where {An} is the sequence
associated with A. And for A = N (An = n, n ~ 1) we have A(z) = sin 7TZj7TZ,
A'(n) = (-l)njn(A n = 2),
<P(s) = L: (-1)n ne- ns .
We are going to indicate properties of the series (V.6). But first we have
to evaluate in a particular case the An in a fairly precise form.
V.2.6. If the step of A is positive and if there is an integer k such that

IAn - nJ ,::; k (n ~ I),
then there exists a constant A such that
An ,::; AA~" (n ~ 2k + 1),
where A depends only on a = inf (An + 1 - An) and on k.
We may assume that k > 0, since we have just seen for A = N that
An=2 (n~JI). So we have:
70 DIRICHLET SERIES
where, on the one hand, for n> 2k + 1,
and on the other hand,
Dn ~ n
n+l",m",n+2k
~) n [1 -
(1 - ~
2"
m"n+2k+l
A2]
n
(m - k)
2'
Let Pn be the nearest integer to An. Evidently,
n - k - 1 < Pn < n + k + 1,
and the preceding inequalities allow us to write
A-I
n
~ n
O<lm-nl",2k
11 - A~ I
A~
n
k+l",m",n-k-l
(A~2
m
- 1)
where for fixed n we have:
which proves that [11> AlA; 4\ with Al depending only on a and k;

also: 4
Finally, we have:
with [14 ~ A3A~\ [15 ~ A 4A; 2\ the constants AI, A 2 , A 3, A4 depending

only on a and on K. From this V.2.6 follows immediately.
Here is the lemma concerning the Dirichlet series with exponents A
which can playa role similar to that of the series L: e-nS, or more pre-
cisely, L: (-l)Rn e- ns when A=N.
V.2.7. If the step of A is positive and if there exists a constant k such that
then the series
where A(z) is defined by (V.S), converges absolutely for a> O.

The function c;P(s) (the analytic continuation of the series) is even, holo-
morphic in the part of the plane from which the half-lines a = 0, It I ~ 7T
have been removed; and in the strip It I< 7T we have
-1 fe-ISY
c;P(s) = ~ A(iy) dy,
the integral converging uniformly and absolutely for It I ~7T-e, for every
O<e<7T. In the strip It I<7T-e the function c;P(s) is bounded.
By V.2.6 and by the equality (see p. 69) IA'(An)1 =2JAnAn. there exist
two constants M and d such that
Since the step of A is positive, we see that the series (V.6) converges
absolutely for a> O.
Now we consider in the plane z=x+iy=r ei9 a closed contour eR
formed from the segments 8=a, 8= -a, 0<a<7TJ2, r~R, and the arc
of the circle r=R, 181 <a, where R is chosen so that no An is equal to R.
By Cauchy's theorem we have:
where N(R) is the number of An less than R. We show that when we set
Rm=(Am+Am+l)J2, there exist two constants C and c such that
(V.7) Am = min IA(Rm e18)1 ~ CR;;;c (m ~ 1).
Evidently
(V.8) Am = A(Rm) = TI I1 - R~I

A~ ,
72 DIRICHLET SERIES
and if we denote by M = {/Ln} the increasing sequence formed by the ele-

ments of A and the quantity Rm (for given m), then we have
IA(Rm)l-l = Mm+h
where {Mn} is the sequence associated with M. Now it follows from the
conditions and the way in which M has been formed that there exists
a positive integer 1 such that:
I/Ln - nl ~ I,
the step of M being positive; consequently, by V.2.6 we have
Mm+1 = IA(Rm)I- 1 :::;; C-l/L~+l = C-IR~
which gives (V.7) (when V.8 is taken into account).
It follows from (V.7) that on the arc r=R m, 181 :::;;a<1T/2, we have
for every e>O and m~m(e):
IA(z) I ~ e- tr ,
and for real positive s we can write, as long as r=R m, 181 <a (m>m(e)),
IA(z)
e-s21 ~ e-(scosa-tlr
~ .
So we see that for real positive r (the integral being taken in the positive
sense)
. 1 i e- S2
hm 21Ti J CR A(z) dz
i
(V.9)
i
m
_ • 1 [ Rm e -srefa la
Rm e -sre-!a _la ]
- hm 21Ti - 0 A(re/a) e dr + 0 A(re ia) e dr
From the conditions on A it follows easily that for a:::;; ():::;; 1T/2 we have
for every ~l> 0
\1 - (n ~2 !)21 : :; 11 - ~ \(1 + ~;:),

provided that rand n are sufficiently large. By using next the canonical
product for cos 1TZ we can state that
Icos 1Tzll A(z)l-l :::;; CT1 (1 + !::) = C21 sin (1T~r) Ii/
1
where C1 and C2 are constants, which proves that for sufficiently large
r and a:::; 101 :::;1T/2:
(V.lO) Icos 1Tzl < IA(z) I e6r .
It is now easy to see that for a:::; 101 :::; 1T/2 the integral
1 roo -sre l6
(V. 11) 18(s) = 21TiJO ~(rei8) ei8 dr
converges absolutely and uniformly in the part of the plane of s = (] +it
defined by
t sin 0 - (] cos 0 :::; 1Tlsin 01 - 20.
Consequently, the integral (V.11) represents in this half-plane a holo-
morphic function. It is sufficient to observe that for given D> 0 by (V.lO)
we have
(V.l2) le-sreI61:::; LIA(rei8 )1 etsln8-O"cos8Hlsln81,
where L is a constant.
Coming back to (V.9) we see that each of the integrals in the expres-
f f
sIOn
1 e-sreia la 1 e-srela -ia
21Ti A(r eia) e dr - 21Ti A(r e - ia) e dr
converges absolutely and uniformly in every part of the plane defined by
(V.l3) It I :::; (] cot a + 1T - e,
and that its sum is -4>(s).

By applying Cauchy's theorem and taking (V.12) into account, we see
that for real positive s
laCs) = LaCs) = 11I/2(S) = L1I12(s) = 10(5) = 111(5).
The integral in our statement, which converges absolutely for It I <1T
(by (V.l3) with a=1T/2), converges uniformly in every strip It I<1T-e
and represents -4>(s) in it; this function is bounded in it. Since A(z) is
an even function, so is 4>(s), and this completes the proof of our proposi-
tion.
Remark. If the sequence A is such that to every positive integer k there

corresponds at most one An for which k < An :::;k+ 1, in other words, if in the
74 DIRICHLET SERIES
notation of V.2.3 en +1 - en ~ 1 (n ~ 1) (with A1> 1), then the series (A, A)

can be written in the form '2 a~ e- L• S, where Len = An and L p , for p dif-
ferent from all the en> can be chosen arbitrarily in ]p, p + 1] with a~n = an
and a~=O whenp is distinct from all en. Consequently, by Theorem V.2.5
the study of the distribution of the singularities of (A, A) can be reduced
to that of the singularities of a function represented by a Dirichlet series
(D, L), L={Ln}, where the coefficients D={dn} are chosen arbitrarily,
and to that of the distribution of singularities of a Taylor-D series
'2 gn e- ns, where G is a generating sequence of a certain order with respect
to the sequence C={cn}, with Cn= and;; 1.
The order of the generating sequence depends, of course, on that of
(C, L), except at the singularities (if such an order exists).
In any case, it seems essential to us to emphasize that in frequent and
important cases one can 'separate' from each other the difficulties of the
problem of studying the singularities of a Dirichlet series which stem
from the double presence of the sequence A and the sequence A, by
studying, on the one hand, the series in which only A takes part (for
example, '2 e- AnS or '2 [A'(An)]-l e- AnS), and on the other hand, by in-
vestigation of the singularities of a Taylor series.
Some of the theorems we have established can also serve in the oppo-
site direction by permitting us to state that a function represented by a
Dirichlet series of general type necessarily has singularities at given
places.
Let f(s) be a function represented by (A, A) and let k be a positive
integer. The n-th coefficient off with respect to the sequence N, of order
k, is called the n-th Taylor coefficient of j, of order k. We denote this
quantity by a~)(n).
THEOREM V.2.S. Let f(s)=(A, A) be uniform in Pal' and suppose that f

is of order Y, except at the singularities, in this half-plane.
Let (J be a singularity of the function F(s) represented by '2 a~)(n) e-nS,
where k is an integer with k > Y, and suppose that {J is not of the form 2m1ri,
where m is an integer. To every e>O there corresponds an integer p such
that there exists a point IX belonging to Sal.! for which
IIX - + (J)I < e.
(2p1Ti
The function <p(s) = L e- ns is uniform in every P a2 and is there bounded,
except at the singularities. We can therefore apply Theorem V.2.1 to f
and rp with k> v. The only possible singularities of F in the entire plane
are consequently the points of SUI.! +Scp = S and the points of Scp. But the
points of S([! are at the places 2m1ri. If f3 is a singularity of F, but not of
the form 2m1Ti, that is, if it does not belong to S([!, then this point must
necessarily belong to S. In other words, either there exists a point
a1 E SUI,! and an integer m such that f3=al +2m1Ti, or there exists a se-
quence an with an E SUI,! (n ~ 1) and a sequence of integers {m n} such that
f3=lim (CX n+2mn1Ti). In the latter case, for given e>O and sufficiently
large n
1f3 - CXn - 2mn1Tij < e.
Thus, the theorem is proved.
The theorem we have just proved enables us to associate with every
theorem that provides a condition for a given point to be the place of a
singularity of a Taylor series a theorem on general Dirichlet series (of
finite order, except at the singularities) indicating the position of some of
its singularities.
V.3. 'Fictitious composition' of singularities

It is important to be able to state theorems on the composition of Dirich-
let series for an arbitrary 11, without reference to the order of the func-
tions (except at the singularities) and, afortiori, without having to assume
the functions to be bounded (of course, except at the singularities). One
can make such statements by adding to the sets S! + S([! and S([! other
sets that are included in the preceding ones when I1=M=N. Before we
come to such a statement, we have to make some new definitions and
some preliminary remarks.
Suppose that f(s)=(A, 11) is uniform in PUI' We denote by EtCa, al)
the set of points in PUI wheref(s)=a. Let rp(s) be uniform in PU2 '
We define a set P=P[J,al;rp,a2] in the following manner: yEP if
and only if to every value of a except at most one and to every e> 0 there
corresponds a Pa=Pa(e)EEtCa,al) and a f3=f3(e)ESU2'CP such that
11'- Pa -f31 < e. If 11 =M =N (that is, if the two series are Taylor-D series),
then P[J, al; rp, a2] is a subset of SUI,! + S(12,([!:
P[J, a1; rp, a2] c SUl.f + SU2,([!'
In fact, let f3n E SU2.CP' and POOn E EtCWn, a1) be such that lim Wn = 00,
lim (POOn +f3n) = 1', which is possible with the given definition of 1'. But
76 DIRICHLET SERIES
since in our case (A = M = N) the functions f and cp are periodic of period

21Ti, by writing PWn = P~n +iP~n we may assume that
and that there exist two integers kl and k2 such that by setting
fJn=fJ~+ifJ~ we have 2kl1TfJ~fJ~~2k21T, where all the equalities and
inequalities are valid for every n ~ 1.
On the other hand, let a'>aa". If Iwnl >SuPO<1' If(s) I for n>no, then
all the corresponding PWn (n > no) are situated in a rectangle Lll defined by
al ~ a ~ a', 0 ~ t ~ 21T; and consequently the fJn (n > no) are likewise situated
in a rectangle Ll 2. Hence there exists a sequence {nj}, apE Lll and a fJ E Ll2
such that limpwnl=p, limfJn,=fJ, y=p+fJ. But then PES<12," fJ E S<12,CP'
which proves our proposition.
It is equally easy to see by Picard's theorem that if al E S<11,' and if al
is an essential singularity of J, then al +fJ E P[J, al; cp, a2] for every fJ
belonging to S<12,CP' In any case, if f and cp are represented by Taylor-D
series and if we denote by S"cp the set S,+Scp, then S"cp U P[J, al; cp, a2]
is the same as S"cp, As to general Dirichlet series, we cannot even claim
that the closure of this union is equal to the closure S"cp of S"cp, To have
composition theorems concerning general Dirichlet series without having
to introduce the order of the functions (except at the singularities), the
introduction of the sets P[J, al; cp, a2] seems natural and indispensable.
If al =a2= -00, the set P[J, al; cp, a2] is denoted by P[f; cpl. Note that
the two sets P[f; cp] and P[cp; f] are distinct, in general.
To grasp the meaning of a set P[f; cp] it is helpful to say that a point y
belonging to it is in the case of a general Dirichlet series 'a fictitious sum'
a + fJ, where a E S" fJ E Scp,
We are going to prove the following theorem.
THEOREM V.3.1. Let F(s) be a function represented by (A, A) (Al > 1),'
suppose that the function cp(s) represented by a series (D, A), where
D={dn} (dn=FO, n~ 1) is uniform and that the sequence {and;l} admits a
generating sequence G= {gn} of integral order k with respect to A. If
lim sup log Ignl < 00

n
and if the function f(s) represented by (G, N) is uniform, then F itself is

uniform and the only possible singular points of F are the points of
S"",=S,+S"" the points of P[cp;!], and the points of S",.
Suppose, first of all, that cp takes all possible values except one. By writ-
ing, as in the proof of Theorem V.2.1, F(z) in a form similar to the integral
in (V.3), where an is replaced by gn and where bn is replaced by dm we see
that Ua.F~Ua"+ua.,,, (ua,,=lim sup log Ignl/n). What we have to show
is that F(s) is holomorphic and uniform in the part of the complement
of the set
S,,'" u S'" u P[cp;f]

with respect to the plane that forms a domain containing a half-plane
Pq ,. In the plane of the variable z=x+iy, let L be a simple Jordan arc,
L=L c A, having one of its extremities z* in the half-plane x>ua"+u a,,,,,
and let Zo E L.
There exist a 8> 0 and two distinct numbers a and b, a = a(zo), b = b(zo) ,
such that for every f1ES", the following relations hold: (1) Izo-f11>8;
for every Pa E E", (a, -00), every Pb E E", (b, -00) and every a E Sf
(2) Izo-a-Pal > 8, (3) IzO-a-Pbl > 8, (4) IZo-f1-al > 8.

First of all, it is evident that there exists a 81 > 0 such that (1) and (4) are
satisfied for the specified a and f1 with 8 = 81 ; for otherwise Zo would be-
long to either S'" or to Sf,rp. On the other hand, if there are no quantities
a and b and a 82 > 0 such that (2) and (3) hold with 8 = 82 , then for every
8>0 we would have the inequality Iz-a-pcl ~8 for every c, with a cer-
tain
Pc E Eic, -00)
and a certain a, except possibly for at most one value c=c(8). But clearly,
c(8) (if this value existed) would have for sufficiently small 8 the same
value, and Zo would therefore belong to P[cp,f]. The smaller of the two
quantities 81 and 82 would be the 8 in (1), (2), (3) and (4). It follows
from these four inequalities that to every Zo E L there corresponds an
e> 0, e = e(zo), such that for z E C(zo, e)
Iz - f11 > e, Iz - a - Pal> e,
Iz - a - Pbl > e, Iz - f1 - al > e.
The quantity e=e(zo) depends on Zo, but if we denote by e'(zo) the greatest
78 DIRICHLET SERIES
lower bound of all the quantities e(zo) (for fixed Zo in L) having the requi-
site properties, we see that
inf e'(zo) = "I > 0
zoeL
(L being compact). In other words, there exists an "I> 0 independent of

Zo such that for a=a(zo) and b=b(zo) every Pa E E", (a(zo), -(0), every
Pb EErp(b(zo), - (0), every 0: ESf and every fJ ES", the following inequali-
ties hold:
(V.l4) Iz - fJl > "I, Iz - 0: - Pal > "I,
Iz - 0: - Pbl > "I, Iz - fJ - al > "I,
as long as z E C(zo, "I) with Zo E L.
We choose finitely many points Zb1 ), Zb2 ), ••• , Zbk ) such that
Zbm+ 1) E C(Zbm), "112) (1~m~k-1), z~)EL (1~j~k), Zb1)=Z*, the other
extremity zt of L being such that zt E C(Zbk >, "112).
Now lim cp(s) = 0 uniformly in t as U -+ 00 (because AI> 0, here we have
Al>l). But it is also known that cp(s):fO for sufficiently large u; this can
be seen by writing
cp(s) = L: dn e- Ans = e- IlIS (1 + 2: d
n~2
n e-(A n -AI)S),
where the last sum tends to zero uniformly as u -+ 00. Consequently,

there exists a Uo such that
ffi(Pa(zb» < Uo, ffi(Pb(Zh» < uo;
and if Ul +e is negative of sufficiently large absolute value and
s=ul+e+it, then for every ZEC(Z~),7J) (l~j~k), every fJES"" and
every a=a(z~», b=b(z~» (1 ~j~k) we have:
(V.IS) Iz - s - Pal > "I, Iz - s - Pbl > "I, Iz - s - fJl > "I.
Now let Ct(e) be the same contour as that defined in (V.4). It follows
from (V.l4) and (V.lS) that if e>O is sufficiently small, the inequalities
(V.lS) hold for every point s of Ct(e) (if Ul is chosen negative and of
sufficiently large absolute value) and also
ffi(z* - s) ~ ffi(z*) - Ua,f - 2e > Ua,'" +e
(for sufficiently small e). If s is a fixed point of Ct(e) and if z varies on
L, the point u=z-s varies on a curve L(s) obtained from L by a
translation. One of the extremities u* = z* - s of L(s) is in the half-plane

a>aa,lP+e; hence we have
Icp(u*) I = Icp(z* - s)1 :::; 2:ld ln e-?'n(G.,Q! +8) = M
for every s E ct(e).

Now by Schottky's theorem we know that if CP(z) is holomorphic on
C(a, R), if ICP(a) I<N, if CP(z) does not assume on this closed disk two
distinct values a and h, and if 0 < y < 1, then on C(a, y R) we have:
ICP(z) I < K(a, h, N, y) < 00.
Consequently, by setting for fixed s:

ui = Zb2 ) - S, ... , u: = Zbk ) - S,
we can write for u E C(u*, TJ/2)

Icp(u) I < K(a(zb1 », h(Zbl », M, ·n = KI ·
that is, also Icp(u~)1 < K I . By applying Schottky's theorem again we have
for u E C(u~, TJ/2):
Icp(u) I < K(a(zb2», h(Zb2», Kl> t) = K2,
that is, also Icp(u:) I:::; K 2 , etc.

Hence, by applying Schottky's theorem k times we see that Icp(z-s)1 <B
when Z varies on a channel whose central line is L(s), where s is an arbi-
trary point of ct(e) and B is independent of z and s.
It is now easy to see that the integral (V.4) again defines a holomorphic
function on L. The rest of the proof is a simple matter.
In the preceding we have assumed that !pes) takes all the possible values
except at most one. If !pes) does not take two distinct values, the proof
becomes easier, because if al and hI are the two values which !pes) does
not take, then the last passage of the proof where Schottky's theorem is
applied can be used directly with a(z~»=al> h(z~»=hl (1 :::;j:::;k).
V.4. Composition of functions of rapid growth

In this section the permitted growth of the analytic continuation of a
series (A, A) (except at the singularities) exceeds that of every power of
Itl. This growth can be expressed by the inequality
log I/(s) I < C(lti),
so DIRICHLET SERIES
where C(t)(t> 0) increases to infinity, and the only condition imposed on

C (as to its growth) is the convergence of the integral of C(t)/t 2 over
[a, oo[ (a>O). The convergence of such an integral plays a very charac-
teristic role in the study of entire functions that are extensions to the
complex plane of the Fourier transform of functions of compact support.
The study of such entire functions is indispensable for what follows.
True, the next lemma is much more precise than the proof of the
theorems on the composition of functions (and their singularities) de-
mands; a simpler and less precise form would be sufficient by and large,
but the form in which we give it here will be essential for the arguments
of Chapter IX.
V.4.l. Let C(x) be a non-decreasing function that vanishes in a neighbour-

hood of the origin. Suppose that
J co
1
C(x) dx <
x2
00
,
and set for a> 0:
h(a) = (CO C(;) dx.

Ja x
Let p and q be two constants such that p>2, 0~q<p-2.
For every a>O there exists an even entire function Fiz) satisfying the
following conditions:
(V.16) iFa(z) I ~ epeh(a)lYI-C(lzI)La(x),
where La(x) is a positive and even function decreasing for x;;:: 0; La E L,
IILal1 = I,· the Fourier transform Fa(u) of Fa is real, non-negative, even, and
satisfies the conditions:
e-C(a-O)
v- for lui ~ qeh(a)
A
(V.17) Fa(u) =
2e (p+ q)/(p - q - 2)
(V.1S) FaCu) = 0 for lui;;:: peh(a).
We recall that the Fourier transform of F (if FE L) is given by
F(u) = I F(x) e- ixu dx
(we do not put a factor before the integral; when the lower and the upper
limits of integration are -00 and +00, respectively, we do not write

them down). Finally, we recall that Ilfll (for f E L) denotes the norm of
fin L, that is, the integral of If I over] - 00,00[.
The proof of Lemma V.4.l is based essentially on the following lemma.
V.4.2. Let {fLn} be a sequence of positive numbers such that
(V.19) L. fLn < 00.

Then the product
(v. 20) F(z) = n sinfLnfLn

Z
Z
converges uniformly on every compact set of the plane of z=x+iy,· it

represents an en tire even function ,. and if we set:
(V.2l) fL = L. fLn
S(u) = inf (fLlfL2 . .. fLnun)-l
n;.O
(V.22) IF(z) I ::;; eIlIY1S(lzj).
If F is the Fourier transform of the restriction of F to R, then

(V.23) F(u) =
where
.
fn(u) =
{2n+l fL,fL2··'fLn+l
1 for lui < fLn+l
o for lui ~ fLn+l.

With the determination limz .... o log sin z/z=O we have:
sin z Z2
log -z- = -6' + O(Z4) (z --+ 0).
Consequently, as z varies on a compact set D,

'"
L.. 1o g
sin-fLn-Z
fLn Z
converges uniformly6 by virtue of (V.19).
82 DIRICHLET SERIES
Since Isin zi ~ ely I, Isin zl ~ Izl e lyl , we write for 1 ~ k ~ n

Isin ,ukzl ~ ellklyl,
and for k>n (n~O):
Isin ,ul;;zl ~ ,uklzl ellklyl.
So we obtain for every integer n > 0:
IF(z) I = III sin ,ukzl

,ukZ
~ 1
,ul,u2 ... ,unlzl
n eilIYI ,
from which (V.22) follows.

As to (V.23), we remark that when Pn(u) denotes the expression follow-
ing 'lim' in this equality, then we have the convolution
Pn = Pn-l *In> (n > 1), Pl = J;.,

where In is, incidentally, the Fourier transform of (sin fLnX/fLnX)(27T)-l.
Thus, 27TPn is the Fourier transform of
(V.24) n sin fLkX.

k';;n fLl;;X
Since the product (V.24) tends to a limit F(x), and since this product
is bounded by a function independent of n and belonging to L, the limit
in (V.23) exists and is equal to the Fourier transform of the limit of
(V.24) (as n""'" 00), that is, to the Fourier transform of F(x).
Now we proceed to the proof of Lemma V.4.l.
Let Ca(x) be the function that is zero for O~x<a and equal to the
restriction of C(x) on [a, 00[; we set Qa(t)=E[Ca(et + l )] (we recall that
E[d] denotes the integral part of (d». It follows from (V.I5) that the
integrals 7
('Xl e-tQa(t) dt = (00 dQa(t)

JIOg a-l JIOg a-l-O
converge, where the equality of the two integrals is a result, after inte-
grating by parts, of the fact that there exists a sequence N j t 00 such
that
Qa(Nj ) e-Nj.....,. O.
Now Qa(t) is a step-function whose jumps are positive integers mk
(k~ 1); if qk are the abscissae of these discontinuities, we set vk=e- q"
and let {fLn} be the set of the Vb each counted mk times.
Then:
(V.2S) 2: fLn = (00 e- t dQa(t) = (00 Qa(t) e- t dt

JIOg a-l-O JIOg a-I
~ eJ oo
a
C(X)
-2-
X
dx = eh(a).
We now construct the function F(z) of Lemma V.4.2 with the fLn we
have just defined. We set {1=(p-q-2)eh(a)J2, y=(p+q)eh(a)J2 and
F(z) = e- C(a-o'._l_.sin{1z.sinyz F(z).

a 21Te ffy z z
Then:
sin yzsin{1zl --- (y+Jl)lyl •
\ yz {1Z "" e mm, {1y 1Z 12
(1 _1_)
& 2e(YHlIYI 1 ,
'<: 1 + {1yx2
which enables us to write (for the chosen values of {1 and y)
(V.26) lFaCz) 1 ~ !e(HYlIYI-c(a-OlIF(z)ILaCz)

e
= ~ e(p -lleh(allyl-C(a -OlIF(z)jLaCx),
where IILal1 = 1.
For a given choice of the fLn we can write (V.20) in the form
log F(z) = i oo
log a-I-O
{log (sin e-tz) +t - log z} dQaCt)
(with the evident determination of the log under the integral sign).
Since S(lzD~(fLb fL2'" fLnlzln)-1 no matter what n is, we can write
for y>O and any B>O:
log IF(z) I ~ y i:a-I-o e-tdQa(t) + i:ga-I-O tdQit)

- log Izl i:ga-l-O dQa(t).
84 DIRICHLET SERIES
By setting B=log Izl and integrating the second integral by parts, we

find:
fro flOg 121
log W(z) I :s;; Y Jloga-l-o e-tdQa(t) - JIOga-l Qa(t)dt
= Y 1 ro
loga-l
Qa(t) e- t dt -
110g 121
log a-I
Qa(t) dt
:s;; ey f a
ro C(u)
-2
U
du -
110g 121
loga-l
E[C(eI+ 1)] dt.
Izl ~ a, we can write

110
And if
log W(z)1 :s;; ey f ro C(u)

-2 du -
8' 121
E[C(et + 1)] dt
a U loglzl-l
:s;; ey fro C(~) du - C(jzi) +1

Ja U
= eh(a)y - C(lzi) + 1.
This inequality combined with (V.26) yields (V.16). For Izl <a (V.16)
follows immediately from (V.26) and from (V.22), written down for the
function F that occurs in our proof (in which we can replace S(lzl) by
1; fI. does not exceed eh(a».
We now define Ay and All as follows:
Ay(u) = I for lul:s;; y, L1iu) = 0 for lui > y
I
L1p(u) = 2fJ for lul:s;; fJ, AIl(u) = 0 for lui > fJ.
Then we have (with our choice of fJ and y):
Fa(u) = v'f!y e- C(a-o)L1 y(u) * L1p(u) * F(u).
But it follows from (V.25) that for fI.= 2. fI.'I> y=(p+q)eh(a)/2,

fJ=(p-q-2)eh(a)/2 we have: y-fJ-fl.~q; and by (V.25) and the defini-
tion of the In :
Ay(u) * L11l(u) * F(u) = L1y(u) for lul:s;; y - fJ - fl..
So we have, in particular, (V.I7). Now (V.IS) follows from (V.16) and
from the Paley-Wiener theorem. 8 Thus, Lemma V.4.1 is proved.
Now we are in a position to state the following theorem.
THEOREM V.4.3. Let M = {P-n} be a sequence of positive step h, and A a

sequence such that
(V.27) IAn - P-nl ~ g < h12.
Let C1(u) and C2(u) be two increasing functions on [0, oo[ with
(V.28) fCO C1(u) du < 00 fCO C2 (u) du < 00

J1 u2 , J1 u2 ,
C2(u) being concave.

Suppose that f(s)=(A, A) is uniform and does not increase in P"l' ex-
cept at the singularities,faster than eCl(ltJ), and that cp(s) = (B, M) is uniform
and does not increase in P"2' except at the singularities, faster than eC2<1t1l.
We set C={cn}, where cn=anb n (n~ 1) and let F(s)=(C, M). Then:
aa.F~aa.f+aa.<p, For a>a2+H(a1) the function F(s) is uniform, its only
possible singularities in this half-plane are the points of S"l.1 +SIJ2.<P,· and
in this half-plane F(s) does not grow, except at the singularities, faster than
eC2(ltll.
The proof of this theorem does not differ materially from that of
Theorem V.2.I. Taking c>aa.f we consider this time the integral
(V.29) I/>(z) = f C
+
ico
f(s)cp(z - s)H( - is) ds,
JC-ICO
where H(z) is an entire function of exponential type which decreases
like L(x) e- Cl (X)-C 2(x), L(x) E L, in every strip Iyl < d. In the proof of
Theorem V.2.1 we considered a similar integral, where instead of H( - is)
we had the function S-k-l (k>v+p-, v and p- being the orders offand of
cp, respectively). It is the strong decrease of H( - is) in every strip lal < d
(as Itl-->- 00) which permits us to compensate the strong increase of f
and of cp in such a strip, except at the singularities; and the fact that H
is an entire function makes it unnecessary to add s=o to the singulari-
ties of f on integration, in other words, there is no need to introduce
SIJ2,<P as subset of the set of possible singular points of F (as this is the
case in Theorem V.2.1). Another important advantage compared with
Theorem V.2.1 is the very simple expression cn=anbn for the coefficients
of the compositum.
On the other hand, Theorem V.2.1 has an equally important advantage
over Theorem V.4.3 in the fact that no assumption is made on the se-
quences A and M; whereas in Theorem V.4.3 we assume that the step
of M is positive and that (V.27) holds.
86 DIRICHLET SERIES
Since the change of finitely many coefficients and exponents in (A, A)

and in (B, M) does not alter anything in the conditions nor in the con-
clusion of the theorem, we may assume without loss of generality that
(V. 30) "n+l - "n > hi = h - S > 0 (n;a: I),
where S is sufficiently small so that also
(V.31) I"n - ftnl ~ g < h1 /2.

We are now going to define the function H in (V.29). According to the
conditions (V.28) and Lemma V.4.1 we may choose the constants
p, q, a so that by setting
C(x) = Cl(x) + C2(x),

there exists an even entire function Fa(z) satisfying (V.16), (V.17) and
(V.18) with
peh(a) ~ hl /2
qeh(a) ;a: g;
that is,
IF(z) I ~ e(h1/2l!Y!-C(!XIlL(x), (L(x) E L),
L(x) being an even function, decreasing on [0,00],

(V. 32) Fa(u) = P for lui ~ g,
(V.33) Fa(u) = 0 for lul;a: h1 /2,
and P a positive constant.
We now set H(z)=Fa(z) and consider the integral (V.29) with c>(Ja",
the integration being taken over the line (J = c. By imitating the argument
in the proof of Theorem V.2.1 we see that (V.29) converges uniformly in
every compact set of the half-plane x> (Ja,(/! +c and that
cfJ(z) = 2,b n e-#n z f C
c-ieo
+ieo
2, am e{IJn-AmlSH(-is) ds.
We have also
and, given the properties of H, by applying Cauchy's theorem we have
Therefore we can write:
But by (V.31), lPn-AnI <g; hence, by (V.32):
H(Pn - An) = P > O.
And by (V.30) and (V.31), when m#n, we have lAm-AnI ~hl/2, that is,
H(Am - An) = 0 (see (V.33».
Thus:
and so
F(z) = - iPCP(z).
Evidently,
(Ja.F = lim sup log Icnl/Pn

~ lim sup log lanl/An + lim sup log Ibnl/Pn
The rest of the proof goes as for Theorem V.2.1. We construct the same
division; we define 'l\(e) in the same way, with one exception: we do
not attribute any special role to the point 8=0: it only occurs if belong-
ing to S"l.f like all the other points of this set.
c
Observe that since 2(u) is an increasing concave function for u>O,
we have:
When we now define ct(e) and A as in the proof of Theorem V.2.l,

except that in the definition of A the inequality
88 DIRICHLET SERIES
is now replaced by
x > H,(Ul) + U2 + 26
and the relation
then the arguments used in the proof of V.2.l lead us first for s E Ct(e)
and Z ELI to the inequality:
Ip(z - s)1 ~ N eC 2<Jt-YI);

that is, by (V .35) to:
and finally, for z ELI, where (/) is holomorphic and uniform, to the in-
equality
which completes the proof of the theorem.

We conclude this section with a theorem that can be obtained from
Theorem V.4.3 in exactly the same manner as Theorem V.2.8 is obtained
from Theorem V.2.1.
THEOREM V.4.4. Suppose that for A
where k is a constant, and that f(s) = (A, A). Suppose also that there exists
a point cc = Ua,' +ito having the following properties:
(1) f is holomorphic in the set of disks Cq = C(cc + 2qrri, S), where q is
an arbitrary integer and S a constant.
(2) There exists a positive increasing sequence {Pm} such that in C ± m
(m ~ 0): If(s)1 ~Pm and that
'" log Pm <

L., - - 2 - 00.
m
The function F(s)=(A, N) (a Taylor-D series with Ua,F=U a,,) then has
an analytic continuation into the Cq • In particular, cc is a regular point for F.
To obtain this theorem it is sufficient to set in V.4.3

1 _
cp(s) = eB _ 1= L: ens,
where f is the function in V.4.4.
V.S. Composition of 'Hurwitz type'

First of all, we recall Hurwitz' original theorem on the composition of
Taylor series. 9 We do this just as incompletely as in the statement of
Hadamard's theorem in V.l.
If El is the set of singularities of the function represented by
L: anz-<n+l), E2 the set of singularities of the function represented by
L bnz-<n+l) and E3 the set of singularities of the function represented
by L cnz-<n+l), where cn=anbo+ G)an- 1b1 + ... +aobn, then to every
point e3 of E3 there corresponds a point el of El and a point e2 of E2 such
that e3 = el +e2'
If z=O is a regular point for the function ifs(z) represented by L: anz-<n+l),
we say that the function f(s) represented by Lan e-<n+l)s is regular at
-00. We say that -00 is a singularity for f(s) if z=O is a singularity of
ifs(z).
By setting z=eS Hurwitz' theorem can be stated in the following way:
If f(s) = L: an e-<n+l)S, cp(s) = L: bn e-<n+l)S and F(s) = L: Cn r<n+l)S with
cn=anbo+ (~)an-lbl + ... +aobno then F(s) has no singularities other
than the points 'Y of the form 'Y = log (e« +eP), where a: is a point of Sf
and f1 a point of S(IJ' Evidently, if a:= -00 is a singularity of J, this point
must also be regarded as belonging to Sf> and every point log (e - <Xl +eP) = f1
belonging to S(IJ must be regarded as a possible singularity of F. Thus,
if f1 = - 00 is a singularity of cp, then every a: of Sf is a possible singularity
of F.
Let O(s) = L dn e-h"s be uniform. We say that s= -00 is a regular point
of 0 if the following three conditions hold:
(1) O(s) is holomorphic in a half-plane (1< (10,
(2) O(s) tends to a finite limit, which we denote by 0(-00), uniformly
in t as (1 tends to - 00.
(3) [O(s)-O(-oo)]e- S tends to a finite limit, which we denote by
0'( - (0), uniformly in t as (1 tends to - 00. If one of these three conditions
does not hold, we say that s = - 00 is a singularity for 0, that is: - 00 E So.
90 DIRICHLET SERIES
We obtain a simple generalization of Hurwitz' theorem if one of the

two series is a Taylor-D series, the other being of arbitrary type.
THEOREM V.S.l. Suppose that the two series

f(s) = Ia n e-<n+llS, cp(s) = Ib n e- An +1 S
are uniform in the whole plane. Let {In} (n ~ I) be the sequence of all the
sums Aq +m (where q ~ 1 and m ~ 0 are integers), these sums being written
in increasing order,lO and set:
where
A(q, m) = Aq+l(\+1 + 1) ... (\+1 + m - l)jm! (q ~ 0, m ~ 1),
A(q,O) = 1,
the double sum being taken over all q~O, m~O,for which Aq+1 +m=ln+l'
We set
Then
F(s) is uniform, and the only possible singularities of F at finite distance

are the points 'Y of the form 'Y = log (e a+en) (with all determinations of
log), where IX and f3 are arbitrary points of Sf and Sq>, respectively, and the
limit points of this setP
If - 00 is a regular point off, then there exists a 0> 0 such that lea + eli I ~ 0
for every IX E Sf and every f3 E Sq>, and if F(s) is holomorphic in a half-plane
a < ao, then F(s) tends to a finite limit as a tends to - 00, uniformly in t
(s=a+it).
Let c>aa,f and let us show that for sufficiently large x:
(V. 36) 1.
F(z) = -2
. Trl
f
C 2111
C
+ f(s)cp[log (e Z - eS)] eSds,
where for a = c and for sufficiently large x
(V. 37) cp[log (e Z - eS)] = I bn(eZ - eS)-An+l

= Ibn e-An+1Z[1 - e<S-Zl]-A n +1
= Ibn e- lIn +1 ZL A(n, m) em<s-Z),
the integration being taken over the straight line segment joining the
points c and C+21Ti. If log Je 2 -eSJ~c>ua,rp, then the integral in (V.36)
can be put in the form
JcfC+ f(s) L: bn e-An+l z L: A(n, m) em(S-Z) eSds

2lti
(V.38)
L: L: bnA(n, m) e-(An+ +m)z Jcf +

C 2lti
= 1 f(s) e(m+1)s ds,
and since the last integral is nothing but 21Tiam, we see that (V.36) holds.
The equalities we have established remain valid when the an and bn
are replaced by their absolute values; consequently F converges abso-
lutely for x>log(ec +eO"a,4», and since c is arbitrary as long as C>Ua,!,
we see that
Ua,F ~ log (eO"a" + eO"a,,,,).

Let U1 < ua,!; we denote by E(U1) the set of singularities of the analytic
continuation of L: anz- n- 1 in JzJ >eO"l; for e>O, let E(ul> e) be the union
of finitely many disks C(z, e) containing E(U1) with z E E(U1)' We write
So-l.tCe)=log E(ul> e); this is therefore the image under log of E(U1' e)
(s E So-l,,(e) if es E E(U1' e».
Let D(ul> e) be the largest domain composed of points of the rectangle
U1~ U~ c, 0 ~ t ~ 21T that do not belong to SO"1.,(e) and whose boundary
contains the points of the interval u=c, O~t~21T (this is possible if
e>O is sufficiently small, which we assume). We write
cp[log (e Z - eS)] = t/J(z, s).
If z is fixed with x sufficiently large, then
4>(z, s) = f(s)t/J(z, s) eS
is a holomorphic fl,lnction of s on D(ul> e), where t/J(z, s) is the analytic

continuation along curves (on which only s varies) situated in D(ul> e),
of the function given on u=c by (V.37). The function 4> (for fixed z) takes
the same value at the points u' and u' +21Ti belonging to the boundary
B(ul> e) of D(ul> e).
If we denote by B*=B*(U1' e) the boundary of B(U1' e) composed of
92 DIRICHLET SERIES
the points that do not belong to O'=c nor to the lines 1=0 and 1=21T
(on which tP, for given z, takes the same value by virtue of the manner in
which SCT1Ae) and D(O'I. e) have been constructed), then by Cauchy's
theorem we have
JrB* !(s)rp(z, s) e
(V.39) 1.
F(z) = -2 S ds.
1T1
Let A be a compact domain containing points of the half-plane

x> log (e C +aCT.,Q», such that for every Z E A, every s E B*, and every fJ E Srp,
except for fJ= -00 in case -00 E S, we have:
(V.40) Iz - log (e S + en)1 > e1

(V.4I) Iz - s + 2k1Til > e2 (k any integer),
where e1 and e2 are two given positive quantities, and all the determina-
tions of the logarithms are taken.
Hence there exists a 81 > 0, an e> 0, and an N> - 00 (8 1 = 81 (e1),
e=e(e1), N=N(e2», such that for Z EA, s E B*, fJ E Srp (fJl: -00) we have:
le z - eS - eli I > 01
Ilog (e - es) - fJl > e

~ log (e 2 - e = log le
S) 2 - esl > N.
The integrand in (V.39) is a holomorphic function of z and of s pro-

vided that z E A, s E B*; therefore F is a holomorphic function in A.
Since e1 and e2 can be chosen arbitrarily small and 0'1 negative, of arbi-
trarily large absolute value, we see that F is holomorphic in every com-
pact domain containing points of a half-plane 0' > log (e C +eO'.'4'), but not
containing any point of the form log (e a +e8). Hence if s= -00 is a singu-
larity for cp, the theorem is proved.
Suppose now that s= -00 is a regular point for cp, If we define A
as above, but omitting the condition (V.4I), we can show that F is holo-
morphic in A.
In fact, rp(z, s) is holomorphic as a function of z in A if for Z EA, z=s,
we give to rp the value of cp( -00); for rp is then a continuous function of z
for fixed s, and this function is differentiable with respect to z as long as
z EA. This is evident for Z=Zo EA, with zol:s (mod 21Ti).
And if Zo =s (mod 27Ti), then:
(V.42) a~~;s)lz=zo = !~~o ~(z,s~:= ~~-oo)

' ~(z,s) - <p(-oo) I' e Z - eZo '( )
= I1m Z z '1m = <p - 00 ez o.
Z-Zo e - e0 Z-Zo z - Zo
From the fact that ~ is holomorphic in Ll for every s E B* it therefore

follows that F is holomorphic in Ll.
Suppose now that - 00 is a regular point off
Observe that the set (S"lof), which is the subset of S"l.f contained in the
strip O~t~27T, contains the points of the interval L(al) given by a=ab
o~ t ~ 27T. Let 2: (ab e) be the union of the images under log of the disks
C(z, s) forming part of E(ab e) and containing the points of L(al)' Since
- 00 is a regular point off, we see that it suffices for - al to be sufficiently
large (e> 0 fixed) for the following facts to hold: on the one hand,
2: (ab e) is separated from the images (under log) of the other disks of
E(aI, e), and on the other hand,! is holomorphic on 2: (ah e). Given that
e > 0 is chosen sufficiently small and that such a choice of al has been
made, let 1= I(aI, e) be the boundary of 2: (ab e) and C* = C*(al> e) =
=B*(ab e)-/(al' e), By (V.39) we can then write:
(V.43) 1 , [ f(s)~(z, s) eS ds
F(z) = -2 1 . [ f(s)~(z, s) e
+ -2 ds.
Jl 7Tl Jc.
S
7Tl
We see that the last conditions of the theorem are satisfied and that
the first integral in (V.43) vanishes for sufficiently large x, because then
the function to be integrated over 1 is holomorphic as a function of s
on 2: (al> e) of which 1 is the boundary. Consequently we have
(V.44) 1 ,[ f(s)~(z, s) eS ds.

F(z) = -2
7Tl Jc.
Now this expression tends to
1 . [ f(S)<p(7Ti
-2
7Tl Jc.
+ s) e
S ds
uniformly in y, as x tends to - 00.

This completes the proof of the theorem.
94 DIRICHLET SERIES
It is easy to verify that Theorem V.S.! contains as a particular case

Hurwitz' theorem in its most complete form (at least for uniform func-
tions): An=n (n~I). It is enough to note that the assertion that the
analytical continuation of §(z) = 2: dnz- n- 1 is holomorphic for 0 < Izl <TJ
and tends to a finite limit as z tends to zero is equivalent to the fact that
this function is holomorphic in Izl < TJ.
A theorem of a more general character concerning two Dirichlet
series (A, A) and (B, M) of general type, the exponents of the series
being distinct a priori, can be obtained by introducing the notion of the
order of each function (except at the singularities).
THEOREM V.5.2. Let f(s) = (A, A) be uniform of order J) in PU1 ' except at the
singularities; and let cp(s)=(B, M) be uniform of order p. in PU2 ' except at
the singularities. Let k be an integer greater than v+p., and let d/f)(n) be
the n-th Taylor coefficient off of order k. Let L = {In} (n ~ 1) be the sequence
of all the sums Aq +m (q ~ 1, m ~ 1), where the m are positive integers and
the sums are written in increasing order,12 and set
dn = 2: 2: a~k) (m)bqA(q - 1, m),
the A(q, m) being defined as in V.SJ and the double sum taken over all
q~2, m~ 1,jor which "\-1 +m=ln'
We write
max (aa", 0) = a,
max (H,(a1), 0) = at.
Let 13 F(s)=(D, L) (D={dn}). Then
(V.4S) aa" ~ log (eUf + eUa,rp),
F is uniform in the half-plane a>log (e U;+eU2 ), and in this half-plane the
only possible singularities of F are the points of the form
y = log (ect + eP), y = log (eli + 1)
where ex E SU1,' and f3 E SU2,rp, and the limit points of this set of points (with
all determinations of log).
Let c> max (0, aa,,) and let us show that
(V.46) F(z) =
k'"
-2
TTl
1
c + 100
c-loo S
ds
f(s)cp[log (e Z - eS)] k+1'
where cp[log (e Z- eS») = !fo(z, s) is given by (V.37) for sufficiently large x.

In fact, the expression on the right-hand side of (V.46) can be written

(see (I.3.2» for sufficiently large x as: 14
= -;;; b -;;; A(n - 1 m) e-(;\n +m)z - .

k! iC+i CO Ji(s) ems _ ds
L. n L. , 2m c-ico Sk+l
= 2: 2: a~k)(m)bnA(n - 1, m) e-(l\n +m)z = 2: dn e- lnz .

These inequalities are valid for x>log(ec+e"a.",). They are valid for
the same values of x when the an and bn are replaced, respectively, by
lanl and Ibnl. This proves that (D, L) converges absolutely for the same
values of z. But since c is chosen only to satisfy the inequality
c> max (0, aa.f)' we see that (V.45) holds.
Now let c> max (0, ah aa,f)' If we define Ct(e) just as we have done it
in writing (V.4), we see, as in the proof of (V.4), that
(V.47) F(s) = k'i

-2
'. f(s)cp[log (e Z - e
S) ] ds
k+l'
m Cr(e) s
Now let LI be a compact domain containing points with x> log (e C + e" a,O, for every z ELI, every a E Sf and every
fJ E S", we have:
Iz - log (e lt + eO)1 > e
(no matter what determination of the logarithm is taken).

Suppose also that for z ELI
(V.48) x > log (e,,;+2t + e"2).
We see that to every s E Ct(e) there corresponds an a E S"l,f such that
where 7)(e) tends to zero with e,
°
From the various inequalities above it follows, on the one hand, that
there exists a 8> such that for sufficiently small e the following in-
equality holds for every z ELI, s E Ct(e) and every fJ E S"2,"':
Ilog (e Z - eS) - fJl > O.
96 DIRICHLET SERIES
On the other hand, it follows from (V.4S) that for Z E.:1, s E et(e):
(V.49) 8i[log (e Z - eS)] = log le z - esl
;;::: log (eX - eO") ;;::: log (eX - e<1/+2t) > 0'2.
In other words, if Z E.:1, s E et(e) and if 8> 0 is given, then for suffi-
ciently large larg (eZ-eS)1
l<p[log (e Z - eS)] I < N(o) I arg (e2 - eS)IIlH,
where arg (eZ-e S) is well-defined, because we see according to the expan-

sion we have given for (eZ_eS)-An that log (ez-e") is taken real for real
Z and s, and that for the other pairs of z and s this function is defined by
analytic continuation.
But from the fact that .:1 is a compact set it follows that there exists a
constant A < CX) such that for Z E.:1, s E etce) and sufficiently large It I:
larg (e~ - eS)1 < Alarg esl = Altl.
Hence, on the one hand, to every 82 > 0 there corresponds an
N 2 =N2(8 2) such that for sufficiently large It I
l<p[log (e Z - eS)] I ~ N 2 (02)ltI IlH2 ,
and on the other hand, to every 81 > 0 there corresponds an Nl8 1 ) such
that for s E et and sufficiently large It I
If(s) I ~ N 1 (Ol)IW H1 •
Therefore, if 01 and 02 are chosen sufficiently small, then (V.47) con-

verges uniformly in.:1 and defines there a holomorphic function.
Our theorem is now proved.
Theorem V.S.2 can be strengthened in the following manner:
THEOREM V.S.3. If s=O is a regular point of f(s), then under the same
assumptions and the same notation as in Theorem V.S.2 the following
conclusion holds:
F(s) = L dn e- lns
is uniform in the half-plane P:
and the only possible singularities of F(s) - J(s), where
J(s) = f(O) dk 9'[log (e: - e~)] I

dg ~=O
+ mf'(O) dk-l9'[I~:,,~: - e~)] ~=O I

+ ... + j<"l(O)9'[log (e S - 1)],
in the half-plane P are the points y of the form y = log (e rt + eB), where a
and f3 are defined as in Theorem V.S.2, and the limit points of this set (all
the determinations of the logarithm must be taken).
The passage from Theorem V.S.2 to Theorem V.S.3 is analogous to
that from Theorem V.2.l to Theorem V.2.2.
Remark. If the functions f and 9' occurring in the various 'composition

theorems' of this chapter are not assumed to be uniform, then the corre-
sponding statements still remain valid, provided that we denote by S"l.!
and S"2.(fJ the singular sets of f and 9' (in the half-planes P"l and P"2'
respectively) to which we add cuts which make these functions uniform
in the exterior of the sets so obtained. But in the conclusions we can no
longer claim that the composite function F is itself uniform in the corre-
sponding domain.
NOTES
1 A great many authors have assumed erroneously that the statement which follows
is valid in all cases.
2 J(Ult e) is clearly closed and connected and contains a half-plane u>u2'
3 If A and B are two point s~ts, then A + B denotes the set of all the points a + fJ
such that a E A, f3 E B.
4 If An=Pn the expression sin 1TAn/A1T,.[l- A~/p~l must be taken as representing
TI
mr;:.Pn
(1 - p~/m2).
5For n=O, by definition (/Ll .•. /Lnun)-l is equal to 1. Thus, S(u)";; 1.

6On D this sum only begins with the term 110 (II ~ 110) so that sin /LnZ does not vanish.
The determination of log is evident.
7 For a Stieltjes integral f-o denotes the limit of f-e if e tends to zero through
positive values.
6 We recall the Paley-Wiener theorem (at least that part which we need): if F(x)
belongs to L on R, then a necessary and sufficient condition for F to be the restric-
tion of an entire function F(z) satisfying the inequality IF(z)I";;A ea1ul (where A is a
constant) is that the support of p(u) should belong to the interval [-a, al.
98 DIRICHLET SERIES
9 In the following statement, as in Theorem V.5.1 the summations 2: extend over

n';;:' 0 (and not over n ;;:. 1, which is our convention if no limit is stated).
10 In this theorem, as in Theorem V.5.2, the In are arranged in non-decreasing order
rather than in increasing order, because two pairs (q1, m1) and (q2, m2) may yield
""1
the same In = + m1 = .\Q2 +m2. In the series for F the same exponent can therefore
occur several times. But this does not affect the proof at all.
11 If - 00 E Sf, then every r =log (e -00 +eP) =fH 2krri must be regarded as a possible
singularity of F. Also, if - 00 E S"" then every ex +2hri, where ex E Sf, must be reo
garded as a possible singularity of F.
12 See Note 10.
13 See Note 9.
14 We recall that a~k)(m)=O for m";;;p.1.
CHAPTER VI
SOME APPLICATIONS OF THE PRINCIPLES

FOR ANALYTIC CONTINUATION
VI.I. Arithmetic properties of the exponents and the

analytic continuation
In this section we give an account of some applications of the arguments
in ILL Our aim is to show that the exponents of a series f(s)=(A, A)
which are' distinguished' from the other exponents, for example, by their
fractional parts, play a role which one could describe as 'autonomous'
in the determination of the properties of the analytic continuation of the
series. The coefficients of the series corresponding to these exponents
already indicate certain important characteristic features of the continua-
tion, which remain in force, no matter what the values of the other co-
efficients are. Thus, if ..1* c A has the property that every Anj E ..1* has
a fractional part different from that of all the other An E A, n =I nf> then
the order of magnitude of the ani determines a set of domains (obtained
from anyone by all the translations 2k-rri, k being an arbitrary integer),
where f cannot be holomorphic and bounded.
One could compare the following theorems with those of Chapter
VII, which while providing conclusions concerning the analytic continua-
tion starting out from the arithmetical character of a subset of exponents,
are nevertheless of an entirely different nature.
THEOREM VI.l.l. Suppose thatf(s)=(A, A) with O'a,,=O and that ..1* c A

is such that every Ani E ..1* has one of the following properties:
(1) Ani =I An (mod 1) for every n =I n;;
(2) Ani is an integer, and if there is another integer An E A, then one of
the two inequalities Ani> 2An or An> 2A nj holds.
We set 1
(VI.1)
and assume that 0'1> -log 2. Let to be arbitrary real, and

o< e < log 2 + 0'1;
100 DIRICHLET SERIES
we denote by 'S) the rectangle

0'1 - e < 0' < e, It -:tol < 3 ar(co(e"l-e ~ 1T.
The function f(s) cannot be holomorphic and bounded in the union of'S)
with all the rectangles obtained by the translations 'S) +2kTTi (where k is
an arbitrary integer).
We set
F(s) = Lan e- lI n("l-s+ItO) e- lI n S = L Cn e- lInS = (C, A),
and denote by D the domain in the plane z=r el8 which consists of the
union of the disk Izl <e"l with the domain defined by r< 1,
101 < 3 arc cos e"l -e.
One easily verifies that the starred curve r* defined by r=cos 0/3
(101<1T) (see II.1.3) is in the interior of D apart from the point z=l,
and in particular, that the part of r* for which 101;;::3 arc cos e"l-e is
situated in the closed disk Izl ~e"l-e.
If the conclusion of the theorem were not satisfied, that is, if f(s) were
holomorphic and bounded in the union of the rectangles defined in the
statement of the theorem, then there would exist a starred curve G*
sheltering r* such that (C, A) would be holomorphic and bounded in
~D(G*).
Suppose, for example, that the condition (1) is satisfied for all An,
belonging to A*. By applying Theorem II.1.3 we would obtain
len"I = Ian I e- lln /"l -6) ~ B1 sup
s e :l'D(GO)
IF(s) I = P,
where P is a constant, and this would give
. log lan,1
bm sup A ~ 0'1 - e;
nJ
this is a contradiction to (V1.1). The arguments used in the proofs of

Theorems II.1.3 and II.1.4 allow us to draw the same conclusion as long
as we assume that one of the conditions (1) or (2) is satisfied for every
Anj E A*.
Theorem VI.l.1 immediately suggests the following more general
theorem:
THEOREM VI.1.2. Letf(s)=(A, A) with O'a,,=O. Let r*

be a starred curve
defined by r=r(O)(I 01 ~1T), where reO) is a (strictly) decreasing curve on
THE PRINCIPLES FOR ANALYTIC CONTINUATION 101
[0,17], with r(rr)=a (O<a< 1), and let

8(r) (r E [a, 1])
be the inverse function to r(8). Let <1>(z) be the function attached to r*,
and let A * c A be the set of An each of which is distinct modulo 1 from
all the other members of A with respect to <1>.
We set (VLl) and assume that U1 > log a. Let to be arbitrary real and
0<e<u1-log a.
The function f(s) cannot be holomorphic and bounded in the union of the
rectangle '5> defined by
U1 - e +2kl7i (where k is an
arbitrary integer).
To prove this theorem it is sufficient to copy the proof of Theorem
VI.1.2, making use of the new curve r*.
From these two theorems we obtain, in particular, the following
theorem:
THEOREM VI. 1.3. If we add to the conditions of one of Theorems VI.l.l

or VI.1.2 the condition U1 = 0, then for any real to and e> 0 the function
f(s) cannot be holomorphic and bounded in the set of disks
q(to + 2k7r)i, e]
(where k is an arbitrary integer).
So we find, for example, the very special case of the classical theorem
of Hadamard: if the An are integers and An + 11 An> 2, then the Taylor series
L anz An of finite radius of convergence has its circle of convergence as
natural boundary. But evidently we can extend in ILl the very definition
of the notion ' A and A' are distinct modulo I with respect to <1>', replacing
it by 'A and A' are distinct modulo lip (where p, for example, is a positive
integer) with respect to <1>'. This will be the case if for any m and m'
such that
d~A) =1= 0,
(see ILl) we always have: p(A-A')i=m-m'. If we then replace in all the
arguments of ILl <1>l(Z)=Z<1>(z) by <1>p(z)=zP<1>(z), we see easily that the
theorems in this part of Chapter II can be generalized. (To simplify the
exposition we have only treated the particular case p = 1.) The theorems
we have just proved here can therefore also be generalized in a corre-

sponding manner. The very special case of the theorem we have just
quoted then becomes Hadamard's theorem, as it was originally stated: if
An+1/An ~ A > 1 (An are integers),
then the circle of convergence is a natural boundary for the Taylor series
L anzAn •And this theorem only appears as a very special case a fairly
general phenomenon.
We do not dwell here on the fact, because in the next section we re-
discover the theorem of Fabry-P6Iya, which has a bearing on general
Dirichlet series, as a very special case of theorems obtained by starting
out from the arguments of n.2.
VI.2. Analytic continuation of general Dirichlet series

The first result of this section concerns the series (A, A) whose abscissa
of convergence is finite, A being of finite step: by the theorems of n.2
these series therefore have singularities. Then we consider series whose
abscissa of convergence is - 00 and which consequently represent entire
functions.
The following theorem is an immediate consequence of Theorem 11.2.3.
THEOREM VI.2.1. Suppose that A is of positive step h. Let D' be its upper
density. Let f(s) = (A, A) with - 00 0, fJ~O there exists a continuous function A(o:, fJ) with A(o:, 0)=0
such that f(s) has for every real value of to a singular point in the rectangle
U a ,! - A(h, D') ~ U ~ U a ,!, It - tol ~ 1TD·.
For A(0:, fJ) we can take the following function:
A(o:, fJ) = 1TfJ - [3 log (o:fJ) - 8.5]fJ for fJ > 0

A(o:,O) = o.
If D' =0, the expression 'singular point in the rectangle .. .' indicates
that U a ,! +ito is itself a singular point. This means that we have the follow-
ing theorem:
THEOREM VI.2.2. If h>O, D'=O, -oo<ua,t<oo, then every point on the

axis of convergence is a singular point off.
One expresses this fact by saying that the axis of convergence of (A, A)
is a natural boundary for (A, A),
Theorem V1.2,2, which is due to P6lya, generalizes to Dirichlet series
the classical theorem of Fabry concerning Taylor series: the circle of con-
vergence of a Taylor series ~ anz iln (where An are integers) with lim n/ An = 0
is a natural boundary, a theorem which in turn generalizes Hadamard's
theorem of which we have talked in the preceding section,
Evidently we can choose for A(a, (3) «(3 > 0) every function of the form
A(a, (3) = 7T(3 + [A(a) - 2a log (a(3)](3 = 7T(3 + L(a, a, (3),
where a is an arbitrary real number greater than 1 and A(a) is the function
A(a) = a {4 + a + log [(a 2 - 1)/a3 ] + log [(a + l)/(a -l)]Ja}
(see 11,2.3),
We have seen that the expression for L(a, h, D') and consequently also
that for A(h, D') contains the term log (hD,), Since hD':::; 1, the expres-
sion - 2aD' log (hD') is positive, and for fixed h (for example h = 1)
tends to zero (as D' tends to zero), like - 2aD'log D', Therefore the
expressions L(a, h, D') and A(h, D') tend to zero in the same manner as
D' tends to zero (for fixed h), We wish to show that the speed with which
these expressions tend to zero is very characteristic, In fact, for fixed h
we necessarily have
lim inf A(h, D')J( - D' log D')
D'-->O
= lim inf L(a, h, D')/( - D' log D') ~ I,
D'-->O
In other words, in A(a, (3), as it occurs in Theorem V1.2,l, the number 3

cannot be replaced by a number smaller than I, A fortiori it is impossible
to take for A(h, D') (with h fixed) an expression of the form cD', where
c is independent of D',
In fact, we can prove the following theorem:
THEOREM VI.2,3, Let p ~ 2 be an integer and let {f-tn} be a sequence of

positive integers for which
f-tn+1/f-tn > p/(p - 1), lim (2 f-tk)/f-tn+l = 0,
k~n
Let {b n} be a sequence for which

(VI. 2) lim sup Ibnl 1/n = 1.
The series
L: bn e- ps
- ;-<P-l)}'n
converges for u > 0, it represents a holomorphic function f(s), which is also
represented in the same half-plane by a series (A, A) having the following
properties:
(1) the An are integers, h = 1,
(2) ua.f=O,
(3) D'=llp,
(4) f(s) is holomorphic in the half-strip It I~7TD', u> -fLD', where fL is
the positive root of
As p tends to infinity, the quantity fL so defined satisfies the relation
(VI.4) fL = -log D' - log (-log D') + log 2 + 0(1) (D' = lip -+ 0),
Remark, In comparing this theorem with Theorem 11.2.3 we see that for
every choice of A(a, (3) corresponding to the statement of Theorem VI ,2,1,
as D' tends to zero, we have:
A(l, D') ~ - [log D' + log (-log D') - log 2 + o(l)]D' "" - D'log D',
Also, in the last paragraph of II,2,3 the phrase' for which the abscissa
of the centre of a disk belonging to it is smaller than [, , ,]D' + u c ' be-
comes incorrect when the number 3 in [,',] is replaced by a number
smaller than I,
Now we proceed to the proof of the theorem,
(1) and (2) are evident: it is sufficient to argue about the series
and to take (VL2) into account.

To see (3) we remark that when we denote by m(x) (x> 0) the number
of An not exceeding x, we have 2
m(PfLn) = 2: (fLk + 1) = n + 2: fLk,

k~n k~n
hence
- = -n
m(PfLn)
- "L.. fLle (pfLn)-1 '
+ P-1 + le';n-1
pfLn PfLn
That is (evidently lim n/fLn = 0) :
D' ~ lim sup m(PfLn)/(PfLn) = !,

P
We also have for PfLn-1 < Ale~PfLn (n~2)
k ~ m(PfLn\
Ale PfLn
and consequently:
D' 1
= '1m sup -k 1
= -,
Ale P
Now we prove (4), First of all we note that for 0 < D < 1 and fL ~ 0:
cos 1TD > 1- (1T D 2)/2, 1- e -IlD < fLD; we can therefore write
(VI. 5) e21l (1 + e- 21lD - 2 e- IlD cos 1TD)
< e21l [1 + e- 21lD - 2 e- IlD + e- IlD (1TD)2]

< e21l [(l - e- IlD)2 + (1TD)2) < e21l(fL2 + 1T2)D2,
le- vs - e-(V-1)SI = e-(V-1)"1 eS - 11

~ eU(V-1lD'(e 2IlD ' +1- 2 euD ' cos 1TD')1f2
= ell (1-D'l(e 2IlD ' - 2 euD ' cos 1TD' + 1)1/2,
that is, by (VI.S),
le- vs - e-(V-llSI 2 ~ e21l(e-2UD' - 2 e- uD ' cos 1TD' + 1)
< e21l (fL2 + 1T2)D'2,
and if fL is a positive root of (VI.3), we have
le-vs - 2e-(V-llSI < l.

By (VI.2) the function is therefore holomorphic on S,

We set:
I-' = -log D' - log (-log D') + log 2 + c(D'),

and show that c=c(D')=o(l) (D' -+0), In fact, we have:
4
D'2 = e2il (1-'2 + 7T 2)
4e 2c , , 21-' 2 + 7T 2 4 e2c
D'2(log D')2 [-log D -log (-log D) + log 2 + c] 1-'2 ,..., D'2'
which shows that c(D')=o(l), The theorem is now proved,
VI.3. Entire functions represented by Dirichlet series
In this section we assume that aa.! = - 00 and, as above, that D' < 00,
ThenJ(s)=(A, A) is an entire function with
lim 10g"lan l = -00,

n
By the classical theorem of Julia we know that every entire function

other than a polynomial has the following property: there exists a ray
through the origin such that in every angle around this ray the function
takes all values, except possibly one, infinitely often, Such a ray is said
to be a Julia line,
Translated into the language of Taylor-D series, the theorem of Julia
can be stated in the following form: ifJ(s)= L: an e- ns is an entire function
and not a polynomial (in e- S), then there exists a to (in every interval
[2k7T, 2(k + 1)7TD such that for every e > 0 the function J(s) takes in the
strip S(to, e) all values, except possibly one, infinitely often,
Let us introduce the notion of such a line for every entire functionJ(s)
represented by a series (A, A), A line t = to is said to be J*-line Jor J if
for every e > 0 the function J(s) takes in S(to, e) all the values, except
possibly one, infinitely often, IfJis not identically zero (we recall that we
assume throughout '\1> 0), then it has a J-line (a classical Julia-line),
but it can also have J*-lines, It is easily seen that if t = to is a J*-line, then
the half-line a<O, t=O, is a J-line, The functionJ(s)=e- s has a=O, t>O
(and t < 0) as a J-line, but it has no J*-line. Besides, no series (A, A)

in which only finitely many an are different from zero has a J*-line.
Theorems 111.2.6 and 111.2.9 lead to the following theorems.
THEOREM VI.3.1. Suppose that the step of A is positive and thatf(s) = (A, A)
is an entire function whose (R)-order is infinite. Then f has a J*-line in every
strip S(to, l7a) with a> D· and any real to.
THEOREM VIJ.2. Let f(s)={A, A) be an entire function. If there exists a

to and a> D· such that the order of fin S(to, l7a) is infinite then for every
d> aJ has in S(to, 17d) a J*-line.
THEOREM VI.3J. Suppose that the step of A is positive and that f(s) = (A, A)
is an entire function whose (R)-order is p > O. Then f has a J*-line in every
strip S(to,l7a) with a>max (D·, 1/2 p) and any real to.
Note that in VI.3.2 we do not postulate that the step should be positive.
If we add this condition, then Theorem VI.3.2 only becomes a special
case of Theorem VI.3.1.
Theorem V1.3.1 is a special case of Theorem VI.3.3. Let us first prove
Theorem VI.3.2.
Let ~ be the square 10'-0'01 <l7d, It-tol <l7d, where 0'0 is an arbitrary
fixed real number, and let us consider in ~ the family of holomorphic
functions Fe(s)=f(s+c), SE~, c::;;O. This is not a normal family. For
if it were, only two cases could arise:
(1) There exists a sequence {c n} with lim cn = -00 such that for a
b(a<b<d)
on the square ~b defined by 10'-0'01 ::;;l7b, It-tol ::;;l7b, which according

to 11.2 (where we set R=b) would imply thatfis identically zero;
(2) lim Fe(s) = 00 as lim c= -00, uniformly on ~b. This would contra-
dict Theorem 111.2.6.
The family Fe(s) is therefore not normal in ~ and consequently has
there at least one irregular point s' = 0" +it'. And no matter what e > 0
is, Fe(s) is not a normal family in C(s', e). This indicates that f takes in
Set' , e) all values, except possibly one, infinitely often. Hence t = t' is
indeed a J*-line.
Theorem VI.3.3 is proved in a completely similar manner, this time

making use of Theorem 111.2.9.
VI.4. Some applications of the composition theorems

Before stating the theorems of this section we have to prove the classical
theorem of Cramer. We also state the extension of this theorem due to
P6lya (which is called the Cramer-P6lya theorem) as well as an impor-
tant contribution to this theorem by V. Bernstein, although only Cramer's
original theorem will really be used.
THEOREM V1.4.l. Let f(s)=(A, A) be uniform with CTa (=CTa,f) finite, and
let p(z) be an entire function of exponential type 8. Let Sf be the singular
set off and Sf,6 the union of the disks of radius 8 around the points of Sf.
Let LJ be the part of the complement of Sf,6 with respect to the s-plane that
is connected with the half-plane Pq " CT' > CTa+8. Then the function
F(s) = (C, A), where C={cn}, Cn=p(An)an, is holomorphic in LJ.
Let <P(z) be the Laplace transform of p(z):
<P(z) = f' p(u) e- UZ duo
It is known that <P(z) is holomorphic for Izl > 8 and that for every e>O:
(VI.6) 1.
p(z) = -2
7Tl
i c.
eZw<p(w) dw,
where Ce is the circle Izl = 0 + e. This can be seen immediately by taking

account of the fact that
p(z) = L dnz n
n .. O
implies
<P(z) = L n~~~
n .. O Z
for Izl > O.
By writing down (A, A) for the variable s- wand making use of (VI.6)
we see that for sufficiently large CT
(VI. 7) 1.
F(s) = -2
7Tl c.
i f(s - w)<P(w) dw.
From this it follows that for sEA we can extend F(s) on every Jordan
curve in LI (having one of its extremities in the half-plane (j> (j' where
CT' is sufficiently large), provided that e is sufficiently small. This proves
the theorem,
Now let h( 0) (0 ~ 0 ~ 27T) by the type of rp along the ray r eW r> 0; that is,
·
h(O) = I1m log Irp(reW ) I
sup .
r ... <Xl r
The conjugate diagram of rp is defined as the closed domain ~ whose
points z=x+iy satisfy the inequality
x cos 0 - y sin 0 ~ h( 0).
According to P6lya ~ is the smallest convex domain in the exterior of
which (j)(z) is holomorphic. Consequently we have
~ c C(0,0).
Thus, denoting by ~ e (e > 0) the union of the disks C(z, e), where z is an
arbitrary point of ~ and by Le the boundary of ~e, we can also write by
virtue of (VI.7):
F(s) =
7T1 JLe
r
1 . f(s - w)(j)(w) dw.
-2
From this we easily deduce the following theorem.
THEOREM VI.4.2. If J, rp and F are defined as in VIA. I and if SF denotes

the singular set of F and ~ the conjugate diagram of rp, then
SF C Sf + ~.
By using the method of Chapter V, that is, by dividing the plane (or
rather part of the plane) suitably in accordance with the singularities of
(j) and by starting out from the equation (VI.7), which is valid for suffi-
ciently large G, only to replace it later, by virtue of Cauchy's theorem, by
an integral concerning the same function, but extended over a set of con-
tours corresponding to the division, we obtain the following statement,
which is due to V. Bernstein.
THEOREM VI.4.3. If the functions J, rp and F are defined as in VI.4.2,

(j) being the Laplace transform of rp and Scp the singular set of (j), then
SF C Sf + Scpo
Theorems V.U and VI.4.1 allow us to state the following theorem
which concerns complementary series.
THEOREM VI.4A. Let P={Pn} and Q={qn} be two disjoint subsequences of

A (P U Q=A, P (l Q=r/»; let !p(z) be an entire function of exponential

type 8.
A and B being two complex sequences, we denote by !pC A), !pCP) and
A!p(P) the sequences {!p(An)}, {!P(Pn)}, {an!p(Pn)}, and we set
f(s) = (A, P)
ifi(s) = (A!p(P), P)
&(s) = (!p(A), A) - (B, Q).
Suppose that f and & are uniform and bounded, except at the singularities,
in every half-plane P (j"
Then the function rp is also uniform and the singular sets Sf> S", and S{J
are linked in the following manner:
(a) for every 8' > 8 every disk C(s, 8') with s E S", contains a point of
Sf>
(b) S", C Sf+S{J.
In particular, if!p is an entire function of minimum type (8=0), then the
assertion (a) can be replaced by the following:
(c) S", C Sf;
and the conclusions on the singular sets can be combined in this case by:
We remark that iff and & are not assumed to be uniform, the statement
still remains valid, provided that we denote (as we have said at the end
of Chapter V for the composition theorems) by Sf and S{J the singular
set off and & to which we add the cuts which make these functions uniform
at the exterior of the sets so obtained. But evidently we can no longer state
that ifi is a uniform function.
To prove the theorem it suffices to observe that as a consequence of
the fact that P and Q are two disjoint parts of A, if we set
f(s) = (A,P) = (C, A), rp(s) = (D, 11), C = {c n}, D = {dn},
we have cn=a m, dn=am!p(Pm) for An=Pm; and cn=O for An=qk' Theorem
V.U then provides the assertion (b). The assertion (a) follows from
Theorem VI.4.1.
Here is a particular case of VI.4.4, stated for the Taylor series
(I1=N, e-s=z).
THEOREM VI.4.5. Let P={Pn}, Q={qn} be two disjoint ascending sequences

of positive integers. If there exists a series
(VI. 8) 2: cnzqn
of finite radius of convergence having a pole among its singularities on the
circle of convergence, then every series
(VI.9) 2: dnz Pn
of finite radius of convergence has at least two singular points on its circle
of convergence.
If IX is a pole of the series (VI.8) on its circle of convergence, and if R
is the radius of convergence of (VI.9), we set
(B, Q) = 2: IXqnCn e-qnS = 2: bn e- qnS
(A, P) = 2: RPndn e- PnS = 2: an e- PnS .
Now clearly there exists a polynomial <p(z) (hence a function of mini-
mum type) such that the series
2: <p(n)zn
represents the principal part of the pole at 1 of the function (B, Q),
with e- S replaced by z. The function &(s) = (<p(A), A)-(B, Q) therefore
has all the points 2k-rri (where k is an arbitrary integer) as regular points
(here A = N). If (A, P) had on its axis of convergence a = 0 only one singu-
lar point (mod 27Ti), for example, the point itoCmod 2rri), then we would
have ito ¢; Sf + Sf}, which would be a contradiction to Theorem VI.4.4.
This proves Theorem VI.4.5.
We remark that Theorem VI.4.5 can be proved more directly, first of
all by taking account of the fact that if <p(z) is a polynomial, then (A<p(P), P)
(P c N) has no singularities other than those of CA, P), and subsequently
by making use of Hadamard's classical composition theorem for the
Taylor series (or rather, in the notation we have used here, the Taylor-D
series (A, P) and (<p(N), N)-(B, Q), Q c N, Q (I P=~.
Theorem VI.4.5 has been generalized in different ways, but all the
generalizations are concerned with Taylor series. For example, the word
'pole' has been replaced by the words' algebraic-logarithmic singularity'.
We are now going to apply the arguments of (V.5). In fact, we shall
only use the classical theorem of Hurwitz on Taylor series to end up with
a theorem which itself only concerns Taylor series.
THEOREM VI.4.6. Let A = {an} be a sequence oj rational numbers such that

there exists an integer N with the property that all the quantities anNn
(n~ 1) are integers.
Suppose that the Junction F(z) = L: anzn is regular at infinity and holo-
morphicJor
Iz - N2/(N 2 - 1)1 ~ N/(N2 - 1),
if N~2; and holomorphicJor x~t (z=x+iy) if N= 1.
Then:
P(z)
(VI.10) F(z) = (1 _ Z)II
where P(z) is a polynomial and h a positive integer.

We set
(VI.11) J(z) = L: an~n

z
( _1)n-1Nn-l
<p(z) = L: zn .
The singularities off are obtained from those of F by the transforma-

tion N/z. Hence all the singularities of J are in the disk C(N, 1).
The theorem of Hurwitz (see (V.5» applied to the series (VI.l1) per-
mits us to state, given that Nn+1Jnal
where Jnal denotes the n-th difference of the sequence {an} with respect
to a1> and where <> is equal to + 1 or -1, is holomorphic for Izl ~ 1.
Consequently we have
lim sup IYnl 1/n = B < 1.
But since the Yn (by assumption) are integers, we see that Yn = 0 for
n>no·
Hence it follows from the identity
L: an yn - 1 n = L yP L an ( p ) ( - l)P - n+ 1
(l+y) p;:'O n';p+l n-l
that
"
L.,anz n-l _
- ""
L,
( z
- _1
)p - _1
1 (l)PA
- ""pal -_ (1PI(z)
_ )h'
O';p';no Z Z Z
where PI(z) is a polynomial. This proves the theorem.

Now we make the following remark. If the positive integer N is such
that the quantities anNn (n ~ 1) are integers, and if we denote by p the
greatest distance between the point 1 and the singularities of F (see VI.4.6),
we must have
1
N ~ - - 1,
p
provided that F(z) = L: anz n is not of the form (VI.10).

So we obtain a lower bound for the integer N for a function repre-
sented by a series with rational coefficients to be, for example, algebraic,
but not rational.
We consider again the function F(z) whose Taylor coefficients are
integers, of non-zero radius of convergence, the function being uniform,
having infinity as a regular point or as a pole. Let S be the complement
of the domain of existence of F in the open complex plane. According to
our preceding definitions, S is therefore the singular set of F (of which we
remove the point infinity if it is a pole). If E is a set in R2, we set H = EnS.
If H is finite and only contains poles of F at algebraic points, then there
exists a polynomial P(z) with integer coefficients such that the function
F(z)P(z) is regular at the points of H. If we now denote by n(F, E) the
set of such polynomials, we define the polar algebraic index of E with
respect to F as the smallest degree of the polynomials in n(F, E); we de-
note this index by p(F, E).
If there is no polynomial P with integer coefficients such that F(z)P(z)
is a regular function at the points of H, that is, if n(F, E) = 4>, we say that
the polar algebraic index p(F, E) of E with respect to F is + 00. Thus,
p(F, E)=oo in one of the following two cases (a) and (b).
(a) The set H is infinite.
(b) The set H is finite but contains at least one point of S which is not
a pole at an algebraic point; in other words, E contains either a point of

S which is not a pole, or a pole at a transcendental point.
The following theorem is an immediate consequence of Theorem V1.4.6
(corresponding to N = 1).
THEOREM VI.4.7. If the analytic continuation of F(z) = .2: anzn, with integer
coefficients, is a uniform function having infinity as a regular point or as a
pole, then the polar algebraic index with respect to F of the (closed) half-
plane x~t (z=x+iy) is not less than the polar index of the (open) half-
plane x>! with the point z = 1 excluded. That is,
(VI.12) p(F, {zlx ~ t}) ~ p(F, {zlx > t, z=/= I}).
More generally,
(VI.13) I/(F, {zlx ~ t}) c I/(F, {zlx > t, z =/= I}).
The relations (VI.12) and (VI.13) remain valid if in the terms on the left
of these relations the half-plane x ~ t is replaced by the same half-plane
with the poles of F at rational points excluded.
Suppose that the relation (VI.13) does not hold when the half-plane
x~t is replaced by the same half-plane from which the poles at rational
points, say Pl/ql' P2/q2, ... , Pm/qm, where the Pi and qj are integers, are
removed. Then there would exist a polynomial with integer coefficients
P(z) of degree less than p(F,{zlx>t,z=/=I}), such that F(z)P(z) has in
the half-plane x ~! as singularities of finite distance (from the origin)
only the poles at the places pj/qj (j= 1,2, ... , m); the same product hav-
ing in the half-plane x> t singularities other than the point 1 at finite
distance. On the other hand, there would exist a polynomial Q(z) with
integer coefficients, of the form
(qlZ - Pl)a1(q2z - P2)a2 ••• (qmz - Pm)am,
where aI, a2,"" am are positive integers, lJI(z)=F(z)P(z)Q(z) is holo-
morphic on every compact set in x ~ t, and where the product has in
x>t the same singularities as F(z)P(z). It is clear that there would also
exist a polynomial (with integer coefficients) R(z) of degree k such that
(lJI(z)-R(Z»Z-k-l is a Taylor series with integer coefficients represent-
ing a function regular at infinity, holomorphic for x~t and having in
the half-plane x> t singularities other than the point 1; but this would
be a contradiction to Theorem VI.4.6.
Here is an immediate corollary to Theorem VI.4.7.
THEOREM VI.4.8. If the non-negative integers an are such that

lim sup a~/n = a with 1< a < 2, and if the analytic continuation F of L: anz n
is a uniform function, with the point at infinity being regular or a pole,
then F has apart from the singular point z = a -1 another singular point at
finite distance situated in the half-plane x:::;!.
NOTES
1It is easy to see that ua,r=O implies that Ul"';O.

2To simplify the formulae we have preferred in this case the function m(x) to N(x),
which denotes the number of An smaller than x.
CHAPTER VII
ON THE BEHAVIOUR OF THE REMAINDERS OF

A DIRICHLET SERIES IN THE DOMAIN OF
EXISTENCE OF THE FUNCTION. APPLICATIONS
VII. I. Evaluation of the remainders of CA, A)

In this chapter we assume that A has a positive step h, and we set
infC"n+l - An) = p.
Then O<p~h. To simplify the statements we also assume that the ab-
scissa of convergence is equal to zero, that is,
aa = a c = lim sup log"lan l = O.

n
Let C = {en} be a real sequence in which infinitely many terms may be

equal to - 00, but infinitely many are finite. Suppose that
Cn = 0.
· sup A
11m
n
Then there exists a concave function C(x) (x ~ "1) having the following
properties:
(1) C("n)~cn (n~ 1)
(2) C(x)~O,
(3) Every other concave function c1 (x) (x ~ "1) satisfying the same
inequalities also satisfies c1(x) ~ C(x). The function C(x) is called the
smallest concave non-negative envelope of the sequence ofpoints Pn= ("n, cn)
(n~ 1).
It is easily seen that lim C(x)/x=O.
Given a series f=(A, A) we denote by fn the function represented by
the sum
that is, by the n-th remainder of (A, A).

We prove the following theorem:
THEOREM VII. I. I. Suppose that f=(A, A) is holomorphic in a simply

connected domain D containing the half-plane of convergence (a>O). Let
THE REMAINDERS OF A DIRICHLET SERIES 117
C(x) be the smallest non-negative concave envelope of the sequence of

points (An' log lanl) (n~ 1).
For every compact domain d contained in D there exists a constant
P = P(Ll) such that when we set qn = eC(A n), we have for s E J:
Ifn(s) I eVJ ::::; Pqn.
We set
(VIl.1)
We see immediately that 7Jn decreases to zero. We consider the function

<Pn(s) = fn(s + 7Jn) eAn(s+nn)q;;l.
For sELl and sufficiently large n we have:
l<Pn(s)I=lf(s + 7Jn)e An(s+nn)q;;l - L a q;;le-(lI
m m-lIn)(s+nn)!
m~n
~
-.. ;;;:
M n eAn(nn+a)q-l
n ~
q-l e-(lI m -lI n)(a+ nn )
'" qmn
m~n
= In(s) + Jis),
where Mn = max If(s+7Jn)1 as long as sEd, n ~no.
Evidently Mn::;:; M < 00 for n ~ no.
Taking account of (VIl.1) and the fact that C(x) is concave, we see that
for l::;:;m<n(n>l)
(VII 2) C( An) - C( Am) >-

• A _ A ~ 7Jn·
n m
Hence, if a < 0, then by making use of (VIl.2) for m = 1 we have:

(VII. 3) In(s)::::; M elln nn e- C(lIn) ::::; M e1l1n1 .
From the same inequality (VIl.2) (and still for a<O) we deduce:
(VIl.4) In(s) = L e - C(An)+ C(7'm)+ (lin -lIm)nn elAn -lI m)a
m~n
where p is the quantity defined at the beginning of this chapter.

The inequalities (VII.3) and (VIl.4) give for a < 0, sEd:
K
(VI1.5) l<Pn(s)!::::; 1 _ epa'
where K is a constant independent of n, n ~ no.

For sELf, 0'>0 we have:
IPn(S) I = 12: amq;;l e-(llm-IInl(s+nnll

m>n
~ 2: eC(lIml - e(lIm -
C(lIn' - (11m - "nlnn "nl(J
m>n
and according to (VII.l):
Thus, both for 0'>0 and 0'<0, S ELf (n~no):
(VII.6) IPn(S) I ~ 11 _K;_P(JI'
where Kl is a constant: Kl =Kl(LI).

We denote by R[so, a] the square defined by 10'-0'01 <a, It-tol <a
(so = 0'0 +ito), and we consider the function
"'n(s) = [1 - e- P(S-l(t o -2d)l][1 - e- P(S-l(t o +26))]Pn(s).
According to (VII.6), f (and "'n(s» is holomorphic in R[ito, 30] for
sufficiently small 0> O. Now "'n(s) satisfies on the contour of R[ito, 20]
the inequality
I"'n(s) I ~ B< 00 (n ~ 1),
where B is independent of n (n ~ no).
Since on the contour of R[ito, 0] for sufficiently small 0
we obtain for S E R[ito, S]:

IPn(S) I ~ Lp. -2 (n ~ 1),
where L is a constant.
Thus, for every So ELf there exists a square with this point at the centre
such that on the square max IPn(s)1 < M o, where Mo does not depend on
so. By the Borel-Lebesgue theorem, the function PII(S) is bounded on Lf
by a quantity independent of n. This completes the proof.
We make the following remark concerning the smallest non-negative
concave envelope of a sequence {cn}. If lim sup cn= 00, then there exists
a sequence {nj} such that C(Anj)=C nj (j~ 1), where the points (AnI' cn)
are the vertices of the polygonal line defined by y = C(x) (x ~ A1)' If
lim sup Cn = I, with - 00 < 1<00, then to every sequence {nj} for which
there corresponds a constant c such that
(VII.7) cn} = C(Anj) ~ cnj + c.

A sequence {n j } satisfying (VII.7) is called sequence of principal indices
for {c n}.
Theorem VII.l.lleads to the following proposition:
VII. 1.2. Let {n j } be a sequence of principal indices for {log lanl}.

Suppose that f = (A, A) is holomorphic in a simply-connected domain D
containing the half-plane a> O.
For every compact domain J contained in D, every s E J and every
j~ 1 we have:
Ifnl(s) I ell.n/ ~ Qlanll,

where Q is a constant.
The following elementary lemma is also useful:
VIl.l.3. Iff= (A, A) has a pole on its axis of convergence, then
(VII. 8) lim sup lanl > O.

If (VIl.8) would not hold, then to every e > 0 there would correspond
an no such that lanl <e for n>no; and then for a>O:
If(s) I ~ L lanl e-lI. =

n" 2:
n ~no
lanl e-lI. n" + 2:
n>no
lanl e-lI. n"
~ A(a) + e L e- np" ~ A(a) + e(1 - e-P")-l,
where p > 0 is the constant used above.

A(a) is bounded for a~O. For at 0 we would also have:
f(a + it) = 0(a- 1)

(uniformly in t). No point ito on the axis of convergence could therefore

be a pole, contrary to our assumption.
Theorem VII.l.l (or its Corollary VII.1.2) enable us to introduce apart
from a function f = (A, A) an auxiliary function which has the effect of
'separating' the singularities of f on its axis of convergence from all its
other singularities.
THEOREM VII.1.4. Let {nil be a sequence ofprincipal indices of the sequence

{log lalll}({all } = A).
Suppose that the axis of convergence (a=O) is not a cut for (A, A), and
let T be the domain composed of the half-planes a < 0, a> 0 and the set of
regular points off on a= O.
From {nil we can extract a sequence {Pk} having the following property:
the sequence
fpk(s) el-."k'a;/ (k ~ 1)
tends, uniformly on every compact set contained in T, to a function g(s)

that is holomorphic in T.
There exist two sequences M = {I-'n} and M' = {I-'~} such that
1-'1 ~ h, 1-'11+1 - 1-'11 ~ h (n ~ 1)
1-'1 ~ h, I-'~+l - I-'~ ~ h (n ~ 1),
where h is the step of A, and two sequences B={bn}, B'={b~} such that for
a>O, a<O we have, respectively,
g(s) = L bn e-P.n S
g(s) = L b~ eU~S - 1.
The function g(s) has singular points on a=O. They are all the singular
points off(s).
Every isolated singular point of g is a simple pole.
We remark that the condition on the n, to be principal indices of the
sequences {log lanl}, that is, that
log lanjl ~ ceAnj) ~ log lan1 1 + c,

where cex) is the smallest non-negative concave envelope of the points
(All, log Ian!), can be replaced by a more general condition. But the general
theorem is more difficult to prove. The statement we have given here

suffices for the applications we have in mind.
The sequence of functions {F;(s)} defined by
(VII.9) F;(s) = fnls) eAn/lanjl-l (j ~ 1)
constitutes a normal family which is bounded on the J defined in VII. 1. 1.
From it we can extract a sequence Fiv(S) that converges uniformly on
every compact set contained in Ll to a holomorphic function, which we
denote by g(s).
From the fact that CCx) is concave and from the relation
lim CCAn)P.n = 0
it follows that
in other words, if as in VII. 1. 1 we set log qn = CCAn) we have:

(VII.10) lim qn+l/qn = 1.
We set niv=m. (v~l). Taking account of the inequality qmv~Clamvl
(where C is a constant) and of (VII.10) we obtain for a positive integer m:
(VII.11) \amv±m\ ~ qmv±m ~ Cqmv±m

amv lamvl qm v
(we consider m.-m only for m<m.), that is:
. sup \am±m\
(VII.12) hm _ v- ~ C.
amv
Note that it is no loss of generality in the theorem if we assume that
there exists a constant H < 00 such that
In fact, if this condition does not hold, it is sufficient to replace, on the

one hand, A by a sequence A' having the same step as A and containing
it and, on the other hand, A by the sequence A' whose elements corre-
sponding to the An of A are those of (A, A) where the other elements of A'
are all zero. We observe immediately that in proving the theorem for
(A', A') we also prove it for (A, A). Hence we assume that (VII.13) holds.
If h is the step of A, let 0 < e < h - e. For sufficiently large v we have:
(0:) m(h-e)~Amv+m-Am/<;.mH,
(fJ) m(h-e)~Amv-Amv-m~mH,
where the inequality between the first and the second term of (fJ) only
holds when mv-m~p=p(e).
It follows that we can extract from {my} a partial sequence {Pk} (for
example, by the diagonal process) having the following property: the
limits
lim (Apk + m ApJ =

- fLm' lim apk + ma;k1 = bm,
lim (A pk - Apk - m) = fL~, lim apk - ma;/ = b~,
exist and satisfy the relations
(VII.14) Ibml ~ C, Ib~1 ~ c

(VII.15) fLl ~ h, fLm +1 - fLm ~ h,
fL~ ~ h, fL~+l - fL~ ~ h.
For a>O and a<O we set, respectively:
(VII.16) g+(s) = 2: bme- llmS ,
(VII.17) g-(s) = 2:b~ell:'s.
According to (VII.14) and (VII.15) the series (VII.16) and (VII.17)
converge, respectively, in the half-planes a> 0 and a < o. And by the
definition of g(s) (which in LI is the limit of the sequence Ft , (nt,=m v »,
for a>O we have:
g(s) = g+(s).
If D is defined as in VII.l.l, we now suppose that the compact set J
contained in D forms part of the half-plane a < O. Then we have uniformly
onJ
limf(s) eAm/a;,! = O.
This allows us to write:
The definition of the sequences {fL~}' {b~} enables us therefore to write

for a<O:
g(s) = g-(s) - 1.
Thus, g(s) cannot have singularities outside the axis of convergence
(a=O); and its singularities are the same as that of f(s). From (VII.14)
and (VII.15) it follows that for a>O:
Ig(s) I = Ig+(s)1 ~ c 2: e- mha = C e- ha (1 - e-ha)-l.
A similar inequality is valid for Ig-(s)1 in the half-plane a<O. Thus,

for a#O, lal ~ 1
Ig(s) I ~ Klal-l,
where K is constant.
Suppose now, contrary to our assertion, that g(s) is regular on a=O.
Let 0 < a < 1 and to be arbitrary real and consider on the square
R[ito,2a] (Ial ~2a, It-tal ~2a) the function
(VII.19) G(s) = g(s)[s - i(to + 2a)][s - i(to - 2a)).
Then G(s) would be hoI om orphic on this square, and on its sides we
would have
On the square R[ito, a) we would therefore have Ig(s) I ~ 20Ka- 1 •

In other words, the function g(s) would then be an entire function and
bounded in the whole plane; consequently, it would be a constant; but
g(s) tends to zero as a tends to + 00 and to - 1 as a tends to - 00. Hence
we have a contradiction. Thus, g(s) necessarily has singularities on a=O.
By the definition of g, these singularities are those off
Let us show finally that every isolated singular point of g is a simple
pole. Let iT be such a point, and consider again the function G(s) defined
by (VII.19), where a>O is such that the closed square R[iT, 3a] contains
no singularities other than iT and where to = T ± 3a.
On to+a=T±2a, lal ~a, we find that Ig(s) I~20Ka-l. Thus,
on the sides of R[iT,2a) we have Ig(s) I ~Pa-l (for lal =2a we have
Ig(s) I~ K(2a) -1), where P is constant. Thus, iT can only be a simple pole.
The proof of the theorem is now complete.
VII.2. Applications of the results of Chapter ill and of VI.l to

series meromorphic on an interval of the axis of convergence
We prove the following theorem.
THEOREM VlI.2.1. Let J5' be the mean upper density of A and suppose that
the sequences A (of step h) and A are such that there exist three constants
H>O, C>O, N~ 1 with
(VII.20) An+l - An < H (n ~ 1)

(VII.21) C- 1 ~ max
I .. v .. N
lan+vl ~ C (n ~ 1).
Suppose that f(s) = (A, A) does not have in a strip S defined by a < t < b,
where b - a > 27T(l5' +h -1), singularities other than simple poles iap
(p= 1,2, ... , q), where the ap are real andf(s) remains bounded in S out-
side every neighbourhood of the points iap (p = 1, 2, ... , q).
Then the function f(s) is holomorphic for a < 0, where it is given by a
series
f(s) = L: a~ eA~S + d. (d a constant)
Furthermore:
Every isolated singularity ia off on a =

q integers ml> m2, ... , mq such that
° is a simple pole, and there exist
In other words, the ap (p= 1,2, ... , q) form an 'integral basis' for all
the a, which are the ordinates of the points of other isolated singularities
of f on the axis of convergence.
We recall that the step h of A in this chapter is supposed to be positive
and that the analytic continuation of (A, A) into S (from which the
points iap (p = 1, 2, ... , q) are removed) is direct.
Remark. The conditions (VII.20) and (VII.21) follow from other con-
ditions in the statement of Theorem VII.2.l, namely from the fact thatf
only has simple poles on the interval in question of the axis of convergence.
But when the An are not integers, the proof of this fact is not immediate.
It follows from (VII.20) and (VII.2l) that there exists a sequence {nil
of principal indices of {log lani} satisfying the conditions
(VII.23) nj+l - nj ~ N
(VII.24) Cil ~ lanil ~ q(n,) = eC(h nj ) ~ Cl>
where C(x) is the smallest non-negative concave envelope of the set
of points Pn=(An, log lani) (n~ 1), and Cl is a constant such that
C(x)~log Cl (X~Al)'
Consider the family of functions
f:(s) = [f(s) - 2:
An",X
an e- hns ] eXS (x > 0).
Let j(x) be the largest integer j such that Anj ~ x, and set m(x) = n,(X)'
Let n(x) be the largest integer n such that An ~ x. Then
o ~ n(x) - m(x) ~ N.
If n(x)~m(x)+ 1, we can write
f:(s) = [fm(xls) - (am(X)+le-hm(x)+lS + ... + an(X)e-hn(X)8)]eXS,
with X-Am(xl+l ~NH; and if n(x)=m(x), then
f:(s) = fm(xls) eXS .
It follows from Theorems VII. 1.4 that the family f:(s) (x> 0) is bounded
on any compact set containing no singularities off and that we can ex-
tract from every sequence {Xk}, Xk > 0, Xk too, a partial sequence {y,}
such that f~(s) tends uniformly on this compact set to a function G(s)
having the following properties:
(1) For a>O and a<O, respectively,
(VII.25) G(s) = L, dn e- VnS + c,

G(s) = L, d~ eV~s + d,
where c and d are constants and where for n ~ 1 :
(VII.26) Vn+l - Vn ~ h
V~+l - v~ ~ h
Idnl ~ C
Id~1 ~ C
(we see easily that C is the constant occurring in (VII.21)).
(2) G(s) is holomorphic outside the singularities of 1 on the axis of

convergence 0'=0.
(3) G(s) has singularities on 0'=0.
(4) The isolated singular points of G(s) are simple poles.
(In fact, (3) follows from VII.1.4 as long as d =f:. c, but we shall prove
this statement directly in the subsequent analysis.)
Let us now show that every simple pole of 1 on 0' = 0 is also a simple
pole for G. Let i{3 be such a pole for I; we can write
r
I(s) = S
--'{J
-l
+ <pes),
where <pes) is a holomorphic function in a disk C(i{3, 8) (8) 0). In this
disk the functions I:;(s) only have i{3 as a singularity, namely as a simple
pole with the residue retBYj ; the functions
relBYj
<PYj = IY~(s) - S - i{J
are therefore holomorphic in C(i{J, 8). If we extract from {yj} a sequence

{yj} tending to infinity and such that eIBY ; tends to a limit I, then the
functions <py; tend uniformly on the boundary of the disk, and conse-
quently on the closed disk, to a limit that is holomorphic in the disk.
The limit of Iy;(s) which is G(s), is therefore of the form
r'
(VII.27) G(s) = --'{J
s - I
+ Go(s),
where r' = Ir, and Go(s) is holomorphic in C(i{3, 8).

Since G(s) is determined by the sequence {y,}, the quantity I, and con-
sequently the quantity r' =rl(1/1 = 1), cannot vary with the choice of the
sequence {yj} C {yj}. It follows that lim etBYj exists and is equal to I.
That is, (VII.27) holds, with r' = r lim elBYj.
Hence if {yj} is a sequence for which/i;(s) tends to G(s) defined above,
then G(s) has the points ia p (p = 1, 2, ... , q) as simple poles. If the residues
of the poles iap of 1 are the quantities
rp (p = 1,2, ... ,q),
then the residues of the same poles of G are rplp, where Ip=lim etapYj
and these limits exist p= 1,2, ... , q.
Now by Dirichlet's approximation theorem we can choose a sequence

y, tending to 00 with the property that
Ip = 1 (p = 1,2, ... ,q),
that is, lim apYj=O (mod 27T) for every
P = 1, 2, ... , q.
Thus, on ]ia, ib[ of a=O the function F(s)=f(s)-G(s) is holomorphic.
We denote by {Yn} the union P·n} U {v n} and set
(VII.28) F(s) = L: en e- YnS = L:ane- AnS - L:dne- vns •
It is easy to see that if we denote by 15~ the mean upper density of Yn,
then
15~ ~ 15' + h- l •
Thus, the function F(s) given by (VII.28) is holomorphic and bounded
in the strip S whose width exceeds 27T15~. We conclude, by Theorem
III. 2.3 , that F(s) is identically zero, that is:
(VII.29) f(s) = G(s) - c.
From this it follows, first of all, (by (4)) that the isolated singular
points off on a = 0 are simple poles and that f(s) is also holomorphic for
a<O. Further, by (VII.25), (VII.26) and (VII.29) An=Vn, hence:
for a<O:
f(s) = L: a~ eA~S + d,
where ,\~ = v~, a~ = d~; hence:
'\~+1 - ,\~ ~ h (n ~ 1)
la~1 < C.
It now remains to show that there exist q integers rnl> rn2, ... , mq such
that (VII.22) holds.
We show that the contrary assertion leads to a contradiction.
If a is such that there exist integers m, Pl, P2, ... , pq satisfying the
relation
(VII.30) rna = Plal + P2a2 + ... + pqaq,
then we denote by p. the smallest possible integer m having the property

(VII.30). If (VII.22) were not possible, we would have p.> 1, and every
other integer m satisfying (VII.30) would be a multiple of p.. Setting
fi = p. -1 « 1), the numbers a, fi, al> a2, ... , aq would have the following
property:
For arbitrary integers n, nl> n2, ... , nq for which
(VII.31) na + n1al + ... + nqaq = 0,
the quantity nfi (0 < fi < 1) is an integer.
From a simple version of Kronecker's approximation theorem it would
follow that for arbitrarily given 8> 0, T> 0 the systems of inequalities
(VII.32) -8 < ta - fi < 8 (mod 1)
-8 < tap < 8 (mod 1), (p = 1,2, ... ,q),
would have a solution t> T.

If (V1I.31) could not be realized with any system of integers
n, nh n2, ... , nq, then (VII.32) would equally have a solution t> T, and
this for an arbitrary fi with 0 < fi < 1.
In other words, if a could not be expressed as a linear combination with
integral coefficients of the a p (p = 1, 2, ... , q), then there would exist a
0< fi < 1 such that (VI1.32) would have a solution t> T.
Hence there would exist an 7]0 with 0 < 7]0 < t such that to every 8> 0,
T> 0 there would correspond at> T and integers n, nl> n2, ... , nq satis-
fying the inequalities
Itap - npl < 8 (p = 1,2, .. . ,q)
7]0 < Ita - nl < 1 - 7]0'
And then we could associate with every sequence {8j}, 8j to, a sequence
{yj} and q+ 1 sequences nJ, n~l), ... , n~q) such that
(V1I.33) IYfap - 217nY') I < 8j (p = 1,2, ... , q;j ~ 1),

27]017 < IYfa - 217n~0) I < 2(1 - 7]0)17 (j ~ 1).
Choosing the function G(s) with the Yj satisfying (VII.33) we would
find, first of all, as above, that G(s) has the poles iah ia2, ... , iaq with the
same residues as /; but if the residue of ia for / is r, say, the residue
of this pole for G would be r' = r e2!t19 #- r (because e2!t19 = lim e1alll, with
7]0::::; B::::; 1 -7]0)' which is a contradiction. Thus, the theorem is proved.
CHAPTER VIII
APPLICATIONS TO THE GENERALIZED

RIEMANN FUNCTIONAL EQUATION
vm.l. Number of independent solutions. Links between exponents

The Riemann zeta function '(s), the analytic continuation of the Dirichlet
series
which is known to play a fundamental role in the theory of numbers

(for example, in the problem of the distribution of prime numbers),
satisfies the so-called Riemann functional equation
Hamburger has shown that if the series
(VIII.2) f(s) = L a~,

n
g(1 - s) = L n~~8
converge absolutely for u> 1, and u<c<O, respectively, where in addi-

tionf(s) is assumed to be of the form G(s)jP(s), with G(s) an entire func-
tion of finite order and pes) a polynomial, then the functional equation
implies thatf(s) =g(s) =al'(s).

Thus, the Equation (VIII.3) characterizes in a way the function '(s),
provided that f and g are represented by Dirichlet series whose sequence
of exponents is {log n} (in the language we have used right through this
book); true, certain simple conditions on the behaviour of these functions
in the entire plane have to be added.
This is where the more general problem arises. As always, let A={An},
M = {f'n} be two positive strictly increasing sequences and let
a b
(VIlI.4) <pes) = L "\~' ifJ(s) = L f'~
be two series having abscissae of absolute convergence. Imitating the
previous notation, we could denote such series by (A, log A), (B, log M),
respectively, where <p and ifJ denote not only the sums of the series but
also their respective analytic continuations. If S is a given positive num-
ber, suppose that the following relation holds:
(VIII.S) 1T- S/2rG)<p(s) = 1T-(6-S)f2r(0 ; S)ifJ(1) - s).
Clearly the behaviour of the function 0, we can ask
for the 'size' of the class of pairs (<p, ifJ) satisfying (VIII.S) We can also
ask to what extent we can choose the sequences {An} and {f'n} themselves
(in general or in relation to a given S). These problems will be treated in
the present chapter.
Now we lay down the explicit meaning which we give to the expression
°°
'the pair (<p, ifJ) satisfies the generalized Riemann equation (VIII.S)'.
If the sequences {An} and {f'n} are defined as above, if > is given, if
<p and ifJ are defined by (VIII.4), where the series have abscissae of absolute
convergence, we suppose that there exists a function xes), not identically
zero, holomorphic at finite distance outside a compact set K, and of
exponential type zero in every vertical strip, that is,
(VIIl.6) lim log IX(<1 + it) I = 0,

Itl-+ CIO t
uniformly in <1 for a <. <1 <. b, a and b being arbitrary; here X is linked
°:
with 
(VIII.7) xes) = 1T-S/2r(~)<p(s).

For sufficiently large - a > 0:
0-
(VIII.8) xes) = 1T-(6-s)/2r ( -2- ifJ(1) - s). S)
THE GENERALIZED RIEMANN FUNCTIONAL EQUATION 131
Then we say that the pair (<<p, ifJ) satisfies (VIII.5).

Now, as we have said above, since we are looking for conditions on
A={.\.n}, M={fLn} and ~l>O under which there exist an and bn (n~l),
not all zero, which make (VIII.5) possible (or, in a more general manner,
since we are looking for the number of such sequences {an}, {bn} from which
all the others can be formed), we agree to say also that the pair (<<p, ifJ)
satisfies the Riemann functional equation corresponding to (A, M, S).
We begin with the following lemma:
VIII. 1. 1. If the pair (<<p, ifJ) defined by (VIII.4) satisfies (VIII.5) and if we
set
(VIII.9) g(s) = (B,2TTM) = L: bn e-2I1JlnS,
then 0'a,g = 0,' for 0' > 0 the following relation is satisfied for sufficiently
large do:
(VIII.10)
where
(VIII.1I) Aj = (-IJ;~~:/ j)
and where K(s) is an analytic function on the Riemann surface of log s
with
(VIII.12) K(s) = O(lslm) (Isl-+ 00, m a constant)
in every angle larg sl ~ 00 (00 )0 positive, m depending on 00 ),
For 0'>0 and do>O we have:
1.
(VIII.13) e- S = -2
TTl
1 c10
clo-leo
+leo r(w)s-W dw.
We also have:
(VIII.14) r(w) = TT-1/22w-lr(i)r(W ; I).

1 f<lo+leo
(VIII. 15) -2. z-Wr(p - w)r(w) dw
TTl clo-leo
= r(p)[(l + z)-P + 2:
O';i<clo-p
Ajz-p-l]
with
(-l)fr(p + j)
Ai = j!r(p) ,for larg zl < TT, do > p > O.
Suppose that a> 0 and that do is sufficiently large. Then (VIII.13) and
(VIII.4) give (for if)
1
= 2---:
TTl
i do +IOO
£10- 100
G(w, s) dw,
where
G(w, s) = r(W)(2TTS)-Wif(w).
Therefore, by (VIII.14) and (VIII.8) we can write:
(VIII.16) G(w, s) = l TT -<w+ll/2s-wr (W ; l)x(O - w).
By setting
(VIII.I7) K(s) = 21 .
TTl
r G(w, s) dw,
Jc
where C is a rectifiable curve enclosing K (the integral being taken in
the positive sense) we can also write for a>O, by applying Cauchy's
theorem (do being sufficiently large), with (VIII.6) taken into account
g(s) = K(s)
1
+ -2 •
TTl
16 -£1 0 +1 00
6-£10-100
G(w, s) dw
1 ido+l
+ -2
OO
= K(s) • G(o - w, s) dw,
TTl £10-1 00
with
G(o - W, s) = lTT-<6+1l/2s-6TTW/2swre - ; + l)x(w).

The formula (VIII.7) allows us to write
G(8 - w, s) = tn--<6+ 1l/2s -6+Wr (i)re - ~ + l)tp(W).

By (VlII.4) (defining cp) and (VIII.15), where we put p=q=(8+ 1)/2,

we have:
-.
1
47Tl
f do
+
do-loo
1OO
-2 r (q - -W)
swr (W) 2
cp(w)dw
A2) -q (S2)Q+f]
= r(q) 2: an [( 1 + ~
s o ..
+
2,2:do - 2Q Aj
<
A2
n
.
This finally gives (VIII.10). As to (VIII.12), we see this immediately

by (VIII.I6) and (VIII.17).
If two pairs (CP1' "'1), (CP2' "'2) satisfy the Riemann functional equation
corresponding to (A, M, 8) then for arbitrary constants Cl and C2, the
pair
(C1CPl + C2CP2, Cl"'l + Cl"'2)

satisfies the same equation. A set of k pairs
(cpi> "',) (j = 1,2, ... , k)

satisfying (VIII.5) is said to be linearly dependent if there exist constants
c, (j= 1,2, ... , k), not all zero, such that
C1CPl + C2({J2 + ... + CkCPk == 0,
Cl"'l + C2"'2 + ... + Ck"'k == O.
If this is not the case, we say that the k pairs are linearly independent.
The following theorem gives an indication of the number of linearly
independent solutions which a Riemann functional equation correspond-
ing to (A, M, 8) (where the sequences A, M and the constant 8>0 are
given) can have:
THEOREM VIII.I.2. Let i5~ be the mean upper density of M and assume
that i5~ < 00.
The number of linearly independent solutions of the Riemann functional
equation corresponding to (A, M, 8) is equal to the smallest number of An
(An E A) lying in an interval on [0, oo[ of length exceeding i5~.
Let [a, b] be an interval a> 0, b - a > i5~, containing the points
Ap+ 11 Ap + 2, ••• , Ap + k and not containing other An' We are going to show
that there exist at most k independent solutions of (VIII. 5) corresponding
to (A, M, 8).
Let (IPq, tPq) (q= 1,2, ... , k) be k solutions of (VIlI.5). If these solutions
were linearly dependent no matter how they are chosen, the theorem would
be proved. Suppose then that this set of solutions can be chosen linearly
independent.
We set
The quantities a~q~ h a~q~ 2, ••• , a~qh, therefore are such that for no set
of k constants Cl, C2, ... , Ck, not all zero, we can have
(VIII.18) L
l",q",k
Cqa~qtl = 0 (j = 1,2, ... , k).
For if (VIII.18) were to hold (for a choice of the constants Ch C2, ... , Ck,
not all zero), then the function
F(s) = L: [ L cqb~q) e - 2nllns] = L: dn e - 2nllnS
l~q(k
would have the following properties:

(1) F(s) would be holomorphic, by VIII.l.l, in every strip a<t<b,
(2) in this strip, again by VIII.l.l, we would have F(s)=O(JaJT) as a
tends to - 00 (r being constant). Denoting by IL a constant 0 < IL < ILl and
setting Fl(S)=F(s) e2nus , we would have
Fl(S) = L:dn e- 2n (U n -Il)S = L:dne- vns ,
where Fl(S) is holomorphic and bounded in the strip a<t<b. But the
upper mean density 15~ of {v n} is equal to 15~/27T. Thus, Fl(s), whose ab-
scissa of absolute convergence is finite (=0), would be holomorphic and
bounded in a horizontal strip of width exceeding 27T 15~. By Theorem III.2.3,
Fl(S) would be identically zero, contrary to our assumption that the pairs
(IPq, tPq) are linearly independent.
We have just shown that the linear independence of the solution
(IPq, tPq) (q= 1, 2, ... , k) implies that of the k systems (a~qt h a~qt2' ... , a~q~k)'
For any (k+ 1)-th system (a~k:F, a~kN), ... , a~kN) there exists constants
al, a2, ... , ak such that
a~\+/) = ala~ltj + a2a~2tj + ... + aka~ktj (j = 1,2, ... , k).
Thus, if we consider a solution (IPk+l, tPk+l) of (VIII. 5) written as
we can state, by applying VIII. I. I once more, that for a certain constant
ft' with 0 < ft' < ftl the function
F2(S) = 2: d~ e-2n(~n-~')S
= 2: [ L aqb~q) e - 2n(~n -"')8] - 2: b~ + 1) e - 2n(/Ln - /L')8
l~q~k
is holomorphic and bounded in the strip a < t < b (whose width exceeds
21T 15v" where 15~, is the mean upper density of {v~} with v~ = 21T(ftn - ft'».
From Theorem III.2.3 we deduce again that F2(S) =0; that is to say,
(<}?1c+1o rPlc+1) = (a1<}?1 + ... + ak<}?lc, a1rP1 + ... + akrPk)'
This proves the theorem.
The next theorem gives conditions which the sequences A, M and the
constant 0 must satisfy so that the Riemann functional equation corre-
sponding to CA, M, 0) should have a solution. This theorem shows essen-
tially that if the step of M is positive (h/L > 0), then A has an 'integral basis'
which incidentally can be taken in an arbitrary interval of length exceed-
ing 15~+h;l.
THEOREM VIII. 1.3. Let M be a sequence of positive step hp.. Let 15~ be the
upper mean density of M. Let 0 = 1 or 0 = 3 and suppose that the sequences
A and A are such that there exist three constants ~, <r, m~ I such that for
n~l
(VIII.19) fLn+1 - fLn < .p

(VIII.20) <r - 1 ~ max Ibn +v I ~ <r, when 1) = 1;
1:::;; v~91
<r- 1 ~ max Ibn+vfL~Jvl ~ <r, when 1) = 3.

l~v~91
Suppose that the pair
<}?(s) = 2: ~~, rP(s) = 2: b:,

n ftn
satisfies the Riemann equation corresponding to (A, M, 0).
Let p and k be two positive integers such that
Ap + k + 1 - Ap > D~ + h;;l.
Then k ~ 2, and every An is of the form
(VIII.21) An = min )Ap +1 + m~n)Ap+2 + ... + mkn)A p + ko
where the mJ (j= 1,2, ... , k) are integers.
There exists at least one An exceeding t(D~ +h; 1). If q is the smallest
integer such that Aq+l>!(D~+h;1), then every An is of the form
(VIIL22) An = p~n) Al + p~n) A2 + ... + p~n) \,
where the Pi (j = 1, 2, ... , q) are integers.
Furthermore,
ftn+1 - ftn ~ hu (n ~ I).
If we set for a>O:
pes) = (E,21TM) = L bn e- 2ltUn S, when 0 = I
pes) = (EM-I, 21TM) = L bnft;;l e-2ltllnS, when 0 = 3,
then the function pes) is holomorphic and uniform outside the axis of con-
vergence a=O, where its only singularities are simple poles at the places
± iAn and s = O. For a < 0 and suitable constants c and d we have:
pes) = - L bn e2ltil nS + c, when 0 = 1
pes) = L bnft;; 1 e2ltilnS + d, when 0 = 3.
We remark that the conditions (VIII.19) and (VIII.20) in the statement
of Theorem VIILI.3 are superfluous: the conclusion is valid without these
assumptions; and these conditions themselves are consequences of the
other assumptions of the theorem. The present theorem, incidentally,
is an almost immediate consequence of Theorem VII.2.1, where, as we
have remarked, the conditions (VIL20) and (VII.21) follow from the
other assumptions of the same theorem, namely from the fact that the
series (A, A) only has simple poles on an interval of a certain length on
the axis of convergence. Here the conditions (VIII.19) and (VIIL20)
follow from the fact that (cp, .p) is a solution of the Riemann functional
equation.
Here is a proof of Theorem VIII.I.3 on a few lines. The fact that
k~2 is an immediate consequence of Theorem IIL2.3. For if we had
O~k~ I, then there would exist a sufficiently small e>O so that
1= (AP+2 - e) - (Ap + e) > D~ + h;;l.

And an interval of length I would contain at most a single An (namely
Ap + 1). And since D~ ~ h; 1, there would exist an interval 11 of length
exceeding D~ not containing any An.
The function P(s), which has on the axis of the convergence a=O only
simple poles at the places ± i'\" and (by Theorems VII.2.1 and VIII. 1.1)
the point s=O, would be holomorphic in the horizontal strip that cuts
out on a=O the interval II; in this strip the function would be O(lalm)
(a~ -(0). Theorem 111.2.1 would then show that ",-=0. We could also
apply Theorem VIII.1.2 directly, which itself is an almost immediate
consequence of Theorems 111.2.3 and VIII.1.1.
Now (VIII.21) follows immediately from VII.2.1. And (VIII.22) fol-
lows from the fact that on an interval [-i(,\q+l-e), i('\q+I-e)] (where
e > 0 is sufficiently small so that the length of this interval exceeds jj~ + h;; I),
pes) has as singularities only the simple poles at the places ± i'\" (1 ~ n ~ q)
and the point s=O (VIII.22) therefore does not differ from (VIII.21). The
other conclusions of the theorem are reached by applying first VI1.2.1
to pes) eP' (0<1L<27r1L1)' Therefore, by VIII.l.l, we see that for J)= 1
or J) = 3, the function pes) (minus a constant) is odd or even, respectively.
Here are some corollaries to Theorem VIII. 1.3. We assume tacitly
that the conditions (VIII.19) and (VIII.20) are satisfied. We also assume
that the pair (rp, "') satisfies the Riemann functional equation correspond-
ing to (A, M, J» with J)= 1 or J)=3.
VIII. 1.4. If h')2> I, then there exists an increasing sequence of positive

integers {m,,} such that '\,,=m"'\1 (n~ I); there exists an integer v such that
ILv,\1 = I, b,,=bqfor n-=q (mod v).
By Theorem VIII.1.3 we have: 0<'\1 <t(jj~+h;;1); and since, by
hypothesis, '\2> h;; 1~ !(jj~ + h;; 1), the same Theorem VIII.1.3 gives
'\" = m,,'\1 (n ~ I).
Thus, the function
<1>(s) = ~ a" e- 2 "A nS
is periodic of period ;'\11. Since the only singularities of <1> are the points
± ilL" and the origin (by VIII.1.1), there exists an integer v such that
ILv='\11. Consequently, for every n we have: 1L,,=lLq+m'\l\ where
O~q<v (lLo=O), and m is an integer. From the periodicity of <1> it also
follows that b,,-=bq(mod v).
Remark. Since the points ± ilL" and the origin are the only singularities
of <1>(s), namely simple poles of <1> (if J)= 1) or of the function
~ a",\;1 e- 2 "il n8 (if 8=3) we conclude easily, owing to the periodicity
of (/> and of that of the places and the residues of the poles, that the
an and the An satisfy the conditions (VIII.l9) and (VIII. 20) , with the bn
and the JLn replaced, respectively, by the an and the An (for a suitable
choice of the constants ~, cr, Qt).
VIII. 1.5. Ifhp.A2> 1 andif2hp.Al> 1, then An=nAl (n~I).

By VIII.1.4 we have: An = mnAl' where mn is an increasing sequence of
positive integers. If the conclusion of VIII.1.5 did not hold, then there
would exist an no such that
and the condition 2hp.Al > 1 would give

Ano +1 - Ano > h;; 1 ~ 15~;
and III.2.3 would again lead to the conclusion (/> == O.
Remark. The conditions of VIII.l.5 are satisfied if hp.Al ~ 1.
VIII.l.6. If to the assumptions of VIII. 1.5 we add JL2Al > 1, then there exists
an increasing sequence of positive integers {m~} such that JL~ = m~JLl and
AIJLl is a rational number.
We refer to the remark after the proof of VIII. 1.4. By VIII. 1.5, we have
An = nAb hence h", the step of ,\, is equal to AI; the equality JLn = m~JLl
(n~ 1) therefore follows from VIII. 1.4. And since JLn=JLq+mA 1 1
(0 ~ fLq ~ Al \ JLo = 0) (see the proof of VIII. 1.4), we conclude that for an
n with JLn> All we have m~fLl = m~fLl + mAl \ with m~ - m~ > 0, that is,
AlfLl = m/(m~ - m~).
VIII. 1.7. If hp.A2 > 1, 2hp.Al> 1, fL2Al> 1, 2fLIAl> 1, then fLn =nfLl (n ~ 1).
This follows immediately from VIII.1.5 and the first part of VIII. 1.6.
In particular, if Al = 1, JLn = n, then also An = n.
VIII. 1.8. If fLn =n (n ~ 1) and if Al ~ 1, then every solution (f{!, !fo) of(l1, M, S)
is equal to a'(s), where a is a constant and, is the Riemann zeta-function. 1
This follows easily from VIII.1.5 and the argument used in the proof
of VIII. 1.4 (which here yields an=an+l =b l (n~ 1)).
NOTE
1 We recall that from Theorem VIII. 1. 3 onwards we assume that ~=1 or ~=3.
ClIAPTER IX
INFLUENCE OF ARITHMETICAL PROPERTIES OF

THE EXPONENTS OF A DIRICHLET SERIES ON
ITS ANALYTIC CONTINUATION
IX.I. Isolated fractional parts of the exponents and

tbe possibility of analytic continuation
There is an interesting link between two apparently distinct phenomena
in the theory of functions: on the one hand, the dependence of the charac-
ter of a function represented by a Dirichlet series (and above all, of the
distribution of its singularities) on the arithmetic structure of its expon-
ents, and on the other hand, the behaviour of a function of class A (that
is, of a Fourier transform of a function belonging to L) with compact
support, or what comes to the same thing, by the Paley-Wiener theorem,
the behaviour of an entire function of exponential type that is integrable
over the real axis.
A very special case of this link between the two phenomena has already
occurred in Chapter V in the proof of Theorem V.4.3, in which Lemma
V.4.1 played an important role. But, in fact, the lemma in question itself
gives a straightforward link between the behaviour of an entire function
of exponential type on the parallels to the real axis, and the type and the
order of magnitude of the maximum of the Fourier transform of the func-
tion will only be fully applied in this chapter, in (IX.2) more completely
than in {Ix. I). Here we are no longer concerned with the question of
'composition of singularities', which was treated in Chapter V; we shall
speak of the 'distribution of singularities', although some important
points of the technique we use are akin to that of Chapter V.
In what follows we assume, as before, that the step of the sequence A
is positive. We also assume that the abscissa of convergence of (A, A)
is ua=O; we recall that when the step is positive,
Uc = Ua = lim sup (log lanl/An).
We suppose that the continuation (which we assume to be direct) of
f(s)=(A, A) into Pq, with u' < U a is uniform. If Sq' (= Sq'.f) is the singular
set off in Pa" we denote by S:, (= S:'.f) the closure of the set of all points
a+2k1Ti, where a is an arbitrary point of Sa' and k is an arbitrary integer,
then we denote by }; the intersection of S:' with the axis of convergence.
This set 1: is independent of a' if a' < aa' If 1: is not the line a= aa, we say
that the series f = (A, A) is continuable modulo 217i. Otherwise we say that
the series has the axis of convergence as natural boundary modulo 217i. If
the points of 1: are isolated points of S:, we say that the singular set off
on the axis of convergence is isolated modulo 217i.
We denote by (An) the fractional part of An, that is, A-EAn (we recall
that Ea is the integral part of a).
The sequence {(An)} of fractional parts can also be denoted by (A).
The theorems which we are going to prove first of all have the follow-
ing meaning: if f=(A, A) can be continued modulo 21ri across its axis of
convergence, and iff does not grow in P (a' < aa), except at the singulari-
(1'
ties, faster than a function A(t) (which itself is subject to certain conditions),
then the values taken by the fractional parts (An) must be sufficiently
'piled up', the 'piling' depending on the growth of A(t). The piling is still
greater (for example, almost all isolated values of the set values taken by
the (An) must each be taken infinitely many times) if the singular set off
on the axis of convergence is isolated modulo 217i.
To make precise statements of these theorems we need some new
definitions.
We denote by D the set of values taken by the members of (A), that is,
by the fractional parts (An) of the An' If wED, then there exists a An E A
such that (An) = w, and 0 ~ w < 1, An - w an integer; but there may be several
An, even infinitely many, whose fractional part is equal to w. We denote
by Aro the set of all the An for which (An)=w. Let Nro) be the smallest term
of Aro and v(ro) the number, possibly infinite, of elements in Aro. If w:;6:0,
where w is an isolated point of D, we denote by 8(ro) the smallest distance
from w to the other points of the closure of D (if such points exist). The
interval pro) = ]w - 8(ro), w +8(ro)[ is called the interval of isolation of w. In
the interval of isolation of w there is no other point of D, and one of its
extremities belongs to n.
In this section we prove the following theorems.
THEOREM IX. I. I. (1) Suppose that f=(A, A) is continuable modulo 217i

across its axis of convergence (a = 0) into Pa' (a' < 0). Suppose that f does
not grow in Pa" except at the singularities, faster than A(t), where A(t)
(t ~ 0) is a positive non-decreasing function satisfying the condition
(IX.1) 1"1
log ACt) d
--"'-;;2:--'-'-
t
t < 00.
ARITHMETICAL PROPERTIES OF THE EXPONENTS 141
(2) Suppose that when we set
(IX.2) h(a) = 1(t) dt

La> IOgt (a > 0),
the set !J of values W taken by the terms of the sequence (A) of fractional
parts of the An contains an infinite sequence {w n} of isolated values in !J
satisfying the condition
(IX.3) h(NWn)) ~ C o(wn) (n ~ 1),
where c is a constant smaller (2e)-1.
(3) In every Awn there exists a Amn such that
(IX.4) lim logA lamni = 0 (= O'a).

mn
The conclusion is:
We recall that lim sup log (lanI/An)=O because the step of A is positive
and O'a=O.
Thus, the order of magnitude of the number of times the isolated value
Wn of!J is taken (where the value of Wn is useful in the sense of (IX.4);
that is, the Wn are the fractional parts of the Amn which with the corre-
sponding coefficients lamnl provide by means of (IX.4) a limit equal to
the superior limit taken over all the An and lanl, which yields the abscissa
of convergence; here 0'a = 0) by the terms of the sequence (A) is at least
equal to that of the smallest term of A whose fractional part is equal to
=
W n•
If A(t) 1, then h(a)=O, and the condition (IX.3) is automatically
satisfied. Hence Theorem IX. 1.1 leads to the following theorem:
THEOREM IX.1.2. The conclusion of Theorem IX.l.1 remains valid if the

conditions in this theorem are replaced by the following condition:
The series f = (A, A) is continuable modulo 2rri across the axis of conver-
gence into the half-plane P (0" < 0), where the continuation off is bounded,
(J'
except at the singularities. The set of values taken by (A) contains an

infinite sequence {w n} of isolated values in !J satisfying condition (3) of
Theorem IX.1.1.
We are going to construct an example of a series (A, A) that satisfies

all the conditions of Theorem IX.1.2 (which, as we have just seen, is
nothing but a particular case of Theorem IX.l.l), but is not a Taylor-D
series.
Consider, on the one hand, a sequence of positive even integers M = {fLn}
such that fLn + 1 > 2fLn, and on the other hand, a positive sequence {con},
with COl < 1, decreasing to zero; in addition, the sequences are such that
for a certain value a < -log 2
(IX.6) 2: COn e- 2llna < 00,
and let us study the function
( e- +2 e- )ll n.
S 2S
(IX.7) f(s) = 2: e- OlnS
Two polynomials (z+z2)lln (corresponding to two different values of n)

have no power in common owing to the fact that fLn + 1 > 2fLn; we also see
thatf(s) for u>O can be put in the form
f(s) = 2: am e- AmS = (A, A),
where the Am range over all the numbers of the form COn +q with
I-'n ~q~2I-'n (n ~ 1). From the inequalities 1> COl > CO2 > ... ; I-'n+l >21-'n + 1
it follows that A has step h>O, because h;;,.I-C01' Since (Am)=co n
for Am=con+q, we see that !.h:{con} and it is clear that NOln)=fLn+COno
v<Oln) = fLn + 1. We also see that when Am=con+q, then
am = 2- lln (l-'n ).
q - fLn
By setting Amn= COn + 3fLn/2 we see that condition (3) of Theorem IX.I.I
is satisfied. We also see that ua=O.
On the other hand, it is easy to see that f is holomorphic in the domain
D bounded by the curve C, given by Je- S +e- 28 J=2 and containing the
half-line u> 0, and that this curve is a natural boundary for f.
In fact, the function
e- S + e- 2S )ll n
f1(S) = 2: ( 2
certainly has these properties, by virtue of Hadamard's 'lacunary'

theorem, which is a very special case of Theorem V1.2.2 (the series 2: zlln
has the circle of convergence JzJ=l as natural boundary); and for u>O
we have:
so that on every compact set contained in Pa, with the chosen value of
a (where a< -log 2 and (IX.6) holds):
(IX.9) I/(s) - 11(S) I ~ K L: wnlsl e-2~na ~ Kl L: Wn e-2~na < 00,
K and Kl being constants.

Since 15 cPa (because a < -log 2), we see that I(s) is holomorphic
in D; it follows from (IX.S) and (IX.9) that I(s) - 11(S) is holomorphic in
Pa; since the boundary C of D, as we have seen above, is a natural bound-
ary for 11> it is also one for f.
The inequality
I/(s) - 11(S) I ~ L: (1 + e- ron(1)8 n,

with O<~kl (8=8(8», which by (IX.S) is valid on every part 158 of D
situated at a distance not less than 8 (8)0) from C, shows that I(s) is
bounded on 158 (because 11(S) is bounded there). Therefore I(s) satisfies
all the conditions of Theorem IX.1.2.
Here evidently
Hence there exists a Dirichlet series that is not a Taylor-D series but
satisfies all the conditions of Theorem IX.l.l and even those of the special
case Theorem IX.l.2. But these theorems indicate that a Dirichlet series
satisfying their conditions has one important point in common with a
Taylor-D series: if there exist isolated fractional values of the exponents
corresponding to 'useful' An (in the sense of condition (3) of Theorem
IX.l.l), then the number of terms of A that have each of these values as
fractional part is at least of the same order of magnitude as the smallest
of these terms.
We now proceed to the proof of Theorem IX. 1. 1. Suppose that the
theorem is not true. Under these conditions we can extract from {w n} a
partial sequence, which we denote by {an}, having the following proper-
ties.
(1) an tends monotonically to a quantity aO; we may assume that an
decreases to a O, because an insignificant change in the subsequent argu-

ment is sufficient to treat the case when an increases to aO.
(2) 0 < an -aO <-!- inf (A·n+1 - An)=r;

(3) y<an)= o(Nan».
(4) 2: y<ak)=o(Nan».
k~n-l
(5) Na n+ 1) is larger than the largest term of Aan' (We recall that Aro
denotes the set of An whose fractional part is equal to w).
(6) a n -8<an»an +l+8<an +l) (n~I).
Denote by A' the sequence {A~}, arranged in order of magnitude, that

is formed by the term of U Aa•. Then A' is composed of all the Am belong-
ing to all the Aa. (Y~ I).
Let mn=EA~. Then A~=mn+e~, with O<e~-ao<r, e~ decreasing to
aO. To every n there corresponds a y such that e~=av; av decreases to aO.
We see that mn increases to infinity, and from (3) and (4) it follows that
(JX.10) lim 2:. =

mn
o.
Denote by AT. a term of Aa. chosen so that
(IX.1I) lim 10gAlar~ = 0,

r.
which is possible by virtue of (IXA).

Now we construct the functions Fa(z) of V.4.1 corresponding to the
elements specified there, in the following manner: if Xo > 0 is such that
A(xo)~I, then we set C(x)=O for O<x<xo, and C(x)=logA(x) for
x~xo. For the quantity p occurring in V.4.1 we take (ce)-l and for q
(in the same theorem) an arbitrary quantity such that 0::;;;q«ce)-1-2.
When we set a=Nan), the corresponding function (of V.4.I) is denoted
by CPn(z).
Since for every a> 0:
(JX.12) h(a) = f' ~~) dt ~ C:),

it follows from the construction of Fa (see (V.16), (V.17) and (V.18»
that tPn(z) is an even entire function satisfying the following conditions:
(IX.13) ItPn(z) I :.: ; e6(an)IYI-C<l21>~n<x),
where ~n is an even positive function decreasing for x ~ 0, ~nE L, I ~nil = 1;

the Fourier transform $n(u) of tPn(x) is real, non-negative, even, and satis-
fies the conditions
(IX. 14) rPn(u) ~ g e-SnA(an) for lui:.::; /8«(9,
(IX.15) if.>n(u) = 0
where en =h(N"n» is a quantity tending to zero, g is a numerical constant,

and 0:'::;/«1-2ce).
We set
Since tP(z) is an entire function, we have by (IX.13) and the definition

of C(x),
(IX.17) ItP(z) I :.: ; [A(lzl)]-l edlllIL(x),
where d is a constant and L(x) decreases to zero as Ixl tends to infinity;
L(x) EL.
Denoting by $(u) the Fourier transform of tP(x) we also have:
Taking (IX. 14) and (IX. IS) into account, we see that when m is a positive
integer,
except when An is a term A~ of A', and m is the integral part of A~. If the
fractional part of A~ is am then
We set
and consider for a> c> 0 the integral

(IX.18) tp(s) = i C
c-loo
+100
f(z)rf;(s - z)~( - iz) dz.
The conditions of the theorem and the inequality (IX.I7) show that
this integral has a meaning for G> c.
For G~C'>C we can write (IX.I8) in the following form:
tp(s) = L e- ns L am 1 C
c-loo
+ 100
e(n-Am)Z~( -iz) dz;
and with the given properties of ~, by applying Cauchy's theorem we

have
Thus, for G ~ c':
According to the remarks on cP made earlier we see now that:
where the quantities m q, nq, Vq are defined as follows: the integers mq

are the integral parts of the A~, the A~ being arranged in order of magni-
tude and forming the set A', which is the union of all the sets Aav (v~ 1);
by virtue of (2), to every mq there corresponds a unique integer V=Vq
such that A~ E A avq' that is, such that
nq denotes the integer for which A~ = Anq'

Thus, we can write tp(s) in the form:
Since ItPn(U) I is bounded, we see that Ga. 'l'~O; we wish to show that
Ga,'l'=O.
There exists an ascending sequence of positive integers {qn} such that
wn=Vqn grows and that with Tv defined by (IX.H) and Twn=Sn,
Since the step of A is positive and {As.} is a subsequence of A, and since

{wn} and {As.} are increasing sequences (by (5)), there exists a constant
A such that
Consequently,
· log Wn
1Im-
A- = 0.
s.
It also follows from (IX.14) that
with lim '7n = 0 ('7n > 0), because ASn ~ ,\<<<wn).

Consequently we have:
·
11m log I1>w.(O) I >- 0
sup A r·
8.
Hence, U a • !p;;'O and therefore U a • !p=O.

The radius of convergence of the series
is, consequently, unity; and since (IX.10) holds, we see, by Fabry's theorem
(a special case of Theorem VI.2.2), that this series has its circle of con-
vergence (a=O) as natural boundary.
Now since !foes) has as singularities only the points 2k7Ti, where k is
an arbitrary integer, a simple repetition of the arguments used in proving
Theorem V.4.3 (a suitable division of part of the plane, use of Cauchy's
theorem, etc.) shows that lJI given by (lX.18) has in P(J' no singularities
other than points of S(J''/' that is, points of the form (X + 2k7Ti, where (X
is any point of S:'.f and k is an arbitrary integer. Now by the conditions
of the theorem the intersection E of S:·.f with a=O is not the entire axis
(a=O). This means that the axis is not a natural boundary of lJI. This
contradicts what we have found above under the assumption that the
conclusion of the theorem is false. Here we have chosen the an decreasing
to aO. We would find the same contradiction assuming that the an increase
to aO: it is sufficient to replace in (IX.l8) €P( -iz) by €P(iz) e2 • Thus, our
theorem is now completely proved.
A set S is said to be an isolated part modulo 21Ti of Sa if it has the follow-
ing properties: S c S: and there exists a closed Jordan curve contained
in Pa- S: and containing S in its interior.
Clearly, in the example we have constructed earlier (after Theorem
IX.1.2) no point of Sa (=Sa,,) on u=o (=u a) can lie in an isolated part
modulo 21Ti of Sa (in fact, this is true, in our example, for every point of
Sa). This remark shows that our example also illustrates the conditions
of the following theorems:
THEOREM IX.1.3. If to the conditions of Theorem IX.l.l we add the fol-

lowing condition:
(d) for infinitely many values of n
then there exists at least one point of Sa, d<O, on U=Ua that does not
belong to any isolated part modulo 21Ti of Sa.
THEOREM IX.I.4. If to the conditions of Theorem IX.I.2 we add the condi-

tion (d) of Theorem IX.1.3, then the conclusion of Theorem IX.1.3 is still
valid.
Suppose that Theorem IX.1.3 does not hold. Then every point of Sa
on u=O belongs to an isolated part modulo 21Ti of Sa. This means that
every point of S:=O is a member of a set that is part of S:
such that
there exists a closed Jordan curve containing the set and contained in
P d , with no point of S:
on the curve.
From the sequence {con} we can now extract a sequence {an} which by
virtue of property (d) in Theorem IX.1.3 has the properties (1) and (6)
on p. 143 and 144/ and if )..({'n) denotes the largest term of A an , also the
property
. Nan + 1)
hm \'(ii'") = 00.
1\1 n
Keeping to the notation and to all the operations in the proof of Theorem
IX. 1.1 , it suffices to make the following remarks: according to what we
have just said about the points of S: on 0'=0 (=O'c) (if the theorem is
supposed to be false), the function T(z) given by (lX.21) has on its circle
of convergence Izl = 1 only singular points each of which can be enclosed
by a Jordan curve inside the disk Izl <e- a and on which T(z) is holo-
morphic. But this time the sequence {m n} has a property which can be
expressed as follows: there exists a subsequence {m np } of {m n} with
Now by a theorem of Ostrowski, the domain of existence of a function

whose Taylor expansion has such gaps is necessarily simply connected.
This is a contradiction, and the theorem is proved. Theorem IX.l.4 is a
special case of Theorem IX.l.3.
IX.2. Relations between isolation of the fractional parts of the

exponents and the distribution of singularities
Here are some supplementary definitions. If Q={qn} and J = {Sn} are two
sequences of positive real numbers, the set J(Q) defined by
is called the neighbourhood J of Q. Thus,
Let D· be the upper density of A and let to be any real number. We

denote by D(A; to) the half-strip defined by
(IX.22) It - tal :( TTD·, a ~ - 3[4 - log D·]D·
if D·>O; D(A; to) is the half-line 1=/0' a~O, if D·=O.
Let S be an arbitrary domain and denote by S* the set composed of
all the points s+2kTTi, where s is any point of Sand k any integer. We
call S* the domain S modulo 2TTi.
If I is any real number, then e[l] denotes the set of all the numbers
1+ k, where k is any integer (hence e[l] is the class of numbers congruent
to I modulo 1).
A real number a is said to belong modulo 1 to a set A if there exists
an integer k such that a+k E A. If P is a sequence and Q c P, then the
set of elements of P belonging modulo 1 to .J[Q] is said to be isolated

modulo 1 by the neighbourhood.J of Q.
We prove the following theorem.
THEOREM IX.2.I. Suppose that the step of A is positive and that

f(s) = (A, A),
with aa=O. Let M={Anp} (cA={AnD, wherefor Anp=/Lp, anp=bp:
(IX.23) lim log Ibnl = O.

/Ln
Let .J={on} a positive sequence such that lim on=O and let A(t) (t~O)
be a positive non-decreasing function satisfying the condition (IX. 1). We
define the function h(a) (a> 0) by (IX.2). Suppose that the following condition
holds:
(IX.24) h(/Ln):::; COn (n ~ 1),
where c is a constant less than (2e)-1.
Then f has no analytic continuation into S* (the domain S modulo 21Ti),
where S is an open domain such that
(IX.25) S => D( A'; to),
to being any fixed real number, A' the set of terms An of A that belong
modulo 1 to .J[(M)], so as to satisfy in S* the condition
(IX.26) f(s) = O[A(ltl)] (It I ~ (0),
uniformly in a as a varies on a compact set.
Observe that this theorem is non-trivial only if for the upper density
D~ of A' we have D~ < 1. For if 1 :::; D~:::; e\ then S* covers, by the defini-
tion of D(A'; to) an entire half-plane P Ii. with d < 0, where f cannot be
continued ipso facto (the step of A being positive, aa is also the abscissa
of holomorphy off); but if D~ > e\ then S*, if it does not contain a half-
plane PIi. with d<O, itself forms part of the half-plane of the convergence
of J, and the question of analytic continuation into S* does not arise.
Intuitively this theorem can be expressed in the following fashion.
For a series f(s)=(A, A) (aa=O) to have an analytic continuation in a
half-strip It I :::;b, a> 3(4-log b)b modulo 21Ti and to grow not faster than
A(ltj), it is necessary that every neighbourhood iJ of the fractional parts
of a useful subsequence M of A (that is: MeA, lim (log Ibpl!ftp) =0)

which isolates modulo 1 as subsequence of A whose upper density does
not exceed b, should be sufficiently small with respect to the useful terms
(that is to say, (IX.24) should not hold).
We proceed to the proof of the theorem.
From M we can extract a sequence M' = {ft~} such that if we set
ft~ = ftmn (n ~ 1), .d' -- {o'}

n,
and denote by .d'[(M')] the set U «ft~)-o~, (ft~)+o~), we have

sup .d'[(M')] - inf .d'[(M')] ~ min (I, hl),
where hl is the step of A'.
The set
A~ = {Ani An E A, ClAn] () .d'[M']} "I (:)
is a subsequence of A'. It follows, by the definition of the sets D, that
D(Al; to) c D(A'; to). Consequently S::> D(Al; to). Therefore, if we
can prove the theorem for the case when M satisfies the additional con-
dition
(IX.27) sup .d[(M)] - inf .d[(M)] < min (1, hl),
where hl is the step of A', then the theorem is proved in the general case.
Consequently we are going to assume that (IX.27) holds.
We suppose that the series f(s)=(A, A), can be continued in S* and
can satisfy there the condition f(s)=O[A{!tj)] (It/-+oo) and we show
that this leads to a contradiction.
As in the proof of IX.l.l we construct the functions Fa(z) of V.4.1
with C(x)=logA(x) for X>Xo, where Xo is such that A(xo)~I, and
C(x)=O for 2 x~xo, the quantities p and q having been chosen as in the
proof of Theorem IX. I. I. We set a = ftn and write
tPn(z) = Fun(z) (n ~ I).
We define tP(z) by (IX.16), with an replaced by (ftn), the fractional part
of ftn' The inequalities (IX.13), (IX.14) and (IX.15) are now replaced,
respectively, by
ItPn(Z) I ~ e6nIYI-C(12J)~n(X)
(IX.28) tPn(u) ~ g e-<u n for lui ~ IOn

where ~n has the same meaning as in (IX. 13), g and I have the same values
as in (lX.14) and (IX.15), and 8~=h(f'n)' We can then verify that the
following relations hold:
(lX.30) J[(M)]::> support of $(u),
(IX.31) lim log <P(fLn) = O.

f'n
If we now set, as in the proof of Theorem IX. 1.1 , lJ'(0 = (e C-1) - \ we

can define lJ'(s) by the integral (IX.lS), where W has the meaning we have
just given to it and f is given in the statement. Since the value of c occur-
ring in (IX.lS) is still arbitrary positive, we see that lJ'(s) is holomorphic
for 0'>0 and bounded for 0'~8>0, where 8>0 is arbitrary.
There exists an R and an a such that if B denotes the half-strip
1t - to 1< 1TR, 0' > a, the following conditions hold:
S ::> E, B::> D(A" , t)

0, R > D~,
where D~ is the upper density of A' and O'o=inf 0' for s E D(A'; to).
Denoting by Dc the contour of Be defined by
we see, by applying the methods already used in this book, that for 0' > 0:
lJ'(s) = i Dc
f(z).p(s - z)W( -iz) dz.
If we set
0< 8 <R- D~
it is easy to see that when s varies in the half-strip B(e) defined by

It- tol <1T(R-e), O'>a+1Te, z varying on Dc, then .p(s-z) is a holomorphic
function of s and of z and that it is bounded. It follows that lJ'(s) is holo-
morphic in B*(e) U Pc + 28' Since e> 0 is arbitrarily small, we conclude
that lJ'(s) is holomorphic in B* U Pc.
Still imitating the argument we have used in proving Theorem IX.I.I
we obtain
Now by (IX.30) we know that «P(Am-n)¥:O only when Am E A', and by

(IX.27) to every Am E A' there corresponds a single n such that
Am - n ELI[(M)].
However, by the same inequality (IX.27), if for sufficiently large n there
exists a Am E A' such that Am-n ELI[(M)], then this Am is the only one
having this property. We denote by {vi} the sequence of such integers
n and let Amj be the term of A' corresponding to Vi (hence Amj - Vi ELI [(M)]).
There can only be finitely many integers n such that to each of them there
corresponds more than one Am with Am-n ELI[(M)]; to each of these
finitely many n there correspond only finitely many such Am.
Therefore 1JI can be written for a>O in the following form:
(IX.32) 1JI(s) = 2: qn e- ns + i L: amj«P(Amj -

nE:;k
Vi) e- VjS ,
where k is a non-negative finite integer. The series occurring in this

expression converges for a> 0, because 1JI(s) is holomorphic in this half-
plane.
For mi=np we have (Amj-Vj) = (Anp) = (/-tp). We also have amj=bp.
Now
where the functions $p are all non-negative; therefore
<PiO) A
(IX.33) -2 ~ (f>«/-tp».
p
So we see, by (IX.2S), (IX.31) and (IX.33), that for the Taylor-D
series
A(s) = L: amj«P(Amj - vi) e-VjS = L: d j e-VjS
the abscissa of convergence is zero.
But we have seen that the function 1JI(s) (hence A(s» can be continued
analytically in the half-strip
It - tol ~ 7TD~, a > -3(4 - log D~)D~,
where D~ is the upper density of A' = {A mj}, hence of {Vi}. This contradicts
Theorem VIo2.1 and thus proves our theorem.
IX.3. Variation of the exponents and the analytic continuation

Let A and M be two sequences of exponents, d a positive number, and
denote the set of all the differences dn,m = An - ttm, with An E A, ttm E M,
\An-ttm\ <d by (A-M)a. If vex) is a positive function for x~O, and Q
a subset of all the dn,m, then the set of dn,m contained in
u,...
ap,qE:ur (dM - v(dp,q), dp,q + v(dP,q»
is called the v-neighbourhood of Q and denoted by v(Q). The set of An
belonging to A such that there exists a ttm E M with dn,m E v(Q) is called
the A-component of v(Q).
If E is a set belonging to the half-plane u::;; 0, then the set of points that
cannot be joined, starting from the half-plane Po, by a Jordan curve
without passing through a point of E is called the envelope of E and is
denoted by E*; it is the complement of the largest domain belonging to
the complement CE that is connected with Po.
We recall that when p=(B, M) is holomorphic in a domain D => Po
and Q(t) is a positive function for t ~ 0, we say that p does not grow in D
faster than Q(t) if in the complement of the union of the disks C(s, e)
with s E CD, e>O arbitrary, we have p(s)=O[Q(I ti)] (\tl--'>- 00) uniformly
in u with u varying on a compact set (so far we have only used this defini-
tion when D was the complement of the singular set of p in a half-plane).
We also recall that if M' c M, we say that M' is a useful subsequence
of M with respect to B if for {ttnj} = M' we have
lim (log \bnjl/ttn) = U a,""
Let S be a closed set in the half-plane u::;;O and M' c M. We say that
S cuts the domain ~ modulo [M, M', Q] if for any domain D => Po for
which there exists a function p = (B, M) holomorphic in D that does
not grow in D faster than Q(t), and where M' is a useful subsequence of
M with respect to B, we have
It is evident that if S cuts ~ modulo [M, M', Q], then S also cuts modulo
[M, M', Q] every set ~y obtained from ~ by the translation iy (where y
is real). To see this it is sufficient to replace bn (where B={bn}) by bn e-1Y"n.
By the theorems stated in this section we shall see that if f=(A, A)
and rp=(B, M) have analytic continuations that are uniform, respectively,
in Dl and D 2 , where these domains contain the half-plane of convergence

(a>O for the two series in our statements), then the upper density of the
A-component of a sufficiently large neighbourhood v(A'-M')a (where
A' and M' are, respectively, useful subsequences of A and of M) deter-
mines an important relation between Dl and D 2 • Here the 'magnitude'
of this neighbourhood depends on the functions pet) and Q(t) which
determine the order of growth off and of q; in their respective domains.
In other words, the sequences A, M, A', M' and the functions pet),
Q(t) impose a 'dispersion law' on [CD l +CD2 ]* no matter what the
functions f=(A, A) and q;=(B, M) are, provided that f and q; can be
continued directly starting from Po into Dl and D 2 , respectively.
THEOREM IX.3.1. Let A and M be two sequences of positive step. Let D

be a domain with D ::> Po, and let PCt) and Q(t) (t~O) be two positive non-
decreasing functions, where log pet) is a concave function.
Suppose thatf=(A, A) does not grow in D faster than pet).
Let M' c M and let A' be a useful subsequence of A with respect to A.
Let O~v(x)~T<hM/2, where hM is the step of M, and let d>O.
If by setting C(t)=log [P(t)Q(t)] we have for infinitely many An and
P-m, with (An - P-m) E (A' - M')a:
f e»
An
C(t)
-2-
t
~ CV(An - P-m),
with 2ec< I, then CD cuts the half-strip

a> -3[4 - log (hD")]D',
modulo [m, M ', Q],' where hand D' are, respectively, the step and the upper
density of the A-component of (A' + M')a.
We do not give here a proof of this proposition. The reader should be
guided by the methods used to prove the other propositions of this chap-
ter (and to exploit the fact that log P(t) is concave in a fashion similar
to that in the proof of Theorem VA.I) and should then be in a position
to reconstruct the proof of Theorem IX.3.1.
NOTES
1 In this proof we also start with the assumption that an taO, and an easy change of
the argument enables us to pass to the case an t ".0.
2 We set C(x) =0 if A(x)~ 1 (x>O).
CHAPTER X
BIBLIOGRAPHICAL NOTES
The results of Chapter I are to a very large extent classical, for example,
all the theorems concerning various domains of convergence. We note,
however, that the upper mean density was introduced by the author in
[36], where there is also the first account of our method to calculate the
terms of the associated sequence; these values also come out from re-
search of Ostrowski [47], taken up by Bernstein [8].
All the results of Chapter II are due to the author. For the theorem of
11.1 we refer to [41], [42], [45]. But the case when all the A are distinct
modulo 1 had already been treated by the author in the older [31], [35].
We add that in our older papers the form itself of the domain (after
translating all the present results to the variable s) is much more restric-
tive than in the more recent papers and in this chapter.
11.2 is essentially devoted to the 'fundamental inequality' on the
coefficients of a Dirichlet series: Theorem 11.2.1. For series having an
abscissa of convergence (which is the case here) this inequality had al-
ready been presented by the author in [32] (under a slightly less general
form). It was then generalized by the author to series that need not
converge, a priori, anywhere: 'adherent series', and this has enabled us
to apply it to a great many different branches of analysis (problems of
moments, weighted approximation, generalized quasi-analyticity, etc.).
See [36] for the general result as well as for the rest of the bibliography
concerning this subject. Later results permitting the evaluation of the
coefficients of the Dirichlet polynomials tending to a function were ob-
tained by Schwartz [54]. Sunyer y Balaguer [55], [56], [57] and [58] has
made a deep generalization of the inequality of the author concerning the
adherent series in transferring it, with some new ideas on the' logarithmic
precision' (of the adherence), to Dirichlet polynomials: 'adherent poly-
nomials'. The results 11.2.2 and 11.2.3 (the more explicit case of 11.2.2),
which are both corollaries of the fundamental inequality, are utilized
a great deal in the rest of the book.
Theorem 111.1.1 is a consequence of the results of the first section of
11.1. This theorem likewise had been announced by the author some years
BIBLIOGRAPHICAL NOTES 157
earlier under a much less general form [31], [35]. In these papers it is
supposed that all the ,\ and the unit element are linearly independent.
Theorem 111.1.2, announced by the author in [45], is a generalization of
111.1.3 (which we call here a corollary to Theorem 111.1.2) proved by
Aronszajn [5], [6] who relied on our theorem in [31] and used the modular
function to pass from our theorem in [31] to his own. As in [31], his
domain is more restrictive than that of the present text.
Theorem 111.2.1, due to Ritt [52], is classical. Theorem 111.2.2 was
proved in a joint paper by the author and Gergen [30] in a less general
form. In the present form it was proved by the author in [36]. The same
remarks apply to 111.2.3, which was also proved by the author in [36].
So was 11.2.4. Theorem 111.2.5 concerning the families of functions tend-
ing to infinity is due to the author [29]. Theorem 111.2.6 is a special case
of 111.2.8, the Valiron-Bieberbach theorem [59]. The proof given here,
which is based on 111.2.5, appears for the first time in [36]. Theorems
111.2.7 and 111.2.9 are due to the author [36].
Chapter IV deals with Taylor series, essentially with Theorem IV.1.1,
published by the author in 1937 [33], and to some of its applications.
This Theorem IV.1.1 has been generalized by many authors, among them
Denjoy [15], Levy [23], Perron [48], Bouligand [12], von Mises [46],
Dvoretzky [16], [17], Bergman [7], to several variables, etc. Theorems
IV.1.5 and IV.I.6, which are both applications of Theorem IV. 1.1 , are
due to the author, the first published in [38], the second in [34]. Theorem
IV.I.6 was taken up and generalized by Valiron [60].
Theorems V.U, V.2.1, the subsequent lemmas, and Theorem V.2.5
are due to the author [27]. The integral (V.3) introduced by the author
has enabled Widder [61] (on the advice of the author) to establish a par-
ticular case of V.2.I. Theorem V.U was generalized by Bochner [10]
who states it for general almost periodic functions. V.2.7 is a particular
case of a theorem of Bernstein [8]. Theorem V.2.8 is due to the author
[27], [35]. The 'fictitious composition' of V.3 and the relevant Theorem
V.3.1 are due to the author [28]. Brunk [13] has proved some interesting
results on such a composition. Blambert has indicated conditions for a
•composite' point to be necessarily singular [9].
The functions Fa of Chapter V with their properties expressed by
V.4.1 were introduced by the author [39] (see also [44]). The version given
in V.4.1 is more precise than that of [39] and of [44]. The precise form of
the inequalities concerning these functions as well as their Fourier trans-
ISS DIRICHLET SERIES
forms, which is expressed by the relations (V.l6), (V.l7) and (V.lS) -

the last one an evident consequence of (V.l6) by the Paley-Wiener
theorem - play an essential role in Chapter IX, above all in the applica-
tions of V.4.l. The existence of a function F satisfying, together with its
Fourier transform, less precise inequalities have been demonstrated by
Levinson [22] and Agmon [2]. VA.2 is due to the author [39]. Theorem
VA.3 is due to Agmon (stated under a slightly different form). Theorem
VAA is also due to Agmon [2]. The results of V.5 (Theorems V.5.l,
V.5.2, and V.5.3) are due to the author [27], [35]. Theorem V.5.l is more
precise than the corresponding theorem in [35]. SchottHinder [53] has
made an interesting synthesis of the composition theorems, mainly for
Taylor series.
Theorems VI. 1.1 and VI. 1.2 have not been published elsewhere.
In contrast, an old theorem of the author [26] and another one given
(under a slightly less general form) by Aronszajn [6] can be regarded as
very closely related to Theorem Vll.3.
Theorem VI.2.l is due to the author [35]. This theorem contains as a
special case a theorem of Ostrowski [47] according to which the function
cannot remain holomorphic on a disk with centre on the axis of con-
vergence, of radius only slightly different from A(h, D').
Theorem VI.2.3 has not been published elsewhere.
The Theorems VI.3.1, VI.3.2, and VI.3.3 are due to the author [35].
For Taylor series (or rather Taylor-D, to stick to our notation) Theorems
Vl3.l and VI.3.2 were proved by P6lya [51]. But in P6lya's theorems
the maximum density occurs, a quantity not inferior to D' (which can,
however, be equal to infinity when D' is finite).
Theorems VI.4.1 is the classical theorem of Cramer (see [S]). Theorem
Vl4.2 is due to P6lya [49], [51] (Theorem of Cramer-P6Iya) and Theorem
Vl4.3 is due to Bernstein [S].
Theorem VIAA has not yet been published. It generalizes Theorem
VI.4.5, due to the author [24]. Theorem VI.4.5 has been generalized by
P6lya [50] and Jungen [20] who replace the word 'pole' by singular
algebraic-logarithmic point. Theorem VIA.6 is due to the author (for
N~ 2 see [25] for N = 1 see [43]). Theorem VIA.7 and its corollary VI.4.S
are due to the author [43].
The theorems ofVII.1 are due to Agmon [1], [3]. As to Theorem VII.2.1,
it forms a modification of a theorem of Agmon [4]. In the form in which
it is stated (apart from the conditions (VII.20) and (VIl2I) which are
BIBLIOGRAPHICAL NOTES 159
satisfied ipso facto, owing to the other conditions of the theorem, as we

have explained in the remark following this statement) it can be found
in a common paper by the author and Chandrasekharan [14]. The given
modification of the original theorem of Agmon appeared interesting for
the applications to the generalized Riemann equation set forth in Chapter
VIII. Incidentally, it is Theorem 111.2.3 that permits the passage from
Agmon's theorem to the theorem stated.
Chapter VIII finds its place in this monograph mainly to illustrate some
new applications of Theorem 111.2.3, which transfers Theorem 11.2.1 to
a strip.
Lemma VIII. 1. 1 is essentially due to Bochner and Chandrasekharan [11].
The form given here to this lemma is that of [21], which was communi-
cated to one of the authors of this paper by Chandrasekharan.
Theorem VIII.1.2 was published in a joint paper of the author and
Chandrasekharan [14]. It improves the conditions of a theorem of Boch-
ner and Chandrasekharan [11]. Theorem VIIL1.3 is also in [14], as well
as the corollaries following Theorem VIII. 1.3. We remark that the condi-
tions (VIII.l9) and (VIIL20) do not occur in the original paper, because
they are superfluous, just like the conditions (VII.20) and (VIL21) in
Theorems VIL2.l.
All the results of Chapter IX are due to the author. The account in this
monograph has less restrictive and simpler conditions than in the original
papers of the author. Thus, condition (IX.3), which occurs in Theorem
IX.l.l, is much simpler than the corresponding condition in [37]. The
same paper also contains Theorems IX.1.2, IX.I.3, and IX.1.4. Theorem
IX.2.1, which is certainly the most important in this chapter, was proved
under less general conditions in [39].
Theorem IX.3.1, for which we do not give a proof in the text, hoping
that the reader will easily be able to reconstruct it if he has had the patience
to read and to study the proofs of the preceding theorems of this chapter,
is published in a slightly more complicated form in [40]. This paper also
contains several other theorems on the same subject, as well as new applica-
tions to the generalized Riemann equation.
NOTE
ON THE DIRICHLET SERIES REPRESENTING

MEROMORPHIC FUNCTIONS
Letf=(A, A) and let ul<ua=ua.I' If S'EP"l' we define the class E,(s')

of S"l.1 or simply, class E(s'), if no ambiguity is to be feared, as the set
of all the points of S"lol that can be put in the form s' +2k1Ti, where k
is an integer.
If E(s') contains only poles, we say that this class is polar. If the orders
of the poles of E(s') have a finite maximum r, we say that E(s') is polar
of order r.
If a=s' +2k1Ti is a point of a polar class of order r, we denote by
Ai~), ... , A~k)(S'), respectively, the coefficients of
(s - a)-1, ... , (s - a)-r
of the principal part of this pole. Then A~k21(S') = ... = A~k)(S') = 0

if q < r, where a is a pole of order q. If for an integer k we have
s' +2k1Ti ¢. E(s'), we set A~k)(S')=O for 1 ~j~r. We set
log IA(k)(S') I
pes') = sup lim sup J •
! Ikl .... '" log Ikl
If f has in P"1 no singularities other than poles belonging to the classes
E(s(1», ... , E(s(m», ... and iff is order v in P"1' except at the singularities,
we say that f is meromorphic in P"l and that the polar order off in P"l
at infinity is p, defined by
p = sup (v, p(s(1», p(S(2», ... , p(s(m», ... ).
In what follows we assume that Ul <Ua=ua.I=O.
The following theorem holds: 1
THEOREM 1. (1) Suppose that f=(A, A) is uniform and meromorphic of

finite order in P"l at infinity, and that f has in this half-plane only finitely
many classes E(/l», ... , E(s(l). Suppose that the classes E(s(1), ... , E(/m»
(m<l) are situated on 00=0 and that rb"" r m, respectively, are their
orders. We set s(f)= U(!) +it(J), and Up=SUpj;'m u(j).
(2) Let p be the polar order at infinity off in P"l'
NOTE 161
We denote by c~g) the n-th Taylor coefficient off of (integral) order

g>p.
We set
N = L
'';'m
rj + g +1
if s = 0 is a pole off or a regular point off with f(O)"# 0; and we set

N = L r, + max (g + 1 -
,.;.m
h, 0)
iff is regular at s = 0 and this point is a zero of positive order h. Setting

c~)
I
D n •N = C()
... C~~N
C()
I
l+N ••• g
n +2N
we have
lim sup log IDn.NI ~ G'p.
n-+oo n
If all the classes E(s(1», E(s(l), ... , E(s(l) are situated on G' = 0 (m = I),
then
Proof. Let c> 0; we set

<p(s) = (es - 1)-\
and consider the integral
(1) F(z) = g!
-2
1T
i C
c-ioo
+ 100 dt (z = x + iy, s =
J(s)<p(z - s) g+l
S
G' + it)
for x> c. For these values of z:
(2) g''. L: L:
F(z) = -2
1Tl
i c~.q
ds
f(s)<p(z - s) Hl
s
g! fl11+s+loo ds
+ -2 .
1Tl 111 +s-ioo
f(s)<p(z - s)-;:IT'
S
the double sum being extended over 1 ~ q ~ I, where k is an arbitrary

integer and over q=O, k=O if E(O) is not a polar class; C~·o is the circle
lsi =e, C~·q, 1 ~q~l, is the circle Is-s<q)-2k1Til =e, where e>O is chosen
so small that for each of closed disks bounded by C~·q contains only the
point s(q)+2k7Ti if q~ 1, and only s=O if q=O, and where all these closed
disks are contained in P"1 + 8 and in the half-plane x < c - B.
If s(q) +2k7Ti does not belong, for a given k, to E(s(q), then
i ck.q
8
J(s)rp(z - s)
S
ds
g+l = o.
On the other hand, it is obvious that the function of z represented by
the last integral in (2) is holomorphic for x> 0'1 + B.
Thus, F(z) can be put for x>c in the form
g!
F(z) = -2.:z::z:
TTl
ic:. q
ds
J(s) ( Z-8 1) g+l
e - s + T(z),
where T(z) is holomorphic in the half-plane x> 0'1 +B.
Since by assumption the polar order ofJat infinity in P"l is less than g,
we see that the sum
Fq(z) = r J(s)rp(z -
L Jc~.q
\k\;.O
s)
s
~: 1
converges absolutely and uniformly on every compact set when x is suffi-

ciently large. If rq is the order of the polar class E(s(q), which is not the
class E(O) (if such a class exists I), we find that Fq(z) is a function that has
as singularities only the points s(q) + 2k7Ti in P"1' which are poles of order
at most rq. However, the points 2k7Ti can be poles of order rq+g + 1 if
E(O) is a polar class of order rq , or of order g+ 1 ifJ(s) is regular at s=O.
On the other hand, the integral in (1) can be represented for suffi-
ciently large x as (27T/g!) :z: c~g) e- nz •
Therefore this Taylor-D series represents a meromorphic function in
P"l which has there at most N poles.
By Theorem V.2.1 the real part of the places of these poles in the half-
planes 0' < 0 does not exceed O'p. It is now enough to apply Hadamard's
theorem on the polar singularities of a Taylor series to arrive at the re-
quired result.
THEOREM 2. Suppose that the sequence A satisfies the condition

IAn - nl ~ a < t (n ~ 1),
NOTE 163
that condition (1) of Theorem I holds, and that condition (2) of Theorem 1
is replaced by the following condition:
(3) f(s) does not grow in Pal' except at the singularities, faster than
exp c(l t I), where the non-decreasing function C(t) (t > 0) satisfies the
condition
[GO C(t) dt < 00

J1 t2 •
By setting
and writing
we have
lim sup log IDn.NI ~ Up.
n
To prove Theorem 2 we utilize instead of (1) the expression
[C+loo
F(z) = JC-IOO f(s)p(z - s)H(-is) ds,
where pes) is the function defined in the proof of Theorem 1 and where
H(z) is defined as in the proof of Theorem V.4.3, with C1(u) replaced by
C(u)+a log u; a > sup (p(P», ... , p(s(l)) and C2(u) == 0, {fLn} being
the sequence of positive integers. We imitate the proof of Theorem V.4.3
and that of Theorem 1 of this note and finish up again by the application
of Hadamard's theorem that was used in the proof of Theorem 1.
NOTE
1 This theorem was proved in a more restrictive form in [271.

BIBLIOGRAPHY
[1] S. Agmon, 'Sur les series de Dirichlet', Ann. £C. Norm. Sup. 66, 1949.
[2] S. Agmon, 'A Composition Theorem for Dirichlet Series', Journal d'Analyse
Math. 1, 1951.
[3] S. Agmon, 'Complex Variable Tauberians', Trans. Am. Math. Soc. 74, 1953.
[4] S. Agmon, 'On the Singularities of a Class of Dirichlet Series', Bull. Research
Counc. of Israel 3, 1954.
[5] N. Aronszajn, 'Sur les series de Dirichlet it exposants lineairement independants "
C.R. Acad. Sci. 199 (1934) 335.
[6] N. Aronszajn, same title, C.R. Acad. Sci. 199 (1934) 1564.
[7] S. Bergman, 'Essential Singularities of a Class of Linear Partial Differential
Equations in three Variables', J. Rat. Mech. Analysis 3, 1954.
[8] V. Bernstein, Le~ons sur les progres recents de la tMorie des series de Dirichlet,
Gauthier-Villars, Paris, 1933.
[9] M. Blambert, 'Un probleme de composition des singularites des series de
Dirichlet generales', Acta. Math. 89, 1953.
[10] S. Bochner, 'Hadamard's Theorem for Dirichlet Series', Annals of Math. 41,
1940.
[11] S. Bochner and K. Chandrasekharan, 'On Riemann's Functional Equation',
Annals of Math. 63, 1956.
[12] G. Bouligand, 'Sur la distance d'un point variable it un ensemble fixe', C.R.
Acad. Sci. 206, 1938.
[13] H. Brunk, 'Some Generalizations for Dirichlet Series of Hadamard's Theorem
with Applications', Thesis, The Rice Institute, 1944.
[14] K. Chandrasekharan and S. Mandelbrojt, 'On Riemann's Functional Equa-
tion', Annals of Math. 66, 1957.
[15] A. Denjoy, 'Etude sur la determination des singularites de la fonction analytique
definie par une serie de Taylor', Ann. Ec. Norm. Sup. 55, 1938.
[16] A. Dvoretzky, 'Sur les arguments des singularites des fonctions analytiques',
C.R. Acad. Sci. 205, 1937.
[17] A. Dvoretzky, 'Sur les singularites des fonctions analytiques', Bull. Soc. Math.
France 66, 1938.
[18] J. Hadamard, 'TMoreme sur les series entieres', Acta Math. 22, 1898.
[19] A. Hurwitz, 'Sur un tMoreme de M. Hadamard', C.R. Acad. Sci. 128, 1899.
[20] R. Jungen, 'Sur les series de Taylor n'ayant que des singularites algebrico-
logarithmiques sur Ie cercle de convergence', Comm. math. Helvet. 3,1931.
[21] J. P. Kahane and S. Mandelbrojt, 'Sur l'equation fonctionnelle de Riemann
et la formule sommatoire de Poisson', Ann. Ec. Norm. Sup. 75,1959.
[22] N. Levinson, 'Gap and Density Theorems', Amer. Math. Soc. Colloq. Publ.
26,1940.
[23] P. Levy, 'Surla variation du maximum d'une fonction', C.R. Acad. Sci. 206, 1938.
[24] S. Mandelbrojt, 'Sur les series de Taylor qui presentent des lacunes', Ann. Ec.
Norm. Sup. 40, 1923.
[25] S. Mandelbrojt, 'La recherche des points singuliers d'une fonction analytique
representee par une serie entiere', Jour. de Liouville 5, 1926.
BIBLIOGRAPHY 165
[26] S. Mandelbrojt, 'Sur une classe de series de Dirichlet', Prace Mat. Fiz. 36,1928.
[27] S. Mandelbrojt, 'Contribution it la theorie du prolongement analytique des
series de Dirichlet', Acta Math. 55, 1929.
[28] S. Mandelbrojt, 'Sur la recherche des points singuliers d'une serie de Dirichlet',
Bull. Soc. Math. de France 57, 1929.
[29] S. Mandelbrojt, 'Sur les suites de fonctions holomorphes, les suites correspon-
dantes de fonctions derivees. Fonctions entieres', Jour. de Liouville 8, 1929.
[30] S. Mandelbrojt and J. Gergen, 'On entire Functions Defined by Dirichlet
Series', Ann. Jour. Math. 53,1931.
[31] S. Mandelbrojt, 'Sur les series de Dirichlet dont les exposants possi:dent quel-
ques proprietes arithmetiques', Bull. Soc. Math. de France 60, 1932.
[32] S. Mandelbrojt, Series lacunaires, Actualites scientifiques et industrielles, No.
305, Hermann, Paris, 1936.
[33] S. Mandelbrojt, 'Theoreme general fournissant l'argument des points singuliers
situes sur Ie cercle de convergence', C.R. Acad. Sci. 204, 1937.
[34] S. Mandelbrojt, 'Applications d'un theoreme sur les arguments des singu-
larites', C.R. Acad. Sci. 206, 1938.
[35] S. Mandelbrojt, 'Dirichlet series', The Rice Institute Pamphlet 31, 1944.
[36] S. Mandelbrojt, Series adMrentes. Regularisation des suites. Applications,
Gauthier-Villars, Paris, 1952.
[37] S. Mandelbrojt, 'Influence des proprietes arithmetiques des exposants dans
une serie de Dirichlet', Ann. Ec. Norm. Sup. 71, 1955.
[38] S. Mandelbrojt, 'Sur les singularites d'une serie de Taylor sur son cercle de
convergence', Annales Univ. Scient. Budapest. Rolando Eotvos, Sectio Mathe-
matica nominatae 3-4, 1960/61.
[39] S. Mandelbrojt, 'Transformees de Fourier de fonctions entieres et series de
Dirichlet: un principe de dualite', Journal d'Analyse math. 10, 1962/1963.
[40] S. Mandelbrojt, 'Dirichlet series', Studies in Math. Analysis and related topics.
Essays in honor 0/ George Palya, Stanford University Press, 1962.
[41] S. Mandelbrojt, 'Considerations arithmetiques dans l'analyse harmonique',
Problemes contemporains dans la tMorie des /onctions analytiques, HAYKA,
Moscow, 1966.
[42] S. Mandelbrojt, 'Arithmetique dans la theorie des fonctions', C.R. Acad. Sci.
262,1966.
[43] S. Mandelbrojt, 'Considerations arithmetiques dans la theorie des fonctions
d'une variable complexe', C.R. Acad. Sci. 262, 1966.
[44] S. Mandelbrojt, Fonctions entieres et trans/ormies de Fourier. Applications,
The Mathematical Society of Japan, 1967.
[45] S. Mandelbrojt, 'Considerations arithmetiques dans la theorie des fonctions
analytiques', Israel Journal 0/ Math. 6, No.1, 1968.
[46] R. de Mises, 'La base geometrique du theoreme de M. Mandelbrojt sur les
points singuliers d'une fonction analytique', C.R. Acad. Sci. 205, 1937.
[47] A. Ostrowski, 'Einige Bemerkungen uber Singularitaten Taylorischer und
Dirichletscher Reihen', Berl. Sitz., 1921.
[48] O. Perron, 'Ober die Lage eines singularen Punktes auf dem Konvergenzkreis',
Sitzgsb. bayr. Akad. Wiss. Math.-naturwiss Kl., 1937.
[49] G. P6lya, 'Uber die Existenz unendlich vieler Singularen Punkte auf der
Konvergenzgerade gewisser Dirichletischer Reihen', Berl. Sitz., 1923.
[50] G. P6lya, 'Sur les singularites des series lacunaires', C.R. Acad. Sci. 184,
1927.
[51] G. P6lya, 'Untersuchungen uber Lucken und Singularitaten von Potenzreihen',
Math. Zeitschri/t 29, 1929.
[52] J. F. Ritt, 'On Certain Points in the Theory of Dirichlet Series', Am. Jour. of
Math. 50, 1928.
[53] S. Schottliinder, 'Der Hadamardsche Multiplikationssatz und weitere Kom-
positionssatze der Funktionnentheorie', Math. Nachr. 11, 1954.
[54] L. Schwartz, Etude des sommes d'exponentielles reelles. Actualites scientifiques
et industrielles No. 959, Paris, 1943.
[55] F. Sunyer y Balaguer, 'Sur des resultats de M. S. Mandelbrojt', C.R. Acad.
Sci. 231, 1950.
[56] F. Sunyer y Balaguer, 'Une generalisation de la precision logarithmique de
M. S. Mandelbrojt', C.R. Acad. Sci. 232, 1951.
[57] F. Sunyer y Balaguer, 'Aproximacion de funciones por sumas de exponen-
ciales', Collectanea Mathematica 5,1952.
[58] F. Sunyer y Balaguer, 'Approximation of Functions by Linear Combinations
of Exponentials', Seminario Matematico de Barcelona, Collectanea Mathe-
matica 17, 1965.
[59] G. Valiron, 'Fonctions entieres et fonctions meromorphes d'une variable',
Mem. Sciences Math. 11, 1925.
[60] G. Valiron, 'Sur les singularites des fonctions definies par les series de Taylor',
Publications math. Univ. Belgrad 6-7, 1937-1938.
[61] D. Widder, 'The Singularities of a Function Defined by a Dirichlet Series',
Am. Jour. Math. 49,1927.

S. Mandelbrojt (Auth.) - Dirichlet Series - Principles and Methods (1971, Springer Netherlands)

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

S. Mandelbrojt (Auth.) - Dirichlet Series - Principles and Methods (1971, Springer Netherlands)

Загружено:

Авторское право:

Доступные форматы

DIRICHLET SERIES

D. REIDEL PUBLISHING COMPANY I DORDRECHT-HOLLAND

ISBN-13: 978-94-010-3136-3 e-ISBN-13: 978-94-010-3134-9

All Rights Reserved

It is not our intention to present a treatise on Dirichlet series. This part

an improved form results contained in our Rice Institute Pamphlet, which

April 1969 s. MANDELBR01T

CHAPTER 1/ SEQUENCES OF EXPONENTS AND ASSOCIATED

CHAPTER II / INEQUALITIES CONCERNING THE COEFFICIENTS 18

CHAPTER III/ THEOREMS OF LIOUVILLE-WEIERSTRASS-PI CARD

CHAPTER IV I SINGULARITIES OF FUNCTIONS REPRESENTED

CHAPTER V / COMPOSITION OF SINGULARITIES 54

CHAPTER VI/SOME APPLICATIONS OF THE PRINCIPLES FOR

CHAPTER VII / ON THE BEHAVIOUR OF THE REMAINDERS OF

CHAPTER VIII / APPLICATIONS TO THE GENERALIZED RIE-

CHAPTER IX / INFLUENCE OF ARITHMETICAL PROPERTIES OF

CHAPTER X / BIBLIOGRAPHICAL NOTES 156

SEQUENCES OF EXPONENTS AND ASSOCIATED

1.1. Ascending sequences of positive numbers

is called the upper density l of A, and the quantity

h = lim inf (An +1 - An)

D' = lim sup D(x),

15(x) = ! ("lC D(y) dy (x > 0)

15' = lim sup 15(x)

is the upper mean density of J,

xD(x) = k~-l i:k+l D(y) dy +i: D(y) dy,

And since log xlx~ lIe for x>O, we can write

15(x) ~~ [J:n D(y) dy + n log (xl An)],

A(a) = a{4 + log [(a 2 - 1)/a3 ] + log [(a + 1)/(a - 1)]/a}.

lim sup An/An ~ [8.5 - 3 log (hD·)]D·.

If m is a positive integer and p is fixed, we set fLn = Ap + n (n ~ 1) and

where in a simplified notation with /Lm = /L:

log Am ~ f-m log [(/L/(/L + pX»2 - I] dx

+ LN-m log [I - (/L/(/L + pX»2] dx

= LN_-mm log [plxl(2/L + pX)/(/L + pX)2] dx

~ p-l j P(N-m) log [Itl/(/L + t)] dt

~ (N - l)[log p - log (/L + (N - m)p)]

+ p-l jP(N-m) log (/tl!/L) dt.

As Np=Nm(a/L)p~a/L, we have It I~a/L for It I~pN. Also

log [1 _(N ~ m)p] ~ (N ~ m)p,

which allows us to write (taking account of the inequality Np~alL):

+ (N - 1) log a - N(N - m)PIL- 1

where ~(x) is the density function of {lLn}.

log Bm ;>-: fIX) log (I - 1L2 /X 2 ) dNm{x)

= -log (I - a- 2)N + p,~'{ap,) log [(a - 1)/(a + 1)],

log Qm ;>-: [2 log (p~{ap,» - (2 +a - log a)]N

lim log (An +p/Mn)/lLn

[pn//Ln(1 + e)e] log [pn//Ln(1 + e)e] ~ (hD'/e) log (hD'/e),

r1m sup log Ao{r)

The uniform convergence of the product is a consequence of the fact

log Ao{r) = L log (1 + ~;) = f' log (1 + ~2) dN(x),

log Ao(r) = 2r 21'"o D(x)+ dxr

This enables us to write for x> 0:

log 11o(r)/r = -2r J: yD(y) d(y2 ~ r2)

1.2. General properties of convergence

In any case R is given by the formula

Moreover, it is known that the series converges uniformly on every com-

(1.5) u > lim sup log lanl = Uc

THEOREM 1.2.1. If Lan e- ilnso (so=uo+it o) converges, then the series

larg (s - so) I : :; y < 11'/2.

a' > 0, It'l/u':::;; tan y = M < co.

For s=so+s' and q>p~2 we have:

+ (Aq - S) e-hqs' - (A p - 1 - S) e-hps'.