Lectures on Spectra of Continuous-Time Signals

Principal questions to be addressed:

1. What, in a general sense, is the "spectrum" of a signal?

2. Why are we interested in spectra? ("spectra" = plural of "spectrum")

3. How does one assess the spectrum of a given signal?

All of these are for continuous-time signals. The next part of the course discusses spectra
of discrete-time signals.

Outline of Coverage of the Spectra of Continuous-Time Signal

A. Rough definition of spectrum and motivation for studying spectra.

B. The spectrum of a signal that is a finite sum of sinusoids

C. The spectrum of a periodic signal via Fourier series

D. Spectra of segments of signals.

Note:

• The spectra plays important two roles:

a. Analysis and design: The spectra is a theoretical tool that enables one to
understand, analyze and design signals and systems.
b. System component: The computation and manipulation of spectra is a component of
many important systems.

May 29, 2002 1 DLN -- spectra of cont-time signals

II. Lectures on Spectra of Continuous-Time Signals

Notes:

• Our coverage of spectra goes significantly beyond the coverage in Chapter 3.

• See the list of errata for Chapter 3.

A. Rough definition of spectrum and motivation for studying spectrum

1 . Introduction to the concept of "spectrum"?

Definition:
Roughly speaking, the "spectrum of a signal" indicates how the signal may be
thought of as being composed of sinusoids.
It describes the frequencies, amplitudes and phases of the sinusoids that "sum"
to yield the signal.
The individual sinusoids that sum to give the signal are called "sinusoidal
components".
Alternatively, the spectum describes the distributions of amplitude and phase
vs. frequency of the sinusoidal components.
Since each sinusoid can be decomposed into the sum of two complex
exponentials, the spectrum equivalently indicates how the signal may be
thought of as being composed of complex exponentials.
It describes the frequencies, amplitudes and phases of the complex
exponentials that "sum" to yield the signal.
The individual complex exponentials that sum to give the signal are called
"complex exponential components".
Alternatively, the spectrum describes distribution of amplitude and phase vs.
frequency of the complex exponential components.
Sinusoidal and complex exponential components are also called "spectral
components".

Plotting the spectra

We like to plot and visualize spectra. We plot lines at the frequencies of the
exponential components (at both positive and negative frequencies). The
height of the line is the magnitude of the component. We label the line with
the complex amplitude of the component, e.g. with 2ej.5.
Alternatively, sometimes we make two line plots, one showing the magnitudes
of the components and the other showing the phases. These are called the
"magnitude spectra" and "phase spectra", respectively.

May 29, 2002 2 DLN -- spectra of cont-time signals

 Important note:
"Spectrum" is a broad collective noun, like "economy" or "health" for which
there is no universal mathematically precise definition. Rather as with
economy and health, there are a variety of specific ways to assess the
spectrum of a signal.
For example, to assess the economy, one can measure gnp, average income,
unemployment rate, poverty rate, djia, nasdaq, money supply, ... ).
For example, to assess health one can measure body temperature, heart rate,
blood pressure, blood chemistry, weight, etc.
Similarly, there are a variety of ways to assess the spectrum of a signal. A
limited set will be discussed in this course: principally, Fourier series (FS)
for periodic continuous-time signals, discrete Fourier transform (DFT) for
periodic discrete-time signals. But there will also be some discussion and use
(mainly in the labs) of FS and DFT to assess the spectra of finite segments of
signals. The Fourier transform, which is another important method of
assessing the spectrum of continuous-time signals, will be discussed in EECS
306.

Reasons for decomposing into sinusoids.

It's mainly that sinusoids into linear systems lead to sinusoids. (No other class
of signals has this property.)
This causes the input-output relationship for linear sytems to be particularly
simple for sinusoidal signals.
So representing signals with sinusoids simplifies analysis greatly.
Because analysis is simplified, efficient design methods can be developed.

May 29, 2002 3 DLN -- spectra of cont-time signals

 2. Why are we interested in spectrum?

Here are some reasons:

• Signals with nonoverlapping spectrum do not interfere with one another.
Thus many information carrying signals can be transmitted over a single
communication medium (wire, fiber, cable, atmosphere, water, etc.).
To design such systems, we need to be able to quantitatively determine the
spectrum of signals to be able to assess whether or not they overlap, and if
they do, by how much. Also, we need to able to develop systems (e.g. filters)
that select one signal over another, based on its spectrum.
• Some signals can be recognized based on their spectra, e.g. vowels (Labs
8,9), touchtone telephone key presses, musical notes and chords, bird
songs, whale sounds, mechanical vibration analysis, atomic/molecular
makeup of sun and other stars, etc. To build systems that automatically
recognize such signals, we need to able to quantatively determine the
spectrum of a signal.
• Communication media, e.g. the atmosphere, the ocean, a wire, an optical
fiber, often limit propagation to signals with components only in a certain
frequency range (atmosphere is high frequency, ocean is low frequency, wire
is low frequency, optical fiber is high frequency, but what is considered
"high" or "low" depends on the media. We need to be able to assess the
spectrum of a signal to see if it will propagate. We need to be able to design
signals to have appropriate spectra for appropriate media.
• In many situations, the behavior of many natural and man-made linear
systems is best analyzed in the "frequency domain", i.e. one determines the
behavior in response to sinusoids (or complex exponentials) at various
frequencies, and from this one can deduce the response to other signals. The
previous bullet is a special case of this.
• In many situations, an undesired signal interferes with a desired signal, e.g.
the desired signal might correspond to someone speaking and the undesired
signal might be background noise. We wish to reduce or eliminate the
background signal. In order to be able to reduce or eliminate the background
signal it must have some characteristic that is distinctly different than the
desired signal. Often it happens that the desired and undesired signals have
distinctly different spectra (e.g. the noise has mostly high frequency
components). In such cases, one can design systems, called "filters", that
selectively reduce certain frequency components. These can be used to
reduce the noise while having little effect on the desired signal.
• Many other signals and systems methods are based on spectra: e.g. control
engineering, data compression, voice recognition, music processing.
• And ....

May 29, 2002 4 DLN -- spectra of cont-time signals

 3. How does one assess the spectrum of a given signal?

The remainder of these notes are intended to make progress on this question,
with occasional references to questions 1 and 2.

There is no single answer.

The answer/answers do not fit into one course.
We address this question in spiral fashion in EECS 206. The answer continues
in EECS 306 and beyond. (Just like you don't learn all there is to know about
the economics in Econ. 101)
We will develop several methods for continuous-time signals, several methods
for discrete-time signals.
There is no universal spectral concept in wide use.
We use different measures of the spectrum for different types of signals.
We will discuss mainly
1. spectra of a sum of sinusoids (with support (-∞,∞))
2. spectra of periodic signals (with support (-∞,∞)) via Fourier series
and briefly discuss
3. spectra of a segment of a signal via Fourier series, which leads to:
• the spectra of signal with finite support
• the spectra of signal with infinite support via Fourier series applied to
successive segments
We won't discuss
4. spectra of a signal with infinite support and finite energy via Fourier
transform. This will be disucssed in EECS 306.
We will have a similar discussion of spectra for discrete-time signals in the next
part of the course.
We won't get rigorous in our treatment of Fourier series. We'll leave that to
future courses such as EECS 306.

May 29, 2002 5 DLN -- spectra of cont-time signals

 B. The spectrum of a finite sum of sinusoids

As in the text Section 3.1, we begin the discussion of how to assess a spectrum by
considering signals that are finite sums of sinusoids, as in

x(t) = 4 + 3 cos(3t+.1) + 5 cos(7t+.3) - 4 cos (9t+.2)

More generally, consider a signal of the form

N
x(t) = A0 + ∑ Ak cos(2πfkt+φk)
k=1

= A0 + A1 cos(2πf1t+φ1) + A2 cos(2πf2t+φ2) + ...

+ AN cos(2πfNt+φN)

where N, A0, A1, f1, φ1, ..., AN, fN, φN are parameters that specify x(t).
Using Euler's formula, we can rewrite x(t) as
N j2πfkt
x(t) = X0 + ∑ Re { X k e }
k=1

where
jφk
X k = Ak e

is the phasor corresponding to Ak cos(2πfkt+φ). (It's a complex number.)

Using the inverse Euler formula, we can also rewrite this
N  X j2πf t X *k - 

x(t) = X0 + ∑  2 e
k k
+ 2 e j2πfkt
k=1

Finally, we can also rewrite this as

N j2πfkt
x(t) = ∑ αk e
k=-N

where α 0 = X0 = A0

X2k, k≥1
αk = X *
 2k, k≤-1

• The (two-sided) spectrum of this signal is the list of pairs

1 1 * 1 1 * 1 1 *
{(X0,0), (2X 1,f1), (2X1,-f1), ( 2X 2,f2), (2X2 ,-f2 ),..., ( 2X N ,fN ), (2XN ,-fN )}

• We like to plot these, i.e we like to plot the spectrum.

• The spectrum, i.e. this list, is considered to be a "compact" representation of the
signal x(t), i.e. just a few numbers.
• The "spectrum" is also called the "frequency-domain representation" of the
signal. In contrast x(t) is the "time domain representation of the signal".
• The terms Ak cos(2πfkt+φ) are called the "sinusoidal components" of x(t).

May 29, 2002 6 DLN -- spectra of cont-time signals

 j2πfkt
• The terms αk e are called the "complex exponential component" or
"spectral components" of x(t).
• It is equally valid to express the frequencies in Hz as in rad/sec.
• Often we're mainly interested in the magnitude of the spectrum.

Example: Sum of several sinusoids. Let

x(t) = 4 + 3 cos(3t+.1) + 5 cos(7t+.3) -4 cos (9t+.2)

Show a plot of x(t)

Problem: Find the spectrum of x(t).

Solution: To do this, we decompose the x(t) into a sum of complex

exponentials using inverse Euler
j(.1) j3t j(.3) j7t j9t
x(t) = 4 + 1.5 e e + 2.5 e e + 2 ej(.1+π) e
-j(.1) -j3t -j(.3) -j7t -j(.1+π) -j9t
+ 1.5 e e + 2.5 e e +2e e

The spectrum is
j(.1) -j(.1) -j(.3)
{(4,0), (1.5 e ,3), (1.5 e ,-3), (2.5 e ,7),
-j(.3) -j(.1+π) -j(.1+π)
(2.5 e ,-7),(2 e ,9),(2 e ,-9)}

Show a plot of the spectrum.

Compare the plot of the spectrum to the plot of the signal. Notice that the plot
of the spectrum is "simpler, more compact and more intuitively informative"
than the plot of x(t). This illustrates what we mean when we say that "the
spectrum is a compact representation of the signal".

Example: Given that the spectrum of the signal x(t) is shown below, find x(t).

Show plot of a spectrum.

Pick off and sum the exponential components. Simplify so as to write x(t)
as a sum of sinusoids in standard form. Plot x(t).

May 29, 2002 7 DLN -- spectra of cont-time signals

 Example: Amplitude Modulation (AM)
Consider the form of a signal transmitted by an AM radio station

x(t) = (v(t)+1) cos(2πfct)

where
v(t) is the audio signal, which is scaled so that -1 ≤ v(t) ≤ 1 for all t
cos(2πfct) is the "carrier signal",
2πfc is usually a high freqeuncy, e.g. xxx khz., fc is the frequency in
hz to tune the radio to

Draw block diagram:

v(t) → O
+→ O
X → x(t) → amplifier → antenna
| |
1 cos(2πfct)

Motivation:
Our audio signal is low frequency typically 0 to 5 khz.
Low frequencies don't propagate through the atmosphere.
Need to generate a high frequency signal that "carries" the audio signal.
cos(2πfct) is called the "carrier signal". It has high frequency.
x(t) is obtained by "modulating" the carrier signal by the audio signal
Specifically, x(t) = (v(t)+1) cos(2πfct)
v(t) becomes the envelope of x(t). (adding the +1 insures this

Plot: Show plot of typical v(t), cos(2πfct), and x(t)

Problem:
Assuming v(t) = cos(2πfvt), find and plot the spectrum of x(t)?
(A real radio station is not usually interested in transmitting a sinusoidal
audio signal. The sinusoidal v(t) is just a stand-in for a genuine audio
signal. We're assuming this choice of v(t), because so far it's all that we can
analyze.)

Solution:
The spectrum has components at frequencies fc-fv, fc, fc+fv .
Find the actual components, plot the spectrum, discuss how it depends on fv
and fc, mention the "bandwidth".

Note: This example is intended as a simple example of using the concept of

"spectrum" to do an "analysis".

May 29, 2002 8 DLN -- spectra of cont-time signals

 Example: frequency multiplexing of AM signals (this example uses spectra to
design a frequency multiplexing parameter)
Suppose:
Radio station 1 wants to transmit audio signal v1(t) at carrier frequency fc,1
Radio station 2 wants to transmit audio signal v2(t) at carrier frequency fc,2 > fc,1.
Question:
How far apart must fc,1 and fc,2 be in order that the two transmitted
signals do not interfere with each other?

For concreteness assume:

v1(t) = cos(2πfa,1t), v2(t) = cos(2πfa,2t)
Then
Radio station 1 transmits: x1(t) = (1+v1(t)) cos(2πfc,1t)
Radio station 2 transmits: x2(t) = (1+v2(t)) cos(2πfc,2t)

Solution:
The spectrum of x1(t) has components at frequencies
fc,1-fa,1, fc,1, fc,1+f a,1 .
The spectrum of xc(t) has components at frequency has components at
frequencies
fc,2-fa,2, fc,2, fc,2+fa,2
We need to choose fc,1 and fc,2 so that
fc,1+fa,1 < fc,2-fa,2
i.e. so that
fc,2-fc,1 > fa,1+fa,2

In a practical AM system, the audio signal has spectrum ranging from 0 khz to
+5 khz. In fact they limit the audio signals to this range. So the AM radio
signal has "bandwidth" about 10khz --- from fc-5000 to fc+5000. Because of
this, AM radio stations are assigned frequencies increments of 10 khz. And the
FCC avoids having two stations in the same area being separated only by 10 khz.
This is because the limited audio signals don't actually have spectra fitting
exactly between 0 and +5 khz. And because even if they did, a radio receiver
cannot pick out the signal components in the range fc-5000 to fc+5000 without
also accepting at a least some signal components outside this band. Sometimes
you can hear two AM radio stations at once, especially if you've tuned to a weak
one and a powerful one is at transmitting at a frequency only 10 khz away,
especially if you have an old/cheap radio receiver.

Note: This example is intended to be a concrete example of the practical use of

the concept of spectrum to do a simple design task.

May 29, 2002 9 DLN -- spectra of cont-time signals

 C. The spectrum of a periodic signal

The main point of this section is the following theorem, which we won't prove, but
which we will illustrate and use.

Fourier Series Theorem: (Fourier, 1768-1830, French mathematician and

Egyptologist, see Oppenheim & Willsky for biosketch)
A periodic signal x(t) with period T can be written as an infinite sum of
sinusoids, all of which have frequencies that are multiples of 1/T.
That is, there are a set of amplitudes and phases (A0,φ0), (A1,φ1), (A2,φ2), ...
such that
∞ 2π
x(t) = A0 + ∑ Ak cos( T kt+φκ)
k=1
Equivalently,
∞  X j 2 πkt X *k -j 2 πkt
∑ 2  T 
k T
x(t) = X0 + e + 2 e
k=1  

where
jφk
X0 = A0 and Xk = Ak e

Eqivalently,
∞ j2 πkt
x(t) = ∑ αk e T
k=-∞

where α 0 = X0 = A0
X2k, k≥1
αk = X *
 2k, k≤-1

Notes:
• This is an amazing, surprising, and deep theorem. (When it was first
discovered, sime important mathematicians and scientists did not believe it.)
• The proof of the theorem is beyond the scope of this class, and EECS 306, too.
• The theorem says that ANY periodic signal can be represented as the sum of
sinusoids. But it may take an infinite number of them.
2π 2π
• Ak cos( T kt+φk) is the sinusoidal component of x(t) at frequency k.
T
2π
• Note that all sinusoids in the above have frequencies that are multiples of T .
• It also says that ANY periodic signal can be represented as a sum of complex
exponentials. (It may take an infinite number.)
j2πkt
• αk e T is the complex exponential component (equivalently, the spectra
2π
component) of x(t) at frequency T k.
• It follows from the theorem that the spectrum of a periodic signal with period T
is concentrated at frequencies
0, ±1/T, ±2/T, ±3/T, ...

May 29, 2002 10 DLN -- spectra of cont-time signals

 or some subset thereof, i.e. x(t) has spectral components only at these
frequencies.
• The frequency 1/T is called the "fundamental" or "first harmonic" frequency,
the frequency k/T is called the "kth-harmonic" frequency. Likewise, the
component at frequency 1/T is called the "fundamental" or "first harmonic"
component, the frequency k/T is called the "kth-harmonic" component.

• The three sums given above are considered to be three forms of the "Fourier
series". (A "series" is an infinite sum.)

• The book introduces the first two forms in Section 3.4 (equation (3.4.1)

• It is most common to use the third form, because it is easier to work with. We'll
primarily use the third form.

• The Ak's, φk's, Xk's and αk's are called "Fourier series coefficients" or just
"Fourier coefficients".

• To "find the spectrum of a periodic signal", we need to find T and we need to

find the Fourier coefficients.

• Here's the formula for the coefficients:

T
1 - j2Tπkt
αk = T ∫ x(t) e dt
0
This is often called the "analysis formula".

• The derivation of the analysis formula is well presented in the new section 3.4.5.
Reading it is strongly recommended.

• Since the integrand is periodic with period T, the integral in the analysis
formula could have limits consisting of any interval of length T.
2π
j kt
• Notice that αk is the correlation of x(t) with e T normalized by 1/T,
which is the energy of one period of the exponential.
• Suggested reading. The discussion of "signal components" at the end of Section
IIIB of "Introduction to Signals" by DLN.

In the terminology of that discussion

2π
j2πkt/T j T kt
αk e is the component of x(t) that is like e

αk measures the similarity of x(t) to the exponential.

There's a similar interpretation that Ak cos(2π

T
kt+φk) is the component of
x(t) that is like a cosine at frequency 2π
T
.

• Illustrate the summing of sinusoids to obtain an arbitrary signal using the

"sinsum" demo program from Lab 3 .
Can also use the Matlab demo program called "xfourier.m".

May 29, 2002 11 DLN -- spectra of cont-time signals

 Summary of Fourier series

Synthesis Formula: shows how x(t) is a sum of complex exponentials

∞ 2π
j T kt
x(t) = ∑ αk e
k=-∞

Analysis Formula: shows how to compute the αk's, i.e. the Fourier coefficients
T
1 -j2πkt/T
αk = T
0
∫ x(t) e dt

Can alternatively integrate over any T-second interval.

Notes:

• Finding the spectrum of a periodic signal involves finding the period and the
α k's.
• Finding the αk's is often called "taking the Fourier series".
• There is a one-to-one relationship between periodic signals with period T and
sequences of Fourier coefficients.
In the Fourier series theorem, it can be shown that there is one and only one set
of coefficients that works in the synthesis formula, i.e.
there is one and only one set of coefficients {αk} such that
∞ 2π
j T kt
x(t) = ∑ αk e
k=-∞
∞ 2π
j T kt
Thus, if you find a set of coefficients αk such that x(t) = ∑ αk e .
k=-∞
Then these are necessarily the Fourier coefficients.
∞ j2Tπkt ∞ j 2Tπkt
Equivalently, if ∑ αk e = ∑ βk e , then α k = β k .
k=-∞ k=-∞
This means that in some cases, the Fourier coefficients can be found by
inspection.
Another statement of the one-to-oneness is that if x1(t) and x2(t) are distinct
signals1, each with period T, then for at least one k, αk for x1(t) does not
equal α k for x2(t).
• If a signal has period T, then it also has period 2T. So when applying Fourier
analysis, we have a choice as to T. Often, but certainly not always, we choose
T to equal the fundamental period. When we want to explicitly specify the
value of T used, we will say "the T-second Fourier series".
• If you wish to find the other forms of the Fourier series, use the formulas:
A 0 = α 0, Ak = |α k|, φk = angle(α k), k = 1, 2, ...
X 0 = α 0, Xk = 2α k, k = 1, 2, ...

1Here, "distinctt" means that their difference has nonzero power.

May 29, 2002 12 DLN -- spectra of cont-time signals

 Example: Find the spectrum of the following signal.
x(t) = some periodic signal like a square or sawtooth wave
find a closed form expression for the coefficients

Example: Find the spectrum of the following example

x(t) = finite sum of sinusoids.
in this example one computes the spectrum (i.e. the Fourier series
coefficients) by inspection just as we did in the section on finite sums of
sinusoids. The one-to-one-ness of the relation between Fourier coefficients
and periodic signals means that the coefficients we obtain by inspection are
the Fourier series coefficients.

Example: Find the signal corresponding to the following spectrum.

Show a spectrum with finite number of spectral lines. This is the same sort
of problem as in the section on finite sums of sinusoids section.

Example: Show a real-world nearly periodic signal, like a vowel.

Show its spectrum, as computed by a computer.

More properties

• α0 = DC value (important)
• One can compute the Fourier coefficients by integrating over any time
interval of length T:
T
1 -j2πkt/T
αk = T ∫ x(t) e dt
0
a+T
1 -j2πkt/T
= T ∫ x(t) e dt for any value of a (important)
a
1 -j2πkt/T
= T ∫ x(t) e dt (new notation)
<T>

• Conjugate symmetry (important)

*
α-k = α k

Derivation:
T T
* 1 -j2πkt/T  * 1 j2πkt/T
αk = T ∫ x(t) e dt = T ∫ x*(t) e dt
 0  0
T
1 j2π(-k)t/T
= T ∫ x(t) e dt because x*(t) = x(t)
0
= α -k
This property does not apply to complex signals.

• Conjugate pairs of coefficients synthesize a sinusoid (this is also important)

j2πkt -j2πkt 2π
αk e T + α-k e T = 2 |αk| cos( T k t + angle(αk))

May 29, 2002 13 DLN -- spectra of cont-time signals

 Thus when looking at a spectrum one "sees" cosines --- one for every
conjugate pair of coefficients.

Derivation:
j2πkt -j2πkt j2πkt 2π
* -j T kt
αk e T + α-k e T = αk e T + αk e by the previous property
j2πkt  j2πkt *
= αk e T + α k e T 
 
j2πkt
= 2 Re(αk e T )
2π
= 2 |αk| cos( T k t + angle(αk))

• Linearity: Suppose x(t) and y(t) are periodic with period T and with αk
and βk as their T-second Fourier coefficients, respectively. Then the T-
second Fourier coefficients of x(t) + y(t) are αk+βk. (useful)

• Parseval's theorem: (useful but not quite as critical)

T
1 2
signal power = T
0
∫x (t)| dt

∞ 2
= ∑ |α k |
k=-∞

Recall: The power of a periodic signal x(t) is

1 S 2 1 T 2
P(x) = lim 2S ∫ x (t) dt = T ∫ x (t) dt
-S 0
S→∞

Though useful, the remaining properties will not be emphasized in this class.
• Approximating the Fourier synthesis formula with a finite number of terms,
i.e.
N j2Tπkt
x(t) = ∑ αk e
k=-N
For practical reasons, this is necessary in many cases. It can be shown that
the difference signal
N 2π
j T kt
e(t) = x(t) - ∑ αk e
k=-N
∞ 2
has power 2 ∑ |α k | which goes to zero as N increases. If possible, we
k=N+1
choose N so large that this is small.

• Time shifting: If x(t) has Fourier coefficients αk, then x'(t) = x(t-to) has
Fourier coefficients
' -j2πkto
αk = α k e T

This shows, not surprisingly, that a time shift causes a phase shift of each
spectral component, where the phase shift is proportional to the frequency of
the component.

May 29, 2002 14 DLN -- spectra of cont-time signals

 • Frequency shifting: If x(t) has Fourier coefficients αk, then x'(t) = x(t)
2π
j T kot
e has Fourier coefficients
'
αk = αk-ko .

This shows that multiplying a signal by a complex exponential has the effect
of shifting the spectrum of the signal.
• Time scaling: Let a>0. If x(t) is periodic with period T with T-second
Fourier coefficients αk, then x'(t) = x(at) is periodic with period T/a and
T/a-second Fourier coefficients
'
αk = α k .
This shows that the Fourier coefficients are not affected by a time scaling.
However, a time scaling does affect the spectrum. Specifically, the Fourier
coefficients of x(t) are spaced at intervals of 1/T hz, whereas the Fourier
coefficients of x'(t) are spaced at intervals of a/T hz. For example, if a>1,
then the Fourier coefficients are more widely spaced, and consequently, the
spectrum of x'(t) is expanded towards higher frequencies. This is consistent
with the fact that making a > 1, means that x'(t) fluctuates more rapidly
than x(t).
• Technicalities (mostly a warning that there are such)
In order that the integral in the anslysis formula be well defined and thta the
synthesi formula holds, one needs to assume
T T
2
0
∫ |x(t)| dt < ∞ and/or ∫ |x(t)|
0
dt < ∞

When mathematicians prove

∞ j2Tπkt
x(t) = ∑ αk e ,
k=-∞
what they really show is that the power in the difference signal
∞ j2Tπkt
T
2
e(t) = x(t) - ∑ αk e is zero, assuming ∫ |x(t)| dt < ∞.
k=-∞ 0
∞ j2Tπkt
So x(t) and ∑ αk e can differ at a isolated points.
k=-∞
T
Moreover, assuming ∫ |x(t)| dt < ∞ and the so-called "Dirichlet
0
conditions2" the only points at which they can differ are points of
discontinuity in x(t).
∞ 2π
j T kt
Specifically, ∑ αk e = x(t) if x(t) is continuous at t.
k=-∞
∞ j2 πkt 1
and ∑ αk e T = 2 (x(t-)+x(t+)) if x(t) is discontinuous at t.
k=-∞

There's more discussion of "technicalities" in EECS 306.

T
2Dirichlet conditions: In addition to ∫ |x(t)| dt < ∞, in any one period x(t) has only a finite number of maxima and minimum
0
and only a finite number of discontinuities.

May 29, 2002 15 DLN -- spectra of cont-time signals

 C. The spectra of segments of a signal

Motivating question: What is the spectra of a signal that is not periodic?

For example, what if the signal has finite support? Or what if the signal has infinite
support, but is not periodic?

Observation: The Fourier series analysis formula works with a finite segment of a
signal.

An approach to assessing the spectra of a signal with finite support:

If the signal has finite support [t1,t2] apply the analysis formula
t2
1 -j2πkt/(t2-t1)
t2-t1 ∫
αk = x(t) e dt
t1
Now let
~ ∞ j2Tπkt
x(t) = ∑ αk e , where T = t2-t1.
k=-∞
~
Then x(t) is periodic with period T, and
~
x(t) = x(t) when t1 ≤ t ≤ t2

Therefore,
∞ j2Tπkt
x(t) = ∑ αk e , t1 ≤ t ≤ t 2
k=-∞

which is a synthesis formula for x(t) that works only for the support interval
t 1 ≤ t ≤ t2 .
~
x(t) is called the "periodic extension" of x(t).

Note: We have considered Fourier series to fundamentally apply to periodic

signals and secondarily apply to signals with finite support. The reverse point of
view is also valid.

An approach to assessing the spectrum of a periodic signal with infinite support:

If the signal has infinite support, choose a time interval length T, divide the
support interval into segments of length T, as in [0,T], [T,2T], [2T,3T], ....
Apply the previous approach to each segment. We obtain a sequence of spectra,
one for each segment. Notice that with this approach the spectra varies with
time. There are lots of issues here --- for example, what choice of T?

May 29, 2002 16 DLN -- spectra of cont-time signals

Homework Problems:

Problems from Chapter 3

Problems from the Chapter 3 section of the CD ROM

Additional questions about spectra and Fourier series

1. Find the spectrum of the signal

x(t) = A cos(2t) sin(3t)

2. Find the To-second Fourier coefficients of the following signals, where To is the
fundamental period of the signal.
(a) x(t) = t, 0 ≤ t ≤ To
(b) x(t) = |sin(t)|
(c) x(t) = sin2(t)

3. Consider the signal x(t) whose spectrum is shown below. (All terms are real.)
3

1 1 1

ω rad/sec
-40 -30 -20 0 30 40

(a) Is the signal periodic? If so, find its fundamental period.

(b) Find its DC value.
(c) Find its power.
(d) Find x(t). (Express the answer as a sum of sinusoids in standard form.)

4. The nonnegative frequency portion of the spectrum of a signal x(t) is shown below

3
jπ/2
2e
-jπ/3
e
ω rad/sec
10 30

(a) Is the signal periodic? If so, find its fundamental period.

(b) Find its DC value.
(c) Find its power.
(d) Find the negative frequency portion of the spectrum of x(t).
(e) Find x(t). (Express the answer as a sum of sinusoids in standard form.)

5. The nonnegative frequency portion of the spectrum of a signal x(t) is shown below
-jπ/3
2e
1

ω rad/sec
4π 15

(a) Is the signal periodic? If so, find its fundamental period.

(b) Find its DC value.

May 29, 2002 17 DLN -- spectra of cont-time signals

 (c) Find its power.
(d) Find the negative frequency portion of the spectrum of x(t).
(e) Find x(t). (Express the answer as a sum of sinusoids in standard form.)

6. Find the power of the signal

x(t) = 4 + 3 cos(3t+.1) + 5 cos(7t+.3) - 4 cos (9t+.2)

7. Do the signals x(t) and y(t) given below have overlapping spectra? That is, is there a
spectral component of one whose frequency lies on top of one spectral component or
between the frequencies of two spectral components of the other?

x(t) = cos(20t) + cos(22t) + cos(23t)

y(t) = 2 + cos(10t)cos(11t)

8. Let x(t) be a periodic signal with fundamental period To, and let αk be the
To-second Fourier coefficients of x(t). Suppose we also calculate the 2To-second
2 2
Fourier coefficients, denoted αk. Derive an expression for α k in terms of α k.

9. (a) Show that if x(t) is an even periodic function, i.e. if x(-t) = x(t), then all of its
Fourier coefficients are real.
(a) Show that if x(t) is an odd periodic function, i.e. if x(-t) = -x(t), then all of its
Fourier coefficients are imaginary.

10. Let x(t) and y(t) be sinusoids with different frequencies and support (-∞,∞).
Show that the power of x(t)+y(t) equals the sum of the powers of x(t) and y(t).
(Do not assume that x(t) and y(t) are such that x(t)+y(t) is periodic. This fact
implies that Parseval's theorem can be used to compute the power of a sum of
sinusoids, even if the sum is not periodic.)

11. Derive the linearity property of Fourier series: Suppose x(t) and y(t) are periodic
with period T and with αk and βk as their T-second Fourier coefficients,
respectively. Then the T-second Fourier coefficients of x(t) + y(t) are αk+βk.

12. Derive the time-shifting property of Fourier series: If x(t) has Fourier coefficients
αk, then x'(t) = x(t-to) has Fourier coefficients
2π
' -j T kto
αk = α k e
13. Derive the frequency-shifting property of Fourier series: If x(t) has Fourier
j2πkot
coefficients αk, then x'(t) = x(t) e T has Fourier coefficients
'
αk = αk-ko .
14. Derive the time-scaling property of Fourier series: Let a>0. If x(t) is periodic
with period T with T-second Fourier coefficients αk, then x'(t) = x(at) is periodic
with period T/a and T/a-second Fourier coefficients
'
αk = α k .

May 29, 2002 18 DLN -- spectra of cont-time signals