ELT2035 Signals & Systems

Lecture 8: Applications of Fourier

representations

Hoang Gia Hung

Faculty of Electronics and Telecommunications
University of Engineering and Technology, VNU Hanoi
 Last lecture review

❑ Properties of Fourier representations

➢ Symmetry
➢ Linearity
➢ Convolution
➢ Differentiation/integration
➢ Time shift/frequency shift
➢ Scaling
➢ Multiplication
➢ Parseval relationship
➢ Time – bandwidth product
➢ Relating FT to FS, DTFT to DTFS

❑ Today’s lecture: apply of Fourier representations to analysis of

signals and systems.
 1. Convolution of periodic and
nonperiodic signals
❑ System analysis: input signal is periodic while the system’s
impulse response is nonperiodic
➢ Suppose that 𝑥(𝑡) is periodic, ℎ(𝑡) is nonperiodic, what is 𝑦(𝑡)?
𝐹𝑇
➢ 𝑦 𝑡 =𝑥 𝑡 ∗ℎ 𝑡 𝑌 𝑗𝜔 = 𝑋(𝑗𝜔)𝐻(𝑗𝜔)

Time domain
?
𝑥(𝑡) ℎ(𝑡) 𝑦(𝑡)

➢ Strictly speaking, neither FT nor DTFT converges for periodic signals

❑ How do we find the FT for 𝑥 𝑡 = σ∞

𝑘=−∞ 𝑋[𝑘]𝑒
𝑗𝑘𝜔0 𝑡 (1)
𝐹𝑇
➢ FT of the unit DC signal: 1 2𝜋𝛿(𝜔). Applying frequency shift property
𝐹𝑇
of the FT, we obtain: 𝑒 𝑗𝑘𝜔0 𝑡 2𝜋𝛿(𝜔 − 𝑘𝜔0 ) (2)
➢ Substituting (2) to (1) and using linear property of the FT, we will arrive at
𝐹𝑇
σ∞
𝑘=−∞ 𝑋[𝑘]𝑒
𝑗𝑘𝜔0 𝑡
2𝜋 σ∞
𝑘=−∞ 𝑋[𝑘]𝛿 𝜔 − 𝑘𝜔0
𝑥(𝑡) 𝑋(𝑗𝜔)
 Mixed-signal convolution
❑ Now,𝑌 𝑗𝜔 can be obtained by
𝐹𝑇
➢ 𝑦 𝑡 =𝑥 𝑡 ∗ℎ 𝑡 𝑌 𝑗𝜔 = 2𝜋 σ∞
𝑘=−∞ 𝑋 𝑘 𝛿 𝜔 − 𝑘𝜔0 𝐻(𝑗𝜔)

Time domain
?
𝑥(𝑡) ℎ(𝑡) 𝑦(𝑡)
∞
FT domain ෍ 𝑋 𝑘 𝛿 𝜔 − 𝑘𝜔0 × 𝐻(𝑗𝜔) = 𝑌(𝑗𝜔)
𝑘=−∞
➢ Applying the sifting property of the impulse:
𝑌 𝑗𝜔 = 2𝜋 σ∞
𝑘=−∞ 𝑋 𝑘 𝛿 𝜔 − 𝑘𝜔0 𝐻(𝑗𝑘𝜔0 ).
➢ The form of 𝑌 𝑗𝜔 corresponds to a periodic signal → 𝑦(𝑡) is periodic with
the same period as 𝑥(𝑡)

❑ Real-life appications: filtering.

Graphical illustration
X 𝑗𝜔
2𝜋𝑋[0]
2𝜋𝑋[2]

−𝜔0 𝜔0
𝜔
−4𝜔0 −2𝜔0 2𝜔0 4𝜔0

2𝜋𝑋[1]

H 𝑗𝜔

𝜔
𝜔0 𝜔0 𝜔0 𝜔0 2𝜔0 4𝜔0

Y 𝑗𝜔
2𝜋𝑋 0 𝐻[𝑗0]

2𝜋𝑋 2 𝐻(𝑗2𝜔0 )
−𝜔0 𝜔0
𝜔
−4𝜔0 −2𝜔0 2𝜔0 4𝜔0

2𝜋𝑋[1]𝐻(𝑗𝜔0 )
Ideal filters
 Example
Let the input signal applied to an LTI system with impulse ℎ 𝑡 =
1
sin 𝜋𝑡 be the periodic square wave. Find the output of the system.
𝜋𝑡

❑ Solution:
1, 𝜔 ≤ 𝜋
➢ The system frequency response is ℱ ℎ(𝑡) = 𝐻(𝑗𝜔) = ቊ .
0, 𝜔 > 𝜋
2 sin 𝑘𝜋 Τ2 𝜋
➢ The FT of the periodic square wave 𝑋 𝑗𝜔 = σ∞
𝑘=−∞ 𝛿 𝜔−𝑘 .
𝑘 2
➢ The FT of the system output is
𝜋 𝜋
𝑌 𝑗𝜔 = 𝑋 𝑗𝜔 𝐻 𝑗𝜔 = 2𝛿 𝜔 + + 𝜋𝛿 𝜔 + 2𝛿 𝜔 − .
2 2
1 2 𝜋
➢ Taking the inverse FT of 𝑌 𝑗𝜔 gives the output 𝑦 𝑡 = + cos 𝑡 .
2 𝜋 2
 Example solution from frequency
domain perspective
H 𝑗𝜔

𝜔
❖ Observation: 𝐻 𝑗𝜔 acts as −𝜋 𝜋

a low-pass filter, passing

the harmonics at −𝜋/2 and X 𝑗𝜔

𝜋/2, while suppressing all 𝜋

others.
𝜔
−𝜋 −𝜋/2 𝜋/2 𝜋

Y 𝑗𝜔

𝜔
−𝜋/2 𝜋/2
 2. Amplitude modulation (AM)
▪ Antenna efficiency
(size, power
radiation, …) for
high range
transmission.
▪ Spectral efficiency
(highest freq.
/lowest freq.)
▪ Interference
prevention

❑ Let 𝑚 𝑡 & 𝑐 𝑡 = 𝐴𝑐 cos 𝜔𝑐 𝑡 be the message signal & carrier

wave, resp.
➢ Full amplitude modulation: 𝑠 𝑡 = 𝐴𝑐 [1 + 𝑘𝑎 𝑚(𝑡)] cos 𝜔𝑐 𝑡 is the process that
varies the amplitude of the carrier in proportion to the message signal.
➢ 𝑘𝑎 is a constant called the amplitude sensitivity factor of the modulator
➢ 𝑎 𝑡 = 𝐴𝑐 |1 + 𝑘𝑎 𝑚 𝑡 | is called the envelope of the AM wave 𝑠 𝑡
 Spectral content of AM wave
❑ Let find ℱ 𝑠(𝑡)
𝐹𝑇
➢ 𝐴𝑐 cos 𝜔𝑐 𝑡 𝜋𝐴𝑐 𝛿 𝜔 − 𝜔𝑐 + 𝛿(𝜔 + 𝜔𝑐 ) .
𝐹𝑇 𝑀 𝑗 𝜔−𝜔𝑐 +𝑀 𝑗 𝜔+𝜔𝑐
➢ 𝑚 𝑡 cos 𝜔𝑐 𝑡
2
➢ Applying linearity & scaling properties of the FT, we obtain 𝑆 𝑗𝜔 =
1
𝜋𝐴𝑐 𝛿 𝜔 − 𝜔𝑐 + 𝛿(𝜔 + 𝜔𝑐 ) + 𝑘𝑎 𝐴𝑐 𝑀 𝑗𝜔 − 𝑗𝜔0 + 𝑀 𝑗𝜔 + 𝑗𝜔0 → AM
2
shifts the modulating signal’s spectrum by ±𝜔0
 Example
❑ Consider a modulating wave 𝑚 𝑡 consisting of a single
frequency component (single tone), i.e. 𝑚 𝑡 = 𝐴0 cos 𝜔0 𝑡 .
Evaluate the time domain & frequency domain characteristics
of the AM wave.
❑ Solution:
➢ 𝑠 𝑡 = 𝐴𝑐 1 + 𝜇 cos 𝜔0 𝑡 cos 𝜔𝑐 𝑡 with 𝜇 = 𝑘𝑎 𝐴0 . To avoid envelope
distortion due to overmodulation, 𝜇 must be less than 1.
1 1
➢ 𝑠 𝑡 = 𝐴𝑐 cos 𝜔𝑐 𝑡 + 𝜇𝐴𝑐 cos[ 𝜔𝑐 + 𝜔0 𝑡] + 𝜇𝐴𝑐 cos[ 𝜔𝑐 − 𝜔0 𝑡]
2 2

➢ Let 𝐴𝑚𝑎𝑥 & 𝐴𝑚𝑖𝑛 denote the maximum & minimum values of the envelope of
𝐴 −𝐴
the modulated wave, then 𝜇 = 𝑚𝑎𝑥 𝑚𝑖𝑛
𝐴𝑚𝑎𝑥 +𝐴𝑚𝑖𝑛

1
➢ 𝑆 𝑗𝜔 = 𝜋𝐴𝑐 𝛿 𝜔 − 𝜔𝑐 + 𝛿(𝜔 + 𝜔𝑐 ) + 𝜋𝜇𝐴𝑐 [𝛿 𝜔 − 𝜔𝑐 − 𝜔0 + 𝛿 (𝜔 −
2
1
𝜔𝑐 + 𝜔0 )] + 𝜋𝜇𝐴𝑐 𝛿 𝜔 + 𝜔𝑐 − 𝜔0 + 𝛿 𝜔 + 𝜔𝑐 + 𝜔0
2
Example solution
 3. Sampling of continuous time signals

❑ A discrete-time signal 𝑔[𝑛] is obtained by sampling a CT signal

𝑥(𝑡) at every interval 𝑇𝑠, i.e. 𝑔[𝑛] = 𝑥(𝑛𝑇𝑠).
➢ The sampled signal is mathematically represented by 𝑥𝛿 𝑡 =
σ∞𝑛=−∞ 𝑥(𝑛𝑇𝑠 )𝛿(𝑡 − 𝑛𝑇𝑠 ). Since 𝑥 𝑛𝑇𝑠 𝛿 𝑡 − 𝑛𝑇𝑠 = 𝑥(𝑡)𝛿 𝑡 − 𝑛𝑇𝑠 , we may
rewrite 𝑥𝛿 𝑡 = 𝑥 𝑡 𝑝(𝑡) where 𝑝 𝑡 = σ∞ 𝑛=−∞ 𝛿(𝑡 − 𝑛𝑇𝑠 ) is an impulse train.
➢ The process of sampling is characterized by the multiplication of the signal
with an impulse train → impulse sampling
➢ What’s the relationship between FT of 𝑥𝛿 𝑡 and FT of 𝑥 𝑡 ?
 Sampling of continuous time signals
(cont.)
❑ Applying multiplication property
1 1 ∞ 2𝜋
➢ 𝑋𝛿 𝑗𝜔 = 𝑋 𝑗𝜔 ∗ 𝑃 𝑗𝜔 = σ 𝑋(𝑗 𝜔 − 𝑘𝜔𝑠 ), where 𝜔𝑠 = is the
2𝜋 𝑇𝑠 𝑘=−∞ 𝑇𝑠
sampling frequency.
➢ The FT of the sampled signal is an infinite sum of shifted versions of the
original signal’s FT.
 Aliasing
❑ If 𝜔𝑠 < 2𝑊(the highest frequency component of 𝑥(𝑡)) → the
shifted replicas of the original spectrum is overlapped → aliasing
➢ Aliasing distorts the spectrum of the sampled signal → impossible to
reconstruct or analyze the original signal
➢ To avoid aliasing, 𝜔𝑠 > 2𝑊 (discussed in detail later in sampling theorem)
 Reconstruction of CT signals from
samples

❑ To reconstruct the signal uniquely from its samples, there must

be a unique correspondence between the FT of the signal and
that of the its sampled signal.
❑ Sampling theorem:
𝐹𝑇
➢ Let 𝑥 𝑡 𝑋(𝑗𝜔) represent a band-limited signal, so that 𝑋 𝑗𝜔 = 0 for
𝜔 > 𝜔𝑚 . If 𝜔𝑠 > 2𝜔𝑚 , where 𝜔𝑠 = 2𝜋/𝑇𝑠 is the sampling frequency, then
𝑥 𝑡 is uniquely determined by its samples 𝑥 𝑛𝑇𝑠 , 𝑛 = 0, ±1, ±2, …
➢ The minimum sampling frequency, 2𝜔𝑚 , is called Nyquist rate.
Example
 Ideal reconstruction

❑ The goal is to apply some operation to 𝑋𝛿 𝑗𝜔 to convert it back

to 𝑋 𝑗𝜔 → must eliminate images of 𝑋 𝑗𝜔 (centered at 𝑘𝜔𝑠 ).
➢ Achieved by spectrum truncation, i.e. multiplying 𝑋𝛿 𝑗𝜔 by 𝐻𝑅 (𝑗𝜔) =
𝑇 , 𝜔 ≤ 𝜔𝑠 Τ2
ቊ 𝑠 , i.e. 𝑋 𝑗𝜔 = 𝑋𝛿 𝑗𝜔 𝐻𝑅 (𝑗𝜔).
0, 𝜔 > 𝜔𝑠 Τ2
➢ Multiplication by 𝐻𝑅 (𝑗𝜔) will not recover 𝑋 𝑗𝜔 from 𝑋𝛿 𝑗𝜔 if aliasing occurs.

❑ Time domain analysis: 𝑥 𝑡 = σ∞

𝑛=−∞ 𝑥 𝑛 sinc 𝜔𝑠 𝑡 − 𝑛𝑇𝑠 Τ 2𝜋
➢ 𝑥 𝑡 is a weighted sum of sinc functions shifted by the sampling interval.

*
 Practical reconstruction

❑ In practice, we use a device called zero-order hold (ZOH),

which holds the value of 𝑥 𝑛 for 𝑇𝑠 seconds to reconstruct 𝑥 𝑡 .
1, 0 < 𝑡 < 𝑇𝑠
➢ The impulse response of the ZOH is ℎ0 (𝑡) = ቊ , and the output is
0, otherwise
𝑥0 𝑡 = ℎ0 (𝑡) ∗ 𝑥𝛿 (𝑡).
➢ In the frequency domain 𝑋0 𝑗𝜔 = 𝐻0 (𝑗𝜔)𝑋𝛿 𝑗𝜔 , where 𝐻0 𝑗𝜔 = ℱ ℎ0 (𝑡) =
sin 𝜔 𝑇𝑠 Τ2
2𝑒 −𝑗𝜔𝑇𝑠Τ2 .
𝜔
 Practical reconstruction (cont.)

❑ Side effects of the ZOH:

➢ A linear phase shift corresponding to a time delay of 𝑇𝑠 Τ2 seconds.
➢ Distortion in the mainlobe of 𝑋0 𝑗𝜔 .
➢ Distorted and attenuated images of 𝑋 𝑗𝜔 , centered at 𝑘𝜔𝑠 .
❑ These side effects are eliminated by further processing by a
compensation/anti-imaging filter 𝐻𝑐 𝑗𝜔 .
 4. Fourier representations of finite-
duration aperiodic signals
❑ DTFS is the only Fourier representation that can be exactly
evaluated numerically.
➢ Apply to DT periodic signals only, can we extend it to nonperiodic signals?

❑ Let 𝑥[𝑛] be a finite-duration signal of length 𝑀 (i.e. 𝑥 𝑛 = 0 ∀𝑛 ∉

[0, 𝑀)) → 𝑋 𝑒 𝑗𝛺 = σ𝑀−1𝑛=0 𝑥[𝑛]𝑒
−𝑗𝛺𝑛

➢ Construct a periodic signal with period 𝑁 ≥ 𝑀, 𝑥[𝑛],

෤ such that one period of
1 𝑁−1
𝑥[𝑛]
෤ is 𝑥[𝑛] by zero-padding → 𝑋෨ 𝑘 = σ𝑛=0 𝑥[𝑛]𝑒
෤ −𝑗𝑘Ω0 𝑛
, where Ω0 = 2𝜋/𝑁.
𝑁

➢ Since 𝑥෤ 𝑛 = 𝑥 𝑛 ∀𝑛 ∉ [0, 𝑁) and 𝑥 𝑛 = 0 for 𝑀 ≤ 𝑛 < 𝑁, the DTFS of 𝑥[𝑛]

෤
1 𝑀−1 1
becomes 𝑋෨ 𝑘 = σ𝑛=0 𝑥[𝑛]𝑒 −𝑗𝑘𝛺0 𝑛 ⟹ 𝑋෨ 𝑘 = 𝑋 𝑒 𝑗𝛺 หΩ=𝑘𝛺 .
𝑁 𝑁 0

➢ The DTFS coefficients of 𝑥෤ 𝑛 are samples of the DTFT of 𝑥 𝑛 , with

sampling intervals of 2𝜋/𝑁, divided by 𝑁.

❑ The effect of sampling the DTFT of a finite-duration nonperiodic

signal is to periodically extend the signal in the time domain.
➢ Duality property: sampling a signal in time domain generates replicas of the
original spectrum in the frequency domain.
Graphical illustration of the relationship
between the DTFS and the DTFT
 Example: sampling the DTFT of a
cosine pulse
3𝜋
cos 𝑛 , 0 ≤ 𝑛 ≤ 31
Consider the signal 𝑥 𝑛 = ቐ 8 .
0, otherwise
Derive both the DTFT and DTFS of 𝑥 𝑛 , assuming a period 𝑁 = 32, 60.
SOLUTION:
❑ First, we evaluate the DTFT of 𝑥[𝑛]. Rewrite 𝑥[𝑛] = 𝑔[𝑛]𝑤[𝑛] with
3𝜋 1, 0 ≤ 𝑛 ≤ 31
𝑔 𝑛 = cos 𝑛 , 𝑤[𝑛] = ቊ .
8 0, otherwise
𝑘
, 𝑘 = ±1
➢ Find the DTFS of 𝑔[𝑛] by inspection: 𝐺[𝑘] = ൝2 . Hence, 𝐺 𝑒 𝑗Ω =
0, 𝑘 ≠ ±1
3𝜋 3𝜋 3𝜋
2𝜋 σ∞
𝑘=−∞ 𝐺 𝑘 𝛿 Ω − 𝑘 = 𝜋𝛿 Ω + + 𝜋𝛿 Ω − , −𝜋 < Ω < 𝜋.
8 8 8
31Ω
1−𝑒 −𝑗32Ω sin 16Ω
➢ 𝑊 𝑒 𝑗Ω = σ31
𝑛=0 𝑒
−𝑗Ω𝑛
= = 𝑒 −𝑗 2
Ω .
1−𝑒 −𝑗Ω sin
2
1
➢ Applying multiplication property: 𝑋 𝑒 𝑗Ω = 𝐺 𝑒 𝑗Ω ⊛ 𝑊 𝑒 𝑗Ω =
2𝜋
31(Ω+3𝜋Τ8) 31(Ω−3𝜋Τ8)
−𝑗 −𝑗
𝑒 2 sin 16(Ω+3𝜋 Τ8) 𝑒 2 sin 16(Ω−3𝜋 Τ8)
(Ω+3𝜋Τ8) + (Ω−3𝜋Τ8)
2 sin 2 sin
2 2
 Example solution
❑ Now, we are going to derive the DTFS of 𝑥[𝑛] by sampling 𝑋 𝑒 𝑗Ω at 𝑁
intervals from – 𝜋 to +𝜋 .
31(𝑘ΩΤ𝑁+3𝜋Τ8) 31(𝑘ΩΤ𝑁−3𝜋Τ8)
−𝑗 −𝑗
𝑒 2 sin 16(𝑘 ΩΤ𝑁+3𝜋 Τ8) 𝑒 2 sin 16(𝑘 ΩΤ𝑁−3𝜋 Τ8)
➢ 𝑋𝑘 = (𝑘ΩΤ𝑁+3𝜋Τ8) + (𝑘ΩΤ𝑁−3𝜋Τ8) (1)
2𝑁 sin 2𝑁 sin
2 2

❑ Let’s check if DTFS derived from sampling the DTFT above is the same
as the DTFS that is computed directly:
1 31 3𝜋𝑛 −𝑗𝑘Ω 𝑛
➢ 𝑋𝑘 = σ𝑛=0 cos 𝑒 0 , with Ω = 2𝜋 Τ𝑁.
𝑁 8 0
1 −𝑗 𝑘Ω0 +3𝜋 Τ8 𝑛 1
➢ 𝑋𝑘 = σ31
𝑛=0 𝑒 + σ31
𝑛=0 𝑒 −𝑗 𝑘Ω0 −3𝜋Τ8 𝑛 ,
after some simple
2𝑁 2𝑁
manipulations, the direct method also yields the same 𝑋[𝑘] as given in (1).
 5. Fast Fourier transform
❑ Computing the DTFS directly from the definition is often too slow
1
➢ 𝑋 𝑘 = σ𝑁−1 𝑥[𝑛]𝑒 −𝑗𝑘Ω0 𝑛 , 𝑥 𝑛 = σ𝑁−1
𝑘=0 𝑋[𝑘]𝑒
𝑗𝑘Ω0 𝑛
→ an algorithm can
𝑁 𝑛=0
compute either relationship with minor changes.
➢ Evaluating 𝑥[𝑛] for a single value of 𝑛 requires 𝑁 complex multiplications &
𝑁 − 1 complex additions → the complete computation of 𝑥[𝑛] requires 𝑁2
complex multiplications and 𝑁2 − 𝑁 complex additions.

❑ Suppose 𝑁 is even
➢ Split 𝑋[𝑘] into even- and odd-indexed coefficients, i.e. 𝑋𝑒[𝑘] = 𝑋[2𝑘], 0 ≤ 𝑘 ≤
𝑁’ − 1 and 𝑋𝑜[𝑘] = 𝑋[2𝑘 + 1], 0 ≤ 𝑘 ≤ 𝑁’ − 1 with 𝑁’ = 𝑁/2, we have:
𝑁 −1 ′ ′
𝑥 𝑛 = σ𝑚=0 𝑋[2𝑚]𝑒 𝑗𝑚2Ω0 𝑛 + σ𝑁 −1
𝑚=0 𝑋[2𝑚 + 1]𝑒
𝑗(𝑚2Ω0 𝑛+Ω0 𝑛)
=
′ ′ ′ ′
σ𝑁 −1
𝑚=0 𝑋𝑒[𝑚]𝑒
𝑗𝑚Ω0 𝑛
+ 𝑒 𝑗Ω0 𝑛 σ𝑁 −1
𝑚=0 𝑋𝑜 [𝑚]𝑒
𝑗𝑚Ω0 𝑛
= 𝑥𝑒 𝑛 + 𝑒 𝑗Ω0 𝑛 𝑥𝑜 𝑛 .
′ )Ω
➢ Since 𝑥𝑒 𝑛 + 𝑁 ′ = 𝑥𝑒 𝑛 , 𝑥𝑜 𝑛 + 𝑁 ′ = 𝑥𝑜 𝑛 , and 𝑒 𝑗(𝑛+𝑁 0 = −𝑒 𝑗𝑛Ω0 ,
𝑥 𝑛 = 𝑥𝑒 𝑛 + 𝑒 𝑗𝑛Ω0 𝑥𝑜 𝑛 , 0 ≤ 𝑛 ≤ 𝑁 ′ − 1 (4.49)
as the first 𝑁 ′ values of 𝑥 𝑛 , and
𝑥 𝑛 + 𝑁 ′ = 𝑥𝑒 𝑛 − 𝑒 𝑗𝑛Ω0 𝑥𝑜 𝑛 , 0 ≤ 𝑛 ≤ 𝑁 ′ − 1 (4.50)
as the second 𝑁 ′ values of 𝑥 𝑛 .
 Fast Fourier transform
(cont.)
❑ Only need to multiply 𝑒 𝑗𝑛Ω0 once in computing both Eqs. (4.49) &
(4.50). The remaining operations are addition and subtraction.
 Fast Fourier transform
(cont.)
❑ The total number of complex multiplications required to evaluate
Eqs. (4.49) and (4.50) are 𝑁2/2 + 𝑁/2.
➢ For large 𝑁, the computation is ~half of that required to evaluate 𝑥[𝑛] directly.
➢ Further reductions can be obtained if we split 𝑋𝑒[𝑘] and 𝑋𝑜[𝑘] again.
 Fast Fourier transform
(cont.)

B
i
t

r
e
v
e
r
s
a
l