Introduction to Signal Processing

Lecture 4: Frequency analysis of signals

Part 2 – discrete time

Filip Elvander

Dept. Information and Communications Engineering

Aalto University
0.
Today’s lecture

March 6, 2025
0.
Today’s lecture

Discrete-time signals • We continue our journey in frequency

(Ch. 2) domain.
Discrete-time systems • Representing and understanding
(Ch. 2) signals in terms of frequency
(sinusoids).
Frequency analysis of signals
• Last lecture: continuous time.
(Ch. 4 and 6.1 (Ch. 7 later))
• Today: discrete time.
z-transform
• Concepts: Discrete-Time Fourier
(Ch. 3)
Transform (DTFT) and its inverse.
Freq. analysis of systems and filters • Also: sampling and reconstruction of
(Ch. 5 and 7) signals.
Implementation of systems • Reading: Chapter 4, 6.1 (aliasing and
(Ch. 9) sampling).

Filter design
(Ch. 10)

March 6, 2025
0.
Four types of signals
1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0

-0.2 -0.2

-0.4 -0.4

-0.6 -0.6

-0.8 -0.8

-1 -1
0 2 4 6 8 10 0 2 4 6 8 10
t t
Continuous and periodic. Continuous and aperiodic.
5 5

4 4

3 3

2 2

1 1

0 0

-1 -1

-2 -2

-3 -3

-4 -4

-5 -5
0 5 10 15 20 0 5 10 15 20
n n
Discrete and periodic. Discrete and aperiodic.
March 6, 2025
0.
Continuous-time periodic signals
Z Tp
1
Analysis equation: ck = xa (t)e− j2πkF0 t dt (Fourier coefficients)
Tp 0
∞
Synthesis equation: xa (t) = ∑ ck e j2πkF0 t (Fourier series)
k=−∞
Z Tp ∞
1
Parseval’s identity: |xa (t)|2 dt = ∑ |ck |2
Tp 0 k=−∞
1 0.025
| ck | 2
0.8

0.6 0.02

0.4

0.2 0.015

-0.2 0.01

-0.4

-0.6 0.005

-0.8

-1 0
0 2 4 6 8 10 -6 -4 -2 0 2 4 6
t F (Hz)

Cont.-time periodic signal. Power spectrum.

March 6, 2025
0.
Continuous-time aperiodic signals

Z ∞
Analysis equation: Xa (F) = xa (t)e− j2πFt dt (Fourier transform)
−∞
Z ∞
Synthesis equation: xa (t) = Xa (F)e j2πFt dF (inverse Fourier transform)
−∞
Z ∞ Z ∞
Parseval’s identity: |xa (t)|2 dt = |X(F)|2 dF
−∞ −∞

1 0.14
|X(F)|2
0.8
0.12
0.6

0.4 0.1

0.2
0.08
0
0.06
-0.2

-0.4 0.04

-0.6
0.02
-0.8

-1 0
0 2 4 6 8 10 -8 -6 -4 -2 0 2 4 6 8
t F
Cont.-time aperiodic signal. Energy spectrum.
March 6, 2025
0.
Four types of signals
1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0

-0.2 -0.2

-0.4 -0.4

-0.6 -0.6

-0.8 -0.8

-1 -1
0 2 4 6 8 10 0 2 4 6 8 10
t t
Continuous and periodic. Continuous and aperiodic.
5 5

4 4

3 3

2 2

1 1

0 0

-1 -1

-2 -2

-3 -3

-4 -4

-5 -5
0 5 10 15 20 0 5 10 15 20
n n
Discrete and periodic. Discrete and aperiodic.
March 6, 2025
0.
Discrete-time periodic signals

• In discrete time we are dealing with 5

sequences 4
x(n) = {. . . , x(−1), x(0), x(1), . . .}. 3

• Periodic with period N ∈ Z if 2

x(n + N) = x(n) for all n. 1

0
• It is enough to be able to explain
-1
x(0), x(1), . . . , x(N − 1). Any description
that is matches these values and -2

have period N will match the full x(n). -3

-4
• What would be a good description?
-5
0 5 10 15 20
n
Discrete and periodic.

March 6, 2025
0.
Recap: periodic signals in continuous time
Recall that for periodic continuous-time signal xa (t) we could write
∞
xa (t) = ∑ ck e j2πkF0 t , F0 = 1/Tp .
k=−∞

• e j2πkF0 t , k = 0, ±1, . . ., are the harmonics of the series.

R Tp j2πkF t − j2πℓF t
• The are orthogonal in the sense 0 dt = 0 if k ̸= ℓ.
0 e
0 e

• Can we ”copy” this idea to discrete-time signals x(n)?

March 6, 2025
0.
Harmonics in discrete time
If x(n + N) = x(n) for all n (with N ∈ N = {1, 2, . . .}), then f0 = 1/N is the
”fundamental frequency”.
• What happens if we create ”harmonics” xk (n) = e j2πk f0 n = e j2πkn/N ,
k = 0, ±1, . . .?
• Then, we could have something like x(n) = ∑∞ j2πkn/N .
k=−∞ ck e
First observation: most harmonics are actually useless. Note that for any
p∈Z
xk+pN (n) = e j2π(k+pN) f0 n = e j2π(k+pN)n/N = e j2πkn/N e j2π pNn/N
= e j2πkn/N e j2π pn = e j2πkn/N = xk (n).

• Any harmonic you choose will be identical to one of the first N

harmonics, i.e., to xk (n) = e j2πkn/N , k = 0, 1, . . . , N − 1.
• You have seen this before: the higher harmonics are aliases of the first
N ones.
• Critical question: is it enough to use only N harmonics to describe any
N-periodic x(n)?
March 6, 2025
0.
Harmonics in discrete time

March 6, 2025
0.
Collecting facts

• As x(n + N) = x(n), we have succeeded if we can find a representation

matching x(0), x(1), . . . , x(N − 1) and that is periodic. Then we match x(n)
for any n.
• ⇒ in fact, we need a basis for N-dimensional vectors that is periodic.
• Can our set of N harmonics xk (n) = e j2πkn/N , k = 0, 1, . . . , N − 1, be a
candidate?

1. Periodic? We already know from the previous lecture that linear

combinations of harmonically related sinusoids are periodic. This is
inherited by discrete-time signals. Answer is YES!
2. Basis for N-dimensional vectors? A basis for N-dimensional space
consists of N elements/vectors that are linearly independent. We do
have N vectors (our harmonics). Are they linearly independent? YES!
because they are orthogonal!
(
N−1 N−1 N−1 N k=ℓ
j2πkn/N − j2πℓn/N j2π(k−ℓ)n/N
∑ xk (n)xℓ (n) = ∑ e e = ∑e =
0 k ̸= ℓ.
n=0 n=0 n=0

March 6, 2025
0.
Representation of discrete-time periodic signals
We have shown that for any N-periodic discrete-time signal x(n)
N−1
∃c0 , c1 , . . . , cN−1 ∈ C : x(n) = ∑ ck e j2πkn/N .
k=0

But what are the coefficients ck ? As in the previous lecture, the orthogonality
of the basis makes it very easy. Take any ℓ ∈ {0, 1, . . . , N − 1}. Then,
N−1 N−1 N−1 N−1 N−1
∑ e− j2πℓn/N x(n) = ∑ e− j2πℓn/N ∑ ck e j2πkn/N = ∑ ck ∑ e j2π(k−ℓ)n/N
n=0 k=0 k=0 k=0 n=0
= Ncℓ .

We can conclude that

N−1
1
ck =
N ∑ x(n)e− j2πkn/N , k = 0, 1, . . . , N − 1.
n=0

We have just figured out Fourier representations for periodic discrete-time

signals!

March 6, 2025
0.
Discrete-time periodic signals

N−1
1
Analysis equation: ck =
N ∑ x(n)e− j2πkn/N
n=0
N−1
Synthesis equation: x(n) = ∑ ck e j2πkn/N
k=0
N−1 N−1
1
Parseval’s identity:
N ∑ |x(n)|2 = ∑ |ck |2 (as an exercise: verify this!)
n=0 k=0

• The coefficients {ck } is a periodic sequence of period N: for ℓ ∈ Z,

N−1 N−1
1 1
ck+ℓN =
N ∑ x(n)e− j2π(k+ℓN)n/N = N ∑ x(n)e− j2πkn/N e|−{z
j2πℓn
} = ck
n=0 n=0
1

• Recall, in discrete time we can only distinguish frequencies on the

interval [−Fs /2, Fs /2], where Fs is the sampling frequency.
• This is reflected in that the power at (normalized) frequency f = k/N can
be explained by frequency (k + ℓN)/N = k/N + ℓ.
• Special case: for k ∈ {0, 1, . . . , N − 1}, c−k = c−k+N .
March 6, 2025
0.
Discrete-time periodic signal and spectrum
5 3

4
2.5
3

2
2
1
| ck |2
0 1.5

-1

-2 1

-3
0.5
-4

-5 0
0 5 10 15 20
-6 -4 -2 0 2 4 6 8 10 12
n k
Disc.-time periodic signal. Power spectrum.

March 6, 2025
0.
Four types of signals
1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0

-0.2 -0.2

-0.4 -0.4

-0.6 -0.6

-0.8 -0.8

-1 -1
0 2 4 6 8 10 0 2 4 6 8 10
t t
Continuous and periodic. Continuous and aperiodic.
5 5

4 4

3 3

2 2

1 1

0 0

-1 -1

-2 -2

-3 -3

-4 -4

-5 -5
0 5 10 15 20 0 5 10 15 20
n n
Discrete and periodic. Discrete and aperiodic.
March 6, 2025
0.
Discrete-time aperiodic signals

• Maybe not so easy as before 5

• N−periodic discrete-time signals are 4

a finite-dimensional space (dimension 3

N). 2

1
• Now, we have infinite dimension:
0
infinitely long sequences with
-1
arbitrary patterns.
-2
• We will need an infinite-dimensional
-3
basis.
-4

-5
0 5 10 15 20
n
Discrete and aperiodic.

March 6, 2025
0.
What did we do in continuous time?
In the last lecture, when the continuous-time signal xa (t) was aperiodic,
Z ∞
Analysis equation: Xa (F) = xa (t)e− j2πFt dt (Fourier transform)
−∞
Z ∞
Synthesis equation: xa (t) = Xa (F)e j2πFt dF (inverse Fourier transform)
−∞

Can we do something similar here? Let’s construct an ”analysis equation”

∞
X(ω) = ∑ x(n)e− jωn , ω ∈ (−π, π] (angular frequency).
n=−∞

• We basically translate the analysis equation from continuous time.

• Why angular frequency (radians/sample)? This is simply a convention.
• Why limit the frequency axis to (−π, π]? Note that for any k ∈ Z,
∞ ∞ ∞
X(ω + k2π) = ∑ x(n)e− j(ω+k2π)n = ∑ x(n)e− jωn e− jk2πn = ∑ x(n)e− jωn
n=−∞ n=−∞ n=−∞
= X(ω), ⇒ X(ω) is 2π − periodic

March 6, 2025
0.
Synthesis?
∞
X(ω) = ∑ x(n)e− jωn , ω ∈ (−π, π].
n=−∞

This transform is only useful if we can synthesize x(n) from X(ω). Note that
sinusoids e jωn for n ∈ Z form an orthogonal set of functions on (−π, π]:


(
2π n = m
Z π Z π
− jωn − jωm jω(m−n)
e e dω = e dω =
−π −π 0 n ̸= m,

so lets try
!
∞
1 1
Z π Z π
jωm − jωn
X(ω)e dω = ∑ x(n)e e jωm dω
2π −π 2π −π n=−∞
∞ ∞
1 1
Z π Z π
”nice” series
= ∑ x(n) e− jωn e jωm dω = ∑ x(n) e jω(m−n) dω = x(n).
2π n=−∞ −π 2π n=−∞ −π

Technically, if ∑∞
n=−∞ |x(n)| < ∞, then x(n) is ”nice” enough for this to work.

March 6, 2025
0.
Signal and its Fourier transform

5 3.5
|X( )|2
4
3
3

2 2.5

1
2
0
1.5
-1

-2 1

-3
0.5
-4

-5 0
0 5 10 15 20 -3 -2 -1 0 1 2 3
n
Discrete and aperiodic signal. Energy spectrum.

March 6, 2025
0.
Discrete-time aperiodic signals

∞
Analysis equation: X(ω) = ∑ x(n)e− jωn (Discrete-time Fourier transform, DTFT)
n=−∞
1 π
Z
Synthesis equation: x(n) = X(ω)e jωm dω
2π −π
∞ Z π
Parseval’s identity: ∑ |x(n)|2 = |X(ω)|2 dω
n=−∞ −π

• Important: X(ω) is 2π-periodic, i.e., X(ω + k2π) = X(ω) for all k ∈ Z.

• If ∑∞ N − jωn
n=−∞ |x(n)| < ∞, then the partial sums XN (ω) = ∑n=−N x(n)e
converge uniformly to X(ω).
2
• If we weaken this to ∑∞
n=−∞ |x(n)| < ∞, then uniform convergence is not
guaranteed, but we have convergence in the sense
Z π
|XN (ω) − X(ω)|2 → 0.
−π

• This subtle difference can be demonstrated by the Gibbs phenomenon.

March 6, 2025
0.
A classic but problematic spectrum
(
1, |ω| ≤ ωc
X(ω) =
0, ωc < |ω| ≤ π.

Then, the corresponding discrete-time signal is

(
ωc
1 1 π , n=0
Z π
jωm
x(n) = X(ω)e dω = sinc(ωc n) = ωc n sin(ωc n)
2π −π π π ωc n , n ̸= 0.

2
You can verify that ∑∞ ∞
−∞ |x(n)| does not converge, but ∑−∞ |x(n)| < ∞.
Rπ 2
⇒ We are guaranteed −π |XN (ω) − X(ω)| → 0 but not uniform convergence.
0.25 1.2
x(n) X( )

0.2 1

0.15
0.8

0.1
0.6

0.05
0.4
0

0.2
-0.05

0
-0.1
-20 -15 -10 -5 0 5 10 15 20
-3 -2 -1 0 1 2 3
n

March 6, 2025
0.
Gibbs phenomenon
Let’s construct the partial sums
N
1
XN (ω) = ∑ x(n)e− jωn , where x(n) = sinc(ωc n).
n=−N π

⇒ can be shown that over and undershoot at ω = ωc never goes to zero.

1.2 1.15
X( ) X( )
XN( ) XN( )
1 1.1

0.8 1.05

0.6 1

0.4 0.95

0.9
0.2

0.85
0

0.8
-0.2 0.6 0.65 0.7 0.75 0.8 0.85
-1.5 -1 -0.5 0 0.5 1 1.5

N = 100.

March 6, 2025
0.
Gibbs phenomenon
Let’s construct the partial sums
N
1
XN (ω) = ∑ x(n)e− jωn , where x(n) = sinc(ωc n).
n=−N π

⇒ can be shown that over and undershoot at ω = ωc never goes to zero.

1.2 1.15
X( ) X( )
XN( ) XN( )
1 1.1

0.8 1.05

0.6 1

0.4 0.95

0.9
0.2

0.85
0

0.8
-0.2 0.6 0.65 0.7 0.75 0.8 0.85
-1.5 -1 -0.5 0 0.5 1 1.5

N = 5000.

March 6, 2025
0.
Sampling
Sampling with sampling period T (i.e., sampling frequency Fs = 1/T ):

x(n) = xa (T n) , n = . . . , −2, −1, 0, 1, 2, . . .

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0

-0.2 -0.2

-0.4 -0.4

-0.6 -0.6

-0.8 -0.8

-1 -1
0 5 10 15 20 0 5 10 15 20
t n

March 6, 2025
0.
Frequency consequences of sampling?
Let xa (t) be a continuous-time aperiodic signal. Consider sampling xa (t)
with sample frequency Fs :
x(n) = xa (nT ), n ∈ Z where T = 1/Fs .
Recall that for the continuous-time signal
Z ∞
Xa (F) = xa (t)e− j2πFt dt
−∞
Z ∞
xa (t) = Xa (F)e j2πFt dF
−∞
and for the discrete-time signal
∞
X(ω) = ∑ x(n)e− jωn
n=−∞
Z 1/2
1
Z π
2π f =ω
x(n) = X(ω)e jωn dω = X( f )e j2π f n d f
2π −π −1/2
Z Fs /2
f =F/Fs 1
= X(F)e j2πnF/Fs dF.
Fs −Fs /2

What can we say about the spectrum X( f ) for the discrete-time signal?
March 6, 2025
0.
Frequency consequences of sampling?
By definition of the sampling, and using T = 1/Fs ,
Z ∞ Z ∞
synthesis
x(n) = xa (nT ) = Xa (F)e j2πFnT dF = Xa (F)e j2πnF/Fs dF
−∞ −∞
∞ Z (k+1/2)Fs
= ∑ Xa (F)e j2πnF/Fs dF (split integration interval)
k=−∞ −(k+1/2)Fs
∞ Z Fs /2
= ∑ Xa (F − kFs )e j2πn(F−kFs )/Fs dF (change of variables F ↔ F − kFs )
k=−∞ −Fs /2
Z Fs /2 ∞
= ∑ Xa (F − kFs )e j2πnF/Fs dF
−Fs /2 k=−∞

Replacing x(n) by its synthesis equation for x(n) yields

Z Fs /2 Z Fs /2 ∞
1
X(F)e j2πnF/Fs dF = ∑ Xa (F − kFs )e j2πnF/Fs dF.
Fs −Fs /2 −Fs /2 k=−∞

What does this tell us?

March 6, 2025
0.
Frequency consequences of sampling!
We have the equality
Z Fs /2 Z Fs /2 ∞
X(F)e j2πnF/Fs dF = Fs ∑ Xa (F − kFs )e j2πnF/Fs dF,
−Fs /2 −Fs /2 k=−∞

n all n ∈ Z.
which holds for o
• The set e j2πnF/Fs as functions of F form an orthogonal basis for (a
certain type of) functions on [Fs /2, Fs /2].
• ⇒ X(F) and Fs ∑∞ k=−∞ Xa (F − kFs ) have to be equal almost everywhere!
For our purposes, we can conclude that
∞
X(F) = Fs ∑ Xa (F − kFs ) for F ∈ [−Fs /2, Fs /2].
k=−∞

• If Xa (F) = 0 for |F| > Fs /2, then X(F) = Fs Xa (F). All good!
• However, if Xa (F) has power above Fs /2, then
X(F) = Fs Xa (F) + Fs Xa (F+Fs ) + Fs Xa (F−Fs ) + . . .
• This is the aliasing effect: power from high frequencies show up at low
frequencies.
March 6, 2025
0.
Aliasing

1.5

0.5

0
-5 0 5
F
March 6, 2025
0.
Aliasing

1.5

0.5

0
-5 0 5
F
March 6, 2025
0.
Aliasing

1.5

0.5

0
-5 0 5
F
March 6, 2025
0.
Aliasing

1.5

0.5

0
-5 0 5
F
March 6, 2025
0.
Aliasing

1.5

0.5

0
-5 0 5
F
March 6, 2025
0.
Aliasing - how to compute X(F)

2 2

1.5 1.5

1 1

0.5 0.5

0 0
-5 0 5 -1 -0.5 0 0.5 1
F F

March 6, 2025
0.
Aliasing - how to compute X(F)

1. Plot Xa (F) and its shifted copies

Xa (F − kFs ) for k ∈ Z. You only need to
consider k such that Xa (F − kFs ) ̸= 0
2
for some F ∈ [−Fs /2, Fs /2].

1.5

0.5

0
-1 -0.5 0 0.5 1
F

March 6, 2025
0.
Aliasing - how to compute X(F)

1. Plot Xa (F) and its shifted copies

Xa (F − kFs ) for k ∈ Z. You only need to
consider k such that Xa (F − kFs ) ̸= 0
2
for some F ∈ [−Fs /2, Fs /2].
2. Ignore everything outside
1.5
[−Fs /2, Fs /2], as X(F) will be periodic.

0.5

0
-1 -0.5 0 0.5 1
F

March 6, 2025
0.
Aliasing - how to compute X(F)

1. Plot Xa (F) and its shifted copies

Xa (F − kFs ) for k ∈ Z. You only need to
consider k such that Xa (F − kFs ) ̸= 0
2
for some F ∈ [−Fs /2, Fs /2].
2. Ignore everything outside
1.5
[−Fs /2, Fs /2], as X(F) will be periodic.
3. Add up all contributions on
1
[−Fs /2, Fs /2].

0.5

0
-1 -0.5 0 0.5 1
F

March 6, 2025
0.
Aliasing - how to compute X(F)

1. Plot Xa (F) and its shifted copies

Xa (F − kFs ) for k ∈ Z. You only need to
consider k such that Xa (F − kFs ) ̸= 0
2
for some F ∈ [−Fs /2, Fs /2].
2. Ignore everything outside
1.5
[−Fs /2, Fs /2], as X(F) will be periodic.
3. Add up all contributions on
1
[−Fs /2, Fs /2].
4. Voilà, you have computed F1s X(F) for
0.5
F ∈ [−Fs /2, Fs /2]! You can extend it
periodically for F ∈ / [−Fs /2, Fs /2].
0
-1 -0.5 0 0.5 1
F

March 6, 2025
0.
Aliasing for sound

-20 -20
xa xa
x x
-40 Fs/2
-40 Fs/2

-60 -60

-80 -80
dB

dB
-100 -100

-120 -120

-140 -140

-160 -160
0 5 10 15 20 0 500 1000 1500 2000
frequency (kHz) frequency (Hz)

Original (blue) Without aliasing (blue) With aliasing (red)

March 6, 2025
0.
Reconstructing a signal from samples
Let’s assume that Xa (F) = 0 for |F| > Fs /2 and sample xa (t) with sampling
frequency Fs :
x(n) = xa (nT ) = xa (n/Fs ) , T = 1/Fs .
Then, for any Zt ∈ R, Z F2 /2
synthesis ∞
xa (t) = Xa (F)e j2πFt dF = Xa (F)e j2πFt dF
−∞ −Fs /2
Z F2 /2
Xa (F)=X(F)
= X(F)e j2πFt dF
−Fs /2
Z F2 /2
!
∞
analysis − j2πFn/Fs
= ∑ x(n)e e j2πFt dF
−Fs /2 n=−∞
∞ Z F2 /2 ∞ Z F2 /2
= ∑ x(n) e− j2πFn/Fs e j2πFt dF = ∑ x(n) e j2πF(t−nT ) dF
n=−∞ −Fs /2 n=−∞ −Fs /2
∞  
t − nT sin(πt)
= ∑ xa (nT )sinc , sinc(t) =
n=−∞ T πt
⇒ xa (t) for any t ∈ R can be perfectly reconstructed from the samples xa (nT )!
This is how an (ideal) digital-to-analog converter (DAC) works.
March 6, 2025
0.
Nyquist-Shannon sampling theorem
Let Xa (F) be bandlimited, i.e., Xa (F) = 0 for |F| > B. Then, xa (t) can be
uniquely recovered from its samples xa (nT ) = xa (n/Fs ) if Fs > 2B. The
reconstruction is given by the sinc-interpolation
∞  
t − nT sin(πt)
xa (t) = ∑ xa (nT )sinc , sinc(t) = .
n=−∞ T πt

1.5

0.5

-0.5

-1

-1.5
-4 -2 0 2 4 6
t

March 6, 2025
0.
Nyquist-Shannon sampling theorem
Let Xa (F) be bandlimited, i.e., Xa (F) = 0 for |F| > B. Then, xa (t) can be
uniquely recovered from its samples xa (nT ) = xa (n/Fs ) if Fs > 2B. The
reconstruction is given by the sinc-interpolation
∞  
t − nT sin(πt)
xa (t) = ∑ xa (nT )sinc , sinc(t) = .
n=−∞ T πt

1.5

0.5

-0.5

-1

-1.5
-4 -2 0 2 4 6
t

March 6, 2025
0.
Nyquist-Shannon sampling theorem
Let Xa (F) be bandlimited, i.e., Xa (F) = 0 for |F| > B. Then, xa (t) can be
uniquely recovered from its samples xa (nT ) = xa (n/Fs ) if Fs > 2B. The
reconstruction is given by the sinc-interpolation
∞  
t − nT sin(πt)
xa (t) = ∑ xa (nT )sinc , sinc(t) = .
n=−∞ T πt

March 6, 2025
0.
Nyquist-Shannon sampling theorem
Let Xa (F) be bandlimited, i.e., Xa (F) = 0 for |F| > B. Then, xa (t) can be
uniquely recovered from its samples xa (nT ) = xa (n/Fs ) if Fs > 2B. The
reconstruction is given by the sinc-interpolation
∞  
t − nT sin(πt)
xa (t) = ∑ xa (nT )sinc , sinc(t) = .
n=−∞ T πt

1.5

0.5

-0.5

-1

-1.5
-4 -2 0 2 4 6
t

March 6, 2025
0.
Nyquist-Shannon sampling theorem
Let Xa (F) be bandlimited, i.e., Xa (F) = 0 for |F| > B. Then, xa (t) can be
uniquely recovered from its samples xa (nT ) = xa (n/Fs ) if Fs > 2B. The
reconstruction is given by the sinc-interpolation
∞  
t − nT sin(πt)
xa (t) = ∑ xa (nT )sinc , sinc(t) = .
n=−∞ T πt

1.5

0.5

-0.5

-1

-1.5
-4 -2 0 2 4 6
t

March 6, 2025
0.
Summary of frequency domain

• Describe signals in terms of frequency ↔ change of basis.

• We do this via Fourier series and the Fourier transform.
• This is an absolutely wonderful toolbox that is fundamental to modern
engineering.
• In this and previous lecture, we have seen how Fourier analysis allows
us to understand signals in terms of frequency.
• Soon we will see how we can exploit this when doing processing.
• On your own: freshen up your memory on the properties of the Fourier
transform (end of Chapter 4).
• For next week, read Chapter 3: ”The z-transform and Its application to
the Analysis of LTI Systems”.

March 6, 2025