Vous êtes sur la page 1sur 9

Lecture 8

White Noise and Power Spectral


Density
8.1

White Noise

White noise is a basic concept underlying the modeling of random disturbances, such as
sensor noise
environmental disturbances
dA
GA

P
GS
dS

In contrast to continuous time, white noise is straightforward to characterize in discrete


time:
Definition (White noise) A noise signal d[n] is white if it has zero mean
E (d[n]) = 0

for all n,

with E (d[n]) being the expected value of the random variable d[n]. We further define white
noise to have unit variance
E (d[n]d[n]) = 1,

for all n,

which may be scaled as appropriate, see the following example with a uniform distribution.
Lastly, a noise signal d[n] is white if it is independent from sample to sample, i.e. not
correlated in time
E (d[n]d[n k]) = 0,

for k 6= 0 and for all n.


1

So far, we have said nothing about the underlying probability distribution. The above is
general and applies to many probability distributions.
Example (Uniform Distribution)
p(x) =

1
ba

axb
otherwise

p(x)

Matlab: The command rand draws uncorrelated pseudorandom numbers from a uniform distribution with a = 0, b = 1.
Zero mean assumption: a = b.
Unit variance assumption:
1
ba
1
2b

x2 dx = 1
a
b

b
x3
b2
=1
x dx =
=
6b b
3
b

= b = 3
2

Therefore, the Matlab command d = 2*sqrt(3)*rand-sqrt(3) draws the pseudorandom number d from a uniform distribution with expected value zero and unit variance.

Example (Normal Distribution)


1 x2
p(x) = e 2
2
0.4

p(x)
0.24

0.3
0.2
0.1

x
4

Matlab: The command randn draws uncorrelated pseudorandom numbers from a normal distribution with mean 0 and variance 1.
Both probability distributions may be used to generate white noise. We often only care
about mean and variance, so the underlying distribution usually does not matter so much.
See the following figure for a time domain example of white noise.

d[n]

-1

3
0

10

15

20

25

30

35

40

45

50

55

60

n
Figure 8.1: Discrete time representation of white noise with uniform distribution.

The analysis of a white noise signal in the frequency domain poses several problems. White
noise
is not periodic. Fourier Series excluded.
has no finite extent. Discrete Fourier Transform excluded.
has infinite energy. Fourier Transform excluded.
Therefore a rigorous way to handle these types of signals in the frequency domain is needed.
A solution is the power spectral density.

8.2

Power Spectral Density Function

We first define the following:


Definition (Auto-Correlation Function) The auto-correlation function is defined as
Rxx [k] = E (x[n]x[n k]) ,
where x[n] is assumed to be wide sense stationary: Rxx [k] does not depend on n (Similar
to time invariance). This is a general definition, and not just for white noise.
When x[n] is white noise,
Rxx [k] = [k],
the unit impulse. We now define the power spectral density function of a signal x[n]:
Definition (Power spectral density function) The power spectral density function is
the Fourier Transform of the auto-correlation function:
FT

Sxx () Rxx [k]

X
Sxx () =
Rxx [k]ejk ,
k=

with on any 2 interval, where < is typical.


When x[n] is white noise,
Sxx () =

[k]ejk = 1.

k=

In Fig. 8.2, we show numerical approximations using finite-length (N = 8192) signals of


the auto-correlation function and power spectral density function of a white noise signal.
The plotted functions were generated using the Matlab script white_noise.m, which you
may find on the course website.
4

Rxx [k]

1
0.5
0
0

10

15

20

25

30

35

40

45

50

55

60

k
Raw
Filtered

Sxx []

3
2
1
0
2

1.5

0.5

0
[rad]

0.5

1.5

2
102

Figure 8.2: Approximated auto-correlation function and power spectral density function
of white noise generated by a normal distribution. For better readability, only 65 of the
total 8192 samples of the respective approximations are plotted. The smoothed power
spectral density function was obtained using a non-causal moving average filter, which will
be covered in future lectures about filtering.

Engineering Examples and Background


The power spectral density function has its name from the fact that in many engineering
systems, power is proportional to the square of a signal.
Example Resistor with current i(t)
+

v(t)

R
i(t)
v(t) = Ri(t)

Power = v(t)i(t) =

v 2 (t)
= i2 (t)R
R

Example Damper with force


x(t)
f (t)
b
f 2 (t)
= x 2 (t)b
b
Incidentally, these simple examples illustrate two powerful design decisions:
f (t) = bx(t)

Power = x(t)f

(t) =

1. High Voltage Lines: 100 000 V with 1 000 000 V coming soon.
i

vs

vd

PS

PD

Source Destination
For a fixed PD = vD i follows vs = vD + iR and
PS = vS i = vD i + i2 R = PD + i2 R,
so its best to make vD really large and i really small.
2. Hydraulics
x

fs

PS

fd
PD

Similar PS = PD +x 2 b. We want really high forces (pressure) and small displacements.


When delivering power, motion is bad.

Power in Frequency Domain


The integral of Sxx () is the expected power of the signal x[n]:
!
Z
Z

X
1
1
jk
Sxx () d =
d
Rxx [k]e
2
2 k=
 Z


X
X

1
jk
=
Rxx [k]
Rxx [k][k] = Rxx [0] = E x2 [n]
e
d =
2
k=
{z
} k=
|
[k]

We may also compute the power contained in a frequency band:


Z 1

Z 2
1
Sxx () d +
Sxx () d
0 1 < 2 .
2
2
1

Some properties of the power spectral density:


Symmetry Sxx () = Sxx (). We only need to consider the range 0 . Consequently, the power in a frequency band is
Z
1 2
Sxx () d
1

Real Sxx () = Sxx


(), where denotes the complex conjugate. That is, if Sxx () = a+bj,
then b = 0. Power is a real quantity.

Non-negative Sxx () 0 if x[n] is real, as the power of a signal must be positive over
any frequency band.

Cross Power Spectral Density Function


The following concept is useful for filtering and system identification, topics that we will
cover in the next lectures.
Definition Given wide sense stationary x[n] and y[n], the cross correlation function is
defined as
Rxy [k] = E (x[n]y[n k]) .
The corresponding Fourier transform results in the cross power spectral density function
FT

Sxy () Rxy [k].


One can show that
Rxy [k] = Ryx [k]
Sxy () = Syx ()
Sxy () is generally complex valued
7

Relation to LTI Systems


Let y[n] = T {x[n]}, T being a linear time invariant system. We want to calculate Syy ()
given Sxx (). Let h[n] be the impulse response of T. Then

y[n] =

l=

h[l]x[n l],

and

Ryy [k] = E (y[n]y[n k]) = E


=

X
X

l= p=

h[l]x[n l]

l=

p=

h[p]x[n k p]

!!

h[l]h[p]E (x[n l]x[n k p])

Introducing a new index m = n l, we obtain


Ryy [k] =

X
X

l= p=

h[l]h[p]E (x[m]x[m + l k p]) .

Comparing to the definition of the auto-correlation function


Rxx [k] = E (x[n]x[n k]) ,
we obtain
Ryy [k] =

X
X

l= p=

But,
Rxy [k] = E (x[n]y[n k]) = E

p=

therefore we obtain
Ryy [k] =

h[l]h[p]Rxx [k + p l].

x[n]h[p]x[n k p]

l=

In terms of convolutions

h[l]Rxy [k l].

Ryy [k] = h[k] Rxy [k]

X
Rxy [k] =
h[p]Rxx [k + p]
=

p=

p=

h[p]Rxx [k p]

= h[k] Rxx [k].


8

p=

h[p]Rxx [k + p],

Applying Fourier transforms, we obtain


Syy () = H()Sxy ()
Sxy () = H()Sxx (),
from which it follows that

Syy () = |H()|2 Sxx (),

since H() = H () and all signals are real. This is a powerful result for the following
topics: filtering and system identification.

Vous aimerez peut-être aussi