Linear Controller Design: Limits of Performance by Stephen Boyd and Craig Barratt - HTML preview

PLEASE NOTE: This is an HTML preview only and some elements such as links or page numbers may be incorrect.
Download the book in PDF, ePub, Kindle for a complete version.

CHAPTER 6 GEOMETRY OF DESIGN SPECIFICATIONS

1:2

1

@

I

@

0:8

@

~13

s

@

I

@

@

0:6

()

( 13 + ~13) 2

s

s

=

t

@

I

@

s

@

0:4

13

s

0:2

0

;0:2

0

0:1

0:2

0:3

0:4

0:5

0:6

0:7

0:8

0:9

1

t

The set in (6.6) is convex, since if the step responses 13 and

Figure

6.6

V

s

~13 do not exceed 1.1, then their average, ( 13 +~13) 2, also does not exceed

s

s

s

=

1.1. By the argument in section 6.2.3, the specication os is convex.

H

6.4.3

A Quasiconvex Functional

We will see in chapter 8 that several important functionals are quasiconvex, for

example, those relating to settling time and bandwidth of a system. We describe

one such example here.

Consider the stability degree of a transfer matrix, de ned by

stab deg( ) = max

is a pole of

H

f<p

j

p

H

g

:

The functional stab deg is quasiconvex since, for each ,

stab deg( )

=

for

fH

j

H

g

fH

j

kH

k

<

1

<

;

g

1

(recall that

is the -shifted

norm described in section 5.2.7), and we

k

k

H

1

1

saw above that the latter speci cation is a ne. stab deg is not convex, however:

for most values of , we have

stab deg(

+ (1 ) ~) = maxn stab deg( ) stab deg( ~)o

H

;

H

H

H

:

6.5

Implications for Tradeoffs and Optimization

When the design speci cations are convex, and more generally, when a family of

design speci cations is given by convex functional inequalities, we can say much

more about the concepts introduced in chapter 3.

index-148_1.png

index-148_2.png

index-148_3.png

index-148_4.png

6.6 CONVEXITY AND DUALITY

139

6.5.1

Performance Space Geometry

Suppose that in the multicriterion optimization problem described in section 3.5,

the hard constraint hard and the objective functionals 1 ...

are convex. Then

D

L

the region of achievable speci cations in performance space,

= n

a

a

L

hard

...

is achievable o

(6.7)

1

L

A

a

2

R

D

^

D

^

^

D

1

L

is also convex.

To see this, suppose that and ~ each correspond to achievable speci cations.

a

a

Then there are transfer matrices and ~ that satisfy hard and also, for each ,

H

H

D

k

1

,

k

L

( )

( ~) ~

H

a

H

a

:

k

k

k

k

Now suppose that 0

1. The transfer matrix

+(1 ) ~ satis es the hard

H

;

H

constraint hard, since hard is convex, and also, for each , 1

,

D

D

k

k

L

( + (1 ) ~)

+ (1 )~

H

;

H

a

;

a

k

k

k

since the functional is convex. But this means that the speci cation +(1 )~

a

;

a

k

corresponds to an achievable speci cation, which veri es that is convex.

A

An important consequence of the convexity of the achievable region in perfor-

mance space is that every Pareto optimal speci cation is optimal for the classical

optimization problem with weighted-sum objective, for some nonnegative weights

(c.f. gure 3.8). Thus the speci cations considered in multicriterion optimization or

classical optimization with the weighted-sum or weighted-max objectives are exactly

the same as the Pareto optimal speci cations.

In the next section we discuss another important consequence of the convexity

of (which in fact is equivalent to the observation above).

A

6.6

Convexity and Duality

The dual function de ned in section 3.6.2 is always concave, meaning that

is

;

convex, even if the objective functionals and the hard constraint are not convex. To

see this, we note that for each transfer matrix that satis es hard, the function

H

D

of de ned by

H

( ) = 1 1( )

( )

;

H

;

;

H

H

L

L

is convex (indeed, linear). Since

can be expressed as the maximum of this family

;

of convex functions, i.e.,

( ) = max

( )

satis es hard

(6.8)

;

f

j

H

D

g

H

it is also a convex function of (see section 6.2.1).

index-149_1.png

index-149_2.png

index-149_3.png

index-149_4.png

140