You are on page 1of 52

EFFECTIVE CORRELATION AND DECORRELATION FOR

NEWFORMS, AND WEAK SUBCONVEXITY FOR L-FUNCTIONS

NAWAPAN WATTANAWANICHKUL
arXiv:2405.05249v1 [math.NT] 8 May 2024

(WITH AN APPENDIX BY JESSE THORNER)

Abstract. Let f and g be spectrally normalized holomorphic newforms of even weight


k ≥ 2 on Γ0 (q). If f 6= g, then assume that q is squarefree. For a nice test function ψ
supported on Γ0 (1)\H, we establish the best known bounds (uniform in k, q, and ψ) for
Z Z
dxdy 3 dxdy
ψ(z)f (z)g(z)y k 2 − 1f =g ψ(z) 2 .
Γ0 (q)\H y π Γ0 (1)\H y
When f = g, our results yield an effective holomorphic variant of quantum unique ergodicity,
refining work of Holowinsky–Soundararajan and Nelson–Pitale–Saha. When f 6= g, our
results extend and improve the effective decorrelation result of Huang for q = 1. To prove our
results, we refine Soundararajan’s weak subconvexity bound for Rankin–Selberg L-functions.

1. Introduction
Let Γ0 (1) = SL2 (Z), and consider the hyperbolic surface Y0 (1) = Γ0 (1)\H. The volume
form is dµ = y −2 dxdy, and µ(Y0 (1)) = π/3. Given bounded measurable functions H1 and H2
on Y0 (1), their Petersson inner product is given by
Z
hH1 , H2i = H1 (z)H2 (z)dµ.
Y0 (1)
∂2 ∂2
Let ∆ = −y 2 ( ∂x2 + ∂y2 ) be the hyperbolic Laplacian on Y0 (1). Let {φj }j be a basis of square-
integrable eigenfunctions of ∆ given by the Hecke–Maaß cusp forms, normalized so that {φj }j
is orthonormal with respect to the Petersson inner product. We order the sequence {φj }j so
that the corresponding sequence of Laplace eigenvalues {λj }j monotonically increases.
A fundamental problem in quantum chaos asks for the limiting behavior of φj . Rudnick
and Sarnak [32] conjectured that if ψ is a bounded measurable function on Y0 (1), then
3
lim hψ, |φj |2 i = hψ, 1i. (1.1)
j→∞ π
This is called quantum unique ergodicity (QUE). The limit (1.1) is equivalent to the con-
vergence in the weak-∗ topology of the probability measures |φj |2 dµ to the uniform measure
(3/π)dµ as j → ∞. The combined work of Lindenstrauss [24] and Soundararajan [33] proves
(1.1). As of now, no rate of convergence is known unconditionally. For all ε > 0, a rate of
−1/4+ε
convergence of Oψ,ε (λj ) would follow from the generalized Lindelöf hypothesis.
In this paper, we focus on a holomorphic variant of QUE. Let z = x + iy, and let
X∞ X∞
k−1 k−1
f (z) = λf (n)n 2 e2πinz = e2πiz + λf (n)n 2 e2πinz
n=1 n=2
be a normalized Hecke eigenform of even weight k ≥ 2 and level 1. In other words, f is a
cusp form of weight k and level 1 that is also an eigenfunction of all the Hecke operators with
1
2 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

λf (1) = 1. Consider a sequence of such f with k → ∞. Write Fk (z) = ρf (1)y k/2f (z), where
ρf (1) is chosen such that h1, |Fk |2 i = 1. The analogue of the QUE conjecture in this setting
is that if ψ is a bounded measurable function on Y0 (1), then
3
lim hψ, |Fk |2 i =
hψ, 1i. (1.2)
k→∞ π
Holowinsky and Soundararajan [6, 7, 34] proved that if ε > 0, then
3 1
hψ, |Fk |2 i = hψ, 1i + Oψ,ε ((log k)− 30 +ε ), (1.3)
π
The proof of (1.3) in [7] relies on two different approaches. In [34], Soundararajan spec-
trally expands the left-hand side of (1.3) in terms of Hecke–Maaß cusp forms and unitary
Eisenstein series, bounding the error term using a weak subconvexity bound (see (1.13)). In
[6], Holowinsky expands the left-hand side of (1.3) in terms of incomplete Poincaré series
and incomplete Eisenstein series, bounding the error term by bounding shifted convolution
sums that arise from the expansion. A rate of convergence of Oψ,ε (k −1/2+ε ) follows from the
generalized Lindelöf hypothesis; up to refinements in the k ε , this is optimal (see [13]). Using
(1.2), Rudnick [31] proved that as k → ∞, the zeros of f equidistribute on Y0 (1).
Let Cc∞ (Y0 (1)) be the space of compactly supported, infinitely differentiable functions on
Y0 (1). Let N denote the set of positive integers. Given M ≥ 1, let Cc∞ (Y0 (1), M) equal
n 1 ∂a ∂b o
ψ ∈ Cc∞ (Y0 (1)) : if a, b ∈ N ∪ {0}, then sup y max(a+b, 2 ) a b ψ(z) ≪a,b M a+b . (1.4)
z∈Y0 (1) ∂x ∂y
A close inspection of the work of Lester, Matomäki, and Radziwill [22, Proof of Theorem 1.1]
shows that if M ≥ 1, ψ ∈ Cc∞ (Y0 (1), M), and there exist absolute and effectively computable
constants A > 0 and δ > 0 such that
3
hψ, |Fk |2 i = hψ, 1i + Oε (M A+ε (log k)−δ+ε ), (1.5)
π
then for any fixed B > 0 and any hyperbolic ball B(z0 , r) ⊆ {z ∈ F : Im(z) ≤ B} centered
at z0 with radius r > 0, the following effective variant of Rudnick’s result holds:
#{̺f ∈ B(z0 , r) : f (̺f ) = 0} 3  r 
= µ(B(z0 , r)) + OB,ε δ . (1.6)
#{̺f ∈ F : f (̺f ) = 0} π (log k) 4+A −ε
Following Iwaniec’s course notes [15] on holomorphic QUE, Lester, Matomäki,
√ and Radziwill
[22, Theorem 1.3] proved that (1.5) holds with A = 2 and δ = 31/2 − 4 15 = 0.008066 . . ..
Iwaniec’s course notes present a proof of holomorphic QUE that does not require Soundarara-
jan’s weak subconvexity from [34]. However, an error was found in [15, Section 11], where
Iwaniec attempts to remove the need for weak subconvexity in the spectral decomposition of
the left-hand side of (1.3). This error propagated to the proof [22, Theorem 1.3], rendering
the proof incomplete. See Appendix A for details. Alternatively, [22, Theorem 1.3] can be
proved by refining Soundararajan’s weak subconvexity bound. We discuss this below.
Kowalski, Michel, and Vanderkam [21, Conjecture 1.4] introduced a variant of holomorphic
QUE in which the levels of holomorphic newforms are allowed to vary. For z = x + iy, let

X k−1
f (z) = λf (n)n 2 e2πinz ∈ Sknew (Γ0 (q))
n=1
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 3

be a holomorphic newform of even weight k ≥ 2 and level q ≥ 1 with trivial nebentypus


(see Section 5.1 for more details). We consider a sequence of such f with kq → ∞. Given
bounded measurable functions H1 and H2 on Y0 (q) = Γ0 (q)\H, we define
Z
hH1 , H2 iq = H1 (z)H2 (z)dµ. (1.7)
Y0 (q)

Write
Fk (z) = ρf (1)y k/2f (z), (1.8)
2
where ρf (1) is chosen such that hψ, |Fk (z)| iq = 1. Kowalski, Michel, and Vanderkam conjec-
tured that if ψ is a bounded measurable function supported on Y0 (1), then
3
lim hψ, |Fk |2 iq = hψ, 1i1 . (1.9)
kq→∞ π
Building on the work in [6, 7, 34], Nelson, Pitale, and Saha [28, 29] proved that there exist
constants δ1 , δ2 > 0 such that if q0 is the largest squarefree divisor of q, then
3  q −δ1 
hψ, |Fk |2 iq = hψ, 1i1 + Oψ (log kq)−δ2 .
π q0
Thus, the rate of convergence is sensitive to whether the level q is powerful.
Let f and g be holomorphic newforms of weight k and level q, and let Gk be defined
similarly to (1.8). In analogy with the variations on holomorphic QUE mentioned above, it
is natural to conjecture that if ψ is a bounded measurable function supported on Y0 (1), then
lim |hψ, Fk Gk iq | = 0. (1.10)
kq→∞

Huang [6] proved that if f and g are normalized Hecke eigenforms of weight k and level 1,
with f 6= g, M ≥ 1, and ψ ∈ Cc∞ (Y0 (1), M), then for all ε > 0, we have
5 −41 +ε
hψ, Fk Gk i1 ≪ε M 3 (log k)−1.19×10 . (1.11)
Constantinescu [3] studied another variation of the decorrelation problem considering nor-
malized Hecke eigenforms f and g of level 1 and weights k1 and k2 , respectively. He showed
that the decorrelation values dissipate as max(k1 , k2 ) tends to infinity.
Now, consider holomorphic newforms f and g of weight k and level q. If f 6= g, then we
assume that q is squarefree. In this paper, we establish the strongest known unconditional
upper bound on |hψ, Fk Gk iq − 1f =g π3 hψ, 1i1 |. When f = g, this addresses the rate of conver-
gence in (1.9) with effective dependence on k, q, and ψ. When f 6= g, this addresses the rate
of convergence in (1.11) and extends such result to squarefree levels. For f = g and q = 1,
our result provides a faster rate of convergence than in (1.3) and [22, Theorem 1.3].
Theorem 1.1. Let q ≥ 1 be an integer, and let q0 be the largest squarefree divisor of q.
Let k ≥ 2 be an even integer. Let f and g be holomorphic newforms of weight k, level q,
and trivial nebentypus. Let Fk and Gk be defined as in (1.8). Assume that if f 6= g, then
q is squarefree. Let θ ∈ [0, 7/64] denote the best bound towards the generalized Ramanujan
conjecture for Hecke–Maaß cusp forms. If ε > 0, M ≥ 1, and ψ ∈ Cc∞ (Y0 (1), M), then
3  − 121 (2√3−3)(1−2θ)+ε √
4 q 7
hψ, Fk Gk iq − 1f =g hψ, 1i1 ≪ε M (log kq)−( 2 −2 3)+ε .
π q0

If q = 1 and f = g, then we can replace M 4 with M (23−2 3)/12
.
4 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

Corollary 1.2. Let f be a normalized à Hecke eigenform of even weight kó 2 and level 1.
1
Equation (1.6) holds with A = 12 (23 − 2 3) = 1.627991 . . . and δ = 72 − 2 3 = 0.035898 . . ..
In particular, if B > 0 is fixed, then the zeros of f equidistribute with respect to dµ within
any hyperbolic ball B(z0 , r) ⊆ {z ∈ F : Im(z) ≤ B} of radius r ≥ (log k)−1/157 .
The main new result that enables us to prove Theorem 1.1 is a refinement of Soundarara-
jan’s weak subconvexity bound. To describe this refinement, let m ≥ 1 be an integer, and let
Fm be the set of all cuspidal automorphic representations of GLm (AQ ) with unitary central
character. For convenience, we normalize each π ∈ Fm so that the central character is trivial
on the diagonally embedded copy of the positive reals. Consider π ∈ Fm and π ′ ∈ Fm′ . Let
C(π × π ′ ) denote the analytic conductor of the Rankin–Selberg L-function L(s, π × π ′ ) (see
Section 3.2 for the definitions). Let ε > 0. The convexity bound for L(s, π × π ′ ) is
1
L( 12 , π × π ′ ) ≪m,m′ ,ε C(π × π ′ ) 4 +ε , (1.12)
while the generalized Riemann hypothesis implies that
L( 21 , π × π ′ ) ≪m,m′ ,ε C(π × π ′ )ε .
If the exponent 1/4 + ε in (1.12) could be replaced by any fixed exponent less than 1/4 (such
a bound is called a subconvexity bound), then Theorem 1.1 would hold with a power saving
rate of convergence in both k and q. Currently, no such exponent is known to exist.
The proofs in [7, 9, 28, 29] crucially rely on saving a power of log C(π × π ′ ) over (1.12).
Soundararajan [34] proved the following weak subconvexity bound: If π satisfies the generalized
Ramanujan conjecture (GRC), then for any ε > 0,
1
′ C(π × π ′ ) 4
L( 12 , π × π ) ≪m,π′ ,ε . (1.13)
[log C(π × π ′ )]1−ε
In particular, if π satisfies GRC, then (1.13) provides a logarithmic improvement over (1.12)
when π ′ is fixed. In order to explicate the dependence on ψ as in Theorem 1.1, one must first
explicate the dependence on π ′ in (1.13). Alternatively, when π satisfies GRC, one might
appeal to the bound
1
′ C(π × π ′ ) 4
L( 21 , π × π ) ≪m,m′ ,
[log C(π × π ′ )]1/(1017 (mm′ )3 )
proved by Soundararajan and Thorner [35]. While this bound is fully uniform in both π and
π ′ , the power of log C(π × π ′ ) that it saves is quite small. In this paper, we prove the following
refinement of (1.13), which we expect to be useful in settings beyond what we consider here.
Theorem 1.3. Let π ∈ Fm and π ′ ∈ Fm′ . Assume that L(s, π × π ′ ) is entire and π satisfies
GRC.1 If 0 < ε < 1, then
1
C(π × π ′ ) 4
L( 21 , π × π ′ ) ≪m,m′ ,ε C(π ′ )ε .
[log C(π × π ′ )]1−ε
Remark. With some additional work, one can show that there exists an effectively computable
constant c1 = c1 (m, m′ ) > 0 such that the factor C(π ′ )ε in Theorem 1.3 can be refined to
exp[c1 (log C(π ′ × π
e′ ))/ log log C(π ′ × π
e′ )].
1Our proof shows that we only require π to satisfy GRC at the non-archimedean places.
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 5

Acknowledgements. The author would like to thank Jesse Thorner for suggesting this
project and providing invaluable feedback that significantly improved this paper. The au-
thor also thanks Peter Humphries, Maksym Radziwill, and Bingrong Huang for very helpful
discussion and comments on the manuscript.

2. Outline of the paper


In Section 3, we record some properties of standard L-functions and Rankin–Selberg L-
functions, and establish useful bounds for partial sums of Dirichlet coefficients.
In Section 4, we prove Theorem 1.3.
In Section 5, we discuss some background on GL(2) automorphic forms, including holo-
morphic newforms, Hecke–Maaß cusp forms, Eisenstein series, incomplete Eisenstein series,
and incomplete Poincaré series. We also recall the spectral decomposition and the Poincaré
decomposition.
In Section 6, we prove Theorem 1.1 assuming Propositions 6.1 and 6.2.
In Section 7, we establish useful lower bounds for the adjoint lift of a holomorphic newform.
We also prove bounds for L-functions that will be useful for proving Propositions 6.1.
In Section 8, we prove Propositions 6.1 and 6.2.

3. Properties of L-functions
In this section and for the rest of the paper, if ν is a parameter, then we write f (z) =
Oν (g(z)), or equivalently, f (z) ≪ν g(z), if there exists an effectively computable constant c,
depending at most on m, m′ , ε, and ν, such that |f (z)| ≤ c|g(z)| for all z under consideration.
If no ν is present, then c depends at most on m, m′ , and ε.
3.1. Standard L-functions. Let m ≥ 1 be an integer, let AQ denote the ring of adèles
over Q, and let Fm be the set of all cuspidal automorphic representations of GLm (AQ ) whose
central characters are normalized to be trivial on the positive reals. If π ∈ Fm , then for
each place v of Q, there exists a smooth admissible representation πv of GLm (Fv ) such that
π can be written as the restricted tensor product ⊗v πv . Given π ∈ Fm , let πe ∈ Fm be the
contragredient representation and Nπ be the conductor of π. When v is non-archimedean
and corresponds with a prime p, then the local L-function L(s, πp ) is defined in terms of the
Satake parameters Aπ (p) = {α1,π (p), . . . , αm,π (p)} by
Ym X∞
−s −1 λπ (pk )
Lp (s, π) = (1 − αj,π (p)p ) = ks
. (3.1)
j=1 k=0
p
If p ∤ Nπ , then for all j, we have that αj,π (p) 6= 0. If p | Nπ , it might be the case that there
exists j such that αj,π (p) = 0. The standard L-function L(s, π) associated to π is
Y X∞
λπ (n)
L(s, π) = Lp (s, π) = .
p n=1
ns
The Euler product and Dirichlet series converge absolutely when Re(s) > 1.
At the archimedean place of Q, there are m Langlands parameters µπ (j) ∈ C from which
we define
Ym  s + µ (j) 
− ms π
L∞ (s, π) = π 2 Γ .
j=1
2
6 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

Luo, Rudnick, and Sarnak [25] and Müeller and Speh [27] proved that there exists θm ∈
[0, 12 − m21+1 ] such that we have the uniform bounds
|αj,π (p)| ≤ pθm and Re(µj,π (v)) ≥ −θm , (3.2)
and GRC asserts that in (3.2), one may take θm = 0. By the combined work of Kim and
Sarnak [20, Appendix 2] and Blomer and Brumley [1], we have the bound θ2 ≤ 7/64. We
have Nπ = Nπe , and we have the equalities of sets
{αj,eπ (p)} = {αj,π (p)}, {µj,eπ (v)} = {µj,π (v)}.
Let rπ be the order of the pole of L(s, π) at s = 1. The completed L-function
Λ(s, π) = (s(1 − s))rπ Nπs/2 L(s, π)L∞ (s, π)
is entire of order 1, and there exists a complex number W (π) of modulus 1 such that for all
s ∈ C, we have the functional equation Λ(s, π) = W (π)Λ(1 − s, π e). The analytic conductor
of π [13] is given by
m
Y
C(π, t) := Nπ (3 + |it + µπ (j)|), C(π) := C(π, 0). (3.3)
j=1

We define aπ (pk ) by the Dirichlet series identity


XX ∞ Pm k XX ∞
L′ j=1 αj,π (p) log p aπ (pk ) log p
− (s, π) = = , Re(s) > 1.
L p k=1
pks p k=1
pks

3.2. Rankin–Selberg L-functions. Let π ∈ Fm and π ′ ∈ Fm′ . For each p ∤ Nπ Nπ′ , define
m Y
m ′
Y 1

Lp (s, π × π ) = −s
.

j=1 j =1
1 − α j,π (p)α j ′ ,π ′ (p)p

Jacquet, Piatetski-Shapiro, and Shalika [17] proved the following theorem.


Theorem 3.1. If (π, π ′ ) ∈ Fm × Fm′ , then there exist
m′
(1) complex numbers (αj,j ′,π×π′ (p))m
j=1 j ′ =1 for each prime p | Nπ Nπ ′ , from which we define

m Y
m ′ m Y
m ′
Y 1 Y 1
′ ′
Lp (s, π × π ) = −s
, Lp (s, π
e×π
e)= ;
j=1 j ′ =1
1 − αj,j ′,π×π ′ (p)p
j=1 j ′ =1 1 − αj,j ′,π×π′ (p)p−s

(2) complex numbers (µπ×π′ (j, j ′ ))m m
j=1 j ′ =1 , from which we define

′s Ym Y m′  s + µ ′ (j, j ′ ) 
′ − mm π×π
L∞ (s, π × π ) = π 2 Γ ,

j=1 j =1
2

′s Ym Ym′  s + µ ′ (j, j ′ ) 
′ − mm π×π
L∞ (s, π
e×π
e)=π 2 Γ ;
j=1 ′
2
j =1

(3) an arithmetic conductor, denoted as Nπ×π′ ; and


(4) a complex number W (π × π ′ ) of modulus 1
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 7

such that the Rankin–Selberg L-function


Y X∞
′ ′ λπ×π′ (n)
L(s, π × π ) = Lp (s, π × π ) =
p n=1
ns
converges absolutely for Re(s) > 1 with a pole at s = 1 of order
(
1 if π ′ = π
e,
rπ×π′ =
0 otherwise,
the completed L-function
s/2
Λ(s, π × π ′ ) = (s(1 − s))rπ×π′ Nπ×π′ L(s, π × π ′ )L∞ (s, π × π ′ ) (3.4)
is entire of order 1, and Λ(s, π × π ′ ) = W (π × π ′ )Λ(1 − s, π e′ ).
e×π
We can explicitly determine the numbers µπ×π′ (j, j ′ ) (resp. αj,j ′,π×π′ (p) at primes p | Nπ Nπ′ )
using the archimedean case of the local Langlands correspondence [27] (resp. [35, Appendix]).
These descriptions and (3.2) yield the bounds
|αj,j ′,π×π′ (p)| ≤ pθm +θm′ , Re(µπ×π′ (j, j ′ )) ≥ −(θm + θm′ ). (3.5)
If k ≥ 1 is an integer, then we define
(
aπ (pk )aπ′ (pk ) if p ∤ Nπ Nπ′ ,
aπ×π′ (pk ) = Pm Pm′ k
(3.6)
j=1 j=1 αj,j ′ ,π×π ′ (p) otherwise.
We define Λπ×π′ (n) by the Dirichlet series identity
XX ∞ ∞
L′ ′ aπ×π′ (pk ) log p X Λπ×π′ (n)
− (s, π × π ) = = , Re(s) > 1. (3.7)
L p
pks n=1
ns
k=1

As with L(s, π), we define the analytic conductor


m Y
m ′
Y

C(π × π , t) := Nπ×π′ (3 + |it + µπ×π′ (j, j ′ )|), C(π × π ′ ) := C(π × π ′ , 0). (3.8)
j=1 j ′ =1

The combined work of Bushnell and Henniart [9] and Brumley [11, Appendix] yields
′ ′
C(π × π ′ , t) ≪ C(π × π ′ )(3 + |t|)mm , C(π × π ′ ) ≪ C(π)m C(π ′ )m . (3.9)
3.3. Bounds for partial sums of Dirichlet coefficients. Consider π ∈ Fm and π ′ ∈ Fm′ .
Brumley [35, Appendix] and Jiang, Lü, and Wang [18] proved useful inequalities for the
coefficients Λπ×π′ (n) and λπ×π′ (n), respectively.
Lemma 3.2 (Lemma 2.2 in [35]). Let π ∈ Fm and π ′ ∈ Fm′ . If n ≥ 1 is an integer, then
p
|Λπ×π′ (n)| ≤ Λπ×eπ (n)Λπ′ ×eπ′ (n).
Lemma 3.3 (Lemma 3.1 in [18]). Let π ∈ Fm and π ′ ∈ Fm . If n ≥ 1 is an integer, then
p
|λπ×π′ (n)| ≤ λπ×eπ (n)λπ′ ×eπ′ (n).
Lemmas 3.2 and 3.3 are useful to us because we assume that π satisfies GRC. To handle
the contribution from π ′ , we use two results, the first of which is due to Li.
8 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

Theorem 3.4 (Theorem 2 in [23]). There exists an absolute and effectively computable con-
stant c2 > 0 such that if π ′ ∈ Fm′ and 1 < σ ≤ 3, then
 ′
e′ ) 
′ ′ −1 ′ 2 log C(π × π
L(σ, π × πe ) ≪ (σ − 1) exp c2 m ′ ′
≪ (σ − 1)−1 C(π ′ )ε .
log log C(π × π e)
The second result is a Brun–Titchmarsh type upper bound with mild dependence on π ′ .
Lemma 3.5. Let π ∈ Fm and π ′ ∈ Fm′ . If x ≥ 2 and π satisfies GRC, then
X p
|Λπ×π′ (n)| ≪ x log C(π ′ ).
n≤x

Proof. By Lemma 3.2 and the Cauchy–Schwarz inequality, we have that


X X  12  X  12
|Λπ×π′ (n)| ≤ Λπ×eπ (n) Λπ′ ×eπ′ (n) . (3.10)
n≤x n≤x n≤x

Since π satisfies GRC, it follows from the prime number theorem that
X  21  X  12 √
Λπ×eπ (n) ≤ m2 Λ(n) ≪ x.
n≤x n≤x

We now estimate the second factor of (3.10). Fix a compactly supported, infinitely differ-
entiable function φ with φ(j) (y) ≪j 1 such that φ is supported on [1/2, 5/2] and is at least 1
on [1, 2]. It follows that
X ∞
X
Λπ′ ×eπ′ (n) ≤ Λπ′ ×eπ′ (n)φ(n/x). (3.11)
x≤n≤2x n=1
′ ′
Let ρ range over all zeros of L(s, π × π
e ). Applying the Mellin inversion formula and pushing
the contour all the way to the left, we find that
X∞ Z 3+i∞ X
1 L′ b ds = xφ(1)
b − b
Λπ′ ×eπ′ (n)φ(n/x) = − (s, π ′ × π
e′ )xs φ(s) xρ φ(ρ). (3.12)
n=1
2πi 3−i∞ L ρ

By (3.2) and (3.5), for any integer k ≥ 0, there are m′2 trivial zeros with Re(ρ) ≤ −2k + 1.
Then the contribution from trivial zeros is ≪ x. For nontrivial zeros, we have that |xρ | ≤ x
b ≪ 1. Hence, (3.12) becomes
trivially, and φ(1)
X ∞ X
Λπ′ ×eπ′ (n)φ(n/x) ≪ x + b
xφ(ρ). (3.13)
n=1 ρ
P b
Now, it remains to bound ρ φ(ρ). For any real number t, we have
|{ρ = β + iγ : |γ − t| ≤ 1}| ≪ log C(π ′ × π
e′ , t) ≪ log C(π ′ × π
e′ ) + log(|t| + 3).
Write ρ = β + iγ. By integration by parts and the derivative estimates of φ, we also have
b
that |φ(ρ)| ≪φ (|γ|2 + 1)−1 . Therefore,
X Z ∞
b ≪ dt
φ(ρ) (log C(π ′ × π
e′ ) + log(|t| + 3)) 2 ≪ log C(π ′ ). (3.14)
ρ −∞ |t| + 1
P
By (3.11), (3.13), and (3.14), if x ≥ 2, then x≤n≤2x Λπ′ ×eπ′ (n) ≪ x log C(π ′ ). Therefore, the
lemma follows upon using dyadic decomposition and (3.10). 
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 9

For future convenience, we record some helpful bounds obtained from applying Lemma 3.3,
the Cauchy–Schwarz inequality, and Theorem 3.4.
Lemma 3.6. Let π ∈ Fm and π ′ ∈ Fm′ . Let π satisfy GRC. If x ≥ 2 and 1 < σ ≤ 2, then
X |λπ×π′ (n)| m2 +1
≪ (log x) 2 C(π ′ )ε , (3.15)
n≤x
n
m2 +1
|L(σ + it, π × π ′ )| ≪ (σ − 1)− 2 C(π ′ )ε , (3.16)
m2 +3
|L′ (σ + it, π × π ′ )| ≪ (σ − 1)− 2 C(π ′ )ε . (3.17)
Proof. We first prove (3.16) Appealing to Lemma 3.3 and the Cauchy–Schwarz inequality, we
have that
X∞ X∞ p
λ π×π ′ (n) λπ×eπ (n)λπ′ ×eπ′ (n)
|L(σ + it, π × π ′ )| = σ+it

n=1
n n=1

X (3.18)
λπ×eπ (n)  12  X λπ′ ×eπ′ (n)  12
∞ ∞
≤ .
n=1
nσ n=1

Let τd (n) denote the d-divisor function, the number of ordered factorizations of n into d
positive integers. Since π ∈ Fm satisfies GRC, we have λπ×eπ (n) ≤ τm2 (n). It follows that
X

λ  12 X
∞  12
π (n)
π×e τ m2 (n) m2 m2
≤ = ζ(σ) 2 ≪ (σ − 1)− 2 . (3.19)
n=1
nσ n=1

By Theorem 3.4, we have that


X

λ  21
π ′ (n)
π ′ ×e 1 1
e′ ) 2 ≪ (σ − 1)− 2 C(π ′ )ε .
= L(σ, π ′ × π (3.20)
n=1

The assertion (3.16) then follows by inserting the bounds from (3.19) and (3.20) into (3.18).
To prove (3.17) , we combine (3.16) with Cauchy’s integral formula for L′ (σ + it, π × π ′ ).
Lastly, to prove (3.15) , we first observe that
X |λπ×π′ (n)| X∞
|λπ×π′ (n)|
≤e 1+1/ log(ex)
.
n≤x
n n=1
n

The result now follows from (3.16) and (3.18)–(3.20) (each with σ = 1 + 1/ log(ex)). 
Lemma 3.7. Let π ∈ F(m) and π ′ ∈ F(m′ ). Let π satisfy GRC. If x ≥ 2, then
X m2 +1
|λπ×π′ (n)| ≪ x(log x) 2 C(π ′ )ε . (3.21)
n≤x

Moreover, if 1 ≤ y ≤ x, then
X m4 +2
|λπ×π′ (n)| ≪ x3/4 y 1/4 (log x) 4 C(π ′ )ε . (3.22)
x≤n≤x+y
10 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

Proof. The first assertion follows from (3.15) of Lemma 3.6 and Rankin’s trick:
X X |λπ×π′ (n)| m2 +1
|λπ×π′ (n)| ≪ x ≪ x(log x) 2 C(π ′ )ε .
n≤x n≤x
n
For the second assertion, it follows from Lemma 3.3 and the Cauchy–Schwarz inequality that
X  X  21  X  12
|λπ×π′ (n)| ≤ λπ×eπ (n) λπ′ ×eπ′ (n) .
x<n≤x+y x<n≤x+y x<n≤x+y

If π satisfies GRC, then λπ×eπ (n) ≤ τm2 (n). Since λπ×eπ (n) is nonnegative, we use the Cauchy–
Schwarz inequality and Rankin’s trick to find that
 X  21 1
 X  14 1
 X τ 2 (n)2  14
2 m
λπ×eπ (n) ≤ y 4 λπ×eπ (n) ≤ (xy) 4 .
x<n≤x+y x<n≤x+y x<n≤x+y
n

Since τm2 (n)2 ≤ τm4 (n), the above display is


1
 X τ 4 (n)  41 1
X∞
τm4 (n)  41
m
≤ (xy) 4 ≤ e(xy) 4
x<n≤x+y
n n=1
n1+1/ log(2ex)
1 m4 1 m4
= e(xy) 4 ζ(1 + 1/ log(2ex)) 4 ≪ (xy) 4 (log x) 4 .
Since λπ′ ×eπ′ (n) is nonnegative, we apply Rankin’s trick and Theorem 3.4 to obtain
 X  21 1
 X λ ′ ′ (n)  21 1
 X∞
λπ′ ×eπ′ (n)  21
π ×e
π
λπ′ ×eπ′ (n) ≤ x2 ≤ x2 e
x<n≤x+y n≤2x
n n=1
n1+1/ log(2ex)
1 1 1 1
= (ex) 2 L(1 + 1/ log(2ex), π ′ × π
e′ ) 2 ≪ x 2 (log x) 2 C(π ′ )ε .
The desired result follows. 

4. Weak subconvexity bound


In this section, we establish our weak subconvexity bound as stated in Theorem 1.3. Our
proof closely follows the framework introduced by Soundararajan in [34]. Consequently,
we shall exclude specific details mirroring Soundararajan’s method and instead focus on
highlighting the distinctive refinements in our approach.
4.1. Notation. Consider π ∈ Fm and π ′ ∈ Fm′ . In this section, we assume that L(s, π × π ′ )
is entire, and π satisfies GRC. For clarity, we provide a list of the conventions and notation
that we will use throughout this section.
• C = C(π × π ′ ).
• x ≥ 2. P
• S(x) = n≤x λπ×π′ (n).
e P
• S(x) = n≤x λπ×π′ (n) log n.
• ε ∈ (0, 1].
• R = ⌊72m2 /ε2⌋ + 1.
• L = ⌊5(m2 + 1)R⌋.
• L = (L, . . . , L).
• X ≥ max(x, 10).
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 11

• w ∈ [2, exp((log X)1/(3R) )] is an integer.


• T = exp((log log X)2 ).
• τ = (τ1 , τ2 , . . . , τR ) ∈ RR with |τj | ≤ T .
• l1 , . . . , lR , j1 , . . . , jR ∈ Z are nonnegative.
• l = (l1 , . . . , lR ) and j = (j1 , . . . , jR ).
• j ≤ l indicates that 0 ≤ j1 ≤ l1 , . . . , 0 ≤ jR ≤ lR .
l
 l1
 lR
 a

• j
= j1
· · · jR
, where b
are the binomial coefficients.
P 
• Ol (x, w) = Ol (x, w; τ ) = j≤l (−1)j1 +···+jR jl w j1(1+iτ1 )+···+jR (1+iτR ) S(x/w j1 +···+jR ).
P  j (1+iτ )+···+j (1+iτ )
• Oel (x, w) = Ol (x, w; τ ) = j1 +···+jR l R e
S(x/w j1 +···+jR ).
j≤l (−1) w1 1 R
j
Q
• Ll (s) = L(s, π × π ′ ) R j=1 (1 − w
1+iτj −s lj
) .

4.2. Main results. As discussed in Soundararajan’s work [34], the essence of his weak sub-
convexity proof, and consequently our proof, is the slow variation of the mean values of
multiplicative functions. In our case, we consider the multiplicative function λπ×π′ (n). Writ-
ing C := C(π × π ′ ), we will use the convexity bound for L(s, π × π ′ ) (see Lemma 4.4) to prove
that there exists a constant B0 = B0 (m, m′ ) > 0 such that
X x 1
λπ×π′ (n) ≪ C(π ′ )ε , x ≥ C 2 (log C)B0 .
n≤x
log x
1
Next, we prove that for all B > 0, similar cancellation holds even when x ≥ C 2 (log C)−B . In
particular, we have the following theorem.
Theorem 4.1. Let π and π ′ be as in Theorem 1.3 and C = C(π × π ′ ). If ε > 0, B > 0, and
1
x ≥ C 2 (log C)−B , then
X x
λπ×π′ (n) ≪B 1−ε
C(π ′ )ε .
n≤x
(log x)

Recall the notation in Section 4.1. Theorem 4.1 follows from our next theorem.
Theorem 4.2. With the notation in Section 4.1, there exist real numbers τ1 , . . . , τR , whose
absolute values are at most exp((log log X)2 ), such that if 2 ≤ x ≤ X, then
x
|OL (x, w; τ1 , . . . , τR )| ≪ (log X)ε C(π ′ )ε .
log x
4.3. Deduction of the weak subconvexity bound. In this section, we prove Theorems 4.1
and 1.3. We start with some useful lemmas. The first lemma is a result from Soundararajan
and Thorner in [35].
Lemma 4.3. Let π ∈ Fm and π ′ ∈ Fm′ , and let π satisfy GRC. We have that
1 mm′
L( 12 + it, π × π ′ ) ≪ C(π × π ′ ) 4 (|t| + 1) 4 .
Proof. By [35, Corollary 1.3], we have
1
′ ′ 2 C(π × π ′ ) 4
L( 12 , π ×π) ≪ |L( 23 , π × π )| .
[log C(π × π ′ )]1/(1017 (mm′ )3 )
12 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

The above result is uniform in the full analytic conductor. Hence, by recognizing the L-
function L(s + it, π × π ′ ) as L(s, π × (π ′ ⊗ | · |it )), we obtain the bound
1
′ ′ 2 C(π × π ′ , t) 4
L( 21 + it, π × π ) ≪ |L( 23 + it, π × π )| . (4.1)
[log C(π × π ′ , t)]1/(1017 (mm′ )3 )
We now write
X∞ 
Λπ×π′ (n) 

|L( 23
+ it, π × π )| = exp Re 3
+it
. (4.2)
n=2 n log n
2

By triangle inequality, Lemma 3.2, and the Cauchy–Schwarz inequality, (4.2) is


 X∞
Λπ×eπ (n)  21  X Λπ′ ×eπ′ (n)  12 

≤ exp 1+ ′21 2− ′21
. (4.3)
n=2 n m +1 log n
n=2 n m +1 log n

Since π satisfies GRC, it follows from (3.2) and (3.5) that (4.3) is O(1), as desired. 

PAs discussed in Section 4.2, the next lemma gives an upper bound for the partial summation
n≤x λπ×π ′ (n) in the standard range of x, which can be obtained via the convexity bound.

Lemma 4.4. Let π ∈ Fm and π ′ ∈ Fm′ . Assume that L(s, π × π ′ ) is entire and π satisfies
1 5 ′
GRC. If C = C(π × π ′ ), in the range x ≥ x0 := C 2 (log C)36m m , we have
X x
λπ×π′ (n) ≪ C(π ′ )ε .
n≤x
log x

Proof. One can follow the proof of [34, Lemma 4.2] almost verbatim. Using the same smooth-
ing kernel, we have that, for any c > 1, x > 0, λ > 0, and any natural number K,
Z c+i∞ s  λs
1 ′ x e − 1 K X  X 
L(s, π × π ) ds = λπ×π (n) + O
′ |λπ×π (n)| . (4.4)

2πi c−i∞ s λs n≤x Kλ x<n≤e x
−(m4 +6)
We shall take K = ⌊mm′ /4⌋ + 2 and λ = (log x) . By (3.22) of Lemma 3.7, the error
term above is
1 m4 +2 x
≪ (eKλ − 1) 4 x(log x) 4 C(π ′ )ε ≪ C(π ′ )ε .
log x
For the left-hand side of (4.4), we move the line of integration to c = 21 and appeal to
Lemma 4.3. The integral is then
Z ∞
1 1
−K mm′ dt 1 1 5 ′
≪ C 4x2λ (1 + |t|) 4 K+1
≪ C 4 x 2 (log x)16m m .
−∞ (1 + |t|)
Putting together the bound for the error term and the above bound in (4.4), we have
X 1 1 5 ′ x
λπ×π′ (n) ≪ C 4 x 2 (log x)16m m + C(π ′ )ε .
n≤x
log x
1 5 m′
If x ≥ C 2 (log C)36m , then the second term dominates. 
Proof of Theorem 4.1, assuming Theorem 4.2. Let x0 be as in Lemma 4.4, and let x0 ≥ x ≥
1
C 2 (log C)−B for any positive constant B. Take w = x0 /x and X = xw LR . Applying Theorem
4.2 for λπ×π′ (n), we obtain
X
|OL (X, w)| ≪ C(π ′ )ε . (4.5)
(log X)1−ε
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 13

By definition, the left-hand side of (4.5) is


X  LR−1
X X 
LR
w λπ×π′ (n) + O wj λπ×π′ (n) . (4.6)
n≤X/w LR j=0 n≤X/w j

Since for 0 ≤ j ≤ LR − 1 we have X/w j ≥ xw = x0 , Lemma 4.4 applies for the range of the
error term above, giving the bound O(C(π ′)ε X/ log X). From 4.5 and 4.6, we conclude that
X X x
λπ×π′ (n) ≪ w −LR 1−ε
C(π ′ )ε ≪ 1−ε
C(π ′ )ε . 
n≤x
(log X) (log x)

Proof of Theorem 1.3. We use an approximation functional equation for L( 21 , π × π ′ ) and


partial summation. Following the proof of [34, Theorem 1], we have that, for c > 1/2
Z c+i∞ 1 ′
1 ′ 1 1 ′ L∞ (s + 2 , π × π ) s2 ds
L( 2 , π × π ) = L(s + 2 , π × π ) e
2πi c−i∞ L∞ ( 12 , π × π ′ ) s
′ Z c+i∞ 1
e′ ) s2 ds
W (π × π ) 1 ′ L∞ (s + 2 , π e×π
+ L(s + 2 , π e×π e) e . (4.7)
2πi c−i∞ L∞ ( 21 , π
e×πe′ ) s
Both integrals in (4.7) can be estimated similarly, so we only consider the first one. By partial
summation, we have
Z ∞X
1 ′ 1 dx
L(s + 2 , π × π ) = (s + 2 ) λπ×π′ (n) s+ 3 .
1 n≤x x 2
Therefore, the first integral in (4.7) equals
Z ∞X  1 Z c+i∞
1
L∞ (s + 21 , π × π ′ ) s2 −s ds  dx
λπ×π′ (n) (s + 2 ) e x . (4.8)
1 n≤x
2πi c−i∞ L∞ ( 21 , π × π ′ ) s x 32

To estimate the inner integral of (4.8), we move the line of integration either to Re(s) = 2,
θ +θ
or to Re(s) = m 2 m′ . Recall that C = C(π × π ′ ). By Stirling’s formula, the inner integral is
n√ θm +θm′ √ o
≪ min ( C/x) 2 , ( C/x)2 .

Thus, (4.8) is
Z √ 2 +3)/(1−θ −θ
C/(log C)(m m m′
)
X
(θm +θm′ )/4 dx
≪C λπ×π′ (n)
1 n≤x
x(3+θm +θm′ )/2
Z √
C X dx
+ C (θm +θm′ )/4 √
λπ×π′ (n)
2
C/(log C)(m +3)/(1−θm −θm′ ) n≤x
x(3+θm +θm′ )/2
Z ∞ X dx
+C √
λπ×π′ (n) 7 .
C n≤x x2
1
The first integral is ≪ C(π ′ )ε C / log C by (3.21) of Lemma 3.7. The second and third integrals
4
1
are ≪ C(π ′ )ε C 4 /(log C)1−ε by Theorem 4.1. The desired result follows. 
14 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

4.4. Successive maxima. In this subsection, we define the points τ1 , . . . , τR appearing in


Theorem 4.2. We define τ1 as the point t in the compact set S1 = [−T, T ] where the maximum
1
of |L(1 + 1/ log X + it, π × π ′ )| is attained. We now remove the interval (τ1 − (log X)− R , τ1 +
1
(log X)− R ) from S1 , and let S2 denote the remaining compact set. We define τ2 to be the
point in S2 where the maximum |L(1 + 1/ log X + it, π × π ′ )| is attained. We continue defining
τ1 , τ2 , . . . , τR similarly and obtain the nested compact sets S1 ⊃ S2 ⊃ . . . ⊃ SR . We observe
1
that if j 6= k, then |τj − τk | ≥ (log X)− R .
Incorporating the above construction, we obtain the following improved version of (3.16)
of Lemma 3.6.
Lemma 4.5. Recall the notation in Section 4.1. Let 1 ≤ j ≤ R and t be a point in Sj . If
ε > 0, then q
3m 1
+ j−1 + ε
|L(1 + 1/log X + it, π × π ′ )| ≪ C(π ′ )ε (log X) 2 j jR 4 .
In particular, if t ∈ SR , we have |L(1 + 1/ log X + it, π × π ′ )| ≪ C(π ′ )ε (log X)ε/2 .
Proof. We apply the same technique as in the proof of Lemma 3.6 and follow the outline of
the proof of [34, Lemma 3.2]. Without any modification, [34, Lemma 3.2] gives
 1 X Λπ×π′ (n) Xj 
′ −iτr
|L(1 + 1/ log X + iτj , π × π )| ≤ exp Re n . (4.9)
j n≥2 n1+1/ log X log n r=1
By Lemma 3.2 and the Cauchy–Schwarz inequality, we have that
j p j
X |Λπ×π′ (n)| X X Λπ×eπ (n)Λπ′ ×eπ′ (n) X
−iτr
1+1/ log X
n ≤ 1+1/ log X
n−iτr
n≥2
n log n r=1 n≥2
n log n r=1
(4.10)
X Λπ×eπ (n) Xj
2  21  X Λπ′ ×eπ′ (n)  21
−iτr
≤ n .
n≥2
n1+1/ log X log n r=1 n≥2
n1+1/ log X log n

Since π satisfies GRC and Λπ×eπ (n) is supported on prime powers, we obtain the bound
Λπ×eπ (n) ≤ m2 Λ(n). Thus, (4.10) is
X Λ(n) Xj
2  12
−iτr ′ ′
≤m 1+1/ log X log n
n log L(1 + 1/log X, π × π
e ) . (4.11)
n≥2
n r=1

By Theorem 3.4, we have that


log C(π ′ × π
e′ )
log L(1 + 1/log X, π ′ × π
e′ ) ≤ log log X + c1 m′2 + O(1). (4.12)
log log C(π ′ × πe′ )
On the other hand, by the calculations on [34, p. 1488], the double sum in (4.11) equals
X
j(log log X + O(1)) + 2 log |(log X)1/R + O(1)|
1≤r<r ′ ≤j
 (4.13)
j(j − 1) 
≤ j+ log log X + O(j).
R
Using (4.11)–(4.13), we bound (4.9) by
h m  j(j − 1)   12  log C(π ′ × π
e′ )  21 i
exp j+ log log X + O(j) log log X + c1 m′2 + O(1) .
j R log log C(π ′ × π
e′ )
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 15
√ √ √ √
For numbers a, b, c ≥ 0, the bound a + b + c ≤ a + b + c holds. Applying this twice,
we bound (4.9) by
" s s !
1 j −1 p log C(π ′×π e ′)
exp m + log log X + log log X c1 m′ 2
j jR log log C(π ′ × π e′ )
s ! s !# (4.14)
1 j −1 m p mm ′ ′
log C(π × π ′
e)
+O m + + log log X + √ c1 ′
.
j jR j j log log C(π × π e′ )
The inequality of arithmetic and geometric means yields
s
p log C(π ′ × π
e′ ) 1 c1 m′2 log C(π ′ × π
e′ )
log log X c1 m′2 ≤ log log X + . (4.15)
log log C(π ′ × π
e′ ) 2 2 log log C(π ′ × π
e′ )
Hence, the lemma follows from (4.14) and (4.15). 
In the proof of Theorem 4.2, there will appear the quantity Ll (s), defined in Section 4.1.
We record some useful bounds for such quantity in the next two lemmas
Lemma 4.6. Recall the notation in Section 4.1. For any σ ≥ 1 + 1/ log X, we have that
max |Ll (σ + it)| ≤ max |Ll (1 + 1/ log X + it)| + O((log X)−1 C(π ′ )ε ).
|t|≤T /2 |t|≤T

Proof. Replacing [34, Lemma 3.2] with (3.16) of Lemma 3.6, we proceed just as in the proof
of [34, Lemma 5.2]. 
Lemma 4.7. Recall the notation and hypotheses in Section 4.1. Let L − 1 = (L−1, . . . , L−1).
If l ≥ L − 1, then
max |Ll (1 + 1/ log X + it)| ≪ (log X)ε/2 C(π ′ )ε/2 .
|t|≤T

Proof. Replacing [34, Lemma 5.1] with Lemma 4.5 and [34, Lemma 3.2] with (3.16) of Lemma
3.6, we proceed just as in the proof of [34, Lemma 5.3]. 
Proposition 4.8. Recall the notation and hypotheses of Lemma 4.7. If 1 ≤ x ≤ X, then
Ol (x, w) ≪ x(log X)2ε/3 C(π ′ )ε .
2
Proof. By (3.21) of Lemma 3.7, we may assume that log x ≥ (log X)ε/(m +1) ; otherwise, the
2 2
desired bound for Ol (x, w) is trivial. If w ≤ exp((2RL)(m +1)/(3Rε−m −1) ), which is a constant
depending only on m, m′ , and ε, then the desired bound is also trivial, so we may assume that
2 2 −1)
w ≥ exp((2RL)(m +1)/(3Rε−m√ ). Upon the second assumption, we then have x ≥ w 2RL . By
(4.4) with K = 1, λ = 1/ T , and c > 1, we find that
Z c+i∞ √
1  s/ T −1  ds X  X 
′ s e
L(s, π × π )z √ = λπ×π′ (n) + O |λπ×π′ (n)| .
2πi c−i∞ s/ T s n≤z

1/ T
z<n≤ze

By (3.22) of Lemma 3.7, the error term above is Oε (C(π ′ )ε z/ log z) in the range T ≤ z ≤ X
where T = exp((log log X)2 ). Using the above formula in the definition of Ol (x, w), we obtain
16 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

that if w 2RL ≤ x ≤ X, then


Z c+i∞ √
1 YR  s/ T −1  ds  x 
′ 1+iτj −s lj s e
Ol (x, w) = L(s, π × π ) (1 − w ) x √ + O C(π ′ )ε .
2πi c−i∞ j=1
s/ T s log x
Choose c = 1 + 1/ log X and split the above integral into two parts, when | Im(s)| ≤ T
and√when | Im(s)|
√ > T . For the first range, we use Lemma 4.7 and apply the fact that
s/ T
(e − 1)/(s/ T ) ≪ 1 when | Im(s)| ≤ T , so this portion contributes
Z √
s/ T
e − 1 |ds|
≪ (log X)ε/2 C(π ′ )ε/2 x1+1/ log X √ ≪ x(log X)2ε/3 C(π ′ )ε/2 .
| Im(s)|≤T s/ T |s|
Re(s)=1+1/ log X
√ √ √
In the range | Im(s)| > T , we use (3.16) of Lemma 3.6. Since (es/ T − 1)/(s/ T ) ≪ T /|s|
in this range, this portion contributes
Z √
s/ T
m2 +1 e − 1 |ds|
≪ C(π ′ )ε x(log X) 2 √ ≪ xC(π ′ )ε ,
| Im(s)|>T s/ T |s|
Re(s)=1+1/ log X

Combining the two contributions gives us the proposition. 


4.5. Proof of Theorem 4.2. In this subsection, we prove Theorem 4.2 with the points
τ1 , . . . , τR defined as in the previous subsection.
Lemma 4.9. Recall the notation and hypotheses in Sections 3 and 4.1. If 0 ≤ log w ≤
1
(log X) 3R and 1 ≤ x ≤ X, then
X
(log x)OL (x, w) = Λπ×π′ (d)OL (x/d, w) + O(x(log X)ε C(π ′ )ε ).
d≤x

Proof. The proof of this lemma is similar to that of [34, Lemma 6.1], where we replace [34,
Proposition 5.4] by our Proposition 4.8. Writing
log x = log(x/w j1 +···+jR ) + (j1 + · · · + jR ) log w,
we observe that OL (x, w) log x equals
X  
j1 +···+jR L
(−1) log(x/w j1 +···+jR )S(x/w j1 +···+jR )w j1 (1+iτj )+···+jR (1+iτR )
j≤L
j
X   (4.16)
j1 +···+jR L
+ log w (−1) (j1 + · · · + jR ) S(x/w j1+···+jR )w j1(1+iτj )+···+jR (1+iτR ) .
j≤L
j

Let ek denote the vector with 1 in the k-th place and 0 elsewhere. The second line in (4.16)
can be written as
XR
− Lw 1+iτk (log w)OL−ek (x/w, w), (4.17)
k=1
Since each component of L − ek is at least L − 1, we can apply Proposition 4.8. It follows
that (4.17) is ≪ x(log X)ε C(π ′ )ε .
We observe that
X X
(log x)S(x) = λπ×π′ (n) log(n) + λπ×π′ (n) log(x/n).
n≤x n≤x
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 17

By partial summation and the Dirichlet convolution identity


X
λπ×π′ (n) log(n) = Λπ×π′ (d)λπ×π′ (n/d),
d|n

it follows that
XX Z Z
x
dt X x
dt
(log x)S(x) = Λπ×π′ (d)λπ×π′ (n/d) + S(t) = Λπ×π′ (d)S(x/d) + S(t) .
n≤x d|n 1 t d≤x 1 t

Hence, the first term of (4.16) equals


X Z x
dt
Λπ×π′ (d)OL (x/d, w) + OL (t, w) .
d≤x 1 t

Again, by applying Proposition 4.8, the integral part is ≪ x(log X)ε C(π ′ )ε . The lemma
follows. 
Lemma 4.10. Recall the notation in Section 4.1. If 1 ≤ z ≤ y and y + z ≤ X, then
LR
X X
2 2 ′ ε ε
||OL (y, w)| − |OL (y + z, w)| | ≪ C(π ) y(log X) wj |λπ×π′ (n)|.
j=0 y/w j <n≤(y+z)/w j

Proof. Replacing [34, Proposition 5.4] with Proposition 4.8 and using our definition of S(x),
we proceed just as in the proof of [34, Lemma 6.2]. 
1
Proposition 4.11. Recall the notation in Section 4.1. If 0 ≤ log w ≤ (log X) 3R and 1 ≤
x ≤ X, then

|OL (x, w)| log x


1 ε
Z x
dy 1
′ ′ ε 2
2
≪ x(log log x) C(π ) 2 4 log(ey)|OL (y, w)| 3 + C(π ) + x(log X)ε C(π ′ )ε .
1 y
Proof. By Lemma 4.9, we are left to estimate the sum
X
|Λπ×π′ (d)||OL (x/d, w)|.
d≤x

We split this sum into two ranges d ≤ D := ⌊exp((log log X)6 )⌋ and d > D. For the first
range of the sum, we use Proposition 4.8 and obtain
X X |Λπ×π′ (d)|
|Λπ×π′ (d)||OL (x/d, w)| ≪ x(log X)2ε/3 . (4.18)
d
d≤D d≤D

By Lemma 3.5 and partial summation, the right-hand side of (4.18) is


≪ x(log X)2ε/3 log(eD)C(π ′ )ε ≪ x(log X)ε C(π ′ )ε . (4.19)
It remains to estimate the contribution of the terms with d > D. Define g(t) = t log(ex/t)
for 1 ≤ t ≤ x. By applying Lemma 3.3 and the Cauchy–Schwarz inequality, we have
18 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)
X
|Λπ×π′ (d)||OL (x/d, w)|
D<d≤x
 X  12  X Λ ′ ′ (d)  12
π ×e
π
≤ Λπ×eπ (d)g(d)|OL(x/d, w)|2 . (4.20)
g(d)
D<d≤x D<d≤x

Using GRC for Λπ×eπ (d), and Lemma 3.5 and partial summation for Λπ′ ×eπ′ (d), we have that
(4.20) is
 X 1 1
2 2 ′ ε/2 2
≪ g(d)Λ(d)|OL(x/d, w)| (log log(ex/D))C(π )
D<d≤x
 X  21 1 ε
2
≪ g(d)Λ(d)|OL(x/d, w)| (log log x) 2 C(π ′ ) 4 .
D<d≤x

PThe remaining of the proof then follows from the proof of [34, Lemma 6.3]. Define √
ψ0 (x) =
n≤x (Λ(n)−1) = ψ(x)−x so that there exists c3 > 0 such that ψ0 (x) = O(x exp(−c3 log x))
by the prime number theorem and
X
g(d)Λ(d)|OL(x/d, w)|2
D<d≤x
X X
= g(d)|OL (x/d, w)|2 + (ψ0 (d) − ψ0 (d − 1))g(d)|OL(x/d, w)|2. (4.21)
D<d≤x D<d≤x

Moreover, we may rewrite the second term of (4.21) as


X  
2 2
ψ0 (d) g(d)|OL (x/d, w)| − g(d + 1)|OL (x/d, w)|
D≤d≤x

− ψ0 (D)g(D + 1)|OL (x/(D + 1), w)|2. (4.22)

For d > D and for any A > 0, we have


d
ψ0 (d) ≪ d exp(−(log log x)2 ) ≪A .
(log X)A+2

The second term in (4.22) is then ≪ D 2 (log X)−1 |OL (x/(D + 1), w)|2 ≪ x2 C(π ′ )ε upon using
Proposition 4.8. On the other hand, the first term in (4.22) is
X d 
≪ A+2
g(d) |OL (x/d, w)|2 − |OL (x/(d + 1), w)|2
D<d≤x
(log X)

+ |g(d + 1) − g(d)||OL (x/(d + 1), w)|2 . (4.23)

By Proposition 4.8, the contribution from the second line of (4.23) is


X d
≪ A+2
|g(d + 1) − g(d)|(x/(d + 1))2 (log X)4ε/3 C(π ′ )ε ≪ x2 C(π ′ )ε .
D<d≤x
(log X)
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 19

On the other hand, by Lemma 4.10, the first line of (4.23) is


X LR
d2 xX j X
≪ w |λπ×π′ (n)|
(log X)A d j=0
D<d≤x x/((d+1)w j )<n≤x/(dw j )
LR
X X
x j x
≪ A
w |λπ×π′ (n)| j ,
(log X) j=0 j
nw
n≤x/w

which is ≪ x2 C(π ′ )ε upon using (3.15) of Lemma 3.6. Thus, the second term in (4.21) is
≪ x2 C(π ′ )ε .
It remains to estimate the first term in (4.21). For any D < d ≤ x and d − 1 ≤ t ≤ d we
have that g(d) = g(t) + O(log x), and by Lemma 4.10 that
 LR
X X 
2 2 ′ εx ε
|OL (x/d, w)| = |OL (x/t, w)| + O C(π ) (log X) wj |λπ×π′ (n)| .
d j=0 x/(dw j )<n≤x/((d−1)w j )

We multiply g(d) to both sides of the above equation and replace g(d) by g(t) + O(log x) on
the right-hand side. Then integrating both sides from d − 1 to d with respect to t and using
Proposition 4.8 gives
Z d  
2 x2
g(d)|OL (x/d, w)| = g(t)|OL (x/t, w)|2 dt + O C(π ′ )ε 2 (log X)1+2ε/3
d−1 d
 LR
X X 
+ O C(π ′ )ε x(log X)1+ε wj |λπ×π′ (n)| . (4.24)
j=0 x/(dw j )<n≤x/((d−1)w j )

Summing (4.24) over all D < d ≤ x and changing the variable t 7→ x/y, the contribution
from the main term on the right-hand side of (4.24) becomes
Z x Z x
2 2 dy
t log(ex/t)|OL (x/t, w)| dt ≤ x |OL (y, w)|2 log(ey) 3 .
D 1 y
The error terms in (4.24) contribute
X LR X
x2
≪ C(π ′ )ε (log X)1+2ε + C(π ′ )ε x(log X)1+ε wj |λπ×π′ (n)|
D j=0 n≤x/(Dw j )
LR
X X
x2 |λπ×π′ (n)|
≪ C(π ′ )ε x2 + C(π ′ )ε (log X)1+ε ≪ C(π ′ )ε x2 ,
D j=0
n
n≤x/(Dw j )

where we use (3.15) of Lemma 3.6 in the last step. This finishes the proof. 
1
Lemma 4.12. Recall the notation in Section 4.1. If 0 ≤ log w ≤ (log X) 3R and 1 ≤ x ≤ X,
then
Z x dt  12 Z x dt  21
2
|OL (t, w)| log(et) 3 ≪ eL (t, w)|2
|O + (log X)7ε/8 C(π ′ )ε .
t t3 log(et)
1 1
20 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

Proof. We start as in the proof of Lemma 4.9. First, we write


R
X
(log t)OL (t, w) = − Lw 1+iτk (log w)OL−ek (t/w, w)
k=1
X   (4.25)
j1 +···+jR L
+ (−1) log(t/w j1 +···+jR )S(t/w j1 +···+jR )w j1 (1+iτ1 )+···+jR (1+iτR ) .
j≤L
j

By writing log z = log n + log(z/n) and using partial summation, we have that
X X Z z
e e S(y)
(log z)S(z) = log z λπ×π′ (n) = S(z) + λπ×π′ (n) log(z/n) = S(z) + dy.
n≤z n≤z 1 y

Then the second line of (4.25) may be written as


Z t
eL (t, w) + OL (y, w)
O dy. (4.26)
1 y
Inserting (4.26) into (4.25) and using Proposition 4.8, we conclude that
Z t R
X
e OL (y, w)
(log t)OL (t, w) = OL (t, w) + dy − Lw 1+iτk (log w)OL−ek (t/w, w)
1 y
k=1
eL (t, w) + O(t(1 + log w)(log X)2ε/3 C(π ′ )ε )
=O
eL (t, w) + O(t(log X)5ε/6 C(π ′ )ε ),
=O
and hence the lemma follows. 
Combining Proposition 4.11 and Lemma 4.12, we have that
Z
x(log log x) 2 C(π ′ )ε/4  x e  21 x(log X)ε C(π ′ )ε
1
dt
|OL (x, w)| ≪ |OL (t, w)|2 3 + . (4.27)
log x 1 t log(et) log x
We estimate the integral in (4.27) as follows.
1
Proposition 4.13. Recall the notation in Section 4.1. If 0 ≤ log w ≤ (log X) 3R and 1 ≤
x ≤ X, then Z x
eL (t, w)|2 dt
|O 3
≪ (log X)3ε/2 C(π ′ )3ε/2 .
1 t log(et)
Proof. We follow the proof of [34, Proposition 6.5]. First, we make substitution t = ey ,
obtaining
Z x Z log x
e 2 dt eL (ey , w)|2e−2y dy
|OL (t, w)| 3 = |O
1 t log(et) 1+y
Z0 ∞ Z ∞ (4.28)
≪ e−2α e y 2 −2y(1+α)
|OL (e , w)| e dydα.
1/ log X 0

As in the proof of [34, Proposition 6.5], the inner integral of (4.28) is


Z ∞ YR
′ ′ dt
≪ |L (1 + α + it, π × π )| |1 − w −α−it+iτk |2L 2
. (4.29)
−∞ k=1
|1 + α + it|
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 21

Recall that T = exp((log log X)2 ). We split the integral above into the ranges when |t| ≤ T /2
and when |t| > T /2. Note that α ≥ 1/ log X in (4.28). For the latter range, we use (3.17) of
Lemma 3.6 to bound |L′ (1 + α + it, π × π ′ )|. Thus, for any A > 0, the portion of the integral
in (4.29) over |t| > T /2 contributes
Z
m2 +3
′ ε dt C(π ′ )ε
≪ (log X) 2 C(π ) 2
≪ A . (4.30)
|t|>T /2 |1 + α + it| (log X)A

When |t| ≤ T /2, Lemmas 4.6 and 4.7 yield


R
Y
′ ′ L′
|L (1 + α + it, π × π )| |1 − w −α−it+iτk |L ≪ (1 + α + it, π × π ′ ) |LL (1 + α + it)|
L
k=1
L′
≪ (1 + α + it, π × π ′ ) (log X)ε/2 C(π ′ )ε/2 .
L
Therefore, the integral in (4.29) over |t| ≤ T /2 contributes
Z
ε ′ ε L′ 2 dt
≪ (log X) C(π ) (1 + α + it, π × π ′ )
|t|≤T /2 L |1 + α + it|2
Z ∞ ′ (4.31)
L 2 dt
≪ (log X)ε C(π ′ )ε (1 + α + it, π × π ′ ) .
−∞ L |1 + α + it|2
P
Observe that the Fourier transform of the function e−y(1+α) n≤ey Λπ×π′ (n) is
Z ∞ X ∞
X Λπ×π′ (n) 1 L′ 1
Λπ×π′ (n)e−y(1+α+it) dy = = − (1 + α + it) · .
−∞ n≤ey n=1
n1+α+it 1 + α + it L 1 + α + it

Thus, by Plancherel’s theorem and Lemma 3.5, (4.31) is


Z ∞ X
ε ′ ε
2 (log X)ε C(π ′ )3ε/2
≪ (log X) C(π ) Λπ×π′ (n) e−(2+2α)y dy ≪ , (4.32)
−∞ n≤ey 2α

where we use that y ≥ 0 in the last step. By (4.30) and (4.32), we then conclude that
Z ∞
e y 2 −y(2+2α) C(π ′ )ε (log X)ε C(π ′ )3ε/2
|OL (e , w)| e dy ≪ + .
0 (log X)A 2α
Inserting the above bound into (4.28), we obtain the bound
Z x Z ∞
e 2 dt C(π ′ )ε ε ′ 3ε/2 e−2α
|OL (t, w)| 3 ≪ + (log X) C(π ) dα
1 t log(et) (log X)A 1/ log X 2α

≪ (log X)ε C(π ′ )3ε/2 log log(X).


The proposition follows. 

Proof of Theorem 4.2. Using the bound from Proposition 4.13 in (4.27), we then obtain The-
orem 4.2 as desired. 
22 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

5. Background on GL(2) automorphic forms


Having established Theorem 1.3 in the previous section, we now shift our focus to laying
some foundation for proving Theorem 1.1. We recall here some standard properties of classical
automorphic forms, including holomorphic newforms, Hecke–Maaß cusp forms, Eisenstein
series, incomplete Eisenstein series, and incomplete Poincaré series. We will also discuss the
spectral decomposition and the Poincaré decomposition. For complete definitions and proofs,
see, for example, [16], [29], and [30].

5.1. Holomorphic newforms. Let k ≥ 2 be an even integer and let α ∈ GL2 (R) with
positive determinant, where the element α acts on H by fractional linear transformations.
Given a function f : H → C, we denote j(( ac db ), z) = cz + d and recall the slash operator
f |k α(z) = (det α)k/2 j(α, z)−k f (αz).
A holomorphic cusp form on Γ0 (q) of weight k with trivial nebentypus is a holomorphic
function f : H → C such that f |k γ = f for all γ ∈ Γ0 (q) and f vanishes at the cusps of Γ0 (q).
We shall denote the vector space of holomorphic cusp forms on Γ0 (q) of weight k with trivial
nebentypus as Sk (Γ0 (q)).
Define Sknew (Γ0 (q)) to be the subspace of Sk (Γ0 (q)) that is orthogonal to the subspace of
oldforms, denoted as Skold (Γ0 (q)), with respect to the Petersson inner product. A holomorphic
newform is a cusp form in Sknew (Γ0 (q)) that is also an eigenform of all of the Hecke operators
and all of the Atkin–Lehner involutions. The space Sknew (Γ0 (q)) has a basis of newforms,
which we shall denote as Bk (q) (see more discussion in [30, Chapter 2, Section 5]).
Consider f ∈ Bk (q). Let e(z) = e2πiz and let λf (n) be the eigenvalue of the n-th Hecke
operator Tn . Then the Fourier expansion of f can be written as
X k−1
f (z) = λf (n)n 2 e(nz), (5.1)
n≥1

where λf (1) = 1. If we let τ (n) denote the number of positive divisors of n, then λf (n) is
real, multiplicative and satisfies Deligne’s bound |λf (n)| ≤ τ (n).
For any γ ∈ Γ0 (q) and f, g ∈ Bk (q), if we write z ′ = γz = x′ + iy ′ , then
y ′k/2 f (z ′ ) = (j(γ, z)/|j(γ, z)|)k y k/2f (z).
This observation implies the Γ0 (q)-invariance for both z 7→ y k |f (z)|2 and z 7→ y k f (z)g(z).
By the Rankin–Selberg theory, we have
(4π)k
|ρf (1)|2 ≍ . (5.2)
kqΓ(k − 1)L(1, Ad f )
5.2. Cusps of Γ0 (q). We now recall some properties of cusps of Γ0 (q) and the Fourier expan-
sion of holomorphic newforms at arbitrary cusps of Γ0 (q) (see more discussion in [29, Sections
3.4.1 and 3.4.2]). This will be crucial when we apply Holowinsky’s strategy in Section 8.2.
Let τ traverse a set of representatives for the double coset space Γ∞ \Γ0 (1)/Γ0 (q). Then a :=
τ −1 ∞ ∈ P1 (Q) traverses a set of inequivalent cusps of Γ0 (q). Denote the set of cusps of Γ0 (q)
as C = C(Γ0 (q)) = {aj }j . The width of aj = τj−1 ∞ is given by wj = [Γ∞ : Γ∞ ∩ τj Γ0 (q)τj−1 ],
and the scaling matrix of aj is
 
−1 wj
σj = τj . (5.3)
1
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 23

By the discussion in [29, Section 3.4.1], if the bottom row of σj is (cj , dj ), then we define
cj as the denominator of the cusp aj . Such cj is a positive divisor of q, and we have that
the width wj equals q/(c2j , q) = [q/c2j , 1]. For each positive divisor c|q, if we denote the set of
cusps of denominator c by
C[c] := {aj ∈ C : cj = c},
then #C[c] equals ϕ((c, q/c)). If q is squarefree, then #C[c] = 1. In that case, the multiset of
the widths of cusps of Γ0 (q) is {d ∈ Z : d|q, d ≥ 1}.
We now collect necessary properties of the Fourier expansion of |f |2 at the cusp aj . For
any positive divisor c|q and d ∈ (Z/(c, q/c))× , we write ad/c ∈ C(Γ0 (q)) for the corresponding
cusp. For aj = ad/c , we may take τj = ( ∗c d∗′ ) for any integer d′ for which (d′ , c) = 1 and
d′ ≡ d (mod (c, q/c)). Such τj ’s then form a set of representatives of Γ∞ \Γ0 (1)/Γ0 (q), and
the corresponding scaling matrix σj is defined as in (5.3). Write zj = xj + iyj for the change
of variable zj := σj−1 z. We observe that |f (z)|2 y k = |f (σj zj )|2 Im(σj zj )k = |f |k σj (zj )|2 yjk . For
some coefficients λj (n) ∈ C, we may write
X k−1
f |k σj (zj ) = λj (n)n 2 e(nzj ). (5.4)
n≥1

In the special case that aj = ∞, we observe that λj (n) = λ(n).


In general, λj is not multiplicative. To circumvent the lack of multiplicativity, we shall
work with the root-mean-square of λj taken over all cusps of a given denominator. For each
positive divisor c of q, we define
 1 X 1/2
λ[c] (n) := |λj (n)|2 . (5.5)
#C[c]
aj ∈C[c]

As discussed in [29, Section 3.4.2], λ[c](n) turns out to be multiplicative in a nonconventional


sense, which allows us to deduce Proposition 8.3. However, when applying Holowinsky’s
method in Section 8.2, we will encounter shifted convolution sums in terms of λj , not λ[c] as
in Proposition 8.3. Hence, we have to apply the following steps. By (5.6) and the Cauchy–
Schwarz inequality, we have
X X X
|λj (n1 )λj (n2 )| = |λj (n1 )λj (n2 )|
aj ∈C c|q aj ∈C[c]
X  X |λ (n )|2 1/2  X |λ (n )|2 1/2
j 1 j 2
≤ #C[c] (5.6)
#C[c] #C[c]
c|q aj ∈C[c] aj ∈C[c]
X
= #C[c]λ[c] (n1 )λ[c] (n2 ).
c|q

The bound (5.6) then allows us to work with λ[c] as desired.

5.3. Hecke–Maaß cusp forms. A Maaß cusp form φ of level 1 is an Γ0 (1)-invariant eigen-
function of the hyperbolic Laplacian ∆ := −y 2 (∂x2 + ∂y2 ) on H that decays rapidly at infinity.
We denote the spectral parameter of φ by tφ , which is given by ∆φ = (1/4 + t2φ )φ. A Hecke–
Maaß cusp form is a Maaß cusp form that is also an eigenfunction of all of the Hecke operators
24 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

and the involution T−1 : φ 7→ [z 7→ φ(−z)], which commute with one another as well as with
∆. A Hecke–Maaß cusp form φ admits a Fourier expansion
X λφ (n)
φ(z) = p κφ (ny)e(nx),
n∈Z6=0
|n|
1−δ
where κφ (y) = 2|y|1/2Kir (2π|y|)sgn(y) 2 with Kir the standard K-Bessel function, sgn(y) =
1 or −1 according as y is positive or negative, and δ ∈ {±1}, the T−1 -eigenvalue of φ. A
normalized Hecke–Maaß cusp form is a Hecke–Maaß cusp form that satisfies λφ (1) = 1. In
that case, the coefficients λφ (n) are real and multiplicative. An orthonormal basis of Maaß
cusp forms can be chosen such that it consists of all normalized Hecke–Maaß cusp forms.
Throughout this paper, we will normalize the Petersson norm of a Hecke–Maaß cusp form
φ to be 1, i.e., hφ, φi1 = 1. We will also assume that φ is an even Hecke–Maaß cusp form;
otherwise, hφ, |Fk |2 iq = 0. Additionally, define NΓ0 (1) (T ) := |{φ : |tφ | ≤ T }|, the number of
Hecke–Maaß cusp forms such that |tφ | ≤ T . By Weyl’s law (see, for example, [14, Corollary
11.2]), we have that
T2
NΓ0 (1) (T ) = + O(T log T ). (5.7)
12
 
5.4. Eisenstein series. Let s ∈ C, z ∈ H, and Γ∞ = ± 1 n1 : n ∈ Z . The real-analytic
Eisenstein series of level 1 is defined as
X
E(s, z) = Im(γz)s .
Γ∞ \Γ0 (1)

The Eisenstein series E(s, z) converges normally for Re(s) > 1 and continues meromorphically
to the half-plane Re(s) ≥ 1/2, where the map s 7→ E(s, z) is holomorphic with the exception
of a unique simple pole at s = 1 of residue ress=1 E(s, z) = 3/π. For all γ ∈ Γ0 (1), the
Eisenstein series satisfies the invariance E(s, γz) = E(s, z).
Let ζ(s)
Pbe the Riemann zeta function and Ks (y) be the standard K-Bessel function. Let
s −s
λs (n) = ab=n (a/b) and θ(s) = π Γ(s)ζ(2s). The Eisenstein series admits the Fourier
expansion

s θ(1 − s) 1−s 2 y X
E(s, z) = y + y + τs−1/2 (|n|)Ks−1/2 (2π|n|y)e(nx).
θ(s) θ(s) n∈Z
6=0

For Re(s) = 1/2, we obtain |θ(1 − s)/θ(s)| = 1, and such E(s, z) is called a unitary Eisenstein
series. Let Et denote E(1/2 + it, ·). The work of Huang–Xu [10, Theorem 1] shows that for
y ≥ 1/2, if z = x + iy, then

|Et (z)| ≪ y(1 + |t|)3/8+ε . (5.8)
Throughout the paper, we will normalize the Petersson norm of Et to be 1, i.e., hEt , Et i1 = 1.

5.5. Spectral decomposition. Now we recall the spectral decomposition, which will be im-
portant when applying Soundararajan’s strategy in Section 8.1. Let Y0 (1) = Γ0 (1)\H, and let
L2 (Y0 (1)) denote the space of square-integrable functions on Y0 (1). Recall the definition of the
Petersson inner product in (1.7). Then one can spectrally decompose any ψ ∈ L2 (Y0 (1)) into
three types of spectra: a constant function, Hecke–Maaß cusp forms, and unitary Eisenstein
series. See, for example, [16, Theorem 15.5].
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 25

Lemma 5.1 (Theorem 15.5 of [16]). Let {φj }j be an orthonormal basis of Hecke–Maaß cusp
forms. For any real number t, let Et := E(1/2 + it, ·) be a unitary Eisenstein series. If
ψ ∈ L2 (Y0 (1)), then
X Z
3 1
ψ(z) = hψ, 1i1 + hψ, φj i1 φj (z) + hψ, Et i1 Et (z)dt.
π j≥1
4π R

The above decomposition gives us that


X Z
3 1
hψ, Fk Gk iq = hψ, 1i11f =g + hψ, φj i1 hφj , Fk Gk iq + hψ, Et i1 hEt , Fk Gk iq dt. (5.9)
π j≥1
4π R

This implies that we need some good estimates for hφj , Fk Gk iq and hEt , Fk Gk iq , along with
estimates for hψ, φj i1 and hψ, Et i1 , to obtain a rate of convergence for hψ, Fk Gk iq . We recall
the following estimates for hψ, φj i1 and hψ, Et i1 .
Lemma 5.2 (Lemma 2.2 of [9]). Let M ≥ 1, ψ ∈ Cc∞ (Y0 (1), M), and φ be a Hecke–Maaß
cusp form with spectral parameter tφ . If A ≥ 0, then
 M A  M A
|hψ, φi1| ≪A and |hψ, Et i1 | ≪A (1 + |t|)3/8+ε .
1 + |tφ | 1 + |t|
In the next subsection, we shall discuss the Poincaré decomposition, which will play an
important role when applying Holowinsky’s method in Section 8.2.
5.6. Incomplete Poincaré series and the Poincaré decomposition. Recall the notation
in Section 1. The discussion in this subsection closely follows [6, Section 4]. Let M ≥ 1 and
ψ ∈ Cc∞ (Y0 (1), M). Let ψ̌ be the function on H to C such that ψ̌(z) =ψ(z) if z ∈ F , and
ψ̌(z) = 0 otherwise. Let Ψ̌(z) be the extension of ψ̌(z) to H by Γ∞ = ± 1 n1 : n ∈ Z .
For any z = x + iy ∈ H, we denote the m-th Fourier coefficients of Ψ̌(z) as
Z 1/2
Ψm (y) = Ψ̌(x + iy)e(−mx) dx.
−1/2

By integration by parts, we have that


|Ψm (y)| ≪A (M/|m|)A . (5.10)
Then the Fourier expansion of Ψ̌(z) can be defined as
X
Ψ̌(z) = Ψm (y)e(mx).
m∈Z

For any Ψ ∈ C (R), a smooth function on R, we define the incomplete Poincaré series as
X
Pm (z, Ψ) = e(m Re(γz))Ψ(Im(γz)). (5.11)
γ∈Γ∞ \Γ0 (1)

When m = 0, P0 (z, Ψ)
P is called the incomplete Eisenstein series, also denoted by E(·|Ψ).
Observe that ψ(z) = γ∈Γ∞ \Γ0 (1) Ψ̌(γz). Then we have that
X X X
ψ(z) = Ψm (Im(γz))e(m Re(γz)) = Pm (z, Ψm ). (5.12)
γ∈Γ∞ \Γ0 (1) m∈Z m∈Z
26 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

Therefore, by (5.12), we obtain that


3 3
hψ, Fk Gk iq − hψ, 1i1 1f =g = hE(·|Ψ0), Fk Gk iq − hE(·|Ψ0), 1i1 1f =g
π π
X 3  (5.13)
+ hPm (·, Ψm), Fk Gk iq − hPm (·, Ψm ), 1i11f =g .
m6=0
π
R∞
b
For any Ψ ∈ C ∞ (R), let Ψ(s) = 0 Ψ(y)y s−1 dy denote the Mellin transform of Ψ(y). By
(5.11) and the Γ0 (1)-invariance of the hyperbolic measure y −2dxdy, we have that
Z X X Z
dxdy dxdy
hE(·|Ψ), 1i1 = Ψ(Im(γz)) 2 = Ψ(Im(z)) 2 .
Γ0 (1)\H y γΓ0 (1)\H y
γ∈Γ∞ \Γ0 (1) γ∈Γ∞ \Γ0 (1)

Upon unfolding and integrating over dx, we obtain that


Z Z ∞ Z 1/2
dxdy dxdy b
hE(·|Ψ), 1i1 = Ψ(y) 2 = Ψ(y) 2 = Ψ(−1). (5.14)
Γ∞ \H y y=0 x=−1/2 y
Similarly, for the incomplete Poincaré series, we have that
Z Z ∞ Z 1/2
dxdy dxdy
hPm (·, Ψ), 1i1 = Ψ(y)e(mx) 2 = Ψ(y)e(mx) 2 = 0. (5.15)
Γ∞ \H y y=0 x=−1/2 y
By (5.13), (5.14), and (5.15), it follows that
3
hψ, Fk Gk iq − hψ, 1i1 1f =g
π
3b X
= hE(·|Ψ0), Fk Gk iq − Ψ0 (−1)1f =g + hPm (·, Ψm ), Fk Gk iq . (5.16)
π m6=0

6. Outline of the proof for Theorem 1.1


In this section, we prove Theorem 1.1 assuming the following two key propositions. The
proofs of our key propositions rely on the bounds in Lemmas 7.3 and 7.4, which are proved
using Theorem 1.3. Define
(
q2, if f 6= g and q is squarefree,
Q = Q(f, g) = (6.1)
NAd f , if f = g.

6.1. Key propositions. To derive Theorem 1.1, we bound hψ, Fk Gk iq − π3 hψ, 1i1 1f =g in two
different ways. Our first result, based on Soundararajan’s approach in [34], relies crucially on
Theorem 1.3 and the variant of Watson’s triple product formula proved in [28, 29].
Proposition 6.1. Keep the notation and hypotheses of Theorem 1.1. Let λf (n) and λg (n)
denote the n-th Hecke eigenvalues of f and g, respectively. If ε > 0, then
3  q −( 21 −θ)+ε Y  λf (p)2 + λg (p)2 − 1 
hψ, Fk Gk iq − hψ, 1i1 1f =g ≪ M δ+ε (log kq)ε √ 1− ,
π Q √ 2p
p≤k Q
23
where δ = 12
when f = g and δ = 2 when f 6= g.
Our second result is based on Holowinsky’s approach in [6].
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 27

Proposition 6.2. Keep the notation and hypotheses of Theorem 1.1. Let λf (n) and λg (n)
denote the n-th Hecke eigenvalues of f and g, respectively. If ε > 0, then
3  q ε Y  (|λ (p)| − 1)2 + (|λ (p)| − 1)2 
4+ε ε f g
hψ, Fk Gk iq − hψ, 1i1 1f =g ≪ M (log kq) √ 1− .
π Q √ 4p
p≤k Q

6.2. Proof of Theorem 1.1. We combine Propositions 6.1 and 6.2 and follow the optimiza-
tion technique in Iwaniec’s course notes. First, observe that
3
hψ, Fk Gk iq − hψ, 1i1 1f =g
π
n  q −( 12 −θ)+ε Y  λf (p)2 + λg (p)2 − 1 
δ+ε ε
≪ min M (log kq) √ 1− ,
Q √ 2p
p≤k Q
 q ε Y  (|λf (p)| − 1)2 + (|λg (p)| − 1)2 o
4+ε ε
M √ (log kq) 1− .
Q √ 4p
p≤k Q

Let α ∈ [0, 1]. If X, Y > 0, then min{X, Y } ≤ X α Y 1−α . By Deligne’s bound, we have that
|λf (p)|, |λg (p)| ≤ 2. Write λf = |λf (p)| and λg = |λg (p)|. Since
  λ2 + λ2 − 1   (λ − 1)2 + (λ − 1)2  7 √ 
f g f g
̟ = min max max − α +(1−α) = − −2 3
α∈[0,1] λf ∈[0,2] λg ∈[0,2] 2 4 2

and the minimum is achieved when α equals α0 = 2/ 3 − 1, it follows from Mertens’ theorem
that
3  q −α0 ( 12 −θ)+ε Y  ̟
4−(4−δ)α0 +ε ε
hψ, Fk Gk iq − hψ, 1i1 1f =g ≪ M (log kq) √ 1+
π Q √ p
p≤k Q
(6.2)
 q −( √2 −1)( 12 −θ)+ε  1  72 −2√3
3
≪ M 4 (log kq)ε √ √ .
Q log(k Q)
√ √
We then invoke the bound (log k Q)−1 ≪ (q/ Q)ε (log kq)−1 and recall the √ definition of Q
in (6.1). Theorem 1.1 then follows upon using the fact that (q/q0 )1/2 ≪ q/ Q, which follows
from [29, Proposition 2.5]. √
1
When q = 1 and f = g, the exponent of M in (6.2) can be refined to 12 (23 − 2 3). To
refine Proposition 6.2 when q = 1, we directly follow the proof of [22, Lemma 4.5] for the
contribution from incomplete Poincaré series and Lemma 8.9 for incomplete Eisenstein series.
Then we obtain
3 5
Y (|λf (p)| − 1)2 
2 +ε ε
hψ, |Fk | i1 − hψ, 1i1 ≪ M 3 (log k) 1− . (6.3)
π p≤k
2p

We now refine Proposition 6.1 when q = 1 and f = g. We follow [22, Lemma 4.4] for the
Eisenstein series contribution, except now we slightly improve the bound for L( 21 + it, f × f )
by using that ζ( 12 + it) ≪ (1 + |t|)13/84+ε . For the contribution from Hecke–Maaß cusp forms,
we improve the exponent of |tφ | in Lemma 7.3 by using that C(Ad f × φ) ≪ k 4 (1 + |tφ |)2 ,
which holds when |tφ | ≤ k. In the proof of Proposition 6.1, we can then improve the exponent
28 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

of |tφ | to 17/12 in the range |tφ | ≤ M(log k)ε/4 (see (8.14)). Hence, we obtain
3 Y 2λf (p)2 − 1 
2 17/12+ε ε
hψ, |Fk | i1 − hψ, 1i1 ≪ M (log k) 1− . (6.4)
π p≤k
2p

Combining (6.3) and (6.4) as in the beginning of this proof, we obtain the desired result.

7. Bounds for L-functions


In this section, we prove useful lower bounds for the adjoint lift of a holomorphic new-
form. We shall also establish some bounds for L-functions that will be useful for proving
Propositions 6.1 and 6.2.
Let π ∈ F2 be non-dihedral with central character ωπ . The adjoint L-function of π is given
by
e)ζ(s)−1.
L(s, Ad π) = L(s, π × π
Since π is non-dihedral, the work of Gelbert and Jacquet [4] shows that Ad π ∈ F3 . Thus,
L(s, Ad π) admits all of the analytic properties of a standard L-function in Section 3.1. Define
the gamma factor of L(s, Ad π) as
3s
s + 1 s + k − 1 s + k 
L∞ (s, Ad π) = π − 2 Γ Γ Γ
2 2 2
if π is a holomorphic newform of weight k and
3s
s s  s 
L∞ (s, Ad π) = π − 2 Γ Γ + it Γ + it
2 2 2
if π is a Hecke–Maaß form with eigenvalue λ = 14 + t2 . Since π is non-dihedral, L(s, Ad π)
has no pole at s = 1, so the completed L-function
s/2
Λ(s, Ad π) = NAd π L(s, Ad π)L∞ (s, Ad π)
is entire of order 1, and satisfies the functional equation Λ(s, Ad π) = Λ(1 − s, Ad π).
Recall the definition of the analytic conductor in (3.3) and write C(π) = Nπ Kπ and
C(Ad π) = NAd π KAd π . In other words, we let Kπ and KAd π denote the contributions from
the spectral parameters. Then we have that
NAd π = Nπ×eπ | Nπ2 and KAd π ≍ Kπ . (7.1)
In the next theorem, we shall use the above properties to refine the work of Goldfeld, Hoffstein,
and Lieman in [5, Appendix].
Theorem 7.1. Let π ∈ F2 be non-dihedral with central character ωπ . There exists an absolute
and effectively computable constant c4 > 0 such that L(s, Ad π) 6= 0 in the region

c4
Re(s) ≥ 1 − , (7.2)
log(NAd π Kπ (|t| + 3))
and
L(1, Ad π) ≫ (log NAd π Kπ )−1 . (7.3)
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 29

Proof. The zero-free region (7.2) holds when t 6= 0 by Theorem A.1 in [11, Appendix]. When
t = 0, the proof is the same as in [5, Appendix] apart from the observation that
ϕ(s) = ζ(s)L(s, Ad π)2 L(s, Ad π × Ad π) = L(s, (1 ⊞ Ad π) × (1 ⊞ Ad π)).
In particular, the analytic conductor of ϕ(s) is O((NAd π Kπ )8 ). This permits us to express
the zero-free region for L(s, Ad π) in terms of NAd π instead of Nπ (as in [5, Appendix]).
Lastly, (7.3) follows from the proof of the main theorem in [5, Appendix], where we replace
their zero-free region with our zero-free region in (7.2). 
Consider f ∈ Bk (q). Recall the notation in Sections 1 and 5.1 and let Q = NAd f denote the
conductor of the adjoint lift of f as in (6.1). Using Theorem 7.1, we establish a lower bound

on L(1, Ad f ) that is sensitive to the distribution of Hecke eigenvalues λf (p) for p ≤ k Q.
This is important for the final steps in our proofs of Theorem 1.1.
Lemma 7.2. If f ∈ Bk (q) and Q is defined as in (6.1), then
1 Y  λf (p)2 − 1 
≪ (log log kq)9 1− .
L(1, Ad f ) √ p
p≤k Q

Proof. Let X ≥ 2. By partial summation, we obtain


X ΛAd f (n) X  Z ∞X  log t + 1
1
=− ΛAd f (n) + ΛAd f (n) 2 2
dt. (7.4)
n≥X
n log n n≤X
X log X X n≤t
t (log t)

By [19, Theorem 2.4] and Theorem 7.1, there exist absolute and effectively computable con-
stants c5 > 0 and c6 > 0 such that if X ≥ (k 2 Q)c5 , then
X  log X 
ΛAd f (n) ≪ X exp − c6 2 Q) +
√ .
n≤X
log(k log X
2 Q)
Applying the above bound in (7.4) and choosing X = (k 2 Q)(6/c6 ) log log(k , we obtain
X ΛAd f (n) (log k 2 Q)−1
≪ 2 Q)(log k 2 Q)
≪ 1.
n≥X
n log n (log log k

Therefore, we have
X ΛAd f (n) X ΛAd f (n) X λAd f (p)
log L(1, Ad f ) = = + O(1) = + O(1), (7.5)
n≥2
n log n 2≤n≤X
n log n p≤X
p

where we used Deligne’s bound to trivially bound the contribution from the composite n ≤ X.
By the Hecke relations, if p ∤ q, then λAd f (p) = λf (p2 ) = λf (p)2 − 1. Applying this fact
and another application of Deligne’s bound in (7.5), we obtain that
X λf (p)2 − 1 X λAd f (p) − λf (p)2 + 1 X λAd f (p)
log L(1, Ad f ) = + + + O(1)
√ p √ p √ p
p≤k Q p≤k Q k Q≤p≤X
p|q
X λf (p)2 − 1 X 6 X 3
≥ − − + O(1).
√ p p √ p
p≤k Q p|q k Q≤p≤X
30 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

Mertens’ theorem in the form


X1  1 
= log log x + 0.26149 . . . + O
p≤x
p log x
P
implies that p|q 1p ≤ log log log q + O(1). The lemma now follows. 
In the next two lemmas, we establish some useful bounds for Section 8.1, when we apply
Soundararajan’s strategy.
Lemma 7.3. Recall the notation in Sections 5.1 and 5.3. Let φ be a Hecke–Maaß cusp form
with spectral parameter tφ . Let f, g ∈ Bk (q) and Q be defined as in (6.1). If ε > 0, then
1
Λ( 12 , f × g × φ) Q 2 (1 + |tφ |)2δ1 +ε
≪ ,
Λ(1, Ad φ)Λ(1, Ad f )Λ(1, Ad g) (log kQ)1−ε L(1, Ad f )L(1, Ad f )
where δ1 = 11/12 when f = g and δ1 = 1 otherwise.
1
Proof. By Stirling’s formula, Γ(s) ≍ |s|σ− 2 exp(− π2 |s|), and the fact that |Γ(k + it)| ≤ Γ(k)
for any real t, it follows that
L∞ ( 21 , f × g × φ) 1
≪ . (7.6)
L∞ (1, Ad φ)L∞ (1, Ad f )L∞ (1, Ad g) k
First, consider the case when f = g. Since f has trivial nebentypus, f is self-dual and we
have the following factorization L( 12 , f × f × φ) = L( 21 , Ad f × φ)L( 21 , φ). Since C(Ad f × φ) ≪
k 4 Q2 (1 + |tφ |)6 and C(φ) ≪ (1 + |tφ |)2 , Theorem 1.3 implies that
1 1 3 ε
ε (k 4 Q2 (1 + |tφ |)6 ) 4 kQ 2 (1 + |tφ |) 2 + 3
L( 12 , Ad f × φ) ≪ (1 + |tφ |) 3 ≪ . (7.7)
(log(k 4 Q2 (1 + |tφ |)6 ))1−ε (log kQ)1−ε
Using the bound L( 12 , φ) ≪ (1 + |tφ |)1/3+ε/3 from [12, Corollary 2], the bound L(1, Ad φ)−1 ≪
log(1 + |tφ |) from [5], and (7.7), we conclude that
1 11
L( 12 , f × f × φ) kQ 2 (1 + |tφ |) 6 +ε
≪ . (7.8)
L(1, Ad φ)L(1, Ad f )2 (log kQ)1−ε L(1, Ad f )2
Hence, combining (7.8) with (7.6), we obtain the lemma when f = g. In the case that f 6= g,
the proof proceeds the same way, except that C(f × g × φ) ≪ k 4 q 4 (1 + |tφ |)8 and we directly
apply Theorem 1.3 with L( 12 , f × g × φ). 
Lemma 7.4. Recall the notation in Section 5.1. Let f, g ∈ Bk (q) and Q be defined as in
(6.1). For any s ∈ C, let ξ(s) be the completed zeta function. If t ∈ R and ε > 0, then we
have 1
Λ( 21 + it, f × g)2 Q 2 (1 + |t|)2δ2 +ε
≪ ,
ξ(1 + 2it)2 Λ(1, Ad f )Λ(1, Ad g) (log kQ)2−ε L(1, Ad f )L(1, Ad g)
where δ2 = 19/21 if f = g and δ2 = 1 otherwise.
1
Proof. By Stirling’s formula, Γ(s) ≍ |s|σ− 2 exp(− π2 |s|), and the fact that |Γ(k + it)| ≤ Γ(k)
for any real t, it follows that
L∞ ( 21 + it, f × g)2 1
1 2
≪ . (7.9)
Γ( 2 + it) L∞ (1, Ad f )L∞ (1, Ad g) k
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 31

We first consider the case when f = g. Since f has trivial nebentypus, f is self-dual and
we have the following factorization L( 12 + it, f × f ) = L( 12 + it, Ad f )ζ( 12 + it). To estimate
L( 21 +it, Ad f ), we invoke the weak subconvexity bound in [34, Theorem 1]. We note that such
theorem also applies for the shifted L-function L(s + it, Ad f ) as discussed in [34, Example
1]. Since C(Ad f, t) ≪ k 2 Q(1 + |t|)3 , we obtain that
1 1 1 3
(k 2 Q(1 + |t|)3 ) 4 k 2 Q 4 (1 + |t|) 4
L( 12 + it, Ad f ) ≪ ε ≪ ε . (7.10)
(log(k 2 Q(1 + |t|)3 ))1− 2 (log kQ)1− 2
Using the bound ζ(1/2 + it) ≪ε (1 + |t|)13/84+ε/3 from [2], the known bound ζ(1 + 2it)−1 ≪
log(1 + |t|), and (7.10), we conclude that
1 38
L( 12 + it, f × f )2 kQ 2 (1 + |t|) 21 +ε
≪ . (7.11)
ζ(1 + 2it)2 L(1, Ad f )2 (log kQ)2−ε L(1, Ad f )2
Hence, combining (7.11) with (7.9), we obtain the lemma when f = g. The proof for the case
that f 6= g proceeds in the same way, except that C(f × g, t) ≪ k 2 q 2 (1 + |t|)4 and we directly
apply [34, Theorem 1] with L( 12 + it, f × g). 

8. Effective correlation and decorrelation


In this section, we prove Propositions 6.1 and 6.2, which are based on Soundararajan’s
approach in [34] and Holowinsky’s approach in [6], respectively. If f 6= g, then we assume, as
stated in Theorem 1.1, that the shared level of f and g is squarefree.

8.1. Soundararajan’s approach. In this subsection, we prove Proposition 6.1. To do so,


we first appeal to the spectral decomposition in (5.9). The problem is then reduced to finding
rates of convergence for hφ, Fk Gk iq and hEt , Fk Gk iq . Following the work of Nelson, Pitale,
and Saha in [29], we derive the next lemma.
Lemma 8.1. Recall the notation and hypotheses as in Theorem 1.1 and the definition of Q
in (6.1). Let φ be a Hecke–Maaß cusp form with spectral parameter tφ and, for any t ∈ R,
let Et = E(1/2 + it, ·) denote the unitary Eisenstein series. If ε > 0, then we have
√ 1
(1 + |tφ |)δ1 +ε (q/ Q)− 2 +θ+ε
hφ, Fk Gk iq ≪ 1 1 1 , (8.1)
L(1, Ad f ) 2 L(1, Ad g) 2 (log kQ) 2 −ε
where δ1 = 11/12 if f = g and δ1 = 1 otherwise, and
√ 1
(1 + |t|)δ2 +ε (q/ Q)− 2 +ε
hEt , Fk Gk iq ≪ 1 1 , (8.2)
L(1, Ad f ) 2 L(1, Ad g) 2 (log kQ)1−ε
where δ2 = 19/21 if f = g and δ2 = 1 otherwise.
Proof. Recall Ichino’s generalization of Watson’s formula and denote Iv∗ the corresponding
normalized local Ichino integral (see [29, Theorem 3.1] and [28, Remark 4.2] for more details
on the correlation and decorrelation cases, respectively). We have that
|hφ, Fk Gk iq |2 1 Λ( 21 , f × g × φ) Y
= I ∗. (8.3)
hφ, φi1h1, |Fk |2 iq h1, |Gk |2 iq 8 Λ(1, Ad φ)Λ(1, Ad f )Λ(1, Ad g) v v
32 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

For any prime p, let np = ordp (q) and mp = ordp ( √qQ ). The values of Iv∗ are defined and

calculated explicitly in [29, Section 2] and [28, Section 4]. In particular, I∞ = 1, and


= 1, if p ∤ q,
∗ 5 −n mp 2 2θmp
Ip < 10 p τ (p ) p
p
, if p | q and f = g,

= p−1 , if p | q and f 6= g.
Using the above calculation of Iv∗ , we obtain that
Y √ p p p
Iv∗ ≪ 105ω(q/ Q) q −1 τ (q/ Q)2 (q/ Q)2θ ≪ q −1 (q/ Q)2θ+2ε , (8.4)
v
√ √ √
upon using that 105ω(q/ Q) τ (q/ Q)2 ≪ (q/ Q)2ε . Recall that hφ, φi1 = h1, |Fk |2 iq =
h1, |Gk |2 iq = 1 by our normalization. Inserting (8.4) and Lemma 7.3 into (8.3) and tak-
ing the square root, we obtain that
1
Q 4 (1 + |tφ |)δ1 +ε − 12
p θ+ε
hφ, Fk Gk iq ≪ 1 1 1 q (q/ Q) , (8.5)
(log kQ) 2 −ε L(1, Ad f ) 2 L(1, Ad g) 2
where δ1 = 11/12 when f = g and δ1 = 1 otherwise. Hence, (8.1) follows.
We now consider the case of unitary Eisenstein series. Let ξ(s) be the completed Riemann
zeta function. Similarly to the case of Hecke–Maaß cusp forms, we have
|Λ(1/2 + it, f × g)|2 Y
|hEt , Fk Gk iq |2 = J ∗, (8.6)
|ξ(1 + 2it)|2 Λ(1, Ad f )Λ(1, Ad g) v v

where J∞ = 1, and 

= 1, if p ∤ q,
∗ 5 −n m p 2
Jp < 10 p τ (p ) , if p | q and f = g,
p

= p−1 , if p | q and f 6= g.
The above formula follows more directly from the Rankin–Selberg method and from the
calculation in [29,
QSection 2]. For√more details, see [29, Remark 3.2]. By the above calculation

for Jv , we have v Jv ≪ q −1 (q/ Q)2ε . Hence, (8.2) follows upon using Lemma 7.4.


Remark. We make the assumption that q is squarefree when f 6= g because to the author’s
knowledge, an analogue of (8.3) that ensures a bound of the shape
Λ( 21 , f × g × φ)
hφ, Fk Gk i2q ≪
Λ(1, Ad φ)Λ(1, Ad f )Λ(1, Ad g)
is only known when q is squarefree. See Hu [8, Section 3] for partial progress in this direction.
Proof of Proposition 6.1. The idea of this proof follows from [6, Proposition 3.3]. Recall the
spectral decomposition in (5.9). We begin with estimating the tail sums when |tφ | ≥ (kQ)ε
and |t| ≥ (kQ)ε . We repeat the proof of Lemma 7.3, but replace the weak subconvexity
bound with the convexity bound. It follows from (8.3) and (7.3) that
p 1
hφ, Fk Gk iq ≪ (1 + |tφ |)δ1 +ε (q/ Q)− 2 +θ+ε log(kQ). (8.7)
Similarly, we repeat the proof of Lemma 7.4, but replace the weak subconvexity bound with
the convexity bound. It follows from (8.6) and (7.3) that
p 1
hEt , Fk Gk iq ≪ (1 + |t|)δ2 +ε (q/ Q)− 2 +ε log(kQ). (8.8)
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 33

By Lemma 5.2, (8.7), and (8.8), for any A > 0, we have that
X Z
1
hψ, φi1hφ, Fk Gk iq + hψ, Et i1 hEt , Fk Gk iq dt
ε
4π |t|≥(kQ)ε
|tφ |≥(kQ)
X  M A p 1
≪A (1 + |tφ |)δ1 +ε (q/ Q)− 2 +θ+ε log(kQ)
1 + |tφ |
|tφ |≥(kQ)ε
Z  M A p
1 3 1
+ (1 + |t|) 8 +ε (1 + |t|)δ2 +ε (q/ Q)− 2 +ε log(kQ)dt
4π |t|≥(kQ)ε 1 + |t|
p 1
≪A M A (q/ Q)− 2 +θ+ε log(kQ) max ε T δ1 +2−A+ε ,
T ≥(kQ)

upon using Weyl’s law in (5.7). If M is less than a fixed power of log(kQ) and A is sufficiently
large with respect to ε, then the contribution in the range |tφ | ≥ (kQ)ε and |t| ≥ (kQ)ε is
p 1
OA ((q/ Q)− 2 +θ+ε (kQ)−A ). (8.9)
Now, we consider when |t|, |tφ | ∈ [M(log kQ)ε/4 , (kQ)ε ]. By Lemma 8.1 and (7.3), we have
the bounds p 1 1 ε
hφ, Fk Gk iq ≪ (1 + |tφ |)δ1 +ε (q/ Q)− 2 +θ+ε (log kQ) 2 + 2 ,
p 1 ε
(8.10)
hEt , Fk Gk iq ≪ (1 + |t|)δ2 +ε (q/ Q)− 2 +ε (log kQ) 2 .
By Lemma 5.2 and the estimates in (8.10), we obtain the bound
X Z
1
hψ, φi1 hφ, Fk Gk iq + hψ, Et i1 hEt , Fk Gk iq dt
ε/4 ε
4π |t|∈[M (log kQ)ε/4 ,(kQ)ε]
|tφ |∈[M (log kQ) ,(kQ) ]
X p 1 1 ε
≪A M A (1 + |tφ |)δ1 −A+ε (q/ Q)− 2 +θ+ε (log kQ) 2 + 2 (8.11)
|tφ |∈[M (log kQ)ε/4 ,(kQ)ε ]
Z p
3 1 ε
+ M A (1 + |t|) 8 +δ2 −A+ε (q/ Q)− 2 +ε (log kQ) 2 dt.
|t|∈[M (log kQ)ε/4 ,(kQ)ε ]

By Weyl’s law in (5.7) and the fact that log(kq)/ log(kQ) ≪ε (q/ Q)ε , (8.11) is
p 1 1 ε
≪A M A (q/ Q)− 2 +θ+ε (log kQ) 2 + 2 max T 2+δ1 −A+ε
M (log kQ)ε/4 ≤T ≤(kQ)ε

p Z (kQ)ε
1 ε 3
A
+ M (q/ Q)− 2 +ε (log kQ) 2 t 8 +δ2 −A+ε dt
M (log kQ)ε/4
p 1 1 ε ε
≪A M 2+δ1 +ε (q/ Q)− 2 +θ+ε (log kQ) 2 + 2 + 4 (2+δ1 −A+ε) .
By taking A > 2 + δ1 sufficiently large with respect to ε, the contribution in the range
|t|, |tφ | ∈ [M(log kQ)ε/4 , (kQ)ε ] is
p 1
OA (M 2+δ1 +ε (q/ Q)− 2 +θ+ε (log kQ)−A ). (8.12)
Next, we bound the contribution from the remaining ranges |tφ |, |t| ≤ M(log kQ)ε/4 . By
the Parseval formula, we have that
X Z
2 3 2 2 1
kψk2 = |hψ, 1i1| + |hψ, φi1| + |hψ, Et i1 |2 dt. (8.13)
π φ
4π R
34 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)
R
By (1.4), we have that kψk22 = Y0 (1) |ψ(z)|2 y −2 dxdy ≪ 1. Applying the Cauchy–Schwarz
inequality and Lemma 8.1, we obtain that
X
hψ, φi1 hφ, Fk Gk iq
|tφ |≤M (log kQ)ε/4
 X  21  X 2
 12
≪ |hψ, φi1|2 hφ, Fk Gk iq
(8.14)
|tφ |≤M (log kQ)ε/4 |tφ |≤M (log kQ)ε/4
 √
X (1 + |tφ |)2δ1 +2ε (q/ Q)−1+2θ+2ε  12
≪ kψk2 .
ε/4
L(1, Ad f )L(1, Ad g) (log kQ)1−ε
|tφ |≤M (log kQ)
Q
By Lemma 7.2, Weyl’s law, and the fact that p≤x (1 − δ/p) ≍ (log x)−δ , we have that (8.14)
is
δ1 +1+ε ε
p − 1 +θ+ε Y  λf (p)2 + λg (p)2 − 1 
≪M (log kq) (q/ Q) 2 1− . (8.15)
√ 2p
p≤k Q
Similarly, for the case of Eisenstein series, it follows from the Cauchy–Schwarz inequality,
Lemma 8.1, and Lemma 7.2 that
Z
hψ, Et i1 hEt , Fk Gk iq dt
|t|≤M (log kQ)ε/4
1 p 1
Y  λf (p)2 + λg (p)2  (8.16)
≪ M δ2 + 2 +ε (log kq)ε (q/ Q)− 2 +ε 1− .
√ 2p
p≤k Q

Putting together (8.9), (8.12), (8.15), and (8.16), we see that |hψ, Fk Gk iq − π3 hψ, 1i11f =g |
starts to decay only if M is smaller than some fixed power of log(kQ). In such case, the
estimates from (8.9) and (8.12) become negligible. We also observe that the estimate from
(8.15) is larger than the estimate from (8.16). Therefore, the rate of convergence is dominated
by the contribution from Hecke–Maaß cups forms in (8.15). 
8.2. Holowinsky’s approach. In this subsection, we prove Proposition 6.2. To apply
Holowinsky’s method, we will need to unfold hPm (·, Ψm), |Fk |2 iq and those alike, for example,
hE(·|Ψ0 ), |Fk |2 iq and hEt , |Fk |2 iq . Since the level of Pm (·, Ψm ) is 1, while the level of f can
be any positive integer q, the application of the unfolding method becomes more complicated
than what we see in Section 5.6. For completeness, we provide a detailed exposition on how
to unfold hPm (·, Ψm ), |Fk |2 iq in the next lemma. For more discussion, see [8, Section 5.2].
Lemma 8.2. Recall the notation in Section 5.2 and let R≥0 denote the set of nonnegative
real numbers. For any m ∈ Z and any Ψ ∈ Cc∞ (R≥0 ), a smooth compactly supported function
on R≥0 , we have that
XZ ∞ Z 1
dxj dyj
2
hPm (·, Ψ), |Fk | iq = Ψ(wj yj ) e(mwj xj )|Fk (z)|2 .
a ∈C yj =0 xj =0 yj2
j

Proof. Let Γ∞ \Γ0 (1)/Γ0 (q) = {τ } be a set of double-coset representatives as in Section 5.2.
Also, set wj = [Γ∞ : Γ∞ ∩ τj Γ0 (q)τj−1 ]. We begin with decomposing Γ∞ \Γ0 (1) into Γ0 (q)-
orbits, namely,
Γ∞ \Γ0 (1) = ⊔Γ∞ \(Γ∞ τ Γ0 (q)) = ⊔τ (τ −1 Γ∞ τ ∩ Γ0 (q)\Γ0 (q)).
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 35

By the above decomposition and the definition of Pm (·, Ψ), we have


Z X
2 dxdy
hPm (·, Ψ), |Fk | iq = Ψ(Im(τ γz))e(m Re(τ γz))|Fk (z)|2 2
Γ0 (q)\H y
τ ∈Γ∞ \Γ0 (1)/Γ0 (q)
γ∈τ −1 Γ∞ τ ∩Γ0 (q)\Γ0 (q)

Since z 7→ |Fk (z)|2 y −2dxdy is Γ0 (q)-invariant, the above integral equals


X Z
dxdy
Ψ(Im(τ z))e(m Re(τ z))|Fk (z)|2 2
−1 y
τ ∈Γ∞ \Γ0 (1)/Γ0 (q) τ Γ∞ τ ∩Γ0 (q)\H
X Z (8.17)
dxdy
= Ψ(y)e(mx)|Fk (τ −1 z)|2 2 .
Γ∞ ∩τ Γ0 (q)τ −1 \H y
τ ∈Γ∞ \Γ0 (1)/Γ0 (q)

Now we index τ with τj and let aj = τj−1 ∞ be the corresponding cusp. If we denote
C = C(Γ0 (q)) = {aj }j , then we have that (8.17) equals
XZ dxdy
Ψ(y)e(mx)|Fk (τj−1 z)|2 2
−1
aj ∈C Γ∞ ∩τj Γ0 (q)τj \H
y
X Z ∞ Z wj (8.18)
−1 2 dxdy
= Ψ(y)e(mx)|Fk (τj z)| ,
a ∈C y=0 x=0
y2
j

upon using that wj = [Γ∞ : Γ∞ ∩ τj Γ0 (q)τj−1 ] in the last step. By the change of variable
z 7→ wj z, (8.18) equals
XZ ∞ Z 1  dxdy
Ψ(Im(wj z))e(m Re(wj z))|Fk (τj−1 wj 1 z)|2 2 .
a ∈C y=0 x=0
y
j

Recall that σj = τj−1 wj 1 is the scaling matrix of aj . For each z ∈ H, we write zj = xj + iyj
for the change of variable zj = σj−1 z. Applying the change of variables x 7→ xj and y 7→ yj
and using that σj zj = z in the above integral, we obtain the lemma. 
Remark. In the decorrelation case, the assumption of q being squarefree helps us simplify
the above unfolding method for several reasons. For each positive divisor cj of q, there is
exactly one cusp in C[cj ], and the corresponding width w = q/(c2j , q) equals q/cj . Moreover,
as discussed in [28], the scaling matrix σj behaves as an Atkin-Lehner operator, and hence
f |k σj = ±f .
Consider hPm (·, Ψ), Fk Gk iq with q squarefree. By the above remark, for any positive divisor
d of q, we have that there is only one cusp with denominator q/d, and the width of such cusp
is d. By Lemma 8.2, we conclude that
XZ ∞ Z 1
dxdy
hPm (·, Ψ), Fk Gk iq = Ψ(dy) e(mdx)Fk (z)Gk (z) 2 . (8.19)
y=0 x=0 y
d|q

With the notation in Section 5.6, we now estimate |hψ, Fk Gk iq − π3 hψ, 1i11f =g |. By the
Poincaré decomposition in (5.16), we have
36 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

3 3b
hψ, Fk Gk iq − hψ, 1i1 1f =g = hE(·|Ψ0), Fk Gk iq − Ψ 0 (−1)1f =g
π  π 
X X
+ + hPm (·, Ψm ), Fk Gk iq . (8.20)
0<|m|<M (log kQ)ε |m|≥M (log kQ)ε

We will estimate the contribution from each range of m individually. In doing so, we need
the following propositions.
Proposition 8.3 (Proposition 3.17 of [29]). Recall the notation in Theorem 1.1 and Section
5.2. Let m 6= 0 be an integer, x ≥ 1 be a real number, and ε ∈ (0, 1). Let q⋄2 denote the largest
square divisor of q. If c is a positive divisor of q, then
Q
X x p≤x (1 + 2|λf (p)|/p)
ε e O(1)
|λ[c] (n1 )λ[c] (n2 )| ≪ q⋄ (log log(e q)) 2−ε
.
n ≥1,n =n +m≥1
(log ex)
1 2 1
max(n1 ,n2 )≤x

When f 6= g, we need the following proposition which follows from [28, Theorem 3.10].
Proposition 8.4. Recall the notation in Theorem 1.1 and in Section 5.2. Let m 6= 0 be an
integer and x ≥ 1 be a real number. If ε ∈ (0, 1), then we have that
Q
X x p≤x (1 + (|λf (p)| + |λg (p)|)/p)
|λf (n1 )λg (n2 )| ≪ 2−ε
.
n ≥1,n =n +m≥1
(log ex)
1 2 1
max(n1 ,n2 )≤x

Proof. We follow closely the proof of [28, Theorem 3.10]. Let λ1 (n) = |λf (n)| and λ2 (n) =
|λg (n)|, so that λ1 and λ2 are nonnegative multiplicative function satisfying
λ1 (n) ≤ τ (n) and λ2 (n) ≤ τ (n). (8.21)
We may assume that 1 ≤ m ≤ x. The following setup is the same as in [28, Theorem 3.10].
Fix α ∈ (0, 1/2) to be chosen later. Set

y = xα , s = α log log(x), z = x1/s .


For x ≫α 1, we have 10 ≤ z ≤ y ≤ x. Let N denote the set of positive integers. For each
n1 ∈ N, write n2 = n1 + m. Define the z-part of a positive integer to be the greatest divisor
of that integer supported on primes p ≤ z. There exist unique positive integers a, b, c such
that gcd(a, b) = 1 and ac (resp. bc) is the z-part of n1 (resp. n2 ); such triples a, b, c satisfy
p|abc ⇒ p ≤ z, c | m, and gcd(a, b) = 1. (8.22)
Write N = ⊔a,b,c Nabc for the fibers of n1 7→ (a, b, c). Then we have
X X X
λ1 (n1 )λ2 (n2 ) = λ1 (n1 )λ2 (n2 ).
n1 ∈N∩[1,x] a,b,c n1 ∈Nabc ∩[1,x]

To estimate the contribution from a, b, c for which ac > y or bc > y, we appeal to [6,
Section 4.2]. Since their proof applies to the shifted convolution sum of distinct multiplicative
functions λ1 and λ2 , we then obtain the same estimate, which is O(x(log x)−2 ).
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 37

We now consider the contribution from a, b, c for which ac ≤ y and bc ≤ y. The assumption
(8.21) implies that λ1 (n1 )λ2 (n2 ) ≤ 4s λ1 (ac)λ2 (bc), so that
X X X
λ1 (n1 )λ2 (n2 ) ≤ 4s λ1 (ac)λ2 (bc) · #(Nabc ∩ [1, x]).
a,b,c n1 ∈Nabc ∩[1,x] a,b,c
ac≤y,bc≤y

An application of the large sieve shows that if ac ≤ y, bc ≤ y, and x ≫ y 2 , then


x + (yz)2 1
#(Nabc ∩ [1, x]) ≪ . (8.23)
(log z) c ϕ(abc−1 m)
2 2

Since (yz)2 ≪α x, (log z)2 ≍α (log log x)−2 (log x)2 , 4s ≪ε (log x)ε (for α ≪ε 1), and ϕ(abc−1 m) ≥
ϕ(c−1 m)ϕ(a)ϕ(b), we see that Proposition 8.4 follows from the bound
X 1 m/c X X λ1 (ac)λ2 (bc) Y λ1 (p) + λ2 (p) 
≪ (log log x)O(1) 1+ .
c ϕ(m/c) ac≤y bc≤y ϕ(a)ϕ(b) p≤z
p (8.24)
c|m
p|c⇒p≤z p|ab⇒p≤z

To establish (8.24), we first note that


X X λ1 (ac)λ2 (bc)  Y X λ1 (pk+ordp (c) )  Y X λ2 (pk+ordp (c) ) 
≤ k) k)
.
ϕ(a)ϕ(b) ϕ(p ϕ(p (8.25)
ac≤y bc≤y p≤z k≥0 p≤z k≥0
p|ab⇒p≤z

Exactly as in the proof of [28, Theorem 3.10], we have that if i ∈ {1, 2}, then
Y X λi (pk+ordp (c) ) Y
k)
≪ (1 + λi (p)/p)(3 ordp (c) + 3). (8.26)
p≤z k≥0
ϕ(p p≤z

The proposition follows once we insert (8.26) into (8.25) and use the fact as in the proof of
[28, Theorem 3.10] that
X 1 m/c Y 6
(3 ordp (c) + 3)2 ≪ (log log x)10 . 
c ϕ(m/c) p≤z
c|m
p|c⇒p≤z

The next proposition will be needed along with Propositions 8.3 and 8.4.
Proposition 8.5 (Proposition 3.18 of [29]). Let x ≥ 2 be a real number and q be a positive
integer. If δ ∈ (0, 1), then
X [q/c2 , 1]ϕ((c, q/c)) q(log log(ee q))O(1)
≪ ,
log([q/c2 , 1]x)2−δ (log qx)2−δ
c|q

with an absolute implied constant.


To prove Proposition 6.2, we first estimate the sum when |m| ≥ M(log kQ)ε/3 in (8.20).
Lemma 8.6. Recall the notation in Theorem 1.1 and Section 5.6. If A > 0 and ε ∈ (0, 1)
are real numbers, then we have
X
hPm (·, Ψm ), Fk Gk iq ≪A,ε M(log kQ)−A .
|m|≥M (log kQ)ε/3
38 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

Proof. By the definition of Pm (z, Ψm ) and Lemma 8.2, we have


Z X dxdy
hPm (·, Ψm ), Fk Gk iq = e(m Re(γz))Ψm (Im(γz))Fk (z)Gk (z) 2 . (8.27)
Γ0 (q)\H y
γ∈Γ∞ \Γ0 (1)

Recall the discussion√ in Section 5.6. Since supp ψ̌ ⊂ F , the function Ψ(Im(γz)) is nonzero
only if Im(γz) ≥ 3/2. Invoking [14, Lemma 2.10], we have #{γ ∈ Γ∞ \Γ0 (1) : Im(γz) ≥

3/2} ≪ 1. Thus, by the rapid decay of Ψ(y) in (5.10) and the inequality of arithmetic and
geometric means, we obtain that
Z √ dxdy
hPm (·, Ψm ), Fk Gk iq ≪A #{γ ∈ Γ∞ \Γ0 (1) : Im(γz) ≥ 23 }(M/|m|)A |Fk (z)Gk (z)| 2
Γ0 (q)\H y
≪A (M/|m|)A (hFk , Fk iq + hGk , Gk iq ) ≪A (M/|m|)A .
Hence, the lemma follows upon summing over m ≥ M(log kQ)ε/3 and taking A sufficiently
large with respect to ε. 
The next lemma estimates the contribution from the mth incomplete Poincaré series, where
we shall sum up over the range 1 ≤ |m| ≤ M(log kQ)ε/3 . With the notation in Section 5.6, we
know that ψ̌(z) = ψ(z)|F and Ψ̌ is the extension of ψ̌ to H by Γ∞ . Since ψ ∈ Cc∞ (Y0 (1), M),
we have that for M ≥ 1
Z 1/2
max(j,1/2) (j) ∂j
y Ψm (y) = y max(j,1/2) j Ψ̌(x + iy)e(−mx) dx ≪j M j . (8.28)
−1/2 ∂y
Lemma 8.7. Recall the notation in Section 5.6 and the estimate in (8.28). Define
Q
p≤x (1 + 2|λf (p)|/p)
Mf (x) := .
L(1, Ad f )(log ex)2
and let q⋄2 denote the largest square divisor of q. If ε ∈ (0, 1), then we have
X 1 1
hPm (·, Ψm ), Fk Gk iq ≪ε M 4+ε (log kq)ε q⋄ε Mf (kq) 2 Mg (kq) 2 .
1≤|m|≤M (log kQ)ε/3

Proof. We first prove the case when f = g. Recall that Fk (z) = ρf (1)y k/2f (z) and |f (z)|2 y k =
|f |k σj |2 (zj )yjk as discussed in Sections 5.1 and 5.2. By Lemma 8.2 and the Fourier expansion
of f in (5.4), we have
X X k−1
hPm (·, Ψm ), |Fk |2 iq = |ρf (1)|2 λj (n1 )λj (n2 )(n1 n2 ) 2
aj ∈C n1 ≥1
n2 =n1 +mwj ≥1
Z ∞
· exp(−2π(n1 + n2 )yj )Ψm (wj yj )yjk−2 dyj , (8.29)
0
upon integrating over xj .
We then use the method in Luo–Sarnak [26, Section 2] to deal with the integral in (8.29).
R
Define the Mellin tranform of Ψm as Ψ b m (s) = ∞ Ψm (y)y s−1 dy. Since Ψm ∈ C ∞ (R≥0 ), Ψ
bm
0 c
is entire. For any integer j ≥ 0 and Re(s) < 0, integrating by parts j times and applying the
bound in (8.28), we have that
b m (s) ≪j Mj
Ψ , (8.30)
(1 + |s|)j
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 39

b m (s),
and hence (8.30) also holds for any real j ≥ 0. We also have the Mellin inversion of Ψ
namely, for any A > 0 and x > 0
Z Z
1 b m (s)x−s ds = 1 b m (−s)xs ds.
Ψm (x) = Ψ Ψ (8.31)
2πi (−A) 2πi (A)

Applying the Mellin inversion, we obtain that the integral in (8.29) becomes
Z Z ∞
b m (−s) ds
Ψ yjs+k−2wjs exp(−2π(n1 + n2 )yj ) dyj
(A) 0 2πi
Z b m (−s)w s Z ∞
Ψ j dyj ds
= s+k−1
(2π(n1 + n2 )yj )s+k−1 exp(−2π(n1 + n2 )yj ) (8.32)
(A) (2π(n1 + n2 )) 0 yj 2πi
Z
Γ(k − 1)  1 k−1 b m (−s)w s Γ(s + k − 1) ds
Ψ j
= ,
(4π)k−1 n1 +n 2
2
(A) (4π( n1 +n2 s
2
)) Γ(k − 1) 2πi

by using the definition of Γ-function in the last step. By Stirling’s formula, if Re(s) = A > 0,
then, as in [26, Eq. (2.3)], we have that

Γ(s + k − 1)   (1 + |s|)2 
A
= (k − 1) 1 + OA ≪A k A (1 + |s|)2 .
Γ(k − 1) k

Using (8.30) with j = 3 + ε, we have that the integral in (8.32) is

Z  max(n , n ) −A
M 3+ε wjA k A 2 3+ε 1 2
≪A,ε 3+ε n1 +n2 A (1 + |s|) |ds| ≪ A M max 1, . (8.33)
(A) (1 + |s|) (4π( 2 )) wj k

Hence, using (8.32) and (8.33) in (8.29), we conclude that

M 3+ε |ρf (1)|2 Γ(k − 1)


hPm (·, Ψm ), |Fk |2 iq ≪A
(4π)k−1
X X  √n n k−1  max(n , n ) −A
1 2 1 2
· λj (n1 )λj (n2 ) n1 +n2 max 1, . (8.34)
aj ∈C n ≥1 2
w j k
1
n2 =n1 +mwj ≥1

Upon using (5.2), (5.6), and applying the inequality of arithmetic and geometric means to

the factor n1 n2 /((n1 + n2 )/2), we have that (8.34) is

M 3+ε X X  max(n , n ) −A


1 2
≪A #C[c] λ[c] (n1 )λ[c] (n2 ) max 1, . (8.35)
L(1, Ad f )kq n1 ≥1
wc k
c|q
n2 =n1 +mwc ≥1
40 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

We decompose the inner sum of (8.35) dyadically and applying Proposition 8.3. Then the
inner sum of (8.35) is

X X  max(n , n ) −A
1 2
≪A |λ[c] (n1 )λ[c] (n2 )| max 1,
N =0 n ≥1,n =n +mw ≥1,
wc k
1 2 1 c
⌊2N−1 ⌋wc k≤max(n1 ,n2 )≤2N wc k
∞ Q
X 2N wc k p≤2N wc k (1 + 2|λf (p)|/p)
≪A q⋄ε (log log(ee q))O(1) ε · 2−N A
N =0
log(e2N wc k)2− 3 (8.36)
Q
wc k p≤wc k (1 + 2|λf (p)|/p)
≪A q⋄ε (log log(ee q))O(1) ε
log(ewc k)2− 3
X∞  log(ew k) 2− ε3 Y 
c
· 2N (1−A) (1 + 2|λf (p)|/p) .
N =0
log(e2N wc k) N wc k≤p≤2 wc k

Using Deligne’s bound |λf (p)| ≤ 2 and choosing A = 2, we have that the summation in the
last line above is O(1). Hence, by inserting the bound from (8.36) into (8.35) and using the
facts that wc = [q/c2 , 1] and #C[c] = ϕ((c, q/c)), we obtain that
q ε (log log(ee q))O(1) X ϕ((c, q/c))[q/c2, 1] Y
hPm (·, Ψm ), |Fk |2 iq ≪ M 3+ε ⋄ ε
2 , 1]k))2− 3
(1 + 2|λf (p)|/p).
L(1, Ad f )q (log([q/c
c|q p≤kq

Applying Proposition 8.5 and the definition of Mf (x), we conclude that


q⋄ε (log log(ee q))O(1) Y
hPm (·, Ψm ), |Fk |2 iq ≪ M 3+ε ε (1 + 2|λf (p)|/p)
L(1, Ad f )(log kq)2− 3 p≤kq

≪ M 3+ε (log kq) 3 q⋄ε Mf (kq).
The lemma then follows by summing the above estimate over m in [1, M(log kQ)ε/3 ].
In the case that f 6= g, we can proceed in the same manner, except now we replace Lemma
8.2 by (8.19) and Proposition 8.3 by Proposition 8.4 in the above proof. 
It remains for us to estimate the contribution from the incomplete Eisenstein series. We
start with writing the Fourier expansion
X
E(z|Ψ0 ) = aΨ0 ,0 (y) + aΨ0 ,l (y)e(lx)
|l|≥1

and recalling the following lemma.


Lemma 8.8 (Lemma 4.2 of [9]). Let M ≥ 1, and let Ψ ∈ Cc∞ (R≥0 ) be a function such that
for all integer j ≥ 0, y max(j,1/2) Ψ(j) (y) ≪j M j . If θ(s) = π −s Γ(s)ζ(2s), then for all y > 0
Z  1
3b 1 b +it θ(1/2 − it) 1 −it 
aΨ,0 (y) = Ψ(−1) + Ψ(−1/2 − it) y 2 + y2 dt. (8.37)
π 2π R θ(1/2 + it)
For any integer l 6= 0, if ε > 0 and A ≥ 0, then for all y > 0
√ 2
 M A  1 ε
aΨ,l (y) ≪A τ (|l|) yM 3 +ε 1+ . (8.38)
|ly| |ly|
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 41

In the next lemma, we state the estimate for the incomplete Eisenstein series. We will see
b 0 (−1)1f =g is obtained in this step.
that the expected main term π3 Ψ
Lemma 8.9. Recall the definitions of Mf (x) and q⋄ in Lemma 8.7, the notation in Section
5.6, and the bound in (8.28). If ε ∈ (0, 1), then we have that
3b 5
+ε ε 1 1
hE(·|Ψ0), Fk Gk iq − Ψ 0 (−1)1f =g ≪ M 3 q⋄ (log kq)ε Mf (kq) 4 Mg (kq) 4 .
π
To prove Lemma 8.9, we shall begin by using the ideas from [6] and [22]. Let 1 ≤ Y ≤
(log kq)5 be chosen later. Let h be a fixed nonnegative smooth compactly supported function
R∞
on R. Assume supp h ⊂ [1, 2], h(j) (y) ≪j 1, and bh(−1) = 0 h(y)y −2 dy = π/3. By integra-
tion by parts, b
h(−s) ≪ (1 + |s|)−A for any s > 0 and A ≥ 0. Define the smoothed incomplete
Eisenstein series X
E Y (z|h) = h(Y Im(γz)).
γ∈Γ∞ \Γ0 (1)
By Mellin inversion, we have
X Z ds
Z
ds
Y
E (z|h) = b
h(−s)Y s Im(γz)s = b
h(−s)Y s E(s, z) .
(2) 2πi (2) 2πi
γ∈Γ∞ \Γ0 (1)

Moving the contour to Re(s) = 1/2 and using that ress=1 E(s, z) = 3/π, we obtain that
Z
Y b ds
E (z|h) = Y + h(−s)Y s E(s, z) . (8.39)
(1/2) 2πi
Define
Z
Y b ds
I(Y ) := hE (·|h)E(·|Ψ0), Fk Gk iq = h(−s)Y s hE(s, ·)E(·|Ψ0), Fk Gk iq .
(2) 2πi
By (8.39), we obtain that
Z
b ds
I(Y ) = Y hE(·|Ψ0 ), Fk Gk iq + h(−s)Y s hE(s, ·)E(·|Ψ0), Fk Gk iq .
(1/2) 2πi
Hence, we have that
I(Y )Y −1 − hE(·|Ψ0), Fk Gk iq
Z ∞
−1 b dt
=Y h(−1/2 − it)Y 1/2+it hE(1/2 + it, ·)E(·|Ψ0), Fk Gk iq . (8.40)
−∞ 2π
We now estimate the term hE(1/2 + it, ·)E(·|Ψ0), Fk Gk iq in the right-hand side of (8.40).
We know from (5.8) that |E(1/2 + it, z)| ≪ y 1/2 (1 + |t|)3/8 for any y ≥ 1/2, and know that
y 1/2 Ψ0√
(y) ≪ 1 due to (8.28). Since E(1/2 + it, z) is Γ0 (1)-invariant and Ψ0 (y) is nonzero only
if y ≥ 3/2, it follows that
X
|E(z|Ψ0 )E(1/2 + it, z)| ≤ |Ψ0 (Im γz)E(1/2 + it, γz)|
γ∈Γ∞ \Γ0 (1)
X p 3
≪ |Ψ0 (Im γz)| Im γz(1 + |t|) 8 (8.41)
γ∈Γ∞ \Γ0 (1)
3 √
≪ (1 + |t|) 8 #{γ ∈ Γ∞ \Γ0 (1) : Im γz ≥ 3/2},
42 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

which is O((1 + |t|)3/8 ) upon using [14, Lemma 2.10]. Using (8.41) and the inequality of
arithmetic and geometric means, we have that
3 3
hE(1/2 + it, ·)E(·|Ψ0), Fk Gk iq ≪ (1 + |t|) 8 (h1, |Fk |2 iq + h1, |Gk |2 iq ) ≪ (1 + |t|) 8 . (8.42)
Thus, inserting (8.42) into (8.40) and using the rapid decay of b
h(−1/2 − it), we have that
1
I(Y )Y −1 − hE(·|Ψ0 ), Fk Gk iq ≪ Y − 2 , (8.43)
which implies that to estimate hE(·|Ψ0), Fk Gk iq , we need to estimate I(Y ).
Similarly to Lemma 8.2, we can apply the unfolding method with I(Y ) as follows.
XZ ∞ Z 1
dxj dyj
I(Y ) = h(Y wj yj ) E(wj zj |Ψ0 )Fk (z)Gk (z) 2
. (8.44)
a ∈C y j =0 x j =0 y j
j

We first consider I(Y ) in the case that f = g. By the fact that |Fk (z)|2 = |ρf (1)|2 |f |k σj (zj )|2 yjk
as in Section 5.1, we have that
XZ ∞ Z 1
dxj dyj
2
I(Y ) = |ρf (1)| h(Y wj yj ) E(wj zj |Ψ0 )|f |k σj (zj )|2 yjk . (8.45)
aj ∈C yj =0 xj =0 yj2
P
In (8.45), we write E(wj zj |Ψ0 ) = l∈Z aΨ0 ,l (wj yj )e(lwj xj ), expand f |k σj (zj ) as in (5.4), and
integrate (8.45) over xj . Then it follows that
XX X k−1
I(Y ) = |ρf (1)|2 λj (n1 )λj (n2 )(n1 n2 ) 2
aj ∈C l∈Z n1 ≥1
n2 =n1 +wj l≥1
Z ∞
dyj
· h(Y wj yj )aΨ0 ,l (wj yj ) exp(−2π(n1 + n2 )yj )yjk . (8.46)
0 yj2
If we let
XX Z ∞
2 2 k−1 dyj
I0 (Y ) = |ρf (1)| λj (n) n h(Y wj yj )aΨ0 ,0 (wj yj ) exp(−4πnyj )yjk , (8.47)
aj ∈C n≥1 0 yj2

and, for l 6= 0, let


X X k−1
Il (Y ) = |ρf (1)|2 λj (n1 )λj (n2 )(n1 n2 ) 2

aj ∈C n1 ≥1
n2 =n1 +lwj ≥1
Z ∞
dyj
· h(Y wj yj )aΨ0 ,l (wj yj ) exp(−2π(n1 + n2 )yj )yjk , (8.48)
0 yj2
P
then I(Y ) = l∈Z Il (Y ).
On the other hand, in the case that f 6= g, by the remark of Lemma 8.2, we obtain the
following simpler expressions for I0 (Y ) and Il (Y ) with l 6= 0:
XX Z ∞
k−1 dy
I0 (Y ) = ρf (1)ρg (1) λf (n)λg (n)n h(Y dy)aΨ0 ,0 (dy) exp(−4πny)y k 2 , (8.49)
n≥1 0 y
d|q
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 43
X X k−1
Il (Y ) = ρf (1)ρg (1) λf (n1 )λg (n2 )(n1 n2 ) 2

d|q n1 ≥1
n2 =n1 +dl≥1
Z ∞
dy
· h(Y dy)aΨ0,l (dy) exp(−2π(n1 + n2 )y)y k . (8.50)
0 y2
Recall that to prove Lemma 8.9, we need to estimate I(Y ). We first estimate the contri-
bution from Il (Y ) when l 6= 0 in the next lemma.
Lemma 8.10. Recall the notation in Lemma 8.9 and the definition of Il (Y ) in (8.48). If
ε ∈ (0, 1), then
X 5 1 1 1
Il (Y )Y −1 ≪ M 3 +ε Y 2 +ε q⋄ε (log kq)ε Mf (kq) 2 Mg (kq) 2 .
l∈Z\{0}

Proof. We first prove the case when f = g. The proof uses similar ideas as in Lemma 8.7.
By applying Mellin inversion for h(Y wj yj ), Stirling’s formula, and the rapid decay of bh(−s),
we obtain that
Z ∞
dyj Γ(k − 1)  1 k−1  max(n , n ) −A
1 2
h(Y wj yj ) exp(−2π(n1 + n2 )yj )yjk 2 ≪A k−1 n1 +n2 max 1, .
0 y j (4π) 2
Y w j k

By using (8.38), the above bound, and the fact that supp h ⊂ [1, 2], we have that the integral
in (8.48) is
2
 MY A Γ(k − 1)  1 k−1  max(n , n ) −A
+ 2ε − 12 + 2ε 1 2
≪A M 3 Y τ (|l|) k−1 n1 +n2 max 1, . (8.51)
|l| (4π) 2
Y wj k
Applying (8.51) in (8.48), we then obtain that

Γ(k − 1) 2 + ε − 1 + ε  MY A
Il (Y ) ≪A |ρf (1)|2 M 3 2Y 2 2 τ (|l|)
(4π)k−1 |l|
X X  √n n k−1  max(n , n ) −A
1 2 1 2
· λj (n1 )λj (n2 ) n1 +n2 max 1, (8.52)
a ∈C n ≥1 2
Y w j k
j 1
n2 =n1 +lwj ≥1

Invoking (5.2) and (5.6), and applying the inequality of arithmetic and geometric means for

n1 n2 /((n1 + n2 )/2) in (8.52), we have that
2 ε 1 ε
MY
A
M 3 + 2 Y − 2 + 2 τ (|l|) |l|
Il (Y ) ≪A
kqL(1, Ad f )
X X  max(n , n ) −A
1 2
· #C[c] λ[c] (n1 )λ[c](n2 ) max 1, . (8.53)
n1 ≥1
Y wc k
c|q
n2 =n1 +lwc ≥1

Similarly to the proof of Lemma 8.7, applying Proposition 8.3 dyacdically in (8.53) and
using the facts that wc = [q/c2 , 1] and #C[c] = ϕ((c, q/c)), we obtain that
44 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

2 ε 1 ε MY
A
M 3 + 2 Y − 2 + 2 τ (|l|) |l|
Il (Y ) ≪A q⋄ε (log log(ee q))O(1)
kqL(1, Ad f )
X ϕ((c, q/c))[q/c2, 1] Y
· 2− 2ε
Yk (1 + 2|λf (p)|/p).
c|q
log([q/c2 , 1]k) p≤Y kq

By Proposition 8.5 and observing that the contribution from the product when kq ≤ p ≤ Y kq
is trivial since we assume 1 ≤ Y ≤ (log kq)5 , we then have
2 ε 1 ε A
M 3 + 2 Y − 2 + 2 τ (|l|) M|l|Y qq⋄ε (log log(ee q))O(1) Y
Il (Y ) ≪A ε Y k (1 + 2|λf (p)|/p)
kqL(1, Ad f ) (log kq)2− 2 p≤kq
2
+ ε 1
+ ε
M Y
A ε ε
≪A M 3 2 Y 2 2 τ (|l|) |l| q⋄ (log kq) Mf (kq),
by the definition of Mf (x) in Lemma 8.7. Therefore, we obtain that
X 2 ε 1 ε
X A
Il (Y )Y −1 ≪A M 3 + 2 Y − 2 + 2 q⋄ε (log kq)ε Mf (kq) τ (|l|) MY
|l|
l∈Z\{0} l∈Z6=0
5 1
+ε +ε
≪M 3 Y 2 q⋄ε (log kq)ε Mf (kq),
when taking A = 1 + ε/2 in the last step.
In the case that f 6= g, we proceed similarly, except that Il (Y ) is now simpler (see (8.50))
and we need to replace Proposition 8.3 by Proposition 8.4 in the above proof. 
Lemma 8.11. Recall the notation in Lemma 8.9 and the definition of I0 (Y ) in (8.47). If
ε ∈ (0, 1), then we have that
3b 1+ε − 12
I0 (Y )Y −1 − Ψ 0 (−1)1f =g ≪ M Y (log kQ)ε .
π
Proof. By (8.44) and the expression of aΨ0 ,0 (y) in (8.37), we have that
3b XZ ∞ Z 1
dxj dyj X ∞
Z
I0 (Y ) = Ψ0 (−1) h(Y wj yj ) Fk (z)Gk (z) + h(Y wj yj )
π aj ∈C yj =0 xj =0 yj2 aj ∈C yj =0
Z 1  Z  1  
·
1 b 0 (− 1 − it) (wj yj ) 21 +it + θ( 2 − it) (wj yj ) 21 −it dt Fk (z)Gk (z) dxj dyj . (8.54)
Ψ
xj =0 2π R
2
θ( 12 + it) yj2
Invoking (8.39), we obtain that the first summation on the right-hand side of (8.54) equals
Z
Y b ds
hE (·|h), Fk Gk iq = Y h1, Fk Gk iq + h(−s)Y s hE(s, ·), Fk Gk iq . (8.55)
(1/2) 2πi
Applying (8.2), the rapid decay of b h(−s), and (7.3), we have that
 p 1 1

hE Y (·|h), Fk Gk iq = Y 1f =g + O (q/ Q)− 2 +ε Y 2 (log kQ)ε . (8.56)
Since supp h ⊂ [1, 2], so h(Y wj yj ) is nonzero only if wj yj ≍ Y −1 . Then, by the rapid decay
b
of Ψ(s) as in (8.30), the second sum on the right-hand side of (8.54) is
XZ ∞ Z 1
dxj dyj
1+ε − 12
≪ (M Y ) h(Y wj yj ) |Fk (z)||Gk (z)| . (8.57)
a ∈C yj =0 xj =0 yj2
j
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 45

By the inequality of arithmetic and geometric means, we have that (8.57) is


1
≪ M 1+ε Y − 2 (hE Y (·|h), |Fk |2 iq + hE Y (·|h), |Gk |2 iq )
1 p 1 1
(8.58)
≪ M 1+ε Y − 2 (Y + (q/ Q)− 2 +ε Y 2 (log kQ)ε ),
where we apply (8.56) in the last step.
Inserting (8.56) and (8.58) in (8.54), we obtain that
3b  p − 1 +ε 1 
ε
I0 (Y ) = Ψ 0 (−1) Y 1 f =g + O (q/ Q) 2 Y 2 (log kQ)
π   
1+ε − 12
p − 1 +ε 1 ε
+O M Y Y + (q/ Q) 2 Y (log kQ)
2 .
b 0 (−1) ≪ 1 and Y ≥ 1.
Hence, the lemma follows from the facts that π3 Ψ 
Proof of Lemma 8.9. Applying Lemmas 8.10 and 8.11 in (8.43), we obtain that
3b
hE(·|Ψ0 ), Fk Gk iq − Ψ 0 (−1)1f =g
π
1 5 1 1 1 1
≪ M 1+ε Y − 2 (log kQ)ε + M 3 +ε Y 2 +ε q⋄ε (log kq)ε Mf (kq) 2 Mg (kq) 2 + Y − 2
5 1 1
≪ M 3 +ε q⋄ε (log kq)ε Mf (kq) 4 Mg (kq) 4 ,
1 1
upon choosing Y = Mf (kq)− 2 Mf (kq)− 2 and noting that Q ≤ q 2 . 
Proof of Proposition
Q 6.2. Consider the factor Mf (kq). When applying
√ Lemma 7.2 and the
facts that p≤x (1 − δ/p) ≍ (log x)−δ and log(kq)/ log(kQ) ≪ (q/ Q)ε , we have that
Y  (|λf (p)| − 1)2  q ε
Mf (kq) ≪ 1− √ (log kq)ε . (8.59)
√ p Q
p≤k Q

Using Lemmas 8.6, 8.7, and 8.9 in (5.16), we obtain that


3 5 1 1
hψ, Fk Gk iq − hψ, 1i1 1f =g ≪A M 3 +ε q⋄ε (log kq)ε Mf (kq) 4 Mg (kq) 4
π
1 1
+ M 4+ε q⋄ε (log kq)ε Mf (kq) 2 Mg (kq) 2 + M(log kQ)−A .

Therefore, by (8.59) and the fact that q⋄ε ≪ (q/ Q)ε , we conclude that
3
hψ, Fk Gk iq − hψ, 1i1 1f =g
π
 q ε Y  (|λf (p)| − 1)2 + (|λg (p)| − 1)2 
≪ M 4+ε (log kq)ε √ 1− . 
Q √ 4p
p≤k Q

Appendix A. The GL2 × GL3 approximate functional equation and QUE


by Jesse Thorner

Let f be a holomorphic cuspidal Hecke eigenform of even weight k ≥ 2 and φ be a Hecke–


Maaß cusp form, both of level 1. Each corresponds with a certain unitary cuspidal automor-
phic representation of GL2 (AQ ). Under the stated hypotheses, the Hecke eigenvalues λf (n)
and λφ (n) of f and φ (respectively) are all real and multiplicative.
46 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

Let Fk (x + iy) = ρf (1)y k/2f (x + iy), where ρf (1) is a normalizing constant so that if
h·, ·i is the Petersson inner product for Y0 (1), then h1, |Fk |2 i = 1. Iwaniec [15, Section 11]
and Lester, Matomäki, and Radziwill [22, Section 4] describe a method for showing that
limk→∞hφ, |Fk |2 i = 0 without a weak subconvexity bound for L( 12 , Ad f × φ). This appendix
shows that their proof is incomplete without assuming bounds for L(s, Ad f ) near Re(s) = 1
that are not yet known. While the dependence on φ is kept effective in [22], we suppress it
here for simplicity.
It is proved in [15, (3.2) and (11.1)] and [22, (4.2)] that
(log log k)3 1
Y λf (p2 )  1
2 1
|hφ, |Fk | i| ≪A,φ,ε √ |L( 2 , Ad f × φ)| 2 1− + . (A.1)
k p≤k
p (log k)A

A bound for L( 12 , Ad f × φ) essentially of the form


X |λf (n2 )λφ (n)| k
|L( 12 , Ad f × φ)| ≪A,φ,ε √ + (A.2)
n (log k)A
n≤k 2 (log k)ε

(a weak form of the approximate functional equation) was asserted without proof in [15,
Section 11.1] and [22, (4.3)]. These appear to neglect the expression for L(s, Ad f × φ) below.
Theorem A.1. Let f be a holomorphic cuspidal Hecke eigenform and φ be a Hecke–Maaß
1
newform, each of level 1. There exists an absolute constant 0 < ω ≤ 10 and a function H(s)
1
such that in the region Re(s) ≥ 2 − ω, H(s) is holomorphic, H(s) ≍ 1, and
X λf (n2 )λφ (n)
L(s, Ad f × φ) = H(s)L(2s, Ad f )L(2s, Ad φ) .
ns
p|n=⇒p≥19

Proof. By multiplicativity of the Hecke eigenvalues, it suffices to work locally. Let p be prime.
For any π ∈ F2 of conductor 1, we have that
Y X λπ (pj )∞
1
Lp (s, π) = = ,
1 − απ (p)j p−s j=0
p js
j∈{−1,1}
Y ∞
X λAd π (pj )
1
Lp (s, Ad π) = = .
1 − απ (p)2j p−s j=0
pjs
j∈{−1,0,1}

We write αp = αf (p), βp = αφ (p), and


Y Y ∞
X
1 λAd f ×φ (pj )
Lp (s, Ad f × φ) = = .
j∈{−1,0,1} ℓ∈{−1,1}
1 − αp2j βpℓ p−s j=0
pjs

Note that
j j
X X
j
λf (p ) = αp2r−j , j
λφ (p ) = λρ (p ) =j
βp2r−j .
r=0 r=0

By (A.3) and the Ramanujan conjecture for π, we have the bounds


|αp | = 1, max{|βp |, |βp |−1 } ≤ p7/64 . (A.3)
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 47

Recall the well-known Ramanujan-type identity


∞ ∞ Pj Pj
X λf (pj )λφ (pj ) X ( r=0 αp2r−j )( r=0 βp2r−j )
js
= js
= (1 − p−2s )Lp (s, f × φ). (A.4)
j=0
p j=0
p
We will now establish a similar identity for Lp (s, Ad f × φ). Consider the rational function
1 + (βp + βp−1 )p−s − (αp2 + 1 + αp−2 )p−2s
. (A.5)
(1 − αp2 βp p−s )(1 − αp2 βp−1 p−s )(1 − αp−2 βp p−s )(1 − αp−2 βp−1 p−s )
When Re(s) > 7/64, we expand (A.5) as a polynomial in p−s times a product of four geometric
sums and conclude that (A.5) equals
∞ P Pj
X ( 2j
r=0 α 2r−2j
p )( 2r−j
r=0 βp ) X ∞
λf (p2j )λφ (pj )
= .
j=0
pjs j=0
pjs
On the other hand, we identify (A.5) as a product of local L-functions, namely
1 + λφ (p)p−s − λf (p2 )p−2s
Lp (s, Ad f × φ) .
Lp (s, φ)
Therefore, if Re(s) > 7/64, then
X∞
Lp (s, φ) λf (p2j )λφ (pj )
Lp (s, Ad f × φ) = , (A.6)
1 + λφ (p) − λf (p2 )p−2s j=0 pjs
provided that the denominator for the leading factor on the right-hand side is nonzero. We
verify that if p ≥ 19 and Re(s) ≥ 25 , then |1 + λφ (p)p−s − λf (p2 )p−2s | > 0.06 using (A.3) and
Mathematica 12. For such p and s, we find that
Lp (s, φ)
= 1 + (λf (p2 ) + λφ (p2 ))p−2s + · · · , (A.7)
1 + λφ (p)p−s − λf (p2 )p−2s
which agrees with the beginning of the expansion
Lp (2s, Ad f )L(2s, Ad φ) = 1 + (λf (p2 ) + λφ (p2 ))p−2s + · · · .
This leads us to define
Lp (s, φ)
Hp (s) = .
(1 + λφ (p)p−s − λf (p )p−2s )Lp (2s, Ad f )Lp (2s, Ad φ)
2

By our above numerical calculation and (A.3), we know that if p ≥ 19, then in the region
Re(s) ≥ 52 , Hp (s) is holomorphic and

X λf (p2j )λφ (pj )
Lp (s, Ad f × φ) = Hp (s)Lp (2s, Ad f )Lp (2s, Ad φ) ,
j=0
pjs
2 −3s
Hp (s) = 1 − λf (p) λφ (p)p + (λf (p) (λφ (p)2 + 1) − 2)p−4s + · · · .
2

Since the coefficients of p−s and p−2s in the expansion of Hp (s) equal zero, (A.3) implies
1
that there exists an absolute constant 0 < ω ≤ 10 such that in the region Re(s) ≥ 12 − ω, the
functions
Y Y Lp (s, Ad f × φ)
H1 (s) = Hp (s), H2 (s) =
p≥19
L (2s, Ad f )Lp (2s, Ad φ)
p<19 p
48 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

are holomorphic and ≍ 1 in the region. To finish, we choose H(s) = H1 (s)H2 (s). 
Using Theorem A.1 and the ideas in [16, Section 5.2], we prove an approximate functional
equation for L(s, Ad f × φ). Let κ be the root number of L(s, Ad f × φ), L∞ (s, Ad f × φ) be
the gamma factor, and
2A + 6 1
A > 2, 0 < ε < 1, + > 3. B= (A.8)
ε 2
Using the functional equation for L(s, Ad f × φ) and Theorem A.1, we derive the identity
X λf (n2 )λφ (n)
L( 21 , Ad f × φ) = (1 + κ) √ W (n), (A.9)
n
p|n=⇒p≥19
Z 2
1 B+i∞
L∞ ( 21 + s, Ad f × φ) es ds
W (n) = H( 21 + s)L(1 + 2s, Ad f )L(1 + 2s, Ad φ) .
2πi B−i∞ L∞ ( 12 , Ad f × φ) ns s
We proceed to estimate W (n).
Lemma A.2. With the notation and hypotheses in Theorem A.1, (A.8), and (A.9), there
holds (
(k 2 /n)B if n > k 2 (log k)ε ,
W (n) ≪A,φ,ε
(log k)3 log log k if n ≤ k 2 (log k)ε .

Proof. Recall (A.8) and (A.9). Write s = σ + it. For −ω < σ ≤ B, define
2
L∞ ( 21 + s, Ad f × φ) es
Y (s) = YAd f ×φ (s) = H( 12 + s)L(1 + 2s, Ad f )L(1 + 2s, Ad φ) ,
L∞ ( 21 , Ad f × φ) ns
so that
Z B+i∞
1 ds
W (n) = Y (s) .
2πi B−i∞ s
Theorem A.1 implies that H( 12 + s) ≪ 1. Using GRC for f and proceeding as in [23, Theorem
2] for φ, we find that
(
1 if σ = B,
|L(1 + 2s, Ad f )L(1 + 2s, Ad φ)| ≪φ 3 1
(log k) (|t| + 1) if σ = 2 log k
.
Also, Stirling’s formula yields
L∞ ( 21 + s, Ad f × φ) es
2  k 2 σ
≪ A,φ,ε (|t| + 1)−200 .
L∞ ( 21 , Ad f × φ) ns n
It follows that
(
(k 2 /n)B (|t| + 1)−100 if σ = B,
Y (σ + it) ≪A,φ,ε 1 (A.10)
n−1/(2 log k) (log k)3 (|t| + 1)−100 if σ = 2 log k
.

When n > k 2 (log k)ε , we estimate the integral in (A.9) on the line σ = B. Otherwise, we
push the contour to σ = 1/(2 log k). The lemma follows from (A.10) in both cases. 
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 49

Remark. In our application of the Lemma A.2, we will sum λf (n2 )λφ (n) over n ≤ k 2 (log k)ε
instead of n ≤ k 2+ε because our result is sensitive to powers of log k. Consequently, our
bounds for W (n) would not improve in a manner useful to our application if we push the
contour for the integral that defines W (n) further to the left. In order to push further
to the left while obtaining something useful for our application, we would need bounds for
L(1+2s, Ad f )L(1+2s, Ad φ) near Re(s) = 1 than are stronger than what is currently known.
Theorem A.1 and Lemma A.2 yield the following bound for L( 12 , Ad f × φ).
Theorem A.3. If f is a holomorphic cuspidal Hecke eigenform of even weight k ≥ 2 and
level 1, φ is a Hecke–Maaß cusp form of level 1, A > 2, and ε > 0, then
X λf (n2 )λφ (n)  k 
1
L( 2 , Ad f × φ) = (1 + κ) √ W (n) + OA,φ,ε
2 ε
n (log k)2A+2
n≤k (log k)
p|n=⇒p≥19
X |λf (n2 )λϕ (n)| k
≪A,φ,ε (log k)3 (log log k) √ + .
n (log k)2A+2
n≤k 2 (log k)ε

Proof. We estimate the tail of the sum in (A.9) using Lemma A.2, the Cauchy–Schwarz
inequality, Rankin–Selberg theory, and (A.8). We have that
X |λf (n2 )λφ (n)| X |λf (n2 )λφ (n)|
√ W (n) ≪A,φ,ε k 2B 1 (A.11)
2 ε
n 2 ε nB+ 2
n>k (log k) n>k (log k)
p|n=⇒p≥19
 X |λf (n2 )|2  12  X |λφ (n)|2  12
≪A,φ,ε k 2B 1 1 .
n>k 2 (log k)ε
nB+ 2 n>k 2 (log k)ε
nB+ 2
Since X X
|λφ (n)|2 ≤ λφ×φ (n) ≪φ x (A.12)
n≤x n≤x
by Lemma 3.3, Theorem 3.1, and contour integration, it follows from partial summation that
X |λφ (n)|2 1−2B ( 12 −B)ε
B+ 2 1 ≪ B,φ,ε k (log k) .
2
n>k (log k) ε n
P P
Using (A.3), we find that n≤x |λf (n2 )|2 ≤ n≤x τ9 (n) ≪ x(log x)8 and
X |λf (n2 )|2 1

B+ 12
≪B,φ,ε k 1−2B (log k)8+( 2 −B)ε .
n>k 2 (log k)ε
n
We conclude that (A.11) is OA,φ,εk(log k)−2A−2 . The condition p | n=⇒p ≥ 19 is dropped by
nonnegativity, so the theorem follows. 
We now apply Theorem A.3 to (A.1).
Theorem A.4. If f is a holomorphic cuspidal Hecke eigenform of even weight k ≥ 2 and
level 1, φ is a Hecke–Maaß cusp form of level 1, A > 2, and ε > 0, then
3
Y λf (p2 ) + 41 (1 − λf (p2 ))2  1
2 +ε
hφ, |Fk | i ≪A,φ,ε (log k) 2 1− + .
p≤k
p (log k)A
50 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

Proof. By Theorem A.3 and the Cauchy–Schwarz inequality, L( 21 , Ad f × φ)1/2 is


3 1
 X λf (n2 )2  14  X λφ (n)2  14 1
≪A,φ,ε (log k) 2 (log log k) 2 √ √ + . (A.13)
2 ε
n 2 ε
n (log k)A+1
n≤k (log k) n≤k (log k)

It follows from (A.12) and partial summation that (A.13) is


1 3
 X λf (n2 )2  14 1
+ 8ε 1
≪A,φ,ε k (log k)
4 2 (log log k) 2 √ + . (A.14)
2 ε
n (log k)A+1
n≤k (log k)

We apply [22, (4.4)] to conclude that (A.14) is


√ 3
Y  λf (p2 ) + 14 (1 − λf (p2 )2 )  1
≪A,φ,ε k(log k) 2 +ε 1− +
2 ε
p (log k)A+1
p≤k (log k)
√ 3
Y λf (p2 ) + 41 (1 − λf (p2 )2 )  1
≪A,φ,ε k(log k) 2 +ε 1− + A+1
.
p≤k
p (log k)

The desired result follows from (7.2), (7.3), and (A.1). 


The quality of the available bounds on hφ, |Fk |2 i affects the rate of convergence in holomor-
phic QUE. By the Hecke relations, λf (p2 ) = λf (p)2 − 1. Assume that there exists an absolute
constant θ ≥ 0 such that
Y λf (p2 ) + 14 (1 − λf (p2 )2 )  1
hφ, |Fk |2 i ≪A,φ,ε (log k)θ+ε 1− +
p≤k
p (log k)A
(A.15)
Y λf (p)2 − 1 + 14 (1 − (λf (p)2 − 1)2 ) − θ  1
ε
≍A,φ,ε (log k) 1− + A
.
p≤k
p (log k)

By a different method, Holowinsky (see [7, (1.2)] or [15, (1.14)]) proved that
Y (|λf (p)| − 1)2 
2 ε
hφ, |Fk | i ≪φ,ǫ (log k) 1− .
p≤k
2p

We synthesize this bound with (A.13) using the fact that if ξ ∈ (0, 1) and a, b > 0, then
min{a, b} ≤ aξ b1−ξ . Let ξ ∈ (0, 1), and let A be sufficiently large with respect to ξ.
ξ  1 
2 −δ+ε 2 2 2 2
hφ, |Fk | i ≪A,φ,ε (log k) , δ = max min (λ−1) +(1−ξ) λ −1+ (1−(λ −1) )−θ .
0≤ξ≤1 0≤λ≤2 2 4
(See [15, §12] and [22, §4.3].) One calculates δ = 0 when θ ≥ 14 , and δ > 0 when θ < 41 .
Since Theorem A.4 establishes (A.15) with θ = 23 , the method in [15, 22] apparently fails
to produce any rate of convergence with which hφ, |Fk |2 i → 0 as k → ∞ without assuming
much better bounds on L(s, Ad f ) near Re(s) = 1 than we currently know (see the remark
following the proof of Lemma A.2).
If we replace λf (n2 ) with λAd f (n), then there exist absolute constants z ≥ 19 and 0 < ω ′ ≤
1
10
such that if Re(s) ≥ 12 − ω ′, then
Y 1 Y Lp (s, Ad f × φ)
H1 (s) = , H 2 (s) =
L (2s, Ad f )(1 − λf (p2 )p−2s + λφ (p)p−3s )
p>z p p≤z
Lp (2s, Ad f )
EFFECTIVE CORRELATION/DECORRELATION AND WEAK SUBCONVEXITY 51

converge absolutely and are each ≍ 1. If H(s) = H1 (s)H2 (s), then


X λAd f (n)λφ (n) 1
L(s, Ad f × φ) = H(s)L(2s, Ad f ) , Re(s) ≥ − ω ′.

ns 2
p|n=⇒p>z

We still arrive at Theorem A.4 because of the contribution of L(2s, Ad f ) in the ensuing
version of Lemma A.2.

References
[1] Valentin Blomer and Farrell Brumley. On the Ramanujan conjecture over number fields. Ann. of Math.
(2), 174(1):581–605, 2011.
[2] J. Bourgain. Decoupling, exponential sums and the Riemann zeta function. J. Amer. Math. Soc.,
30(1):205–224, 2017.
[3] Petru Constantinescu. Dissipation of correlations of holomorphic cusp forms. arXiv e-prints, page
arXiv:2112.01427, December 2021.
[4] Stephen Gelbart and Hervé Jacquet. A relation between automorphic representations of GL(2) and
GL(3). Ann. Sci. École Norm. Sup. (4), 11(4):471–542, 1978.
[5] Jeffrey Hoffstein and Paul Lockhart. Coefficients of Maass forms and the Siegel zero. Ann. of Math. (2),
140(1):161–181, 1994. With an appendix by Dorian Goldfeld, Jeffrey Hoffstein and Daniel Lieman.
[6] Roman Holowinsky. Sieving for mass equidistribution. Ann. of Math. (2), 172(2):1499–1516, 2010.
[7] Roman Holowinsky and Kannan Soundararajan. Mass equidistribution for Hecke eigenforms. Ann. of
Math. (2), 172(2):1517–1528, 2010.
[8] Yueke Hu. Triple product formula and mass equidistribution on modular curves of level N . Int. Math.
Res. Not. IMRN, (9):2899–2943, 2018.
[9] Bingrong Huang. Effective decorrelation of Hecke eigenforms. arXiv e-prints, page arXiv:2201.12481,
January 2022.
[10] Bingrong Huang and Zhao Xu. Sup-norm bounds for Eisenstein series. Forum Math., 29(6):1355–1369,
2017.
[11] Peter Humphries and Farrell Brumley. Standard zero-free regions for Rankin-Selberg L-functions via
sieve theory. Math. Z., 292(3-4):1105–1122, 2019.
[12] Aleksandar Ivić. On sums of Hecke series in short intervals. J. Théor. Nombres Bordeaux, 13(2):453–468,
2001.
[13] H. Iwaniec and P. Sarnak. Perspectives on the analytic theory of L-functions. Number Special Volume,
Part II, pages 705–741. 2000. GAFA 2000 (Tel Aviv, 1999).
[14] Henryk Iwaniec. Spectral methods of automorphic forms, volume 53 of Graduate Studies in Mathematics.
American Mathematical Society, Providence, RI; Revista Matemática Iberoamericana, Madrid, second
edition, 2002.
[15] Henryk Iwaniec. Notes on quantum unique ergodicity for holomorphic cusp forms (after Holowinsky and
Soundararajan). April 2010. Rutgers lectures.
[16] Henryk Iwaniec and Emmanuel Kowalski. Analytic number theory, volume 53 of American Mathematical
Society Colloquium Publications. American Mathematical Society, Providence, RI, 2004.
[17] H. Jacquet, I. I. Piatetskii-Shapiro, and J. A. Shalika. Rankin-Selberg convolutions. Amer. J. Math.,
105(2):367–464, 1983.
[18] Yujiao Jiang, Guangshi Lü, and Zhiwei Wang. Exponential sums with multiplicative coefficients without
the Ramanujan conjecture. Math. Ann., 379(1-2):589–632, 2021.
[19] Ikuya Kaneko and Jesse Thorner. Highly Uniform Prime Number Theorems. arXiv e-prints, page
arXiv:2203.09515, March 2022.
[20] Henry H. Kim. Functoriality for the exterior square of GL4 and the symmetric fourth of GL2 . J. Amer.
Math. Soc., 16(1):139–183, 2003. With appendix 1 by Dinakar Ramakrishnan and appendix 2 by Kim
and Peter Sarnak.
[21] Emmanuel Kowalski, Philippe Michel, and Jeffrey VanderKam. Rankin-Selberg L-functions in the level
aspect. Duke Mathematical Journal, 114(1):123 – 191, 2002.
52 NAWAPAN WATTANAWANICHKUL (WITH AN APPENDIX BY JESSE THORNER)

[22] Stephen Lester, Kaisa Matomäki, and Maksym Radziwill. Small scale distribution of zeros and mass of
modular forms. J. Eur. Math. Soc. (JEMS), 20(7):1595–1627, 2018.
[23] Xiannan Li. Upper bounds on L-functions at the edge of the critical strip. Int. Math. Res. Not. IMRN,
(4):727–755, 2010.
[24] Elon Lindenstrauss. Invariant measures and arithmetic quantum unique ergodicity. Ann. of Math. (2),
163(1):165–219, 2006.
[25] Wenzhi Luo, Zeév Rudnick, and Peter Sarnak. On Selberg’s Eigenvalue Conjecture, pages 387–401.
Birkhäuser Basel, Basel, 1995.
[26] Wenzhi Luo and Peter Sarnak. Mass equidistribution for Hecke eigenforms. volume 56, pages 874–891.
2003. Dedicated to the memory of Jürgen K. Moser.
[27] W. Müller and B. Speh. Absolute convergence of the spectral side of the Arthur trace formula for GLn .
Geom. Funct. Anal., 14(1):58–93, 2004. With an appendix by E. M. Lapid.
[28] Paul D. Nelson. Equidistribution of cusp forms in the level aspect. Duke Math. J., 160(3):467–501, 2011.
[29] Paul D. Nelson, Ameya Pitale, and Abhishek Saha. Bounds for Rankin-Selberg integrals and quantum
unique ergodicity for powerful levels. J. Amer. Math. Soc., 27(1):147–191, 2014.
[30] Ken Ono. The web of modularity: arithmetic of the coefficients of modular forms and q-series, volume 102
of CBMS Regional Conference Series in Mathematics. Conference Board of the Mathematical Sciences,
Washington, DC; by the American Mathematical Society, Providence, RI, 2004.
[31] Zeév Rudnick. On the asymptotic distribution of zeros of modular forms. Int. Math. Res. Not., (34):2059–
2074, 2005.
[32] Zeév Rudnick and Peter Sarnak. The behaviour of eigenstates of arithmetic hyperbolic manifolds. Com-
munications in Mathematical Physics, 161(1):195 – 213, 1994.
[33] Kannan Soundararajan. Quantum unique ergodicity for SL2 (Z)\H. Ann. of Math. (2), 172(2):1529–1538,
2010.
[34] Kannan Soundararajan. Weak subconvexity for central values of L-functions. Ann. of Math. (2),
172(2):1469–1498, 2010.
[35] Kannan Soundararajan and Jesse Thorner. Weak subconvexity without a Ramanujan hypothesis. Duke
Math. J., 168(7):1231–1268, 2019. With an appendix by Farrell Brumley.

(Nawapan Wattanawanichkul) 1409 West Green Street, University of Illinois Urbana-Champaign,


Urbana, IL 61801, USA
Email address: nawapan2@illinois.edu

(Jesse Thorner) 1409 West Green Street, University of Illinois Urbana-Champaign, Urbana,
IL 61801, USA
Email address: jesse.thorner@gmail.com

You might also like