ON SOME APPLICATIONS OF PROBABILITY TO ANALYSIS
AND NUMBER THEORY
P . ERDÖS
It would be quite impossible to give a survey of these subjects in a
short article or lecture, and I will only succeed by making some arbitrary
restrictions on the topics with which I will deal . First of all, I will restrict
myself to problems and results on which I worked, and secondly, I will
not discuss subjects which have been discussed in recently appeared
review articles [1] .
Probabilistic methods have been used in analysis for several decades ;
it suffices to name Paley, Wiener, Kolmogoroff, Zygmund, Salem,
Steinhaus, Kac, Dvoretzky, Kahane, and many others . I will restrict
myself to some questions my collaborators and I worked on for several
ar
years . Hardy was the first to give an example of a power series E a,,, z"k
k=1
cA
which converges uniformly in z a < 1 but for which a k = oo . Piranian
k=1
asked me for what sequences of integers n 1 < n 2 < .. . does there exist a
power series a k zn k which converges uniformly in I z < 1 but for which
k=1
z
Z I ak = 00 . I proved [2] by probabilistic methods that if the sequence
k=1
{n.1} satisfies
lim inf (n;-n 1 ) 1i( i - 0 = 1 where j-i-->oo (1)
then such a power series exists . Zygmund [3] proved that if n k+1 /n k > c > 1
then if Y a,. z111-- converges for z 1, Z I a, < oo . Thus (1) is certainly
k--1 k=1
not far from being best possible, and it is quite likely that it is, in fact,
best possible ; in fact, Zygmund's theorem may remain true for every
sequence which does not satisfy (1), in other words, for every sequence
{n k } for which there exists an absolute constant c so that for every i < j
n ;-nti > ( 1+c)' - ti . (2)
Curiously enough, (1) occurred in a seemingly different context .
Gaier and Meyer-König [4] call the radius defined by z = re 2 O, 0 < it < 1,
U)
singular for f (z) = Z ak z n if f (z) is unbounded in every sector IzI < 1,
71=1
Co
~- E < arg z < 0-f- E where E > 0. They showed that if f (z) _ a,, znk
k=1
and nk+1 /n k > c > 1, and if f (z) is unbounded in ( z ~ < 1, then every radius
Received 1 May, 1963 .
[JOURNAL LONDON MATH . Soc ., 39 (1964), 692-696]
APPLICATIONS OF PROBABILITY TO ANALYSIS AND NUMBER THEORY 693
is a singular radius . Rényi and I [5] showed by probabilistic methods
W
that if {n,;} satisfies (1) then there exists a power series f (z) = E ak znk
k=1
m
for which ak > 0, ak = oo, thus the positive real axis is a singular
k=7
radius but no other radius is singular. In fact, f (z) is bounded in I z < 1
if a region I z-1 < e is excluded (for every e > 0) . It again seems quite
possible that our theorem is best possible ; in fact, perhaps if {n k} satisfies
(2) then the theorem of Gaier and Meyer-Köng remains true, but we
could prove nothing in this direction .
Finally using the methods of [5], Rényi and I solved the following
problem of Zygmund [6] : A well-known theorem of Wiener [7] states
that if ~, (a,,, cos A,, x+bk, sin i\k x) satisfies Ak+1-Ak' 3o and is the Fourier
k=I
series of a function f (x) in L1 , and if f (x) is in L 2 in (a, f3), 0 < a < 0 < 27r,
then it is in L 2 in (0, 27r) . Zygmund now asked whether the same result
remains true for L1, instead of L 2 . We proved that for p > 2 the answer
is negative . In fact, we showed that there exists for every E > 0 a function
,f (x) in L 2 (0 < x < 27T) with the Fourier series
m
Z (ak cos Ak x--b k sin A k x), Ak I-1 - A2 ' 00
k=1
which is bounded for E < x < 27r- E but which does not belong to any
L 2 E ,j for q > 0 in (0, 2Tr) . For p < 2 we could not make any contribution
to the problem of Zygmund .
W
Let E I ak 12 = OC) . Put
k=1
W W
1~-Ek
f, (z) _ E Ek ak z k , Ek = + 1, t = L.r
k=1 k=1 2 k+1
It is well known that for almost all 1, Ek ak zk diverges almost every-
where on the unit circle . Dvoretzky and I proved [8] that if C k is a mono-
where
tone sequence of positive numbers tending to zero and satisfying
2
Jim sup log > 0 (3)
k=W (11/ck)
W
and if I a k > ck then for almost all t E k a k zk diverges everywhere on
k-1
z = 1 . In particular, our theorem holds if ak > c/ki (c > 0) . Further,
we showed that there is a sequence a k satisfying j ak+1 < I ak and I I
W UO
E I ak 2 = 00 so that for almost all t the series E Ek ak zk has on every arc
k=1 k=1
69 4 P . ERDÖS
of I z I = 1 points of convergence whose power is that of the continuum .
We could not decide whether (3) is best possible ; in other words, is it
true that if (3) is false then there exists a sequence {ak} for which I ak I - c k
and for which E Ek ak zk has at least one point of convergence for almost
k=1
all t ?
I would just like to call attention to a problem in the probabilistic
theory of polynomials and power series which I tried several times to
solve, unfortunately without any success . Put
co n I+Ek
fi (z) _ E Ek zk, ft, n (z) = E Ek zk , Ek = + 1, t = k-1 2k+1
k=1 k=1
Is it true that for almost all t
limmaxJ f, +n (z)I/(nlogn)i=C, (4)
n=~ Izl=1
where C is independent oft? The proof of (3) seems very difficult . Salem
and Zygmund [9] proved the following slightly weaker result : There
exist two constants c l and c2 so that for almost all t and sufficiently large
n > no (t)
c l (n logn)l < maxJ ff,n (z) < c 2 (n logn)i . (5)
Izl=1
In (5) the proof of the upper bound is simple ; the real difficulty is
the proof of the lower bound .
A well-known theorem states the following [10]
Let n 1 < n 2 < . . ., nk+1/nk > c > 1, be an infinite sequence of real
numbers and (a k2 +b k2 ) a divergent series of real numbers satisfying
k=1
lim (av+b 2 ( ( ak 2 +bk 2 )) = 0.
N=~ \k=1
Then
N
lim E„ <<', (a k cos 2 7 r nk x+b k sin 2a nk a7 )
N= . k=1
1 w e -,4 212
<w( 1 Z (a k2 +bk2 ) 1 )} du.
k=1 '0 2ir) f- " (6)
Recently I weakened the lacunarity condition in the case a,k = b k = 1 .
In fact I proved [11 ] (using the method of moments) that if n l < n2 < . . .
is an infinite sequence of integers satisfying
c
nk+1 > n k 1+ (7)
where ck -> co, then (6) holds if we assume ak = b k = 1 . It is not hard to
see that the theorem is no longer true for all sequences {nk} if Ck 00 is
APPLICATIONS OF PROBABILITY TO ANALYSIS AND NUMBER THEORY 6 0 5
no longer assumed, but for special sequences, say n k = [ck x ], a > 0, the
theorem probably remains true, but probably could only be proved by
deep number theoretical methods . (For a > 2 this follows from our
theorem (i .e . (7) is satisfied), but I cannot prove it for a = 2 .
I would like to mention, finally, some number theoretic results which
I have recently obtained by probabilistic methods and which are not yet
published .
1 . To every EI and E2 there exists an n o so that if n > no and
m < 2( 1- G, ) log l og ", then all but E2 n integers 1 < u < m have divisors in
every residue class mod m . The result is best possible in the following
sense : If m > 2(1+61) '()g log" then the number of integers u < n which have
a divisor in any given residue class mod m is less than E2 n if n > n o (E l , e 2 ) .
The proof of the second statement is comparatively simple and does not
require probabilistic arguments .
The proof cf the first statement depends on the following result, which
seems to have independent interest : Let G,, be an abelian group having
r(I + E) log n~
n elements, let k -- and choose k elements a l , . . . , a k at
log 2
random . Then for all but (k)) choices of a1, . . ., ak every element of
o(
Gn can be represented in the form H aifi where ci = 0 or 1 .
i=I
I was led to these questions by the following result of Sivasankaranara-
yanu Pillai : Denote by Q(n) the number of integers m < n which do not
have a divisor of the form p ( kp+ 1) . Then
Q(n) < en/log log log n.
Using the above results, I proved
e 'n
Q(n ) = (1 fo (1 )) log 2
. log log n'
Let a l <a,< . . . be an infinite sequence of integers and denote by
f (n) the number of solutions of n = ai +a; . Sidon asked the following
question (in connection with his work on lacunary trigonometric series)
How slowly can the sequence {a k } grow so that f(n) should be bounded?
Rényi and I [12] proved by probabilistic methods that for every e there
exists a sequence ak < k 2+e Ic = 1, 2, . . . for which f (n) < ce . We could
not give an explicit construction of such a sequence and we could not
decide if a, . < k2 +6 can be improved . An old and probably very difficult
conjecture of Turán and myself states that if a k < ck2 then lim sup,f (n) = oo .
n--'0
We can only prove that the sums ai +a; cannot be all different .
2 . I proved several years ago [13] that the density of integers n which
have two divisors dI and d 2 satisfying d1 < d 2 < 2d 1 exists, but I could not
690 APPLICATIONS OF PROBABILITY TO ANALYSIS AND NUMBER THEORY
prove that it is 1 . Unless I made a mistake I proved this recently ; in
fact, I showed that for every q > 0 the density of integers n which have
two divisors d l and d 2 satisfying
V (1-rt) log log n
d1 <d 2 <d l 1 (8)
3
is 1, but the density of integers n which have two divisors d 1 and d2 for
which
(1+q) log log n\
dl < d 2 < d l 1 -{-
3 (9)
is 0 . The proof of (8) is comparatively simple and does not require
probabilistic methods .
References
1 . Recently several review articles appeared on applications of probability- to number
theory, e .g ., 11 . Kac, " Probability methods in some problems of analysis and
number theory ", Bull . American Math . Soc ., 55 (1949), 641-665 ; see also :
Kubilijus, Uspehi Matem . Nauk ., 11 (1956), 31-66 ; P . Erdös, Proc . Internat .
Congress of Math ., Amsterdam (1959) Vol . 3. 13-19 .
2 . P . Erdös, " On the uniform but not absolute convergence of power series with gaps ",
Ann . Soc . Pol. Math ., 25 (1952), 162-168 .
3 . A. Zygmund, Studia Math ., 3 (1931), 77-91 .
4 . U . Gaier and W . Meyer-König, " Singuläre Radien bei Potenzreihen ", Jahresbericht
d .Dm .V ., 59 (1956), 36-48 .
5 . P . Erdös and A . Rényi, " On singular radii of power series ", Publ . Math . Inst . Flung .
Acid., 3 (1959), 159-169 .
6. , " On a problem of Zygmund ", Stanford Studies in Math ., and Stat . 1V .
(Essays in Honor of G . Pólya) (1962), 110-116 ; see also for a slightly weaker
result P . Turin, " On a certain problem in the theory of power series with gaps",
ibid, 404-409 .
7 . N . Wiener, " A class of gap theorems ", Annali di Pisa, 3 (1934), 367-372 .
8 . A . Dvoretzky and P . Erdös, " Divergence of random power series ", Michigan Math .
J ., 6 (1959), 343-347 ; see also A. Dvoretzky, " On the covering of the circle
by randomly placed arcs, Proc . Nat . Acad. Sci . U .S . A ., 42 (1956), 199-203 .
9 . 11 . Salem and A . Zygmund, " Some properties of trigonometric series whose terms have
random signs ", Acta Math ., 91 (1959), 245-301 .
10 . - , "On lacunary trigonometric series (1) and (II)", P roc. Nat . Acid . Sci . U .S. A .,
33 (1947), 333-338 and 34 (1948), 59-62 . For the history of this problem see the
paper of Kac quoted in [1 ] .
11 . P . Erdös, " On trigonometric sums with gaps ", Publ . Math . Inst . Hung . Acid .,
7 (1962), 37-42 .
12 . 1' . Erdös and A . Rényi, " Additive properties of random sequences of positive integers ",
Acta Arithmetica, 6 (1960), 83-110 ; see also " Problems and results in additive
number theory ", Coll . theorie des numbres, Bruxelles (1955), 127-137 .
13 . P . Erdös, " Density of some sequences of integers ", Bull. American Math . Soc .,
64 (1948), 685-692 .
Németvölgyi
Budapest XII,
Hungary .
Printed by C. F . Hodgson & Son, Ltd., Pakenham Street, London, W .C .1
