Uncertainty Principles in Framed Hilbert Spaces
Uncertainty Principles in Framed Hilbert Spaces
TigerPrints
December 2020
Recommended Citation
Li, Haodong, "Uncertainty Principles in Framed Hilbert Spaces" (2020). All Dissertations. 2748.
https://tigerprints.clemson.edu/all_dissertations/2748
This Dissertation is brought to you for free and open access by the Dissertations at TigerPrints. It has been
accepted for inclusion in All Dissertations by an authorized administrator of TigerPrints. For more information,
please contact [email protected].
Uncertainty Principles in Framed Hilbert Spaces
A Dissertation
Presented to
the Graduate School of
Clemson University
In Partial Fulfillment
of the Requirements for the Degree
Doctor of Philosophy
Mathematical Science
by
Haodong Li
December 2020
Accepted by:
Mishko Mitkovski, Committee Chair
Shitao Liu
Martin Schmoll
Jeong-Rock Yoon
Introduction
The uncertainty principle is one of the most fundamental concepts in harmonic analysis. It
has many facets and appears in many different (non-equivalent) forms. However, a common theme
of all uncertainty principles can be vaguely summarized into two points of view. From Fourier
analysis perspective: In 1925, Wiener first formulate the uncertainty principle [19] in his Göttingen
lecture. It says a function and its Fourier transform cannot both be sharply localized. From operator
theoretical perspective: In 1927, two years after the Wiener’s lecture, Heisenberg formulated his
famous uncertainty principle [7], saying that the position and the momentum of a quantum particle
cannot be measure simultaneously. Mathematically, this can be expressed as the multiplication and
1
[D, X] := DX − XD = I.
2πi
Which shows a strong non-commutativity of X and D. These two points of view are related with
each other through the fact that the Fourier transform F conjugates the operators X and D. Namely,
X = FDF ∗ .
The thesis mainly consists of three projects whose common theme is the uncertainty prin-
ciple. Two of them concern interpolation and sampling problems, while the third one is about the
so-called generalized Balian-Low theorem. We now briefly describe each of them, and give an outline
The main results of the thesis are included in the last three chapters. The main purpose
of Chapter 1 is to introduce the well-known concepts that will be used throughout the thesis, and
set up the notations. Besides this, we state the classical uncertainty principle and the classical
Balian-Low theorem that will be generalized in the later chapters. We also include proofs of these
ii
results which are based on the similar ideas with our proofs of more general results. The reason for
In Chapter 2, we introduce the concept of framed Hilbert spaces as a general setting for
studying problems of uncertainty principle type. These spaces could be viewed as a special type of
reproducing kernel Hilbert spaces which include many function spaces that play an important role
in harmonic analysis. In the chapter, we prove some basic results about framed Hilbert spaces that
will be used in the rest of the thesis. Many of those are new results in this level of generality.
principle in the following way. Take the classical Paley-Wiener space as an example. This space
fixed interval. As a consequence of the classical uncertainty principle, these functions cannot be
localized on any finite intervals. However, we have more than that. It turns out that the oscillation
of the functions get controlled by their band limit. Therefore, it is difficult to solve interpolation
problems for such functions. More precisely, to be able to solve interpolation problems, the density
of interpolating points must be sparse enough relative to the band limit. However, at the same time,
it is easy to solve sampling problems. In other words, such functions could be completely determined
on the set (sampling set) which is sufficiently dense relative to the band limit.
In Chapter 3, we define a very general interpolation set called d-approximate (weak) in-
terpolation set in framed Hilbert spaces. This type of sets were relatively recently introduced by
Olevskii and Ulanovskii in the classical Paley-Wiener space. And we will prove a necessary density
condition for those sets similar to the one for usual interpolation sets.
Like Chapter 3, sampling problems could be also studied in the general framework of framed
Hilbert spaces. However, we will concentrate on a more precise problem of estimating a sampling
constant in the space of multiband-limited functions. A sharp estimate of this constant in classi-
cal Paley-Wiener spaces is given by Kovrijkine, who also provides a quantitative estimate for this
constant in multiband cases. In both of these cases, the dependence of this sampling constant on
the length of the (multi)band is exponential. In Chapter 4, we will impose additional (suboptimal)
conditions on the sampling set which allow us to obtain a much-improved sampling constant which
The classical uncertainty principle gives a bound on how close a function could become a
joint eigenvector of X and D. This classical version of uncertainty principles also gives us the optimal
iii
approximate joint eigenvectors, which are the time-frequency shifts of the Gaussian. However, it
turns out that the time-frequency shifts of the Gaussian, unfortunately, cannot form an orthonormal
basis (even a Riesz basis). Moreover, if we use an integer lattice of time-frequency shifts of some
generating function to form an orthonormal basis, it turns out this generating function must have
a very poor time-frequency localization. This result of uncertainty principle type is the so-called
Balian-Low theorem.
In Chapter 5, we will give a very general version of Balian-Low theorem, which only requires
the operator pair to satisfy a weak non-commutativity property and does not necessarily require the
orthonormal (or Riesz) basis to be the time-frequency shifts of a single function. Furthermore, we
examine the relationship between diagonalization and possibility of joint orthonormal (or Riesz)
iv
Table of Contents
Title Page . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Multiplication and Differentiation Operators . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Classical Uncertainty Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Classical Balian–Low Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Frames and Riesz Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .100
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .101
v
Chapter 1
Preliminaries
Arguably, the two most important operators in analysis are the multiplication and differen-
tiation operators:
Xf (x) := xf (x), ∀x ∈ R.
1 0
Df (x) := f (x), ∀x ∈ R.
2πi
X ∗ = X; D∗ = D.
Using X and D, we could also define the following two groups of operators:
1
The group of translations Ta : L2 (R) → L2 (R) is given by
Ta := e−2πiaD , ∀a ∈ R.
Mb := e2πibX , ∀b ∈ R.
Moreover, we could show a → Ta and b → Mb are two unitary representations from (R, +)
to U(H), which means Ta and Mb are the unitary operators such that
XMb = Mb X; DTa = Ta D.
However, simple calculation shows that X and D do not commute but satisfy the so-called canonical
commutation relation:
1
[D, X] := DX − XD = I, (1.1)
2πi
where I : D(X) ∩ D(D) → D(X) ∩ D(D) is the identity map. Rewrite (1.1) in terms of Ta and Mb ,
it is equivalent to
Ta Mb = e−2πiba Mb Ta .
2
As a consequence of the canonical commutation relation, we obtain
In next section, we will see the canonical commutation relation plays an important role in the
Before we state the classical uncertainty principle, we need to introduce a very important
Z
fb(ξ) := f (x)e−2πixξ dx, ∀ξ ∈ R.
R
And we will also denote fb by Ff . This notation indicates that we would consider the Fourier
from L2 (R) onto L2 (R). Throughout the thesis, we would view the Fourier transform F as an
unitary operator on L2 (R). Since F is unitary on L2 (R), its inverse Fourier transform F −1
Z
f (x) = fb(ξ)e2πixξ dξ, ∀x ∈ R.
R
For us, the importance of the Fourier transform F mainly comes from the fact that it connects
FX = −DF; FD = XF.
3
We now could state the classical uncertainty principle (which is often referred to as the
Theorem 1.2.1 ([5], Theorem 1.1). For any f ∈ L2 (R) and any a, b ∈ R,
Z Z
1 4 2
kf k2 ≤ |(x − a)f (x)| dx |(ξ − b)fb(ξ)|2 dξ. (1.3)
16π 2 R R
2
Equality holds if and only if f (x) = Ce2πibx e−c(x−a) for some C ∈ C and c > 0.
Note the inequality (1.3) can be written as the following equivalent form in terms of X and
1 4 2 2
kf k2 ≤ k(X − a)f k2 k(D − b)f k2 . (1.4)
16π 2
Proof. Here is a simple proof for Schwartz class S(R) using the canonical commutation relation.
1 2 1
kf k2 = f, f
2πi 2πi
= h[D, X]f, f i
= h(DX − XD)f, f i
1 4 2
kf k2 = |Im h(X − a)f, (D − b)f i|
16π 2
2
≤ |h(X − a)f, (D − b)f i|
2 2
≤ k(X − a)f k2 k(D − b)f k2 . (1.5)
Notice equality in (1.5) holds if and only if (D − b)f = ic(X − a)f for some c ∈ R. This is the
4
differential equation:
2
Its solutions are Ce2πibx e−c(x−a) for every C ∈ C. Finally, f ∈ L2 (R) requires c > 0.
The proof for f ∈ L2 (R) is similar. Instead of using (1.1), we need the generalized
1 2
kf k2 = hXf, Df i − hDf, Xf i . (1.6)
2πi
Since (1.6), X and D cannot have a common eigenvector in L2 (R) (otherwise we would get
a contradiction). And (1.4) could be viewed as a quantitive version of this argument. It tell us there
is no sequence of L2 (R) functions which approximates to and performs like a common eigenvector
for X and D. However, we could find some functions in L2 (R) which minimize the right hand side in
(1.4). Naturally, we might call these functions the “best approximate common eigenvectors”. And
2
it turns out these approximants are the multiples of Mb Ta g, where g(x) = e−cx is the Gaussian
Recall the spectral theorem in functional analysis, if there are two compact self-adjoint
operators on a Hilbert space commuting with each other, then we could find a sequence of common
eigenvectors which forms an orthonormal basis for the Hilbert space. Go back to operators X and
D, the classical uncertainty principle shows that instead of common eigenvectors, we could find some
“best approximate common eigenvectors” for X and D. The natural question is: can we use these
“best approximants” to form an orthonormal basis just like in the spectral theorem. Unfortunately,
If we set g(x) := 1[0,1] (x), the sequence {Mm Tn g}m,n∈Z which does not consist of the “best
approximants” does form an orthonormal basis of L2 (R). However, in this case, the Fourier transform
of g has a very poor localization property. The question is: can we find some g ∈ L2 (R) such that
{Mm Tn g}m,n∈Z forms an orthonormal basis, in addition, g and gb are both well localized. Again,
the answer is No! It is given by the so-called Balian-Low theorem. In order to state the theorem,
5
we need to introduce the Gabor systems.
Definition 1.3.1. A Gabor system is a sequence in L2 (R) of the form {Mmb Tna g}m,n∈Z , where
The set Λ = {(na, mb)}m,n∈Z is called the lattice with volume ab. Actually, we could generalize the
definition for high dimensional cases, i.e., we could define the Gabor system {Mmb Tna g}m,n∈Zn for
Theorem 1.3.2 ([2], Theorem 1.1). Let g ∈ L2 (R) and let a, b > 0 satisfy ab = 1. If the Gabor
Z Z
2
|xg(x)| dx g (ξ)|2 dξ = ∞.
|ξb (1.7)
R R
Z Z
|(x − x0 )g(x)|2 dx g (ξ)|2 dξ = ∞.
|(ξ − ξ0 )b
R R
Z Z
|xg(x)|2 dx g (ξ)|2 dξ < ∞.
|ξb
R R
6
Which implies Xg, Dg ∈ L2 (R), then by (1.2) we have
hXg, Dgi
X
= hXg, Mmb Tna gi hMmb Tna g, Dgi
m,n
X
= hT−na M−mb Xg, gi hg, T−na M−mb Dgi
m,n
X
= h(XM−mb T−na + naM−mb T−na )g, gi hg, (DM−mb T−na + mbM−mb T−na )gi
m,n
X
= hXM−mb T−na g, gi hg, DM−mb T−na gi
m,n
X
= hM−mb T−na g, Xgi hDg, M−mb T−na gi
m,n
X
= hDg, M−mb T−na gi hM−mb T−na g, Xgi
m,n
= hDg, Xgi .
It contradicts with the fact that {Mmb Tna g}m,n∈Z forms an orthonormal basis.
The classical Balian-Low theorem tells us that there is no well-localized window function
g in both time and frequency, which generates an orthonormal basis. Furthermore, in some sense
(1.7) is sharp. Actually, Benedetto et al. [1] shows that for any ε > 0, there exists an orthonormal
2 2
1 + |x| 1 + |ξ|
Z Z
|g(x)|2 dx g (ξ)|2
|b dξ < ∞.
R log1+ε (2 + |x|) R
2+ε
log (2 + |ξ|)
If we ask for the Gabor system to be a weaker type of basis such that the Balian-Low theorem
still holds. The answer would be the Riesz basis which is slightly weaker than the orthonormal basis.
7
basis, if there exists an invertible operator U : H → H such that
f k = U ek , ∀k ∈ N,
where {ek }∞
k=1 is an orthonormal basis of H.
Throughout the thesis, we always view H as a complex and separable Hilbert space. Just
as any basis, the Riesz basis has two dual features: spanning and independence. The sequence in
H which is usually viewed as a representative of spanning property is the so-called frame. And the
Riesz sequence. In this section, for completeness we will give their definitions and state their main
properties. And we will not include proofs which can be found in [4, 20].
The (optimal: maximal for A and minimal for B) constants are called the (optimal) lower
and upper frame bounds. A frame is said to be tight if we can pick A = B as frame bounds. And
∞
X
2 2
kf k = |hf, fk i| , ∀f ∈ H.
k=1
span{fk }∞
k=1 = H.
It is not hard to see that if a sequence satisfies the inequality to the left in (1.8), then it
has to be complete. As a consequence, any frame is a complete sequence. If a sequence satisfies the
8
The (optimal) constant B is called the (optimal) Bessel bound. By the definition, we can
prove every Bessel sequence is a bounded sequence. Another equivalent definition of Bessel sequences
is as the following.
∞
X
T : l2 (N) → H, T {ck }∞
k=1 := ck fk .
k=1
∞
X
S : H → H, Sf := T T ∗ f = hf, fk i fk .
k=1
Actually, {fk }∞
k=1 being a Bessel sequence not only guarantees the operators are well-defined,
but also make them being bounded. A very interesting fact is that if the operator T or T ∗ is well
From the definition, it is easy to see that the frame operator S is positive and self-adjoint.
In addition, if {fk }∞
k=1 is a frame of H, then the frame operator S is invertible, i.e., S
−1
exists. It
9
follows from the definition that
∞
X
f, S −1 fk fk ,
f= ∀f ∈ H;
k=1
X∞
f= hf, fk i S −1 fk , ∀f ∈ H.
k=1
Notice there are some elements fek := S −1 fk (for any k ∈ N) appearing in the last two series. The
sequence {fek }∞ ∞
k=1 is called the canonical dual frame of {fk }k=1 . And one could show the canonical
Furthermore, if {fk }∞
k=1 is a Parseval frame, then S = I. So the canonical dual frame is
∞
X
f= hf, fk i fk , ∀f ∈ H.
k=1
such that
2
∞
X
X∞
∞
X
2 2
A |ck | ≤
ck fk
≤ B |ck | , (1.9)
k=1 k=1 k=1
The (optimal) constants A and B are called the (optimal) lower and upper Riesz bounds.
By Proposition 1.4.5, we have already known a sequence satisfies the inequality to the right
in (1.9) is a Bessel sequence. Now we introduce another important sequence which satisfies the
inequality to the left in (1.9), such sequence is the so-called Riesz-Fischer sequence.
10
About Riesz-Fischer sequences, we also have the following two useful equivalent definitions.
of sequence {fk }∞
k=1 in H is solvable, i.e., there exists an element f ∈ H such that
hf, fk i = ck , ∀k ∈ N,
sequence {gk }∞
k=1 of H such that
sequence). Go back to the Riesz basis, we have the following two important propositions:
sequence.
11
Chapter 2
Spaces
In Chapter 2, we will introduce a very important concept for the thesis which is the so-called
“continuous frame”, and make some preparations for the rest of the chapters.
12
3. there exist constants A, B > 0 such that
Z
2 2 2
A kf k ≤ |hf, kx i| dσ(x) ≤ B kf k , ∀f ∈ H.
X
Throughout the thesis, we may omit the measure σ for some continuous frame {kx }x∈X ,
Namely,
Z
2 2
kf k = |hf, kx i| σ(x), ∀f ∈ H.
X
kkx k = 1, ∀x ∈ X.
Note that if the measure space X = N with σ being the counting measure, then the continu-
ous frame {kx }x∈N becomes a discrete frame (see 1.4.2). So continuous frames are the generalization
of discrete frames, and some mathematicians use the phrase the generalized frame instead of the
continuous frame.
Proposition 2.1.2. Let {kx }x∈(X,σ) be a continuous frame of H, then {kx }x∈(X,σ) is complete, i.e.,
span{kx }x∈X = H.
Definition 2.1.3. Let {kx }x∈(X,σ) be a continuous frame of H, the frame operator S : H → H is
given by
Z
Sf := hf, kx i kx dσ(x),
X
where the right-hand side is defined in the weak sense, i.e., as the unique element in H such that
Z
hSf, gi = hf, kx i hkx , gi dσ(x), ∀g ∈ H.
X
Just like the discrete case, the frame operator is invertible for continuous frames. Let
of {kx }x∈X . In addition, the canonical dual frame is also a continuous frame of H.
13
Again, if {kx }x∈X is a continuous Parseval frame, the corresponding frame operator is the
this case
Z
f= hf, kx i kx dσ(x), ∀f ∈ H.
X
Definition 2.1.4. A continuous Gabor frame is a family in L2 (Rn ) of the form {Mb Ta g}(a,b)∈R2n ,
Proposition 2.1.5. The continuous Gabor frame {Mb Ta g}(a,b)∈(R2n ,m) is a continuous frame of
L2 (Rn ), where m is the Lebesgue measure on R2n . If kgk2 = 1, then {Mb Ta g}(a,b)∈R2n is a normalized
Proof. We give the following proof for n = 1. In high dimensional case, the proof is similar. For any
f ∈ L2 (R),
Z Z
2
|hf, Mb Ta gi| dbda
R R
Z Z Z 2
= 2πibx g(x − a)dx dbda
f (x)e
R R R
Z Z Z 2
f (x)g(x − a)e−2πibx dx dbda
=
ZR ZR R 2
= F f (·)g(· − a) (b) dbda
R R
Z Z 2
= f (b)g(b − a) dbda
ZR ZR
2
= |f (b)g(b − a)| dadb
ZR R Z
2 2
= |f (b)| |g(b − a)| dadb
ZR ZR
2 2
= |f (b)| |g(a)| dadb
R R
2 2
= kgk2 kf k2 . (2.1)
14
2
Denote kgk2 by A, then (2.1) shows that
Z Z
2
A kf k2 = | hf, Mb Ta gi |2 dbda, ∀f ∈ L2 (R).
R R
Z Z
2 2
kf k2 = |hf, Mb Ta gi| dbda, ∀f ∈ L2 (R).
R R
Another example of continuous frames is the exponential functions, which is very familiar
Proposition 2.1.6. Let E be a subset of R with finite Lebesgue measure, then the exponential func-
tions {e2πiξx }ξ∈(R,m) is a continuous Parseval frame of L2 (E). If m(E) = 1, then {e2πiξx }ξ∈(R,m)
Proof. Let f be any function in L2 (E), and we view f as a L2 (R) function supported on E. Notice
Z D E 2
f, e2πiξ(·)
dξ
2
L (E)
R
Z Z 2
= f (x)e−2πiξx dx dξ
R E
Z Z 2
−2πiξx
= f (x)e
dx dξ
ZR R 2
= fb(ξ) dξ
ZR
2
= |f (x)| dx
ZR
2
= |f (x)| dx
E
2
= kf kL2 (E) .
15
It follows that {e2πiξx }ξ∈(R,m) is a continuous Parseval frame of L2 (E). Furthermore, if m(E) = 1,
Z
2πiξ(·)
2
2πiξx 2
e
2 = e dx = m(E) = 1.
L (E) E
Definition 2.2.1. A framed Hilbert space (FHS) is a pair (H, {kx }x∈(X,σ) ), where H is a complex
Actually, one can view a FHS as another more familiar space, which is the so-called repro-
is said to be a reproducing kernel Hilbert space (RKHS) if for any x ∈ X, the evaluation
functional δx given by
δx : H → C, δx (f ) := f (x),
is bounded.
such that
f (x) = hf, Kx i , ∀f ∈ H.
The collection of elements {Kx }x∈X is called the reproducing kernel of H; the element Kx is
called the reproducing kernel of H at point x. And the collection {kx := Kx / kKx k}x∈X is called
the normalized reproducing kernel of H; the element kx is called the normalized reproducing
kernel of H at point x.
The following two propositions show that every RKHS which is embedded in a L2 (X, γ) for
Proposition 2.2.3. Let H ⊆ L2 (X, γ) be a RKHS, and {Kx }x∈X be its reproducing kernel. Then
16
2
(H, {kx }x∈(X,σ) ) forms a FHS, where dσ(x) := kKx k dγ(x), and {kx }x∈X is the normalized repro-
ducing kernel.
Z
2 2
kf k = |f (x)| dγ(x)
ZX
2
= |hf, Kx i| dγ(x)
ZX
2 2
= |hf, kx i| kKx k dγ(x)
X
Z
2
= |hf, kx i| dσ(x),
X
for any f ∈ H. It implies {kx }x∈(X,σ) is a normalized continuous Parsevel frame, and (H, {kx }x∈(X,σ) )
forms a FHS.
Proposition 2.2.4. Let (H, {kx }x∈(X,σ) ) be a framed Hilbert space. Then H can be viewed as a
H
e := {fe : X → C| fe(x) := hf, kx i , f ∈ H}.
Z
e 2
kfek22 : = f (x) dσ(x)
ZX
2
= |hf, kx i| dσ(x)
X
2
= kf k .
17
boundedness property:
δx (fe) : = fe(x)
= |hf, kx i|
≤ kf k kkx k
= kf k
= kfek2 ,
for any fe ∈ H.
e By δx is bounded, H
e is a RKHS. Since f → fe is unitary, for any x ∈ X we have
D E
fe(x) = hf, kx i = fe, e
kx , ∀fe ∈ H.
e
2
So {e
kx }x∈X is the reproducing kernel of H.
e
Define L2 (Rn ) as the Hilbert space as usual. And let {Mb Ta g}(a,b)∈R2n be the continuous
The classical Bargmann-Fock space Fα (Cn ) with the parameter α > 0 is the space of
αn
Z
2 2 2
kf kα := |f (z)| e−α|z| dm(z) < ∞,
πn Cn
18
where m is the Lebesgue measure on Cn . It is known that Fα (Cn ) is a reproducing kernel Hilbert
α 2
kzα (w) = eαhw,zi− 2 |z| , ∀w ∈ Cn .
Define measure mα by
2 αn −α|z|2 αn
dmα (z) := kKz kα n
e dm(z) = n dm(z),
π π
then the classical Bargmann-Fock space (Fα , {kzα }z∈(Cn ,mα ) ) forms a framed Hilbert space.
We will use k·k , Kz (w), and kz (w) to denote the norm, the reproducing kernel, and the normalized
Because there is an unitary operator B (the Bargmann transform) from L2 (Rn ) onto F(Cn ),
which maps the continuous Gabor frame {Mb Ta g}(a,b)∈R2n generated by the normalized Gaussian
window function g onto the normalized reproducing kernel {kz }z∈Cn of F(Cn ) (up to some complex
numbers with modulus 1). One could view the classical Bargmann-Fock space (Fα , {kzα }z∈(Cn ,mα ) ) as
a special case of the FHS (L2 (Rn ), {Mb Ta g}(a,b)∈(R2n ,m) ), that is when g is the normalized Gaussian
function.
Let E be a subset of R with finite Lebesgue measure, we define the general Paley-Wiener
2πiξx
e 1
let { √ }ξ∈R be the √ multiple of exponential functions. As shown in Proposition 2.1.6,
m(E) m(E)
2πiξx
e
we obtain { √ }ξ∈(R,mE ) is a normalized continuous Parseval frame of L2 (E), where dmE :=
m(E)
19
2πiξx
e
The general Paley-Wiener space (L2 (E), { √ }ξ∈(R,mE ) ) forms a FHS.
m(E)
The classical Paley-Wiener space PWα with the parameter α > 0 is given by
sin α(y − x)
Kxα (y) = , ∀y ∈ R;
π(y − x)
√
α sin α(y − x)
kxα (y) = √ , ∀y ∈ R.
π α(y − x)
Define measure mα by
2 α
dmα (x) := kKxα k2 dm(x) = dm(x),
π
Then the classical Paley-Wiener space (PWα , {kxα }x∈(R,mα ) ) forms a FHS.
Because the Fourier transform F is an unitary operator from PWα onto L2 [−α, α], which
2πiξx
maps the normalized reproducing kernel {kxα }x∈R of PWα onto the continuous frame { e√2α }ξ∈R
2πiξx
of L2 [−α, α]. One could view the space (L2 [−α, α], { e√2α }ξ∈(R,m[−α,α] ) ) as the classical Paley-
Wiener space (PWα , {kxα }x∈(R,mα ) ). In addition, for any finite interval I ⊆ R with length 2α, there
2πiξx
exists a translation operator mapping from L2 [−α, α] onto L2 (I), which also maps { e√2α }ξ∈R onto
2πiξx
e
{√ }ξ∈R up to some constants with modulus 1. Based on these arguments, we could also view
m(I)
2πiξx
e
the space (L2 (I), { √ }ξ∈(R,|·|,mI ) ) as a classical Paley-Wiener space.
m(I)
Definition 2.3.1. Let µ be a σ-finite positive measure on the same measurable sets with σ, such
20
that
Z
|hkx , ky i| dµ(y) < ∞, ∀x ∈ X.
X
Z
Tµ : span{kx }x∈X → H, Tµ f := hf, kx i kx dµ(x),
X
Proposition 2.3.2. The Toeplitz operator Tµ with symbol µ is a positive symmetric linear operator.
1. Tµ is positive:
Z
hTµ f, f i = hf, kx i hkx , f i dµ(x)
X
Z
2
= |hf, kx i| dµ(x)
X
≥ 0,
2. Tµ is symmetric:
Z
hTµ f, gi = hf, kx i hkx , gi dµ(x)
X
Z
= hg, kx i hkx , f i dµ(x)
X
= hTµ g, f i
= hf, Tµ gi ,
Proposition 2.3.3. Let Tµ be a Toeplitz operator with a finite measure µ. Then Tµ can be extend
21
Proof. By Proposition 2.3.2, we only need to show the boundedness of Tµ on span{kx }x∈X . For any
2 2
kTµ f k = sup |hTµ , gi|
kgk=1
Z 2
= sup hf, kx i hkx , gi dµ(x)
kgk=1 X
Z Z
2 2
≤ sup |hf, kx i| dµ(x) |hkx , gi| dµ(x)
kgk=1 X X
Z Z
2 2 2 2
≤ sup kf k kkx k dµ(x) kkx k kgk dµ(x)
kgk=1 X X
2 2
= sup µ(X)2 kf k kgk
kgk=1
2
= µ(X)2 kf k .
Now we could view the Toeplitz operator Tµ as a bounded operator on the whole H, if we
Proposition 2.3.4. Let Tµ be a Toeplitz operator with a finite measure µ. Then Tµ is in the trace
class, and its trace and Hilbert-Schmidt norm satisfy the following identities:
Z Z
2
kTµ kT r = µ(X) = |hkx , ky i| dσ(y)dµ(x); (2.2)
X X
Z Z
2 2
kTµ kHS = |hkx , ky i| dµ(y)dµ(x). (2.3)
X X
22
Proof. Let {en }∞
n=1 be an orthonormal basis of H. On one hand,
kTµ kT r = T r(|Tµ |)
= T r(Tµ )
∞
X
= hTµ en , en i
n=1
X∞ Z
2
= |hen , kx i| dµ(x)
n=1 X
∞
Z X
2
= |hen , kx i| dµ(x)
X n=1
Z
2
= kkx k dµ(x)
X
= µ(X) < ∞.
Which implies Tµ is in the trace class, and kTµ kT r = µ(X). Then (2.2) comes from the following
Z Z
2
|hkx , ky i| dσ(y)dµ(x)
X X
Z
2
= kkx k dµ(x)
X
=µ(X).
23
On the other hand,
2
kTµ kHS = T r(Tµ∗ Tµ )
∞
X
Tµ∗ Tµ en , en
=
n=1
X∞
= hTµ en , Tµ en i
n=1
X∞ Z Z
= hen , kx i kx dµ(x), hen , ky i ky dµ(y)
n=1 X X
X∞ Z Z
= hen , kx i kx , hen , ky i ky dµ(y) dµ(x)
n=1 X X
X∞ Z Z
= hen , kx i hky , en i hkx , ky i dµ(y)dσ(x)
n=1 X X
X∞ Z Z
= hky , en i hen , kx i hkx , ky i dµ(y)dµ(x)
n=1 X X
∞
Z Z X
= hky , en i hen , kx i hkx , ky i dµ(y)dµ(x)
X X n=1
Z Z
= hky , kx i hkx , ky i dµ(y)dµ(x)
X X
Z Z
2
= |hkx , ky i| dµ(y)dµ(x).
X X
In general, it is not that easy to characterize the boundedness, compactness and the trace
class membership for Toeplitz operators without the assumption µ(X) < ∞. We will continue to do
Let (H, {kx }x∈(X,σ) ) be a FHS, we now introduce a very special Toeplitz operator by letting
the measure dµ := 1E dσ for some measurable subset E of X, such operator is called the concentration
operator.
over E by
Z
TE : H → H, TE f := hf, kx i kx dσ(x),
E
24
where the right-hand side is defined in the weak sense.
Proposition 2.3.6. Let TE be a concentration operator over the set E, then TE is bounded with
kTE k ≤ 1.
2 2
kTE f k = sup |hTE , gi|
kgk=1
Z 2
= sup hf, kx i hkx , gi dσ(x)
kgk=1 E
Z Z
2 2
≤ sup |hf, kx i| dσ(x) |hkx , gi| dσ(x)
kgk=1 E E
Z Z
2 2
≤ sup |hf, kx i| dσ(x) |hkx , gi| dσ(x)
kgk=1 X X
2 2
= sup kf k kgk
kgk=1
2
= kf k .
So kTE k ≤ 1.
Proposition 2.3.6 shows that as a special Toeplitz operator, the concentration operator TE
(equals Tµ , dµ := 1E dσ) will always being bounded. In addition, if we assume σ(E) is finite, we
have µ(X) = σ(E) < ∞. By Proposition 2.3.3 and 2.3.4, we obtain the following corollary.
Corollary 2.3.7. Let TE be the concentration operator over the set E with σ(E) < ∞. Then TE is
a positive compact self-adjoint operator, and its trace and Hilbert-Schmidt norm satisfy the following
identities:
Z Z
2
kTE kT r = σ(E) = |hkx , ky i| dσ(y)dσ(x); (2.4)
E X
Z Z
2 2
kTE kHS = |hkx , ky i| dσ(y)dσ(x). (2.5)
E E
Let E be a finite measure set, Corollary 2.3.7 tell us the concentration operator TE is a
positive compact self-adjoint operator. Combine Proposition 2.3.6, we know all the eigenvalues of
Definition 2.3.8. Let E be a measurable subset of X with σ(E) < ∞. Given a number 0 < c < 1,
25
we say that a subspace G of H is c-concentrated on set E if
Z
2 2
c kf k ≤ |hf, kx i| dσ(x),
E
for every f ∈ G.
Lemma 2.3.9. Let E be a Borel subset of X with σ(E) < ∞. Given a number 0 < c < 1, if G is a
σ(E)
dimG ≤ .
c
Proof. Let TE be the concentration operator over E. Denote all the eigenvalues of TE in the
principle,
≥ min hTE f, f i
f ∈G 0 ,kf k=1
Z
2
= min |hf, kx i| dσ(x)
f ∈G 0 ,kf k=1 E
2
≥ min c kf k
f ∈G 0 ,kf k=1
= c.
∞
X
σ(E) = kTE kT r = T r(TE ) = lj ≥ klk ≥ kc.
j=1
So we obtain
σ(E)
dimG 0 = k ≤ .
c
26
Lemma 2.3.9 shows that, given a finite measure set E, we could find a “inverse” relation
between the dimension of a concentrated subspace G on E and its concentration level c. Notice,
from the proof, if we knew the distribution of eigenvalues of TE , we could get a more precise relation
In this section, we will impose additional conditions on FHS to be able to obtain additional
properties. All these conditions are natural, and almost all of them are satisfied in the previous
examples.
Definition 2.4.1. We say (H, {kx }x∈(X,d,σ) ) is a framed Hilbert space with metric d, if
(H, {kx }x∈(X,σ) ) is a FHS and σ is a Borel measure with respect to the imposed metric d on X.
2πiξx
e
3. the general Paley-Wiener space (L2 (E), { √ }ξ∈(R,mE ) ),
m(E)
Because all of the measures in these FHS are the Borel measure with respect to the Euclidean metric
|·|. If we impose the Euclidean metric |·| on their index space respectively, then all of them are the
For convenience, we will continue using the phrase “framed Hilbert space” or “FHS ” to call
Let (H, {kx }x∈(X,d,σ) ) be a FHS, and we continue to impose conditions on it.
27
We say the metric measure space (X, d, σ) satisfies the doubling measure property
Simple calculation shows that the Euclidean metric space with the Lebesgue measure (Rn , |·| , m)
The metric measure space (X, d, σ) is said to satisfy the annular decay property (ADP ),
Easily see (Rn , |·| , m) satisfies ADP as well. For other metric measure spaces, we provide a
Proposition 2.4.2 ([3]). Let (X, d, σ) be a metric measure space. If (X, d, σ) satisfies DMP and
We say (X, d, σ) satisfies the uniform measure distribution (UMD), if for every r > 0
We say H satisfies the mean value property (MVP ), if for every r > 0 there exists Cr > 0
such that
Z
2 2
|hf, ka i| ≤ Cr |hf, kx i| dσ(x),
B(a,r)
28
Combining MVP with UMD, we have
Z
2 1 2
|hf, ka i| . |hf, kx i| dσ(x).
σ(B(a, r)) B(a,r)
It tell us that the averaging value of the function x → hf, kx i over a ball is greater or equal to
the value of the center. That’s the reason we borrow the terminology “mean value property” from
complex analysis here. Actually, the MVP usually holds when the Hilbert space is an analytic
function space.
We say {kx }x∈(X,d,σ) satisfies the approximate orthogonality property (AOP ), if the
followings hold:
1. there exists δ > 0 and c > 0 such that for any x, y ∈ X with d(x, y) < δ,
c ≤ |hkx , ky i| ;
2.
|hkx , ky i| → 0, as d(x, y) → ∞.
Notice AOP tell us that the angle between continuous frame kx and ky is small when x is
The continuous frame {kx }x∈(X,d,σ) is said to satisfy the uniform localization property
(ULP ), if
Z
2
lim sup |hka , kx i| dσ(x) = 0.
r→∞ a∈X B(a,r)c
In the rest part of the thesis, for convenience we just say (H, {kx }x∈(X,d,σ) ) satisfies DMP,
ADP, UMD, MVP, AOP and ULP, instead of mentioning (X, d, σ), H or {kx }x∈(X,d,σ) respectively.
1. The space (L2 (Rn ), {Mb Ta g}(a,b)∈(R2n ,|·|,m) ) satisfies DMP, ADP, UMD and AOP.
2. The classical Bargmann-Fock space (Fα , {kzα }z∈(Cn ,|·|,mα ) ) satisfies DMP, ADP, UMD, MVP,
2πiξx
e
3. The general Paley-Wiener space (L2 (E), { √ }ξ∈(R,|·|,mE ) ) satisfies DMP, ADP, UMD and
m(E)
AOP.
29
4. The classical Paley-Wiener space (PWα , {kxα }x∈(R,|·|,mα ) ) satisfies DMP, ADP, UMD, MVP, AOP
and ULP.
Proposition 2.4.3. Let (H, {kx }x∈(X,d,σ) ) be a FHS, which satisfies ADP, UMD and ULP. Then
Z Z
1 2
lim sup |hkx , ky i| dσ(y)dσ(x) = 0. (2.7)
r→∞ a∈X σ(B(a, r)) B(a,r) B(a,r)c
Z Z Z Z
+ .
B(a,r) B(a,r+ρ)c B(a,r) B(a,r+ρ)\B(a,r)
Estimate each term separately, and in both estimates we will divide by σ(B(a, r)) ( σ(B(a, r)) < ∞,
by UMD). About the first term, we have for any a ∈ X and r > 0,
Z Z
1 2
|hkx , ky i| dσ(y)dσ(x)
σ(B(a, r)) B(a,r) B(a,r+ρ)c
Z Z
1 2
≤ |hkx , ky i| dσ(y)dσ(x)
σ(B(a, r)) B(a,r) B(x,ρ)c
Z Z !
1 2
≤ sup |hkx , ky i| dσ(y) dσ(x)
σ(B(a, r)) B(a,r) x∈X B(x,ρ)c
Z
1 2
= σ(B(a, r)) sup |hkx , ky i| dσ(y)
σ(B(a, r)) x∈X B(x,ρ)c
Z
2
= sup |hkx , ky i| dσ(y)
x∈X B(x,ρ)c
Z
2
sup |hkx , ky i| dσ(y) < ε.
x∈X B(x,ρ)c
Z Z
1 2
lim sup sup |hkx , ky i| dσ(y)dσ(x)
r→∞ a∈X σ(B(a, r)) B(a,r) B(a,r+ρ)c
Z
2
≤ sup |hkx , ky i| dσ(y) < ε. (2.8)
x∈X B(x,ρ)c
30
Now we estimate the second term, using the positive ρ > 0 from above,
Z Z
1 2
|hkx , ky i| dσ(y)dσ(x)
σ(B(a, r)) B(a,r) B(a,r+ρ)\B(a,r)
Z Z
1 2
= |hkx , ky i| dσ(x)dσ(y)
σ(B(a, r)) B(a,r+ρ)\B(a,r) B(a,r)
Z Z
1 2
≤ |hkx , ky i| dσ(x)dσ(y)
σ(B(a, r)) B(a,r+ρ)\B(a,r) X
Z
1 2
= kky k dσ(y)
σ(B(a, r)) B(a,r+ρ)\B(a,r)
σ(B(a, r + ρ) \ B(a, r))
= .
σ(B(a, r))
Z Z
1 2
lim sup sup |hkx , ky i| dσ(y)dσ(x)
r→∞ a∈X σ(B(a, r)) B(a,r) B(a,r+ρ)\B(a,r)
σ(B(a, r + ρ) \ B(a, r))
≤ lim sup sup = 0. (2.9)
r→∞ a∈X σ(B(a, r))
Z Z
1 2
lim sup sup |hkx , ky i| dσ(y)dσ(x) < ε.
r→∞ a∈X σ(B(a, r)) B(a,r) B(a,r)c
Notice (2.7) is very similar with ULP, for some particular spaces, one implies another. Out
of the thesis, we might sometimes name (2.7) as ULP. As a direct consequence of Proposition 2.4.3,
Corollary 2.4.4. Let (H, {kx }x∈(X,d,σ) ) be a FHS satisfying ADP, UMD and ULP, and TB(a,r) be
the concentration operator over the ball B(a, r). Denote all its eigenvalues in the decreasing order
by 1 ≥ l1 (TB(a,r) ) ≥ l2 (TB(a,r) ) ≥ · · · ≥ 0, where entries are repeated with multiplicity. Then for any
∞
X ∞
X
(1 − ) li (TB(a,r) ) ≤ li (TB(a,r) )2 ,
i=1 i=1
31
Proof. By UMD, we have σ(B(a, r)) < ∞ for any a ∈ X and all r > 0. So we can apply Corol-
∞ ∞
!
1 X X
2
lim sup sup li (TB(a,r) ) − li (TB(a,r) ) = 0.
r→∞ a∈X σ(B(a, r))
i=1 i=1
∞
X ∞
X
li (TB(a,r) ) − li (TB(a,r) )2
i=1 i=1
2
=
TB(a,r)
T r −
TB(a,r)
HS
Z Z Z Z
2 2
= |hkx , ky i| dσ(y)dσ(x) − |hkx , ky i| dσ(y)dσ(x)
B(a,r) X B(a,r) B(a,r)
Z Z
2
= |hkx , ky i| dσ(y)dσ(x).
B(a,r) B(a,r)c
Because ADP and UMD are common properties, we usually view Corollary 2.4.4 as a
consequence of ULP. Study Corollary 2.4.4, it tell us every eigenvalue li (TB(a,r) ) of TB(a,r) should
be closed to either 1 or 0, and the closeness depends on how large of r. It somehow tell us the
information about the distribution of eigenvalues of TB(a,r) when r is large. Based on this concern,
we have the following Lemma which improves the inequality in Lemma 2.3.9 when r is large.
Lemma 2.4.5. Given a number 0 < c < 1, then for every 0 < θ < 1, there exists R > 0 such that
σ(B(a, r))
dimG < ,
θ
Proof. Let TB(a,r) be the concentration operator over an arbitrary ball B(a, r). Denote all the eigen-
subspace of H on B(a, r) and denote dimG by k (k < ∞, by Lemma 2.3.9). By the min-max
32
principle,
D E
lk (TB(a,r) ) = max min TB(a,r+ δ ) g, g
dimF =k g∈F ,kgk=1 2
D E
≥ min TB(a,r+ δ ) g, g
g∈G,kgk=1 2
≥ c.
dimG = k
≤ # i : li (TB(a,r) ≥ c
= # i : li (TB(a,r) ) > 1 − ε + # i : c ≤ li (TB(a,r) ) ≤ 1 − ε
X li (TB(a,r) ) X li (TB(a,r) )
≤ +
1−ε c
li >1−ε c≤li ≤1−ε
∞
1 X 1 X
≤ li (TB(a,r) ) + li (TB(a,r) ). (2.10)
1−ε c
i=1 li ≤1−ε
Let = cε2 . By Corollary 2.4.4, there exists R > 0 such that for any a ∈ X and all r > R, we have
∞
X
(1 − ) li (TB(a,r) )
i=1
∞
X
≤ li (TB(a,r) )2
i=1
X X
= li (TB(a,r) )2 + li (TB(a,r) )2
li ≤1−ε li >1−ε
X X
≤(1 − ε) li (TB(a,r) ) + li (TB(a,r) )
li ≤1−ε li >1−ε
∞
X X
= li (TB(a,r) ) − ε li (TB(a,r) ).
i=1 li ≤1−ε
∞
1 X X
li (TB(a,r) ) ≤ ε li (TB(a,r) ). (2.11)
c
li ≤1−ε i=1
33
Combining (2.10) and (2.11), we obtain that for any a ∈ X and all r > R,
∞
1−ε X
dimG ≤ li (TB(a,r) ). (2.12)
1 + ε − ε2 i=1
Notice as ε → 0+ ,
1−ε
↑ 1.
1 + ε − ε2
1−ε
> θ.
1 + ε − ε2
P∞ P∞
i=1 li (TB(a,r) ) i=1 li (TB(a,r) ) σ(B(a, r))
dimG ≤ 1−ε < = ,
1+ε−ε2
θ θ
In this section, based on the additional conditions we impose on FHS, we could give some
criteria to characterize the boundedness, compactness and trace class membership of Toeplitz oper-
ators without assuming µ(X) is finite. All of the results in this section are based on my master’s
thesis [11].
Definition 2.5.1. Let {kx }x∈(X,σ) be a continuous frame of H, and µ be another positive measure
for the same measurable sets with measure σ. We define the Berezin transform µ
e of measure µ
by
Z
2
e : X → R,
µ µ
e(x) := |hkx , ky i| dµ(y).
X
R
It is easy to check that if X
|hkx , ky i| dµ(y) < ∞ for every x ∈ X, then
e(x) = hTµ kx , kx i ,
µ ∀x ∈ X.
34
Therefore, the Berezin transform µ
e(x) of measure µ can be viewed as the “continuous diagonal
Definition 2.5.2. Let µ be a positive Borel measure on a metric space (X, d). For any r > 0, we
br : X → R,
µ µ
br (x) := µ(B(x, r)).
We now list the following theorems to characterize the boundedness, compactness and trace
Theorem 2.5.3 ([11], Theorem 5). Let (H, {kx }x∈(X,d,σ) ) be a FHS which satisfies DMP, UMD,
MVP and AOP. And Let µ be a σ-finite positive Borel measure on X that satisfies
Z
|hkx , ky i| dµ(y) < ∞, ∀x ∈ X.
X
Theorem 2.5.4 ([11], Theorem 6). Let (H, {kx }x∈(X,d,σ) ) be a FHS which satisfies DMP, UMD,
MVP and AOP. And Let µ be a σ-finite positive Borel measure on X that satisfies
Z
|hkx , ky i| dµ(y) < ∞, ∀x ∈ X.
X
a. Tµ : H → H is a compact operator.
b. µ̃(x) → 0 as x → ∞.
35
Theorem 2.5.5 ([11], Theorem 8). Given 1 ≤ p < ∞, Let (H, {kx }x∈(X,d,σ) ) be a FHS which
satisfies DMP, UMD, MVP and AOP. And Let µ be a σ-finite positive Borel measure on X that
satisfies
Z
|hkx , ky i| dµ(y) < ∞, ∀x ∈ X.
X
For RKHS, one could define the sampling and interpolation set. Similarly, we could give
Definition 2.6.1. Let (H, {kx }x∈(X,σ) ) be a FHS. A countable subset Λ := {λ} of X is called a
2
X 2 2
A kf k ≤ |hf, kλ i| ≤ B kf k , ∀f ∈ H.
λ∈Λ
From the definition, Λ is a sampling set for H if and only if the corresponding sequence of
Definition 2.6.2. Let (H, {kx }x∈(X,σ) ) be a FHS. A countable subset Λ := {λ} of X is called an
interpolation set for H, if for any complex sequence (cλ ) ∈ l2 (Λ), there exists f ∈ H such that
hf, kλ i = cλ , ∀λ ∈ Λ.
36
2.6.2 Density Theorems in Particular Framed Hilbert Spaces
Now we are going to list some important results about sampling and interpolation sets which
will be applied in the rest of the chapters. All of them are related to the so-called Beurling densities.
Definition 2.6.3. Let Λ be a countable subset of a metric measure space (X, d, σ). The lower
If σ is the Lebesgue measure, by convention, we will drop the subscript σ in the above notations. We
will use D− (Λ) and D+ (Λ) to denote the lower and upper Beurling densities in this case.
For example, all the integers Z is a countable subset of (R, |·| , m). It is not hard to see
D− (Z) = D+ (Z) = 1. If we pick all the natural numbers N as the countable subset, then D− (N) = 0
and D+ (N) = 1.
The following density theorem is based on the general Paley-Wiener space which is proved
by Landua in 1967.
Theorem 2.6.4 ([10]). Let E be a bounded measurable set on R, and Λ be a countable subset of R.
2πiξx
e
For the general Paley-Wiener space (L2 (E), { √ }ξ∈(R,|·|,mE ) ),
m(E)
−
Dm E
(Λ) ≥ 1;
+
Dm E
(Λ) ≤ 1.
Note Theorem 2.6.4 gives us a necessary density condition for sampling and interpolation
sets in the general Paley-Wiener space. Besides this, Beurling and Kahane proved a sufficient density
37
condition for sampling and interpolation sets in the classical Paley-Wiener space by the following
theorem:
Theorem 2.6.5 ([8]). Let I be a finite interval of R, and Λ be countable subset of R. For the
2πiξx
e
classical Paley-Wiener space (L2 (I), { √ }ξ∈(R,|·|,mI ) ),
m(I)
−
1. if Λ is relatively separated with Dm I
(Λ) > 1, then Λ is a sampling set for L2 (I), i.e. there exist
X D E 2
2 ≤ B kf k2 2 ,
f, e2πiλ(·) ∀f ∈ L2 (I);
A kf kL2 (I) ≤ L (I)
L2 (I)
λ∈Λ
+
2. if Λ is separated with Dm I
(Λ) < 1, then Λ is an interpolation set for L2 (I), i.e. there exists a
Similar results for the one-dimensional classical Bargmann-Fock space are proved by Seip
Theorem 2.6.6 ([17, 18]). Let Λ be a countable subset in C. For the one-dimensional classical
Bargmann-Fock space (Fα (C), {kzα }z∈(C,|·|,mα ) ), Λ is a sampling set for Fα (C) if and only if Λ is
Theorem 2.6.7 ([17, 18]). Let Λ be a countable subset in C. For the one-dimensional classical
Bargmann-Fock space (Fα (C), {kzα }z∈(C,|·|,mα ) ), Λ is an interpolation set for Fα (C) if and only if Λ
+
is separated and Dm α
(Λ) < 1.
For high dimensional Bargmann-Fock spaces, we don’t have the perfect if and only if density
conditions for sampling and interpolation sets. But still, Lindholm gives a very similar necessary
density condition for high dimensional weighted Bargmann-Fock spaces which implies the following
theorem:
Theorem 2.6.8 ([12]). Let Λ be a countable subset of Cn . For the classical Bargmann-Fock space
38
1. if Λ is a sampling set for Fα (Cn ), then Λ is relatively separated and contains a separated subset
−
Dm α
(Λ0 ) ≥ 1;
+
Dm α
(Λ) ≤ 1.
In general cases, we have the following theorem proved by Mitkovski and Ramirez which
Theorem 2.6.9 ([13], Theorem 4.5). Let (H, {kx }x∈(X,d,σ) ) be a FHS which satisfies the ADP,
Dσ+ (Λ) ≥ 1;
Dσ+ (Λ) ≤ 1.
39
Chapter 3
Approximate Interpolation in
3.1 Introduction
Let (H, {kx }x∈(X,σ) ) be a FHS. A countable subset Λ := {λ} of X is called an interpolation
set for H, if the corresponding sequence of continuous frame {kλ }λ∈Λ is a Riesz-Fischer sequence of
By Proposition 1.4.11, Λ is an interpolation set if and only if there exists a Bessel sequence
If we replace the Bessell sequence condition by the condition of bounded sequence, we would
Definition 3.1.1. Let (H, {kx }x∈(X,σ) ) be a FHS. A countable subset Λ := {λ} of X is called a
weak interpolation set for H, if there exists a bounded sequence {fλ }λ∈Λ such that
40
Because any Bessel sequence is bounded, we have every interpolation set is a weak in-
terpolation set. And for particular FHS, the converse is also true. For example, in the classical
Recall Theorem 2.6.9: let (H, {kx }x∈(X,d,σ) ) be a FHS which satisfies the ADP, UMD, MVP
and ULP , if a separated set Λ (will be defined in next section) is an interpolation set for H, then
Dσ+ (Λ) ≤ 1.
The goal of this chapter is to provide a similar type necessary density condition for an even
larger class of “interpolation sets”. This type of sets (to be defined momentarily) were relatively
recently introduced by Olevskii and Ulanovskii (see [15]) in the classical Paley-Wiener space, where
it was shown that all such sets have to satisfy a Beurling density condition similar to the one for
usual interpolation sets. Our results can be viewed as the generalization of their results.
Definition 3.1.2. Given 0 ≤ d < 1, we will say that a countable subset Λ := {λ} of X is a
d-approximate interpolation set for H, if there exists a Bessel sequence {hλ }λ∈Λ of H such that
X 2
|hhλ , kν i − δλν | ≤ d2 , ∀λ ∈ Λ.
ν∈Λ
Namely, the l2 (Λ) distance between the sequences (hhλ , kν i) and (δλν ) is no greater than d for any
λ.
Note that 0-approximate interpolation sets coincide with interpolation sets. Again, using a
bounded sequence instead of a Bessel sequence in Definition 3.1.2, we could define a weak type of
Definition 3.1.3. Given 0 ≤ d < 1, we will say that a countable subset Λ = {λ} of X is a d-
approximate weak interpolation set for H, if there exists a bounded sequence {fλ }λ∈Λ in H
such that
X 2
|hfλ , kν i − δλν | ≤ d2 , ∀λ ∈ Λ.
ν∈Λ
Again, 0-approximate weak interpolation sets coincide with weak interpolation sets in H.
41
And every d-approximate interpolation set is a d-approximate weak interpolation set. But we don’t
Our first result is based on FHS which gives a necessary density condition for d-approximate
interpolation sets.
Theorem 3.2.1. Let (H, {kx }x∈(X,d,σ) ) be a FHS which satisfy ADP, UMD, MVP and ULP. Given
Then
1
Dσ+ (Λ) ≤ .
1 − d2
Definition 3.2.2. A subset Λ in a metric space (X, d) is said to be separated or uniformly discrete
if
δ := inf{d(λ, ν) : λ 6= ν ∈ Λ} > 0.
Λ = ∪K
k=1 Λk ,
The following two propositions will show some simple but important properties for Λ being
relatively separated.
42
Proposition 3.2.4. Let (X, d, σ) be a metric measure space satisfying UMD, and Λ be a relatively
separated set in X. Then for any ball B(a, r) in X with center a ∈ X and radius r > 0,
Λ = ∪K
k=1 Λk ,
where Λk is a separated set with the separation constant δk for any k = 1, · · · , K. Let δ :=
min1≤k≤K δk , and {λki }i∈Ik := {Λk ∩ B(a, r)} for any k = 1, · · · , K. For every 1 ≤ k ≤ K, by Λk is
separated,
δ δ
B(λki , ) ∩ B(λkj , ) = ∅, ∀i 6= j.
2 2
δ δ
∪i∈Ik B(λki , ) ⊆ B(a, r + ).
2 2
X δ δ
σ(B(λki , )) ≤ σ(B(a, r + )).
2 2
i∈Ik
δ X δ δ
#{Λk ∩ B(a, r)}c( ) ≤ σ(B(λki , )) ≤ σ(B(a, r + )).
2 2 2
i∈Ik
K K
X X σ(B(a, r + 2δ )) mD(r + 2δ )
#{Λ ∩ B(a, r)} ≤ #{Λk ∩ B(a, r)} ≤ ≤ < ∞.
k=1 k=1
c( 2δ ) c( 2δ )
43
Combine ADP,
Actually, Theorem 3.2.1 tell us under additional interpolation assumptions on Λ, this trivial
density upper bound (3.1) could be significantly improved (especially when δ is close to 0).
The second consequence of Λ being relatively separated is that it will generate a Bessel
Proposition 3.2.5. Let (H, {kx }x∈(X,d,σ) ) be a FHS satisfying ADP, UMD, MVP and ULP. If Λ
is a relatively separated set in X, then there exists a constant C > 0 such that
X 2 2
|hf, kλ i| ≤ C kf k , ∀f ∈ H.
λ∈Λ
Λ = ∪K
k=1 Λk ,
where Λk is a separated set with the separation constant δk for any k = 1, · · · , K. Let δ :=
min1≤k≤K δk , and {λki }i∈Ik := {Λk ∩ B(a, r)} for any k = 1, · · · , K. Then by MVP, there exists
44
Cδ > 0 such that
X K X
X
2 2
|hf, kλ i| ≤ |hf, kλ i|
λ∈Λ k=1 λ∈Λk
K X
X Z
2
≤ Cδ |hf, kx i| dσ(x)
k=1 λ∈Λk B(λ, δ2 )
K X Z
X 2
= Cδ |hf, kx i| dσ(x)
k=1 λ∈Λk B(λ, δ2 )
K Z
X 2
≤ Cδ |hf, kx i| dσ(x)
k=1 X
2
= KCδ kf k .
In order to prove Theorem 3.2.1, we will adopt the proof strategy of Olevskii and Ulanovskii [15].
The argument has two crucial ingredients. The first one (essentially going back to Landau [10]) says
that any subspace of H which is c-concentrated on a fixed finite measure set cannot have arbitrar-
ily large dimension. We have these results as Lemma 2.3.9 and Lemma 2.4.5 in Chapter 2, for
Lemma. 2.3.9 Let E be a Borel subset of X with σ(E) < ∞. Given a number 0 < c < 1, if G is a
σ(E)
dimG ≤ .
c
Lemma. 2.4.5 Given a number 0 < c < 1, then for every 0 < θ < 1, there exists R > 0 such that
σ(B(a, r))
dimG < ,
θ
The following lemma is the second important ingredient in our proofs. It allows us to
45
generate a fairly high-dimensional subspace which is concentrated on some ball so that we can apply
the above two lemmas. Note the following result is a finite-dimensional result, which can be applied
Lemma 3.2.6 ([15], Lemma 2). Let {uj }1≤j≤N be an orthonormal basis in an N-dimensional
complex Euclidean space U. Given 0 < d < 1, suppose that {vj }1≤j≤N is a set of vectors in U
satisfying
kvj − uj k2 ≤ d2 , 1 ≤ j ≤ N.
Then for any b with 1 < b < 1/d, there is a subspace X of CN , such that
1. (1 − b2 d2 )N − 1 < dimX;
2. the estimate
N N
1 X X
(1 − )2 |cj |2 ≤ k cj vj k2 ,
b j=1 j=1
Proof. By Λ = {λ} ⊆ X is a d-approximate interpolation set for H, there exists a Bessel sequence
Let B(a, r) be an arbitrary open ball in X. Since Λ is relatively separated, Λ ∩ B(a, r) is finite set.
vj := ( hλj , kλ1 , · · · , hλj , kλN ), 1 ≤ j ≤ N,
uj := (δλj λ1 , · · · , δλj λN ), 1 ≤ j ≤ N.
46
Notice that
N
hλj , kλi − δλj λi 2
2
X
kvj − uj k =
i=1
hλ , kν − δλ ν 2
X
≤ j j
ν∈Λ
≤ d2 , 1 ≤ j ≤ N.
By Lemma 3.2.6, for any 1 < b < 1/d, there exists a subspace U of CN , such that
1.
2. the inequality
N N
1 X X
(1 − )2 |cj |2 ≤ k cj vj k2 , (3.3)
b j=1 j=1
PN
Let Y := { j=1 cj vj |c = (c1 , ..., cN ) ∈ U } be the subspace of CN . we want to show dimU ≤ dimY .
The idea is that using a basis of U to create some linearly independent vectors of Y .
M
X
0= a i wi
i=1
M
X N
X
= ai cij vj
i=1 j=1
N M
!
X X
= ai cij vj
j=1 i=1
47
P
M PM
Notice the scalars of coefficients i=1 ai cij = i=1 ai ci ∈ U , by (3.3) we have
1≤j≤N
2
!
N X M
X
ai cij vj
0=
j=1 i=1
N
M 2
1 2 X X i
≥ (1 − ) ai cj ,
b j=1 i=1
PM
which implies the linear combination i=1 ai ci = 0. By {ci }M
i=1 is a basis, ai = 0, 1 ≤ i ≤ M. So
{wi }M
i=1 are the linearly independent vectors of Y . Then
PN
Let G := { j=1 cj hλj |c = (c1 , ..., cN ) ∈ U } be the subspace of H, now we want to prove dimY ≤
dimG.
Let T ∗ be the analysis operator of the finite sequence {kλ1 , kλ2 , ..., kλN } on H. By T ∗ hλj =
PN
Let g = j=1 cj hλj ∈ G for some (c1 , ..., cN ) ∈ U . Using that {hλ }λ∈Λ is a Bessel sequence and
48
(3.3), we get
N
*
N X N
+2
X 2
X
|hg, kλi i| =
cj hλj , kλi
i=1 i=1 j=1
N
X
=k cj vj k2
j=1
N
1 X
≥ (1 − )2 |cj |2
b j=1
N
1 X
≥ (1 − )2 Ck cj hλj k2
b j=1
2
= C1 kgk , (3.7)
for any g ∈ G.
Let δ be the separation constant of Λ. Then B(λ, δ/2) ∩ B(ν, δ/2) = ∅ for any λ 6= ν ∈ Λ.
N
X N
X Z
2 2
|hg, kλi i| ≤ Cδ |hg, kx i| dσ(x)
i=1 i=1 B(λi , δ2 )
Z
2
≤ Cδ |hg, kx i| dσ(x), (3.8)
B(a,r+ δ2 )
Z D E
2 2
c kgk ≤ |hg, kx i| dσ(x) = TB(a,r+ δ ) g, g , (3.9)
2
B(a,r+ δ2 )
for any g ∈ G, where 0 < c := C1 /Cδ < 1 is independent of a and r. It follows that G is a
For every 0 < θ < 1, apply Lemma 2.4.5, there exists R > 0 such that for any ball B(a, r+ 2δ )
with r > R,
σ(B(a, r + 2δ ))
dimG < . (3.10)
θ
σ(B(a, r + 2δ ))
(1 − b2 d2 )#{Λ ∩ B(a, r)} − 1 < .
θ
49
Then for every a ∈ X and r > R,
σ(B(a, r + 2δ )) 1
#{Λ ∩ B(a, r)} < + .
θ(1 − b2 d2 ) (1 − b2 d2 )
σ(B(a, r + 2δ ))
= lim sup sup
r→∞ a∈X θ(1 − b2 d2 )σ(B(a, r))
1 σ(B(a, r + 2δ ))
= lim sup sup
θ(1 − b2 d2 ) r→∞ a∈X σ(B(a, r))
1
= .
θ(1 − b2 d2 )
Fock Spaces
Our next result is based on the classical Bargmann-Fock space (F(Cn ), |·| , m). In the
classical one-dimensional Bargmann-Fock space F(C), it was shown by Seip and Schuster [16] that
the class of weak interpolation sets coincides with the class of interpolation sets.
Another important result (Theorem 2.6.7) of Seip and Wallstén shows that, in the classical
in terms of the upper Beurling density D+ (Λ). Namely, Λ is an interpolation set (or equivalently
50
3.3.1 Main Theorem
For the classical Bargmann-Fock space (F(Cn ), |·| , m), as a special FHS, we could improve
Theorem 3.2.1 by applying a weaker condition that Λ is a d-approximate weak interpolation set.
Theorem 3.3.1. Let (F(Cn ), |·| , m) be the classical Bargmann-Fock space. Given 0 ≤ d < 1,
suppose a relatively separated subset Λ of Cn is a d-approximate weak interpolation set for F(Cn ).
Then
1
D+ (Λ) ≤ .
1 − d2
This result can be easily extended to all of the classical Bargmann-Fock spaces Fα (Cn ).
Recall in Chapter 2, the classical Bargmann-Fock space (Fα (Cn ), {kz }z∈(Cn ,|·|,mα ) ) is a FHS
satisfying DMP, ADP, UMD, MVP, AOP and ULP. In addition, Fα (Cn ) is a RKHS. Its reproducing
α 2
kzα (w) = eαhw,zi− 2 |z| , ∀w ∈ Cn .
About the reproducing kernels, we have the following important equalities hold:
2 2
kKzα kα = hKzα , Kzα iα = eα|z| , ∀z ∈ Cn ,
α 2
|hkzα , kw
α
iα | = e− 2 |z−w| , ∀z, w ∈ Cn .
Proof. By Λ = {λ} ⊆ Cn is a d-approximate weak interpolation set for F(Cn ), there exists a
X 2
|hfλ , kν i − δλν | ≤ d2 , ∀λ ∈ Λ.
ν∈Λ
51
Let ε > 0. For any λ ∈ Λ define gλ (z) := fλ (z)kλε (z). Clearly gλ : Cn → C is entire as a product of
2 2
|kλε (z)| = |hkλε , Kzε i|
2 2
≤ kkλε kε kKzε kε
2
= eε|z| ,
Z
2
|gλ (z)|2 e−(π+ε)|z| dm(z)
Cn
Z
2
= |fλ (z)kλε (z)|2 e−(π+ε)|z| dm(z)
n
ZC
2
≤ |fλ (z)|2 e−π|z| dm(z) < ∞.
Cn
X
2
gλ , kνπ+ε π+ε − δλν
ν∈Λ
X
−1 2
gλ , Kνπ+ε π+ε
Kνπ+ε
π+ε − δλν
=
ν∈Λ
X
−1 2
gλ (ν)
Kνπ+ε
π+ε − δλν
=
ν∈Λ
X
−1 2
fλ (ν)kλε (ν)
Kνπ+ε
π+ε − δλν
=
ν∈Λ
X
−1 2
hfλ , Kν i hkλε , Kνε iε
Kνπ+ε
π+ε − δλν
=
ν∈Λ
X
−1 2
hfλ , kν i kKν k hkλε , kνε iε kKνε kε
Kνπ+ε
π+ε − δλν
=
ν∈Λ
X 2
= |hfλ , kν i hkλε , kνε iε − δλν |
ν∈Λ
X 2
≤ |hfλ , kν i − δλν | ≤ d2 , (3.11)
ν∈Λ
π+ε
Kν
π+ε
= kKν k kKνε kε .
52
Let B(a, r) be any ball in Cn . Since Λ is relatively separated, Λ ∩ B(a, r) is a finite set. Let
N
X 2
2
gλj , kλπ+ε
kvj − uj k = − δ
i π+ε λj λi
i=1
X
2
gλj , kνπ+ε π+ε − δλj ν
≤
ν∈Λ
≤ d2 , 1 ≤ j ≤ N.
By Lemma 3.2.6, for any 1 < b < 1/d, there exists a subspace U of CN , such that
1. (1 − b2 d2 )N − 1 < dimU ;
2. the inequality
N N
1 X X
(1 − )2 |cj |2 ≤ k cj vj k2 , (3.12)
b j=1 j=1
PN
Let G := { j=1 cj gλj |c = (c1 , ..., cN ) ∈ U }. By the same argument of (3.6) in the proof of
Theorem 3.2.1,
(1 − b2 d2 )N − 1 ≤ dimG. (3.13)
53
PN
Let g = j=1 cj gλj ∈ G for some (c1 , ..., cN ) ∈ U . Using that {kλπ+ε }λ∈Λ is a Bessel sequence (due
X
2
2
g, kλπ+ε π+ε
kgkπ+ε ≥ C
λ∈Λ
* + 2
N X
N
X
π+ε
≥C
cj gλj , kλi
i=1 j=1 π+ε
2
N X
N
X
π+ε
≥C
cj gλj , kλi π+ε
i=1 j=1
2
N
X
=C
cj vj
j=1
N
X
≥ C(1 − 1/b)2 |cj |2
j=1
N
X
= C1 |cj |2 , (3.14)
j=1
Fix a small ρ > 0. A simple application of the Cauchy-Schwarz inequality and the already
Z
2 2 π
|g(z)| e−(π+ε)|z| −nlog( π+ε )
dm(z)
B(a,r+ρr)c
Z
2 2
= |g(z)| e−(π+ε)|z| dmπ+ε (z)
B(a,r+ρr)c
2
N
Z X
ε
−(π+ε)|z|2
=
c f
j λj (z)kλj (z) e
dmπ+ε (z)
B(a,r+ρr)c j=1
N Z N 2
X 2
X 2
≤ |cj | fλj (z)kλε j (z) e−(π+ε)|z| dmπ+ε (z)
j=1 B(a,r+ρr)c j=1
N Z N E 2
X 2
X D
−2
fλj , Kz kλε j , Kzε
Kzπ+ε
π+ε dmπ+ε (z)
= |cj |
B(a,r+ρr)c j=1 ε
j=1
N Z N E 2
X 2
X D
fλj , kz kλε j , kzε dmπ+ε (z),
= |cj | (3.15)
B(a,r+ρr)c j=1 ε
j=1
for any g ∈ G. We now estimate the integral term in (3.15). Applying {fλ }λ∈Λ is bounded and
54
doing a simple change of variables, we obtain
Z N E 2
X D
fλj , kz kλε j , kzε dmπ+ε (z)
B(a,r+ρr)c j=1 ε
Z N
X D E 2
fλj
kkz k kλε j , kzε dmπ+ε (z)
≤
B(a,r+ρr)c j=1 ε
Z N D
ε ε 2
X E
≤C kλj , kz dmπ+ε (z)
B(λj ,ρr)c j=1 ε
N Z
ε ε 2
X D E
=C kλj , kz dmπ+ε (z)
B(λj ,ρr)c ε
j=1
N Z
X 2
=C e−ε|z−λj | dmπ+ε (z)
j=1 B(λj ,ρr)c
Z
2
=CN e−ε|z| dmπ+ε (z). (3.16)
B(0,ρr)c
Λ = ∪K
k=1 Λk ,
where Λk is a separated set with the separation constant δk for any k = 1, · · · , K. Let δ :=
min1≤k≤K δk , and {λki }i∈Ik := {Λk ∩ B(a, r)} for any k = 1, · · · , K. The simple counting argument
55
And doing change of variables by surface coordinates, we get (3.16) is bounded by
Z
2
CN e−ε|z| dmπ+ε (z)
B(0,ρr)c
Z
2 2n 2n 2
<CK(1 + ) r e−ε|z| dmπ+ε (z)
δ B(0,ρr)c
Z ∞
2
=C 0 r2n e−εt t2n−1 dt
ρr
where C 0 does not depend on r (which depends on n, δ and K). Denote the last expression by C2 (r).
Observe that C2 (r) → 0 as r → ∞ (to be used in a moment). Combine (3.15), (3.16) and (3.17), we
Z N
2 2 π X 2
|g(z)| e−(π+ε)|z| −nlog( π+ε )
dm(z) ≤ C2 (r) |cj | . (3.18)
B(a,r+ρr)c j=1
Z
C2 (r) 2 2 2 π
(1 − ) kgkπ+ε ≤ |g(z)| e−(π+ε)|z| −nlog( π+ε )
dm(z)
C1 B(a,r+ρr)
Z
2
g, kzπ+ε π+ε dmπ+ε (z)
=
B(a,r+ρr)
= TB(a,r+ρr) g, g π+ε .
Let 0 < < 1. Since C2 (r)/C1 → 0 as r → ∞, there exists R > 0, such that
2
(1 − ) kgkπ+ε ≤ TB(a,r+ρr) g, g π+ε ,
for every g ∈ G when r > R. In other words, the subspace G is (1 − )-concentrated on B(a, r + ρr)
56
Combining (3.13) and (3.19), we obtain for every a ∈ Cn and r > R,
Therefore,
1
D+ (Λ) ≤ .
1 − d2
57
Chapter 4
4.1 Introduction
Recall Beurling and Kahane’s density theorem for sampling sets in the classical Paley-Wiener
space.
Theorem. 2.6.5 Let I be a finite interval of R, and Λ be countable subset of R. For the classical
2πix(·)
−
Paley-Wiener space (L2 (I), { e√ }x∈(R,|·|,mI ) ), if Λ is relatively separated with Dm I
(Λ) > 1, then
m(I)
Λ is a sampling set for L2 (I). Namely, there exist constants A, B > 0 such that
X D E 2
2 ≤ B kf k2 2 ,
f, e2πiλ(·) ∀f ∈ L2 (I).
A kf kL2 (I) ≤ 2 L (I)
L (I)
λ∈Λ
The following theorem could be viewed as an analogue for “continuous sampling sets” in
Theorem 4.1.1 ([9], Theorem 1 Logvinenko-Sereda). Let E be an interval with the length b and let
S be a measurable subset of R. If there exist a > 0 and γ > 0 such that |S ∩ I| := m(S ∩ I) ≥ γa
−C(ab+1)
Z
C
b
2 b 2
f
2 ≤ f (x) dx, (4.1)
γ L (R) S
58
for any function f ∈ L2 (E).
−C(ab+1) Z D 2
C 2 f, e2πix(·)
E
dx ≤ kf k2 2
kf kL2 (E) ≤ L (E) .
γ
S L2 (E)
2πix(·)
e
It shows us the exponentials { √ }x∈(S,mE ) forms a continuous frame of L2 (E) with lower frame
m(E)
−C(ab+1)
bound Cγ and upper frame bound 1. By this observation, Theorem 4.1.1 tell us that
when the subset S is pretty “dense” in R, S could be viewed as a “continuous sampling set” of the
2πix(·)
e
classical Paley-Wiener space (L2 (E), { √ }x∈(R,|·|,mE ) ).
m(E)
In 1973, Logvinenko and Sereda first proved inequality (4.1) for some constant of lower
frame bound. Then Kovrijkine improved it by giving a formula for that lower frame bound. In
addition, Kovrijkine developed and generalized Theorem 4.1.1 even for multiband-limited functions.
Theorem 4.1.2 ([9], Theorem 2 Kovrijkine). Let E := ∪ni=1 Ii be the union of n intervals with the
same length b. And let S be a measurable subset of R. If there exist a > 0 and γ > 0 such that
−2ab( Cγ )n −2n+1
Z
C
b
2 b 2
f
≤ f (x) dx, (4.2)
γ L2 (R) S
Notice the inverses of lower frame bound appearing in (4.1) and (4.2) both grow exponen-
tially with the length of E. The goal of this chapter is to provide a better lower frame bound whose
inverse would grow linearly with the length of E under certain conditions.
Exponentials
In order to state our theorems in an easier way, it is better to impose the conditions on set
Σ, which plays the role of the complement of set S in Theorem 4.1.1 and Theorem 4.1.2. In those
59
theorems, we impose conditions to make set S “dense” in R. For our theorems, we would like to
When Σ is a countable union of bounded intervals, we find two different ways to make set
Σ “sparse”. One way is to require that the length of intervals and the density of centers of intervals
are small; another way is to require the density of Σ itself is small. Based on these, we formulate
Theorem 4.2.1. Let E := ∪ni=1 Ii be the union of n bounded intervals. And let Λ := {λk }k∈Z be
a sequence, and Σ := ∪k∈Z (λk − 2b , λk + 2b ) be the countable union of intervals with center λk and
1
length b. If [b]D+ (Λ) < n, then there exists a constant C such that
2 Z
C
fb
2 ≤ |fb(x)|2 dx,
L (R) Σc
for any function f ∈ L2 (E). And C −1 grows linearly with |E| when |E| is large. Note: [b] denote
Theorem 4.2.2. Let E := ∪ni=1 Ii be the union of n bounded intervals. And let Σ = ∪k∈Z (λk −
bk bk
2 , λk + 2 ) be the countable disjoint union of intervals with the center λk and the length bk . If
1
inf k {bk } ≥ b > 0 for some constant b, and D+ (Σ) < n, then there exists a constant C such that
2 Z
C
fb
≤ |fb(x)|2 dx,
L2 (R) Σc
for any function f ∈ L2 (E). And C −1 grows linearly with |E| when |E| is large. Note: D+ (Σ) =
|Σ∩[a−r,a+r]|
lim supr→∞ supa∈R 2r as usual.
In order to prove Theorem 4.2.1 and Theorem 4.2.2, we will adopt the proof strategy of
Definition 4.2.3. Given any function f ∈ L1 (R) and any positive number ε. We define the random
periodization g εν of f by
1 X k+t
g εν (t) := √ f( ), t ∈ [0, 1)
εν εν
k∈Z
60
The series in the definition of g εν converges in L1 [0, 1), and for every ν this series represents
a 1-periodic function on R. The random periodization g εν allows some useful properties which will
Proposition 4.2.4. Given any function f ∈ L1 (R), let g εν be the random periodization of f . Then
εν (k) =
√
gc εν fb(ενk),
for any k ∈ Z.
Z 1
εν (k) =
gc g εν (t)e−2πikt dt
0
Z 1
1 X n + t −2πikt
= √ f( )e dt
0 εν εν
n∈Z
XZ 1 1 n + t −2πikt
= √ f( )e dt
εν εν
n∈Z 0
X Z n+1 1 t
= √ f ( )e−2πikt dt
εν εν
n∈Z n
Z ∞
1 t
= √ f ( )e−2πikt dt
−∞ εν εν
√
= εν fb(ενk).
Let Σ be a measurable subset of R and α be a positive number. Define the following subset
of Z
Λ(α, Σ) := {k ∈ Z : kα ∈ Σ},
Proposition 4.2.5 ([14], Proposition 2.2). Let g εν be the random periodization of the function f ,
X Z
E εν (k)|2 ≤ 2
|gc |fb(x)|2 dx.
06=k∈Λ
/ Σc
61
Note: E denote the expectation of random variable.
Lemma 4.2.6 ([14], Lemma 2.1). Let ϕ : R → [0, ∞) be a positive function, and let ε > 0 be a fixed
number. Then
Z 2 Z
X 1
ϕ(kεν)dν ≤ ϕ(x)dx.
1 k6=0 ε R
Z 2 X Z 2 X Z 2 X
ϕ(kεν)dν = ϕ(kεν)dν + ϕ(kεν)dν.
1 k6=0 1 k>0 1 k<0
Z 2 X XZ 2
ϕ(kεν)dν = ϕ(kεν)dν
1 k>0 k>0 1
X 1 Z 2εk
= ϕ(x)dx (x = kεν)
εk εk
k>0
1 X 2εk ϕ(x)
Z
= dx
ε k
k>0 εk
1 ∞ X ϕ(x)
Z
= dx
ε ε x x k
2ε <k< ε
1 ∞
Z X 1
= ϕ(x) dx
ε 0 x x k
2ε <k< ε
1 ∞
Z
≤ ϕ(x)dx, (4.3)
ε 0
1
P
where the last inequality comes from x x ≤ 1 for any x > 0. Similarly, for the second term,
2ε <k< ε k
we have
Z 2 Z 0
X 1
ϕ(kεν)dν ≤ ϕ(x)dx. (4.4)
1 k<0 ε −∞
Z 2 Z
X 1
ϕ(kεν)dν ≤ ϕ(x)dx.
1 k6=0 ε R
62
Proof of Proposition 4.2.5. By Proposition 4.2.4 and Lemma 4.2.6, we have
X X
E εν (k)|2 = E
|gc εν|fb(kεν)|2
06=k∈Λ
/ 06=k∈Λ
/
X
=E εν|fb(kεν)|2 1Σc (kεν)
k6=0
X
≤ 2εE |fb(kεν)|2 1Σc (kεν)
k6=0
Z 2 X
= 2ε |fb(kεν)|2 1Σc (kεν)dν
1 k6=0
Z
≤2 |fb(x)|2 dx.
Σc
To prove Theorem 4.2.1, we need Beurling and Kahane’s interpolation theorem (see Theo-
rem 2.6.5). For convenience, we modify the statement and restate as the following:
Theorem. 2.6.5 Let Λ be a sequence of real numbers, and I be an interval of torus T := R/Z. If Λ
is separated with D+ (Λ) < |I|, then there exists a constant C such that
X X
|cλ |2 ≤ Ck cλ e2πiλ(·) k2L2 (I)
λ∈Λ λ∈Λ
for any {cλ }λ∈Λ ∈ l2 (Λ). Where the constant C only depends on D+ (Λ) and |I|, i.e., C =
Besides of Beurling and Kahane’s interpolation theorem, we also need the following lemmas.
Consider any function f ∈ L2 (R) supported on the δ-neighbourhood Eδ of E for some δ > 0. Let
Let Eyεν := {t ∈ T : gyεν (t) 6= 0} be the support of gyεν , and Fyεν := {t ∈ T : gyεν (t) = 0} be
Lemma 4.2.7.
D+ (Λεν +
y ) ≤ ([b] + 2ε)D (Λ).
63
Proof. Without loss of generality, consider the case when y = 0. Let ενΛεν
0 := {kεν|kεν ∈ Σ, k ∈ Z}
#{ενΛεν
0 ∩ (a − r, a + r)}
b b b
=d e#{Λ ∩ (a − r − , a + r + )},
εν 2 2
b b
where d εν e denote the smallest integer bigger or equal to εν . Then
D+ (ενΛεν
0 )
#{ενΛεν 0 ∩ (a − r, a + r)}
= lim sup
r→∞ a 2r
b
d εν e#{Λ ∩ (a − r − 2b , a + r + 2b )}
≤ lim sup
r→∞ a 2r
b #{Λ ∩ (a − r − 2b , a + r + 2b )} 2r + b
=d e lim sup
εν r→∞ a 2r + b 2r
b
=d eD+ (Λ).
εν
D+ (Λεν + εν
0 ) = ενD (ενΛ0 )
b
≤ ενd eD+ (Λ)
εν
≤ ([b] + 2ε)D+ (Λ).
1
Lemma 4.2.8. For ε < 2|Eδ | , Fyεν contains an interval I εν with length |I εν | > (1 − 2ε|Eδ |) n1 .
64
Proof. By gyεν is the random periodization of fy ,
1 X k + t −2πi( k+t )y
gyεν (t) = √ f( )e εν .
εν εν
k∈Z
⇒ t ∈ ενEδ (mod1),
> 1 − |ενEδ |
> 1 − 2ε|Eδ |.
1−2ε|Eδ |
So T \ ενEδ (mod1) contains an interval I with length |I| > n . By Fyεν = T \ Eyεν ⊃ T \
1
Lemma 4.2.9. If [b]D+ (Λ) < n, then there exists a constant C such that
Z
sup |fb(y)|2 ≤ C |fb(x)|2 dx,
y∈Σ Σc
Proof. Let gyεν be the random periodization of f ∈ L2 (R) supported on Eδ . Rewrite gyεν as the sum
65
of functions pεν εν
y and qy ,
X
gyεν (t) = εν
gc
y (k)e
2πikt
k∈Z
= pεν εν
y (t) + qy (t),
D+ (Λεν + +
y ) ≤ ([b] + 2ε)D (Λ) ↓ [b]D (Λ).
On the other hand, by lemma 4.2.8, Fyεν contains an interval I εν with length |I εν | > (1 − 2ε|Eδ |) n1 .
Let ε → 0+ ,
1 1
|I εν | > (1 − 2ε|Eδ |) ↑ .
n n
1 1
Since [b]D+ (Λ) < n, let d := n − [b]D+ (Λ) > 0, and ε0 := min{ 6D+d(Λ) , 6|E
nd
δ|
}. Then
d 1 d
D+ (Λεy0 ν ) ≤ [b]D+ (Λ) + < − < |I ε0 ν |.
3 n 3
2
X 2
X
ε0 ν ε0 ν 2πik(·)
gy (k) ≤ C0
gy (k)e .
d
d
ε ν
k∈Λεy0 ν
k∈Λy0
L2 (I ε0 ν )
Notice gyε0 ν (t) = pεy0 ν (t) + qyε0 ν (t) = 0 on its zero set Fyε0 ν which contains interval I ε0 ν . Combin-
66
ing (4.5), we have for any y ∈ Σ,
1 X d
|fb(y)|2 ≤ |gyε0 ν (k)|2
ε0 ε0 ν
k∈Λy
C0 X d
≤ k gyε0 ν (k)e2πik(·) k2L2 (I ε0 ν )
ε0 ε0 ν
k∈Λy
C0 ε0 ν 2
= kp k 2 ε ν
ε0 y L (I 0 )
C0 ε0 ν 2
= kq k 2 ε ν
ε0 y L (I 0 )
C0 ε0 ν 2
≤ kq k (4.6)
ε0 y
C0 X d
= |gyε0 ν (k)|2 .
ε0 ε0 ν
k∈Λ
/ y
By Proposition 4.2.5,
X Z
Ekqyε0 ν k2 =E ε0 ν (k)|2 ≤ 2
|gd |fb(x)|2 dx.
ε ν Σc
/ y0
k∈Λ
By Chebyshev’s inequality,
Z
P{kqyε0 ν k2 > 4 |fb|2 }
Σc
Then,
Z
1
P{kqyε0 ν k2 ≤ 4 |fb|2 } > .
Σc 2
Z
kqyε0 ν0 k2 ≤4 |fb(x)|2 dx.
Σc
67
Since y ∈ Σ is arbitrary,
Z
4C0
sup |fb(y)|2 ≤ |fb(x)|2 dx,
y∈Σ ε0 Σc
nd
where ε0 = min{ 6D+d(Λ) , 6|E δ|
} with d = 1
n − [b]D+ (Λ), and C0 does not depend on |E|.
Proof of theorem 4.2.1. For every function f ∈ L2 (E) and every y ∈ Σ, define
e2πiyt
hy (t) := 1(δ,δ) (t), t ∈ R,
2δ
and
Fy (t) := f ∗ hy (t), t ∈ R,
Z
cy (x)|2 ≤ 4C0
sup |F cy (x)|2 dx.
|F (4.8)
x∈Σ ε0 Σc
Z
4C0
|fb(y)|2 = |F
cy (y)|2 ≤ sup |F
cy (x)|2 ≤ cy (x)|2 dx.
|F
x∈Σ ε0 Σc
Z Z Z
4C0
|fb(y)|2 dy ≤ cy (x)|2 dxdy
|F
Σ Σ ε 0 c
Z ZΣ
4C0 sin(2πδ(x − y)) 2
= |fb(x)|2 | | dxdy
ε0 Σ Σc 2πδ(x − y)
sin(2πδ(x − y)) 2
Z Z
4C0
= |fb(x)|2 | | dydx
ε0 Σc Σ 2πδ(x − y)
sin(2πδ(x − y)) 2
Z Z
4C0
= |fb(x)|2 | | dydx
ε0 Σc Σ 2πδ(x − y)
sin(2πδ(x − y)) 2
Z Z
4C0
≤ |fb(x)|2 | | dydx
ε0 Σc R 2πδ(x − y)
Z
2C0
= |fb(x)|2 dx.
ε0 δ Σc
68
Then we obtain
2 Z
ε0 δ
f
2 ≤ |fb(x)|2 dx.
b
ε0 δ + 2C0 L (R) Σc
nd
Since ε0 = min{ 6D+d(Λ) , 6|E δ|
} with d = 1 +
n −[b]D (Λ), and C0 does not rely on |E| (see Lemma 4.2.9).
1
Let δ = 2n , for large enough |E| we have
d ε0 δ
C := ≤ .
(24C0 + 1)(|E| + 1) ε0 δ + 2C0
It follows that
2 Z
C
fb
2 ≤ |fb(x)|2 dx,
L (R) Σc
In order to prove Theorem 4.2.2, we need the following new lemmas which play the same
role with Lemmas 4.2.7 and 4.2.9 for the proof of Theorem 4.2.1.
Again, consider any function f ∈ L2 (R) supported on the δ-neighbourhood Eδ of E for some
δ > 0. Let fy := M−y f be the modulation of f , and gyεν be the random periodization of fy .
Let Eyεν := {t ∈ T : gyεν (t) 6= 0} be the support of gyεν , and Fyεν := {t ∈ T : gyεν (t) = 0} be
Lemma 4.2.10. If inf k {bk } > b for some positive number b, then
2ε +
D+ (Λεν
y ) ≤ (1 + )D (Σ).
b
Proof. Without loss of generality, consider the case when y = 0. Let ενΛεν
0 := {kεν : kεν ∈ Σ, k ∈ Z}
69
those intersections by {l1 , l2 , · · · , ln }, then
#{ενΛεν
0 ∩ (a − r, a + r)}
|Σ ∩ (a − r, a + r)|
n≤ + 2.
b
Then
#{ενΛεν
0 ∩ (a − r, a + r)}
D+ (ενΛεν
0 ) = lim sup sup
r→∞ a 2r
|Σ∩(a−r,a+r)|
εν +n
≤ lim sup sup
r→∞ a 2r
|Σ∩(a−r,a+r)|
εν+ |Σ∩(a−r,a+r)|
b +2
≤ lim sup sup
r→∞ a 2r
1 1 |Σ ∩ (a − r, a + r)| 2
= lim sup sup ( + ) +
r→∞ a εν b 2r 2r
1 1 |Σ ∩ (a − r, a + r)|
= ( + ) lim sup sup
εν b r→∞ a 2r
1 1 +
= ( + )D (Σ).
εν b
By simple computations,
D+ (Λεν + εν
0 ) = ενD (ενΛ0 )
1 1
≤ εν( + )D+ (Σ)
εν b
2ε +
≤ (1 + )D (Σ).
b
70
1
Lemma 4.2.11. If D+ (Σ) < n, then there exists a constant C such that
Z
sup |fb(y)|2 ≤ C |fb(x)|2 dx,
y∈Σ Σc
Proof. Let gyεν be the random periodization of f ∈ L2 (R) supported on Eδ . We can rewrite gyεν as
X
gyεν (t) = εν
gc
y (k)e
2πikt
k∈Z
= pεν εν
y (t) + qy (t),
When y ∈ Σ, 0 ∈ Λεν
y . Then for y ∈ Σ,
2ε
By lemma 4.2.10, D+ (Λεν
y ) ≤ (1 +
+
b )D (Σ). Let ε → 0+ ,
2ε +
D+ (Λεν
y ) ≤ (1 + )D (Σ) ↓ D+ (Σ).
b
By lemma 4.2.8, Fyεν contains an interval I εν with length |I εν | > (1 − 2ε|Eδ |) n1 . Let ε → 0+ ,
1 1
|I εν | > (1 − 2ε|Eδ |) ↑ .
n n
71
1 1
Since D+ (Σ) < n, let d := n − D+ (Σ) > 0, and ε0 := min{ 6Dbd nd
+ (Σ) , 6|E | }. Then
δ
d 1 d
D+ (Λεy0 ν ) ≤ D+ (Σ) + < − < |I ε0 ν |.
3 n 3
Again, by Theorem 2.6.5, there exists a constant C0 = C0 (D+ (Σ) + d3 , n1 − d3 ) such that
2
2
X X
ε0 ν ε0 ν 2πik(·)
gy (k) ≤ C0
gy (k)e .
d
d
ε ν
k∈Λεy0 ν
k∈Λy0
L2 (I ε0 ν )
By gyε0 ν (t) = pεy0 ν (t) + qyε0 ν (t) = 0 on its zero set Fyε0 ν which contains the interval I ε0 ν . Combin-
1 X d
|fb(y)|2 ≤ |gyε0 ν (k)|2
ε0 ε0 ν
k∈Λy
C0 X d
≤ k gyε0 ν (k)e2πik(·) k2L2 (I ε0 ν )
ε0 ε0 ν
k∈Λy
C0 ε0 ν 2
= kp k 2 ε ν
ε0 y L (I 0 )
C0 ε0 ν 2
= kq k 2 ε ν
ε0 y L (I 0 )
C0 ε0 ν 2
≤ kq k (4.10)
ε0 y
C0 X d
= |gyε0 ν (k)|2 .
ε0 ε0 ν
k∈Λ
/ y
By Proposition 4.2.5,
X Z
Ekqyε0 ν k2 =E ε0 ν (k)|2 ≤ 2
|gd |fb(x)|2 dx.
ε ν Σc
/ y0
k∈Λ
72
By Chebyshev’s inequality,
Z
P{kqyε0 ν k2 >4 |fb|2 }
Σc
Then,
Z
1
P{kqyε0 ν k2 ≤4 |fb|2 } > .
Σc 2
Z
kqyε0 ν0 k2 ≤ 4 |fb(x)|2 dx.
Σc
Z
C0 ε0 ν0 2 4C0
|fb(y)|2 ≤ kq k ≤ |fb(x)|2 dx.
ε0 y ε0 Σc
Since y ∈ Σ is arbitrary,
Z
4C0
sup |fb(y)|2 ≤ |fb(x)|2 dx,
y∈Σ ε0 Σc
Proof of theorem 4.2.2. For every function f ∈ L2 (E) and every y ∈ Σ, define
e2πiyt
hy (t) := 1(δ,δ) (t), t ∈ R,
2δ
and
Fy (t) := f ∗ hy (t), t ∈ R.
Then
cy (x) = fb(x) sin(2πδ(x − y)) ,
F
2πδ(x − y)
73
and suppFy ⊆ Eδ . By Lemma 4.2.11
Z
cy (x)|2 ≤ 4C0
sup |F cy (x)|2 dx.
|F
x∈Σ ε0 Σc
Let x = y ∈ Σ, we obtain
Z
|fb(y)|2 = |F cy (x)|2 ≤ 4C0
cy (y)|2 ≤ sup |F cy (x)|2 dx.
|F
x∈Σ ε0 Σc
Z Z Z
4C0
|fb(y)|2 dy ≤ cy (x)|2 dxdy
|F
Σ Σ ε0 Σc
sin(2πδ(x − y)) 2
Z Z
4C0
= |fb(x)|2 | | dxdy
ε0 Σ Σc 2πδ(x − y)
sin(2πδ(x − y)) 2
Z Z
4C0
= |fb(x)|2 | | dydx
ε0 Σc Σ 2πδ(x − y)
sin(2πδ(x − y)) 2
Z Z
4C0
= |fb(x)|2 | | dydx
ε0 Σc 2πδ(x − y)
ZΣ
sin(2πδ(x − y)) 2
Z
4C0
≤ |fb(x)|2 | | dydx
ε0 Σc R 2πδ(x − y)
Z
2C0
= |fb(x)|2 dx.
ε0 δ Σc
Then we obtain
2 Z
ε0 δ
f
2 ≤ |fb(x)|2 dx.
b
ε0 δ + 2C0 L (R) Σc
d ε0 δ
C := ≤ .
(24C0 + 1)(|E| + 1) ε0 δ + 2C0
It follows that
2 Z
C
fb
≤ |fb(x)|2 dx,
L2 (R) Σc
74
Chapter 5
Hilbert Spaces
5.1 Introduction
Let X and D be the multiplication and differentiation operators on their domain D(X) and
D(D) (see Section 1.1). For any f ∈ L2 (R) and any a, b ∈ R, we define the notations:
Z
2 2
k(X − a)f k2 := |(x − a)f (x)| dx = ∞, when f ∈
/ D(X),
R
Z 2
2
k(D − a)f k2 := (ξ − b)fb(ξ) dξ = ∞, when f ∈
/ D(D).
R
Recall the classical uncertainty principle (Theorem 1.2.1): for any f ∈ L2 (R) and any a, b ∈ R,
2 2 1 4
k(X − a)f k2 k(D − b)f k2 ≥ kf k2 . (5.1)
16π 2
And the classical Balian-Low theorem (Theorem 1.3.2) could be restated as: Let g ∈ L2 (R) and
α, β > 0 with αβ = 1. If the Gabor system {Mmβ Tnα g}m,n∈Z forms a Riesz basis for L2 (R), then
for any a, b ∈ R,
2 2
k(X − a)f k2 k(D − b)f k2 = ∞, (5.2)
75
or equivalently
2 2
k(X − a)f k2 + k(D − b)f k2 = ∞.
Comparing the similarity of (5.1) and (5.2), the Balian-Low theorem could be view as the uncertainty
It is natural for mathematicians to release the Gabor system structures and work on more
general bases in next step. So Meyer asks the following question: is it true that for any orthonormal
basis {fn }∞ 2
n=1 ⊆ L (R) (which may not be a Gabor system) and any sequences of real numbers
{an }∞ ∞
n=1 and {bn }n=1 , we have
2 2
sup k(X − an )fn k2 + k(D − bn )fn k2 = ∞. (5.3)
n∈N
If this is true, then it would be a “generalized Balian-Low theorem” for general orthonormal bases.
Based on this fact, the next question will be: can we impose some extra conditions to
make (5.3) still hold. In 2011, Gröchenig and Malinnikova proved the following very similar equality
Z Z
sup |x − an |2s |fn (x)|2 dx + |ξ − bn |2s |fbn (ξ)|2 dξ = ∞,
n∈N R R
By studying Theorem 5.1.1, we get some hints to formulate the conditions for our “gener-
76
5.2 Generalized Balian-Low Theorem
In order to introduce our main theorem, we set up the following definition and notation:
Definition 5.2.1. We say a countable subset Λ of a metric space (X, d) decays annularly around
Let A and B be two operators on a Hilbert space H, denote their domain by D(A) := {f ∈
countable subset of some metric space (X, d) which decays annularly around point a ∈ X and satisfies
#{Λ ∩ B(a, r)} < ∞ for any r > 0. And let A : D(A) → H and B : D(B) → H be two symmetric
operators such that for every n ∈ N, hi[B, A]gn , fn i ≥ c for some c > 0. If
P 2
1. limr→∞ supn∈N d(λm −λn )>r |hAgn , fm i| = 0,
P 2
2. limr→∞ supn∈N d(λm −λn )>r |hBgn , fm i| = 0.
2 2
sup k(A − an )fn k + k(B − bn )fn k = ∞.
n∈N
Notice if we let Λ := Zd embedded in the Euclidean metric space (Rd , |·|), then Λ would
satisfy all the requirements in Theorem 5.2.2. Based on this fact, we have the following corollary:
Corollary 5.2.3. Let {fn }n∈Zd be an orthonormal basis of H. And let A : D(A) → H and B :
D(B) → H be two symmetric operators such that for every n ∈ Zd , hi[B, A]fn , fn i ≥ c for some
c > 0. If
77
P 2
1. limr→∞ supn∈Zd |m−n|>r |hAfn , fm i| = 0,
P 2
2. limr→∞ supn∈Zd |m−n|>r |hBfn , fm i| = 0.
Then for any sequences of complex numbers {an }n∈Zd , {bn }n∈Zd ,
2 2
sup k(A − an )fn k + k(B − bn )fn k = ∞.
n∈Zd
2 2
sup k(A − an )fn k + k(B − bn )fn k ≤ C < ∞.
n∈N
Since {fn }n∈N and {gn }n∈N are dual Riesz bases, for any n ∈ N we have
c ≤ hi[A, B]gn , fn i
Let B := B(a, r) be the ball with the fixed center a (for which Λ decays annularly around)
X X
c#{Λ ∩ B} ≤ i hBgn , fm i hgm , Afn i − hAgn , fm i hgm , Bfn i
λn ∈Λ∩B m∈N
X X
=i hBgn , fm i hgm , Afn i − hAgn , fm i hgm , Bfn i
λn ∈Λ∩B λm ∈Λ∩B
X X
+i hBgn , fm i hgm , Afn i − hAgn , fm i hgm , Bfn i .
λn ∈Λ∩B λm ∈Λ∩B c
78
By A, B are symmetric, the first term vanishes. So we have
X X
c#{Λ ∩ B} ≤ i hBgn , fm i hgm , Afn i − hAgn , fm i hgm , Bfn i .
λn ∈Λ∩B λm ∈Λ∩B c
X X
c#{Λ ∩ B} ≤ |hBgn , fm i hgm , Afn i − hAgn , fm i hgm , Bfn i|
λn ∈Λ∩B λm ∈Λ∩B c
X X
≤ |hBgn , fm i hgm , Afn i| + |hAgn , fm i hgm , Bfn i|
λn ∈Λ∩B λm ∈Λ∩B c
X X
≤ |hBgn , fm i hgm , Afn i|
λn ∈Λ∩B λm ∈Λ∩B c
X X
+ |hAgn , fm i hgm , Bfn i| . (5.4)
λn ∈Λ∩B λm ∈Λ∩B c
X X
|hBgn , fm i hgm , Afn i|
λn ∈Λ∩B λm ∈Λ∩B c
! 12 ! 21
X X 2
X X 2
≤ |hBgn , fm i| |hgm , Afn i| .
λn ∈Λ∩B λm ∈Λ∩B c λn ∈Λ∩B λm ∈Λ∩B c
79
assumption, the first factor satisfies
X X 2
|hBgn , fm i|
λn ∈Λ∩B λm ∈Λ∩B c
X X 2
X X 2
= |hBgn , fm i| + |hBgn , fm i|
λn ∈Λ∩B λm ∈Λ∩Bρc λn ∈Λ∩B λm ∈Λ∩(Bρ \B)
X X 2
X X 2
≤ |hBgn , fm i| + |hgn , Bfm i|
λn ∈Λ∩B λm ∈Λ∩B(λn ,ρ)c λm ∈Λ∩(Bρ \B) λn ∈Λ∩B
X X 2
X X 2
= |hBgn , fm i| + |hgn , (B − bm )fm i|
λn ∈Λ∩B λm ∈Λ∩B(λn ,ρ)c λm ∈Λ∩(Bρ \B) λn ∈Λ∩B
X X 2
X 2
≤ sup |hBgn , fm i| + C k(B − bm )fm k
n∈N
λn ∈Λ∩B λm ∈Λ∩B(λn ,ρ)c λm ∈Λ∩(Bρ \B)
X 2
X
≤ #{Λ ∩ B} sup |hBgn , fm i| + C
n∈N
λm ∈Λ∩B(λn ,ρ)c λm ∈Λ∩(Bρ \B)
X 2
= #{Λ ∩ B} sup |hBgn , fm i| + C#{Λ ∩ (Bρ \ B)}.
n∈N
λm ∈Λ∩B(λn ,ρ)c
X X 2
X X 2
|hgm , Afn i| = |hgm , (A − an )fn i|
λn ∈Λ∩B λm ∈Λ∩B c λn ∈Λ∩B λm ∈Λ∩B c
X 2
X
≤ C k(A − an )fn k ≤ C ≤ C#{Λ ∩ B}.
λn ∈Λ∩B λn ∈Λ∩B
Then for some C > 0, we get an estimate for the first term,
X X
|hBgn , fm i hgm , Afn i|
λn ∈Λ∩B λm ∈Λ∩B c
21
1
X 2
≤C#{Λ ∩ B} 2 #{Λ ∩ B} |hBgn , fm i| + #{Λ ∩ (Bρ \ B)} . (5.5)
λm ∈Λ∩B(λn ,ρ)c
Similarly, for some C > 0, we get an estimate for the second term,
X X
|hAgn , fm i hgm , Bfn i|
λn ∈Λ∩B λm ∈Λ∩B c
12
1
X 2
≤C#{Λ ∩ B} #{Λ ∩ B}
2 |hAgn , fm i| + #{Λ ∩ (Bρ \ B)} . (5.6)
λm ∈Λ∩B(λn ,ρ)c
80
So for any r > 0 and ρ > 0, combine (5.4) (5.5) (5.6), we have
1 X X
c≤ |hBgn , fm i hgm , Afn i|
#{Λ ∩ B}
λn ∈Λ∩B λm ∈Λ∩B c
1 X X
+ |hAgn , fm i hgm , Bfn i|
#{Λ ∩ B}
λn ∈Λ∩B λm ∈Λ∩B c
21
X 2 #{Λ ∩ (Bρ \ B)}
≤ C sup |hBgn , fm i| +
n∈N c
#{Λ ∩ B}
λm ∈Λ∩B(λn ,ρ)
12
X 2 #{Λ ∩ (B ρ \ B)}
+ C sup |hAgn , fm i| + . (5.7)
n∈N c
#{Λ ∩ B}
λm ∈Λ∩B(λn ,ρ)
X 2
X 2
sup |hBgn , fm i| < ε2 , sup |hAgn , fm i| < ε2 . (5.8)
n∈N n∈N
λm ∈Λ∩B(λn ,ρ)c λm ∈Λ∩B(λn ,ρ)c
1 1
c 2 #{Λ ∩ (Bρ \ B)} 2 2 #{Λ ∩ (Bρ \ B)} 2
≤ ε + + ε + ,
C #{Λ ∩ B} #{Λ ∩ B}
1 1
c 2 #{Λ ∩ (Bρ \ B)} 2 2 #{Λ ∩ (Bρ \ B)} 2
≤ lim ε + + ε +
C r→∞ #{Λ ∩ B} #{Λ ∩ B}
1 1
= ε2 + 0 2 + ε2 + 0 2 = 2ε.
5.2.3 Applications
Apply Corollary 5.2.3, we could give another proof of the classical Balian-Low theorem.
Proof of Theorem 1.3.2. For convenience, we let a = b = 1. The multiplication operator X and
differentiation operator D are two symmetric operators which will play the role of operators A and
81
Again, we will prove the theorem by contradiction. Suppose
Z Z
2 2 2
|xg(x)| dx g (ξ)|2 dξ = kXgk2 kDgk2 < ∞.
|ξb (5.9)
R R
Z
2 2
kXgm,n k2 := |Xgm,n (x)| dx
R
Z
2πimx 2
= xe g(x − n) dx
ZR
2 2
= |x + n| |g(x)| dx
R
Z Z
2 2 2 2
≤2 |x| |g(x)| dx + 2 |n| |g(x)| dx
R R
2 2 2
= 2 kXgk2 + 2 |n| kgk2 < ∞,
which implies {gm,n }m,n∈Z ⊆ D(X). Similarly, we could obtain {gm,n }m,n∈Z ⊆ D(D). By general-
82
By (1.2), we obtain
X 2
|hXgm,n , gk,l i|
k(k,l)−(m,n)k>r
X 2
= |hXMm Tn g, gk,l i|
k(k,l)−(m,n)k>r
X 2
= |hMm XTn g, gk,l i|
k(k,l)−(m,n)k>r
X 2
= |hMm (Tn X + nTn )g, gk,l i|
k(k,l)−(m,n)k>r
X 2
= |hMm Tn Xg, gk,l i|
k(k,l)−(m,n)k>r
X 2
= |hMm Tn Xg, Mk Tl gi|
k(k,l)−(m,n)k>r
X 2
lim sup |hXgm,n , gk,l i|
r→∞ m,n∈Z
k(k,l)−(m,n)k>r
X 2
= lim sup |hXg, gk,l i|
r→∞ m,n∈Z
k(k,l)k>r
X 2
= lim |hXg, gk,l i| = 0.
r→∞
k(k,l)k>r
Similarly,
X 2
lim sup |hDgm,n , gk,l i| = 0.
r→∞ m,n∈Z
k(k,l)−(m,n)k>r
Apply Corollary 5.2.3, we have for any sequences {am,n }m,n∈Z and {bm,n }m,n∈Z ,
2 2
sup k(X − am,n )gm,n k2 + k(D − bm,n )gm,n k2 = ∞ (5.10)
m,n∈Z
83
On the other hand, let am,n = n and bm,n = m for any m, n ∈ Z. Again by (1.2), we have
2 2
k(X − n)gm,n k2 + k(D − m)gm,n k2
2 2
= k(X − n)Mm Tn gk2 + k(D − m)Mm Tn gk2
2 2
= kMm (X − n)Tn gk2 + kMm DTn gk2
2 2
= kMm Tn Xgk2 + kMm Tn Dgk2
2 2
= kXgk2 + kDgk2 ,
2 2
sup k(X − n)gm,n k2 + k(D − m)gm,n k2
m,n∈Z
2 2
= sup kXgk2 + kDgk2
m,n∈Z
2 2
= kXgk2 + kDgk2 < ∞.
Our next result is based on FHS. In order to state the theorem, we introduce the following
Z
2
lim sup |hfn , kx i| dσ(x) = 0.
r→∞ n∈N B(λn ,r)c
Recall the ULP (in Section 2.4.2), it is not hard to see if {kx }x∈(X,d,σ) satisfies the ULP, then
84
sequence {fn }∞
n=1 of H is said to be uniformly w-localized on Λ if
Z
2
sup |hfn , kx i| w(d(λn , x))dσ(x) < ∞.
n∈N X
The following proposition proves that uniform w-localization implies uniform localization.
Proof. By {fn }∞
n=1 is uniformly w-localized on Λ,
Z
2
sup |hfn , kx i| w(d(λn , x))dσ(x) < ∞,
n∈N X
Z
2
sup |hfn , kx i| dσ(x)
n∈N B(λn ,r)c
Z
2 w(d(λn , x))
= sup |hfn , kx i| dσ(x)
n∈N B(λn ,r)c w(d(λn , x))
Z
2 w(d(λn , x))
≤ sup |hfn , kx i| dσ(x)
n∈N B(λn ,r)c w(r)
Z
1 2
≤ sup |hfn , kx i| w(d(λn , x))dσ(x)
w(r) n∈N X
1
. .
w(r)
By the property of w,
Z
2 1
lim sup |hfn , kx i| dσ(x) . lim = 0.
r→∞ n∈N B(λn ,r)c r→∞ w(r)
say w is localized on Λ if
X 1
lim sup = 0.
r→∞ n∈N w(d(λn , λm ))
d(λn ,λm )≥r
One example of localized weights is the power weight on the lattice of integers. Let Λ := Zn
85
be the subset of the Euclidean metric space (Rn , |·|), and let the weight function w(d) := ds be the
By Euclidean metric is equivalent to maximum metric, i.e., |x| ' maxi |xi |,
X 1 X 1
s .
|m| max{|m1 | , |m2 | , · · · , |mn |}s
m∈Zn \{0} m∈Zn \{0}
∞
X 1
= ((l + 1)n − ln )
ls
l=1
∞
X (2n − 1)ln−1
≤
ls
l=1
∞
X 1
.
ls−n+1
l=1
< ∞. (5.11)
It implies
X 1
lim sup s
r→∞ n∈Zn |n − m|
|n−m|≥r
X 1
= lim s
r→∞ |m|
|m|≥r
=0.
So in the Euclidean metric space (Rn , |·|), the weight function w(d) = ds with s > n is localized on
Λ = Zn .
Let (H, {kx }x∈(X,d,σ) ) be a FHS which satisfies ADP and UMD, then our next result is the
following:
{λn }∞ +
n=1 is relatively separated with 0 < D (Λ). Let A : D(A) → H and B : D(B) → H be two
symmetric operators such that for every n ∈ N, hi[B, A]gn , fn i ≥ c for some c > 0. If there exist
conditions hold:
86
2 2
1. supn∈N k(A − an )gn k + k(B − bn )gn k < ∞,
2. {(A − an )gn }∞ ∞
n=1 and {(B − bn )gn }n=1 are uniformly localized on Λ,
3. {fn }∞
n=1 is uniformly w-localized on Λ,
Then
2 2
sup k(A − an )fn k + k(B − bn )fn k = ∞.
n∈N
5.3.3 Prerequisites
Proposition 5.3.6. Let (X, d, σ) be a metric measure space satisfying ADP and UMD, and Λ be a
Λ = ∪K
k=1 Λk ,
where Λk is a separated set with the separation constant δk for any k = 1, · · · , K. Let δ :=
87
Combine ADP,
=0.
The next proposition is very important for us. We will use it to verify
X 2
lim sup |hAgn , fm i| = 0,
r→∞ n∈N
d(λm ,λn )>r
and
X 2
lim sup |hBgn , fm i| = 0,
r→∞ n∈N
d(λm ,λn )>r
Proposition 5.3.7. Let w : [0, ∞) → [0, ∞) be a weight function satisfies w(d) ↑ ∞ as d → ∞. Let
{hn }∞ ∞
n=1 be a sequence in H, and {fn }n=1 be a frame of H. If the following conditions hold
2
1. supn∈N khn k < ∞,
2. {hn }∞
n=1 is uniformly localized on Λ,
3. {fn }∞
n=1 is uniformly w-localized on Λ,
then
X 2
lim sup |hhn , fm i| = 0.
r→∞ n∈N
d(λm ,λn )>r
88
Proof. Fix any λn ∈ Λ, and define the ball B := B(λn , 2r ) for some r > 0. And let TB , TB c be two
Z Z
TB f := hf, kx i kx dσ(x), T Bc f := hf, kx i kx dσ(x).
B Bc
Z
f= hf, kx i kx dσ(x)
ZX Z
= hf, kx i kx dσ(x) + hf, kx i kx dσ(x)
B Bc
= TB f + TB c f.
2 2
|hhn , fm i| = |h(TB + TB c )hn , fm i|
2
= |hTB hn , fm i + hTB c hn , fm i|
2 2
≤ 2 |hTB hn , fm i| + 2 |hTB c hn , fm i| . (5.12)
Z
|hTB hn , fm i| = hhn , kx i hkx , fm i dσ(x)
ZB Z
2 2
≤ |hhn , kx i| dσ(x) |hkx , fm i| dσ(x)
X B
Z
2 2 w(d(λm , x))
= khn k |hfm , kx i| dσ(x)
B w(d(λm , x))
Z
2 1 2
≤ sup khn k sup |hfm , kx i| w(d(λm , x))dσ(x)
n∈N x∈B w(d(λm , x)) X
1
. sup . (5.13)
x∈B w(d(λm , x))
89
Notice when x ∈ B(λn , r/2) and d(λm , λn ) > r, we have
≤ d(λm , x) + r/2
X 2
lim sup |hTB hn , fm i|
r→∞ n∈N
d(λm ,λn )>r
X 1
. lim sup sup
r→∞ n∈N x∈B(λn , r2 ) w(d(λm , x))
d(λm ,λn )>r
X 1
≤ lim sup d(λm ,λn )
r→∞ n∈N
d(λm ,λn )>r w( 2 )
X 1
. lim sup
r→∞ n∈N w(d(λm , λn ))
d(λm ,λn )>r
=0. (5.15)
90
continuous Parseval frame,
X 2
lim sup |hTB c hn , fm i|
r→∞ n∈N
d(λm ,λn )>r
X 2
≤ lim sup |hTB c hn , fm i|
r→∞ n∈N
m
2
. lim sup kTB c hn k
r→∞ n∈N
Z
2
= lim sup
hhn , kx i kx dσ(x)
r→∞ n∈N
Bc
Z 2
= lim sup sup hhn , kx i kx dσ(x), g
r→∞ n∈N kgk=1 B c
Z 2
= lim sup sup hhn , kx i hkx , gi dσ(x)
r→∞ n∈N kgk=1 Bc
Z Z
2 2
≤ lim sup sup |hhn , kx i| dσ(x) |hkx , gi| dσ(x)
r→∞ n∈N kgk=1 Bc X
Z
2 2
= lim sup sup |hhn , kx i| dσ(x) kgk
r→∞ n∈N kgk=1 Bc
Z
2
= lim sup |hhn , kx i| dσ(x)
r→∞ n∈N Bc
Z
2
= lim sup |hhn , kx i| dσ(x)
r→∞ n∈N B(λn , r2 )c
=0, (5.16)
where the last equality comes from condition 2. Combining (5.12), (5.15) and (5.16), we have
X 2
lim sup |hhn , fm i|
r→∞ n
d(λm ,λn )>r
X 2 2
. lim sup |hTB hn , fm i| + |hTB c hn , fm i|
r→∞ n
d(λm ,λn )>r
=0.
91
5.3.4 Proof of Theorem 5.3.5
X 2
lim sup |hAgn , fm i|
r→∞ n∈N
d(λm ,λn )>r
X 2
= lim sup |h(A − an )gn , fm i| = 0; (5.17)
r→∞ n∈N
d(λm ,λn )>r
and
X 2
lim sup |hBgn , fm i|
r→∞ n∈N
d(λm ,λn )>r
X 2
= lim sup |h(B − bn )gn , fm i| = 0. (5.18)
r→∞ n∈N
d(λm ,λn )>r
Notice we have different assumptions on Λ from Theorem 5.2.2, we cannot use Theorem 5.2.2 to
prove Theorem 5.3.5 directly. However, we could borrow the first half of the proof of Theorem 5.2.2.
Again, we assume
2 2
sup k(A − an )fn k + k(B − bn )fn k < ∞,
n∈N
Let B := B(a, r) be any ball with the center a ∈ X and the radius r > 0, as in (5.4) we
have
X X
c#{Λ ∩ B} ≤ |hBgn , fm i hgm , Afn i|
λn ∈Λ∩B λm ∈Λ∩B c
X X
+ |hAgn , fm i hgm , Bfn i| . (5.19)
λn ∈Λ∩B λm ∈Λ∩B c
X X
|hBgn , fm i hgm , Afn i|
λn ∈Λ∩B λm ∈Λ∩B c
12
1
X 2
≤C#{Λ ∩ B} 2 #{Λ ∩ B} |hBgn , fm i| + #{Λ ∩ (Bρ \ B)} . (5.20)
λm ∈Λ∩B(λn ,ρ)c
92
Similarly, as in (5.6), for some C > 0 we have
X X
|hAgn , fm i hgm , Bfn i|
λn ∈Λ∩B λm ∈Λ∩B c
12
1
X 2
≤C#{Λ ∩ B} #{Λ ∩ B} 2 |hAgn , fm i| + #{Λ ∩ (Bρ \ B)} . (5.21)
λm ∈Λ∩B(λn ,ρ)c
Divide (5.19) by σ(B) on both sides. Combine (5.20) and (5.21), we have for every a ∈ X, r > 0
and ρ > 0,
c#{Λ ∩ B}
σ(B)
1 X X
≤ |hBgn , fm i hgm , Afn i|
σ(B)
λn ∈Λ∩B λm ∈Λ∩B c
1 X X
+ |hAgn , fm i hgm , Bfn i|
σ(B)
λn ∈Λ∩B λm ∈Λ∩B c
21
2
#{Λ ∩ B} X 2 #{Λ ∩ B} #{Λ ∩ (Bρ \ B)}
. sup |hBgn , fm i| +
σ(B)2 n∈N σ(B) σ(B)
λm ∈Λ∩B(λn ,ρ)c
21
2
#{Λ ∩ B} X 2 #{Λ ∩ B} #{Λ ∩ (Bρ \ B)}
+ sup |hAgn , fm i| + . (5.22)
σ(B)2 n∈N σ(B) σ(B)
λm ∈Λ∩B(λn ,ρ)c
Let ε > 0, by (5.17) and (5.18), we can find a ρ > 0 such that
X 2
X 2
sup |hBgn , fm i| < ε2 , sup |hAgn , fm i| < ε2 . (5.23)
n∈N n∈N
λm ∈Λ∩B(λn ,ρ)c λm ∈Λ∩B(λn ,ρ)c
21
#{Λ ∩ B}2 2 #{Λ ∩ B} #{Λ ∩ (Bρ \ B)}
c#{Λ ∩ B(a, r)}
. ε + ,
σ(B(a, r)) σ(B)2 σ(B) σ(B)
93
for any a ∈ X and r > 0. Then by Proposition 5.3.6, we obtain for such ρ
= εD+ (Λ).
Since Λ is relatively separated with D+ (Λ) > 0, by Proposition 3.1, we also have D+ (Λ) < ∞. By
5.3.5 Applications
Now we could give another proof of the main result in [6] applying Theorem 5.3.5.
2
2s 2 2s b
dξ ≤ S 2 < ∞ for every n,
R R
(a) |x − an | |fn (x)| dx + |ξ − b | f (ξ)
R R n n
2s 2 2s 2
gn (ξ)| dξ ≤ T 2 < ∞ for every n,
R R
(b) R
|x − an | |gn (x)| dx + R
|ξ − bn | |b
Then s ≤ 1.
Proof. Let (L2 (R), {Mb Ta g}(a,b)∈(R2 ,|·|,m) ) be our FHS which satisfies ADP and UMD, where g(x) :=
π 2
e− 2 x is normalized Gaussian function. And multiplication X and differentiation D will play the
Again, we will prove the theorem by contradiction. Suppose all the above assumptions
(a),(b),(c) hold and s > 1. Our proof strategy is the following: We will verify the following conditions
1-4 of Theorem 5.3.5 hold. Namely, there exists a weight function w such that
94
2 2
1. supn∈N k(X − an )gn k + k(D − bn )gn k < ∞,
2. {(X − an )gn }∞ ∞
n=1 and {(D − bn )gn }n=1 are uniformly localized on Λ,
3. {fn }∞
n=1 is uniformly w-localized on Λ,
2 2
sup k(X − an )fn k + k(D − bn )fn k < ∞.
n∈N
Let w(d) := d2s be the power weight function with exponent 2s. Obviously, w(2d) . w(d)
By assumption (a),
Z
2 2 2
k(X − an )fn k = |x − an | |fn (x)| dx
R
Z
2s 2
≤ (|x − an |
+ 1) |fn (x)| dx
R
Z Z
2s 2 2
= |x − an | |fn (x)| dx + |fn (x)| dx
R R
2
≤ S + C,
2 2
k(D − bn )fn k = kF(D − bn )fn k
2
= k(X − bn )Ffn k
Z 2
2
= |ξ − bn | fbn (ξ) dξ
R
Z 2
2s
≤ (|ξ − bn | + 1) fbn (ξ) dξ
R
≤ (S 2 + C).
95
So we have
2 2
sup k(X − an )fn k + k(D − bn )fn k < ∞. (5.24)
n∈N
2 2
sup k(X − an )gn k + k(D − bn )gn k < ∞.
n∈N
Z
2 2s
|hfn , gb,a i| |a − an | dadb
R2
Z
2 2s
= |F(fn Ta g)(b)| |a − an | dbda
R2
Z
2 2s
= |(fn Ta g)(b)| db |a − an | da
R2
Z
2 2 2s
= |fn (b)| |g(b − a)| |a − b + b − an | dbda
R2
Z
2 2 2s 2s
. |fn (b)| |g(b − a)| |a − b| + |b − an | dadb
R2
Z Z
2 2s 2
= |g(x)| |x| dx |fn (x)| dx
R R
Z Z
2 2 2s
+ |g(x)| dx |fn (x)| |x − an | dx. (5.25)
R R
Z
2 2s
|hfn , gb,a i| |b − bn | dadb
R2
ZD E2
2s
= fn , gb−a,b |b − bn | dadb
b
R2
Z Z 2
2 2s
. |b g (ξ)| |ξ| dξ fbn (ξ) dξ
ZR Z R 2
2 2s
+ |b g (ξ)| dξ fbn (ξ) |ξ − bn | dξ. (5.26)
R R
96
Combine (5.25) and (5.26), by g ∈ S(R) and assumption (a), we have
Z
2
sup |hfn , gb,a i| d2s ((a, b), (an , bn ))dadb
n∈N R2
Z s
2 2 2
= sup |hfn , gb,a i| |a − an | + |b − bn | dadb
n∈N R2
Z
2 2s 2s
. sup |hfn , gb,a i| |a − an | + |b − bn | dadb
n∈N R2
<∞. (5.27)
Z
2
sup |hgn , gb,a i| d2s ((a, b), (an , bn ))dadb < ∞. (5.28)
n∈N R2
In order to verify condition 2, let kb,a := Mb Ta Xg be the continuous Gabor frame generated by the
Z
2
|h(X − an )gn , gb,a i| dadb
B((an ,bn ),r)c
Z
2
= |hgn , (X − an )gb,a i| dadb
B((an ,bn ),r)c
Z
2
= |hgn , (X − an )Mb Ta gi| dadb
B((an ,bn ),r)c
Z
2
= |hgn , (a − an )Mb Ta g + Mb Ta Xgi| dadb
B((an ,bn ),r)c
Z
2
≤2 |hgn , (a − an )Mb Ta gi| dadb
B((an ,bn ),r)c
Z
2
+2 |hgn , Mb Ta Xgi| dadb
B((an ,bn ),r)c
Z
2 2
=2 |hgn , gb,a i| |a − an | dadb
B((an ,bn ),r)c
Z
2
+2 |hgn , kb,a i| dadb. (5.29)
B((an ,bn ),r)c
97
Estimate the first term in (5.29),
Z
2 2
|hgn , gb,a i| |a − an | dadb
B((an ,bn ),r)c
Z
2 2 2
≤ |hgn , gb,a i| |a − an | + |b − bn | dadb
B((an ,bn ),r)c
Z
2
= |hgn , gb,a i| d2 ((a, b), (an , bn ))dadb
B((an ,bn ),r)c
d2s ((a, b), (an , bn ))
Z
2
= |hgn , gb,a i| dadb
B((an ,bn ),r)c d2s−2 ((a, b), (an , bn ))
Z
2−2s 2
≤r |hgn , gb,a i| d2s ((a, b), (an , bn ))dadb
B((an ,bn ),r)c
Z
2
≤r2−2s |hgn , gb,a i| d2s ((a, b), (an , bn ))dadb.
R2
By (5.28), we have
Z
2 2
lim sup |hgn , gb,a i| |a − an | dadb = 0. (5.30)
r→∞ n∈N B((an ,bn ),r)c
Z
2
|hgn , kb,a i| dadb
B((an ,bn ),r)c
d2s ((a, b), (an , bn ))
Z
2
= |hgn , kb,a i| dadb
B((an ,bn ),r)c d2s ((a, b), (an , bn ))
Z
2
≤r−2s |hgn , kb,a i| d2s ((a, b), (an , bn ))dadb
B((an ,bn ),r)c
Z
2
≤r−2s |hgn , kb,a i| d2s ((a, b), (an , bn ))dadb.
R2
Repeat the same argument of (5.25), (5.26) and (5.27), by Xg ∈ S(R) and assumption (b), we
obtain
Z
2
sup |hgn , kb,a i| d2s ((a, b), (an , bn ))dadb < ∞.
n∈N R2
So we have
Z
2
lim sup |hgn , kb,a i| dadb = 0. (5.31)
r→∞ n∈N B((an ,bn ),r)c
98
Combine (5.29), (5.30) and (5.31),
Z
2
lim sup |h(X − an )gn , gb,a i| dadb = 0.
r→∞ n∈N B((an ,bn ),r)c
Similarly,
Z
2
lim sup |h(D − bn )gn , gb,a i| dadb = 0.
r→∞ n∈N B((an ,bn ),r)c
Which imply condition 2 holds as well. Since we collect all the conditions 1-4, we could apply
2 2
sup k(X − an )fn k + k(D − bn )fn k = ∞.
n∈N
99
Acknowledgment
First, I would like to thank my advisor, Dr. Mishko Mitkovski. Without his guidance and
knowledge, I couldn’t have done it. Thank you for leading me to begin my journey of analysis, and
also thank you for providing suggestions and helps when I met a problem in daily life.
Besides, I want to say thank you to all of my professors at Clemson, especially to: Dr.
Shitao Liu, Dr. Martin Schmoll, Dr. Jeong-Rock Yoon, and Dr. Brian Fralix. Thank all of you
for teaching me with great patience. Thank you again to Dr. Liu for sharing your experience of
studying abroad and treating me more like a friend. It was a huge support and encouragement to
me.
sister Xiaoming Wang for taking care of me all the time. Thank you to my parents. You give me
everything you can give. I feel sorry about not being at home for a long time. Because of you, I
100
Bibliography
[1] John J Benedetto, Wojciech Czaja, Przemystaw Gadziński, and Alexander M Powell, The balian-low theorem
and regularity of gabor systems, The Journal of Geometric Analysis 13 (2003), no. 2, 239. ↑7
[2] John J Benedetto, Christopher Heil, and David F Walnut, Differentiation and the balian-low theorem, Journal
of Fourier Analysis and Applications 1 (1994), no. 4, 355–402. ↑6
[3] Stephen M Buckley, Is the maximal function of a lipschitz function continous?, Annales academiæ scientiarum
fennicæ mathematica, 1999, pp. 519–528. ↑28
[4] Ole Christensen et al., An introduction to frames and riesz bases, Springer, 2016. ↑8
[5] Gerald B Folland and Alladi Sitaram, The uncertainty principle: a mathematical survey, Journal of Fourier
analysis and applications 3 (1997), no. 3, 207–238. ↑4
[6] Karlheinz Gröchenig and Eugenia Malinnikova, Phase space localization of riesz bases for lˆ 2 (rˆ d), arXiv
preprint arXiv:1102.3097 (2011). ↑76, 94
[7] Werner Heisenberg, Über den anschaulichen inhalt der quantentheoretischen kinematik und mechanik, Original
scientific papers wissenschaftliche originalarbeiten, 1985, pp. 478–504. ↑ii
[8] Jean-Pierre Kahane, Sur les fonctions moyenne-périodiques bornées, Annales de l’institut fourier, 1957, pp. 293–
314. ↑38
[9] Oleg Kovrijkine, Some results related to the logvinenko-sereda theorem, Proceedings of the American Mathemat-
ical Society 129 (2001), no. 10, 3037–3047. ↑58, 59
[10] HJ Landau et al., Necessary density conditions for sampling and interpolation of certain entire functions, Acta
Mathematica 117 (1967), 37–52. ↑37, 45
[11] Haodong Li, Multiplier operator on framed hilbert spaces (2015). ↑34, 35, 36
[12] Niklas Lindholm, Sampling in weighted lp spaces of entire functions in cn and estimates of the bergman kernel,
Journal of Functional Analysis 182 (2001), no. 2, 390–426. ↑38
[13] Mishko Mitkovski and Aaron Ramirez, Density results for continuous frames, Journal of Fourier Analysis and
Applications 26 (2020), no. 4, 1–26. ↑39
[14] Fedor L’vovich Nazarov, Local estimates for exponential polynomials and their applications to inequalities of the
uncertainty principle type, Algebra i analiz 5 (1993), no. 4, 3–66. ↑60, 61, 62
[15] Alexander Olevskii and Alexander Ulanovskii, Approximation of discrete functions and size of spectrum, St.
Petersburg Mathematical Journal 21 (2010), no. 6, 1015–1025. ↑41, 45, 46
[16] Alexander P Schuster and Kristian Seip, Weak conditions for interpolation in holomorphic spaces, Publicacions
matematiques (2000), 277–293. ↑50
[17] Kristian Seip, Density theorems for sampling and interpolation in the bargmann-fock space, Bulletin of the
American Mathematical Society 26 (1992), no. 2, 322–328. ↑38
[18] Kristian Seip and Robert Wallstén, Density theorems for sampling and interpolation in the bargmann-fock space.
ii, J. reine angew. Math 429 (1992), 107–113. ↑38
[19] Norbert Wiener, I am a mathematician, the later life of a prodigy: An autobiographical account of the mature
years and career of norbert wiener and a continuation of the account of his childhood in ex-prodigy, Vol. 20,
Doubleday, 1956. ↑ii
[20] Robert M Young, An introduction to nonharmonic fourier series, Academic press, 1981. ↑8
101