EE531 (Semester II, 2010)

6. Input signals

Common input signals in system identification

step function
sum of sinusoids
ARMA sequences
Pseudo random binary sequence

Spectral characteristics
Persistent excitation

6-1

Step function
A step function is given by
u(t) =

0,
u0 ,

t<0
t0

where the amplitude u0 is arbitrarily chosen

The step response can be related to rise time, overshoots, static gain, etc.

It is useful for systems with a large signal-to-noise ratio

Input signals

6-2

Sum of sinusoids
The input signal u(t) is given by
u(t) =

m
X

ak sin(k t + k )

k=1

where the angular frequencies {k } are distinct,

0 1 < 2 < . . . < m
and the amplitudes and phases ak , k are chosen by the user

Input signals

6-3

Characterization of sinusoids
Let SN be the average of a sinusoid over N points
SN

N
1 X
a sin(t + )
=
N t=1

Let be the mean of the sinusoidal function

= lim SN
N

Therefore, u(t) =

(
a sin ,
=
0,

Pm

k=1 ak sin(k t

= 2n, n = 0, 1, 2, . . .
otherwise
+ k ) has zero mean if 1 > 0

Otherwise, we can always subtract the mean from u(t)

WLOG, assume zero mean for u(t)
Input signals

6-4

Spectrum of sinusoidal inputs

The autocorrelation function can be computed by
N
1 X
R( ) = lim
u(t + )u(t)
N N
t=1

m
X

Ck cos(k )

k=1

with Ck = a2k /2 for k = 1, 2, . . . , m

If m = , the coefficient Cm should be modified by
Cm = a2m sin2 m
Therefore, the spectrum is
S() =

m
X

(Ck /2) [( k ) + ( + k )]

k=1
Input signals

6-5

Autoregressive Moving Average sequence

Let e(t) be a pseudorandom sequence similar to white noise in the sense
that
N
1 X
e(t)e(t + ) 0, as N
N t=1
A general input u(t) can be obtained by linear filtering
u(t) + c1u(t 1) + . . . cpu(t p) = e(t) + d1e(t 1) + . . . + dq e(t p)

u(t) is called ARMA (autoregressive moving average) process

When all ci = 0 it is called MA (moving average) process
When all di = 0 it is called AR (autoregressive) process
The user gets to choose ci, di and the random generator of e(t)
Input signals

6-6

ARMA sequence (cont.)

The transfer function from e(t) to u(t) can be written as
D(z)
U (z) =
E(z)
C(z)
where
C(z) = 1 + c1z 1 + c2z 2 + . . . cpz p
D(z) = 1 + d1z 1 + d2z 2 + . . . dq z q
The distribution of e(t) is often chosen to be Gaussian
ci, di are chosen such that C(z), D(z) have zeros outside the unit circle
Different choices of ci, di lead to inputs with various spectral
characteristics
Input signals

6-7

Spectrum of ARMA process

Let e(t) be a white noise with variance 2
The spectral density of ARMA process is

D(z)
C(z)


2


2  D() 
S() = 
C() 

D(z)
C(z)

= 1/(z 2 1.5z + 0.7)

60

= 1/(z 2 + 0.5)

3.5

50
3

40

S()

S()

2.5

30

1.5

20
1

10
0.5

0.5

1.5

frequency (rad/s)

Input signals

2.5

0.5

1.5

2.5

frequency (rad/s)

6-8

Pseudorandom binary sequence (PRBS)

Clock

State

a1

State

State

n
a2

an1

y(t)
an

a1 a2 . . . an
1
0 ... 0

x(t)
x(t + 1) = . . .
.
.
.
.
0 ... 1
0


y(t) = 0 . . . 0 1 x(t)
Input signals

6-9

PRBS (cont.)
Every initial state is allowed except the all-zero state
The feedback coefficients a1, a2, . . . , an are either 0 or 1
All additions is modulo-two operation
The sequences are two-state signals (binary)
There are possible 2n 1 different state vectors x (all-zero state is
excluded)
A PRBS of period equal to M = 2n 1 is called a maximum length
PRBS (ML PRBS)
For maximum length PRBS, its characteristic resembles white random
noise (pseudorandom)

Input signals

6-10

Influence of the feedback path on the period

Let n = 3 and initialize x with x(0) = (1, 0, 0)
With a = (1, 1, 0), the state vectors x(k), k = 1, 2, . . . are

1
0
0

1
1
0

0
1
1

1
0
1

1
1
0

The sequence has period equal to 3

With a = (1, 0, 1), the state vectors x(k), k = 1, 2, . . . are

1
0
0

1
1
0

1
1
1

0
1
1

1
0
1

0
1
0

0
0
1

1
0
0

The sequence has period equal to 7 (the maximum period, 23 1)

Input signals

6-11

Maximum length PRBS

Denote q 1 the unit delay operator and let
A(q 1) = 1 a1q 1 a2q 2 . . . anq n
The PRBS y(t) satisfies the homogeneous equation:
A(q 1)y(t) = 0
This equation has only solutions of period M = 2n 1 if and only if
1. The binary polynomial A(q 1) is irreducible, i.e., there do not exist any
two polynomial A1(q 1) and A2(q 1) such that
A(q 1) = A1(q 1)A2(q 1)
2. A(q 1) is a factor of 1 q M but is not a factor of 1 q p for any
p<M
Input signals

6-12

Generating Maximum length PRBS

Examples of polynomials A(z) satisfying the previous two conditions
n
3
4
5
6
7
8
9
10

Input signals

A(z)
1 z z3
1 z2 z3
1 z z4
1 z3 z4
1 z2 z5
1 z3 z5
1 z z6
1 z5 z6
1 z z7
1 z3 z7
1 z z2 z7 z8 1 z z6 z7 z8
1 z4 z9
1 z5 z9
1 z 3 z 10
1 z 7 z 10

6-13

Properties of maximum length PRBS

Let y(t) be an ML PRBS of period M = 2n 1
Within one period y(t) contains (M + 1)/2 = 2n1 ones and
(M 1)/2 = 2n1 1 zeros
For k = 1, 2, . . . , M 1,
y(t) y(t k) = y(t l)
for some l [1, M 1] that depends on k
Moreover, for any binary variables x, y,
1
xy = (x + y (x y))
2
These properties will be used to compute the covariance function of
maximum length PRBS
Input signals

6-14

Covariance function of maximum length PRBS

The mean is given by counting the number of outcome 1 in y(t):


M
1
1 M +1
1
1 X
y(t) =
= +
m=
M t=1
M
2
2 2M
The mean is slightly greater than 0.5
Using y 2(t) = y(t), we have the covariance function at lag zero as
M
1 X 2
M2 1
2
2
C(0) =
y (t) m = m m =
M t=1
4M 2

The variance is therefore slightly less than 1/4

Input signals

6-15

Covariance function of maximum length PRBS

For = 1, 2, . . .,

C( ) = (1/M )

M
X

y(t + )y(t) m2

t=1

M
1 X
[y(t + ) + y(t) (y(t + ) y(t))] m2
=
2M t=1

M
1 X
=m
y(t + l) m2 = m/2 m2
2M t=1

Input signals

M +1
4M 2

6-16

Asymptotic behavior of the covariance function of PRBS

Define y(t) = 1 + 2y(t) so that its outcome is either 1 or 1

m
= 1 + 2m = 1/M 0

C(0)
= 4C(0) = 1 1/M 2 1
) = 4C( ) = 1/M 1/M 2 1/M,
C(

= 1, 2, . . . , M 1

When M is large, the covariance function of PRBS has similar properties

to a white noise

However, their spectral density matrices can be drastically different

Input signals

6-17

Spectral density of PRBS

The output of PRBS sequence is shifted to values a and a with period M
The autocorrelation function is also periodic and given by
(
a2 ,
= 0, M, 2M, . . .
R( ) =
a2
, otherwise
M
Since R( ) is periodic with period M , it has a Fourier representation:
R( ) =

M
1
X

Ck ei2 k/M ,

with Fourier coefficients Ck

k=0

Therefore, the spectrum of PRBS is an impulse train:

S() =

M
1
X
k=0

Input signals

2k
)
Ck (
M

6-18

Spectral density of PRBS

Hence, the Fourier coefficients
M 1
1 X
R( )ei2 k/M
Ck =
M =0

are also the spectral coefficients of S()

Using the expression of R( ), we have
a2
C0 = 2 ,
M

a2
Ck = 2 (M + 1),
M

k = 1, 2, . . .

Therefore,
"

a2
S() = 2 () + (M + 1)
M

M
1
X
k=1

( 2k/M )

It does not resemble spectral characteristic of a white noise (flat spectrum)

Input signals

6-19

Comparison of the covariances between filtered inputs

Define y1(t) as the output of a filter:
y1(t) ay1(t 1) = u1(t),
with white noise u(t) of zero mean and variance 2
Let y2(t) be the output of the same filter:
y2(t) ay2(t 1) = u2(t),
where u2(t) is a PRBS of period M and amplitude

What can we say about the covariances of y1(t) and y2(t) ?

Input signals

6-20

Comparison of the correlations between filtered inputs

The correlation function of y1(t) is given by
2
R1( ) =
a ,
2
1a

The correlation function of y2(t) can be calculated as

R2( ) =

Sy2 ()ei d

2

 i

1
 e d

=
Su2 () 
1 aei 

"
#
M
1
2
X
cos(2 k/M )
1

+ (M + 1)
=
M (1 a)2
1 + a2 2a cos(2k/M )
Z

k=1

Input signals

6-21

Plots of the correlation functions

3

white noise
PRBS (M=20)
PRBS (M=50)
PRBS (M=200)

2.5

R( )

1.5

0.5

0.5

10

The filter parameter is a = 0.8

R( ) of white noise and PRBS inputs are very close when M is large
Input signals

6-22

Persistent excitation
A signal u(t) is persistently exciting of order n if
1. The following limit exists:
N
1 X
u(t + )u(t)
R( ) = lim
N N
t=1

2. The following matrix is positive definite

R(0)
R(1)
. . . R(n 1)
R(1)
R(0)
. . . R(n 2)

R(n) =
..
..
..
...

R(1 n) R(2 n) . . .
R(0)
(if u(t) is from an ergodic stochastic process, then R(n) is the usual
covariance matrix (assume zero mean))
Input signals

6-23

Examining the order of persistent excitation

White noise input of zero mean and variance 2
R( ) = 2 ( ),

R(n) = 2 In

Thus, white noise is persistently exciting of all orders

Step input of magnitude
R( ) = 2, ,

R(n) = 21n

A step function is persistently exciting of order 1

Impulse input: u(t) = 1 for t = 0 and 0 otherwise
R( ) = 0, ,

R(n) = 0

An impulse is not persistently exciting of any order

Input signals

6-24

Example 1: FIR models

Recall the problem of estimating an FIR model where
h(k) = 0,

kM

The coefficients h(k) are the solution to the following equation

Ryu
(0)

Ryu
(1)

..

Ryu
(M

To solve the equation, the matrix R(M + 1) must be nonsingular

Input signals

Ru(0)
Ru(1)
Ru(M 1)
h (0)
Ru(1)
h(1)
Ru(0)
Ru(M 2)
=

.
.
.
.
.

..
.
.
.
.
Ru(1 M ) Ru(2 M )
Ru(0)
h(M 1)
1)

6-25

Example 2: Estimating noisy linear models

Consider a least-squares problem of estimating a first-order model
y(t) = ay(t 1) + bu(t) + e(t)
where u(t) is an input signal, and e(t) is an i.i.d. noise of zero mean
We can show that
If u(t) is a PRBS or step input, the consistent estimates are obtained,
i.e.,
(
a, b) (a, b), as N
If u(t) is an impulse, a
a but b does not converge to b as N increases
In loose terms, the impulse input does not provide enough information
on y(t) to estimate b

Input signals

6-26

Properties of persistently exciting signals

Let u(t) be a multivariable ergodic process. Assume that Su() is positive
definite in at least n distinct frequencies (within the interval (, ))
We have the following two properties
Property 1 u(t) is persistently exciting of order n
Property 2 The filtered signal y(t) = H(q 1)u(t) is persistently exciting
of order n
where H(z) is an asymptotically stable linear filter and det H(z) has no
zero on the unit circle
From above facts, we can imply
An ARMA process is persistently exciting of any finite order
Input signals

6-27

Examining the order of PRBS

Consider a PRBS of period M and magnitude a, a
The matrix containing n-covariance sequenes (where n M ) is
2

a
a2/M
R(n) =
..
a2/M

a /M
a2
..
a2/M

Therefore, for any x Rn,

T

x R(n)x = x

a
a
(a + )I 11T
M
M
2

. . . a /M
. . . a2/M
..
...

...
a2



2
1
a
(1

n)
a2(1 + )xT x xT x1T 1 = a2kxk2 1 +
0
M
M
M
A PRBS with period M is persistently exciting of order M
Input signals

6-28

Examining the order of sum of sinusoids

Consider the signal u(t) =

Pm

k=1 ak sin(k t

+ k )

where 0 1 < 2 < . . . < m

The spectral density of u is given by
S() =

m
X
Ck

k=1

[( k ) + ( + k )]

Therefore S() is nonzero (in the interval (, ]) in exactly n points

where

0 < 1, m <
2m,
n = 2m 1, 0 = 1, or m =

2m 2, 0 = 1 and m =
It follows from Property 1 that u(t) is persistently exciting of order n
Input signals

6-29

Summary
The choice of input is imposed by the type of identification method
The input signal should be persistently exciting of a certain order to
ensure that the system can be identified
Some often used signals include PRBS and ARMA processes

Input signals

6-30

References
Chapter 5 in
T. S
oderstrom and P. Stoica, System Identification, Prentice Hall, 1989

Input signals

6-31