An Introduction to Frames

Brody Dylan Johnson

St. Louis University

12 October 2010

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 1 / 36
 Overview

This goal of this talk is to introduce students to an area of mathematics called

frame theory, which draws heavily on linear algebra and finds application in
many real-world settings. In essence, frames provide a means for storing the
numeric data found in digital signals, such as those originating from images,
audio, and video. Frames can be designed for a variety of uses, e.g., data
compression, noise reduction, frequency analysis, etc.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 2 / 36
 Outline

Introduction

Elements of Linear Algebra

Frame Fundamentals

A Simple Tight Frame

A Bigger Example

Intuition for 2-D signals

Show & Tell

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 3 / 36
 Introduction

The Olden Days

AUDIO
Audio was recorded and stored in a continuous, or analog, format. The
earliest versions imprinted the signal in wax that could be retraced afterwards
to recover the recorded sound. Modern versions of analog recording imprint
the signal on magnetic tape (as shown below).

Figure: An 8mm tape reel from the early 1970’s.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 4 / 36
 Introduction

The Olden Days

I MAGES
Early photography made use of silver compounds that would undergo a
chemical reaction when exposed to light. The image was then captured on a
copper plate. This later developed into modern film photography where the
silver compounds are bonded to a plastic sheet (as shown below).

Figure: An 35mm negative from the early 2000’s.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 5 / 36
 Introduction

The Digital Revolution

AUDIO
A digital audio signal consists of a discrete sequence of numbers and a sample
rate. The sample rate describes how many digital samples are taken from the
analog signal in a given period of time.

Figure: 1000 of the 8,087,552 samples of the digitized version of the 1970’s tape
recording. This corresponds to about 0.0227 seconds.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 6 / 36
 Introduction

The Digital Revolution

I MAGES
A digital image typically consists of a rectangular array of numbers. The
numeric values in the array describe the intensity of light in the image at the
corresponding location. For grayscale images the intensity ranges from 0
(black) to 255 (white), while color images combine three separate intensity
values (red, green, and blue channels).

Figure: Digitized version of previously shown film negative.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 7 / 36
 Introduction

The Digital Revolution

S AMPLING
One obtains a discrete signal from a continuous one by a sampling procedure.
In the case of digital recording and digital photography the sampling is
typically performed by combining special hardware and software.

Figure: Representations of one- and two-dimensional sampling.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 8 / 36
 Introduction

The Digital Revolution

S AMPLING
One obtains a discrete signal from a continuous one by a sampling procedure.
In the case of digital recording and digital photography the sampling is
typically performed by combining special hardware and software.

Figure: Representations of one- and two-dimensional sampling.

In either case, one ends up with a vector representation of the original signal.
This allows us to use our knowledge of Linear Algebra.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 8 / 36
 Elements of Linear Algebra

Vector Spaces (Dimension 3)

V ECTORS
A vector has the form x = x1~i + x2~j + x3~k.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 9 / 36
 Elements of Linear Algebra

Vector Spaces (Dimension 3)

V ECTORS
A vector has the form x = x1~i + x2~j + x3~k.
I NNER P RODUCTS
The inner product of two vectors x and y is given by

hx, yi = x · y = x1 y1 + x2 y2 + x3 y3 = kxk kyk cos θ,

where kxk2 = x12 + x22 + x32 and θ is the angle between x and y.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 9 / 36
 Elements of Linear Algebra

Vector Spaces (Dimension 3)

V ECTORS
A vector has the form x = x1~i + x2~j + x3~k.
I NNER P RODUCTS
The inner product of two vectors x and y is given by

hx, yi = x · y = x1 y1 + x2 y2 + x3 y3 = kxk kyk cos θ,

where kxk2 = x12 + x22 + x32 and θ is the angle between x and y.
BASIS
A basis consists of three linearly independent vectors {u, v, w}, where
linearly independent means: the only solution of c1 u + c2 v + c3 w = 0 is
c1 = c2 = c3 = 0.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 9 / 36
 Elements of Linear Algebra

Vector Spaces (Dimension 3)

Given a basis {u, v, w}, how can one find the coefficients of a given vector x?
I.e., what values c1 , c2 , c3 achieve

x = c1 u + c2 v + c3 w?

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 10 / 36
 Elements of Linear Algebra

Vector Spaces (Dimension 3)

Given a basis {u, v, w}, how can one find the coefficients of a given vector x?
I.e., what values c1 , c2 , c3 achieve

x = c1 u + c2 v + c3 w?

Notice that one can write the above as a matrix equation:

    
x1 u1 v1 w1 c1
x2  = u2 v2 w3  c2  .
x3 u3 v3 w3 c3

Hence, one can find the coefficients by inverting the 3 × 3 matrix. Note that
this is possible because the vectors are linearly independent, implying that the
determinant is nonzero.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 10 / 36
 Elements of Linear Algebra

Vector Spaces (Dimension 3)

The basis is orthonormal if it satisfies

u·v=v·w=w·u=0 (orthogonality)

and
u · u = v · v = w · w = 1. (unit length)

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 11 / 36
 Elements of Linear Algebra

Vector Spaces (Dimension 3)

The basis is orthonormal if it satisfies

u·v=v·w=w·u=0 (orthogonality)

and
u · u = v · v = w · w = 1. (unit length)

Orthogonality allows us to find the coefficients using the inner product:

x · u = c1 u · u + c2 v · u + c3 w · u = c1 .

x · v = c1 u · v + c2 v · v + c3 w · v = c2 .
x · w = c1 u · w + c2 v · w + c3 w · w = c3 .

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 11 / 36
 Elements of Linear Algebra

Vector Spaces (Dimension 3)

The basis is orthonormal if it satisfies

u·v=v·w=w·u=0 (orthogonality)

and
u · u = v · v = w · w = 1. (unit length)

Orthogonality allows us to find the coefficients using the inner product:

x · u = c1 u · u + c2 v · u + c3 w · u = c1 .

x · v = c1 u · v + c2 v · v + c3 w · v = c2 .
x · w = c1 u · w + c2 v · w + c3 w · w = c3 .

Recall that the inner product x · y can also be written hx, yi.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 11 / 36
 Elements of Linear Algebra

Vector Spaces (Dimension 3)

The quantity kxk2 = x12 + x22 + x32 = hx, xi is commonly referred to as the
squared length of the vector x. However, in many applications it is reasonable
to consider this quantity as a measure of the energy in the signal x.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 12 / 36
 Elements of Linear Algebra

Vector Spaces (Dimension 3)

The quantity kxk2 = x12 + x22 + x32 = hx, xi is commonly referred to as the
squared length of the vector x. However, in many applications it is reasonable
to consider this quantity as a measure of the energy in the signal x.
If {u, v, w} is an orthonormal basis, then

kxk2 = hx, xi
= hc1 u + c2 v + c3 w, c1 u + c2 v + c3 wi
= c21 hu, ui + c22 hv, vi + c23 hw, wi + (zero terms)
= c21 + c22 + c23
= hx, ui2 + hx, vi2 + hx, wi2 .

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 12 / 36
 Elements of Linear Algebra

Vector Spaces (Dimension 3)

The quantity kxk2 = x12 + x22 + x32 = hx, xi is commonly referred to as the
squared length of the vector x. However, in many applications it is reasonable
to consider this quantity as a measure of the energy in the signal x.
If {u, v, w} is an orthonormal basis, then

kxk2 = hx, xi
= hc1 u + c2 v + c3 w, c1 u + c2 v + c3 wi
= c21 hu, ui + c22 hv, vi + c23 hw, wi + (zero terms)
= c21 + c22 + c23
= hx, ui2 + hx, vi2 + hx, wi2 .

This shows that the inner products “capture” the energy of the signal.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 12 / 36
 Elements of Linear Algebra

Vector Spaces (Arbitrary Dimension)

V ECTORS :

x = x1 x2 · · · xn

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 13 / 36
 Elements of Linear Algebra

Vector Spaces (Arbitrary Dimension)

V ECTORS :

x = x1 x2 · · · xn
I NNER P RODUCT:
n
X
hx, yi = x1 y1 + · · · + xn yn = xk yk
k=1

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 13 / 36
 Elements of Linear Algebra

Vector Spaces (Arbitrary Dimension)

V ECTORS :

x = x1 x2 · · · xn
I NNER P RODUCT:
n
X
hx, yi = x1 y1 + · · · + xn yn = xk yk
k=1

BASES :
As in the case of dimension 3, bases consist of n linearly independent
vectors {v1 , . . . , vn } and matrix inversion can be used to determine the
coefficients in a basis expansion of the form
n
X
x= ck vk .
k=1

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 13 / 36
 Frame Fundamentals

Definition of a Frame

Consider the following alternative to a basis, which focuses on the idea of

capturing the energy of a signal through inner products.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 14 / 36
 Frame Fundamentals

Definition of a Frame

Consider the following alternative to a basis, which focuses on the idea of

capturing the energy of a signal through inner products.

Definition
A collection of vectors e1 , e2 , . . . , em is a frame for an n-dimensional vector
space if there exist 0 < A ≤ B < ∞ such that for all vectors x,
m
X
2
Akxk ≤ hx, ek i2 ≤ Bkxk2 .
k=1

The numbers hx, ek i will be referred to as frame coefficients.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 14 / 36
 Frame Fundamentals

Definition of a Frame

Consider the following alternative to a basis, which focuses on the idea of

capturing the energy of a signal through inner products.

Definition
A collection of vectors e1 , e2 , . . . , em is a frame for an n-dimensional vector
space if there exist 0 < A ≤ B < ∞ such that for all vectors x,
m
X
2
Akxk ≤ hx, ek i2 ≤ Bkxk2 .
k=1

The numbers hx, ek i will be referred to as frame coefficients.

Okay, but so what? How can one know that this will be useful?

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 14 / 36
 Frame Fundamentals

Question #1:

Does this definition include bases?

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 15 / 36
 Frame Fundamentals

Question #1:

Does this definition include bases?

Yes. Notice that for an orthonormal basis one has
n
X
2
kxk = hx, ek i2 ,
k=1

which means the vectors form a frame with A = B = 1.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 15 / 36
 Frame Fundamentals

Question #1:

Does this definition include bases?

Yes. Notice that for an orthonormal basis one has
n
X
2
kxk = hx, ek i2 ,
k=1

which means the vectors form a frame with A = B = 1.

One can also prove the following result.
Proposition (see [1])
A collection of vectors is a frame for an n-dimensional vector space if and
only if it contains a basis.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 15 / 36
 Frame Fundamentals

Question #2:

How does one use a frame to represent a signal?

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 16 / 36
 Frame Fundamentals

Question #2:

How does one use a frame to represent a signal?

Recall that if the basis is orthonormal and we write x = c1 e1 + · · · + cm em ,
then taking inner products with ej on both sides gives
m
X
hx, ej i = ck hek , ej i = cj hej , ej i = cj .
k=1

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 16 / 36
 Frame Fundamentals

Question #2:

How does one use a frame to represent a signal?

Recall that if the basis is orthonormal and we write x = c1 e1 + · · · + cm em ,
then taking inner products with ej on both sides gives
m
X
hx, ej i = ck hek , ej i = cj hej , ej i = cj .
k=1

In other words,
m
X vector
hx, ek i ek
z}|{
x=
| {z }
k=1
coefficient

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 16 / 36
 Frame Fundamentals

Question #2:

How does one use a frame to represent a signal?

Recall that if the basis is orthonormal and we write x = c1 e1 + · · · + cm em ,
then taking inner products with ej on both sides gives
m
X
hx, ej i = ck hek , ej i = cj hej , ej i = cj .
k=1

In other words,
m
X vector
hx, ek i ek
z}|{
x=
| {z }
k=1
coefficient

This motivates a slightly different question.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 16 / 36
 Frame Fundamentals

Question #2: (modified)

How does one recover x from the frame coefficients {hx, ek i}mk=1 ?

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 17 / 36
 Frame Fundamentals

Question #2: (modified)

How does one recover x from the frame coefficients {hx, ek i}mk=1 ?
Well, we can try the same solution. Define the frame operator S by
m
X
Sx = hx, ek i ek .
k=1

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 17 / 36
 Frame Fundamentals

Question #2: (modified)

How does one recover x from the frame coefficients {hx, ek i}mk=1 ?
Well, we can try the same solution. Define the frame operator S by
m
X
Sx = hx, ek i ek .
k=1

If Sx = x, then it follows that A = B = 1 because

* m + m
X X
kxk2 = hx, xi = hSx, xi = hx, ek iek , x = hx, ek i2 .
k=1 k=1

Thus, when A 6= B recovery of x cannot be this easy.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 17 / 36
 Frame Fundamentals

Question #2: (modified)

However, using a little bit of advanced linear algebra one can prove the
following theorem about recovery from frame coefficients.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 18 / 36
 Frame Fundamentals

Question #2: (modified)

However, using a little bit of advanced linear algebra one can prove the
following theorem about recovery from frame coefficients.

Theorem (Frame Algorithm – see [1])

Given a frame {ek }m
k=1 one may recover x from its frame coefficients as
follows. Define x0 = 0 and
2
xj = xj−1 + S(x − xj−1 ).
A+B
 j
B−A
Then, kx − xj k ≤ B+A kxk.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 18 / 36
 Frame Fundamentals

Question #2: (modified)

However, using a little bit of advanced linear algebra one can prove the
following theorem about recovery from frame coefficients.

Theorem (Frame Algorithm – see [1])

Given a frame {ek }m
k=1 one may recover x from its frame coefficients as
follows. Define x0 = 0 and
2
xj = xj−1 + S(x − xj−1 ).
A+B
 j
B−A
Then, kx − xj k ≤ B+A kxk.

Notice that we only need the coefficients hx, ek i to compute Sx.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 18 / 36
 Frame Fundamentals

Some Remarks

The Frame Algorithm converges geometrically (error is reduced by the

same factor with each iteration) and in the case that A = B, convergence
is immediate. Frames where A = B are called tight frames.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 19 / 36
 Frame Fundamentals

Some Remarks

The Frame Algorithm converges geometrically (error is reduced by the

same factor with each iteration) and in the case that A = B, convergence
is immediate. Frames where A = B are called tight frames.
It is also possible to construct a dual frame consisting of vectors {ẽk }m
k=1
so that for all signals x one has
m
X
x= hx, ek iẽk .
k=1

One can even find the ẽk vectors using the Frame Algorithm, since
ẽk = S−1 ek , or, ek = Sẽk .

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 19 / 36
 A Simple Tight Frame

A tight-frame for R2

Define {e1 , e2 , e3 } by
√ √
1 3 1 3
e1 = (1, 0) e2 = (− , ) e3 = (− , − ).
2 2 2 2

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 20 / 36
 A Simple Tight Frame

A tight-frame for R2
To see that the collection is a frame, let x = (x1 , x2 ).

3
X
hx, ek i2 = hx, e1 i2 + hx, e2 i2 + hx, e3 i2
k=1
√ !2 √ !2
1 3 1 3
x12
= + − x1 + x2 + − x1 − x2
2 2 2 2
√ √
2 1 2 3 3 2 1 2 3 3
= x1 + x1 − x1 x2 + x2 + x1 + x1 x2 + x22
4 2 4 4 2 4
3 2
= (x1 + x22 )
2
3
= kxk2 .
2

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 21 / 36
 A Simple Tight Frame

A tight-frame for R2
To see that the collection is a frame, let x = (x1 , x2 ).

3
X
hx, ek i2 = hx, e1 i2 + hx, e2 i2 + hx, e3 i2
k=1
√ !2 √ !2
1 3 1 3
x12
= + − x1 + x2 + − x1 − x2
2 2 2 2
√ √
2 1 2 3 3 2 1 2 3 3
= x1 + x1 − x1 x2 + x2 + x1 + x1 x2 + x22
4 2 4 4 2 4
3 2
= (x1 + x22 )
2
3
= kxk2 .
2
So, it’s actually a tight frame.
Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 21 / 36
 A Bigger Example

A Random Frame of R10 of Size 25

The first 15 vectors:

-0.29 0.24 -0.18 0.33 0.61 -0.11 0.95 -0.29 0.70 -0.85 -0.91 -0.47 -0.81 0.64 -0.66
-0.62 -0.51 -0.43 0.45 0.66 -0.27 -0.56 -0.90 -0.58 -0.60 0.20 1.00 -0.97 -0.47 0.08
-0.02 0.17 -0.21 -0.44 -0.67 -0.39 0.41 0.51 -0.09 -0.90 0.90 -0.58 -0.42 0.51 0.25
-0.18 0.01 0.01 -0.48 -0.21 0.70 0.04 0.79 -0.84 0.13 -0.42 -0.00 0.63 0.32 0.37
-0.07 -0.07 0.44 0.42 0.04 0.52 0.87 -0.43 0.70 -0.76 0.78 -0.42 0.97 -0.57 0.35
0.22 0.08 -0.39 0.57 0.44 0.90 0.43 -0.50 0.12 0.04 -0.80 0.35 -0.97 0.20 0.75
-0.86 0.88 -0.78 0.97 0.14 0.12 -0.54 0.87 -0.36 -0.77 -0.87 0.92 0.64 0.21 -0.97
-0.37 -0.32 -0.11 -0.05 -0.08 -0.97 -0.10 -0.74 -0.25 0.54 -0.53 0.53 0.24 0.32 -0.38
0.22 -0.20 -0.07 0.81 -0.11 0.19 -0.66 0.88 0.74 -0.25 0.87 0.33 0.12 -0.63 0.56
-0.65 -0.38 -0.97 -0.10 -0.82 0.63 0.94 0.40 -0.26 0.65 -0.87 -0.74 -0.51 0.27 -0.39

The last 10 vectors:

0.85 0.02 0.90 0.30 -0.32 0.06 0.94 0.04 -0.20 0.33
0.36 -0.85 0.66 0.51 -0.07 -0.64 -0.95 0.79 -0.28 -0.73
-0.85 -0.61 0.84 0.33 0.83 0.00 0.74 0.88 -0.43 -0.96
-0.86 -0.24 -0.77 0.77 -0.54 -0.16 -0.95 -0.33 0.74 -0.48
-0.98 -0.45 0.62 -0.46 0.72 0.32 0.04 -0.13 0.25 -0.77
-0.55 0.54 0.82 -0.16 0.31 0.35 -0.62 -0.06 -0.52 -0.86
0.03 -0.37 -0.69 -0.57 0.78 0.91 0.43 -0.70 0.96 0.71
-0.08 0.28 -0.76 -0.93 -0.02 -0.62 -0.50 -0.73 0.28 -0.64
0.41 0.97 0.53 -0.84 0.99 -0.78 0.87 0.06 -0.54 -0.94
0.16 0.01 0.44 0.70 -0.25 0.13 -0.73 0.45 0.36 0.47

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 22 / 36
 A Bigger Example
10
A frame of R of size 25

Let F be the 10 × 25 matrix having these vectors as its columns.

 
| | |
F = e1 e2 · · · e25  .
| | |

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 23 / 36
 A Bigger Example
10
A frame of R of size 25

Let F be the 10 × 25 matrix having these vectors as its columns.

 
| | |
F = e1 e2 · · · e25  .
| | |

To compute the frame coefficients one multiplies the transpose by a

given vector:
   
− e1 −   hx, e1 i
− e2 − |
    hx, e2 i 
 
T
F x=  x =  ..  .

..
 . 
|
 . 
− e25 − hx, e25 i

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 23 / 36
 A Bigger Example
10
A frame of R of size 25

The frame operator can also be computed using F. In fact, S = FF T and is

shown below: (to two decimal places)
8.18 0.30 0.83 -3.62 -0.68 1.36 0.70 -1.75 -1.04 1.94
0.30 9.67 0.86 -0.88 -0.69 2.38 -1.75 0.62 -0.00 -0.21
0.83 0.86 8.79 -0.74 3.20 0.41 -2.32 -2.85 2.20 -0.07
-3.62 -0.88 -0.74 7.05 -0.35 -0.18 0.87 0.66 -3.39 2.17
-0.68 -0.69 3.20 -0.35 7.89 1.34 -0.66 -1.27 2.33 -1.99
1.36 2.38 0.41 -0.18 1.34 7.20 -1.59 -0.27 1.04 1.21
0.70 -1.75 -2.32 0.87 -0.66 -1.59 12.34 1.05 -0.55 -0.26
-1.75 0.62 -2.85 0.66 -1.27 -0.27 1.05 6.16 -0.75 -1.10
-1.04 -0.00 2.20 -3.39 2.33 1.04 -0.55 -0.75 9.83 -3.82
1.94 -0.21 -0.07 2.17 -1.99 1.21 -0.26 -1.10 -3.82 7.78

The minimum and maximum eigenvalues of this matrix (Matlab) actually

determine the frame bounds.
1.725 2.785 3.777 5.413 6.72 8.69 11.00 11.94 14.32 18.50

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 24 / 36
 A Bigger Example

Implementing the Frame Algorithm

The movie below shows 20 iterations of the frame algorithm.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 25 / 36
 A Bigger Example

Implementing the Frame Algorithm

The movie below shows 20 iterations of the frame algorithm.

B−A
The convergence is slow because ≈ 0.8294.
B+A
Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 25 / 36
 A Bigger Example
10
Another frame of R of size 25

The first 15 vectors:

1.00 0.97 0.87 0.72 0.52 0.28 0.03 -0.22 -0.46 -0.67 -0.84 -0.95 -1.00 -0.98 -0.90
1.00 0.87 0.52 0.03 -0.46 -0.84 -1.00 -0.90 -0.57 -0.10 0.40 0.80 0.99 0.93 0.62
1.00 0.72 0.03 -0.67 -1.00 -0.76 -0.10 0.62 0.99 0.80 0.16 -0.57 -0.98 -0.84 -0.22
1.00 0.52 -0.46 -1.00 -0.57 0.40 0.99 0.62 -0.35 -0.98 -0.67 0.28 0.97 0.72 -0.22
1.00 0.28 -0.84 -0.76 0.40 0.99 0.16 -0.90 -0.67 0.52 0.97 0.03 -0.95 -0.57 0.62
1.00 0.03 -1.00 -0.10 0.99 0.16 -0.98 -0.22 0.97 0.28 -0.95 -0.35 0.93 0.40 -0.90
1.00 -0.22 -0.90 0.62 0.62 -0.90 -0.22 1.00 -0.22 -0.90 0.62 0.62 -0.90 -0.22 1.00
1.00 -0.46 -0.57 0.99 -0.35 -0.67 0.97 -0.22 -0.76 0.93 -0.10 -0.84 0.87 0.03 -0.90
1.00 -0.67 -0.10 0.80 -0.98 0.52 0.28 -0.90 0.93 -0.35 -0.46 0.97 -0.84 0.16 0.62
1.00 -0.84 0.40 0.16 -0.67 0.97 -0.95 0.62 -0.10 -0.46 0.87 -1.00 0.80 -0.35 -0.22

The last 10 vectors:

-0.76 -0.57 -0.35 -0.10 0.16 0.40 0.62 0.80 0.93 0.99
0.16 -0.35 -0.76 -0.98 -0.95 -0.67 -0.22 0.28 0.72 0.97
0.52 0.97 0.87 0.28 -0.46 -0.95 -0.90 -0.35 0.40 0.93
-0.95 -0.76 0.16 0.93 0.80 -0.10 -0.90 -0.84 0.03 0.87
0.93 -0.10 -0.98 -0.46 0.72 0.87 -0.22 -1.00 -0.35 0.80
-0.46 0.87 0.52 -0.84 -0.57 0.80 0.62 -0.76 -0.67 0.72
-0.22 -0.90 0.62 0.62 -0.90 -0.22 1.00 -0.22 -0.90 0.62
0.80 0.16 -0.95 0.72 0.28 -0.98 0.62 0.40 -1.00 0.52
-1.00 0.72 0.03 -0.76 0.99 -0.57 -0.22 0.87 -0.95 0.40
0.72 -0.98 0.93 -0.57 0.03 0.52 -0.90 0.99 -0.76 0.28

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 26 / 36
 A Bigger Example

Implementing the Frame Algorithm

The movie below shows 20 iterations of the frame algorithm.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 27 / 36
 A Bigger Example

Implementing the Frame Algorithm

The movie below shows 20 iterations of the frame algorithm.

B−A
The convergence is faster because ≈ 0.1723. (This frame was not
B+A
chosen randomly.)
Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 27 / 36
 Intuition for 2-D signals

Vector Spaces for 2-D Signals

The following representation is used for two-dimensional signals, e.g.,
images.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 28 / 36
 Intuition for 2-D signals

Vector Spaces for 2-D Signals

The following representation is used for two-dimensional signals, e.g.,
images.
Vectors:  
x11 ··· x1n
 .. .. .. 
x= . . . 
xm1 ··· xmn

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 28 / 36
 Intuition for 2-D signals

Vector Spaces for 2-D Signals

The following representation is used for two-dimensional signals, e.g.,
images.
Vectors:  
x11 ··· x1n
 .. .. .. 
x= . . . 
xm1 ··· xmn
Inner Product:
m X
X n
hx, yi = xj,k yj,k
j=1 k=1

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 28 / 36
 Intuition for 2-D signals

Vector Spaces for 2-D Signals

The following representation is used for two-dimensional signals, e.g.,
images.
Vectors:  
x11 ··· x1n
 .. .. .. 
x= . . . 
xm1 ··· xmn
Inner Product:
m X
X n
hx, yi = xj,k yj,k
j=1 k=1

Basis: A collection {uj,k : 1 ≤ j ≤ m, 1 ≤ k ≤ n} is a basis provided that

the only solution of
X m X n
cj,k uj,k = 0
j=1 k=1

is the trivial solution cj,k = 0, 1 ≤ j ≤ m, 1 ≤ k ≤ n.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 28 / 36
 Intuition for 2-D signals

3 × 3 Image Space

An example signal:

 
0.95 0.49 0.45
x = 0.23 0.89 0.02
0.61 0.76 0.82

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 29 / 36
 Intuition for 2-D signals

3 × 3 Image Space

An example signal:

 
0.95 0.49 0.45
x = 0.23 0.89 0.02
0.61 0.76 0.82

An orthonormal basis:

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 29 / 36
 Show & Tell

Wavelet Basis

Below we see an image and its wavelet basis coefficients. (no redundancy)

Original

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 30 / 36
 Show & Tell

Wavelet Basis

Below we see an image and its wavelet basis coefficients. (no redundancy)

Original Single Scale

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 30 / 36
 Show & Tell

Wavelet Basis

Below we see an image and its wavelet basis coefficients. (no redundancy)

Original Single Scale Full

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 30 / 36
 Show & Tell

Wavelet Basis

Below we see an image and its wavelet basis coefficients. (no redundancy)

Original Single Scale Full

The gray pixels correspond to coefficients close to zero. Black and white
pixels correspond to - and + coefficients, respectively.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 30 / 36
 Show & Tell

Wavelet Basis – Compression

By ignoring small coefficients one can “compress” the image.

Original

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 31 / 36
 Show & Tell

Wavelet Basis – Compression

By ignoring small coefficients one can “compress” the image.

Original Reconstructed

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 31 / 36
 Show & Tell

Wavelet Basis – Compression

By ignoring small coefficients one can “compress” the image.

Original Reconstructed

The compression ratio here is 33.8 with a mean-squared error of 2.66.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 31 / 36
 Show & Tell

Wavelet Frame – Algorithme À Trous

Below we see an image and its wavelet frame coefficients (highly redundant).

Original

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 32 / 36
 Show & Tell

Wavelet Frame – Algorithme À Trous

Below we see an image and its wavelet frame coefficients (highly redundant).

Original Single Scale

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 32 / 36
 Show & Tell

Wavelet Frame – Algorithme À Trous

Below we see an image and its wavelet frame coefficients (highly redundant).

Original Single Scale

Each component of the coefficient image has the same size as the original.
Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 32 / 36
 Show & Tell

Wavelet Frame – Denoising

By ignoring small coefficients one can remove noise from the image. The
greater redundancy helps preserve the original signal features.

Noisy

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 33 / 36
 Show & Tell

Wavelet Frame – Denoising

By ignoring small coefficients one can remove noise from the image. The
greater redundancy helps preserve the original signal features.

Noisy Reconstructed

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 33 / 36
 Show & Tell

Wavelet Frame – Denoising

By ignoring small coefficients one can remove noise from the image. The
greater redundancy helps preserve the original signal features.

Noisy Reconstructed

As the level of the noise increases denoising will begin to affect important
signal features, resulting in a blurring of the image.
Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 33 / 36
 Show & Tell

Discrete Cosine Transform (DCT) – Basis

The JPEG image standard originally made use of a basis whose elements are
described by discrete cosine functions. (Newer JPEG standards use wavelets.)

Original

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 34 / 36
 Show & Tell

Discrete Cosine Transform (DCT) – Basis

The JPEG image standard originally made use of a basis whose elements are
described by discrete cosine functions. (Newer JPEG standards use wavelets.)

Original DCT

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 34 / 36
 Show & Tell

Discrete Cosine Transform (DCT) – Basis

The JPEG image standard originally made use of a basis whose elements are
described by discrete cosine functions. (Newer JPEG standards use wavelets.)

Original DCT

The periodic nature of the basis elements lends itself to representation of

signals with repeating patterns.
Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 34 / 36
 Conclusion

Concluding Remarks

The rise of computers has opened new avenues of research in

mathematics and other fields. Frame theory is just one example.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 35 / 36
 Conclusion

Concluding Remarks

The rise of computers has opened new avenues of research in

mathematics and other fields. Frame theory is just one example.
Frame theory is very accessible to students with some exposure to linear
algebra. A little analysis doesn’t hurt.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 35 / 36
 Conclusion

Concluding Remarks

The rise of computers has opened new avenues of research in

mathematics and other fields. Frame theory is just one example.
Frame theory is very accessible to students with some exposure to linear
algebra. A little analysis doesn’t hurt.
Computational platforms like Matlab are great for working with real data
(like audio or images). There is also an open-source Matlab emulator
called Octave:
http://www.octave.org/

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 35 / 36
 Conclusion

Concluding Remarks

The rise of computers has opened new avenues of research in

mathematics and other fields. Frame theory is just one example.
Frame theory is very accessible to students with some exposure to linear
algebra. A little analysis doesn’t hurt.
Computational platforms like Matlab are great for working with real data
(like audio or images). There is also an open-source Matlab emulator
called Octave:
http://www.octave.org/

T HANK YOU !

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 35 / 36
 Conclusion

References

B.D. Johnson,
Frames in Rn , expository notes,
http://mathcs.slu.edu/~johnson/public/maths/frames.pdf.

Brody Dylan Johnson (St. Louis University) An Introduction to Frames 12 October 2010 36 / 36