Académique Documents
Professionnel Documents
Culture Documents
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
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
TERENCE TAO
n
n
pjq q
pjdq pdq
j1
j1
3 pjq
j1
i0 1mn{3j
3 p3i mq.
3 p3i mq 1
1mn{3j
DISCREPANCY PROBLEM
THE ERDOS
j1
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
?
j1
H
ei`l
3 pmd q
f pjdq
j1
i0:3i n
mn{3i
H
1{2
1
i0:3i n
log n.
?
log N.
TERENCE TAO
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.
DISCREPANCY PROBLEM
THE ERDOS
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
log .
(1.2)
n
x{nx
TERENCE TAO
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)
DISCREPANCY PROBLEM
THE ERDOS
|
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
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.
1michaelnielsen.org/polymath1/index.php?title=Fourier
reduction
DISCREPANCY PROBLEM
THE ERDOS
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
10
TERENCE TAO
F px ` pjqq !C 1
j1
H
r1
F
px
`
pjqq
nX
j1
H
(2.2)
F px ` pjqq !C 1.
r
M
r j1
xPpZ{M Zq
F pxq
F pqe
M
r
PpZ{M Zq
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
}F pq}2H 1
PpZ{M Zqr
2
n
pjq
E e
!C 1
j1
DISCREPANCY PROBLEM
THE ERDOS
11
E gpjq !C 1
j1
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
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
E
gpjq !C log X
n j1
?
XnX
n j1
?
XnX
1 n`H
n j1
?
XnX
1 n`H
n jn`1
?
XnX
which we rewrite as
XnX
n h1
(3.1)
DISCREPANCY PROBLEM
THE ERDOS
13
h1 ,h2 Pr1,Hs
XnX
gpn ` h1 qgpn ` h2 q
.
n
`
h
q
gpn
`
h
qgpn
2
1
"C,,H log X.
?
n
XnX
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=
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.
(4.2)
pX
DISCREPANCY PROBLEM
THE ERDOS
15
pX
pX
gpnq : pnqn
hpnq
(4.3)
1 Re hppq
(4.4)
! 1.
pX
p
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
H HH 1 2H m1
` 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
pn
`
mqhpn
`
mq
m1
1
! log X.
E
H HH 1 2H n
n1`1{ log X
pn
`
mqhpn
`
mq
m1
1
! log X.
H HH 1 2H n
n1`1{ log X
DISCREPANCY PROBLEM
THE ERDOS
17
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
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)
n1`1{ log X
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)
n1`1{ log X
hppq1 ppq
1 1`1{ log X
.
p
p
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
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
hpnq
S
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
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
a ` m1
d1
a ` m2
d2
! qk .
(4.8)
DISCREPANCY PROBLEM
THE ERDOS
21
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
1
d|qk1
a ` m
1
a ` m
2
(4.10)
! qk .
22
TERENCE TAO
side as
1
H
k1
d|q
a ` m
a
`
m
! qk .
i
H
q
?
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
? H 1
1
i
k 1
i
H PrH,3H{2s
i:q H
2
a ` m
! qk .
i
q
a
`
m
! qk .
i
q
?
a ` m
k
! q .
i
q
?
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
1 ! 1,
?
i:qi H
[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