You are on page 1of 37

Notes of Szemerédi’s Theorem for

length 4

1 Some notations
In this paper, C will denote various large absolute constants, and c will
denote various small absolute constants.
[a, b] denotes the integers from a to b inclusive. Similarly define the
[a, b), etc. If E is a set, |E| will denote the cardinality of E. N will be

0-0
a large integer, and 0 < δ < 1 will be a parameter, and ε = ε(δ) > 0 be
a quantity depending continuously on δ.
Let M = M (N ) be a slowly growing function of N such that ε−C 
M  N c ; one could take for instance M = log(N ), although one can
get away with a much larger M actually. The exact choice of M is only
important if one is interested in the dependence of N on δ to be chosen
later.
For any function f on ZN , define the Fourier transform fˆ by
X
fˆ := f (m) exp(2πimn/N ),
m∈ZN

and define the Wiener norm kf kA∞ by


kf kA∞ := kfˆkl∞ (ZN ) = sup |fˆ(n)|,
n∈ZN

we also define the modified A∞ norm kf kȦ∞ by

kf kȦ∞ := kfˆkl∞ (ZN −{0}) .


We write Y & X if Y is not of the form o(X).

0-1
2 The main idea of the proof
Szemerédi asserts that for any integers k, if N is sufficiently large de-
pending on δ and k, then every set E ⊂ [0, N ] with |E| > δN must
contain a non-trivial arithmetic progression of length k.
We are going to deal with the case k = 4.
The main idea is to show by induction. Let S be the set of all
0 < δ < 1 such that the Szemerédi theorem hold for k = 4. It is
easy to see that all the δ > 3/4 belong to S, since there is at least one
of a, a + 1, a + 2, a + 3 lies in A, furthermore, we shall prove that if
δ + ε(δ) ∈ S then δ ∈ S.
Assume these, we can give the proof of the theorem.

Proof. Consider the infimum δ0 of S, now claim δ0 = 0. If not, δ0 > 0,


and since δ0 + ε(δ0 ) ∈ S, we get δ0 ∈ S by induction. By the continuity
of ε, we get
δ + ε(δ) → δ0 + ε(δ0 ), δ → δ0 − .

0-2
Therefore there exists δ1 < δ0 such that

δ1 + ε(δ1 ) = δ0 ,

we get δ1 ∈ S by induction, contradict to the infimum of δ0 .The proof is


done.

Fix δ, let N be a prime sufficiently large depending on δ and ε, and


let E ⊂ [0, N ) be such that |E| > δN , by refining E if necessary we may
assume that
|E| = δN (1)
we can assume that
|E ∩ P | 6 (δ + ε)|P | (2)
for all arithmetic progressions P ⊂ [0, N ) of length  M by induction.
This is the only place where we shall use the induction hypothesis.

0-3
3 Proof of the induction above
Lemma 3.1. We have

|E ∩ P | 6 δ|P | + O(ε|P |) (3)

for all ZN arithmetic progression P of length |P |  M 3

Proof. Let r 6= 0 be the spacing of an arithmetic progression P in Z/N Z.


(i) if 0 < |r|  N/M . Then we can partition P into the disjoint union of
genuine arithmetic progressions of the same step size r, simply by making
a cut every time the progression wraps around the end of 1, · · · , N .
Except possibly for the first and last progression, each of the genuine
progressions has length at least M , and so the density of A in those
genuine progressions is at most δ + ε by (2). Adding up all these density
estimates for those genuine progressions of length at least M , we obtain
the bound
|E ∩ P | 6 (δ + ε)|P | + 2M.

0-4
(ii) Assume r is arbitrary. Consider the first O(M ) multiples of r in Z/N Z.
By the pigeonhole principle, two of them must be  N/M apart, thus
we can find a non-zero j = O(M ) such that |jr mod N |  N/M . We
can then partition P into |j| disjoint progressions, all of length ∼ |P |/|j|,
with step size jr mod N . Applying the previous estimate to all of these
progressions and adding up, we obtain

|E ∩ P | 6 (δ + ε)|P | + 2|j|M 6 (δ + ε)|P | + O(M 2 ),

and the claim follows if |P | is large enough.

We can show that (1) and (3) will imply

k1E kȦ∞ = O(εN ), (4)

and X
k1E∩(E+k) kȦ∞ = O(εN 2 ). (5)
k∈ZN

0-5
1
Assume (4) and (5) hold, we show that E contains ∼ δ 4 N 2 + O(ε 4 N 2 )
genuine arithmetic progression of length 4, and the induction follows.
It suffices to show that
1
X
1E 0 (a)1E 0 (a + r)1E (a + 2r)1E (a + 3r) ∼ δ 4 N 2 + O(ε 4 N 2 ) (6)
a,r∈ZN

where E 0 = E ∩ [0.4N, 0.6N ], since the r = 0 contribution is negligible


if N is sufficiently large. From (1) and (3), we see that
|E\E 0 | = |E ∩ (0.6N, 1.4N )| 6 (δ + ε)0.8N
hence we have
|E 0 | = δ 0 N
for some δ 0 ∼ δ. Write
X
f (a) = 1E 0 (a + r)1E (a + 2r)1E (a + 3r).
r∈ZN

It suffices to show that


1
X
0 2
(f (a) − δ δ N ) = O(ε N 2 ).
4

a∈E 0

0-6
By Cauchy-Schwarz, it suffices to show that
1
X
0 2
(f (a) − δ δ N ) = O(ε 2 N 3 ).
2

a∈E 0

We need to show that


1
X
(f (a) − δ 0 δ 2 N )2 = O(ε 2 N 3 ).
a∈ZN

Lemma 3.2. For any a, b, c ∈ ZN − {0} we have


X
f1 (k)f2 (l)f3 (m) 6 kf1 k2 kf2 k2 kf3 kA∞ (7)
k,l,m∈ZN :ak+bl+cm=0

Similarly for permutations.

Proof. We can rewrite the left hand side of (7) as


1 X ˆ
f1 (an)fˆ2 (bn)fˆ3 (cn),
N
n∈ZN

0-7
which can then be estimated using Cauchy-Schwarz by
1 ˆ
kf1 k2 kfˆ2 k2 kf3 kA∞ .
N
The claim then follows from Plancherel’s theorem.
Corollary 3.3. For any a, b, c ∈ ZN \{0}, we have
X
f1 (k)f2 (l)f3 (m) = N 2 E(f1 )E(f2 )E(f3 )+O(kf1 k2 kf2 k2 kf3 kA
k,l,m∈ZN :ak+bl+cm=0
(8)
Similarly for permutations.
Proof. Note that
kf kȦ∞ = kf − E(f )kA∞ . (9)
Then the corollary can be got from Lemma 3.2, (9)and the identity
X
(f1 (k) − E(f1 ))(f2 (l) − E(f2 ))(f3 (m) − E(f3 ))
k,l,m∈ZN :ak+bl+cm=0
X
= f1 (k)f2 (l)f3 (m) − N 2 E(f1 )E(f2 )E(f3 )
k,l,m∈ZN :ak+bl+cm=0

0-8
.
Proposition 3.4. For any F ⊂ ZN , we have
X
||F ∩ (E + u)| − δ|F || = O(εN 2 ). (10)
u∈ZN

Proof. It suffices to show


X
(|F ∩ (E + u)| − δ|F |) = O(εN 2 )
u∈A

for arbitrary A, F ⊂ ZN . From (4) and Lemma 3.2, we see:


X X
| (|F ∩ (E + u)| − δ|F |)| = | 1A (u)1F (k)(1E (k − u) − δ)|
u∈A u,k∈ZN

6 k1A k2 k1F k2 k1E − δkA∞


6 N k1E kȦ∞
= O(εN 2 ),
and the claim follows.

0-9
Remark: In Lemma 3.4, if we take F = E, we get
X
||E ∩ (E + u)| − δ 2 N | = O(εN 2 ).
u∈ZN

1
Hence we get for all but at most O(ε 2 N ) values of u such that
1
|E ∩ (E + u)| = δ 2 N + O(ε 2 N ). (11)

By Lemma 3.4, take F = 2E = {2e : e ∈ E}, we get

0-10
X X X
0 2
(f (a) − δ δ N ) = ( 1E 0 (a + r)1E (a + 2r)1E (a + 3r) − δ 0 δ 2 N )
a∈ZN a∈ZN r∈ZN
X X
= ( 1E 0 (b)1E (b + r)1E (b + 2r) − δ 0 δ 2 N )
b∈ZN r∈ZN
X X
= 1E 0 (b)1E (b + r)1E (b + 2r) − δ 0 δ 2 N 2
b∈ZN r∈ZN
X X
= ( 1E (b + r)1E (b + 2r) − δ 2 N )
b∈E 0 r∈ZN
X X
= ( 1E (u)1E (2u − b) − δ 2 N )
b∈E 0 u∈ZN
X
= (|2E ∩ (E + b)| − δ 2 N )
b∈E 0

= O(εN 2 ).

0-11
Hence we have X
f (a) = δ 0 δ 2 N 2 + O(εN 2 ).
a∈ZN

Therefore
1
X
(f (a) − δ 0 δ 2 N )2 = O(ε 2 N 3 )
a∈ZN
1
X X
2 0 2 02 4
⇐⇒ f (a) − 2δ δ N f (a) + δ δ N = O(ε N 3 )
3 2

a∈ZN a∈ZN
1
X
02 4
⇐⇒ 2 3
f (a) = δ δ N + O(ε N 3 ) 2

a∈ZN

0-12
By Corollary 3.3 and (5),
X
f 2 (a)
a∈ZN
X X X
= 1E 0 (a + r)1E 0 (a + s)1E (a + 2r)1E (a + 2s)1E (a + 3r)1E (a + 3s
a∈ZN r∈ZN s∈ZN
X X X
= 1E 0 ∩(E 0 +u) (a + r)1E∩(E+2u) (a + 2r)1E∩(E+3u) (a + 3r)
a∈ZN u∈ZN r∈ZN
X X
= 1E 0 ∩(E 0 +u) (b)1E∩(E+2u) (b + r)1E∩(E+3u) (b + 2r)
u∈ZN b,r∈ZN
X X
−1 0 0
= N |E ∩ (E + u)||E ∩ (E + 2u)||E ∩ (E + 3u)| + O(N k1E∩(E+3u)
u∈ZN u∈ZN
X
−1
=N |E 0 ∩ (E 0 + u)||E ∩ (E + 2u)||E ∩ (E + 3u)| + O(εN 3 )
u∈ZN

Hence it suffices to show


1
X
|E 0 ∩ (E 0 + u)||E ∩ (E + 2u)||E ∩ (E + 3u)| = δ 02 δ 4 N 4 + O(ε 2 N 4 )
u∈ZN

0-13
By the remark after the lemma 3.4,
1
|E ∩ (E + 2u)| = δ 2 N + O(ε 2 N ),
1
|E ∩ (E + 3u)| = δ 2 N + O(ε 2 N )
1
for all but at most O(ε 2 N ) values of u. Therefore
X
|E 0 ∩ (E 0 + u)||E ∩ (E + 2u)||E ∩ (E + 3u)|
u∈ZN
1 1
X
0 0
= |E ∩ (E + u)|(δ N + O(ε N )) + O(ε N 4 )
2 2 2 2

u∈ZN
1 1
X
4 2 2 0 0
=(δ N + O(ε N )) 2 |E ∩ (E + u)| + O(ε N 4 ) 2

u∈ZN
1 1
=(δ 4 N 2 + O(ε 2 N 2 ))|E 0 |2 + O(ε 2 N 4 )
1 1
=(δ 4 N 2 + O(ε 2 N 2 ))δ 02 N 2 + O(ε 2 N 4 )
1
=δ 4 δ 02 N 4 + O(ε 2 N 4 ).
That’s the end of the proof, all the left is to show (4) and (5).

0-14
3.1 The proof of (4)
Proposition 3.5. We have
X X
| exp(2πiφs (j)/N )| = O(εN |P0 |) (12)
s∈ZN j∈E∩(P0 +s)

for any ZN arithmetic progression P0 of length  M C and any non-


constant quadratic functions φ0 , φ1 , · · · , φN −1 with integer coefficients.

Assume this proposition, we show that (4) is the corollary of it.

Corollary 3.6. k1E kȦ∞ = O(εN ).

Proof : Assume Proposition 3.5, we have


X X
| exp(2πiφs (j)/N )| = O(εN |P0 |).
s∈ZN j∈E∩(P0 +s)

0-15
For a fixed 1 6 n 6 N − 1, take φs (j) = nj, we get:
X X
O(εN |P0 |) = | exp(2πinj/N )|
s∈ZN j∈E∩(P0 +s)
X X
=| 1E (j) exp(2πinj/N )|
s∈ZN j∈P0 +s
X
= ||P0 | 1E (j) exp(2πinj/N )|
j∈ZN

= ||P0 |1̂E (n)|,

and therefore we get (4) since n is non-zero. The lemma is done.


We shall prove Proposition 3.5. Fix the functions φ0 , φ1 , · · · , φN −1
and the ZN arithmetic progression P0 , we can reduce to the case when
P0 is the arithmetic progression [0, r) for some M C  r < N since N is
prime.

Definition 3.7. We say Z N arithmetic progressions P1 , P2 , · · · , Pn evenly


P n
cover ZN if the function j=1 1Pj is a constant on ZN .

0-16
Lemma 3.8. Let P1 , P2 , · · · , Pn be ZN arithmetic progression of length
 M 3 which evenly cover ZN . Then we have
n
X n
X
||E ∩ Pj | − δ|Pj || = O(ε |Pj |). (13)
j=1 j=1

Proof. From (3), we have

|E ∩ Pj | − δ|Pj | = O(ε|Pj |),


Pn
if |E ∩ Pj | − δ|Pj | > 0. Write f = j=1 1Pj , then

n n N n
X X X |E| X |E| X
(|E ∩ Pj | − δ|Pj |) = f −δ |Pj | = f− |Pj | = 0
j=1 j=1
N N j=1
x∈E k=1

with the identity


n
X n X
X N N X
X n N
X
|Pj | = 1Pj (k) = 1Pj (k) = f.
j=1 j=1 k=1 k=1 j=1 k=1

0-17
Hence,
n
X n
X n
X
||E ∩ Pj | − δ|Pj || = 2 (|E ∩ Pj | − δ|Pi |)+ = O(ε |Pi |).
j=1 j=1 j=1

Lemma 3.9. For every s ∈ ZN , we can partition P0 + s = [0, r) + s into


arithmetic progressions Ps,1 , · · · , Ps,ns of length |Ps,j |  M 3 such that
for each 1 6 j 6 ns , the function exp(2πiφs (·)/N ) is within o(1) of a
constant on Ps,j , j = 1, · · · , ns .

Assume this lemma for the moment and complete the proof of (12).
Proof of (12): Firstly we observe that when s ranges over ZN and
j ranges from 1 to ns , the progressions Ps,j evenly cover ZN . By Lemma
3.8, we have
ns
X X X
||E ∩ Ps,j | − δ|Ps,j || = O(ε |Ps,j |) = O(εN |P0 |). (14)
s∈ZN j=1 s∈ZN

0-18
By Lemma 3.9, we have
X
exp(2πiφs (k)/N )
k∈E∩Ps,j

|E ∩ Ps,j | X
= exp(2πiφs (k)/N ) + o(|Ps,j |)
|Ps,j |
k∈Ps,j

|E ∩ Ps,j | − δ|Ps,j | + δ|Ps,j | X


= exp(2πiφs (k)/N ) + o(|Ps,j |)
|Ps,j |
k∈Ps,j
X
=δ exp(2πiφs (k)/N ) + O(||E ∩ Ps,j | − δ|Ps,j ||) + o(|Ps,j |).
k∈Ps,j

Summing this in j and taking absolute value, we get

0-19
X
| exp(2πiφs (k)/N )|
k∈E∩(P0 +s)
X Xns
=δ| exp(2πiφs (k)/N )| + O( ||E ∩ Ps,j | − δ|Ps,j ||) + o(|P0 |).
k∈P0 +s j=1

Summing the above equation in s and using (14) we have


X X
| exp(2πiφs (k)/N )|
s∈ZN k∈E∩(P0 +s)
X X ns
X X
=δ | exp(2πiφs (k)/N )| + O( ||E ∩ Ps,j | − δ|Ps,j ||) + o(N |P0 |)
s∈ZN k∈P0 +s s∈ZN j=1
X X
=δ | exp(2πiφs (k)/N )| + O(εN |P0 |)
s∈ZN k∈P0 +s

From standard exponential sum estimates and the non-constancy of φs

0-20
we observe that
X
exp(2πiφs (k)/N ) = o(|P0 |). (15)
k∈P0 +s

and (12) follows from that.


It remains to prove Lemma 3.9. Fix s, without loss of generality we
can assume that s = 0. It suffices to partition [0, r) into ZN arithmetic
progression P0,j of length ∼ rc for some small c such that
φ0 (k 0 ) − φ0 (k)
|| || = o(1)
N
for all k, k 0 ∈ P0,j . Let dj be the step size and kj the initial element of
P0,j . Assume φ0 (k) = A2 k 2 + A1 k + A0 , then

φ0 (kj + adj ) − φ0 (kj + bdj ) = (a2 − b2 )A2 d2j + (a − b)dj (2A2 kj + A1 ).

It suffices to find dj such that

a2 A2 d2j
|| || = o(1),
N

0-21
(2A2 kj + A1 )dj a
|| || = o(1)
N
hold for all |a| . rc . It thus suffices to obtain the conditions
A2 d2j (2A2 kj + A1 )dj
|| ||, || || = O(r−3c ). (16)
N N
By a quantitative version of Weyl’s theorem (Weyl: Let k > 2,
2
232k
t > 2 , N > t and a ∈ ZN . Then there exists p 6 t such that
k+1 c
|pk a/N | 6 t−1/k2 . Here we take k = 2, t = r 7 .)
c
We can find d 6 r 7 such that
A2 d2
|| || = o(r−100c ).
N
Fix this d and partition [0, r) into arithmetic progressions P 1 , P 2 · · ·
of step d and length ∼ r5c . For each such progression P t , let k t be an
element of it. By pigeonhole principle we can find an integer mt = O(r3c )
such that
(2A2 kt + A1 )mt d
|| || = O(r−3c ).
N

0-22
Now partition each P t into arithmetic progressions P0,j of step mt d and
length ∼ r5c /mt  M 3 . The first condition of (16) is clear. To verify
the latter we observe that

kj = kt + ud

for some integer u = O(r5c ), and thus

(2A2 kj + A1 )mt d (2A2 kt + A1 )mt d 2A2 umt d2


|| || = || + ||
N N N
(2A2 kt + A1 )mt d 2A2 umt d2
6 || || + || ||
N N
−3c A2 d2 5c 3c
= O(r ) + O(|| ||r r )
N
= O(r−3c )

as desired.

0-23
3.2 The proof of (5)
Lemma 3.10. For any ZN arithmetic progression P of length M C 
|P | = o(N ) and (λ, µ) ∈ ZN × ZN − {(0, 0)} we have
X
|1̂E∩(E+k) (λk + µ)|2 = O(εN 2 |P |). (17)
k∈P

Assume this lemma, we can now prove (5). Suppose for contradiction
that (5) failed, we get ∀c > 0, ∀N0 > 0, ∃N > N0 such that
X
k1E∩(E+k) kȦ∞ > 2cεN 2 .
k∈ZN

For an arbitrary fixed c and N . We can find a set B ⊂ ZN of size


|B| > cεN such that
||1E∩(E+k) ||Ȧ∞ > cεN.
In fact, let B be the set of all such k that
||1E∩(E+k) ||Ȧ∞ > cεN.

0-24
Hence we have
X
2
2cεN 6 k1E∩(E+k) kȦ∞
k∈ZN
X X
= k1E∩(E+k) kȦ∞ + k1E∩(E+k) kȦ∞
||1E∩(E+k) ||Ȧ∞ <cεN ||1E∩(E+k) ||Ȧ∞ >cεN
X X
6 cεN + N 1
k∈ZN ||1E∩(E+k) ||Ȧ∞ >cεN

= cεN 2 + N |B|.
Therefore |B| > cεN .
We can therefore find a function φ : B → ZN − {0} such that
|1̂E∩(E+k) (φ(k))| > cεN (18)
for all k ∈ B. Fix B, φ, and let Ω denote the graph of φ:
Ω := {(k, φ(k)) : k ∈ B}.
Ω has considerable arithmetic structure:

0-25
Lemma 3.11. We have
]{(a, b, c, d) ∈ Ω4 : a + b = c + d} > (cε)12 N 3
Proof. Square (18) and use the identity
X X
2
|1̂A (n)| = exp(2πiun/N ) (19)
u∈ZN s∈A∩(A+u)

for all A ⊂ ZN and n ∈ ZN , we have


X X
exp(2πiuφ(k)/N ) > c2 ε2 N 2 .
u∈ZN s∈E∩(E+k)∩(E+u)∩(E+k+u)

Summing this over k ∈ B,


X X X
3 3 3
c ε N 6 exp(2πiuφ(k)/N )
k∈B u∈ZN s∈E∩(E+k)∩(E+u)∩(E+k+u)
X X X
= 1B (k) exp(2πiuφ(k)/N )
u∈ZN k∈ZN s∈E∩(E+k)∩(E+u)∩(E+k+u)
X X X
= 1E∩(E+u) (s)1E∩(E+u) (s − k)1B (k) exp(2πiuφ(k)/N )
u∈ZN k∈ZN s∈ZN

0-26
By Lemma 3.2, we have
X X X
3 3 3
c ε N 6 exp(2πiuφ(k)/N )
k∈B u∈ZN s∈E∩(E+k)∩(E+u)∩(E+k+u)
X
6 ||1E∩(E+u) ||2 ||1E∩(E+u) ||2 ||1B (·) exp(2πiuφ(·)/N )||A∞
u∈ZN
X
6N ||1B (·) exp(2πiuφ(·)/N )||A∞ .
u∈ZN

Hence, X
||1B (·) exp(2πiuφ(·)/N )||A∞ > c3 ε3 N 2 .
u∈ZN

0-27
Since
X
||1B (·) exp(2πiuφ(·)/N )||A∞ > c3 ε3 N 2
u∈ZN
X X
= max | ||1B (s) exp(2πiuφ(s)/N ) exp(2πims/N )|
m∈ZN
u∈ZN s∈ZN
X X
= max | 1Ω (s, t) exp(2πi(ut + ms)/N )|
m∈ZN
u∈ZN (s,t)∈ZN ×ZN
X
= max |1̂Ω (m, u)|,
m∈ZN
u∈ZN

we have X
sup |1̂Ω (m, u)| > c3 ε3 N 2 .
m∈ZN
u∈ZN

From Hölder, we thus have


X
sup |1̂Ω (m, u)|4 > (cε)12 N 5
m∈ZN
u∈ZN

0-28
and hence X
|1̂Ω (m, u)|4 > (cε)12 N 5
u,m∈ZN

which by Plancherel becomes

||1Ω ∗ 1Ω || > (cε)12 N 3

as desired.

With this lemma and use Balog -Szemerédi theorem which asserts
that we can find a subset Ω0 ⊂ Ω with |Ω0 | > cε N and

|Ω0 + Ω0 | 6 Cε N.

Remark. Balog-Szemerédi theorem: Let A = {a1 , a2 , · · · , an } be a


finite set of integers and if ]{(ai , aj , ak , al ) ∈ A4 |ai +aj = ak +al } > cN 3 ,
then there are positive integers c1 and c2 and a subset A0 ⊂ A such that
|A0 | > c1 N and |A0 + A0 | 6 c2 N .
Apply Freiman’s theorem to conclude that Ω0 is contained in a proper
multi-dimensional ZN ×ZN arithmetic progression P1 +P2 +· · ·+Pd = Q

0-29
of dimension dε  1 and cardinality at most Cε0 N . Assume |P1 | > |Pi |
for all 1 6 i 6 d.
Claim that there exists a ∈ ZN × ZN such that

|(a + P1 ) ∩ Ω| > cε Cε0−1 |P1 |,

or else, we have
X
|Ω ∩ Q| 6 |(a + P1 ) ∩ Q|
a∈P2 +···+Pd
X
6 cε Cε0−1 |P1 |
a∈P2 +···+Pd

= cε Cε0−1 |P1 ||P2 | · · · |Pd |


6 cε Cε0−1 Cε0 N
= cε N.

However observe that Ω0 ⊂ Ω ∩ Q we get |Ω ∩ Q| > |Ω0 | > cε N , which


is a contradiction. Fix such a a, since Ω is a graph, a + P1 can not be a

0-30
vertical line segment, hence there exists a ZN arithmetic progression P0
1
of length  N ε , and
d

X
|1̂E∩(E+k) (φ(k))|2
k∈P0
X
= |1̂E∩(E+k) (λk + µ)|2
k∈P0

>c2 ε2 N 2 |P0 |
contradict to Lemma 3.10.
We have already shown (5) , all the remains is to prove lemma 3.10.
Proof of lemma 3.10: In the case λ 6= 0. From (19) and rewrite
the left hand side of (17) as
X X X
exp(2πiu(λk + µ)/N ).
k∈P u∈ZN s∈E∩(E+k)∩(E+u)∩(E+k+u)

Using the identity


X 1 X X
f (u) = f (u + l),
|P |
u∈ZN u∈ZN l∈P

0-31
X X X X X
[ f (u + l) = f (k) 1 = |P | f (k)]
u∈ZN l∈P k∈ZN u∈k−P k∈ZN

we get
X
|1̂E∩(E+k) (λk + µ)|2
k∈P
1 X X X X
= exp(2πi(u + l)(λk + µ)/N )
|P |
k∈P u∈ZN l∈P s∈E∩(E+k)∩(E+u+l)∩(E+k+u+l)
1 X X XX
= 1E+k (s)1E+u+l (s)1E+k+u+l (s) exp(2πi(u + l)(λk + µ)/N
|P |
s∈E u∈ZN k∈P l∈P
1 X X X X
= 1(s−E)∩P (k)1(s−u−E)∩P (l)1(s−u−E)∩(P +P ) (k + l) exp(2
|P |
s∈E u∈ZN k∈ZN l∈ZN
1 X X
6 | f1,s (k)f2,s−u (l)f3,s−u (k + l)|
|P |
s∈E,u∈ZN k∈ZN ,l∈ZN
1 X
6 ||f1,s ||A∞ ||f2,s−u ||2 ||f3,s−u ||2 ,
|P |
s∈E,u∈ZN

0-32
where
f1,s (k) = 1(s−E)∩P (k) exp(2πiφ1 (k)/N ),

f2,s−u (l) = 1(s−u−E)∩P (l) exp(2πiφ2 (l)/N ),

f3,s−u (k + l) = 1(s−u−E)∩(P +P ) (k + l) exp(2πiφ3 (k + l)/N ),

φ1 (k), φ2 (l), φ3 (k + l) are some non-affine quadratic functions such that

φ1 (k) + φ2 (l) + φ3 (k + l) = (u + l)(λk + µ).

Since X
||f2,s−u ||22 =| exp(4πiφ2 (l)/N )| 6 |P |
l∈P ∩(s−u−E)

and
||f3,s−u ||22 6 |P + P | 6 2|P |

0-33
we get
X
2 1 X X X
|1̂E∩E+k (λk + µ)| . |P |||f1,s ||A∞ = N ||f1,s ||A∞ .
|P |
k∈P s∈E u∈ZN s∈E

It suffices to show X
||f1,s ||A∞ = O(εN |P |).
s∈E
From proposition 3.5, we see
X
||f1,s ||A∞
s∈E
X X
= | f1,s (m) exp(2πimn(s)/N )|
s∈E m∈ZN
X X
6 | exp(2πi(φ1 (m) + mn(s))/N )|
s∈ZN m∈(s−E)∩P
X X
= | exp(2πi(φ1 (s − j) + (s − j)n(s))/N )|
s∈ZN j∈(s−P )∩E

=O(εN |P |)

0-34
since φ1 (s − ·) + (s − ·)n(s) is a non-constant quadratic function for every
s.
In the case λ = 0 and µ 6= 0, By a rescaling, we may assume that
the progression P = [0, r). It then suffices to show that
X m
ψ( )|1̂E∩(E+m) (µ)|2 = O(εN 2 r)
r
m∈ZN

for some suitable real even bump function ψ. We can rewrite the left
hand side as
1 X r(n − n0 )
2
rψ̂( )1̂E (n)1̂E (µ − n)1̂E (n0 )1̂E (µ − n0 )
N 0
N
n,n ∈ZN

where ψ̂ is the Fourier transform of ψ on R. From (4) we have


1̂E (n0 )1̂E (µ − n0 ) = O(εN 2 )
and so it suffices to show that
X r(n − n0 )
|ψ̂( )||1̂E (n)||1̂E (µ − n)| = O(N 2 )
0
N
n,n ∈ZN

0-35
But this follows from Schur’s test, Plancherel, and the rapid decay of ψ̂.

0-36