.

SUBCONVEXITY BOUND FOR GL(3) × GL(2) L-FUNCTIONS IN

GL(2) SPECTRAL ASPECT

SUMIT KUMAR

Abstract
Let π be a Hecke-Maass cusp form for SL(3, Z) and f be a holomorphic cusp form
with weight k or Hecke-Maass cusp form corresponding to the Laplacian eigenvalue
1/4 + k 2 , k ≥ 1 for SL(2, Z). In this paper, we prove the following subconvexity
arXiv:2007.05043v1 [math.NT] 9 Jul 2020

bound:
 
1 3 1
L , π × f ≪π,ǫ k 2 − 51 +ǫ ,
2
for the central values L(1/2, π × f ) of the Rankin-Selberg L-function of π and f .
Using the same method, by taking π to be the Eisenstein series, we also obtain the
following subconvexity bound:
 
1 1 1
L , f ≪ǫ k 2 − 153 +ǫ .
2

1. Introduction
A general automorphic L-function of degree d ≥ 1 is a Dirichlet series with an Euler
product of degree d. It satisfies some “nice” properties. It has a meromorphic con-
tinuation in the complex plane C and its completed L-function satisfies a functional
equation relating its values at s to 1 −s. One may apply Phragmén-Lindelöf principle
together with the functional equation to get an upper bound ((Conductor)1/4+ǫ ) for
the L-function on the critical line ℜs = 1/2. The resulting bound is usually referred
to as the convexity bound (or the trivial bound) for the L-function. It is conjectured
(Lindelöf Hypothesis) that the exponent 1/4 can be reduced to 0. While Lindelöf
Hypothesis is still out of reach, breaking the convexity bounds for L-functions is an
interesting problem.

For degree one L-functions, such subconvexity estimates are known due to Weyl
([22]) in the t-aspect and Burgess [4] in the q-aspect. For degree two L-functions, it
was achieved by Good [6] in the t-aspect and by Iwaniec [7] in the spectral aspect.
In the case of degree three L-functions, first subconvexity estimates, in the t-aspect,
were achieved by Xiaoqing Li in the groundbreaking work [15]. She dealt with the
self-dual forms for SL(3, Z). We will discuss this paper in detail later. For any degree
three L-functions, R. Munshi introduced an ingenious method to obtain subconvexity
bounds (in the t and twist aspect) in a series of papers [16], [17]. In the GL(3) spec-
tral aspect, such subconvexity bounds were obtained by V. Blomer and J. Buttcane
[3].

Date: July 13, 2020.

2010 Mathematics Subject Classification. Primary 11F66, 11M41; Secondary 11F55.
Key words and phrases. Maass forms, subconvexity, Rankin-Selberg L-functions.
1
2 SUMIT KUMAR

For higher degree L-functions, the subconvexity problem becomes more challenging
and out of reach, so it is still open except for a few particular cases. For Rankin-
Selberg L-functions on GL(2) × GL(2), subconvexity is known due to Sarnak [20],
Michel and Venkatesh [11], Lau, Liu and Ye [13] and a few others (see [11]). Impressive
subconvexity estimates for triple L-functions on GL(2) were obtained by Bernstein
and Reznikov [1]. We will now discuss a few known results for the degree six Rankin-
Selberg L-functions on GL(3) × GL(2). To start with, let π be a normalized Hecke-
Maass cusp form of type (ν1 , ν2 ) for SL(3, Z). Let f be a holomorphic cusp form
with weight k or Maass Hecke cusp form with the Laplace eigenvalue 1/4 + k 2 for
SL(2, Z). The associated Rankin-Selberg L-series is given by
X X λπ (n, r) λf (n)
(1) L(s, π × f ) = 2 )s
, ℜ(s) > 1.
n,r≥1
(nr

In her pioneer work [15], Xiaoqing Li considered L(s, π×f ), and she obtained subcon-
vexity bounds for GL(3) × GL(2) L-functions in the GL(2) spectral aspect together
with the t-aspect subconvexity bounds for self-dual GL(3) forms in one shot. Her
main theorem was the following:
Theorem (X. Li.). Let π be a fixed self-dual Hecke-Maass cusp form for SL(3, Z) and
uj be an orthonormal basis of even Hecke-Maass cusp form for SL(2, Z) corresponding
3
to the Laplacian eigenvalue 1/4 + t2j with tj ≥ 0; then for ǫ > 0, T large and T 8 +ǫ ≤
1
M ≤ T 2 , we have
X′ − (tj −T )2  1  Z ∞ (t −T )2   2
1 − j 1
− it, π  dt ≪ǫ,π T 1+ǫ M,
 
e M2 L , π × uj + e M 2 L
j
2 4π −∞ 2

where ′ means summing over the orthonormal basis of even Hecke-Maass cusp forms.
As a corollary, she obtained
 
1 3 1
L , π × uj ≪ǫ,π (1 + |tj |) 2 − 8 +ǫ .
2
She adapted
Conrey-Iwaniec’s moment method to prove her theorem. The fact that
“L 12 , π × uj ≥ 0” is playing a crucial role in her approach to obtain the subconvex-
ity bounds. Unfortunately, the above fact is not valid for non-self-dual GL(3) forms;
that is why she only dealt with self-dual forms. Recently, R. Munshi [18] obtained
the subconvexity bounds for L(s, π × f ) in the t-aspect. He proved the following
bound:
 
1 3 1
L + it, π × f ≪ǫ,f,π (1 + |t|) 2 − 51 +ǫ .
2
His method is insensitive to the self-duality of GL(3) forms; hence he obtained the
above bound for any GL(3) forms. Using a similar method, P. Sharma [21] and
the author, K. Mallesham and Saurabh Singh [9] proved the subconvexity bounds in
the twist and the GL(3) spectral aspect, respectively. It is natural to ask a similar
question in other aspects. As suggested by R. Munshi, we consider the case of GL(2)
spectral aspect. In this paper, we establish subconvexity bounds for the Rankin-
Selberg L-functions on GL(3) × GL(2) in the GL(2) spectral aspect. Our main
theorem is the following:
 SUBCONVEXITY BOUND FOR GL(3) × GL(2) L-FUNCTIONS IN GL(2) SPECTRAL ASPECT 3

Theorem 1. Let π be a fixed Hecke-Maass cusp form for SL(3, Z) and f be a holo-
morphic cusp form with weight k or Hecke-Maass cusp form corresponding to the
Laplacian eigenvalue 1/4 + k 2 with k ≥ 1 for SL(2, Z). Then for ǫ > 0, we have
 
1 3 1
L , π × f ≪π,ǫ k 2 − 51 +ǫ .
2
Remarks. 1. We follow Munshi’s Method [18] to prove the above theorem. As we
mentioned before that his method is insensitive to the self-duality of GL(3) forms, we
also obtain our result for any GL(3) form. Thus we generalize X. Li’s main result
[15]. Although our bound is weaker than her’s, it yields a subconvexity.
2. We mention here that our arguments in the proof work for both Maass and holomor-
phic forms f . For the exposition of the method, we will give details for holomorphic
forms.
If we take π to be the minimal Eisenstein series for SL(3, Z) in (1), we note that
X X λmin(n, r) λf (n)
(2) L (1/2, π × f ) = 2 1/2
= L3 (1/2, f ) ,
n,r≥1
(nr )
where X
λmin (n, r) = µ(d)d3 (r/d)d3(n/d).
d|(n,r)
Our method of proof of the Theorem 1 also fits well for the above L-function. Hence
we also obtain the following result:
Theorem 2. Let f be a holomorphic cusp form with weight k or Hecke-Maass cusp
form corresponding to the Laplacian eigenvalue 1/4 + k 2 with k ≥ 1 for SL(2, Z).
Then for ǫ > 0, we have
 
1 1 1
L , f ≪ǫ k 2 − 153 +ǫ .
2
We end the introduction by mentioning a few facts about the method. As a first
step, we use the ‘conductor lowering’ trick as a device for the separation of oscilla-
tions using the Duke, Friedlander, and Iwaniec’s Delta method. We now apply GL(3)
and GL(2) Voronoi summation formulae to shorten the length of the corresponding
sums. Stationary phase analysis of the integral transforms also plays a vital role in
the proof. Finally, an application of the Cauchy inequality coupled with the Poisson
summation formula gives us the desired result.

Notations. Throughout the paper, e(x) means e2πix and negligibly small means
O(k −A ) for any A > 0. In particular, we will take A = 2020. The notation α ≪ A
will mean that for any ǫ > 0, there is a constant c such that |α| ≤ cAk ǫ . By α ≍ A,
we mean that k −ǫ A ≤ α ≤ k ǫ A, also α ∼ A means A ≤ α < 2A. For absolute explicit
constant, we will write c or ci in the whole paper.

Acknowledgement. The author is grateful to Prof. Ritabrata Munshi for sharing

his beautiful ideas, explaining his ingenious method in full detail, and his kind support
throughout the work. He would also like to thank Prof. Satadal Ganguly, Saurabh
Kumar Singh and Kummari Mallesham for their encouragement and constant support
and Stat-Math Unit, Indian Statistical Institute, Kolkata for the excellent research
environment.
4 SUMIT KUMAR

2. Preliminaries
In this section, we will recall some known results which we need in the proof.
2.1. Holomorphic forms on GL(2). Let f be a holomorphic Hecke eigenform of
weight k for the full modular group SL(2, Z). The Fourier expansion of f at ∞ is
∞
X
f (z) = λf (n) n(k−1)/2 e(nz),
n=1
for z ∈ H. We have a well-known Deligne’s bound for the Fourier coefficients which
says that
(3) |λf (n)| ≤ d(n),
for n ≥ 1, where d(n) is the divisor function. We now state the Voronoi summation
formula for f in the following lemma.
Lemma 1. Let λf (n) be as above and g be a smooth, compactly supported function
on (0, ∞). Let a, q ∈ Z with (a, q) = 1. Then we have
∞ ∞
2πik X
   
X an dn
λf (n) e g(n) = λf (n) e − h(n),
n=1
q q n=1 q
where ad ≡ 1(mod q) and
∞ √ 
4π xy
Z 
h(y) = g(x) Jk−1 dx.
0 q
Proof. Proof can be found in the book of Iwaniec-Kowalski [8]. 
2.2. Automorphic forms on GL(3). In this subsection, we will recall some back-
ground on the Maass forms for GL(3). This subsection, except for the notations, is
taken from [15]. Let π be a Hecke-Maass cusp form of type (ν1 , ν2 ) for SL(3, Z). Let
λπ (n, r) denote the normalized Fourier coefficients of π. Let
α1 = −ν1 − 2ν2 + 1, α2 = −ν1 + ν2 , α3 = 2ν1 + ν2 − 1
be the Langlands parameters for π (see Goldfeld [5] for moreRdetails). Let g be a
∞
compactly supported smooth function on (0, ∞) and g̃(s) = 0 g(x)xs−1 dx be its
Mellin transform. For ℓ = 0 and 1, we define
3 3
π −3s− 2 Y Γ 1+s+α i +ℓ


2
γℓ (s) := −s−αi +ℓ

.
2 i=1
Γ 2

Set γ± (s) = γ0 (s) ∓ γ1 (s) and let

1
Z
G± (y) = y −s γ± (s) g̃(−s) ds,
2πi (σ)

where σ > −1 + max{−ℜ(α1 ), −ℜ(α2 ), −ℜ(α3 )}. With the aid of the above termi-
nology, we now state the GL(3) Voronoi summation formula in the following lemma:
Lemma 2. Let g(x) and λπ (n, r) be as above. Let a, ā, q ∈ Z with q 6= 0, (a, q) = 1,
and aā ≡ 1(mod q). Then we have
∞   ∞  2 
X an XX X λπ (n1 , n2 ) n1 n2
λπ (n, r)e g(n) = q S (rā, ±n2 ; qr/n1 ) G± ,
n=1
q ± n =1
n1 n2 q3r
n1 |qr 2
 SUBCONVEXITY BOUND FOR GL(3) × GL(2) L-FUNCTIONS IN GL(2) SPECTRAL ASPECT 5

where S(a, b; q) is the Kloosterman sum which is defined as follows:

X⋆  ax + bx̄ 
S(a, b; q) = e .
x mod q
q

Proof. See [12] for the proof. 

To apply the Lemma 2 in our setup, we need to extract the integral transform’s
oscillations G± (y). To this end, we state the following lemma.
Lemma 3. Let G± (x) be as above, with support in the interval [X, 2X]. Then for
any fixed integer K ≥ 1 and xX ≫ 1, we have
Z ∞ K
cj (±)e 3(xy)1/3 + dj (±)e −3(xy)1/3


 
X −K+2
G± (x) = x g(y) dy +O (xX) 3 ,
0 j=1 (xy)j/3
where cj (±) and dj (±) are absolute constants depending on αi , for i = 1, 2, 3.
Proof. See Lemma 6.1 of [14]. 
The following lemma, which gives the Ramanujan conjecture on average, is also
well-known.
Lemma 4. We have XX
|λπ (n1 , n2 )|2 ≪ x1+ǫ ,
n21 n2 ≤x

where the implied constant depends on the form π and ǫ.

Proof. For the proof, we refer to Goldfeld’s Book [5]. 

2.3. The Delta method. Let δ : Z → {0, 1} be defined by

(
1 if n = 0;
δ(n) =
0 otherwise.
The above delta symbol can be used to separate the oscillations involved in a sum.
Further, we seek a Fourier expansion of δ(n). We mention here an expansion for δ(n)
which is due to Duke, Friedlander and Iwaniec. Let L ≥ 1 be a large number. For
n ∈ [−2L, 2L], we have
 Z  
1 X 1 X⋆ na nx
δ(n) = e g(q, x) e dx,
Q 1≤q≤Q q a mod q q R qQ

where Q = 2L1/2 . The ⋆ on the sum indicates that the sum over a is restricted by
the condition (a, q) = 1. The function g is the only part in the above formula which
is not explicitly given. Nevertheless, we only need the following two properties of g
in our analysis.

 B !
1 q
(4) g(q, x) = 1 + h(q, x), with h(q, x) = O + |x| ,
qQ Q
g(q, x) ≪ |x|−B
6 SUMIT KUMAR

for any B > 1. Using the second property of g(q, x) we observe that the effective
range of the integration in (5) is [−Lǫ , Lǫ ]. We record the above observations in the
following lemma.
Lemma 5. Let δ be as above and g be a function satisfying (4). Let L ≥ 1 be a large
parameter. Then, for n ∈ [−2L, 2L], we have
 Z  
1 X 1 X⋆ na nx
δ(n) = e W (x)g(q, x) e dx + O(L−2020 ),
Q 1≤q≤Q q a mod q q R qQ

where Q = 2L1/2 and W (x) is a smooth bump function supported in [−2Lǫ , 2Lǫ ], with
W (x) = 1 for x ∈ [−Lǫ , Lǫ ] and W (j) ≪j 1.
Proof. For the proof, we refer to chapter 20 of the book [8]. 
2.4. Bessel function. In this subsection, we will recall some familiar expansions and
formulae of J-Bessel functions. We start with the following integral representation.
Lemma 6. Let k ≥ 2 be an integer. Let Jk−1 (x) be the Bessel function of the first
kind of order k − 1. Then for any x ∈ R, we have
Z π  
1 (k − 1)τ − x sin τ
Jk−1 (x) = e dτ.
2π −π 2π
Proof. It is a standerd result. Proof can be found in any book on Bessel functions. 
In our analysis of integrals, we need an expansion of Jk−1 (x) for large values of
k and x. In the following lemma, we state an asymptotic expansion of the J-bessel
function which is known due to Langer.
Lemma 7. Let x ≥ k 1+ǫ . Let Pj denotes any polynomial of degree of j whose
coefficients are bounded functions of k − 1 and log(x/(k − 1)). Then, as (k − 1)(w −
tan−1 w) tends to ∞, we have
1/2 "  ∞ 1

#
P

2 π X j w−tan w
−1
Jk−1 (x) = cos (k − 1)(w − tan−1 w) −
π(k − 1)w 4 j=0 (k − 1)j
1/2 "  ∞ 1

#
P

2 π X j w−tan −1 w
+ sin (k − 1)(w − tan−1 w) − ,
π(k − 1)w 4 j=1 (k − 1)j
 1/2
x2
where w = (k−1)2
−1 .

Proof. See Langer’s paper [10]. 

We can truncate the above series at any stage. Thus the above expansion essentially
looks like
Corollary 1. Under the same assumptions as in the Lemma 7, as k → ∞, we have
 
(k−1)(w−tan−1 w) 1


Xe ±
2019
2π
P j w−tan w
−1

1

Jk−1 (x) = +O .
j=0
w 1/2 (k − 1)j+1/2 k 2020

Proof. Proof follows directly from Lemma 7. 

For x ≪ k 1−ǫ , we have the following lemma.
 SUBCONVEXITY BOUND FOR GL(3) × GL(2) L-FUNCTIONS IN GL(2) SPECTRAL ASPECT 7

Lemma 8. Let x ≤ k 1−ǫ . Then, as k → ∞, we have

Jk−1 (x) ≪ e−(k−1) .
Proof. See Lemma 4.2-4.3 of [19]. 
2.5. Stationary phase method. In this subsection, we will recall some facts about
the exponential integrals of the following form:
Z b
I= g(x)e(f (x))dx,
a

where f and g are smooth real valued functions on [a, b].

Lemma 9. Let I, f and g be as above. Then, for r ≥ 1, we have
Var g
I≪ ,
min |f (r) (x)|1/r
where Var is the total variation of g on [a, b]. Moreover, let f ′ (x) ≥ B and f (j) (x) ≪
B 1+ǫ for j ≥ 2 together with Supp(g) ⊂ (a, b) and g (j)(x) ≪a,b,j 1. Then we have
I ≪a,b,j,ǫ B −j+ǫ.
Proof. First part of the lemma is standard. For the second part, we use integration
by parts. See [16] for more details. 
We will use the above lemma later at several places whenever the phase function
does not have any stationary point. In case there is a single stationary point, we use
the following stationary phase expansion.
Lemma 10. Let I, f and g be as above. Let 0 < δ < 1/10, X, Y , U, Q > 0,
Z := Q + X + Y + b − a + 1, and assume that
δ
3δ QZ 2
Y ≥Z , b−a≥U ≥ √ .
Y
Further, assume that g satisfies
X
g (j) (x) ≪j for j = 0, 1, 2, . . . .
Uj
Suppose that there exists a unique x0 ∈ [a, b] such that f ′ (x0 ) = 0, and the function
f satisfies
Y Y
f ′′ (x) ≫ 2 , f (j) (x) ≪j j for j = 1, 2, 3, . . . .
Q Q
Then we have
−1
3δ A
eif (x0 ) X
pn (x0 ) + OA,δ Z −A ,


I=p
f ′′ (x0 ) n=0
√ n
2πeπi/4

i
pn (x0 ) = G(2n) (x0 ),
n! 2f ′′ (x0 )
where A > 0 is arbitrary, and
1
G(x) = g(x)eiF (x) , F (x) = f (x) − f (x0 ) − f ′′ (x0 )(x − x0 )2 .
2
8 SUMIT KUMAR

Furthermore, each pn is a rational function in f ′ , f ′′ , . . . , satisfying

  2 −n !
j

d 1 1 U Y n

j pn (x0 ) ≪j,n X j
+ j 2
+ Y −3 .
dx0 U Q Q

Proof. See [2] for the proof. 

3. The set-up
Let π and f be as defined in Theorem 1. Let λπ (n, r) denote the normalized Fourier
coefficients of the form π (see Chapter 6 of [5]) and let λf (n) denote the normalised
Fourier coefficients of the form f (see [8]). We consider the
Rankin-Selberg L-series
attacted to π and f , as given in (1). To study L 21 , π × f , we first express it as a
weighted Dirichlet series.
Lemma 11. For any θ : 0 < θ < 3/2 and k → ∞, we have
 
1 |Sr (N)|
(5) L , π × f ≪π,f,ǫ k ǫ sup sup 1/2
+ k (3−θ)/2 ,
2 r≤k θ k 3−θ
≤N ≤ k 3+ǫ N
r2 r2

where Sr (N) is a sum of the form

∞
X n
(6) Sr (N) := λπ (n, r)λf (n)V ,
n=1
N
(j)
Rfor some smooth function V supported in [1, 2] and satisfying V (x) ≪j 1 with
V (x)dx = 1.
Proof. On applying the approximate functional equation (Theorem 5.3 from [8]) to
L(s, π × f ), we get that
  XX  2
1 λπ (n, r) λf (n) nr
L ,π × f = 2 1/2
V 3+ǫ
+ ǫ(1/2, π × f )
2 n,r≥1
(nr ) k
X X λπ̄ (n, r) λf¯(n)  nr 2 
× 2 )1/2
V′ ,
n,r≥1
(nr k 3+ǫ

where ǫ(1/2, π × f ) = ǫ(π × f ) is the root number of absolute value 1 and V (x), V ′ (x)
are smooth functions which are negligibly small if x ≫ k ǫ (see [15] for more details).
We proceed with the first sum as the calculations for the dual sum are same. Hence
we have

1
 X X λπ (n, r) λf (n)  nr 2 
L ,π × f ≪ 2 1/2
V 3+ǫ
+ k −2020 .
2 (nr ) k
nr 2 ≪k 3+ǫ
X 1 X λπ (n, r) λf (n)  nr 2 
= 1/2
V 3+ǫ
+ k −2020 .
(3+ǫ)/2
r 3+ǫ 2
n k
r≤k n≤p /r

We split the above sum as follows:

X X X X X X X X
= + + ,
r≤k (3+ǫ)/2 n≤p3+ǫ /r 2 r≤k θ k3−θ 3+ǫ
≤n≤ k 2 r≤k θ n≤ k3−θ r≥k θ n≤ k3+ǫ
r2 r r 2 r 2
 SUBCONVEXITY BOUND FOR GL(3) × GL(2) L-FUNCTIONS IN GL(2) SPECTRAL ASPECT 9

where θ > 0 is a constant which will be choosen later optimally. Note that last two
sums are bounded by k (3−θ)/2 . In fact,
X 1 X λπ (n, r) λf (n)  nr 2  sX X s
X 1
1/2
V 3+ǫ
≪ |λ π (n, r)| 2
2
≪ k (3−θ)/2 .
θ
r k3+ǫ
n k 2 3+ǫ θ
r
r≥k n≤ nr ≪k r≥k
r2

Here we have used Cauchy’s inequality coupled with Lemma 4. Similarly we can
treat second sum. Hence, we arrive at
   2
1 X1 X λπ (n, r) λf (n) nr
L ,π ×f ≪ 1/2
V 3+ǫ
+ k (3−θ)/2 .
2 θ
r 3−θ 3+ǫ
n k
r≤k k
≤n≤ k
r2 r2

Taking a smooth dyadic subdivision (see [17] for more details), we get that
   2  
1 X1 X X λπ (n, r) λf (n) nr n
L ,π ×f ≪ 1/2
V 3+ǫ
U + k (3−θ)/2
2 θ
r 3−θ 3+ǫ
n k R
r≤k (U,R) k
≤n≤ k
r2 r2
X1 |Sr (N)|
≪ sup + k (3−θ)/2 ,
θ
r k3−θ
≤N ≤ k
3+ǫ N 1/2
r≤k r2 r2

where
∞
nr 2
X   n
Sr (N) = λπ (n, r) λf (n)V U ,
n=1
k 3+ǫ N
and N = 2α/2 with α ∈ [−1, ∞)∩Z. In the last inequality, we used partial summation
 2 
and the fact that there are only k ǫ -many R. Note that Vr,N (x) := U(x)V Nk3+ǫ
r x
is
j
supported on [1, 2] and satisfies Vr,N ≪ 1. By the abuse of notations, we will use V
in place of Vr,N hereafter. Finally, taking supremum over r, we get the lemma. 
1
Remark. Upon estimating (5) trivially, we get L 12 , π × f ≪ N 2 ≪ k 3/2+ǫ . Hence,


to establish subconvexity, we need to show some cancellations in the sum Sr (N). In
the rest of paper, we will study Sr (N).
3.1. An application of delta method. There are two oscillatory factors, namely,
λπ (n, r) and λf (n), in the sum Sr (N) (6). We will use the delta method of Duke,
Friedlander and Iwaniec to separate these oscillations. As mentioned in Munshi [18],
just this does not suffice, we need to lower the conductor as well. For this purpose,
we introduce an extra t-integral. In fact, we rewrite the sum Sr (N) in (6) as
  X∞X
1 t  n it  n   m 
Z
Sr (N) = V λπ (n, r)λ(m) V U dt
T R T n,m=1
m N N
n=m
  X∞X
1 t  n it  n   m 
Z
(7) = V δ(n − m)λπ (n, r)λf (m) V U dt.
T R T n,m=1
m N N
where k ǫ < T < k 1−ǫ is a parameter which will be chosen later optimally, and U is a
smooth function supported in [1/2, 5/2] with U(x) = 1 for x ∈ [1, 2], and U (j) ≪j 1.
Consider the t-integral
  
1 t m it
Z
V dt.
T R T n
10 SUMIT KUMAR

By repeated integration by parts, we observe that the above integral is negligibly

small if |n − m| ≫ k ǫ N/T . We now apply Lemma 5 to (7) with
r
ǫN ǫ N
L=k , Q=k .
T T
Thus Sr (N) transforms as
  X
1 t g(q, x) X⋆
Z Z
(8) Sr (N) = W (x) V
QT R R T 1≤q≤Q q
a mod q
∞    
X an nx  n
× λπ (n, r)e e nit V
n=1
q qQ N
∞      
X
−it −am −mx m
× λf (m)m e e U dtdx + O(k −2020 ).
m=1
q qQ N
3.2. Sketch of the proof. We end this section with a sketch of the proof of Theorem
1. For clarity, we will consider the generic case, i.e., N = k 3 , r = 1 and q ≍ Q =
k 3/2 /T 1/2 . Thus Sr (N) in (8) looks like
Z 2T X X X   X  
⋆
it an −it −am
λπ (n, 1)n e λf (m) m e dt.
T q∼Q a mod q n∼N
q m∼N
q

Note that, at this stage, we trivially have Sr (N) ≪ N 2 . Hence, to get subconvexity,
we need to save N plus a ‘little more’. In the next step, we dualize n-sum.
GL(3) Voronoi. We apply GL(3) Voronoi summation formula to the n-sum. It
converts the n-sum into another n⋆ -sum with the dual length n⋆ ∼ T 3/2 N 1/2 and we
save N 1/4 /T 3/4 in this step.
GL(2) Voronoi. Now we dualize m-sum using GL(2) Voronoi formula. Here, we
save (NT )1/2 /k and the dual length becomes ⋆ 2
√ m ∼ k /T .
a-sum and t-integral . We also save Q from the a-sum and T 1/2 from the t
integral. Thus total savings so far is
N 1/4 N 1/2 T 1/2 p N
3/4
× × Q × T 1/2 = .
T k k
Hence it remains to save k and a little more in the transformed sum
X X X
λπ (1, n) λf (m)C J
q∼Q n⋆ ∼T 3/2 N 1/2 m⋆ ∼k 2 /T

where J is an integral transform which oscillates like niT with respect to n, and the
character sum is given by
   
X∗ ām m̄n
C= S (ā, n; q) e qe − .
q q
a mod q

Cauchy and Poisson. Next applying the Cauchy’s inequality, we arrive at

  
X X X m̄n 2
λf (m) e − I .

q

⋆ 3/2 1/2 q∼Q ⋆ 2 m ∼k /T
n ∼T N

Opening the absolute value square, we apply the Poisson summation formula on the
n⋆ -sum. In the zero-frequency (n⋆ = 0), we save the whole length, i.e., k 2 Q2 /T
which suffices if k 2 Q2 /T > k 2 which implies that T < k. On the other hand, we
 SUBCONVEXITY BOUND FOR GL(3) × GL(2) L-FUNCTIONS IN GL(2) SPECTRAL ASPECT11

save T 3/2 N 1/2 /T 1/2 in the non-zero frequency (n⋆ 6= 0), which is good enough if
T 3/2 N 1/2 /T 1/2 > k 2 which boils down to T > k 1/2 . Hence for the whole method to
work, we must ensure that k 1/2 < T < k. In fact, we will choose T = k 41/51 . Thus
we get the theorem.

4. APPLICATIONS OF SUMMATION FORMULAE

4.1. GL(3) Voronoi. We now apply GL(3)-Voronoi summation formula to the n-
sum in (8). To this end, we will use Lemma 2 coupled with Lemma 3. In the present
n2 n
setup, we have g(n) = e (nx/qQ) nit V (n/N) and X = N. Thus for q13 r2 N ≫ k ǫ and
K large enough, n-sum (upto a negligibly small error term) transforms into
∞
XX X λπ (n1 , n2 )
q n1 S (rā, ±n2 ; qr/n1 )
± n1 |qr n2 =1
n 1 n2
  1/3    1/3 
n21 n2 z n21 n2 z
K cj (±)e 3 + dj (±)e −3 q3 r
n21 n2 ∞ q3 r
Z X
× 3 g(z)  2 j/3 dz.
q r 0 j=1
n1 n2 z
q3 r

We will work the terms corresponding to j = 1, as other terms gives us even better
estimates. Thus we essentially arrive at
∞ ∞
N 2/3+it X X 1/3 X λπ (n1 , n2 )
 
X an
λπ (n, r)e g(n) = n1 1/3
S(rā, ±n2 ; qr/n1)
n=1
q qr 2/3 ± n2 =1 n2
n1 |qr
Z ∞
Nxz 3(Nn21 n2 z)1/3
 
it
× V (z)z e ± 1/3
dz + O(k −2020 ).
0 qQ qr
Next applying integration by parts repeatedly, we see that the z-integral is negligibly
small if
( )
3
(qT ) r 3/2 1/2 3
(9) n21 n2 ≫ k ǫ max , T N rx =: N0 .
N
n2 n
Now consider the range q13 r2 N ≪ k ǫ . Analysis for this case is exactly similar to that
of Munshi’s paper (see page 8, [18]). Hence in this case, we get
∞ √
q 2 Qr
   
X an nx it
n
(10) λπ (n, r)e e n V ≪ √ .
n=1
q qQ N N
Finally we summarize the above discussion in the following lemma.
Lemma 12. Let g(n) = e (nx/qQ) nit V (n/N) and N0 be as defined in (9). Then we
have
∞
N 2/3+it X X 1/3 X λπ (n1 , n2 )
 
X an
λπ (n, r)e g(n) = n1 1/3
S(rā, ±n2 ; qr/n1 )
n=1
q qr 2/3 ± 2 n2
n1 |qr n2 n1 ≪N0
Z ∞  2√ 
Nxz 3(Nn21 n2 z)1/3
 
it q Qr
× V (z)z e ± 1/3
dz + O √ .
0 qQ qr N
We will further proceed with the main term and get back to error term at the end.
12 SUMIT KUMAR

4.2. GL(2) voronoi. We will now consider the sum over m in Sr (N) in (8). We will
dualize it using GL(2) voronoi summation formula. On applying Lemma 1 to the
m-sum, we see that it transforms into
∞ Z ∞  √ 
2πik X 4π my
    
ām −it −xy y
λf (m)e y e U Jk−1 dy
q m=1 q 0 qQ N q
∞ Z ∞  √
2πik N 1−it X
   
ām −it −Nxy 4π mNy
= λf (m)e U(y)y e Jk−1 dy.
q m=1
q 0 qQ q
√
4π mN y
We observe that if q
≤ k 1−ǫ , then using Lemma 8, we have
 √ 
4π mNy
Jk−1 < e−(k−1) ,
q
√
and hence the m-sum is negligibly small. On the other hand, if 4π qmN y ≥ k 1+ǫ , we
apply Langer’s formula to the J- Bessel function. On applying Corollary (1) to Jk−1,
the above integral looks like
 −1

2019 e ± (k−1)(w−tan w) P 1
Z ∞  X
 
−it −Nxy 2π j w−tan−1 w 1
U(y)y e 1/2 (k − 1)j+1/2
dy + O 2020
.
0 qQ j=0
w k

Consider the terms corresponding to j = 0, i.e.,

Z ∞
(k − 1)(w − tan−1 w)
   
1 −it −1/2 −Nxy
√ U(y)P0 y w e e ± dy.
k−1 0 qQ 2π
On applying repeated integration by parts, we see that the integral is negligibly small
if
 2 2 
ǫ T q 2
m ≥ k max ,Tx .
N
2 2
But, initially, we had started with the condition m ≥ k ǫ q (k−1)
N
. Hence the effective
range of m is given by
2 2
(k − 1)2 q 2
 
−ǫ q (k − 1) ǫ 2
(11) M := k ≤ m ≤ k max , T x =: M0 .
N N
Similarly we can deal with the terms corresponding to other j’s. Lastly, we end this
subsection by summarizing the above discussion in the following lemma.
Lemma 13. Let M and M0 be as defined in (11) and ψ(m) = m−it e (−mx/qQ) U (m/N).
Then we have
∞
2πik N 1−it X
   
X −am ām
λf (m)e ψ(m) = λf (m)e h(m) + O(k −2020 ),
m=1
q q M ≤m≤M
q
0

where h(m) is the integral transform

Z ∞    √ 
−it −Nxy 4π mNy
h(m) = U(y)y e Jk−1 dy.
0 qQ q
 SUBCONVEXITY BOUND FOR GL(3) × GL(2) L-FUNCTIONS IN GL(2) SPECTRAL ASPECT13

5. SIMPLIFYING THE FOUR-FOLD INTEGRAL

On applying Lemma 12 and Lemma 13 to (8), we see that Sr (N) has the following
expression:
2πik N 5/3 X 1 X⋆ X X 1/3 X λπ (n1 , n2 )
(12) n1 1/3
S(rā, ±n2 ; qr/n1)
QT r 2/3 1≤q≤Q q 3a mod q ± n2
n1 |qr n2 n21 ≪N0
 Z  Z ∞ Z ∞
ām t
X Z
−it
× λf (m)e W (x)g(q, x) V U(y)y V (z)z it
m≪M0
q R R T 0 0
 √
Nx(z − y) 3(Nn21 n2 z)1/3
  
4π mNy
×e ± Jk−1 dzdydtdx.
qQ qr 1/3 q
In this section, we will simplify above integrals. At first, we analyze the integral over
x, which is given as
 
Nx(z − y)
Z
W (x)g(q, x)e dx.
R qQ
Using the property g(q, x) = 1 + h(q, x) from (4), we split it as a sum of the following
two integrals:
   
Nx(z − y) Nx(z − y)
Z Z
W (x)e dx + W (x)h(q, x)e dx.
R qQ R qQ
On applying integration by parts repeatedly in the first x-integral, we see that it is
negligibly small if |z − y| ≫ k ǫ q/QT . Next, writing z = y + u with u ≪ k ǫ q/QT , the
above four-fold integral in (12) transforms as
 Z  Z ∞
t Nxu
Z Z
(13) W (x) V e U(y)V (y + u)(1 + u/y)it
R R T R qQ 0
 2 1/3
  √ 
3(Nn1 n2 (y + u)) 4π mNy
×e ± Jk−1 dydudtdx.
qr 1/3 q
Note that (1 + u/y)it = eit log(1+u/y) is not oscillating as a function of y, i.e.,

(j)
(1 + u/y)it (y) ≪ k ǫj .
So, there is no harm in combining it with the weight function U. We will bound x,
t and u-integral trivially later, also we will see that y-integral bound is independent
of other variables. Thus the four-fold integral in (13) will reduce to
q
× T × I± (m, n21 n2 , q),
QT
where
∞  √
3(Nn21 n2 (y + u))1/3
  
4π mNy
Z
(14) I± (m, n21 n2 , q) := Uu,t (y)e ± Jk−1 dy,
0 qr 1/3 q
with Uu,t (y) = U(y)V (y + u)(1 + u/y)it. Next we consider the second integral
 
Nx(z − y)
Z
W (x)h(q, x)e dx.
R qQ
14 SUMIT KUMAR

We don’t know much about the weight function h except its size, i.e., h ≪ 1/qQ.
In this case, we analyze the t-integral to get a restriction on z − y. Consider the
t-integral
   it
t z
Z
V dt.
R T y
Upon applying integration by parts, we see that the t-integral is negligibly small
ǫ
unless |z − y| ≪ kT . Thus, like above, four-fold integral in this case will reduce to
1 1
× × T × I± (m, n21 n2 , q).
qQ T
Here we have used h(q, x) ≪ 1/qQ. We observe that we obtain a much better bound
in this case (We are saving q 2 more then that of the first case). So we will continue our
further analysis with (13). Lastly, we sum up the above arguments in the following
lemma for future reference.
Lemma 14. We have
2πik N 5/3 X 1 X⋆
Sr (N) =
QT r 2/3 1≤q≤Q q 3
a mod q
XX 1/3
X λπ (n1 , n2 )
× n1 1/3
S(rā, ±n2 ; qr/n1)
± n1 |qr n2 n21 ≪N0
n2
 
X ām
× λf (m)e J± (m, n21 n2 , q),
m≪M0
q

where J± (m, n21 n2 , q) is the four-fold integral given in (13).

6. CAUCHY AND POISSON

In this section, we will further analyze Sr (N) given in Lemma 14.

6.1. Cauchy inequality. Spliting the sum over q in dyadic blocks q ∽ C and writing
q = q1 q2 with q1 |(n1 r)∞ , (n1 r, q2 ) = 1, we see that Sr (N) in Lemma 14 is bounded
by
N 5/3 log Q X X 1/3
X X |λπ (n1 , n2 )|
(15) sup 2/3 C 3
n1 1/3
C≪Q QT r ± n1 n1 N0 n2
(n1 ,r)
≪C |q |(n1 r)∞
(n1 ,r) 1 n2 ≪
n2
1
 X X 
× λf (m)C± (q, r, n1 , m)J± (m, n21 n2 , q),
 
q2 ∽C/q1 m≪M0

where the character sum C± (q, r, n1 , m) is defined as

  X    
X⋆ ām q X⋆ ᾱn2
C± (...) := S(rā, ±n2 ; qr/n1 )e = dµ e ± .
a mod q
q d qr/n1
d|q α mod qr/n1
n1 α≡−m mod d

We now analyze the sum inside | |. We split the m-sum into dyadic blocks m ∼ M1 .
On applying the Cauchy’s inequality to the n2 -sum in (15), we get the following
 SUBCONVEXITY BOUND FOR GL(3) × GL(2) L-FUNCTIONS IN GL(2) SPECTRAL ASPECT15

bound for Sr (N):

N 5/3 (QM0 )ǫ X X 1/3
X p
(16) Sr (N) ≪ sup 2/3 C 3
n1 Θ1/2 Ω± ,
M1 ≪M0 QT r ± n1 n1
C≪Q (n1 ,r)
≪C |q |(n1 r)∞
(n1 ,r) 1

where
X |λπ (n1 , n2 )|2
(17) Θ= 2/3
,
n2 ≪N0 /n21
n2

and
X  X X 2
(18) Ω± = λf (m)C± (q, r, n1 , m)J± (m, n21 n2 , q) .
 

n2 ≪N0 /n21 q2 ∽C/q1 m∽M1

with
(k − 1)2 C 2 −ǫ (k − 1)2 C 2
 
(19) k = M ≪ M1 ≪ M0 = k ǫ max 2
,Tx ,
N N
( )
3
(CT ) r
N0 = k ǫ max , T 3/2 N 1/2 rx3 .
N

6.2. Poisson summation. We now apply the Poisson summation formula to the n2 -
sum in (18). To this end, we smooth out the n2 -sum, i.e., we plug in an appropriate
smooth bump function, say, W . Opening the absolute value square, we get
XX XX
Ω± = λf (m)λf (m′ )
q2 ,q2′ ∽C/q1 m,m′ ∽M1
 
X n2
× W C± (...)C∓ (...)J± (m, n21 n2 , q)J± (m′ , n21 n2 , q ′ ),
n2 ∈Z
N0 /n21

where q ′ = q1 q2′ . Using the change of variable n2 n2 q1 q2 q2′ r/n1 + β with 0 ≤ β <
′
q1 q2 q2 r/n1 followed by the Poisson summation formula, we arrive at
N0 X X X X X
(20) Ω± = 2 λf (m)λf (m′ ) C± J± ,
n1 ′ m,m′ ∽M n ∈Z
q2 ,q2 ∽C/q1 1 2

where
XX  q   q′  X⋆ X⋆
′
(21) C± = dd µ µ 1
d d′
d|q α mod qr/n1 α′ mod q ′ r/n
1
d′ |q ′ n1 α≡−m mod d n1 α′ ≡−m′ mod d′
±ᾱq2′ ∓ᾱ′ q2 ≡−n2 mod q1 q2 q2′ r/n1

and
(22)
       
t t1 Nxu −Nx1 u1
Z Z Z Z Z Z
J± = W (x)W (x1 )V V e e
R R R R R R T T qQ qQ
 
N0 n2 w
Z
× W (w)I± (m, N0 w, q)I± (m′ , N0 w, q ′)e − dwdudtdxdu1 dt1 dx1 .
R q1 q2 q2′ rn1
16 SUMIT KUMAR

Now applying the Ramanujan bound (3) on the Fourier coefficients λf (m) and
λf (m′ ), we finally get
N0 X X X X X
(23) Ω± ≪ 2 |C± ||J± |.
n1 n ∈Z
′ ′ q2 ,q2 ∽C/q1 m,m ∽M1 2

On applying integration by parts repeatedly on the w-integral in (22), we observe

that it is negligibly small if
CN 1/3 r 2/3 n1
(24) |n2 | ≫ k ǫ 2/3
:= N2 .
q1 N0

7. BOUND FOR THE INTEGRAL

In this section, we will analyze w-integral given in (22). Let’s recall that it has the
following expression:
 
N0 n2 w
Z
(25) J± := ′ ′
W (w)I± (m, N0 w, q)I± (m , N0 w, q )e − ′ dw.
R q2 q2 q1 rn1
Our aim is to prove the following proposition:
√
4π M1 N
Proposition 1. Let J± be as above. Let B = C
. Then we have
1
J± ≪.
B2
Furthermore, if C ≫ k 1+ǫ and n2 =
6 0, then B ≍ k and we have
Cr 1/3 B 2/3
J± ≪ .
B 2 (NN0 )1/3
Before proving the above propositon, we will analyze the y-integral, I± (m, N0 w, q),
which is defined in (14) as,
(26)
5/2  √
3(NN0 w(y + u))1/3
  
4π mNy
Z
I± := I± (m, N0 w, q) = U(y)e ± Jk−1 dy.
1/2 qr 1/3 q
We have the following lemma.
√
4π mN
Lemma 15. Let I± (m, N0 w, q) be as in (26). Let B = q
and A = A(w) =
3(N N0 w)1/3
qr 1/3
. If A ≫ k ǫ B, then I± (m, N0 w, q) is negligibly small. In the other case,
i.e., A ≪ B, we have
1
I± (m, N0 w, q) ≪ .
B
(k−1) sin−1 τ 16π 2 A3
Furthermore, let f (τ ) = 2π
+ 27B 2 τ 2
with f ′ (τ0 ) = 0. If q ∼ C ≫ k 1+ǫ , then
we have
6 !
cA9/2

1 A
(27) I± (m, N0 w, q) = p 5
U
f (2) (τ0 ) B 5 τ0 cos τ0 Bτ0
(k − 1) sin−1 τ0 16π 2A3
 
−2020


×e + + O k .
2π 27B 2 τ02
 SUBCONVEXITY BOUND FOR GL(3) × GL(2) L-FUNCTIONS IN GL(2) SPECTRAL ASPECT17

Remark. The condition C ≫ k 1+ǫ assures that B ≍ k, also in this case, A =

3(N N0 w)1/3
qr 1/3
≪ (kT )1/2 = k 1−η/2 < B, where T = k 1−η .

Proof. Firstly, we consider the term e(±A(y + u)1/3 ) in I± , which can be expanded
as
e(±A(y + u)1/3 ) = e(±Ay 1/3 )e(±Au/3y 2/3 + ...).
We observe that the second term of the right hand side does not oscillate with respect
to y. So we can insert it into the weight function. Thus we essentially have the
following expression:
Z 5/2
√
I± = U(y)e(±Ay 1/3 )Jk−1 (B y)dy.
1/2

To analyze it further, we use an integral representation of Jk−1 (x). On applying

Lemma 6 to Jk−1(x), we see that the above y-integral becomes
Z π Z 5/2 √
B y sin τ
 
1 i(k−1)τ 1/3
I± = e U(y)e ±Ay − dydτ
2π −π 1/2 2π
We further split the above τ -integral as follows:
Z π Z π/2 Z π Z 0 Z −π/2
= + + + .
−π 0 π/2 −π/2 −π
(i)
Let I± denotes the part of I± corresponding to the i-th term of the above right side
(i)
for i = 1, ..., 4. We investigate each I± separately to analyze I± . To this end, we
(1)
first consider I± which is defined as follows:
Z π/2 Z 5/2 √
B y sin τ
 
(1) 1 i(k−1)τ 1/3
(28) I± = e U(y)e ±Ay − dydτ.
2π 0 1/2 2π
In the next step, we apply stationary phase analysis to the y integral. Using the
change of variable y y 3, y-integral transfers into
Z √3
5/2
By 3/2 sin τ
 
2 3
√ 3y U(y )e ±Ay − dy.
3
1/2 2π
Note that if we have negative sign with A, then the above y-integral is negligibly small
(1)
by Lemma 9. So, we now proceed with the y-integral of I+ . The phase function is
Ay − By 3/2 sin τ /2π.

2
The stationary point occurs at y0 = 3B4πA
sin τ
. Note that
p3
p 4π A 4π A
1/2 ≤ y0 ≤ 3 5/2 ⇔ b1 := 1/6
≤ sin τ ≤ =: b2 .
3 B(2.5) 3 B(0.5)1/6
Let’s assume that 0 < b1 /2 < 2b2 < 1 ( when 1 ≤ b1 /2, y-integral is negligibly small
due to absence of stationary point and the case b1 /2 < 1 < 2b2 can be analyzed
similarly). So that A ≪ B. Based on the above observations, we further split the
(1)
τ -range of I+ in (28) as follows:
Z π/2 Z sin−1 (b1 /2) Z sin−1 2b2 Z π/2
= + + .
0 0 sin−1 (b1 /2) sin−1 2b2
18 SUMIT KUMAR

Note that the first and third range of the right side makes y-integral negligibly small.
Hence we focus on the second range which is given as
Z sin−1 2b2 Z √3
5/2
By 3/2 sin τ
 
i(k−1)τ 2 3
e √ 3y U(y )e Ay − dydτ.
sin−1 (b1 /2) 3
1/2 2π
Since U is supported on [1/2, 5/2], we can adjust y-integral boundary as
Z sin−1 2b2 Z 16 √3
5/2
By 3/2 sin τ
 
i(k−1)τ 2 3
e √ 3y U(y )e Ay − dydτ.
sin−1 b1 /2 1 3
16
1/2 2π
(1)
On applying Lemma 10 to the above y-integral, we see that I+ , upto a negligibly
small error term, is given by
Z −1 6 ! 
cA9/2 sin 2b2 16π 2 A3
 
1 4πA (k − 1)τ
(29) U e + dτ.
B 5 sin−1 (b1 /2) sin5 τ 3B sin τ 2π 27B 2 sin2 τ
where c is some explicit absolute constant. Note that we have only worked with the
leading term of Lemma 10 as other terms gives us a better approximation. Next, we
apply second derivative bound on the above τ -integral. Applying Lemma 9 to (29),
we get
(1) 1
I± ≪ .
B
(i)
We can treat other I± in a similar manner. In fact, we get the same bound

(i) 1
I± ≪ ,
B
for i = 1, 2, 3, 4. Hence we get the first part of the lemma.
Now we proceed to prove the second part. Let’s assume that C ≫ k 1+ǫ . We will
(1) (i)
give details for I± only, as the analysis for other I± is similar. We start by analyzing
the τ -integral in (29). Using the change of variable sin τ τ in (29), we arrive at
Z 2b2 6 !
cA9/2 (k − 1) sin−1 τ 16π 2 A3
  
A
(30) 5 5
U e + dτ.
b1 /2 B τ cos τ Bτ 2π 27B 2 τ 2
We apply stationary phase expansion again. The stationary point of the phase func-
tion, f (τ ), occurs at τ0 , where τ0 satisfies
 3
τ03 4π A3
= .
B 2 (k − 1)
p
1 − τ02 3
The above equation can be solved using Cardano’s method to yield a solution of the
following form
(31) τ0 = c1 h(w) + c3 (h(w))3 + c3 (h(w))5 ... + c2n−1 (h(w))2n−1 + ...,
A
where each ci is a non-zero explicit absolute constant and h(w) = B 2/3 (k−1)1/3
. On
computing higher order derivatives of f (τ ), we get
f (2) (τ ) ≫ B 2 /A = A(A/B)−2 , f (j) (τ ) ≪ A(A/B)−j , j = 1, 3, 4, ...
 SUBCONVEXITY BOUND FOR GL(3) × GL(2) L-FUNCTIONS IN GL(2) SPECTRAL ASPECT19

On computing derivatives of the weight function g(τ ), we get

 −i
(i) 1 A
g (τ ) ≪ 1/2 , i = 0, 1, 2, ...
A B
Thus, on applying Lemma 10 to (30), we get the following expression:
6 ! 
cA9/2 (k − 1) sin−1 τ0 16π 2 A3
 
1 1 A
+ O k −2020 .


p 5 5
U e + 2 2
f (2) (τ0 ) B τ0 cos τ0 Bτ0 2π 27B τ0
Like before, here also, we have only considered the leading term of Lemma 10 as
other terms gives us a better approximation. This proves the lemma. 
Next, we consider the y ′-integral I± (m′ , N0 w, q ′ ) in (25). Analysis for I± (m′ , N0 w, q ′)
is exactly the same. We mention it in the following lemma.
√
4π m′ N
Lemma 16. Let I± (m′ , N0 w, q ′) be as above. Let B ′ = q′
and A′ = A′ (w) =
3(N N0 w)1/3
q ′ r 1/3
. If A′ ≫ k ǫ B ′ , then I± (m′ , N0 w, q ′) is negligibly small. Otherwise, we have
1
I± (m′ , N0 w, q ′) ≪ .
B′
(k−1) sin−1 τ0′ 16π 2 A′3
Furthermore, Let f1 (τ ) = 2π
+ 27B ′2 τ0′2
with f1′ (τ0′ ) = 0. If q ′ ∼ C ≫ k 1+ǫ ,
then we have
′9/2
 ′ 6 !
1 cA 1 A
(32) I± (m′ , N0 w, q ′ ) = p ′5 ′
U
′5
f ′(2) (τ0 ) B τ0 cos τ0
′ B ′ τ0′
(k − 1) sin−1 τ0′ 16π 2 A′3
 
×e + + O(k −2020 ).
2π 27B ′2 τ0′2
Proof of Proposition 1. First part of the proposition is easily deduced from the
first part of Lemma 15 and Lemma 16. To prove the second part, we substitute (27)
and (32) in place of I± (m, N0 w, q) and I± (m′ , N0 w, q ′ ) respectively in (25) to get the
following expression:

(33)
6 ! 6 !
c2 (AA′ )9/2 A′
 
1 A
Z
W (w) q U U
(2) (BB ′ )5 (τ0 τ0′ )5 cos τ0 cos τ0′ Bτ0 B ′ τ0′
R
f (2) (τ0 )f1 (τ0′ )
(k − 1)(sin−1 τ0 − sin−1 τ0′ ) 16π 2
 3
A′3
  
A N0 n2 w
×e + − − dw.
2π 27 B 2 τ02 B ′2 τ0′2 q2 q2′ q1 rn1
Now we apply the third derivative bound on the above w-integral. Consider the phase
function
(k − 1)(sin−1 τ0 − sin−1 τ0′ ) 16π 2
 3
A′3

A N0 n2 w
+ 2
− ′2 ′2 −
2π 27 2
B τ0 B τ0 q2 q2′ q1 rn1
(k − 1)(sin−1 τ0 − sin−1 τ0′ ) 16π 2 (k − 1) N0 n2 w
= + ((h(w))3 τ0−2 − (h′ (w))3 τ0′−2 ) − .
2π 27 q2 q2′ q1 rn1
Recall that
A 3(NN0 w)1/3 ′ 3(NN0 w)1/3
h(w) = 2/3 , A = , and A = .
B (k − 1)1/3 qr 1/3 q ′ r 1/3
20 SUMIT KUMAR

Using the Taylor’s series expansion and the stationary point expansion (31), we arrive
at
N0 n2 w
(k − 1)(c1 h(w) − c′1 h′ (w) + c3 (h(w))3 − c′3 (h′ (w))3 + ...) − .
q2 q2′ q1 rn1
Using the change of variable w w 3, and computing the third derivative of the
above phase function we arrive at
c3 (NN0 ) c′3 (NN0 )
 
6N0 n2 6N0 n2 6N0 n2 NN0
− ′ +− ′ + 3 2 − ′3 ′2 + ... = − ′ +O .
q2 q2 q1 rn1 q2 q2 q1 rn1 q rB q rB q2 q2 q1 rn1 C 3 rB 2
Since n2 6= 0, we note that
6N0 n2 N0 NN0
′
≫ 2
≫ kǫ 3 2 .
q2 q2 q1 rn1 (n1 , r)C r C rB
In the first inequality we have used the fact that n1 q2 ≪ C(n1 , r), while for the second
inequality, C ≫ k 1+ǫ , and Nr 2 ≪ k 3 is being used. Moreover, as τ0 , τ0′ ≍ A/B and
(2)
f (2) (τ0 ) ≍ B 2 /A, f1 (τ0′ ) ≍ B ′2 /A′ ,
we observe that weight function in (33) has total variation of size 1/B 2 . Hence using
the third derivative bound, i.e., Lemma 9 in (33) we get
1/3
1 n1 q2 q2′ q1 r Cr 1/3 B 2/3

(1)
J± ≪ 2 ≪ 2 ,
B N0 n2 B (NN0 )1/3
(1) (1)
where J± denotes the part of the w-integral (25) corresponding to I± . Analysis of
(i)
integrals corresponding to other I± gives us the same bound as the above. Hence we
have proved the second part of Proposition 1. 
We conclude this section by giving a bound on the main integral J± in (22).
Corollary 2. Let J± be as defined in (22). Under the same assumptions as in
Proposition (1), we have
C2
J± ≪ .
Q2 B 2
Furthermore, if C ≫ k 1+ǫ , n2 6= 0, then B ≍ k and we have
C 2 Cr 1/3 B 2/3
J± ≪ .
Q2 B 2 (NN0 )1/3
Proof. Proof follows by applying Proposition 1 to the w-integral and estimating the
rest of the integrals in (22) trivially. 

8. ANALYSIS OF THE ZERO FREQUENCY

With all the ingredients in hands, we will now estimate Sr (N) in (16) in the present
and coming sections. We start by considering Ω± in (20). Let Ω0± denotes the part of
Ω± corresponding to n2 = 0 and let Sr (N)0 denotes the part of Sr (N) corresponing
to Ω0± . We will prove the following lemma in this section.
Lemma 17. let Ω0± and Sr (N)0 be as above. Let T = k 1−η . Then we have
N0 C 2 rM1
Ω0± ≪ (M1 + C),
n21 kT q1
 SUBCONVEXITY BOUND FOR GL(3) × GL(2) L-FUNCTIONS IN GL(2) SPECTRAL ASPECT21

and
Sr (N)0 ≪ r 1/2 N 1/2 k 3/2−η/2 .
Proof. Firstly, let’s recall from (23) that
N0 X X X X
Ω0± ≪ 2 |C± ||J± |,
n1 ′ ′q2 ,q2 ∽C/q1 m,m ∽M1

In the case n2 = 0, it follows from the congruence conditions in the definition (21) of
C± that ±ᾱq2′ ∓ ᾱ′ q2 ≡ 0 mod q1 q2 q2′ r/n1 , which implies that q2 = q2′ and α = α′ . So
we can bound the character sum C± as
XX X⋆ XX qr
C± ≪ dd′ 1≪ dd′ ,
′
[d, d′ ]
d,d |q α mod qr/n1 ′ d,d |q
n1 α≡−m mod d (d,d′ )|(m−m′ )
n1 α≡−m′ mod d′

and hence we get

N0 X XX XX
Ω0± ≪ 2
qr (d, d′ ) |J± |
n1 ′
q2 ∼C/q1 ′ d,d |q m,m ∼M1
(d,d′ )|(m−m′ )

N0 C 2 X XX
≪ 2 2 2
qr (M1 (d, d′) + M12 ).
n1 Q B ′ q2 ∼C/q1 d,d |q

In the last inequality we have used Corollary 2. Upon executing the remaining sums
trivially we get
N0 C 4 rM1
Ω0± ≪ (M1 + C).
n21 q1 Q2 B 2
Hence first part of the lemma follows. We now substitute the above bound in place
of Ω± in (16) to get
1/2
N 5/3 X X 1/3 1/2
X C 2 (N0 r)1/2 M1 p √ 
sup 2/3 C 3
n 1 Θ 1/2 QB
M1 + C .
M1 ≪M0 QT r n n
n1 q1
C≪Q ± 1 ≪Cr 1 ∞ |q1 |(n1 r)
(n1 ,r) (n1 ,r)

Executing the q1 -sum trivially and replacing the range for n1 by the longer range
n1 ≪ Cr, we get
0 N 2/3 (N0 rM1 )1/2 X (n1 , r)1/2 1/2 p √ 
Sr (N) ≪ sup Θ M1 + C .
M1 ≪M0 r 2/3 CB n1 ≪Cr
7/6
n1
C≪Q

We evaluate n1 -sum, using the Cauchy’s inequality and the Ramanujan bound (see
Lemma 4), as
" #1/2  1/2
X (n1 , r) 1/2 X (n1 , r) X X |λπ (n1 , n2 )| 2
(34) Θ 1/2
≪   ≪ N01/6 .
7/6 2 2/3
n ≪Cr n1 n ≪Cr
n1 2
(n1 n2 )
1 1 n1 n2 ≤N0

Finally using N0 ≪ T 3/2 N 1/2 r, B = (M1 N)1/2 /C, we see that Sr (N)0 is bounded by
(NN0 )2/3 p √  √ 1/2 p p 
Sr (N)0 ≪ sup 1/6 N 1/2
M 1 + C ≪ sup T Nr M 0 + Q .
M1 ≪M0 r C≪Q
C≪Q
22 SUMIT KUMAR

Using trivial bound for M0 , i.e., M0 ≪ Q2 k 2 /N, we get

0
√ 1/2
 √ p  √ 1/2
 √ 
Sr (N) ≪ N T r Qk/ N + Q ≪ N T r Qk/ N ≪ r 1/2 N 1/2 k 3/2−η/2 .
Hence the lemma follows. 

9. ANALYSIS OF NON-ZERO FREQUENCIES

In this section, we will consider the cases which are compliment to Section 8. To
start with, we first analyze the character sum C± in (21).

9.1. The character sum. We estimate C± in the following lemma. This lemma is
taken from [18].
Lemma 18. Let C± be as in (21). Then we have
q 2 r(m, n1 ) XX
C± ≪ 1 d2 d′2 .
n1 ′ d2 |(q2 ,n1 q2 ∓mn2 )
d′2 |(q2′ ,n1 q2 ±m′ n2 )

Proof. Firstly, let’s recall (21) that

XX  q   q′  X⋆ X⋆
′
C± = dd µ µ 1.
d d′
d|q α mod qr/n1 α′ mod q ′ r/n
1
d′ |q ′ n1 α≡−m mod d n1 α′ ≡−m′ mod d′
±ᾱq2′ ∓ᾱ′ q2 ≡−n2 mod q1 q2 q2′ r/n1

Using the Chinese Remainder theorem, we observe that C± can be dominated by a

(1) (2)
product of two sums C± ≪ C± C± , where
(1)
XX X⋆ X⋆
C± = d1 d′1 1
q r q r
d1 ,d′1 |q1 β mod n1 β ′ mod n1
1 1
n1 β ≡ −m mod d1 n1 β ≡ −m′ mod d′1
′

±βq2′ ∓β ′ q2 +n2 ≡ 0 q1 r/n1

and X⋆ X⋆
(2)
XX
C± = d2 d′2 1.
d2 |q2 β mod q2 mod β′ q′2
d′2 |q2′ n1 β ≡ −m mod d2 n1 β ′ ≡ −m′ mod d′
2
±βq2′ ∓β ′ q2 +n2 ≡ 0 q2 q2′
(2)
In the second sum C± , since (n1 , q2 q2′ ) = 1, we get β ≡ −mn¯1 mod d2 and β ′ ≡
−m′ n¯1 mod d′2 . Then using the congruence modulo q2 q2′ , we conclude that
(2)
XX
C± ≪ d2 d′2 .
d2 |(q2 ,n1 q2′ ±mn2 )
d′2 |(q2′ ,n1 q2 ±m′ n2 )
(1)
In the first sum C± , the congruence condition determines β ′ uniquely in terms of β,
and hence
(1)
XX
′
X⋆ r q12 (m, n1 )
C± ≪ d1 d1 1≪ .
′
n1
d1 ,d1 |q1 β mod q1 r/n1
n1 β ≡ −m mod d1

Hence we have the lemma. 

 SUBCONVEXITY BOUND FOR GL(3) × GL(2) L-FUNCTIONS IN GL(2) SPECTRAL ASPECT23

9.2. Estimates for small q. In this subsection, we will estimate Sr (N) for small
value of q. Let Ω6= 0
± denotes the part of Ω± in (20) which is compliment to Ω± (
0

contribution of n2 6= 0) and let Sr6=0 (N) denotes the part of Sr (N) corresponding to
Ω6= 0
± . We will prove the following lemma.

Lemma 19. Ω6= 0 6=0

± and Sr (N) be as above. Then we have

r 2 C 5 (T N)1/2 C 2 n1 Cn1 M1
 
6=0 2
Ω± ≪ + + M1 .
n21 q1 Q2 B 2 q12 q1
6=0
Furthermore, let Sr,small (N) denotes the contribution of C ≪ k 1+ǫ to Sr6=0 (N). Then
we have
6=0
Sr,small (N) ≪ r 1/2 k 3−η/2 .

Proof. We start by estimating Ω6= 0

± . On using Lemma 18 in (23), we get

q12 N0 r X X XX X X X
Ω6= 0
± ≪ d 2 d ′
2 (m, n1 )|J± |.
n31 C
d2 |q2
′ ′
q2 ,q2 ∼ q m,m ∼M1 n2 ∈Z−{0}
d′2 |q2′ 1
n1 q2′ d′2 ∓mn2 ≡ 0 mod d2
n1 q2 d2 ±m′ n2 ≡ 0 mod d′2

Further writing q2 d2 in place of q2 and q2′ d′2 in place of q2′ , we arrive at

q12 N0 r X X XX X X X
(35) Ω6=
±
0
≪ d 2 d ′
2 (m, n1 )|J± |.
n31 ′ C
d2 ,d2 ≪C/q1 ′
q2 ∼ d m,m ∼M1 n2 ∈Z−{0}
2 q1
q2′ ∼ d′Cq n1 q2′ d′2 ∓mn2 ≡ 0 mod d2
2 1 n1 q2 d2 ±m′ n2 ≡ 0 mod d′2
′
Next, we count the number of m and m in the above expression. We have
 
X X X M1
(m, n1 ) = ℓ 1 ≪ (d2 , n2 ) n1 + .
m∼M
d2
1 ℓ|n1 m∼M1 /ℓ
n1 q2′ d′2 ∓mn2 ≡ 0 mod d2 n1 q2′ d′2 ℓ̄∓mn2 ≡ 0 mod d2

In the above estimate we have used the fact (d2 , n2 ) = 1. Counting the number of m
in a similar fashion we get that m-sum and m′ -sum in (35) is dominated by
  
ǫ ′ M1 M1
k (d2 , n1 q2 d2 ) (d2, n2 ) n1 + 1+ ′ .
d2 d2
Now substituting the above bound in (35), we arrive at
q12 N0 r|J± | X X
  
′
XX X
′ M1 M1
d2 d2 (d2 , n1 q2 d2 ) (d2, n2 ) n1 + 1+ ′ .
n31 ′ C
d2 d2
d2 ,d2 ≪C/q1 q2 ∼ d 1<|n2 |≪N2
2 q1
q2′ ∼ d′Cq
2 1

Now summing over n2 , q2′ and using Corollary 2, we get the following expression:
q1 N0 rN2 C 3 X X
  
6=0
X
′ M1 M1
Ω± ≪ d2 (d2 , n1 q2 d2 ) n1 + 1+ ′ .
n31 Q2 B 2 ′ C
d2 d2
d2 ,d2 ≪C/q1 q2 ∼ d
2 q1

Next we sum over d′2 to get

q1 N0 rC 3 N2 X   
X M1 C
Ω6=
±
0
≪ d2 n1 + + M1 .
n31 Q2 B 2 d2 q1
d2 ≪C/q1 q2 ∼ d Cq
2 1
24 SUMIT KUMAR

Finally executing the remaining sums, we get

(36)
N0 rC 4 N2 N0 rC 4 N2 C 2 n1 Cn1 M1
    
Cn1 C
Ω6=
±
0
≪ 3 2 2 + M1 + M1 ≪ 3 2 2 + 2
+ M1 .
n1 Q B q1 q1 n1 Q B q12 q1
Using the bounds for N0 and N2 from (19) and (24) respectively, we get
Cr 2/3 n1 rn1
N0 N2 ≪ (NN0 )1/3 ≪ (T N)1/2 C.
q1 q1
Hence
r 2 C 5 (T N)1/2 C 2 n1 Cn1 M1
 
(37) Ω6=
±
0
≪ + 2
+ M1 .
n21 q1 Q2 B 2 q12 q1
Thus we get the first part of the lemma. Let’s consider the third term in the right
side of the expression (37). Substituting it in place of Ω± in Sr (N) in (16), we get
N 5/3 M1 r(T N)1/4 C 5/2 X X −2/3 1/2 X 1
sup n1 Θ 1/2
M1 ≪M0 QT r 2/3 C 3 BQ ± n ≪Cr n1 q1
C≪k 1 |q |(n1 r)∞
(n1 ,r) 1
p
N 5/3 M1 rT 3/4 C 5/2 X X (n1 , r)
≪ sup 2/3 C 3 7/6
Θ1/2
M1 ≪M0 QT r BN 1/4 ± n ≪Cr n1
C≪k 1

N 5/3 M1 rT 3/4 C 5/2 1/6 1/2

≪ sup 2/3 3 1/4
N0 ≪ sup (rT N)1/2 M1 C 1/2 ≪ r 1/2 k 3−η/2 .
M1 ≪M0 QT r C BN M1 ≪M0
C≪k C≪k

In the above estimate we have used (34) and the fact M0 ≪ k 4 /N as C ≪ k. We now
consider the second term in the right hand side of (37). We see that it’s contribution
to Sr (N) in (16) is given by
N 5/3 (CM1 )1/2 r(T N)1/4 C 5/2 X X −2/3+1/2 1/2 X 1
sup n1 Θ
M1 ≪M0 QT r 2/3 C 3 BQ ± n ≪Cr n1
q1
C≪k 1 |q |(n1 r)∞
(n1 ,r) 1

N 5/3 (CM1 )1/2 rT 3/4 C 5/2 X X −2/3+1/2−1

≪ sup 2/3 3 1/4
n1 (n1 , r)Θ1/2
M1 ≪M0 QT r C BN ± n ≪Cr
C≪k 1

The above n1 -sum can be evalauted as follows:

" #1/2  1/2
X (n1 , r) X (n1 , r)2 X X |λπ (n1 , n2 )|2
Θ1/2 ≪   ≪ r 1/2 N01/6 .
7/6 2 2/3
n ≪Cr n1 n ≪Cr
n1 2
(n1 n2 )
1 1 n1 n2 ≤N0

Thus, using the above bound, we get

N 5/3 (CM1 )1/2 rT 3/4 C 5/2 1/2 1/6
sup r N0
M1 ≪M0 QT r 2/3 C 3 BN 1/4
C≪k

1/2 C 1/2
≪ sup r(T N)1/2 M1 C 1/2 ≪ k 3−η/2 .
M1 ≪M0 M 1/2
C≪k

In the last inequality we have used Nr 2 ≪ k 3+ǫ .

 SUBCONVEXITY BOUND FOR GL(3) × GL(2) L-FUNCTIONS IN GL(2) SPECTRAL ASPECT25

Finally, the contribution of the first term of (37) to Sr (N) is given by

N 5/3 M1 r(T N)1/4 C 5/2 X X 1/3−1/2 1/2 X 1
sup 2/3 C 3
n 1 Θ 3/2
M1 ≪M0 QT r BQ ± n ≪Cr n1 q1
C≪k 1
(n1 ,r)
|q1 |(n1 r)∞

N 5/3 CrT 3/4 C 5/2 X X (n1 , r)3/2 1/2

≪ sup 5/3
Θ .
M1 ≪M0 QT r 2/3 C 3 BN 1/4 ± n ≪Cr n1
C≪k 1

Using (34), we evaluate the n1 -sum as follows:

X (n1 , r)3/2 X (n1 , r) (n1 , r)1/2 1/6
1/2
5/3
Θ ≪ 1/2 7/6
Θ1/2 ≪ r 1/2 N0 .
n1 ≪Cr n1 n1 ≪Cr n1 n1
On using the above bound, we get
N 5/3 CrT 3/4 C 5/2 1/2 1/6 1/2 C
sup 2/3 3 1/4
r N0 ≪ sup r(T N)1/2 M1 C 1/2 ≪ k 3−η/2 .
M1 ≪M0 QT r C BN M1 ≪M0 M
C≪k C≪k

Finally, combining all the estimates we get the lemma. 

9.3. Estimates for generic q. It now remains to tackle the case when C ≫ k 1+ǫ
6=0
and n2 6= 0. Let Sr,generic (N) denotes the part of Sr6=0 (N) which is compliment to
6=0
Sr,small (N), i.e., the contribution of C ≫ k 1+ǫ to Sr6=0 (N). We have the following
lemma.
Lemma 20. Let T = k 1−η . For C ≫ k 1+ǫ , we have
6=0
(38) Sr,generic (N) ≪ N 1/2 k 3/2−1/6+3η/4 .
Proof. Let’s recall from (36) that
N0 rC 4 N2 C 2 n1 Cn1 M1
 
Ω6=
±
0
≪ 3 2 2 + 2
+ M1 .
n1 Q B q12 q1
In it’s proof, we had used J± ≪ C 2 /Q2 B 2 . For C ≫ k, we have a better bound for
the integral (see Corollary 2), i.e.,
C 2 Cr 1/3 B 2/3
(39) J± ≪ .
Q2 B 2 (NN0 )1/3
Cr 2/3 n1
Using the above bound coupled with N0 N2 ≪ q1
(NN0 )1/3 , we see that Ω6= 0
± is
bounded by
Cr 1/3 B 2/3 rC 4 Cr 2/3 n1 C 2 n1 Cn1 M1
 
Ω6=
±
0
≪ 2 3 2
(NN0 )1/3 + 2
+ M1 .
1/3
B (NN0 ) n1 Q q1 q12 q1
Note that
Cr 1/3 B 2/3 rC 4 Cr 2/3 n1 1/3 rC 4 rn1 Ck 2/3 C
(NN0 ) ≪
B 2 (NN0 )1/3 n31 Q2 q1 n31 B 2 Q2 q1
4
rC rn1 (T N)1/2 C
≪ 3 2 2 × k η−1/3
n1 B Q q1
r 2 C 5 (T N)1/2
= 2 2 2
× k η−1/3 .
n1 q1 Q B
26 SUMIT KUMAR

Here we used the fact that C ≪ (N/T )1/2 and T = k 1−η . Hence we get
r 2 C 5 (T N)1/2
 2 
6=0 η−1/3 C n1 Cn1 M1 2
Ω± ≪ ×k + + M1 .
n21 q1 Q2 B 2 q12 q1
Comparing it with the previous subsection bound (37), we observe that we have an
extra k η−1/3 factor in this case. Taking this into account and substituting C ≪
(N/K)1/2 , and accordingly M0 ≪ k 2 /T , in place of the corresponing bounds in the
6=0
proof of Lemma 19, we get the lemma. Let’s estimate Sr,generic (N) in this case.
Considering the contribution of the third term to Sr (N), we see that it is bounded
by
1/2
sup (rT N)1/2 M1 C 1/2 × k η/2−1/6 ≪ r 1/2 N 1/2 k(N/T )1/4 k η/2−1/6
M1 ≪M0
C≪Q

≪ k 7/12+3η/4 N 3/4 r 1/2 ≪ N 1/2 k 3/2−1/6+3η/4 .

Terms of Sr (N) corresponding to the second term are bounded by
sup T 1/2 N 1/2 rC × k η/2−1/6 ≪ T 1/2 N 1/2 r(N/T )1/2 × k η/2−1/6 ≪ N 1/2 k 3/2−1/6+η/2 .
C≪Q

Finally, in the last case, terms of Sr (N) corresponding to the first term are bounded
by
√ √ η/2−1/6
1/2 C QrN Tk
sup r(T N)1/2 M1 C 1/2 × k η/2−1/6 ≪ ≪ N 1/2 k 3/2−1/6+η/4 .
M1 ≪M0 M k
C≪k
6=0
Thus combining all the cases, we conclude that Sr,generic (N) is dominated by
6=0
Sr,generic (N) ≪ N 1/2 k 3/2−1/6+3η/4 .
Hence, we get the lemma. 

9.4. Estimates for the error term. In this subsection, we estimate Sr (N) corre-
sponding to the error term that was obtained from GL(3) Voronoi formula. Recall
from (10) that
∞ √
q 2 Qr
   
X an nx it
n
λπ (n, r)e e n V ≪ √ .
n=1
q qQ N N
√ p
Thus we saved N 3/2 /q 2 Qr. From GL(2) Voronoi formula, we saved N/M0 .
√
Estimation of a-sum also saves q. So in the present case we are able to save
√
N 2 /q 3/2 M0 Qr over the trivial bound N 2 of Sr (N) (8) and consequently the total
n2 n N
contribution of the terms 1q32r ≪ k ǫ to Sr (N) is bounded by
√ √ √ r 1/2 kN √ 3/2−1/2+3η/2
   
2 kQ kN N
(40) Q r √ + T ≪ r + √ ≪ ≪ Nk .
N T 3/2 T T 3/2
9.5. Conclusion. We now pull together the bounds from Lemma 17, Lemma 19,
Lemma 20 and the bound for the error term (40) to get that
3/2−η/2
Sr (N) −1/2+3η/2 1/2 −η/2 1/2 k
≪ k + r k + r + k −1/6+3η/4 .
N 1/2 k 3/2 N 1/2
 SUBCONVEXITY BOUND FOR GL(3) × GL(2) L-FUNCTIONS IN GL(2) SPECTRAL ASPECT27

Using k 3−θ < Nr 2 ≪ k 3 and r ≪ k θ , we further get

Sr (N)
≪ k −1/2+3η/2 + k θ/2−η/2 + k 2θ−η/2 + k −1/6+3η/4 .
N 1/2 k 3/2
Hence to get subconvexity, we need all of the above exponents to be negative. So
first and third term gives 1/3 > η > 4θ, and consequently third and fourth terms
dominate the rest. Thus above bound reduces to
Sr (N)
1/2 3/2
≪ k 2θ−η/2 + k −1/6+3η/4 .
N k
The optimal choice for η is given by η = 8θ/5 + 2/15. Plugging this in Lemma 11 we
get that
L(1/2, π × f ) ≪ k 3/2+6θ/5−1/15 + k 3/2−θ/2 ,
and with the optimal choice θ = 2/51, we obtain the bound given in Theorem 1.

References
[1] J. Bernstein and A. Reznikov: Periods, subconvexity of L-functions and representation theory,
J. Differential Geom. 13 (2005), 129-141.
[2] V. Blomer, R. Khan and M. Young: Distribution of mass of holomorphic cusp forms, Duke.
Journal (2013).
[3] V. Blomer and J. Buttcane: On the subconvexity problem for L-functions on GL(3), to appear
in Ann. Sci. ENS.
[4] D. A. Burgess: On character sums and L-series.II, Proc. London. Soc. 13 (1963), 524–536.
[5] D. Goldfeld: Automorphic forms and L-functions for the group GL(n, R), Cambridge Stud.
Adv. Math. , 99, Cambridge Univ. Press, Cambridge, 2006.
[6] A. Good: The square mean of Dirichlet series associated with cusp forms, Mathematika 29
(1982), 278–295.
[7] H. Iwaniec: The spectral growth of automorphic L-functions, . reine angew. Math. 428 (1992),
139-159.
[8] H. Iwaniec and M. Kowalski: Analytic number theory, American Mathematical society collo-
quium publications, Vol. 53, American Mathematical Society, Providence, RI, 2004.
[9] S. Kumar; Kummari M.; S. K. Singh: Sub-convexity bound for GL(3) × GL(2) L-functions:
GL(3)-spectral aspect, https://arxiv.org/abs/2006.07819, 2020.
[10] R. E. Langer : On the asymptotic solutions of ordinary differential equations, with an ap-
plication to the Bessel functions of large order, Transactions of the American Mathematical
Society, Volume 33, (1931), no. 1, 23-64.
[11] P. Michel; A. Venkatesh: The subconvexity problem for GL(2). Publ. Math. IHES 111 (2010),
171–280.
[12] S. D. Miller and W. Schmid: Automorphic distribution, L-functions, and Voronoi summation
for GL(3), Ann. of Math. 164 (2006), 423-488.
[13] Y.Lau, J. Liu, and Y. Ye: A new bound k 2/3+ǫ for Rankin-Selberg L-functions for Hecke
congruence subgroups, IMRP Int. Math. Res. Pep. (2006), Art. ID 35090, 78. MR 2235495.
[14] X. Li: The central value of the Rankin-Selberg L-functions. Geom. funct. anal. 18 (2009),
1660-1695.
[15] X. Li: Bounds for GL(3) × GL(2) L-functions and GL(3) L-functions, Annals of Math. 173
(2011), 301-336.
[16] R. Munshi: The circle method and bounds for L-functions-III, Journal of the American Math-
ematical Society 28 (2015), 913-938.
[17] R. Munshi: The circle method and bounds for L-functions-IV: subconvexity for twist of GL(3)
L-functions, Ann. Math 182 (2015), 617–672.
[18] R. Munshi: Subconvexity for GL(3) × GL(2) L-functions in t-aspect, arxiv 2018.
[19] Robert Rankin, The vanishing of Poincaré series, Proc. Edin. Math. Soc. 23 (1980), no. 2,
151-161.
[20] P. Sarnak: Estimates for Rankin-Selberg L-functions and quantum unique ergodicity, J. Funct.
Anal. 184 (2001), 419-453. MR 1851004.
28 SUMIT KUMAR

[21] P. Sharma: Subconvexity for GL(3) × GL(2) twists in level aspect, arxiv 2019.
[22] H. Weyl : Zur abschätzung von ζ(1 + it), Math. Z. 10 (1921), 88-101.

Sumit Kumar
Stat-Math Unit, Indian Statistical Institute, 203 B.T. Road, Kolkata 700108, India;
Email: [email protected]