THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II

NED CARMICHAEL

Abstract. Let f be a Hecke cusp form of weight k for SL2 (Z), and let (λf (n))n≥1 denote its
(suitably normalised) sequence
P of Hecke eigenvalues. We compute the first and second moments
of the sums S(x, f ) = x≤n≤2x λf (n), on average over forms f of large weight k. It is proved
that when the length of the sums x is larger than k2 , the second moment is roughly of size
x1/2 . This is in sharp contrast to the regime where x is slightly smaller than k2 , where it was
shown in preceding work [3] that the second moment is of size x.
arXiv:2502.03436v1 [math.NT] 5 Feb 2025

1. Introduction
Let k be an even positive integer, and let Bk denote a basis for the space of weight k cusp
forms for SL2 (Z) consisting of orthogonal Hecke eigenforms. Given f ∈ Bk , write
X
f (z) = λf (n)n(k−1)/2 e(nz),
n≥1

and normalise so that λf (1) = 1. We study the sums of Hecke eigenvalues

X
S(x, f ) := λf (n),
x≤n≤2x

as f varies over eigenforms in the basis Bk with k large (and even).

Similar problems have already been considered. For example, Lester and Yesha [12] study
the distribution of sums of Hecke eigenvalues over short intervals. Sums of eigenvalues in
progressions have also been investigated in [1], [6], [11], [12] and [13]. Notably, Lau and Zhao
[11] prove asymptotics for the variance of these sums (on average over the congruence class)
which demonstrate an interesting transition in the average size of the sums as the length of the
sums varies relative to the modulus.
The sums S(x, f ) themselves have been studied previously, for fixed f (and k). Indeed, a
1989 paper of Hafner and Ivić [7] gives
(1.1) S(x, f ) ≪f x1/3 .
(This can be slightly improved, see [16], [20] and [21].) Moreover, the following mean-square
estimate is known:
1 X X 2
Z
λf (n) dx = cf X 1/2 + O(log2 X).
X 0
n≤x

This is easily deduced from a result of Chandrasekharan and Narasimhan [4, Theorem 1] (see
also [18]). From these results, one may expect the sums S(x, f ) to be of size roughly x1/4 on
average.
None of the above results are uniform in the weight, however. Deligne’s bound states |λf (n)| ≤
d(n) (where d(n) denotes the divisor function), and therefore shows S(x, f ) ≪ x log x uniformly
in f . One can also easily derive
P uniform bounds using Perron’s formula and standard properties
of the L-function L(s, f ) = n≥1 λf (n)n−s . This yields S(x, f ) ≪ k 1+ǫ + x1/2+ǫ , which beats
the bound x log x if x ≥ k1+ǫ . Moreover, if one assumes GRH for L(s, f ) then one obtains
S(x, f ) ≪ x1/2+ǫ kǫ , improving the unconditional bound when x ≤ k2−ǫ .

Date: February 5, 2025.

1
2 NED CARMICHAEL

These bounds are reasonable in view of results in the preceding work [3], which we now
review. Define the f -averaging operator
X
(1.2) h·i = ω(f )·,
f ∈Bk

where ω(f ) are the harmonic weights

Γ(k − 1)
ω(f ) = .
(4π)k−1 kf k2
The harmonic weights arise naturally from the Petersson trace formula. We remark that the
1 + O(e−k ). In [3], the first and
P
f -averages are ‘normalised’ in the sense h1i = f ω(f ) =
second moments of the sums S(x, f ) were studied. It was proved that for x satisfying x → ∞
with k, but x = o(k2 / log6 k), one has
√
hS(x, f )i ≪ e− k
and hS(x, f )2 i ∼ c(x)x,
where c(x) = ck (x) is an explicit function satisfying 1/100 ≤ c(x) ≤ 2. In fact, c(x) = 1
provided x ∈ / [k/(8π), k/(4π)]. Therefore one expects S(x, f ) to be roughly of size x1/2 for
x = o(k2 / log6 k), and thus the GRH bound S(x, f ) ≪ x1/2+ǫ kǫ is sharp (at least, up to the
factor of (xk)ǫ ) in this range of x.
However, the sums S(x, f ) transition in size approximately when k2 /(32π 2 ) ≤ x ≤ k2 /(16π 2 ),
and we expect the sums to be considerably smaller after this transition (i.e. for x ≥ k2 /(16π 2 )).
In this paper, we consider the regime where x ≥ k2 /(8π 2 ), and demonstrate that the average
size of S(x, f ) is around x1/4 in this range of x.

1.1. Statement of Results. We now state the theorems. Our first result confirms that the
size of S(x, f ) is considerably smaller than x1/2 once x ≥ k2 /(8π 2 ). Indeed, in this range we
prove essentially the bound (1.1), but uniformly for f ∈ Bk when k is large.
Theorem 1.1. Let f be a Hecke eigenform of weight k for SL2 (Z), normalised so that λf (1) = 1.
Then for x ≥ k2 /(8π 2 ) and any ǫ > 0, we have
S(x, f ) ≪ x1/3+ǫ ,
where the implied constant depends only on ǫ.
We next evaluate the first and second moments (on average over f , see (1.2)) of the sums
S(x, f ). First, we must introduce some notation. Throughout this paper, we write κ = k − 1
for convenience. Define
ω(z) = ωκ (z) = (z 2 − κ2 )1/2 − κ arctan (z 2 /κ2 − 1)1/2 − π/4,

and (provided nx > κ2 /(16π 2 )) we define

(1.3) Ω(n, x) = Ωκ (n, x)

√ √
:= 2(32π 2 − κ2 /(nx))−3/4 sin ω(4π 2nx) − (16π 2 − κ2 /(nx))−3/4 sin ω(4π nx).
Importantly, Ω satisfies Ω(n, x) ≪ 1 for all integers n ≥ 1 whenever x ≥ k2 /(8π 2 ). We have the
following estimate for the first moment.
Theorem 1.2. Let ǫ > 0. If k2 /(8π 2 ) ≤ x ≤ k4 , then
√
hS(x, f )i = (−1)k/2 4 2πΩ(1, x)x1/4 + O(x1/2 k−1+ǫ ).
Finally, we evaluate the second moment.
Theorem 1.3. Let ǫ > 0. If k2 /(8π 2 ) ≤ x ≤ k12/5 , then
X Ω(n, x)2
hS(x, f )2 i = 32πx1/2 3/2
+ O(x3/4 k−3/5+ǫ ) + O(k29/30+ǫ ).
n≥1
n
 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 3

Moreover, the above main term satisfies

 log x  X Ω(n, x)2
x1/2 exp − ≪ 32πx1/2 ≪ x1/2 .
log log x n3/2 n≥1

Remark. The asymptotic behaviour of the variance

h(S(x, f ) − hS(x, f )i)2 i = hS(x, f )2 i − hS(x, f )i2 (1 + O(e−k ))
is easily deduced from the above. Indeed, if k2 /(8π 2 ) ≤ x ≤ k12/5 , one combines Theorems 1.2
& 1.3 to see
X Ω(n, x)2
h(S(x, f ) − hS(x, f )i)2 i = 32πx1/2 3/2
+ O(x3/4 k−3/5+ǫ ) + O(k29/30+ǫ ).
n≥2
n
As before, the main term satisfies
 log x  X Ω(n, x)2
x1/2 exp − ≪ 32πx1/2 ≪ x1/2 .
log log x n3/2n≥2

We now offer some explanation for the significant transition in the size of the sums S(x, f )
(occurring approximately when k2 /(32π 2 ) ≤ x ≤ k2 /(16π 2 )). This phenomenon may be ex-
plained with reference to the Voronoı̈ type summation formula for S(x, f ). This is formulated
precisely in Lemma 2.4, but very roughly speaking states that for a normalised eigenform f ,
Z 2
k/2
X √
(1.4) S(x, f ) ≈ 2π(−1) x λf (n) Jk−1 (4π nxt)dt.
n≥1 1

Here Jk−1 denotes the Bessel function. The behaviour of Jk−1 (z) is roughly as follows. When
the argument z is much smaller than the index k − 1, Jk−1 (z) is negligibly small. Jk−1 (z) then
increases to a large global maximum when z ≈ k − 1. To the right of this peak, once z is larger
than k − 1, the Bessel function oscillates with decaying amplitude.
Note that if x ≤ k2 /(32π 2 ) then there √
exist integers n ≥ 1 with k2 /(32π 2 x) ≤ n ≤ k2 /(16π 2 x).
For these n (and some 1√≤ t ≤ 2), 4π nxt ≈ k, and therefore one expects the peak of the
Bessel function Jk−1 (4π nxt) to produce a large contribution to the corresponding weight
R2 √
1 Jk−1 (4π nxt)dt found in (1.4). Consequently, we expect that the sums S(x, f ) can be large
when x ≤ k2 /(32π 2 ). √
On the other hand, if x ≥ k 2 /(16π 2 ), then 4π nxt ≥ k for all n ≥ 1 and 1 ≤ t ≤ 2. Thus all
Bessel functions appearing in (1.4) are in their oscillatory regime, and consequently one expects
R2 √
a lot of cancellation in the weights 1 Jk−1 (4π nxt)dt. Consequently, once x ≥ k2 /(16π 2 ), we
expect the size of the sums S(x, f ) to be relatively small.
1.2. Acknowledgements. I would like to thank Stephen Lester for many helpful discussions
and useful comments an earlier draft of this paper. I also thank Bingrong Huang for pointing
out the relevant work [8]. This work was supported by the Additional Funding Programme for
Mathematical Sciences, delivered by EPSRC (EP/V521917/1) and the Heilbronn Institute for
Mathematical Research.

2. Preliminaries
In this section we summarise some results for later use.
2.1. Bessel Functions. The Bessel functions Jν (z) will appear throughout this paper. They
are defined [19, p.40] by
X (−1)l  z ν+2l
(2.1) Jν (z) = .
l!Γ(ν + 1 + l) 2
l≥0

Throughout this section, we assume that the index ν and the argument z are real and positive.
The behaviour of Jν (z) is roughly as follows. When the argument z is small relative to the
4 NED CARMICHAEL

index ν, Jν (z) is small. The transition regime is where |z − ν| ≪ ν 1/3 , i.e. the argument is
approximately equal to the index. In this regime Jν (z) reaches its global maximum of size
≈ ν −1/3 . Finally, once the argument becomes larger than the index, the Bessel functions
oscillate, with the amplitude of the oscillations decaying roughly like z −1/2 .
A detailed discussion of the Bessel function (with references) is given in [3, Appendix A].
Here it suffices to summarise some useful facts in the following lemma.
Lemma 2.1. Let ν > 0 be large. We have the following.
(i) If 0 ≤ z ≤ (ν + 1)/4, then
 14ν 
(2.2) Jν (z) ≪ z 2 exp − .
13
(ii) If 0 ≤ z ≤ (ν + 1) − (ν + 1)1/3+δ for some 0 < δ ≤ 2/3, then
(2.3) Jν (z) ≪ exp(−ν δ ).
(iii) Uniformly for z ≥ 0, we have the bound
(2.4) Jν (z) ≪ ν −1/3 .
(iv) If z ≥ ν + ν 1/3+ǫ for any ǫ > 0, then
r r
2 2 2 2 1 5 ν2 
(2.5) Jν (z) = (z − ν 2 )−1/4 cos ω(z) + (z − ν 2 )−3/4 + sin ω(z)
π π 8 24 z 2 − ν 2
 z4 
+O ,
(z 2 − ν 2 )13/4
where ω is given by
ω(z) = ων (z) = (z 2 − ν 2 )1/2 − ν arctan (z 2 /ν 2 − 1)1/2 − π/4.


(2.6)
Remark. For convenience, we record
(z 2 − ν 2 )1/2
(2.7) ω ′ (z) = ,
z
and
ν2
(2.8) ω ′′ (z) = .
z 2 (z 2 − ν 2 )1/2
Proof. The bounds given in parts (i), (ii) and (iii) may be found in [3, Lemma A.2]. We now
address part (iv). The full asymptotic expansion in the oscillatory regime is computed, for
example, by Olver [15] - see §10.8, ex.8.2 (cf. §10.7, §10.8). Taking n = 1 in Olver’s expansion,
i.e. truncating after the first two terms gives that for z ≥ ν + ν 1/3+ǫ ,
r  z 2 −1/2  sin ω(z)  z 2
2 2 n −1/2 
(2.9) Jν (z) = (z − ν 2 )−1/4 cos ω(z)Û0 − 1 + Û1 − 1
π ν2 ν ν2
 2   z2 −1/2    z2 −1/2 o
+ O ν −2 exp Var Û1 ; 0, 2 − 1 Var Û2 ; 0, 2 − 1 .
ν ν ν
Here Var(f ; a, b) denotes the total variation of a function f on the interval [a, b], and the Ûi are
polynomials (cf. §10.7, (7.11)) given by
1 1
Û0 (t) = 1, Û1 (t) = (3t + 5t3 ) and Û2 (t) = (81t2 + 462t4 + 385t6 ).
24 1152
Since Û1 and Û2 are both increasing and Û1 (0) = Û2 (0) = 0, one has
  z2 −1/2   z 2 −1/2  ν 5ν 3
Var Û1 ; 0, 2 − 1 = Û1 − 1 = + .
ν ν2 8(z 2 − ν 2 )1/2 24(z 2 − ν 2 )3/2
Our assumption z ≥ ν + ν 1/3+ǫ =⇒ z 2 − ν 2 ≫ ν 4/3+ǫ implies
  z2 1/2 
Var Û1 ; 0, 2 − 1 ≪ ν 1/3−ǫ/2 + ν 1−3ǫ/2 ≪ ν 1−3ǫ/2 .
ν
 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 5

Similarly
  z2 −1/2   z 2 −1/2 
Var Û2 ; 0, 2 − 1 = Û2 − 1
ν ν2
 z2 −1  z 2 −2  z 2 −3 ν2  ν4 
≪ − 1 + − 1 + − 1 ≪ 1 + .
ν2 ν2 ν2 (z 2 − ν 2 ) (z 2 − ν 2 )2
So the error term in (2.9) is
2   z2 −1/2    z2 −1/2 
(2.10) ν −2 exp Var Û1 ; 0, 2 − 1 Var Û2 ; 0, 2 − 1
ν ν ν
1  ν4  z4
≪ exp(O(ν −3ǫ/2 )) 2 1 + ≪ .
(z − ν 2 ) (z 2 − ν 2 )2 (z 2 − ν 2 )3
In the last step, we used that for any z > ν,
z4 ν4 z4
1≪ and ≪ .
(z 2 − ν 2 )2 (z 2 − ν 2 )2 (z 2 − ν 2 )2
The asymptotic given in part (iv) now follows immediately from (2.9) and the bound (2.10). 

2.2. The Petersson Trace Formula. The key tool for computing f -averages is the Petersson
trace formula, which we now formulate. This gives an expression for the averages hλf (n)λf (m)i
as diagonal and off-diagonal terms. Conveniently, the off-diagonal is negligibly small in certain
ranges of m and n.
Lemma 2.2 (Petersson Trace Formula). Let S(m, n; c) denote the Kloosterman sums
X  a∗ m + an 
S(m, n; c) = e ,
c
a (mod c)
(a,c)=1

where a∗ denotes the multiplicative inverse of a modulo c. Set

(
1 if m = n,
δmn =
0 otherwise.
Let k be large, and h·i as given in (1.2). We have the following.
(i) [9, Theorem 3.6] For any positive integers m and n,
X  4π √mn 
k/2 −1
hλf (n)λf (m)i = δmn + 2π(−1) c S(m, n; c)Jk−1 .
c
c≥1

(ii) [17, Lemma 2.1] If m and n are positive integers satisfying mn ≤ k2 /104 , then
hλf (n)λf (m)i = δmn + O(e−k ).
2.3. A Voronoı̈ Summation Formula. We now state a Voronoı̈ type summation formula for
the sums S(x, f ). A version of this for smoothed sums of Hecke eigenvalues is given in [10,
ex.9, p.83]; the version here is adapted for the sharp cut-off sums S(x, f ). One first requires a
suitable smoothing function, given as follows.
Definition 2.3. Given ∆ ≥ 1, denote by w = w∆ a smooth function w : R → R satisfying the
following:
• supp w = [1, 2],
• w(ξ) = 1 for 1 + ∆−1 ≤ ξ ≤ 2 − ∆−1 ,
• for all integers j ≥ 0 and all ξ, we have w(j) (ξ) ≪j ∆j .
We now state the Voronoı̈ formula. This is [3, Lemmas 2.4 & 2.5].
6 NED CARMICHAEL

Lemma 2.4. Let ∆ be a large parameter satisfying ∆ ≤ x1−ǫ for some ǫ > 0. Let w = w∆ be
the associated smooth function given in Definition 2.3 above. Then for a Hecke eigenform f of
weight k for SL2 (Z), normalised so that λf (1) = 1, we have
X  nx   x log x 
S(x, f ) = 2π(−1)k/2 x λf (n)w̃ 2 + O ,
k + ∆2 ∆
n≥1
where
Z ∞ p
(2.11) w̃(ξ) = w̃∆ (ξ) = w(t)Jk−1 (4π (k2 + ∆2 )ξt)dt.
0
Moreover, for any integer A ≥ 0, w̃ satisfies the bound
(2.12) w̃(ξ) ≪A ξ −A .
Throughout this paper, we will require several estimates for the weights (2.11). Observe
 nx  Z ∞ √
(2.13) w̃ 2 = w(t)J k−1 (4π nxt)dt.
k + ∆2 0
√
Under our assumption x ≥ k2 /(8π 2 ), the Bessel function Jk−1 (4π nxt) above is always in its
oscillatory regime. This leads to cancellation in (2.13) (as discussed in the introduction). By
applying asymptotics for the Bessel function valid in the oscillatory regime, we capture this
cancellation in the following lemma.
Lemma 2.5. For ∆ ≥ 1, let w = w∆ be given as in Definition 2.3 and w̃ = w̃∆ as in (2.11).
Suppose x ≥ k2 /(8π 2 ). Then
 nx  r 2 Z ∞ √
w̃ 2 2
= w(t)(16π 2 nxt − κ2 )−1/4 cos ω(4π nxt)dt + O((nx)−5/4 ).
k +∆ π 0
Proof. We will apply the asymptotic (2.5) of √ Lemma√2.1 and integrate by parts. Indeed, since
we assume x ≥ k2 /(8π 2 ) it is clear that 4π nxt ≥ √ 2k for all n ≥ 1 and t ∈ supp w = [1, 2].
Thus (2.5) gives an asymptotic expansion for Jk−1 (4π nxt) valid in this range of x. Replacing
this in (2.13) (and noting 16π 2 nxt ≥ 2k2 =⇒ (16π 2 nxt − κ2 ) ≍ nx) shows
 nx  r 2 Z ∞ √
(2.14) w̃ 2 2
= w(t)(16π 2 nxt − κ2 )−1/4 cos ω(4π nxt)dt
k +∆ π 0
r Z ∞
2 1 5 κ2  √
+ w(t)(16π 2 nxt − κ2 )−3/4 + 2 2
sin ω(4π nxt)dt
π 0 8 24 (16π nxt − κ )
+ O((nx)−5/4 ).
We now integrate by parts in the latter integral above. First note that (2.7) implies
∂ √ 1
(2.15) ω(4π nxt) = (16π 2 nxt − κ2 )1/2 .
∂t 2t
This shows the second integral in (2.14) is
Z ∞ √
1 5 κ2 
w(t)(16π 2 nxt − κ2 )−3/4 + 2 2
sin ω(4π nxt)dt
0 8 24 (16π nxt − κ )
Z ∞ 1 5 κ2  √
tw(t)(16π 2 nxt − κ2 )−5/4


=−2 + 2 2
d cos ω(4π nxt)
0 8 24 (16π nxt − κ )
Z ∞ n
5 16π 2 nx  1 5 κ2 
=2 w(t) + tw′ (t) − tw(t) +
0 4 (16π 2 nxt − κ2 ) 8 24 (16π 2 nxt − κ2 )
5 16π 2 nxκ2 o
2 2 −5/4
√
− tw(t) (16π nxt − κ ) cos ω(4π nxt)dt ≪ (nx)−5/4 .
24 (16π 2 nxt − κ2 )2
The final bound above follows from the bounds w ≪ 1, w′ ≪ ∆ and the observation supp w′ ⊂
[1, 1 + ∆−1 ] ∪ [2 − ∆−1 , 2] (which has measure 2∆−1 ) (see Definition 2.3). The lemma now
follows immediately from (2.14). 
 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 7

2.4. Oscillatory Integrals. Finally, we record two lemmas that will be used later to handle
oscillatory integrals. The first of these is essentially a very general formulation of integration
by parts, due to Blomer, Khan and Young [2, Lemma 8.1].
Lemma 2.6 (Integration by Parts). Let Q, U, R, X > 0 and Y ≥ 1 be some parameters. Let
ρ and φ be two smooth functions. Assume ρ is compactly supported on the interval [α, β] and
satisfies
ρ(j) (t) ≪j XU −j for j = 0, 1, 2 . . .
Assume φ satisfies
φ′ (t) ≫ R and φ(j) (t) ≪j Y Q−j for j = 2, 3, . . .
Then
Z β n QR −B o
(2.16) ρ(t)eiφ(t) dt ≪B (β − α)X + (RU )−B
,
α Y 1/2
for any integer B ≥ 0.
We will apply this in situations where the terms QRY −1/2 and RU are both large, and thus
the integral (2.16) is shown to be negligible. Very roughly speaking, these conditions ensure
that the exponential phase eiφ(t) in the integral (2.16) oscillates rapidly compared to the weight
function ρ(t). One then expects this oscillation to create lots of cancellation in the integral.
On the other hand, if the phase function φ has a stationary point (that is, a point t0 for which
φ′ (t0 ) = 0) within the region of integration, one does not expect the integral to be negligible,
owing to a large contribution from a neighbourhood of the stationary point. In this situation,
Blomer, Khan and Young apply the method of stationary phase (in a very general setting) to
asymptotically expand the integral around the stationary point [2, Proposition 8.3]. We state
a simplified version of that result here, giving only an upper bound for the integral.
Lemma 2.7 (Stationary Phase). Let Q, V, X, Y > 0 be some parameters. Let ρ and φ be
two smooth functions, with ρ compactly supported on an interval [α, β] where β − α ≥ V . Set
Z := Q + X + Y + (β − α) + 1, and suppose there exists a fixed η > 0 such that
V Y 1/2
(2.17) Y ≥ Z η and ≥ Zη.
Q
Assume
ρ(j) (t) ≪j XV −j for j = 0, 1, 2, . . .
Additionally, suppose there exists a unique point t0 ∈ [α, β] satisfying φ′ (t0 ) = 0. Assume
further that
φ′′ (t) ≍ Y Q−2 and φ(j) (t) ≪j Y Q−j for j = 1, 2, 3, . . .
Then for any integer B ≥ 0,
Z β
QX
(2.18) ρ(t)eiφ(t) dt ≪B 1/2 + Z −B .
α Y
3. Bounds for Sums of Hecke Eigenvalues: Proof of Theorem 1.1
To prove Theorem 1.1, we will apply Lemma 2.4 to transform the sums (shortening their
effective length). We first give the following bound for the weights appearing in the transformed
sums, from which the theorem will follow straightforwardly.
Lemma 3.1. For ∆ ≥ 1, let w = w∆ be given as in Definition 2.3 and w̃ = w̃∆ as in (2.11).
Suppose x ≥ k2 /(8π 2 ). Then
 nx 
(3.1) w̃ 2
k + ∆2
√ Z ∞n
2 2 12π 2 nx ′
o
2 2 −3/4
√
= √ tw(t) − w(t) − tw (t) (16π nxt − κ ) sin ω(4π nxt)dt
π 0 (16π 2 nxt − κ2 )
+ O (nx)−5/4 .


8 NED CARMICHAEL

In particular, we have the bound

 nx 
(3.2) w̃ ≪ (nx)−3/4 .
k 2 + ∆2
Proof. Integrating the expression given in Lemma 2.5 by parts (using (2.15)), we have
 nx  r 2 Z ∞ √
w̃ 2 2
= w(t)(16π 2 nxt − κ2 )−1/4 cos ω(4π nxt)dt + O((nx)−5/4 )
k +∆ π
√ Z ∞0
2 2 √
tw(t)(16π 2 nxt − κ2 )−3/4 d sin ω(4π nxt) + O((nx)−5/4 )


= √
π 0
√ Z ∞n
2 2 3 16π 2 nx o √
=−√ w(t) + tw′ (t) − tw(t) (16π 2
nxt − κ2 −3/4
) sin ω(4π nxt)dt
π 0 4 (16π 2 nxt − κ2 )
+ O((nx)−5/4 ),
proving (3.1). To deduce (3.2) from (3.1), one uses the bounds w ≪ 1, w′ ≪ ∆ and notes that
w′ is supported on a set of measure ≤ 2/∆. 
Theorem 1.1 is now a simple consequence of Lemma 2.4 and the bound (3.2) of Lemma 3.1.
Proof of Theorem 1.1. Set ∆ = x2/3 throughout this proof. We apply Lemma 2.4 with this
choice of ∆, which shows
X  nx   x log x 
(3.3) S(x, f ) = 2π(−1)k/2 x λf (n)w̃ 2 + O
k + ∆2 ∆
n≥1
X  nx 
≪x d(n) w̃ 2 + x1/3 log x.
k + ∆2
n≥1

In the last line, we used Deligne’s bound |λf (n)| ≤ d(n).

If n ≥ x1/3+ǫ , then since x1/3+ǫ = ∆2 /x1−ǫ ≥ (k2 + ∆2 )/(2x1−ǫ ) (from our assumption
x ≥ k2 /(8π 2 ) =⇒ ∆2 = x4/3 ≥ k2 ), (2.12) of Lemma 2.4 gives the bound w̃(nx/(k2 + ∆2 )) ≪
n−2 x−1000 , say. If n ≤ x1/3+ǫ , we use the bound w̃(nx/(k2 + ∆2 )) ≪ (nx)−3/4 of Lemma 3.1.
From (3.3), we conclude the bound
X d(n) X d(n)
S(x, f ) ≪ x1/4 3/4
+ x −999
2
+ x1/3 log x ≪ x1/3+ǫ .
1/3+ǫ
n 1/3+ǫ
n
n≤x n≥x

4. The First Moment: Proof of Theorem 1.2

To compute the first moment hS(x, f )i, our strategy is to transform the sums using Lemma
2.4 and then apply the Petersson trace formula. We first need a more explicit form of the
smoothings w̃ appearing in the transformed sums, given in the following lemma.
Lemma 4.1. For ∆ ≥ 1, let w = w∆ be given as in Definition 2.3 and w̃ = w̃∆ as in (2.11).
Suppose x ≥ k2 /(8π 2 ). Then
 nx  2√2
w̃∆ 2 = √ Ω(n, x)(nx)−3/4 + O((nx)−5/4 ) + O((nx)−1/4 ∆−1 ).
k + ∆2 π
Proof. We use the expression given in Lemma 2.5, and first unsmooth the integral there. From
Definition 2.3 and Lemma 2.5 we obtain
 nx  r 2 Z 2 √
(4.1) w̃ 2 2
= (16π 2 nxt − κ2 )−1/4 cos ω(4π nxt)dt
k +∆ π 1
n Z 1+∆−1 Z 2 o 
+O + (16π 2 nxt − κ2 )−1/4 dt + O((nx)−5/4 ).
1 2−∆−1
 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 9

Since we assume x ≥ k2 /(8π 2 ) (which implies (16π 2 nxt − κ2 ) ≍ nx), the first error term here is
O((nx)−1/4 ∆−1 ). Integrating the main term by parts (recall (2.15)) we obtain
Z 2 √ Z 2 √
2 2 −1/4
(16π nxt − κ ) cos ω(4π nxt)dt = 2 t(16π 2 nxt − κ2 )−3/4 d(sin ω(4π nxt))
1 1
Z 2 2 √
3 16π nxt 
= 2Ω(n, x)(nx)−3/4 − 2 1− 2 2
(16π 2 nxt − κ2 )−3/4 sin ω(4π nxt)dt.
1 4 16π nxt − κ
Integrating by parts once again, the remaining integral is easily shown to be O((nx)−5/4 ).
Replacing this expression in (4.1), the lemma is proved. 
Equipped with this expression for the weights, Theorem 1.2 now follows straightforwardly
from the Voronoı̈ formula (Lemma 2.4) and the Petersson trace formula (Lemma 2.2).
Proof of Theorem 1.2. Throughout this proof, we fix ǫ > 0 and set ∆ = x1/2 k1−ǫ/2 . We first
wish to apply Lemma
√ 2.4 with this choice of ∆ (applicable since k4 ≥ x ≥ k2 /(8π 2 ) =⇒ ∆ =
1/2
x k 1−ǫ/2 ≤ 2π 2xk −ǫ/2 ≤ x1−ǫ/10 , say). Since for n ≥ (k2 + ∆2 )kǫ/2 /x, (2.12) provides the
bound w̃(nx/(k + ∆ )) ≪ n−2 k−1100 , Lemma 2.4 (with ∆ = x1/2 k1−ǫ/2 ) shows that for f ∈ Bk ,
2 2

X  nx   x log x 
S(x, f ) = 2π(−1)k/2 x λf (n)w̃ 2 + O
k + ∆2 ∆
n≥1
X  nx 
= 2π(−1)k/2 x λf (n)w̃ 2 2
+ O(k−1000 ) + O(x1/2 k−1+ǫ/2 log x).
2 2 ǫ/2
k + ∆
n≤(k +∆ )k /x

Since (k2 + ∆2 )kǫ/2 /x = k2+ǫ/2 /x + k2−ǫ/2 ≤ k2 /104 , we may apply part (ii) of Lemma 2.2 to
compute
X  nx 
hS(x, f )i = 2π(−1)k/2 x hλf (n)λf (1)iw̃ 2 2
+ O(x1/2 k−1+ǫ/2 log x)
k + ∆
n≤(k 2 +∆2 )k ǫ/2 /x
 x   X  nx  
= 2π(−1)k/2 xw̃ 2 + O x w̃ · e−k
+ O(x1/2 k−1+ǫ ).
k + ∆2 2 2 ǫ/2
k 2 + ∆2
n≤(k +∆ )k /x

It is clear (from the bound (3.2), say) that the first error term is O(e−k/2 ). Now applying
Lemma 4.1, we conclude
√
hS(x, f )i = 4 2π(−1)k/2 Ω(1, x)x1/4 + O(x−1/4 ) + O(x1/4 k−1+ǫ/2 ) + O(x1/2 k−1+ǫ ).


5. The Second Moment: Proof of Theorem 1.3

In this section, we compute the second moment hS(x, f )2 i in the range k2 /(8π 2 ) ≤ x ≤ k12/5 .
We begin with the following lemma, which applies the Petersson trace formula to naturally split
the second moment into diagonal and off-diagonal terms.
Lemma 5.1. Let ǫ > 0. Let ∆ be a large parameter satisfying x1/2 ≤ ∆ ≤ x1−ǫ . Let w = w∆
and w̃ = w̃∆ be the corresponding smoothings, given in Definition 2.3 and (2.11) respectively.
Suppose k2 /(8π 2 ) ≤ x ≤ k4 . Then we have
 σx log x   x2 log2 x 
hS(x, f )2 i = σ 2 + O +O ,
∆ ∆2
where
(5.1) σ 2 = (D) + (OD).
Here (D) denotes the diagonal terms
X  nx 2
(5.2) (D) = 4π 2 x2 w̃ .
k 2 + ∆2
n≤∆2 k ǫ /x
10 NED CARMICHAEL

The off-diagonal terms are given by

(5.3)
X  nx   mx  X −1  4π √mn 
(OD) = 8π 3 (−1)k/2 x2 w̃ w̃ 2 c S(m, n; c)Jk−1 .
k 2 + ∆2 k + ∆2 c
m,n≤∆2 k ǫ /x c≥1

Proof. Recall Lemma 2.4, which shows that for f ∈ Bk ,

X  nx   x log x 
(5.4) S(x, f ) = 2π(−1)k/2 x λf (n)w̃ 2 + O .
k + ∆2 ∆
n≥1

We claim we may truncate the sum over n in (5.4) at ∆2 kǫ /x. Indeed, since we assume ∆2 ≥
x ≥ k2 /(8π 2 ), we have
∆2 k ǫ ∆2 k ǫ k2+ǫ (k2 + ∆2 )kǫ  nx 
n≥ ≥ + ≥ =⇒ w̃ ≪ n−2 k−1100 ,
x 2x 16π 2 x 16π 2 x k 2 + ∆2
say, by the bound (2.12) of Lemma 2.4. It follows from (5.4) that
X  nx   x log x 
S(x, f ) = 2π(−1)k/2 x λf (n)w̃ 2 + O .
2 ǫ
k + ∆2 ∆
n≤∆ k /x

This allows us to compute

X  nx   mx 
(5.5) hS(x, f )2 i = 4π 2 x2 hλf (n)λf (m)iw̃ w̃ 2
k 2 + ∆2 k + ∆2
m,n≤∆2 k ǫ /x
D X  nx  x log x E  x2 log2 x 
+O x λf (n)w̃ · + O .
k 2 + ∆2 ∆ ∆2
n≤∆2 k ǫ /x

Set
X  nx   mx 
(5.6) σ 2 = 4π 2 x2 hλf (n)λf (m)iw̃ w̃ 2 .
k 2 + ∆2 k + ∆2
m,n≤∆2 k ǫ /x

Then
 σx log x   x2 log2 x 
hS(x, f )2 i = σ 2 + O ,+O
∆ ∆2
as claimed (we have obtained the first error term here by applying the Cauchy-Schwarz inequality
to the first error term of (5.5)). The explicit expression (5.1) for σ 2 follows immediately, upon
applying the Petersson trace formula (part (i) of Lemma 2.2) to (5.6). 
The diagonal terms (D) given in (5.2) will contribute the main term in Theorem 1.3; the
off-diagonal terms (OD) given in (5.3) will contribute only an error term. In Section 5.1, we
will evaluate the diagonal contribution, and prove the following.
Lemma 5.2 (Evaluation of (D)). Let ǫ > 0, and assume k2 /(8π 2 ) ≤ x ≤ k4 and x1/2 ≤ ∆ ≤
x1−ǫ . Then the following hold.
(i) We have
X Ω(n, x)2
(D) = 32πx1/2 3/2
+ O(xk ǫ ∆−1 ).
n≥1
n
(ii) Moreover, the main term above satisfies
 log x  X Ω(n, x)2
x1/2 exp − ≪ 32πx1/2 ≪ x1/2 .
log log x n3/2 n≥1

The expression for the main term in part (i) is a straightforward consequence of Lemma 4.1
(which gives an explicit expression for the smoothings w̃ appearing in (5.2)). It is perhaps
reasonable to expect that this main term (appearing in Lemma 5.2 and Theorem 1.3) should be
 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 11

of size x1/2 . Indeed, for x ≥ k2 /(8π 2 ) we have the bound Ω(n, x) ≪ 1, and so it is immediate
that in this range X
(D) ≪ x1/2 n−3/2 ≪ x1/2 .
n≥1
The more difficult part of the proof of Lemma 5.2 is establishing the lower bound (D) ≫
x1/2 exp(− log x/ log log x). The main term behaves roughly like
X sin2 ϕ(n)
x1/2 ,
n≥1
n3/2
for some complicated phase function ϕ depending on x. Since all terms in this sum are positive,
for the lower bound it suffices to show that the phase ϕ(n) is asymptotically equidistributed
modulo 2π (since then not all the oscillatory terms can be small simultaneously). This is
achieved by a standard application of the Erdős-Turán inequality.
Bounding the off-diagonal contribution (5.3) is the most difficult part of the proof of Theorem
1.3. In Section 5.2, we obtain the following.
Lemma 5.3 (Bound for (OD)). Let 0 < ǫ < 1/1000, and assume x1/2 ≤ ∆ ≤ x2/3 k1/3−ǫ .
Provided k2 /(8π 2 ) ≤ x ≤ k4 , we have
(OD) ≪ k−8/3 ∆2 + k−3/2+ǫ ∆3/2 + x−1/2 k1/6+2ǫ ∆ + x−3/2 k−5/6+4ǫ ∆3 .
We now give a brief sketch of the proof of Lemma 5.3. The basic idea is to reorder summation
in (5.3) and apply Poisson summation in the n variable. Very roughly speaking, after reordering
summation we have
X X  mx  X  nx   4π √mn 
2 −1
≈x c w̃ 2 S(m, n; c)w̃ 2 Jk−1 .
2 2
k + ∆2 2
k + ∆2 c
c≤∆ /(kx) m≤∆ /x n≤∆ /x

(The effective
√ range of c is a consequence of the effective support of the Bessel function
Jk−1 (4π mn/c).) Opening the Kloosterman sum (writing a∗ for the multiplicative inverse
of a modulo c), we can rewrite the above as
X X X  a∗ m   mx 
≈ x2 c−1 e w̃ 2
c k + ∆2
2
c≤∆ /(kx) 1≤a≤c m≤∆ /x
2
(a,c)=1
X  an   nx   4π √mn 
e w̃ 2 Jk−1 .
c k + ∆2 c
n≤∆2 /x

We will use
mx  −3/4
 nx 
(5.7) w̃ ≪ (mx) and w̃ ≈ (nx)−3/4 sin ϕ(n),
k 2 + ∆2 k 2 + ∆2
where again ϕ is some complicated phase function depending on x. The first bound is (3.2) of
Lemma 3.1, and the latter is a very rough approximation of (3.1) (which suffices for this sketch).
Applying the triangle inequality and (5.7), it suffices to bound a sum of shape
X X X X  an   4π √mn 
1/2 −1 −3/4 −3/4
≈x c m e n sin ϕ(n)Jk−1 .
c c
2
c≤∆ /(kx) 2
m≤∆ /x 1≤a≤c n≤∆ /x
2
(a,c)=1

The difficult part of the proof is bounding the inner sum over n. One would expect to find
cancellation in this sum, coming from the oscillations of the exponential terms, and this is
captured by applying Poisson summation. Indeed, this shows that the inner sum over n is
roughly equal to the dual sum
X Z ∆2 /x  4π √mt   ñt 
−3/4
t sin ϕ(t)Jk−1 e − dt.
0 c c
ñ∈Z
ñ≡−a (mod c)
12 NED CARMICHAEL

√
Integrating by parts, one finds that this dual sum is effectively of length ≈ mx/k(≪ ∆/k 1−ǫ ).
This is shorter than the original sum, and therefore we find some savings.
It remains to estimate the oscillatory integrals appearing
√ in the dual sum, which requires
some careful analysis of the Bessel function Jk−1 (4π mt/c). Handling the integrals using the
method of stationary phase, we are able to show that the contribution of the off-diagonal terms
is asymptotically smaller than that of the diagonal main term in the range x ≤ k 12/5−ǫ .
Our proof is similar to work of Hough [8], in which a twisted second moment estimate (also
in weight aspect) is established for central values of L(s, f ). After approximating the central
values by Dirichlet polynomials (with coefficients λf (n)) and averaging with the Petersson trace
formula, the off-diagonal contribution arising is handled in the same way.
Proof of Theorem 1.3, assuming Lemma 5.2 and Lemma 5.3. We will apply Lemma 5.3 with
∆ = x1/2 k3/5 (and 0 < ǫ < 1/1000) to bound the off-diagonal contribution. One first checks
the condition
∆ = x1/2 k3/5 ≤ x2/3 k1/3−ǫ ⇐⇒ x ≥ k24/15+6ǫ ,
which is assured by our assumptions x ≥ k2 /(8π 2 ) and ǫ < 1/1000. Lemma 5.3 therefore shows
(OD) ≪ xk −22/15 + x3/4 k−3/5+ǫ + k23/30+2ǫ + k29/30+4ǫ .
Noting that our assumption x ≤ k 12/5 =⇒ xk −22/15 ≤ x3/4 k−13/15 ≤ x3/4 k−3/5+ǫ , so the above
bound may be simplified to
(OD) ≪ x3/4 k−3/5+ǫ + k29/30+4ǫ .
Observe from (5.1) and Lemma 5.2 (applied with our choice of ∆ = x1/2 k3/5 ) that
X Ω(n, x)2
(5.8) σ 2 = (D) + (OD) = 32πx1/2 3/2
+ O(x3/4 k−3/5+ǫ ) + O(k29/30+4ǫ ).
n≥1
n

Note x ≤ k12/5 =⇒ x3/4 k−3/5+ǫ ≤ x1/2 kǫ . In particular, we have σ 2 ≪ x1/2 + x3/4 k−3/5+ǫ =⇒
σ ≪ x1/4 kǫ/2 .
Finally, from (5.8) and Lemma 5.1 we have for k2 /(8π 2 ) ≤ x ≤ k12/5 ,
 σx log x   x2 log2 x 
hS(x, f )2 i = σ 2 + O +O
∆ ∆2
X Ω(n, x)2
= 32πx1/2 + O(x3/4 k−3/5+ǫ ) + O(k29/30+4ǫ ).
n≥1
n3/2

(We used that x2 log2 x/∆2

= x log 2
xk −6/5
≤ x3/4 k−3/5+ǫ , since x ≤ k12/5 .) This completes
the proof of Theorem 1.3 (upon replacing ǫ by ǫ/4). 
5.1. The Diagonal Terms. In this section we investigate the diagonal contribution, which
gives the main term in Theorem 1.3. Recall the definition (5.2), which states
X  nx 2
(D) = 4π 2 x2 w̃ 2 .
2 ǫ
k + ∆2
n≤∆ k /x

Our goal is the following.

Lemma 5.2 (Evaluation of (D)). Let ǫ > 0, and assume k2 /(8π 2 ) ≤ x ≤ k4 and x1/2 ≤ ∆ ≤
x1−ǫ . Then the following hold.
(i) We have
X Ω(n, x)2
(D) = 32πx1/2 3/2
+ O(xk ǫ ∆−1 ).
n≥1
n
(ii) Moreover, the main term above satisfies
 log x  X Ω(n, x)2
x1/2 exp − ≪ 32πx1/2 ≪ x1/2 .
log log x n3/2 n≥1
 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 13

Proof of Lemma 5.2, part (i). The explicit expression for the diagonal terms is a simple conse-
quence of Lemma 4.1, which states
 nx  2√2
w̃ 2 = √ Ω(n, x)(nx)−3/4 + O((nx)−5/4 ) + O((nx)−1/4 ∆−1 ).
k + ∆2 π
From the definition (5.2) and the bound Ω(n, x) ≪ 1 we compute
√
X 2 2 2
2 2
(D) = 4π x √ Ω(n, x)(nx)−3/4 + O((nx)−5/4 ) + O((nx)−1/4 ∆−1 )
2 ǫ
π
n≤∆ k /x
X Ω(n, x)2  X 
= 32πx1/2 + O x2 (nx)−1 ∆−1
n3/2
n≤∆2 k ǫ /x n≤∆2 k ǫ /x
 X 
+ O x2 (nx)−1/2 ∆−2 + O(1)
n≤∆2 k ǫ /x
X Ω(n, x)2  X   x log ∆   xkǫ/2 
= 32πx1/2 + O x1/2 n−3/2 + O +O + O(1)
n≥1
n3/2 ∆ ∆
n≥∆2 k ǫ /x
X Ω(n, x)2  xk ǫ 
= 32πx1/2 +O ,
n≥1
n3/2 ∆

as claimed. 
Proving part (ii) is much more difficult, and we devote the remainder of the section to this
task. It is immediate that x1/2 n Ω(n, x)2 /n3/2 ≪ x1/2 , but it is more difficult to establish
P

an inequality in the opposite direction. It is reasonable to guess x1/2 n Ω(n, x)2 /n3/2 ≍ x1/2 .
P
To prove this, it would of course suffice to show that for at least one n ≤ 1000, say, we√have
|Ω(n, x)| ≥ 1/1000.
√ Unfortunately, due to the presence of the oscillatory terms sin ω(4π nx)
and sin ω(4π 2nx) in the definition (1.3) of Ω(n, x), it is possible that no such condition holds
for some perverse choice of x.
InPthe remainder of this section, we apply standard techniques to establish the lower bound
x1/2 n Ω(n, x)2 /n3/2 ≫ x1/2 exp(− log x/ log log x). It is certainly reasonable to expect that
for at least some n, we have
(5.9) Ω(n, x) ≫ 1.
Indeed, sufficient conditions for (5.9) to hold are n ≥ 4 and
√  
ω(4π 2nx) 9 11
(5.10) h(n) := ∈ , (mod 1).
2π 40 40
(In other words the fractional part {h(n)} := h(n)−⌊h(n)⌋ ∈ [9/40, 11/40].) To see this, observe
√
 
9 11 h 9π 11π i
h(n) ∈ , (mod 1) =⇒ ω(4π 2nx) ∈ , (mod 2π)
40 40 20 20
√  9π 
=⇒ sin ω(4π 2nx) ≥ sin ≥ c,
20
where (for convenience later on) we define the constant
101
c := 21/2 · (7/4)−3/4 ·
= 0.938 . . .
100
(sin(9π/20) = 0.987 . . . for comparison). Consequently, if (5.10) is satisfied we have
√ √
|Ω(n, x)| = |2(32π 2 − κ2 /(nx))−3/4 sin ω(4π 2nx) − (16π 2 − κ2 /(nx))−3/4 sin ω(4π nx)|
√
≥ 2(32π 2 − κ2 /(nx))−3/4 sin ω(4π 2nx) − (16π 2 − κ2 /(nx))−3/4
≥ 2c(32π 2 − κ2 /(nx))−3/4 − (16π 2 − κ2 /(nx))−3/4 .
14 NED CARMICHAEL

We next use the assumptions x ≥ k2 /(8π 2 ) and n ≥ 4 ⇐⇒ 2 − 1/n ≥ 7/4. If (5.10) is satisfied,
it follows from these assumptions that
(5.11) |Ω(n, x)| ≥ 2c(32π 2 )−3/4 − (16π 2 − 8π 2 /n)−3/4
≥ 2c(32π 2 )−3/4 − (8π 2 )−3/4 (7/4)−3/4
= (8π 2 )−3/4 (2−1/2 c − (7/4)−3/4 )
≥ (8π 2 )−3/4 (7/4)−3/4 (101/100 − 1) ≫ 1.
Our strategy to prove the lower bound in part (ii) of Lemma 5.2 is to show that the sequence
(h(n))n≥1 is asymptotically equidistributed modulo 1. This ensures that the condition (5.10)
(hence (5.9)) holds for a positive proportion of n. It is these n which will give a large contribution
to the main term x1/2 n Ω(n, x)2 /n3/2 , leading to the desired lower bound.
P
For any integer N , we define
   
9 11
Z(N ) := 4 ≤ n ≤ N : h(n) ∈ , (mod 1) .
40 40
Considering n ∈ Z(N ) =⇒ Ω(n, x) ≫ 1, we seek lower bounds for #Z(N ). We have the
following estimate (see [14, §1]):
(5.12) #Z(N ) = (N − 4)/20 + O(D(N )).
Here (N − 4)/20 is the expected size of #Z(N ) assuming the sequence h(n) is asymptotically
equidistributed modulo 1, and D(N ) ≥ 0 is the discrepancy of the sequence (h(n))n≥4 . The
discrepancy is a measure of the extent to which this assumption fails, and is defined as
D(N ) := sup #{4 ≤ n ≤ N : h(n) ∈ [α, β]} − (β − α)(N − 4) .
[α,β]⊂[0,1]

We will use the following bound for the discrepancy (which is essentially a quantitative version
of Weyl’s equidistribution theorem).
Lemma 5.4 (Erdős-Turán Inequality [14, Corollary 1.1]). For any positive integer R, we have
N X1 X
(5.13) D(N ) ≤ +3 e(rh(n)) .
R+1 r
r≤R n≤N

To bound the exponential sums appearing in (5.13), we apply van der Corput’s method.
Specifically, we will require the following standard lemma.
Lemma 5.5 (van der Corput [10, Theorem 8.20]). Let f be a real-valued function which is p
times continuously differentiable, with p ≥ 2. Let b − a ≥ 1, and suppose that if ξ ∈ [a, b],
λ ≤ f (p) (ξ) ≤ µλ for some λ > 0 and µ ≥ 1. Then
X
e(f (n)) ≪ (b − a)µ2/P λ1/(2P −2) + (b − a)1−2/P λ−1/(2P −2) ,
a<n≤b

where P = 2p−1 and the implied constants are independent of p.

Equipped with these preliminary results, we now prove the key lower bound in Lemma 5.2.
Proof of Lemma 5.2, part (ii). For any positive integer N , we have the following lower bound
for the diagonal main term:
X Ω(n, x)2 X Ω(n, x)2
32πx1/2 ≥ 32πx 1/2
.
n≥1
n3/2 n3/2
n∈Z(N )

Since n ∈ Z(N ) =⇒ n ≤ N and Ω(n, x) ≫ 1 (see (5.11)), it follows

X Ω(n, x)2
(5.14) 32πx1/2 3/2
≫ x1/2 N −3/2 #Z(N ).
n≥1
n
 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 15

To optimise the bound (5.14), we wish to choose N small enough that x1/2 N −3/2 is of comparable
size to x1/2 . However, we must also take N large enough to ensure that #Z(N ) is non-zero. To
achieve the latter condition, we will choose N large enough that D(N ) = o(N ), whence (5.12)
implies #Z(N ) ∼ N/20.
In order to bound the
P discrepancy D(N ), we will apply Lemma 5.4. But first, we must bound
the exponential sums n≤N e(rh(n)) arising from (5.13). To do so, we apply Lemma 5.5. Recall
√
ω(4π 2xξ)
h(ξ) = .
2π
For ξ ≥ 1, one computes (see (2.7))
(32π 2 xξ − κ2 )1/2 √  −1/2 X
  2 
l 1/2 κ l 
′ −1/2−l
(5.15) h (ξ) = = 2x ξ + (−1) ξ ,
4πξ l 32π 2 x
l≥1

where
1 1 3−2l




− ···
 
1/2 2 2 2
Y  3 − 2s 
(5.16) = = .
l l! 2s
1≤s≤l

For p ≥ 2, one computes

dp−1  −1/2  1  3   2p − 3 
(5.17) ξ = − − · · · − ξ 1/2−p
dξ p−1 2 2 2
(−1)p (2p − 3)! 1/2−p  p p−1 1/2−p
= 2(p−2) ξ ≍ ξ ,
2 (p − 2)! e
where the last step follows from Stirling’s formula. Similarly, for l ≥ 1 one has
dp−1  −1/2−l  2l + 1  2(l + 1) + 1   2(l + p − 1) − 1 
ξ = − − · · · − ξ 1/2−p−l
dξ p−1 2 2 2
dp−1 
= ap (l) · ξ −l · p−1 ξ −1/2 ,
dξ
where we define
(2p − 1)(2(p + 1) − 1) · · · (2(l + p − 1) − 1)
ap (l) := .
1 · 3 · · · (2l − 1)
From (5.15), it follows
√  dp−1  −1/2 
  2 
l 1/2 κ
X l 
(p) −l
(5.18) h (ξ) = 2x ξ 1 + (−1) ap (l)ξ .
dξ p−1 l 32π 2 x
l≥1

If ξ ≥ 2p we observe
    
1 1 2l−3
1 − 2p 1 + 2p · 1+ 2p
 1  1 1   1 2l − 3 
ap (l)ξ −l ≤ ≤ 1− + ··· + ≤ 1.
1 · 3 · · · (2l − 1) 2p 3 6p 2l − 1 (2l − 1)(2p)
Note our assumption x ≥ k2 /(8π 2 ) implies κ2 /(32π 2 x) ≤ 1/4. Moreover, it is clear from (5.16)
that 1/2


l ≤ 1. Consequently, if ξ ≥ 2p ≥ 4 we bound
  2 
l 1/2 κ
X l X 1 1
(5.19) (−1) 2
ap (l)ξ −l ≤ = .
l 32π x 16l 15
l≥1 l≥1

We now conclude from (5.17), (5.18) and (5.19) that for any p ≥ 2 and ξ ≥ 2p,
dp−1   p p−1
h(p) (ξ) ≍ x1/2 p−1 ξ −1/2 ≍ x1/2 ξ 1/2−p .
dξ e
We may now apply Lemma 5.5. Suppose p ≥ 2. If ξ ∈ [a, b] with a ≥ 2p and b ≤ 2a, we have
 p p−1  p p−1  p p−1
x1/2 b1/2−p ≪ h(p) (ξ) ≪ x1/2 a1/2−p ≪ 2p x1/2 b1/2−p .
e e e
16 NED CARMICHAEL

So there exist positive (absolute) constants A1 and A2 such that for ξ ∈ [a, b] and r ≥ 1,
 p p−1  p p−1
A1 r x1/2 b1/2−p ≤ rh(p) (ξ) ≤ A2 2p r x1/2 b1/2−p .
e e
Applying Lemma 5.5 with λ = A1 r(p/e)(p−1) x1/2 b1/2−p and µ = A2 2p /A1 , we find
 A 2p 2/P   p p−1 1/(2P −2)
2
X
e(rh(n)) ≪ b A1 r x1/2 b1/2−p
A1 e
a<n≤b
  p p−1 −1/(2P −2)
+ b1−2/P A1 r x1/2 b1/2−p
e
1 2p−1 2p−1
1− 2(2P 1 − 2(2P1−2) 1+ 2(2P − P2 − 2P1−2
≪ x 2(2P −2) b −2) r 2P −2 + x b −2) r ,
where P = 2p−1 . In the last line, we used that 1/2 ≤ (p/e) ≤ 2 and 1 ≤ 22p/2 (p−1)/(2p −2)
≤5
p−1

for all p ≥ 2. Summing over dyadic intervals and bounding the contribution of n ≤ 2p trivially,
we obtain
1
1− 2p−1 1 − 1
1+ 2p−1 − 2 1
e(rh(n)) ≪ x 2(2P −2) N 2(2P −2) r 2P −2 + x 2(2P −2) N 2(2P −2) P r − 2P −2 + p.
X

n≤N

Returning to Lemma 5.4 and applying the estimate above, for any integer R ≥ 1 we obtain
(5.20)
1 2p−1 2p−1
1− 2(2P 1 − 2(2P1−2) 1+ 2(2P − P2
D(N ) ≪ N R−1 + x 2(2P −2) N −2) (2P − 2)R 2P −2 + x N −2) log R + p log R
1 2p−1 2p−1
− 2(2P 1 − 2(2P1−2) − P2
≪ N R−1 + 2p x log R + pN −1 log R .


2(2P −2) N −2) R 2P −2 +x N 2(2P −2)

Choose
j1 k j 1
k j 1
k
(5.21) p= (log log x + 3) , N = x (2p−3) , and R = x (2p−3)(2P −1) .
2
We claim that with these choices of parameters, the expression given in (5.20) is o(N ). Indeed,
we now check each term appearing in (5.20). Firstly, it is simple to note
1
 log x   (log x)1−log 2/2 
R ≥ x (2p−3)(2P −1) − 1 ≥ exp − 1 = exp − 1 → ∞,
log log x · 2(log log x+3)/2 23/2 log log x
in other words R−1 = o(1). Next considering the second term in (5.20), one has
1 2p−1 1 1 1 2p−1 1
− 2(2P −
2p x 2(2P −2) N R 2P −2 ≤ 2(log log x+3)/2 x 2(2P −2) (x 2p−3 − 1) 2(2P −2) x (2p−3)(2P −2)(2P −1)
−2)

  1 2p − 1 1 
≪ (log x)log 2/2 exp log x − +
2(2P − 2) 2(2p − 3)(2P − 2) (2p − 3)(2P − 2)(2P − 1)
 log x log 2 
= exp − + log log x
(2p − 3)(2P − 1) 2
 (log x)1−log 2/2 log 2 
≤ exp − 3/2 + log log x = o(1).
2 log log x 2
For the third term in (5.20), one checks
 
1 2p−1
− 2(2P1−2) 2p−1
− P2 − 2(2P1−2) 1
− P2
x N log R ≤ x
2(2P −2) x log x (2p−3)(2P −1)
2p−3 2(2P −2)

log x   1 2p − 1 2 
= exp − log x − +
(2p − 3)(2P − 1) 2(2P − 2) 2(2p − 3)(2P − 2) (2p − 3)P
  3P − 4  
≤ exp − log x + log log x
P (2P − 2)(2p − 3)
 2 log x 
≤ exp − + log log x
(2p − 3)(2P − 2)
 (log x)1−log 2/2 
≤ exp − 1/2 + log log x = o(1).
2 log log x
 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 17

For the final term in (5.20), we note

1 1 p log x  log x 
pN −1 log R ≤ p(x 2p−3 − 1)−1 log x (2p−3)(2P −1) ≪


exp −
(2p − 3)(2P − 1) 2p − 3
 log x 
≪ exp − + log log x = o(1).
log log x
So D(N ) = o(N ) for the N given in (5.21). Thus (5.12) shows #Z(N ) ∼ N/20 for this N ,
which (from (5.21)) is easily seen to satisfy N ≤ exp(log x/(log log x − 2)). Thus from (5.14),
we conclude that for this choice of N ,
X Ω(n, x)2  log x   log x 
32πx1/2 ≫ x 1/2 −1/2
N ≫ x 1/2
exp − ≫ x 1/2
exp − .
n≥1
n3/2 2(log log x − 2) log log x


5.2. The Off-diagonal Terms. Our goal is to prove the following off-diagonal bound.
Lemma 5.3 (Bound for (OD)). Let 0 < ǫ < 1/1000, and assume x1/2 ≤ ∆ ≤ x2/3 k1/3−ǫ .
Provided k2 /(8π 2 ) ≤ x ≤ k4 , we have
(OD) ≪ k−8/3 ∆2 + k−3/2+ǫ ∆3/2 + x−1/2 k1/6+2ǫ ∆ + x−3/2 k−5/6+4ǫ ∆3 .
We now fix 0 < ǫ < 1/1000 for the remainder of this section (§5.2). Recall (5.3), which states
(5.22)
X  nx   mx  X  4π √mn 
3 k/2 2 −1
(OD) = 8π (−1) x w̃ 2 w̃ 2 c S(m, n; c)Jk−1 .
k + ∆2 k + ∆2 c
2 ǫ
m,n≤∆ k /x c≥1

Throughout this section, we will always assume x ≥ k2 /(8π 2 ) and x1−ǫ ≥ ∆ ≥ x1/2 . Moreover,
considering the ranges of m and c in (5.22) above, we will also frequently assume c ≥ 1 and
m ≤ ∆2 kǫ /x. Several lemmas appearing in this section will depend on m and ∆, and these will
hold uniformly for m and ∆ in these ranges (and for x ≥ k 2 /(8π 2 )), unless otherwise stated.
As discussed in the introduction to Section 5, our strategy is roughly to interchange the order
of summation in (5.22) and apply Poisson summation to the n-sum. This will capture some
cancellation, coming from the Kloosterman sums, the oscillations of the smoothings
√ w̃, and the
(in certain ranges of n, m and c) oscillations of the Bessel function Jk−1 (4π mn/c).
In the following lemma, we perform some initial simplifications. We interchange the order of
summation in (5.22) and smooth the resulting sums (with a view to applying Poisson summation
later on).
Lemma 5.6. Let g be a smooth, compactly supported function satisfying the following:
• supp(g) ⊂ [k/10, 10∆2 kǫ /x],
• g(ξ) = 1 for k/2 ≤ ξ ≤ 2∆2 kǫ /x,
• for all integers j ≥ 0 and all ξ, we have g (j) (ξ) ≪j ξ −j .
We then have1
X X  mx 
(OD) = 8π 3 (−1)k/2 x2 c−1 w̃ 2
k + ∆2
c≤100∆2 /(xk 1−ǫ ) m≤∆2 k ǫ /x
X  nx   4π √mn   4π √mn 
S(m, n; c)w̃ 2 g Jk−1 + O(k−1000 ).
k + ∆2 c c
n∈Z

Proof. We first truncate the inner sum over c in (5.22) at 100∆2 /(xk 1−ǫ ). If c ≥ 100∆2 /(xk1−ǫ )
then for any m, n ≤ ∆2 kǫ /x one has
√  4π √mn 
4π mn 4π∆2 kǫ 4πk k mn
≤ ≤ ≤ =⇒ Jk−1 ≪ 2 e−14k/13 ,
c cx 100 4 c c
1For notational convenience we also implicitly set g(4π √mn/c) = 0 for any n < 0.
18 NED CARMICHAEL

by the bound given in (2.2) of Lemma 2.1. Using the trivial bound |S(m, n; c)| ≤ c and the
bound (3.2) of Lemma 3.1 for the weights w̃, the contribution of c ≥ 100∆2 /(xk 1−ǫ ) to (5.22)
is therefore seen to be
X X
≪ x1/2 (mn)1/4 c−2 e−14k/13 ≪ e−k .
m,n≤∆2 k ǫ /x c≥100∆2 k ǫ /x

Truncating the sum over c and interchanging the order of summation, we thus rewrite (5.22) as

X X  mx 
(5.23) (OD) = 8π 3 (−1)k/2 x2 c−1 w̃
k 2 + ∆2
c≤100∆2 /(xk 1−ǫ ) m≤∆2 k ǫ /x
X nx 
  4π √mn 
S(m, n; c)w̃ 2 Jk−1 + O(e−k ).
k + ∆2 c
n≤∆2 k ǫ /x

The lemma now follows from (5.23). In other words, introducing g to the√ sums (5.23) produces
an error that is ≪ k −1000 . Indeed, the contribution of the terms with 4π mn/c < k/2 to (5.23)
√ 1/2
is negligible: for these values of m, n and c the bound (2.3) shows Jk−1 (4π mn/c) ≪ e−k ,
say, so their contribution is clearly ≪ k−1000 . On the other hand, since m ≤ ∆2 kǫ /x,
√
4π mn 2∆2 kǫ n 1  ∆2 kǫ 2 1 ∆2 k ǫ nx kǫ
> =⇒ n ≥ 2 ≥ 2 ≥ 2 =⇒ 2 ≥ ,
c x c 4π m x 4π x k + ∆2 4π 2 (8π 2 + 1)

since we assume k2 ≤ 8π 2 x ≤ 8π 2 ∆2 =⇒ k2 + ∆2 ≤ (8π 2 + 1)∆2 . The bound (2.12) of Lemma

2.4 now shows
√
4π mn 2∆2 kǫ  nx 
> =⇒ w̃ 2 ≪ n−2 k−1100 .
c x k + ∆2

Thus the contribution of these terms to (5.23) is indeed ≪ k−1000 . 

Our idea is now to replace w̃ by the explicit expression given in Lemma 3.1. This leads to
the following lemma, which is the key starting point for our proof of Lemma 5.3.

Lemma 5.7. We have the bound

X X
(OD) ≪ x5/4 c−1 m−3/4 max |S1 | + max |S2 | + x−5/4 k−4/3+2ǫ ∆5/2 ,


t∈[1,2] t∈[1,2]
c≤100∆2 /(xk 1−ǫ ) m≤∆2 k ǫ /x

where S1 = S1 (t; m, c) is given by

X X √  4π √mn   4π √mn   an 
(5.24) S1 = (16π 2 nxt − κ2 )−3/4 sin ω(4π nxt)g Jk−1 e ,
c c c
1≤a≤c n∈Z
(a,c)=1

and S2 = S2 (t; m, c) is given by2

(5.25)
X X √  4π √mn   4π √mn   an 
2 2 −7/4
S2 = x n(16π nxt − κ ) sin ω(4π nxt)g Jk−1 e .
c c c
1≤a≤c n∈Z
(a,c)=1

2The factor of x in the definition of S is included for convenience, so that S and S are of comparable sizes.
2 1 2
 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 19

Proof. Replacing w̃(nx/(k2 + ∆2 )) by the expression given in Lemma 3.1 in Lemma 5.6 and
interchanging the order of summation and integration shows
X X  mx  Z ∞ n
9/2 5/2 k/2 2 −1
(5.26) (OD) = 2 π (−1) x c w̃ 2 2
12π 2 xtw(t)
k + ∆ 0
c≤100∆2 /(xk 1−ǫ ) m≤∆2 k ǫ /x
√  4π √mn   4π √mn 
1 1
X
2 2 −7/4
S(m, n1 ; c)n1 (16π n1 xt − κ ) sin ω(4π n1 xt)g Jk−1 − (w(t)
c c
n1 ∈Z
√  4π √mn   4π √mn o
2 2
X
′ 2 2 −3/4
+ tw (t)) S(m, n2 ; c)(16π n2 xt − κ ) sin ω(4π n2 xt)g Jk−1 dt
c c
n2 ∈Z
 X X  mx 
+ O x2 c−1 w̃ 2
k + ∆2
c≤100∆2 /(xk 1−ǫ ) m≤∆2 k ǫ /x
X  4π √mn   4π √mn  
−5/4
(nx) S(m, n; c)g Jk−1 + O(k−1000 ).
c c
n≥1
To bound the error term above, we use the following bounds. From Lemma 3.1, we have
|w̃(mx/(k2 + ∆2 ))| ≪ (mx)−3/4 . Trivially |S(m, n; c)| ≤ c. By the construction of g, we have
|g(ξ)| ≪ 1 for all ξ. Finally, we use the Bessel function bound (2.4) of Lemma 2.1, which gives
√
|Jk−1 (4π mn/c)| ≪ k−1/3 . These bounds show that the error in (5.26) is
X X X
≪ k−1/3 m−3/4 n−5/4 ≪ x−5/4 k−4/3+2ǫ ∆5/2 .
c≤100∆2 /(xk 1−ǫ ) m≤∆2 k ǫ /x n≥1

Opening the Kloosterman sums in (5.26) and interchanging the order of summation, we now
obtain
X X  mx 
(5.27) (OD) = 29/2 π 5/2 (−1)k/2 x2 c−1 w̃ 2
k + ∆2
c≤100∆2 /(xk 1−ǫ ) m≤∆2 k ǫ /x
Z ∞n X  a∗ m  X √  4π √mn 
2 1 2 2 −7/4 1
12π xtw(t) e n1 (16π n1 xt − κ ) sin ω(4π n1 xt)g
0 c c
1≤a1 ≤c n1 ∈Z
(a1 ,c)=1
 4π √mn   a n  X  a∗ m  X
1 1 1
Jk−1 e − (w(t) + tw′ (t)) e 2 (16π 2 n2 xt − κ2 )−3/4
c c c
1≤a2 ≤c n2 ∈Z
(a2 ,c)=1
√  4π √mn   4π √mn   a n o
2 2 2 2
sin ω(4π n2 xt)g Jk−1 e dt + O(x−5/4 k−4/3+2ǫ ∆5/2 ).
c c c
Using the bound w̃(mx/(k2 + ∆2 )) ≪ (mx)−3/4 (which is (3.2) of Lemma 3.1), the triangle
inequality shows
X X
(5.28) (OD) ≪ x5/4 c−1 m−3/4
c≤100∆2 /(xk 1−ǫ ) m≤∆2 k ǫ /x
Z ∞
|tw(t)||S2 (t; m, c)| + |w(t) + tw′ (t)||S1 (t; m, c)| dt + x−5/4 k−4/3+2ǫ ∆5/2
0
X X  Z ∞
5/4 −1 −3/4
≪x c m max |S2 | |tw(t)|dt
t∈[1,2] 0
c≤100∆2 /(xk 1−ǫ ) m≤∆2 k ǫ /x
Z ∞ 
+ max |S1 | |w(t) + tw′ (t)|dt + x−5/4 k−4/3+2ǫ ∆5/2 .
t∈[1,2] 0
In the last line, we observed that the range of integration is t ∈ supp w = [1, 2]. Recall also
(from Definition 2.3) that w ≪ 1, supp w′ = [1, 1 + ∆−1 ] ∪ [2 − ∆−2 , 2] and w′ ≪ ∆. It follows
Z ∞ Z ∞
|tw(t)|dt ≪ 1 and |w(t) + tw′ (t)|dt ≪ 1.
0 0
20 NED CARMICHAEL

The lemma now follows from (5.28). 

Remark. Considering t ∈ [1, 2] in Lemma 5.7, from now on we implicitly assume t ∈ [1, 2]
wherever this variable t appears. Indeed, the following lemmas hold uniformly for t ∈ [1, 2].

We wish to bound the sums S1 and S2 . To do so, we first apply Poisson summation, which
gives the following.

Lemma 5.8 (Poisson Summation). For i = 1, 2 we have the bounds

X
Si ≪ c1/2 m−1/4 |Ii (n)|.
n≥0
(n,c)=1

Here
Z ∞  c√xt   cn 
2
Ii (n) = Gi (y) sin ω √ y e y dy,
0 m 16π 2 m

where
 c2 xt −3/4
(5.29) G1 (y) = c3/2 m−3/4 y y 2 − κ2 g(y)Jκ (y),
m

and
 c2 xt −7/4
(5.30) G2 (y) = c7/2 m−7/4 xy 3 y 2 − κ2 g(y)Jκ (y).
m
Proof. We consider only S1 here, since S2 may be handled in exactly the same way. Splitting
into progressions, write

X X √  4π √mn   4π √mn   an 
2 2 −3/4
(5.31) S1 = (16π nxt − κ ) sin ω(4π nxt)g Jk−1 e
c c c
1≤a≤c n∈Z
(a,c)=1
X X  ab  X
(16π 2 (b + lc)xt − κ2 )−3/4 sin ω(4π (b + lc)xt)
p
= e
c
1≤a≤c b (mod c) l∈Z
(a,c)=1
 4π pm(b + lc)   4π pm(b + lc) 
g Jk−1 .
c c

Applying Poisson summation to the inner sum over l, we find

 4π p p
X m(b + lc)   4π m(b + lc) 
(16π 2 (b + lc)xt − κ2 )−3/4 sin ω(4π (b + lc)xt)g
p
Jk−1
c c
l∈Z
XZ ∞
(16π 2 (b + vc)xt − κ2 )−3/4 sin ω(4π
p
= (b + vc)xt)
0
l̃∈Z
 4π pm(b + vc)   4π pm(b + vc) 
g Jk−1 e(−˜lv)dv
c c
XZ ∞ √  4π √mu   4π √mu   ˜l(u − b)  du
2 2 −3/4
= (16π uxt − κ ) sin ω(4π uxt)g Jk−1 e − .
0 c c c c
l̃∈Z
 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 21

(In the last line, we have set u = b + vc.) Consequently, we obtain from (5.31)
(5.32)
X XZ ∞  4π √mu 
√  4π √mu 
2 2 −3/4
S1 = (16π uxt − κ ) sin ω(4π uxt)g Jk−1
0 c c
1≤a≤c l̃∈Z
(a,c)=1
 ˜lu n 1
 b(˜l + a) o
X
e − e du
c c c
b (mod c)
X X Z ∞ √  4π √mu 
2 2 −3/4
= (16π uxt − κ ) sin ω(4π uxt)g
0 c
1≤a≤c l̃∈Z
(a,c)=1 l̃≡−a (mod c)
 4π √mu   ˜lu 
Jk−1 e − du
Z ∞ √ c √ c
X √  4π mu   4π mu   nu 
≤ (16π 2 uxt − κ2 )−3/4 sin ω(4π uxt)g Jk−1 e − du .
0 c c c
n∈Z
(n,c)=1
√
The last line is a consequence of the triangle inequality. Setting y = 4π mu/c, we find
Z ∞
√  4π √mu   4π √mu   nu 
2 2 −3/4
(16π uxt − κ ) sin ω(4π uxt)g Jk−1 e − du
0 c c c
c1/2
= I1 (−n).
8π 2 m1/4
Noting I1 (−n) = I1 (n), the result now follows from (5.32). 
Remark. Since the integrals Ii (n) are taken over y ∈ supp g = [k/10, 10∆2 kǫ /x], from now on
we will assume the k/10 ≤ y ≤ 10∆2 kǫ /x wherever this variable appears.
Our goal is now to bound the integrals Ii (n) appearing in Lemma 5.8 via the method of
stationary phase. Before proceeding further, in the following lemma we first handle some very
simple cases where Ii (n) is negligibly small. After some minor aesthetic rearrangements, we
thus truncate the sum over n in Lemma 5.8.
Lemma 5.9. Assume ∆ ≤ x2/3 k1/3−ǫ . For i = 1, 2 we have
X
(5.33) Si ≪ c1/2 m−1/4 |Ii′ (n)| + k−1000 ,
√
m1/2 x3/2
≤n≤ 1000k mx
10∆2 kǫ

where
Z ∞
(5.34) Ii′ (n) = Gi (y)eiF (y) dy.
0
Here G1 and G2 are the functions given in Lemma 5.8, and
cn 2  c√xt 
(5.35) F (y) = y −ω √ y .
8πm m
Proof. Starting from Lemma 5.8, we first note
Z ∞  c√xt   cn 
2
(5.36) Ii (n) = Gi (y) sin ω √ y e y dy
0 m 16π 2 m
1 ∞
Z n  cn  c√xt o
2
= Gi (y) exp i y +ω √ y dy
2i 0 8πm m
1 ∞
Z n  cn  c√xt o
2
− Gi (y) exp i y −ω √ y dy.
2i 0 8πm m
22 NED CARMICHAEL

Using Lemma 2.6, we will show that the integrals above contribute ≪ k−1000 to Si in many
cases. One computes (using (2.7))
d n cn 2  c√xt o cn  c2 xt κ2 1/2
(5.37) y ±ω √ y = y± − 2 ,
dy 8πm m 4πm m y
and (using (2.8))
d2 n cn 2  c√xt o cn κ2  c2 xt κ2 −1/2
(5.38) y ± ω √ y = ± − 2 .
dy 2 8πm m 4πm y 3 m y
Recall we assume ∆ ≤ x2/3 k1/3−ǫ . From this, using also the standing assumptions that x ≥
k2 /(8π 2 ), c ≥ 1, 1 ≤ t ≤ 2, m ≤ ∆2 kǫ /x and k/10 ≤ y ≤ 10∆2 kǫ /x, one has
(5.39)
c2 xt x2 κ2  c2 xt κ2  c2 xt c2 x
≫ 2 ǫ ≫ x2/3 k−2/3+ǫ ≫ k2/3 , say, and 2 ≪ 1 =⇒ − 2 ∼ ≍ .
m ∆ k y m y m m
In particular (using m ≤ ∆2 kǫ /x, y ≥ k/10 and c ≥ 1),
κ2  c2 xt κ2 −1/2 k2 m1/2 ∆
− ≍ ≪ .
y3 m y2 3
y cx 1/2 xk 1−ǫ/2

Now the assumption ∆ ≤ x2/3 k1/3−ǫ (and m ≤ ∆2 kǫ /x) guarantees

∆ x −3ǫ/2 1 −3ǫ/2
1
≤ · k ≪ · k = o .
xk1−ǫ/2 ∆2 k ǫ m m
Combining the two calculations above, we have shown (provided n 6= 0)
√
d2 n  c xt o κ2  c2 xt κ2 −1/2 k2 m1/2  cn 
(5.40) ω √ y = − ≍ = o .
dy 2 m y3 m y2 y 3 cx1/2 4πm
Consequently, for n 6= 0 we have established the second derivative estimate
d2 n cn 2  c√xt o cn
(5.41) y ±ω √ y ≍ .
dy 2 8πm m m
We now claim that for n 6= 0 and k/10 ≤ y ≤ 10∆2 kǫ /x in the region of integration, we have
dj n cn 2  c√xt o cn 2−j cn 2−j
(5.42) j
y ±ω √ y ≪j y ≪j k for j = 2, 3, . . . .
dy 8πm m m m
Indeed, the j = 2 case is included in (5.41). For y ≥ k/10, one has (using (5.39)) that for any
a > 0 and an integer j ≥ 1,
dj n c2 xt κ2 −a o  c2 xt −a−1 k2
(5.43) − ≪ j .
dy j m y2 m y 3+j
Thus, upon each differentiation of
√
d2 n  c xt o κ2  c2 xt κ2 −1/2
ω √ y = 3 − 2 ,
dy 2 m y m y
one saves at least y. That is to say, for j > 2
dj n cn 2  c√xt o √
dj n  c xt o
√
2 n  c xt o
−(j−2) d cn 2−j
j
y ±ω √ y =± j ω √ y ≪j y 2
ω √ y ≪j y ,
dy 8πm m dy m dy m m
(using (5.40)) as claimed.
Finally, we must bound the derivatives of G1 and G2 . Recall (from (5.29) and (5.30))
 c2 xt κ2 −3/4
G1 (y) = c3/2 m−3/4 y −1/2 − 2 g(y)Jκ (y),
m y
and
 c2 xt κ2 −7/4
G2 (y) = c7/2 m−7/4 xy −1/2 − 2 g(y)Jκ (y).
m y
 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 23

Recall also g(j) (y) ≪j y −j (by our construction, see Lemma 5.6). For any ν ≥ 0, one has
1
(5.44) Jν′ (y) = (Jν−1 (y) − Jν+1 (y)).
2
This is easily checked from the definition (2.1) (see also [19, §3.2]). Using the derivative bound
(5.43), and using (5.44) to differentiate the Bessel functions (≤ j times), we obtain the bound
(j)
(5.45) Gi (y) ≪j x−3/4 y −1/2 (y −j |Jκ (y)| + |Jκ+Oj (1) (y)|) ≪j x−3/4 k−5/6 for j = 0, 1, 2 . . .
Here we also used the bound Jκ+O(1) (y) ≪ k−1/3 given in (2.4) (and our assumption y ≥ k/10).
Using these calculations, we now apply Lemma 2.6 to show that the integrals appearing in
(5.36) are negligible (i.e. contribute ≪ k−1000 ) in the following four simple cases.
(i) Firstly, consider the term Ii (0) (which appears in Lemma 5.8 only if c = 1). Here, it
follows from (5.39) that the phase satisfies
dn  √xt o  xt κ2 1/2  x 1/2
(5.46) ±ω √ y =± − 2 ≫ .
dy m m y m
Turning our attention to the second derivative, the c = 1 case of (5.40) gives that for y ≥ k/10
d2 n  √xt o κ2 m1/2 1  m 1/2
± ω √ y ≍ ≪ .
dy 2 m y 3 x1/2 k x
Differentiating further (cf. (5.42)) and using the above we have for y ≥ k/10
dj n  √xt o 2 n  √xt o
−(j−2) d 1−j m
 1/2
(5.47) ± ω √ y ≪ j y ± ω √ y ≪ j k for j = 2, 3 . . .
dy j m dy 2 m x
Consequently, using the estimates (5.45), (5.46) and (5.47), we may apply Lemma 2.6 with
α = k/10, β = ∆2 kǫ /x, X = x−3/4 k−5/6 , U = 1, R = (x/m)1/2 , Y = k(m/x)1/2 and Q = k.
This gives the bound
n x 3/4 −B  x −B/2 o
(5.48) Ii (0) ≪B ∆2 x−7/4 k−5/6+ǫ k1/2 + ,
m m
valid for any integer B ≥ 0. Together with m ≤ ∆2 kǫ /x, our assumption ∆ ≤ x2/3 k1/3−ǫ implies
1 x k2 x  x 2/3
(5.49) ≥ 2 ǫ ≥ x−1/3 k−2/3+ǫ . In particular, x ≥ 2 =⇒ ≥ kǫ ≥ k2/3 .
m ∆ k 8π m k
So taking B large enough in (5.48), we easily obtain Ii (0) ≪ k−1000 .
(ii) Next, we consider the former integral appearing in (5.36):
Z ∞ n  cn  c√xt o
2
Gi (y) exp i y +ω √ y dy.
0 8πm m
In this case one has for y ≥ k/10 (see (5.37) and (5.39)),
d n cn 2  c√xt o cn  c2 xt κ2 1/2 cnk cx1/2
(5.50) y +ω √ y = y+ − 2 ≫ + 1/2 .
dy 8πm m 4πm m y m m
Equipped with (5.42), (5.45) and (5.50),pwe apply Lemma 2.6 with α = k/10, β = ∆2 kǫ /x,
X = x−3/4 k−5/6 , U = 1, R = cnk/m + c x/m, Y = cnk 2 /m and Q = k. This shows
Z ∞ n  cn  c√xt o
2
(5.51) Gi (y) exp i y +ω √ y dy
0 8πm m
n cnk c√x −B  √cnk √cx −B o
2 −7/4 −5/6+ǫ
≪B ∆ x k + √ + √ + √ ,
m m m n
valid for any integer B ≥ 0. We will use the bound
 cnk c√x −B  √cnk √cx −B  nk −2  x −(B−2)/2  nk2 −2  nk2 x −(B−4)/2
+√ + √ + √ ≤ + + .
m m m n m m m m n
24 NED CARMICHAEL

For any n, we have

nk2 x kx1/2
+ ≥ 2 1/2 ≥ k4/3 ,
m n m
since x/m ≥ k 2/3 by (5.49). So we obtain
 cnk c√x −B  √cnk √cx −B  m2 m2 
+√ + √ + √ ≤ n−2 2 · k−(B−2)/3 + 4 k−2(B−4)/3 .
m m m n k k
Thus taking B large enough in (5.51) shows
Z ∞ n  cn  c√xt o
2
Gi (y) exp i y +ω √ y dy ≪ n−2 k−1100 ,
0 8πm m
say, and consequently these integrals contribute ≪ k−1000 to Si .
(iii) Finally, we consider the latter integral in (5.36):
Z ∞ n  cn  c√xt o
2
Gi (y) exp i y −ω √ y dy := Ii′ (n).
0 8πm m
This is more difficult than the previous case, since now the derivative of the oscillatory phase
F (y) in Ii′ (n) can vanish:
d n cn 2  c√xt o cn  c2 xt κ2 1/2
′
F (y) = y −ω √ y = y− − 2 .
dy 8πm m 4πm m y
√
But clearly if n is large enough, F ′ (y) must be also be large. Indeed, suppose n ≥ 1000 mx/k.
Then, for 1 ≤ t ≤ 2 and y ≥ k/10 (in the range of integration) one has
 c2 xt κ2 1/2 √ √ 
′ cn cnk c xt cnk  1 2 cnk
F (y) = y− − 2 ≥ − √ ≥ − ≫ .
4πm m y 40πm m m 40π 1000 m
Using this estimate (and also (5.42) and (5.45)), we may apply Lemma 2.6 with α = k/10,
β = ∆2 kǫ /x, X = x−3/4 k−5/6 , U = 1, R = cnk/m, Y = cnk 2 /m and Q = k, we obtain
Z ∞ n  cn  c√xt o
′ 2
(5.52) Ii (n) = Gi (y) exp i y −ω √ y dy
0 8πm m
n cnk 2 −B/2  cnk −B o
≪B ∆2 x−7/4 k−5/6+ǫ + ,
m m
√
valid for any integer B ≥ 0. Using (5.49) (which states x/m ≥ k2/3 ) one has for n ≥ 1000 mx/k
that cnk 2 /m ≥ cnk/m ≫ cx1/2 /m1/2 ≫ k1/3 . Taking B large √ enough, we concludes from (5.52)
that Ii′ (n) ≪ n−2 k−1100 . Thus these Ii′ (n) with n ≥ 1000 mx/k contribute ≪ k−1000 to Si .
(iv) Finally, we show that the integrals Ii′ (n) are also negligible when n is very small. Indeed,
if n ≤ m1/2 x3/2 /(10∆2 kǫ ) then using (5.39) and our assumption y ≤ 10∆2 kǫ /x we obtain
 c2 xt κ2 1/2 cn 1 cx1/2 t1/2 cn 10∆2 kǫ cx1/2  1 1   x 1/2
|F ′ (y)| ≥ − 2 − y ≥ − ≥ − ≫ c .
m y 4πm 2 m1/2 4πm x m1/2 2 4π m
Using this estimate together with (5.42) and (5.45) as before,pwe are able to apply Lemma 2.6
with α = k/10, β = ∆2 kǫ /x, X = x−3/4 k−5/6 , U = 1, R = c x/m, Y = cnk2 /m and Q = k.
This provides the bound
n cx −B/2  c2 x −B/2 o
(5.53) Ii′ (n) ≪B ∆2 x−7/4 k−5/6+ǫ + ,
n m
valid for any integer B ≥ 0. Since m ≤ ∆2 kǫ /x, we have
m1/2 x3/2 cx 10∆2 kǫ
n≤ =⇒ ≥ ≥ 10∆kǫ/2 .
10∆2 kǫ n m1/2 x1/2
Because c2 x/m ≥ k2/3 is also large (see (5.49)), taking B sufficiently large in (5.53) shows that
those Ii′ (n) with n ≤ m1/2 x3/2 /(10∆2 kǫ ) contribute ≪ k−1000 to Si .

 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 25

Remark. We record the following useful facts from the above proof. Firstly, we established in
(5.45) that G1 and G2 satisfy
(j)
(5.54) Gi (y) ≪j x−3/4 k−5/6 for j = 0, 1, 2 . . .
Considering the phase F (y), we showed (see (5.37) and (5.38))
cn  c2 xt κ2 1/2
(5.55) F ′ (y) = y− − 2 ,
4πm m y
and
cn κ2  c2 xt κ2 −1/2
(5.56) F ′′ (y) = − 3 − 2 .
4πm y m y
Finally, we showed (see (5.41) and (5.42)) that under the condition ∆ ≤ x2/3 k1/3−ǫ , for k/10 ≤
y ≤ ∆2 kǫ /x in the region of integration
cn cn 2−j
(5.57) F ′′ (y) ≍ , and F (j) (y) ≪j y for j = 2, 3, . . . .
m m
√
It remains to bound the contribution of the Ii′ (n) with m1/2 x3/2 /(10∆2 kǫ ) ≤ n ≤ 1000 mx/k.
These integrals are not negligible, since in this case a larger contribution appears from the sta-
tionary phase. The analysis is therefore more involved.
Let y0 = y0 (n) be the stationary phase of F in the region of integration, i.e. y0 satisfies
F ′ (y0 ) = 0 and y0 ∈ [k/10, 10∆2 kǫ /x]. From (5.55), one finds
1 n 16π 2 mxt  16π 2 mxt 2 64π 2 m2 κ2 1/2 o 8π 2 mxt   κ2 n2 1/2 
y02 = ± − = 1 ± 1 − .
2 n2 n2 c2 n2 n2 4π 2 c2 x2 t2
√
Since m ≤ ∆2 kǫ /x and 1 ≤ t ≤ 2, if n ≤ 1000 mx/k then
κ2 n2 106 m ∆2 k ǫ
≤ ≪ = o(1),
4π 2 c2 x2 t2 4π 2 c2 xt2 x2
under the standing assumption ∆ ≤ x1−ǫ , say. So a Taylor expansion shows
 κ2 n2 1/2  k 2 n2  16π 2 mxt  k2 m   k2 m 
1− 2 2 2 2 = 1 + O 2 2 =⇒ y02 = + O , or y 2
0 = O .
4π c x t c x n2 c2 x c2 x
But using m ≤ ∆2 kǫ /x and ∆ ≤ x1−ǫ , we observe
k2 m ∆2 k2+ǫ 2+ǫ −2ǫ
 k 2
≤ ≤ k x < .
c2 x x2 10
So y02 = O(k2 m/(c2 x)) is impossible, since we assume y0 ≥ k/10 (so that y0 lies in the region
of integration). It follows
√
2 16π 2 mxt  k2 m  4π mxt  k2 nm1/2 
y0 = + O ⇐⇒ y 0 = + O .
n2 c2 x n c2 x3/2
√
We bound the error term above by using that m ≤ ∆2 kǫ /x, and assuming n ≤ 1000 mx/k.
From this, we conclude
√  k1+ǫ ∆2 
4π mxt
(5.58) y0 = +O .
n x2
We will show that the only significant contribution to the integrals Ii′ (n) comes from a neigh-
bourhood of y0 . To this end, we now introduce a smooth partition of unity in the following
lemma. This result is standard, and similar to (for example) [5, Lemme 2].
Lemma 5.10 (A Smooth Partition of Unity). Let α ∈ R and L > 0. There exists a sequence
of real-valued smooth functions (bL,α
l )l∈Z satisfying the following.
P L,α
(i) For any ξ, l∈Z bl (ξ) = 1.
(ii) For any l ≥ 1 we have supp bL,α l = [α + 2l−1 L, α + 2l+1 L], and for any l ≤ −1 we
L,α
have supp bl = [α − 2|l|+1 L, α − 2|l|−1 L]. Finally, for l = 0 we have supp bL,α
0 =
[α − 2L, α + 2L].
26 NED CARMICHAEL

dj L,α
(iii) For a non-negative integer j and any l ∈ Z, we have b (ξ)
dξ j l
≪j 2−j|l| L−j .

Proof. One can construct a smooth transition function h : [0, 1] → R satisfying h(0) = 0,
h(1) = 1, and limξ→0+ h(j) (ξ) = limξ→1− h(j) (ξ) = 0 for any j ≥ 0. Equipped with such a h,
define 
0

   if ξ ≤ α − 2L
 α−ξ
1 − h L − 1 if α − 2L ≤ ξ ≤ α − L,



L,α
b0 (ξ) = 1
   if α − L ≤ ξ ≤ α + L,
ξ−α
1−h L −1 if α + L ≤ ξ ≤ α + 2L,






0 if α + 2L ≤ ξ.


For l ≥ 1 we define
if ξ ≤ α + 2l−1 L,


 0 
h ξ−α if α + 2l−1 L ≤ ξ ≤ α + 2l L,

− 1

2l−1
L
bL,α
l (ξ) =  ξ−α


 1 − h l
2L
− 1 if α + 2l L ≤ ξ ≤ α + 2l+1 L,

0 if α + 2l+1 L ≤ ξ.


Finally, for l ≤ −1 we define

if ξ ≤ α − 2|l|+1 L,


 0  
1 − h α−ξ if α − 2|l|+1 L ≤ ξ ≤ α − 2|l| L,

− 1

|l|
bL,α L,α
l (ξ) = b−l (2α − ξ) = 

α−ξ
2 L 

h 2|l|−1 L − 1
 if α − 2|l| L ≤ ξ ≤ α − 2|l|−1 L,

0 if α − 2|l|−1 L ≤ ξ.


It is easy to check that the properties (i) and (iii) hold for this choice of (bL,α
l )l∈Z . 
We use this construction to split the range of integration in Ii′ (n) into intervals surrounding
the stationary point y0 . We now set L = L(n) = k ǫ max{1, m/(cn)}, and write
Z ∞X  Z y0 +2L
L,y0
′
(5.59) Ii (n) = bl (y) Gi (y)e iF (y)
dy = bL,y
0
0
(y)Gi (y)eiF (y) dy
0 l∈Z y0 −2L

XZ y0 +2l+1 L XZ y0 −2|l|−1 L
+ bL,y
l
0
(y)Gi (y)eiF (y) dy + bL,y
l
0
(y)Gi (y)eiF (y) dy.
l≥1 y0 +2l−1 L l≤−1 y0 −2|l|+1 L

In the following lemma, we show that all but the l = 0 term is negligible.
Lemma 5.11. Let L be as above, and assume ∆ ≤ x2/3 k1/3−ǫ . Then for i = 1, 2 we have
Z y0 +2L
′
Ii (n) = bL,y
0
0
(y)Gi (y)eiF (y) dy + O(k−1000 ).
y0 −2L

Proof. For i = 1, 2, we apply Lemma 2.6 to bound the l 6= 0 terms appearing in (5.59), which
are: Z ∞
bL,y
l
0
(y)Gi (y)eiF (y) dy.
0
dj L,y0
From the estimate b
dy j l
(y) ≪j 2−j|l| L−j and (5.54), we deduce

dj L,y0
(5.60) {b (y)Gi (y)} ≪j x−3/4 k−5/6 , for j = 0, 1, 2, . . .
dy j l
From (5.57), we also have
cn 2−j cn cn
(5.61) F (j) (y) ≪j y for j = 2, 3, . . . , and F ′′ (y) ≍ =⇒ F ′ (y) ≍ |y − y0 |.
m m m
 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 27

Equipped with (5.60) and (5.61), and noting that |y − y0 | ≍ 2|l| L for y ∈ supp bL,y
l
0
(provided
L,y0
l 6= 0), we are ready to apply Lemma 2.6. We take [α, β] to be the interval supp bl (so that
|l|
β − α ≍ 2 L), and take X = x −3/4 k −5/6 |l| 2
, U = 1, R = cn2 L/m, Y = cnα /m and Q = α
(where α ≍ β is the infimum of supp bL,y l
0
). This yields
Z ∞ n  cn −B/2  cn −B o
bL,y
l
0
(y)Gi (y)eiF (y) dy ≪B 2|l| Lx−3/4 k−5/6 2−|l|B L2 + L ,
0 m m
valid for any non-negative integer B. But with the above choice of L, one always has cnL2 /m >
cnL/m ≥ kǫ . So by taking B large enough, the contribution of the integrals with l 6= 0 to (5.59)
is seen to be ≪ k−1000 . 
Remark. From now on, we will denote b = bL,y
0
0
for simplicity.
Our final task is to bound the integrals appearing in Lemma 5.11. We consider three cases
separately, based on the value of n, or equivalently y0 (n). The first (and easiest) is the case
where the region of integration [y0 (n) − 2L, y0 (n) + 2L] lies in the region in which the Bessel
function Jk−1 (y) is negligibly small. Secondly, we consider the case in which y0 (n) is close to
the transition of the Bessel function. The final case is that in which the region of integration
lies within the oscillatory regime of the Bessel function.
First recall L = kǫ max{1, m/(cn)} ≤ kǫ (1 + m/(cn)). Note that for n ≥ m1/2 x3/2 /(10∆2 kǫ )
in the range of summation, using m ≤ ∆2 kǫ /x and assuming ∆ ≤ x2/3 k1/3−ǫ we have
 10m1/2 ∆2 kǫ   ∆3 k3ǫ/2 
(5.62) L = L(n) ≤ kǫ 1 + 3/2
≪ k ǫ
1 + 2
≪ k1−ǫ/2 .
cx cx
We also remark that assuming ∆ ≤ x2/3 k1/3−ǫ , the error term in (5.58) is k1+ǫ ∆2 /x2 ≤
√
(k2 /x)2/3 k1/3−ǫ ≪ k1/3−ǫ . Thus for ∆ ≤ x2/3 k1/3−ǫ and n ≤ 1000 mx/k, (5.58) shows
√ √
4π mxt 1/3−ǫ 4π mxt
(5.63) y0 = y0 (n) = + O(k )∼ ≫ k.
n n
In particular, combining (5.62) and (5.63) shows that for y0 − 2L ≤ y ≤ y0 + 2L in the region
of integration of Ii′ (n), one has y = y0 + O(L) ≍ y0 .
√
Lemma 5.12. Assume ∆ ≤ x2/3 k1/3−ǫ . For m1/2 x3/2 /(10∆2 kǫ ) ≤ n ≤ 1000 mx/k and
i = 1, 2 we have the following bounds.
ǫ
(i) If y0 (n) + 2L ≤ k − k 1/3+ǫ , then Ii′ (n) ≪ e−k .
(ii) If y0 (n) + 2L ≥ k − k 1/3+ǫ and y0 (n) − 10L ≤ k + k1/3+ǫ , then
Ii′ (n) ≪ x−3/4 k−5/6+ǫ + x−5/4 k1/6+ǫ c−1 m1/2 .
(iii) If y0 (n) − 10L ≥ k + k1/3+ǫ , then
 √  √
−9/8 −1/4 −1/2 1/8 1/2 4π mxt −5/8 −9/4 1/8 2 4π mxt
−1/4 −9/4
′
Ii (n) ≪ x k c m n −n +x k m n −n .
k k
Proof. Parts (i) and (ii) are relatively simple. For both of these, we use the trivial bound for
the integrals Ii′ (n) (given in Lemma 5.11), which shows
Z y0 +2L
′
(5.64) Ii (n) ≤ |b(y)Gi (y)|dy ≪ L max |Gi (y)|.
y0 −2L y∈[y0 −2L,y0 +2L]

It follows from the definitions (5.29) and (5.30) (cf. (5.54)) that for i = 1, 2 and y ∈ [y0 −
2L, y0 + 2L] (these y satisfy y ≍ y0 ≫ k, see (5.62) and (5.63)),
(5.65) Gi (y) ≪ x−3/4 k−1/2 |Jk−1 (y)|.
ǫ
For part (i), (2.3) shows Jk−1 (y) ≪ e−k for y ≤ y0 + 2L ≤ k − k1/3+ǫ . So we conclude from
(5.64) and (5.65) that
ǫ ǫ
Ii′ (n) ≪ Lx−3/4 k−1/2 e−k ≪ e−k ,
as claimed (since we assume x ≥ k2 /(8π 2 ) and L ≪ k1−ǫ/2 by (5.62)).
28 NED CARMICHAEL

For part (ii), we instead use the bound Jk−1 (y) ≪ k−1/3 given in (2.4) (which is sharp in the
transition regime). In this case, (5.64) and (5.65) therefore show
 m
(5.66) Ii′ (n) ≪ x−3/4 k−5/6 L ≪ x−3/4 k−5/6+ǫ 1 + .
cn
√
Using that y0 ≍ mx/n from (5.63) and L ≪ k 1−ǫ/2 = o(y0 ) from (5.62), we have
√
1/3+ǫ mx m km1/2
y0 − 10L ≤ k + k =⇒ ≪ k =⇒ ≪ .
n cn cx1/2
So the required bound follows from (5.66) and the preceding estimate.
Part (iii) corresponds to the case where the region of integration is entirely contained in the
oscillatory regime of the Bessel function Jk−1 (y), and is considerably more involved. We first
replace the Bessel functions by the asymptotic (2.5) of Lemma 2.1, which is available in this
regime. For y ≥ k + k1/3+ǫ , (2.5) states
r r
2 2 2 −1/4 2 2 1 5 κ2 
(5.67) Jk−1 (y) = (y − κ ) cos ω(y) + (y − κ2 )−3/4 + sin ω(y)
π π 8 24 y 2 − κ2
 y4 
+O
(y 2 − κ2 )13/4
(
1  1 5 2 2 
= √ (y 2 − κ2 )−1/4 1 + (y 2 − κ2 )−1/2 + κ (y − κ2 )−3/2 eiω(y)
2π 8i 24i
)
 1 2 2 −1/2 5 2 2 2 −3/2

−iω(y)
 y4 
+ 1 − (y − κ ) − κ (y − κ ) e +O .
8i 24i (y 2 − k2 )13/4
Replacing this in the expressions (5.29) and (5.30) for G1 and G2 (given in Lemma 5.8), we
have from Lemma 5.11 that
Z y0 +2L
′
(5.68) Ii (n) = b(y)Gi (y)eiF (y) dy + O(k−1000 )
y0 −2L
Z y0 +2L Z y0 +2L
i(F (y)+ω(y))
= Gi+ (y)e dy + Gi− (y)ei(F (y)−ω(y)) dy
y0 −2L y0 −2L
Z y0 +2L 
+O |b(y)g(y)x−3/4 y −1/2 | · y 4 (y 2 − k2 )−13/4 dy + O(k−1000 ),
y0 −2L

where the functions Gi± are given by

1  c2 xt −3/4
G1+ (y) = √ c3/2 m−3/4 b(y)g(y)y y 2 − κ2 (y 2 − κ2 )−1/4
2π m
 1 5 2 2 
1 + (y 2 − κ2 )−1/2 + κ (y − κ2 )−3/2 ,
8i 24i
1  2
c xt 2  −3/4
G1− (y) = √ c3/2 m−3/4 b(y)g(y)y y − κ2 (y 2 − κ2 )−1/4
2π m
 1 5 2 2 
1 − (y 2 − κ2 )−1/2 − κ (y − κ2 )−3/2 ,
8i 24i
1 7/2 −7/4  c2 xt −7/4
G2+ (y) = √ c m xb(y)g(y)y 3 y 2 − κ2 (y 2 − κ2 )−1/4
2π m
 1 5 2 2 
1 + (y 2 − κ2 )−1/2 + κ (y − κ2 )−3/2 ,
8i 24i
1 7/2 −7/4  c2 xt −7/4
G2− (y) = √ c m xb(y)g(y)y 3 y 2 − κ2 (y 2 − κ2 )−1/4
2π m
 1 5 2 2 
1 − (y 2 − κ2 )−1/2 − κ (y − κ2 )−3/2 .
8i 24i
 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 29

We now bound the error term in (5.68). Since the smoothing functions b and g are bounded
(by their construction), this is
(5.69)
Z y0 +2L Z y0 +2L
−3/4 −1/2 4 2 2 −13/4 −3/4
|b(y)g(y)x y |·y (y −k ) dy ≪ x y 7/2 (y+k)−13/4 (y−k)−13/4 dy
y0 −2L y0 −2L
Z y0 +2L
−3/4 7/2 −13/4
≪x (y0 + 2L) (y0 − 2L + k) (y − k)−13/4 dy
y0 −2L
1/4
≪ x−3/4 y0 (y0 − 2L − k)−9/4 .
In the last 1/2 3/2 2 ǫ
√ step, we used y0 ≍ y0 + 2L ≍ y0 − 2L + k for n in the range m x /(10∆ k ) ≤
n ≤ 1000 mx/k (see (5.62) and (5.63)). Additionally, using (5.63) we have
 4π √mxt  √
4π mxt
(y0 − 2L − k) − − k ≤ 2L + y0 − ≤ 2L + O(k1/3−ǫ ).
n n
√
Since for part (iii) we assume y0 −2L−k ≥ 8L+k1/3+ǫ , it follows (y0 −2L−k) ≍ (4π mxt/n−k).
Consequently (5.69) is
 √  √
−3/4 1/4 4π mxt −5/8 −9/4 1/8 2 4π mxt
−9/4 −9/4
≪x y0 −k ≪x k m n −n .
n k
√
(We also used y0 ≍ mx/n, see (5.63).) Replacing this error bound in (5.68), we have estab-
lished
Z y0 +2L Z y0 +2L
(5.70) Ii′ (n) = Gi+ (y)ei(F (y)+ω(y)) dy + Gi− (y)ei(F (y)−ω(y)) dy
y0 −2L y0 −2L
 √
−5/8 −9/4 1/8 2 4π mxt
 −9/4 
+O x k m n −n .
k
The final task is to bound the oscillatory integrals appearing in (5.70). This will be done using
Lemma 2.7. In order to apply this, we first require a bound for the derivatives of the functions
Gi± . This follows relatively straightforwardly from the (above) definitions. Indeed, firstly one
recalls that for any j ≥ 0, we have b(j) (y) ≪j L−j and g(j) (y) ≪j y −j (by construction).
Secondly, if y ∈ [y0 − 2L, y0 + 2L] then y ≍ y0 (see (5.62) and (5.63)). For these y, the bound
c2 xt/m ≥ x/m ≥ k2/3 (valid for m ≤ ∆2 kǫ /x, c ≥ 1, 1 ≤ t ≤ 2 and ∆ ≤ x2/3 k1/3−ǫ , see (5.49))
also shows (c2 xty 2 /m − κ2 ) ≍ c2 xy02 /m. For y ∈ [y0 − 2L, y0 + 2L] and j ≥ 0 one now computes
dj  c2 xt 2 2
−3/4  2
−j c xt 2 2
−3/4  c2 x −3/4
−3/2−j
y − κ ≪ j y y − κ ≪ j y0 ,
dy j m m m
dj  c2 xt 2 −7/4  c2 xt −7/4  c2 x −7/4
−7/2−j
and similarly j y − κ2 ≪j y −j y 2 − κ2 ≪j y0 .
dy m m m
Finally, for y ∈ [y0 − 2L, y0 + 2L] and j ≥ 0 we have
dj 2 2 −1/4
 y j
−1/4
(y − κ ) ≪ j (y 2 − κ2 )−1/4 ≪j y0 (y0 − k)−1/4−j .
dy j y 2 − κ2
In the last step, we used that y − κ ≍ y0 − k, which follows from our assumption y0 − k ≥
10L + k1/3+ǫ and the fact that |y − y0 | ≤ 2L in this range of y. The above calculations
show that upon differentiating Gi± , we save at least L each time (note L ≪ y0 − k by our
assumption y0 − k ≥ 10L + k1/3+ǫ ). In other words (assuming ∆ ≤ x2/3 k1/3−ǫ ) we obtain for
y ∈ [y0 − 2L, y0 + 2L] and integers j ≥ 0 the bound
dj −3/4
(5.71) Gi± (y) ≪j x−3/4 y0 (y0 − k)−1/4 L−j .
dy j
We next turn our attention to the new oscillatory phase. Set
F± (y) = F (y) ± ω(y).
30 NED CARMICHAEL

It turns out that despite the extra ω(y) term, F± behaves essentially the same as F on the
interval [y0 − 2L, y0 + 2L]. Indeed, by (2.7) one has
 κ2 1/2
(5.72) F±′ (y) = F ′ (y) ± 1 − 2 = F ′ (y) + O(1),
y
Furthermore, by (2.8)
κ2 2
(5.73) F±′′ (y) = F ′′ (y) ± (y − κ2 )−1/2 .
y2
Recall from (5.57) that F ′′ (y) ≍ cn/m. Under our assumption y0 − k ≥ k1/3+ǫ + 10L, for
y ∈ [y0 − 2L, y0 + 2L] one has y − k ≍ y0 − k ≫ k1/3+ǫ , which implies
κ2 2 κ2  k 3/2
(y − κ2 −1/2
) ≪ (y 0 − k)−1/2
≪ y0−1 k1/2 (y0 − k)−1/2 ≪ y0−1 k1/3−ǫ/2 .
y2 y0
5/2 y0
√
On the other hand, using (5.63) (which gives y0 ≍ mx/n) and the standing assumptions
x ≥ k2 /(8π 2 ) and m ≤ ∆2 kǫ /x, the assumption ∆ ≤ x2/3 k1/3−ǫ of the lemma implies
cn cx1/2 x −1 1/3 −1/3+ǫ/2 −1 x
 1/3
≍ ≫ ≫ y 0 x k ≫ y 0 k1/3+ǫ/2 ≫ y0−1 k1/3+ǫ/2 .
m y0 m1/2 y0 ∆kǫ/2 k2
Combining the above two inequalities, we obtain κ2 y −2 (y 2 − κ2 )−1/2 = o(cn/m). So from (5.57)
and (5.73), we conclude
cn
(5.74) F±′′ (y) ≍ F ′′ (y) ≍ .
m
Differentiating repeatedly, we obtain that for y ∈ [y0 − 2L, y0 + 2L] and j > 2,
(j) (j) dj−2  κ2 2 2 −1/2
 cn 2−j k2
F± (y) = F (y) + j−2 2 (y − κ ) ≪j y + 2 (y + k)−1/2 (y − k)−1/2−(j−2) .
dy y m y
Here we used the bound (5.57) for F (j) (y). Since κ2 y −2 (y 2 − κ2 )−1/2 = o(cn/m), for y ∈
[y0 − 2L, y0 + 2L] (satisfying y − k ≍ y0 − k) we have
cn 2−j cn
(5.75) F (j) (y) ≪j (y + (y − k)2−j ) ≪j (y0 − k)2−j , for j = 2, 3, 4 . . .
m m
We also deduce that F± has a single stationary point in the interval [y0 −2L, y0 +2L]. Indeed,
F±′′ does not change sign on this interval, since F±′′ ≍ cn/m for all y ∈ [y0 − 2L, y0 + 2L] by
(5.74). Thus F±′ is monotonic on this interval. Moreover, the mean value theorem shows
(
′ ′′ [0, y − y0 ] if y ≥ y0 ,
(5.76) F (y) = (y − y0 )F (y0 + ξ), for some ξ = ξ(y) ∈
[y − y0 , 0] if y ≤ y0 .
In particular, by (5.72)
F±′ (y0 + 2L) = F ′ (y0 + 2L) + O(1) = 2LF ′′ (y0 + ξ+ ) + O(1)
and F±′ (y0 − 2L) = −2LF ′′ (y0 − ξ− ) + O(1), for some 0 ≤ ξ+ , ξ− ≤ 2L.
Now since ±2LF ′′ (y0 + ξ± ) ≍ cnL/m ≫ kǫ (by (5.74) and the fact L = kǫ max(1, m/(cn))), we
deduce that F±′ has a unique zero in this interval, call this y0± .
Finally, if y ∈ [y0 − 2L, y0 + 2L], from (5.74) and (5.76) we obtain the derivative bound
cnL cnL cn
F ′ (y) ≪ |y − y0 ||F ′′ (y + ξ(y))| ≪ =⇒ F±′ (y) = F ′ (y) + O(1) ≪ ≪ (y0 − k).
m m m
In the last step, we used (5.72) and our assumption y0 − k ≥ 10L + k1/3+ǫ ≫ L. Thus we can
extend (5.75) to the case j = 1 also. In other words
(j) cn
(5.77) F± (y) ≪j (y0 − k)2−j , valid for j = 1, 2, . . .
m
Equipped with the estimates (5.71), (5.74) and (5.77) and the fact that F±′ has a unique
zero y0± in [y0 − 2L, y0 + 2L], we now apply Lemma 2.7 with α = y0 − 2L, β = y0 + 2L,
 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 31

−3/4
X = x−3/4 y0 (y0 − k)−1/4 , V = L, Y = cn(y0 − k)2 /m and Q = y0 − k. In this case,
Z = Q + (β − α) + X + Y + 1 clearly satisfies kǫ ≤ Z ≤ k100 , say, so to check the conditions
(2.17) it suffices to find an η > 0 such that
V Y 1/2
Y ≥ kη and ≥ kη .
Q
Since we assume y0 − k ≥ 10L + k1/3+ǫ ≥ L, we have
cn cn 2 V Y 1/2  cn 1/2
Y = (y0 − k)2 ≥ L , and = L.
m m Q m
Since L = kǫ max{1, m/(cn)}, we have (cn/m)1/2 L = kǫ max{(cn/m)1/2 , (cn/m)−1/2 } ≥ kǫ , so
the conditions required in (2.17) of Lemma 2.7 are easily satisfied. We thus conclude the bound
Z y0 +2L
−3/4
Gi± (y)eiF± (y) dy ≪ x−3/4 y0 (y0 − k)−1/4 c−1/2 m1/2 n−1/2 + k−1000 .
y0 −2L
√ √
Recall (5.63) states y0 = 4π √mxt/n + O(k1/3−ǫ ) ≍ mx/n. Since y0 − k ≥ 10L + k1/3+ǫ ≥
k1/3+ǫ , it follows y0 − k ≍ 4π mxt/n − k. Therefore the above bound shows
Z y0 +2L  √mx −3/4  4π √mxt −1/4
iF± (y) −3/4
Gi± (y)e dy ≪ x −k c−1/2 m1/2 n−1/2
y0 −2L n n
 √
−9/8 −1/4 −1/2 1/8 1/2 4π mxt
−1/4
≪x k c m n −n .
k
Part (iii) now follows from (5.70). 
Finally, we apply the previous lemma to deduce the following bound for the sums Si .
Lemma 5.13. Assume ∆ ≤ x2/3 k1/3−ǫ . Then for i = 1, 2 one has

Si ≪ x−1/4 k−13/6−ǫ c1/2 m1/4 + x−1/2 k−3/2 m1/2 + x−3/4 k−3/2+2ǫ c−1/2 m3/4
+ x−3/4 k−5/6+ǫ c1/2 m−1/4 + x−5/4 k1/6+ǫ c−1/2 m1/4 + x−5/4 k−5/6+2ǫ c−3/2 m5/4 .
Proof. Recall from Lemma 5.9 that
X
Si ≪ c1/2 m−1/4 |Ii′ (n)| + k−1000 .
√
m1/2 x3/2
≤n≤ 1000k mx
10∆2 kǫ

We now apply the bounds of Lemma 5.12. To do so, we must split the sum over n into three
parts based on the value of y0 (n) - recall from (5.63)
√ √ √
4π mxt 1/3−ǫ 4π mxt 1/3 4π mxt
(5.78) y0 = + O(k ) =⇒ −k ≤ y0 ≤ + k1/3 .
n n n
Firstly, we can see from part (i) of Lemma 5.12 that all terms with y0 + 2L ≤ k − k1/3+ǫ
ǫ √
contribute ≪ e−k /2 , say, which is negligible. (There are ≤ 1000 mx/k of these terms, all of
ǫ
which are ≪ c1/2 m−1/4 e−k .)
We next consider the contribution of Ii′ (n) where
(5.79) y0 (n) + 2L ≥ k − k1/3+ǫ and y0 (n) − 10L ≤ k + k1/3+ǫ .
We first bound the number of n for which (5.79) holds. Since L ≤ kǫ (1 + m/(cn)), using (5.78)
we first observe
√
1/3+ǫ 4π mxt  m
y0 + 2L ≥ k − k =⇒ + k1/3 + 2kǫ 1 + ≥ k − k1/3+ǫ
√ n cn √
4π mxt + 2kǫ m/c 1/3+ǫ 4π mxt + 2kǫ m/c
=⇒ ≥ k − 2k ⇐⇒ n ≤ .
n k − 2k1/3+ǫ
32 NED CARMICHAEL

Similarly,
√
1/3+ǫ4π mxt  m
y0 − 10L ≤ k + k =⇒ − k1/3 − 10kǫ 1 + ≤ k + k1/3+ǫ
√ n cn √
4π mxt − 10kǫ m/c 1/3+ǫ 4π mxt − 10kǫ m/c
=⇒ ≤ k + 2k ⇐⇒ n ≥ .
n k + 2k1/3+ǫ
Consequently, there are
√ √
4π mxt + 2kǫ m/c 4π mxt − 10kǫ m/c
≤ 1/3+ǫ
− 1/3+ǫ
+ 1 ≪ x1/2 k−5/3+ǫ m1/2 + k−1+ǫ c−1 m + 1
k − 2k k + 2k
integers n satisfying (5.79). For these n, we use the bound from part (ii) of Lemma 5.12. This
states
Ii′ (n) ≪ x−3/4 k−5/6+ǫ + x−5/4 k1/6+ǫ c−1 m1/2 .
We thus conclude that the overall contribution to Si from n satisfying (5.79) is

(5.80)
≪ c1/2 m−1/4 x−3/4 k−5/6+ǫ + x−5/4 k1/6+ǫ c−1 m1/2 x1/2 k−5/3+ǫ m1/2 + k−1+ǫ c−1 m + 1

≪ x−1/4 k−5/2+2ǫ c1/2 m1/4 + x−3/4 k−11/6+2ǫ c−1/2 m3/4 + x−3/4 k−5/6+ǫ c1/2 m−1/4
+ x−3/4 k−3/2+2ǫ c−1/2 m3/4 + x−5/4 k−5/6+2ǫ c−3/2 m5/4 + x−5/4 k1/6+ǫ c−1/2 m1/4 .
(Note the second term is dominated by the fourth term, so can be ignored.)
Finally, we bound the contribution of the remaining n, for which y0 (n) − 10L ≥ k + k1/3+ǫ .
Note (using (5.78))
√ √
1/3+ǫ 4π mxt 1/3 1/3+ǫ 4π mxt
y0 − 10L ≥ k + k =⇒ +k ≥k+k =⇒ n ≤ .
n k + 12 k1/3+ǫ
For these n, we use the bound given in part (iii) of Lemma 5.12. This states
 √  √
−9/8 −1/4 −1/2 1/8 1/2 4π mxt −5/8 −9/4 1/8 2 4π mxt
−1/4 −9/4
′
Ii (n) ≪ x k c m n −n +x k m n −n .
k k
It follows that the overall contribution to Si from n satisfying y0 (n) − 10L ≥ k + k1/3+ǫ is
 √
1/2 4π mxt
 X −1/4
1/2 −1/4 −9/8 −1/4 −1/2 1/8
(5.81) ≪ c m x k c m n −n
√ k
4π mxt
n≤
k+ 1
2k
1/3+ǫ

X  4π √mxt −9/4 
−5/8 −9/4 1/8 2
+x k m n −n .
√ k
4π mxt
n≤
k+ 1
2k
1/3+ǫ

(We have dropped the restriction n ≥ m1/2 x3/2 /(10∆2 kǫ ) from these sums at no cost.) It is
simple to bound the remaining sums over n. Firstly,
 4π √mxt Z 4π√mxt  √
1/2 4π mxt
−1/4 1 k1/3+ǫ
−1/4
k+ 2
X
1/2
n −n ≪ u −u du
√ k 1 k
4π mxt
n≤
k+ 1
2k
1/3+ǫ
√
Z 4π
k
mxt
−1  4π √mxt 1/2
= √ √ +v v −1/4 dv
4π mxt
− 4π1 1/3+ǫ
mxt k
k k+ 2 k
√
 √mx  1/2
4π mxt
Z
k
−1  √mx 5/4
−1/4
≪ √ √ v dv ≪ .
k 4π mxt mxt
− 4π1 1/3+ǫ k
k k+ 2 k
 THE SECOND MOMENT OF SUMS OF HECKE EIGENVALUES II 33

Secondly, we easily bound

X  4π √mxt −9/4  √mx 2  4π √mxt √

4π mxt −5/4
2
n −n ≪ −
√
4π mxt
k k k k + 21 k1/3+ǫ
n≤
k+ 1
2k
1/3+ǫ

≪ x3/8 k1/12−5ǫ/4 m3/8 .

Consequently, the contribution of all n with y0 (n) − 10L ≥ k + k1/3+ǫ , given in (5.81) is

(5.82) ≪ x−1/2 k−3/2 m1/2 + x−1/4 k−13/6−ǫ c1/2 m1/4 .

Combining the bounds (5.80) and (5.82) for the two (non-negligible) contributions to the sums
Si (noting also that the first term of (5.80) is dominated by the latter term of (5.82), so can be
ignored), we obtain the lemma. 

Equipped with this bound for the sums Si , the main result of this section (Lemma 5.3) follows
straightforwardly from Lemma 5.7.

Proof of Lemma 5.3. Recall Lemma 5.7, which states

X X
(OD) ≪ x5/4 c−1 m−3/4 max |S1 | + max |S2 | + x−5/4 k−4/3+2ǫ ∆5/2 .


t∈[1,2] t∈[1,2]
c≤100∆2 /(xk 1−ǫ ) m≤∆2 k ǫ /x

Lemma 5.13 (which holds uniformly for t ∈ [1, 2]) now yields
X X X X
(OD) ≪ xk−13/6−ǫ c−1/2 m−1/2 + x3/4 k−3/2 c−1 m−1/4
c m c m
X X X X X X
1/2 −3/2+2ǫ −3/2 1/2 −5/6+ǫ −1/2
+x k c 1+x k c m−1 + k1/6+ǫ c−3/2 m−1/2
c m c m c m
X X
−5/6+2ǫ −5/2 1/2 −5/4 −4/3+2ǫ
+k c m +x k ∆5/2 ,
c m

where c denotes that the sum is taken over c ≤ 100∆2 /(xk1−ǫ ), and m denotes that the
P P
sum is taken over m ≤ ∆2 kǫ /x. One easily bounds these sums, and obtains

(5.83) (OD) ≪ k−8/3 ∆2 + k−3/2+ǫ ∆3/2 + x−1/2 k−3/2+3ǫ ∆2

+ k−4/3+2ǫ ∆ + x−1/2 k1/6+2ǫ ∆ + x−3/2 k−5/6+4ǫ ∆3 + x−5/4 k−4/3+2ǫ ∆5/2 .

We may simplify this expression somewhat, using our assumptions x ≥ k2 /(8π 2 ) and x1/2 ≤
∆ ≤ x2/3 k1/3−ǫ . These imply

x−1/2 k−3/2+3ǫ ∆2 ≪ x−1/6 k−4/3+5ǫ/2 ∆3/2 = k−3/2+ǫ ∆3/2 · x−1/6 k1/6+3ǫ/2 ≪ k−3/2+ǫ ∆3/2
and k−4/3+2ǫ ∆ ≤ x−1/4 k−4/3+2ǫ ∆3/2 ≪ k−11/6+2ǫ ∆3/2 ≪ k−3/2+ǫ ∆3/2 .

Therefore both the third and fourth error term in (5.83) can be absorbed into the second. We
also have
x−5/4 k−4/3+2ǫ ∆5/2 ≤ x−3/2 k−4/3+2ǫ ∆3 ≤ x−3/2 k−5/6+4ǫ ∆3 ,
so the final error term in (5.83) can be absorbed into the penultimate one. We thus obtain the
required bound

(OD) ≪ k−8/3 ∆2 + k−3/2+ǫ ∆3/2 + x−1/2 k1/6+2ǫ ∆ + x−3/2 k−5/6+4ǫ ∆3 .


34 NED CARMICHAEL

References
[1] V. Blomer. The average value of divisor sums in arithmetic progressions. The Quarterly Journal of Mathe-
matics, 59(3):275–286, 10 2007.
[2] V. Blomer, R. Khan, and M. Young. Distribution of mass of holomorphic cusp forms. Duke Mathematical
Journal, 162(14):2609–2644, 2013.
[3] N. Carmichael. The variance of sums of Hecke eigenvalues I. arXiv preprint, 2 2025.
[4] K. Chandrasekharan and R. Narasimhan. On the mean value of the error term for a class of arithmetical
functions. Acta Mathematica, 112:41–67, 1964.
[5] E. Fouvry. Sur le problème des diviseurs de Titchmarsh. Journal für die reine und angewandte Mathematik,
357:51–76, 1985.
[6] E. Fouvry, S. Ganguly, E. Kowalski, and P. Michel. Gaussian distribution for the divisor function and Hecke
eigenvalues in arithmetic progressions. Commentarii Mathematici Helvetici, 89:979–1014, 2014.
[7] J. Hafner and A. Ivić. On sums of Fourier coefficients of cusp forms. Enseign. Math., 35:375–382, 01 1989.
[8] B. Hough. Zero-density estimate for modular form L-functions in weight aspect. Acta Arithmetica,
154(2):187–216, 2012.
[9] H. Iwaniec. Topics in Classical Automorphic Forms, volume 17 of Graduate Studies in Mathematics. Amer-
ican Mathematical Society, 1997.
[10] H. Iwaniec and E. Kowalski. Analytic Number Theory. Colloquium Publications - American Mathematical
Society. American Mathematical Society, 2013.
[11] Y. K. Lau and L. Zhao. On a variance of Hecke eigenvalues in arithmetic progressions. Journal of Number
Theory, 132(5):869–887, 2012.
[12] S. Lester and N. Yesha. On the distribution of the divisor function and Hecke eigenvalues. Israel Journal of
Mathematics, 212:443–472, 2016.
[13] G. Lü. The average value of Fourier coefficients of cusp forms in arithmetic progressions. Journal of Number
Theory, 129:488–494, 02 2009.
[14] H. L. Montgomery. Ten Lectures on the interface between number theory and harmonic analysis, volume 84
of CBMS Regional Conference Series in Mathematics. American Mathematical Society, 1994.
[15] F. W. J. Olver. Asymptotics and Special Functions. A K Peters, 1997.
[16] R. A. Rankin. Sums of cusp form coefficients. In Automorphic Forms and Analytic Number Theory (Montreal,
PQ, 1989), pages 115–121. Univ. Montreal, Montreal, QC, 1990.
[17] Z. Rudnick and K. Soundararajan. Lower bounds for moments of L-functions: Symplectic and orthogo-
nal examples. In Proceedings of the Bretton Woods Workshop on Multiple Dirichlet Series, Proceedings of
Symposia in Pure Mathematics, volume 75. American Mathematical Society, 2006.
[18] A. Walfisz. Über die koeffizientensummen einiger modulformen. Mathematische Annalen, 108:75–90, 1933.
[19] G. N. Watson. A Treatise on the Theory of Bessel Functions. Cambridge University Press, second edition,
1942.
[20] J. Wu. Power sums of Hecke eigenvalues and application. Acta Arithmetica, 137(4):333–344, 2009.
[21] J. Wu and H. Tang. Fourier coefficients of symmetric power L-functions. Journal of Number Theory, 167:147–
160, 2016.

Department of Mathematics, King’s College London, London, WC2R 2LS, UK

Email address: [email protected]