A survey on frame theory

Frames in general Hilbert spaces

Construction of dual Gabor frames

Construction of tight wavelet frames

Ole Christensen
Technical University of Denmark
Department of Mathematics
Building 303, 2800 Lyngby
Denmark
Email: Ole.Christensen@mat.dtu.dk

1
1 Goal and scope

(H, h·, ·i): Hilbert space.

Want: Expansions
X
f= ck e k
of signals f ∈ H in terms of convenient building
blocks ck .

Desirable properties could be:

• Easy to calculate the coefficients ck
• Only few large coefficients ck for the relevant
signals f .

2
Part I: Expansions in general Hilbert spaces
• Bases;
• Shortcomings of bases;
• Frames;
• Central themes: tight frames, pairs of dual
frames.
Part II: Expansions in L2(R) of the
(i) Gabor type:
ek ∼ e2πimbxg(x − na), m, n ∈ Z,
for some g ∈ L2(R), a, b > 0;
(i) Wavelet type:
ek ∼ 2j/2ψ(2j x − k), j, k ∈ Z,
for some ψ ∈ L2(R).
Central themes:
• Why are frames needed?
• Explicit construction of tight frames and dual
frame pairs.
3
Operators on L2(R):

Translation by a ∈ R:
Ta : L2(R) → L2(R), (Taf )(x) = f (x − a).

Modulation by b ∈ R
Eb : L2(R) → L2(R), (Ebf )(x) = e2πibxf (x).

Dilation by a 6= 0:
1 x
Da : L2(R) → L2(R), (Daf )(x) = p f ( ).
|a| a
With this notation, the Gabor system can be
written

{EmbTnag}m,n∈Z = {e2πimbxg(x − na)}m,n∈Z.

The wavelet system has the form (D := D1/2)

{Dj Tk ψ}j,k∈Z = {2j/2ψ(2j x − k)}j,k∈Z.

4
1.1 The Fourier transform

For f ∈ L1(R), the Fourier transform is de-

fined by
Z ∞
Ff (γ) = fˆ(γ) := f (x)e−2πixγ dx, γ ∈ R.
−∞
The Fourier transform can be extended to a uni-
tary operator on L2(R).

Plancherel’s equation:

hfˆ, ĝi = hf, gi, ∀f, g ∈ L2(R), and ||fˆ|| = ||f ||.

Important commutator relations:

FTa = E−aF, FEa = TaF,

FDa = D1/aF, FD = D −1F.

5
2 Bessel sequences

Let H be a separable Hilbert space.

Definition 2.1 A sequence {fk }∞ k=1 in H is
called a Bessel sequence if there exists a con-
stant B > 0 such that
X∞
|hf, fk i|2 ≤ B ||f ||2, ∀f ∈ H.
k=1

Theorem 2.2 Let {fk }∞ k=1 be a sequence in

H, and B > 0 be given. Then {fk }∞ k=1 is a
Bessel sequence with Bessel bound B if and
only if
X∞
T : {ck }∞
k=1 → ck f k
k=1
2
defines a bounded
√ operator from ` (N) into
H and ||T || ≤ B.
Corollary P 2.3 If {fk }∞k=1 is a Bessel sequence
in H, then ∞ k=1 ck fk converges uncondition-
ally for all {ck }∞ 2
k=1 ∈ ` (N).

6
Pre-frame operator associated to a Bessel se-
quence:
∞
X
T : `2(N) → H, T {ck }∞
k=1 = ck f k
k=1

T is also called the synthesis operator.

The adjoint operator - the analysis operator:

T ∗ : H → `2(N), T ∗f = {hf, fk i}∞
k=1 .

The frame operator:

∞
X
S : H → H, Sf = T T ∗f = hf, fk ifk .
k=1

The series defining S converges unconditionally

for all f ∈ H by Corollary 2.3.

7
 3 Various bases

Definition 3.1 Consider a sequence {ek }∞

k=1
of vectors in H.
(i) The sequence {ek }∞ k=1 is a basis for H if
there for each f ∈ H exist unique scalar
coefficients {ck (f )}∞k=1 such that
∞
X
f= ck (f )ek .
k=1

(ii) A basis {ek }∞k=1 is an unconditional basis

if the series in (i) converges uncondition-
ally for each f ∈ H.
(iii) A basis {ek }∞ k=1 is an orthonormal basis
if {ek }∞
k=1 is an orthonormal system, i.e.,
if
hek , ej i = δk,j .

8
 3.1 Orthonormal bases

Theorem 3.2 For an orthonormal system {ek }∞

k=1 ,
the following are equivalent:
(i) {ek }∞
k=1 is an orthonormal basis.
P∞
(ii) f = k=1hf, ek iek , ∀f ∈ H .
P
(iii) hf, gi = ∞ k=1 hf, ek ihek , gi, ∀f, g ∈ H.
P∞
(iv) k=1 |hf, ek i|2 = ||f ||2, ∀f ∈ H.
(v) span{ek }∞
k=1 = H.

(vi) If hf, ek i = 0, ∀k ∈ N, then f = 0.

Corollary 3.3 If {ek }∞

k=1 is an orthonormal
basis, then each f ∈ H has an uncondition-
ally convergent expansion
∞
X
f= hf, ek iek .
k=1

9
Theorem 3.4 Let {ek }∞ k=1 be an orthonor-
mal basis for H. Then the orthonormal bases
for H are precisely the sets {U ek }∞
k=1 , where
U : H → H is a unitary operator.

Proposition 3.5 Assume that {ek }∞ k=1 is a

sequence of normalized vectors in H and that
∞
X
|hf, ek i|2 = ||f ||2, ∀f ∈ H.
k=1

Then {ek }∞
k=1 is an orthonormal basis for H.

10
3.2 Riesz bases

Definition 3.6 A Riesz basis for H is a fam-

ily of the form {U ek }∞
k=1 , where {e k } ∞
k=1 is an
orthonormal basis for H and U : H → H is
a bounded bijective operator.

Theorem 3.7 If {fk }∞ k=1 = {U e k } ∞

k=1 is a
Riesz basis for H, there exists a unique se-
quence {gk }∞
k=1 in H such that
∞
X
f= hf, gk ifk , ∀f ∈ H.
k=1

The sequence {gk }∞k=1 is also a Riesz basis,

and the series converges unconditionally for
all f ∈ H.

{gk }∞
k=1 is called the dual Riesz basis.

11
From the proof:
{gk }∞
k=1 = {(U −1 ∗
) e k } ∞
k=1 .

Dual of {gk }∞
k=1 :
n
∗ o ∞
−1
(U −1)∗ ek = {U ek }∞ ∞
k=1 = {fk }k=1 .
k=1

∞
So {fk }∞
k=1 and {g k } k=1 are duals of each other:
they are called a pair of dual Riesz bases.

12
Definition 3.8 Two sequences {fk }∞ k=1 and
{gk }∞
k=1 in a Hilbert space are biorthogonal if
hfk , gj i = δk,j .

Corollary 3.9 For a pair of dual Riesz bases

∞
{fk }∞
k=1 and {g k } k=1 the following hold:
∞
(i) {fk }∞
k=1 and {g k } k=1 are biorthogonal.
(ii) For all f ∈ H,
X∞ ∞
X
f= hf, gk ifk = hf, fk igk .
k=1 k=1

Proposition 3.10 If {fk }∞ k=1 = {U e k } ∞

k=1 is
a Riesz basis for H, there exist constants
A, B > 0 such that
∞
X
A ||f ||2 ≤ |hf, fk i|2 ≤ B ||f ||2, ∀f ∈ H.
k=1
The largest possible value for the constant A
1
is ||U −1 ||2
, and the smallest possible value for
B is ||U ||2.

13
Theorem 3.11 For a sequence {fk }∞ k=1 in
H, the following conditions are equivalent:
(i) {fk }∞
k=1 is a Riesz basis for H.

(ii) {fk }∞
k=1 is complete in H, and there exist
constants A, B > 0 such that for every
finite scalar sequence {ck } one has
X X  2 X
2   
A |ck | ≤  ck f k  ≤ B |ck |2.

14
4 The expansion property for a non-basis:

Example 4.1 The family

{ek }k∈Z = {e2πikx}k∈Z
forms an ONB for L2(0, 1).

Consider an open subinterval I ⊂]0, 1[ with

|I| < 1.

Identify L2(I) with the subspace of L2(0, 1) con-

sisting of the functions which are zero on ]0, 1[\I.

For a function f ∈ L2(I):

X
f= hf, ek iek in L2(0, 1). (1)
k∈Z

Since
  2
 
 X 
 f − hf, ek iek  → 0 as n → ∞,

 |k|≤n  2
L (I)

15
we also have
X
f= hf, ek iek in L2(I). (2)
k∈Z
That is, the (restrictions to I of the) functions
{ek }k∈Z also have the expansion property in
L2(I). However, they are not a basis for L2(I)!
To see this, define the function

f (x) if x ∈ I,
f˜(x) =
1 if x ∈
/ I.
Then f˜ ∈ L2(0, 1) and we have the representa-
tion
X
f˜ = hf˜, ek iek in L2(0, 1). (3)
k∈Z
By restricting to I, the expansion (3) is also
valid in L2(I); since f = f˜ on I, this shows
that
X
f= hf˜, ek iek in L2(I). (4)
k∈Z
Thus, (2) and (4) are both expansions of f in
L2(I), and they are non-identical; the argument
16
is that since f 6= f˜ in L2(0, 1), the expansions
(1) and (3) show that
{hf, ek i}k∈Z 6= {hf˜, ek i}k∈Z.

Conclusion: The restriction of the functions

{ek }k∈Z to I is not a basis for L2(I), but the
expansion property is preserved. 

It is natural to consider sequences which

are not bases, but nevertheless have
the expansion property.

17
5 Frames in Hilbert spaces

Definition 5.1 A sequence {fk }∞

k=1 of ele-
ments in H is a frame for H if there exist
constants A, B > 0 such that
∞
X
A ||f ||2 ≤ |hf, fk i|2 ≤ B ||f ||2, ∀f ∈ H.
k=1

The numbers A and B are called frame bounds.

The optimal upper frame bound is the infimum
over all upper frame bounds, and the optimal
lower frame bound is the supremum over all
lower frame bounds.
Lemma 5.2 Suppose that {fk }∞ k=1 is a se-
quence of elements in H and that there exist
constants A, B > 0 such that
∞
X
A ||f ||2 ≤ |hf, fk i|2 ≤ B ||f ||2
k=1

for all f in a dense subset V of H. Then

{fk }∞
k=1 is a frame for H with bounds A, B.

18
Definition 5.3 A sequence {fk }∞ k=1 in H is
a tight frame if there exist a number A > 0
such that
X∞
|hf, fk i|2 = A ||f ||2, ∀f ∈ H.
k=1
The (exact) number A is called the frame
bound.

Recall: the pre-frame operator is

X∞
T : H → `2(N), T {ck }∞ k=1 = ck f k .
k=1

Frame operator associated to general frame {fk }∞

k=1 :
X∞
S : H → H, Sf = T T ∗f = hf, fk ifk .
k=1

{fk }∞
k=1 is a Bessel sequence, so the series defin-
ing S converges unconditionally for all f ∈ H
by Corollary 2.3.
19
 Some important properties of S:
Lemma 5.4 Let {fk }∞ k=1 be a frame with frame
operator S and frame bounds A, B. Then the
following holds:
(i) S is bounded, invertible, self-adjoint, and
positive.
(ii) {S −1fk }∞
k=1 is a frame with frame opera-
tor S −1 and frame bounds B −1, A−1.
(iii) If A, B are the optimal frame bounds for
{fk }∞
k=1 , then the bounds B −1
, A −1
are op-
timal for {S −1fk }∞ k=1 .

{S −1fk }∞
k=1 is called the canonical dual frame
of {fk }∞
k=1 .

20
Frame decomposition:
Theorem 5.5 Let {fk }∞ k=1 be a frame with
frame operator S. Then
X∞
f= hf, S −1fk ifk , ∀f ∈ H,
k=1
and
∞
X
f= hf, fk iS −1fk , ∀f ∈ H. (5)
k=1
Both series converge unconditionally for all
f ∈ H.
Proof. Let f ∈ H. Via Lemma 5.4,
∞
X ∞
X
f = SS −1f = hS −1f, fk ifk = hf, S −1fk ifk .
k=1 k=1
Since {fk }∞
k=1 is a Bessel sequence and
{hf, S −1fk i}∞
k=1 ∈ ` 2
(N),
the series converges unconditionally by Corol-
lary 2.3. The expansion (5) follows from
f = S −1Sf. 

21
Corollary 5.6 If {fk }∞
k=1 is a tight frame with
frame bound A, then the canonical dual frame
is {A−1fk }∞
k=1 , and
∞
1X
f= hf, fk ifk , ∀f ∈ H. (6)
A
k=1

Proof. For a tight frame,

hSf, f i = A||f ||2 = hAf, f i;
since S is self-adjoint, this implies that
S = AI.

By scaling of the vectors {fk }∞
k=1 in a tight
frame, we can always obtain that A = 1; in
that case (6) has exactly the same form as the
representation via an orthonormal basis.

Tight frames can be used without any

additional computational effort compared
to the use of ONB’s.

22
Other advantages of tight frames:
• If {fk }∞
k=1 consists of functions with com-
pact support or fast decay, the same is the
case for the functions in the canonical dual
frame.
• If {fk }∞
k=1 consists of functions with a spe-
cial structure (Gabor structure or wavelet
structure) the same is the case for the func-
tions in the canonical dual frame.

The corresponding statements do not

hold for a general frame {fk }∞
k=1 and its
canonical dual frame {S −1fk }∞
k=1 !

23
To each frame {fk }∞
k=1 one can associate a tight
frame:
Theorem 5.7 Let {fk }∞ k=1 be a frame for H
with frame operator S. Denote the positive
square root of S −1 by S −1/2. Then {S −1/2fk }∞k=1
is a tight frame with frame bound equal to 1,
and
X∞
f= hf, S −1/2fk iS −1/2fk , ∀f ∈ H.
k=1

Problems:
• Not easy to find {S −1/2fk }∞
k=1 ;

• “Nice properties” of {fk }∞

k=1 not necessarily
inherited by {S −1/2fk }∞
k=1 .

24
5.1 Dual frames

Another way to avoid the problem of

inverting the frame operator S:

A frame which is not a Riesz basis is said to be

overcomplete.
Theorem 5.8 Assume that {fk }∞ k=1 is an over-
complete frame. Then there exist frames
{gk }∞
k=1 6
= {S −1
f k } ∞
k=1

for which
∞
X
f= hf, gk ifk , ∀f ∈ H.
k=1

{gk }∞ ∞
k=1 is called a dual frame of {fk }k=1 .

If the canonical dual frame is difficult

to find, maybe there exist other duals
which are easy to find???

25
 ∞
Lemma 5.9 Assume that {fk }∞ k=1 and {g k } k=1
are Bessel sequences in H. Then the follow-
ing are equivalent:
P∞
(i) f = k=1hf, gk ifk , ∀f ∈ H.
P∞
(ii) f = k=1hf, fk igk , ∀f ∈ H.
P∞
(iii) hf, gi = k=1hf, fk ihgk , gi, ∀f, g ∈ H.
In case the conditions are satisfied, {fk }∞
k=1
∞
and {gk }k=1 are dual frames for H.

Characterization of all dual frames:

Theorem 5.10 Let {fk }∞ k=1 be a frame for
H. The dual frames of {fk }∞ k=1 are precisely
the families
 ∞
 X∞ 
{gk }∞
k=1 = S −1fk + hk − hS −1fk , fj ihj ,
 
j=1
k=1
where {hk }∞
k=1 is a Bessel sequence in H.
Does not (immediately) help!

26
5.2 Frames and Riesz bases

Theorem 5.11 A Riesz basis {fk }∞ k=1 for H

is a frame for H, and the Riesz basis bounds
coincide with the frame bounds. The dual
Riesz basis equals the canonical dual frame
{S −1fk }∞
k=1 .

Theorem 5.12 Let {fk }∞ k=1 be a frame for

H. Then the following are equivalent.
(i) {fk }∞
k=1 is a Riesz basis for H.
P∞
(ii) If k=1 ck fk = 0 for some {ck }∞ 2
k=1 ∈ ` (N),
then ck = 0, ∀k ∈ N.

27
5.3 Characterizations of frames

Theorem 5.13 A sequence {fk }∞ k=1 in H is

a frame for H if and only if
∞
X
T : {ck }∞
k=1 → ck f k
k=1

is a well-defined mapping of `2(N) onto H.

Compare to the characterization of Bessel

sequences in terms of T (T well-defined)!

Theorem 5.14 Let {ek }∞ k=1 be an arbitrary

orthonormal basis for H. The frames for
H are precisely the families {U ek }∞
k=1 , where
U : H → H is a bounded and surjective op-
erator.

Compare to the characterization of

• ONB’s (U unitary);
• Riesz bases (U bounded and bijective).

28
5.4 A frame where no subsequence is a basis

Intuitively: A frame consists of a basis +

some extra elements (redundance). Good as in-
tuitive feeling - but wrong in the technical sense:

Example 5.15 Let {ek }∞ k=1 be an orthonor-

mal basis for H. Let
1 1 1 1 1
{fk }∞
k=1 := {e 1 , √ e 2 , √ e 2 , √ e 3 , √ e 3 , √ e3, · · · };
2 2 3 3 3
that is, each vector √1` e`, ` ∈ N, is repeated `
times. Then {fk }∞k=1 is a tight frame for H with
frame bound A = 1. No subfamily is a Riesz
basis. 

More complicated: In each separable Hilbert

space, there exists a frame for which no sub-
family is a basis!

29
Part II: Constructions in L2(R).

Final goal:

Obtain convenient expansions in L2(R),

X
f= hf, gk ifk , ∀f ∈ L2(R).
k∈I

Convenient means:
• fk , gk are given explicitly;
• fk , gk have compact support (no truncation
necessary in numerical calculations);
• fk , gk decay fast in the Fourier domain;
• If fk has a certain structure (e.g., Gabor
structure), gk has the same structure.
In concrete applications, other requirements might
be relevant (differentiability, zero-means, van-
ishing moments, approximation order etc.)
30
 6 Special functions

Typical generators:
B-splines: The B-splines Nn, n ∈ N, are
given inductively by
N1 = χ[0,1], Nn+1(x) = Nn ∗ N1(x)
Z 1
= Nn(x − t) dt.
0
The functions Nn are called B-splines, and n
is the order.

Fundamental properties:
Theorem 6.1 Given n ∈ N, the B-spline
Nn has the following properties:
(i) supp Nn = [0, n] and Nn > 0 on ]0, n[.
R∞
(ii) −∞ Nn(x)dx = 1.
P
(iii) k∈Z Nn(x − k) = 1

31
(iv) For any n ∈ N,
 −2πiγ
n
cn(γ) = 1−e
N .
2πiγ
.

For any function f : R → R, let

f (x)+ = max{0, f (x)}.

Theorem 6.2 For each n = 2, 3, . . . , the B-

spline Nn can be written
Xn  
1 n
Nn(x) = (−1)j (x − j)n−1
+ .
(n − 1)! j=0 j

32
 

 

  

Figure 1: The B-spline N2 .

     


Figure 2: The B-spline N3 .

33
6.1 Symmetric B-splines:

For n ∈ N, let
n
Bn(x) := T − n2 Nn(x) = Nn(x + ).
2

The symmetric B-splines have similar proper-

ties:
Corollary 6.3 For n ∈ N, the B-spline Bn
has the following properties:
P
(i) k∈Z Bn(x − k) = 1.
 πiγ −πiγ n  n
cn(γ) =
(ii) B e −e
= sin(πγ)
.
2πiγ πγ

34
 

       

Figure 3: The B-spline B2 .

         

Figure 4: The B-spline B3 .

35
7 Gabor systems

Gabor systems, have the form

{e2πimbxg(x − na)}m,n∈Z
for some g ∈ L2(R), a, b > 0. Short notation:
{EmbTnag}m,n∈Z = {e2πimbxg(x − na)}

Gabor bases exist:

Example 7.1 Let χ[0,1] denote the character-
istic function for the interval [0, 1]. Then
{e2πikxχ[0,1](x)}k∈Z
is an ONB for L2(0, 1). By translation, for each
n ∈ Z the space L2(n, n + 1) has the ONB
{e2πik(x−n)χ[0,1](x−n)}k∈Z = {e2πikxχ[0,1](x−n)}k∈Z.
We conclude that L2(R) has the ONB
 2πikx
e χ[0,1](x − n) k,n∈Z .


36
Problem: The function χ[0,1] is discontinuous
and has very slow decay in the Fourier domain:
Z 1
sin πγ
b[0,1](γ) =
χ e−2πixγ dx = e−πiγ .
0 πγ

Thus, the function is not suitable for time-frequency

analysis.

Question: Can we obtain more suitable Ga-

bor bases by replacing χ[0,1] by a smoother func-
tion g?

No! A continuous function with com-

pact support can not generate a Gabor
Riesz basis.

37
A related short coming - the Balian–Low The-
orem:
Theorem 7.2 Assume that {EmbTnag}m,n∈Z
is 
a ZRiesz basis. Then
 Z 
|xg(x)|2dx |γ ĝ(γ)|2dγ = ∞.
R R
In words:
A function g generating a Gabor
Riesz basis
can not be well localized
in both time and frequency.

For example: impossible that the estimates

C
|g(x)| ≤ ,
(1 + x2)1/2
C
|ĝ(γ)| ≤
(1 + γ 2)1/2
hold simultaneously.

This motives the construction of

Gabor frames!
38
Example 7.3
• The Gaussian
−x2
g(x) = e
generates a Gabor frame {EmbTnag}m,n∈Z
for all a, b ∈]0, 1[. The Fourier transform
√ −π2x2
ĝ(x) = πe
has exponential decay. Also, the dual gen-
erator S −1g has exponential decay in time
and frequency.
• The function
2
h(x) = S −1/2e−x
generates a tight Gabor frame {EmbTnah}m,n∈Z
for all a, b ∈]0, 1[, and h as well as ĥ decay
exponentially.

2 2
Problem: S −1/2e−x and S −1e−x are not given
explicitly.
39
Solution: Don’t construct a nice frame and
expect the canonical dual to be nice.

Construct simultaneously dual pairs

{EmbTnag},{EmbTnah} such that g and h
have the required properties, and
X
f= hf, EmbTnahiEmbTnag, ∀f ∈ L2(R).
m,n∈Z

40
7.1 Sufficient and necessary conditions

Given a function g ∈ L2(R) and two numbers

a, b > 0, consider the matrix-valued function

M (x) := (g(x − na − m/b))m,n∈Z , x ∈ R.

Theorem 7.4 Let A, B > 0 and the Gabor

system {EmbTnag}m,n∈Z be given. Then the
following holds:
(i) {EmbTnag}m,n∈Z is a Bessel sequence with
bound B if and only if M (x) for a.e. x ∈
2
R defines a bounded√ operator on ` (Z)
with norm at most bB.
(ii) Assuming that {EmbTnag}m,n∈Z is a Bessel
sequence, it is a frame for L2(R) with
lower frame bound A if and only if
bAI ≤ M (x)M (x)∗ a.e. x ∈ R,
where I is the identity operator on `2(Z).
41
The following necessary condition for {EmbTnag}m,n∈Z
to be a frame for L2(R) depends on the inter-
play between the function g and the translation
parameter a, and is expressed in terms of the
function
X
G(x) := |g(x − na)|2, x ∈ R.
n∈Z

Proposition 7.5 Let g ∈ L2(R) and a, b > 0

be given, and assume that {EmbTnag}m,n∈Z is
a frame with bounds A, B. Then
X
bA ≤ |g(x − na)|2 ≤ bB, a.e. x ∈ R.
n∈Z

42
Lemma 7.6 Suppose that f is a bounded mea-
surable function with compact support and
that the function
X
G(x) = |g(x − na)|2, x ∈ R
n∈Z

is bounded. Then
X
|hf, EmbTnagi|2
m,n∈Z
Z ∞ X
1 2
= |f (x)| |g(x − na)|2 dx
b −∞ n∈Z
1 XZ ∞
+ f (x)f (x − k/b)
b −∞
k6=0
X
× g(x − na)g(x − na − k/b) dx.
n∈Z

43
Theorem 7.7 Let g ∈ L2(R), a, b > 0 and
suppose that
 
1 
X X 

B := sup  g(x − na)g(x − na − k/b) < ∞.
b x∈[0,a]  
k∈Z n∈Z

Then {EmbTnag}m,n∈Z is a Bessel sequence

with bound B. If also

A := "
1 X
inf |g(x − na)|2
b x∈[0,a]
n∈Z
 
X X 

−  g(x − na)g(x − na − k/b) > 0,
 
k6=0 n∈Z

then {EmbTnag}m,n∈Z is a frame for L2(R)

with bounds A, B.

44
Corollary 7.8 Let g ∈ L2(R) be bounded and
compactly supported. Then {EmbTnag}m,n∈Z
is a Bessel sequence for any a, b > 0.

Recall:

X
G(x) = |g(x − na)|2, x ∈ R.
n∈Z

Corollary 7.9 Let a, b > 0 be given. Sup-

pose that g ∈ L2(R) has support in an in-
terval of length 1b and that the function G is
bounded above and below. Then {EmbTnag}m,n∈Z
is a frame for L2(R) with bounds A, B. The
frame operator S and its inverse S −1 are
given by
G b
Sf = f, S −1f = f, f ∈ L2(R).
b G

45
Corollary 7.10 Suppose that g ∈ L2(R) is a
continuous function with support on an in-
terval I with length |I| and that g(x) > 0 on
the interior of I. Then {EmbTnag}m,n∈Z is a
1
frame for all (a, b) ∈]0, |I|[×]0, |I| [.
Corollary 7.11 For n ∈ N, the B-splines
Bn and Nn generate Gabor frames for all
(a, b) ∈]0, n[×]0, 1/n[.

Question: Characterization of (a, b) ∈ R2 for

which Bn generates a Gabor frame?

The exact answer is unknown, and is bound to

be complicated:

1) {EmbTnaB2}m,n∈Z can not be a frame for

any b > 0 whenever a ≥ 2.

1) [Gröchenig, Janssen, Kaiblinger, Pfander]:

For b = 2, 3, . . . , {EmbTnaB2}m,n∈Z can not be
a frame for any a > 0.
46
Example 7.12 Consider characteristic functions,
g := χ[0,c[, c > 0.
Surprisingly complicated to find the exact range
of c > 0 and parameters a, b > 0 for which
{EmbTnag}m,n∈Z is a frame!

A complete answer has not been obtained yet.

47
Via a scaling, assume that b = 1.

Janssen showed that

(i) {EmTnag}m,n∈Z is not a frame if c < a or
a > 1.
(ii) {EmTnag}m,n∈Z is a frame if 1 ≥ c ≥ a.
(iii) {EmTnag}m,n∈Z is not a frame if a = 1 and
c > 1.
Assuming now that a < 1, c > 1, we further
have
(iv) {EmTnag}m,n∈Z is a frame if a ∈
/ Q and
c ∈]1, 2[.
(v) {EmTnag}m,n∈Z is not a frame if a = p/q ∈
Q, gcd(p, q) = 1, and 2 − 1q < c < 2.
(vi) {EmTnag}m,n∈Z is not a frame if a > 43 and
c = L − 1 + L(1 − a) with L ∈ N, L ≥ 3.
The graphical illustration of this result is known
as Janssen’s tie. 

48
Theorem 7.13 Let g ∈ L2(R) and a, b > 0
be given. Then the following holds:
(i) If ab > 1, then {EmbTnag}m,n∈Z can not
be a frame for L2(R).
(ii) If {EmbTnag}m,n∈Z is a frame, then
ab = 1 ⇔ {EmbTnag}m,n∈Z is a Riesz basis.

49
Duality principle: [Janssen], [Daubechies,
Landau and Landau], [Ron and Shen].

Concerns the relationship between frame prop-

erties for a function g with respect to the lattice
{(na, mb)}m,nZ
and the so-called dual lattice
{(n/b, m/a)}m,nZ.

Theorem 7.14 Let g ∈ L2(R) and a, b > 0

be given. Then the Gabor system
{EmbTnag}m,n∈Z
is a frame for L2(R) with bounds A, B if and
only if
{Em/aTn/bg}m,n∈Z
is a Riesz sequence with bounds abA, abB.

50
7.2 Time-frequency localization of Gabor expansions

No function g 6= 0 can have compact support si-

multaneously in the time-domain and the frequency-
domain.

But most signals appearing in practice are es-

sentially localized in the time-frequency plane:

The interesting part of the signal

takes place on a finite time-interval,
with frequencies belonging to a
certain finite interval.

How does this affects the Gabor frame expan-

sion?

51
Given a number T > 0, define the operator
QT : L2(R) → L2(R), (QT f )(x) = χ[−T,T ](x)f (x).

Will use ||(I − QT )f || as a measure for the

content of the function f outside the interval
[−T, T ]: that f essentially is localized on the
interval [−T, T ] means that ||(I − QT )f || is
small compared to ||f ||.

Similarly, for Ω > 0 define PΩ by

PΩ : L2(R) → L2(R), Pd ˆ
Ω f (ν) = χ[−Ω,Ω] (ν)f (ν).

That fˆ essentially is localized on [−Ω, Ω] means

that ||(I − PΩ)f || is small compared to ||f ||.

Now assume that f is essentially localized in

both domains, i.e., on [−T, T ] × [−Ω, Ω] for
some T, Ω > 0. Let
B(T, Ω)
= {(m, n) ∈ Z2 : mb ∈ [−Ω, Ω], na ∈ [−T, T ]}.
52
Consider the frame expansion of f in terms of
dual frames {EmbTnag}m,n∈Z and {EmbTnah}m,n∈Z,
X
f= hf, EmbTnahiEmbTnag.
m,n∈Z

Do we obtain a reasonable approximation of f

if we replace the sum over (m, n) ∈ Z2 with a
sum over (m, n) ∈ B(T, Ω)?

Positive answer if we replace B(T, Ω) by a cer-

tain enlargement B(T + Λ, Ω + Γ) :

53
Theorem 7.15 Assume that the Gabor sys-
tems {EmbTnag}m,n∈Z and {EmbTnah}m,n∈Z form
a pair of dual frames for L2(R) with upper
frame bounds B and D respectively, and that
for some constants C > 0, α > 1/2, the de-
cay conditions
|h(x)| ≤ C(1 + x2)−α, |ĥ(ν)| ≤ C(1 + ν 2)−α
hold. Then, for any  > 0 there exist num-
bers Λ, Γ > 0 such that for all T, Ω > 0,
 
 
 X 
f − hf, EmbTnahiEmbTnag 

 {(m,n)∈B(T +Λ,Ω+Γ)} 
√
≤ BD (||I − QT )f || + ||(I − PΩ)f || +  ||f ||)
for all f ∈ L2(R).

54
7.3 Tight Gabor frames

Theorem 7.16 Let g ∈ L2(R) and a, b > 0

be given. Then the following are equivalent:
(i) {EmbTnag}m,n∈Z is a tight frame for L2(R)
with frame bound A = 1.
(ii) For a.e. x ∈ R the following conditions
hold:
P
G(x) = n∈Z |g(x − na)|2 = b;
P
n∈Z g(x − na)g(x − na − k/b) = 0 for
all k 6= 0.
Moreover, when the equivalent conditions
hold, {EmbTnag}m,n∈Z is an orthonormal ba-
sis for L2(R) if and only if kgk = 1.

55
Corollary 7.17 Let a, b > 0 be given. As-
sume that ϕ ∈ L2(R) is a real-valued non-
negative function with support in an interval
of length 1/b, and that
X
ϕ(x + na) = 1, a.e. x ∈ R.
n∈Z

Then the function

p
g(x) := bϕ(x)
generates a tight Gabor frame {EmbTnag}m,n∈Z
with frame bound A = 1.

Example 7.18 For any n ∈ N, the B-spline

ϕ = Nn satisfies the requirements in Corollary
7.17 with a = 1 and any b ∈]0, 1/n]. Thus, for
any b ∈]0, 1/n], the function
p
g(x) = bNn(x)
generates a tight Gabor frame {EmbTng}m,n∈Z
with frame bound A = 1. 

56
7.4 The duals of a Gabor frame

For a Gabor frame {EmbTnag}m,n∈Z with asso-

ciated frame operator S, the frame decomposi-
tion shows that
X
f= hf, S −1EmbTnagiEmbTnag, ∀f ∈ L2(R).
m,n∈Z

We need to be able to calculate the

canonical dual frame {S −1EmbTnag}m,n∈Z
– difficult!

A simplification can be obtained via

Lemma 7.19 Let g ∈ L2(R) and a, b > 0 be
given, and assume that {EmbTnag}m,n∈Z is a
Bessel sequence with frame operator S. Then
the following holds:
(i) SEmbTna = EmbTnaS for all m, n ∈ Z.
(ii) If {EmbTnag}m,n∈Z is a frame, then also
S −1EmbTna = EmbTnaS −1, ∀m, n ∈ Z.

57
Theorem 7.20 Let g ∈ L2(R) and a, b > 0
be given, and assume that {EmbTnag}m,n∈Z is
a Gabor frame. Then the following holds:
(i) The canonical dual frame also has the Ga-
bor structure and is given by {EmbTnaS −1g}m,n∈Z.
(ii) The canonical tight frame associated with
{EmbTnag}m,n∈Z is given by {EmbTnaS −1/2g}m,n∈Z.

Proposition 7.21 Let g ∈ L2(R), and as-

sume that g as well as ĝ decay exponentially.
Let a, b > 0 be given and assume that {EmbTnag}m,n∈Z
is a frame. Then {EmbTnaS −1/2g}m,n∈Z is
a tight frame, for which S −1/2g as well as
F(S −1/2g) decay exponentially.

Theoretically perfect!
2
Can be applied to the Gaussian g(x) = e−x /2
– but the resulting generators are not given ex-
plicitly.
58
7.5 Characterizations of pairs of dual Gabor frames

The dual Gabor pairs with Gabor structure are

characterized in the Wexler–Raz Theorem.
Theorem 7.22 Let g, h ∈ L2(R) and a, b >
0 be given. Then, if the two Gabor systems
{EmbTnag}m,n∈Z and {EmbTnah}m,n∈Z are Bessel
sequences, they are dual frames if and only
if

hh, Em/aTn/bgi = 0 for all (m, n) 6= (0, 0)
hh, gi = ab.

59
Janssen proved the following:

Theorem 7.23 Two Bessel sequences

{EmbTnag}m,n∈Z and {EmbTnah}m,n∈Z form dual
frames if and only if
X
g(x − n/b − ka)h(x − ka) = bδn,0, a.e. x ∈ [0, a].
k∈Z

Consequence of Theorem 7.23:

Lemma 7.24 Let g, h be real-valued, bounded,
and compactly supported functions. Then for
b > 0 sufficiently small the following two
conditions are equivalent:
(i) {EmbTng}m,n∈Z and {EmbTnh}m,n∈Z form
dual frames for L2(R);
(ii)
X
g(x − k)h(x − k) = b, a.e. x ∈ [0, 1].
k∈Z

60
7.6 Explicit construction of dual pairs of Gabor frames

Dual Gabor frames with a = 1:

Theorem 7.25 (C.) Let N ∈ N. Let g ∈
L2(R) be a real-valued bounded function with
supp g ⊆ [0, N ], for which
X
g(x − n) = 1.
n∈Z

Let b ∈]0, 2N1−1 ]. Then the function g and the

function h defined by
N
X −1
h(x) = bg(x) + 2b g(x + n)
n=1

generate dual frames {EmbTng}m,n∈Z and

{EmbTnh}m,n∈Z for L2(R).
Proof. Note that
supp g ⊆ [0, N ], supp h ⊆ [−N + 1, N ],
so Lemma 7.24 applies for b ≤ 1/(2N − 1).

61
Also, g and h have compact support, so {EmbTng}m,n∈Z
and {EmbTnh}m,n∈Z are Bessel sequences.
We check the condition in Lemma 7.24. For
x ∈ [0, 1],
N −1
!2
X
1 = g(x + k)
k=0
= (g(x) + g(x + 1) + · · · + g(x + N − 1)) ×
(g(x) + g(x + 1) + · · · + g(x + N − 1))
= g(x) ×
[g(x) + 2g(x + 1) + 2g(x + 2) + · · · + 2g(x + N − 1)]
+g(x + 1) ×
[g(x + 1) + 2g(x + 2) + 2g(x + 3) + · · · + 2g(x + N − 1)]
+g(x + 2) ×
[g(x + 2) + 2g(x + 3) + 2g(x + 4) + · · · + 2g(x + N − 1)]
+···
+···
+g(x + N − 2) [g(x + N − 2) + 2g(x + N − 1)]
+g(x + N − 1) [g(x + N − 1)]
N −1
1X
= g(x + k)h(x + k).
b
k=0

Thus the conclusion follows. 

62
Corollary 7.26 (C.) For any N ∈ N and
b ∈]0, 2N1−1 ], the functions NN and
N
X −1
hN (x) := bNN (x) + 2b NN (x + n)
n=1

generate a pair of dual frames {EmbTnNN }m,n∈Z

and {EmbTnhN }m,n∈Z.

Important features of the dual pair of frame gen-

erators (NN , hN ) in Corollary 7.26:

• The functions NN and hN are splines;

• NN and hN have compact support, i.e., per-
fect time–localization;
• By choosing N sufficiently large, polynomial
cN and hc
decay of N N of any desired order can
be obtained.

63
Example 7.27 For the B-spline


x x ∈ [0, 1[,
N2(x) = 2 − x x ∈ [1, 2[,


0 x∈ / [0, 2[,
we can use Corollary 7.26 for b ∈]0, 1/3]. For
b = 1/3 we obtain the dual generator

1 2
h(x) = N2(x) + N2(x + 1)
3
 3
2

 3 (x + 1) x ∈ [−1, 0[,
= 31 (2 − x) x ∈ [0, 2[,


0 x∈/ [−1, 2[.

64
 

          

Figure 5: The B-spline N2 and the dual generator h for b = 1/3.

         !

Figure 6: The B-spline N3 and the dual generator h with b = 1/5.

65
Dual Gabor frames with general a > 0:

Theorem 7.28 (C.) Let N ∈ N. Let g ∈

L2(R) be a real-valued bounded function with
supp g ⊆ [0, N ], for which
X
g(x − n) = 1.
n∈Z

Let a, b > 0 and choose J ∈ N such that

J ≥ ab(2N − 1). Define the function h by
N
X −1
h(x) = abg(x) + 2ab g(x + n).
n=1

Then the functions

gk = T Ja k Da/J g, hk = T Ja k Da/J h, k = 0, . . . , J − 1
generate dual multi-Gabor frames
{EmbTnagk }m,n∈Z,k=0,...,J−1, {EmbTnahk }m,n∈Z,k=0,...,J−1
for L2(R).

66
Other choices?

Yes!
Theorem 7.29 (C., Kim, 2007) Let N ∈
N. Let g ∈ L2(R) be a real-valued bounded
function with supp g ⊆ [0, N ], for which
X
g(x − n) = 1.
n∈Z

Let b ∈]0, 2N1−1 ]. Define h ∈ L2(R) by

N
X −1
h(x) = ang(x + n),
n=−N +1

where
a0 = b, an + a−n = 2b, n = 1, 2, · · · , N − 1.
Then g and h generate dual frames {EmbTng}m,n∈Z
and {EmbTnh}m,n∈Z for L2(R).

67
Example 7.30

1) Take


 a0 = b,
an = 0 for n = −N + 1, . . . , −1

 a = 2b, n = 1, . . . N − 1
n

This is Theorem 7.25. This choice gives the

shortest support.

2) Take
a−N +1 = a−N +2 = · · · = aN −1 = b :
if g is symmetric, this leads to a symmetric dual
generator
N
X −1
h(x) = b g(x + n).
n=−N +1

Note: h(x) = 1 on supp g.

68
 1

0.8

0.6

0.4

0.2

0
-2 -1 0 1 2 3 4
x

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
-2 0 2 4 6
x

The generators N2 and N3 and their dual generators via

Figure 7:
2) in Example 7.30.

69
8 Polynomial generators and duals [C., Kim]

The restriction of any polynomial to a suffi-

ciently large interval will generate a Gabor frame
for small modulation parameters:
Proposition 8.1 Let N ∈ N, and consider
any bounded interval I ⊂ R with |I| ≥ N .
Then any (nontrivial) polynomial
N −1
!
X
g(x) = ck xk χI (x)
k=0

generates a Gabor frame {EmbTng}m,n∈Z for

b ∈]0, 1/|I|].

70
The condition on the relationship between the
support size and the degree of the polynomial
is necessary in Proposition 8.1:
Example 8.2 Let
g(x) = (x − 1/2)(x − 3/2)χ[0,2](x)
= (x2 − 2x + 3/4)χ[0,2](x).
Then {EmbTng}m,n∈Z is not a Gabor frame for
any b > 0. In fact, for x ∈ [0, 1],
X
|g(x − k)|2 = |g(x)|2 + |g(x + 1)|2,
k∈Z
so
X
|g(1/2 − k)|2 = 0.
k∈Z

71
In applications of frames: crucial that the frame
generator as well as the dual frame generator
have a convenient form. The polynomials are
very convenient - but no pair of dual generators
of this form exists:
Theorem 8.3 Two compactly supported poly-
nomials
N
!
X 1

g(x) = ak xk χI1 (x)

k=0

and
N2
!
X
h(x) = b k xk χI2 (x)
k=0

of degree ≥ 1 can not generate dual Gabor

frames {EmbTng}m,n∈Z and {EmbTnh}m,n∈Z for
any b > 0, no matter how the intervals I1
and I2 are chosen.

72
Remark: The proof shows that even finite lin-
ear combinations of the type
X X
g̃(x) = ck Tk g, h̃(x) = d k Tk h
with g and h as in Theorem 8.3 can’t gen-
erate dual Gabor frames {EmbTng̃}m,n∈Z and
{EmbTnh̃}m,n∈Z

Conclusion: for polynomial generators

we must search for other simple gen-
erators for the dual. Natural choices:
B-splines!

73
 P −1
Theorem 8.4 Let N ∈ N and g(x) := N k=0 ck x
k

be a polynomial. Given an interval I ⊂ R,

let
g̃(x) := g(x)χI (x).
Let h ∈ L2(R) be such that supp h ⊂ I. Then
the following are equivalent:
(i) There exists β ∈ R such that g̃(x) and
h̃(x) := βh(x) generate dual frames {EmbTng̃}m,n∈Z
and {EmbTnh̃}m,n∈Z for b > 0 sufficiently
small;
(ii)


 P D k
ĥ(n)
 N −1
 k=0 ck = 0, n ∈ Z \ {0};
(−2πi)k

 PN −1 Dk ĥ(0)

 k=0 ck 6= 0.
(−2πi) k

74
 P 0
Theorem 8.5 Let N ∈ N and g(x) := N k=0 ck x
k

be a polynomial of degree N0 < N . Given an

interval I ⊂ R, let
g̃(x) := g(x)χI (x).
Let h ∈ L2(R) be such that supp h is an
interval with supp h ( I. If h satisfy the
Strang-Fix conditions, that is,
ĥ(0) 6= 0;
Dk ĥ(n) = 0, n ∈ N \ {0}, k = 0, 1, · · · , N − 1,
then there always exist α, β ∈ R such that
g̃(x) and h̃(x) := βh(x − α) generate dual
frames {EmbTng̃}m,n∈Z and {EmbTnh̃}m,n∈Z for
b > 0 sufficiently small.

75
 P −1
Corollary 8.6 Let N ∈ N and g(x) := N k=0 ck x
k

be a polynomial. Given an interval I con-

taining [0, N ], let
g̃(x) := g(x)χI (x).
Then the following hold:
(a) If I = [0, N ] and
N
X −1 cN (0)
Dk N
κ := ck 6= 0,
(−2πi)k
k=0

then g̃(x) and h(x) := κb NN (x) generate

dual frames {EmbTng̃}m,n∈Z and {EmbTnh}m,n∈Z
for b > 0 sufficiently small.
(b) If I ) [0, N ], then there exists α, β ∈ R
such that g̃ and
h(x) := βNN (x − α)
generate dual frames {EmbTng̃}m,n∈Z and
{EmbTnh}m,n∈Z for b > 0 sufficiently small.

76
Conclusion: polynomial generators on suffi-
ciently large intervals almost always generate
Gabor frames with B-spline duals for sufficiently
small modulation parameters.

77
Example 8.7 Let g(x) = (x − 1)χI (x). First,
let I = [0, 2]. A direct calculation shows that
c2(0) = 1,
N c2(0) = −2πi.
DN
Thus
1
X c2(0)
Dk N
κ= ck k
= 0.
(−2πi)
k=0
One can prove that
X
g(x − k)N2(x − k) = 0.
k∈Z

Hence N2 is not a dual of g for any b > 0 by

Lemma 7.24.
Now, let I = [−1, 3]. Then βN2(x − α0) is
dual generator of g for b > 0 sufficiently small
for some constant β (which can be explicitly
calculated).

78
Example 8.8 Let
g(x) = x2(5 − x)2χ[0,5](x)
2 3 4


= 25x − 10x + x χ[0,5](x)
4
!
X
=: ck xk χ[0,5](x).
k=0

For the B-spline N5,


c5(0)  20/3, k = 2;
Dk N
k
= 75/4, k = 3;
(−2πi) 
331/6, k = 4.
Thus
4
X c5(0)
Dk N
κ := ck k
= 103/3 6= 0.
(−2πi)
k=0

Hence h(x) := κb N5 is a dual of g(x) for b >

0 sufficiently small by Corollary 8.6 (a). See
Figure 8.

79
 30

20

10

0
-1 0 1 2 3 4 5 6
x

0.0035

0.003

0.0025

0.002

0.0015

0.001

0.0005

0
-1 0 1 2 3 4 5 6
x

Figure 8: The generators g and h for b = 1/5 in Example 8.8.

80
Example 8.9 Let
g(x) = x3(6 − x)2χ[0,6](x)
3 4 5


= 36x − 12x + x χ[0,6](x)
5
!
X
= ck xk χ[0,6](x).
k=0

For N6, we have


c6(0)  63/2, k = 3;
Dk N
k
= 1087/10, k = 4;
(−2πi) 
777/2, k = 5.
Thus
5
X c6(0)
Dk N
κ := ck k
= 2181/10 6= 0.
(−2πi)
k=0

Hence h(x) := κb N6 is a dual of g(x) for b >

0 sufficiently small by Corollary 8.6 (a). See
Figure 9.

81
 250

200

150

100

50

0
0 2 4 6
x

0.0004

0.0003

0.0002

0.0001

0
0 2 4 6
x

Figure 9: The generators g and h for b = 1/6 in Example 8.9.

82
Extensions:
1) Higher dimensions: (C., R. Y. Kim)

For f ∈ L2(Rd) and d×d matrices B, C, and

m, n ∈ Zd, let

EBmf (x) = e2πiBm·xf (x), x ∈ Rd

TCnf (x) = f (Cn − x), x ∈ Rd.

83
Theorem 8.10 Let N ∈ N and assume that
g ∈ L2(Rd) has supp g ⊆ [0, N ]d, and that
X
g(x − n) = 1.
n∈Zd

Assume that the d × d matrix B is invertible

1
and ||B|| ≤ √d(2N −1)
. For i = 1, . . . , d, let Fi
be the set of lattice points {kj }dj=1 ∈ Zd for
which the coordinates kj , j = 1, . . . , d, satisfy


if j = 1, . . . , i − 1, then |kj | ≤ N − 1;
if j = i, then 1 ≤ kj ≤ N − 1;


if j = i + 1, . . . , d, then kj = 0.
Define h ∈ L2(Rd) by
 
d X
X
h(x) := | det B| g(x) + 2 g(x + k) .
i=1 k∈Fi

Then {EBmTng}m,n∈Zd and {EBmTnh}m,n∈Zd

are dual frames for L2(Rd).

84
2) Tight frames (C., R.Y. Kim):
Theorem 8.11 Let N ∈ N. Let g ∈ L2(Rd)
be a non-negative function with supp g ⊆
[0, N ]d, for which
X
g(x − n) > 0, a.e. x ∈ Rd.
n∈Zd
Assume that the d × d matrix B is invertible
and ||B|| ≤ √d1 N . Define h ∈ L2(Rd) by
s
g(x)
h(x) := | det B| P .
n∈Zd g(x − n)
Then the function h generates a Parseval
frame {EBmTnh}m,n∈Zd for L2(Rd).

The generator in Theorem 8.11 has fast decay

in the Fourier domain if it is smooth:
Lemma 8.12 Let k ∈ N and let f ∈ C dk (Rd)
be compactly supported. Then
|fˆ(γ)| ≤ A(1 + |γ|2)−k/2.

85
3) Irregular B-splines (C., Wenchang
Sun):
Theorem 8.13 Let {xn : n ∈ Z} ⊂ R be a
sequence such that
lim xn = ±∞, xn−1 ≤ xn,
n→±∞

and
xn+2N −1 − xn ≤ M, n ∈ Z,
for some constants N ∈ N and M > 0.
Let gn be the N -th order normalized B-spline
with knots (xn, xn+1, · · · , xn+N ) and
N
X −1
hn(x) = bgn(x) + 2b gn−k (x).
k=1

Then {Embgn}m,n∈Z and {Embhn}m,n∈Z are a

pair of dual frames for L2(R) provided that
0 < b ≤ 1/M .

86
 9 Wavelet frames

9.1 Wavelet bases

Given a function ψ ∈ L2(R) and j, k ∈ Z, let

ψj,k (x) := 2j/2ψ(2j x − k), x ∈ R.
In terms of Tk and Df (x) = 21/2f (2x),
ψj,k = Dj Tk ψ, j, k ∈ Z.
If {ψj,k }j,k∈Z is an orthonormal basis for L2(R),
the function ψ is called a wavelet.
Definition 9.1 A multiresolution analysis for
L2(R) consists of a sequence of closed sub-
spaces {Vj }j∈Z of L2(R) and a function φ ∈
V0, such that the following conditions hold:
(i) · · · V−1 ⊂ V0 ⊂ V1 · · · .
(ii) ∪j Vj = L2(R) and ∩j Vj = {0}.
(iii) f ∈ Vj ⇔ [x → f (2x)] ∈ Vj+1.
(iv) f ∈ V0 ⇒ Tk f ∈ V0, ∀k ∈ Z.
(v) {Tk φ}k∈Z is an orthonormal basis for V0.

87
An MRA can be used to construct an ONB for
L2(R):

For j ∈ Z, let Wj denote the orthogonal com-

plement of Vj in Vj+1. Then
M
2
L (R) = Wj .
j∈Z

The spaces Wj satisfy the same dilation rela-

tionship as Vj , i.e.,
ψ ∈ W0 ⇔ [x → ψ(2j x)] ∈ Wj .

In order to obtain an orthonormal basis {ψj,k }j,k∈Z

for L2(R) it is now enough to find ψ ∈ W0 such
that {ψ(· − k)}k∈Z is an orthonormal basis for
W0 .

88
We will now explain how to find a suitable func-
tion ψ. The condition φ ∈ V0 ⊂ V1 implies by
(iii) that
1
√ D−1φ ∈ V0.
2
Since {T−k φ}k∈Z is an orthonormal basis for V0,
there exist coefficients {ck }k∈Z ∈ `2(Z) such
that
1 −1 X
√ D φ= ck T−k φ. (7)
2 k∈Z
We will now rewrite (7) in terms of the Fourier
transform of φ. Let
Ek (x) = e2πikx, x ∈ R.
Now, applying the Fourier transform on both
sides of (7),
1 X
√ Dφ̂ = ck Ek φ̂.
2 k∈Z
P
Defining the 1-periodic function H0 := k∈Z ck Ek ,
this can be written as
φ̂(2γ) = H0(γ)φ̂(γ), a.e. γ ∈ R. (8)
89
The equation (8) is called a scaling equation
or refinement equation. Now, a certain choice
of a 1-periodic function H1, implies that the
function ψ defined via
ψ̂(2γ) = H1(γ)φ̂(γ) (9)
generates a wavelet orthonormal basis {D j Tk ψ}j,k∈Z.
One choice of H1 is to take
1 −2πiγ
H1(γ) = H0(γ + )e .
2

Note that (9) leads to an explicit expression of

the function ψ in terms of the given function φ:

90
Lemma 9.2 Assume that
ψ̂(2γ) = H1(γ)φ̂(γ)
holds for a 1-periodic and bounded
P function
H1 with Fourier expansion H1 = k∈Z ck Ek .
Then
√ X
ψ(x) = 2 ck DT−k φ(x)
k∈Z
X
= 2 ck φ(2x + k), a.e. x ∈ R.
k∈Z

In particular, if H
P1 Nis2 a trigonometric poly-
nomial, H1(x) = k=N1 ck e2πikx, then
N2
√ X
ψ(x) = 2 ck DT−k φ(x)
k=N1
N2
X
= 2 ck φ(2x + k), ∀x ∈ R.
k=N1

91
The Haar basis can be constructed via the mul-
tiresolution analysis defined by φ = χ[0,1[ and
Vj = {f : f is const. on [2−j k, 2−j (k + 1)[, ∀k ∈ Z}.
In terms of the function φ, the Haar function is
1 1
ψ= √ φ1,0 − √ φ1,1.
2 2

The Haar function is a special case of a spline

wavelet.

One can consider higher order splines Nn and

define associated multiresolution analyses, which
leads to wavelets of the type
X
ψ(x) = ck Nn(2x − k).
k∈Z
These wavelets are called Battle–Lemarié wavelets.
Except for the case n = 1, all coefficients ck are
non-zero, which implies that the wavelet ψ has
support equal to R. However, the wavelets have
exponential decay.
92
For φ ∈ L2(R), let
Vj = span{D j Tk φ}k∈Z.
Then the MRA-conditions (iii) and (iv) are sat-
isfied.

When does Vj generate an MRA?

Lemma 9.3 Let φ ∈ L2(R). Then the fol-

lowing holds:
(i) ∩j Vj = {0}.
(ii) Assume that the spaces Vj are nested. If
|φ̂| > 0 on a neighborhood of 0, then ∪j Vj
is dense in L2(R).

Thus, if the spaces Vj are nested and the condi-

tion in (ii) is satisfied, then the MRA-conditions
(i)–(iv) are satisfied. We thus need a condition
ensuring that Vj are nested:

93
Lemma 9.4 Assume that φ ∈ L2(R) and that
{Tk φ}k∈Z is a Bessel sequence. Define the
spaces Vj by
Vj = span{Dj Tk φ}k∈Z.
Then the following holds:
(i) If ψ ∈ L2(R) and there exists a bounded
1-periodic function H1 such that
ψ̂(2γ) = H1(γ)φ̂(γ),
then ψ ∈ V1.
(ii) If there exists a bounded 1-periodic func-
tion H0 such that
φ̂(2γ) = H0(γ)φ̂(γ),
then Vj ⊆ Vj+1 for all j ∈ Z.

94
Via Lemma 9.3 and Lemma 9.4 we obtain the
following:
Theorem 9.5 Let φ ∈ L2(R), and assume
that |φ̂| > 0 on a neighborhood of 0. Assume
further that
φ̂(2γ) = H0(γ)φ̂(γ),
is satisfied for a bounded 1-periodic function
H0. Define the spaces Vj by
Vj = span{Dj Tk φ}k∈Z.
Then the following holds:
(i) If {Tk φ}k∈Z is an orthonormal system,
then φ and the spaces Vj form a multires-
olution analysis.
(i) If {Tk φ}k∈Z is a Bessel sequence, then the
spaces Vj satisfy the conditions (i)–(iv) in
Definition 9.1.

95
Important properties for wavelet bases
{ψj,k }j,k∈Z:
• that ψ has a computationally convenient form,
for example that ψ is a piecewise polynomial
(a spline);
• regularity of ψ;
• symmetry (or anti-symmetry) of ψ, i.e., that
ψ(x) = ψ(−x) or ψ(x) = −ψ(−x), x ∈ R;

• compact support of ψ, or at least fast decay;

• that ψ has vanishing moments, i.e., that
for a certain m ∈ N,
Z ∞
x`ψ(x) dx = 0 for ` = 0, 1, . . . , m.
−∞

96
Short description of the role of these
conditions:
• Vanishing moments: needed if we want smooth
wavelets. If ψ is an m times differentiable
function with bounded derivatives with rea-
sonable decay and {ψj,k }j,k∈Z is an orthonor-
mal system, then ψ has m − 1 vanishing
moments.
• Vanishing moments are essential in the con-
text of compression. Assuming that {ψj,k }j,k∈Z
is an orthonormal basis for L2(R), every
f ∈ L2(R) has the representation
X
f= hf, ψj,k iψj,k . (10)
j,k∈Z

All information about f is stored in the co-

efficients {hf, ψj,k i}j,k∈Z, and (10) tells us
how to reconstruct f based on the coeffi-
cients. In practice one can not store an in-
finite sequence of non-zero numbers, so one
has to select a finite number of the coeffi-
cients to keep. Done by
97
Thresholding: Chooses a certain  > 0 and
keep only the coefficients hf, ψj,k i for which
|hf, ψj,k i| ≥ .
If ψ has a large number of vanishing moments,
then only relatively few coefficients hf, ψj,k i will
be large:

Theorem 9.6 Assume that the function ψ ∈

L2(R) is compactly supported and has N − 1
vanishing moments. Then, for any N times
differentiable function f ∈ L2(R) for which
f (N ) is bounded, there exists a constant C >
0 such that
|hf, ψj,k i| ≤ C 2−jN 2−j/2, ∀j, k ∈ Z.

98
Further relevant properties:

• Compact support (or at least fast decay) of

ψ is essential for the use of computer-based
methods, where a function with unbounded
support always has to be truncated. For the
same reason we often want the support to be
small.
• The condition of ψ being symmetric is help-
ful in image processing, where a non-symmetric
wavelet will generate non-symmetric errors,
which are more disturbing to the human eye
than symmetric errors.

99
One can not combine the classical multiresolu-
tion analysis with the desire of having a sym-
metric wavelet ψ:
Proposition 9.7 Assume that φ ∈ L2(R) is
real-valued and compactly supported, and let
Vj = span{Dj Tk φ}k∈Z, j ∈ Z.
Assume that (φ, {Vj }) constitute a multireso-
lution analysis. Then, if the associated wavelet
ψ is real-valued and compactly supported and
has either a symmetry axis or an antisym-
metry axis, then ψ is necessarily the Haar
wavelet.
Thus, under the above assumptions we are
back at the function we want to avoid!

100
Definition 9.8 Let ψ ∈ L2(R). A frame for
L2(R) of the form {D j Tk ψ}j,k∈Z is called a
dyadic wavelet frame.

The associated frame operator:

X
2 2
S : L (R) → L (R), Sf = hf, D j Tk ψiDj Tk ψ.
j,k∈Z

The frame decomposition:

X
f= hf, S −1Dj Tk ψiDj Tk ψ, f ∈ L2(R).
j,k∈Z

Inconvenient - one needs to calculate

hf, S −1Dj Tk ψi, ∀j, k ∈ Z.

101
Improvement: can show that
S −1Dj Tk ψ = Dj S −1Tk ψ.

Unfortunately, in general
Dj S −1Tk ψ 6= Dj Tk S −1ψ.

We can not expect the canonical dual

frame of a wavelet frame to have wavelet
structure.

Bownik and Weber: example of a wavelet sys-

tem {ψj,k }j,k∈Z for which
• The canonical dual does not have the wavelet
structure;
• There exist infinitely many functions ψ̃ for
which {ψ̃j,k }j,k∈Z is a dual frame.

102
Example 9.9 Let {ψj,k }j,k∈Z be a wavelet ONB
for L2(R). Given  ∈]0, 1[, let
θ = ψ + Dψ.
Then {θj,k }j,k∈Z is a Riesz basis and the canon-
ical dual frame is given by

S −1θj,2k+1 = ψj,2k+1 for all j, k ∈ Z.

Also, for any k 6= 0,
S −1θj,2k = ψj,2k − ψj−1,k + · · · + 0, j ∈ Z, k 6= 0.
For k = 0,
∞
X
S −1θj,0 (−)nψj−n,0, j ∈ Z.
n=0

In particular, the canonical dual frame of {θj,k }j,k∈Z

does not have the wavelet structure.

For a Riesz basis, the dual is unique: so no dual

with wavelet structure exists!
103
Other properties which are not inher-
ited by the canonical dual frame:

If ψ has compact support, then θ also has com-

pact support, and all the functions {θj,k }j,k∈Z
have compact support.

For the canonical dual frame {S −1θj,k }j,k∈Z, the

functions have compact support when k 6= 0.
However, the functions S −1θj,0 do not have com-
pact support. 

104
Fom general frame theory:

Two ways to awoid inconvenient frame expan-

sions:
• Look for tight frames;
• Look for convenient dual frame pairs.

Some more aspects are relevant in the wavelet

case:

105
The popular wavelet bases are based on mul-
tiresolution analysis, which leads to a very con-
venient algorithmic structure. This implies a
special form for the function ψ generating the
wavelet basis:
X
ψ= ck DTk φ
k∈Z

for a certain function φ satisfying a scaling equa-

tion,
φ̂(2γ) = H0(γ)φ̂(γ)
for some 1-periodic function H0. The algorith-
mic structure offered by a multiresolution anal-
ysis is a great advantage compared to the use
of general wavelet orthonormal bases. Thus:

While constructing wavelet frames, it

is desirable to maintain the important
aspects of the multiresolution analysis.

106
Therefore, it is natural to require the generator
ψ for a tight frame {D j Tk ψ}j,k∈Z to have the
form
X
ψ= ck DTk φ
k∈Z

for some function φ; and, if we want to con-

struct two dual wavelet frames {D j Tk ψ}j,k∈Z
and {D j Tk ψ̃}j,k∈Z, it is natural to require that
both ψ and ψ̃ have that form.

We want the coefficients in these formulas to be

finite sequences.

The B-splines Bm are obvious candidates for the

function φ. However, we can not obtain all of
these properties simultaneously:
Theorem 9.10 Consider Bm for m > 1. Then
there does not exist pairs of dual wavelet frames
{Dj Tk ψ}j,k∈Z and {Dj Tk ψ̃}j,k∈Z for which ψ
and ψ̃ are finite linear combinations of func-
tions DTk Bm, j, k ∈ Z.
107
Thus, neither the approach of looking at tight
frames, nor the idea of considering wavelet frame
pairs work if we want the generator (respec-
tively, generators) to have the form
X
ψ= ck DTk Bm
k∈Z

for a finite sequence {ck }.

Solution: we will consider systems of

the wavelet-type, but generated by more
than one function.

108
Definition 9.11 Consider two sequences of
functions
ψ1, . . . , ψn ∈ L2(R) and ψ̃1, . . . , ψ̃n ∈ L2(R).
We say that {D j Tk ψ`}j,k∈Z,`=1,...,n and
{Dj Tk ψ̃`}j,k∈Z,`=1,...,n are a pair of dual mul-
tiwavelet frames if both are Bessel sequences
and
Xn X
f= hf, D j Tk ψ`iDj Tk ψ̃`, ∀f ∈ L2(R).
`=1 j,k∈Z

A pair of dual multiwavelet frames is also

called sibling frames or bi-frames. The frame
{Dj Tk ψ`}j,k∈Z,`=1,...,n itself is called a multi-
wavelet frame.

109
Characterization of all dual wavelet frame pairs:

Theorem 9.12 Assume that {D j Tk ψ`}j,k∈Z,`=1,...,n

and {Dj Tk ψe`}j,k∈Z,`=1,...,n are Bessel sequences.
Then {Dj Tk ψ`}j,k∈Z,`=1,...,n and {Dj Tk ψe`}j,k∈Z,`=1,...,n
are a pair of dual wavelet frames if and only
if the two equations

Xn X



 b`(2j γ)ψb
ψ e`(2j γ) = 1,


`=1 j∈Z
Xn X ∞



 b`(2j γ)ψb
ψ e`(2j (γ + q)) = 0, q ∈ 2N + 1


`=1 j=0

hold for a.e. γ ∈ R.

110
Characterization of tight wavelet frames gener-
ated by one generator ψ:

Theorem 9.13 A function ψ ∈ L2(R) gen-

erates a tight wavelet frame {ψj,k }j,k∈Z with
frame bound A if and only if the equations
 X

 | ψ̂(2 j
γ)| 2
= A,


 j∈Z
∞
X

 j

 ψ̂(2 γ)ψ̂ (2j (γ + q)) = 0, q ∈ 2N + 1

j=0

hold for a.e. γ ∈ R.

111
10 The unitary extension principle

We will state the unitary extension principle of

Ron and Shen, which enables us to construct
tight frames for L2(R) of the form
{Dj Tk ψ`}j,k∈Z,`=1,...,n.

Terminology:

The interval ] − 21 , 12 [ is identified with the torus

T

The class of 1-periodic functions on R whose

restriction to ] − 12 , 12 [ belongs to Lp(− 21 , 12 ) is
denoted by Lp(T).

The functions ψ1, . . . , ψn will be constructed on

the basis of a function satisfying a refinement
equation. It is convenient to denote the refin-
able function by ψ0 instead of φ.

112
Standing assumptions and conventions:
Let ψ0 ∈ L2(R) and assume that
(i) There exists a function H0 ∈ L∞(T) such
that
ψb0(2γ) = H0(γ)ψb0(γ).
(ii) limγ→0 ψb0(γ) = 1.

Further, let H1, . . . , Hn ∈ L∞(T), and define

ψ1, . . . , ψn ∈ L2(R) by
ψb`(2γ) = H`(γ)ψb0(γ), ` = 1, . . . , n.

Finally, let H denote the (n + 1) × 2 matrix-

valued function defined by
 
H0(γ) T1/2H0(γ)
 H1(γ) T1/2H1(γ) 
 

H(γ) =  · ·  , γ ∈ R.

 · · 
Hn(γ) T1/2Hn(γ)

113
 We want to find conditions on the functions
H1, . . . , Hn such that ψ1, . . . , ψn generate a mul-
tiwavelet frame for L2(R).

Explicit expression for ψ`: expanding H` in a

Fourier series,
X
H`(γ) = ck,`e2πikγ ,
k∈Z

we have

√ X
ψ`(x) = 2 ck,`DT−k ψ0(x)
k∈Z
X
= 2 ck,`ψ0(2x + k).
k∈Z

Note: If H` are trigonometric polynomials, the

sums are finite. Therefore the functions ψ` have
compact support if ψ0 has compact support.

114
The general setup presented here preserves the
algorithmic structure of a multiresolution anal-
ysis: by Theorem 9.5, the spaces
Vj := span{D j Tk ψ0}k∈Z, j ∈ Z,
satisfy the conditions for a multiresolution anal-
ysis in Definition 9.1, except (v). Also,
ψ1 , . . . , ψ n ∈ V 1 .

The unitary extension principle:

Theorem 10.1 Let {ψ`, H`}n`=0 be as in the

general setup on page 113, and assume that
H(γ)∗H(γ) = I for a.e. γ ∈ T. Then the
multiwavelet system {D j Tk ψ`}j,k∈Z,`=1,...,n con-
stitutes a tight frame for L2(R) with frame
bound equal to 1.

115
 The matrix H(γ)∗H(γ) has four entries, but
it is enough to verify two sets of equations:

Corollary 10.2 Let {ψ`, H`}n`=0 be as in the

general setup on page 113, and assume that
 n
 X

 |H (γ)| 2
= 1,

 `
`=0
n
X



 H`(γ)T1/2H`(γ) = 0,

`=0

for a.e. γ ∈ T. Then {D j Tk ψ`}j,k∈Z,`=1,...,n

constitutes a tight frame for L2(R) with frame
bound equal to 1.

116
Example 10.3 For any m = 1, 2, . . . , we con-
sider the B-spline
ψ0 := B2m
of order 2m. Then
 2m
sin(πγ)
ψb0(γ) = .
πγ
It is clear that limγ→0 ψb0(γ) = 1, and by direct
calculation,
 2m
sin(2πγ)
ψb0(2γ) =
2πγ
 2m
2 sin(πγ) cos(πγ)
=
2πγ
= cos2m(πγ)ψb0(γ).
Thus ψ0 satisfies a refinement equation with
two-scale symbol
H0(γ) = cos2m(πγ).

117
Now, consider the binomial coefficient
 
2m (2m)!
:= ,
` (2m − `)!`!
and define the functions H1, . . . , H2m ∈ L∞(T)
by
s 
2m
H`(γ) = sin`(πγ) cos2m−`(πγ).
`
Using that cos(π(γ − 1/2)) = sin(πγ) and
sin(π(γ − 1/2)) = − cos(πγ),
s 
2m
T1/2H`(γ) = (−1)` cos`(πγ) sin2m−`(πγ).
`

118
Thus, the matrix H is given by
 
H0(γ) T1/2H0(γ)
 H1(γ) T1/2H1(γ) 
 

H(γ) =  · · =

 · · 
H2m(γ) T1/2H2m(γ)
 
s cos2m (πγ) s sin2m (πγ)
   
 2m 2m 
 sin(πγ) cos2m−1 (πγ) − cos(πγ) sin2m−1 (πγ) 
 1 1 
 s  s  
 
 2m 2 2m 
 sin (πγ) cos 2m−2 (πγ) cos2 (πγ) sin2m−2 (πγ) 
 2 2 .
 
 · · 
 
 · · 
 s s 
   
 2m 2m 
sin2m (πγ) cos2m (πγ)
2m 2m

119
We now verify the conditions in Corollary 10.2.
Using the binomial formula
X2m  
2m
(x + y)2m = x`y 2m−`,
`
`=0

2m
X 2m 
X 
2m
|H`(γ)|2 = sin2`(πγ) cos2(2m−`)(πγ)
`
`=0 `=0
2 2

2m
= sin (πγ) + cos (πγ)
= 1, γ ∈ T.
Using the binomial formula with x = −1, y =
1,
2m
X
H`(γ)T1/2H`(γ)
`=0
2m
X  
2m
= sin2m(πγ) cos2m(πγ) (−1)`
`
`=0
= sin2m(πγ) cos2m(πγ)(1 − 1)2m
= 0.

120
Now Corollary 10.2 implies that the 2m func-
tions ψ1, . . . , ψ2m defined by

ψb`(γ) = H`(γ/2)ψ0(γ/2)
s  2m+`
2m sin (πγ/2) cos2m−`(πγ/2)
=
` (πγ/2)2m

generate a tight frame {D j Tk ψ`}j,k∈Z,`=1,...,2m

for L2(R). 

121
A small modification:
Example 10.4 We continue Example 10.3, but
now we define
s 
2m
H`(γ) = i` sin`(πγ) cos2m−`(πγ).
`

H` only differs from the choice in Example 10.3

by a constant of absolute value 1, so the func-
tions ψ1, . . . , ψ2m given by
ψb`(2γ) = H`(γ)ψ c0(γ), ` = 1, . . . , 2m,
also generate a tight multiwavelet frame.

Rewrite H`(γ) using Euler’s formula:

H`(γ)
s  `  2m−`
πiγ −πiγ πiγ −πiγ
2m e −e e +e
= i`
` 2i 2
s 
−2m 2m πiγ


−πiγ ` πiγ


−πiγ 2m−`
= 2 e −e e +e .
`

122
Via the binomial formula we see that H`(γ) is
a finite linear combination of terms
e−2πimγ , e−2πi(m−1)γ , . . . , e2πi(m−1)γ , e2πimγ .
All coefficients in the linear combination are
real. Writing
m
X
H`(γ) = ck,`e2πikγ ,
k=−m

Lemma 9.2 shows that

m
√ X
ψ` = 2 ck,`DT−k ψ0.
k=−m

That is, ψ` is a real-valued spline. Since DTmψ0

has support in [0, m] and DT−mψ0 has sup-
port in [−m, 0], the spline ψ` has support in
[−m, m]. The splines ψ` inherit other proper-
ties from ψ0: they have degree 2m − 1, belong
to C 2m−2(R), and have knots at Z/2. 

123
Explicitly, in the case m = 1:
Example 10.5 In the case m = 1, the con-
struction in Example 10.4 leads to two genera-
tors ψ1 and ψ2. Via the expression for H1,
s 
1 2m
H1(γ) = (eπiγ − e−πiγ )(eπiγ + e−πiγ )
4 1
1
= √ (e2πiγ − e−2πiγ ).
2 2
By Lemma 9.2 we conclude that
1
ψ1(x) = √ (B2(2x + 1) − B2(2x − 1)).
2
Similarly,
1
ψ2(x) = (B2(2x + 1) − 2B2(2x) + B2(2x − 1)) .
2


124
 $

" # " $ % $ #

" $

Figure 10: The function ψ1 .

& ' & ( ) ( '

& (

Figure 11: The function ψ2 .

125
Example 10.6 Can also construct spline frames
with support on [0, 2m]. Letting ψ0 := N2m,
one can prove that
c c0(γ)
ψ0(2γ) = H0(γ)ψ
with
 −2πiγ
2m
1+e
H0(γ) = = e−2πimγ cos2m(πγ).
2
Since H0 appears from the corresponding func-
tion for B2m simply by multiplication with e−2πimγ ,
the functions
s 
2m
H`(γ) = e−2πimγ sin`(πγ) cos2m−`(πγ)
`
satisfy the conditions in the unitary extension
principle. We prefer to consider
s 
2m
H`(γ) = i`e−2πimγ sin`(πγ) cos2m−`(πγ);
`
with this choice, the functions ψ1, . . . , ψ2m

126
defined by
ψb`(γ)
= H`(γ/2)ψ0(γ/2)
s  2m+`
` −2πimγ 2m sin (πγ/2) cos2m−`(πγ/2)
= ie
` (πγ/2)2m

generate a tight frame for L2(R). The spline

functions ψ1, . . . , ψ2m now have support on [0, 2m].


127
Shortcomings for the unitary extension
principle:
• The computational effort increases with the
order of the B-spline B2m: For higher orders,
we need more generators, and more non-zero
coefficients appear in ψ`.
• There is a limitation on the possible number
of vanishing moments ψ` can have.

If {ψ`}n`=1 is constructed via the unitary exten-

sion principle,then
ψb`(γ) = H`(γ/2)ψ
c0(γ/2)

and
c0(0) = 1.
ψ
Thus, the number of vanishing moments for the
function ψ` is equal to the order of zero for H`
at γ = 0.

128
Example 10.7 Consider B2m; it satisfies a re-
finement equation with two-scale symbol
H0(γ) = cos2m(πγ).
If we want to construct a frame via the unitary
extension principle, then
n
X
1= |H`(γ)|2,
`=0
i.e.,
n
X
|H`(γ)|2 = 1 − cos4m(πγ). (11)
`=1

The order of the zero at γ = 0 for the function

1 − cos4m(πγ) is 2, so also on the left-hand side
of (11) we can only factor γ 2 out; this implies
that at least one of the functions |H`|2 can at
most have a zero at γ = 0 of order 2.

Therefore at least one of the functions ψ` can

at most have one vanishing moment. 

129
11 The oblique extension principle

An important reformulation of Theorem 10.1

was simultaneously obtained by Daubechies, Han,
Ron and Shen in [4] and Chui, He and Stöckler
in [3]. It gives a more flexible recipe for con-
struction of frames than Theorem 10.1, and is
called the oblique extension principle:
Theorem 11.1 Let {ψ`, H`}n`=0 be as in the
general setup. Assume that there exists a
strictly positive function θ ∈ L∞(T) for which
lim θ(γ) = 1
γ→0

and such that for a.e. γ ∈ T,

Xn
H0(γ)H0(γ + ν)θ(2γ) + H`(γ)H`(γ + ν)
`=1

θ(γ) if ν = 0,
=
0 if ν = 21 .
Then the functions {D j Tk ψ`}j,k∈Z,`=1,...,n con-
stitute a tight frame for L2(R) with frame
bound equal to 1.
130
Proof. Assume that the conditions in The-
orem 11.1 are satisfied, and define the function
f0 ∈ L2(R) by
ψ
c p
f0(γ) = θ(γ)ψ
ψ c0(γ).
f0, . . . , H
Define the 1-periodic functions H fn by
s s
f θ(2γ) f 1
H0(γ) = H0(γ), H`(γ) = H`(γ).
θ(γ) θ(γ)

We now prove that ψf0, Hf0, . . . , H

fn satisfy the
conditions in the general setup. First,
c p p
f0(2γ) = θ(2γ)ψ
ψ c0(2γ) = θ(2γ)H0(γ)ψ c0(γ)
s
θ(2γ) c
f0(γ)
= H0(γ)ψ
θ(γ)
f0(γ)ψ c
f0(γ).
= H
Also,
p 
c
f0(γ) = lim c0(γ) = 1.
lim ψ θ(γ)ψ
γ→0 γ→0

131
Via the definition,
Xn Xn 2
f`(γ)| =
2 θ(2γ) 2 |H ` (γ)|
|H |H0(γ)| +
θ(γ) θ(γ)
`=0 `=1
= 1, a.e. γ ∈ T.
f0, . . . , H
Thus, H fn ∈ L∞(T). Similarly,
n
X
f`(γ + 1 ) = 0, a.e. γ ∈ T.
f`(γ)H
H
2
`=0

f1, . . . , ψ
Defining the functions ψ fn by
b c
f0(γ), ` = 1, . . . , n,
ψe`(2γ) = H̃`(γ)ψ
it follows from Theorem 10.1 that the functions
{Dj Tk ψe`}j,k∈Z,`=1,...,n constitute a tight frame
for L2(R) with frame bound equal to 1. Now,
ψb`(2γ) = H`(γ)ψc0(γ)
p
= θ(γ)H f`(γ) p 1 ψ c
f0(γ)
θ(γ)
b
= ψe`(2γ),
which shows that ψ` = ψe`. 
132
The UEP and the OEP are equivalent:

Taking θ = 1 in the OEP we obtain the UEP.

The proof of Theorem 11.1 shows that all OEP

constructions can also be obtained via the UEP.

However, in practice the OEP is more flexible:

Suppose that
• ψ0 is a compactly supported function such
that for some function H0 ∈ L∞(R),
ψb0(2γ) = H0(γ)ψb0(γ);
• the functions θ, H` are trigonometric poly-
nomials satisfying the conditions in the OEP
Then the generators ψ` for the frame {D j Tk ψ`}
have the form
X
ψ`(x) = ck,`ψ0(2x − k)
(finite sum) and thus compact support.
133
The same frame can be constructed via the UEP:
f0 by
if we define ψ
c p
f
ψ0(γ) = θ(γ)ψ c0(γ),

then the functions ψe` constructed in the proof

of the UEP will satisfy the conditions in the
unitary extension principle, and ψ` = ψe`.

However, in general f ψ0 is not compactly sup-

ported, and H` is not a trigonometric polyno-
mial, so the fact that the resulting frame {D j Tk ψe`}
is generated by compactly supported functions
is somewhat miraculous and could certainly not
be predicted in advance.

There are constructions which appear

naturally via the OEP, but one would
not even think about constructing them
via the UEP.

134
In practice: desirable to have as few
generators as possible.

135
Corollary 11.2 Let ψ0 and H0 be as in the
general setup. Let θ ∈ L∞(T) be a strictly
positive function for which limγ→0 θ(γ) = 1,
chosen such that the function
 
1
η(γ) := θ(γ) − θ(2γ) |H0(γ)|2 + |H0(γ + )|2
2
is positive as well. Fix an integer n ≥ 2
and let {G`}n`=2 be 1-periodic trigonometric
polynomials for which
X n X n
2 1
|G`(γ)| = 1, and G`(γ)G`(γ + ) = 0.
2
`=2 `=2
Let ρ, σ be 1-periodic functions such that
|ρ(γ)|2 = θ(γ), |σ(γ)|2 = η(γ), (12)
and define the 1-periodic functions {H`}n`=1
by
1
H1(γ) = e2πiγ ρ(2γ)H0(γ + ), H`(γ) = G`(γ)σ(γ).
2
Then the functions {ψ`}n`=1 in the general
setup generate a tight frame {D j Tk ψ`}j,k∈Z,`=1,...,n.
136
If the condition on η is satisfied, Corollary 11.2
makes it easy to obtain frames with for example
three generators.

For example, the conditions on G` are satisfied

with
1 1
G2(γ) = √ , G3(γ) = √ e2πiγ .
2 2
Thus, in order to apply Corollary 11.2, the re-
maining work consists in finding ρ, σ such that
(12) is satisfied.

If the functions θ and η are trigonometric poly-

nomials, then the functions ρ and σ can be
chosen to be trigonometric polynomials as well
(spectral factorization).

The assumption on η even implies that we can

construct a frame generated by two functions:

137
Corollary 11.3 Let ψ0 and H0 be as in the
general setup. Let θ ∈ L∞(T) be a strictly
positive function for which limγ→0 θ(γ) = 1,
chosen such that the function
 
1
η(γ) := θ(γ) − θ(2γ) |H0(γ)|2 + |H0(γ + )|2
2
is positive as well. Define the functions ρ, σ
by
|ρ(γ)|2 = θ(γ), |σ(γ)|2 = η(γ),
and let
2πiγ 1
H1(γ) = e ρ(2γ)H0(γ + ),
2
H2(γ) = H0(γ)σ(2γ).
Then the functions {ψ`}2`=1 generate a tight
frame {Dj Tk ψ`}j,k∈Z,`=1,2 for L2(R).

If θ and H0 are trigonometric polynomials, then

η is also a trigonometric polynomial.

138
The assumption that θ and η are positive im-
plies that we can choose ρ, σ in (12) to be trigono-
metric polynomials.

In this case, the generators ψ` are finite linear

combinations of functions DTk ψ0.

139
 The conditions for reduction of the num-
ber of generators are satisfied for B-
splines:

Theorem 11.4 Let B2m denote the B-spline

of order 2m with two-scale symbol H0(γ) =
cos2m(πγ). Then, for each positive integer
M ≤ 2m, there exists a trigonometric poly-
nomial θ of the form
M
X −1
θ(γ) = 1 + cj sin2j (πγ),
j=1

for which the following hold:

(i) cj ≥ 0 for all j = 1, . . . , M − 1, i.e.,
θ(γ) > 0 for all γ ∈ R;
(ii) The function η is positive;
(iii) The generators in the tight wavelet frames
constructed via the oblique extension prin-
ciple and its corollaries have M vanishing
moments.
140
 The coefficients cj , j = 1, . . . , M − 1 can be
determined via the requirement that
 4m
X ∞
1 + (2j − 1)! j 
y
j=1
(2j)! (2j + 1)
M
X −1
= 1+ cj y j + O(|y|M ) as y → 0.
j=1

141
Example 11.5 Consider the B-spline B2m and
M = 2. Note that
 4m
X ∞
1 + (2j − 1)! j 
y
j=1
(2j)! (2j + 1)
 4m
1 1
= 1 + y + y2 + · · ·
6 20
2m
= 1+ y + O(|y|2).
3
This proves that for M = 2, the condition is
satisfied with c1 = 2m/3. Thus, the desired
trigonometric polynomial is
2m 2
θ(γ) = 1 + sin (πγ)
3
2m 1 − cos 2πγ
= 1+
3 2
3+m m
= − cos 2πγ.
3 3


142
Conclusion:
• We can construct multiwavelet frames with
two generators based on any B-spline B2m.
• If we choose ρ, σ in (12) to be trigonometric
polynomials, then H1, H2 in Theorem 11.3
are trigonometric polynomials, and the as-
sociated frame generators ψ1, ψ2 are finite
linear combinations of functions
B2m(2x − k), k ∈ Z.

• Choosing m large enough, we can obtain

generators belonging to any prescribed smooth-
ness class C N (R).

143
Frame constructions via the UEP and the OEP:
Example 11.6 Let ψ0 = N2. Then
c2(2γ) = H0(γ)N
N c2(γ),

where
(1 + e−2πiγ )2
H0(γ) = = e−2πiγ cos2(πγ).
4
We first revisit Example 10.6 and then give con-
structions via the oblique extension principle
and its corollaries.
(i) Defining H1 and H2 by
−2πiγ
√
H1(γ) = ie 2 sin(πγ) cos(πγ)
√
2
= (1 − e−4πiγ ),
4
−2πiγ 2 (1 − e−2πiγ )2
H2(γ) = −e sin (πγ) = ,
4
it follows from Example 10.6 that the associated
(i)
functions ψ1 := ψ1 and ψ2 generate a tight

144
frame for L2(R). They are given by
(i) 1
ψ1 (x) = √ (N2(2x) − N2(2x − 2)),
2
1
ψ2(x) = (N2(2x) − 2N2(2x − 1) + N2(2x − 2)) .
2
See Figures 9-10.
(ii) An alternative construction can be ob-
tained via the oblique extension principle. Let
4 − cos(2πγ)
θ(γ) := .
3
In this example we keep the choice of H2 in (i).
Thus, if we want to use the oblique extension
principle, we have to choose H1 such that
|H1(γ)|2 = θ(γ) − |H0(γ)|2θ(2γ) − |H2(γ)|2,
and
1
H1(γ)H1(γ + )
2
1 1
= −H0(γ)H0(γ + )θ(2γ) − H2(γ)H2(γ + ).
2 2

145
One can take
(ii) 1
ψ1 (γ) = √ (N2(2γ − 4) + 2N2(2γ − 3))
2 6
6
− N2(2γ − 2)
2
1
+ √ (2N2(2γ − 1) + N2(2γ)) ,
2 6
(ii)
which has support on [0, 3]. The function ψ1
is shown in Figure 11. 

* + , + - .

* +

(i)
Figure 12: The function ψ1 .

146
 0

/ 0 1 0 2 3

/ 0

Figure 13: The function ψ2 .

4 5 6 5 7 8

4 5

(ii)
Figure 14: The function ψ1 .

147
12 Approximation orders

More reasons for constructing frames

via the oblique extension principle:

Assume that {H`, ψ`}n`=0 is as in the general

setup, and that {D j Tk ψ`}j,k∈Z,`=1,...,n is a tight
frame constructed via the oblique extension prin-
ciple and ψ0 = B2m for some m ∈ N.Let
Vj = span{D j Tk ψ0}j,k∈Z.

For s > 0, consider the Sobolev space

 Z ∞ 
Hs(R) = f | |fˆ(γ)|2(1 + |γ|2)sdγ < ∞ .
−∞

Hs(R) is a Banach space with respect to the

natural norm,
Z ∞ 1/2
||f ||Hs = |fˆ(γ)|2(1 + |γ|2)sdγ .
−∞

148
We say that ψ0 provides approximation order
s if for all f in the Sobolev space H s(R),
dist(f, Vj ) = O(2−js),
i.e., if there exists a constant C > 0 such that
dist(f, Vj ) ≤ C2−js, ∀j ∈ Z.
For the tight frame {D j Tk ψ`}j,k∈Z,`=1,...,n, we
know that for all f ∈ L2(R),
n XX
X
f= hf, D j Tk ψ`iDj Tk ψ`.
`=1 j∈Z k∈Z

As an approximation of f we can use

n XX
X
QJ f := hf, D j Tk ψ`iDj Tk ψ`
`=1 j<J k∈Z

for a reasonably large value of J ∈ Z. We say

that the frame {D j Tk ψ`}j,k∈Z,`=1,...,n provides
approximation order s if for all f ∈ H s(R),
||f − QJ f || = O(2−sJ ).

149
 We know that ψ1, . . . , ψn ∈ V1, so QJ f ∈ VJ
for all J ∈ Z; thus, the approximation order
of the frame {D j Tk ψ`}j,k∈Z,`=1,...,n can not ex-
ceed the approximation order of the underlying
refinable function ψ0.

We usually want the approximation order of

{Dj Tk ψ`}j,k∈Z,`=1,...,n to be as large as possible.

Can prove that

• The refinable function ψ0 = B2m provides
approximation order 2m.
• With the function θ chosen as in the OEP,
the approximation order of {D j Tk ψ`} is
min(2m, 2M ).

• Choosing M sufficiently large, we can obtain

the approximation order 2m, which is the
best possible one can hope for with the given
function ψ0 = B2m.

150
13 Construction of pairs of dual wavelet frames

Extension of the OEP to construction of dual

multiwavelet pairs:
Theorem 13.1 Let {H`, ψ`}n`=0 and {K`, ψe`}n`=0
be two sets of functions, satisfying the con-
ditions in the general setup on page 113, and
such that for some C > 0 and ρ > 21 ,

c c
f C
|ψ0(γ)|, |ψ0(γ)| ≤ ρ , a.e.
|γ|
Assume that there exists a function θ ∈ L∞(T)
such that limγ→0 θ(γ) = 1 and
n
X
H0(γ)K0(γ + ν)θ(2γ) + H`(γ)K`(γ + ν)
`=1

θ(γ) if ν = 0,
=
0 if ν = 21 .
Then {Dj Tk ψ`}j,k∈Z,`=1,...,n and {Dj Tk ψe`}j,k∈Z,`=1,...,n
are a pair of dual multiwavelet frames.

151
Easier to find a pair of dual frames via
Theorem 13.1 than to construct tight
frames via the OEP:
• The function θ is not required to be positive.
• We have freedom to chose two sets of trigono-
metric polynomials H` and K`: in fact, the
condition in the OEP corresponds exactly
to the condition in Theorem 13.1 with H` =
K`, and is more complicated to satisfy.

152
Setup for construction of pairs of dual
wavelet frames:

f0} be as in the general setup.

Let {ψ0, H0}, {K0, ψ
Let θ ∈ L∞(T) be a real-valued function for
which limγ→0 θ(γ) = 1, and assume that the
function
η(γ) := θ(γ)
 
1 1
− θ(2γ) H0(γ)K0(γ) + H0(γ + )K0(γ +
2 2
is real-valued and has a zero of order at least
2 at the origin. Choose real-valued functions
η1, η2 ∈ L∞(T) such that
η(γ) = 2η1(γ)η2(γ), and η1(0) = η2(0) = 0,
and choose two 21 -periodic and real-valued func-
tions θ1, θ2 such that
θ(2γ) = θ1(γ)θ2(γ).


153
In the construction of tight multiwavelet frames:
we had to perform a spectral factorization of the
functions θ and η.

The choices of the functions η1, η2, θ1, θ2 will

replace the spectral factorization - much easier!

154
Corollary 13.2 Assume the setup on page
153 and define {H`}3`=1 and {K`}3`=1 by

2πiγ 1
H1(γ) = e θ1(γ)K0(γ + ),
2
H2(γ) = η1(γ),
H3(γ) = e2πiγ η1(γ),
1
K1(γ) = e2πiγ θ2(γ)H0(γ + ),
2
K2(γ) = η2(γ),
K3(γ) = e2πiγ η2(γ).
Define the associated functions {ψ`}3`=1 and
{ψe`}3`=1 as in the general setup on page 113.
Then {Dj Tk ψ`}j,k∈Z,`=1,2,3 and {Dj Tk ψe`}j,k∈Z,`=1,2,3
constitute a pair of dual multiwavelet frames.

155
The number of generators can be reduced to
two:
Corollary 13.3 Assume the setup on page
153 and let
2πiγ 1
H1(γ) = e θ1(γ)K0(γ + ),
2
H2(γ) = η1(2γ)H0(γ),
1
K1(γ) = e2πiγ θ2(γ)H0(γ + ),
2
K2(γ) = η2(2γ)K0(γ).
Then {Dj Tk ψ`}j,k∈Z,`=1,2 and {Dj Tk ψe`}j,k∈Z,`=1,2
constitute a pair of dual multiwavelet frames.
If θ, H0 and K0 are trigonometric polynomi-
als and η1, η2 and θ1, θ2 are real-valued trigono-
metric polynomials, then the frame generators
{ψ`}3`=1 and {ψe`}3`=1 are symmetric if the re-
finable functions ψ0 and ψ f0 are symmetric real-
valued functions. Thus, the above process will
lead to symmetric dual wavelet pairs when ap-
plied to even-order B-splines.

156
Example 13.4 Frame construction with two
f0 = N2; the associ-
generators, based on ψ0 = ψ
ated two-scale symbol is
(1 + e−2πiγ )2
H0(γ) = = e−2πiγ cos2(πγ).
4
We again take
4 − cos(2πγ)
θ(γ) = ;
3
as proved in Example 11.6 this leads to
η(γ)
2
= (8 cos4(πγ) + 1)(cos(πγ) − 1)2(cos(πγ) + 1)2.
3
If we want to apply Corollary 13.3, we need to
find functions η1, η2, θ1, θ2 such that

η(γ) = 2η1(γ)η2(γ), and η1(0) = η2(0) = 0,
θ(2γ) = θ1(γ)θ2(γ)

Easy!

157
For example, we can take
η1(γ)

1
= (8 cos4(πγ) + 1)(cos(πγ) − 1)(cos(πγ) + 1)2,
3
η2(γ) = (cos(πγ) − 1),

4 − cos(4πγ)
θ1(γ) = 1, θ2(γ) = θ(2γ) = .
3

158
The functions in Corollary 13.3 are now
1
H1(γ) = e2πiγ θ1(γ)K0(γ + )
2
2πiγ 2
(1 − e )
= e2πiγ ,
4
2πiγ 1
K1(γ) = e θ2(γ)H0(γ + )
 2 
4πiγ −4πiγ
2πiγ 4 e +e (1 − e2πiγ )2
= e − ,
3 6 4
H2(γ) = η1(2γ)H0(γ) !
 2πiγ  4
1 e + e−2πiγ
= 8 +1
3 2
 2πiγ −2πiγ

e +e
× −1
2
 2πiγ 2
e + e−2πiγ (1 + e−2πiγ )2
× +1 ,
2 4
K2(γ) = η2(2γ)K0(γ)
 2πiγ −2πiγ

e +e (1 + e−2πiγ )2
= −1 .
2 4

159
With these choices, {D j Tk ψ`}j,k∈Z,`=1,2 and
{Dj Tk ψe`}j,k∈Z,`=1,2 constitute a pair of dual mul-
tiwavelet frames. 

160
14 The signal processing perspective

Goal: Formulate the UEP in signal processing

terms.

Formally, the Z-transform of a sequences {hk }k∈Z

is defined as the infinite series (depending on a
variable z ∈ C)
X
H̃(z) := hk z −k .
k∈Z

For finite sequences {hk }k∈Z, the Z-transform

is defined for all z 6= 0.

For `2(Z)-sequences (Fourier coefficients), the

Z-transform converges for a.e. z ∈ C with
|z| = 1.

The sequence {hk }k∈Z is often called a filter.

161
Consider the 1-periodic functions H`, ` = 0, . . . , n,
in the general setup. Write H` in terms of
Fourier series, with Fourier coefficients hk,`:
X
H`(γ) = hk,`e2πikγ .
k∈Z
In terms of the Z-transform, this means that
X

−2πiγ −k f`(e−2πiγ ).
H`(γ) = hk,` e =H
k∈Z
Reformulation of the UEP:
Theorem 14.1 Assume that the functions H`, ` =
0, . . . , n, have real Fourier coefficients hk,`, k ∈
Z. Then the UEP-conditions hold if and only
if the equations
 n
 X

 f`(z)H
H f`(z −1) = 1,


`=0
n
X



 f`(z)H
H f`(−z −1) = 0

`=0
hold for a.e. z ∈ C for which |z| = 1.

162
The polyphase decomposition

We can decompose a sequence {hk }k∈Z into “even“

and “odd“ parts:
(. . . , h−2, h−1, h0, h1, h2, . . . )
= (. . . , h−2, 0, h0, 0, h2, . . . )
+ (. . . , 0, h−1, 0, h1, 0, . . . ).
Thus, the Z-transformation of {hk }k∈Z is
 2 −2

H̃(z) = · · · + h−2z + h0 + h2z + · · ·
 −1 −3

+ · · · + h−1z + h1z + h3z + · · ·
 2 −2

= · · · + h−2z + h0 + h2z + · · ·
−1
 2 −2

+z · · · + h−1z + h1 + h3z + · · ·
X X
−2k −1
= h2k z +z h2k+1z −2k .
k∈Z k∈Z

The polyphase components of H̃(z) are now

defined as the two functions
X X
f0(z) :=
H −k
h2k z , Hf1(z) = h2k+1z −k .
k∈Z k∈Z

163
Thus, the Z-transformation has the polyphase
decomposition
f0(z 2) + z −1H
H̃(z) = H f1(z 2).

Consider now a given sequence of 1-periodic

functions H`, ` = 0, . . . , n, or, equivalently, a
sequence of filters {hk,`}k∈Z, ` = 0, . . . n. As-
sociated to the filter {hk,`}k∈Z, we denote the
polyphase components of H f` by H g g
`,0 and H`,1 .
Define the (n + 1) × 2 matrix of polyphase com-
ponents Hp by
 
g
H g
0,0 (z) H0,1 (z)
 g g 
 H1,0(z) H 1,1 (z) 
 
Hp(z) =  · · .
 
 · · 
g
H g
n,0 (z) Hn,1 (z)

164
The UEP in terms of the polyphase matrix:
Theorem 14.2 Assume that the functions H`,
` = 0, . . . , n, have real Fourier coefficients
hk,`, k ∈ Z. Then the UEP condition is sat-
isfied if and only if
1
HpT (z −1)Hp(z) = I (13)
2
for almost all z ∈ C with |z| = 1.

The condition (13) is well known in the

context of filter banks!

165
An analysis filter bank is a “black box”, which
performs operations on an incoming signal (i.e.,
a sequence of numbers).

Typically, a filter bank splits the incoming sig-

nal into certain subsignals, which contain par-
ticular information about the signal.

After processing the subsequences coming out of

the analysis filter bank, engineers usually wish
to get back to the original input sequence.

For this reason, an analysis filter bank is fol-

lowed by another filter bank, which reconstructs
the original signal from the subsignals; such a
filter bank is called a synthesis filter bank.

In that case the entire system consisting of the

two filter banks is said to have the perfect re-
construction property.

166
The filter banks considered here will contain
three operations on the incoming sequence {xk }k∈Z:
• Convolution with a sequence {hk }k∈Z:
The outcome is a new sequence,
P whose k-th
coordinate is given by n∈Z hnxk−n.
• Downsampling: The outcome is the se-
quence
↓ {xk }k∈Z := (· · · x−2, x0, x2, · · · }.
Thus, downsampling removes each second
element in the sequence.
• Upsampling: The outcome is the sequence
↑ {xk }k∈Z := (· · · x−1, 0, x0, 0, x1, · · · }.
Thus, upsampling inserts zeroes between the
elements in the sequence.

Note that downsampling is the left-inverse of

upsampling, but not the right-inverse.

167
Description of particular filter bank:

The analysis filter bank splits the incoming sig-

nal {xk }k∈Z into n + 1 subsignals: each of these
signals is obtained by convolving {xk }k∈Z with
a sequence hk,`, ` = 0, . . . , n, followed by a
downsampling.

The synthesis filter bank first upsamples each

of the incoming n + 1 subsignals, then con-
volves the resulting sequences with sequences
{gk,`}k∈Z, ` = 0, . . . , n, and finally add the
outcoming n + 1 signals.

See Figure 11.6.

We will assume that the sequences {hk,`}k∈Z

and {gk,`}k∈Z, ` = 0, . . . , n, are related by
gk,` = h−k,`, k ∈ Z, ` = 0, . . . , n.

168
 Figure 15:

For the above system , the perfect reconstruc-

tion property can be formulated in terms of the
polyphase components associated to the filters
{hk,`}k∈Z:
Theorem 14.3 For the considered filter bank,
the perfect reconstruction property is equiva-
lent to the condition
HpT (z −1)Hp(z) = I for z ∈ C with |z| = 1.

Note that the condition in Theorem 14.3 is “iden-

tical” to the UEP condition!
169
References

[1] Christensen, O.: Frames and bases in

mathematics and engineering. An intro-
ductory course. Birkhäuser 2007.
[2] Christensen, O. and Kim, R. Y.: On dual
Gabor frame pairs generated by polyno-
mials. Preprint, 2007.
[3] Chui, C., He, W., and Stöckler, J.:
Compactly supported tight and sibling
frames with maximum vanishing mo-
ments. Appl. Comp. Harm. Anal. 13 no.3
(2002), 226-262.
[4] Daubechies, I., Han, B., Ron, A. and Shen,
Z.: Framelets: MRA-based constructions
of wavelet frames. Appl. Comp. Harm.
Anal. 14 no.1 (2003), 1-46. 2001.
[5] Feichtinger, H. G. and Strohmer, T. (eds.):
Gabor Analysis and Algorithms: The-
ory and Applications. Birkhäuser, Boston,
1998.
170
[6] Feichtinger, H. G. and Strohmer, T. (eds.):
Advances in Gabor Analysis. Birkhäuser,
Boston, 2002.
[7] Janssen, A.J.E.M.: The duality condi-
tion for Weyl-Heisenberg frames. In ”Ga-
bor analysis: theory and applications”
(eds. H.G. Feichtinger and T. Strohmer).
Birkhäuser, Boston, 1998.
[8] Gröchenig, K.: Foundations of time-
frequency analysis. Birkhäuser, Boston,
2000.

171