15

On a Problem of A . Zygmund
P . Erdős and A . Rényi

Dedicated to Professor G . Pólya on has 75th birthday

1. Introduction
We shall consider in the present paper the theory of lacunary power
series (and also Fourier series), an area of study in which Professor Pólya
has made a number of important contributions-see, for example, his
beautiful and already classical paper [1] . Many results of this theory may be
characterized (somewhat vaguely) as follows : The behavior of a function
f (z) having a "sufficiently" lacunary power series is essentially the same
on every arc of its circle of convergence (if the series has a finite radius
convergence), or in every angle a < arg z <_ i as I z I - +- (if it is an
entire function) . Among results of this type we mention only one .
Wiener [2] proved that if a lacunary power series
CkZ4 + 00) ,
f (Z) _ ( ) k+1 - ~k
k=1

satisfies lim r -, f (ret') = f(e"') for a < t~ < E and f (e" 9 ) E L2 on the interval
(a, O, then lim, f (re") = f (e 29 ) exists almost everywhere, and f (e' 9 ) E L2 on
the interval [-a, r] . This result can be formulated in the language of
Fourier series as follows : consider a lacunary trigonometic series with gaps
tending to +-, i .e ., a series of the form

(1 .1) Z a k cos ~kx + bk sin 2kx ,

k=1

where )k is an increasing sequence of positive integers such that

(1 .2) lim ().k+1 4) +'~>0
k_T

If such a series is Abel-summable almost everywhere in some interval

(a, p) toa function f(#) which belongs to the class L2 (a, () for -rr <_ a < R <
a, then (1 .1) is the Fourier series of a function f(o) in L'(-7r, ;-L) ; i .e, the
series Zk',(ak + bk) is convergent . Wiener's inequality for trigonometric
polynomials, from which he deduced this result, has been generalized by
Ingham [3] and by Turán [4] . For an elementary proof of a somewhat
110 MATH - ,r : AL
,_ ; .i by
y
 ON A PROBLEM OF A . ZYGMUND 111

weaker inequality see [5] . It follows particularly from Wiener's theorem

that if (1 .1) is the Fourier series of an L-integrable function f(o) and f(tY)
belongs to L 2 in (a, (), then it belongs to L 2 in the whole interval .
About 20 years ago Zygmund suggested to the first-named author that he
consider the problem of whether a similar result holds for the space LQ
with q 2 (q > 1), instead of the space L 2 . In other words, he asked, if
the Fourier series (1 .1) of f (?Y) E L is lacunary in the sense of (2 .2) and if f (,Y)
belongs in the subinterval (a, ~) to the space L°, then does it follow that
f (6) belongs to L° in the whole interval [-r, r]? The problem has so far
remained unsolved, and Zygmund mentions it in his book [6, vol . I, p . 380]
as an open question .
In Sec . 3 of this paper we shall show that the answer to Zygmund's query
is negative for q > 2 ; i .e ., the theorem of Wiener cannot be generalized for
the space L9 (q > 2) . We shall prove even more : There exist functions
f(o) E L 2 having a lacunary Fourier series which do not belong to any class
LQ (q > 2) in the full interval [-r, r] but which, however, are bounded
in every closed subinterval of the interval [-r, r] not containing the point
6 = 0.
We shall prove this by the use of probability theory . We consider a
class of random lacunary Fourier series and prove that almost all series of
this class have the above-mentioned properties . Thus our proof is not
constructive ; only the existence of a function (as a matter of fact, of an
infinity of functions) having the required properties will be proved . Such
a method has often been used in similar situations . Usually in such proofs
the coefficients ak, b k are taken as random variables and the exponents Ak
are explicitly given (not random) numbers . In our proof, however, the
coefficients ak, bk will be given numbers and the exponents 2k will be positive,
integer-valued random variables .
Lacunary random power series in which the exponents are random vari-
ables have already been used for a similar purpose in a previous joint
paper [7] of the authors . The method used in this paper is essentially the
same as that developed in [7], only it has been modified to some extent .
The proof is based on a lemma presented in Sec . 2, similar to the lemma
of [7] . The modification of the method is as follows : In the lemma of [7]
we considered random exponents, each of which is uniformly distributed on
a set of consecutive integers ; in the lemma of Sec . 3, the exponents are
random integers having a binomial distribution .
In Sec . 4 we consider some additional questions . For another related prob-
lem where a probabilistic method was also applied with success, see [10] .

2. A Lemma on Random Cosine Polynomials

In this section we prove the following
LEMMA . Let d > 2 and let s < m, < m 2 < . . . < and be arbitrary positive
integers . Let V1, V2, - - -, vd be independent random variables, each of which
takes on the values 0, ±1, ±2, - - -, ± s with the corresponding probabilities *
* Here and in what follows P( ) denotes the probability of the event in the brackets .
 112 P . ERDŐS AND A . RENYI

(2 .1) P(v; = 1) _ (s . . . , ±s ; j = 1 , . . . , d)
(t = 0 , ±l,
s l) 22a
Put
d
(2 .2) C = Z cos (m ; + vj)T ,
=1

i
where 0 < I ~p I <_ x . Let s be an arbitrary positive number with 0 < s < 2 .
Then we have

(2 .3) P(C > d"/2+e) < 2 exp {-ó 2e/16} ,

Provided that

(2 .4) s > 8log

(p 2 d

and
(2 .5) dE >_ 2 .

PROOF . The proof can be carried out by the method of S . Bernstein [9, pp .
162-65] (see also [7] and [8], where this method has been similarly applied) .
Let t be a real number, I t I <_ 1 . Then we have
d
M(et° ) = rj M(exp Q cos (m ; + v ;)<p}) ,
=1

where M(*) denotes the mean value of the random variable enclosed in
parentheses. Since I ez - 1 - x I <_ I x 12 for x I <_ 1, we have
(2 .7) M(exp {t cos (m; + v;)~o}) <_ 1 + I t I • M (cos (m ; + v;)~o) I + t2 .

Since by (2 .1) clearly

)2S
(2 .8) M(cos (m ; + v ;)-o) _ (cos m;~0) • (cos
2
we obtain
lp )2s
(2 .9) M(et0 ) <_ rl + I t I (cos )2a + t2 a << exp Id[ I t I cos
2 2

Clearly, for 0 < t < 1 \,

`1/2)+2)
(2 .10) P(C >_ d
(tó11/2'+8[) < exp [-td ` 1 / 21+e} M[exp {Ct}] .
= P(exp (Ct[ >_ exp

From (2 .9) and (2 .10), we obtain

~ d d2
P(C tó"/21 +a} .
(2 .lla) 1/2 ) s)} < exp {d(t(cos cp/2)2a + t 2 ) -

Thus for any t with 0 <_ t < 1 we have

(2 .11b) P(I C I > d"/2)+e) <_ 2 exp {t(d(cos (p/2)2' - ó '1/2) +e) + ót 2 }
 ON A PROBLEM OF A . ZYGMUND 1 13

Let us now choose t so as to minimize the right-hand side of (2 .11b) ; that

is, choose t = 12(d' - ' I' - (COs X12)`) . We obtain
)
(2 .12) 4 - /
P(I C I > d "/2'+8 < 2exp {-[d' d(cos (o12) "I ' j .

By the inequality (cos rp12) 2 <_ exp {-áp 2 /16}, valid for tp <_ 7r, and in view I I

of (2 .4) and (2 .5), we have 1/ (p/2)2s <- 1/ d exp {-s~0 2 /16} < 1 _<- d'12, and
d(cos

thus we have from (2 .12)

(2 .13) P(I C I > d (1/2) +P) <_ 2 exp {-d 2e/16} ;

thus our Lemma is proved .

3 . A Class of Random Lacunary Fourier Series

We shall prove the following
THEOREM . There exist real, even functions f (o) defined in the interval
[-7r, 7r] that have the following four properties:
(a) f(z9) belongs to the class L'(-7r, 7r) ;
(b) the Fourier series of f (z9) is of the form

(3 .1) f (z9) - Y, a ; cos 2 ;z9 ,

3=1

where the 2 ; are positive integers such that

(3 .2) lim (2 ;+1 - 2 ;) _ +0 ;

(e) f (t9) is bounded for 8 <_ z9 I <_ 7r for every 8 > 0 ;

I

(d) f (z9) does not belong to any class L°(-7r, 7r) with q > 2 .
PROOF . Put
dk
(3 .3) Ck(z9) _ COS (nk + mkj + Vkj)79 >
j=1
and

(3 .4) Í (r9) 1/2 1

k=a dk +e k1
^ r

where Mk, dk, and mk ; are positive integers, defined as follows : n, = 4,

nk+1 = 2nk (k = 1, 2, • • •) ; dk = nklnk-, (k = 4, 5, • ) [evidently dk is an integer
for each k >_ 4] ; and mk; 1)nk_, ( j = 1, , dk ; k = 4, 5, ) . Let Ek
be defined by

ck = nk-2 for k=4,5,— .

nk-1
Suppose also that Vk ; (k = 1, 2, --- ; j = 1, 2, • , dk) are independent random
variables having the distribution
 114 P . ERDŐS AND A . RENYI

= l s2
(3 .5) P(vk j = 1) U = 0 , ±1, . . . , ±Sk) ,
sk l l 2l k

where Sk = nk_ 1 - Clearly, we have

dk
(3 .6) 2 (z9) dz9 = = Z exp {log 2(4nÁ-21 - 20n Á
1 J -1
z,1 F k -4 dk +4FÁ k-4 171 k _,)}

Since lim k - + - nk_V .j,- ' = 0, it follows that f(z9) e Let us put
(3 .7) Nk = nk/2 = 23 .2nk-2 -1

and consider the polynomial Ck(z9) for z9 = z9h = ahlNk, where h is any integer
such that NkBk < I h I <_ Nk and 8k = link-1 . Clearly, dkk > 2 . Since 7t2 > 8,
we have 8 log dÁ /6 2 < sk for k >> 4 . Thus we may apply our Lemma and
obtain

(3 .8) P(I Ck(r9h) I > dÁ1/21-2k) < 2 exp {-dkek/16}

for Nkók I h I < N1, and therefore

(3 .9) P(max I Ck(6h) I > dÁ1/2)+1k) < 4NÁ exp {-dk ek /16} .
NkőkS_IhIN k

Now evidently for I z9h - 0 I <_ 7rlNk we have, in view of mkd k + Sk < n k ,

2zr nk
(3 .10) I CkW - CÁ0Á) I < max I Ck(z9) I < < dk/ 2 ,
Nk

and thus we obtain

(3 .11) P(max I Ck(z9) I >_ 2dk1/z1+ek) :-5 4Nk exp { -dk ek /16} .
rő k 5_ 19125 .

As the series Zk-4 Nk exp {-dÁ 8Á /16} is clearly convergent, it follows by

the Borel-Cantelli lemma that with probability 1 the inequalities

(3 .12) max I Ck(z9) I < Z dkl/2)+ek

xők s_ 1915x.

are satisfied except for a finite number of values of k ; i .e ., inequality (3 .12)

holds for almost all series of the form (3 .2) (equally for almost all choices
of the random integers vk;) .
It follows, in view of the convergence of the series 1/dk k and of 8k --> 0,
that the series (3 .2) is convergent for every z9 with 0 < 10 I < rr, also that
the series is uniformly convergent in every closed subinterval of [ - 7i, n]
which does not contain the point z9 = 0, and that its sum is bounded in
every such interval . Thus we have already shown that almost all functions
f(z9) defined by (3 .2) satisfy conditions (a) and (c) of our Theorem . It is
easy to see that condition (b) is always satisfied ; as a matter of fact, the
gaps of Ck(z9) are all at least equal to nÁ_ 1 - 2nÁ-1, and the gap between
the greatest exponent of Ck-,(z9) and the least exponent of Ck(?9) is at least
 ON A PROBLEM OF A . ZYGMUND 1 15

nk. - 21ík- 1 , and

thus (3 .2) is evidently satisfied . We even have ,l j - ij-, >
It remains to show that f(0) also has property (d) .
c log' ;,j (c > 0) .
To prove this let us consider f(79) for rlnk <_ o < 5,14n k . For such values
of 0 we obviously have
k-1 ,l
(3 .13) E d;1/2)++21j < kdk 1-1 < nk/2 1

We have also in view of 7rln k >_ a8j+i = z/nj for j >_ k, with probability 1
for k sufficiently large,

(3 .14) ' 1 •'C2) +2 1

- < 1
7- d'j

Finally, we have for ;zlnk <_ 0 <_ 5rrl4n k and for sufficiently large k

dk dk dk 20ir > dk
(3 .15) 1 Ck(o) I >
1/ 2
- 2
j-1
Z I sin (rnkj + vkj)6 I > 1/ 2 -
nk- 1 - 2 .

Thus we obtain from (3 .13)-(3 .15)

1 (112)-22k 1 X1/2)-2Ek ij2
(3 .16) I f(~) I = zdk - 1 - ylk-1 > 4dk

for a/n k < r9 < 5r/4nk and sufficiently large k .

Now let q > 2 be arbitrary but fixed . It follows that for sufficiently
large k one has q(2 - 2e k ) > 1 + p, where p = (q - 2)l4 > 0, and thus
5 xj 1n a
(3 .17) k f«9) I° dt9 >
x/nk 4q+ nkk1 1' 1

As the right-hand side of (3 .16) tends to +~ for k-> +oo, it follows that

~~ 1 f(?9) I° d?~ _ + co . Q .E .D .
One can even prove somewhat more . As a matter of fact, it follows
from (3 .16) that if a > 9,
5x/9nk 2
lim f (r9) log' If () 1 do = +c .
xfn k

Thus not even f 2 (o) log"' I f(?Y) I is integrable if e > 0.

4 . Some Remarks on Additional Problems

It seems that our method is not applicable in the case 1 < q < 2 . As a
matter of fact, to settle this case one has to consider series of the form
(1 .1) with Y, (ak + bk) _ +oo, and in this case if we choose the exponents at
random we cannot even be sure of obtaining a Fourier series .
Another open question is how rapidly the exponents of a series of the
form (1 .1) can increase so that the series will still have properties (a)-(d) .
As is well known (see [6]), if ; k+112k ? A > 1, that is, if (1 .1) is a series
with Hadamard gaps, then if f (0) belongs to L2 (i .e ., Z ( ak + bk) < +00), it
1 16 P . ERDŐS AND A . RENYI

also belongs to L° for every q . Thus certainly J,k cannot increase ex-
ponentially . By modifying our construction, we could get series having
properties (a) - (d) with .i ;+, - a ; > (log 2 ;)" with arbitrarily large A . Our
method is not applicable in the case where Aj, - A ; > ~B with 0 < B < 1 .
It may be true, however, that if the function is bounded in a subinterval
(a, ,6), then it belongs to L°(-s, 7r) for some q > 2 if Aj, - A ; > d; for some
a > 0 . If 2 ;+1 - ). ; > ;, for every a < 1 if j > jo(a), then perhaps it belongs
to L°(- ;r, a) for every q .
Finally, we may ask whether for every function a)(x) tending monotonically
to +- for x-> +- there exists a function f(?9) which has properties (a),
(b), and (c) and is such that f2(?9)w(j f(0) 1) is not integrable . As we men-
tioned at the end of Sec . 3, our functions f(o) are such that

z
JR f2(0) log' I f(0) I dz~
is divergent for a > 9 . By modifying the construction, the value 9 could
be replaced by a smaller one, but our method is not suitable to deal with
the case of arbitrarily slowly increasing functions o_)(x) .
Mathematical Institute of the Hungarian Academy of Sciences

REFERENCES
[11 PÓLYA, G ., Lücken und Singularitaten von Potenzreihen, Math . Zeit ., 29 (1929),
549-640 .
[21 WIENER, N ., A Class of Gap Theorem, Annali di Pisa, 3 (1934), 367-72 .
[31 INGHAM, A . E ., Some Trigonometric Inequalities with Applications to the Theory
of Series, Math . Zeit ., 41 (1936), 367-69 .
[4] TURÁN, P ., On an Inequality, Annal . Univ . Sci . R . Eötvös, Budapest, 1 (1958),
3-6 .
[51 MARCZINKIEWICZ, J ., and A . ZYGMUND, Proof of a Gap Theorem, Duke Math . J.,
4 (1938), 469-72 .
[61 ZYGMUND, A ., Trigonometric Series . Cambridge : Cambridge Univ . Press, 1959 .
[71 ERDŐS, P ., and A . RÉNYI, On Singular Radii of Power Series, Pub . Math . Inst .
Hung . Acad . Sci ., 3 (1958), 159-70 .
[8] ERDŐS, P ., and A . RÉNYI, Probabilistic Approach to Some Problems of Diophantine
Approximation, Illinois J. Math ., 1 (1957), 303-15 .
[91 BERNSTEIN, S ., Teoria Veroiatnostei (Theory of probability), 4th ed . Moscow : 1946 .
[101 NORDLANDER, A ., Studies in Exponential Polynomials (Ph .D . Dissertation), Univ .
Uppsala, Sweden, 1961 .