Signals and Systems

DR TANIA STATHAKI
READER (ASSOCIATE PROFESSOR) IN SIGNAL PROCESSING
IMPERIAL COLLEGE LONDON
 Continuous time versus discrete time
• Continuous time systems.
 Good for analogue signals and general understanding of signals and systems.
 Appropriate mostly to analogue electronic systems.

𝑥(𝑡) Electronic devices 𝑦(𝑡)

• Electronic devices are increasingly digital.

 E.g. mobile phones are all digital, TV broadcast will be 100% digital in UK.
 We use digital ASIC chips, FPGAs and microprocessors to implement systems
and to process signals.
 Continuous signals are converted to numbers (discrete signals), they are
processed and then converted back to continuous signals.

𝑥 𝑡 𝑥[𝑛] 𝑦𝑛 𝑦(𝑡)
Digital
A-to-D D-to-A
Signal Processing
 Sampling

• The sampling process converts continuous signals 𝑥(𝑡) into a sequence of

numbers 𝑥[𝑛].
• A sample is kept every 𝑇𝑠 units of time. This process is called uniform
sampling and 𝑥 𝑛 = 𝑥(𝑛𝑇𝑠 ).
𝑥(𝑡) 𝑇𝑠 = 1/𝑓𝑠 𝑥[𝑛]
 Sampling theorem

• Sampling theorem is the bridge between continuous-time and discrete-

time signals.
• It states how often we must sample in order not to loose any information.
𝑥(𝑡) 𝑇𝑠 = 1/𝑓𝑠 𝑥[𝑛]

Sampling theorem

A continuous-time lowpass signal 𝑥(𝑡) with frequencies no higher than

𝑓𝑚𝑎𝑥 𝐻𝑧 can be perfectly reconstructed from samples taken every 𝑇𝑠 units
of time, 𝑥 𝑛 = 𝑥(𝑛𝑇𝑠 ), if the samples are taken at a rate 𝑓𝑠 = 1/𝑇𝑠 that is
greater than 2𝑓𝑚𝑎𝑥 𝐻𝑧.
 Sampling theorem cont.

• If a lowpass signal has a spectrum bandlimited to 𝐵𝐻𝑧, i.e., 𝑋 𝜔 = 0 for

𝜔 > 2𝜋𝐵, it can be reconstructed from its samples without error if these
samples are taken uniformly at a rate 𝑓𝑠 > 2𝐵 samples per second.

• The minimum sampling rate 𝑓𝑠 = 2𝐵 required to reconstruct 𝑥(𝑡) from its

samples is called the Nyquist rate for 𝑥(𝑡) and the corresponding
1
sampling interval 𝑇𝑠 = is called the Nyquist interval. Samples of a
2𝐵
continuous signal taken at its Nyquist rate are the Nyquist samples of that
signal.

• In other words the minimum sampling frequency is 𝑓𝑠 = 2𝐵𝐻𝑧.

• A bandpass signals whose spectrum exists over a frequency band

𝐵 𝐵
𝑓𝑐 − < 𝑓 < 𝑓𝑐 + also has a bandwidth of 𝐵𝐻𝑧. Such a signal is still
2 2
uniquely determined by 2𝐵 samples per second but the sampling scheme
is a bit more complex compared to the case of a lowpass signal.
 Sampling theorem: mathematical proof

• Consider a signal bandlimited to 𝐵𝐻𝑧 with Fourier transform 𝑋(𝜔).

• The sampled version of the signal 𝑥(𝑡) at a rate 𝑓𝑠 𝐻𝑧 can be expressed as
the multiplication of the original signal with an impulse train as follows:
𝑥 𝑡 = 𝑥 𝑡 𝛿𝑇𝑠 𝑡 = 𝑛 𝑥(𝑛𝑇𝑠 )𝛿(𝑡 − 𝑛𝑇𝑠 ), 𝑇𝑠 = 1/𝑓𝑠
• We can express the impulse train using Fourier series as follows:
1 2𝜋
𝛿𝑇𝑠 𝑡 = [1 + 2cos𝜔𝑠 𝑡 + 2cos2𝜔𝑠 𝑡 + 2cos3𝜔𝑠 𝑡 + ⋯ ], 𝜔𝑠 = = 2𝜋𝑓𝑠
𝑇𝑠 𝑇𝑠

• Therefore,
1
𝑥 𝑡 = 𝑥(𝑡)𝛿𝑇𝑠 𝑡 = [𝑥(𝑡) + 2𝑥(𝑡)cos𝝎𝒔 𝑡 + 2𝑥(𝑡)cos𝟐𝝎𝒔 𝑡 + 2cos𝟑𝝎𝒔 𝑡 + ⋯ ]
𝑇𝑠
1
• Since the following holds: 𝑥 𝑡 cos𝜔𝑠 𝑡 ⇔ [𝑋 𝜔 + 𝜔𝑠 + 𝑋 𝜔 − 𝜔𝑠 ]
2
we have
1 ∞
𝑋 𝜔 = 𝑛=−∞ 𝑋 𝜔 − 𝑛𝜔𝑠
𝑇𝑠
 Depiction of previous analysis

• The previous analysis is depicted below. Consider a signal, bandlimited

to 𝐵𝐻𝑧, with Fourier transform 𝑋(𝜔) (depicted real for convenience).

• The sampled signal has the following spectrum.

 Depiction of previous analysis

• We graphically illustrate below the collection of the above mentioned

processes.
 The signal is multiplied by a train of impulses (in reality these are
very narrow pulses).
 The sampled signal is generated.
 A lowpass filter is required in order to isolate the main period of the
Fourier spectrum.

𝑥(𝑡) 𝑇𝑠 = 1/𝑓𝑠 𝑥[𝑛]

 Depiction of previous analysis cont.

• A continuous signal with its Fourier transform is depicted on the left.

• The sampled signal and its Fourier transform is depicted on the right.
 Note that, for simplicity, the Fourier transform is considered to
be real in this case.
 Reconstruction of the original signal

• The gap between two adjacent spectral repetitions is (𝑓𝑠 −2𝐵)𝐻𝑧.

• In order to reconstruct the original signal 𝑥(𝑡) we can use an ideal
lowpass filter on the sampled spectrum which has a bandwidth of any
value between 𝐵 and (𝑓𝑠 − 𝐵𝐻𝑧).

• Reconstruction process is possible only if the shaded parts do not

overlap. This means that 𝑓𝑠 must be greater that twice 𝐵.
• We can also visually verify the sampling theorem in the above figure.
 Reconstruction generic example

• The signal 𝑥 𝑛𝑇𝑠 has a spectrum 𝑋𝑠 (𝜔) which is multiplied with a

rectangular pulse in frequency domain in order to isolate the main
period which is the spectrum of the original continuous time signal.
▪ Note that, for simplicity, the Fourier transform is considered to
be real in this case.

𝑋𝑠 (𝜔)
 Reconstruction generic example cont.

• The signal 𝑥 𝑛𝑇𝑠 is convolved with a sinc function, which is the time
domain version of a rectangular pulse in frequency domain centred at
the origin.
𝑥 𝑛𝑇𝑠 ∗ sinc 𝑎𝑡 = 𝑛 𝑥[𝑛] sinc[𝑎(𝑡 − 𝑛𝑇𝑠 )]

𝑥 𝑛𝑇𝑠 sinc 𝑎𝑡
 Example

• Consider the signal 𝑥 𝑡 = sinc 2 5𝜋𝑡 .

𝑊
• The property ( )sinc 2 𝑊𝑡/2 ⇔ Δ(𝜔/2𝑊) holds with:
2𝜋
1 0.2
0 𝑥 ≥
2
• Δ 𝑥 = 1
1−2 𝑥 𝑥 <
2
• Using the above property with
𝑊 = 10𝜋 we obtain
𝑋 𝜔 = 0.2Δ(𝜔/20𝜋)
• The bandwidth is 𝐵 = 5𝐻𝑧(10𝜋𝑟𝑎𝑑/𝑠).
• Consequently, the Nyquist sampling
rate is 𝑓𝑠 = 10𝐻𝑧;
we require at least 10 samples
per second. −10𝜋 10𝜋
1
• The Nyquist interval is 𝑇𝑠 = = 0.1𝑠.
10
 Example cont.
Nyquist sampling: Just about the correct sampling rate
• In that case we use the Nyquist sampling rate of 10𝐻𝑧.
• The spectrum 𝑋 𝜔 consists of back-to-back, non-overlapping
1
repetitions of 𝑋 𝜔 repeating every 𝑓𝑠 = 2𝐵 = 10𝐻𝑧.
𝑇𝑠
• In order to recover 𝑋(𝜔) from 𝑋 𝜔 we must use an ideal lowpass filter
of bandwidth 5𝐻𝑧. This is shown in the right figure below with the dotted
line.
• Obviously, it is not possible to design and implement such a filter.
 Example cont.
Oversampling: What happens if we sample too quickly?
• Sampling at higher than the Nyquist rate (in this case 20𝐻𝑧 ) makes
reconstruction easier.
1
• The spectrum 𝑋 𝜔 consists of non-overlapping repetitions of 𝑋 𝜔 ,
𝑇𝑠
repeating every 20𝐻𝑧 with empty bands between successive cycles.
• In order to recover 𝑋(𝜔) from 𝑋 𝜔 we can use a practical lowpass filter
and not necessarily an ideal one. This is shown in the right figure below
with the dotted line.

• The filter we use for reconstruction must have gain 𝑇𝑠 and bandwidth of
any value between 𝐵 and (𝑓𝑠 − 𝐵)𝐻𝑧.
 Example cont.
Undersampling: What happens if we sample too slowly?

• Sampling at lower than the Nyquist rate (in this case 5𝐻𝑧) makes
reconstruction impossible.
1
• The spectrum 𝑋 𝜔 consists of overlapping repetitions of 𝑋 𝜔
𝑇𝑠
repeating every 5𝐻𝑧.
• 𝑋(𝜔) is not recoverable from 𝑋 𝜔 .
• Sampling below the Nyquist rate corrupts the signal. This type of
distortion is called aliasing.
 Aliasing

• Consider what happens when a 1𝐻𝑧 and a 6𝐻𝑧 sinewaves are sampled
at a rate of 5Hz.

• The 1𝐻𝑧 and 6𝐻𝑧 sinewaves are indistinguishable after sampling. The
two discrete signals produced are identical.
 Aliasing cont.

• The two original signals are shown together with the sampled values. As
verified, the sampled values are not efficient to recover the original
shape of the high frequency signal (shown in blue).
 Aliasing example (recall Signals and Systems Lab)

• Consider a signal bandlimited to 𝐵𝐻𝑧 with Fourier transform 𝑋(𝜔). The sampled
version of the signal 𝑥(𝑡) at a rate 𝑓𝑠 𝐻𝑧 can be expressed as the multiplication of the
original signal with an impulse train as follows:
𝑥 𝑡 = 𝑥 𝑡 𝛿𝑇𝑠 𝑡 = 𝑛 𝑥(𝑛𝑇𝑠 )𝛿(𝑡 − 𝑛𝑇𝑠 ), 𝑇𝑠 = 1/𝑓𝑠

• Assume that 𝑥 𝑡 = sin(2𝜋𝑓𝑡). Consider two frequencies 𝑓1 and 𝑓2 which satisfy the
condition −𝑓1 + 𝑓2 = 𝑓𝑠 . This happens in the previous example.

• Take the signal 𝑥 𝑡 = sin(2𝜋𝑓1 𝑡). Its sampled version is

𝑥 𝑡 = 𝑥 𝑡 𝛿𝑇𝑠 𝑡 = 𝑛 𝑥(𝑛𝑇𝑠 )𝛿(𝑡 − 𝑛𝑇𝑠 ), 𝑇𝑠 = 1/𝑓𝑠
𝑥 𝑛𝑇𝑠 = sin 2𝜋𝑓1 𝑛𝑇𝑠 = sin 2𝜋(−𝑓𝑠 + 𝑓2 )𝑛𝑇𝑠 = sin 2𝜋𝑓2 𝑛𝑇𝑠 − 2𝜋𝑓𝑠 𝑛𝑇𝑠

• We know that 𝑓𝑠 𝑇𝑠 = 1. Therefore, the above becomes:

𝑥 𝑛𝑇𝑠 = sin 2𝜋𝑓2 𝑛𝑇𝑠 − 2𝜋𝑛 = sin 2𝜋𝑓2 𝑛𝑇𝑠

• We see that if 𝑓2 = 𝑓𝑠 +𝑓1 the two sinusoidals look the same after sampling.
 Aliasing and the wagon wheel effect

• Consider making a video of a clock face.

1
• The second hand makes one revolution per minute (𝑓𝑚𝑎𝑥 = 𝐻𝑧).
60
1
• Critical Sampling: 𝑓𝑠 > 𝐻𝑧 (𝑇𝑠 < 30sec), anything below that sampling
30
frequency will create problems.
For example:
 When 𝑇𝑠 = 60sec (𝑓𝑠 ~0.167𝐻𝑧), the second hand will not move.
 When 𝑇𝑠 = 59sec ( 𝑓𝑠 ~0.169𝐻𝑧 ), the second hand will move
backwards.

• Watch the videos (optional)

https://www.youtube.com/watch?v=VNftf5qLpiA
https://www.youtube.com/watch?v=QOwzkND_ooU
 Anti-aliasing filter
• To avoid distortion of a signal after sampling, one must ensure that the
𝑓
signal being sampled at 𝑓𝑠 is bandlimited to a frequency 𝐵, where 𝐵 < 𝑠 .
2
• If the signal does not obey the above restriction, we may apply a lowpass
𝑓
filter with cut-off frequency 𝑠 before sampling.
2
Sampling Reconstruction

𝑓
• If the original signal 𝑥(𝑡) is not bandlimited to 𝑠 , perfect reconstruction is
2
not possible when sampling at 𝑓𝑠 . However, the reconstructed signal 𝑥 (𝑡) is
the best bandlimited approximation to 𝑥(𝑡) in the least-square sense.
 Practical sampling (Lathi page 778)
• The impulse train is not a very practical sampling signal.
• In practice we may use a train of pulses 𝑝𝑇 (𝑡) as the one shown below.
 The pulse height is 𝐴 = 1, its width is 𝜏 = 0.025𝑠𝑒𝑐 and the period is
𝑇 = 0.1sec.
• We use the same signal as previously, i.e.,
𝑥 𝑡 = sinc 2 (5𝜋𝑡) with 𝑋 𝜔 = 0.2Δ(𝜔/20𝜋)
• For the train of pulses we use Fourier Series (Lathi, Chapter 6):
2 𝑛𝜋
𝑝𝑇𝑠 𝑡 = 𝐶0 + ∞ 𝑛=1 𝑛𝐶 cos𝑛𝜔 𝑠 𝑡 , 𝐶0 = 1/4, 𝐶𝑛 = sin( )
𝑛𝜋 4
1
• 𝑥 𝑡 = 𝑥 𝑡 𝑝𝑇𝑠 𝑡 = 𝑥𝑡 + 𝐶1 𝑥 𝑡 cos20𝜋𝑡 + 𝐶2 𝑥 𝑡 cos40𝜋𝑡 +
4
𝐶3 𝑥 𝑡 cos60𝜋𝑡 + ⋯
1 𝐶
• 𝑋 𝜔 = 𝑋 𝜔 + 𝟐𝝅{ 1 𝑋 𝜔 − 20𝜋 + 𝑋 𝜔 + 20𝜋 +
4 2
𝐶2 𝐶
𝑋 𝜔 − 40𝜋 + 𝑋 𝜔 + 40𝜋 + 3 𝑋 𝜔 − 60𝜋 + 𝑋 𝜔 + 60𝜋 + ⋯}
2 2
• Lowpass filter can be used to recover 𝑋 𝜔 .
 Practical sampling cont.
𝑥 𝑡 = sinc 2 (5𝜋𝑡) 𝑋 𝜔 = 0.2Δ(𝜔/20𝜋)

×
 Ideal signal reconstruction

• The process of reconstructing a continuous-time signal 𝑥(𝑡) from its

samples is also known as interpolation.
• As already mentioned, the filter we use for reconstruction must have gain
𝑇𝑠 and bandwidth of any value between 𝐵 and (𝑓𝑠 −𝐵)𝐻𝑧.
𝑓𝑠 1 𝜋
• A good choice is the middle value = 𝐻𝑧 or 𝑟𝑎𝑑/𝑠. This gives a filter
2 2𝑇𝑠 𝑇𝑠
with frequency response:
𝜔 𝜔𝑇𝑠
𝐻 𝜔 = 𝑇𝑠 rect = 𝑇𝑠 rect( )
2𝜋𝑓𝑠 2𝜋
• In time domain, the impulse response of the above filter is given by:
𝜋𝑡
ℎ 𝑡 = sinc
𝑇𝑠
1
• For the Nyquist sampling rate, 𝑇𝑠 = ,
2𝐵
ℎ 𝑡 = sinc 2𝜋𝐵𝑡
 Ideal signal reconstruction cont.

• Observe the interesting fact that ℎ 𝑡 = 0

𝑛
for all Nyquist sampling instants 𝑡 = ± ,
2𝐵
except at 𝑡 = 0.

• Each sample in 𝑥 𝑡 , being an impulse, generates a sinc pulse of height

equal to the strength of the sample when is applied at the input of this filter
(recall convolution with an impulse). Addition of all sinc impulses results in
𝑥(𝑡) (look at next slide).

• The filter output is:

𝜋
𝑥 𝑡 = 𝑛 𝑥 𝑛𝑇𝑠 ℎ 𝑡 − 𝑛𝑇𝑠 = 𝑛 𝑥 𝑛𝑇𝑠 sinc (𝑡 − 𝑛𝑇𝑠 )
𝑇𝑠

1
• In the case of Nyquist sampling rate 𝑇𝑠 = . The above equation becomes:
2𝐵
𝑥 𝑡 = 𝑛 𝑥 𝑛𝑇𝑠 ℎ 𝑡 − 𝑛𝑇𝑠 = 𝑛𝑥 𝑛𝑇𝑠 sinc 2𝜋𝐵𝑡 − 𝑛𝜋
This is called the interpolation formula.
 Ideal signal reconstruction cont.

• The above analysis is depicted below.

 Practical signal reconstruction

• Ideal reconstruction system is therefore:

• In practise, we normally sample at higher frequency than Nyquist rate, so

that we don’t have to use an ideal lowpass filter (there isn’t one anyway!).
 Practical signal reconstruction cont.

• The disappointing truth is that any filter that has a frequency response
which is zero above a certain frequency is not realizable (Paley-Wiener
criterion; for more information look at Lathi, Chapter 7, equation 7.61).

• The above implies that in practice it is impossible to perfectly reconstruct

a bandlimited signal from its samples!

• However, a filter with gradual cutoff characteristics is easier to realize

compared to an ideal filter.

• Furthermore, as the sampling rate increases, the recovered signal

approaches the desired signal more closely.
 Scalar quantisation

• The sampling theorem allows us to represent signals by means of

samples.

• In order to store these samples we need to convert the real values 𝑥[𝑛]
into a format which fits the memory model of a computer.

• The process that maps the real line to a countable set is called
quantisation.

• Quantisation is irreversible. This means that quantisation introduces

approximation errors.

• Usually the number of quantisation levels 𝑁 is a power of 2, i.e., 𝑁 = 2𝑅

so that each symbol can be expressed using a stream of bits.
 Scalar quantisation cont.

• Typically each sample is quantized independently (scalar quantisation).

• When the quantisation intervals have equal width we say that
quantisation is ‘uniform’.
 Scalar quantisation cont.

• Sampling followed by quantisation.

 Quantisation examples

• A few numbers (in absence of compression)

 Low-quality audio signal (e.g., voice)
 Frequency from 300𝐻𝑧 to 3400𝐻𝑧, we assume 𝐵 = 4𝑘𝐻𝑧.
 Sampling Frequency 𝑓𝑠 = 2𝐵 = 8𝑘𝐻𝑧.
 Each sample is quantized with a 8-bit quantiser. The signal-to-noise-
ratio due to quantisation error is 𝑆𝑁𝑅 ≈ 48𝑑𝐵.
64𝑘𝑏𝑖𝑡𝑠
 This produces .
𝑠

 High Fidelity Audio Signals (e.g., CD)

 Bandwidth 15𝐻𝑧, we assume 𝐵 = 22.5𝑘𝐻𝑧.
 Sampling frequency 𝑓𝑠 = 2𝐵 = 44.1𝑘𝐻𝑧.
 16-bit quantiser (𝑆𝑁𝑅 ≈ 96𝑑𝐵).
706𝑘𝑏𝑖𝑡𝑠
 This produces ≈ .
𝑠
 Quantisation error

• The figure depicts a 2-bit quantiser.

• Low-rate quantization introduces correlation between the quantisation

error and the original signal.
• This correlation leads to audible artifacts in audio signals or artificial
edges in images.
 Quantisation error cont.

• In the example below the quantized image has been allocated a number
of intensity levels which is not large enough. Therefore, the quantized
image does not look “real” to the human eye; instead it looks like a
segmented image.

Original Quantized

• Images have been taken from http://en.wikipedia.org/wiki/Dither.

For your own interest: Who discovered the sampling theorem?

• The sampling theorem is usually known as the Shannon Sampling

Theorem due to Claude E. Shannon’s paper “A mathematical theory of
communication” in 1948.

• The minimum required sampling rate 𝑓𝑠 (i.e. 2𝐵) is known as the Nyquist
sampling rate or Nyquist frequency because of H. Nyquist’s work on
telegraph transmission in 1924 with K. Küpfmüller.

• The first formulation of the sampling theorem which was applied to

communication is probably due to a Russian scientist by the name of V.
A. Kotelnikov in 1933.

• However, mathematicians already knew about this in a different form

and called this the interpolation formula. E. T. Whittaker published the
paper “On the functions which are represented by the expansions of the
interpolation theory” back in 1915.