ON THE SET OF POINTS OF CONVERGENCE

OF A LACUNARY TRIGONOMETRIC SERIES

AND THE EQUIDISTRIBUTION PROPERTIES
OF RELATED SEQUENCES
By P . ERDÖS and S. J. TAYLOR
[Received 3 May 1957 .-Read 16 May 1957]

1 . Introduction
WE consider the convergence of the series

sin(n k x+1t k ), (1)

k=1
where {µk} ( k = 1, 2, . . .) is a sequence of constants satisfying 0 < µk < 2zr
and {n k } ( k = 1, 2, . . .) is an increasing sequence of integers satisfying

tk = nk+1 > p > 1 . (2)

nk
It is well known from the classical theory of trigonometric series that the
series (1) cannot converge except possibly for values of x in a set of zero
Lebesgue measure . Our object is to discover how `thin' this set is for various
types of sequence {n k } . The first result of this kind is due to P . Turan who
proved, in 1941, that the series (1) converges absolutely in a set of positive
logarithmic capacity in the case
nk = (k!) 2 , Pk = 0 (k = 1, 2, . . .) . (3)
It will follow from the results of the present paper that in the case (3) con-
sidered by Turan, the set of values of x for which (1) converges absolutely
has dimension 2, whereas (1) converges in a set of dimension 1 .
The convergence, or absolute convergence, of the lacunary trigonometric
series is intimately related to that of the series
I {((nkx))-cxk},
k=1
(4)

where {a k } ( k = 1, 2, . . .) is a sequence of real numbers satisfying 0 < ak < 1,

{n k} satisfies (2) and, for a positive real number ~,

((S)) = fl-[a]
denotes the non-integer part of P. We will consider the set of values of x
which make (4) convergent or absolutely convergent concurrently with the
corresponding sets for the series (1) . Absolute convergence is considered
in § 2, and convergence in § 3 .
Proc. London Math. Soc. (3) 7 (1957)
 CONVERGENCE OF A LACUNARY TRIGONOMETRIC SERIES 899
Our discussion of the convergence of the series (4) leads naturally to the
problem of equidistribution of the sequence {((n k x))} (k = 1, 2, . . .) . By the
result of Weyl (3), given any increasing sequence Ink} (k = 1, 2, . . .) of posi-
tive integers, the set of values of x such that ((nk x)) (k = 1, 2, . . .) is equi-
distributed in (0, 1) has full measure in the Lebesgue sense . Our object in
§ 4 is to examine the exceptional set of values of x for which ((nk x))
(k = 1, 2, . . .) is not equidistributed for different types of sequence {nk} .
Among other results obtained, we prove that if {nk} satisfies (2), then
((nk x)) (k = 1, 2, . . .) is not equidistributed for values of x in a set of
dimension 1 .
For notation and definitions relating to the theory of Hausdorff measures,
see, for example, (2) . We need the following theorem which is a special
case of a result due to Eggleston (1) .
THEOREM A. Suppose Ik (k = 1, 2, .. .) is a linear set consisting of Nk closed
intervals each of length 8 k. Let each interval of Ik contain nk+l > 2 closed
intervals of Ik+l so distributed that their minimum distance apart is Pk+1 > Sk+1 •
Let
P = (~ Ik.
k=1
Then, if lim inf Nk+1 Pk+1 5ks-1 > 0,
k_m
the set P has positive As-measure .

2 . Absolute convergence
We first consider the cardinal number of the set of absolute convergence
of (1) or (4). For sets known to have the cardinal number of the continuum,
we consider the dimension in the sense of Besicovitch . This classifies linear
sets of zero Lebesgue measure by considering their measures with respect
to the class of Hausdorff s-dimensional measures 0 < s < 1 .
The connexion between the absolute convergence of the series (1) and
(4) is given by

LEMMA 1 . If the series converges, then the series

\nk 2a)) 2a
sin(n k x-tik ) converges absolutely .
The proof is trivial . The converse is not true, but we will see that the
infinite cardinal or dimension of the sets of absolute convergence of (1) and
(4) will always be the same . To save space we will state the theorems only
for the series (1) . Theorem nA will be the theorem for series (4) correspond-
ing to the Theorem n for series (1).
We will see that the `size' of the set of absolute convergence depends on
the rate at which t k increases . First we consider the case of t k bounded.
600 P . ERDÖS AND S . J. TAYLOR
In Theorem 16 of (1), Eggleston proves that, given y (0 < y < 1), there
cannot be more than a countable set of x for which ((n k x)) -* y as k -> oo .
Trivial modifications of his proof give :
THEOREM 1 . If {µk} (k = 1, 2, . . .) satisfies 0 < µk < 27r and
{n k } ( k = 1, 2, . . .)
is an increasing sequence of integers such that 1 < t k < K < oo, then there
is at most a countable set of values x such that
sin(n k x-µ k) --> 0 as k - oo .
COROLLARY . Under the conditions of Theorem 1, there is at most a countable
set of values x such that sin(nk x-µk )
converges .
This theorem is best possible in the sense that the set of points where
(1) converges absolutely can have power N o. For example, take n k = 2k,
µk = 0 (k = 1, 2, . . .), then (1) converges absolutely if x = pir2 -4, for any
positive integers p, q . However, the result can be sharpened in the sense
that given {nk} with p < tk < K, for `almost all' sequences {µk} there will
be no points x for which (1) converges . We content ourselves with proving
THEOREM 2 . Suppose the sequence of integers {n k} is such that
1<p<t k <K<oo .
Then there are at most enumerably many pairs (x, y) (0 < x < 29r, 0 < y < 27r)
such that sin(nk x-y) -->0 as k - oo .
COROLLARY . Under the conditions of Theorem 2, there is a linear set Q with
cardinal number < N o such that, when y is not in Q, there is no x with
sin(nk x-y) --> 0 ; and therefore the series 2: sin(nk x-y) does not converge for
any x, unless y c Q .
P-1
Proof. Let E satisfy 0< E (5)
< 8 K2
Let Ek be the subset of the closed square 0 < x < 27T, 0 < y < 21r such
that
sin(nkx-y) sin < E-
Since Since E < 8, it follows that, for every (x, y) in Ek , there must exist an
integer r such that
ink x-y-rir < 2E .
Hence Ek is a subset of the set Fk consisting of the closed strips of the plane
nk x+17T-2E < y < nk x+liT }-2E,
where 1 takes all integer values satisfying
-2(nk +1) < 1 < 2(n k +1) .
 CONVERGENCE OF A LACUNARY TRIGONOMETRIC SERIES 601
Then Fk fl Fk+1 consists of (4nk+5)(4nk+1+5) = Ck closed parallelograms
which are congruent, equally spaced, and similarly situated . Let the pro-
jection on the x-axis of one of these parallelograms have length 8 . By
elementary trigonometry it follows that, if nk > 10,
16E _ 16E
8 C nk+1-nk nk(tk-1) .

Hence, by (2), we have 8 < 16E (6)

nk(P-1)
Now the parallel strips making up Fk+2 are separated by a horizontal
distance d ='r-4E . Thus
n'k+2

d > 2 > 2 1 > 8 by (5) and (6) .

nk+2 K2 nk
The gradient of the parallel strips of Fk+2 is greater than that of either of the
sides of the parallelograms of Fk n Fk+1 • Hence not more than one strip of
Fk+2 can have a non-void intersection with a single parallelogram of Fk n Fk+1 •
M
Suppose, if possible, (x1, y1), (x2, Y2) are two points of (l Fi which are in
i=k
a single parallelogram of Fk n Fk+1 • Then for every positive integer r, the
two points must be in a single parallelogram of Fk+r n Fk+r+1 • But the pro-
jection of such a parallelogram on the x-axis has length which tends to zero
as r -* oo, by (6) . Hence x1 = x2 . Thus the projection n Fi on the x-axis
i=k 00
has at most Ck points . But Ei c Fi (i = 1, 2, . . .), so the projection of f Ei
on the x-axis has at most Ck points .
Now if (x, y) is such that
sin(nkx-y) - 0 ask--co, (7)
m co
then (x, y) is in Ei for all sufficiently large i ; that is (x, y) is in U f Ei = E .
k=1 i=k
oo
Since the projection of n Ei on the x-axis is finite for each k, it follows
i=k:
that the projection of E on the x-axis is at most countable . Thus the set
of x for which there is a y such that (x, y) satisfies (7) is at most countable .
For each such x there are at most 3 values of y in 0 < y < 2i such that
(7) is satisfied . Hence the set of pairs (x, y) satisfying (7) has cardinal
number < No .
We now consider the case of a sequence {nk} such that tk -* oo . In this
case, there can be a set of power continuum for which (1) or (4) converges
absolutely.
602 P . ERDÖS AND S . J. TAYLOR
THEOREM 3 . I f {n k } is such that tk is an integer for large values of k, and
tk - co as k - co, then the set of x such that I sin n k x converges absolutely
has power continuum .

Proof. It is clearly sufficient to prove that I ((n k x)) converges for x in

a set of power continuum . Let k 1 be such that tk is an integer for k > k 1 -1.
Define a sequence {ki } (i = 1, 2, .. .) inductively by letting k i be the smallest
integer such that k i > k i _ 1 , and tk > 2i for every k > ki - 1. Let
Co

i=2

where 0 or 1 for each i. Let E be the set of values taken by x for all
such sequences {-]i} (i = 2, 3, . . . ) . Clearly E has the power of the continuum.
Suppose x is in E . Then if k > k2 ,
Z nk,
((nkx)) =
nk
where the summation extends over those i for which ki > k . Hence
co ki-1

A ((nk X ))
z
k=ka ki>k
nki k
i=3
nki n
k '

Now when k = ki - 1, nk/nk, < 2 - i, and, for k > k 2 , nk _ 1 /nk < 2 . Hence
k; -1

n2
nk < 2i[1+2+ . ..] = 2 11 ;
k=ka k.
Co 1 1
and so ((nkx)) < 2i -1 = 2'
k- s
i=3

Thus for all x in E, the series I ((n k x)) converges .

Remark . The condition that tk be an integer in the above theorem cannot
be omitted . For, if {n k } is defined inductively by
n1 = 1, nk = kn k _ 1 + 1 for k = 2, 3,...,
then it can be shown that sin(nk x-y) converges absolutely for no
pairs (x, y) .
Theorem 3 shows that, when tk is an integer, and tk -->- oo, there are some
sequences {µ k } for which (1) converges absolutely in a set of power con-
tinuum . In fact if tk increases smoothly in some sense and I ti,-1 diverges,
one can prove that for almost all y (in the Lebesgue sense) there is no
value of x such that I sin(nk x-y) converges absolutely. This means that
if tk increases slowly and smoothly the series (1) will not converge absolutely
for any x unless {µ k } is a special sequence . We do not make this idea precise
because we have not obtained nice conditions which are best possible .
 CONVERGENCE OF A LACUNARY TRIGONOMETRIC SERIES 603
Instead we state a result which shows that the conclusion of Theorem 3 is
more or less `best possible' .
THEOREM 4 . Suppose nk = k! (k = 1, 2, . . .) and 0 < y <,, or 7r < y < 2,r .
Then the series I sin(n k x-y) converges absolutely for no value of x .

This is easily proved using the comparison test .

COROLLARY . Suppose nk = k! (k = 1, 2, . . .), then the series I ~cosnk x1
diverges for every x .

The above theorems show that if :E tk 1 diverges, then the series (1) is
unlikely to converge for a set of values of x of power continuum . The
situation is completely different when tk increases rapidly enough to make
ti-1 convergent . This is given by

THEOREM 5 . Suppose {n k } is such that I tk1 converges ; then for any {µ k }

the series I sin (nk x-µk ) converges absolutely for values of x in a set of power
continuum .

Proof. In view of Lemma 1, it is sufficient to prove the absolute con-

vergence of I {((nkx))-ak} for continuum many values of x. Let k o be
such that t k > 10 fork > k o . Let Ik be the set of closed intervals of x such
that 0 < x < 1, and
sin ((nkx)) - aksin < 4 .
tk
Then Ik contains at least (n k -2) closed intervals of length lk, whose cen-
tres are distance nk 1 apart, where
4 =
4 < lk < 8
tk nk nk+1 nk+1
Each interval of Ik contains at least 2 intervals of Ik+1 of length 1k+1, whose
ka+r
centres are distance apart nk+1 . Hence (l Ik contains at least 2r(nka -2)
k=ko
disjoint closed intervals of length lka +r. Thus E = k=ko
n Ik has a perfect
subset, and therefore has power continuum . But, if x is in E, then

sin ((nkx)) - aksin for k > ko ,

< tk
and so I sin ((n k x))-a k I converges .
We now study the dimension in the sense of Besicovitch of the sets of
absolute convergence of (1) and (4) . The dimension is interesting only in
the case where the set has power continuum since enumerable sets neces-
sarily have dimension 0 . Various results can be obtained showing how the
dimension depends on the rate at which tk ->_ oo . We prove only
604 P. ERDÖS AND S . J . TAYLOR
THEOREM 6 . Suppose A > 0, µ > 0, p > 0 are constants, and {n k} is an
increasing sequence of integers such that AkP < tk < akP for each integer k,
and {fck} is any sequence of constants 0 < µk < 27r. Then
(i) if 0 < p < 1, the dimension of the set of x for which I sin(nk x-µk)
converges absolutely is zero;
(ii) if p > 1, the dimension of the set of x for which I sin(n k x-(L k ) con-
verges absolutely is 1 - 1/p .
To obtain Theorems 6 A (ii) and 6 (ii) it is sufficient to prove
(a) the set where (4) converges absolutely has dimension at least (1-1/p) ;
and
(b) the set where (1) converges absolutely has dimension at most (1-1/p) .
Proof of (a) . Let e satisfy 0 < E < p-1 and s satisfy
.
1+E
0 < s < 1-
P
For any such s we will prove that the set of x for which (4) converges
absolutely has positive As-measure . Then the result (a) will follow by taking
values of E which are arbitrarily small .

Put (k = 1, 2, . . .), (8)

'tk = [2 klk-]

and let ko be a fixed integer such that

ktk€- 2 > "ik, fork > ko . (9)

ko exists since t k > AkP and p > 1+e . Let Pk be the set of x such that
0<x<1and
sin ((nkx))-akl < kl+e ( 10 )

Then Pk consists of a finite number of closed intervals, at least (nk- 1) of

which are of the same length yk . If ak = 0 or 1, Yk = (nk kl+E)-1 while if
a k is not near 0 or 1, yk = 2(n k kl+E) -1 : in any case

1nk < Yk <

k12nk
kl

The centres of the intervals of Pk are distant apart 1/n k . Now define a
subset Ik of Pk (k > ko ) as follows . Let
5k = (nkk1+e ) -1 < Yk • ( 11 )
Let Iko be a set of (n,,,-1) closed intervals each of length 8ko, concentric
with intervals of Pko of length yka . Each interval of Iko contains at least
(tko ko 1- E-2) intervals of Pko+1 of length Yko+v by (10) . Define Iko+1 as a
set consisting of closed intervals of length 8k,+1 concentric with some of
 CONVERGENCE OF A LACUNARY TRIGONOMETRIC SERIES 605
the complete intervals of Iko n Pka+1 so chosen that each interval of Iko
contains precisely 71ko intervals of Iko+i : this is possible by (9).
For k > ko , suppose Ik has been defined and consists of closed intervals
of length Sk . Define Ik+l as a set consisting of closed intervals of length
Sk+1 concentric with some of the intervals of length Yk+1
of Ik n Pk+l so
chosen that each interval of Ik contains precisely '/k intervals of Ik+i •
Write p = k=ko
n 1k . The conditions of Theorem A are satisfied with
I(ko-1)!11+E
Nk+1 = (nko - 1 )~ ko~]ko+1 • • • ~ k >i 2 kn
k+1
kl
by (8) ; ( 12 )

Pk+1 > 1Tlnk+1 (13)

for large k, by (10) since k -1- E < 41 . Thus, by (11), (12), (13),
)k+1(k I )-1-E(nk k1+6)1-8
Nk+1 Pk+1 Sk 1 > (2
-1-E}
> 2(2A)kka+Ex1- S~(k!){Pa - S>

since nk > Ak(k!)P . Since p(1-s)-(1+E) > 0, Nk+1 Pk+1 Sk 1 -* ce ask -- co .

By Theorem A, ASP > 0 . But P is a subset of the set of x such that
< k-1 E
sin ((n'kx))-akl

for sufficiently large k. Hence this set has positive AS-measure . But for x
in this set, the series (4) converges absolutely. Hence the set of x for which
(4) converges absolutely has positive AS-measure . This completes the proof.

Proof of (b) .Let E be the set of x such that I I sin(nk x-µk ) I converges .
Suppose 1 > s > 1-1/p ; then it is sufficient to prove that AS(E) = 0 for
all such s . Let E satisfy
0 < e < 1-p(1-s) . (14)
For each x in E, let k 1 , k 2 , . . ., kq , . . . be the sequence of values of the integer k
for which
sin(n k x - µk)1 > 4k • ( 15 )

Then since (1) converges absolutely for this x, the sequence {ki} ( i = 1, 2, . . .)
must have zero density . This implies that there exists an integer N (depend-
ing on x in E) such that, when n > N, the number of integers k < n which
satisfy (15) is less than En .
Let Qk be the set of x such that 0 < x <_ 27T, and

j sin(n k x - µk)1 < ( 16 )

4k

Then Qk C'k , where Ik consists of (2n k +3) closed intervals of length 1/knk
and centres at the points (µ k +l7r)/nk with 1 taking integer values between
- 2 and 2nk • Since the centres of the intervals of Ik are distance apart 7rnk 1 ,
606 P . ERDÖS AND S . J. TAYLOR
the number of intervals of Ik+1 which have a non-void intersection with a
single interval of Ik is at most

m
kink { 2 < k tk .

Letn
k Ik consist of tN,m closed intervals of length not greater than 1/mnm .

Then
tN,m < (2nN } ~N+1 . . nm
3)Nnv .(m-1)nm_l'

so that tN,m < 3nm1) ! (17)

km - .
Now suppose {ki} is an increasing sequence of integers such that
N<kl<k2< . . . <kq < . .<
and, if kq < r, then q < Er
The number of such sequences which differ in the range N < k < m is
certainly less than 2m .
For each ki (i = 1, 2,.. .) the number of intervals inn Ik differs from
k#ki
N<k<m
the number in n Ik by a factor < 27rki since the intervals of Ik, have
N<k<m
length 1/ki nk{ and centres distance apart 7r/nka . Hence for a fixed sequence
satisfying (18), the number of intervals of length 1/mnm needed to cover
n Ik is not greater than
k#k; (1<i<q)
N<k<m ~q
WNm, = tN,m 11 (21Tk i )

(MN!
3(27Tm)q 1) m,
!n
by (17) . Hence, by (18), we have
WN,m < 3(27rm)Em (mN!1) n„t . ( 19 )
!
Thus, if Ev is the set of values of x such that x is in Qk for k > N, except
for a sequence {ki } satisfying (18) where (16) is not known to be satisfied,
then EN can be covered by rN,m intervals of length 1/mnm = lm , where
!
rN,m < 2mWN,m < 15 mm Em nm ,
(m-1)!
by (19) . Hence
N! !
s < 15mrymem n 1 = 15mmEm+1-sr~l-s
rN,m m m mnm m m !'
(m-1)! /S
But nm < µm (m!)P, and hence
rN,m lm < fGm(1-s)15mmem+1N! (m!)P(1-s)-1 < vmmcmm[E+P(1 -s) -11
 CONVERGENCE OF A LACUNARY TRIGONOMETRIC SERIES 607
where v, c are suitable finite constants, since m! > (4m)m . By (14), it
follows that E+p(1-s)-1 < 0, and hence
rNm ln,-±0 as m-~oo .
M
Thus AS(EN ) = 0 . But E c U EN. Hence As(E) = 0 as required .
N=1
To prove Theorems 6 A (i) and 6 (i) it is sufficient to prove
(c) When 0 < p < 1, the set where (1) converges absolutely has dimension
zero .
This can be proved by making a few obvious modifications to the
proof of (b) .
Remark 1 . If this theorem is applied to the case n k = k! of Theorem 4
we see that, if y is not a multiple of -,T, then I sin(n k x-y) converges
absolutely for no value of x, while if y = 0, 7r, or 27r, I sin nk x converges
absolutely in a set of power continuum but zero dimension .
Remark 2 . In the hypothesis of Theorem 6, if one assumes only one-sided
inequalities to be satisfied by tk , then one obtains by the above method of
proof upper or lower bounds for the dimension of the appropriate set of
absolute convergence . In particular, if
n 1 = 1, n2 =2, tk = kk (k = 2, 3, . . .),
the series (1) and (4) each converge absolutely on a set of dimension 1 .
The same methods can be used to prove the following
THEOREM 7 . If {µk} is any sequence of constants 0 < µk < 2-,T and h(z) is
any measure function of class 1,t there exists an increasing sequence {nk} of
integers such that the set of values of x for which I sin(nk x-µk ) converges
absolutely has infinite measure with respect to h(z) .

3. Convergence
Clearly the series (1) or (4) may converge without converging absolutely,
so that, in general the set of points x making (1) or (4) converge will be
larger than the set making the series converge absolutely . However, if
(1) is to converge, sin(n k x-µk ) must tend to zero as k - cc)
. Thus, by
Theorem 1, if tk is bounded for all k, then the set of convergence of (1) is
at most enumerable . We now see that if t k - oc, however slowly, as k -* cc,
then the set of convergence has dimension 1 . Thus the situation is much
simpler than for absolute convergence .
THEOREM 8 . Suppose {µk} is any sequence of constants, 0 < µk < 27r,
and {nk } is an increasing sequence of integers such that tk _> ee, then the set
of values of x such that I sin (nk x-µk ) converges has dimension 1 .
t See (2) for a definition of measure function of class 1 .
608 P . ERDÖS AND S . J . TAYLOR
Proof. Let E be the set of x such that the series (1) converges . Suppose
s is fixed and 0 < s < 1 . It is sufficient to prove that, for any such s,
A8(E) > 0 . We prove this by defining a perfect set P and an integer j such
that, when x is in P,
sin(nk x-11k )-> 0 as k -->oo ; (20)
and, fork > j,
k
sin(nk+lx - µk+1) and sin(nr x-µr) have opposite signs . (21)
r=7

Clearly P is a subset of E, and so it is sufficient to prove that

A8(P) > 0 . (22)
Now nk Ilk -* 0 as k - oo, since tk -> oo . Fork = 1, 2, . . ., let

g(k) = max{40 nk-, 8 nk (1-s)Ikl (23)

nk+1
h(k) = sup g(i) . (24)
i k
Then g(k) ->- 0, h(k) -* 0 as k -> oo, h(k) decreases as k increases, and
h(k) > g(k) (k = 1, 2, . . .) . Put
bk - Stk h(k) . (25)

Then h()nk+l - 4 , (26)

~k <
k
by (23) and (24) .
Choose an integer j so large that

nj > 100 and h(k) < 1'-o for k > j . (27)

By (23) and (24),
nk -8)Ik > {1g(k)J -1 > { 1h(7c)j -1 ,
and therefore n1-8 > {Sh(k)} - k .
Since h(k) decreases, there is a constant C such that
n,1- > C{$h(j)lh(j+1) . . .gh(j+r)}-1 (r = 0, 1, 2, . . .) . (28)
Let Rk be the set of closed intervals given by

x =n (0 < 6 < h(k)),

k
with l taking all integer values . Let Sk be the corresponding set given by

x = nk {µ k ~hr { (-1)i } (-h(k) < 6 < 0) .

0 for x in Rk ,
Then sin(nkx-µk) { (29)
< 0 for x in Sk .
Each of the sets Rk , Sk consists of closed intervals of length h(k)nk'separated
 CONVERGENCE OF A LACUNARY TRIGONOMETRIC SERIES 609

by distances of either 7rnk 1 or {7r-2h(k)}nk 1 . By (27), there are at least

(2nj -1) complete intervals of Rj in 0 < x < 27r . Choose precisely nj of
these, and call this set I 0 . In each interval of I0 there are at least

nj+1
(h(j)
7rnj -2 I
complete intervals of Sj+1 • By (25), we can choose exactly ~j such intervals
in each interval of Io : call this set Il . Then for any x in 10 fl Il , (21) is true
for k = j . We proceed inductively . Suppose Ir (r > 1) has been defined
and consists of closed intervals of length
8, = h( .l +r)nj+r • ( 30 )
Let (1, m) be a typical interval of Ir . There exists a point y (1 < y < m)
such that +r
I sin (n k x-µ k ) < 0 when l< x< y
k=)* (31)
j+r
but sin (n k x-Nt k ) > 0 when y < x < m
k=j
In (1,y) there are at least {(y-l)(nj+r+1/7r)-2} complete intervals of
Rj+r+1, and in (y, m) there are at least {(m-y)(nj+r+1/7r)-2} complete
intervals of Sj+r+1 • Hence, by (26) and (30), since 8, = m-l, we can choose
precisely ~j+r intervals of length 8r+1 such that some of these are complete
intervals of (1, y) n Rj+r+l and others are complete intervals of (y, m) n Sj+r+1 •
Call the set obtained by treating each interval of Ir in this way Ir+1 • Then,
oo
by (31), for x in Ir+l (21) is satisfied with k = j+r . Thus the set P Ir = fl
r=0
satisfies the conditions (20) and (21) . We now apply Theorem A to this set :

y y 7r-2h(j+r+1) l
Nr+1 = njb jbj+1 ~j+r ; Pr+1 > > ~j+r+1>
nj+r+1
and 8, is given by (29) . Hence

> n.
Nr+iPr+1 88-1
r i - bj . bj+rn1-s{h(j+r)}-c1-s>
y. . y 9+r
nj+r+1

>nj tj tj+l . . .tj+r C{h(j+r)}-(1-s),

nj+r+1
by (28) and (25) . Thus

Nr+1 Pr+1 8r' -1 > C{h(j+r)}-(1-s) > C, by (27) ;

and P satisfies (22), as required .

Remark 1 . Theorem 8A is not true for completely arbitrary folk}
: for
example it is not true if a k - 0, since in this case the series (4) converges
only if it converges absolutely . However, if 0 < 8 < ak < 1-8 (k = 1, 2, . . .)
and tk -* ao, it can be proved that (4) converges for x in a set of dimension 1 .
5388 .3 .7 R r
610 P . ERDÖS AND S . J . TAYLOR
Remark 2 . By making some modifications to the argument of Theorem 8,
it can be shown that, with the same hypothesis, and any real number K,
the set of values of x such that
W
sin(n k x-µ k) = K
k=1
has dimension 1 .

4. Equidistribution of ((n k x))

We say that the sequence z1 , z2 , . . ., z,, . . . is equidistributed in (0, 1) if, for
every 1, m satisfying 0 < 1 < m < 1, the density of integers r for which
l < zr < m is exactly (m-l) : that is, if

- 1 when 1 < z,. < m,

E' 0 otherwise,
t
then lim (1
t-.o t l
r=1
THEOREM 9 . The sequence z1 , z 2 , . . . of real numbers is equidistributed in
(0, 1) if and only if t
lim 1 exp(gzr 2,ri) = 0
t--, t r=1
for every positive integer q .

This result is due to Weyl (3) . By the method of Weyl one can prove
easily

THEOREM 10 . If {n k} (k = 1, 2, . . .) is an increasing sequence of integers,

then the set of values of x such that ((n k x)) (k = 1, 2, . . .) is not equidistributed
in (0, 1) has zero Lebesgue measure .

The `size' of the exceptional set of x for which ((nk x)) is not equidistri-
buted depends on the sequence {nk} . For example, it is known that if n k
is given by a polynomial in k with integer coefficients, then the set of x
for which ((nk x)) is not equidistributed is enumerable . In this case t k - 1
as k -~- oo . However, we first see that tk --> 1 is not a sufficient condition
to ensure that the exceptional set has power N o .

THEOREM 11 . There exists a finite constant C, and an increasing sequence

of integers {n k} such that

nk+l - nk < C (k = 1, 2, . . .)
and the set of x such that ((n k x)) is not equidistributed is not enumerable .

Proof. We define a sequence {nk } for which there is a G S-set E, such that
E is dense in an interval, and for x in E, ((nk x)) is not equidistributed .
By the Baire category theorem, a G S-set which is dense in an interval
 CONVERGENCE OF A LACUNARY TRIGONOMETRIC SERIES 611

cannot have power < X o so that it is clearly sufficient for the truth of the
theorem to define such a sequence .
Suppose {Az} (i = 1, 2, . . .) contains all the rationals p satisfying s < p < ,
and each rational occurs in the sequence infinitely often . Let
ks = 51, (s = 0, 1, 2, . . .) . (32)
Put n1 = 1 . Suppose for some positive integer r, n k has been defined for
k < kr _ l . We define nk by induction in the range
kr _ 1 < k < k r (r = 1, 2, . . .)
as follows . Suppose nk _ 1 has been defined . Let n k be the smallest integer
greater than nk _ 1 for which
cos(nk Ar .21r) > 2. (33)
Since 2 > A r > 8, it is clear that

n k -nk-1 < ~ < 24rr

A,
so that nk+l-nk < 100 (k = 1, 2, . . .) . (34)
By (33), kr
cos(n k Ar 2rr) > 2(k r -kr _ 1 ) = 2kr _1, by (32) .
k=k,_,+1
k,
Hence k cos(nk Ar 2rr) >
r k=1 k
k r1 = 5 .
Let Ir be an open interval containing , such that if x is in Ir, then

I k
cos(nk x 27r) > 5 . (35)
k=1

Let E = q=1
n r=q
u Ir .

Then E contains all points x which are in infinitely many Ir . Thus E

contains every rational in s < p < s, and is therefore everywhere dense
in the interval Further, E is a Gs -set .
If x E, then given N, x is in Ir for some
E r > N . By (35), there is an
integer t = k r > N such that

1
- cos{((nk x))2rr} = cos(nk x2rr) >
t
k=1 k=1
Hence, for any x in E,
>5, 2r} x) k cos{(n 1 limsup
t-o 1,
k=1
and therefore, by Theorem 9, ((n k x)) is not equidistributed in (0, 1) . By
(34), the sequence {n k } satisfies the required conditions with C = 100 .
612 P . ERDÖS AND S . J . TAYLOR
It now becomes interesting to ask what is the dimension of the excep-
tional set of non-equidistribution for a sequence satisfying the conditions
of Theorem 11 .
THEOREM 12 . Suppose C is a constant, and {n k} an increasing sequence
of integers such that

nk+i - nk < C (k = 1, 2, . .),

then the set of points x for which ((n k x)) is not equidistributed has dimension
zero .
Proof. Under the given conditions, there is a constant A such that
n k < Ak (k = 1, 2, . . .) . (36)
F or s = 1, 2, . . ., t = 1, 2, . . ., put
t
fm(x) = cos(nk x 27rs) . (37)
k=1~
For any rational µ > 0, let FS µ be the set of x such that
Ifs,l(x)I > µt . (38)
Put a m = [exp(m/logm)] (m = 3, 4, . . .) : then {a m } ( m = 3, 4, . . .) is an
increasing sequence of integers such that
am+i --> 1 as m -* oo .
am
Then if (38) is satisfied for infinitely many integers t (fixed s, µ), it must
be satisfied for infinitely many integers of the sequence {a m}, that is

ES, µ = 1 1 U Fs .a>n ./L (39)

1=3 m=l

contains the set of x such that

I
lim sup fs.t(x) > µ.
t_~
Given a satisfying 1 > a > 0, we prove that
A"&µ = 0 . (40)
This, together with the corresponding result for polynomials of sines
instead of cosines, implies the truth of the theorem by taking the union
of Es,µ for s = 1, 2, . . ., and all positive rationals µ, and applying Theorem 9 .
2n
nsi f',, ( X ) 12 dx = art,
Now f
0

so sinFs,hµ < µ2
t ,
(41)
 CONVERGENCE OF A LACUNARY TRIGONOMETRIC SERIES 613

by (38) . Now
e
• 27r nk
dx f8' t(x) 1
t
• 27rAkI k, by (36),

• 7TAt . 2
Hence, if x o E FS,t,,,, there is an interval containing x 0 of length at least
VA t

which is a subset of F3,, ,1µ . The total length of F,, , ,1µ is less than w2
t , by
(41), so that F3 ,, µ can be enclosed in a finite set of not more than K3 ,µ
42
intervals of total length
µ t where
KS,µ= L -}- 1 .
µ42 -
µ
If 11, 1 2 , . . ., lp (p < KS, µ ) are the lengths of a set of intervals covering F,, ,t,µ ,
but contained in FS,t,1µ , we have
p
Z~ li
< sinFs.~,kµ < 2t,
P l« < KS 47r 1 1
and therefore
µ2 t KS,µ Yl

since the real function z" is convex . It follows that there is a constant LS,µ ,
and a covering by a finite set of intervals of lengths 11, 12 , . . ., l2 of the set
FS such that

oo
By (39) the set ES,µ c U FS,a.,µ (1 = 3, 4, . . .) and therefore E,,,, can be
m=l
covered by a sequence of intervals of lengths 1, 1 2 , . . . such that

l« < L.,11 I am
m=q

Since the series on the right-hand side of this expression converges (40) is
proved . This completes the proof of the theorem .

Remark . The method of proof used is good enough to strengthen the

result in Theorem 12 . Thus the exceptional set of x for which ((nk x)) is
not equidistributed has zero measure with respect to the measure function
[log(1/z)]-1_E for every e > 0 . We have been unable to decide whether or
not this set must have zero capacity-this would follow if the measure
with respect to [log(1/z)]-1 is finite .
614 P . ERDÖS AND S . J . TAYLOR
By the same method of proof used in Theorem 12 one can prove

THEOREM 13 . Suppose C > 0, p > 1 are constants and {n k } is an increasing

sequence of integers such that
n k < CkP (k = 1, 2, . . .),
then the set of points x for which ((n k x)) is not equidistributed has dimension
not greater than (1-1/p) .
By constructing a special sequence {nk} one can prove that the bound
(1-1/p) of this theorem can be attained .
So far, in the present section, we have been considering sequences {n k }
which do not increase too quickly . They are certainly not lacunary, for
under the hypotheses of Theorem 12 or 13
lim inf tk = 1
k-w
The case t -- oc is easily decided . For in § 3 we proved that the set E of
values of x such that {((nk x))-a} converges, 0 < a < 1, has dimension 1 .
For x in this set E, a,
((nkx))
and therefore ((nk x)) is certainly not equidistributed . We now show that
the condition (2) that the sequence {nk} be lacunary is sufficient to imply
that the exceptional set of x for which ((nk x)) is not equidistributed has
dimension 1 . Zygmund, in (4), proved that the set of x for which

lim sup cos(n k x 2lr) > 0

t~00 t
is everywhere dense in (0, 1), but his method does not seem to give the
dimension .

THEOREM 14 . If {n k} is an increasing sequence of integers such that

tk > p > 1, then the set E of values of x such that ((n k x)) is not equidistributed
in (0, 1) has dimension 1 .
Proof. It is sufficient to show that E has positive As-measure for any s
satisfying 0 < s < 1 . Let / > 1/(1-s) . Choose C so that io > C > 0,
and Ci-fla-s) > (42)
3

Let V= Rl0g(1/C) ~ -F-1, (43)

IL log p
where p satisfies (2) . Then for any positive integer k,

nk+v > C-9 > 10 . (44)

nk

Let Qr be the set of x such that 0 < x < 1, and

0 < ((n'rv x )) < C (r = l, 2, . . .) .

I
 CONVERGENCE OF A LACUNARY TRIGONOMETRIC SERIES 615
Then Qr contains at least nrv closed intervals of length

8, = C, ( 45)
nr ,,,
whose centres are distance n,.,,1 apart . Put
Cn(r+1)u -2 (r = 1, 2, . . .) .
Yr =
nr,

Then, by (44), yr > 2Cn(r+1)v > 2C1- S, (46)

n rv
and each interval of Qr contains at least y, complete intervals of Qr+1 •
Let I1 consist of n v complete intervals of Q1 . Suppose Ir has been defined,
r > 1, so that it consists of some of the intervals of Qr of length 8r . In
each interval of Ir choose yr complete intervals of Qr+1, and call this set Ir+1 •
Now let us apply Theorem A to P = r=1
n Ir . 8, is given by (45) ;
Cr+1
Pr+1 > 2 n(r+1)v ; Nr+1 = nvY1Y2 . . .Yr > 2r n(r+i)v, by (46)-

Hence N1-+1 Pr+18s-1 >

(2)1 -1 (nrv) 1-3
By (44), n,.,, > n„ C-g(r-1) > (;-mr-1), and so
1-g(1-s) rC1+P-ps
C
Nr+1 Pr+1 sr -1 > { 2 } 2
as r --)- oo by (42) .
Thus As(P) > 0 . Now, if x is in P, the lower density of integers q such
that 0 < ((nx)) < C is at least 1/v . By (43), if C is small enough,

1v > 10C.

Thus, for x in P, ((nk , x)) has too many members in the interval (0, C),
and is therefore not equidistributed in (0, 1) . Hence P c E, and As(E) > 0,
as required .

REFERENCES
1 . H . G . EGGLESTON, `Sets of fractional dimensions which occur in some problems
of number theory', Proc . London Math . Soc . 54 (1951-2) 42-93 .
2 . S . J . TAYLOR, `On cartesian product sets', J . London Math . Soc . 27 (1952)
295-304 .
3 . H . WEYL, `Über die Gleichwertteilung von Zahlen mod Eins', Math . Ann. 77
(1916) 313-54 .
4. A . ZYGMUND, 'Quelques théorémes sur les series trigonometriques et celles de
puissances', Studia Math . 3 (1931) 77-91 .

Birmingham University