11 Waveforms

(November 26, 2022)
Waveforms, I
Paul Garrett garrett@math.umn.edu http://www.math.umn.edu/egarrett/
[This document is http://www.math.umn.edu/˜garrett/m/mfms/notes 2013-14/11 waveforms I.pdf]
1. Waveform Eisenstein series Es
2. Heegner point periods of Eisenstein series
3. Closed geodesic periods of Eisenstein series
4. Cuspidal waveforms, Fourier expansions
5. Fourier expansions of Eisenstein series
6. Standard L-functions attached to cuspforms
7. The good argument for non-vanishing of ζ(1 + it)
Hecke had observed examples where the zeta functions of rings of algebraic integers in complex quadratic
extensions of Q are expressible as Mellin transforms of binary theta series. For example, with o = Z[i] the
ring of Gaussian integers,
X 1
ζQ(i) (s) = (with N the Galois norm Q(i) → Q)
(N α)s
06=α∈o/o×
the simplest binary theta series

2
+d2 )z
X
θ2 (z) = eπi(c
c,d∈Z
gives Z ∞
dy
y s θ2 (iy) = π −s Γ(s) ζQ(i) (s)
0 y
√
−1+ −3
Similarly, for the Eisenstein integers o = 2 ,
X 1 √
ζQ(√−3) (s) = (with N the Galois norm Q( −3) → Q)
(N α)s
06=α∈o/o×
and binary theta series

2
+mn+n2 )z
X
θ(z) = eπi(m
m,n∈Z
we have Z ∞
dy
y s θ(iy) = π −s Γ(s) ζQ(√−3) (s)
0 y
As with Riemann’s treatment of ζ(s) of the rational integers Z, these identities give analytic continuations
and functional equations of the zeta functions. The theta functions are holomorphic weight-one modular
forms for various congruence subgroups of SL2 (Z). Similarly grossencharacter L-functions such as
X (α/|α|)4m
LQ(i) (s, χ4m ) =
(N α)s
06=α∈o/o×
useful in proving equidistribution of Gaussian primes in angular sectors, appeared as Mellin transforms of
harmonic binary theta series
2 2
 P
 c,d∈Z (ci + d)4m eπi(c +d )z (for m ≥ 0)
θ2,4m (z) =
2
+d2 )z
(ci − d)4|m| eπi(c
P
c,d∈Z (for m ≤ 0)
1
Paul Garrett: Waveforms, I (November 26, 2022)
by Z ∞
dy
y s θ2,4m (iy) = π −s Γ(s) LQ(i) (s − 2|m|, χ4m )
0 y
Again, θ2,4m is a holomorphic modular form of weight 1 + 4m, and is a cuspform for m 6= 0.
At the time, such devices
√ seemed only to work for complex quadratic extensions of Q, not real quadratic
extensions such as Q( 2). Although [Hecke 1918/20] had already exhibited other means to produce such
zeta functions by Mellin transforms, Hecke proposed to Maaß to find some kind of automorphic form for
real quadratic fields analogous to the holomorphic binary theta series for complex quadratic fields, and for
grossencharacter L-functions.
Maaß succeeded in finding the special waveforms to function as hoped. They are a sort of binary theta series
with an indefinite form, such as x2 − 2y 2 , rather than the definite forms such as x2 + y 2 in Hecke’s treatment.
Rather than being holomorphic functions of z ∈ H with an automorphy relation for suitable Γ ⊂ SL2 (Z),
Maaß special waveforms satisfy the analytic condition
∂2 ∂2
∆H u = λ · u

(for some λ ∈ C, with ∆H = y 2 ∂x2 + ∂y 2 )
and are invariant under suitable Γ ⊂ SL2 (Z). Here ∆H is the SL2 (R)-invariant Laplacian on H. [1] In
fact, the special waveforms exhibited by Maaß to address Hecke’s question indeed prove to be special among
waveforms. One special aspect is that none of the cuspidal waveforms among Maass’ special waveforms is of
level one. They are only of higher level, depending on the real quadratic field extension. [Selberg 1956] used
a trace formula to prove that there are infinitely-many cuspidal waveforms of level one. [2]
This part treats the basic features of waveforms, postponing several more sophisticated discussions. The
first emphasis is on a family of Eisenstein series, each an eigenfunction for the invariant Laplacian. These
are the most explicit waveforms, and have several important roles: in integral representations of L-functions
as periods of Eisenstein seriess, in giving the most conceptual proof of non-vanishing of ζ(s) and other L-
functions on the edges of the critical strip, and in Rankin-Selberg integral representations of L-functions
attached to cuspforms.
Later, we will see the fundamental role of Eisenstein series in the spectral decomposition of L2 (Γ\H).
1. Waveform Eisenstein series Es

Let G = SL2 (R) act on the upper half-plane H by linear fractional transformations, as usual. Let P be
upper-triangular matrices in G. Use coordinates z = x + iy on H.
[1.1] Invariant Laplacian The differential operator

∂2 ∂2
∆ = ∆H = y 2 +
∂x2 ∂y 2

[1] The SL (R)-invariance of ∆H = y 2 ∂ 2 + ∂ 2 can be proven by direct computation, but this is silly. Indeed,
2 ∂x2 ∂y 2
we want to determine such an invariant operator in this and other circumstances, rather than merely verify it after
someone gives it to us. A dignified, broadly applicable approach to expressing an invariant Laplacian in coordinates
is given in a supplement.
[2] A modern form of Selberg’s argument will be given later. Elementary heuristic versions of the argument
unfortunately fail to suggest how to legitimize many details, producing the hazard that similar-sounding arguments
produce nonsense.
2
is G-invariant: the P -invariance is visible, but, again, the G-invariant should not be verified by direct
computation, although such verification is possible in principle. The supplement shows how to obtain
invariant operators. The simplest ∆ eigenfunctions are the functions z = x + iy −→ y s for complex s,
because ∂2 ∂2 s
∆y s = y 2 + (y ) = y 2 · s(s − 1)y s−2 = s(s − 1) · y s
∂x2 ∂y 2
With Γ = SL2 (Z), since Z is a principal ideal domain, and Z× = {±1} has cardinality 2,

∗ ∗
(P ∩ Γ)\Γ ←→ {±1} \ {(c d) : coprime c, d} by ←→ ( c d)
c d
The SL2 (R) invariant measure on H is

dx dy
y2
As with ∆, the P -invariance is easy to see, but the G-invariance is not, and should not be verified after-the-
fact by brute force computation, but derived from generally-applicable considerations as in the supplement.
By coincidence, the symmetry property
Z Z
dx dy dx dy
(∆f )(z) · g(z) 2
= f (z) · (∆g)(z)
H y H y2
is visible from Euclidean integration by parts for the Euclidean Laplacian and usual measure on R2 : for
f, g ∈ Cc∞ (H), Z 2
∂2
Z
dx dy ∂
(∆f )(z) · g(z) = + f (z) · g(z) dx dy
H y2 H ∂x
2 ∂y 2
Z ∂2 ∂2
Z
dx dy
= f (z) · 2
+ 2
g(z) dx dy = f (z) · (∆g)(z)
H ∂x ∂y H y2
The convenient coincidence of cancellation of the y 2 factors is not essential for symmetry. [3] Also, ∆ = ∆H
has the negative-semi-definite property that
Z
dx dy
(∆f )(z) · f (z) ≤ 0
H y2
since, integrating once by parts,

Z Z 2 2
dx dy ∂f ∂f
(∆f )(z) · f (z) = − + dx dy ≤ 0
H y2 H ∂x ∂y
Integrals of Γ-invariant functions on H on the quotient Γ\H can be understood in an elementary way as
integrals over the standard fundamental domain
{z = x + iy ∈ H : |z| ≥ 1, − 21 ≤ x ≤ 21 }
[3] The same coincidence does not occur in other, analogous situations, such as n-dimensional hyperbolic space for
n ≥ 3: with coordinates (x, y) with x ∈ Rn−1 and y > 0 on hyperbolic n-space, the invariant Laplacian and invariant
measure are found to be 2
∂2

∂ ∂ dx dy
y2 + − (n − 2) y and
∂x2 ∂y 2 ∂y y n+1
Nevertheless, invariant Laplacians are always symmetric with respect to the corresponding invariant measures, for
very general reasons. See the supplement.
3
Apart from other, subtler problems, such a viewpoint complicates understanding of why integration by parts
for ∆ = ∆H is correct for f ∈ Cc∞ (Γ\H). Thus, instead of integrating over a fundamental domain, one might
eventually discern an intrinsic integral on the quotient, as is done in a supplement. In summary, first prove
surjectivity of the averaging maps
X
α : Cco (H) −→ Cco (Γ\H) by (αf )(z) = (f ◦ γ)(z)
γ∈Γ
and X
α : Cc∞ (H) −→ Cc∞ (Γ\H) by (αf )(z) = (f ◦ γ)(z)
γ∈Γ
where Co∞ (Γ\H) is construed as right K-invariant compactly-supported smooth functions on the smooth
manifold Γ\G, using H ≈ G/K. Second, show that there is a unique measure on Γ\H such that, for every
f ∈ Cco (H), Z Z X
f = f ◦γ
H Γ\H γ∈Γ
Indeed, similar iterated-integral identities are essential to unwinding arguments.
[1.2] Eisenstein series The Eisenstein series Es for Γ is

X 1 X ys
Es (z) = Im (γz)s = ·
2 |cz + d|2s
γ∈P ∩Γ\Γ coprime c,d
As with the holomorphic Eisenstein series c,d 1/(cz + d)2k , the coprimality condition and quotient by {±1}
P
can be avoided by inserting all possible common factors and multiplying by 2:
X ys
2 ζ(2s) · Es (z) =
|cz + d|2s
(0,0)6=(c,d)∈Z2
For Re (s) > 1 the series defining

P Es converges absolutely, uniformly for z in compacts, much as for the
holomorphic Eisenstein series c,d 1/(cz + d)2k , as follows. Recall

a b Im (z) a b
Im (z) = (for ∈ SL2 (R))
c d |cz + d|2 c d
Since Es is Γ-invariant, it suffices to consider z in a fixed compact C inside the usual fundamental domain
{z = x + iy ∈ H : |z| ≥ 1, − 12 ≤ x ≤ 21 }
For such z,
(cx + d)2 + (cy)2 = (x2 + y 2 )c2 + 2x · cd + d2 ≥ c2 − |cd| + d2 ≥ 1 2

2 (c + d2 )
The sum over coprime (c, d) is certainly dominated by the sum over all (c, d), not both 0. Thus, the Eisenstein
series is uniformly (for z in compacts inside the fundamental domain) dominated by
X 1
c,d not both 0
(c2 + d2 )Re (s)
An adaptation of an integral test proves that this converges for Re (s) > 1.
4
Since ∆ is G-invariant, it is certainly Γ-invariant. Thus, at least in the region of convergence Re (s) > 1, the
Eisenstein series Es is a ∆-eigenfunction: [4]
X X X
∆ Es (z) = ∆ Im (γz)s = ∆ Im (γz)s = ∆ Im (z)s ◦ γ
γ∈P ∩Γ\Γ γ∈P ∩Γ\Γ γ∈P ∩Γ\Γ
X X
= s(s − 1) · Im (z)s ◦ γ = s(s − 1) · Im (z)s ◦ γ
γ∈P ∩Γ\Γ γ∈P ∩Γ\Γ
X
= s(s − 1) · Im (γz)s = s(s − 1) · Es (z)
γ∈P ∩Γ\Γ
This eigenfunction property persists for the meromorphically-continued Eisenstein series (below), by the
identity principle from complex analysis.
Note that the eigenvalue λs = s(s−1) is real and non-positive, as would be expected of a negative-semi-definite
symmetric/self-adjoint operator, only for either Re (s) = 12 or s ∈ [0, 1], but that the series characterization
of the Eisenstein series does not converge there.
[1.3] Meromorphic continuation and functional equation We follow [Godement 1966a]’s Poisson
summation argument, descended from [Rankin 1939], attributed by Rankin to his advisor Ingham. This
argument is elementary, but less informative than arguments that engage with the spectral theory.
[Epstein 1903] had already meromorphically continued Es (z) by expressing it as a Mellin transform.
The usual ζ(s) with its gamma factor is ξ(s) = π −s/2 Γ( 2s )ζ(s), with functional equation ξ(1 − s) = ξ(s).
[1.3.1] Theorem: For each fixed z ∈ H, s(1 − s)ξ(2s) · Es (z) has an analytic continuation to an entire
function [5] of s. The functional equation is
ξ(2s) Es = ξ(2 − 2s) E1−s
Es (z) has no pole in Re (s) > 21 other than at s = 1. The pole of Es at s = 1 is simple with residue the
constant function 3/π. For fixed δ > 0, via Phragmén-Lindelöf, letting s = σ + it,

−s
y ζ(2s) Es (z) δ y 1+2δ · t1+2δ (for −δ ≤ Re (s) ≤ 1 + δ)

For any fixed s, the function Es (z) is of moderate growth, in the sense that there is a sufficiently large
exponent A depending on s such that
Es (z) s y A (as y → +∞)
[1.3.2] Remark: Further, with z fixed, away from poles, if we are willing to invoke some convexity results
such as Hadamard’s three-circle theorem, there is the convexity bound
Es (z) t1−Re (s)+ε (for 1

2 ≤ Re (s) ≤ 1, for every ε > 0)
The argument for this is essentially as follows. For a holomorphic function f (s) with polynomial vertical
growth, let
ψ(σ) = inf{A : |f (σ + it)| tA+ε , for all ε > 0}
[4] Term-wise application of ∆ is justified by checking that the series defining E converges in the C k topology on
s
smooth functions on H, for all k ≥ 0.
[5] The Eisenstein series s → E is a function-valued function of s. Nevertheless, we mostly consider the scalar-valued
s
functions s → Es (z) with fixed z ∈ H.
5
Hadamard’s three-circle and other convexity results show that ψ is convex. Since ψ(1 + δ) = 0 for every
δ > 0, and ψ(−δ) = 1 + 2δ, necessarily ψ( 12 ) = 12 , and ψ(σ) = 1 − σ for 21 ≤ σ ≤ 1.
Proof: For v = (c d) ∈ R2 , consider the Gaussian

2 2
+d2 )
ϕ(v) = e−π|v| = e−π(c
with v → |v| the usual length function on R2 . For g ∈ SL2 (R), define
X X 2
Θ(g) = ϕ(v · g) = e−π|(c,d)g|
v∈Z2 (c,d)∈Z2
where v ∈ R2 is a row vector. Consider the integral (a Mellin transform)

Z ∞
dt
t2s (Θ(tg) − 1)
0 t
where the t in the argument of Θ simply acts by scalar multiplication on g ∈ SL2 (R). On one hand,
integrating term-by-term gives
Z ∞ X Z ∞
dt 2 dt
2s
t (Θ(tg) − 1) = t2s e−π|tvg|
0 t 0 t
v6=(0,0)
Since √
π|tvg|2 = (t · π|vg|)2
√
we can change variables by replacing t by t/( π|vg|) to obtain
X √ Z ∞ Z ∞
−2s 2s −t2 dt 1 −s X −2s dt
( π|vg|) t e = π |vg| ts e−t
0 t 2 0 t
v6=(0,0) v6=(0,0)
1 −s X
= π Γ(s) |vg|−2s
2
v6=(0,0)
We want g ∈ SL(2, R) to map i → x + iy. One reasonable choice is

√
1 x y 0
g = gz = √
0 1 0 1/ y
Using this choice of g and writing out v = (c, d) gives

√
1 x y 0 √ cx + d
vg = (c, d)g = ( c d ) √ = (c y, √ )
0 1 0 1/ y y
and thus
X X √ cx + d X (cx + d)2 −s
|vg|−2s = |(c y, √ )|−2s = (c2 y + )
v v
y v
y
s
X y X ys X ys
= = =
v
(c2 y 2 2
+ (cx + d) )s
v
|ciy + cx + d|2s
v
|cz + d|2s
Letting 1 ≤ δ = gcd(c, d), this is

X ys X 1 X ys
2s
= = 2 ζ(2s) · Es (z)
v
|cz + d| δ 2s |cz + d|2s
δ coprime c,d
6
The expression
X ys
2 ζ(2s) Es (z) = (summing (c, d) over all non-zero vectors in Z2 )
|cz + d|2s
(c,d)6=(0,0)
is convenient, being a sum over a lattice with 0 removed.

Thus, the integral representation yields the Eisenstein series with a leading power of π, a gamma function,
and a factor of ζ(2s): Z ∞
dt
t2s (Θ(tg) − 1) = π −s Γ(s) ζ(2s) Es (g(i))
0 t
To prove the meromorphic continuation, use that integral representation as in Riemann’s argument for ζ(s),
first breaking the integral into two parts, one from 0 to 1, and the other from 1 to +∞. Keep g ∈ SL(2, R)
in a compact subset of SL(2, R). Since elementary estimates show the absolute convergence uniformly on
compacts, Z ∞
dt
t2s (Θ(tg) − 1) = entire in s
1 t
2
Apply Poisson summation to the kernel: first, the Gaussian ϕ(v) = e−π|v| is its own Fourier transform, and
Fourier transform of (v → ϕ(tvg)) = (v → t−2 det(g)−1 · ϕ(t−1 v >g −1 ))
where >g is g-transpose. Then Poisson summation asserts
Θ(tg) = t−2 det(g)−1 · Θ(t−1 >g −1 )
Noting det g = 1, the modification for the kernel gives
Θ(tg) − 1 = t−2 · [Θ(t−1 >g −1 ) − 1] + t−2 − 1
Transform the integral from 0 to 1 into an integral from 1 to +∞: at first only for Re (s) > 1,
Z 1 Z 1
dt dt
t2s (Θ(tg) − 1) = t2s t−2 · [Θ(t−1 >g −1 ) − 1] + t−2 − 1
0 t 0 t
Replacing t by 1/t turns this into

Z ∞ dt
t−2s t2 · [Θ(t >g −1 ) − 1] + t2 − 1
1 t
Explicitly evaluating the last two elementary integrals of powers of t from 1 to ∞, using Re (s) > 1, this is
Z ∞
dt 1 1
t2−2s (Θ(t >g −1 ) − 1) + −
1 t 2s − 2 2s
For g in SL(2),
> −1
g = wgw−1
where w is the long Weyl element
0 −1
w =
1 0
Since Z2 − (0, 0) is stable under w, and since the length function v → |v|2 is invariant under w,
Θ(g) = Θ(wg) = Θ(gw−1 )
7
so
Θ( >g −1 ) = Θ(g)
Thus, the original integral from 0 to 1 becomes
Z ∞
dt 1 1
t2−2s (Θ(tg) − 1) + −
1 t 2s − 2 2s
and the whole equality, with g of the special form above, is

Z ∞ Z ∞
dt dt 1 1
π −s Γ(s) ζ(2s) Es (z) = t2s (Θ(tg) − 1) + t2−2s (Θ(tg) − 1) + −
1 t 1 t 2s − 2 2s
The integrals from 1 to ∞ are nicely convergent for all s ∈ C, uniformly for g in compacts. The elementary
rational expressions of s have meromorphic continuations. Thus, the right-hand side gives a meromorphic
continuation of the Eisenstein series, and is visibly invariant under s → 1 − s.
The only poles are at s = 1, 0. The residue at s = 1 is the constant function 12 , and at s = 0 the residue is
the constant function − 12 . At s = 1 the factor π −s Γ(s) is holomorphic and has value 1/π, so
π
Ress=1 ζ(2s) Es =
2
Now we recover the assertions for Es itself. The convergence of the infinite product
X 1 Y 1
ζ(2s) = =
n
n2s 1 − p−2s
p prime
for Re (s) > 1/2 assures that ζ(2s) is not zero for Re (s) > 1/2. And ζ(2) = π 2 /6. These facts and the
previous discussion give the result for Es .
To see the moderate growth, first observe that on bounded vertical strips s(s − 1) · ξ(2s) · Es easily admits
N
a bound of the form e|s| , from the integral representation in terms of Θ. Thus, Phragmén-Lindelöf is
applicable.
Fix δ > 0 and yo > 0. On the vertical line Re (s) = 1 + δ, the convergent series for |y −s · ζ(2s) · Es (z)| shows
that it is bounded. The functional equation gives
y −(1−s) y −(1−s)
y −(1−s) ζ(2 − 2s) E1−s (z) = ξ(2 − 2s) E1−s (z) = −(1−s) ξ(2s) Es (z)
π −(1−s)Γ(1 − s) π Γ(1 − s)
y −(1−s) π −s Γ(s) y −(1−2s) π −s Γ(s)

= −(1−s)
ζ(2s) Es (z) = −(1−s) · y −s ζ(2s) Es (z)
π Γ(1 − s) π Γ(1 − s)
On the vertical line Re (s) = 1 + δ, using asymptotics of Γ(s) [6] this is bounded by
Γ(1 + δ + it)
y 1+2δ π −1−2δ y 1+2δ π −1−2δ t1+2δ

Γ(−δ − it)
Thus, with large N , getting rid of the poles by multiplying by s(s − 1) and compensating via the Gamma
function,
Γ(s + N )
× s(s − 1) × y −s ζ(2s) Es (z)
Γ(s + N + 3 + 2δ)
[6] The asymptotic Γ(a + it)/Γ(b + it) ∼ ta−b follows from Stirling’s formula, but even more easily from asymptotics
of integrals, sometimes called Watson’s lemma.
8
is holomorphic on the strip −δ ≤ Re (s) ≤ 1+δ, and bounded by a constant multiple of y 1+2δ . By Phragmén-
Lindelöf,
Γ(s + N )
× s(s − 1) × y −s ζ(2s) Es (z) ≤δ,N y 1+2δ (for −δ ≤ Re (s) ≤ 1 + δ)

Γ(s + N + 3 + 2δ)

That is, away from poles,

−s
y ζ(2s) Es (z) ≤δ y 1+2δ · t1+2δ (for −δ ≤ Re (s) ≤ 1 + δ)

Moderate growth Es (z) y Re (s) is clear in Re (s) ≥ 1 + δ for fixed δ > 0, and from the previous discussion
in −δ ≤ Re (s) ≤ 1 + δ. Similarly, in −B ≤ Re (s) ≤ −δ, the functional equations of Es and ξ(2s) give
ξ(2s) Es (z) ξ(2 − 2s) E1−s (z)

Es (z) = = s y 1−Re (s) (as y → +∞)
ξ(2s) ξ(1 − 2s)
Thus, Es (z) is of moderate growth for all s. ///
[1.4] Constant term of Eisenstein series By definition, the constant term is a sort of 0th Fourier
th
coefficient in x = Re (z). Literally, it is the 0 Fourier component in x, a function of y = Im (z):
Z 1
cP Es (iy) = Es (x + iy) dx
0
We will see later that the form of the constant term of Es dictates the functional equation and other features
of Es .
[1.4.1] Claim: The constant term of Es is

ξ(2s − 1) 1−s
cP Es (iy) = y s + y
ξ(2s)
Proof: By direct computation,

1
y s dx
X Z
cP 2ζ(2s) · Es (x + iy) =
0 |cz + d|2s
(0,0)6=(c,d)∈Z2
1 X X XZ 1
XZ y s dx y s dx
= +
d6=0 0 |d| 2s
c6=0 d mod c `∈Z 0
|c| · |z + dc + `|2s
2s
X 1 X XZ 1 1 dx
= 2ζ(2s) y s + y s s
|c|2s (x + d
+ `)2 + y 2 )
c6=0 d mod c `∈Z 0 c
Unwinding the sum over ` and integral over [0, 1] to make an integral over R, this becomes
X 1 X Z 1 dx
s s
2ζ(2s) y + y
|c|2s d 2 2 s

c6=0 R (x + c ) + y )
d mod c
The |c| translates by d/c for d mod c do not alter the integral, so this is
X 1 Z dx
2ζ(2s) y s + y s 2s−1 2 2 s
|c| R (x + y )
c6=0
9
Replacing x by xy in the integral gives

X 1 Z dx
Z
dx
2ζ(2s) y s + y 1−s 2s−1 2 s
= 2ζ(2s) y s + 2y 1−s ζ(2s − 1)
|c| R (x + 1) R (x2 + 1)s
c6=0
The remaining integral is evaluated in terms of Γ(s) by a standard device, as follows. At first just for real
s > 0, but then for complex s by the identity principle,
Z ∞ Z ∞
dt dt
ty s
e t = y −s et ts = y −s Γ(s)
0 t 0 t
Thus,
Z Z ∞ Z Z ∞
1 √
Z
dx 1 −t(x2 +1) s dt 1 2 dt
2 s
= e t dx = π √ e−(πx +t) ts dx
R (x + 1) Γ(s) R 0 t Γ(s) R 0 t t
Z ∞
1 √ 1 dt Γ(s − 12 ) √
= π e−t ts− 2 = π
Γ(s) 0 t Γ(s)
Thus, the whole constant term of 2ζ(2s) Es is
Γ(s − 12 ) √
2ζ(2s) y s + 2y 1−s ζ(2s − 1) π
Γ(s)
Dividing by 2ζ(2s) and re-attributing powers of π gives
ξ(2s − 1)
y s + y 1−s
ξ(2s)
as claimed. ///
[1.4.2] Remark: In 0 < Re (s) < 12 , the Eisenstein series has poles at ρ/2 for all non-trivial zeros ρ of ζ(s).
[Epstein 1903] studied s → Es (zo ) for fixed zo ∈ H as a generalized zeta function, subsequently called an
Epstein zeta function. For generic zo ∈ H the function s → Es (zo ) does not have an Euler product, although
it can be construed as a generalized Dirichlet series. Indeed, in a great variety of cases, these Epstein
zeta functions substantially violate a Riemann Hypothesis: their zeros do not all lie on Re (s) = 21 . See
[Potter-Titchmarsh 1935], [Davenport-Heilbronn 1936], [Stark 1967], [Chowla-Selberg 1967], [Voronin 1976],
[Bombieri-Hejhal 1995].
2. Heegner point periods of Eisenstein series

Linear combinations or integrals of functions s → Es (zo ) are called periods of the Eisenstein series Es . Those
periods admitting Euler products are of special interest. [7]
Hecke and Maaß and others were aware of facts such as
ζQ(i) (s)
Es (i) = 2
ζ(2s)
[7] Although some such computations can be done in a seemingly elementary context, as here, they are significantly
more intelligible and persuasive when done using the Iwasawa-Tate adele-group description of L-functions, and using
a description of Eisenstein series as functions on adele groups GL2 (A). We return to these methods a little later.
10
where ζQ(i) (s) is the Dedekind zeta function of the Gaussian integers Z[i]: the ring of Gaussian integers Z[i]
is a principal ideal domain with just four units {±1, ±i}, and its Dedekind zeta function can be correctly
described in a naive fashion:
X 1 1 X 1 X 1
ζQ(i) (s) = 2s
= × 2s
= 14 2 2 )s
= 21 · ζ(2s) · Es (i)
×
|α| #Z[i] |α| 2
(c + d
α∈Z[i] \(Z[i]−0) 06=α∈Z[i] (0,0)6=(c,d)∈Z
Because Z[i] is Euclidean, it is a principal ideal domain, and has unique factorization, so ζQ(i) (s) has an
Euler product, analogous to ζ(s):
X 1 Y 1
ζQ(i) (s) = =
|α|2s 1 − |$|−2s
α∈Z[i]× \(Z[i]−0) Gaussian primes $
A similar straightforward result holds for rings of algebraic integers Z[ω] for ω complex quadratic whenever
Z[ω] happens to be a principal ideal domain. For example, rings Z[ω] with
√ √ √
√ √ −1 + −3 −1 + −7 −1 + −11
ω = −1, −2, , ,
2 2 2
are demonstrably Euclidean, so are principal ideal domains.
√
[2.1] Heegner points and complex-quadratic
√
periods The ring of algebraic integers o = Z[ −5], inside
the complex quadratic field k = Q( −5), is not a principal ideal domain: [8]
√ √
2 · 3 = (1 + −5) · (1 − −5)
√
The zeta function ζk (s) of the ring of algebraic integers o = ok = Z[ −5] is
X 1
ζk (s) =
N as
06=a⊂o
where a is summed over non-zero ideals in o, and N a = #o/a. Since o is not a principal ideal domain, ζk (s)
is not a single value Es (zo ), but happnes to be essentially the sum of two values of Epstein zetas: quite
unobviously,
√ ζk (s) √ 1 + √−5
| −5|s · = Es ( −5) + Es
ζ(2s) 2
√ X 1 X 1
= | −5|s · +
(m2 + 5n2 )s (2m2 + 2mn + 3n2 )s
coprime m,n mod ±1 coprime m,n mod ±1
Note that the discriminants of both quadratics x + 5 and 2x2 + 2x + 3 are −20.
2
√
[2.1.1] Theorem: For a complex quadratic field extension k = Q( −D) with ring of algebraic integers
o = ok having h different o-isomorphism classes of non-zero ideals, [9] there are z1 , . . . , zh in a strict version
of the fundamental domain,
F str = {z = x + iy ∈ H : |z| ≥ 1, − 21 < x ≤ 21 , and |z| > 1 if x < 0}
[8] The Galois norm N (a + b√−5) = a2 + 5b2 is multiplicative, and (easily checked) N (α) = 1 for α ∈ Z[√−5] if and
only α is a unit. For integers a, b, we have a2 + 5b2 = 1 exactly for b = 0, a = ±1. Since N (2) = 4, 2 is irreducible
√
since there is no α ∈ o with N (α) = 2. Similarly, 3 and 1 ± −5 are irreducible. Similarly, none of the other three
factors appearing can be further factored. The left-hand factors 2, 3 do not differ by units from the right-hand factors
√
1 ± −5, so 6 has two different factorizations.
[9] The number of o-isomorphism classes of non-zero ideals in o is the class number of o.
11
for SL2 (Z) on H, such that

√
(| −D|/2)s
ζk (s) = Es (z1 ) + Es (z2 ) + . . . + Es (zh )
ζ(2s)
√
[2.1.2] Remark: This characterizes the Heegner points z1 , . . . , zh associated to k = Q( −D). [10]
Proof: Such expressions are obtained as follows. The sum expression for the usual zeta-function of o breaks
into subsums according to o-isomorphism classes of ideals:
X X 1 X X 1 X 1 X 1
ζk (s) = = =
(N b)s N (ba−1 )s · (N a)s (N a)s N (ba−1 )s
classes a b≈a classes a b≈a classes a b≈a
An o-isomorphism f : a → b from one non-zero o-ideal to another is given by some θ ∈ k × , since the k-linear
extension to f : k → k is a k-linear automorphism of a one-dimensional k-vectorspace. That is, b = θ · a.
Since θ · a = b ⊂ o, necessarily θ ∈ a−1 , where we adopt the standard convention
a−1 = {β ∈ k : β · a ⊂ o}
In such terms,
X X 1
ζk (s) =
N (θ · a)s
classes a 06=θ∈a−1 , mod o×
The quadratic form

Qa (θ) = N (θ · a) (on θ ∈ a−1 )
on the Z-module a−1 is positive-definite and turns out to have discriminant −D, in the following sense. Fix
a Z-basis e1 , e2 of a−1 , and take integers A, B, C so that for all m, n ∈ Z
N (me1 + ne2 ) · a−1 = Am2 + Bmn + Cn2

It turns out that B 2 − 4AC = −D: proof of this point requires a little algebraic number theory, which we
postpone. Put √
B ± B 2 − 4AC
z = za =
2A
The quadratic form Qa is recovered from z = x + iy:
B
|m + zn|2 m2 + 2xmn + |z|2 n2 m2 + 2 2A mn + 4AC
4A2 n
2
Am2 + Bmn + Cn2 Qa (m, n)
= = √ = √ = √
y y | −D|/2A | −D|/2 | −D|/2
That is,
|√−D| s X 1 √
B± B 2 −4AC
= Es (za ) (with za = 2A )
2 N (θ · a)s
06=θ∈a−1 mod ±1
Change-of-basis in the Z-module a−1 changes the

coefficients
A, B, C of the quadratic form Qa , which changes
a b
za ∈ H by the action of SL2 (Z) on H: for γ = ∈ Γ,
c d
[10] We should restrict attention to fundamental discriminants −D < 0, meaning that square-free −d < 0 is replaced
by −D = −4d for −d = 2, 3 mod 4. Thus, for example, −1 is replaced by −4, −2 is replaced by −8, and −5 is
replaced by −20. In terms of algebraic number theory, this makes the discriminant the square of the volume of the
√
quotient C/o, where o is the ring of algebraic integers in Q( −D). The ring o = ok of algebraic integers in k is the
collection of α ∈ k satisfying a monic equation α2 + bα + c = 0 with a, b ∈ Z.
12
Qa ((m, n) · γ) Qa (ma + nc, mb + nd) |(ma + nc) + za (mb + nd)|2 |m(a + bza ) + (c + dza )n|2
√ = √ = =
| −D|/2 | −D|/2 ya ya
dza +c dza +c
|m + bza +a n|2 |m + bz a +a
n|2
= =
ya /|bza + a|2 Im bzdza +c
a +a
In particular, by changing the Z-basis of a−1 we can put za into the strict fundamental domain F str for Γ,
and uniquely so for each isomorphism class a of non-zero o-ideals. These points za are the Heegner points
attached to −D < 0. For √
B ± B 2 − 4AC
za = ∈ F str
2A
the corresponding quadratic form is said to be reduced. In terms of the coefficients A, B, C in the quadratic
form Am2 + Bmn + Cn2 , the reduced condition is
−A < B ≤ A, A ≤ C, and A < C if B < 0 (all integers, with B 2 − 4AC = −D)
From −A < B ≤ A ≤ C and B 2 − 4AC = −D follows
D = 4AC − B 2 ≥ 3AC ≥ A2
Thus, there are only finitely-many values of A, hence also of B, C, for given −D.
This proves that there are only finitely-many Heegner points za ∈ F str for given −D < 0. This is finiteness
of class number, that is, finiteness of the o-isomorphism classes of non-zero ideals in o. ///
[2.1.3] Remark: Proof that the discriminants of all quadratic forms θ → N (θ · a) for θ ∈ a−1 are −D was
omitted in the above discussion.
3. Closed geodesic periods of Eisenstein series

√
For real quadratic fields k = Q( D) with discriminants D > 0 and ring of algebraic integers o = ok , the
Dedekind zeta function X 1
ζk (s) =
|N a|s
06=a⊂o
defined in terms of the non-zero ideals in o are integrals of Es (z) over z in a finite sum of closed geodesic
curves on Γ\H. For real quadratic fields, the norm can be negative, hence the necessity of the absolute value.
The number of geodesics proves to be the class number of o. Apparently, the Heegner points for the complex
quadratic
√ case become closed geodesics for the real quadratic case. To isolate the new phenomenon, take
k = Q(√D) with ring of√integers o = ok a principal ideal domain, that is, with class number 1. For example,
k = Q( 2) with o = Z[ 2] is provably Euclidean, so is a principal ideal domain.
As a very special case of Dirichlet’s Units Theorem: in real-quadratic rings of integers the units group is
infinite. Specifically, modulo torsion (which is just ±1) the units group is free on one generator. For example,
√ √
Z[ 2]× = {± powers of 1 + 2}
The useful description of the relevant geodesic curves is not in metric-geometry terms, [11] but group-theoretic,
more essentially so than the complex-quadratic case. Imbed k × in GL2 (Q) by
√

a Db
a + b D −→ (for a, b ∈ Q)
b a
[11] An elementary description of all full geodesics on H is that they are arcs of circles that are orthogonal to the real
axis, together with vertical lines orthogonal to the real axis.
13
Let √
a Db 2 2 cosh t D sinh t
H = { : a, b ∈ R, a − Db = 1} = { sinh
√ t
: t ∈ R}
b a D
cosh t
For brevity, write √

cosh t D sinh t
ht = { sinh
√ t
D
cosh t
The fact that there are non-trivial units in o is equivalent to the compactness of the quotient (Γ ∩ H)\H.
Noting that √ √
√ cosh t · i D + D sinh t √ i cosh t + sinh t
ht · i D = sinh t
√ = D·
√ · i D + cosh t i sinh t + cosh t
D
√
the (Γ ∩ H)\H-orbit of the point zo = i D ∈ H in Γ\H is
/n√ i cosh t + sinh t o
Y = (Γ ∩ H)\{h · zo : h ∈ H} = (Γ ∩ H) D· : t∈R
i sinh t + cosh t
√
In fact, for to > 0 such that eto is the smallest non-trivial unit a + b D > 1, the corresponding matrix hto
is in Γ ∩ H, and so closes-up the line swept out by H in its action on zo ∈ H.
[3.0.1] Theorem: Take√D > 0 square-free,

√ D = 2 mod 4 or D = 3 mod 4, so that the ring of algebraic
integers o = ok in k = Q( D) is o = Z[ D]. Suppose for simplicity that o is a principal ideal domain. Then
the geodesic period is √ s s
D Γ( 2 ) Γ( 2s ) ζk (s)
Z
s
Es (h · zo ) dh = 2 ·
(Γ∩H)\H Γ(s) ζ(2s)
Proof: The period is

Z Z X
Es (h · zo ) dh = Im (γh(zo ))s dh
(Γ∩H)\H (Γ∩H)\H γ∈(Γ∩P )\Γ
X Z X
= Im (αβh(zo ))s dh
α∈(Γ∩P )\Γ/(Γ∩H) (Γ∩H)\H β∈(α−1 (Γ∩P )α∩H)\(Γ∩H)
Since eigenvalues of elements of Γ ∩ P are integers, and eigenvalues of (non-trivial) elements of Γ ∩ H are
(non-trivial) units in o,
α−1 (Γ ∩ P )α ∩ H = {±1}
and
Z X Z X
Es (h · zo ) dh = Im (αβh(zo ))s dh
(Γ∩H)\H α∈(Γ∩P )\Γ/(Γ∩H) (Γ∩H)\H β∈{±1}\(Γ∩H)
Im (hzo )s
X Z X Z
s
= Im (αh(zo )) dh = dh (unwinding!)
{±1}\H {±1}\H |c · hzo + d|2s
α∈(Γ∩P )\Γ/(Γ∩H) coprime (c,d)/(Γ∩H)
Im (hzo )s
Z
1 X
= dh (removing coprimality condition on c, d)
2ζ(2s) {±1}\H |c · hzo + d|2s
{(c,d)6=(0,0)}/(Γ∩H)
√ √

a Db
Identifying (c, d) ∈ Z2 with c D + d is compatible with the identification a + b D → , in the
b a
sense that
14
√ √

Db a
(c D + d) × (a + b D) −→ ( c d)
a b
√ √ √
= ( ca + db cbD + da ) ←− (ca + db) D + (cbD + da) = (c D + d) · (a + b D)
Thus, {(c, d) 6= (0, 0)}/(Γ ∩ H) is identified with (o − 0)/o1 , where o1 ⊂ o× is the group of units with norm
1 (as opposed to −1). Rearrange a little:
√ s
Im (hzo )s D
= √ 2s
|c · hzo + d|2s
c D(i cosh t + sinh t) + d(i sinh t + cosh t)

and
√ 2 √ √
c D(i cosh t + sinh t) + d(i sinh t + cosh t) = (c D cosh t + d sinh t)2 + (c D sinh t + d cosh t)2

√
= c2 D(cosh2 t + sinh2 t) + 4cd D cosh t sinh t + d2 (sinh2 t + cosh2 t)
√ √
= 21 · (c2 D + 2cd D + d2 )e2t + (c2 D − 2cd D + d2 )e−2t
The integral is the seemingly-daunting

√ s
2s D dt
Z
√ √ s
R (c2 D + 2cd D + d2 )e2t + (c2 D − 2cd D + d2 )e−2t
Up to constants, this is of the form Z

dt
R (Ae2t + Be−2t )s
Recall the trick
Z ∞ Z ∞
du du
y −s · Γ(s) = y −s us e−u = us e−u·y (for y > 0)
0 u 0 u
Thus, Z Z Z ∞
dt 2t
+Be−2t ) du
Γ(s) 2t + Be−2t )s
= us e−u(Ae dt
R (Ae R 0 u
log v
Letting t = 2 makes this
Z ∞ Z ∞ Z ∞ Z ∞
−u(Av+Bv −1 ) du dv 2 du dv
1
2
s
u e = 1
2 us v s e−u(Av +B)
0 0 u v 0 0 u v
√
by replacing u by uv. Replacing v by v makes this
Z ∞ Z ∞ Z ∞ Z ∞
1 s du dv 1 s s du dv
us v 2 e−u(Av+B) = u 2 v 2 e−(Av+Bu)
4 0 0 u v 4 0 0 u v
by replacing v by v/u. Replacing v by v/A and u by u/B gives
Γ( 2s ) Γ( 2s ) − s − s
Z
dt
−2t
= A 2B 2
R (Ae2t + Be ) s 4 · Γ(s)
Thus, the seemingly-daunting integral is
15
√ s Γ( 2s ) Γ( 2s ) 2 √ s √ s √ s Γ( s ) Γ( s ) 1
2s D (c D + 2cd D + d2 )− 2 (c2 D − 2cd D + d2 )− 2 = 2s D 2 2
4 · Γ(s) 4 · Γ(s) |c2 D − d2 |s
Thus, up to a constant factor, the geodesic period is

Z √ s Γ( s ) Γ( s ) 1 X 1
Es (h · zo ) dh = 2s D 2 2
(Γ∩H)\H 4 · Γ(s) ζ(2s) |c D − d2 |s
2
c,d
where the sum is, in effect, over integers modulo units, as in the Dedekind zeta for a principal ideal domain.
That is, √ s s
D Γ( 2 ) Γ( 2s ) ζk (s)
Z
Es (h · zo ) dh = 2s ·
(Γ∩H)\H Γ(s) ζ(2s)
The Gamma factors in the numerator are the appropriate ones for the real quadratic field k, and the
Gamma factor in the denominator is appropriate for ζ(2s). The corresponding powers of π happen to cancel
completely in the ratio. ///
4. Cuspidal waveforms, Fourier expansions

The idea of a waveform cuspform f for Γ = SL2 (Z) is that f is Γ-invariant and vanishing constant term: [12]
Z 1
f (x + iy) dx = 0
0
We may also require that f ∈ L2 (Γ\H) with respect to the invariant measure dxy2dy , or, alternatively, be of
moderate growth. Any similar conditions characterize vector spaces of cuspforms with additional properties.
Relations among the additional properties are typically non-trivial, so should not be expected to be obvious
or elementary.
We may also require that f be a ∆ = ∆H eigenfunction. Linear combinations of eigenfunctions with different
eigenvalues usually fail to produce eigenfunctions, so the eigenfunction requirement is often weakened to a
condition that does allow finite sums, namely, ∆-finiteness: f is ∆-finite when the ideal If ⊂ C[∆] of
polynomials in ∆ annihilating f is not {0}, equivalently, C[∆]/If is finite-dimensional over C.
Exactly which properties are assumed of a cuspform R 1 can only be determined from context. The essential
point is vanishing of the constant term cp f (iy) = 0 f (x + iy) dx.
[4.0.1] Theorem: Let f ∈ C ∞ (Γ\H) be a ∆-eigenfunction with eigenvalue λ, and of moderate growth in
the sense that f (x + iy) y A as y → +∞, for some exponent A. Assume the cuspidal condition cP f = 0.
Then f has a Fourier expansion
X
f (x + iy) = cn uλ (|n|y) e2πinx
n6=0

where y → uλ (y) is the unique (up to scalar multiples) solution of u00 − λ
y2 + 4π 2 n2 · u = 0 going to 0 as
y → +∞. [13] Such a cuspform f is of exponential decrease as y → +∞, in the sense that f (x+iy) f e−2πy
as y → +∞. The eigenvalue λ is non-positive real. The numerical coefficients cn are bounded.
[12] A minor technical issue is that for f merely measurable, for example, the integral defining the constant term c f
P
is not defined everywhere. That is, the sense of cP f = 0 may require elaboration, depending on context.
[13] This function u is known in the special-functions literature as essentially a Bessel function, and the differential
λ
equation it satisfies is essentially a Bessel equation. Many special things are known about these functions, but it is
preferable to develop their properties from more-standard ideas, as we attempt.
16
Proof: The smoothness assumption guarantees a nicely-converging Fourier expansion in x, since f is

invariant under z → z + 1: X
f (x + iy) = Cn (y) e2πinx
n∈Z
for some functions Cn (y). The cuspidal condition is C0 (y) = 0. The assumption of moderate growth
f (x + iy) y A implies that the coefficient functions Cn are of moderate growth:
Z 1 Z 1
Cn (y) = e−2πinx f (x + iy) dx y A dx = y A
0 0
The condition ∆f = λ · f gives

X X
y 2 Cn00 (y) e2πinx + Cn (y) (2πin)2 e2πinx = λ · Cn (y) e2πinx
n6=0 n6=0
By uniqueness of Fourier expansions, this is equivalent to

y 2 Cn00 (y) − 4π 2 n2 Cn (y) = λ · Cn (y) (for all n ∈ Z)
That is, Cn is a solution to [14]

λ
u00 − + 4π 2 2
n ·u = 0
y2
By changing variables, we see that for u1 a solution with n = 1, y → u1 (|n|y) is a solution for general n 6= 0.
Thus, to understand solutions for general n 6= 0 it suffices to treat n = 1.

To understand the behavior of solutions of u00 − yλ2 + 4π 2 · u = 0 as y → +∞, a natural heuristic is to
replace the coefficients of the equation by their limiting values as y → +∞, sometimes termed freezing the
equation (at y = +∞). As discussed in detail in the supplement on differential equations with irregular
singular points, this heuristic is correct: for n 6= 0, there is a unique (up to scalars) solution uλ on (0, +∞)
with asymptotic
uλ (y) ∼ e−2πy
and another solution vλ with vλ (y) ∼ e2πy . Any linear combination of uλ , vλ non-trivially involving vλ blows
up exponentially. Since Cn (y) is of moderate growth, up to scalars it must be the rapidly-decreasing solution
uλ (|n|y).
Thus, there are constants cn such that
X
f (x + iy) = cn uλ (|n|y) e2πinx
n6=0
The moderate growth condition gives

Z 1 Z 1
cn uλ (|n|y) = e−2πinx f (x + iy) dx y A dx = y A (as y → +∞)
0 0
and the asymptotics of uλ give a bad initial estimate for large yo :
yoA yoA
|cn | yo λ −2π|n|y yoA e2π|n|yo
uλ (|n|yo ) e o
[14] For n = 0 the differential equation becomes u00 − λ · u = 0. This is Eulerian, in the sense that it has solutions
y2
of the form u(y) = y s , where s ∈ C can be found by substituting and solving the resulting equation s(s − 1) − λ = 0
for s in terms of λ.
17
Then
X X 2 e−2π(y−yo )
|f (x + iy)| yoA e2π|n|yo · e−2π|n|y e−2π|n|(y−yo ) = e−2πy
n6=0 n6=0
1 − e−2π(y−yo )
Since f is exponentially decreasing as y → +∞, integration by parts twice leaves no boundary terms, and
Z Z Z Z
dx dy dx dy dx dy dx dy
λ· f ·f 2
= ∆f · f 2
= f · ∆f 2
= λ · f ·f
Γ\H y Γ\H y Γ\H y Γ\H y2
Thus, λ ∈ R. Further, integration by parts once gives

Z Z ∂f 2 ∂f 2
dx dy
λ· f ·f = − − dx dy ≤ 0
Γ\H y2 Γ\H ∂x ∂y
so λ ≤ 0. Since f is bounded on a fundamental domain and Γ-invariant, it is bounded on H, and

Z 1 Z 1
|cn uλ (|n|y)| = e−2πinx f (x + iy) dx 1 dx = 1

0 0
and |cn | 1/|uλ (|n|y)| for all y > 0. Taking y = yo /|n| for some yo such that uλ (yo ) 6= 0 gives
1 1
|cn | yo = λ 1
|uλ (|n| · |n| )| uλ (yo )
5. Fourier expansions of Eisenstein series

The explicit and relatively elementary nature of Eisenstein series gives further information about the solution
uλ to the eigenfunction equation y 2 (u00 − 4π 2 u) = λ · u, in the course of determining the Fourier expansion
of Es (x + iy) in x.
[5.1] Fourier expansion of Es We already computed the 0th Fourier component: it is

1
ξ(2s − 1) 1−s
Z
Es (x + iy) dx = y s + y (with ξ(2s) = π −s Γ(s) ζ(2s))
0 ξ(2s)
As on other occasions, it is easiest to treat

X ys
2 ζ(2s) Es (z) =
|cz + d|2s
(0,0)6=(c,d)∈Z2
For 0 6= n ∈ Z, the nth Fourier component is

Z 1
2 ζ(2s) · e−2πinx Es (x + iy) dx
0
Since n 6= 0, the c = 0 subsum does not contribute to this. In the region of convergence of the sum expression
for Es , we break the sum into fragments stable under x → x + 1:
X 1 X 1 X 1 X X 1
2 ζ(2s) · Es (x + iy) = y s d
= ys
c6=0
|c|2s
d∈Z
|x + c + iy|2s
c6=0
|c| 2s
d mod c `∈Z
|x + ` + dc + iy|2s
18
The inner sum over ` ∈ Z is periodic in x. Its nth Fourier component is

Z 1 Z
−2πinx
X 1 1
e d
dx = e−2πinx d
dx (by unwinding)
0
`∈Z
|x + ` + c + iy|2s R |x + c + iy|2s
The sum over d mod c is

X Z 1 X Z
1
e−2πinx d
dx = e 2πind/c
e−2πinx dx
R |x + c + iy|2s R |x + iy|2s
d mod c d mod c
The sum over d is |c| for c|n and 0 otherwise, by the cancellation lemma.
Unlike the holomorphic Eisenstein series, the latter integral is not an elementary function for n 6= 0.
Nevertheless, it can be rearranged to give an expression that meromorphically continues in s, and shows
other symmetries. First, replacing x by xy,
Z Z Z
1 1 1
e−2πinx 2s
dx = y 1−2s
e −2πinxy
2s
dx = y 1−2s
e−2πinxy 2 dx
R |x + iy| R |x + i| R (x + 1)s
Using the trick

Z ∞ Z ∞
dt dt
A−s · Γ(s) = A−s · ts e−t = ts e−tA (for A > 0)
0 t 0 t
we have Z Z Z ∞
1 1 2 dt
e−2πinxy dx = e −2πinxy
ts e−t(x +1) dx
R (x2 + 1)s Γ(s) R 0 t
Changing the order of integration, and using the good behavior of suitable Gaussians under Fourier transform,
Z
2 2
e−2πiξx e−πx dx = e−πξ
R
√
replace x by x · √π and take the Fourier transform, to obtain
t
√ Z ∞ Z √ √ Z ∞
π s− 12 −t −2πinxy √
π
−πx2 dt π 1 π 2 dt
t e e t e dx = ts− 2 e−t−π· t (ny)
Γ(s) 0 R t Γ(s) 0 t
Replacing t by t · π|n|y gives

√ 1 ∞
π · (π|n|y)s− 2
Z
1 1 dt
ts− 2 e−(t+ t )π|n|y
Γ(s) 0 t
Restoring the factors of y s and y 1−2s dropped along the way, this is
1 √ Z ∞
π s |n|s− 2 y 1 1 dt
Γ(s) 0 t
The integral is invariant under s → 1 − s, by replacing t by 1/t, and is nicely convergent for all s ∈ C.
Thus, the nth Fourier component of 2 ζ(2s) Es (x + iy) is
X 1 π s |n|s− 12 √y Z ∞ 1 1 dt
nth Fourier component 2 ζ(2s) Es (x + iy) = 2s−1
|c| Γ(s) 0 t
c|n
19
Summing over just the positive divisors of n, replacing c by |n|/c, and dividing through by 2 ζ(2s) gives
√ Z
th σ2s−1 (|n|) π s y ∞ s− 1 −(t+ 1 )π|n|y dt
n Fourier component Es (x + iy) = 1 · t 2 e t
ζ(2s) |n|s− 2 Γ(s) 0 t
where σ2s−1 (|n|) is the sum of (2s − 1)th powers of positive divisors. Altogether,
ξ(2s − 1) 1−s 1 X σ2s−1 (|n|) √ Z ∞ 1 1 dt 2πinx
Es (x + iy) = y s + y + −s s− 21
· y ts− 2 e−(t+ t )π|n|y ·e
ξ(2s) π Γ(s) ζ(2s) |n| 0 t
n6=0
[5.1.1] Remark: The Fourier expansion can be used to give another proof of the meromorphic continuation
and functional equation of Es .
[5.2] Integral representation of uλ Let λ = λs = s(s − 1). Crude estimates show that
Z ∞
√ 1 1 dt
vλ (|n|y) = y ts− 2 e−(t+ t )π|n|y
0 t
is exponentially decreasing as y → +∞. Since ∆Es = s(s − 1) · Es , separation of variables in the Fourier
expansion shows that vλ integral is a solution of the differential equation
y 2 (u00 − 4π 2 u) = λ · u
The earlier discussion of the asymptotics at ∞ of the solutions of this differential equation observed that up
to scalars there is a unique solution asymptotic to e−2πy , while all other solutions blow up like e2πy . Thus,
up to a normalizing constant, this integral expression vλ is the function uλ appearing in Fourier expansions
of cuspforms with eigenvalue λ.
6. Standard L-functions attached to cuspforms

A natural, naive normalization Lnf (s, f ) of L-functions attached to cuspidal waveforms f is proven to have
analytic continuation and functional equation. With hindsight, the set-up is renormalized for compatibility
with other conventions.
[6.1] Naive normalization, analytic continuation and functional equation The naive form of the
standard L-function attached to a cuspidal waveform f has two descriptions. When f is presumed a λ-
eigenfunction for ∆, it has an expansion
X
f (x + iy) = cn uλ (|n|y) e2πinx (where uλ solves y 2 (u00 − 4π 2 u) = λ · u)
n6=0
Assume for simplicity that f is even in x, meaning f (−x + iy) = f (x + iy), so cn = c|n| , and no information
is lost in putting X cn
Lnf (s, f ) =
ns
n≥1
For λ < − 41 , the bound cn ε 1 proven above guarantees convergence of this Dirichlet series in every right
half-plane Re (s) > 1 + ε. The superscript indicates that this normalization is naive, although arguably
natural.
On the other hand, if we can justify application of the usual integral transform, then
Z ∞ X Z ∞ X cn Z ∞ Z ∞
dy dy dy dy
y s f (iy) = cn y s uλ (n y) = s
· y s
uλ (y) = L nf
(s, f ) · y s uλ (y)
0 y 0 y n 0 y 0 y
n≥1 n≥1
20
By now, we know that a cuspidal ∆-eigenfunction f is rapidly decreasing, so the integral behaves well at +∞.
To be sure of sufficiently good behavior as y → 0+ , note that the differential equation y 2 u00 −(λ+4π 2 y 2 )u = 0
for uλ has a regular singular point at y = 0, meaning that it is of the form y 2 u00 + yb(y)u0 + c(y)u = 0 with
b, c holomorphic at y = 0. To express uλ (y) = y s · ϕ(y) with ϕ holomorphic at y = 0, solve
0 = s(s − 1) + b(0)s + c(0) = s(s − 1) − λ
for the two values of s which make this possible. For λ ≤ 41 , s = 12 ± iν for some ν ∈ R. Thus, for some
constants A, B,
1 1 √
|uλ (y)| ∼ |Ay 2 +iν + By 2 −iν | = y · |Ay iν + By −iν | (as y → 0+ )
Thus, for example,

√
|uλ (y)| y (for 0 < y ≤ 1)
and |uλ (y)| e−2πy for y ≥ 1, from above. Also, from above, |cn | 1. Thus,
1
X X p X √
Z y √
cn uλ (|n|y) |n|y + e−2π|n|y y t dt + e−2πy

n 1
|n|≤ y1 |n|> y1
√ 1 1
y· = (as y → 0+ )
y 3/2 y
Thus, the integral converges nicely for Re (s) > 1.
Below, we will show that the resulting integral of uλ (y) is a product of Gamma functions: after renormalizing
by constants depending only on λ, not s,
Z ∞ s + µ s + 1 − µ
dy
y s uλ (y) = π −s Γ Γ
0 y 2 2
Granting that, the form of the functional equation disclosed by the following theorem provides motivation
to renormalize:
[6.1.1] Theorem: For an even cuspidal waveform f with ∆H f = λ · f with λ = µ(µ − 1), the completed
L-function s + µ s + 1 − µ
Λnf (f, s) = π −s Γ Γ · Lnf (s, f )
2 2
has an analytic continuation to an entire function, with functional equation
Λnf (−s, f ) = Λnf (s, f )
Proof: As in Riemann’s argument with θ(iy) in place of f (iy), the integral representation will yield the
analytic continuation, beginning by breaking the integral into two pieces:
Z ∞ Z ∞ Z 1
dy dy dy
Λnf (s, f ) = y s f (iy) = y s f (iy) + y s f (iy)
0 y 1 y 0 y
The estimate f (x + iy) f e−2πy from the previous section shows that the integral from 1 to ∞ converges
nicely for all s ∈ C, and gives an entire function.
The functional equation f (−1/z) = f (z) gives f (i y1 ) = f (iy), and converts the integral from 0 to 1 to an
integral from 1 to ∞: replacing y by 1/y,
Z 1 Z ∞ Z ∞
dy 1 dy dy
y s f (iy) = y −s f (i ) = y −s f (iy)
0 y 1 y y 1 y
21
The latter integral is entire, by the decay estimate f (iy) f e−2πy as y → +∞. Thus,
Z ∞
dy
Λnf (s, f ) = y s + y −s f (iy)

1 y
expresses Λnf (s, f ) as an entire function with visible symmetry under s → −s. ///
[6.2] Standard normalization and functional equation We might prefer the functional equation to be
s → 1 − s rather than s → −s, especially if this can be arranged so that the correct notion of critical strip
for these L-functions is 0 ≤ Re (s) ≤ 1. As we verify later, to achieve this effect put
X cn X cn · √n
L(f, s) = 1 =
ns− 2 ns
n≥1 n≥1
and, correspondingly,
Z ∞ s − 1 + µ s + 1 − µ
1 dy
Λ(s, f ) = y s− 2 f (iy) = π −s Γ 2
Γ 2
L(s, f )
0 y 2 2
With this normalization, the functional equation is s → 1 − s.
[6.3] The gamma factor The explicit nature of the Eisenstein series suggests an approach to
understanding the Gamma factor Z ∞
dy
y s uλ (y)
0 y
by using the integral representation for uλ obtained in the course of determining the Fourier expansion of
Es (z). Proceeding directly, putting λ = µ(µ − 1), up to a normalizing constant depending on µ, invoking
the Gamma-function trick,
∞ ∞
yµ
Z Z Z
dy dy
y s uλ (y) = ys e−2πix dx
0 y 0 R (x2 + y 2 )µ y
Z ∞Z Z ∞ √ Z ∞Z ∞
1 2
+y 2 ) dy dt π 1 π2 2 dy dt
= y s+µ e−2πix tµ e−t(x dx = y s+µ tµ− 2 e−( t +ty )
Γ(µ) 0 R 0 y t Γ(µ) 0 0 y t
√
Replacing y by y, and then y by y/t, gives
√ Z ∞ Z ∞ √ Z ∞Z ∞
π s+µ 1 π2 dy dt π s+µ 1 s+µ π2 dy dt
y 2 tµ− 2 e−( t +ty)
= y 2 tµ− 2 − 2 e−( t +y)
2 Γ(µ) 0 0 y t 2Γ(µ) 0 0 y t
√ ∞ √
π s + µ −(s+1−µ) s + 1 − µ
Z
π s + µ µ 1 s
− πt
2 dt
= Γ t 2 −2−2 e = Γ π Γ
2Γ(µ) 2 0 t 2Γ(µ) 2 2
Since we are at liberty to adjust by things depending on µ but not on s, we can renormalize uλ by constants
depending only on λ = µ(µ − 1) to achieve the effect that
Z ∞ s + µ s + 1 − µ
dy
y s uλ (y) = π −s Γ Γ (naive normalization)
0 y 2 2
Renormalizing by replacing s by s − 12 , and adjusting by a constant, makes this

Z ∞ s − 1 + µ s + 1 − µ
1 dy
y s− 2 uλ (y) = π −s Γ 2
Γ 2
(renormalization)
0 y 2 2
22
[6.3.1] Remark: The above computation is sensible when all the integrals converge nicely. Unfortunately,
that regime is not where values of λ = µ(µ − 1) are non-positive real, corresponding to L2 (Γ\H)-
eigenfunctions. The region of interest is reached only by analytic continuation.
[6.3.2] Remark: Cuspforms whose L-functions have Euler products are distinguished by Hecke operators,
in a fashion essentially identical to holomorphic modular forms.
[6.3.3] Remark: Hecke operators on waveforms commute with ∆ = ∆H , so stabilize each λ-eigenspace for
∆ in L2 (Γ\H). Conjecturally, for Γ = SL2 (Z), for each λ appearing as an eigenvalue, the λ-eigenspace is
one-dimensional. Thus, conjecturally, every square-integrable ∆-eigenfunction is also a Hecke eigenfunction.
7. Good argument for non-vanishing of ζ(1 + it)

Non-vanishing of ζ(s) on the edge Re (s) = 1 of the critical strip 0 ≤ Re (s) ≤ 1 is essentially equivalent to
the prime number theorem. Similarly, non-vanishing of other Euler products on the edges of the critical strip
is critical in many number-theoretic applications.
Clever ad hoc arguments for this non-vanishing, using trigonometric identities, can succeed for the simplest
classes of Euler products, but are inadequate in general. Even in simple cases, the old non-vanishing
arguments are unilluminating.
Instead, the appearance of ζ(2s) in the denominators of the Fourier components of the Eisenstein series
Es for SL2 (Z) gives a meaningful argument that generalizes to many other situations. From the Fourier
expansion obtained above,
ξ(2s − 1) 1−s 1 X σ2s−1 (|n|) √ Z ∞ 1 1 dt 2πinx

s
Es (x + iy) = y + y + −s 1 · y ts− 2 e−(t+ t )π|n|y ·e
ξ(2s) π Γ(s) ζ(2s) |n|s− 2 0 t
n6=0
a zero 2so = 1 + 2ito of ζ(s) on Re (s) = 1 would give a zero so = 21 + ito of the ζ(2s) appearing in the
denominators. In the constant term, because of the functional equation,
ξ(2so − 1) = ξ(2 − 2so ) = ξ(1 − 2ito ) = ξ(1 + 2ito ) = 0
so the constant term has no pole at so . Thus, for ζ(1 + 2ito ) = 0, the residue of Es at so would be a
cuspform f , and an eigenfunction with ∆f = so (so − 1) · f . Thus, f would be of rapid decay, and its integral
against any Eisenstein series Es would converge nicely, even for Es meromorphically continued, since the
meromorphic continuation is still of moderate growth, meaning Es (x + iy) y A for some exponent A as
y → +∞. On one hand, for Re (s) > 1, unwinding Es ,
Z Z Z
dx dy s dx dy s
X dx dy
Es · f 2
= y f 2
= y cn (y) e2πinx
Γ\H y Γ∞ \H y Γ∞ \H y2
n6=0
where we write the nth Fourier component as e2πinx with coefficient a function cn (y) of y. A great virtue of
the unwinding to an integral over Γ∞ \H is that the latter quotion has very nice representatives allowing a
separation of variables: a fundamental domain for Γ∞ is [0, 1] × (0, +∞) in the x, y coordinates. The basic

Z 1  1 (for m = 0)
e2πimx dx =
0 
0 (for m 6= 0)
and the vanishing of the 0th Fourier component of a cuspform f gives

Z Z
dx dy dy
Es · f 2
= ys 0 2 = 0
Γ\H y N \H y
23
Yet, on the other hand, presuming that evaluation-of-residue commutes with the integral [15]
Z Z Z
ress=so Es · f = ress=so Es · f = f ·f
Γ\H Γ\H Γ\H
Thus, f = 0, there is no such residue, and no such zero so = 1 + 2ito on the edge of the critical strip.
[7.0.1] Remark: The residue of Es at s = so can be understood as the collection of residues pointwise
ress=so Es (zo ) for all values zo , since the function s → Es (zo ) is meromorphic in s for every fixed zo . This
would not promise anything about the collection of pointwise residues. However, we have further information
about the collection of pointwise residues from the explicit Fourier expansion. Even in situations without
explicit formulas, we can use the notion of function-valued meromorphic function s → Es , and rely upon
the good behavior of residue, that is, that taking residues commutes with application of continuous linear
functionals of the form Z
Es −→ Es · F
Γ\H
with F of rapid decay on the fundamental domain to give a (continuous) linear functional on moderate-growth
functions such as Es . This viewpoint is indispensable in more complicated, less-formulaic situations.
[15] Yes, evaluation-of-residue commutes with such integrals. This is a special case of the general good behavior of
holomorphic function-valued functions. This idea was more-than-fully developed by Grothendieck, building upon
Schwartz’ distribution theory. The happy news is that things turn out as well as possible. The lack of dramatic
tension, and opaque exposition, may account for the unfortunate lack of wider awareness of these results. Thanks to
Bill Casselman some years ago for bringing the Grothendieck papers to my attention!
24
Bibliography
[Bombieri-Hejhal 1995] E. Bombieri, D. Hejhal, On the distribution of zeros of linear combinations of Euler
products, Duke J. Math. 80 (1995), 821-862.
[Chowla-Selberg 1967] S. Chowla, A. Selberg, On Epstein’s zeta-function, J. reine und angew. Math. 227
(1967), 86-110.
[Davenport-Heilbronn 1936] H. Davenport, H. Heilbronn, On the zeros of certain Dirichlet series, I, II, J.
London Math. Soc. 11 (1936), 181-185 and 307-312.
[Epstein 1903] P. Epstein, Zur Theorie allgemeiner Zetafunktionen, I, Math. Ann. 56 (1903), 614-644.
[Godement 1966a] R. Godement, Decomposition of L2 (Γ\G) for Γ = SL2 (Z), in Proc. Symp. Pure Math.
9 (1966), AMS, 211-24.
[Grothendieck 1953a,b] A. Grothendieck, Sur certaines espaces de fonctions holomorphes, I, J. Reine Angew.
Math. 192 (1953), 35-64; II, 192 (1953), 77-95.
[Hecke 1918/20] E. Hecke, Eine neue Art von Zetafunktionen und ihre Beziehungen der Verteilung der
Primzahlen, I, II, Math. Z. 1 no. 4 (1918), 357-376; ibid 6 no. 1-2 (1920), 11-51.
[Hecke 1937] E. Hecke, Über Modulfunktionen und die Dirichletschen Reihen mit Eulerscher Produktentwick-
lungen, I, II, Math. Ann. 114 (1937) no. 1, 1-28, 316-351.
[Maaß 1949] H. Maaß, Über eine neue Art von nicht analytischen automorphen Funktionen und die
Bestimmung Dirichletscher Reihen durch Funktionalgleichungen, Math. Ann. 121 (1949), 141-183.
[Potter-Titchmarsh 1935] H.S. Potter, E.C. Titchmarsh, The zeros of Epstein’s zeta functions, Proc. London
Math. Soc. 39 (1935), 372-384.
[Rankin 1939] R. Rankin, Contributions to the theory of Ramanujan’s function τ (n) and similar arithmetic
functions, I, Proc. Cam. Phil. Soc. 35 (1939), 351-372.
[Schwartz 1950/51] L. Schwartz, Théorie des Distributions, I,II Hermann, Paris, 1950/51, 3rd edition, 1965.
[Selberg 1940] A. Selberg, Bemerkungen über eine Dirichletsche Reihe, die mit der Theorie der Modulformen
nahe verbunden ist, Arch. Math. Naturvid 43 (1940), pp. 47-50.
[Stark 1967] H.M. Stark, On the zeroes of Epstein’s zeta function, Mathematika 14 (1967), 47-55.
[Voronin 1976] M. Voronin, On the zeros of zeta functions of quadratic forms, Trudy Mat. Inst. Steklov 142
(1976), 135-147.
25

11 Waveforms

Uploaded by

Copyright:

Available Formats

You might also like

11 Waveforms

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

11 Waveforms

Uploaded by

Copyright:

Available Formats

(November 26, 2022)

the simplest binary theta series

and binary theta series

1. Waveform Eisenstein series Es

[1.1] Invariant Laplacian The differential operator

The SL2 (R) invariant measure on H is

since, integrating once by parts,

Indeed, similar iterated-integral identities are essential to unwinding arguments.

[1.2] Eisenstein series The Eisenstein series Es for Γ is

For Re (s) > 1 the series defining

(cx + d)2 + (cy)2 = (x2 + y 2 )c2 + 2x · cd + d2 ≥ c2 − |cd| + d2 ≥ 1 2

ξ(2s) Es = ξ(2 − 2s) E1−s

Es (z) s y A (as y → +∞)

Es (z)  t1−Re (s)+ε (for 1

Proof: For v = (c d) ∈ R2 , consider the Gaussian

where v ∈ R2 is a row vector. Consider the integral (a Mellin transform)

We want g ∈ SL(2, R) to map i → x + iy. One reasonable choice is

Using this choice of g and writing out v = (c, d) gives

Letting 1 ≤ δ = gcd(c, d), this is

is convenient, being a sum over a lattice with 0 removed.

Fourier transform of (v → ϕ(tvg)) = (v → t−2 det(g)−1 · ϕ(t−1 v >g −1 ))

where >g is g-transpose. Then Poisson summation asserts

Θ(tg) = t−2 det(g)−1 · Θ(t−1 >g −1 )

Noting det g = 1, the modification for the kernel gives

Θ(tg) − 1 = t−2 · [Θ(t−1 >g −1 ) − 1] + t−2 − 1

Replacing t by 1/t turns this into

Θ(g) = Θ(wg) = Θ(gw−1 )

and the whole equality, with g of the special form above, is

y −(1−s) π −s Γ(s) y −(1−2s) π −s Γ(s)

That is, away from poles,

ξ(2s) Es (z) ξ(2 − 2s) E1−s (z)

Thus, Es (z) is of moderate growth for all s. ///

[1.4.1] Claim: The constant term of Es is

Proof: By direct computation,

Replacing x by xy in the integral gives

Dividing by 2ζ(2s) and re-attributing powers of π gives

2. Heegner point periods of Eisenstein series

F str = {z = x + iy ∈ H : |z| ≥ 1, − 21 < x ≤ 21 , and |z| > 1 if x < 0}

for SL2 (Z) on H, such that

The quadratic form

N (me1 + ne2 ) · a−1 = Am2 + Bmn + Cn2

Change-of-basis in the Z-module a−1 changes the

3. Closed geodesic periods of Eisenstein series

For brevity, write  √ 

[3.0.1] Theorem: Take√D > 0 square-free,

Proof: The period is

The integral is the seemingly-daunting

Up to constants, this is of the form Z

by replacing v by v/u. Replacing v by v/A and u by u/B gives

Thus, the seemingly-daunting integral is

Thus, up to a constant factor, the geodesic period is

4. Cuspidal waveforms, Fourier expansions

Proof: The smoothness assumption guarantees a nicely-converging Fourier expansion in x, since f is

The condition ∆f = λ · f gives

By uniqueness of Fourier expansions, this is equivalent to

That is, Cn is a solution to [14]

The moderate growth condition gives

Es (z) s y A (as y → +∞)

Es (z) t1−Re (s)+ε (for 1

For brevity, write √