Convolution and Filtering: The Convolution Theorem

Convolution and Filtering:

The Convolution Theorem
CS 450: Introduction to Digital Signal and Image Processing

Bryan Morse
BYU Computer Science
Convolution and Filtering: The Convolution Theorem
Introduction

Linear Systems and Responses

Time/Spatial Frequency
Input f F
Output g G
Impulse Response h
Transfer Function H
Relationship g=f∗ h G = FH

Is there a relationship?
Convolution and Filtering: The Convolution Theorem
Convolution Theorem

The Convolution Theorem

Let F , G, and H denote the Fourier Transforms of signals f , g,

and h respectively.

g=f∗ h g = fh

implies implies

G = FH G=F∗ H

Convolution in one domain is multiplication in the other

and vice versa.
Convolution and Filtering: The Convolution Theorem
Convolution Theorem

The Convolution Theorem

Thus,
F(f (t) ∗ g(t)) = F(f (t))F(g(t))

Likewise,
F(f (t)g(t)) = F(f (t)) ∗ F(g(t))
Convolution and Filtering: The Convolution Theorem
Convolution Theorem

Filtering: Frequency-Domain vs. Spatial (Convolution)

Convolution and Filtering: The Convolution Theorem
Convolution Theorem

Linear Systems and Responses

Time/Spatial Frequency
Input f F
Output g G
Impulse Response h
Transfer Function H
Relationship g=f∗ h G = FH

Relationship: the Transfer Function H(u)

is the Fourier Transform of the impulse response h(u)
Convolution and Filtering: The Convolution Theorem
Applications of the Convolution Theorem

Example 1: Frequency Analysis of Convolution

If the impulse response is a Gaussian with standard deviation σ,

what does this do to frequencies going through the system?

0.6
0.4
0.5
0.3
0.4

0.3 0.2
0.2
0.1
0.1

-10 -5 5 10 20 40 60 80 100 120

Time Domain Frequency Domain

Impulse Response Transfer Function
Convolution and Filtering: The Convolution Theorem
Applications of the Convolution Theorem

Example 2: Truncating a Signal

What is the Fourier Transform of a truncated signal?

10 cos(2πst) + 10 if −π ≤ t ≤ π
f (t) =
0 otherwise
20

15

10

-4 -2 2 4

(10 cos(2πst) + 10)(Rect2π (t))

Convolution and Filtering: The Convolution Theorem
Applications of the Convolution Theorem

Example 2 (cont’d)

20 200
175
15 150
125
10 100
75
5 50
25

-4 -2 2 4 50 100 150 200 250

Time Domain Frequency Domain

Convolution and Filtering: The Convolution Theorem
Applications of the Convolution Theorem

Example 3: Frequency Multiplexing

Suppose that we modulate (multiply) a carrier signal of a
particular frequency s by a signal f (t):

f (t) cos(2πst)

Fourier Transform:

F(f (t) cos(2πst)) = F(f (t)) ∗ F(cos(2πst))

1 1
= F (u) ∗ [ δ(u − s) + δ(u + s)]
2 2
1 1
= F (u − s) + F (u + s)]
2 2
This just shifts the signal in the frequency domain by the
frequency s of the carrier (splits between s and −s).
Convolution and Filtering: The Convolution Theorem
Applications of the Convolution Theorem

Example 3 (cont’d)

Signal Carrier Frequency 20 Carrier Frequency 40

1 1 1
0.75 0.75
0.8
0.5 0.5
0.6 0.25 0.25

0.4 -3 -2 -1 1 2 3 -3 -2 -1 1 2 3
-0.25 -0.25
0.2 -0.5 -0.5
-0.75 -0.75
-3 -2 -1 1 2 3 -1 -1

4 4 4
3.5 3.5 3.5
3 3 3
2.5 2.5 2.5
2 2 2
1.5 1.5 1.5
1 1 1
0.5 0.5 0.5

100 200 300 400 500 50 100 150 200 250 50 100 150 200 250
Convolution and Filtering: The Convolution Theorem
Applications of the Convolution Theorem

Example 3 (cont’d)

What if we add together multiple signals, each modulated by a

different carrier?

Can carry multiple signals through a single medium, each

carried by a different carrier frequency.
I Frequency multiplexing
I AM Radio

Just have to make sure the frequency-shifted signals don’t

overlap in the frequency-domain.
Convolution and Filtering: The Convolution Theorem
Applications of the Convolution Theorem

Convolution in the Frequency Domain

One application of the Convolution Theorem is that we can

perform time-domain convolution using frequency-domain
multiplication:
f ∗ g = F −1 (F(f )F(g))
What is the computational complexity of doing it this way?
Convolution and Filtering: The Convolution Theorem
Convolution and Correlation

Correlation
Convolution is
Z ∞
f (t) ∗ g(t) = f (τ ) g(t − τ ) dτ
−∞

Correlation is
Z ∞
f (t) ∗ g(−t) = f (τ ) g(t + τ ) dτ
−∞

Correlation is the “unflipped kernel” version of convolution that

we saw earlier as weighted neighborhood masks.
Convolution and Filtering: The Convolution Theorem
Convolution and Correlation

Correlation in the Frequency Domain

Can also do correlation in the frequency domain:

Convolution
f (t) ∗ g(t) ↔ F (s) G(s)

Correlation
f (t) ∗ g(−t) ↔ F (s) G∗ (s)
Convolution and Filtering: The Convolution Theorem
Convolution and Correlation

Autocorrelation

Autocorrelation is the correlation of a function with itself:

f (t) ∗ f (−t)

Useful for finding periodicity in a function (as the shifted copies

match the original, the output of the correlation is high).

In the Frequency Domain:

F (u) F ∗ (u) = |F (u)|2

This is the same as the power spectrum.

Convolution and Filtering: The Convolution Theorem
Deconvolution

Deconvolution - Basic Idea

If G = FH, can’t you reverse the process by F = G/H?

This is called deconvolution—the “undoing” of a convolution.

Problem: most systems have noise, which limits deconvolution.

(More on this later.)