Sampling

What is really happening with sampling?

r∗ (t) r(t)
d @ 

Recall that sampling maps strips of the s-plane onto the z-plane.
s-plane Imaginary z-plane Imaginary

π/T Mapping via

sampling
-1 1
Real Real
−π/T

What does this look like from a Fourier Transform point of view?

Roy Smith: ECE 147b 6: 1

Sampling
Model of the sampling process
Continuous-time signal: r(t) Model sampling as time-domain multiplication by
a train of impulses.
r∗ (t) = r(t)v(t)
X∞
= r(t)δ(t − kT )
time: t
k=−∞

Taking the Laplace transform,

Sampling impulse train: v(t)

Z ∞
∗
L {r (t)} = r∗ (t)e−st dt
0 T 2T 3T 4T 5T 6T 7T time: t −∞
Z ∞ ∞
X
= r(t)δ(t − kT )e−st dt
Sampled signal: r(t)v(t) −∞ k=−∞

∞
X
= r(kT )e−skT
k=−∞

0 T 2T 3T 4T 5T 6T 7T time: t =: R∗ (s)

Roy Smith: ECE 147b 6: 2

 Sampling
Impulse trains
Fourier series representation

Sampling impulse train

(time-domain) Impulse trains are periodic (with period T ) and so
1 have a Fourier series:
∞ ∞
Cn ejn( T )t
X X 2π
v(t) = δ(t − kT ) =
k=−∞ n=−∞
-3T -2T -T 0 T 2T 3T
time: t The Fourier coefficients, Cn , are given by,
∞
1 T /2 X
Z
δ(t − kT )e−jn( T )t dt
2π
Sampling impulse train Cn =
T −T /2
(frequency domain) k=−∞
2π/T T /2
1
Z
δ(t)e−jn( T )t dt
2π
=
T −T /2
T /2
1
Z
−2ωs -ωs 0 ωs 2ωs = δ(t)dt
frequency: ω
T −T /2

1
=
T

Roy Smith: ECE 147b 6: 3

Sampling
Impulse trains
Fourier series representation:
∞
X
Sampling impulse train v(t) = δ(t − kT )
(time-domain) k=−∞
1
∞
1 X jn( 2πT )t
= e
T n=−∞
∞
-3T -2T -T 0 T 2T 3T 1 X jnωs t
time: t = e
T n=−∞

Sampling impulse train Fourier Transform:

(frequency domain)
∞
2π/T 2π X
v(s) = δ(ω − nωs )
T n=−∞

(note: ωs = 2π/T ).
−2ωs -ωs 0 ωs 2ωs
frequency: ω
A train of impulses is equivalent to an infinite sum
of (equal) sinusoids.

Roy Smith: ECE 147b 6: 4

 Sampling
Spectrum of a sampled signal

Z ∞
R∗ (s) = L {r∗ (t)} = r∗ (t)e−st dt
−∞
∞
!
∞
1 X jnωs t −st
Z
= r(t) e e dt
−∞ T n=−∞
∞ Z ∞
1 X
= r(t)ejnωs t e−st dt
T n=−∞ −∞
∞ Z ∞
1 X
= r(t)e−(s−jnωs )t dt
T n=−∞ −∞
∞
1 X
= R(s − jnωs ).
T n=−∞

The spectrum of the sampled signal is an infinite sum of shifted versions of the original
spectrum.

Roy Smith: ECE 147b 6: 5

Sampling
Spectrum of a sampled signal
Nyquist
|R(jω)| frequency

−2π/T −π/T 0 π/T 2π/T

frequency: ω

Sampling
Aliased signal components
|R (jω)|
* frequency

−3π/T −2π/T −π/T 0 π/T 2π/T 3π/T

frequency: ω

Aliasing: If |R(jω)| =
6 0 for |w| > π/T then the shifted parts of the spectrum overlap.

Roy Smith: ECE 147b 6: 6

 Sampling
Spectrum of a sampled signal

Sampling impulse train: v(t) Sampling impulse train, v(jω)

(time-domain) (frequency domain)
1
2π/T

-3T -2T -T 0 T 2T 3T −2ωs -ωs 0 ωs 2ωs

time: t frequency: ω

∞ ∞ ∞
X 1 X jnωs t FT 2π X
v(t) = δ(t − kT ) = e ⇐⇒ v(s) = δ(ω − nωs )
T n=−∞ T n=−∞
k=−∞

FT
1
Multiplication/convolution duality: r∗ (t) = r(t)v(t) ⇐⇒ R∗ (jω) = 2π R(jω) ∗ v(jω)

Sampling impulse train, v(jω)

(frequency domain)
|R(jω)|
2π/T

∗
−2π/T −π/T 0 π/T 2π/T
−2ωs -ωs 0 ωs 2ωs
frequency: ω
frequency: ω

Roy Smith: ECE 147b 6: 7

Sampling
Spectrum of a sampled signal
Frequency domain convolution
|R(jω)|
1
R∗ (jω) = R(jω) ∗ v(jω)
2π
Z ∞ ∞
−2π/T −π/T 0 π/T 2π/T 1 2π X
frequency: ω = δ(ν − jnωs )R(jω − ν)dν
2π −∞ T n=−∞
Sampling impulse train, v(jω)
(frequency domain) ∞ Z ∞
1 X
2π/T = δ(ν − jnωs )R(jω − ν)dν
T n=−∞ −∞
∞
1 X
−2ωs -ωs 0 ωs 2ωs = R(jω − jnωs )
frequency: ω
T n=−∞

|R*(jω)|

−3π/T −2π/T −π/T 0 π/T 2π/T 3π/T

frequency: ω

Roy Smith: ECE 147b 6: 8

 Sampling
Aliasing, what can we do?

We have seen that if R(jω) 6= 0 for |ω| > π/T then the signal will be aliased.

The frequency components of the original signal cannot be deduced from the sampled signal.

The control system will react (incorrectly) to aliased errors and disturbances.

Options

1. Sample faster. T is decreased and ωs is increased.

– This costs more money, and may also degrade the resolution.
2. Include a low-pass anti-aliasing filter to remove the frequency components |ω| > π/T .
– We lose all information about the higher frequency components.
– In a closed-loop system the extra phase lag due to F (s) degrades (or even
destabilizes) the closed-loop operation.
– Signal reconstruction is not the most critical part of closed-loop operation.

Roy Smith: ECE 147b 6: 9

Sampling
Aliasing: The effects of an anti-aliasing filter: F (s)

rF (k) rF (t) r(t)

 @
@  F (s) 
T

|F(jω)| 1

|F(jω)R(jω)| |R(jω)|

−π/T 0 π/T
frequency: ω

In control systems the phase effects of F (s) are usually much more important (i.e. potentially
destabilizing).

Roy Smith: ECE 147b 6: 10

 Sampling
Reconstruction

How do we recover r(t) from the sampled signal, r∗ (t)?

Reconstruction Filter

y(k) u(k) u(t)

  @
@ 
L(s)
T

Ideal case:

1, |ω| < π/T
L(jω) =
0, |ω| ≥ π/T

T
But the impulse response of L(s) is: sin(πt/T ) = sinc(πt/T )
πt
This is acausal and unstable!

Roy Smith: ECE 147b 6: 11

Sampling
Reconstruction

Sampling
L(jω) 1 frequency
|R *(jω)|

−3π/T −2π/T −π/T 0 π/T 2π/T 3π/T

frequency: ω

|R(jω)| Nyquist
frequency

−2π/T −π/T 0 π/T 2π/T

frequency: ω

Roy Smith: ECE 147b 6: 12

 Sampling
How good is a ZOH at reconstruction?

The ZOH is the cheapest and most readily available reconstruction filter. How good is it?

The ZOH frequency response (for the fundamental frequency only) is given by:

1 − e−jωT
ZOH(jω) =
jω
 jωT /2
− e−jωT /2 2j

−jωT /2 e
= e
2j jω
sin(ωT /2)
= T e−jωT /2
ωT /2
= T e−jωT /2 sinc(ωT /2)

So the ZOH looks something like a low pass filter in cascade with a delay of T /2 seconds.

It is not very close to the ideal filter, and it has quite a lot of phase lag.

Roy Smith: ECE 147b 6: 13

Sampling
How good is a ZOH at reconstruction?
1
10 Magnitude

0
10
ZOH(jω)

-1
10
Ideal
-2
10

-3
10
-1 0 1
10 10 10
Frequency [rad/sec]

50
Phase (degrees)
Ideal
0

-50

-100
ZOH(jω)
-15 0

-200

-25 0
-1 0 1
10 10 10
Frequency [rad/sec]

Roy Smith: ECE 147b 6: 14