Harmonic Analysis as found in Analytic

Number Theory

Hugh L. Montgomery
Department of Mathematics
University of Michigan
Ann Arbor, MI 48109-1109 USA
phone: 734-763-3269
[email protected]

ABSTRACT. A wide variety of questions of Harmonic Analysis arise naturally in var-

ious contexts of Analytic Number Theory; in what follows we consider a number of
examples of this type.
The author is grateful to Dr. Ulrike Vorhauer for advice and assistance at all stages
of preparation of this paper.

1. Uniform Distribution
The definition of uniform distribution is fairly intuitive:
DEFINITION 1. A sequence {Un} E T is uniformly distributed iffor any a, 0 ~ a < I,
we have
lim NI card{1
N.-;oo
~n~N : Un E [0, a) (mod I)} = a.
Let UN be a measure with unit masses at the points Un for I ~ n ~ N. Then the
Fourier transform of UN is the exponential sum
N
UN(k) = Le(-kun )
n=l

where e(O) = (This notation was introduced by I. M. Vinogradov.) H. Weyl [36,37]

e27ri8 •
introduced an important criterion for uniform distribution in terms of the size of the UN,
namely that the following are equivalent statements concerning a sequence {Un}:
Research supported in part by NSF Grant OMS 0070720.

271
J.S. Byrnes (ed.), Twentieth Century Harmonic Analysis - A Celebration, 271-293.
© 2001 Kluwer Academic Publishers.
272 HL Montgomery / Harmonic Analysis as found in Analytic Number Theory

(a) The sequence {un} is uniformly distributed;

(b) For each integer k I- 0, UN(k) = o(N) as N ---+ 00;
(c) If F is properly Riemann-integable on 'll' then

lim N1
N--+oo
LN
n=l
F(u n ) = 1 'll'
F(x) dx.

From the formula for the value of a geometric series it is immediate that

N 1 1
j ~ e(n8)j ~ Isin 71"81 ~ 211811
where 11811 is the distance from 8 to the nearest integer, 11811 = miDnEz 18 - nl. If 8 is
irrational and k is a non-zero integer then k8 is not an integer, and hence by the above with
8 replaced by k8 we see that UN(k) = Ok(1) when Un = n8. In this way we see easily
that the sequence n8 is uniformly distributed if 8 is irrational.
The proof of Weyl's Criterion depends on the existence of one-sided approximations to
the characteristic function XI of an interval by trigonometric polynomials; these approxima-
tions should be close in the Ll norm. Of course the existence of such trigonometric polyno-
mials follows easily from the uniform approximation to continuous functions by trigonomet-
ric polynomials, but it is also useful to put this in a quantitative form. Erdos and Tunm [9]
showed that there exist trigonometric polynomials T_ and T+ of degree at most K such
that T_(x) ~ XI(x) ~ T+(x) for all x and such that I'll'T± = a + 0(1/ K), and thus that

N K IU(k)1
j card{1 ~ n ~ N : Un E [0, a) (mod I)} - Naj ~ C K +CL-k-'
k=l

In the 1970's Selberg considered how the large sieve could be refined, and in doing so discov-
ered more natural functions that yield very sharp constants (see Selberg [32], pp. 213-226).
Indeed, Beurling [5] defined the entire function

( sin 71" Z)2 ( 2 ~ 1 ~ 1 )

B(z) = -71"- ; + ~ (z - n)2 - ~ (z + n)2 ,

and observed that B(z) = 0(e2"I~zl), that B(x) ~ sgn(x) for all real x, and that

l B(x) - sgn(x) dx = 1.

Indeed, Beurling showed that among functions with the prior properties, this function unique-
ly minimizes the above integral.
 HL Montgomery / Harmonic Analysis as found in Analytic Number Theory 273

-3 -2 -1

-0.5

FIGURE 1. Beurling's function B(x).

If I = [a, b] is an interval of the real line and 0 > 0 then the Selberg majorant and
minorant are

1 1
S+(z) = ZB(c5(z - a)) + ZB(c5(b - z)),
1 1
S - (z) = --B(o(a
2 - z)) - -B(c5(z
2 - b)) .

Then S_(x) ~ XI (x) ~ S+(x) for all real x, Set) = 0 for It I ? 0, and IJR S±(x) dx =
b - a ± 1/0. This approximation is optimal when c5(b - a) is an integer, and in any case
is quite good. To obtain corresponding approximations for the circle group we apply the
Poisson summation formula. Suppose that b - a < 1, so that the interval [a, b] defines
an arc of 1l'. We take 0 = K + 1 and set T±(x) = En S±(n + x). Then T± is a
trigonometric polynomial of degree not exceeding K, T_(x) ~ XI(x) ~ T+(x) for all x,
and 111' T±(x) dx = b - a ± l/(K + 1). This is optimal when (b - a)(K + 1) is an integer,
and is close to optimal in any case.
Weyl's criterion has since been vastly generalized to describe the weak convergence of a
sequence of measures J-tn to a limiting measure J-t in terms of the convergence of the Fourier
transforms of these measures. One fruitful generalization is to 1rt. Suppose that {un} is a
sequence of points in 1rt and let 13 = lab bd x ... x [ad, bd] denote a box in 1rt. Then the
following are equivalent:
(a) limN-+oo t card{l ~ n ~ N: Un E 13} = vol 13 for every box 13 in 1rt ;
(b) If k E Zd, k =f. 0, then EZ=l e(k . un) = o(N) as N -7 00;
(c) If F is properly Riemann-integrable on 1rt then limN-+oo t EZ=l F(u n) =
ITd F(a;) da;.
274 H.L Montgomery / Harmonic Analysis astound in Analytic Number Theory

Quantitative majorants in ']1"l are easily obtained by fonning a product of one-dimen-

sional majorants. Minorants are a little more elusive, but Barton, Vaaler and Montgomery [2]
have given a construction that works pretty well.
In the same way that we used Weyl's Criterion to see that the sequence {nO} is uniformly
distributed if 0 is irrational, we can use Weyl's Criterion for ']1"l to obtain a sharpening of
Kronecker's Theorem. Suppose that 1,01 " " , Od are linearly independent over the field Q
of rational numbers. Then the points nO = (nOl, n02, .. . , nOd) are not only dense in ']1"l
(Kronecker's Theorem) but are actually uniformly distributed. Indeed, suppose that 13 is a
box in 'II'" , and let T (z) be a trigonometric polynomial that minorizes X~ with JTd T > 0 .
Then
M+N
card{M + 1 ~ n ~ M + N : nO E 13} ~ L T(nO)
n=M+l
M+N
= L1'(k) L e(nk· 0)
Ie n=M+l

~ 1 ~ 11'(k) I
~ NT(O) - '2 ~ Ilk· 011'
Ie#O

Here the last sum is finite because there are only finitely many k for which 1'(k) =f:. 0 and
because k . 0 is never an integer. Since 1'(0) > 0, the above is positive if N is sufficiently
large. The remarkable feature here is that the expression above is independent of M . That is,
if nl < n2 < na < . .. are the n for which no E 13 then the gaps ~+1 - ni are uniformly
bounded above. This insight is critical to Bohr's definition of almost periodicity.
DEFINITION 2. Let f : lR --t C be continuous. We say that a real number t is an e
-almost period of f if If(x + t) - f(x)1 < e for all real x. The junction f is almost
periodic iffor every e > 0 there is a number C = C(e) such that any interval of length at
least C contains an e -almost period.
It is by the strengthened form of Kronecker's Theorem that we see that the almost periodic
polynomial
N
T(x) = L ane(An X)
n=l
is indeed an almost periodic function. It can also be shown that the almost periodic poly-
nomials are dense in the space of almost periodic functions. Let ((8) = E:=l n- S be the
Riemann zeta function. We write 8 = o+it. If a is fixed, a> 1, then ((a+it) is an almost
periodic function of t. The concept of almost periodicity can be generalized to other norms.
For example, when a is fixed, 1/2 < a < 1 the function ((a + it) is a mean-square almost
periodic function, even though it is not an almost periodic function in the uniform norm. One
 HL Montgomery / Harmonic Analysis asfound in Analytic Number Theory 275

might expect that by almost periodicity one could show that if the zeta function has a zero
with real part > 1/2 then it would have many such zeros, all with approximately the same
real part. However, attempts to prove such an assertion have so far been unsuccessful. If one
could prove such a thing then the Riemann Hypothesis would likely follow, since we have
fairly good upper bounds on the number of zeros of the zeta function with large real part. Let
N(a, T) denote the number of zeros of ((s) in the rectangle a ::::; a ::::; 1, 0::::; t ::::; T. Then
N(a, T) «: T.p(u) logT where </J(a) < 1 if a> 1/2. (Following Vinogradov, we say that
1 «: 9 if 1 = 0 (g). This notation saves a set of parentheses when there is no main term.)
Although we have not succeeded so far to use almost periodicity to produce many zeros from
one, we do know that translates of the zeta function are universal among non-zero analytic
functions, in the following sense: Let:R be a rectangle, :R = {s : al ::::; a ::::; a2, tl ::::; t ::::; t2}
with 1/2 < al < a2 < 1. Let 1 be analytic and non-zero on a domain containing:R. Then
for any c > 0 there exists a real number r such that I/(s) - ((s + ir)1 < c uniformly for
s E:R.
almost periodicity also arises in the error term in the Prime Number Theorem. Let A(n)
be von Mangoldt's lambda function, which is to say that A(n) = logp when n = pk for
some positive integer k, and A(n) = 0 otherwise. We put t/J(x) = En.:::!: A(n). By
integration by parts we see that the Prime Number Theorem in the form 1T(;j '" li(x) is
equivalentto the assertion that t/J(x) '" x. Since -('(s)/((s) = E:'=l A(n)n-· for a> 1,
we can recover t/J( x) from the logarithmic derivative of the zeta function by an inverse Mellin
transform:

-1
t/J(x) = -2.
1TZ
l c+ioo ('

c-ioo'"
x·
7(s)- ds.
S
By moving the contour to the left we see that this is

1
=x- L-p - -(O)+2
p
xP

(
L -k-·
(' 00

k=l
x- 2k

Here p runs over all the non-trivial zeros of the zeta function, which is to say all those zeros
with positive real part. This explicit formula for the error term in the Prime Number Theorem
is essentially one that Riemann stated and von Mangoldt later proved. Write p = f3 + h. In
the quantity x P = xfJXi'Y, the second factor oscillates, but more slowly as x increases. To
make it periodic we make an exponential change of variables. Suppose also that the Riemann
Hypothesis is true, which is to say that f3 = 1/2 for all p. Then

t/J( ell) - ell _ ' " ei'YlI

--'-~/2- - - ~ -
ell p
+ 0(1) .
P
276 H.L Montgomery / Harmonic Analysis as found in Analytic Number Theory

This expression is mean-square almost periodic, and the sum on the right is its Fourier ex-
pansion. It is generally believed that the 'Y > 0 are linearly independent over Q, so that the
terms ei-yy behave like independent random variables.

2. Exponential Sums
In his seminal work [36,37], Hermann Weyl not only gave his criterion for uniform distri-
bution, but also a useful method for estimating exponential sums of the form 'E:=l e(P(n))
where P is a polynomial with real coefficients. Indeed, today such an exponential sum is
called a Weyl sum. Weyl's basic observation was that

N N N
1Le(P(n))1
2

=L Le(P(m) - P(n))
n=l n=l m=l
N N-n
= L L e(P(n + h) - P(n))
n=lh=l-n

L
N-l

= L e(P(n + h) - P(n))
h=-N+1 l::;;n::;;N
l-h::;;n::;;N-h
N-l N-h
=N + 2~ L L e(P(n + h) - P(n)).
h=l n=l

This manipulation is known as 'Weyl differencing'. If P(x) has degree d then P(x + h) -
P( x) has degree d - 1 when h "I o. Hence if we perform this differencing d - 1 times
then we are left with a geometric series, which we know how to estimate. In this way, Weyl
showed that if P (x) = 'Et=o CLixi is a polynomial with real coefficients, and if at least one of
the numbers al, a2, ... ,ad is irrational, then the sequence {P( n)} is uniformly distributed
modulo 1.
In the Weyl differencing, it is somewhat a disadvantage that the parameter h runs all the
way from 1 to N - 1 . Later, van der Corput found that h can be restricted. For 1 ~ n ~ N
let Yn be a complex number, and suppose that Yn = 0 if n < 1 or n > N . Then
2 H 2
(H + 1)21 LYnl = 1LLYn+hl
n h=O n
H 2
= ILLYn+hl
n h=O
 H.L Montgomery / Harmonic Analysis as found in Analytic Number Theory 277

By Cauchy's inequality this is

H 2
~ (N + H) L
n
IL
h=O
Yn+hl
H H
= (N + H) L L L Yn+hYn+k
h=O k=O n

= (N + H)(H + 1) L IYnl 2
n
H

+ 2(N + H)!R L(H + 1 - r) L Yn+rYn .

r=l n

Thus we obtain van der Corput's inequality, which asserts that

N 2 N +H N 2 2(N + H) H h N-h _
ILYnl
n=l
~ H+1 LIYnl + H+1 L ( l - H+1)ILYn+ hYn l·
n=l h=l n=l

On taking Yn = e(ku n) and applying Weyl's Criterion twice, we obtain

THEOREM 1. (van der Corput) Iffor each positive integer h the sequence {Un+h - un}
is uniformly distributed (mod 1), then the sequence {Un} is uniformly distributed (mod 1).
More recently it has been recognized that the above remains true even when h is re-
stricted to lie in certain subsets of positive integers; we say that 1C is a van der Corput set
if the above is true when h is restricted to lie in 1C. This is equivalent to the existence of
non-negative cosine polynomials
T(x) = ao + L ah cos 27rhx
l~h~H
hE~

with T(O) = 1 and ao arbitrarily small. In this context it is no accident that we see the
coefficients of the Fejer kernel in van der Corput's inequality. Since the set of perfect squares
constitute a van der Corput set, there exist non-negative cosine polynomials of the form
H
T(x) = ao + LahCOs27rh2X
h=l

with T(O) = 1 and ao arbitrarily small, but it is not known how rapidly ao tends to 0 as
H -+ 00.
278 H.L Montgomery / Harmonic Analysis as found in Analytic Number Theory

van der Corput devised a general method for estimating exponential sums of the form
Eas:nS:b e (J (n)) where f is a sufficiently smooth real valued function. As an example of a
simple"'first result of this type, we mention that if 0 < A2 ~ f"(x) ~ AA2 for a ~ x ~ b
then
L e(J(n)) «A A~/2(b - a) + A;-1/2 .
a:;:;n:;:;b
In Process A of van der Corput, one exponential sum is made to give rise to another by a
suitable application of the van der Corput inequality. This method is destructive in the sense
that usually some cancellation is lost. In van der Corput's Process B, one takes r(x) =
e(J(x)) for a ~ x ~ b, and r(x) = 0 otherwise. Then by the Poisson summation formula,

L e(J(n)) = L r(n) = Lr(lI) = Lib e(J(x) - IIX) dx.

a:;:;n:;:;b n II II a

If f' is strictly increasing with f'(a) = a, f'(b) = f3, then for a ~ II ~ f3 we obtain a
stationary phase at XII where f'(x lI ) = II. If f"(x) > 0 for a ~ x ~ b then the above is
approximately
L e(J(x lI ) - IIX II + 1/8) .
a:;:;l1:;:;p v!f"(xll )
Thus the problem of estimating one sum is reduced to that of estimating another. Process B is
non-destructive, since a second application of it takes us back to our initial sum. For a certain
class of functions f , these two processes lead to estimates of the form

L e(J(n))« (max 1f'l)k(b - a)l

a:;:;n:;:;b

for certain pairs (k, e) in the square 0 ~ k ~ 1/2 ~ e ~ 1. If (k, e) is an exponent pair
then Process A gives the exponent pair (2k:2' k2t~+:} ) , and Process B gives the exponent pair
(e - 1/2, k + 1/2). It is trivial that (0,1) is an exponent pair. By Process B it follows that
(1/2,1/2) is an exponent pair, and by Process A this yields the further pair (1/6,2/3). The
collection of exponent pairs that can be obtained in this way is indicated in Figure 2 below.
Recently, Huxley [17], building on work of Bombieri and Iwaniec [6], has slightlyen-
larged the region of known exponent pairs, but we are still far from proving the conjecture
that (k, e) is an exponent pair if k > 0 and e > 1/2. This is a quite deep conjecture,
since the special case f(x) = tlogx yields the LindelOf Hypothesis, which asserts that
((1/2 + it) « to for every c > o. Some useful exponent pairs are given in Table 1.
Quantitative estimates can also be derived for the Weyl sum E:=l e(P(n)) in terms of
the rational approximations to the coefficients of P. For example, by Weyt's method we find
 H.L Montgomery / Harmonic Analysis as found in Analytic Number Theory 279

1.r-------------------------------------~

0.8

0.7

0.6

0.5 0.1 0.2 0.3 0.4 0.5

k

FIGURE 2. Exponent Pairs derived by van der Corput's processes.

Here the most favorable circumstance is when N ~ q ~ N d - 1 , in which case the upper
bound is of the order Nl-21-d+E . This is only slightly better than the trivial bound N if d
is large, and it falls far short of what we conjecture, which is that

N
Ee{F(n)) «-d Nl+ E
(1 + Ndq )l/d .
-
n=l q

Here the most favorable situation arises when q ~ N d/2 , and then the upper bound is of the
order N 1/2+E.
This leads to a non-trivial estimate for IS (a ) I.
280 HL Montgomery / Harmonic Analysis as found in Analytic Number Theory

TABLE 1. Some Exponent Pairs.

(k,l) Operation

(0,1)
(1/254,247/254) AAAAAAB
(1/126,20/21) AAAAAB
(1/86,161/172) AAAABAAB
(1/62,57/62) AAAAAB
(1/50,181/200) AAABABAAB
(1/42,25/28) AAABAAB
(2/53,181/212) AABABAAAB
(1/24,27/32) AABABAAB
(1/22,101/121) AABABABAAB
(1/20,33/40) AABAAB
(13/238,97/119) AABAABAAB
(11/186,25/31) AABAAAB
(4/49,75/98) ABABAAAB
(11/128,97/128) ABABAABAAB
(1/11,3/4) ABABAAB
(1/10,81/110) ABABABAAB
(13/106,75/106) ABAABAAB
(11/86,181/258) ABAABABAAB
(11/78,161/234) ABAAABAAB
(22/117,25/39) BABAAABAAB
(26/129,27/43) BABAABABAAB
(11/53,33/53) BABAABAAB
(13/55,3/5) BABABABAAB
(1/4,13/22) BABABAAB
(33/128,75/128) BABABAABAAB
(13/49,57/98) BABABAAAB
(19/62,52/93) BAABAAAB
(75/238,66/119) BAABAABAAB
(13/40,11/20) BAABAAB
(81/242,6/11) BAABABABAAB
(11/32,13/24) BAABABAAB
(75/212,57/106) BAABABAAAB
(11/28,11/21 ) BAAABAAB
(81/200,13/25) BAAABABAAB
(13/31,16/31) BAAAAB
(75/172,22/43) BAAAABAAB
(19/42,32/63) BAAAAAB
(60/127,64/127) BAAAAAAB
(1/2,1/2) B
 H.L. Montgomery / Harmonic Analysis as found in Analytic Number Theory 281

3. Power Sums
Although we spend a lot of effort to estimate exponential sums, in the opposite direction it is
sometimes useful to show that a cancelling sum is not always too cancelling. Turan's method
of power sums provides tools of exactly this sort. Since error terms in analytic number theory
are often expressed as a sum of oscillatory terms (as we have already seen in the case of the
error term in the Prime Number Theorem), Turan's method assists us in proving that such
error terms are sometimes large. To exemplify the method, we describe Turan's First Main
Theorem. Let
N

and suppose that IZnl ~ 1 for all n. Suppose that M is a given non-negative integer. We
wish to show that there is a v, M + 1 ~ v ~ M + N, such that ISIII is not too small
compared with Isol. To this end we employ a simple duality argument, which is typical of
Toran's method. Suppose that numbers all have been determined so that
N-l
(3.1) So = L a 8M+l+II'
11=0
Il

Then

By the definition of 8 11 we see that (3.1) asserts that

N N N-l

Lbn = Lbnz~+1 La"z~.

n=l n=l ,,=0

This identity certainly holds for arbitrary bn provided that

N-l
1 = zM+1
n L...J aII Z"n
~
11=0

for 1 ~ n ~ N. That is, P{z} = E~:r} a"z" should be a polynomial of degree at most
N - 1 that satisfies the N conditions P{zn} = Z;;M-l. Without loss of generality the Zn
are distinct, and hence P{z} is uniquely determined. It can be shown that E~:Ol lalll ~
",N-l (M+k)2k and thus we find that
L."k=O k '
282 HL Montgomery / Harmonic Analysis as found in Analytic Number Theory

where

c(M,N) = (~(M :k)2kr1

The constant here is best-possible, but it is disappointingly small, since it is only a little larger
than
( N )N-1
2e(M +N) .
Suppose that T(x) is an exponential polynomial of N terms and period 1, say
N
T(x) = 2)ne(Anx)
n=1
where the An are integers. Let I be a closed arc of the circle group '][' , and let L denote the
length of I. Then by Tunin's First Main Theorem, it is easy to show that

max IT(x)l ;;3 (~)N-1 max IT(x)l.

xEI 2e xET
Although the constant here depends on the number N of terms in T(x) , it is noteworthy
that it is independent of the size of the frequencies An. This inequality makes it possible to
give a simple and motivated proof of the Fabry Gap Theorem.
The small constant c(M, N) can be replaced by a larger constant if one is prepared to
allow v to run over a range longer than N. In more restricted situations the lower bound
can be very good indeed. For example, suppose that
N
(3.2) 8" = L e(vOn).
n=1

K: 1)1 , 1 K:
Then
K N N K
L (1-
11=1
8 2
= LLL (1-
m=1 n=1 11=1
l)e(v(Om - On)).
Since the expression is real, we may take real parts to see that the above is

=L LL (1- K: 1)
NNK NN1
cos 27rv(Om - On) = LL 2(~K+1(Om - On) -1)
m=1n=111=1 m=1n=1

where ~K+1(O) denotes the Fejer kerne1. Since ~K+1(O) ;;3 0 for all 0, and ~K+1(O) =
K + 1 , it follows that the above is
1 1 2
~-(K+1)N--N
".... 2 2 .
 H.L Montgomery / Harmonic Analysis as found in Analytic Number Theory 283

Thus if K ~ (1 + c)N then maxl~v~K Isvl ~ C(c).,fN, and in particular

max Isvl ~
l~v~2N
IN/2.

By working similarly with higher moments one can obtain still better lower bounds over
longer ranges of II: If Sv is given by (3.2) and 1 ::::; m ::::; N /2 then there is a II, 1 ::::; II ::::;
(12N/m)m such that Isvl ~ lvmN. It would be useful to have examples to show that this
is close to best-possible.
For a more extensive survey of Turan's method one may consult Montgomery [21], pp.
85-107. For a detailed account of the subject and important applications, one should see the
book of Turlin [33].

4. Irregularities of Distribution
We now consider how well-distributed N points can be. If the points are to fall in [0, 1] then
we could take Un = n/N. These points are extremely well-distributed in the sense that the
number of them in a subinterval [0, a] is N a + 0(1) . However, we find that it is not so easy
to distribute points well in 'f2 . Suppose that our points are 'Un = (Ul, U2) , let :R( a) be the
rectangle :R(a) = [0, al) x [0, a2], and let D(a) be the discrepancy function
D(a) = card{l ::::; n ::::; N: 'Un E :R(a)} - Nala2.
Roth [30] used a construction suggestive of wavelets to show that

r D(a?da;:}>
iT2
logN.

Since it is also possible to construct points that are this well-distributed, this solves the prob-
lem as to distribution in mean-square. Ostrowski [27] had observed that D(a) « 10gN
when 'Un = (n/ N, nv'2). The problem of showing that in any case liD 1100 ;:}> log N was
solved by Schmidt [31] by means of a complicated induction. However, a curious difference
arises here. Roth's argument generalizes easily to 1* to show that IIDII2 ;:}>k (log N)(k-l)/2
for any k > 1. However, Schmidt's approach has not been extended to k > 2. It has
been conjectured that IIDlloo ;:}>k (log N)k-l . This would be best possible, in view of con-
structions of Halton [14]. On the other hand, Pollington has recently mounted a wavelet
approach to this problem that has led him to conclude that one should be able to show that
IIDlloo ;:}>k (log N)k/2 , and that this sup norm need not be larger. Hence one should regard
the question of how large IIDlloo need be to be a wide open unsolved problem when k > 2.
Halasz [13] devised a variant of Roth's method that gives Schmidt's Theorem concerning
IIDlloo when k = 2, and also gives the lower bound IIDlll ;:}> v'Jog N when k = 2. This
°
is best possible, since Chen [8] has shown that if < p < 00 and k is given, k ~ 2, then
there exists a configuration of N points in 1* for which IIDllp «p,k (log N)(k-l)/2 .
284 H.L Montgomery / Harmonic Analysis as found in Analytic Number Theory

Roth obtained his lower bound by the Cauchy-Schwarz inequality,

hk D(o)F(o) do ~ (hk D(O)2 dO) 1/2 (hk F(o? dO) 1/2

where F is a test function defined to be a sum, F = E Fr , of more basic orthogonal
functions. Halasz similarly used the inequality

( D(o)F(o) do ~ IIDllooIIF1l1
i"r2
where F = nr(1 + eFr ). This bears a strong resemblance to a Riesz product, as occurs in
the theory of lacunary trigonometric series. For the lower bound of II Dill, he wrote

( D(o)F(o) do ~ IIDlhllFlloo
i"r2
where

F= g (1 + JRF
R .
r ).

Since -1 ~ Fr ~ 1, it follows that IFI ~ (1 + 1/ R)R «: 1 .

Let e(x1)' e(x2)' . .. be an infinite sequence of unimodular complex numbers, and put
PN(Z) = n:=l (z - e(xn)). Erdos asked whether it is possible to choose the Xn in such a
way that the numbers maxlzl~lIPN(z)1 are bounded as N --t 00. Note that ~logPN(e(x))
is just the discrepancy of the first N points, while the problem now being considered involves
the harmonic conjugate lRlog PN(e(x)) , but with the important difference that we need this
quantity to be large and positive, not just large in absolute value. That this sequence can not
remain bounded was first proved by Wagner [35], by means of a modified form of Schmidt's
method. Later Beck [3] used Halasz's modified form of Roth's method to obtain this in the
following sharp quantitative form: There is an absolute constant 8 > 0 such that
max IPN(Z) I > N 6
Izl~l

for infinitely many N.

Measuring the distribution of points in 'll'k relative to rectangles with sides parallel to the
coordinate axes is only one of many possibilities. If S is a measurable set then the quantity
D(S) = card{1 ~ n ~ N: 'Un E S} - NvolS
provides a measure of the distribution of the 'Un. But when we consider S we would also
include its translates, so we put d( 0) = D( S + 0) . Then d( 0) = 0 , but for k =f. 0 we have
d(k)
-
= Xs (-k)UN(k)
-
where UN(k)
- N
= En=l e( -k . 'Un) . Hence by Parseval's identity,

{ d(ofdo =L Ixs(kWIUN(kW.
i"rk k#O
 H.L Montgomery / Harmonic Analysis as found in Analytic Number Theory 285

The rate that Xs (k) tends to 0 as k tends to infinity in a particular direction depends on
how much of the boundary of S is orthogonal to the direction in question. Thus in the case
of a rectangle the Fourier coefficients decay slowly in the direction of the coordinate axes,
but comparatively rapidly in other directions. Thus one may expect that points may be found
so that fjN (k) is small when Xs (k) is large, and vice versa. By using the Fejer kernel as in
our discussion of power sums we can show that

Ikll~Xl
Ik21~X2
k;to

for any positive real numbers Xl> X 2 • In this way we can recover Roth's lower bound for
IIDII2. By averaging over disks (and averaging over the radius as well) we find that there
is a disk 1> for which D(1)) » N 1/ 4 • More generally, if we start with a set S, and are
allowed to shrink, translate, and rotate S, then we obtain this larger order of magnitude, in
view of a general principle governing the mean square decay of the Fourier transform of the
characteristic function of a set: If e is a simple, closed, piecewise a1 curve in R.2 , and S is
its interior, then
r
Jltl~R S
Ix (t)1 2 dt
211" R
rv ~
as R -+ 00. This is due to Montgomery [19], [20] pp. 114-119; see also Herz [15,16].

5. The Large Sieve

The large sieve was originated by Linnik [18] in a somewhat obscure form. It gained new life
in the hands of Renyi [28], who viewed it as a statement about almost independent vectors.
Today we usually think of this as an extension of Bessel's inequality for vectors in an inner
product space: Let ifJl> ... , ifJR be arbitrary vectors in an inner product space. Then the
following three assertions concerning the constant a are equivalent:
(a) For any vector ~ in the space,
R
L: I(~, ifJrW ~ all~112;
r=l
(b) For any complex numbers Ur we have
R
I L: UrU.(ifJr,ifJ.)1 ~ aL:Ic,.12;
l~r,.~R r=l

(c) a = P([(ifJr, ifJ.)]) .

286 HL Montgomery / Harmonic Analysis as found in Analytic Number Theory

(Here p(A) denotes the spectral radius of the matrix A.) Renyi took the coordinates of his
vectors to depend on arithmetic progressions or on Dirichlet characters, but the vectors he
obtained in doing so were not very close to orthogonal, so the estimates he obtained were
imperfect. Roth [29] had the excellent idea of taking tPr = (e(na r )) where the a r are well-
spaced in l' . These vectors are quite close to being orthogonal, and we now think of the most
basic form of the large sieve as being a statement about the mean square of a trigonometric
polynomial at well-spaced points. That is, we take
M+N
8(a) = E ane(na),
n=M+1
and we consider inequalities of the shape
R M+N
E18(a W~ A E lanl 2 •
r
r=l n=M+1
We suppose that liar - as II ~ 6 for r f. 8, and want A to depend on N and 6. If
an =e(-na1) then 8(a1) = N, and thus we must have A ~ N. By averaging over
translations of the a r we can also show that A ~ 6- 1 - 1. We find that A does not have to
be much larger than is required by these considerations.
Gallagher [10] used an inequality of the Sobolev type,

11Ot+6/2 11Ot+O/2
I/(a)1 ~ 8 I/(x)1 dx + 2 1/'(x)1 dx,
Ot-O/2 Ot-O /2
to show that one can take A = 1/6 + 1fN. This is the best constant with respect to 6,
but in arithmetic settings the coefficient of N is more important. The main advantage of
Gallagher's approach is that it generalizes readily to other families of functions. To obtain
good dependence on N we note that by duality the stated inequality is equivalent to the
inequality
M+N R 2 R
E 1 EYre(nar)1 ~AEIYrI2.
n=M+1 r=l r=l
On the left hand side we square out and take the sum over n inside. The diagonal terms give
N Er IYrI2, so it remains to consider the non-diagonal terms. This brings us to Hilbert's
Inequality, which asserts that

1E :::s81 ~
r,8
1f E IYrI2.
r
r#s
 HL Montgomery / Harmonic Analysis asfound in Analytic Number Theory 287

Montgomery and Vaughan [23], with some assistance from Selberg, found that this can be
generalized, so that

IL T,8
AY~aAI ~
r S
i L IYrl2
r
r¢s
provided that IAr - A81 ~ 8 whenever r =I s. Moreover for the circle group we have,
correspondingly, the inequality

I"~ sin7r(aYrYs r -
I ~ ~ "I
as) "" 8 ~ Yr
12
r¢a
where liar - aall ~ 8 for r =I s. This gives the large sieve with the factor ~ = N + 1/8.
With a little more care one can obtain ~ = N + 1/8 - 1 , and with this constant there are
situations in which equality can occur.
In arithmetic situations the a r are usually taken to be the Farey fractions a/q, in which
(a,q) = 1 and q ~ Q. Since lIa/q - a'/q'll ~ l/(qq') ~ 1/Q2 when a/q and a'/q' are
distinct modulo 1, we find that
Q M+N
L L IS(a/q)1 2 ~ (N + Q2) L lan l2 •
q=l a=l n=M+l
(a,q)=l

The generalized Hilbert Inequality can also be established in a weighted form, in which
we find that
IL YrYs ~ ~7r
Ar - As
I
IYrl2
2
L 8r
T,B r
r¢s
where IAr - As I ~ 8r when s =I r. Here the constant !7r is certainly not best possible. The
above also has a counterpart for the circle group, and hence we have a weighted form of the
large sieve,
~ IS(ar)1 2 ~ ~ la 12
LN+..L"'-L n'
r=l 26r n=M+l
Many other variants of the large sieve have been derived, involving, for example, maxi-
mal partial sums (via the Carleson-Hunt Theorem), or the Hardy-Littlewood Maximal Theo-
rem. As for further generalizations of Hilbert's Inequality, Montgomery and Vaaler [22] have
shown that if Pr = f3r + i'Yr with f3r ~ 0 for all r and l'Yr - 'Y.I ~ 8r for s =I r, then

IL Yr+Ys I ~ 84 L 1~12 .
r,s Pr Ps r r
r¢.
288 H.L Montgomery / Harmonic Analysis as found in Analytic Number Theory

The proof depends on the theory of H2 functions that are analytic in a half-plane.

6. Dirichlet Series
Let D(s) = L::=l ann- s • By Hilbert's Inequality we see that
T N
fo ID(itWdt = (T + O(N))?; lan l2 ,
and the weighted Hilbert Inequality gives

?; la l (T + O(n)).
T N
fo ID(itW dt = n 2

Thus we have some limitation on the amount of time that ID I is large. Suppose that Ian I ~ 1
for all n. Then by applying the above to D2 we find that

foT ID(it)1 4dt« (T + N 2)N2+o.

It would be very useful if we could interpolate between these estimates, in the sense that

foT ID(itW dt« (T + NQ/2)Nq/2+o

for real q, 2 ~ q ~ 4. Indeed, this implies the Density Hypothesis concerning the Riemann
zeta function. Many years ago, Hardy and Littlewood had conjectured that if lj(k)1 ~ F(k)
for all k and q ~ 2 then IIfllQ «Q IlFllq, and it can be shown that this majorant conjecture
would imply the conjecture above. However, Bachelis [1] used a method of Katznelson to
show that this is true only when q is an even integer (in which case it holds with constant 1).
In any case there is more that can be said about the number of times a Dirichlet polynomial
can be large than follows from moment estimates. For example, if 0 ~ h < t2 < ... < t R ~
T and tr+l - tr ~ 1 for all r, if ID(itr ) I ~ V for all r, and if lanl ~ 1 for all n, then
it is known that R « N 2 V- 2To provided that V 2 > NTl/2+o. It has been conjectured that
this estimate for R holds when the last condition is weakened to read V 2 ~ NTo . However,
Bourgain [7] has shown that if this is so then every Kakeya set in R,d , d ~ 2, has Hausdorff
dimension d. Thus there are those that doubt such a strong conjecture.

7. Spectral Characteristics of Zeros of the Zeta Function

Let h(d) denote the class number of the quadratic number field with discriminant d. In an
effort to derive a useful lower bound for h( d) when d is negative, it was recognized that
it would suffice to have a good supply of pairs of zeros of the Riemann zeta function that
are < Cl~:t apart where C < 1/2. With this motivation, an attempt was made in 1971 to
determine the distribution of 'Y - 'Y' as 'Y and 'Y' run over nearby ordinates of zeros of the
 H.L Montgomery / Harmonic Analysis as found in Analytic Number Theory 289

zeta function. One begins with a generalization of the explicit fonnula noted earlier. We
observe that

L
~
-1
A(n)n- S = -2.
lrZ
l c+ioo ('

c-ioo
7(8 + w)- dw
'>
XW
W
n::::;x

(' x1-s xp-s 00 x-2n- s

=--(s)+--L-+L--'
( 1- 8 P P- 8 n=1 2n + 8

We assume the Riemann Hypothesis and combine two such explicit fonnulre to see that

2 '"'
L...J Xi"! = -x -1/2 ( 'L...JA(n)
"' (X)-1/2+it
- + ""'
L...JA(n) (X)3/2+it)
-
'Y
1+(t-')')2 ~
n:::::x
n n>x
n
+ x-l+it ( logt + 0(1)) + 0(X 1/ 2/t) .

We take the modulus squared of both sides, and integrate over 0 ~ t ~ T. We write x = TOt
for 0 ~ 0: ~ 1, and note that the expression on the right hand side can be asymptotically
evaluated by using our mean square estimate for Dirichlet polynomials. Let

F(o:) = (~ logT) -1 L TiOt("!-"!')wb - ')")

O<,,!~T
O<"!'~T

where w(u) = 4/(4 + u2 ). Then we find that F is real, even, non-negative, and F(o:) =
(1 + o(I))T- 2Ot log T + 0: + 0(1) as T ~ 00, uniformly for 0 ~ 0: ~ 1. When 0: > 1
the method fails because the Dirichlet polynomial is too long. But only the tenns near the
diagonal contribute, and one can use the Hardy-Littlewood quantitative fonn of the Twin
Prime Conjecture to estimate that those tenns contribute. In this way one is led to guess that
F(o:) = 1+0(1) unifonnlyfor 1 ~ 0: ~ A,forany A > 1. This is known as the Strong Pair
Correlation Conjecture. By taking Fourier transfonns, we are led to a conjecture concerning
the distribution of the frequencies ')' - ')": The number of pairs ,)" i of ordinates of zeros,
o ~ ')' ~ T, 0 ~ ')" ~ T , for which 211'0: flog T ~ ')' - i ~ 211'f3 /log T is asymptotic to
(~+ laP 1- (Si::Ur dU) ~ logT.
Here 8 = 1 if 0 E [0:, f31, and 8 = 0 otherwise. That is, 8 is a Dirac point mass at O.
This arises because of the possibility that ')' = i when 0: < 0 < f3. This is the Weak
Pair Correlation Conjecture. Freeman Dyson observed that the density function here is that
of a random hermitian matrix of unitary type, and thus we take the above, although only a
conjecture, as evidence that the zeros of the zeta function are spectral in nature.
290 H.L. Montgomery / Harmonic Analysis as found in Analytic Number Theory

Goldston and Montgomery [12] showed that if the Riemann Hypothesis is true then the
Strong Pair Correlation Conjecture is equivalent to the estimate

l X
('I/J(x + h) - 'I/J(x) - h? dx '" hX 10gXjh

for x· ~ h ~ x 1- • • A heuristic argument in favor of this can be obtained by expanding out

and using the Hardy-Littlewood Conjecture concerning the number of d -twin primes not
exceeding x.
Recently Montgomery and Soundararajan [26] considered the higher moments

On expanding, one encounters enumerations of prime k-tuples. The Hardy-Littlewood Prime

k -tuple Conjecture asserts that if d1 , d2 , ••• dk are distinct integers then
k
E II A(n + d i) = 6('D)X + E(X, 'D)
n:!l;X ;=1

where 6('D) is the 'singular series'

k k
6('D) = E II Il-(q~) E e(E ai~i).
I:!l;Qi<OO ;=1 ¢(q.) l:!l;ai:!l;Qi ;=1 q.
(ai,Qi)=1
'Eai/QiEZ

Gallagher [11] considered the moments Il-k(X, h) for smaller h of the form h = clog X ,
and for that purpose showed that

E 6('D) '" Xk
O:!l;di:!l;X
d; distinct
(l:!l;i:!l;k)

as X ~ 00. For larger h, the mean value h is much larger than the usual size of the
difference 'I/J(x + h) - 'I/J(x) - h, and so it is useful to consider the arithmetic function
Ao(n) = A(n) - 1, whose mean value is asymptotically zero. For this function we have an
alternative formulation of the prime k-tuple Conjecture, which asserts that
k
E II Ao(n + d i) = 6 o('D)X + Eo (X, 'D)
n~X ;=1
 HL Montgomery / Harmonic Analysis as found in Analytic Number Theory 291

IT Jt(q~)
where now
6 0 (1)) = L L e(t ai~i).
l<q.<oo i=l t/>(q.) l",a.",q. i=l q.
(l"'i",k) (a.,q.)=l
L.a;jq.EZ
Thus 6 0 is the same as 6 except that the possibility that qi = 1 is now excluded. The mean
value of 6 0 is of course smaller, and hence more difficult to determine, but by elaborating
on work of Montgomery and Vaughan [24] concerning the distribution of reduced residues
modulo q in short intervals it can be shown that

L 6 0 (1)) = { (k/2~;2k/2 (-X logX)k/2 + O(Xk/2(logX)k/4) if k is even,

O",d;",X
d; distinct
°(X k / 2 - 1/(7k)+e) if k is odd.
(l",i",k)

This is established unconditionally; when combined with plausible hypotheses concerning

the size and behavior of the error terms E(X, h) we are led to expect that
Jtk(X, h) = (Ck + o(1))Xhk/2(logX/h)k/2
where the Ck are the moments of the normalized normal variable,

k' if k is even,
Ck = { (k/20)!2k/2
if k is odd.
Since these moments occur uniquely in the case of normal distribution, we are led to expect
that the distribution of 'IjJ(x + h) - 'IjJ(x) , for 0 ~ x ~ X, is approximately normal with
mean h and variance h log X / h .
Suppose that X > T. In terms of zeros of the zeta function, the Strong Pair Correlation
Conjecture seems to be telling us that the mean square size of the sum
L cos"Ylogx (0 ~ x ~ X)
Oq",T

is the same as if the terms were uncorrelated random variables. (We recall from the theory
of probability that if Xi are uncorrelated variables then Var (E Xi) = E Var(Xi ).) It
seems that our new speculations concerning the Jtk can similarly be interpreted as asserting
that the above sum is distributed as if the terms are independent random variables (as in the
Central Limit Theorem). How this relates to the spectral nature of the zeros, or any possible
underlying operators remains to be seen.
292 H.L Montgomery / Harmonic Analysis as found in Analytic Number Theory

References
[1] G.F. Bachelis, On the upper and lower majorant properties in U(G) , Quart. J. Math. (Oxford) (2) 24
(1973),119-128.
[2] J. T. Barton, H. L. Montgomery and J. D. Vaaler, Note on a Diophantine inequality in several variables, to
appear.
[3] J. Beck, The modulus of polynomials with zeros on the unit circle: a problem of Erdos, Ann. of Math. (2)
134 (1991), 609-651.
[4] J. Beck and W. W. L. Chen, Irregularities ofdistribution, Cambridge Tract 89, Cambridge University Press,
Cambridge, 1987.
[5] a
A. Beurling, Sur les integrales de Fourier absolument convergentes et leur application une transforma-
tionfonctionelle, Neuvieme congres des mathematiciens scandinaves, Helsingfors, 1938
[6] E. Bombieri and H.Iwaniec, On the order of «( ~ + it) , Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) 13
(1986),449-472.
[7] J. Bourgain, On the distribution of Dirichlet sums, J. Anal. Math. 60 (1993), 21-32.
[8] W. W. L. Chen, On irregularities of distribution, I, Mathematika 27 (1980), 153-170.
[9] P. Erd6s and P. Tunin, On a problem in the theory of uniform distribution I, Nederl. Akad. Wetensch. Proc.
51 (1948), 1146-1154; II, 1262-1269, (= Indag. Math. 10, 370-378; 406-413).
[10] P. X. Gallagher, The large sieve, Mathematika 14 (1967), 14-20.
[11] ,On the distribution ofprimes in short intervals, Mathematika 23 (1976), 4-9.
[12] D. A. Goldston and H. L. Montgomery, On pair correlations of zeros and primes in short intervals, An-
alytic number theory and Diophantine problems (Stillwater, 1984), Birkhiiuser Verlag, Boston-Basel-
Berlin, 1987, 183-203.
[13] G. Halasz, On Roth's method in the theory of irregularities of point distributions, Recent Progress in
Analytic Number Theory (Durham, 1979), Vol. 2, Academic Press, London, 1981,79-84.
[14] J. H. Halton, On the efficiency of certain quasirandom sequences of points in evaluating multidimensional
integrals, Num. Math. 2 (1960), 84-90.
[15] C. S. Herz, Fourier transforms related to convex sets, Ann. of Math. (2) 75 (1961),81-92.
[16] C. S. Herz, On the number of lattice points in a convex set, Amer. J. Math. 84 (1962),126-133.
[17] M. N. Huxley, Area, Lattice Points and Exponential Sums, Clarendon Press, Oxford, 1996.
[18] Ju. V. Linnik, The large sieve, Dokl. Akad. Nauk SSSR 30 (1941), 292-294.
[19] H. L. Montgomery, The analytic principle of the large sieve, Bull. Amer. Math. Soc. 84 (1978), 547-567.
[20] H. L. Montgomery, Irregularities of distribution, Congress of Number Theory (Zarautz, 1984), Universi-
dad del Pais Vasco, Bilbao, 1989,11-27.
[21] H. L. Montgomery, Ten lectures on the interface between analytic number theory and harmonic analysis,
CBMS No. 84, Amer. Math. Soc., Providence, 1994.
[22] H. L. Montgomery and J. D. Vaaler, A further generalization of Hilbert's inequality, Mathematika 45
(1999),35-39.
[23] H. L. Montgomery and R. C. Vaughan, Hilbert's inequality, J. London Math. Soc. (2) 8 (1974),73-82.
[24] , On the distribution of reduced residues, Annals of Math. 123 (1986), 311-333.
[25] H. L. Montgomery and K. Soundararajan, Beyond pair correlation, to appear.
[26] H. L. Montgomery and K. Soundararajan, Primes in short intervals, to appear.
[27] A. Ostrowski, Bemerkungen zur Theorie der Diophantischen Approximationen I, Abh. Math. Sem. Ham-
burg 1 (1922), 77-98; II, 1 (1922),250-251; III, 41 (1926), 224.
[28] A. Renyi, Un nouveau tMoreme concernant les fonctions independantes et ses applications a la tMorie
des nombres, 1. Math. Pures AppJ. 28 (1949),137-149.
[29] K. F. Roth, On irregularities of distribution, Mathematika 1 (1954), 73-79.
 H.L Montgomery / Harmonic Analysis as found in Analytic Number Theory 293

[30] K. F. Roth, On the large sieves of Unnik and Renyi, Mathematika 12 (1965), 1-9.
[31] W. M. Schmidt, Irregularities of distribution, VII, Acta Arith. 21 (1972), 49-50.
[32] A. Selberg, Collected Papers, Volume n, Springer-Verlag, Berlin, 1991.
[33] P. Thran, On a new method of analysis and its applications, Wiley-Interscience, New York, 1984.
[34] J. D. Vaaler, Some extremal functions in Fourier analysis, Bull. Amer. Math. Soc. 12 (1985), 183-216.
[35] G. S. Wagner, On a problem of ErdlJs in Diophantine approximation, Bull. London Math. Soc. 12 (1980),
81-88.
[36] H. Weyl, Ober ein Problem aus dem Gebiete der diophantischen Approximationen, Nachr. Ges. Wiss.
Gottingen, Math.-phys. Kl. (1914), 234-244; Gesammelte Abhandlungen, Band I, Springer-Verlag,
Berlin-Heidelberg-New York, 1968,487-497.
[37] Oberdie Gleichverteilung von Zahlen mod. Eins, Math. Ann. 77 (1916), 313-352; Gesammelte Abhand-
lungen, Band I, Springer-Verlag, Berlin-Heidelberg-New York, 1968,563-599; Selecta Hermann Weyl,
BirkhiiuserVerlag, Basel-Stuttgart,1956,lll-147.