Bernstein and Reznikov Subconvexity
Bernstein and Reznikov Subconvexity
REPRESENTATION THEORY
1. Introduction
1.1. Maass forms. Let H denote the upper half plane equipped with the standard
Riemannian metric of constant curvature −1. We denote by dv the associated volume
element and by ∆ the corresponding Laplace-Beltrami operator on H.
Fix a discrete group Γ of motions of H and consider the Riemann surface Y = Γ\H. For
simplicity we assume that Y is compact (the case of Y of finite volume is discussed at the
end of the introduction). According to the uniformization theorem any compact Riemann
surface Y with the metric of constant curvature −1 is a special case of this construction
Consider the spectral decomposition of the operator ∆ in the space L2 (Y, dv) of functions
on Y . It is known that the operator ∆ is non-negative and has a purely discrete spectrum;
we will denote the eigenvalues of ∆ by 0 = µ0 < µ1 ≤ µ2 ≤ ... .
For these eigenvalues, we always use a natural (from the representation-theoretic point of
1−λ2
view) parametrization µi = 4 i , where λi ∈ C. We denote by φi = φλi the corresponding
eigenfunctions (normalized to have L2 -norm one).
In the theory of automorphic forms, the functions φλi are called automorphic functions
or Maass forms (after H. Maass, [M]). The study of Maass forms plays an important
role in analytic number theory, analysis and geometry. We are interested in their analytic
properties and will present a new method of bounding some important quantities arising
from the φi .
A particular problem we are going to address in this paper belongs to an active area of
research in the theory of automorphic functions studying an interplay between periods,
1991 Mathematics Subject Classification. Primary 11F67, 22E45; Secondary 11F70, 11M26.
Key words and phrases. Representation theory, Periods, Automorphic L-functions, Subconvexity.
1
2 JOSEPH BERNSTEIN AND ANDRE REZNIKOV
special values of automorphic L-functions and representation theory. One of the central
features of this interplay is the uniqueness of invariant functionals associated to corre-
sponding periods. The discovery of this interplay goes back to classical works of E. Hecke
and H. Maass.
It is well-known that the uniqueness plays a central role in the modern theory of au-
tomorphic functions (see [PS]). The impact uniqueness has on the analytic behavior of
periods and L-functions is yet another manifestation of this principle.
1.2. Triple products. For any three Maass forms φi , φj , φk , we define the following
triple product or triple period:
Z
cijk = φi φj φk dv .
Y
1.3. Results. In this paper we consider the following problem. We fix two Maass forms
φ = φτ and φ′ = φτ ′ as above and consider the coefficients defined by the triple period:
Z
ci = φφ′ φi dv (1.2)
Y
as the φi run over an orthonormal basis of Maass forms.
Thus we see from (1.1) that the estimates of the coefficients ci are essentially equivalent
to the estimates of the corresponding L-functions. One would like to have a general
SUBCONVEXITY OF TRIPLE L-FUNCTIONS 3
method of estimating the coefficients ci and similar quantities. This problem was raised
by Selberg in his celebrated paper [Se].
The first non-trivial observation is that the coefficients ci have exponential decay in |λi |
as i → ∞. Namely, as we have shown in [BR2], it is natural to introduce normalized
coefficients
di = γ(λi )|ci |2 . (1.3)
Here γ(λ) is given by an explicit rational expression in terms of the standard Euler
Γ-function (see [BR2]) and, for purely imaginary λ, |λ| → ∞, it has an asymptotic
γ(λ) ∼ β|λ|2 exp( π2 |λ|) with some explicit β > 0. It turns out that the normalized
coefficients di have at most polynomial growth in |λi |, and hence the coefficients ci decay
exponentially. This is consistent with (1.1) and general experience from the analytic
theory of automorphic L-functions (see [BR2], [Wa]). In Section 4 we explain a more
conceptual way to introduce the coefficients di which is based on considerations from
representation theory.
In [BR2] we proved the following mean value bound
X
di ≤ AT 2 , (1.4)
|λi |≤T
[I]), this leads to the following subconvexity bound for the triple L-function (for more on
the relation between triple period and special values of L-functions, see [Wa]).
Corollary. Let φ and φ′ be fixed Hecke-Maass cusp forms. For any ε > 0, there exists
Cε > 0 such that the bound
L( 21 , φ ⊗ φ′ ⊗ φλi ) ≤ Cε |λi |5/3+ε (1.6)
holds for any Hecke-Maass form φλi .
The convexity bound for the triple L-function corresponds to (1.6) with the exponent
5/3 replaced by 2. We refer to [IS] for a discussion of the subconvexity problem which is
in the core of modern analytic number theory. We note that the above bound is the first
subconvexity bound for an L-function of degree 8. All previous subconvexity results were
obtained for L-functions of degree at most 4.
Recently, A. Venkatesh [V] obtained a subconvexity bound for the triple L-function in
the level aspect (i.e., with respect to a tower of congruence subgroups Γ(N) as N → ∞).
His method is quite different from the method we present in this paper and is based on
ergodic theory.
We formulate a natural
Conjecture. For any ε > 0 we have di ≪ |λi |ε .
For Hecke-Maass forms on congruence subgroups, this conjecture is consistent with the
Lindelöf conjecture for the triple L-functions (for more details, see [BR2] and [Wa]).
1.3.1. Remarks. 1. Our results can be generalized to the case of a general finite co-
volume lattice Γ ⊂ G. In this case the spectral decomposition of the Laplace-Beltrami
operator on Y = Γ\H is given by a collection of eigenfunctions φz , where the parameter z
runs through some set Z with the Plancherel measure dµ. The spectral set Z has discrete
points which correspond to eigenfunctions (Maass forms) φz ∈ L2 (Y ) and continuous
part which corresponds to eigenfunctions coming from the unitary Eisenstein series. The
collection {φz }z∈Z defines a transform û(z) =< u, φz > for every u ∈ Cc∞ (Y ). Main
property of this transform is the Plancherel formula ||u||2L2(Y ) = Z |û(z)|2 dµ.
R
Let us fix two Maass cusp forms φ and φ′ on Y . For every z ∈ Z we define the parameter
λz ∈ C and the coefficient dz in the same way as before. In this case we can prove the
bound Z
dz dµ ≤ BT 5/3 +ε , where ZT = {z ∈ Z | |λz | ∈ IT }
ZT
2. First results on the exact exponential decay of triple products for a general lattice Γ
were obtained by A. Good [Go] and P. Sarnak [Sa] using ingenious analytic continuation of
Maass form to the complexification of the Riemann surface Y (for representation-theoretic
approach to this method and generalizations, see [BR1] and [KS]). Our present method
seems to be completely different and avoids analytic continuation.
SUBCONVEXITY OF TRIPLE L-FUNCTIONS 5
3. We would like to stress that the bound for the triple product in Theorem 1.3 is
valid for a general lattice Γ, including non-arithmetic lattices. In fact, in our method we
do not use Fourier coefficients or Hecke eigenvalues through which one usually accesses
values of L-functions for congruence subgroups. Our method gives estimates for periods of
automorphic functions directly and L-functions appear only through the Watson formula
(1.1) (the same is true for the method of Venkatesh).
1.4. The method. We describe now the general ideas behind our proof. It is based on
ideas from representation theory (for a detailed account of the corresponding setting, see
[BR2] and Section 3 below). In what follows we sketch the method of the proof with the
complete details appearing in the rest of the paper.
1.4.1. Automorphic representations. Let G denote the group of all motions of H. This
group is naturally isomorphic to P GL2 (R) and as a G-space H is naturally isomorphic to
G/K, where K = P O(2) is the standard maximal compact subgroup of G.
By definition, Γ is a subgroup of G. The space X = Γ\G with the natural right action
of G is called an automorphic space. We will identify the Riemann surface Y = Γ\H with
X/K.
We use the standard language of automorphic representations (see [G6] and Section 2
below). Let (π, G, V ) be an irreducible smooth representation of G. An automorphic
structure on V is a continuous G-morphism ν : V → C ∞ (X).
The pair (π, ν) consisting of an abstract representation (π, V ) and the automorphic
structure ν will be called an automorphic representation. This terminology is slightly
more precise then the standard one. We find it more convenient for our purposes.
We will usually present the abstract representation (π, V ) by their explicit model. We
will deal mostly with class one irreducible representations of G (i.e., those with a non-
zero K-fixed vector). For such representations we use the model Vλ , where λ ∈ iR ∪ (0, 1)
and Vλ is the space of smooth even homogeneous functions on R2 \ 0 of homogeneous
degree λ − 1 (see [G5]). We always assume that (π, V ) is unitary (i.e., equipped with a
positive definite G-invariant Hermitian form P ), and that the automorphic structure ν
is compatible with P . We denote by eλ ∈ Vλ the function taking constant value 1 on
S 1 ⊂ R2 \ 0. This gives a K-invariant vector in the representation Vλ .
It is the theorem of Gelfand and Fomin that all Maass forms (or more generally au-
tomorphic functions) could be obtained as special vectors in appropriate automorphic
representations (see [G6]). In particular a Maass form φ corresponding to an automor-
2
phic representation with a model Vλ has the eigenvalue µ = 1−λ 4
. For λ ∈ iR we may
assume that φ = ν(eλ ); for λ ∈ (0, 1), we have to re-normalize eλ .
We translate various questions about Maass forms into corresponding questions about
associated automorphic representations. This allows us to employ powerful methods of
representation theory.
6 JOSEPH BERNSTEIN AND ANDRE REZNIKOV
1.4.3. Hermitian forms. In order to estimate the quantities di , we consider the space
E = Vτ ⊗ Vτ ′ and use the fact that the coefficients di appear in the spectral decomposition
of the following geometrically defined non-negative Hermitian form H∆ on E (for a detailed
discussion, see [BR2]).
Consider the space C ∞ (X × X). The diagonal ∆ : X → X × X gives rise to the
restriction morphism r∆ : C ∞ (X × X) → C ∞ (X). We define a non-negative Hermitian
form H∆ on C ∞ (X × X) by setting H∆ = (r∆ )∗ (PX ), where PX is the standard L2
Hermitian form on C ∞ (X) i.e.,
Z
H∆ (w) = PX (r∆ (w)) = |r∆ (w)|2dµX
X
SUBCONVEXITY OF TRIPLE L-FUNCTIONS 7
for any w ∈ C ∞ (X ×X). We call the restriction of the Hermitian form H∆ to the subspace
E ⊂ C ∞ (X × X) the diagonal Hermitian form and denote it by the same letter.
We will describe the spectral decomposition of the Hermitian form H∆ in terms of
Hermitian forms corresponding to trilinear functionals. Namely, if L is a pre-unitary
representation of G with G-invariant norm || ||L then every G-invariant trilinear functional
l : V ⊗ V ′ ⊗ L → C, defines a Hermitian form H l on E by H l (w) = sup |l(w ⊗ u)|2 .
||u||L =1
Of course, one can introduce similar model trilinear functionals for the discrete series
representations Vκ and the corresponding coefficients dκ via Hκaut = dκ Hκ . We will not
need these in this paper (in fact, we are trying to avoid computations with the discrete
series representations; see Remark 7.1).
We will mostly use the fact that for every vector w ∈ E this basic spectral identity gives
us an inequality
X
di Hλi (w) ≤ H∆ (w) (1.9)
i
which turns into an equality if the vector r∆ (w) does not have projection to discrete series
representations (for example, if the vector w is invariant with respect to the diagonal
action of K on E).
We can use this inequality to bound coefficients di . Namely, for a given vector w ∈ E we
usually can compute the values of the weight function Hλ (w) by explicit computations in
the model of representations V, V ′ , Vλ . It is usually much more difficult to get reasonable
8 JOSEPH BERNSTEIN AND ANDRE REZNIKOV
estimates of the right hand side H∆ (w) since it refers to the automorphic picture. In cases
when we manage to do this we get some bounds for the coefficients di .
1.4.4. Mean-value estimates. In [BR2], using the geometric properties of the diagonal
form
P and explicit estimates of forms Hλ , we established the mean-value bound (1.4):
di ≤ AT 2 . Roughly speaking, the proof of this bound is based on the fact that while
|λi |≤T
the value of the form H∆ on a given vector w ∈ E is very difficult to control, we can
show that for many vectors w the value H∆ (w) can be bounded by PE (w), where PE is
the Hermitian form which defines the standard unitary structure on E.
More precisely, consider the natural representation σ = π ⊗ π ′ of the group G × G on the
space E. Then for a given compact neighborhood U ⊂ G×G of the identity element, there
exists a constant C such that for any vector w ∈ E, the inequality H∆ (σ(g)w) ≤ CPE (w)
holds for at least half of the points g ∈ U. This follows from the fact that the average
over U of the quantity H∆ (σ(g)w) is bounded by CPE (w)/2.
This allows us for every T ≥ 1, to find a vector w ∈ E such that H∆ (w) ≤ CT 2 while
Hλ (w) ≥ c for all |λ| ≤ T .
1.4.5. Bounds for sums over shorter intervals. The main starting point of our approach
to the subconvexity bound is the inequality (1.9) for Hermitian forms. For a given T > 1,
we construct a test vector wT ∈ E such that the weight function λ 7→ Hλ (wT ) has a sharp
peak near |λ| = T (i.e., a vector satisfying the condition (1.12) below).
The problem is how to estimate effectively H∆ (wT ). The idea is that the Hermitian form
H∆ is geometrically defined and, as a result, satisfies some non-trivial bounds, symmetries,
etc. None of the explicit model Hermitian forms Hλ satisfies similar properties. By
applying these symmetries to the vector wT we construct a new vector w̃T and from the
geometry of the automorphic space X we deduce the bound H∆ (wT ) ≤ H∆ (w̃T ).
POn the other hand, the weight function Hλ (w̃T ) in the spectral decomposition H∆ (w̃T ) =
di Hλi (w̃T ) for w̃T behaves quite differently from the weight function Hλ (wT ) for wT .
Namely, the function Hλ (w̃T ) behaves regularly (i.e., satisfies condition (1.13) below),
while the weight function Hλ (wT ) has a sharp peak near |λ| = T .
The regularity of the function Hλ (w̃T ) coupled with the mean-value bound (1.4) allows
us to prove a sharp upper bound on the value of H∆ (w̃T ) by purely spectral considerations
(in cases we consider there is no contribution from discrete series). We do not see how
to get such sharp bound by geometric considerations working on the automorphic space
X × X.
Using this bound for H∆ (w̃T ) and the inequality H∆ (wT ) ≤ H∆ (w̃T ) we obtain a non-
trivial bound for H∆ (wT ) and, as a result, the desired bound for the coefficients di .
In next section we describe this strategy in more detail.
SUBCONVEXITY OF TRIPLE L-FUNCTIONS 9
1.5. Formulas for test vectors. Let us describe the construction of vectors wT , w̃T .
We assume for simplicity that V ′ ≃ V̄ – the complex conjugate representation; it is also
an automorphic representation with the realization ν̄ : V̄ → C ∞ (X). It is easy to see
that the upper bound estimate that we need in the general case can be reduced to this
special case (see Appendix D).
We only consider the case of representations of the principal series, i.e. we assume that
V = Vτ , V ′ = V̄ = V−τ for some τ ∈ iR; the case of representations of the complementary
series can be treated similarly.
Let {en }n∈2Z be a K-type orthonormal basis in V . We denote by {e′n = ē−n } the complex
conjugate basis in V̄ .
For a given T ≥ 1 we choose even n such that |T − n| ≤ 10 and set
wT = en ⊗ e′−n and w̃T = en ⊗ e′−n + en+2 ⊗ e′−n−2 . (1.10)
With such a choice of test vectors we have the following bounds.
Geometric bound:
H∆ (wT ) ≤ H∆ (w̃T ) . (1.11)
Spectral bounds:
(i) There exist constants b, c > 0 such that
Hλ (wT ) ≥ c|λ|−5/3 for |λ| ∈ IT (1.12)
where IT is the interval of length bT 1/3 centered at point 2T .
(ii) There exists a constant c′ such that
(
c′ T −1 (1 + |λ|)−1 for all |λ| ≤ 4T ,
Hλ (w̃T ) ≤ (1.13)
c′ |λ|−3 for all |λ| > 4T .
Using the bound (1.13) we can get the following sharp estimate of H∆ (w̃) (see Proposi-
tion 5.4):
H∆ (w̃T ) ≤ D (1.14)
with some explicit constant
P D > 0 (for the proof, see Section 6). This bound follows from
the identity H∆ (w̃) = di Hλi (w̃) (see (1.8)), the spectral bound (1.13) and the mean-
value bound (1.4) for the coefficients di . Here we use the fact that there are no contribution
to H∆ (w̃) coming from the discrete series since the vector w̃ is ∆K-invariant.
The bound (1.14) allows us to complete the proof of Theorem 1.3. Namely, the geometric
inequality (1.11) implies that H∆ (wT ) ≤ D. Using the spectral bound (1.12) we finally
obtain X X
di cT −5/3 ≤ di Hλi (wT ) ≤ H∆ (wT ) ≤ D .
|λi |∈IT i
10 JOSEPH BERNSTEIN AND ANDRE REZNIKOV
1.5.1. Proof of the geometric bound (1.11). The inequality (1.11) easily follows from the
pointwise bound on X. Namely, in the automorphic realization, the vector en ⊗ e′−n is
represented by a function which restriction un = r∆ (en ⊗ e′−n ) to the diagonal is non-
negative
un (x) = ν(en )(x) · ν̄(e′−n )(x) = |ν(en )(x)|2 ≥ 0.
From this we see that
Z Z
2
H∆ (wT ) = |un (x)| dµX ≤ |un (x) + un+2(x)|2 dµX = H∆ (w̃T ) .
X X
1.5.2. Sketch of proof of the spectral bounds (1.12) and (1.13). Proof of these bounds
carried out by the standard application of the stationary phase method and the Van der
Corput lemma. It constitutes the main technical bulk of the paper. We will use the explicit
form of the kernel defining Hermitian forms Hλ in the model realizations of representations
V , V ′ and Vλ . Namely, we use the standard realization of these representations in the space
∞
Ceven (S 1 ) of even functions on S 1 (see [BR2] and Section 1.4.1). Under this identification
the basis {en }n∈2Z becomes the standard basis of exponents {en = einθ }, where 0 ≤ θ < 2π
is the standard parameter on S 1 .
As was shown in [BR2], Section 5, in such realization the invariant functional lλmod on
the space V ⊗ V ′ ⊗ Vλ ≃ C ∞ ((S 1 )3 ) is given by the following kernel on (S 1 )3
−1+λ −1+2τ −λ −1−2τ −λ
Kλ (θ, θ′ , θ′′ ) = | sin(θ − θ′ )| 2 | sin(θ − θ′′ )| 2 | sin(θ′ − θ′′ )| 2 ,
′
where V = Vτ , V = V−τ with τ ∈ iR. From this it follows that the Hermitian forms Hλ
on E ≃ C ∞ (S 1 × S 1 ) are given by oscillatory integrals (over (S 1 )4 ) and the verification of
conditions (1.12) and (1.13) is reduced to the integration by parts and the stationary phase
method. Here the main difference between test vectors wT and w̃T appears. It manifests
itself in the form of oscillating integrals computing Hλ (wT ) and Hλ (w̃T ). Namely, both
of these integrals have the same phase which posses a degenerate critical point. The
main difference then is that for the vector wT the corresponding integral have a non-zero
amplitude at the critical point (this gives the crucial lower bound (1.12)) and for w̃ the
amplitude vanishes (resulting in bounds (1.13)).
In fact, we will use the values of Hλ (w) only for ∆K-invariant vectors w ∈ E. This
considerably simplifies our computations since we can reduce them to two repeated inte-
grations in one variable and use the stationary phase method in one variable.
Remark. The existence of vectors satisfying spectral conditions (1.12) and (1.13) allows
us to shorten the summation over the spectrum, comparatively to the range of the summa-
tion in the convexity bound (1.4). This is necessary if one wants to deduce a subconvexity
bound from the Bessel inequality of Hermitian forms (1.9) since the convexity bound (1.4)
is essentially sharp (see [Re1]). This approach to the subconvexity is reminiscent of the
classical amplification method introduced by Selberg (see [Mi], [MiV] for the review of
SUBCONVEXITY OF TRIPLE L-FUNCTIONS 11
the state of the art subconvexity results). Usually one uses a variant of a trace formula to
control the so-called off-diagonal terms arising after shortening the sum. In our approach
there is no use of the Selberg or the Kuznetsov trace formulas. Instead, we use the hidden
symmetries of the diagonal form H∆ .
1.6. L4 -norm. Our arguments also prove the following result on the L4 -norm of K-types
in irreducible automorphic representations of P GL2 (R). This result is of independent
interest.
Theorem. For a fixed class one automorphic representation ν : V → C ∞ (X), there
exists D > 0 such that ||ν(en )||L4 (X) ≤ D for all n.
One would expect that a similar fact holds for representations of the discrete series as
well. It is a very interesting question to study dependence of the constant D on the
parameter of the automorphic representation and on the subgroup Γ.
1.7. Gelfand pairs and subconvexity. In this section we briefly discuss another pos-
sible approach to bounds for triple periods (see [Re2] for a detailed discussion of this
approach). It is based on the notion of strong Gelfand pairs (see [Gr] and references
therein).
We call a pair (A, B), of a group A and a subgroup B, a strong Gelfand pair if for any
smooth irreducible representations V of A and W of B, the condition dim MorB (V, W ) ≤ 1
holds.
As we have shown in Section 1.4.3, the main object of our study is the diagonal Her-
mitian form H∆ . Our main observation is that the form H∆ has two different spectral
decompositions. These decompositions are associated with two different strong Gelfand
triples in the ambient group, and the corresponding arrangement of closed automorphic
orbits. Let us explain this construction.
Let G = G × G × G × G be the product of four copies of the group G. Consider the
following three subgroups of G:
H1 = {(g, g, h, h)| g, h ∈ G} ⊂ G ,
H2 = {(g, h, g, h)| g, h ∈ G} ⊂ G ,
F = ∆G = {(g, g, g, g)| g ∈ G} ⊂ Hi ⊂ G .
Note, that the pairs (G, Hi ) and (Hi , F) are strong Gelfand pairs as follows from the
uniqueness of trilinear invariant functionals.
12 JOSEPH BERNSTEIN AND ANDRE REZNIKOV
X ✛j2
j 1✲ ⊃
⊂
X ×X
✛ X ×X (1.15)
i1 ⊃ ✲
⊂
i2
X
Here j1 = ∆1,2 × ∆2,3 , j2 = ∆1,3 × ∆2,4 (∆i,j is the diagonal imbedding of X into the
i’s and j’s copies of X inside of X = X 4 ) and i1 = i2 = ∆ is the diagonal imbedding. We
endow each space in (1.15) with a measure invariant under the corresponding subgroup
(for simplicity, we assume that the space X is compact, and hence, these measures can
be normalized to have mass one).
Let V ⊂ C ∞ (X) be the space of smooth vectors in an irreducible automorphic repre-
sentation of G and let E = V ⊗ V̄ ⊗ V̄ ⊗ V = E ⊗ Ē be the space of smooth vectors in
the corresponding tensor product (see Section 1.4.3).
The integration over the diagonal copy ∆X ⊂ X defines on the vector space E a ∆G-
invariant form I∆ : E → C polarizing the Hermitian form H∆ . Namely, for any w ∈ E we
have
Z
H∆ (w) = I∆ (w ⊗ w̄) = |νE (w)|2 dµX .
∆X
We would like to describe the spectral decomposition of the functional I∆ and to deduce
from it the spectral decomposition of H∆ . The uniqueness principle is not applicable
directly here since the pair (G, F) is not a Gelfand pair. Instead, we can write two
different spectral expansions for I∆ using the intermediate groups H1 and H2 .
Namely, for any e ∈ E, we have two different ways to compute the value of the functional
I∆ :
Z Z
I∆ (e) = (i1 )∗ (j1 )∗ (e)dµX = (i2 )∗ (j2 )∗ (e)dµX .
X X
This gives rise to a non-trivial spectral identity. Namely, we use the uniqueness of trilinear
functionals (i.e., the fact that (G×G, ∆G) is a strong Gelfand pair). As a result, we obtain
the following spectral identity (it could be viewed as a Rankin-Selberg type spectral
identity; see [Re2])
X X
dω · Tωmod (e) = I∆ (e) = dω · Smod
ω (e) (1.16)
Vω ⊂L2 (X) Vω ⊂L2 (X)
for any e ∈ E. Here the summation is over all irreducible smooth automorphic represen-
tations (of class one and of the discrete series), Tωmod is the model ∆G-invariant functional
on E associated with a choice of the model invariant trilinear functional lωmod : E ⊗Vω∗ → C
SUBCONVEXITY OF TRIPLE L-FUNCTIONS 13
and the corresponding intertwining map Tωmod : V ⊗ V̄ → Vω (see Section 1.4.3) via the
relation
Tωmod (w1 ⊗ w2 ⊗ w3 ⊗ w4 ) = hTωmod (w1 ⊗ w2 ), T̄ωmod (w3 ⊗ w4 )iVω
and similarly
Smod mod
ω (w1 ⊗ w2 ⊗ w3 ⊗ w4 ) = hTω (w1 ⊗ w3 ), T̄ωmod (w2 ⊗ w4 )iVω
(i.e., these two functionals are obtained from each other by applying the ∆G-intertwining
map i2,3 : E → E, i2,3 (w1 ⊗ w2 ⊗ w3 ⊗ w4 ) = w1 ⊗ w3 ⊗ w2 ⊗ w4 ). In particular, for the
class one automorphic representations Vω ≃ Vλi we have dω = di (for the description of
the triple product coefficients di see Section 1.4.3).
Our main observation is that the resulting model functionals Tωmod and Smod ω look com-
pletely different on the space E. This is because the “flip” map i2,3 has nothing to do with
the action of G and constitutes a hidden symmetry of the functional I∆ . In particular,
for a vector w ∈ E and Vω ≃ Vλi , we have
Tωmod (w ⊗ w̄) = Hλi (w)
while the quantity Smod
ω (w ⊗ w̄) is impossible to express in terms of the forms Hλi . As
the result, the spectral density of a vector e ∈ E with respect to the functionals Tωmod and
Smod
ω can have significantly different properties. This leads to highly non-trivial identities
among coefficients dω .
One can use the spectral identity (1.16) in order to bound coefficients dω (see [Re2]
for another application of similar spectral identities to the subconvexity problem). For
this one have to exhibit a family of vectors which have “singular” behavior with respect
to functionals Tωmod (i.e., having the spectral measure Tωmod (e) concentrated at a short
interval of ω’s), but have regular behavior with respect to functionals Smod
ω (i.e., have
spectral measure such that the summation in the spectral identity (1.16) is possible to
estimate with a help of the mean-value bound for the coefficients dω ). Existence of such
a family of vectors would provide a proof of the subconvexity bound in Theorem 1.3.
There is one technical complication, though. We where not able to produce the desired
family of vectors in E which is also ∆K × ∆K-invariant. Without this property one has
to consider terms in the spectral decomposition (1.16) coming from the discrete series
representations. It is more cumbersome to study model trilinear functionals on discrete
series as these representations do not have nice geometric models. As a result we use in
the proof another property of the form H∆ , the extra positivity provided by the Cauchy-
Schwartz inequality (see Section 1.5.1), instead of the associated Gelfand pairs structure.
We hope to return to this subject elsewhere.
Research was partially supported by BSF grant, by Minerva Foundation and by the
Excellency Center “Group Theoretic Methods in the Study of Algebraic Varieties” of the
Israel Science Foundation, the Emmy Noether Institute for Mathematics (the Center of
Minerva Foundation of Germany), and by RTN “Liegrits”. The results of this paper
were obtained during our visits to Max-Planck Institute in Bonn and in Leipzig, and to
Courant Institute.
2.1. We recall the standard connection between Maass forms and representation theory
of P GL2 (R) (see [G6]). Most of the material in the next three sections is taken from
[BR2], where it is discussed in more detail.
2.1.1. Automorphic space. Let H be the upper half plane with the hyperbolic metric of
constant curvature −1. The group SL2 (R) acts on H by fractional linear transformations.
This action allows to identify the group P SL2 (R) with the group of all orientation pre-
serving motions of H. For reasons explained below (see Remark 3.2), we would like to
work with the group G of all motions of H; this group is isomorphic to P GL2 (R). Hence
throughout the paper we denote G = P GL2 (R) and K = P O(2) the standard maximal
compact subgroup. We have G/K = H.
We fix a discrete co-compact subgroup Γ ⊂ G and set X = Γ\G. The group G acts on X
(from the right) and hence on the space of functions on X. We fix the unique G-invariant
measure µX on X of total mass one. Let L2 (X) = L2 (X, dµX ) be the space of square
integrable functions and (ΠX , G, L2 (X)) the corresponding unitary representation. We
will denote by PX the Hermitian form on L2 (X) given by the inner product. We denote
by || · ||X or simply || · || the corresponding norm and by hf, giX the corresponding inner
product.
2.1.4. Decomposition of the representation (ΠX , G, L2 (X)). It is well known that for a
compact X, the representation (ΠX , G, L2 (X)) decomposes into a direct (infinite) sum of
irreducible representations of G with finite multiplicities (see [G6]). We will fix one such
decomposition and call it the automorphic spectrum of X. We can write
L2 (X) = (⊕i Li ) ⊕ (⊕κ Lκ ) ,
where Li irreducible representations corresponding to Maass forms (including the trivial
representation), and Lκ some irreducible representations of discrete series.
For us it will be convenient to write this decomposition as the following decomposition
of the Hermitian form PX on C ∞ (X)
X X
PX = Pi + Pκ , (2.1)
i κ
3. Triple products
3.1. Automorphic triple products. We introduce now our main object of study.
3.1.1. Automorphic triple products. Suppose we are given three automorphic representa-
tions (πj , Vj , νj ), j = 1, 2, 3 of G
νj : Vj → C ∞ (X) .
Z
lVaut
1 ,V2 ,V3
(v1 ⊗ v2 ⊗ v3 ) = φv1 (x)φv2 (x)φv3 (x)dµX ,
X
In particular, the triple periods ci in (1.2) can be expressed in terms of this form as
aut ′
ci = lV,V ′ ,V (e ⊗ e ⊗ eλi ) ,
λ
(3.1)
i
3.2. Uniqueness of triple products. The central fact about invariant trilinear func-
tionals is the following uniqueness result:
Theorem. Let (πj , Vj ), where j = 1, 2, 3, be three irreducible smooth admissible repre-
sentations of G. Then dim HomG (V1 ⊗ V2 ⊗ V3 , C) ≤ 1.
Remark. The uniqueness statement was proven by A. Oksak in [O] for the group
SL(2, C) and the proof could be adopted for P GL2 (R) as well (see also [Mo] and [Lo]).
For the p-adic GL(2) more refined results were obtained by D. Prasad (see [P]). He also
proved the uniqueness when at least one representation is a discrete series representation
of GL2 (R).
There is no uniqueness of trilinear functionals for representations of SL2 (R) (the space
is two-dimensional). This is the reason why we prefer to work with P GL2 (R).
We note however, that the absence of uniqueness does not pose any serious problem
for the method we present. All what is really needed for our method is the fact that the
space of invariant functionals is finite dimensional.
3.3. Model triple products. In Section 7.1, we use an explicit model for representations
(π, V ), (π ′ , V ′ ) and (πi , Vi ) to construct a model invariant trilinear functional. Model
functionals will be given by an explicit formula. We call it the model triple product and
mod
denote it by lV,V ′ ,V
λ
or simply lλmod
i
if π and π ′ are fixed.
i
3.3.1. Exponential decay. Relations (3.1) and (3.2) give rise to a formula for the triple
product coefficients ci
ci = lλaut
i
(e ⊗ e′ ⊗ eλi ) = ai · lλmod
i
(e ⊗ e′ ⊗ eλi ) .
Let us explain how one can deduce the exponential decay for the coefficients ci using this
identity.
The value of the model triple product functional lλmodi
(e ⊗ e′ ⊗ eλi ) constructed in Section
7.1 is given by an explicit integral. In [BR2], Appendix A, we evaluated this integral in
terms of the standard Euler Γ-function by a direct computation in the model and showed
that |lλmod (eτ ⊗ eτ ′ ⊗ eλ )|2 = 1/γ(λ), where γ(λ) is as in Section 1.3. After applying the
Stirling formula to that expression one sees that it has an exponential decay in |λ|. Hence,
in order to obtain bounds on the coefficients ci one needs to bound coefficients di = |ai |2 .
It turns out that the coefficients di are at most polynomial. This explains the exponential
decay of coefficients ci . Note that the coefficients di encode deep arithmetic information
– special values of L-functions.
4. Hermitian forms
4.1. Hermitian forms and trilinear coefficients di . We explain now how to obtain
bounds for the coefficients di
Our method is based on the fact that these coefficients appear in the spectral decompo-
sition of some geometrically defined Hermitian form on the space E which is essentially
the tensor product of spaces V and V ′ . This form plays a crucial role in what follows.
More precisely, denote by L and L′ the Hilbert completions of spaces V and V ′ , consider
the unitary representation (Π, G × G, L ⊗ L′ ) of the group G × G and denote by E its
smooth part; so E is a smooth completion of V ⊗ V ′ .
Denote by H(E) the (real) vector space of continuous Hermitian forms on E and by
H+ (E) the cone of nonnegative Hermitian forms.
We will describe several classes of Hermitian forms on E; some of them have spectral
description, others are described geometrically.
18 JOSEPH BERNSTEIN AND ANDRE REZNIKOV
Here the summation on the right is over all irreducible unitary automorphic representa-
tions appearing in the decomposition of L2 (X) (see (2.1)). Among these representations
there are representations of discrete series besides those corresponding to Maass forms we
described in Section 2.1.2.
SUBCONVEXITY OF TRIPLE L-FUNCTIONS 19
Remark. For the most of the proof we will need just the inequality (the Bessel inequality)
X
di Hλi ≤ H∆ . (4.1)
Vλi of class one
Here the summation is over all automorphic representations of class one (i.e., correspond-
ing to Maass forms), including the trivial representation.
This follows from the simple fact that for a ∆K-invariant vector w ∈ E, the restriction
onto the diagonal ∆X of the automorphic realization ν ⊗ ν̄(w) is a K-invariant function
on X, and hence orthogonal to discrete series representations appearing in L2 (X).
In this section we prove the subconvexity bound (1.5) in Theorem 1.3. We assume, first,
that V ′ ≃ V̄ . We also assume, for simplicity, that V ≃ Vτ , τ ∈ iR, is a representation of
the principal series. In Appendix D we show how easily dispose with these restrictions.
5.2. K-types. All the necessary computations will be done in the circle model Vτ ≃
∞
Ceven (S 1 ) (i.e., smooth functions such that f (t+π) = f (t)). For a principal series
R represen-
tation V , the invariant unitary Hermitian form on V is given by ||f || = 1/2π S 1 |f (t)|2 dt,
2
where 0 ≤ t < 2π is the standard angular parameter. Let en = exp(int), where n ∈ 2Z,
be an orthonormal basis of K-types in the space Vτ (all weights are even as we work with
P GL2 (R)).
Consider the space V̄τ . We have a natural identification V̄τ ≃ V−τ induced by the
realization of these spaces as spaces of functions on R2 \ 0.
20 JOSEPH BERNSTEIN AND ANDRE REZNIKOV
We denote by {e′n = ē−n }n∈2Z the corresponding complex conjugate basis for V̄τ ≃ V−τ .
∞
Under the natural identification V−τ ≃ Ceven (S 1 ), we have e′n = exp(int) as before.
5.3. Test vectors. In the Introduction (see formula (1.10)) we defined two families of
test vectors central for our proof of the subconvexity. Namely, for any real parameter
T ≥ 1, we consider two vectors in E = Eτ = Vτ ⊗ V−τ given by
wT = en ⊗ e′−n ,
and
w̃T = en ⊗ e′−n + en+2 ⊗ e′−n−2 ,
where |T − n| ≤ 10 (in fact, all we need is that |n − T | remains bounded as T → ∞). We
∞
note that in the model Vτ ⊗ V−τ ≃ Ceven,even (S 1 × S 1 ) these vectors are represented by
the functions wT = ein(x−y) and w̃T (x, y) = (1 + ei2(x−y) )ein(x−y) .
In Section 1.5.1 we have proven the basic geometric bound (1.11) for these vectors
H∆ (wT ) ≤ H∆ (w̃T ) . (⋆)
5.5. Proof of Theorem 1.3. For any T ≥ 1, we consider the pair of test vectors wT
and w̃T described above. We proved a variety of inequalities for this pair. In particular,
inequalities (⋆) and (♮) above imply the main inequality needed in the proof of Theorem
1.3:
H∆ (wT ) ≤ D .
On the other hand we claim that the test function Hλ (wT ) is not small for λ in some
range. Namely, we claim that there are constants b, c > 0 such that
Hλ (wT ) ≥ c|λ|−5/3 for |λ| ∈ IT ,
where IT is the interval of length bT 1/3 centered at point 2T (this is the bound (1.12)
from the Introduction; see also Lemma 6.1).
SUBCONVEXITY OF TRIPLE L-FUNCTIONS 21
These two inequalities together with the Bessel inequality (4.1) imply the bound in
Theorem 1.3:
X X
di cT −5/3 ≤ di Hλi (wT ) ≤ H∆ (wT ) ≤ D.
λi ∈IT i
Hence λi ∈IT di ≤ BT 5/3 for some B > 0. We established the bound in Theorem 1.3
P
6.1. Spectral Lemma. We start with the following spectral bounds needed in the proof
of Proposition 5.4 and in the proof of Theorem 1.3 (these are the bounds (1.12) and (1.13)
from the Introduction).
Lemma. There are positive constants b, c, C such that, for any T ≥ 1, the following
spectral bounds hold
(I): Hλ (wT ) ≥ c|λ|−5/3 for all λ satisfying ||λ| − 2T | ≤ bT 1/3 ,
(II1 ): Hλ (w̃T ) ≤ C · T −1 (1 + |λ|)−1 + C(1 + |λ|)−3 for all 0 < |λ| ≤ 4T , λ ∈ iR,
(II2 ): Hλ (w̃T ) ≤ C|λ|−3 for all |λ| ≥ 4T .
The model Hermitian forms Hλ on E are defined explicitly for every λ ∈ iR as in Section
7.1. The proof of the lemma amounts to a routine application of the stationary phase
method and the van der Corput lemma (see Appendices A-E).
Remark. While the formulation of bounds (II1,2 ) is quite technical it might be inter-
preted in following more conceptual terms related to Sobolev norms and the proof of
convexity bound in [BR2].
In [BR2] the convexity bound (1.4) was deduced from the Bessel bound (4.1) of non-
negative Hermitian forms with the help of the notion of non-negative functionals on the
space of Hermitian forms H(E) on E (see Section 4.1). Namely, in [BR2], Proposition 4.6
we introduced non-negative test functionals ρt : H+ (E) → R+ ∪∞ such that ρt (Hλ ) ≤ ct2
for |λ| ≤ 2t and ρt (H∆ ) ≤ 1. The existence of such functionals implied the convexity
bound.
RT
Consider the non-negative functional σT on H+ (E) defined by σT = T1 1 ρt dt. The
bound σT (H∆ ) ≤ 1 still holds.
On the other hand, any vector w ∈ E defines a non-negative functional ρw on H+ (E) by
the formula ρw (H) = H(w). The meaning of bounds in (II1,2 ) is that ρw̃T (Hλ ) ≤ c′′ σT (Hλ )
for all λ and from this follows that ρw̃T (H∆ ) ≤ c′′ (see Section 5.5 for a detailed argument).
We note, however, that while the construction of forms ρt is geometric the proof of the
bound ρw̃T (Hλ ) ≤ c′′ σT (Hλ ) is spectral, and we do not know of a geometric proof of the
bound ρw̃T (H∆ ) ≤ c′′ .
22 JOSEPH BERNSTEIN AND ANDRE REZNIKOV
6.2. Proof of Proposition 5.4. As could be seen from our construction in Section 5.3,
vectors w̃T are ∆K-invariant. It follows from the discussion in Remark 4.2 that for such
vectors, we have the following Parseval identity (4.2)
X X
H∆ (w̃T ) = Hiaut (w̃T ) = di Hλi (w̃T ) .
i i
We split the summation over the spectrum according to bounds (II1,2 ), Lemma 6.1
into three ranges: the low lying spectrum λ ∈ [0, 1], the intermediate spectrum λ ∈ iR
satisfying |λ| ≤ 4T , and the high lying spectrum |λ| ≥ 4T . We have accordingly
X
H∆ (w̃T ) = Hiaut (w̃T ) =
i
X X X
= Hiaut (w̃T ) + di Hλi (w̃T ) + di Hλi (w̃T )
λi ∈[0,1] 0<|λi |≤4T, λi ∈iR |λi |≥4T, λi ∈iR
which is bounded by
X X X
≤ Hiaut (w̃T ) + di CT −1 (1 + |λi |)−1 + di (C + C)(1 + |λi |)−3
λi ∈[0,1] |λi |≤4T, λi ∈iR λi ∈iR
6.2.1. Low spectrum. The vector w̃T is ∆K-invariant and hence the corresponding func-
tion b = νE (w̃T )|∆X = |φn |2 + |φn+2|2 is a K-invariant function, and hence we can view it
as a function on Y . Moreover, we can see that it has the L1 -norm on Y
Z
||b||L1 (Y ) = |φn |2 + |φn+2 |2 dv = 2
Y
6.2.2. Intermediate spectrum. The convexity bound (1.4) and the summation by parts
imply that for 4T = 2N ,
X X X X
di 2−N (1 + |λi |)−1 ≤ 2−N di 2−k ≤ 2−N A22k 2−k
0≤|λi |≤2N , λi ∈iR 1≤k≤N 2k ≤|λi |≤2k+1 1≤k<N
≤ A2−N 2N +1
is uniformly bounded.
is uniformly bounded.
The rest of the paper is devoted to the proof of spectral bounds (I) and (II1,2 ). This will
be done using explicit computations in the explicit model of irreducible representations.
As a preparation we start with an explicit construction of model Hermitian forms Hλ .
7.1. Model trilinear functionals. In this section we briefly recall our construction from
[BR2] of model trilinear invariant functionals.
For every λ ∈ C, we denote by (πλ , Vλ ) the smooth class one representation of the
generalized principle series of the group G = P GL2 (R) described in Section 2.1.3. There
24 JOSEPH BERNSTEIN AND ANDRE REZNIKOV
We describe the model invariant trilinear functional using this geometric model. Namely,
for three given complex numbers τ, τ ′ , λ, we explicitly construct nontrivial trilinear
functional lmod : Vτ ⊗ Vτ ′ ⊗ Vλ → C by means of its kernel. In the circle model, the
trilinear functional on the triple Vτ , Vτ ′ , Vλ is given by the following integral:
Z
mod −3
lπ,π′ ,πλ (f1 ⊗ f2 ⊗ f3 ) = (2π) f1 (x)f2 (y)f3(z)Kτ,τ ′ ,λ (x, y, z)dxdydz ,
(S 1 )3
Here x, y, z are the standard angular parameters on the circle. As we verified in [BR2]
this defines a non-zero G-invariant functional.
Remark. 1. The integral defining the trilinear functional is often divergent and the
functional should be defined using regularization of this integral. There are standard
procedures how to make such a regularization (see [G1]).
Fortunately, in the case of class one unitary representations, all integrals converge ab-
solutely, so we will not discuss the regularization procedure.
2. We do not have a similar simple formula for the trilinear invariant functional when
at least one representation is a representation of discrete series. This is because we do not
know of a simple “geometric” model for representations of discrete series. As a result it is
more cumbersome to carry out explicit computations in that case. Another problem we
have to face is that the results of [BR2] have not been extended yet to cover the discrete
series.
Nevertheless, we expect our methods to carry out for discrete series as well and to
produce corresponding subconvexity bounds.
8.1. Reduction to integrals of one variable. In what follows, we only need to deal
with ∆K-invariant vectors in E. For such vectors, we can simplify the integral repre-
senting the Hermitian form Hλ . Namely, let lλmod : E ⊗ Vλ → C be the model trilinear
functional introduced in Section 7.1, Tλ = Tλmod : E ∗ → V−λ be the corresponding map,
and Hλ the model Hermitian form on E obtained from the composition of Tλ with the
SUBCONVEXITY OF TRIPLE L-FUNCTIONS 25
where e∗λ = Tλ∗ (eλ ) ∈ E ∗ is the distribution with the kernel given by
Z
∗ − 21
eλ (x, y) = (2π) K−τ,τ,λ (x, y, z)dz . (8.1)
S1
Here the kernel K−τ,τ,λ is given by the kernel of the model trilinear functional as in (7.1),
Section 7.1.
The kernel e∗λ (x, y) is ∆K-invariant, and hence it is a function of x − y. It is not given
by an elementary function, but we claim that it has a simple approximate representation.
Claim. For all λ ∈ iR, |λ| ≥ 1 and any w ∈ E, we have the following decomposition
Z 2
−1
Hλ (w) = |λ| w(c)kλ(c)dc + rλ,τ (w) , (8.2)
S1
where the kernel in the integral is given by kλ (c) = A(c)mλ (c) with
π λ 1
A(c) = (π)−1 ei 4 2 2 · | sin(c)|− 2 ,
λ λ
mλ (c) = | sin(c/2)|− 2 | cos(c/2)| 2 ,
and the remainder satisfies the estimate rλ,τ (w) = O(| sup(w)|2 (1 + |λ|)−3 ).
Here sup(w) is the supremum of the function w(x, y) ∈ C ∞ (S 1 × S 1 ), and the implied
constant in the O-terms depends on τ .
We prove the claim in Appendix C.1 by a straightforward application of the stationary
phase method, and estimate the reminder via van der Corput lemma.
From the approximate representation (8.2) of the Hermitian form Hλ we deduce spectral
bounds (I) and (II1,2 ), Lemma 6.1.
8.2. Proof of (I), Lemma 6.1. Let wT (x, y) = ein(x−y) , n ≍ T . We need to prove a
lower bound for the value of Hλ (wT ). According to the asymptotic formula (8.2), we need
to estimate the following integral
26 JOSEPH BERNSTEIN AND ANDRE REZNIKOV
Z
1 λ λ
I(n, λ) = | sin(c)|− 2 | sin(c/2)|− 2 | cos(c/2)| 2 einc dc (8.3)
S1
8.3. Proof of (II1,2), Lemma 6.1. As in the proof of the bound (I), an estimate of
Hλ (w̃T ) is reduced to an estimate of the following integral
Z
1 λ λ
J(n, λ) = (1 + ei2c )| sin(c)|− 2 | sin(c/2)|− 2 | cos(c/2)| 2 einc dc . (8.5)
S1
This time we are looking for uniform in n upper bounds valid for all λ ∈ iR.
Claim B.
1
(a): There exists a constant C ′′ > 0 such that |J(n, λ)| ≤ C ′′ |n|− 2 for all λ ∈ iR,
|λ| ≥ 1,
(b): |J(n, λ)| ≪ (1 + |λ|)−N for all |λ| > 4|n| and any N > 0.
From the approximate formula (8.2) it follows that
Hλ (w̃T ) ≤ (1 + |λ|)−1 |J(n, λ)|2 + O((1 + |λ|)−3) ,
for all λ ∈ iR. Hence, Claim B implies bounds (II1,2 ), Lemma 6.1.
SUBCONVEXITY OF TRIPLE L-FUNCTIONS 27
Claim B is verified in Appendix B. The reason for the bound in (b) to hold is that
under the condition |λ| > a|n|, for any a > 2, the phase function in the corresponding
integral (8.5) has no critical points. In the complementary case, |λ| ≤ b|n|, for any b < 2,
one discovers that the phase has only Morse type critical points. In the resonance case,
n ≍ |λ|/2, we have two colliding Morse type critical points of the phase. However, these
points collide at the zero set of the amplitude. This is a classical situation in the theory of
the stationary phase method for which the uniform bound is well-known (e.g., bounds on
the Airy function and its derivative). The vanishing of the amplitude is achieved by the
inclusion of the factor 1 + ei2c in the amplitude. This makes the amplitude to vanish at
the degenerate critical point c = π/2. Such a vanishing constitutes the crucial difference
between the integral J(n, λ) and the integral I(n, λ) considered in Section 8.2. We also
have to treat the case |λ| ≪ |n|. In this case, we show that the contribution from the
1
singularities of the amplitude in the integral J(n, λ) is uniformly bounded by |n|− 2 , as
directly follows from the van der Corput lemma (see Appendix E.3).
Appendix A.
In this Appendix we prove Claim A from Section 8.4. We will use results from Appendix
E on the Airy integral and the stationary phase method.
λ λ
A.1. Proof of Claim A. We denote by Sn,λ (c) = ln(| sin(c/2)|− 2 | cos(c/2)| 2 ) + inc .
The integral (8.3) then takes the form
Z
1
I(n, λ) = | sin(c)|− 2 eSn,λ (c) dc . (A.1)
S1
A.1.1. Critical points. A direct computation shows that the critical points of the phase
Sn,λ are the solutions of the equation
− λ2 sin−1 (c) + in = 0 . (A.2)
We are interested in n and λ satisfying |in − λ2 | ≤ b|λ|1/3 for some fixed b > 0 to be
specified later. In that case, we see that there is a pair, c+ and c− , of critical points of
Sn,λ which are close to π/2 and collide for in = λ2 . We apply the method of Appendix
E.1.
First we split the integral into the sum of two integrals corresponding to the contribution
from a fixed neighborhood of critical points and the rest.
Let U = [π/2 − 0.1, π/2 + 0.1] and let χ be a smooth function such that supp(χ) ⊂ U
and χ|U ′ ≡ 1 for U ′ = [π/2 − 0.01, π/2 + 0.01]. We have then I(n, λ) = I1 (n, λ) + I2 (n, λ)
where
Z Z
− 21 Sn,λ (c) 1
I1 (n, λ) = | sin(c)| e χ(c)dc and I2 (n, λ) = | sin(c)|− 2 eSn,λ (c) (1 − χ(c))dc.
S1 S1
28 JOSEPH BERNSTEIN AND ANDRE REZNIKOV
We claim that |I1 (n, λ)| ≥ c|λ|−1/3 for |in − λ2 | ≤ b|λ|1/3 with some explicit b, c > 0, and
that |I2 (n, λ)| ≤ c|λ|−1/2 . This clearly implies the bound in Claim A: |I(n, λ)| ≥ 21 c|λ|−1/3 .
The upper bound |I2 (n, λ)| ≤ c|λ|−1/2 immediately follows from the van der Corput
lemma as explained in Appendix E.3. We show now that the lower bound for the integral
I1 follows from well-know properties of the Airy integral.
A.1.2. Estimate from below for I1 . To reduce the integral I1 to the Airy integral from
Appendix E, we compute the 3-jet of Sn,λ near the point c0 = π/2 :
Sn,λ(c0 + s) = − λ2 (s + s3 + O(s5)) + in(s0 + s) = Sn,λ (s0 ) − λ2 (βs + s3 + O(s5)) , (A.3)
where we denote by β = (in − λ2 )/ λ2 and view it as a real parameter. We consider here
only |β| ≤ 1 (in fact, in this Appendix we only need |β| ≤ b|λ|−2/3 for some b > 0
to be specified later). Let Sβ be a family of functions on S 1 defined by the formula
λ
S (s) = Sn,λ (s0 ) − Sn,λ(s0 + s). We have then
2 β
Z
1 λ
|I1 (n, λ)| = | sin(c)|− 2 e 2 Sβ (c) χ(c)dc .
S1
We see that the 3-jet of Sβ near s = 0 is given by Sβ (s) = βs + s3 + O(s5 ), and higher
order terms are independent of β. Hence, the phase function Sβ has the unique critical
point of the cubic type for β = 0 (i.e., λ2 = in ), two Morse type critical points for β < 0
and no critical points for β > 0. Moreover, it is easy to see that for any |β| ≤ 1 the
family of smooth functions Sβ is a non-degenerate family of the type A1 with respect to
the parameter β according to the classification in [Ar] (i.e., Sβ is a generic family with the
unique non-degenerate cubic point for β = 0). Hence, from the general theory of normal
forms, it follows that there exists a smooth change of variables s = s(x, α), β = β(x, α)
such that
Sα (x) = αx + x3 /3 .
The change of variables does not depend on λ, and it is smooth in x and α.
Clearly, we can choose a change of variable so that s(0, 0) = 0 and β(0, 0) = 0. Moreover,
it is easy to see that, that the image of the set |β| ≤ t for any 0 ≤ t ≤ 1 is contained
in the set |α| ≤ at for some fixed a > 0. In this appendix we only deal with the case
|β| ≤ b|λ|−2/3 , |λ| ≥ 1, for some b > 0 to be specified later. For such β, we have
|(in − λ2 )/ λ2 | = |β| ≤ |aα| + O(β 2), and hence |in − λ2 | ≤ |aα λ2 | + O((1 + |λ|)1/3 ).
1
Let φ(x, α) = sin(s(x, α))− 2 χ(s(x, α)) · |D(s(x, α))|, where D(s(x, α)) is the Jacobian of
the change of variables s = s(x, α). Note that the amplitude function sin(c) have no zeroes
in U and hence the amplitude in the integral I1 is a smooth function. Also φ(0, 0) 6= 0 as
the amplitude in the integral I1 have no zeroes. Hence we see that
Z
λ 3
|I1 (β, λ)| = φ(x, α)e− 2 (αx+x /3) dx , (A.4)
S1
where φ(0, 0) 6= 0.
SUBCONVEXITY OF TRIPLE L-FUNCTIONS 29
Applying the bound (Ai1) from Appendix E we arrive at the lower bound |I1 (n, λ)| =
|A λ (φ(x, α))| ≥ c′ |φ1 (α)||λ|−1/3 ≥ c′′ |λ|−1/3 for all |α| ≤ b′ |λ|−2/3 and some constants
2
b′ , c′ , c′′ > 0. Translating this back into the parameters n and λ we see that there are
constants b, c > 0 such that |I1 (n, λ)| ≥ c|λ|−1/3 for all n satisfying |in − λ2 | ≤ b|λ|1/3 .
Appendix B.
B.1. Proof of Claim B. In this appendix we prove Claim B from Section 8.3. The
integral we are going to consider have the same phase function as the integral which we
considered in Appendix A, and hence, we will borrow necessary computations from there.
The main difference with the Appendix A is that now we are interested in the uniform in
n and λ bound for the integral. The crucial fact we are going to exploit in this appendix
is the vanishing of the amplitude at the degenerate critical point of the relevant phase
function.
We need to estimate the integral (8.5)
Z
1 λ λ
J(n, λ) = (1 + ei2c )| sin(c)|− 2 einc | sin(c/2)|− 2 | cos(c/2)| 2 dc (B.1)
S1
B.1.1. Large λ: no critical points. Let |λ| ≥ 4|n|. Under this condition an easy compu-
tation shows that the derivative of the phase in the integral (B.1) have no zeroes and
the bound |J(n, λ)| ≪ |λ|−N immediately follows from the van der Corput lemma as
explained in Appendix E.3.
B.1.2. Intermediate λ: degenerate critical points. Let λ satisfy the condition 0.01|n| ≤
|in− λ2 | ≤ 4|n|. As we have seen in Appendix A.1.1, for the resonance value of parameters,
(i.e., in = λ2 ), there is a degenerate critical point of the phase at c0 = π/2. Let β =
(in − λ2 )/ λ2 . We can assume that |β| ≤ 1. For this range of the parameter β, we have
shown in Appendix A.1.1 that all critical points of the phase belong to some interval U ′′
such that the points of singularity of the amplitude (i.e., zeroes of sin(c)) are outside of
U ′′ . We again split the integral J(n, λ) into the sum of two integrals J1 (n, λ) (over U ′′ )
and J2 (n, λ) (over the complement of U ′′ ). We note that the integral J2 is negligible (i.e.,
is of order of |λ|−N for any N > 0) for the reasons explained in Appendix E.3. Hence we
1
need to bound the integral J1 . We claim that for |β| ≤ 1, we have |J2 (n, λ)| = O(|λ|− 2 ).
Let s be local coordinate centered at π/2 and |β| ≤ 1. As was shown in Appendix A.1.2,
in a fixed neighborhood of π/2 there are new coordinates s = s(x, α) and β = β(x, α)
30 JOSEPH BERNSTEIN AND ANDRE REZNIKOV
such that
Z
λ 3 /3)
|J1 (n, λ)| = |A λ (φ̂(x, α)| = φ̂(x, α)e− 2 (αx+x dx . (B.2)
2
S1
Moreover, this time we have φ̂(0, 0) = 0 since the amplitude in the integral (B.1) had zero
at π/2. Applying the bound (Ai3) from Appendix E.1.1 we arrive at the bound desired:
1
|J(n, λ)| ≤ c|λ|− 2 for 0.01|n| ≤ |λ| ≤ 4|n| and some c.
B.1.3. Small λ. Let 1 ≤ |λ| ≤ 0.01|n|. As we have seen, in that range of λ there are
no degenerate critical points of the phase in the integral J(n, λ). The integral is of the
R1 1
form 0 a(x)|x|− 2 +s esφ(x) dx, a(x) smooth, 1 ≤ |φ′(x)| ≤ 2 and s ∈ iR. The bound claimed
follows immediately from the van der Corput lemma as explained in the Appendix E.3.
Appendix C.
C.1. The kernel reduction. We consider the integral (8.1), Section 8.1:
Z
− 21
e∗λ (x, y) = (2π) K−τ,τ,λ(x, y, z)dz =
S1
Z
− 21 − 21 + λ 1 λ 1 λ
(2π) | sin(x − y)| 2 · | sin(x − z)|− 2 +τ − 2 | sin(y − z)|− 2 −τ − 2 dz
S1
1 λ
= | sin(x − y)|− 2 + 2 Kλ,τ (x − y) ,
where the kernel K−τ,τ,λ is as in (7.1), and we denoted by
Z
λ λ
− 21 1 1
Kλ,τ (c) = (2π) | sin(t)|− 2 +τ − 2 | sin(c − t)|− 2 −τ − 2 dt . (C.1)
S1
Such an asymptotic expression follows from the stationary phase method (see Appendix
E.2). The phase of the oscillating kernel in the integral (C.1) has two non-degenerate
critical points. Hence, the leading term in the asymptotic is given as a sum of two terms.
Singularities of the amplitude at c = 0, π are responsible for the logarithmic term in the
remainder. For |λ| → ∞, the contribution from the singularities of the amplitude is of
order of O((1 + |λ|)−N ) for any N > 0 due to the logarithmic oscillation of the phase at
the same points (see Appendix E.3).
After elementary manipulations with (C.3) we arrive at
1 λ π λ 1
e∗λ (c) = | sin(c)|− 2 + 2 Kλ (c) = ei 4 2 2 |λ|− 2 ×
1
λ λ λ λ
× | sin(c)|− 2 | sin(c/2)|− 2 | cos(c/2)| 2 + | sin(c/2)| 2 | cos(c/2)|− 2 + rλ,τ
′
(c)
with c = x − y.
λ λ
Let mλ (c) = | sin(c/2)|− 2 | cos(c/2)| 2 . We note that mλ (c + π) = m−λ (c). As we
integrate in (8.2) against even function w and the amplitude A(c) is also even, we obtain
the asymptotic formula in Claim 8.1.
Appendix D.
D.1. Proof of Theorem 1.6. As a by-product of the bound (♮), Proposition 5.4 an the
geometric bound (1.11), we established the bound in Theorem 1.6 on L4 -norm of K-types.
Namely, we have
||ν(en )||4L4 (X) = H∆ (en ⊗ e−n ) ≤ H∆ (en ⊗ e−n + en+2 ⊗ e−n−2 ) = H∆ (w̃T ) ≤ D , (D.1)
for some D independent of n.
D.2. Proof of Theorem 1.3 for V ′ 6≃ V̄ . After examining the proof of Claim A in
Appendix A.1, we see that the condition V ≃ V ′ is not essential in the proof of the lower
bound (I) in Lemma 6.1. Hence in fact we have proved there the following more general
bound.
Let V ≃ Vτ and V ′ ≃ Vτ ′ be automorphic representations of the principal series and set
E ′ = V ⊗ V ′ . Let Hλ′ be the corresponding model Hermitian forms on E constructed in
Section 7.1. Let {en } ∈ V and {e′n } ∈ V ′ , n ∈ 2Z be corresponding orthonormal basses
of K-types. Denote by wT′ = en ⊗ e′−n , |T − n| ≤ 10. There are constants b, c > 0,
depending on τ, τ ′ only, such that
For the sake of completeness only, in this appendix we collect classical bounds for stan-
dard integrals used in the paper.
E.1. The Airy integral. Let fα (x) = αx + x3 /3 and hα (x) = fα′ (x) = α + x2 with α a
real parameter. For any λ ∈ iR and −1 ≤ α ≤ 1, consider the functional given by the
integral
Z
Aλ (φ; α) = φ(x)eλfα (x) dx . (E.1)
R
Corollary. For φ(x, α) as above and |λ| ≥ 1, we have the following identity
−1 d
Aλ (φ(x, α); α) = φ1 (α)A(λ, α) + φ2 (α)B(λ, α) − λ Aλ φ3 (x, α); α ,
dx
where A(λ, α) = Aλ (1; α) and B(λ, α) = Aλ (x; α).
d
Since the L1 -norm (in x) of φ (x, α)
dx 3
is bounded for all α, we have
Aλ (φ(x, α); α) = φ1 (α)A(λ, α) + φ2 (α)B(λ, α) + O(|λ|−1) . (E.3)
The integrals A(λ, α) and B(λ, α) are classical integrals representing the Airy function
and its derivative. Namely, we have the following elementary lemma which is proved by
a simple change of variables argument.
Lemma. Let real ν and µ be such that λ = iν 3 and µ = ν 2 α. The following relations
are satisfied
A(λ, α) = ν −1 Ai(µ) and B(λ, α) = −iν −2 Ai′ (µ) ,
3 /3)
ei(µx+x
R
where Ai(µ) = R
dx is the Airy function.
Various properties of the Airy function and of its derivative are well-known (see [Mag]).
In particular, the Airy function Ai is smooth and non-zero at the origin. We denote by b0
the first zero of Ai (it is known that |b0 | > 2). For µ ≤ −1, Ai(µ) is rapidly decreasing.
Namely, for any N > 0, we have |Ai(µ)| ≪ |µ|−N and |Ai′ (µ)| ≪ |µ|−N . For positive µ,
we have uniform bounds
1 1
|Ai(µ)| ≤ c1 (1 + |µ|)− 4 and |Ai′ (µ)| ≤ c2 (1 + |µ|)− 4 .
Combining these asymptotic with the expansion (E.3) we obtain the following estimates.
There exists a constant c > 0 depending on φ(x, α) such that in the notations of the
proposition above
(Ai1): |Aλ (φ(x, α); α)| ≥ c · |φ1 (α)| · |λ|−1/3 for all |α| ≤ 2|λ|−2/3 ,
1
(Ai2): |Aλ (φ(x, α); α)| ≤ c · |φ1 (α)| · |λ|−1/3 (1 + |λ2/3 α|)− 4 for all α and |λ| ≥ 1,
1
(Ai3): if φ1 (0) = 0 then |Aλ (φ(x, α); α)| ≤ c| · |λ|− 2 for all α and |λ| ≥ 1.
E.1.2. Proof of Proposition E.1.1. Let φ(x, α) ∈ C 5 (R × [−1, 1]). We decompose φ into
the even part φ+ = 21 (φ(x, α) + φ(−x, α)) and the odd part φ− = 21 (φ(x, α) − φ(−x, α)).
As φ+ and φ− are in C 5 (R × [−1, 1]), we can write φ+ (x, α) = ψ1 (x2 , α) and φ− (x, α) =
x · ψ2 (x2 , α) where ψ1 (z, α), ψ2 (z, α) are in C 4 (R+ × [−1, 1]). Obviously we can extend
ψ1 (z, α), ψ2 (z, α) to R × [−1, 1] and obtain new functions ψ̂1 (z, α), ψ̂2 (z, α) ∈ C 4 (R ×
[−1, 1]) of compact support. Let u1 (α) = ψ̂1 (α, α) and u2 (α) = ψ̂2 (α, α). There are
functions v1 , v2 ∈ C 3 (R × [−1, 1]) such that ψ̂1 (z, α) = u1 (α) + (α + z)v1 (, z, α) and
ψ̂2 (z, α) = u2 (α) + (α + z)v2 (, z, α). Setting φ1 (α) = u1 (α) and φ2 (α) = u2 (α) and
φ3 (x, α) = v1 (x, α) + xv2 (x, α) we obtain the required representation φ(x, α) = φ1 (α) +
xφ2 (α) + (α + x2 )φ3 (x, α).
34 JOSEPH BERNSTEIN AND ANDRE REZNIKOV
(b): |Hλ,c (ψ)| ≪ (1 + |λ|)−N for any eventually polynomial ψ vanishing in the neigh-
borhood of the interval [−1, 1] and any N > 0.
(c): The distribution Hλ,c is well-defined on smooth eventually polynomial functions.
Proof. The claim (a) is obvious. To prove the claim in (b), we note that for any eventually
polynomial function there exists a polynomial P ∈ C[t] such that |P (ξ0 )ψ| ≪ |x|−1 for
|x| ≥ 1. On the other hand Q(1 + 2λ) < F, ψ >=< P ∗ (ξ0 )F, ψ >=< F, P (ξ0)ψ > with
Q a polynomial of the same degree as P . Hence we have the inequality | < F, ψ > | =
|Q(1 + 2λ)|−1 · | < F, P (ξ0)ψ > | ≤ Cψ |λ|− deg P for |λ| sufficiently large. Iterating this
we obtain the claim. Note that in fact all what is needed for this argument to work is
the fact that ψ has zeroes of high order at the singularities of ξ0 . Namely, that ξ0 ψ is
smooth and bounded at infinity by |x|−1 . We will use this extension below. The claim in
(c) immediately follows from (a) and (b).
E.2.3. Reduction of integrals. We explain now how to choose the necessary change of
′
variables in the integral R |h(t − α)|σ+λ |h(t + α)|σ +λ φ(t)dt (formula (E.4)) in order to
R
36 JOSEPH BERNSTEIN AND ANDRE REZNIKOV
′
|x − c|σ+λ |x + c|σ +λ ψc (x) (formula
R
transform it to the standard integral Hλ,c (ψc ) = R
(E.5)) we dealt with above.
Let g ∈ Cc∞ (R) be a function such that h(t) = tg(t) and g(0) > 0. We denote by
f (t, c) = h(t − c)h(t + c). The necessary change of variables is given by the following
Lemma. There exists a change of variables (y, b) = (y(t, c), b(t, c)) in a neighborhood
of the point (0, 0) such that
(1) The variable b is a function of c only,
(2) f (t, c) = (y + b)(y − b) in new coordinates, and
(3) h(t − c) = (y − b)g1 (y, b) and h(t + c) = (y + b)g2 (y, b), where g1 and g2 are smooth
functions not vanishing near the point (0, 0).
(E.5) we dealt with above.
Proof. The proof is based on repeated application of the Hadamard’s lemma. For every
c denote by m(c) be the value of t where f (t, c) has its minimum. By implicit function
theorem the function m(c) is well-defined and is smooth near c = 0 since it is given by
the equation ft (0) = 0 and we assumed that ftt (x) > 0 in the neighborhood of 0.
This also shows that the minimal value M(c) = f (m(c), c) of function f is a smooth
function. By Hadamard’s lemmap M(c) = c2 N(c), where the function N(c) is strictly
negative. Let us set b(c) = c −N(c). The function b(c) is smooth and satisfies b(c)2 =
−M(c).
Similarly, consider the function F (t, c) = f (t, c) − M(c). By Hadamard’s lemma we can
write F (t, c) = (t − m(c))2 H(t, c). The function H(t, c) is smooth and strictly positive,
so its square root h(t, c) is a smooth function.
Let us set y(t, c) = (t − m(c))h(t, c). Then f (y, b) = y 2 − b2 . It is easy to see that this
is a change of variables near the point 0.
Now we can end the proof of the proposition. By Hadamard’s lemma h(t − c) is divisible
by (y − b) since these functions have the same zeroes (one of the brunches of zero set for
the function f (y, b) = y 2 − b2 ). Hence we can write h(t − c) = (y − b)g1 (y, b). It is clear
that g1 is invertible near 0. Similarly for the function h(t + c).
E.3. Van der Corput lemma. In this section we explain how to obtain upper bounds
for integrals of the form u(x)|x|−s eiα(x) dx. These bounds are used in various places
R
throughout the paper. We claim that the necessary type of bound follows directly from
the van der Corput lemma.
where φ and f are real valued and smooth in the interval [a, b] functions. We will bound
this integral in terms of the variation of the amplitude and the minimum value of
R bthe′ phase.
(k)
For an integer k ≥ 1 denote by mk (φ) = min |φ (x)| and let M(f ) = |f (b)|+ a |f (x)|dx.
x∈[a,b]
We have the following general estimate essentially due to van der Corput (see [St], p. 332).
Lemma. Let k ≥ 1 be such that mk (φ) > 0. There exists a constant ck such that the
following bound
1
|I(φ, f )| ≤ ck · mk (φ)− k · M(f )
holds when
(1) k ≥ 2, or
(2) k = 1 and φ′ is monotonic.
The bound for ck is independent of φ, f and the inteval [a, b].
We use this lemma with k = 1 or 2, so we can assume that ck is a universal constant.
E.3.2. We apply the van der Corput lemma to obtain bounds for the following family of
integrals appearing throughout the paper:
Z 1
1
I(µ, λ) = u(x)|x|− 2 −iλ eiµα(x) dx ,
−1
where we assume that 1 ≤ λ ≤ µ, α is smooth and monotonic, 0.99 < α′ (x) < 1.01 (i.e.,
′′
bounded away from 0 and ∞) and |α (x)| ≤ 21 for all x (this insures that there is no
degenerate critical points of the phase), and u is smooth of compact support in (−1, 1).
Lemma. Under the above assumptions the following uniform bound holds
1
|I(µ, λ)| ≤ bµ− 2 ,
where the constant b is independent of µ and λ.
Proof. We denote by a the ratio a = λ/µ and consider the integral over the interval (0, 1)
(and the similar integral over (−1, 0))
Z 1
1
I(µ, a) = u(x)|x|− 2 eiµ(α(x)−a ln |x|)dx .
0
We are interested in the uniform (in µ) bound for this integral for the values of the
parameter a satisfying the bound µ−1 ≤ a ≤ 1.
In order to apply the van der Corput lemma we break the interval (0, 1) into 4 intervals
J1 = (2a, 1), J2 = ( 21 a, 2a), J3 = ( 12 µ−1 , 21 a) and J4 = (0, 12 µ−1 ) (for a ≥ 21 the first interval
is missing).R Denote by φa (x) = α(x) − a ln |x| and consider the corresponding integrals
1
Ij (µ, a) = Jj u(x)|x|− 2 eiµφa (x) dx.
On the interval J1 we have |φ′a (x)| ≥ 1. Hence from the van der Corput lemma (with
k = 1) we have |I1 (µ, a)| ≤ b1 µ−1 .
38 JOSEPH BERNSTEIN AND ANDRE REZNIKOV
On the interval J2 the phase φa has zero of the first derivative, but satisfies the bound
|φa (x)| > 12 a−1 . Hence on the interval J2 the van der Corput lemma with k = 2 implies
′′
1
|I2 (µ, a)| ≤ b2 µ− 2 .
To bound the integral I3 (µ, a) with note that |φ′a (x)| ≥ 12 and that the variation of the
1 Rb 1
amplitude satisfies | 21 a|− 2 + a |x|−3/2 dx ≤ cµ 2 on J3 . The van der Corput lemma implies
1
that |I3 (µ, a)| ≤ b3 µ− 2 .
Bounding the integral over J4 by the integral of the absolute value, we see that trivially
1
|I4 (µ, a)| ≤ b4 µ− 2 .
References
[Ar] V. I. Arnol′ d, S. M. Guseı̆n-Zade, A. N. Varchenko, Singularities of differentiable maps. Vol. I.
Monographs in Mathematics, 82. Birkhuser, 1985.
[BR1] J. Bernstein, A. Reznikov, Analytic continuation of representations, Ann. Math., 150 (1999),
329–352.
[BR2] J. Bernstein, A. Reznikov, Estimates of automorphic functions, Moscow Math. J. 4 (2004), no.
1, 19–37, arXiv: math.RT/0305351.
[Ca] W. Casselman, Canonical extensions of Harish-Chandra modules to representations of G, Canad.
J. Math. 41 (1989), no. 3, 385–438.
[G1] I. Gelfand, G. Shilov , Generalized Functions. vol. 1, Academic Press, 1964.
[G5] I. Gelfand, M. Graev, N. Vilenkin, Generalized Functions. vol. 5, Academic Press, 1966.
[G6] I. Gelfand, M. Graev, I. Piatetski-Shapiro, Representation Theory and Automorphic Forms.
Saunders, 1969.
[Go] A. Good, Cusp forms and eigenfunctions of Laplacian, Math. Ann., 255 (1981), 523–548.
[Gr] B. Gross, Some applications of Gel’fand pairs to number theory. Bull. Amer. Math. Soc. (N.S.)
24 (1991), no. 2, 277–301.
[I] H. Iwaniec, Small eigenvalues of Laplacian on Γ(N ), Acta Arith. 16 (1990), 65–82.
[IS] H. Iwaniec, P. Sarnak, Perspectives on the analytic theory of L-functions. GAFA 2000. Geom.
Funct. Anal. 2000, Special Volume, Part II, 705–741.
[HK] M. Harris, S. Kudla, The central critical value. Ann. of Math. 133, (1991), 605–672.
[KS] B. Krötz, R. Stanton, Holomorphic extensions of representations. I, Ann. of Math. (2) 159 (2004),
no. 2, 641–724.
[Lo] H. Loke, Trilinear forms of GL(2), Pacific J. Math. 197 (2001), no. 1, 119–144.
[M] H. Maass, Über eine neue Art von nichtanalytischen automorphen Funktionen und die Bestim-
mung Dirichletscher Reihen durch Funktionalgleichungen, Math. Ann. 121, (1949). 141–183.
[Mag] W. Magnus et al., Formulas and Theorems for the Special Functions, Springer, 1966.
[Ma] B. Malgrange, Ideals of differentiable functions, Tata Institute of Fundamental Research, Bom-
bay; Oxford University Press, London 1967.
[Mi] P. Michel, Familles de fonctions L de formes automorphes et applications. J. Théor. Nombres
Bordeaux 15 (2003), no. 1, 275–307.
[MiV] P. Michel, A. Venkatesh, Equidistribution, L-functions and ergodic theory: on some problems of
Yu. V. Linnik, preprint, 2006.
[Mo] V. Molchanov, Tensor products of unitary representations of the three-dimensional Lorentz
group, Izv. Akad. Nauk SSSR Ser. Mat. 43 (1979), no. 4, 860–891.
[O] A. Oksak, Trilinear Lorenz invariant forms. Comm. Math. Phys. 29 (1973), 189–217.
[P] D. Prasad, Trilinear forms for representations of GL(2), Composito Math., 75 (1990), 1–46.
SUBCONVEXITY OF TRIPLE L-FUNCTIONS 39
[PS] I. Piatetsky-Shapiro, Euler subgroups, in Lie groups and their representations, Halsted, (1975),
597–620.
[Ra] R. Rankin, Contributions to the theory of Ramanujan’s function τ (n), Proc. Camb. Philos. Soc.
35 (1939), 357–372.
[Re1] A. Reznikov, Non-vanishing of periods of automorphic functions, Forum Math. 13, No.4 (2001),
485–493.
[Re2] A. Reznikov, Rankin-Selberg without unfolding and bounds for spherical Fourier coefficients of
Maass forms, preprint (2005). arXiv: math.NT/0509077.
[Sa] P. Sarnak, Integrals of products of eigenfunctions, IMRN, no. 6, (1994), 251–260.
[Se] A. Selberg, On the estimation of Fourier coefficients, in Collected works, Springer-Verlag, New
York (1989), 506–520.
[St] E. Stein, Harmonic analysis: real-variable methods, orthogonality, and oscillatory integrals.
Princeton Mathematical Series, 43. Princeton University Press, Princeton, NJ, 1993.
[V] A. Venkatesh, Sparse equidistribution problems, period bounds, and subconvexity, preprint
(2005). arXiv: math.NT/0506224.
[W] N. Wallach, Real reductive groups. I. Pure and Applied Mathematics, 132. Academic Press,
Boston, MA, 1988.
[Wa] T. Watson, Thesis, Princeton, 2001.