Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

CDS 110b Norms of Signals and Systems

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

CALIFORNIA INSTITUTE OF TECHNOLOGY

Control and Dynamical Systems


CDS 110b
R. M. Murray Norms of Signals and Systems 13 February 2006
In this set of lectures we introduce the concept of the norm of a (linear) system and
show how this can be used to specify stability and performance measures.
Reading:
DFT, Chapters 2 and 3
1 Introduction
One of the key uses of feedback is to provide robustness to uncertainty. We have already
seen several examples of this in the previous chapters, from tracking a reference signal whose
evolution is not known ahead of time, to rejecting disturbances that might aect the process
or our measurements of it. In this chapter we generalize the classes of uncertainty that we
can analyze to include uncertainty in the process that we are trying to control.
Robustness to process variations is the ability of a closed loop system to be insensitive to
component variations that make up the system. It is one of the most useful properties of
feedback and is what make it possible to design feedback systems based on strongly simplied
models.
One of the key issues in studying robustness is to describe variations in system dynamics.
Using state space representations for a system, uncertainty can be captured by varying the
parameters of a system (e.g., the elements of the A, B, C and D matrices). There are
however many variations that are not captured by such an approach. For example, some
dynamic phenomena may have been neglected during the modeling stage to keep the the
size of the model tractable. These unmodeled dynamics cannot be accounted for by simply
modifying the parameters of the simplied system. Similarly, there may also be small time
delays that have been neglected in a model, in part because these are dicult to analyze
using ODEs.
As an example of this type of uncertainty, consider the speed controller that has been dis-
cussed throughout CDS 110a. We have designed several dierent controllers for this system
using a model that was derived in Chapter 2 of AM05. The original model considered the dy-
namics of the forward motion of the vehicle and the torque characteristics of the engine and
transmission. It did not, however, include a detailed model of the engine dynamics (whose
F C P
1

r
e u v
d
w
n
y
Figure 1: Block diagram of a basic feedback loop.
combustion processes are extremely complex), nor the slight delays that can occur in modern
electronically controlled engines (due to the processing time of the embedded computers).
Rather than try to include all of these details in our model, and how they change over time,
we can instead try to show that the controllers that we have designed using our simplied
models are robust with respect to the unmodeled dynamics. Under the assumption that the
dynamics that we initially ignored are small compared with the dominant dynamics of the
system, we expect that our controllers will work when implemented on the full system.
Transfer functions provide a natural way of modeling these types of uncertainty (and many
others). The basic idea is to model the unmodeled dynamics by including a transfer func-
tion in the system desription whose frequency response is bounded, but otherwise unspecied.
So, for example, we might model the engine dynamics in the speed control example as a sys-
tem that very quickly provides the torque that is requested through the throttle, giving a
small deviation from the simplied model, which assumed the torque response was instan-
taneous. This technique can also be used in many instances to model parameter variations,
allowing a quite general approach to uncertainty management. One limitation of the tools
we present here is that they are usually restricted to linear systems, although some nonlinear
extensions have been developed.
A block diagram representation of this approach is shown below: In this gure, the block
represents the unmodeled dynamics of the system and we think of as small (in a sense
to be made precise later). Our goal is to design a controller C such that the closed loop
performance of the system satises a given specication for any M. We say that such
a controller is robust with respect to the (bounded) perturbation . In order to make these
concepts precise, we dene the norms of signals and systems.
2 Norms of Signals
Denition 1. A set V is a linear (vector) space over R if we can dene operations of addition
(of vectors) and multiplication (by scalars) such that for every x, y V and , R we
2
have
x +y V (x +y) = x +y V
and there exists a zero element

0 V such that

0 = xx = 0 x (where 0 is the real number


zero).
Examples:
1. V = R
n
2. C
n
[t
0
, t
1
] = the space of continous functions mapping the interval [t
0
, t
1
] to R
n
.
3. P
n
(, ) = the set of piecewise continous functions on (, ) taking values in
R
n
.
These last two examples are innite dimensional, in the sense that there is no nite basis
for the space.
Denition 2. Let V be a vector space over R. A mapping : V R
n
is a norm if it
satises:
1. x 0 for all x V
2. x = 0 if and only if x = 0
3. x = || x for all x V and R
4. x +y x +y (triangle inequality)
The denitions are correct even if V is innite dimensional. For example, if we take
C(, ) = {x : (, ) R, continuous}, then the zero element is the zero func-
tion and scaling and addition are done pointwise in time:
(u +v)(t) = u(t) +v(t) (u)(t) = u(t)
Examples of norms:
Name V = R
n
V = {u : (, ) R}
1-norm x
1
=

n
i=1
|x
i
| u
1
=
_

|u(t)| dt
2-norm x
2
=
_

n
i=1
|x
i
|
2
u
2
=
_
_

|u(t)|
2
dt
_
1/2
p-norm x
p
=
p
_

n
i=1
|x
i
|
p
u
p
=
_
_

|u(t)|
p
dt
_
1/p
-norm x

= max
i
|x
i
| u

= sup
t
|u(t)|
3
The operation sup is the least upper bound of a function.
Example: 2 norm on C(, )
1. u
2
=
_
_

|u(t)|
2
dt
_
1/2
> p
2.
u
2
= 0 =
_

|u(t)|
2
= 0
= u(t) = 0 on any interval
= u(t) = 0 for all t (by continuity) = u = 0
3. u
2
=
_
_

|u(t)|
2
dt
_
1/2
= u
2

4. Triangle inequality: u +v
2
u
2
+v
2
Proof: Exercise. Use the fact that |u(t) +v(t)| |u(t)| +|v(t)|.
Denition 3. A normed linear space V is a linear (vector) space V equipped with a norm
.
Examples:
L
1
V = P(, )
1
L
2
V = P(, )
2
L

V = P(, )

3 Linear Mappings and Induced Norms


Let V and W be linear spaces over R.
Denition 4. A mapping A : V W is a linear map if
A(
1
v
1
+
2
v
2
) =
1
Av
1
+
2
Av
2
for all v
1
, v
2
V and
1
,
2
R.
Examples:
1. The set of n n matrices are a linear map from R
n
to R
n
using the usual rules of
multiplying a vector by a matrix.
2. Let V = P[0, 1] and W = R. Then the operator given by
Av =
_
1
0
v(t) dt
is a linear mapping from V to W.
4
3. Let V = P[0, ] and W = P[0, ]. Then the operator given by
w(t) =
_
t
0
e
a(t)
v() d
is a linear mapping from V to W.
Denition 5. Let V be a vector space with norm
a
and W be a vector space with norm

b
. Then the induced norm
a,b
of A : V W is given by
A
a,b
= sup
va1
Avb
Example: matrix norms
V = R
n
W = R
n
A : V W
v
2
=

v
T
v w
2
=

w
T
w A
2,2
=
_

max
(A
T
A)
4 Norms of Systems
u P y y(t) = Ce
At
x(0)
. .
Assume zero
+
_
t
0
Ce
A(t)
Bu() d
. .
Linear map
Q: what is the induced norm of the system P, thought of as an input/output mapping of
signals?
A: Depnds on the norm of the signals
Theorem 1. Let be a stable system with state space representation (A, B, C, 0), transfer
function H(s) = C(sI A)
1
B. Then the following table gives the input/output norm of
u
2
=
_
_

u
2
(t)dt
_
1/2
u

= sup
t
|u(t)|
y
2
sup

|H(j)|
y

1
2
_
_

|H(j)|
2
d
_
1/2 _

|Ce
At
B|dt
If we let h(t) = Ce
At
B be the impulse response for the system and dene H

= sup

|H(j|
and H
2
=
1
2
_
_

|H(j)|
2
d
_
1/2
then
u
2
u

y
2
H

H
2
h
1
5
These dene the three major approaches to linear control:
H

= L
2
to L
2
H
2
= L
2
to L

l
1
= L

to L

Remarks:
1. These system norms only work for linear systems since otherwise looking at u 1
is not sucient.
2. Can extend to nonlinear systems if you constrain the inputs to be u U = {u : u
2

1}. Get nonlinear H

this way.
3. The focus of DFT is H

control = L
2
L
2
induced norm. Think of this as the
energy of the input (or noise) to the energy of the output.
4. Proofs of these norms are given in the book. OK to skip, but here is the idea (for H

):
Let y(s) =
_

y(t)e
st
dt (Laplace transform)
Can show that y
2
2
=
1
2
_

| y(j)|
2
d (Parsevals theorem)
Can also show that y(s) = H(s)u(s) (convolution multiplication)
y
2
2
= y
2
2
=
1
2
_

H(j)|
2
| u(j)|
2
d
H
2

1
2
_

| u(j)|
2
d = H
2

u
2
2
= y
2
H

u
2
.
To complete the proof, show that there exists u(t) that achieves this bound so that
max
u
2
1
y
2
= max
u
2
1
Hu
2
= H

5 Stability and Performance


We now show how to write conditions for stability and performance in terms of the norms
of systems. We assume that the process and controller are LIT systems. In addition, we will
make the simplifying assumption that the disturbances and noise enter linearly at the same
point as the inputs and outputs of the process (as shown in Figure 1 on page 2).
Remarks:
6
1. H(s) is the Laplace transform of the impulse response h(t)
2. H(s) is rational if the system is nite dimensional
H(s) =
n(s)
d(s)
n(s), d(s) polynomials
3. Standard terms:
H(s) is stable if it is analytic in the closed right half plane (i.e., no RHP poles)
H(s) is proper if H(j) is nite (deg d deg n)
H(s) is strictly proper if H(j) 0 (deg d > deg n)
H(s) is biproper if H(s) and H
1
(s) are proper (deg d = deg n)
4. Parsevals theorem
1
2
_

|H(j)|
2
d =
_

|h(t)|
2
d
Stability
Proposition 2. If a transfer function is stable then bounded inputs yield bound outputs.
Proof. Assume that H is stable. Then h
1
is bounded and hence y

h
1
u

=
bounded inputs yield bounded outputs. The converse is also true.
Denition 6. A feedback interconnection is internally stable if the closed loop transfer
function from any input to any other signal is stable.
Remarks:
1. Internal stability = signals inside the feedback loop remain bounded. Example of
something that is not internally stable:

x3
1
s1
s1
(s+1)
2
r
n
d
u y H
yr
=
PC
1 +PC
=
1
(s + 1)
2
(work out)
H
ur
=
C
1 +PC
=
(s + 1)
2
(s 1)(s + 1)
2
(unstable)
2. Consider the basic feedback loop
7

C P
F
r
n
d
u y x
2
x
3
x
3
v
_
_
x
1
x
2
x
3
_
_
=
1
1 +PCF
_
_
1 PF F
C 1 CF
PC P 1
_
_
_
_
r
d
n
_
_
Internally stable all 9 transfer functions are stable (other signals are simple
linear combinations, eg y = x
3
n, etc).
3. If F = 1, can reduce the number of transfer functions which need to be checked
(homework).
If P, C and F are rational transfer functions, we can write them as
P =
n
P
d
P
C =
n
C
d
C
F =
n
F
d
F
where each n
X
and d
X
re coprime polynomials (no common zeros).
Note that the closed loop poles of the basic feedback loop are given by the zeros of n
P
n
C
n
F
+
d
P
d
C
d
F
= (s) = characteristic polynomial.
Theorem 3. The basic feedback loop is internally stable if and only if there are no zeros of
(s) = n
P
n
C
n
F
+d
P
d
C
d
F
in the closed right half plane.
Proof. Consider the case when F = 1 (for simplicity). Can show
_
_
x
1
x
2
x
3
_
_
=
1
n
P
n
C
+d
P
d
C
_
_
d
P
d
C
n
P
d
C
d
P
d
C
d
P
n
C
d
P
d
C
d
P
n
C
n
P
n
C
n
P
d
C
d
P
d
C
_
_
_
_
r
d
n
_
_
and the poles are all contained in the common factor 1/(n
P
n
C
+d
P
d
C
).
Theorem 4. The basic feedback loop is internally stable if and only if the following conditions
hold:
1. 1 +PCF has no zeros in the right half plane (Re s 0)
2. There are no pole zero cancellations in Re s 0 when PCF is formed
Proof. See DFT
8
Performance
We choose our performance goal to try to get good tracking for a set of signals.
Sensitivity function: S =
1
1 +PC
= H
er
We would like to keep this transfer function small ( = good tracking)
Assume internal stability = S is stable and proper
Assume P is strickly proper = S(j) = 1 = bad tracking at high frequency.
Complementary sensitivity function: T ==
PC
1 +PC
H
yr
Note that T is equal to H
du
= gives the transfer function between disturbances
and input
Would like to keep this transfer function small as well
Cant make both S and T small since
S +T =
1
1 +PC
+
PC
1 +PC
= 1 for all
= tradeo between input magnitude and tracking error.
Suppose we want performance to mean small error for unit sized inputs. Try
S

= tolerated error
= for r(t) = sin(t), |e(t)| .
Problem: S(j) = 1 = S

1 = cant make S arbitarily small.


Fix: Specify performance in terms of weighted sensitivity function.
W
1
S

1 when W
1
is small, allow larger errors.
9
W
1
1
|W
1
|

S should lie below W


1
1
good perf
dont
care
large error
Remarks:
1. Lots of ways of justifying why W
1
S

makes sense (L
2
-induced norm, etc). Big
reason: it works well.
2. Other possibilities:
_
e
u
_
=
_
PS S
T CS
_ _
d
n
_
S = sensitivity function
T = complementary sensitivity function
Can shoose any of these (or combination) for performance spec.
Nyquisti plot interpretation:
W
1
S

< 1

W
1
(j)
1 +L(j)

< 1
|W
1
(j)| < |1 +L(j)| 1 +|L(j)|
|W
1
|
1
The size of the disk at 1 changes size depending on frequency = |W
1
gives the distance
that we want to stay away from 1 (where we go unstable).
10

You might also like