arXiv:1509.05363v2 [math.

CO] 25 Sep 2015

DISCREPANCY PROBLEM
THE ERDOS
TERENCE TAO
Abstract. We show that for any sequence f p1q, f p2q, . . . taking
values in t1, `1u, the discrepancy

f pjdq
sup

n,dPN j1

of f is innite. This answers a question of Erd

os. In fact the
argument also applies to sequences f taking values in the unit
sphere of a real or complex Hilbert space.
The argument uses three ingredients. The rst is a Fourieranalytic reduction, obtained as part of the Polymath5 project on
this problem, which reduces the problem to the case when f is replaced by a (stochastic) completely multiplicative function g. The
second is a logarithmically averaged version of the Elliott conjecture, established recently by the author, which eectively reduces
to the case when g usually pretends to be a modulated Dirichlet
character. The nal ingredient is (an extension of) a further argument obtained by the Polymath5 project which shows unbounded
discrepancy in this case.

1. Introduction
Given a sequence f : N H taking values in a real or complex
Hilbert space H, define the discrepancy of f to be the quantity

sup f pjdq .

n,dPN j1
H

In other words, the discrepancy is the largest magnitude of a sum

of f along homogeneous arithmetic progressions td, 2d, . . . , ndu in the
natural numbers N t1, 2, 3, . . . u.
The main objective of this paper is to establish the following result:
Theorem 1.1 (Erdos discrepancy problem, vector-valued case). Let
H be a real or complex Hilbert space, and let f : N H be a function
such that }f pnq}H 1 for all n. Then the discrepancy of f is infinite.
Specalising to the case when H is the reals, we thus have
Corollary 1.2 (Erdos discrepancy problem, original formulation). Every sequence f p1q, f p2q, . . . taking values in t1, `1u has infinite discrepancy.
1

TERENCE TAO

This answers a question of Erdos [6] (see also Cudakov

[4] for related
questions), which was recently the subject of the Polymath5 project
[14]; see the recent report [7] on the latter project for further discussion.
It is instructive to consider some near-counterexamples to these results - that is to say, functions that are of unit magnitude, or nearly so,
which have surprisingly small discrepancy - to isolate the key difficulty
of the problem.
Example 1.3 (Dirichlet character). Let : N C be a non-principal
Dirichlet character of period q. Then is completely multiplicative
(thus pnmq pnqpmq for any n, m P N) and has mean zero on any
interval of length q. Thus for any homogeneous arithmetic progression
td, 2d, . . . , ndu one has

n
n

pjq q
pjdq pdq

j1
j1

and so the discrepancy of is at most q (indeed one can refine this

bound further using character sum bounds such as the Burgess bound
[2]). Of course, this does not contradict Theorem 1.1 or Corollary 1.2,
even when the character is real, because pnq vanishes when n shares
a common factor with q. Nevertheless, this demonstrates the need to
exploit the hypothesis that f has magnitude 1 for all n (as opposed to
merely for most n).
Example 1.4 (Borwein-Choi-Coons example). [1] Let 3 be the nonprincipal Dirichlet character of period 3 (thus 3 pnq equals `1 when
n 1 p3q, 1 when n 2 p3q, and 0 when n 0 p3q), and define
the completely multiplicative function 3 : N t1, `1u by setting
3 ppq : 3 ppq when p 3 and
3 p3q `1. This is about the simplest
modification one can make to Example 1.3 to eliminate the zeroes. Now
consider the sum
n

3 pjq
j1

with n : 1 ` 3 ` 3 ` ` 3 for some large k. Writing j 3i m with

m coprime to 3 and i at most k, we can write this sum as
2

i0 1mn{3j

3 p3i mq.

Now observe that 3 p3i mq

3 p3qi 3 pmq 3 pmq. The function 3
has mean zero on every interval of length three, and tn{3j u is equal to
1 mod 3, hence

3 p3i mq 1
1mn{3j

 DISCREPANCY PROBLEM
THE ERDOS

for every i 0, . . . , k. Summing in i, we conclude that

n

j1

3 pjq k ` 1 " log n.

More generally, for natural numbers n, nj1 3 pjq is equal to the number of 1s in the base 3 expansion of n. Thus 3 has infinite discrepancy,
but the divergence is only logarithmic in the n parameter; indeed from
the above calculations
n and the complete multiplicativity of 3 we see
that supnN ;dPN | j1
3 pjdq| is comparable to log N for N 1. This
can be compared with random sequences f : N t1, `1u, whose discrepancy would be expected to diverge like N 1{2`op1q . See the paper of
Borwein, Choi, and Coons [1] for further analysis of functions such as
3 .
Example 1.5 (Vector-valued Borwein-Choi-Coons example). Let H
be a real Hilbert space with orthonormal basis e0 , e1 , e2 , . . . , let 3 be
the character from Example 1.4, and let f : N H be the function
defined by setting f p3a mq : 3 pmqea whenever a 0, 1, 2 . . . and m
is coprime to 3. Thus f takes values in the unit sphere of H. Repeating
the calculation in Example 1.4, we see that if n 1 ` 3 ` 32 ` ` 3k ,
then
n

f pjq e0 ` ` ek
j1

and hence

a
?

f pjq k ` 1 " log n.

j1
H

Conversely, if n is a natural number and d 3l d1 for some l 0, 1, . . .

and d1 coprime to 3, we see from Pythagorass theorem that

ei`l
3 pmd q
f pjdq

j1

i0:3i n
mn{3i
H

1{2

1
i0:3i n

log n.

Thus the discrepancy of this function is infinite and diverges like

?
log N.

Example 1.6 (Random Borwein-Choi-Coons example). Let g : N

t1, `1u be the stochastic (i.e. random) multiplicative function defined
by setting gpnq 3 pnq is coprime to 3, and gp3j q : j for j
1, 2, 3, . . . , where 3 is as in Example 1.4 and 1 , 2 , P t1, `1u are

TERENCE TAO

independently identically distributed signs, attaining 1 and `1 with

equal probability. Arguing similarly to 1.5, we have

1
i`l
3 pmd q
E gpjdq E

j1
i0:3i n

mn{3i

1{2

1
i0:3i n

a
! log n.

?
Thus g in some sense has discrepancy growth like log N on the
average. (Note that one can interpret this example as a special case
of Example 1.5, by setting H to be the Hilbert space of real-valued
square-integrable random variables.) However, by carefully choosing
the base 3 expansion of n depending on the signs 1 , . . . , k (similarly
to Example 1.4) one can show that

sup gpjq k ` 1

k`1
n3
j1

and so the actual discrepancy grows like log N. So this random example
actually has essentially the same discrepancy growth as Example 1.4.
We do not know if scalar sequences of significantly slower discrepancy
growth than this can be constructed.

Example 1.7 (Numerical examples). In [9] a sequence f pnq supported

on n N : 1160, with values in t1, `1u in that range, was constructed with discrepancy 2 (and a SAT solver was used to show that
1160 was the largest possible value of N with this property). A similar
sequence with N 130 000 of discrepancy 3 was also constructed in
that paper, as well as a sequence with N 127 645 of discrepancy 3
that was the restriction to t1, . . . , Nu of a completely multiplicative
sequence taking values in t1, `1u (with the latter value of 127 645
being the best possible value
? of N). This slow growth in discrepancy
may be compared with the log N type divergence in Example 1.5.
The above examples suggest that completely multiplicative functions
are an important test case for Theorem 1.1 and Corollary 1.2. Indeed
the Polymath5 project [14] obtained a number of equivalent formulations of the Erdos discrepancy problem and its variants, including the
logical equivalence of Theorem 1.1 with the following assertion involving such functions. We define a stochastic element of a measurable
space X to be a random variable g taking values in X, or equivalently
a measurable map g : X from an ambient probability space p, Pq
(known as the sample space) to X.

 DISCREPANCY PROBLEM
THE ERDOS

Theorem 1.8 (Equivalent form of vector-valued Erdos discrepancy

problem). Let g : N S 1 be a stochastic completely multiplicative
function taking values in the unit circle S 1 : tz P C : |z| 1u
(where we give the space pS 1 qN of functions from N to S 1 the product -algebra). Then
2

sup E gpjq `8.

j1
n
The equivalence between Theorem 1.1 and Theorem 1.8 was obtained
in [14] using a Fourier-analytic argument; for the convenience of the
reader, we reproduce this argument in Section 2. The close similarity
between Example 1.5 and Example 1.6 can be interpreted as a special
case of this equivalence.
It thus remains to establish Theorem 1.8. To do this, we use a
recent result of the author [15] regarding correlations of multiplicative
functions:

Theorem 1.9 (Logarithmically averaged nonasymptotic Elliott conjecture). [15, Theorem 1.3] Let a1 , a2 be natural numbers, and let b1 , b2
be integers such that a1 b2 a2 b1 0. Let 0, and suppose that A is
sufficiently large depending on , a1 , a2 , b1 , b2 . Let x A, and let
g1 , g2 : N C be multiplicative functions with |g1 pnq|, |g2 pnq| 1 for
all n, with g1 non-pretentious in the sense that
1 Re g1 ppqppqpit
A
p
px

(1.1)

for all Dirichlet characters of period at most A, and all real numbers
t with |t| Ax. Then

g1 pa1 n ` b1 qg2 pa2 n ` b2 q

log .
(1.2)

n
x{nx

This theorem is a variant of the Elliott conjecture [5] (as corrected

in [12]), which in turn is a generalisation of a well known conjecture of
Chowla [3]. See [15] for further discussion of this result, the proof of
which relies on a number of tools, including the recent results in [11],
[12] on mean values of multiplicative functions in short intervals. It
can be viewed as a sort of inverse theorem for pair correlations of
multiplicative functions, asserting that such correlations can only be
large when both of the multiplicative functions pretend to be like
modulated Dirichlet characters n pnqnit .
Using this result and a standard van der Corput argument, one can
show that the only counterexamples to (1.8) come from (stochastic)
completely multiplicative functions that usually pretend to be like

TERENCE TAO

modulated Dirichlet characters (cf. Examples 1.3, 1.4, 1.6). More

precisely, we have
Proposition 1.10 (van der Corput argument). Let g : N S 1 be a
stochastic completely multiplicative function, such that
2

E gpjq C 2
(1.3)

j1

for some finite C 0 and all natural numbers n (thus, g is a counterexample to Theorem 1.8). Let 0, and suppose that X is sufficiently
large depending on , C. Then with probability 1 Opq, one can find
a (stochastic) Dirichlet character of period q OC, p1q and a (stochastic) real number t OC, pXq such that
1 Re gppqppqpit
!C, 1.
p
pX

(1.4)

(See Section 1.1 below for our asymptotic notation conventions.)

We give the (easy) derivation of Proposition 1.10 from Theorem 1.9 in
Section 3.
It remains to demonstrate Theorem 1.8 for random completely multiplicative functions g that obey (1.4) with high probability for large
X and small . Such functions g can be viewed as (somewhat complicated) generalisations of the Borwein-Choi-Coons example (Example
1.4), and it turns out that a more complicated version of the analysis
in Example
n 1.4 2(or Example 1.5) suffices to establish a lower bound
for E| j1 gpjq| (of logarithmic type, similar to that in Example 1.5)
which is enough to conclude Theorem 1.8 and hence Theorem 1.1 and
Corollary 1.2. We give this argument in Section 4.
In principle, the argments in [15] provide an effective value for A
as a function of , a1 , a2 , b1 , b2 in Theorem 1.9, which would in turn
give an explicit lower bound for the divergence of the discrepancy in
Theorem 1.1 or Corollary
? 1.2. However, this bound is likely to be far
too weak to match the log N type divergence observed in Example
?
1.5. Nevertheless, it seems reasonable to conjecture that the log N
order of divergence is best possible for Theorem 1.1 (although it is
unclear to the author whether such a slowly diverging example can
also be attained for Corollary 1.2).
Remark 1.11. In [14, 7], Theorem 1.1 was also shown to be equivalent
to the existence
of sequences

pcm,d qm,dPN , pbn qnPN of non-negative reals

such that m,d cm,d 1, n bn 8, and such that the real quadratic
form

cm,d pxd ` x2d ` ` xmd q2 bn x2n

m,d

 DISCREPANCY PROBLEM
THE ERDOS

is positive semi-definite. The arguments of this paper thus abstractly

show that such sequences exist, but do not appear to give any explicit
construction for such a sequence.
Remark 1.12. In [7, Conjecture 3.12], the following stronger version
of Theorem 1.1 was proposed: if C 0 and N is sufficiently large
depending on C, then for any matrix paij q1i,jN of reals with diagonal
entries equal to 1, there exist homogeneous arithmetic progressions
P td, 2d, . . . , ndu and Q td1 , 2d1, . . . , n1 d1 u in t1, . . . , Nu such that

|
aij | C.
iPP jPQ

Setting aij : xf piq, f pjqyH , we see that this would indeed imply Theorem 1.1 and thus Corollary 1.2. We do not know how to resolve this
conjecture, although it appears that a two-dimensional variant of the
Fourier-analytic arguments in Section 2 below can handle the special
case when aij 1 for all i, j (which would still imply Corollary 1.2
as a special case). We leave this modification of the argument to the
interested reader.
Remark 1.13. As a further near-counterexample to Corollary 1.2, we
present here an example of a sequence f : N t1, `1u for which

sup f pjdq 8

nPN j1

for each d (though of course the left-hand side must be unbounded in

d, thanks to Corollary 1.2). We set f p1q : 1, and then recursively for
D 1, 2, . . . we define
f pjD! ` kq : p1qj f pkq

for k 1, . . . , D! and j 1, . . . , D. Thus for instance the first few

elements of the sequence are
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, . . . .
If d is a natural number and D is any odd number larger than d, then
we see that the block f pdq, f p2dq, . . . , f ppD ` 1q!q is D ` 1 alternating
copies of f pdq, f p2dq, . . . , f pD!q and thus sums to zero; in fact if we
divide f pdq, f p2dq, . . . into consecutive
blocks
of length pD ` 1q!{d then
all such blocks sum to zero, and so n | nj1 f pjdq| is finite for all d.

1.1. Notation. We adopt the usual asymptotic notation of X ! Y ,

Y " X, or X OpY q to denote the assertion that |X| CY for
some constant C. If we need C to depend on an additional parameter
we will denote this by subscripts, e.g. X O pY q denotes the bound
|X| C Y for some C depending on Y . For any real number , we
write epq : e2i .

TERENCE TAO

All sums and products will be over the natural numbers N t1, 2, . . . u
unless otherwise specified, with the exception of sums and products
over p which is always understood to be prime.
We use d|n to denote the assertion that d divides n, and n pdq to
denote the residue class of n modulo d. We use pa, bq to denote the
greatest common divisor of a and b.
We will frequently use probabilistic notation such as the expectation
EX of a random variable X or a probability PpEq of an event E. We
will use boldface symbols such as g to refer to random (i.e. stochastic)
variables, to distinguish them from deterministic variables, which will
be in non-boldface.

1.2. Acknowledgments. The author is supported by NSF grant DMS0649473 and by a Simons Investigator Award. The author thanks Uwe
Stroinski for suggesting a possible connection between Elliott-type results and the Erdos discrepancy problem, leading to the blog post at
terrytao.wordpress.com/2015/09/11 in which it was shown that a
(non-averaged) version of the Elliott conjecture implied Theorem 1.1.
Shortly afterwards, the author obtained the averaged version of that
conjecture in [15], which turned out to be sufficient to complete the
argument. The author also thanks Timothy Gowers for helpful discussions and encouragement, as well as Andrew Granville, Gergely Harcos,
Gil Kalai, Joseph Najnudel, Royce Peng, Uwe Stroinski, and anonymous blog commenters for corrections and comments on the abovementioned blog post and on other previous versions of this manuscript.

2. Fourier analytic reduction

In this section we establish the logical equivalence between Theorem
1.1 and Theorem 1.8. The arguments here are taken from a website1
of the Polymath5 project [14].
The deduction from Theorem 1.8 from Theorem 1.1 is straightforward: if g a random completely multiplicative function taking values
in S 1 , then for each natural number n, gpnq is a random variable taking
values in S 1 . If we write p, Pq for the sample space that the random
variables are defined on, then gpnq can thus be viewed as a measurable
function from to S 1 that we will denote as f pnq. In particular, f pnq
is a unit vector in the complex Hilbert space H : L2 p, Pq. For any

1michaelnielsen.org/polymath1/index.php?title=Fourier

reduction

 DISCREPANCY PROBLEM
THE ERDOS

homogeneous arithmetic progression td, 2d, . . . , ndu, one has

2

n
n

gpjdq

E
f
pjdq

j1

j1
H
2

E gpjq

j1

and on taking suprema in n and d we conclude that Theorem 1.8 follows from Theorem 1.1. (Note that this argument also explains the
similarity between Example 1.6 and Example 1.5.)
Now suppose for contradiction that Theorem 1.8 was true, but Theorem 1.1 failed. Thus we can find a function f : N H taking values
in the unit sphere of a Hilbert space H and a finite quantity C such
that

(2.1)
f pjdq C

j1
H

for all homogeneous arithmetic progressions d, 2d, . . . , nd. By complexifying H if necessary, we may take H to be a complex Hilbert space. To
obtain the required contradiction, it will suffice to construct a random
completely multiplicative function g taking values in S 1 , such that
2

E gpjq !C 1

j1

for all n.
The space of completely multiplicative functions g of magnitude 1
can be identified with the infinite product pS 1 qZ since g is determined by
its values gppq P S 1 at the primes. In particular, this space
is compact
metrisable in the product topology. The functions g | nj1 gpjq|2 are
continuous in this topology for all n. By vague compactness of probability measures on compact metrisable spaces (Prokhorovs theorem),
it thus suffices to construct, for each X 1, a stochastic completely
multiplicative function g gX taking values in S 1 such that
2

E gpjq !C 1

j1
for all n X, where the implied constant is uniform in n and X.
Let X 1, and let p1 , . . . , pr be the primes up to X. Let M X
be a natural number that we assume to be sufficiently large depending
on C, X. Define a function F : pZ{MZqr H by the formula
F pa1 pMq, . . . , ar pMqq : f ppa11 . . . par r q

for a1 , . . . , ar P t0, . . . , M 1u, thus F takes values in the unit sphere

of H. We also define the function : r1, Xs pZ{MZqr by setting

10

TERENCE TAO

ppa11 . . . par r q : pa1 , . . . , ar q whenever pa11 . . . par r is in r1, Xs : tn P N :

1 Xu; note that is well defined for M X. Applying (2.1) for
n X and d of the form pa11 . . . par r with 1 ai M X, we conclude
that

F px ` pjqq !C 1

j1
H

r1

for all n X and all but OX pM q of the M r elements x px1 , . . . , xr q

of pZ{MZqr . For the exceptional elements, we have the trivial bound

F
px
`
pjqq
nX

j1
H

from the triangle inequality. Square-summing in x, we conclude (if M

is sufficiently large depending on C, X) that
2

(2.2)
F px ` pjqq !C 1.
r

M
r j1
xPpZ{M Zq

By Fourier expansion, we can write

F pxq
F pqe
M
r
PpZ{M Zq

where px1 , . . . , xr q p1 , . . . , r q : x1 1 ` ` xr r , and the Fourier

transform F : pZ{MZqr H is defined by the formula

1
x

.
F pq : r
F pxqe
M xPpZ{M Zqr
M
A routine Fourier-analytic calculation (using the Plancherel identity)
then allows us to write the left-hand side of (2.2) as

2
n

pjq

}F pq}2H e
.
j1

M
r
PpZ{M Zq

On the other hand, from a further application of the Plancherel identity

we have

}F pq}2H 1
PpZ{M Zqr

and so we can interpret }F pq}2H as the probability distribution of a

random frequency p1 , . . . , r q P pZ{MZqr . The estimate (2.2)
now takes the form

2
n

pjq

E e
!C 1
j1

 DISCREPANCY PROBLEM
THE ERDOS

11

for all n X. If we then define the stochastic completely multiplicative

function g by setting gppj q : epj {Mq for j 1, . . . , r, and gppq : 1
for all other primes, we obtain
2

E gpjq !C 1

j1

for all n X, as desired.

Remark 2.1. It is instructive to see how the above argument breaks

down when one tries to use the Dirichlet character example in Example
1.3. While often has magnitude 1 in the ordinary (Archimedean)
sense, the function pa1 , . . . , ar q ppa11 . . . par r q is almost always zero,
since the argument pa11 . . . par r of is likely to be a multiple of q. As
such, the quantity }F pq}2H sums to something much less than 1, and
one does not generate a stochastic completely multiplicative function
g with bounded discrepancy.
Remark 2.2. The above arguments also show that Theorem 1.1 automatically implies an apparently stronger version2 of itself, in which
one assumes }f pnq}H 1 for all n, rather than }f pnq}H 1. Indeed,
if f has bounded discrepancy then it must be bounded (since f pnq is

n1
the difference of nj1 f pjq and j1
f pjq), and the above arguments

then carry through; the sum PpZ{M Zqr }F pq}2H is now greater than or
equal to 1, but one can still define a suitable probability distribution
from the }F pq}2H by normalising.
Remark 2.3. If Theorem 1.8 failed, then we could find a constant
C 0 and a stochastic completely multiplicative function g : N S 1
such that
2

E gpjq C 2

j1
for all n. In particular, by the triangle inequality we have
2

n
N

E
gpjq
C2

N n1 j1

and hence for each N, there exists a deterministic completely multiplicative function gN : N S 1 such that
2

n
N

gN pjq C 2 .

N n1 j1
2This

version was suggested by Harrison Brown in the Polymath5 project.

12

TERENCE TAO

Thus, to prove Theorem 1.8 (and hence Theorem 1.1 and Corollary
1.2), it would suffice to obtain a lower bound of the form
2

n
N

(2.3)
gpjq pNq

N n1 j1

for all deterministic completely multiplicative functions g : N S 1 , all

N 1, and some function pNq of N that goes to infinity as N 8.
This was in fact the preferred form of the Fourier-analytic reduction
obtained by the Polymath5 project [14], [7]. It is conceivable that
some refinement of the analysis in this paper in fact yields a bound of
the form (2.3), though this seems to require removing the logarithmic
averaging from Theorem 1.9, as well as avoiding the use of Lemma 4.1
below.
3. Applying the Elliott-type conjecture
In this section we prove Proposition 1.10. Let g, C, be as in that
proposition. Let H 1 be a moderately large natural number depending on to be chosen later, and suppose that X is sufficiently large
depending on H, . From (1.3) and the triangle inequality we have
2

E
gpjq !C log X

n j1
?
XnX

so from Markovs inequality we see with probability 1 Opq that

2

gpjq !C, log X.

n j1
?
XnX

Let us condition to this event. Shifting n by H, we conclude (for X

large enough) that
2

1 n`H

gpjq !C, log X

n j1
?
XnX

and hence by the triangle inequality

2

1 n`H

gpjq !C, log X,

n jn`1
?
XnX

which we rewrite as

XnX

gpn ` hq !C, log X.

n h1

(3.1)

 DISCREPANCY PROBLEM
THE ERDOS

13

We can expand out the left-hand side of (3.1) as

h1 ,h2 Pr1,Hs

XnX

gpn ` h1 qgpn ` h2 q
.
n

The diagonal term h1 , h2 contributes a term of size " H log X to this

expression. Thus, choosing H to be a sufficiently large quantity depending on C, , we can apply the triangle inequality and pigeonhole
principle to find distinct (and stochastic) h1 , h2 P r1, Hs such that

`
h
q
gpn
`
h
qgpn
2
1

"C,,H log X.
?

n
XnX

Applying Theorem 1.9 in the contrapositive, we obtain the claim. (It

is easy to check that the quantities , t produced by Theorem 1.9 can
be selected to be measurable, for instance one can use continuity to
restrict t to be rational and then take a minimal choice of p, tq with
respect to some explicit well-ordering of the countable set of possible
pairs p, tq.)

Remark 3.1. The same argument shows that the hypothesis |gpnq|
1 may be relaxed to |gpnq| 1, and g need only be multiplicative
rather than completely multiplicative, provided that one has a lower

2
" log X. Thus the Dirichlet charbound of the form ?XnX |gpnq|
n
acter example in Example 1.3 is in some sense the only example of a
bounded multiplicative function with bounded discrepancy that is large
for many values of n, in that any other such example must pretend
to be like a (modulated) Dirichlet character. (We thank Gil Kalai for
suggesting this remark.)
4. A generalised Borwein-Choi-Coons analysis
We can now complete the proof of Theorem 1.8 (and thus Theorem
1.1 and Corollary 1.2). Our arguments here will be based on those
from a website3 of the Polymath5 project [14], which treated the case
in which the functions g and appearing in Proposition 1.10 were
real-valued (and the quantity t was set to zero).
Suppose for contradiction that Theorem 1.8 failed, then we can find
a constant C 0 and a stochastic completely multiplicative function
g : N S 1 such that
2

E gpjq C 2

j1
3michaelnielsen.org/polymath1/index.php?title=

Bounded discrepancy multiplicative functions do not correlate with characters

14

TERENCE TAO

for all natural numbers n. We now allow all implied constants to depend
on C, thus
2

E gpjq ! 1

j1

for all n.
We will need the following large and small parameters, selected in
the following order:
A quantity 0 1{2 that is sufficiently small depending on
C.
A natural number H 1 that is sufficiently large depending on
C, .
A quantity 0 1{2 that is sufficiently small depending on
C, , H.
A natural number k 1 that is sufficiently large depending on
C, , H.
A real number X 1 that is sufficiently large depending on
C, , H, , k.
We will implicitly assume these size relationships in the sequel to simplify the computations, for instance by absorbing a smaller error term
into a larger if the latter dominates the former under the above assumptions. The reader may wish to keep the hierarchy
1
1
C ! ! H ! ,k ! X

in mind in the arguments that follow. One could reduce the number
of parameters in the argument by setting : 1{k, but this does not
lead to significant simplifications in the arguments below.
By Proposition 1.10, we see with probability 1 Opq that there
exists a Dirichlet character of period q O p1q and a real number
t O pXq such that
1 Re gppqppqpit
! 1.
(4.1)
p
pX
By reducing if necessary we may assume that is primitive.
It will be convenient to cut down the size of t.

Lemma 4.1. With probability 1 Opq, one has

t O pX q.

(4.2)

Proof. By Proposition 1.10 with X replaced by X , we see that with

probability 1 Opq, one can find a Dirichlet character 1 of period
q1 O p1q and a real number t1 O pX q such that
1 Re gppq1 ppqpit1
! 1.
p

pX

 DISCREPANCY PROBLEM
THE ERDOS

15

We may restrict to the event that |t1 t| X , since we are done

otherwise. Applying the pretentious triangle inequality (see [8, Lemma
3.1]), we conclude that
1 Re ppq1 ppqpipt1 tq
! 1.
p

pX

However, from the Vinogradov-Korobov zero-free region for Lp, 1 q

(see [13, 9.5]) and the bounds X |t1 t| ! X, it is not difficult to
show that
1 Re ppq1 ppqpipt1 tq
" log log X
p

pX

if X is sufficiently large depending on , , a contradiction. The claim

follows.

Let us now condition to the probability 1Opq event that , t exist
obeying (4.1) and the bound (4.2).
The bound (4.1) asserts that g pretends to be like the completely
multiplicative function n pnqnit . We can formalise this by making
the factorisation
it

gpnq : pnqn
hpnq

(4.3)

is the completely multiplicative function of magnitude 1 dewhere

fined by setting ppq

: ppq for p q and ppq
: gppqpit for p|q,
and h is the completely multiplicative function of magnitude 1 defined
by setting hppq : gppqppqpit for p q, and hppq 1 for p|q. The
should be compared with the function
function
3 in Example 1.4
and the function g in Example 1.6.
With the above notation, the bound (4.1) simplifies to

1 Re hppq

(4.4)
! 1.

pX
p

The model case to consider here is when t 0 and h 1, in which

case g .
In this case, one could skip directly ahead to (4.7) below.
Of course, in general t will be non-zero (albeit not too large) and h will
not be identically 1 (but pretends to be 1 in the sense of (4.4)). We
will now perform some manipulations to remove the nit and h factors
factor, which
from g and isolate medium-length sums (4.7) of the
are more tractable to compute with than the corresponding sums of
g; then we will perform more computations to arrive at an expression
(4.10) just involving which we will be able to control fairly easily.

16

TERENCE TAO

We turn to the details. The first step is to eliminate the role of nit .
From (1.3) and the triangle inequality we have
2
1
H

gpn
`
mq
E
!1

H HH 1 2H m1

for all n (even after conditioning to the 1 Opq event mentioned

above). The H1 HH 1 2H averaging will not be used until much later
in the argument, and the reader may wish to ignore it for the time
being.
By (4.3), the above estimate can be written as
1
2
H

` mqpn ` mqit hpn ` mq ! 1.

pn
E

H HH 1 2H m1

For n X 2 , we can use (4.2) and Taylor expansion to conclude that

pn ` mqit nit ` O,H, pX q. The contribution of the error term is
negligible, thus
1
2
H

` mqnit hpn ` mq ! 1
pn
E

H HH 1 2H m1
for all n X 2 . We can factor out the nit factor to obtain
1
2
H

` mqhpn ` mq ! 1.
E
pn

H HH 1 2H m1

For n X 2 we can crudely bound the left-hand side by H 2 . If is

sufficiently small, we can then sum weighted by n1`1{1log X and conclude
that
2
1

pn
`
mqhpn
`
mq

m1
1
! log X.
E
H HH 1 2H n
n1`1{ log X

(The zeta function type weight of n1`1{1log X will be convenient later in

the argument when one has to perform some multiplicative number
theory, as the relevant sums can be computed quite directly and easily
using Euler products.) Thus, with probability 1 Opq, one has from
Markovs inequality that
2
1

pn
`
mqhpn
`
mq

m1
1
! log X.
H HH 1 2H n
n1`1{ log X

We condition to this event. From this point onwards, our arguments

will be purely deterministic in nature (in particular, one can ignore the
boldface fonts in the arguments below if one wishes).

 DISCREPANCY PROBLEM
THE ERDOS

17

We have successfully eliminated the role of nit ; we now work to

and
eliminate h. To do this we will have to partially decouple the
h factors in the above expression, which can be done4 by exploiting
as follows. Call a residue class
the almost periodicity properties of
a pqk q bad if a ` m is divisible by pk for some p|q and 1 m H, and
good otherwise. We restrict n to good residue classes, thus
2
1

pn
`
mqhpn
`
mq

m1
1
H HH 1 2H
n1`1{ log X
k
aPr1,qk s, good na pq q
! log X.
By Cauchy-Schwarz, we conclude that

H 1

pn
`
mqhpn
`
mq
1
m1

1`1{ log X
H HH 1 2H
n

k
aPr1,qk s, good na pq q
!

log2 X
.
qk

Now we claim that for n in a given good residue class a pqk q, the
` mq does not depend on n. Indeed, by hypothesis,
quantity pn
pn ` m, qk q pa ` m, qk q is not divisible by pk for any p|q and is thus a
n`m
n`m
factor of qk1 , and pn`m,q
k q pa`m,qk q is coprime to q. We then factor

n`m
k
` mq ppn

pn
` m, q qq
pn ` m, qk q

n`m
k

ppa
` m, q qq
pa ` m, qk q

a`m
k

ppa
` m, q qq
pa ` m, qk q
`
where in the last line we use the periodicity of . Thus we have pn
` mq, and so
mq pa

2
H1

hpn ` mq

` mq
pa

1`1{ log X
H HH 1 2H
n

na pqk q
aPr1,qk s, good m1

log2 X
.
qk
Shifting n by m, we see that
hpn ` mq

n1`1{ log X
k
!

na pq q

4The

na`m

hpnq
pqk q

n1`1{ log X

` OH p1q

argument here was loosely inspired by the Maier matrix method [10].

18

TERENCE TAO

and thus (for X large enough)

1
H

HH 1 2H aPr1,qk s,

H1

hpnq

` mq
pa

n1`1{ log X

m1
na`m pqk q
good

log X
.
qk

(4.5)

Now, we perform some multiplicative number theory to understand

the innermost sum in (4.5), with the aim of showing that the summand
here is approximately equidistributed modulo qk . From taking Euler
products, we have
hpnq

n1`1{ log X

where S is the Euler product

S :
1
p

hppq
p1`1{ log X

From (4.4) and Mertens theorem one can easily verify that
log X ! |S| ! log X.

(4.6)

More generally, for any Dirichlet character 1 we have

1 pnqhpnq
n

n1`1{ log X

hppq1 ppq

1 1`1{ log X
.
p
p

If 1 is a non-principal character of period dividing qk , then the L

is analytic near s 1, and in particular
function Lps, q : n 1npnq
s
we have
Lp1 `

1
, 1 q !q,k 1.
log X

 DISCREPANCY PROBLEM
THE ERDOS

19

We conclude that
1 pnqhpnq
n

n1`1{ log X

1 ppq
1
hppq1 ppq
1 1`1{ log X
Lp1 `
, 1 q
1 1`1{ log X
log X
p
p
p

|1 hppq|
!q,k exp
p1`1{ log X
p

|1 hppq|
!q,k exp
p
pX

Op1 Re hppqq1{2
!q,k exp
p
pX

1{2
1 Re hppq

!q,k exp O plog log Xq

p
pX
!q,k exppO pplog log Xq1{2 qq

where we have used the Cauchy-Schwarz inequality, Mertens theorem,

and (4.4). For a principal character 0 of period r dividing qk we have
1

0 hpnq
hppq

1 1`1{ log X
1`1{ log X
n
p
n
pr

1
1 1`1{ log X
S
p
p|r

1
1
1
S 1 ` Ok
log X
p
p|r

prq
S ` Ok p1q
r

thanks to (4.6) and the fact that hppq 1 for all p|r. By expansion
into Dirichlet characters we conclude that

nb prq

hpnq
n1`1{ log X

S
` Oq,k pexppO pplog log Xq1{2 qqq
r

for all r|q k and primitive residue classes b prq. For non-primitive residue
classes b prq, we write r pb, rqr 1 and b pb, rqb1 . The previous
arguments then give

nb1 prq1

hpnq
n1`1{ log X

S
` Oq,k pexppO pplog log Xq1{2 qqq
r1

20

TERENCE TAO

which since hppb, rqq 1 gives (again using (4.6))

hpnq
S

` Oq,k pexppO pplog log Xq1{2 qqq

1`1{
log
X
n
r
nb prq

for all b prq (not necessarily primitive). Inserting this back into (4.5)
we see that
1

2
H

S
log2 X
1

1{2
` mq
pa
!
`
O
pexppO
pplog
log
Xq
qqq
.

q,k

k
k
m1

H HH 1 2H
q
q
aPr1,qk s good

The contribution of the Oq,k pexppO pplog log Xq1{2 qqq error term here
can be shown by (4.6) to be O plog2 X{qk q if X is large enough. Applying (4.6) again to cancel off the S term, we conclude that
1
2
H

1
1

(4.7)
pa
`
mq

! 1.
k
m1

q
H
1
HH
2H
aPr1,qk s good
We have now eliminated both t and h. The remaining task is to
establish some lower bound on the discrepancy of medium-length sums
that will contradict (4.7). As mentioned above, this will be a
of
more complicated variant of the analysis in Examples 1.4, 1.5, 1.6 in
which the perfect orthogonality in Example 1.5 is replaced by an almost
orthogonality argument.
We turn to the details. We first expand (4.7) to obtain

1
` m2 q ! qk .
` m1 qpa
pa
H HH 1 2H m ,m Pr1,H 1 s
1
2
aPr1,qk s, good

Write d1 : pa ` m1 , qk q and d2 : pa ` m2 , qk q, thus d1 , d2|qk1 and

for i 1, 2 we have

a ` mi
` mi q pd
i q
pa
.
di

We thus have

1
2q
1 qpd
pd
H
k1
d1 ,d2 |q

aPr1,qk s,

HH 1 2H m1 ,m2 Pr1,H 1 s

good:pa`m1 ,qk qd1 ,pa`m2 ,qk qd2

a ` m1

d1

a ` m2

d2

! qk .
(4.8)

We reinstate the bad a. The number of such a is at most OpH2k qk q,

so their total contribution here is OH p2k qk q which is negligible, thus
we may drop the requirement in (4.8) that a is good.
Note that as is already restricted to numbers coprime to q, and
d1 , d2 divide qk1 , we may replace the constraints pa ` mi , qk q di

 DISCREPANCY PROBLEM
THE ERDOS

21

with di |a ` mi for i 1, 2. Summarising these modifications, we have

arrived at the estimate

1
2q
1 qpd
pd
H HH 1 2H m ,m Pr1,H 1 s
d1 ,d2 |qk1
1
2

(4.9)

a ` m1
a ` m2
k
! q .

d1
d2
k
aPr1,q s:d1 |a`m1 ;d2 |a`m2

Consider the contribution to the left-hand side of (4.9) of an offdiagonal term d1 d2 for a fixed choice of m1 , m2 . To handle these
terms we use the Fourier transform to expand the character pnq
(which, as mentioned before Lemma 4.1, can be taken to be primitive) as a linear combination
of epn{qq for P pZ{qZq . Thus, the

function n 1d1 |n dn1 can be written as a linear combination of
n 1d1 |n epn{d1 qq for P Z coprime to q, which by Fourier expansion
of the 1d1 |n factor (and the fact that all the prime factors of d1 also divide q) can in turn be written as a linear combination of n epn{d1 qq
for p, d1qq 1. Translating, we see that the function

a ` m1
a 1d1 |a`m1
d1
can be written as a linear combination of a epa{d1 qq for p, d1qq
1. Similarly

a ` m2
a 1d2 |a`m2
d2
can be written as a linear combination of a epa{d2 qq for p, d2qq
1. If d1 d2 , then the frequencies involved here are distinct; since qk
is a multiple of both d1 q and d2 q, we conclude the perfect cancellation

a ` m1
a ` m2

0.
d1
d2
k
aPr1,q s:d1 |a`m1 ;d2 |a`m2

Thus we only need to consider the diagonal contribution d1 d2 to

(4.9). For these diagonal terms we do not perform a Fourier expansion
1 qpd
2 q terms helpfully cancel, and we
of the character . The pd
obtain the bound

1
d|qk1

HH 1 2H m1 ,m2 Pr1,H 1 s aPr1,qk s:d|a`m1 ,a`m2

a ` m
1

a ` m
2

(4.10)

! qk .

leaving only the Dirichlet character which

We have now eliminated ,
is much easier to work with. We gather terms and write the left-hand

22

TERENCE TAO

side as
1
H
k1

d|q

a ` m

HH 1 2H aPr1,qk s mPr1,H 1 s:d|a`m

The summand in d is now non-negative. We

? can thus discard all the d
i
i
that are not of the form d q with q H, to conclude that

a
`
m

! qk .

i
H
q
?

HH 1 2H aPr1,qk s mPr1,H 1 s:qi |a`m

i:qi H

It is now that we finally take advantage of the averaging H1 HH 1 2H

to simplify the m summation. Observe from the triangle inequality
that for any H 1 P rH, 3H{2s and a P r1, qk s one has

a
`
m

qi
H 1 aH 1 `qi :qi |a`m

2
2

a ` m
a`m

qi
qi

1
i
i
1
i
mPr1,H `q s:q |a`m

mPr1,H s:q |a`m

summing over i, H 1 , a we conclude that

? H 1
1
i
k 1
i
H PrH,3H{2s

aPr1,q s H mH `q :qi |a`m

i:q H

2
a ` m
! qk .

i
q

In particular, by the pigeonhole principle there exists H1 P rH, 3H{2s

such that
2

a
`
m
! qk .

i
q
?

i:qi H aPr1,qk s H1 mH1 `qi :qi |a`m

Shifting a by H1 and discarding some terms, we conclude that

a ` m
k

! q .

i
q
?

i:qi H aPr1,qk {2s 0mqi :qi |a`m

ObserveY that
] for a fixed a there is exactly one m in the inner sum, and
a`m
a
qi ` 1. Thus we have
qi

?
i:qi H aPr1,qk {2s

Z ^
2

! qk .
`
1

i
q

 DISCREPANCY PROBLEM
THE ERDOS

23

Y ]
Making the change of variables b : qai ` 1 and discarding some
terms, we thus have

qi
|pbq|2 ! qk .
?
i:qi H

bPr1,qki {4s

But b |pbq|2 is periodic of period q with mean " 1, thus

|pbq|2 " qki

bPr1,qki {4s

which when combined with the preceding bound yields

1 ! 1,
?
i:qi H

which leads to a contradiction for H large enough (note the logarithmic

growth in H here, which is consistent with the growth rates in Example
1.5). The claim follows.
References
[1] P. Borwein, S. Choi, M. Coons, Completely multiplicative functions taking
values in t1, 1u, Trans. Amer. Math. Soc. 362 (2010), no. 12, 62796291.
[2] D. A. Burgess, On character sums and primitive roots, Proc. London Math.
Soc. (3), 12 (1962), 179192.
[3] S. Chowla, The Riemann hypothesis and Hilberts tenth problem, Gordon
and Breach, New York, 1965.

[4] N. G. Cudakov,
Theory of the Characters of Number Semigroups, J. Ind.
Math. Soc. 20 (1956), 1115.
[5] P. D. T. A. Elliott, On the correlation of multiplicative functions, Notas Soc.
Mat. Chile, Notas de la Sociedad de Matem
atica de Chile, 11 (1992), 111.
[6] P. Erd
os, Some unsolved problems, Michigan Math. J. 4 (1957), 299300.
[7] W. T. Gowers, Erd
os and arithmetic progressions, Erd
os Centennial, Bolyai
Society Mathematical Studies, 25, L. Lovasz, I. Z. Ruzsa, V. T. Sos eds.,
Springer 2013, pp. 265287.
[8] A. Granville, K. Soundararajan, Large character sums: pretentious characters
and the P
olya-Vinogradov theorem, J. Amer. Math. Soc. 20 (2007), no. 2,
357384.
[9] B. Konev, A. Lisitsa, Computer-aided proof of Erd
os discrepancy properties,
Articial Intelligence 224 (2015), 103118.
[10] H. Maier, Chains of large gaps between consecutive primes, Advances in Mathematics 39 (1981), 257269.
[11] K. Matom
aki, M. Radziwill, Multiplicative functions in short intervals,
preprint. arXiv:1501.04585.
[12] K. Matom
aki, M. Radziwill, T. Tao, An averaged form of Chowlas conjecture,
preprint. arXiv:1503.05121.
[13] H. Montgomery, Ten lectures on the interface between analytic number theory and harmonic analysis, volume 84 of CBMS Regional Conference Series in
Mathematics. Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence,
RI, 1994.

24

TERENCE TAO

[14] D.H.J. Polymath, michaelnielsen.org/polymath1/index.php?title=

The Erd%C5%91s discrepancy problem
[15] T. Tao, The logarithmically averaged Chowla and Elliott conjectures for twopoint correlations, preprint. arXiv:1509.05422
Department of Mathematics, UCLA, 405 Hilgard Ave, Los Angeles
CA 90095, USA
E-mail address: tao@math.ucla.edu