You are on page 1of 177

4 Multivariable Control System Design

Overview Design of controllers for multivariable systems requires an assessment of structural properties of transfer matrices. The zeros and gains in multivariable systems have directions. With norms of multivariable signals and systems it is possible to obtain bounds for gains, bandwidth and other system properties. A possible approach to multivariable controller design is to reduce the problem to a series of single loop controller design problems. Examples are decentralized control and decoupling control. The internal model principle applies to multivariable systems and, for example, may be used to design for multivariable integral action.

4.1 Introduction
Many complex engineering systems are equipped with several actuators that may inuence their static and dynamic behavior. Commonly, in cases where some form of automatic control is required over the system, also several sensors are available to provide measurement information about important system variables that may be used for feedback control purposes. Systems with more than one actuating control input and more than one sensor output may be considered as multivariable systems or multi-input-multi-output (MIMO). The control objective for multivariable systems is to obtain a desirable behavior of several output variables by simultaneously manipulating several input channels.

4.1.1 Examples of multivariable feedback systems


The following two examples discuss various phenomena that specically occur in MIMO feedback systems and not in SISO systems, such as interaction between loops and multivariable non-minimum phase behavior. 153

154

Chapter 4. Multivariable Control System Design

h1 2 h2

Figure 4.1: Two-tank liquid ow process with recycle

h1,ref

k1

1 4 P

h1

h2,ref

k2

h2

Figure 4.2: Two-loop feedback control of the two-tank liquid ow process

4.1. Introduction

155

Example 4.1.1 (Two-tank liquid ow process). Consider the ow process of Fig. 4.1. The incoming ow 1 and recycle ow 4 act as manipulable input variables to the system, and the outgoing ow 3 acts as a disturbance. The control objective is to keep the liquid levels h1 and h2 between acceptable limits by applying feedback from measurements of these liquid levels, while accommodating variations in the output ow 3 . As derived in Appendix 4.6, a dynamic model, linearized about a given steady state, is h1 (t) 1 2 (t) = 1 h 0 0 h1 (t) 1 + h2 (t) 0 1 1 1 (t) 0 + (t). 4 (t) 1 3 (4.1)

Here the h i and i denote the deviations from the steady state levels and ows. Laplace transformation yields H1 (s) s+1 = H2 (s) 1 0 s
1

1 0

1 1

1 (s) 4 (s)

0 1

3 (s)

.,

which results in the transfer matrix relationship H1 (s) = H2 (s)


1 s+1 1 s(s+1) 1 s+1 1 s+1 1 (s) 4 (s)

0 1 s

3 (s).

(4.2)

The example demonstrates the following typical MIMO system phenomena:

Each of the manipulable inputs 1 and 4 affects each of the outputs to be controlled h1 and h2 , as a consequence the transfer matrix
P(s) =
1 s+1 1 s(s+1) 1 s+1 1 s+1

(4.3)

has nonzero entries both at the diagonal and at the off-diagonal entries.

Consequently, if we control the two output variables h1 and h2 using two separate control loops, it is not immediately clear which input to use to control h1 , and which one to control h2 . This issue is called the input/output pairing problem. A possible feedback scheme using two separate loops is shown in Fig. 4.2. Note that in this control scheme, there exists a coupling between both loops due to the non-diagonal terms in the transfer matrix (4.3). As a result, the plant transfer function in the open upper loop from 1 to h1 , with the lower loop closed with proportional gain k2 , depends on k2 ,
P11 (s)
llc

k2 1 1 . s+1 s(s + 1 k2 )

(4.4)

Here llc indicates that the lower loop is closed. Owing to the negative steady state gain in the lower loop, negative values of k2 stabilize the lower loop. The dynamics of P11 (s) llc changes with k2 : k2 = 0 : k2 = 1 : k2 = : P11 (s) P11 (s) P11 (s)
llc llc llc

= = =

1 s+1 , s+1 s(s+2) , 1 s.

156

Chapter 4. Multivariable Control System Design

The phenomenon that the loop gain in one loop also depends on the loop gain in another loop is called interaction. The multiple loop control structure used here does not explicitly acknowledge interaction phenomena. A control structure using individual loops, such as multiple loop control, is an example of a decentralized control structure, as opposed to a centralized control structure where all measured information is available for feedback in all feedback channels.

h1,ref

k1 K pre

1 4 P

h1

h2,ref

k2
controller K

h2

Figure 4.3: Decoupling feedback control of two-tank liquid ow process

A different approach to multivariable control is to try to compensate for the plant interaction. In Example 4.1.1 a precompensator K pre in series with the plant can be found which removes the interaction completely. The precompensator K pre is to be designed such that the transfer matrix PK pre is diagonal. Figure 4.3 shows the feedback structure. Suppose that the precompensator K pre is given the structure
K pre (s) = 1
pre K21 (s)

K12 (s) 1

pre

(4.5)

then PK pre is diagonal by selecting K12 (s) = 1,


pre

K21 (s) =

pre

1 . s

(4.6)

The compensated plant PK pre for which a multi-loop feedback is to be designed as a next step, reads P(s)K pre (s) =
1 s

0 . 1 s

(4.7)

Closing both loops with proportional gains k1 > 0 and k2 < 0 leads to a stable system. In doing so, a multivariable controller K has been designed having transfer matrix K(s) = K pre (s) k1 0 0 k1 k2 = . k1 /s k2 k2 (4.8)

This approach of removing interaction before actually closing multiple individual loops is called decoupling control. The next example shows that a multivariable system in which the transfer matrix contains rstorder entries may show non-minimum phase behavior.

4.1. Introduction

157

Example 4.1.2 (Multivariable non-minimum phase behavior). Consider the two input, two output linear system y1 (s) = y2 (s)
1 s+1 1 s+1 2 s+3 1 s+1

u1 (s) . u2 (s)

(4.9)

If the lower loop is closed with constant gain, u2 (s) = k2 y2 (s), then for high gain values (k2 ) the lower feedback loop is stable but has a zero in the right-half plane, y1 (s) = 2 s1 1 u1 (s) = u1 (s). s+1 s+3 (s + 1)(s + 3) (4.10)

Thus under high gain feedback of the lower loop, the upper part of the system exhibits nonminimum phase behavior. Conversely if the upper loop is closed under high gain feedback, u1 = k1 y1 with k1 , then y2 (s) = s1 u2 (s). (s + 1)(s + 3) (4.11)

Apparently, the non-minimum phase behavior of the system is not connected to one particular input-output relation, but shows up in both relations. One loop can be closed with high gains under stability of the loop, and the other loop is then restricted to have limited gain due to the non-minimum phase behavior. Analysis of the transfer matrix P(s) =
1 s+1 1 s+1 2 s+3 1 s+1

1 s+3 (s + 1)(s + 3) s + 3

2s + 2 s+3

(4.12)

shows that it loses rank at s = 1. In the next section it will be shown that s = 1 is an unstable transmission zero of the multivariable system and this limits the closed loop behavior irrespective of the controller design method used. Consider the method of decoupling precompensation. Express plant and precompensator as P(s) = P11 (s) P21 (s) P12 (s) , P22 (s) K(s) = K11 (s) K21 (s) K12 (s) . K22 (s) (4.13)

Decoupling means that PK is diagonal. Inserting the matrix (4.12) into PK, a solution for a decoupling K is K11 (s) = 1, K22 (s) = 1, K12 (s) = 2 s+1 , s+3 K21 (s) = 1. (4.14)

Then the compensated system is P(s)K(s) =


1 s+1 1 s+1 2 s+3 1 s+1 s1 s+1 (s+1)(s+3) 1 2 s+3 = 1 1 0

0
s1 (s+1)(s+3)

(4.15)

The requirement of decoupling has introduced a right-half plane zero in each of the loops, so that either loop now only can have a restricted gain in the feedback loop. The example suggests that the occurrence of non-minimum phase phenomena is a property of the transfer matrix and exhibits itself in a matrix sense, not connected to a particular input-output pair. It shows that a zero of a multivariable system has no relationship with the possible zeros of individual entries of the transfer matrix.

158

Chapter 4. Multivariable Control System Design

In the example the input-output pairing has been the natural one: output i is connected by a feedback loop to input i. This is however quite an arbitrary choice, as it is the result of our model formulation that determines which inputs and which outputs are ordered as one, two and so on. Thus the selection of the most useful input-output pairs is a non-trivial issue in multiloop control or decentralized control, i.e. in control congurations where one has individual loops as in the example. A classical approach towards dealing with multivariable systems is to bring a multivariable system to a structure that is a collection of one input, one output control problems. This approach of decoupling control may have some advantages in certain practical situations, and was thought to lead to a simpler design approach. However, as the above example showed, decoupling may introduce additional restrictions regarding the feedback properties of the system.

4.1.2 Survey of developments in multivariable control


The rst papers on multivariable control systems appeared in the fties and considered aspects of noninteracting control. In the sixties, the work of Rosenbrock (1970) considered matrix techniques to study questions of rational and polynomial representation of multivariable systems. The polynomial representation was also studied by Wolovich (1974). The books by Kailath (1980) and Vardulakis (1991) provide a broad overview. The use of Nyquist techniques for multivariable control design was developed by Rosenbrock (1974a). The generalization of the Nyquist criterion and of root locus techniques to the multivariable case can be found in the work of Postlethwaite and MacFarlane (1979). The geometric approach to multivariable state-space control design is contained in the classical book by Wonham (1979) and in the book by Basile and Marro (1992). A survey of classical design methods for multivariable control systems can be found in (Korn and Wilfert 1982), (Lunze 1988) and in the two books by Tolle (1983), (1985). Modern approaches to frequency domain methods can be found in (Raisch 1993), (Maciejowski 1989), and (Skogestad and Postlethwaite 1995). Interaction phenomena in multivariable process control systems are discussed in terms of a process control formulation in (McAvoy 1983). A modern, process-control oriented approach to multivariable control is presented in (Morari and Zariou 1989). The numerical properties of several computational algorithms relevant to the area of multivariable control design are discussed in (Svaricek 1995).

4.2 Poles and zeros of multivariable systems


In this section, a number of structural properties of multivariable systems are discussed. These properties are important for the understanding of the behavior of the system under feedback. The systems are analyzed both in the time domain and frequency domain.

4.2.1 Polynomial and rational matrices


Let P be a proper real-rational transfer matrix. Real-rational means that every entry Pi j of P is a ratio of two polynomials having real coefcients. Any rational matrix P may be written as a fraction P= 1 N d

where N is a polynomial matrix1 and d is the monic least common multiple of all denominator polynomials of the entries of P. Polynomial matrices will be studied rst. Many results about
1 That

is, a matrix whose entries are polynomials.

4.2. Poles and zeros of multivariable systems

159

polynomial matrices may be found in, e.g., the books by MacDuffee (1956), Gohberg, Lancaster, and Rodman (1982) and Kailath (1980). In the scalar case a polynomial has no zeros if and only if it is a nonzero constant, or, to put it differently, if and only if its inverse is polynomial as well. The matrix generalization is as follows. Denition 4.2.1 (Unimodular polynomial matrix). A polynomial matrix is unimodular if it square and its inverse exists and is a polynomial matrix. It may be shown that a square polynomial matrix U is unimodular if and only if det U is a nonzero constant. Unimodular matrices are considered having no zeros and, hence, multiplication by unimodular matrices does not affect the zeros. Summary 4.2.2 (Smith form of a polynomial matrix). For every polynomial matrix N there exist unimodular U and V such that 1 0 0 0 0 0 0 2 0 0 0 0 0 0 . . . 0 0 0 (4.16) U NV = 0 0 0 r 0 0 0 0 0 0 0 0 0 0 0 S where i , (i = 1, . . . , r) are monic polynomials with the property that i /i1 is polynomial. The matrix S is known as the Smith form of N and the polynomials i the invariant polynomials of N. In this case r is the normal rank of the polynomial matrix N, and the zeros of N are dened as the zeros of one or more of its invariant polynomials. The Smith form S of N may be obtained by applying to N a sequence of elementary row and column operations, which are: 0 0 0

multiply a row/column by a nonzero constant; interchange two rows or two columns; add a row/column multiplied by a polynomial to another row/column.
Example 4.2.3 (Elementary row/column operations). The following sequence of elementary operations brings the polynomial matrix N(s) = s+1 s+2 s1 s2 s1 s2
(1)

to diagonal Smith-form: s+1 s+2 s+1 1 s1 1


(2)

s + 1 2s 1 0

(3)

0 2s 1 0

(4)

1 0

0 . s

Here, in step (1) we subtracted row 1 from row 2; in step (2) we added the rst column to the second column; in step (3), s + 1 times the second row was subtracted from row 1. Finally, in step (4), we interchanged the rows and then divided the (new) second row by a factor 2. Elementary row operations correspond the premultiplication by unimodular matrices, and elementary column operations correspond the postmultiplication by unimodular matrices. For the above four elementary operations these are

160

Chapter 4. Multivariable Control System Design

(1) premultiply by U L1 (s) := (2) postmultiply by V R2 (s) := (3) premultiply by U L3 (s) := (4) premultiply by U L4 (s) :=

1 1 1 0 1 0

0 ; 1 1 ; 1 (s + 1) ; 1

0 1 . 1/2 0

Instead of applying sequentially multiplication on N we may also rst combine the sequence of unimodular matrices into two unimodular matrices U = U L4 U L3 U L1 and V = V R2 , and then apply the U and V, 1 1 1 + s/2 1/2 s/2 U(s) = U L4 (s)U L3 (s)U L1 (s) s+1 s+2 s1 s2 1 1 0 1 V (s) = V R2 (s) = 1 0 0 . s

N(s)

S(s)

This clearly shows that S is the Smith-form of N. The polynomial matrix N has rank 2 and has one zero at s = 0. Now consider a rational matrix P. We write P as P= 1 N d (4.17)

where N is a polynomial matrix and d a scalar polynomial. We immediately get a generalization of the Smith form. Summary 4.2.4 (Smith-McMillan form of a rational matrix). For every rational matrix P there exist unimodular polynomial matrices U and V such that 1 0 0 0 0 0 1 0 2 0 0 0 0 2 .. 0 . 0 0 0 0 (4.18) U PV = 0 0 0 rr 0 0 0 0 0 0 0 0 0 0 0 0 0 0 M where the i and i are coprime such that i = i i d (4.19)

and the i , (i = 1, . . . , r) are the invariant polynomials of N = d P. The matrix M is known as the Smith-McMillan form of P. In this case r is the normal rank of the rational matrix P. The transmission zeros of P are dened as the zeros of i , (i = 1, . . . , r) and the poles are the zeros of i , (i = 1, . . . , r).

4.2. Poles and zeros of multivariable systems

161

Denitions for other notions of zeros may be found in (Rosenbrock 1970), (Rosenbrock 1973) and (Rosenbrock 1974b), and in (Schrader and Sain 1989). Example 4.2.5 (Smith-McMillan form). Consider the transfer matrix P(s) =
1 s 1 s+1 1 s(s+1) 1 s+1

1 1 s+1 1 = N(s). s s s(s + 1) d(s)

The Smith form of N is obtained by the following sequence of elementary operations: s+1 1 s+1 s s s2 1 1 0 0 1 s+1 0 s2 0 s2

Division of each diagonal entry by d yields the Smith-McMillan form


1 s(s+1)

0
s s+1

The set of transmission zeros is {0}, the set of poles is {1, 1, 0}. This shows that transmission zeros of a multivariable transfer matrix may coincide with poles without being canceled if they occur in different diagonal entries of the Smith-McMillan form. In the determinant they do cancel. Therefore from the determinant det P(s) = 1 (s + 1)2

we may not always uncover all poles and transmission zeros of P. (Generally P need not be square so its determinant may not even exist.) An s0 is a pole of P if and only if it is a pole of one or more entries Pi j of P. An s0 that is not a pole of P is a transmission zero of P if and only if the rank of P(s0 ) is strictly less than the normal rank r as dened by the Smith-McMillan form of P. For square invertible matrices P there further holds that s0 is a transmission zero of P if and only it is pole if P1 . For example P(s) = 1 1/s 0 1

1 has a transmission zero at s = 0 even though det P(s) = 1 because P1 (s) = 0 1/s has 1 a pole at s = 0. If the normal rank of an n y nu plant P is r then at most r entries of the output y = Pu be given independent values by manipulating the input u. This generally means that r entries of the output are enough for feedback control. Let K be any nu n y controller transfer matrix, then the sensitivity matrix S and complementary sensitivity matrix T as previously dened are for the multivariable case (see Exercise 1.5.4)

S = (I + PK )1 ,

T = (I + PK )1 PK.

(4.20)

If r < n y , then the n y n y loop gain PK is singular, so that S has one or more eigenvalues = 1 for any s in the complex plane, in particular everywhere on the imaginary axis. In such cases S( j) in whatever norm does not converge to zero as 0. This indicates an undesired situation which is to be prevented by ascertaining the rank of P to equal the number of the be

162

Chapter 4. Multivariable Control System Design

controller outputs. This must be realized by proper selection and application of actuators and sensors in the feedback control system. Feedback around a nonsquare P can only occur in conjunction with a compensator K which makes the series connection PK square, as required by unity feedback. Such controllers K are said to square down the plant P. We investigate down squaring. Example 4.2.6 (Squaring down). Consider P(s) =
1 s 1 s2

Its Smith-McMillan form is M(s) = 1/s2 0 . The plant has no transmission zeros and has a double pole at s = 0. As pre-compensator we propose K0 (s) = 1 . a s+a . s2

This results in the squared-down system P(s)K0 (s) =

Apparently squaring down may introduces transmission zeros. In this example the choice a > 0 creates a zero in the left-half plane, so that subsequent high gain feedback can be applied, allowing a stable closed-loop system with high loop gain. In the example, P does not have transmission zeros. A general 1 2 plant P(s) = a(s) b(s) has transmission zeros if and only if a and b have common zeros. If a and b are in some sense randomly chosen then it is very unlikely that a and b have common zeros. Generally it holds that nonsquare plants have no transmission zeros assuming the entries of Pi j are in some sense uncorrelated with other entries. However, many physical systems bear structure in the entries of their transfer matrix, so that these cannot be considered as belonging to the class of generic systems. Many physically existing nonsquare systems actually turn out to possess transmission zeros. The theory of transmission zero placement by squaring down is presently not complete, although a number of results are available in the literature. Sain and Schrader (1990) describe the general problem area, and results regarding the squaring problem are described in (Horowitz and Gera 1979), (Karcanias and Giannakopoulos 1989), (Le and Safonov 1992), (Sebakhy, ElSingaby, and ElArabawy 1986), (Stoorvogel and Ludlage 1994), and (Shaked 1976). Example 4.2.7 (Squaring down). Consider the transfer matrix P and its Smith-McMillan form 1 2 1 1 s 1 s s(s21) 1 1 s(s2 1) s = P(s) = 1 (4.21) s2 1 s(s2 1) s 0 s2 0 s2 1 1 0 1 0 0 s(s21) s2 1 1 0 1 0 . (4.22) = s s 1 0 s2 1 0 0 0
U 1 (s) M(s) V 1 (s)

There is one transmission zero {0} and the set of poles is {1, 1, 0}. The postcompensator K0 is now 2 3, which we parameterize as K0 = a d b e c . f

4.2. Poles and zeros of multivariable systems

163

We obtain the newly formed set of transmission zeros as the set of zeros of 1 0 a b c s 0 = (cd a f ) + (ce b f )s, det d e f s2 1 where the second matrix in this expression are the rst two columns of U 1 . Thus one single transmission zero can be assigned to a desired location. Choose this zero at s = 1. This leads for example to the choice a = 0, b = 0, c = 1, d = 1, e = 1, f = 0, K0 (s) := 0 1 0 1 1 0

and the squared-down system K0 (s)P(s) = 0


s+1 s s s2 1 1 s(s1)

s(s2

1 0 1) (s + 1)(s2 1)

s2 . s+1

This squared-down matrix may be shown to have Smith-McMillan form


1 s(s2 1)

0 s(s + 1)

and consequently the set of transmission zeros is {0, 1}, the set of poles remain unchanged as {1, 1, 0}. These values are as expected, i.e., the zero at s = 0 has been retained, and the new zero at 1 has been formed. Note that s = 0 is both a zero and a pole.

4.2.2 Transmission zeros of state-space realizations


Any proper n y nu rational matrix P has a state-space realization P(s) = C(sIn A)1 B + D, A nn , B nnu , C n y n , D n y nu .

Realizations are commonly denoted as a quadruple ( A, B, C, D). There are many realizations ( A, B, C, D) that dene the same transfer matrix P. For one, the order n is not xed, A realization of order n of P is minimal if no realization of P exists that has a lower order. Realizations are minimal if and only of the order n equals the degree of the polynomial 1 2 . . . r formed from the denominator polynomials in (4.18), see (Rosenbrock 1970). The poles of P then equal the eigenvalues of A, which, incidentally, shows that computation of poles is a standard eigenvalue problem. Numerical computation of the transmission zeros may be more difcult. In some special cases, computation can be done by standard operations, in other cases specialized numerical algorithms have to be used. Lemma 4.2.8 (Transmission zeros of a minimal state-space system). Let ( A, B, C, D) be a minimal state-space realization of order n of a proper real-rational n y nu transfer matrix P of rank r. Then the transmission zeros s0 C of P as dened in Summary 4.2.4 are the zeros s0 of the polynomial matrix A s0 I C B D
As0 I B C D

(4.23) < n + r.

That is, s0 is transmission zero if and only if rank

164

Chapter 4. Multivariable Control System Design

The proof of this result may be found in Appendix 4.6. Several further properties may be derived from this result. Summary 4.2.9 (Invariance of transmission zeros). The zeros s0 of (4.23) are invariant under the following operations:

nonsingular input space transformations (T2 ) and output space transformations (T3 ):
A sI C B A sI D T3 C BT2 T3 DT2

nonsingular state space transformations (T1 ):


A sI C B T AT 1 sI 1 1 1 D CT1 T1 B D

Static output feedback operations, i.e. for the system x = Ax + Bu, y = Cx + Du, we apply the feedback law u = Ky + v where v acts as new input. Assuming that det(I + DK ) = 0, this involves the transformation:
A sI C B A BK(I + DK )1 C sI D (I + DK )1 C B(I + K D)1 (I + DK )1 D

State feedback (F) and output injection (L) operations,


A sI C B A BF sI D C DF B A BF LC + LDF sI D C DF B LD D

The transmission zeros have a clear interpretation as that of blocking certain exponential inputs, (Desoer and Schulman 1974), (MacFarlane and Karcanias 1976). Example 4.2.10 (Blocking property of transmission zeros). Suppose that the system realization P(s) = C(sI A)1 B + D is minimal and that P is left-invertible, that is rank P = nu . Then B s0 is a transmission zero if and only if As0 I D has less than full column rank, in other words if C and only if there exist vectors xs0 , us0 not both2 zerothat satisfy A s0 I C B D x s0 = 0. u s0 (4.24)

Then the state and input dened as x(t) = xs0 es0 t , u(t) = u s0 es0 t

satisfy the system equations x(t) = Ax(t) + Bu(t), Cx(t) + Du(t) = 0. (4.25)

We see that y = Cx + du is identically zero. Transmission zeros s0 hence correspond to certain exponential inputs whose output is zero for all time.
2 In

fact here neither xs0 nor us0 is zero.

4.2. Poles and zeros of multivariable systems

165

Under constant high gain output feedback u = Ky the nite closed loop poles generally approach the transmission zeros of the transfer matrix P. In this respect the transmission zeros of a MIMO system play a similar role in determining performance limitations as do the zeros in the SISO case. See e.g. (Francis and Wonham 1975). If P is square and invertible, then the transmission zeros are the zeros of the polynomial det A sI C B D (4.26)

If in addition D is invertible, then det A sI C B D I D1 C 0 ( A BD1 C) sI = det 0 I B . D (4.27)

So then the transmission zeros are the eigenvalues of A BD1 C. If D is not invertible then computation of transmission zeros is less straightforward. We may have to resort to computation of the Smith-McMillan form, but preferably we use methods based on state-space realizations. Example 4.2.11 (Transmission zeros of the liquid ow system). Consider the Example 4.1.1 with plant transfer matrix P(s) =
1 s+1 1 s(s+1) 1 s+1 1 s+1

1 s s . s(s + 1) 1 s

Elementary row and column operations successively applied to the polynomial part lead to the Smith form s s (1) s + 1 1 s 1 0 (2) s + 1 s 1 s(s + 1) (3) 1 0 0 0 . s(s + 1)

The Smith-McMillan form hence is M(s) = 1 1 s(s + 1) 0 0 = s(s + 1)


1 s(s+1)

0 . 1

The liquid ow system has no zeros but has two poles {0, 1}. We may compute the poles and zeros also from the state space representation of P, 1 s 0 1 1 1 s 0 1 A sI B . = P(s) = C(sI A)1 B + D, 1 C D 0 0 0 0 1 0 0 The realization is minimal because both B and C are square and nonsingular. The poles are therefore the eigenvalues of A, which indeed are {0, 1} as before. There are no transmission B zeros because det As I D is a nonzero constant (verify this). C Example 4.2.12 (Transmission zeros via state space and transfer matrix representation). Consider the system with state space realization ( A, B, C, D), with 0 0 0 0 1 0 1 0 0 0 A = 1 0 0 , B = 1 0 , C= , D= . 0 0 1 0 0 0 1 0 0 0

166

Chapter 4. Multivariable Control System Design

The systems transfer matrix P equals P(s) = C(sI A)1 B + D = 1/s 1/s2 1 s2 1/s2 3 = 3 1/s s s s . 1

Elementary operations on the numerator polynomial matrix results in s2 s s 0 s 1 1 0 1 0 0 0

so the Smith-McMillan form of P is M(s) = 1 1 s3 0 0 1/s3 = 0 0 0 . 0

The system therefore has three poles, all at zero: {0, 0, 0}. As the matrix A 33 has an equal number of eigenvalues, it must be that the realization is minimal and that its eigenvalues coincide with the poles of the system. From the Smith-McMillan form we see that there are no transmission zeros. This may also be veried from the matrix pencil s 0 0 0 1 1 s 0 1 0 A sI B 1 s 0 0 . = 0 C D 0 1 0 0 0 0 0 1 0 0 Elementary row and column operations do not the rank equals that of s 0 s 0 0 0 1 1 1 s 0 1 0 0 0 0 1 s 0 0 0 0 1 0 1 0 0 0 0 0 0 0 1 0 0 affect the rank of this matrix pencil. Therefore 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 1 0 0 . 0 0

0 0 0 0 1

0 1 0 0 0

0 0 0 0 1

0 1 0 0 0

As s has disappeared from the matrix pencil it is direct that the rank of the matrix pencil does not depend on s. The system hence has no transmission zeros.

4.2.3 Numerical computation of transmission zeros


If a transfer matrix P is given, the numerical computation of transmission zeros is in general based on a minimal state space realization of P because reliable numerical algorithms for rational and polynomial matrix operations are not yet numerically as reliable as state space methods. We briey discuss two approaches. Algorithm 4.2.13 (Transmission zeros via high gain output feedback). The approach has been proposed by Davison and Wang (1974), (1978). Let ( A, B, C, D) be a minimal realization of P and assume that P is either left-invertible or right-invertible. Then: 1. Determine an arbitrary full rank output feedback matrix K nu n y , e.g. using a random number generator.

4.3. Norms of signals and systems

167

2. Determine the eigenvalues of the matrix 1 Z := A + BK( I DK )1 C for various large real values of , e.g. = 1010 , . . . , 1020 at double precision computation. 3. For square systems, the transmission zeros equal the nite eigenvalues of Z . For nonsquare systems, the transmission zeros equal the the nite eigenvalues of Z for almost all choices of the output feedback matrix K. 4. In cases of doubt, vary the values of and K and repeat the calculations.

Algorithm 4.2.14 (Transmission zeros via generalized eigenvalues). The approach has been proposed in (Laub and Moore 1978) and makes use of the QZ algorithm for solving the generalized eigenvalue problem. Let ( A, B, C, D) be a minimal realization of P of order n having the additional property that it is left-invertible (if the system is right-invertible, then use the dual system ( A T , C T , B T , D T )). Then: 1. Dene the matrices M and L as M= In 0 0 , 0 L= A C B . D

2. Compute a solution to the generalized eigenvalue problem i.e. determine all values s and r n+nu satisfying [sM L]r = 0 3. The set of nite values for s are the transmission zeros of the system. Although reliable numerical algorithms exist for the solution of generalized eigenvalue problems, a major problem is to decide which values of s belong to the nite transmission zeros and which values are to be considered as innite. However, this problem is inherent in all numerical approaches to transmission zero computation.

4.3 Norms of signals and systems


For a SISO system with transfer function H the interpretation of |H( j)| as a gain from input to output is clear, but how may we dene gain for a MIMO system with a transfer matrix H? One way to do this is to use norms H( j) instead of absolute values. There are many different norms and in this section we review the most common norms for both signals and systems. The theory of norms leads to a re-assessment of stability.

4.3.1 Norms of vector-valued signals


We consider continuous-time signals z dened on the time axis . These signals may be scalarvalued (with values in or ) or vector-valued (with values in n or n ). They may be added and multiplied by real or complex numbers and, hence, are elements of what is known as a vector space. Given such a signal z, its norm is a mathematically well-dened notion that is a measure for the size of the signal.

168

Chapter 4. Multivariable Control System Design

Denition 4.3.1 (Norms). Let X be a vector space over the real or complex numbers. Then a function : X (4.28)

that maps X into the real numbers is a norm if it satises the following properties: 1. 2. x 0 for all x X (nonnegativity), x = 0 if and only if x = 0 (positive-deniteness),

3. x = || x for every scalar and all x X (homogeneity with respect to scaling), 4. x + y x + y for all x X and y X (triangle inequality).

The pair (X, ) is called a normed vector space. If it is clear which norm is used, X by itself is often called a normed vector space. A well-known norm is the p-norm of vectors in
n

Example 4.3.2 (Norms of vectors in n ). Suppose that x = (x1 , x2 , , xn ) is an n-dimensional complex-valued vector, that is, x is an element of n . Then for 1 p the p-norm of x is dened as x
p

n p 1/ p i=1 |x i |

for 1 p < , for p = .

maxi=1,2, ,n |xi |

(4.29)

Well-known special cases are the norms


n n 1/2

x x
2

=
i=1

|xi |,

=
i=1

|xi |

i=1,2, ,n

max |xi |.

(4.30)

is the familiar Euclidean norm.

Lp

The p-norm for constant vectors may easily be generalized to vector-valued signals. Denition 4.3.3 (L p -norm of a signal). For any 1 p the p-norm or L p -norm z continuous-time scalar-valued signal z, is dened by z
Lp

of a

p |z(t)|

dt

1/ p

for 1 p < , for p = .


n

supt |z(t)|

(4.31)

If z is vector-valued with values in n or z


Lp

, this denition is generalized to (4.32)


n

z(t)

dt

1/ p

for 1 p < , for p = , .

supt z(t)

where is any norm on the n-dimensional space n or The signal norms that we mostly need are the L2 -norm, z
L2

1/2

z(t)

2 2

dt

(4.33)

4.3. Norms of signals and systems

169

and the L -norm, z


L

= sup z(t)
t

(4.34)

2 The square z L2 of the L2 -norm is often called the energy of the signal z, and the L -norm z L its amplitude or peak value.

Example 4.3.4 (L2 and L -norm). Consider the signal z with two entries z(t) = et (t) . 2 e3t (t)

Here (t) denotes the unit step, so z(t) is zero for negative time. From t = 0 onwards both entries of z(t) decay to zero as t increases. Therefore z
L

1 2

= 2.

The square of the L2 -norm follows as e6t e2t +4 2 6 0 The energy of z equals 7/6, its L2 -norm is 7/6. z
L2
2

e2t +4 e6t dt =

t=

=
t=0

7 1 4 + = 2 6 6

4.3.2 Singular values of vectors and matrices


We review the notion of singular values of a matrix3 . If A is an n m complex-valued matrix then the matrices AH A and A AH are both nonnegative-denite Hermitian. As a result, the eigenvalues of both AH A and A AH are all real and nonnegative. The min(n, m) largest eigenvalues i ( AH A), i ( A AH ), i = 1, 2, . . . , min(n, m), (4.35)

of AH A and A AH , respectively, ordered in decreasing magnitude, are equal. The remaining eigenvalues, if any, are zero. The square roots of these min(n, m) eigenvalues are called the singular values of the matrix A, and are denoted i ( A) = 1/2 ( AH A) = 1/2 ( A AH ), i i i = 1, 2, . . . , min(n, m). (4.36)

Obviously, they are nonnegative real numbers. The number of nonzero singular values equals the rank of A. Summary 4.3.5 (Singular value decomposition). Given n m matrix A let be the diagonal n m matrix whose diagonal elements are i ( A), i = 1, 2, . . ., min(n, m). Then there exist square unitary matrices U and V such that A = U V H. (4.37)

A (complex-valued) matrix U is unitary if U H U = UU H = I, with I a unit matrix of correct dimensions. The representation (4.37) is known as the singular value decomposition (SVD) of the matrix A.
3 See

for instance Section 10.8 of Noble (1969).

170

Chapter 4. Multivariable Control System Design

The largest and smallest singular value 1 ( A) and min(n,m) ( A) respectively are commonly denoted by an overbar and an underbar ( A) = 1 ( A), ( A) = min(n,m) ( A).

Moreover, the largest singular value ( A) is a norm of A. It is known as the spectral norm.

There exist numerically reliable methods to compute the matrices U and V and the singular values. These methods are numerically more stable than any available for the computation of eigenvalues. Example 4.3.6 (Singular value decomposition). The singular value decomposition of the 3 1 matrix A given by 0 (4.38) A = 3 4 is A = U V H, where 0 0.6 0.8 U = 0.6 0.64 0.48 , 0.8 0.48 0.36 5 = 0 , 0

V = 1.

(4.39)

The matrix A has a single nonzero singular value (because it has rank 1), equal to 5. Hence, the spectral norm of A is 1 ( A) = 5. Exercise 4.3.7 (Singular value decomposition). Let A = U V H be the singular value decomposition of the n m matrix A, with singular values i , i = 1, 2, , min(n, m). Denote the columns of the n n unitary matrix U as u i , i = 1, 2, , n, and those of the m m unitary matrix V as vi , i = 1, 2, , m. Prove the following statements: 1. For i = 1, 2, , min(n, m) the column vector ui is an eigenvector of A AH corresponding to the eigenvalue i2 . Any remaining columns are eigenvectors corresponding to the eigenvalue 0. 2. Similarly, for i = 1, 2, , min(n, m) the column vector vi is an eigenvector of AH A corresponding to the eigenvalue i2 . Any remaining columns are eigenvectors corresponding to the eigenvalue 0. 3. For i = 1, 2, , min(n, m) the vectors ui and vi satisfy Avi = i ui , AH u i = i vi . (4.40)

Exercise 4.3.8 (Singular values). Given is a square n n matrix A. Prove the following properties (compare p. R-5 of (Chiang and Safonov 1988)): 1. ( A) = maxx 2. ( A) = min x
n

Ax 2 x|2 . Ax 2 x 2 .

3. ( A) |i ( A)| ( A), with i ( A) the ith eigenvalue of A.

4.3. Norms of signals and systems

171

4. If A1 exists then ( A) = 1/( A1 ) and ( A) = 1/( A1 ). 5. ( A) = ||( A), with any complex number. 6. ( A + B) ( A) + (B). 7. ( A B) ( A)(B). 8. ( A) (B) ( A + B) ( A) + (B). 9. max(( A), (B)) ([ A B]) 2 max(( A), (B)). 10. maxi, j | Ai j| ( A) n maxi, j | Ai j|, with Ai j the (i, j) element of A. 11.
n 2 i=1 i ( A)

= tr( AH A).

4.3.3 Norms of linear operators and systems


On the basis of normed signal space we may dene norms of operators that act on these signals. Denition 4.3.9 (Induced norm of a linear operator). Suppose that is a linear map : U Y from a normed space U with norm U to a normed space Y with norm Y . Then the norm of the operator induced by the norms U and Y is dened by = sup
u U =0

u Y . u U

(4.41)

Exercise 4.3.10 (Induced norms of linear operators). The space of linear operators from a vector space U to a vector space Y is itself a vector space. 1. Show that the induced norm as dened by (4.41) is indeed a norm on this space satisfying the properties of Denition 4.3.1. Prove that this norm has the following additional properties: 2. If y = u, then y
Y

for any u U .

3. Submultiplicative property: Let 1 : U V and 2 : V Y be two linear operators, with U , V and Y normed spaces. Then 1 2 1 2 , with all the norms induced.

Constant matrices represent linear maps, so that (4.41) may be used to dene norms of matrices. The spectral norm (M) is the norm induced by the 2-norm (see Exercise 4.3.8(1)). Exercise 4.3.11 (Matrix norms induced by vector norms). Consider the m k complex-valued matrix M as a linear map k m dened by y = Mu. Then depending on the norms dened on k and m we obtain different matrix norms. 1. Matrix norm induced by the 1-norm. Prove that the norm of the matrix M induced by the 1-norm (both for U and Y ) is the maximum absolute column sum M = max
j i

|Mi j|,

(4.42)

with Mi j the (i, j) entry of M.

172

Chapter 4. Multivariable Control System Design

2. Matrix norm induced by the -norm. Prove that the norm of the matrix M induced by the -norm (both for U and Y ) is the maximum absolute row sum M = max
i j

|Mi j|,

(4.43)

with Mi j the (i, j) entry of M. Prove these statements.

4.3.4 Norms of linear systems


We next turn to a discussion of the norm of a system. Consider a system as in Fig. 4.4, which maps the input signal u into an output signal y. Given an input u, we denote the output of the system as y = u. If is a linear operator the system is said to be linear. The norm of the system is now dened as the norm of this operator. u y

Figure 4.4: Input-output mapping system We establish formulas for the norms of a linear time-invariant system induced by the L2 - and

L -norms of the input and output signals.


Summary 4.3.12 (Norms of linear time-invariant systems). Consider a MIMO convolution system with impulse response matrix h, so that its IO map y = u is determined as y(t) =

h()u(t ) d,

t .

(4.44)

Moreover let H denote the transfer matrix, i.e., H is the Laplace transform of h. 1. L -induced norm. The norm of the system induced by the L -norm (4.33) if it exists, is given by
k

i=1,2, ,m j=1

max

|h i j (t)| dt,

(4.45)

where h i j is the (i, j) entry of the m k impulse response matrix h. 2. L2 -induced norm. Suppose H is a rational matrix. The norm of the system induced by the L2 -norm exists if and only if H is proper and has no poles in the closed right-half plane. In that case the L2 -induced norm (4.34) equals the H -norm of the transfer matrix dened as H
H

= sup (H( j)).

(4.46)

A sketch of the proof is found in 4.6. For SISO systems the expressions for the two norms obtained in Summary 4.3.12 simplify considerably.

4.3. Norms of signals and systems

173

Summary 4.3.13 (Norms for SISO systems). The norms of a SISO system with (scalar) impulse response h and transfer function H induced by the L - and L2 -norms are successively given by action of the impulse response h
L1

|h(t)| dt,

(4.47)

and the peak value on the Bode plot if H has no unstable poles H
H

= sup |H( j)|.

(4.48)

If H has unstable poles then the induced norms do not exist.

Example 4.3.14 (Norms of a simple system). As a simple example, consider a SISO rst-order system with transfer function H(s) = 1 , 1 + s
1

(4.49)

with a positive constant. The corresponding impulse response is h(t) = et/ 0


0

for t 0, for t < 0.

(4.50)

It is easily found that the norm of the system induced by the L -norm is h
1

1 t/ e dt = 1. 1 1 = sup = 1. j| 1 + 2 2

(4.51)

The norm induced by the L2 -norm follows as H


H

= sup

|1 +

(4.52)

For this example the two system norms are equal. Usually they are not.

Exercise 4.3.15 (H2 -norm and Hankel norm of MIMO systems). Two further norms of linear time-invariant systems are commonly encountered. Consider a stable MIMO system with transfer matrix H and impulse response matrix h. 1. H2 -norm. Show that the H2 -norm dened as H
H2

tr

H T (j2 f )H( j2 f ) d f

tr

hT (t)h(t) dt

(4.53)

is a norm. The notation tr indicates the trace of a matrix. 2. Hankel norm. The impulse response matrix h denes an operator y(t) =
0

h(t )u() d,

t 0.

(4.54)

which maps continuous-time signals u dened on the time axis (, 0] to continuoustime signals y dened on the time axis [0, ). The Hankel norm H H of the system with impulse response matrix h is dened as the norm of the map given by (4.54) induced by the L2 -norms of u : (, 0] nu and y : [0, ) n y . Prove that this is indeed a norm.

174

Chapter 4. Multivariable Control System Design

3. Compute the H2 -norm and the Hankel norm of the SISO system of Example 4.3.14. Hint: To compute the Hankel norm rst show that if y satises (4.54) and h is given by (4.50) 0 1 then y 2 = 2 ( e/ u() d)2 . From this, prove that H H = 1 . 2 2

Summary 4.3.16 (State-space computation of common system norms). Suppose a system has proper transfer matrix H and let H(s) = C(sIn A)1 B + D be a minimal realization of H. Then 1. H
H2

is nite if and only if D = 0 and A is a stability matrix4 . In that case H H2 = tr BT Y B = tr C XC T

where X and Y are the solutions of the linear equations A X + X AT = B BT , AT Y + Y A = C T C.

The solutions X and Y are unique and are respectively known as the controllability gramian and observability gramian. 2. The Hankel norm H H
H H

is nite only if A is a stability matrix. In that case

max (XY )

where X and Y are the controllability gramian and observability gramian. The matrix XY has real nonnegative eigenvalues only, and max (XY ) denotes the largest of them. 3. The H -norm H bound H
H H

is nite if and only if A is a stability matrix. Then is an upper

<

if and only if (D) < and the 2n 2n matrix A C T C B 0 T ( 2 I DT D)1 DT C AT C D BT (4.55)

has no imaginary eigenvalues. Computation of the H -norm involves iteration on . Proofs are listed in Appendix 4.6.

4.4 BIBO and internal stability


Norms shed a new light on the notions of stability and internal stability. Consider the MIMO input-output mapping system of Fig. 4.4 with linear input-output map . In 1.3.2 we dened the system to be BIBO stable if any bounded input u results in a bounded output y = u. Now bounded means having nite norm, and so different norms may yield different versions of BIBO stability. Normally the L - or L2 -norms of u and y are used. We review a few fairly obvious facts. Summary 4.4.1 (BIBO stability).
4A

constant matrix A is a stability matrix if it is square and all its eigenvalues have negative real part.

4.4. BIBO and internal stability

175

1. If is nite, with the norm induced by the norms of the input and output signals, then the system is BIBO stable. 2. Suppose that the system is linear time-invariant with rational transfer matrix H. Then the system is BIBO stable (both when the L - and the L2 -norms of u and y are used) if and only if H is proper and has all its poles in the open left-half complex plane.

Exercise 4.4.2 (Proof). 1. Prove the statements 1 and 2. 2. Show by a counterexample that the converse of 1 is not true, that is, for certain systems and norms the system may be BIBO stable in the sense as dened while is not necessarily nite. 3. In the literature the following better though more complicated denition of BIBO stability is found: The system is BIBO stable if for every positive real constant N there exists a positive real constant M such that u N implies y M. With this denition the system is BIBO stable if and only if < . Prove this.

Given the above it will be no surprise that we say that a transfer matrix is stable if it is proper and has all its poles in the open left-half complex plane. In 1.3.2 we already dened the notion of internal stability for MIMO interconnected systems: An interconnected system is internally stable if the system obtained by adding external input and output signals in every exposed interconnection is BIBO stable. Here we specialize this for the MIMO system of Fig. 4.5.

Figure 4.5: Multivariable feedback structure

w r e K u P y

Figure 4.6: Inputs r, w and outputs e, u dening internal stability Assume that feedback system shown in Fig. 4.5 is multivariable i.e., u or y have more than one entry. The closed loop of Fig. 4.5 is by denition internally stable if in the extended closed

176

Chapter 4. Multivariable Control System Design

loop of Fig. 4.6 the maps from (w, r) to (e, u) are BIBO stable. In terms of the transfer matrices, these maps are e u = = I K P I
1

r w P(I + K P)1 (I + K P)1 r w

(4.56) (4.57)

I P(I + K P)1 K (I + K P)1 K

Necessary for internal stability is that det(I + P()K()) = 0. Feedback systems that satisfy this nonsingularity condition are said to be well-posed, (Hsu and Chen 1968). The following is a variation of Eqn. (1.44). Lemma 4.4.3 (Internal stability). Suppose P and K are proper, having minimal realizations ( A P , B P , C P , D P ) and ( A K , B K , C K , D K ). Then the feedback system of Fig. 4.5 is internally stable if and only if det(I + D P D K ) = 0 and the polynomial det(I + P(s)K(s)) det(sI A P ) det(sI A K ) has all its zeros in the open left-half plane.

If the n y nu system P is not square, then the dimension n y n y of I + PK and dimension n u nu of I + K P differ. The fact that det(I + PK ) = det(I + K P) allows to evaluate the stability of the system at the loop location having the lowest dimension. Example 4.4.4 (Internal stability of closed-loop system). Consider the loop gain L(s) =
1 s+1 1 s1 1 s+1

The unity feedback around L gives as closed-loop characteristic polynomial: det(I + L(s)) det(sI A L ) = s+2 s+1
2

(s + 1)2 (s 1) = (s + 2)2 (s 1).

(4.58)

The closed loop system hence is not internally stable. Evaluation of the sensitivity matrix (I + L(s))1 =
s+1 s+2 s+1 (s+2)2 s+1 s+2

shows that this transfer matrix in the closed-loop system is stable, but the complementary sensitivity matrix (I + L(s))1 L(s) =
1 s+2 s2 +2s+3 (s1)(s+2)2 1 s+2

is unstable, conrming that the closed-loop system indeed is not internally stable.

4.4. BIBO and internal stability

177

Example 4.4.5 (Internal stability and pole-zero cancellation). Consider P(s) =


1 s 2 s+1 1 s 1 s

K(s) =

1 0

2s s1 2s s1

The unstable pole s = 1 of the controller does not appear in the product PK, P(s)K(s) =
1 s 2 s+1

0
2 s+1

This may suggest as in the SISO case that the map from r to u will be unstable. Indeed the lower-left entry of this map is unstable, K(I + PK )1 = 4s2 (s1)(s+1)(s+3) .

Instability may also be veried using Lemma 4.4.3; it is readily veried that det(I + P(s)K(s)) det(sI A P ) det(sI A K ) = det and this has a zero at s = 1.
s+1 s 2 s+1

0
s+3 s+1

s2 (s + 1)(s 1) = s(s 1)(s + 1)(s + 3)

Exercise 4.4.6 (Non-minimum phase loop gain). In Example 4.1.2 one decoupling precompensator is found that diagonalizes the loop gain PK with the property that both diagonal elements of PK have a right half-plane zero at s = 1. Show that every internally stabilizing decoupling controller K has this property. v r e K u P z

Figure 4.7: Disturbance rejection for MIMO systems Exercise 4.4.7 (Internal model control & MIMO disturbance rejection). Consider the MIMO system shown in Fig. 4.7 and suppose that the plant P is stable. 1. Show that a controller with transfer matrix K stabilizes if and only if Q := K(I + PK )1 is a stable transfer matrix. 2. Express K and S := (I + PK )1 in terms of P and Q. Why is it useful to express K and S in terms of P and Q?
1 1 s 3. Let P be the 2 2 system P(s) = s+1 s 1 and suppose that v is a persistent harmonic j0 t for some known 0 . disturbance v(t) = 1 e 2

178

Chapter 4. Multivariable Control System Design

(a) Assume 0 = 1. Find a stabilizing K (using the parameterization by stable Q) such that in closed loop the effect of v(t) on y(t) is asymptotically rejected (i.e., e(t) 0 as t for r(t) = 0). (b) Does there exist such a stabilizing K if 0 = 1? (Explain.)

4.5 MIMO structural requirements and design methods


During both the construction and the design of a control system it is mandatory that structural requirements of plant and controller and the control conguration are taken into account. For example, the choice of placement of the actuators and sensors may affect the number of righthalf plane zeros of the plant and thereby may limit the closed loop performance, irrespective which controller is used. In this section a number of such requirements will be identied. These requirements show up in the MIMO design methods that are discussed in this section.

4.5.1 Output controllability and functional reproducibility


Various concepts of controllability are relevant for the design of control systems. A minimal requirement for almost any control system, to be included in any control objective, is the requirement that the output of the system can be steered to any desired position in output space by manipulating the input variables. This property is formalized in the concept of output controllability. Summary 4.5.1 (Output controllability). A time-invariant linear system with input u(t) and output y(t) is said to be output controllable if for any y1 , y2 p there exists an input u(t), t [t1 , t2 ] with t1 < t2 that brings the output from y(t1 ) = y1 to y(t2 ) = y2 . In case the system has a strictly proper transfer matrix and is described by a state-space realization ( A, B, C), the system is output controllable if and only if the constant matrix CB C AB C A2 B ... C An1 B (4.59)

has full row rank

If in the above denition C = In then we obtain the denition of state controllability. The concept of output controllability is generally weaker: a system ( A, B, C) may be output controllable and yet not be (state) controllable5. The property of output controllability is an input-output property of the system while (state) controllability is not. Note that the concept of output controllability only requires that the output can be given a desired value at each instant of time. A stronger requirement is to demand the output to be able to follow any preassigned trajectory in time over a given time interval. A system capable of satisfying this requirement is said to be output functional reproducible or functional controllable. Functional controllability is a necessary requirement for output regulator and servo/tracking problems. Brockett and Mesarovic (1965) introduced the notion of reproducibility, termed output controllability by Rosenbrock (1970). Summary 4.5.2 (Output functional reproducibility). A system having proper real-rational n y nu transfer matrix is said to be functionally reproducible if rank P = n y . In particular for functionally reproducible it is necessary that n y nu .
5 Here we assume that C has full row rank, which is a natural assumption because otherwise the entries of y are linearly dependent.

4.5. MIMO structural requirements and design methods

179

Example 4.5.3 (Output controllability and Functional reproducibility). Consider the linear time-invariant state-space system of Example 4.2.12, 0 0 0 0 1 0 1 0 A = 1 0 0 , B = 1 0 , C= . (4.60) 0 0 1 0 1 0 0 0 This forms a minimal realization of the transfer matrix P(s) = C(sI A)1 B =
1 s 1 s2 1 s2 1 s3

(4.61)

The system ( A, B, C) is controllable and observable, and C has full rank. Thus the system also is output controllable. However, the rank of P is 1. Thus the system is not functionally reproducible.

4.5.2 Decoupling control


A classical approach to multivariable control design consists of the design of a precompensator that brings the system transfer matrix to diagonal form, with subsequent design of the actual feedback loops for the various single-input, single-output channels separately. This allows the tuning of individual controllers in separate feedback loops, and it is thought to provide an acceptable control structure providing ease of survey for process operators and maintenance personnel. The subject of noninteracting or decoupling control as discussed in this section is based on the works of Silverman (1970), Williams and Antsaklis (1986). The presentation follows that of Williams and Antsaklis (1996). In this section we investigate whether and how a square P can be brought to diagonal form by applying feedback and/or static precompensation. Suppose that P has state-space representation x(t) y(t) = = Ax(t) + Bu(t), Cx(t) + Du(t)

with u and y having equally many entries, n y = nu = m, i.e., P(s) = C(sI A)1 B + D square. Assume in what follows that P is invertible. The inverse transfer matrix P1 necessarily a rational matrixmay be viewed as a precompensator of P that diagonalizes the series connection P P1 . If the direct feedthrough term D = P() of P is invertible, then the inverse P1 is proper and has realization P1 (s) = D1 C(sI A + BD1 C)1 BD1 + D1 . Exercise 4.5.4. Verify this.
1

If D is singular then the inverse P assuming it existsis not proper. In this case we proceed as follows. Dene the indices f i 0 (i = 1, . . . m) such that D f dened as D f (s) = diag(s f1 , s f2 , . . . , s fm ) is such that (4.62)

D := lim D f (s)P(s)
|s|

(4.63)

is dened and every row of D has at least one nonzero entry. This identies the indices f i uniquely. The f i equal the maximum of the relative degrees of the entries in the ith row of P.

180

Chapter 4. Multivariable Control System Design

Indeed, then by construction the largest relative degree in each row of D f P is zero, and as a consequence lim|s| D f (s)P(s) has a nonzero value at precisely the entries of relative degree zero. The indices f i may also be determined using the realization of P: If the ith row of D is not identical to zero then f i = 0, otherwise f i = min{ k > 0 | row i of C Ak1 B is not identical to zero }. It may be shown that f i n, where n is the state dimension. The indices f i so dened are known as the decoupling indices. Summary 4.5.5 (Decoupling state feedback). If D as dened in (4.63) is nonsingular, then the regular state feedback u(t) = D 1 (C x(t) + v(t)) where C1 A f1 C2 A f2 C = . . . Cm A fm renders the system with output y and new input v decoupled with diagonal transfer matrix D1 (s) = diag(s f1 , s f2 , . . . , s fm ). The compensated plant is shown in Fig. 4.8(b). f Here Ci is denotes the ith row of the matrix C. A proof is given in Appendix 4.6. The decoupled plant D1 has all its poles at s = 0 and it has no transmission zeros. We also know f B that regular state-feedback does not change the zeros of the matrix As I D . This implies that C after state feedback all non-zero transmission zeros of the system are canceled by poles. In other words, after state feedback the realization has unobservable modes at the open loop transmission zeros. This is undesirable if P has unstable transmission zeros as it precludes closed loop stability using output feedback. If P has no unstable transmission zeros, then we may proceed with the decoupled plant D1 and use (diagonal) K to further the design, see Fig. 4.8. A decoupled plant f

C
e K u P y e K v

x u P y

D 1

Decoupled plant D1 f

(a)

(b)

Figure 4.8: (a) closed loop with original plant, (b) with decoupled plant that is stable may offer better perspectives for further performance enhancement by multiloop feedback in individual channels. To obtain a stable decoupled plant we take instead of (4.62) a D f of the form D f = diag( p1 , . . . , pm )

4.5. MIMO structural requirements and design methods

181

where the p i are strictly Hurwitz6 if monic7 polynomials of degree f i . The formulae are now more messy, but the main results holds true also for this choice of D f : Summary 4.5.6 (Decoupling, stabilizing feedback). Suppose the transfer matrix P of the plant is proper, square m m and invertible, and suppose it has no unstable transmission zeros. Let ( A, B, C, D) be a minimal realization of P and let f i be the decoupling indices and suppose that p i are Hurwitz polynomials of degree f i . Then there is a regular state feedback u(t) = Fx(t) + Wv(t) for which the loop from v(t) to y(t) is decoupled. Its realization is controllable 1 and detectable and has a stable transfer matrix diag( p1 , . . . , p1m ). Example 4.5.7. (Diagonal decoupling by state feedback precompensation). Consider a plant with transfer matrix P(s) = 1/s 1/s2 1/s2 . 1/s2 .

A minimal realization of P is 0 0 1 0 0 0 A B = 0 0 C D 0 1 0 0

0 0 0 1 0 0

0 0 0 0 0 1

0 1 1 0 0 0

1 0 1 0 0 0

Its transmission zeros follow after a sequence of elementary row and column operations P(s) = 1 s s2 1 1 1 1 s+1 1 0 1 s2 1 1 s2 0 0 1/s2 = 0 s+1 0 . (s + 1)/s2

Thus P has one transmission zeros at s = 1. It is a stable zero hence we may proceed. The direct feedthrough term D is the zero matrix so the decoupling indices f i are all greater than zero. We need to compute Ci Ak1 B. 0 1 1 0 C1 B = 0 1 0 0 1 1 = 1 0 0 0 it is not identically zero so that f 1 0 1 C2 B = 0 0 0 1 1 0 = 1. Likewise 1 0 = 0 0 . 1 0

It is zero hence we must compute C2 AB, C2 A B = 1 1 .


6A 7A

strictly Hurwitz polynomial is a polynomial whose zeros have strictly negative real part. monic polynomial is a polynomial whose highest degree coefcient equals 1.

182

Chapter 4. Multivariable Control System Design

As it is nonzero we have f 2 = 2. This gives us

D=

C1 B 1 = C2 A B 1

0 . 1

The matrix D is nonsingular so a decoupling state feedback exists. Take D f (s) = s+1 0 0 . (s + 1)2

Its diagonal entries have the degrees f 1 = 1 and f 2 = 2 as required. A realization of D f (s)P(s) = is ( A, B, C , D ) with (s + 1)/s (s + 1)2 /s2 (s + 1)/s2 (s + 1)2 /s2

C=

1 0

1 0 0 2

0 , 1

D=

1 0 . 1 1

Now the regular state feedback u(t) = 1 1 0 1


1

1 0

1 0 0 2

0 x(t) + v(t) 1

renders the system with input v and output y, decoupled. After applying this state feedback we arrive at the state-space system 1 1 1 1 2 1 0 1 0 0 x(t) + 1 0 v(t) x(t) = A + BD 1 (C x(t) + v(t)) = 0 1 0 0 2 1 0 0 0 0 1 0 0 1 0 0 y(t) = Cx(t) + Dv(t) = x(t) 0 0 0 1 (4.64) and its transfer matrix by construction is D1 (s) = f 1/(s + 1) 0 0 1/(s + 1)2

Note that a minimal realization of D1 has order 3 whereas the realization (4.64) has order 4. f Therefore (4.64) is not minimal. This is typical for decoupling procedures. Exercise 4.5.8 (An invertible system that is not decouplable by state feedback). Consider the plant with transfer matrix P(s) = 1/(s + 1) 1/s . 1/(s 1) 1/s

1. Find the transmission zeros. 2. Show that the methods of Summary 4.5.5 and Summary 4.5.6 are not applicable here. 3. Find a stable precompensator C0 that renders the product PC0 diagonal.

4.5. MIMO structural requirements and design methods

183

4.5.3 Directional properties of gains


In contrast to a SISO system, a MIMO system generally does not have a unique gain. A trivial example is the 2 2 system with constant transfer matrix P(s) = 1 0 . 0 2

The gain is in between 1 and 2 depending on the direction of the input. There are various ways to dene gains for MIMO systems. A natural generalization of the SISO gain |y( j)|/|u( j)| from input u( j) to output y( j) is to use norms instead of merely absolute values. We take the 2-norm. Summary 4.5.9 (Bounds on gains). For any given input u and xed frequency there holds ( P( j)) P( j)u( j) u( j) 2
2

( P( j))

(4.65)

and the lower bound ( P( j)) and upper bound ( P( j)) are achieved for certain inputs u.
10
1

10

10

-1

10

-2

10

-3

10

-4

10

-2

10

-1

10

10

10

Figure 4.9: Lower and upper bound on the gain of T ( j) Example 4.5.10 (Upper and lower bounds on gains). Consider the plant with transfer matrix P(s) =
1 s 1 2(s+1) 1 s2 + 2s+1 1 s

In unity feedback the complementary sensitivity matrix T = P(I + P)1 has dimension 2 2. So at each frequency the matrix T ( j) has two singular values. Figure 4.9 shows the two singular values ( j) and ( j) as a function of frequency. Near the crossover frequency the singular values differ considerably. Let P( j) = Y ( j) ( j)U ( j) be an SVD (at each frequency) of P( j), that is, Y ( j) and U( j) are unitary and ( j) = diag(1 ( j), 2 ( j), . . . , min(n y ,nu ) ( j)) (4.66)

184

Chapter 4. Multivariable Control System Design

with 1 ( j) 2 ( j) . . . min(n y ,nu ) ( j) 0. The columns uk ( j) of U( j) the input principal directions, (Postlethwaite, Edmunds, and MacFarlane 1981) and they have the special property that their gains are precisely the corresponding singular values P( j)uk ( j) uk ( j) 2
2

= k ( j)

and the response yk ( j) = P( j)u k ( j) in fact equals k ( j) times the kth column of Y ( j). The columns of Y ( j) are the output principal directions and the k ( j) the principal gains. The condition number dened as ( P( j)) = ( P( j)) ( P( j)) (4.67)

is a possible measure of the difculty of controlling the system. If for a plant all principal gains are the same, such as when P( j) is unitary, then ( j) = 1. If ( j) 1 then the loop gain L = PK around the crossover frequency may be difcult to shape. Note that the condition number is not invariant under scaling. Describing the same system in terms of a different set of physical units leads to a different condition number. A high condition number indicates that the system is close to losing its full rank, i.e. close to not satisfying the property of functional controllability. Example 4.5.11 (Multivariable control systemdesign goals). The performance of a multivariable feedback system can be assessed using the notion of principal gains as follows. The basic idea follows the ideas previously discussed for single input, single output systems. The disturbance rejecting properties of the loop transfer require the sensitivity matrix to be small: (S( j)) 1.

At the same time the restriction of the propagation of measurement noise requires (T ( j)) 1.

As in the SISO case these two requirements are conicting. Indeed, using the triangle inequality we have the bounds |1 (S( j))| (T ( j)) 1 + (S( j)) and |1 (T ( j))| (S( j)) 1 + (T ( j)). (4.69) (4.68)

It is useful to try to make the difference between (T ) and (T ), and between (S) and (S), not too large in the cross over region. Making them equal would imply that the whole plant behaves identical in all directions and this is in general not possible.

4.5. MIMO structural requirements and design methods

185

We must also be concerned to make the control inputs not too large. Thus the transfer matrix (I + K P)1 K from r to u must not be too large. u( j) r( j)
2 2

(I + K( j)P( j))1 K( j)r( j) r( j) 2 ((I + K( j)P( j))1 K( j))

((I + K( j)P( j))1 )(K( j)) (K( j)) (I + K( j)P( j)) (K( j)) 1 (K( j)P( j)) (K( j)) . (4.70) 1 (K( j))( P( j)) Here several properties of Exercise 4.3.8 are used and in the last two inequalities it is assumed that (K( j))( P( j)) < 1. The upperbound shows that where ( P( j)) is not too large that (K( j)) 1 guarantees a small gain from r( j) to u( j). The upper bound (4.70) of this gain is easily determined numerically. The direction of an important disturbance can be taken into consideration to advantage. Let v( j) be a disturbance acting additively on the output of the system. Complete rejection of the disturbance would require an input u( j) = P1 ( j)v( j). Thus u( j) v( j)
2 2

(4.71)

P1 ( j)v( j) v( j) 2

(4.72)

measures the magnitude of u needed to reject a unit magnitude disturbance acting in the direction v. The most and least favorable disturbance directions are those for which v is into the direction of the principal output directions corresponding to ( P( j)) and ( P( j)), respectively. This leads to the disturbance condition number of the plant, v ( P( j)) = P1 ( j)v( j) v( j) 2
2

( P( j))

(4.73)

which measures the input needed to reject the disturbance v relative to the input needed to reject the disturbance acting in the most favorable direction. It follows that 1 v ( P) ( P) (4.74) and v ( P) generalizes the notion of condition number of the plant dened in (4.67). 5rejection properties with respect to these disturbance directions.

4.5.4 Decentralized control structures


Dual to decoupling control where the aim is make the plant P diagonal, is decentralized control where it is the controller K that is restricted to be diagonal or block diagonal, K1 0 K2 (4.75) K = diag{Ki } = . .. . 0 Km

186

Chapter 4. Multivariable Control System Design

This control structure is called multiloop control and it assumes that there as many control inputs as control outputs. In a multiloop control structure it is in general desirable to make an ordering of the input and output variables such that the interaction between the various control loops is as small as possible. This is the input-output pairing problem. Example 4.5.12 (Relative gain of a two-input-two-output system). Consider a two-input-twooutput system y1 u =P 1 y2 u2 and suppose we decide to close the loop by a diagonal controller u j = K j (r j y j ), that is, the rst component of u is only controlled by the rst component of y, and similarly u2 is controlled only by y2 . Suppose we leave the second control loop open for the moment and that we vary the rst control loop. If the open loop gain from u2 to y2 does not change a lot as we vary the rst loop u1 = K1 (r1 y1 ), it is save to say that we can design the second control loop independent from the rst, or to put it differently, that the second control loop is insensitive to tuning of the rst. This is desirable. In summary, we want the ratio := gain from u2 to y2 if rst loop is open gain from u2 to y2 if rst loop is closed (4.76)

preferably close to 1. To make this more explicit we assume now that the reference signal r is constant and that all signals have settled to their constant steady state values. Now if the rst loop is open then the rst input entry u1 is zero (or constant). However if the rst loop is closed thenassuming perfect controlit is the output y1 that is constant (equal to r1 ). This allows us to express (4.76) as = dy2 /du2 dy2 /du2
u1 y1

where u1 expresses that u1 is considered constant in the differentiation. This expression for exists and may be computed if P(0) is invertible: dy2 du2 =
u1

d P21 (0)u1 + P22 (0)u2 du2

= P22 (0)
u1

and because u = P1 (0)y we also have dy2 du2 =


y1

1 du2 /dy2

=
y1

1
1 1 d P21 (0)y1 +P22 (0)y2 y1 dy2

1
1 P22 (0)

1 = P22 (0).

1 The relative gain hence equals P22 (0)P22 (0).

For general square MIMO systems P we want to consider the relative gain array (RGA) which is the (rational matrix) dened element-wise as
ij

= Pi j P1 ji

or, equivalently as a matrix, as = P ( P1 )T (4.77)

where denotes the Hadamard product which is the entry-wise product of two matrices of the same dimensions.

4.5. MIMO structural requirements and design methods

187

Summary 4.5.13 (Properties of the RGA). 1. In constant steady state there holds that
i j (0)

dyi du j all loops (uk , yk ), k = j open dyi du j all loops (uk , yk ), k = j closed

2. The sum of the elements of each row or each column of

is 1

3. Any permutation of rows and columns in P results in the same permutations of rows and columns in 4. is invariant under diagonal input and output scaling = Im

5. If P is diagonal or upper or lower triangular then

6. The norm ( ( j)) of the RGA is believed to be closely related to the minimized condition number of P( j) under diagonal scaling, and thus serves as an indicator for badly conditioned systems. See (Nett and Manousiouthakis 1987), (Grosdidier, Morari, and Holt 1985), and (Skogestad and Morari 1987). 7. The RGA measures the sensitivity of P to relative element-by-element uncertainty. It indicates that P becomes singular by an element perturbation from Pi j to Pi j[1 1 ]. ij If deviates a lot from Im then this is an indication that interaction is present in the system.

Example 4.5.14 (Relative Gain Array for 2 2 systems). For the two-input, two-output case P= P11 P21 P12 , P22 ( P1 )T = 1 P22 P11 P22 P21 P12 P12 P21 , P11

the relative gain array = 1

= P PT equals := P11 P22 . P11 P22 P21 P12

1 ,

We immediate see that rows and columns add up to one, as claimed. Some interesting special cases are P= P= P= P= P11 0 P11 P21 0 P21 p p 0 P22 0 P22 P12 0 p p = = = = 1 0 1 0 0 1 0 1 0 1 1 0

0.5 0.5 0.5 0.5

In the rst and second example, the relative gain suggests to pair u1 with y1 , and u2 with y1 . In the rst example this pairing is also direct from the fact that P is diagonal. In the third example the RGA is antidiagonal, so we want to pair u1 with y2 and u2 with y1 . In the fourth example all entries of the RGA are the same, hence no pairing can be deduced from the RGA.

188

Chapter 4. Multivariable Control System Design

If P is close to singular then several entries of are large. The corresponding pairings are to be avoided, and possibly no sensible pairing exists. For example P= with 0. a a a a(1 + ) =
1+ 1 1 1+

Although the RGA has been derived originally by Bristol (1966) for the evaluation of P(s) at steady-state s = 0 assuming stable steady-state behavior, the RGA may be useful over the complete frequency range. The following i/o pairing rules have been in use on the basis of an heuristic understanding of the properties of the RGA: 1. Prefer those pairing selections where ( j) is close to unity around the crossover frequency region. This prevents undesirable stability interaction with other loops. 2. Avoid pairings with negative steady-state or low-frequency values of ii ( j) on the diagonal. The following result by Hovd and Skogestad (1992) makes more precise the statement in (Bristol 1966) relating negative entries in the RGA, and nonminimum-phase behavior. See also the counterexample against Bristols claim in (Grosdidier and Morari 1987). Summary 4.5.15 (Relative Gain Array and RHP transmission zeros). Suppose that P has stable elements having no poles nor zeros at s = 0. Assume that the entries of are nonzero and nite for s . If i j shows different signs when evaluated at s = 0 and s = , then at least one of the following statements holds: 1. The entry Pi j has a zero in the right half plane 2. P has a transmission zero in the right half plane 3. The subsystem of P with input j and output i removed has a transmission zero in the right half plane

As decentralized or multiloop control structures are widely applied in industrial practice, a requirement arises regarding the ability to take one loop out of operation without destabilizing the other loops. In a problem setting employing integral feedback, this leads to the notion of decentralized integral controllability (Campo and Morari 1994). Denition 4.5.16 (Decentralized integral controllability (DIC)). The system P is DIC if there exists a decentralized (multiloop) controller having integral action in each loop such that the feedback system is stable and remains stable when each loop i is individually detuned by a factor i for 0 i 1. The denition of DIC implies that the system P must be open-loop stable. It is also assumed that the integrator in the control loop is put out of order when i = 0 in loop i. The steady-state RGA provides a tool for testing on DIC. The following result has been shown by Grosdidier, Morari, and Holt (1985). Summary 4.5.17 (Steady-state RGA and stability). Let the plant P be stable and square, and consider a multiloop controller K with diagonal transfer matrix having integral action in each diagonal element. Assume further that the loop gain PK is strictly proper. If the RGA (s) of the plant contains a negative diagonal value for s = 0 then the closed-loop system satises at least one of the following properties:

4.5. MIMO structural requirements and design methods

189

1. The overall closed-loop system is unstable 2. The loop with the negative steady-state relative gain is unstable by itself 3. The closed-loop system is unstable if the loop corresponding to the negative steady-state relative gain is opened

A further result regarding stability under multivariable integral feedback has been provided by Lunze (1985); see also (Grosdidier, Morari, and Holt 1985), and (Morari 1985). Result 4.5.18 (Steady-state gain and DIC). Consider the closed loop system consisting of the open-loop stable plant P and the controller K having transfer matrix K(s) := Kss s (4.78)

Assume unity feedback around the loop and assume that the summing junction in the loop introduces a minus sign in the loop. Then a necessary condition for the closed loop system to be stable for all in an interval 0 with > 0 is that the matrix P(0)Kss must have all its eigenvalues in the open right half plane. Further results on decentralized i/o pairing can be found in (Skogestad and Morari 1992), and in (Hovd and Skogestad 1994). The inherent control limitations of decentralized control have been studied by Skogestad, Hovd, and Lundstr m (1991). Further results on decentralized o integral controllability can be found in (Le, Nwokah, and Frazho 1991) and in (Nwokah, Frazho, and Le 1993).

4.5.5 Internal model principle and the servomechanism problem


The servomechanism problem is to design a controller called servocompensator that renders the plant output as insensitive as possible to disturbances while at the same asymptotically track certain reference inputs r. We discuss a state space solution to this problem; a solution that works for SISO as well as MIMO systems. Consider a plant with state space realization x(t) y(t) = = Ax(t) + Bu(t) Cx(t) + v(t) u(t) nu , y n y , x(0) = x0 n (4.79)

and assume that v is a disturbance that is generated by a dynamic system having all of its poles on the imaginary axis xv (t) v(t) = = Av xv (t), Cv xv (t) xv (0) = xv0 (4.80)

The disturbance is persistent because the eigenvalues fof Av are assumed to lie on the imaginary axis. Without loss of generality we further assume that ( Av , Cv ) is an observable pair. Then a possible choice of the servocompensator is the compensator (controller) with realization xs (t) e(t) = As xs (t) + Bs e(t), xs (0) = xs0 (4.81) = r(t) y(t)

where A s is block-diagonal with n y blocks As = diag( Av , Av , . . . , Av ) (4.82)

190

Chapter 4. Multivariable Control System Design

and where Bs is of compatible dimensions and such that ( As , Bs ) is controllable, but otherwise arbitrary. The composite system, consisting of the plant (4.79) and the servocompensator (4.81) yields x(t) xs (t) y(t) = = A Bs C Bs C 0 As 0 x(t) B 0 (r(t) v(t)) + u(t) + xs (t) 0 Bs x(t) + v(t) xs (t)

as its stands the closed loop is not stable, but it may be stabilized by an appropriate state feedback u(t) = Fx(t) + Fs xs (t) (4.83)

where F and Fs are chosen by any method for designing stabilizing state feedback, such as LQ theory. The resulting closed loop is shown in Fig. 4.10. v r e K xs
Fs F

P x

Figure 4.10: Servomechanism conguration Summary 4.5.19 (Servocompensator). Suppose that the following conditions are satised. 1. ( A, B) is stabilizable and ( A, C) is detectable, 2. n u n y , i.e. there are at least as many inputs as there are outputs 3. rank
Ai B C 0

= n + n y for every eigenvalue i of Av .

Then F, Fs can be selected such that the closed loop system matrix A + BF Bs C B Fs As (4.84)

has all its eigenvalues in the open left-half plane. In that case the controller (4.81),(4.83) is a servocompensator in that e(t) 0 as t for any of the persistent v that satisfy (4.80). The system matrix As of the servocompensator contains several copies of the system matrices Av that denes v. This means that the servocompensator contains the mechanism that generates v; it is an example of the more general internal model principle. The conditions as mentioned imply the following, see (Davison 1996) and (Desoer and Wang 1980).

No transmission zero of the system should coincide with one of the imaginary axis poles of the disturbance model The system should be functionally controllable

4.5. MIMO structural requirements and design methods

191

The variables r and v occur in the equations in a completely interchangeable role, so that the disturbance model used for v can also be utilized for r. That is, for any r that satises
xv (t) = Av xv (t), r(t) = Cr xv (t)

the error signal e(t) converges to zero. Hence the output y(t) approaches r(t) as t .

Asymptotically, when e(t) is zero, the disturbance v(t) is still present and the servocompensator acts in an open loop fashion as the autonomous generator of a compensating disturbance at the output y(t), of equal form as v(t) but of opposite sign. If the state of the system is not available for feedback, then the use of a state estimator is possible without altering the essentials of the preceding results.
Example 4.5.20 (Integral action). If the servocompensator is taken to be an integrator xs = e then constant disturbances v are asymptotically rejected and constant reference inputs r asymptotically tracked. We shall verify the conditions of Summary 4.5.19 for this case. As before the plant P is assumed strictly proper with state space realization and corrupted with noise v, x(t) = Ax(t) + Bu(t), y(t) = Cx(t) + v(t).

Combined with the integrating action of the servocompensator xs (t) = e(t), with e(t) = r(t) y(t), we obtain x(t) x s (t) = = A 0 A C 0 0 0 0 B x(t) 0 + u(t) + e(t) 0 xs (t) I B 0 x + u(t) + (r(t) v(t)). 0 I xs

Closing the loop with state feedback u(t) = Fx(t) + Fs xs (t) renders the closed loop state space representation x(t) xs (t) y(t) = = A + BF C C 0 B Fs 0 x(t) 0 + (r(t) v(t)) xs (t) I

x(t) . xs (t)

All closed-loop poles can be arbitrarily assigned by state feedback u(t) = Fx(t) + Fs xs (t) if and A only if C 0 , B is controllable. This is the case if and only if 0 0 B 0 AB CB A2 B C AB

has full row rank. The above matrix may be decomposed as A C B 0 0 Im B 0 AB 0

and therefore has full row rank if and only if 1. ( A, B) is controllable, 2. n u n y ,

192

Chapter 4. Multivariable Control System Design

3. rank

A C

B = n + n y. 0

(In fact the third condition implies the second.) The three conditions are exactly what Summary 4.5.19 states. If the full state x is not available for feedback then we may replace the feedback law u(t) = Fx(t) + Fs xs (t) by an approximating observer, x(t) u(t) = = ( A KC) x(t) + Bu(t) + Ky(t) F x(t) + Fs xs (t).

4.6 Appendix: Proofs and Derivations


Derivation of process dynamic model of the two-tank liquid ow process. A dynamic process model for the liquid ow process of gure 4.1 is derived under the following assumptions:

The liquid is incompressible; The pressure in both vessels is atmospheric; The ow 2 depends instantaneously on the static pressure dened by the liquid level H1 ; The ow 4 is the instantaneous result of a ow-controlled recycle pump; 3 is determined by a valve or pump outside the system boundary of the process under consideration, independent of the variables under consideration in the model.
Let denote density of the liquid expressed in kg/m3 , and let A1 and A2 denote the crosssectional areas (in m2 ) of each vessel, then the mass balances over each vessel and a relation for outow are: A1 h1 (t) A2 h2 (t) = = [1 (t) + 4 (t) 2 (t)] [2 (t) 4 (t) 3 (t)] (4.85) (4.86)

The ow rate 2 (t) of the rst tank is assumed to be a function of the level h1 of this tank. The common model is that 2 (t) = k h1 (t). (4.87)

where k is a valve dependent constant. Next we linearize these equations around an assumed equilibrium state (h k,0 , i,0 ). Dene the deviations from the equilibrium values with tildes, that is, i (t) h k (t) = = i,0 + i (t), h k,0 + hk (t), i = 1 . . .4 k = 1, 2.

(4.88)

Linearization of equation (4.87) around (h k,0 , i,0 ) gives 2 (t) = k 1 2,0 h1 (t). h1 (t) = 2 2h1,0 h1,0 (4.89)

4.6. Appendix: Proofs and Derivations

193

Inserting this relation into equations (4.85), (4.86) yields the linearized equations
2 A12,01,0 h1 (t) h = 2,0 (t) 2 h 2 A1 h1,0

0 0

h1 (t) 2 (t) + h

1 A1

1 A1 1 A2

0 1 (t) 1 4 (t) + A 3 (t) 2

(4.90)

Assuming that 4,0 = 1,0 , and thus 2,0 = 21,0 , and taking numerical values 1,0 = 1, A1 = A2 = 1, and h1,0 = 1 leads to Equation (4.1). Proof of Summary 4.3.16. 1. If D is not zero then the impulse response matrix h(t) = C e At B(t) + (t)D contains Dirac delta functions and as a result the H2 -norm can not be nite. So D = 0 and h(t) = C e At B(t). If A is not a stability matrix then h(t) is unbounded, hence the H2 -norm is innite. So D = 0 and A is a stability matrix. Then H
2

H2

= = =

tr

h(t)T h(t) dt BT e A t C T C e At B dt
0 Y
T

tr
0

tr BT

e A t C T C e At dt B .

The so dened matrix Y satises AT Y + Y A =


0

AT e A t C T C e At + e A t C T C e At A dt = e A t C T C e At

t= t=0

= C T C. (4.91)

That is, Y satises the Lyapunov equation AT Y + Y A = C T C, and as A is a stability matrix the solution Y of (4.91) is well known to be unique. 2. Let X and Y be the controllability and observability gramians. By the property of state, the output y for positive time is a function of the state x0 at t = 0. Then
0 T

y(t)T y(t) dt =
0

xT e A t C T C e At x0 dt = xT Y x0 . 0 0

If X satises A X + X A = BBT then its inverse Z := X 1 satises Z A + AT Z = Z B BT Z. By completion of the square we may write d T x Zx dt = = = 2xT Z x = 2xT (Z Ax + Z Bu) = xT (Z A + AT Z)xT + xT (2Z Bu) xT (Z B BT Z)xT + 2xT Z Bu uT u (u BT Zx)T (u BT Zx).

Therefore, assuming x() = 0,


0

u(t)

2 2

u(t) BT Zx(t)

2 2

dt = xT (t)Zx(t)

t=0 t=

= xT Zx0 = xT X 1 x0 . 0 0

194

Chapter 4. Multivariable Control System Design

From this expression it follows that the smallest u (in norm) that steers x from x() = 0 to x(0) = x0 is u(t) = BT Zx(t) (verify this). This then shows that sup
u T 0 y (t)y(t) dt 0 T u (t)u(t) dt

xT Y x0 0 . T X 1 x x0 0

Now this supremum is less then some 2 for some x0 if and only if xT Y x0 < 2 xT X 1 x0 . 0 0 There exist such x0 iff Y < 2 X 1 , which in turn is equivalent to that i (Y X) = i (X 1/2 Y X 1/2 ) < 2 . This proves the result. 3. H H < holds if and only if H is stable and 2 I H H is positive denite on j . This is the case iff it is positive denite at innity (i.e., (D) < , i.e. 2 I DT D > 0) and nowhere on the imaginary axis 2 I H H is singular. We will show that 2 I H H has imaginary zeros iff (4.55) has imaginary eigenvalues. It is readily veried that a realization of 2 I H H is B A sI 0 C T C AT sI CT D T T 2 T D C B ID D By the Schur complement results applied to this matrix we see that det A sI C T C 0 det( 2 I H H) = det( 2 I DT D) det( Aham sI ). AT sI

where Aham is the Hamiltonian matrix (4.55). As A has no imaginary eigenvalues we have that 2 I H H has no zeros on the imaginary axis iff Aham has no imaginary eigenvalues.

Proof of Summary 4.5.5. We make use of the fact that if x(t) y(t) y(t) = = = Ax(t) + Bu(t), Cx(t) C x(t) = C Ax(t) + C Bu(t) x(0) = 0,

then the derivative y (t) satises The matrix D f (s) = diag(s f1 , . . . , s fm ) is a diagonal matrix of differentiators. Since the plant y = Pu satises x(t) y(t) = = Ax(t) + Bu(t), Cx(t) + Du(t) x(t) = y1 (t) y2 (t) = . . . ym (t) Ax(t) + Bu(t), x(0) = 0, f1 C1 A C2 A f2 . x(t) + D u(t) . . Cm A fm
C

x(0) = 0, (4.92)

it follows that the component-wise differentiated v dened as v := D f y satises v(t) =


d f1 dtff 1 d 2 dt f 2 d fm dt fm

4.6. Appendix: Proofs and Derivations

195

By assumption D is nonsingular, so u(t) = D 1 (C x(t) + v(t)). Inserting this in the state realization (4.92) yields x(t) y(t) = = ( A BD 1C )x(t) + BD 1 v(t), (C DD 1 C )x(t) + DD 1 v(t) x(0) = 0,

Since v = D f y it must be that the above is a realization of the D1 . f Proof of Lemma 4.2.8 (sketch). We only prove it for the case that none of the transmission zeros s0 are poles. Then ( A s0 In )1 exists for any transmission zero, and s0 is a transmission zero if and only if P(s0 ) drops rank. Since A s 0 In C B D In 0 (s0 In A)1 B A s 0 In = I C 0 P(s0 )

B we see that the rank of As0 I D equals n plus the rank of P(s0 ). Hence s0 is a transmission zero C As0 In B of P if and only if C D drops rank.

Proof of Norms of linear time-invariant systems. We indicate how the formulas for the system norms as given in Summary 4.3.12 are obtained. 1. L -induced norm. In terms of explicit sums, the ith component yi of the output y = h u of the convolution system may be bounded as |yi (t)| = so that y
L
j j

hi j ()u j (t ) d |h i j ()| d u
L ,

|h i j ()| |u j (t )| d t , (4.93)

max
i

|h i j ()| d u

L .

(4.94)

This shows that max


i j

|h i j ()| d.

(4.95)

The proof that the inequality may be replaced by equality follows by showing that there exists an input u such that (4.94) is achieved with equality. 2. L2 -induced norm. By Parsevals theorem (see e.g. (Kwakernaak and Sivan 1991)) we have that y u
L2 L2
2 2

= = =

H y (t)y(t) dt H u (t)u(t) dt 1 H 2 y ( j) y( j) d 1 H ( j)u( j) d 2 u H H u ( j)H ( j)H( j)u( j) d H ( j)u( j) d u

(4.96) (4.97)

196

Chapter 4. Multivariable Control System Design

where y is the Laplace transform of y and u that of u. For any xed frequency we have that uH ( j)H H ( j)H( j)u( j) = 2 (H( j)). uH ( j)u( j) u( j) sup Therefore sup
u

y u

L2 L2
2

sup 2 (H( j)).

(4.98)

The right hand side of this inequality is by denition H 2 . The proof that the inequality H may be replaced by equality follows by showing that there exists an input u for which (4.98) is achieved with equality within an arbitrarily small positive margin .

5 Uncertainty Models and Robustness

Overview Various techniques exist to examine the stability robustness of control systems subject to parametric uncertainty. Parametric and nonparametric uncertainty with varying degree of structure may be captured by the basic perturbation model. The size of the perturbation is characterized by bounds on the norm of the perturbation. The small gain theorem provides the tool to analyze the stability robustness of this model. These methods allow to generalize the various stability robustness results for SISO systems of Chapter 2 in several ways.

5.1 Introduction
In this chapter we discuss various paradigms for representing uncertainty about the dynamic properties of a plant. Moreover, we present methods to analyze the effect of uncertainty on closed-loop stability and performance. Section 5.2 is devoted to parametric uncertainty models. The idea is to assume that the equations that describe the dynamics of the plant and compensator (in particular, their transfer functions or matrices) are known, but that there is uncertainty about the precise values of various parameters in these equations. The uncertainty is characterized by an interval of possible values. We discuss some methods to analyze closed-loop stability under this type of uncertainty. The most famous recent result is Kharitonovs theorem. In 5.3 we introduce the so-called basic perturbation model for linear systems, which admits a much wider class of perturbations, including unstructured perturbations. Unstructured perturbations may involve changes in the order of the dynamics and are characterized by norm bounds. Norm bounds are bounds on the norms of operators corresponding to systems. In 5.4 we review the xed point theorem of functional analysis to the small gain theorem of system theory. With the help of that we establish in 5.5 sufcient and necessary conditions for the robust stability of the basic perturbation model. Next, in 5.6 the basic stability robustness 197

198

Chapter 5. Uncertainty Models and Robustness

result is applied to proportional, proportional inverse perturbations and fractional perturbations of feedback loops. The perturbation models of 5.6 are relatively crude. Doyles structured singular value allows much ner structuring. It is introduced in 5.7. An appealing feature of structured singular value analysis is that it allows studying combined stability and performance robustness in a unied framework. This is explained in 5.8. In 5.9 a number of proofs for this chapter are presented.

5.2 Parametric robustness analysis


5.2.1 Introduction
In this section we review the parametric approach to robustness analysis. In this view of uncertainty the plant and compensator transfer functions are assumed to be given, but contain several parameters whose values are not precisely known. We illustrate this by an example.

reference input r prefilter F

error signal e + compensator C

plant input u plant P output z

Figure 5.1: Feedback system Example 5.2.1 (Third-order uncertain plant). Consider a feedback conguration as in Fig. 5.1 where the plant transfer function is given by P(s) = g . s2 (1 + s) (5.1)

The gain g is not precisely known but is nominally equal to g0 = 1 [s2 ]. The number is a parasitic time constant and nominally 0 [s]1 . A system of this type can successfully be controlled by a PD controller with transfer function C(s) = k + Td s. The latter transfer function is not proper so we modify it to C(s) = k + Td s , 1 + T0 s (5.2)

where the time constant T0 is small so that the PD action is not affected at low frequencies. With these plant and compensator transfer functions, the return difference J = 1 + L = 1 + PC is given by J(s) = 1 + g s2 (1 + s) T0 s4 + ( + T0 )s3 + s2 + gTd s + gk k + Td s = . 1 + T0 s s2 (1 + s)(1 + T0 s) (5.3)

Hence, the stability of the feedback system is determined by the locations of the roots of the closed-loop characteristic polynomial (s) = T0 s4 + ( + T0 )s3 + s2 + gTd s + gk.
1 All

(5.4)

physical units are SI. From this point on they are usually omitted.

5.2. Parametric robustness analysis

199

Presumably, the compensator can be constructed with sufcient precision so that the values of the parameters k, Td , and T0 are accurately known, and such that the closed-loop system is stable at the nominal plant parameter values g = g0 = 1 and = 0. The question is whether the system remains stable under variations of g and , that is, whether the roots of the closed-loop characteristic polynomial remain in the open left-half complex plane. Because we pursue this example at some length we use the opportunity to demonstrate how the compensator may be designed by using the classical root locus design tool. We assume that the ground rules for the construction of root loci are known.

k=

Im k=0

k = Re

Figure 5.2: Root locus for compensator consisting of a simple gain Example 5.2.2 (Compensator design by root locus method). Given the nominal plant with transfer function g0 (5.5) P0 (s) = 2 , s the simplest choice for a compensator would be a simple gain C(s) = k. Varying k from 0 to or from 0 to results in the root locus of Fig. 5.2. For no value of k stability is achieved. Modication to a PD controller with transfer function C(s) = k + sTd amounts to addition of a zero at k/ Td . Accordingly, keeping k/ Td xed while varying k the root locus of Fig. 5.2 is altered to that of Fig. 5.3 (only the part for k 0 is shown). We assume that a closed-loop bandwidth of 1 is required. This may be achieved by placing the two closed-loop poles at 1 2 2(1 j). The distance of this pole pair from the origin is 1, resulting in the desired closedloop bandwidth. Setting the ratio of the imaginary to the real part of this pole pair equal to 1 ensures an adequate time response with a good compromise between rise time and overshoot.

Im

Re zero at k/Td double pole at 0

Figure 5.3: Root locus for a PD compensator The closed-loop characteristic polynomial corresponding to the PD compensator transfer function C(s) = k + sTd is easily found to be given by s2 + g0 Td s + g0 k. (5.6)

200

Chapter 5. Uncertainty Models and Robustness

Choosing g0 Td = 2 and g0 k = 1 (i.e., Td = 2 and k = 1) places the closed-loop poles at the 1 desired locations. The zero in Fig. 5.3 may now be found at k/ Td = 2 2 = 0.7071.

Im

Re pole at 1/To zero at k/Td double pole at 0

Figure 5.4: Root locus for the modied PD compensator The nal step in the design is to make the compensator transfer function proper by changing it to C(s) = k + Td s . 1 + T0 s (5.7)

This amounts to adding a pole at the location 1/ T0. Assuming that T0 is small the root locus now takes the appearance shown in Fig. 5.4. The corresponding closed-loop characteristic polynomial is T0 s3 + s2 + g0 Td s + g0 k. (5.8) Keeping the values Td = 2 and k = 1 and choosing T0 somewhat arbitrarily equal to 1/10 (which places the additional pole in Fig. 5.4 at 10) results in the closed-loop poles 8.4697. (5.9) The dominant pole pair at 1 2(1 j) has been slightly shifted and an additional non-dominant 2 pole at 8.4697 appears. 0.7652 j0.7715,

5.2.2 Routh-Hurwitz Criterion


In the parametric approach, stability robustness analysis comes down to investigating the roots of a characteristic polynomial of the form (s) = n ( p)sn + n1 ( p)sn1 + + 0 ( p), (5.10)

whose coefcients n ( p), n1 ( p), , 0 ( p) depend on the parameter vector p. Usually it is not possible to determine the dependence of the roots on p explicitly. Sometimes, if the problem is not very complicated, the Routh-Hurwitz criterion may be invoked for testing stability. For completeness we summarize this celebrated result (see for instance Chen (1970)). Recall that a polynomial is Hurwitz if all its roots have zero or negative real part. It is strictly Hurwitz if all its roots have strictly negative real part. Summary 5.2.3 (Routh-Hurwitz criterion). Consider the polynomial (s) = a0 sn + a1 sn1 + + a n1 s + an (5.11)

5.2. Parametric robustness analysis

201

with real coefcients a0 , a1 , , an . From these coefcients we form the Hurwitz tableau a0 a1 b0 b1 The rst two rows of the tableau are directly taken from the even and odd coefcients of the polynomial , respectively. The third row is constructed from the two preceding rows by letting b0 b2 b4 = a2 a4 a6 a0 a3 a1 a5 a7 . (5.13) a2 a3 b2 b3 a4 a5 b4 b5 a6 a7 (5.12)

Note that the numerator of b0 is obtained as minus the determinant of the 2 2 matrix formed by the rst column of the two rows above and the column to the right of and above b0 . Similarly, the numerator of b2 is obtained as minus the determinant of the 2 2 matrix formed by the rst column of the two rows above and the column to the right of and above b2 . All further entries of the third row are formed this way. Missing entries at the end of the rows are replaced with zeros. We stop adding entries to the third row when they all become zero. The fourth row b1 , b3 , is formed from the two rows above it in the same way the third row is formed from the rst and second rows. All further rows are constructed in this manner. We stop after n + 1 rows. Given this tableau, a necessary and sufcient condition for the polynomial to be strictly Hurwitz is that all the entries a0 , a1 , b0 , b1 , of the rst column of the tableau are nonzero and have the same sign. A useful necessary condition for stability (known as Descartes rule of signs) is that all the coefcients a0 , a1 , , an of the polynomial are nonzero and all have the same sign. In principle, the Routh-Hurwitz criterion allows establishing the stability region of the parameter dependent polynomial as given by (5.10), that is, the set of all parameter values p for which the roots of all have strictly negative real part. In practice, this is often not simple. Example 5.2.4 (Stability region for third-order plant). By way of example we consider the third-order plant of Examples 5.2.1 and 5.2.2. Using the numerical values established in Example 5.2.2 we have from Example 5.2.1 that the closed-loop characteristic polynomial is given by (s) = 4 1 s + ( + )s3 + s2 + g 2s + g. 10 10 (5.14)

The parameter vector p has components g and . The Hurwitz tableau may easily be found to be given by
10

+ b0 b1 g

1 10

1 g 2 g

(5.15)

202

Chapter 5. Uncertainty Models and Robustness

where b0 = +
1 10

2 10 g

+ ( +
1 10

1 10

, +
1 2 10 )

(5.16)

b1

2 10 g) 2 ( 1 2 + 10 10 g

g.

(5.17)

Inspection of the coefcients of the closed-loop characteristic polynomial (5.14) shows that a necessary condition for closed-loop stability is that both g and be positive. This condition ensures the rst, second and fth entries of the rst column of the tableau to be positive. The third entry b0 is positive if g < g3 (), where g3 is the function 1 5 2( + 10 ) . (5.18) g3 () = The fourth entry b1 is positive if g < g4 (), with g4 the function 1 1 5( + 10 )( 2 10 ) . g4 () =

(5.19)

Figure 5.5 shows the graphs of g3 and g4 and the resulting stability region. The case = 0 needs to be considered separately. Inspection of the root locus of Fig. 5.4 (which applies if = 0) shows that for = 0 closed-loop stability is obtained for all g > 0.

40 g 20 g3 g4 0 0 1 2

Figure 5.5: Stability region Exercise 5.2.5. Stability margins The stability of the feedback system of Example 5.2.4 is quite robust with respect to variations in the parameters and g. Inspect the Nyquist plot of the nominal loop gain to determine the various stability margins of 1.4.2 and Exercise 1.4.9(b) of the closed-loop system.

5.2.3 Gridding
If the number of parameters is larger than two or three then it seldom is possible to establish the stability region analytically as in Example 5.2.4. A simple but laborious alternative method is

5.2. Parametric robustness analysis

203

known as gridding. It consists of covering the relevant part of the parameter space with a grid, and testing for stability in each grid point. The grid does not necessarily need to be rectangular. Clearly this is a task that easily can be coded for a computer. The number of grid points increases exponentially with the dimension of the parameter space, however, and the computational load for a reasonably ne grid may well be enormous. Example 5.2.6 (Gridding the stability region for the third-order plant). Figure 5.6 shows the results of gridding the parameter space in the region dened by 0.5 g 4 and 0 1. The small circles indicate points where the closed-loop system is stable. Plus signs correspond to unstable points. Each point may be obtained by applying the Routh-Hurwitz criterion or any other stability test.

4 g 2 0

+ ++ + ++ ++ +

Figure 5.6: Stability region obtained by gridding

5.2.4 Kharitonovs theorem


Gridding is straightforward but may involve a tremendous amount of repetitious computation. In 1978, Kharitonov (1978a) (see also Kharitonov (1978b)) published a result that subsequently attracted much attention in the control literature2 because it may save a great deal of work. Kharitonovs theorem deals with the stability of a system with closed-loop characteristic polynomial (s) = 0 + 1 s + + n sn . Each of the coefcients i is known to be bounded in the form i i i , (5.21) (5.20)

with i and i given numbers for i = 0, 1, , n. Kharitonovs theorem allows verifying whether all these characteristic polynomials are strictly Hurwitz by checking only four special polynomials out of the innite family. Summary 5.2.7 (Kharitonovs theorem). Each member of the innite family of polynomials (s) = 0 + 1 s + + n sn , with i i i ,
2 For

(5.22)

i = 0, 1, 2, , n,

(5.23)

a survey see Barmish and Kang (1993).

204

Chapter 5. Uncertainty Models and Robustness

is strictly Hurwitz if and only if each of the four Kharitonov polynomials k1 (s) k2 (s) k3 (s) k4 (s) = = = = 0 + 1 s + 2 s 2 + 3 s 3 + 4 s 4 + 5 s 5 + 6 s 6 + , 0 + 1 s + 2 s 2 + 3 s 3 + 4 s 4 + 5 s 5 + 6 s 6 + , 0 + 1 s + 2 s + 3 s + 4 s + 5 s + 6 s + ,
2 2 3 3 4 4 5 5 6 6

(5.24) (5.25) (5.26) (5.27)

0 + 1 s + 2 s + 3 s + 4 s + 5 s + 6 s +

is strictly Hurwitz.

Note the repeated patterns of under- and overbars. A simple proof of Kharitonovs theorem is given by Minnichelli, Anagnost, and Desoer (1989). We see what we can do with this result for our example. Example 5.2.8 (Application of Kharitonovs theorem). From Example 5.2.1 we know that the stability of the third-order uncertain feedback system is determined by the characteristic polynomial (s) = T0 s4 + ( + T0 )s3 + s2 + gTd s + gk, (5.28) where in Example 5.2.2 we took k = 1, Td = 2 and T0 = 1/10. Suppose that the variations of the plant parameters g and are known to be bounded by g g g, 0 . (5.29)

Inspection of (5.28) shows that the coefcients 0 , 1 , 2 , 3 , and 4 are correspondingly bounded by g 0 g, g 2 1 g 2, 1 2 1, T0 3 T0 + , 0 4 T0 . We assume that 0 0.2, so that = 0.2, but consider two cases for the gain g: 1. 0.5 g 5, so that g = 0.5 and g = 5. This corresponds to a variation in the gain by a factor of ten. Inspection of Fig. 5.5 shows that this region of variation is well within the stability region. It is easily found that the four Kharitonov polynomials are k1 (s) k2 (s) k3 (s) k4 (s) = 0.5 + 0.5 2s + s2 + 0.3s3 , = 5 + 5 2s + s2 + 0.1s3 + 0.02s4 , = 5 + 0.5 2s + s2 + 0.3s3 + 0.02s4 , = 0.5 + 5 2 + s2 + 0.1s3 . (5.32) (5.33) (5.34) (5.35) (5.31) (5.30)

5.2. Parametric robustness analysis

205

By the Routh-Hurwitz test or by numerical computation of the roots using M ATLAB or another computer tool it may be found that only k1 is strictly Hurwitz, while k2 , k3 , and k4 are not Hurwitz. This shows that the polynomial 0 + 1 s + 2 s2 + 3 s3 + 4 s4 is not strictly Hurwitz for all variations of the coefcients within the bounds (5.30). This does not prove that the closed-loop system is not stable for all variations of g and within the bounds (5.29), however, because the coefcients i of the polynomial do not vary independently within the bounds (5.30). 2. 0.5 g 2, so that g = 0.5 and g = 2. The gain now only varies by a factor of four. Repeating the calculations we nd that each of the four Kharitonov polynomials is strictly Hurwitz, so that the closed-loop system is stable for all parameter variations.

5.2.5 The edge theorem


Example 5.2.8 shows that Kharitonovs theorem usually only yields sufcient but not necessary conditions for robust stability in problems where the coefcients of the characteristic polynomial do not vary independently. We therefore consider characteristic polynomials of the form (s) = 0 ( p)sn + 1 ( p)sn1 + + n ( p), (5.36)

where the parameter vector p of uncertain coefcients p i , i = 1, 2, , N, enters linearly into the coefcients i ( p), i = 0, 1, , n. Such problems are quite common; indeed, the example we are pursuing is of this type. We may rearrange the characteristic polynomial in the form
N

(s) = 0 (s) +
i=1

pi i (s),

(5.37)

where the polynomials i , i = 0, 1, , N are xed and given. Assuming that each of the parameters pi lies in a bounded interval of the form pi p i pi , i = 1, 2, , N, the family of polynomials (5.37) forms a polytope of polynomials. Then the edge theorem (Bartlett, Hollot, and Lin 1988) states that to check whether each polynomial in the polytope is Hurwitz it is sufcient to verify whether the polynomials on each of the exposed edges of the polytope are Hurwitz. The exposed edges are obtained by xing N 1 of the parameters pi at their minimal or maximal value, and varying the remaining parameter over its interval. The edge theorem actually is stronger than this. Summary 5.2.9 (Edge theorem). Let D be a simply connected domain in the complex plane. Then all the roots of each polynomial (5.37) are contained in D if and only if the roots of each polynomial on the exposed edges of the polytope are contained in D . A simply connected domain is a domain such that every simple closed contour (i.e., a contour that does not intersect itself) inside the domain encloses only points of the domain. Although obviously application of the edge theorem involves much more work than that of Kharitonovs theorem it produces more results. Example 5.2.10 (Application of the edge theorem). We apply the edge theorem to the thirdorder uncertain plant. From Example 5.2.1 we know that the closed-loop characteristic polynomial is given by (s) = = T0 s4 + ( + T0 )s3 + s2 + gTd s + gk (T0 s + s ) + (T0 s + s ) + g(Td s + k).
3 2 4 3

(5.38) (5.39)

206

Chapter 5. Uncertainty Models and Robustness

Assuming that 0 and g g g this forms a polytope. By the edge theorem, we need to check the locations of the roots of the four exposed edges =0: =: g=g: g=g: T0 s3 + s2 + gTd s + gk, T0 s + ( + T0 )s + s + gTd s + gk,
4 3 2

g g g, g g g, 0 , 0 . (5.40)

T0 s4 + ( + T0 )s3 + s2 + gTd s + gk, T0 s4 + ( + T0 )s3 + s2 + gTd s + gk,

Figure 5.7 shows the patterns traced by the various root loci so dened with T0 , Td and k as determined in Example 5.2.2, and = 0.2, g = 0.5, g = 5. This is the case where in Example 5.2.10 application of Kharitonovs theorem was not successful in demonstrating robust stability. Figure 5.7 shows that the four root loci are all contained within the left-half complex plane. By the edge theorem, the closed-loop system is stable for all parameter perturbations that are considered.
First edge
10 5 0 -5 -10 -10 1 0.5 0 -0.5 -1 -150 -5 -15 10 5 0 -5 -10 -400 0 5

Second edge

-5

-10

-5

Third edge

Fourth edge

-100

-50

-300

-200

-100

Figure 5.7: Loci of the roots of the four exposed edges

5.2.6 Testing the edges


Actually, to apply the edge theorem it is not necessary to grid the edges, as we did in Example 5.2.10. By a result of Biaas (1985) the stability of convex combinations of stable polynomials3 p1 and p2 of the form p = p1 + (1 ) p2 , [0, 1], (5.41)

may be established by a single test. Given a polynomial q(s) = q0 sn + q1 sn1 + q2 sn2 + + q n ,


3 Stable

(5.42)

polynomial is the same as strictly Hurwitz polynomial.

5.2. Parametric robustness analysis

207

dene its Hurwitz matrix Q (Gantmacher 1964) as the n n matrix q1 q3 q5 q0 q2 q4 0 q1 q3 q5 . Q= 0 q0 q2 q4 0 0 q1 q3 q5 Biaas result may be rendered as follows.

(5.43)

Summary 5.2.11 (Biaas test). Suppose that the polynomials p1 and p2 are strictly Hurwitz with their leading coefcient nonnegative and the remaining coefcients positive. Let P1 and P2 be their Hurwitz matrices, and dene the matrix
1 W = P1 P2 .

(5.44)

Then each of the polynomials p = p1 + (1 ) p2 , [0, 1], (5.45)

is strictly Hurwitz iff the real eigenvalues of W all are strictly negative. Note that no restrictions are imposed on the non-real eigenvalues of W.

Example 5.2.12 (Application of Biaas test). We apply Biaas test to the third-order plant. In Example 5.2.10 we found that by the edge theorem we need to check the locations of the roots of the four exposed edges T0 s3 + s2 + gTd s + gk, T0 s4 + ( + T0 )s3 + s2 + gTd s + gk, T0 s + ( + T0 )s + s + gTd s + gk,
4 3 2

g g g, g g g, 0 , 0 . (5.46)

T0 s4 + ( + T0 )s3 + s2 + gTd s + gk,

The rst of these families of polynomials is the convex combination of the two polynomials p1 (s) p2 (s) = = T0 s3 + s2 + gTd s + gk, T0 s + s + gTd s + gk,
3 2

(5.47) (5.48)

which both are strictly Hurwitz for the given numerical values. The Hurwitz matrices of these two polynomials are 0 1 gk 1 gk 0 P1 = T0 gTd 0 , (5.49) P2 = T0 gTd 0 , 0 1 gk gk 0 1 respectively. Numerical evaluation, with T0 = 1/10, Td = 2, k = 1, g = 0.5, and g = 5, yields 1.068 0.6848 0 1 0 . (5.50) W = P1 P2 = 0.09685 0.03152 0.01370 0.1370 0.1

208

Chapter 5. Uncertainty Models and Robustness

The eigenvalues of W are 1, 0.1, and 0.1. They are all real and negative. Hence, by Biaas test, the system is stable on the edge under investigation. Similarly, it may be checked that the system is stable on the other edges. By the edge theorem, the system is stable for the parameter perturbations studied. Exercise 5.2.13 (Application of Biaas test). Test the stability of the system on the remaining three edges as given by (5.46).

5.3 The basic perturbation model


5.3.1 Introduction
This section presents a quite general uncertainty model. It deals with both parametric or structured and unstructured perturbations, although it is more suited to the latter. Unstructured perturbations are perturbations that may involve essential changes in the dynamics of the system, characterized by norm bounds. We rst consider the fundamental perturbation paradigm, and then apply it to a number of special cases.

5.3.2 The basic perturbation model


Figure 5.8 shows the block diagram of the basic perturbation model. The block marked H is the system whose stability robustness under investigation. Its transfer matrix H is sometimes called the interconnection matrix. The block H represents a perturbation of the dynamics of the system. This perturbation model is simple but the same time powerful4. We illustrate it by some applications.

H
Figure 5.8: Basic perturbation model

5.3.3 Application: Perturbations of feedback systems


The stability robustness of a unit feedback loop is of course of fundamental interest. Example 5.3.1 (The additive perturbation model). Figure 5.9(a) shows a feedback loop with loop gain L. After perturbing the loop gain from L to L + L , the feedback system may be represented as in Fig. 5.9(b). The block within the dashed lines is the unperturbed system, denoted H, and L is the perturbation H in the basic model. To compute H, denote the input to the perturbation block L as q and its output as p, as in Fig. 5.8. Inspection of the diagram of Fig. 5.9(b) shows that with the perturbation block L taken away, the system satises the signal balance equation q = p Lq, so that q = (I + L)1 p.
4 It

(5.51)

may be attributed to Doyle (1984).

5.3. The basic perturbation model

209

p + H L

(a)

(b)

(c)

Figure 5.9: Perturbation models of a unit feedback loop. (a) Nominal system. (b) Additive perturbation. (c) Multiplicative perturbation It follows that the interconnection matrix H is H = (I + L)1 = S, (5.52)

with S the sensitivity matrix of the feedback system. The model of Fig. 5.9(b) is known as the additive perturbation model. Example 5.3.2 (The multiplicative perturbation model). An alternative perturbation model, called multiplicative or proportional perturbation model, is shown in Fig. 5.9(c). The transfer matrix of the perturbed loop gain is (1 +
L )L,

(5.53)

which is equivalent to an additive perturbation L L. The quantity L may be viewed as the relative size of the perturbation. From the signal balance q = L( p q) we obtain q = (I + L)1 Lp, so that the interconnection matrix is H = (I + L)1 L = T. T is the complementary sensitivity matrix of the feedback system. (5.54)

5.3.4 Application: Parameter perturbations


The basic model may also be used for perturbations that are much more structured than those in the previous application. Example 5.3.3 (Parameter perturbation). Consider the third-order plant of Example 5.2.1 with transfer function g . (5.55) P(s) = 2 s (1 + s) The parameters g and are uncertain. It is not difcult to represent this transfer function by a block diagram where the parameters g and are found in distinct blocks. Figure 5.10 shows one way of doing this. It does not matter that one of the blocks is a pure differentiator. Note the way the factor 1 + s in the denominator is handled by incorporating a feedback loop. After perturbing g to g + g and to + and including a feedback compensator with transfer function C the block diagram may be arranged as in Fig. 5.11. The large block inside the dashed lines is the interconnection matrix H of the basic perturbation model.

210

Chapter 5. Uncertainty Models and Robustness

u g 1/ s
2

Figure 5.10: Block diagram for third-order plant

q1

p1

p2 +

q2

C(s)

+ 1/s
2

H
Figure 5.11: Application of the the basic perturbation model to the thirdorder plant

To compute the interconnection matrix H we consider the block diagram of Fig. 5.11. Inspection shows that with the perturbation blocks g and omitted, the system satises the signal balance equation q2 = s ( p2 + q2 ) + Solution for q2 yields q2 = 1/s2 s p1 p2 . 1 + s + C(s)g/s2 1 + s + C(s)g/s2 (5.57) 1 ( p1 gC(s)q2 ) . s2 (5.56)

Further inspection reveals that q1 = C(s)q2, so that q1 = C(s)/s2 sC(s) p + p2 . 2 1 1 + s + C(s)g/s 1 + s + C(s)g/s2 (5.58)

Consequently, the interconnection matrix is H(s) = = 1 1 + s +


1 1+s C(s)g s2

C(s) s2
1 s2

sC(s) s s ,

(5.59) (5.60)

1 + L(s)

C(s) 1

s12

where L = PC is the (unperturbed) loop gain.

5.4. The small gain theorem

211

5.4 The small gain theorem


5.4.1 Introduction
In 5.5 we analyze the stability of the fundamental perturbation model using what is known as the small gain theorem. The small gain theorem is an application of the xed point theorem, one of the most important results of functional analysis5 . Summary 5.4.1 (Fixed point theorem). Suppose that U is a complete normed space6 . Let L : U U be a map such that L(u) L(v) k u v for all u, v U , (5.61)

for some real constant 0 k < 1. Such a map L is called a contraction. Then: 1. There exists a unique solution u U of the equation u = L(u). The point u is called a xed point of the equation. 2. The sequence un , n = 0, 1, 2, , dened by un+1 = L(u n ), n = 0, 1, 2, , (5.63) (5.62)

converges to u for every initial point u0 U . This method of approaching the xed point is called successive approximation.

p + + q

(a)

(b)

Figure 5.12: Left: feedback loop. Right: feedback loop with internal input and output.

5.4.2 The small gain theorem


With the xed point theorem at hand, we are now ready to establish the small gain theorem (see for instance Desoer and Vidyasagar (1975)).
an easy introduction to this material see for instance the well-known though no longer new book by Luenberger (1969). 6 The normed space U is complete if every sequence u U , i = 1, 2, such that lim i n,m un um = 0 has a limit in U .
5 For

212

Chapter 5. Uncertainty Models and Robustness

Summary 5.4.2 (Small gain theorem). In the feedback loop of Fig. 5.12(a), suppose that L :

U U is a BIBO stable system for some complete normed space U . Then a sufcient condition
for the feedback loop of Fig. 5.12(a) to be internally stable is that the input-output map L is a contraction. If L is linear then it is a contraction if and only if L < 1, where is the norm induced by the signal norm, L := sup
uU

(5.64)

Lu U . u U

The proof may be found in 5.9. The small gain theorem gives a sufcient condition for internal stability that is very simple but often also very conservative. The power of the small gain theorem is that it does not only apply to linear time-invariant systems but also to nonlinear time-varying systems. The signal spaces of signals of nite energy u L2 and the signals of nite amplitude u L are two examples of complete normed spaces, hence, for these norms the small gain theorem applies. Example 5.4.3 (Small gain theorem). Consider a feedback system as in Fig. 5.12(a), where L is a linear time-invariant system with transfer function L(s) = k . 1 + s (5.65)

is a positive time constant and k a positive or negative gain. We investigate the stability of the feedback system with the help of the small gain theorem. 1. Norm induced by the L -norm. First consider BIBO stability in the sense of the L norm on the input and output signals. By inverse Laplace transformation it follows that the impulse response corresponding to L is given by l(t) =
k et/

for t 0, otherwise.

(5.66)

The norm L of the system induced by the L -norm is (see Summary 4.3.13) L = l
L1

|l(t)| dt = |k|.

(5.67)

Hence, by the small gain theorem a sufcient condition for internal stability of the feedback system is that 1 < k < 1. (5.68)

2. Norm induced by the L2 -norm. For positive the system L is stable so according to Summary 4.3.13 the L2 -induced norm exists and equals L
H

= sup |L( j)| = sup |


k | = |k|. 1 + j

(5.69)

Again we nd from the small gain theorem that (5.68) is a sufcient condition for closedloop stability.

5.5. Stability robustness of the basic perturbation model

213

In conclusion, we determine the exact stability region. The feedback equation of the system of Fig. 5.12(b) is w = v + Lw. Interpreting this in terms of Laplace transforms we may solve for w to obtain w= 1 v, 1 L (5.70)

so that the closed-loop transfer function is 1 1 1 + s = . = k 1 L(s) (1 + k) + s 1 + 1+s (5.71)

Inspection shows that the closed-loop transfer function has a single pole at (1 + k)/. The corresponding system is BIBO stable if and only if 1 + k > 0. The closed-loop system is internally stable if and only if k > 1. This stability region is much larger than that indicated by (5.68). (5.72)

Remark. In the robust control literature the H -norm of a transfer matrix H is commonly denoted by H and not by H H as we have done so far. In the remainder of the lecture notes we will adopt the notation predominantly used and hence write H to mean the H norm, and we refer to this norm simply as the -norm. This will not lead to confusion as the p-norm for p = dened for vectors in Chapter 4 will not return in the remainder of the lecture notes.

5.5 Stability robustness of the basic perturbation model


5.5.1 Introduction
We study the internal stability of the basic perturbation model of Fig. 5.8, which is repeated in Fig. 5.13(a).

w1 + + + H H w2 + v2

v1

(a)

(b)

Figure 5.13: (a) Basic perturbation model. (b) Arrangement for internal stability

214

Chapter 5. Uncertainty Models and Robustness

5.5.2 Sufcient conditions


We assume that the model is nominally stable, that is, internally stable if H = 0. This is equivalent to the assumption that the interconnection system H by itself is BIBO stable. Similarly, we assume that also H by itself is BIBO stable. Introducing internal inputs and outputs as in Fig. 5.1(b), we obtain the signal balance equations w 1 = v1 +
H (v2

+ Hw1 ),

w2 = v2 + H(v1 +

H w1 ).

(5.73)

Rearrangement yields w1 w2 = = H
H Hw1 H w2

+ v1 +

H v2 ,

(5.74) (5.75)

+ Hv1 + v2 .

Inspection shows that by the xed point theorem the equations have bounded solutions w1 and w2 for any bounded v1 and v2 if the inequalities
HH

< 1 and

<1

(5.76)

both hold. Following the work of Doyle we use the system norm induced by the L2 -norm7. This is the -norm of the systems transfer matrix as introduced in Summary 4.3.12. We reformulate the conditions H H < 1 and H H < 1 in two ways. First, since H = sup (H( j) H ( j)) and by property 7 of 4.3.8 (H( j) H ( j)) (H( j)) H ( H ( j)) it follows that H < 1 is implied by H (H( j)) (
H ( j))

<1

for all .

(5.77)

The same condition implies H H < 1. Hence, (5.77) is a sufcient condition for stability of the perturbed system. It provides a frequency-dependent bound of the perturbations that are guaranteed not to destabilize the system. Another way of reformulating the conditions H H < 1 and H H < 1 is to replace them with the sufcient condition
H

< 1.

(5.78)

This follows from the submultiplicative property of the induced norm of Exercise 4.3.103.

5.5.3 Necessity
The inequality (5.77) is a sufcient condition for internal stability. It is easy to nd examples that admit perturbations which violate (5.77) but at the same time do not destabilize the system. It may be proved, though8, that if robust stability is desired for all perturbations satisfying (5.77) then the condition is also necessary. This means that it is always possible to nd a perturbation that violates (5.77) within an arbitrarily small margin but destabilizes the system. The condition (5.78) is equivalent to
H

<

1 H

(5.79)

7 Use of the norm induced by the L -norm has also been considered in the literature (see e.g. Dahleh and Ohda (1988)). 8 See Vidysagar (1985) for a proof for the rational case.

5.5. Stability robustness of the basic perturbation model

215

Also this condition is necessary for robust stability if stability is required for all perturbations satisfying the bound. The latter condition is particularly elegant if the perturbations are scaled such that they satisfy H 1. Then a necessary and sufcient condition for robust stability is that H < 1. Summary 5.5.1 (Stability of the basic perturbation model). Suppose that in the basic perturbation model of Fig. 5.13(a) both H and H are BIBO stable. 1. Sufcient for internal stability is that (
H ( j))

<

1 (H( j))

for all ,

(5.80)

with denoting the largest singular value. If stability is required for all perturbations satisfying this bound then the condition is also necessary. 2. Another sufcient condition for internal stability is that
H

<

1 H

(5.81)

If stability is required for all perturbations satisfying this bound then the condition is also necessary. 3. In particular, a sufcient and necessary condition for robust stability under all perturbations satisfying H 1 is that H < 1.

5.5.4 Examples
We consider two applications of the basic stability robustness result. Example 5.5.2 (Stability under perturbation). By way of example, we study which perturbations of the real parameter preserve the stability of the system with transfer function P(s) = 1 . 1 + s (5.82)

Arranging the system as in Fig. 5.14(a) we obtain the perturbation model of Fig. 5.14(b), with 0

p + y (a) u + +

u +

s y

(b)

Figure 5.14: (a) Block diagram. (b) Perturbation model the nominal value of and we need 0 to be positive.

its perturbation. Because the system should be nominally stable

216

Chapter 5. Uncertainty Models and Robustness

The interconnection matrix H of the standard perturbation model is obtained from the signal balance equation q = s( p + 0 q) (omitting u and y). Solution for q yields q= s p. 1 + s0 s0 1 s = , 1 + s0 1 + s0 0 (5.83)

It follows that H(s) = so that (H( j)) = |H( j)| = and H

(5.84)

1 0

2 2 0 2 2 1+ 0

(5.85)

= sup (H( j)) =

1 . 0

(5.86)

Since ( H ( j)) = | | = H , both (1) and (2) of Summary 5.5.1 imply that internal stability is guaranteed for all perturbations such that | | < 0 , or 0 <

< 0 .

(5.87) such that (5.88)

Obviously, though, the system is stable for all > 0, that is, for all perturbations 0 <

< .

Hence, the estimate for the stability region is quite conservative. On the other hand, it is easy to nd a perturbation that violates (5.87) and destabilizes the system. The perturbation = 0 (1 + ), for instance, with positive but arbitrarily small, violates (5.87) with an arbitrarily small margin, and destabilizes the system (because it makes negative). Note that the class of perturbations such that 0 is much larger than just real perturbations satisfying (5.87). For instance, could be a transfer function, such as
(s)

= 0

, 1 + s

(5.89)

with any positive time constant, and a real number with magnitude less than 1. This parasitic perturbation leaves the system stable. Example 5.5.3 (Two-parameter perturbation). A more complicated example is the feedback system discussed in Examples 5.2.1, 5.2.2 and 5.3.3. It consists of a plant and a compensator with transfer functions P(s) = g , s2 (1 + s) C(s) = k + Td s , 1 + T0 s (5.90)

respectively. In Example 5.3.3 we found that the interconnection matrix H with respect to perturbations in the parameters g and is H(s) =
1 1+s0

1 + L0 (s)

C(s) 1

s12

s ,

(5.91)

5.5. Stability robustness of the basic perturbation model

217

q1

p1

p2

q2

1 + 1/s2 + s

+ +

C(s)

H
Figure 5.15: Scaled perturbation model for the third-order plant with g0 and 0 denoting the nominal values of the two uncertain parameters g and . Before continuing the analysis we modify the perturbation model of Fig. 5.11 to that of Fig. 5.15. This model includes scaling factors 1 and 1 for the perturbation g and scaling factors 2 and 2 for the perturbation . The scaled perturbations are denoted g and , respectively. The product 1 1 = 1 is the largest possible perturbation in g, while 2 2 = 2 is the largest possible perturbation in . It is easily veried that the interconnection matrix corresponding to Fig. 5.15 is H(s) = 1 C(s) 1 + L0 (s) 2
1 1+s0

21 s

2 s .

(5.92)

It may also easily be worked out that the largest eigenvalue of H T (j)H( j) is () =
2 1 1+2 2 0

|1 + L0 ( j)|2

2 |C( j)|2 + 2 1 2

2 1 + 2 2 , 2 4

(5.93)

The other eigenvalue is 0. The nonnegative quantity () is the largest singular value of H( j). By substitution of L0 and C into (5.93) it follows that 2 () =
2 2 (k2 + 2 Td ) + 2 (1 + 2 T02 ) (2 + 2 6 ) 1 2 1 2 , |( j)|2

(5.94)

with the closed-loop characteristic polynomial (s) = 0 T0 s4 + (0 + T0 )s3 + s2 + g0 Td s + g0 k. (5.95)

First we note that if we choose the nominal value 0 of the parasitic time constant equal to zero which is a natural choice and 2 = 2 2 = 0 then () = , so that stability is not guaranteed. Indeed, this situation admits negative values of , which destabilize the closed-loop system. Hence, we need to choose 0 positive but such that the nominal closed-loop system is stable, of course. Second, inspection of (5.94) shows that for xed uncertainties 1 = 1 1 and 2 = 2 2 the function depends on the way the individual values of the scaling constants 1 , 1 , 2 , and

218

Chapter 5. Uncertainty Models and Robustness

2 are chosen. On rst sight this seems surprising. Reection reveals that this phenomenon is caused by the fact that the stability robustness test is based on full perturbations H . Full perturbations are perturbations such that all entries of the perturbation transfer matrix H are lled with dynamical perturbations. We choose the numerical values k = 1, Td = 2, and T0 = 1/10 as in Example 5.2.2. In Example 5.2.8 we consider variations in the time constant between 0 and 0.2. Correspondingly, we let 0 = 0.1 and 2 = 0.1. For the variations in the gain g we study two cases as in Example 5.2.8: 1. 0.5 g 5. Correspondingly, we let g0 = 2.75 and 1 = 2.25. 2. 0.5 g 2. Correspondingly, we take g0 = 1.25 and 1 = 0.75. For lack of a better choice we select for the time being 1 = 1 = 1 , 2 = 2 = 2 . (5.96)

Figure 5.16 shows the resulting plots of () for the cases (1) and (2). In case (1) the peak value is about 67.1, while in case (2) it is approximately 38.7. Both cases fail the stability robustness test by a very wide margin, although from Example 5.2.4 we know that in both situations stability is ensured. The reasons that the test fails are that (i) the test is based on full perturbations rather than the structured perturbations implied by the variations in the two parameters, and (ii) the test also allows dynamical perturbations rather than the real perturbations implied by the parameter variations.

80 (H( j)) 40 (a) (b)

0 10

10

10

10

10

10

Figure 5.16: Plots of (H( j)) for the two-parameter plant Less conservative results may be obtained by allowing the scaling factors 1 , 1 , 2 , and 2 to be frequency-dependent, and to choose them so as to minimize () for each . Substituting 1 = 1 /1 and 2 = 2 /2 we obtain () =
2 2 2 (k2 + 2 Td ) + 1
2 2 (k 2 +2 Td )6 1

+ 2 (1 + 2 T02 ) + 2 (1 + 2 T02 )6 2 2

|( j)|2

(5.97)

where = 2 /2 . It is easy to establish that for xed the quantity 2 () is minimized for 1 2 = 3 1 2
2 k2 + 2 Td , 1 + 2 T02

(5.98)

5.5. Stability robustness of the basic perturbation model

219

and that for this value of () =


2 1 k2 + 2 Td + 2 3 1 + 2 T02

|( j)|

0.

(5.99)

Figure 5.17 shows plots of for the same two cases (a) and (b) as before. In case (a) the peak

2 (H( j)) 1 (a) (b) 0 10


2

10

10

10

10

10

Figure 5.17: Plots of (H( j)) with optimal scaling for the two-parameter plant value of is about 1.83, while in case (b) it is 1. Hence, only in case (b) robust stability is established. The reason that in case (a) robust stability is not proved is that the perturbation model allows dynamic perturbations. Only if the size of the perturbations is downsized by a factor of almost 2 robust stability is guaranteed. In Example 5.7.11 in the section on the structured singular value this example is further discussed.

+ + +

1 sI

Figure 5.18: Perturbation of the state space system Exercise 5.5.4 (Stability radius). Hinrichsen and Pritchard (1986) investigate what perturbations of the matrix A make the system described by the state differential equation x(t) = Ax(t), t , (5.100)

unstable. Assume that the nominal system x(t) = Ax(t) is stable. The number r( A) = inf { A + has at least one eigenvalue in the closed right-half plane (5.101)

220

Chapter 5. Uncertainty Models and Robustness

is called the stability radius of the matrix A. The matrix norm used is the spectral norm. If only real-valued perturbations are considered then r( A) is the real stability radius, denoted r( A). If the elements of may also assume complex values then r( A) is the complex stability radius, denoted r ( A). 1. Prove that r( A) r ( A) 1 , F (5.102)

with F the rational matrix given by F(s) = (sI A)1 . Hint: Represent the system by the block diagram of Fig. 5.18. 2. A more structured perturbation model results by assuming that A is perturbed as A A + B C, (5.103)

with B and C matrices of suitable dimensions, and the perturbation. Adapt the block diagram of Fig. 5.18 and prove that the associated complex stability radius r ( A, B, C) satises r ( A, B, C) 1/ H , with H(s) = C(sI A)1 B. Hinrichsen and Pritchard (1986) prove that actually r ( A) = 1/ F and r ( A, B, C) = 1/ H . Thus, explicit formulas are available for the complex stability radius. The real stability radius is more difcult to determine. Only very recently results have become available (Qiu, Bernhardsson, Rantzer, Davison, Young, and Doyle 1995).

5.5.5 Nonlinear perturbations


The basic stability robustness result also applies for nonlinear perturbations. Suppose that the perturbation H in the block diagram of Fig. 5.13(a) is a nonlinear operator that maps the signal e(t), t , into the signal ( H e)(t), t . Assume that there exists a positive constant k such that
H e1

H e2

k e1 e2

(5.104)

for every two input e1 , e2 to the nonlinearity. We use the L2 -norm for signals. By application of the xed point theorem it follows that the perturbed system of Fig. 5.13(a) is stable if k H

< 1.
H

(5.105) is a static nonlinearity described by (5.106)

Suppose for instance that the signals are scalar, and that (
H e)(t)

= f (e(t)),

with f : . Let f satisfy the inequality f (e) c, e e = 0, e , (5.107)

with c a nonnegative constant. Figure 5.19 illustrates the way the function f is bounded. We call f a sector bounded function. It is not difcult to see that (5.104) holds with k = c. It follows that the perturbed system is stable for any static nonlinear perturbation satisfying (5.107) as long as c 1/ H . Similarly, the basic stability robustness result holds for time-varying perturbations and perturbations that are both nonlinear and time-varying, provided that these perturbations are suitably bounded.

5.6. Stability robustness of feedback systems

221

ce f (e) e ce
Figure 5.19: Sector bounded function

5.6 Stability robustness of feedback systems


5.6.1 Introduction
In this section we apply the results of 5.5 to various perturbation models for single-degreeof-freedom feedback systems. We successively discuss proportional, proportional inverse and fractional perturbations for MIMO systems. The results of 1.4 emerge as special cases.

5.6.2 Proportional perturbations


In 5.3 we consider additive and multiplicative (or proportional) perturbation models for singledegree-of-freedom feedback systems. The proportional model is preferred because it represents the relative size of the perturbation. The corresponding block diagram is repeated in Fig. 5.20(a). It represents a perturbation

L +

p + L

W +

V +

(a)

(b)

Figure 5.20: Proportional and scaled perturbations of a unit feedback loop L (I +


L )L

(5.108)

Since q = L( p + q), so that q = (I + L)1 Lp, the interconnection matrix is H = (I + L)1 L = T, (5.109)

222

Chapter 5. Uncertainty Models and Robustness

with T = (I + L)1 L = L(I + L)1 the complementary sensitivity matrix of the closed-loop system. To scale the proportional perturbation we modify the block diagram of Fig. 5.20(a) to that of Fig. 5.20(b). This diagram represents a perturbation L (I + V L W )L, (5.110)

where V and W are available to scale such that L 1. For this conguration the interconnection matrix is H = W T V. Application of the results of Summary 5.5.1 leads to the following conclusions. Summary 5.6.1 (Robust stability of feedback systems for proportional perturbations). Assume that the feedback system of Fig. 5.20(a) is nominally stable. 1. A sufcient condition for the closed-loop system to be stable under proportional perturbations of the form L (1 + is that (
L ( j)) L )L

(5.111)

<

1 (T ( j))

for all .

(5.112)

with T = (I + L)1 L = L(I + L)1 the complementary sensitivity matrix of the feedback loop. If stability is required for all perturbations satisfying the bound then the condition is also necessary. 2. Under scaled perturbations L (1 + V L W )L the system is stable for all L WTV

(5.113) 1 if and only if (5.114)

< 1.

For SISO systems part (1) of this result reduces to Doyles robustness criterion of 1.4. In fact, Summary 5.6.1 is the original MIMO version of Doyles robustness criterion (Doyle 1979).

5.6.3 Proportional inverse perturbations


The result of the preceding subsection conrms the importance of the complementary sensitivity matrix (or function) T for robustness. We complement it with a dual result involving the sensitivity function S. To this end, consider the perturbation model of Fig. 5.21(a), where the perturbation L1 is included in a feedback loop. The model represents a perturbation L (I +
L1 ) 1

L.

(5.115)

Assuming that L has an inverse, this may be rewritten as the perturbation L1 L1 (I +


L1 ).

(5.116)

5.6. Stability robustness of feedback systems

223

L1

L1

q V + W

(a)

(b)

Figure 5.21: Proportional inverse perturbation of unit feedback loop Hence, the perturbation represents a proportional perturbation of the inverse loop gain. The interconnection matrix H follows from the signal balance q = p Lq so that q = (I + L)1 p. Hence, the interconnection matrix is H = (I + L)1 = S. (5.117) S = (I + L)1 is the sensitivity matrix of the feedback loop. We allow for scaling by modifying the model to that of Fig. 5.21(b). This model represents perturbations of the form L1 L1 (I + V L1 W ), (5.118)

where V and W provide freedom to scale such that L1 1. The interconnection matrix now is H = V SW. Application of the results of Summary 5.5.1 yields the following conclusions. Summary 5.6.2 (Robust stability of feedback systems for proportional inverse perturbations). Assume that the feedback system of Fig. 5.21(a) and (b) is nominally stable. 1. Under proportional inverse perturbations of the form L1 L1 (I +
L1 )

(5.119)

a sufcient condition for stability is that (


L1 ( j))

<

1 (S( j))

for all ,

(5.120)

with S = (I + L)1 the sensitivity matrix of the feedback loop. If stability is required for all perturbations satisfying the bound then the condition is also necessary. 2. Under scaled inverse perturbations of the form L1 L1 (I + V L1 W ) the closed-loop system is stable for all L1 W SV

(5.121) 1 if and only if (5.122)

< 1.

224

Chapter 5. Uncertainty Models and Robustness

5.6.4 Example
We illustrate the results of Summaries 5.6.1 and 5.6.2 by application to an example. Example 5.6.3 (Robustness of a SISO closed-loop system). In Example 5.2.1 we considered a SISO single-degree-of-freedom feedback system with plant and compensator transfer functions P(s) = g , s2 (1 + s) C(s) = k + Td s , 1 + T0 s (5.123)

respectively. Nominally the gain g equals g0 = 1, while the parasitic time constant is nominally 0. In Example 5.2.2 we chose the compensator parameters as k = 1, Td = 2 and T0 = 1/10. We use the results of Summaries 5.6.1 and 5.6.2 to study what perturbations of the parameters g and leave the closed-loop system stable. 1. Loop gain perturbation model. Starting with the expression L(s) = P(s)C(s) = g s2 (1 + s) k + Td s 1 + T0 s (5.124)

it is easy to nd that the proportional loop gain perturbations are


L (s) =

L(s) L0 (s) = L0 (s)

gg0 g0

1 + s

(5.125)

Figures 5.22(a) and (b) show the magnitude plot of the inverse 1/ T0 of the nominal complementary sensitivity function. Inspection shows that 1/ T0 assumes relatively small values in the low-frequency region. This is where the proportional perturbations of the loop gain need to be the smallest. For = 0 the proportional perturbation (5.125) of the loop gain reduces to
L (s)

g g0 , g0

(5.126)

and, hence, is constant. The minimal value of the function 1/|T0 | is about 0.75. Therefore, for = 0 the robustness criterion (a) of Summary 5.6.1 allows relative perturbations of g up to 0.75, so that 0.25 < g < 1.75. For g = g0 , on the other hand, the proportional perturbation (5.125) of the loop gain reduces to
L (s)

s . 1 + s

(5.127)

Figure 5.22(a) shows magnitude plots of this perturbation for several values of . The robustness criterion (a) of Summary 5.6.1 permits values of up to about 1.15. For this value of the gain g can be neither increased nor decreased without violating the criterion. For smaller values of the parasitic time constant a wider range of gains is permitted. The plots of of | L | of Fig. 5.22(b) demonstrate that for = 0.2 the gain g may vary between about 0.255 and 1.745. The stability bounds on g and are conservative. The reason is of course that the perturbation model allows a much wider class of perturbations than just those caused by changes in the parameters g and .

5.6. Stability robustness of feedback systems

225

(a) 1000 1/|To| g = 1.745 1 = 0.2 = 0.6 = 1.15 1 100 1 g = 1.2 g = 0.2 .001 .01 1000

(b) 1/|To|

.001 .01

100

Figure 5.22: Robust stability test for the loop gain perturbation model: (a) 1/|T0 | and the relative perturbations for g = 1. (b) 1/|T0| and the relative perturbations for = 0.2. 2. Inverse loop gain perturbation model. The proportional inverse loop gain perturbation is given by
L1 (s)

L1 (s) L1 (s) g0 g0 g 0 + s . = g g L1 (s) 0

(5.128)

We apply the results of Summary 5.6.2. Figure 5.23 gives the magnitude plot of the inverse 1/S0 of the nominal sensitivity function. Inspection shows that for the inverse loop gain perturbation model the high-frequency region is the most critical. By inspection of (5.128) we see that if = 0 then | L1 ()| = , so that robust stability is not ensured. Apparently, this model cannot handle high-frequency perturbations caused by parasitic dynamics. For = 0 the proportional inverse loop gain perturbation reduces to
L1 (s)

g0 g , g

(5.129)

and, hence, is constant. The magnitude of 1/S0 has a minimum of about 0.867, so that g for = 0 stability is ensured for | g0g | < 0.867, or 0.536 < g < 7.523. This range of variation is larger than that found by application of the proportional loop gain perturbation model. Again, the result is conservative. It is disturbing that the model does not handle parasitic perturbations.

The derivation of the unstructured robust stability tests of Summary 5.6.1 is based on the small gain theorem, and presumes the perturbations L to be BIBO stable. This is highly restrictive if the loop gain L by itself represents an unstable system, which may easily occur (in particular, when the plant by itself is unstable). It is well-known (Vidysagar 1985) that the stability assumption on the perturbation may be relaxed to the assumption that both the nominal loop gain and the perturbed loop gain have the same number of right-half plane poles. Likewise, the proofs of the inverse perturbation tests of Summary 5.6.2 require the perturbation of the inverse of the loop gain to be stable. This is highly restrictive if the loop gain by itself

226

Chapter 5. Uncertainty Models and Robustness

10

10 1

1/|So|

10

10

10

10

10

Figure 5.23: Magnitude plot of 1/S0 has right-half plane zeros. This occurs, in particular, when the plant has right-half plane zeros. The requirement, however, may be relaxed to the assumption that the nominal loop gain and the perturbed loop gain have the same number of right-half plane zeros.

5.6.5 Fractional representation


The stability robustness analysis of feedback systems based on perturbations of the loop gain or its inverse is simple, but often overly conservative. Another model that is encountered in the literature9 relies on what we term here fractional perturbations. It combines, in a way, loop gain perturbations and inverse loop gain perturbations. In this analysis, the loop gain L is represented as L = N D1 , (5.130)

where the denominator D is a square nonsingular rational or polynomial matrix, and N a rational or polynomial matrix. Any rational transfer matrix L may be represented like this in many ways. If D and N are polynomial then the representation is known as a (right)10 polynomial matrix fraction representation. If D and N are rational and proper with all their poles in the open left-half complex plane then the representation is known as a (right) rational matrix fraction representation . Example 5.6.4 (Fractional representations). For a SISO system the fractional representation is obvious. Suppose that g . (5.131) L(s) = 2 s (1 + s) Clearly we have the polynomial fractional representation L = N D1 with N(s) = g and D(s) = s2 (1 + s). The fractional representation may be made rational by letting D(s) = s2 (1 + s) , d(s) N(s) = g , d(s) (5.132)

with d any strictly Hurwitz polynomial.


9 The

idea originates from Vidyasagar (Vidyasagar, Schneider, and Francis 1982; Vidysagar 1985). It is elaborated in McFarlane and Glover (1990). 10 Because the denominator is on the right.

5.6. Stability robustness of feedback systems

227

Right fractional representation may also be obtained for MIMO systems (see for instance Vidysagar (1985)).

p1

q1

q2

N +

p2

H
Figure 5.24: Fractional perturbation model

5.6.6 Fractional perturbations


Figure 5.24 shows the fractional perturbation model. Inspection shows that the perturbation is given by L (I + or N D1 (I +
N )N D 1 N )L(I

D)

(5.133)

(I +

D)

(5.134)

Hence, the numerator and denominator are perturbed as N (I +


N )N,

D (I +

D )D.

(5.135)

Thus, D and N represent proportional perturbations of the denominator and of the numerator, respectively. By block diagram substitution it is easily seen that the conguration of Fig. 5.24 is equivalent to that of Fig. 5.25(a). The latter, in turn, may be rearranged as in Fig. 5.25(b). Here p= with
L D q1

N q2

q1 , q2

(5.136)

= [

N] .

(5.137)

To establish the interconnection matrix H of the conguration of Fig. 5.25(b) we consider the signal balance equation q1 = p Lq1. It follows that q1 = (I + L)1 p = Sp, with S = (I + L)1 the sensitivity matrix. Since q2 = Lq1 we have q2 = L(I + L)1 p = T p, (5.139) (5.138)

228

Chapter 5. Uncertainty Models and Robustness

q2

N +

p2

p1 +

q1

(a)

+ H q2

L + p

q1

(b)

+ H

Figure 5.25: Equivalent congurations with T = L(I + L)1 the complementary sensitivity matrix. Inspection of (5.1385.139) shows that the interconnection matrix is H= S . T (5.140)

Investigation of the frequency dependence of the greatest singular value of H( j) yields information about the largest possible perturbations L that leave the loop stable. It is useful to allow for scaling by modifying the conguration of Fig. 5.25(b) to that of Fig. 5.26. This modication corresponds to representing the perturbations as
D

= V D W1 ,

= V N W2 ,

(5.141)

where V , W1 , and W2 are suitably chosen (stable) rational matrices such that L

1, with L = [ D N ].

(5.142)

Accordingly, the interconnection matrix changes to H= W1 SV . W2 T V (5.143)

Application of the results of Summary 5.5.1 yields the following conclusions. Summary 5.6.5 (Numerator-denominator perturbations). In the stable feedback conguration of Fig. 5.27, suppose that the loop gain is perturbed as L = N D1 [(I + V N W2 )N][(I + V D W1 )D]1 (5.144)

5.6. Stability robustness of feedback systems

229

q2

L p

q1

W2

V +

W1

H
Figure 5.26: Perturbation model with scaling with V, W1 and W2 stable transfer matrices. Then the closed-loop system is stable for all stable perturbations P = [ D N ] such that P 1 if and only if H

< 1.

(5.145)

Here we have H= W1 SV , W2 T V (5.146)

with S = (I + L)1 the sensitivity matrix and T = L(I + L)1 the complementary sensitivity matrix.

Figure 5.27: Feedback loop The largest singular value of H( j), with H given by (5.146), equals the square root of the largest eigenvalue of H T (j)H( j) =
T V T (j)[S T (j)W1 (j)W1 ( j)S( j) T + T T (j)W2 (j)W2 ( j)T ( j)]V ( j).

(5.147)

For SISO systems this is the scalar function H T (j)H( j) = |V ( j)|2 [|S( j)|2 |W1 ( j)|2 + |T ( j)|2 |W2 ( j)|2 ]. (5.148)

5.6.7 Discussion
We consider how to arrange the fractional perturbation model. In the SISO case, without loss of generality we may take the scaling function V equal to 1. Then W1 represents the scaling factor

230

Chapter 5. Uncertainty Models and Robustness

for the denominator perturbations and W2 that for the numerator perturbations. We accordingly have H
2

= sup |S( j)|2 |W1 ( j)|2 + |T ( j)|2 |W2 ( j)|2 .

(5.149)

For well-designed control systems the sensitivity function S is small at low frequencies while the complementary sensitivity function T is small at high frequencies. Figure 5.28 illustrates this. Hence, W1 may be large at low frequencies and W2 large at high frequencies without violating

|S|

|T|

.01

.01

.001 .01 1 100

.001 .01 1 100

Figure 5.28: Sensitivity function and complementary sensitivity function for a well-designed feedback system the robustness condition H < 1. This means that at low frequencies we may allow large denominator perturbations, and at high frequencies large numerator perturbations. The most extreme point of view is to structure the perturbation model such that all low frequency perturbations are denominator perturbations, and all high frequency perturbations are numerator perturbations. Since we may trivially write L= 1
1 L

(5.150)

modeling low frequency perturbations as pure denominator perturbations implies modeling low frequency perturbations as inverse loop gain perturbations. Likewise, modeling high frequency perturbations as pure numerator perturbations implies modeling high frequency perturbations as loop gain perturbations. This amounts to taking |W1 ( j)| |W2 ( j)| = = WL1 () 0 at low frequencies, at high frequencies, (5.151)

0 at low frequencies, WL () at high frequencies,

(5.152)

with WL1 a bound on the size of the inverse loop perturbations, and WL a bound on the size of the loop perturbations. Obviously, the boundary between low and high frequencies lies in the crossover region, that is, near the frequency where the loop gain crosses over the zero dB line. In this frequency region neither S nor T is small.

5.6. Stability robustness of feedback systems

231

Another way of dealing with this perturbation model is to modify the stability robustness test to checking whether for each |
L1 ( j)|

<

1 or | |S( j)|

L ( j)|

<

1 . |T ( j)|

(5.153)

This test amounts to verifying whether either the proportional loop gain perturbation test succeeds or the proportional inverse loop gain test. Obviously, its results are less conservative than the individual tests. Example 5.6.6 (Numerator-denominator perturbations). Again consider the feedback system that we studied before, most recently in Example 5.6.3, with nominal plant P0 (s) = g0 /s2 , perturbed plant g , (5.154) P(s) = 2 s (1 + s) and compensator C(s) = (k + sTd )/(1 + sT0 ). We use the design values of Example 5.2.2. Plots of the magnitudes of the nominal sensitivity and complementary sensitivity functions S and T are given in Fig. 5.28. S is small up to a frequency of almost 1, while T is small starting from a slightly higher frequency. The crossover region is located about the frequency 1. Figure 5.29 illustrates that the perturbation from the nominal parameter values g0 = 1 and 0 = 0 to g = 5, = 0.2, for instance, fails the two separate tests, but passes the test (5.153). Hence, the perturbation leaves the closed-loop system stable.

(a) 10000
1/|S|

(b)

(c)

100

1/|T| L

1/|S| L1

1/|T| L

L1

.01

100 .01

100 .01

100

Figure 5.29: (a), (b) Separate stability tests. (c) Combined test. Feedback systems are robustly stable for perturbations in the frequency regions where either the sensitivity is small (at low frequencies) or the complementary sensitivity is small (at high frequencies). In the crossover region neither sensitivity is small. Hence, the feedback system is not robust for perturbations that strongly affect the crossover region. In the crossover region the uncertainty therefore should be limited. On the one hand this limitation restricts structured uncertainty caused by load variations and environmental changes that the system can handle. On the other hand, unstructured uncertainty deriving from neglected dynamics and parasitic effects should be kept within bounds by adequate modeling.

5.6.8 Plant perturbation models


In the previous subsections we modeled the uncertainty as an uncertainty in the loop gain L, which results in interconnection matrices in terms of this loop gain, such as S and T. It is important to

232

Chapter 5. Uncertainty Models and Robustness

realize that we have a choice in how to model the uncertainty and that we need not necessarily do that in terms of the loop gain. In particular as the uncertainty is usually present in the plant P only, and not in controller C, it makes sense to model the uncertainty as such. Table 5.1 lists several ways to pull out the uncertainty from the plant.
plant P P P P P D1 N N D1 (D + Z ( N + W1 perturbed plant P+V (I + V P(I + V (I + V P(I + V
D W1 ) 1 PW PW)P PW) 1

interconnection matrix H W K SV W T V W K S PV W SV W (I + K P)1 V


N W2 ) D Z) 1

perturbation matrix
P

PW)

PW)

(N + Z

Z D1

W1 S W2 K S K SW1

D1 Z
(I+K P)1 W2

Z )( D + W2

N D

Table 5.1: A list of perturbation models with plant P and controller K

5.7 Structured singular value robustness analysis


5.7.1 Introduction
We return to the basic perturbation model of Fig. 5.8, which we repeat in Fig. 5.30(a). As

H (a)

H (b) 1

Figure 5.30: (a) Basic perturbation model. (b) Structured perturbation model

5.7. Structured singular value robustness analysis

233

demonstrated in 5.3, the model is very exible in representing both structured perturbations (i.e., variations of well-dened parameters) and unstructured perturbations. The stability robustness theorem of Summary 5.5.1 (p. 215) guarantees robustness under all perturbations such that ( ( j)) (H( j)) < 1 for all . Perturbations ( j) whose norm ( ( j)) does not exceed the number 1/(H( j)), however, are completely unstructured. As a result, often quite conservative estimates are obtained of the stability region for structured perturbations. Several examples that we considered in the preceding sections conrm this. Doyle (1982) proposed another measure for stability robustness based on the notion of what he calls structured singular value. In this approach, the perturbation structure is detailed as in Fig. 5.30(b) (Safonov and Athans 1981). The overall perturbation has the block diagonal form 0 0 1 0 0 2 = (5.155) . 0 0 K Each diagonal block entry
i

has xed dimensions, and has one of the following forms:

i = I, with a real number. If the unit matrix I has dimension 1 this represents a real parameter variation. Otherwise, this is a repeated real scalar perturbation .

= I, with a stable11 scalar transfer matrix. This represents a scalar or repeated scalar dynamic perturbation .
i i is a (not necessarily square) stable transfer matrix. This represents a multivariable dynamic perturbation .

A wide variety of perturbations may be modeled this way.

Im

Re

Figure 5.31: Nyquist plot of det(I H ) for a destabilizing perturbation. We study which are the largest perturbations with the given structure that do not destabilize the system of Fig. 5.30(a). Suppose that a given perturbation destabilizes the system. Then by the generalized Nyquist criterion of Summary 1.3.13 the Nyquist plot of det(I H ) encircles the origin at least once, as illustrated in Fig. 5.31. Consider the Nyquist plot of det(I H ), with a real number that varies between 0 and 1. For = 1 the modied Nyquist plot coincides with that of Fig. 5.31, while for = 0 the plot reduces to the point 1. Since obviously the plot depends continuously on , there must exist a value of in the interval (0, 1] such that the Nyquist plot of det(I H ) passes through the origin. Hence, if destabilizes the perturbed
11 A

transfer function or matrix is stable if all its poles are in the open left-half plane.

234

Chapter 5. Uncertainty Models and Robustness

system, there exist (0, 1] and such that det(I H( j) ( j)) = 0. Therefore, does not destabilize the perturbed system if and only if there do not exist (0, 1] and such that det(I H( j ( j)) = 0. Fix , and let (H( j)) be the largest real number such that det(I H( j) ( j)) = 0 for all ( j) (with the prescribed structure) such that ( ( j)) < (H( j)). Then, obviously, if for a given perturbation ( j), , ( ( j)) < (H( j)) for all (5.156) then det(I H( j) ( j)) = 0 for (0, 1] and . Therefore, the Nyquist plot of det(I H ) does not encircle the origin, and, hence, does not destabilize the system. Conversely, it is always possible to nd a perturbation that violates (5.156) within an arbitrarily small margin that destabilizes the perturbed system. The number (H( j)) is known as the multivariable robustness margin of the system H at the frequency (Safonov and Athans 1981). We note that for xed (H( j)) = sup{ | ( ( j)) < det(I H( j) ( j)) = 0} = inf{( ( j)) | det(I H( j) ( j)) = 0}. (5.157)

If no structure is imposed on the perturbation then (H( j)) = 1/(H( j)). This led Doyle to terming the number (H( j)) = 1/(H( j)) the structured singular value of the complexvalued matrix H( j). Besides being determined by H( j) the structured singular value is of course also dependent on the perturbation structure. Given the perturbation structure of the perturbations of the structured perturbation model, dene by D the class of constant complex-valued matrices of the form 0 0 1 0 0 2 = (5.158) . 0 0 K The diagonal block i has the same dimensions as the corresponding block of the dynamic perturbation, and has the following form:

i = I, with a real number, if the dynamic perturbation is a scalar or repeated real scalar perturbation. i = I, with a complex number, if the dynamic perturbation is a scalar dynamic or a repeated scalar dynamic perturbation. i is a complex-valued matrix if the dynamic perturbation is a multivariable dynamic perturbation.

5.7.2 The structured singular value


We are now ready to dene the structured singular value of a complex-valued matrix M: Denition 5.7.1 (Structured singular value). Let M be an n m complex-valued matrix, and

D a set of m n structured perturbation matrices. Then the structured singular value of M given the set D is the number (M) dened by
1 = (M)
D , det( IM )=0

inf

( ).

(5.159)

If det(I M ) = 0 for all

D then 1/(M) = 0.

5.7. Structured singular value robustness analysis

235

The structured singular value (M) is the inverse of the norm of the smallest perturbation (within the given class D ) that makes I + M singular. Thus, the larger (M), the smaller the perturbation is that is needed to make I + M singular. Example 5.7.2 (Structured singular value). By way of example we consider the computation of the structured singular value of a 2 2 complex-valued matrix M= m11 m21 m12 , m22 (5.160)

under the real perturbation structure = with


1 1

0
2 2

(5.161)

and

. It is easily found that


1 |, | 2 |),

( ) = max(|

det(I + M ) = 1 + m11

+ m22

+m

2,

(5.162)

with m = det(M) = m11 m22 m12 m21 . Hence, the structured singular value of M is the inverse of (M) =
{
1 , 2 :

1+m11

inf

1 +m22

2 +m

2 =0}

max(|

1 |,

2 |).

(5.163)

Elimination of, say,


1

results in the equivalent expression


1 |,

(M) = inf max(|

1 + m11 m22 + m

1 1

).

(5.164)

Suppose that M is numerically given by M= 1 3 2 . 4 (5.165)

Then it may be found (the details are left to the reader) that 5 + 33 , (M) = 4 so that 4 5 + 33 = 5.3723. (M) = = 2 5 + 33 Exercise 5.7.3 (Structured singular value). Fill in the details of the calculation.

(5.166)

(5.167)

Exercise 5.7.4 (Alternative characterization of the structured singular value). Prove (Doyle 1982) that if all perturbations are complex then (M) =
D : ( )1

max

(M ),

(5.168)

with denoting the spectral radius that is, the magnitude of the largest eigenvalue (in magnitude). Show that if on the other hand some of the perturbations are real then (M) =
D : ( )1

max

(M ).

(5.169)

H1ere ( A) is largest of the magnitudes of the real eigenvalues of the complex matrix A. If A has no real eigenvalues then ( A) = 0.

236

Chapter 5. Uncertainty Models and Robustness

5.7.3 Structured singular value and robustness


We discuss the structured singular value in more detail in 5.7.4, but rst summarize its application to robustness analysis. Summary 5.7.5 (Structured robust stability). Given the stable unperturbed system H the perturbed system of Fig. 5.30(b) is stable for all perturbations such that ( j) D for all if ( ( j)) < 1 for all , (H( j)) (5.170)

with the structured singular value with respect to the perturbation structure D . If robust stability is required with respect to all perturbations within the perturbation class that satisfy the bound (5.170) then the condition is besides sufcient also necessary. Given a structured perturbation

such that

( j) D for every , we have (5.171)

= sup ( ( j)).

Suppose that the perturbations are scaled, so that the perturbed system is stable if H < 1, where H = sup (H( j)).

1. Then Summary 5.7.5 implies that (5.172)

With some abuse of terminology, we call H the structured singular valueof H. Even more can be said: Summary 5.7.6 (Structured robust stability for scaled perturbations). The perturbed system of Fig. 5.30(b) is stable for all stable perturbations such that ( j) D for all and 1 if and only if H < 1. Clearly, if the perturbations are scaled to a maximum norm of 1 then a necessary and sufcient condition for robust stability is that H < 1.

5.7.4 Properties of the structured singular value


Before discussing the numerical computation of the structured singular value we list some of the principal properties of the structured singular value (Doyle 1982; Packard and Doyle 1993). Summary 5.7.7 (Principal properties of the structured singular value). The structured singular value (M) of a matrix M under a structured perturbation set D has the following properties: 1. Scaling property: If all perturbations are complex, then (M) = || (M) (5.173)

for every . If none of the perturbations are real then this holds as well for every complex . 2. Upper and lower bounds: Suppose that M is square. Then (M) (M) (M), (5.174)

5.7. Structured singular value robustness analysis

237

with dened as in Exercise 5.7.4. If none of the perturbations is real then (M) (M) (M), (5.175)

with denoting the spectral radius. The upper bounds also apply if M is not square. 3. Preservation of the structured singular value under diagonal scaling: Suppose that the ith diagonal block of the perturbation structure has dimensions mi ni . Form two block diagonal matrices D and D whose ith diagonal blocks are given by D i = d i In i , D i = d i Im i , (5.176)

respectively, with d i a positive real number. Ik denotes a k k unit matrix. Then (M) = (DM D1 ). (5.177)

4. Preservation of the structured singular value under unitary transformation: Suppose that the ith diagonal block of the perturbation structure has dimensions mi ni . Form the block diagonal matrices Q and Q whose ith diagonal blocks are given by the mi mi unitary matrix Qi and the ni ni unitary matrix Qi , respectively. Then (M) = (QM) = (M Q). (5.178)

The scaling property is obvious. The upper bound in (b) follows by considering unrestricted perturbations of the form mn (that is, is a full m n complex matrix). The lower bound in (b) is obtained by considering restricted perturbations of the form = I, with a real or complex number. The properties (c) and (d) are easily checked. Exercise 5.7.8 (Proof of the principal properties of the structured singular value). Fill in the details of the proof of Summary 5.7.7.

5.7.5 A useful formula


The following formula for the structured singular value of a 2 2 dyadic matrix has useful applications. Summary 5.7.9 (Structured singular value of a dyadic matrix). The structured singular value of the 2 2 dyadic matrix M= a1 b1 a2 b1 a1 b2 a = 1 a2 b2 a2 b1 b2 , (5.179)

with a1 , a2 , b1 , and b2 complex numbers, with respect to the perturbation structure = is (M) = |a1 b1 | + |a2 b2 |. The proof is given in 5.9. (5.181)
1

0
2

(5.180)

238

Chapter 5. Uncertainty Models and Robustness

5.7.6 Numerical approximation of the structured singular value


Exact calculation of the structured singular value is often not possible and in any case computationally intensive. The numerical methods that are presently available for approximating the structured singular value are based on calculating upper and lower bounds for the structured singular value. Summary 5.7.10 (Upper and lower bounds for the structured singular value). 1. D-scaling upper bound. Let the diagonal matrices D and D be chosen as in (c) of Summary 5.7.7. Then with property (b) we have (M) = (DM D1 ) (DM D1 ). (5.182)

The rightmost side may be numerically minimized with respect to the free positive num bers d i that determine D and D. Suppose that the perturbation structure consists of m repeated scalar dynamic perturbation blocks and M full multivariable perturbation blocks. Then if 2m + M 3 the minimized upper bound actually equals the structured singular value12 2. Lower bound. Let Q be a unitary matrix as constructed in (d) of Summary 5.7.7. With property (b) we have (for complex perturbations only) (M) = (M Q) (M Q). Actually (Doyle 1982), (M) = max (M Q),
Q

(5.183)

(5.184)

with Q varying over the set of matrices as constructed in (d) of Summary 5.7.7.

Practical algorithms for computing lower and upper bounds on the structured singular value for complex perturbations have been implemented in the M ATLAB -Analysis and Synthesis Toolbox (Balas, Doyle, Glover, Packard, and Smith 1991). The closeness of the bounds is a measure of the reliability of the calculation. Real perturbations are not handled by this version of the -toolbox. Algorithms for real and mixed complex-real perturbations are the subject of current research (see for instance Young, Newlin, and Doyle (1995b) and Young, Newlin, and Doyle (1995a)). The M ATLAB Robust Control Toolbox (Chiang and Safonov 1992) provides routines for calculating upper bounds on the structured singular value for both complex and real perturbations.

5.7.7 Example
We apply the singular value method for the analysis of the stability robustness to an example. Example 5.7.11 (SISO system with two parameters). By way of example, consider the SISO feedback system that was studied in Example 5.2.1 and several other places. A plant with transfer function g (5.185) P(s) = 2 s (1 + s)
12 If there are real parametric perturbations then this result remains valid, provided that we use a generalization of D-scaling called ( D, G)-scaling which allows to exploit real perturbations.

5.7. Structured singular value robustness analysis

239

is connected in feedback with a compensator with transfer function k + sTd . (5.186) 1 + sT0 We use the design values of the parameters established in Example 5.2.2. In Example 5.3.3 it was found that the interconnection matrix for scaled perturbations in the gain g and time constant is given by C(s) = H(s) =
1 1+s0

1 C(s) 2

1 + L0 (s)

21 s

2 s ,

(5.187)

where the subscript o indicates nominal values, and L0 = P0 C is the nominal loop gain. The numbers 1 , 2 , 1 , and 2 are scaling factors such that |g g0 | 1 with 1 = 1 1 , and | 0 | 2 with 2 = 2 2 . The interconnection matrix H has a dyadic structure. Application

2 (H( j )) 1 (a) (b) 0 10


2

10

10

10

10

10

Figure 5.32: Structured singular values for the two-parameter plant for two cases of the result of Summary 5.7.9 shows that its structured singular value with respect to complex perturbations in the parameters is (H( j)) =
1 1+2 2 0

|1 + L0 ()|

1 1

|C( j)| + 2 2 2 ,

2 1 k2 + 2 Td + 2 3 1 + 2 T02

|0 ( j)|

(5.188)

with 0 (s) = 0 T0 s4 + (0 + T0 )s3 + s2 + g0 Td s + g0 k the nominal closed-loop characteristic polynomial. Inspection shows that the right-hand side of this expression is identical to that of (5.99) in Example 5.5.3, which was obtained by a singular value analysis based on optimal scaling. In view of the statement in Summary 5.7.10(a) this is no coincidence. Figure 5.32 repeats the plots of of Fig. 5.17 of the structured singular value for the two cases of Example 5.5.3. For case (a) the peak value of the structured singular value is greater than 1 so that robust stability is not guaranteed. Only in case (b) robust stability is certain. Since the perturbation analysis is based on dynamic rather than parameter perturbations the results are conservative. Exercise 5.7.12 (Computation of structured singular values with M ATLAB). Reproduce the plots of Fig. 5.32 using the appropriate numerical routines from the -Toolbox or the Robust Control Toolbox for the computation of structured singular values.

240

Chapter 5. Uncertainty Models and Robustness

5.8 Combined performance and stability robustness


5.8.1 Introduction
In the preceding section we introduced the structured singular value to study the stability robustness of the basic perturbation model. We continue in this section by considering performance and its robustness. Figure 5.33 represents a control system with external input w, such as disturbances, measurement noise, and reference signals, and external output z. The output signal z represents an error signal, and ideally should be zero. The transfer matrix H is the transfer matrix from the external input w to the error signal z.

Figure 5.33: Control system. Example 5.8.1 (Two-degree-of-freedom feedback system). To illustrate this model, consider the two-degree-of-freedom feedback conguration of Fig. 5.34. P is the plant, C the compensator, and F a precompensator. The signal r is the reference signal, v the disturbance, m the measurement noise, u the plant input, and y the control system output. It is easy to nd that y = = (I + PC)1 PC Fr + (I + PC)1 v (I + PC)1 PCm T Fr + Sv T m, (5.189)

where S is the sensitivity matrix of the feedback loop and T the complementary sensitivity matrix. The tracking error z = y r is given by z = y r = (T F I )r + Sv T m. Considering the combined signal r w = v m as the external input, it follows that the transfer matrix of the control system is H = TF I S T . (5.192) (5.190)

(5.191)

The performance of the control system of Fig. 5.33 is ideal if H = 0, or, equivalently, H = 0. Ideals cannot always be obtained, so we settle for the norm H to be small, rather than zero. By introducing suitable frequency dependent scaling functions (that is, by modifying H to W HV) we may arrange that satisfactory performance is obtained if and only if H

< 1.

(5.193)

Example 5.8.2 (Performance specication). To demonstrate how performance may be specied this way, again consider the two-degree-of-freedom feedback system of Fig. 5.34. For simplicity we assume that it is a SISO system, and that the reference signal and measurement noise

5.8. Combined performance and stability robustness


v + y + + m

241

Figure 5.34: Two-degree-of-freedom feedback control system. are absent. It follows from Example 5.8.1 that the control system transfer function reduces to the sensitivity function S of the feedback loop: H= 1 = S. 1 + PC (5.194)

Performance is deemed to be adequate if the sensitivity function satises the requirement

10 |Sspec| 1 .1 .01 .001 .0001 .01 .1 1 10 100

Figure 5.35: Magnitude plot of Sspec . |S( j)| < |Sspec ( j)|, , (5.195)

with Sspec a specied rational function, such as Sspec (s) = s2 . s2 + 20 s + 2 0 (5.196)

The parameter 0 determines the minimal bandwidth, while the relative damping coefcient species the allowable amount of peaking. A doubly logarithmic magnitude plot of Sspec is shown in Fig. 5.35 for 0 = 1 and = 1 . 2 The inequality (5.195) is equivalent to |S( j)V ( j)| < 1 for , with V = 1/Sspec. Redening H as H = SV this reduces to H

< 1.

(5.197)

242

Chapter 5. Uncertainty Models and Robustness

By suitable scaling we thus may arrange that the performance of the system of Fig. 5.33 is considered satisfactory if H

< 1.

(5.198)

In what follows we exploit the observation that by (1) of Summary 5.5.1 this specication is equivalent to the requirement that the articially perturbed system of Fig. 5.36 remains stable under all stable perturbations 0 such that 0 1.

Figure 5.36: Articially perturbed system

5.8.2 Performance robustness


Performance is said to be robust if H remains less than 1 under perturbations. To make this statement more specic we consider the conguration of Fig. 5.37(a). The perturbation may be structured, and is scaled such that 1. We describe the system by the interconnection equations z v H11 =H = q p H21 H12 H22 v . p (5.199)

H11 is the nominal control system transfer matrix. We dene performance to be robust if 1. the perturbed system remains stable under all perturbations, and 2. the -norm of the transfer matrix of the perturbed system remains less than 1 under all perturbations.

0 w p z q w p z q

(a)

(b)

Figure 5.37: (a) Perturbed control system. (b) Doubly perturbed system.

5.8. Combined performance and stability robustness

243

Necessary and sufcient for robust performance is that the norm of the perturbed transfer matrix from w to z in the perturbation model of Fig. 5.37(a) is less than 1 for every perturbation with norm less than or equal to 1. This, in turn, is equivalent to the condition that the augmented perturbation model of Fig. 5.37(b) is stable for every full perturbation 0 and every structured perturbation , both with norm less than or equal to 1. Necessary and sufcient for this is that H < 1, (5.200)

with the structured singular value of H with respect to the perturbation structure dened by
0

(5.201)

Summary 5.8.3 (Robust performance and stability). Robust performance of the system of Fig. 5.37(a) is achieved if and only if H < 1, where H is the structured singular value of H under perturbations of the form
0

(5.202)

0
0

. structured as specied.

(5.203)

is a full perturbation, and

w u + + z (a)

w q W2 C u P P + p + + + y

W1

(b)

Figure 5.38: SISO feedback system. (a) Nominal. (b) Perturbed.

5.8.3 SISO design for robust stability and performance


We describe an elementary application of robust performance analysis using the structured singular value (compare Doyle, Francis, and Tannenbaum (1992)). Consider the feedback control system conguration of Fig. 5.38(a). A SISO plant P is connected in feedback with a compensator C.

244

Chapter 5. Uncertainty Models and Robustness

Performance is measured by the closed-loop transfer function from the disturbance w to the control system output z, that is, by the sensitivity function S = 1/(1 + PC). Performance is considered satisfactory if
|S( j)W1 ( j)| < 1, , (5.204)

with W1 a suitable weighting function.

Plant uncertainty is modeled by the scaled uncertainty model


P P + P W2 , (5.205)

with W2 a stable function representing the maximal uncertainty, and P a scaled stable perturbation such that P 1. The block diagram of Fig. 5.38(b) includes the weighting lter W1 and the plant perturbation model. By inspection we obtain the signal balance equation y = w + p PCy, so that y= 1 (w + p). 1 + PC (5.206)

By further inspection of the block diagram it follows that z q Here S= 1 , 1 + PC T= PC 1 + PC (5.209) = = W1 y = W1 S(w + p), W2 PCy = W2 T (w + p). (5.207) (5.208)

are the sensitivity function and the complementary sensitivity function of the feedback system, respectively. Thus, the transfer matrix H in the conguration of Fig. 5.37(b) follows from q W2 T = z W1 S H H has the dyadic structure H= W2 T W1 S 1 1 . (5.211) W2 T W1 S p . w (5.210)

With the result of Summary 5.7.9 we obtain the structured singular value of the interconnection matrix H as H = =

sup (H( j)) = sup (|W1 ( j)S( j)| + |W2 ( j)T ( j)|)

(5.212) (5.213)

|W1 S| + |W2 T|

By Summary 5.8.3, robust performance and stability are achieved if and only if H < 1.

5.8. Combined performance and stability robustness

245

10 1 .1 .01 |S|

10 1 .1 .01 .001 .01 |T |

.001 .01

100

100

Figure 5.39: Nominal sensitivity and complementary sensitivity functions Example 5.8.4 (Robust performance of SISO feedback system). We consider the SISO feedback system we studied in Example 5.2.1 (p. 198) and on several other occasions, with nominal plant and compensator transfer functions P0 (s) = g0 , s2 C(s) = k + sTd , 1 + sT0 (5.214)

respectively. We use the design values of Example 5.2.2 (p. 199). In Fig. 5.39 magnitude plots are given of the nominal sensitivity and complementary sensitivity functions of this feedback system. The nominal sensitivity function is completely acceptable, so we impose as design specication that under perturbation S( j) 1 + , S0 ( j) , (5.215)

with S0 the nominal sensitivity function and the positive number a tolerance. This comes down to choosing the weighting function W1 in (5.204) as W1 = 1 . (1 + )S0 (5.216)

We consider how to select the weighting function W2 , which species the allowable perturbations. With W1 chosen as in (5.216) the robust performance criterion H < 1 according to (5.212) reduces to 1 + |W2 ( j)T0 ( j)| < 1, 1+
1+

(5.217)

with T0 the nominal complementary sensitivity function. This is equivalent to |W2 ( j)| < |T0 ( j)| , . (5.218)

Figure 5.40(a) shows a plot of the right-hand side of (5.218) for = 0.25. This right-hand side is a frequency dependent bound for the maximally allowable scaled perturbation P . The plot shows that for low frequencies the admissible proportional uncertainty is limited to /(1 + ) = 0.2,

246

Chapter 5. Uncertainty Models and Robustness

10 1 .1

10

bound

1 .1 .01 .001 .01

|(1+ )S o |
| S|

| P |

.01 .001 .01

100

100

(a) Uncertainty bound and admissible perturbation

(b) Effect of perturbation on S

Figure 5.40: The bound and its effect on S but that the system is much more robust with respect to high-frequency perturbations. In the crossover frequency region the allowable perturbation is even less than 0.2. Suppose, as we did before, that the plant is perturbed to P(s) = g . s2 (1 + s) (5.219)

The corresponding proportional plant perturbation is


P (s) =

P(s) P0 (s) = P0 (s)

gg0 g0

1 + s

(5.220)

At low frequencies the perturbation has approximate magnitude |g g0 |/g0 . For performance robustness this number denitely needs to be less than the minimum of the bound on the right hand side of (5.218), which is about 1+ 0.75. If = 0.25 then the latter number is about 0.15. Figure 5.40(a) includes a magnitude plot of P for |g g0 |/g0 = 0.1 (that is, g = 0.9 or g = 1.1) and = 0.1. For this perturbation the performance robustness test is just satised. If g is made smaller than 0.9 or larger than 1.1 or is increased beyond 0.1 then the perturbation fails the test. Figure 5.40(b) shows the magnitude plot of the sensitivity function S for g = 1.1 and = 0.1. Comparison with the bound |(1 + )S0 | shows that performance is robust, as expected. Exercise 5.8.5 (Bound on performance robustness). In Example 5.8.4 we used the performance specication (in the form of W1 ) to nd bounds on the maximally allowable scaled perturbations P . Conversely, we may use the uncertainty specication (in the form of W2 ) to estimate the largest possible performance variations. 1. Suppose that g varies between 0.5 and 1.5 and between 0 and 0.2. Determine a function W2 that tightly bounds the uncertainty P of (5.220). 2. Use (5.218) to determine the smallest value of for which robust performance is guaranteed.

5.9. Appendix: Proofs

247

3. Compute the sensitivity function S for a number of values of the parameters g and within the uncertainty region. Check how tight the bound on performance is that follows from the value of obtained in 2.

5.9 Appendix: Proofs


In this Appendix to Chapter 5 the proofs of certain results in are sketched.

5.9.1 Small gain theorem


We prove Summary 5.4.2. Proof of Small gain theorem. The proof of the small gain theorem follows by applying the xed point theorem to the feedback equation w = v + Lw that follows from Fig. 5.12(b). Denote the normed signal space that L operates on as U . For any xed input v U , if L is a contraction, so is the map dened by w v + Lw. Hence, if L is a contraction, for every v U the feedback equation has a unique solution w U . It follows that the system with input v and output w is BIBO stable. If L is a contraction with contraction constant k then it follows from Lu U k u U that L k. Hence, L < 1. Conversely, if L < 1, for any u and v in U we have Lu Lv U = L(u v) U L u v U , so that L is a contraction. This proves that L is a contraction if and only if its norm is strictly less than 1.

5.9.2 Structured singular value of a dyadic matrix


We prove Summary 5.7.9. In fact we prove the generalization that for rank 1 matrices a1 ai , b j M = a2 b 1 b2 , . . . and perturbation structure 0 1 = 0 , 2 there holds that (M) =
i

(5.221)

(5.222)

|ai bi |.

Proof of structured singular value of dyadic matrix. Given M as the product of a column and row vector M = ab we have det(I M ) = det(I ab ) = det(1 b a) = 1
i

ai bi

i.

(5.223)

This gives a lower bound for the real part of det(I M ), det(I M ) = 1
i

ai bi

1
i

|ai bi | max |
i

i|

=1
i

|ai bi |( ).

(5.224)

For det(I M ) to be zero we need at least that the lower bound is not positive, that is, that ( ) 1/ i |ai bi |. Now take equal to 0 sgn(a1 b1 ) 1 0 sgn(a2 b2 ) . = (5.225) .. i |ai bi | .

248

Chapter 5. Uncertainty Models and Robustness

Then ( ) = 1/

|ai bi | and it is such that a jb j


j j

det(I M ) = 1

= 1
j

a jb j

sgn a j b j = 0. i ai bi

(5.226)

Hence a smallest (in norm) that makes I M singular has norm 1/ i |ai bi |. Therefore (M) = |ai bi |. i It may be shown that the D-scaling upper bound for rank 1 matrices equals (M) in this case.

H-Optimization and -Synthesis

Overview Design by H -optimization involves the minimization of the peak magnitude of a suitable closed-loop system function. It is very well suited to frequency response shaping. Moreover, robustness against plant uncertainty may be handled more directly than with H2 optimization. Design by -synthesis aims at reducing the peak value of the structured singular value. It accomplishes joint robustness and performance optimization.

6.1 Introduction
In this chapter we introduce what is known as H -optimization as a design tool for linear multivariable control systems. H -optimization amounts to the minimization of the -norm of a relevant frequency response function. The name derives from the fact that mathematically the problem may be set in the space H (named after the British mathematician G. H. Hardy), which consists of all bounded functions that are analytic in the right-half complex plane. We do not go to this length, however. H -optimization resembles H2 -optimization, where the criterion is the 2-norm. Because the 2- and -norms have different properties the results naturally are not quite the same. An important aspect of H optimization is that it allows to include robustness constraints explicitly in the criterion. In 6.2 (p. 250) we discuss the mixed sensitivity problem. This special H problem is an important design tool. We show how this problem may be used to achieve the frequency response shaping targets enumerated in 1.5 (p. 27). In 6.3 (p. 258) we introduce the standard problem of H -optimization, which is the most general version. The mixed sensitivity problem is a special case. Several other special cases of the standard problem are exhibited. Closed-loop stability is dened and a number of assumptions and conditions needed for the solution of the H -optimization problem are reviewed. The next few sections are devoted to the solution of the standard H -optimization problem. In 6.4 (p. 264) we derive a formula for so-called suboptimal solutions and establish a lower 249

250

Chapter 6. H -Optimization and -Synthesis

bound for the -norm. We also establish how all stabilizing suboptimal solutions, if any exist, may be obtained. The key tool in solving H -optimization problems is J-spectral factorization. In 6.5 (p. 272) we show that by using state space techniques the J-spectral factorization that is required for the solution of the standard problem may be reduced to the solution of two algebraic Riccati equations. In 6.6 (p. 275) we discuss optimal (as opposed to suboptimal) solutions and highlight some of their peculiarities. Section 6.7 (p. 279) explains how integral control and high-frequency rolloff may be handled. Section 6.8 (p. 288) is devoted to an introductory exposition of -synthesis. This approximate technique for joint robustness and performance optimization uses H optimization to reduce the peak value of the structured singular value. Section 6.9 (p. 292) is given over to a rather elaborate description of an application of -synthesis. Section 6.10 (p. 304), nally, contains a number of proofs.

6.2 The mixed sensitivity problem


6.2.1 Introduction
In 5.6.6 (p. 227) we studied the stability robustness of the feedback conguration of Fig. 6.1. We considered fractional perturbations of the type

Figure 6.1: Feedback loop L = N D1 (I + V N W2 )N D1 (I + V D W1 )1 . (6.1)

The frequency dependent matrices V , W1 , and W2 are so chosen that the scaled perturbation p = [ D N ] satises P 1. Stability robustness is guaranteed if H where H= W1 SV . W2 T V (6.3)

< 1,

(6.2)

S = (I + L)1 and T = L(I + L)1 are the sensitivity matrix and the complementary sensitivity matrix of the closed-loop system, respectively, with L the loop gain L = PC. Given a feedback system with compensator C and corresponding loop gain L = PC that does not satisfy the inequality (6.2) one may look for a different compensator that does achieve inequality. An effective way of doing this is to consider the problem of minimizing H with respect to all compensators C that stabilize the system. If the minimal value of H is greater than 1 then no compensator exists that stabilizes the systems for all perturbations such that P 1. In this case, stability robustness is only obtained for perturbations satisfying P 1/.

6.2. The mixed sensitivity problem

251

The problem of minimizing W1 SV W2 T V (6.4)

(Kwakernaak 1983; Kwakernaak 1985) is a version of what is known as the mixed sensitivity problem (Verma and Jonckheere 1984). The name derives from the fact that the optimization involves both the sensitivity and the complementary sensitivity function. In what follows we explain that the mixed sensitivity problem cannot only be used to verify stability robustness for a class of perturbations, but also to achieve a number of other important design targets for the one-degree-of-freedom feedback conguration of Fig. 6.1. Before starting on this, however, we introduce a useful modication of the problem. We may write W2 T V = W2 L(I + L)1 V = W2 PC(I + L)1 V = W2 PUV , where U = C(I + PC)1 (6.5)

is the input sensitivity matrix introduced in 1.5 (p. 27) . For a xed plant we may absorb the plant transfer matrix P (and the minus sign) into the weighting matrix W2 , and consider the modied problem of minimizing W1 SV W2 UV (6.6)

with respect to all stabilizing compensators. We redene the problem of minimizing (6.6) as the mixed sensitivity problem. We mainly consider the SISO mixed sensitivity problem. The criterion (6.6) then reduces to the square root of the scalar quantity sup |W1 ( j)S( j)V ( j)|2 + |W2 ( j)U( j)V ( j)|2 .

(6.7)

Many of the conclusions also hold for the MIMO case, although their application may be more involved.

6.2.2 Frequency response shaping


The mixed sensitivity problem may be used for simultaneously shaping the sensitivity and input sensitivity functions. The reason is that the solution of the mixed sensitivity sensitivity problem often has the equalizing property (see 6.6.1, p. 275). This property implies that the frequency dependent function |W1 ( j)S( j)V ( j)|2 + |W2 ( j)U( j)V ( j)|2 , (6.8)

whose peak value is minimized, actually is a constant (Kwakernaak 1985). If we denote the constant as 2 , with nonnegative, then it immediately follows from |W1 ( j)S( j)V ( j)|2 + |W2 ( j)U( j)V ( j)|2 = 2 that for the optimal solution |W1 ( j)S( j)V ( j)|2 2 , |W2 ( j)U( j)V ( j)|2 2 , , . (6.10) (6.9)

252

Chapter 6. H -Optimization and -Synthesis

Hence, |S( j)| |U( j)| , |W1 ( j)V ( j)| , |W2 ( j)V ( j)| , . (6.11) (6.12)

By choosing the functions W1 , W2 , and V correctly the functions S and U may be made small in appropriate frequency regions. This is also true if the optimal solution does not have the equalizing property. If the weighting functions are suitably chosen (in particular, with W1 V large at low frequencies and W2 V large at high frequencies), then often the solution of the mixed sensitivity problem has the property that the rst term of the criterion dominates at low frequencies and the second at high frequencies: |W1 ( j)S( j)V ( j)|2 dominates at low frequencies As a result, |S( j)| |U( j)| |W1 ( j)V ( j)| |W2 ( j)V ( j)| for small, for large. (6.14) (6.15) + |W2 ( j)U( j)V ( j)|2 dominates at high frequencies = 2. (6.13)

This result allows quite effective control over the shape of the sensitivity and input sensitivity functions, and, hence, over the performance of the feedback system. Because the -norm involves the supremum frequency response shaping based on minimization of the -norm is more direct than for the H2 optimization methods of 3.5 (p. 127). Note, however, that the limits of performance discussed in 1.6 (p. 40) can never be violated. Hence, the weighting functions must be chosen with the respect due to these limits.

6.2.3 Type k control and high-frequency roll-off


In (6.146.15), equality may often be achieved asymptotically. Type k control. Suppose that |W1 ( j)V ( j)| behaves as 1/k as 0, with k a nonnegative integer. This is the case if W1 (s)V (s) includes a factor sk in the denominator. Then |S( j)| behaves as k as 0, which implies a type k control system, with excellent low-frequency disturbance attenuation if k 1. If k = 1 then the system has integrating action. High-frequency roll-off. Likewise, suppose that |W2 ( j)V ( j)| behaves as m as . This is the case if W2 V is nonproper, that is, if the degree of the numerator of W2 V exceeds that of the denominator (by m). Then |U( j)| behaves as m as . From U = C/(1 + PC) it follows that C = U/(1 + U P). Hence, if P is strictly proper and m 0 then also C behaves as m , and T = PC/(1 + PC) behaves as (m+e), with e the pole excess1 of P. Hence, by choosing m we pre-assign the high-frequency roll-off of the compensator transfer function, and the roll-offs of the complementary and input sensitivity functions. This is important for robustness against high-frequency unstructured plant perturbations.
1 The pole excess of a rational transfer function P is the difference between the number of poles and the number of zeros. This number is also known as the relative degreeof P

6.2. The mixed sensitivity problem

253

Similar techniques to obtain type k control and high-frequency roll-off are used in 3.5.4 (p. 130) and 3.5.5 (p. 132) , respectively, for H2 optimization.

6.2.4 Partial pole placement


There is a further important property of the solution of the mixed sensitivity problem that needs to be discussed before considering an example. This involves a pole cancellation phenomenon that is sometimes misunderstood. The equalizing property of 6.2.2 (p. 251) implies that W1 (s)W1 (s)S(s)S(s)V (s)V (s) + W2 (s)W2 (s)U(s)U(s)V (s)V (s) = 2 (6.16) for all s in the complex plane. We write the transfer function P and the weighting functions W1 , W2 , and V in rational form as P= A1 A2 M N , W1 = , W2 = , V= , D B1 B2 E (6.17)

with all numerators and denominators polynomials. If also the compensator transfer function is represented in rational form as C= Y X (6.18)

then it easily follows that S= DX , DX + NY U= DY . DX + NY (6.19)

The denominator Dcl = DX + NY (6.20)

is the closed-loop characteristic polynomial of the feedback system. Substituting S and U as given by (6.19) into (6.16) we easily obtain
D D M M ( A A1 B2 B2 X X + A A2 B1 B1 Y Y ) 1 2 = 2. E B B B B D D E 1 2 cl 1 2 cl

(6.21)

If A is any rational or polynomial function then A is dened by A (s) = A(s). Since the right-hand side of (6.21) is a constant, all polynomial factors in the numerator of the rational function on the left cancel against corresponding factors in the denominator. In particular, the factor D D cancels. If there are no cancellations between D D and E E B1 B1 B2 B2 then the closed-loop characteristic polynomial Dcl (which by stability has left-half plane roots only) necessarily has among its roots those roots of D that lie in the left-half plane, and the mirror images with respect to the imaginary axis of those roots of D that lie in the right-half plane. This means that the open-loop poles (the roots of D), possibly after having been mirrored into the left-half plane, reappear as closed-loop poles. This phenomenon, which is not propitious for good feedback system design, may be prevented by choosing the denominator polynomial E of V equal to the plant denominator polynomial D, so that V = M/D. With this special choice of the denominator of V , the polynomial E cancels against D in the left-hand side of (6.21), so that the open-loop poles do not reappear as closed-loop poles.

254

Chapter 6. H -Optimization and -Synthesis

Further inspection of (6.21) shows that if there are no cancellations between M M and E E B1 B1 B2 B2 , and we assume without loss of generality that M has left-half plane roots only, then the polynomial M cancels against a corresponding factor in Dcl . If we take V proper (which ensures V ( j) to be nite at high frequencies) then the polynomial M has the same degree as D, and, hence, has the same number of roots as D. All this means that by letting

V=

M , D

(6.22)

where the polynomial M has the same degree as the denominator polynomial D of the plant, the open-loop poles (the roots of D) are reassigned to the locations of the roots of M. By suitably choosing the remaining weighting functions W1 and W2 these roots may often be arranged to be the dominant poles. This technique, known as partial pole placement (Kwakernaak 1986; Postlethwaite, Tsai, and Gu 1990) allows further control over the design. It is very useful in designing for a specied bandwidth and good time response properties. In the design examples in this chapter it is illustrated how the ideas of partial pole placement and frequency shaping are combined. A discussion of the root loci properties of the mixed sensitivity problem may be found in Choi and Johnson (1996).

6.2.5 Example: Double integrator


We apply the mixed sensitivity problem to the same example as in 3.6.2 (p. 133), where H2 optimization is illustrated2 . Consider a SISO plant with nominal transfer function P0 (s) = 1 . s2 (6.23)

The actual, perturbed plant has the transfer function P(s) = g s2 (1 + s) , (6.24)

where g is nominally 1 and the nonnegative parasitic time constant is nominally 0. Perturbation analysis. We start with a preliminary robustness analysis. The variations in the parasitic time constant mainly cause high-frequency perturbations, while the low-frequency perturbations are primarily the effect of the variations in the gain g. Accordingly, we model the effect of the parasitic time constant as a numerator perturbation, and the gain variations as denominator perturbations, and write P(s) = N(s) = D(s)
1 1+s s2 g

(6.25)

Correspondingly, the relative perturbations of the denominator and the numerator are 1 D(s) D0 (s) = 1, D0 (s) g
2 Much

N(s) N0 (s) s = . N0 (s) 1 + s

(6.26)

of the text of this subsection has been taken from Kwakernaak (1993).

6.2. The mixed sensitivity problem

255

The relative perturbation of the denominator is constant over all frequencies, also in the crossover region. Because the plant is minimum-phase trouble-free crossover may be achieved (that is, without undue peaking of the sensitivity and complementary sensitivity functions). Hence, we expect that in the absence of other perturbations values of |1/g 1| up to almost 1 may be tolerated. The size of the relative perturbation of the numerator is less than 1 for frequencies below 1/, and equal to 1 for high frequencies. To prevent destabilization it is advisable to make the complementary sensitivity small for frequencies greater than 1/. As the complementary sensitivity starts to decrease at the closed-loop bandwidth, the largest possible value of dictates the bandwidth. Assuming that performance requirements specify the system to have a closedloop bandwidth of 1, we expect that in the absence of other perturbations values of the parasitic time constant up to 1 do not destabilize the system. Thus, both for robustness and for performance, we aim at a closed-loop bandwidth of 1 with small sensitivity at low frequencies and a sufciently fast decrease of the complementary sensitivity at high frequencies with a smooth transition in the crossover region. Choice of the weighting functions. To accomplish this with a mixed sensitivity design, we successively consider the choice of the functions V = M/D (that is, of the polynomial M), W1 and W2 . To obtain a good time response corresponding to the bandwidth 1, which does not suffer from sluggishness or excessive overshoot, we assign two dominant poles to the locations 1 2(1 j). 2 This is achieved by choosing the polynomial M as 1 1 2(1 + j)][s 2(1 j)] = s2 + s 2 + 1, (6.27) M(s) = [s 2 2 so that s2 + s 2 + 1 V (s) = . (6.28) s2 We tentatively choose the weighting function W1 equal to 1. Then if the rst of the two terms of the mixed sensitivity criterion dominates at low frequencies from we have from (6.14) that |S( j)| ( j)2 = |V ( j)| ( j)2 + j 2 + 1 at low frequencies. (6.29)

Figure 6.2 shows the magnitude plot of the factor 1/ V. The plot implies a very good lowfrequency behavior of the sensitivity function. Owing to the presence of the double open-loop pole at the origin the feedback system is of type 2. There is no need to correct this low frequency behavior by choosing W1 different from 1. Next contemplate the high-frequency behavior. For high frequencies V is constant and equal to 1. Consider choosing W2 as W2 (s) = c(1 + rs), (6.30)

with c and r nonnegative constants such that c = 0. Then for high frequencies the magnitude of W2 ( j) asymptotically behaves as c if r = 0, and as cr if r = 0. Hence, if r = 0 then the high-frequency roll-off of the input sensitivity function U and the compensator transfer function C is 0 and that of the complementary sensitivity T is 2 decades/decade (40 dB/decade). If r = 0 then U and C roll off at 1 decade/decade (20 dB/decade) and T rolls off at 3 decades/decade (60 dB/decade).

256

Chapter 6. H -Optimization and -Synthesis

1/|V |

10 1 .1 .01 .001
.0001

.01

.1

10

100

Figure 6.2: Bode magnitude plot of 1/ V

sensitivity |S| 10 1 .1 .01 .001 c=10 10 1 .1

complementary sensitivity |T |

c=1/100

c=1 c=1/10 c=1/100

c=1/10 .01 .001 c=1 c=10

.01

.1

10

100

.01

.1

10

100

Figure 6.3: Bode magnitude plots of S and T for r = 0

6.2. The mixed sensitivity problem

257

Solution of the mixed sensitivity problem. We rst study the case r = 0, which results in a proper but not strictly proper compensator transfer function C, and a high-frequency roll-off of T of 2 decades/decade. Figure 6.3 shows3 the optimal sensitivity function S and the corresponding complementary sensitivity function T for c = 1/100, c = 1/10, c = 1, and c = 10. Inspection shows that as c increases, |T| decreases and |S| increases, which conforms to expectation. The smaller c is, the closer the shape of |S| is to that of the plot of Fig. 6.2. We choose c = 1/10. This makes the sensitivity small with little peaking at the cut-off frequency. The corresponding optimal compensator has the transfer function C(s) = 1.2586 s + 0.61967 , 1 + 0.15563s (6.31)

and results in the closed-loop poles 1 2(1 j) and 5.0114. The two former poles dominate 2 the latter pole, as planned. The minimal -norm is H = 1.2861. Robustness against high-frequency perturbations may be improved by making the complementary sensitivity function T decrease faster at high frequencies. This is accomplished by taking the constant r nonzero. Inspection of W2 as given by (6.30) shows that by choosing r = 1 the resulting extra roll-off of U, C, and T sets in at the frequency 1. For r = 1/10 the break point is shifted to the frequency 10. Figure 6.4 shows the resulting magnitude plots. For r = 1/10
sensitivity |S| complementary sensitivity |T |

10 1 .1 .01 .001 r=1

10 1 .1 r = 1/10 r=0 .01 .001

r=1 r = 1/10 r=0

.01

.1

10

100

.01

.1

10

100

Figure 6.4: Bode magnitude plots of S and T for c = 1/10 the sensitivity function has little extra peaking while starting at the frequency 10 the complementary sensitivity function rolls off at a rate of 3 decades/decade. The corresponding optimal compensator transfer function is s + 0.5987 , (6.32) 1 + 0.20355s + 0.01267s2 which results in the closed-loop poles 1 2(1 j) and 7.3281 j1.8765. Again the former 2 two poles dominate the latter. The minimal -norm is H = 1.3833. Inspection of the two compensators (6.31) and (6.32) shows that both basically are lead compensators with high-frequency roll-off. C(s) = 1.2107
3 The

actual computation of the compensator is discussed in Example 6.6.1 (p. 276).

258

Chapter 6. H -Optimization and -Synthesis

Robustness analysis. We conclude this example with a brief analysis to check whether our expectations about robustness have come true. Given the compensator C = Y/ X the closedloop characteristic polynomial of the perturbed plant is Dcl (s) = D(s)X(s) + N(s)Y (s) = (1 + s)s2 X(s) + gY (s). By straightforward computation, which involves xing one of the two parameters g and and varying the other, the stability region of Fig. 6.5 may be established for the compensator (6.31). That for the other compensator is similar. The diagram shows that

6 4 g 2 0 0 .5 1 1.5

Figure 6.5: Stability region for = 0 the closed-loop system is stable for all g > 0, that is, for all 1 < 1 1 < . This g stability interval is larger than predicted. For g = 1 the system is stable for 0 < 1.179, which also is a somewhat larger interval than expected. The compensator (6.31) is similar to the modied PD compensator that is obtained in Example 5.2.2 (p. 199) by root locus design. Likewise, the stability region of Fig. 6.5 is similar to that of Fig. 5.5.

6.3 The standard H optimal regulation problem


6.3.1 Introduction
The mixed sensitivity problem is a special case of the so-called standard H optimal regulation problem (Doyle 1984). This standard problem is dened by the conguration of Fig. 6.6. The

plant w u G z y

K compensator
Figure 6.6: The standard H optimal regulation problem. plant is a given system with two sets of inputs and two sets of outputs. It is often referred to as the generalized plant. The signal w is an external input, and represents driving signals that generate disturbances, measurement noise, and reference inputs. The signal u is the control input.

6.3. The standard H optimal regulation problem

259

The output z has the meaning of control error, and ideally should be zero. The output y, nally, is the observed output, and is available for feedback. The plant has an open-loop transfer matrix G such that z w G11 =G = y u G21 u = Ky G12 G22 w . u (6.33)

By connecting the feedback compensator (6.34)

we obtain from y = G21 w + G22 u the closed-loop signal balance equation y = G21 w + G22 Ky, so that y = (I G22 K )1 G21 w. From z = G11 w + G12 u = G11 x + G12 Ky we then have z = [G11 + G12 K(I G22 K )1 G21 ] w. H Hence, the closed-loop transfer matrix H is H = G11 + G12 K(I G22 K )1 G21 . (6.36) (6.35)

The standard H -optimal regulation problem is the problem of determining a compensator with transfer matrix K that 1. stabilizes the closed-loop system, and 2. minimizes the -norm H w to the control error z.

of the closed-loop transfer matrix H from the external input

z2 W2 y

V + v

W1

z1

Figure 6.7: The mixed sensitivity problem. To explain why the mixed sensitivity problem is a special case of the standard problem we consider the block diagram of Fig. 6.7, where the compensator transfer function now is denoted K rather than C. In this diagram, the external signal w generates the disturbance v after passing through a shaping lter with transfer matrix V . The control error z has two components, z1 and z2 . The rst component z1 is the control system output after passing through a weighting lter with transfer matrix W1 . The second component z2 is the plant input u after passing through a weighting lter with transfer matrix W2 . It is easy to verify that for the closed-loop system z= z1 W1 SV = w, W2 UV z2 H (6.37)

260

Chapter 6. H -Optimization and -Synthesis

so that minimization of the -norm of the closed-loop transfer matrix H amounts to minimization of W1 SV W2 UV . (6.38)

In the block diagram of Fig. 6.7 we denote the input to the compensator as y, as in the standard problem of Fig. 6.6. We read off from the block diagram that z1 z2 y = = = V w W1 V w + W1 Pu, W2 u, Pu. (6.39)

Hence, the open-loop transfer matrix G for the standard problem is W1 V G= 0 V W1 P W2 P

(6.40)

Other well-known special cases of the standard problem H problem are the model matching problem and the minimum sensitivity problem (see Exercises 6.3.2 and 6.3.3, respectively.) The H -optimal regulation problem is treated in detail by Green and Limebeer (1995) and Zhou, Doyle, and Glover (1996). Exercise 6.3.1 (Closed-loop transfer matrix). Verify (6.37).

Exercise 6.3.2 (The model matching problem). The model matching problem consists of nding a stable transfer matrix K that minimizes PK
,

(6.41)

with P a given (unstable) transfer matrix. Find a block diagram that represents this problem and show that it is a standard problem with generalized plant G= P I I . 0 (6.42)

The problem is also know as the (generalized) Nehari problem. It is more of theoretical than of practical interest. Exercise 6.3.3 (Minimum sensitivity problem). A special case of the mixed sensitivity problem is the minimization of the innity norm W SV of the weighted sensitivity matrix S for the closed-loop system of Fig. 6.1. Show that this minimum sensitivity problem is a standard problem with open-loop plant G= WV V WP . P (6.43)

The SISO version of this problem historically is the rst H optimization problem that was studied. It was solved in a famous paper by Zames (1981).

6.3. The standard H optimal regulation problem

261

v + zo + + m

Figure 6.8: Two-degree- of-freedom feedback conguration Exercise 6.3.4 (Two-degree-of-freedom feedback system). As a further application consider the two-degree-of-freedom conguration of Fig. 6.8. In Fig. 6.9 the diagram is expanded with a shaping lter V1 that generates the disturbance v from the driving signal w1 , a shaping lter V2 that generates the measurement noise m = V2 w2 from the driving signal w2 , a shaping lter V0 that generates the reference signal r from the driving signal w3 , a weighting lter W1 that produces the weighted tracking error z1 = W1 (z0 r), and a weighting lter W2 that produces the weighted plant input z2 = W2 u. Moreover, the compensator C and the prelter F are combined into a single block K.

z2

w1

W2

V1

w3 y2 V0 r y1 K u P + + v z0 + + V2 w2 + W1 z1

Figure 6.9: Block diagram for a standard problem Dene the control error z with components z1 and z2 , the observed output y with components y1 and y2 as indicated in the block diagram, the external input w with components w1 , w2 , and w3 , and the control input u. Determine the open-loop generalized plant transfer matrix G and the closed-loop transfer matrix H.

6.3.2 Closed-loop stability


We need to dene what we mean by the requirement that the compensator K stabilizes the closed-loop system of Fig. 6.6. First, we take stability in the sense of internal stability, as dened in 1.3.2 (p. 12). Second, inspection of the block diagram of Fig. 6.10 shows that we

262

Chapter 6. H -Optimization and -Synthesis

certainly may require the compensator to stabilize the loop around the block G22 , but that we cannot expect that it is always possible to stabilize the entire closed-loop system. The loop

G11

z +

G12

G21 + + G

G22 u K

Figure 6.10: Feedback arrangement for the standard problem around G22 is stable means that the conguration of Fig. 6.11 is internally stable.

G22

K
Figure 6.11: The loop around G22 Exercise 6.3.5 (Closed-loop stability). Prove that the conguration of Fig. 6.11 is internally stable if and only if the four following transfer matrices have all their poles in the open left-half complex plane: (I K G22 )1 , (I K G22 )1 K, (I G22 K )1 , (I G22 K )1 G22 . (6.44)

Prove furthermore that if the loop is SISO with scalar rational transfer functions K = Y/ X, with X and Y coprime polynomials4 and G22 = N0 /D0 , with D0 and N0 coprime, then the loop is internally stable if and only if the closed-loop characteristic polynomial Dcl = D0 X N0 Y has all its roots in the open left-half complex plane. Since we cannot expect that it is always possible or, indeed, necessary to stabilize the entire closed-loop system we interpret the requirement that the compensator stabilizes the closed-loop system as the more limited requirement that it stabilizes the loop around G22 .
4 That

is, X and Y have no common polynomial factors.

6.3. The standard H optimal regulation problem

263

Example 6.3.6 (Stability for the model matching problem). To illustrate the point that it is not always possible or necessary to stabilize the entire system consider the model matching problem of Exercise 6.3.2 (p. 260). Since for this problem G22 = 0, internal stability of the loop around G22 is achieved iff K has all its poles in the open left-half plane, that is, iff K is stable. This is consistent with the problem formulation. Note that if P is unstable then the closed-loop transfer function H = P K can never be stable if K is stable. Hence, for the model matching problem stability of the entire closed-loop system can never be achieved (except in the trivial case that P by itself is stable). Exercise 6.3.7 (Mixed sensitivity problem). Show that for the mixed sensitivity problem stability of the loop around G22 is equivalent to stability of the feedback loop. Discuss what additional requirements are needed to ensure that the entire closed-loop system be stable.

6.3.3 Assumptions on the generalized plant and well-posedness


Normally, various conditions are imposed on the generalized plant G to ensure its solvability. This is what we assume at this point: (a) G is rational but not necessarily proper, where G11 has dimensions m1 k1 , G12 has dimensions m1 k2 , G21 has dimensions m2 k1 , and G22 has dimensions m2 k2 . (b) G12 has full normal column rank5 and G21 has full normal row rank. This implies that k1 m2 and m1 k2 . These assumptions are less restrictive than those usually found in the H literature (Doyle, Glover, Khargonekar, and Francis 1989; Stoorvogel 1992; Trentelman and Stoorvogel 1993). Indeed the assumptions need to be tightened when we present the state space solution of the H problem in 6.5 (p. 272). The present assumptions allow nonproper weighting matrices in the mixed sensitivity problem, which is important for obtaining high frequency roll-off. Assumption (b) causes little loss of generality. If the rank conditions are not satised then often the plant input u or the observed output y or both may be transformed so that the rank conditions hold. We conclude with discussing well-posedness of feedback systems. Let K = Y X 1 , with X and Y stable rational matrices (possibly polynomial matrices). Then K(I G22 K )1 = Y (X G22 Y )1 , so that the closed-loop transfer matrix is H = G11 + G12 Y (X G22 Y )1 G21 . (6.45)

The feedback system is said to be well-posed as long as X G22 Y is nonsingular (except at a nite number of points in the complex plane). A feedback system may be well-posed even if D0 or X is singular, so that G and K are not well-dened. This formalism allows considering innite-gain feedback systems, even if such systems cannot be realized. The compensator K is stabilizing iff X G22 Y is strictly Hurwitz, that is, has all its zeros6 in the open left-half plane. We accordingly reformulate the H problem as the problem of minimizing the -norm of H = G11 + G12 X(X G22 Y )1 G21 with respect to all stable rational X and Y such that X G22 Y is strictly Hurwitz.
rational matrix function has full normal column rank if it has full column rank everywhere in the complex plane except at a nite number of points. 6 The zeros of a polynomial matrix or a stable rational matrix are those points in the complex plane where it loses rank.
5A

264

Chapter 6. H -Optimization and -Synthesis

Example 6.3.8 (An innite-gain feedback system may be well-posed). Consider the SISO mixed sensitivity problem dened by Fig. 6.7. From (6.40) we have G22 = P. The system is well-posed if X + PY is not identical to zero. Hence, the system is well-posed even if X = 0 (as long as Y = 0). For X = 0 the system has innite gain.

6.4 Frequency domain solution of the standard problem


6.4.1 Introduction
In this section we outline the main steps that lead to an explicit formula for so-called suboptimal solutions to the standard H optimal regulation problem. These are solutions such that H

(6.46)

with a given nonnegative number. The reason for determining suboptimal solutions is that strictly optimal solutions cannot be found directly. Once suboptimal solutions may be obtained optimal solutions may be approximated by searching on the number . Suboptimal solutions may be determined by spectral factorization in the frequency domain. This solution method applies to the wide class of problems dened in 6.3 (p. 258). In 6.5 (p. 272) we specialize the results to the well-known state space solution. Some of the proofs of the results of this section may be found in the appendix to this chapter ( 6.10, p. 304). The details are described in Kwakernaak (1996). In Subsection 6.4.7 (p. 269) an example is presented. In what follows the -norm notation is to be read as H

= sup (H( j)),

(6.47)

with the largest singular value. This notation is also used if H has right-half plane poles, so that the corresponding system is not stable and, hence, has no -norm. As long as H has no poles on the imaginary axis the right-hand side of (6.47) is well-dened. We rst need the following result. Summary 6.4.1 (Inequality for -norm). Let be a nonnegative number and H a rational transfer matrix. Then H is equivalent to either of the following statements: (a) H H 2 I on the imaginary axis; (b) H H 2 I on the imaginary axis.

H is dened by H (s) = H T (s). H is called the adjoint of H. If A and B are two complex-valued matrices then A B means that B A is a nonnegative-denite matrix. On the imaginary axis means for all s = j, . The proof of Summary 6.4.1 may be found in 6.10 (p. 304). To motivate the next result we consider the model matching problem of Exercise 6.3.2 (p. 260). For this problem we have H = P K, with P a given transfer matrix, and K a stable transfer matrix to be determined. Substituting H = P K into H H 2 I we obtain from Summary 6.4.1(a) that H is equivalent to P P P K K P + K K 2 I on the imaginary axis. (6.48)

6.4. Frequency domain solution of the standard problem

265

This in turn we may rewrite as I K 2 I P P P

P I

I 0 K

on the imaginary axis.

(6.49)

This expression denes the rational matrix . Write K = Y X 1 , with X and Y rational matrices with all their poles in the left-half complex plane. Then by multiplying (6.49) on the right by X and on the left by X we obtain [X Y ] 2 I P P P

P I

X 0 Y

on the imaginary axis.

(6.50)

The inequality (6.50) characterizes all suboptimal compensators K. A similar inequality, less easily obtained than for the model matching problem, holds for the general standard problem: Summary 6.4.2 (Basic inequality). The compensator K = Y X 1 achieves H standard problem, with a given nonnegative number, if and only if [X Y ] where

for the

X 0 Y

on the imaginary axis,

(6.51)

is the rational matrix = 0


G12

I
G22

G11 G11 2 I G21 G11

G11 G21 G21 G21

0 G12 I G22

(6.52)

An outline of the proof is found in 6.10 (p. 304).

6.4.2 Spectral factorization


A much more explicit characterization of all suboptimal compensators (stabilizing or not) than that of Summary 6.4.2 follows by nding a spectral factorization of the rational matrix , if one exists. Note that is para-Hermitian, that is, = . A para-Hermitian rational matrix has a spectral factorization if there exists a nonsingular square rational matrix7 Z such that both Z and its inverse Z 1 have all their poles in the closed left-half complex plane, and = Z J Z, (6.53)

with J a signature matrix. Z is called a spectral factor. J is a signature matrix if it is a constant matrix of the form J=
7 That

In 0

0 , Im

(6.54)

is, Z is nonsingular except in a nite number of points in the complex plane.

266

Chapter 6. H -Optimization and -Synthesis

with In and Im unit matrices of dimensions n n and m m, respectively. On the imaginary axis is Hermitian. If has no poles or zeros8 on the imaginary axis then n is the number of positive eigenvalues of on the imaginary axis and m the number of negative eigenvalues. The factorization = Z J Z, if it exists, is not at all unique. If Z is a spectral factor, then so is U Z, where U is any rational or constant matrix such that U JU = J. U is said to be J-unitary. There are many such matrices. If J= 1 0 0 , 1 (6.55)

for instance, then all constant matrices U, with U= c1 cosh c1 sinh c2 sinh , c2 cosh (6.56)

, c1 = 1, c2 = 1, satisfy U T JU = J.

6.4.3 All suboptimal solutions


Suppose that

has a spectral factorization (6.57)

= Z J Z ,

with J= Im 2 0 0 . Ik2 (6.58)

It may be proved that such a spectral factorization generally exists provided is sufciently large. Write X 1 A = Z . Y B (6.59)

with A an m2 m2 rational matrix and B a k2 m2 rational matrix, both with all their poles in the closed left-half plane. Then (6.51) reduces to A A B B on the imaginary axis. (6.60)

Hence, (6.59), with A and B chosen so that they satisfy (6.60), characterizes all suboptimal compensators. It may be proved that the feedback system dened this way is well-posed iff A is nonsingular. This characterization of all suboptimal solutions has been obtained under the assumption that has a spectral factorization. Conversely, it may be shown that if suboptimal solutions exist then has a spectral factorization (Kwakernaak 1996). Summary 6.4.3 (Suboptimal solutions). Given , a suboptimal solution of the standard prob lem exists iff has a spectral factorization = Z J Z , with J=
8 The

Im 2 0

0 . Ik2
are the poles of
1 .

(6.61)

zeros of

6.4. Frequency domain solution of the standard problem

267

If this factorization exists then all suboptimal compensators are given by K = Y X 1 , with X 1 A , = Z B Y (6.62)

with A nonsingular square rational and B rational, both with all their poles in the closed left-half plane, and such that A A B B on the imaginary axis. There are many matrices of stable rational functions A and B that satisfy A A B B on the imaginary axis. An obvious choice is A = I, B = 0. We call this the central solution The central solution as dened here is not at all unique, because as seen in 6.4.2 (p. 265) the spectral factorization (6.57) is not unique.

6.4.4 Lower bound


Suboptimal solutions exist for those values of for which is spectrally factorizable. There exists a lower bound 0 so that is factorizable for all > 0 . It follows that 0 is also a lower bound for the -norm, that is, H

(6.63)

for every compensator. Summary 6.4.4 (Lower bound). Dene 1 as the rst value of as decreases from for which the determinant of
G11 G11 2 I G21 G11 G11 G21 G21 G21

(6.64)

has a -dependent9 zero on the imaginary axis. Similarly, let 2 be the rst value of as decreases from for which the determinant of
G11 G11 2 I G12 G11 G11 G12 G12 G12

(6.65) is spectrally factorizable for > 0 ,

has a -dependent zero on the imaginary axis. Then with 0 = max(1 , 2 ). Also, H

(6.66)

0 .

(6.67)

The lower bound 0 may be achieved.

The proof may be found in Kwakernaak (1996). Note that the zeros of the determinant (6.64) are the poles of , while the zeros of the determinant of (6.65) are the poles of 1 .
9 That

is, a zero that varies with .

268

Chapter 6. H -Optimization and -Synthesis

6.4.5 Stabilizing suboptimal compensators


In 6.4.4 a representation is established of all suboptimal solutions of the standard problem for a given performance level . Actually, we are not really interested in all solutions, but in all stabilizing suboptimal solutions, that is, all suboptimal compensators that make the closed-loop system stable (as dened in 6.3.2, p. 261). Before discussing how to nd such compensators we introduce the notion of a canonical spectral factorization. First, a square rational matrix M is said to be biproper if both M and its inverse M 1 are proper. M is biproper iff lim|s| M(s) is nite and nonsingular. The spectral factorization = Z J Z of a biproper rational matrix is said to be canonical if the spectral factor Z is biproper. Under the general assumptions of 6.4 (p. 264) the matrix is not necessarily biproper. If it is not biproper then the denition of what a canonical factorization is needs to be adjusted (Kwakernaak 1996). The assumptions on the state space version of standard problem that are introduced in 6.5 (p. 272) ensure that is biproper. We consider the stability of suboptimal solutions. According to 6.3.3 (p. 263) closed-loop stability depends on whether the zeros of X G22 Y all lie in the open left-half plane. Denote
1 Z = M =

M,11 M,21

M,12 M,22

(6.68)

Then by (6.62) all suboptimal compensators are characterized by X = M,11 A + M,12 B, Correspondingly, X G22 Y = = (M,11 G22 M,21 ) A + (M,12 A G22 M,22 )B (M,11 G22 M,21 )(I + TU) A. (6.70) Y = M,21 A + M,22 B. (6.69)

Here T = (M,11 G22 M,21 )1 (M,12 G22 M,22 ) and U = BA1 . Suppose that the factorization of is canonical. Then it may be proved (Kwakernaak 1996) that as the real number varies from 0 to 1, none of the zeros of (M,11 G22 M,21 )(I + TU) A (6.71)

crosses the imaginary axis. This means that all the zeros of X G22 Y are in the open left-half complex plane iff all the zeros of (M,11 G22 M,21 ) A are in the open left-half complex plane. It follows that a suboptimal compensator is stabilizing iff both A and M,11 G22 M,21 have all their roots in the open left-half plane. The latter condition corresponds to the condition that the central compensator, that is, the compensator obtained for A = I and B = 0, be stabilizing. Thus, we are led to the following conclusion: Summary 6.4.5 (All stabilizing suboptimal compensators). Suppose that the factorization = Z J Z is canonical. 1. A stabilizing suboptimal compensator exists if and only if the central compensator is stabilizing. 2. If a stabilizing suboptimal compensator exists then all stabilizing suboptimal compensators K = Y X 1 are given by X 1 A , = Z B Y (6.72)

6.4. Frequency domain solution of the standard problem

269

with A nonsingular square rational and B rational, both with all their poles in the open left-half complex plane, such that A has all its zeros in the open left-half complex plane and A A B B on the imaginary axis.

Exercise 6.4.6 (All stabilizing suboptimal solutions). Prove that if the factorization is canonical then all stabilizing suboptimal compensators K = Y X 1 , if any exist, may be characterized by X 1 I = Z , Y U with U rational stable such that U

(6.73) 1.

6.4.6 Search procedure for near-optimal solutions


In the next section we discuss how to perform the spectral factorization of . Assuming that we know how to do this numerically analytically it can be done in rare cases only the search procedure of Fig. 6.12 serves to obtain near-optimal solutions. Optimal solutions are discussed in 6.6 (p. 275).
Compute the lower bound o , or guess a lower bound Choose a value of > o Decrease Does a canonical spectral factorization of exist? Yes No Increase

Compute the central compensator

Is the compensator stabilizing? Yes No

Is the compensator close enough to optimal? No Yes Exit

Figure 6.12: Search procedure

6.4.7 Example: Model matching problem


By way of example, we consider the solution of the model matching problem of Exercise 6.3.2 (p. 260) with 1 . (6.74) P(s) = 1s

270

Chapter 6. H -Optimization and -Synthesis

Lower bound. G(s) =

By (6.42) we have
1 s1

1 0

(6.75)

It may easily be found that the determinants of the matrices (6.64) and (6.65) are both equal to 2 . It follows that 1 = 2 = 0 so that also 0 = 0. Indeed, the compensator K(s) = P(s) = 1/(1 s) achieves H = P K = 0. Note that this compensator is not stable. Spectral factorization. It is straightforward to nd from (6.52) and (6.75) that for this problem is given by 1 1 2 2 1 1s2 1+s (6.76) . (s) = 1 2 1 2 1s We need > 0. It may be veried that for = signature matrix 2 3 Z (s) =
4 2 1 4

1 2

the matrix

has the spectral factor and

+s

1 2 1 4

1+s

1+s
1 2 + 1 4 2 1 4

The inverse of Z is 1 Z (s) =

1 4 ( 2 1 ) 4

+s

J=

1 0

0 . 1

(6.77)

1+s

1+s

2 + 1 4 2 1 4 1 4 2 1 4

+s

1+s

2 1 4

. (6.78)

1+s +s)

2 3 4 ( 2 1 4

1+s

1+s

1 is biproper. For = 1 both Z and Z are proper, so that the factorization is canonical. For 2 1 = 2 the spectral factor Z is not dened. It may be proved that for = 1 actually no canonical 2 factorization exists. Noncanonical factorization do exist for this value of , though.

Suboptimal compensators. It follows by application of Summary 6.4.3 (p. 266) after some algebra that for = 1 all suboptimal compensators have the transfer function 2 K(s) =
1 4 2 1 4 2+ 1 4 2 1 4

2 3 4 2 1 4

+ s U(s)
2 1 4

+s

(6.79)

U(s)

with U = B/ A such that A A B B on the imaginary axis. We consider two special cases. Central compensator. Let A = 1, B = 0, so that U = 0. We then have the central compensator K(s) =
1 4 2 1 4 2+ 1 4 2 1 4

+s

(6.80)

6.4. Frequency domain solution of the standard problem

271

For > 1 the compensator is stable and for 0 < 2 transfer function may be found to be H(s) = P(s) K(s) =
2 2 1 4 2+ 1
4

1 2

it is unstable. The resulting closed-loop

(1 + s)

( 2 4 + s)(1 s) 1

(6.81)

The magnitude of the corresponding frequency response function takes its peak value at the frequency 0, and we nd H

= |H(0)| =

2 . 2 + 1 4

(6.82)

It may be checked that H

, with equality only for = 0 and = 1 . 2

Equalizing compensator. Choose A = 1, B = 1, so that U = 1. After simplication we obtain from (6.79) the compensator s+ K(s) = s+
1 + 1 2 2

+ 1 2 1 2 + 1 2

(6.83)

For this compensator we have the closed-loop transfer function H(s) = (1 + s)(s (1 s)(s +
1 2 + 1 2 1 2 + 1 2

) )

(6.84)

The magnitude of the corresponding frequency response function H( j) is constant and equal to . Hence, the compensator (6.83) is equalizing with H = . Stabilizing compensators. The spectral factorization of exists for > 0, and is canonical for = 1 . The central compensator (6.80) is stabilizing iff the compensator itself is stable, 2 which is the case for > 1 . Hence, according to Summary 6.4.5 (p. 268) for > 1 stabilizing 2 2 suboptimal compensators exist. They are given by K(s) =
1 4 2 1 4 2+ 1 4 2 1 4

2 3 4 2 1 4

+ s U(s)
2 1 4

+s

(6.85)

U(s)

with U scalar stable rational such that U 1. In particular, the equalizing compensator (6.83), obtained by letting U(s) = 1, is stabilizing. Since the central compensator is not stabilizing for < 1 , we conclude that no stabilizing 2 compensators exist such that H < 1 . 2 The optimal compensator. Investigation of (6.79) shows that for 1 the transfer function 2 of all suboptimal compensators approaches K0 (s) = 1 . This compensator is (trivially) stable, and 2 the closed-loop transfer function is H0 (s) = P(s) K0 (s) =
1 2

1+s , 1s

(6.86)

272

Chapter 6. H -Optimization and -Synthesis

so that H0 = 1 . Since for < 1 no stabilizing compensators exist we conclude that the 2 2 compensator K0 = 1 is optimal, that is, minimizes H with respect to all stable compensators. 2

6.5 State space solution


6.5.1 Introduction
In this section we describe how to perform the spectral factorization needed for the solution of the standard H optimal regulator problem. To this end, we rst introduce a special version of the problem in state space form. In 6.5.5 (p. 274) a more general version is considered. The state space approach leads to a factorization algorithm that relies on the solution of two algebraic Riccati equations.

6.5.2 State space formulation


Consider a standard plant that is described by the state differential equation x = Ax + Bu + Fw1 , (6.87)

with A, B, and F constant matrices, and the vector-valued signal w1 one component of the external input w. The observed output y is y = Cx + w2 , (6.88)

with w2 the second component of w, and C another constant matrix. The error signal z, nally, is assumed to be given by z= Hx , u (6.89)

with also H a constant matrix. It is easy to see that the transfer matrix of the standard plant is H(sI A)1 F 0 H(sI A)1 B (6.90) G(s) = . I 0 0 C(sI A)1 F I C(sI A)1 B This problem is a stylized version of the standard problem that is reminiscent of the LQG regulator problem. In this and other state space versions of the standard H problem it is usual to require that the compensator stabilize the entire closed-loop system, and not just the loop around G22 . This necessitates the assumption that the system x = Ax + Bu, y = Hx be stabilizable and detectable. Otherwise, closed-loop stabilization is not possible.

6.5.3 Factorization of

In 6.10 (p. 304) it is proved that for the state space standard problem described in 6.5.2 the rational matrix as dened in Summary 6.4.2 (p. 265) may be spectrally factored as = 1 Z J Z , where the inverse M = Z of the spectral factor is given by M (s) = I 0 C 1 0 (sI A2 )1 (I 2 X1 X2 )1 [X1 C T B]. + I BT X2 (6.91)

6.5. State space solution


The constant matrix A2 is dened as 1 A2 = A + ( 2 F F T B BT )X2 .

273

(6.92)

The matrix X1 is a symmetric solution if any exists of the algebraic Riccati equation 0 = A X1 + X1 AT + F F T X1 (C T C such that 1 A1 = A + X1 ( 2 H T H C T C) (6.94) 1 T H H)X1 2 (6.93)

is a stability matrix10 . The matrix X2 is a symmetric solution again, if any exists of the algebraic Riccati equation 0 = AT X2 + X2 A + H T H X2 (BBT 1 F F T )X2 2 (6.95)

such that A2 is a stability matrix. The numerical solution of algebraic Riccati equations is discussed in 3.2.9 (p. 114).

6.5.4 The central suboptimal compensator


Once the spectral factorization has been established, we may obtain the transfer matrix of the central compensator as K = Y X 1 , with X 1 I = M = Z 0 Y I M,11 = . 0 M,21 (6.96)

1 After some algebra we obtain from (6.91) the transfer matrix K = M,21 M,11 of the central compensator in the form

K(s) = BT X2 sI A

1 F F T X2 + BBT X2 2
1

1 + (I 2 X1 X2 )1 X1 C T C

1 (I 2 X1 X2 )1 X1 C T .

(6.97)

In state space form the compensator may be represented as x u = = 1 1 F F T X2 ) x + (I 2 X1 X2 )1 X1 C T (y C x) + Bu, 2 BT X2 x. (A + (6.98) (6.99)

The compensator consists of an observer interconnected with a state feedback law. The order of the compensator is the same as that of the standard plant.
10 A

matrix is a stability matrix if all its eigenvalues have strictly negative real part.

274

Chapter 6. H -Optimization and -Synthesis

6.5.5 A more general state space standard problem


A more general state space representation of the standard plant is x z y = = = Ax + B1 w + B2 u, C1 x + D11 w + D12 u, C2 x + D21 w + D22 u. (6.100) (6.101) (6.102)

In the -Tools MATLAB toolbox (Balas, Doyle, Glover, Packard, and Smith 1991) and the Robust Control Toolbox (Chiang and Safonov 1992) a solution of the corresponding H problem based on Riccati equations is implemented that requires the following conditions to be satised: 1. ( A, B2 ) is stabilizable and (C2 , A) is detectable. 2. D12 and D21 have full rank. 3. 4. A jI C1 A jI C2 B2 has full column rank for all (hence, D12 is tall11 ). D12 B1 has full row rank for all (hence, D21 is wide). D21

The solution of this more general problem is similar to that for the simplied problem. It consists of two algebraic Riccati equations whose solutions dene an observer cum state feedback law. The expressions, though, are considerably more complex than those in 6.5.3 and 6.5.4. The solution is documented in a paper by Glover and Doyle (1988). More extensive treatments may be found in a celebrated paper by Doyle, Glover, Khargonekar, and Francis (1989) and in Glover and Doyle (1989). Stoorvogel (1992) discusses a number of further renements of the problem.

6.5.6 Characteristics of the state space solution


We list a few properties of the state space solution. 1. The assumption of stabilizability and detectability ensures that the whole feedback system is stabilized, not just the loop around G22 . There are interesting H problems that do not satisfy the stabilizability or detectability assumption, and, hence, cannot be solved with this algorithm. 2. The eigenvalues of the Hamiltonian corresponding to the Riccati equation (6.93) are the poles of . Hence, the number 1 dened in the discussion on the lower bound in 6.4 may be determined by nding the rst value of as decreases from such that any of these eigenvalues reaches the imaginary axis. Likewise, the eigenvalues of the Hamiltonian corresponding to the Riccati equation (6.95) are the poles of the inverse of . Hence, the number 2 dened in the discussion on the lower bound in 6.4 may be determined by nding the rst value of as decreases from such that any of these eigenvalues reaches the imaginary axis. 3. The required solutions X1 and X2 of the Riccati equations (6.93) and (6.95) are nonnegative-denite.
11 A

matrix is tall if it has at least as many rows as columns. It is wide if it has at least as many columns as rows.

6.6. Optimal solutions to the H problem

275

4. The central compensator is stabilizing iff max (X2 X1 ) < 2 , (6.103)

where max denotes the largest eigenvalue. This is a convenient way to test whether stabilizing compensators exist. 5. The central compensator is of the same order as the generalized plant (6.1006.102). 6. The transfer matrix K of the central compensator is strictly proper. 7. The factorization algorithm only applies if the factorization is canonical. If the factorization is noncanonical then the solution breaks down. This is described in the next section.

6.6 Optimal solutions to the H problem


6.6.1 Optimal solutions
In 6.4.6 (p. 269) a procedure is outlined for nding optimal solutions that involves a search on the parameter . As the search advances the optimal solution is approached more and more closely. There are two possibilities for the optimal solution: Type A solution. The central suboptimal compensator is stabilizing for all 0 , with 0 the lower bound discussed in 6.4.4 (p. 267). In this case, the optimal solution is obtained for = 0 . Type B solution. As varies, the central compensator becomes destabilizing as decreases below some number opt with opt 0 . In type B solutions a somewhat disconcerting phenomenon occurs. In the example of 6.4.7 (p. 269) several of the coefcients of the spectral factor and of the compensator transfer matrix grow without bound as approaches opt . This is generally true. In the state space solution the phenomenon is manifested by the fact that as approaches opt either of the two following eventualities occurs (Glover and Doyle 1989): (B1.) The matrix I
1 2

X2 X1 becomes singular.

(B2.) In solving the Riccati equation (6.93) or (6.95) or both the top coefcient matrix E1 of 3.2.9 (p. 114) becomes singular, so that the solution of the Riccati equation ceases to exist. In both cases large coefcients occur in the equations that dene the central compensator. The reason for these phenomena for type B solutions is that at the optimum the factorization of is non-canonical. As illustrated in the example of 6.4.7, however, the compensator transfer matrix approaches a well-dened limit as opt , corresponding to the optimal compensator. The type B optimal compensator is of lower order than the suboptimal central compensator. Also this is generally true. It is possible to characterize and compute all optimal solutions of type B (Glover, Limebeer, Doyle, Kasenally, and Safonov 1991; Kwakernaak 1991). An important characteristic of optimal solutions of type B is that the largest singular value (H( j)) of the optimal closed-loop frequency response matrix H is constant as a function of the frequency . This is known as the equalizing property of optimal solutions.

276

Chapter 6. H -Optimization and -Synthesis

Straightforward implementation of the two-Riccati equation algorithm leads to numerical difculties for type B problems. As the solution approaches the optimum several or all of the coefcients of the compensator become very large. Because eventually the numbers become too large the optimum cannot be approached too closely. One way to avoid numerical degeneracy is to represent the compensator in what is called descriptor form (Glover, Limebeer, Doyle, Kasenally, and Safonov 1991). This is a modication of the familiar state representation and has the appearance Ex y = = Fx + Gu, Hx + Lu. (6.104) (6.105)

If E is square nonsingular then by multiplying (6.104) on the left by E 1 the equations (6.104 6.105) may be reduced to a conventional state representation. The central compensator for the H problem may be represented in descriptor form in such a way that E becomes singular as the optimum is attained but all coefcients in (6.1046.105) remain bounded. The singularity of E may be used to reduce the dimension of the optimal compensator (normally by 1). General descriptor systems (with E singular) are discussed in detail in Aplevich (1991).

6.6.2 Numerical examples


We present the numerical solution of two examples of the standard problem. Example 6.6.1 (Mixed sensitivity problem for the double integrator). We consider the mixed sensitivity problem discussed in 6.2.5 (p. 254). With P, V, W1 and W2 as in 6.3, by (6.39) the standard plant transfer matrix is M(s) 1 G(s) =
s2

0 M(s) s2

c(1 + rs) , s12

s2

(6.106)

where M(s) = s2 + s 2 + 1. We consider the case r = 0 and c = 0.1. It is easy to check that for r = 0 a state space representation of the plant is x= 1 0 0 1 x+ w+ u, 1 0 0 2 A z= 0 0 C1 y= 0 B1 B2 (6.108) (6.107)

1 0 1 x+ w+ u, 0 c 0 D11 D12

1 x + 1 w. C2 D21

(6.109)

Note that we constructed a joint minimal realization of the blocks P and V in the diagram of Fig. 6.7. This is necessary to satisfy the stabilizability condition of 6.5.5 (p. 274). A numerical solution may be obtained with the help of the -Tools M ATLAB toolbox (Balas, Doyle, Glover, Packard, and Smith 1991). The search procedure for H state space problems

6.6. Optimal solutions to the H problem

277

implemented in -Tools terminates at = 1.2861 (for a specied tolerance in the -search of 108 ). The state space representation of the corresponding compensator is 0.3422 109 x= 1 1.7147 109 0.3015 x+ y, 1.4142 0.4264 (6.110) (6.111)

u = [1.1348 109 5.6871 109 ] x. Numerical computation of the transfer function of the compensator results in K(s) = 2.7671 109 s + 1.7147 109 . s2 + 0.3422 109 s + 2.1986 109

(6.112)

The solution is of type B as indicated by the large coefcients. By discarding the term s2 in the denominator we may reduce the compensator to K(s) = 1.2586 s + 0.6197 , 1 + 0.1556s (6.113)

which is the result stated in (6.31)(p. 257). It may be checked that the optimal compensator possesses the equalizing property: the mixed sensitivity criterion does not depend on frequency. The case r = 0 is dealt with in Exercise 6.7.2 (p. 288). Example 6.6.2 (Disturbance feedforward). Hagander and Bernhardsson (1992) discuss the example of Fig. 6.13. A disturbance w acts on a plant with transfer function 1/(s + 1). The

w K(s) y + u z2
Figure 6.13: Feedforward conguration plant is controlled by an input u through an actuator that also has the transfer function 1/(s + 1). The disturbance w may be measured directly, and it is desired to control the plant output by feedforward control from the disturbance. Thus, the observed output y is w. The control error has as components the weighted control system output z1 = x2 /, with a positive coefcient, and the control input z2 = u. The latter is included to prevent overly large inputs. In state space form the standard plant is given by x= 1 0 1 0 w+ u, x+ 0 1 1 1 A B1 B2 (6.114)

1 s+1 actuator

+ x1

1 s+1 plant

x2

z1

278

Chapter 6. H -Optimization and -Synthesis


1

z=

0 0 C1

x+

0 0 w+ u, 0 1 D11 D12

(6.115)

y= 0 C2

0 x + 1 w. D21

(6.116)

The lower bound of 6.4.4 (p. 267) may be found to be given by 0 = 1 1 + 2 . (6.117)

Dene the number c = 5+1 = 1.270. 2 (6.118)

Hagander and Bernhardsson prove this (see Exercise 6.6.3): 1. For > c the optimal solution is of type A. 2. For < c the optimal solution is of type B. For = 2, for instance, numerical computation using -Tools with a tolerance 108 results in 1 opt = 0 = = 0.4472, 5 with the central compensator 1.6229 0.4056 0 x= x+ y, 1 1 0.7071 u = [0.8809 0.5736] x. (6.120) (6.121) (6.119)

The coefcients of the state representation of the compensator are of the order of unity, and the compensator transfer function is K(s) = 0.4056s 0.4056s . s2 + 2.6229s + 2.0285 (6.122)

There is no question of order reduction. The optimal solution is of type A. For = 1 numerical computation leads to opt = 0.7167 > 0 = 0.7071, with the central compensator 6.9574 107 x= 1 4.9862 107 0 x+ y, 1 1 (6.124) (6.125) (6.123)

u = [6.9574 107 4.9862 107 ] x.

6.7. Integral control and high-frequency roll-off

279

The compensator transfer function is K(s) = 4.9862 107 s 4.9862 107 . + 0.6957 108 s + 1.1944 108 (6.126)

s2

By discarding the term s2 in the denominator this reduces to the rst-order compensator K(s) = 0.7167 s+1 . s + 1.7167 (6.127)

The optimal solution now is of type B. Exercise 6.6.3 (Disturbance feedforward). 1. Verify that the lower bound 0 is given by (6.117). 2. Prove that c as given by (6.118) separates the solutions of type A and type B.

3. Prove that if < c then the minimal -norm opt is the positive solution of the equation 4 + 2 3 = 1/2 (Hagander and Bernhardsson 1992). 4. Check the frequency dependence of the largest singular value of the closed-loop frequency response matrix for the two types of solutions. 5. What do you have to say about the optimal solution for = c ?

6.7 Integral control and high-frequency roll-off


6.7.1 Introduction
In this section we discuss the application of H optimization, in particular the mixed sensitivity problem, to achieve two specic design targets: integral control and high-frequency roll-off. Integral control is dealt with in 6.7.2. In 6.7.3 we explain how to design for high-frequency roll-off. Subsection 6.7.4 is devoted to an example. The methods to obtain integral control and high frequency roll-off discussed in this section for H design may also be used with H2 optimization. This is illustrated for SISO systems in 3.5.4 (p. 130) and 3.5.5 (p. 132).

6.7.2 Integral control


In 2.3 (p. 64) it is explained that integral control is a powerful and important technique. By making sure that the control loop contains integrating action robust rejection of constant disturbances may be obtained, as well as excellent low-frequency disturbance attenuation and command signal tracking. In 6.2.3 (p. 252) it is claimed that the mixed sensitivity problem allows to

Figure 6.14: One-degree-of-freedom feedback system

280

Chapter 6. H -Optimization and -Synthesis

design for integrating action. Consider the SISO mixed sensitivity problem for the one-degree-offreedom conguration of Fig. 6.14. The mixed sensitivity problem amounts to the minimization of the peak value of the function |V ( j)W1 ( j)S( j)|2 + |V ( j)W2 ( j)U( j)|2 , . (6.128)

S and U are the sensitivity function and input sensitivity function S= 1 , 1 + PC U= C , 1 + PC (6.129)

respectively, and V , W1 and W2 suitable frequency dependent weighting functions. If the plant P has natural integrating action, that is, P has a pole at 0, then no special provisions are needed. In the absence of natural integrating action we may introduce integrating action by letting the product V W1 have a pole at 0. This forces the sensitivity function S to have a zero at 0, because otherwise (6.128) is unbounded at = 0. There are two ways to introduce such a pole into V W1 : 1. Let V have a pole at 0, that is, take V (s) = V0 (s) , s (6.130)

with V0 (0) = 0. We call this the constant disturbance model method. 2. Let W1 have a pole at 0, that is, take W1 (s) = W1o (s) , s (6.131)

with W1o (0) = 0. This is the constant error suppression method. We discuss these two possibilities. Constant disturbance model. We rst consider letting V have a pole at 0. Although V W1 needs to have a pole at 0, the weighting function V W2 cannot have such a pole. If V W2 has a pole at 0 then U would be forced to be zero at 0. S and U cannot vanish simultaneously at 0. Hence, if V has a pole at 0 then W2 should have a zero at 0. This makes sense. Including a pole at 0 in V means that V contains a model for constant disturbances. Constant disturbances can only be rejected by constant inputs to the plant. Hence, zero frequency inputs should not be penalized by W2 . Therefore, if V is of the form (6.130) then we need W2 (s) = sW2o (s), (6.132)

with W2o (0) = . Figure 6.15(a) denes the interconnection matrix G for the resulting standard problem. Inspection shows that owing to the pole at 0 in the block V0 (s)/s outside the feedback loop this standard problem does not satisfy the stabilizability condition of 6.5.5 (p. 274). The rst step towards resolving this difculty is to modify the diagram to that of Fig. 6.15(b), where the plant transfer matrix has been changed to s P(s)/s. The idea is to construct a simultaneous minimal state realization of the blocks V0 (s)/s and s P(s)/s. This brings the unstable pole at 0 inside the loop.

6.7. Integral control and high-frequency roll-off

281

(a)

w Vo(s) s + v

sW2o (s)

K(s)

P(s)

W1(s)

z1

z2 (b) sW2o(s)

w Vo(s) s sP(s) s + v + W1(s) z1

K(s)

(c)

w Vo (s) s P(s) s + v + W1(s) z1

W2 o(s)

sK(s) Ko(s)

Figure 6.15: Constant disturbance model

282

Chapter 6. H -Optimization and -Synthesis

The difculty now is, of course, that owing to the cancellation in s P(s)/s there is no minimal realization that is controllable from the plant input. Hence, we remove the factor s from the numerator of s P(s)/s by modifying the diagram to that of Fig. 6.15(c). By a minimal joint realization of the blocks V0 (s)/s and P(s)/s the interconnection system may now be made stabilizable. This is illustrated in the example of 6.7.4. The block diagram of Fig. 6.15(c) denes a modied mixed sensitivity problem. Suppose that the compensator K0 solves this modied problem. Then the original problem is solved by compensator K(s) = K0 (s) . s (6.133)

Constant error suppression. We next consider choosing W1 (s) = W1o (s)/s. This corresponds to penalizing constant errors in the output with innite weight. Figure 6.16(a) shows the corresponding block diagram. Again straightforward realization results in violation of the stabilizability condition of 6.5.5 (p. 274). The offending factor 1/s may be pulled inside the loop as in Fig. 6.16(b). Further block diagram substitution yields the arrangement of Fig. 6.16(c). Integrator in the loop method. Contemplation of the block diagrams of Figs. 6.15(c) and 6.16(c) shows that both the constant disturbance model method and the constant error rejection method may be reduced to the integrator in the loop method. This method amounts to connecting an integrator 1/s in series with the plant P as in Fig. 6.17(a), and doing a mixed sensitivity design for this augmented system. After a compensator K0 has been obtained for the augmented plant the actual compensator that is implemented consists of the series connection K(s) = K0 (s)/s of the compensator K0 and the integrator as in Fig. 6.17(b). To prevent the pole at 0 of the integrator from reappearing in the closed-loop system it is necessary to let V have a pole at 0. Replacing V (s) with V0 (s)/s in the diagram of Fig. 6.17(a) produces the block diagrams of Figs. 6.15(c) and 6.16(c).

6.7.3 High-frequency roll-off


As explained in 6.2.3 (p. 252), high-frequency roll-off in the compensator transfer function K and the input sensitivity function U may be pre-assigned by suitably choosing the high-frequency behavior of the weighting function W2 . If W2 (s)V (s) is of order sm as s approaches then K and U have a high-frequency roll-off of m decades/decade. Normally V is chosen biproper12 such that V () = 1. Hence, to achieve a roll-off of m decades/decade it is necessary that W2 behaves as sm for large s. If the desired roll-off m is greater than 0 then the weighting function W2 needs to be nonproper. The resulting interconnection matrix W1 V W1 P W2 (6.134) G= 0 V P is also nonproper and, hence, cannot be realized as a state system of the form needed for the state space solution of the H -problem. The makeshift solution usually seen in the literature is to cut off the roll-on at high frequency. Suppose by way of example that the desired form of W2 is W2 (s) = c(1 + rs), with c and r
12 That

is, both V and 1/ V are proper. This means that V() is nite and nonzero.

6.7. Integral control and high-frequency roll-off

283

(a)

z2 W2 (s)

V(s) + v

K(s)

P(s)

W1o(s) s

z1

z2 (b) W2 (s)

V(s) + v

K(s)

P(s)

W1o(s) s s W1o (s)

z2 (c) W2 (s)

w V(s) W (s) 1o s P(s)W1o (s) s + v + z1

sK(s) W1o (s) Ko(s)

Figure 6.16: Constant error suppression

284

Chapter 6. H -Optimization and -Synthesis

(a)

z2

w V(s) Po (s) + P(s) + v W1(s) z1

W2 (s)

Ko (s)

1 s

(b) Ko(s)

K(s) 1 s P(s)

Figure 6.17: Integrator in the loop method positive constants. Then W2 may be made proper by modifying it to W2 (s) = c(1 + rs) , 1 + s (6.135)

with r a small positive time constant. This contrivance may be avoided by the block diagram substitution of Fig. 6.18 (Krause 1 1992). If W2 is nonproper then, in the SISO case, W2 is necessarily proper, and there is no difculty in nding a state realization of the equivalent mixed sensitivity problem dened by 1 Fig. 6.18 with the modied plant P0 = W2 P. Suppose that the modied problem is solved by the compensator K0 . Then the original problem is solved by
1 K = K0 W2 .

(6.136)

Inspection suggests that the optimal compensator K has the zeros of W2 as its poles. Because this is in fact not the case the extra poles introduced into K cancel against corresponding zeros. This increase in complexity is the price of using the state space solution. The example of 6.7.4 illustrates the procedure.

6.7.4 Example
We present a simple example to illustrate the mechanics of designing for integrating action and high-frequency roll-off. Consider the rst-order plant P(s) = p , s+ p (6.137)

with p = 0.1. These are the design specications: 1. Constant disturbance rejection.

6.7. Integral control and high-frequency roll-off

285

w z2 Ko y K W2 u W21 Po P +

V + v W1 z1

Figure 6.18: Block diagram substitution for nonproper W2 2. A closed-loop bandwidth of 1. 3. A suitable time response to step disturbances without undue sluggishness or overshoot. 4. High-frequency roll-off of the compensator and input sensitivity at 1 decade/decade. We use the integrator in the loop method of 6.7.2 to design for integrating action. Accordingly, the plant is modied to p . (6.138) P0 (s) = s(s + p) Partial pole placement (see 6.2.4, p. 253) is achieved by letting s2 + as + b . (6.139) s(s + p) The choice a = 2 and b = 1 places two closed-loop poles at the roots 1 2(1 j) (6.140) 2 of the numerator of V . We plan these to be the dominant closed-loop poles in order to achieve a satisfactory time response and the required bandwidth. To satisfy the high-frequency roll-off specication we let V (s) = W2 (s) = c(1 + rs), (6.141)

with the positive constants c and r to be selected13 . It does not look as if anything needs to be accomplished with W1 so we simply let W1 (s) = 1. Figure 6.19 shows the block diagram of the mixed sensitivity problem thus dened. It is obtained from Fig. 6.17(a) with the substitution of Fig. 6.18. We can now see why in the eventual design the compensator K has no pole at the zero 1/r of the weighting factor W2 . The diagram of Fig. 6.19 denes a standard mixed sensitivity problem with P(s) V (s)
13 See

= =

p rc

s(s + 1 )(s + p) r

(6.142)

(s + 1 )(s2 + as + b) s2 + as + b r = , s(s + p) s(s + 1 )(s + p) r

(6.143)

also Exercise 6.7.1 (p. 287).

286

Chapter 6. H -Optimization and -Synthesis

z2 Po(s) y Ko(s) uo 1 c(1+rs) u p s(s+p)

s2+as+1 s(s+p) + + z1

Figure 6.19: Modied block diagram for the example and W1 (s) = W2 (s) = 1. If V is in the form of the rightmost side of (6.143) then its numerator has a zero at 1/r. By the partial pole assignment argument of 6.2.4 (p. 253) this zero reappears as a closed-loop pole. Because this zero is also an open-loop pole it is necessarily a zero of the compensator K0 that solves the mixed sensitivity problem of Fig. 6.19. This zero, in turn, cancels in the transfer function K of (6.136). The system within the larger shaded block of Fig. 6.19 has the transfer matrix representation
2 z1 = s + as + b s(s + p)

p s(s + p)

w . u

(6.144)

This system has the minimal state realization x1 x2 z1 The block u= 1 u0 c(1 + rs) (6.147) = = 0 0 1 p 0 1 x1 b p + w+ u, a p 0 x2 x1 + w. x2 (6.145) (6.146)

may be represented in state form as x3 u = = 1 1 x3 + u 0 , r c 1 x3 . r (6.148) (6.149)

Combining the state differential equations (6.145) and (6.148) and arranging the output equations we obtain the equations for the interconnection system that denes the standard problem as

6.7. Integral control and high-frequency roll-off

287

follows: x1 x2 = x3

0 0 1 p 0 0 A

0 x1 b 0 x 2 + a p w + 0 u 0 , 1 x3 0 1 r c
p r

(6.150)

B1 x 0 0 1 0 1 1 x2 + w+ u 1 0 0 0 0 0 x3 D12 D11 C1 x1 0 1 0 x2 + (1) w. x3 D21 C2

B2 (6.151)

(6.152)

Let c = 1 and r = 0.1. Numerical computation of the optimal compensator yields the type B solution K0 (s) = 1.377 1010 s2 + 14.055 1010 s + 2.820 1010 s3 + 0.1142 1010 s2 + 1.3207 1010 s + 1.9195 1010 (6.153)

Neglecting the term s3 in the denominator results in K0 (s) = 1.377s2 + 14.055s + 2.820 0.1142s2 + 1.3207s + 1.91195 (s + 10)(s + 0.2047) = 12.056 . (s + 9.8552)(s + 1.7049) K0 (s) (s + 10)(s + 0.2047) 10 = 12.056 sW2 (s) (s + 9.8552)(s + 1.7049) s(s + 10) 120.56(s + 0.2047) . s(s + 9.8552)(s + 1.7049) (6.154) (6.155)

Hence, the optimal compensator for the original problem is K(s) = = (6.156) (6.157)

The pole at 10 cancels, as expected, and the compensator has integrating action, as planned. The compensator has a high-frequency roll-off of 2 decades/decade. Exercise 6.7.1 (Completion of the design). Complete the design of this example. (a) For the compensator (6.157), compute and plot the sensitivity function S, the input sensitivity function U, and the complementary sensitivity function T. (b) Check whether these functions behave satisfactorily, in particular, if S and T do not peak too much. See if these functions may be improved by changing the values of c and r by trial and error. (c) The roll-off of 2 decades/decade is more than called for by the design specications. The desired roll-off of 1 decade/decade may be obtained with the constant weighting function W2 (s) = c. Recompute the optimal compensator for this weighting function with c = 1. (d) Repeat (a) and (b) for the design of (c).

288

Chapter 6. H -Optimization and -Synthesis

(e) Compute and plot the time responses of the output z and the plant input u to a unit step disturbance v in the conguration of Fig. 6.20 for the designs selected in (b) and (d). Discuss the results.

v + z

Figure 6.20: One-degree-of- freedom system with disturbance v Exercise 6.7.2 (Roll-off in the double integrator example). In Example 6.6.1 (p. 276) the numerical solution of the double integrator example of 6.2.5 (p. 254) is presented for r = 0. Extend this to the case r = 0, and verify (6.32).

6.8

-Synthesis

6.8.1 Introduction
In 5.8 (p. 240) the use of the structured singular value for the analysis of stability and performance robustness is explained. We briey review the conclusions. In Fig. 6.21(a), the interconnection block H represents a control system, with structured perturbations represented by the block . The signal w is the external input to the control system and z the control error. The input and output signals of the perturbation block are denoted p and q, respectively. The perturbations are assumed to have been scaled such that all possible structured perturbations are bounded by 1. Let H denote the transfer function from w to z (under perturbation ). Again, the transfer function is assumed to have been scaled in such a way that the control system performance is deemed satisfactory if H 1. The central result established in 5.8 is that the control system 1. is robustly stable, that is, is stable under all structured perturbations and 2. has robust performance, that is, H 1, if and only if H < 1. (6.158)

such that

1,

1 under all structured perturbations

such that

The quantity H , dened in (5.172), is the structured singular value of H with respect to perturbations as in Fig. 6.21(b), with structured and 0 unstructured. Suppose that the control system performance and stability robustness may be modied by feedback through a compensator with transfer matrix K, as in Fig. 6.22(a). The compensator feeds the measured output y back to the control input u. Denote by H K the structured singular

6.8. -Synthesis

289

value of the closed-loop transfer matrix H K of the system of Fig. 6.22(b) with respect to structured perturbations and unstructured perturbations 0 . Then -synthesis is any procedure to construct a compensator K (if any exists) that stabilizes the nominal feedback system and makes H K < 1. (6.159)

(a)

(b)

q w
H

p z w q
H

p z

Figure 6.21: -Analysis for robust stability and performance.

(a)

(b)

q w u K G y p z w u q

p G y K z

Figure 6.22: -Synthesis for robust stability and performance.

6.8.2 Approximate -synthesis for SISO robust performance


In 5.8 the SISO single-degree-of-freedom conguration of Fig. 6.23 is considered. We analyze its robustness with respect to proportional plant perturbations of the form P P(1 + p W2 ) with P

(6.160)

1. The system is deemed to have satisfactory performance robustness if

W1 S

(6.161)

for all perturbations, with S the sensitivity function of the closed-loop system. W1 and W2 are suitable frequency dependent functions.

290

Chapter 6. H -Optimization and -Synthesis


w K u P + + z

Figure 6.23: Single-degree- of-freedom conguration By determining the structured singular value it follows in 5.8 that the closed-loop system is robustly stable and has robust performance if and only if H K = sup (|W1 ( j)S( j)| + |W2 ( j)T ( j)|) < 1.

(6.162)

T is the complementary sensitivity function of the closed-loop system. Hence, the robust performance and stability problem consists of determining a feedback compensator K (if any exists) that stabilizes the nominal system and satises H K < 1. One way of doing this is to minimize H K with respect to all stabilizing compensators K. Unfortunately, this is not a standard H -optimal control problem. The problem cannot be solved by the techniques available for the standard H problem. In fact, no solution is known to this problem. By the well-known inequality (a + b)2 2(a2 + b2 ) for any two real numbers a and b it follows that (|W1 S| + |W2 T|)2 2(|W1 S|2 + |W2 T|2 ). Hence, sup (|W1 ( j)S( j)|2 + |W2 ( j)T ( j)|2 ) <

1 2

(6.163)

(6.164)

implies H K < 1. Therefore, if we can nd a compensator K such that

sup (|W1 ( j)S( j)|2 + |W2 ( j)T ( j)|2 )1/2 =

W1 S W2 T

<

1 2

(6.165)

then this compensator achieves robust performance and stability. Such a compensator, if any exists, may be obtained by minimizing W1 S W2 T , (6.166)

which is nothing but a mixed sensitivity problem. Thus, the robust design problem has been reduced to a mixed sensitivity problem. This reduction is not necessarily successful in solving the robust design problem see Exercise 6.9.1 (p. 294). Exercise 6.8.1 (Proof of inequality). Prove the inequality (a + b)2 2(a2 + b2 ) for any two real numbers a and b, for instance by application of the Cauchy-Schwarz inequality to the vectors (a, b) and (1, 1). Exercise 6.8.2 (Minimization of upper bound). Show that the upper bound on the right-hand side of (6.163) may also be obtained from Summary 5.7.7(b) (p. 236).

6.8. -Synthesis

291

6.8.3 Approximate solution of the -synthesis problem


More complicated robust design problems than the simple SISO problem we discussed cannot be reduced to a tractable problem so easily. Various approximate solution methods to the synthesis problem have been suggested. The best known of these (Doyle 1984) relies on the property (see Summary 5.7.723) (M) [DM D1 ], with D and D suitable diagonal matrices. The problem of minimizing H K = sup (H K ( j))

(6.167)

(6.168)

is now replaced with the problem of minimizing the bound sup [D( j)H K ( j) D1 ( j)] = DH K D1
,

(6.169)

where for each frequency the diagonal matrices D( j) and D( j) are chosen so that the bound is the tightest possible. Minimizing (6.169) with respect to K is a standard H problem, provided D and D are rational stable matrix functions. Doyles method is based on D-K iteration: Summary 6.8.3 (D-K iteration). 1. Choose an initial compensator K that stabilizes the closed-loop system, and compute the corresponding nominal closed-loop transfer matrix H K .
0 One way of nding an initial compensator is to minimize H K with respect to all sta0 bilizing K, where H K is the closed-loop transfer matrix of the conguration of Fig. 6.22 from w to z with = 0 = 0.

2. Evaluate the upper bound


D( j), D( j)

min

[D( j)H K ( j) D1 ( j)],

(6.170)

with D and D diagonal matrices as in Summary 5.7.7(c), for a number of values of on a suitable frequency grid. The maximum of this upper bound over the frequency grid is an estimate of H K . If H K is small enough then stop. Otherwise, continue. 3. On the frequency grid, t stable minimum-phase rational functions to the diagonal entries of D and D. Because of the scaling property it is sufcient to t their magnitudes only. The resulting extra freedom (in the phase) is used to improve the t. Replace the original matrix functions D and D with their rational approximations. 4. Given the rational approximations D and D, minimize DH K D1 with respect to all stabilizing compensators K. Denote the minimizing compensator as K and the corresponding closed-loop transfer matrix as H K . Return to 2.

292

Chapter 6. H -Optimization and -Synthesis

Any algorithm that solves the standard H -problem exactly or approximately may be used in step (d). The procedure is continued until a satisfactory solution is obtained. Convergence is not guaranteed. The method may be implemented with routines provided in the -Tools toolbox (Balas, Doyle, Glover, Packard, and Smith 1991). A lengthy example is discussed in 6.9. The method is essentially restricted to complex perturbations, that is, perturbations corresponding to dynamic uncertainties. Real perturbations, caused by uncertain parameters, need to be overbounded by dynamic uncertainties.

6.9 An application of -synthesis


6.9.1 Introduction
To illustrate the application of -synthesis we consider the by now familiar SISO plant of Example 5.2.1 (p. 198) with transfer function P(s) = g . s2 (1 + s) (6.171)

Nominally, g = g0 = 1 and = 0, so that the nominal plant transfer function is P0 (s) = g0 . s2 (6.172)

In Example 5.2.2 (p. 199) root locus techniques are used to design a modied PD compensator with transfer function C(s) = k + sTd , 1 + sT0 k = 1, Td = 2 = 1.414, T0 = 0.1. (6.173)

The corresponding closed-loop poles are 0.7652 j 0.7715 and 8.4679. In this section we explore how DK iteration may be used to obtain an improved design that achieves robust performance.

6.9.2 Performance specication


In Example 5.8.4 (p. 245) the robustness of the feedback system is studied with respect to the performance specication |S( j)| 1 , |V ( j)| . (6.174)

The function V is chosen as 1 = (1 + )S0 . V S0 is the sensitivity function of the nominal closed-loop system and is given by S0 (s) = s2 (1 + sT0 ) , 0 (s) (6.176) (6.175)

with 0 (s) = T0 s3 + s2 + g0 Td s + g0 k the nominal closed-loop characteristic polynomial. The number is a tolerance.

6.9. An application of -synthesis

293

In Example 5.8.4 it is found that robust performance is guaranteed with a tolerance = 0.25 for parameter perturbations that satisfy 0.9 g 1.1, 0 < 0.1. (6.177)

The robustness test used in this example is based on a proportional loop perturbation model. In the present example we wish to redesign the system such that performance robustness is achieved for a modied range of parameter perturbations, namely for 0.5 g 1.5, 0 < 0.05. (6.178)

For the purposes of this example we moreover change the performance specication model. Instead of relating performance to the nominal performance of a more or less arbitrary design we choose the weighting function V such that 1 s2 = (1 + ) 2 . V (s) s + 20 s + 2 0 (6.179)

Again, is a tolerance. The constant 0 species the minimum bandwidth and determines the maximally allowable peaking. Numerically we choose = 0.25, 0 = 1, and = 1 . 2 The design (6.173) does not meet the specication (6.174) with V given by (6.179) for the range of perturbations (6.178). Figure 6.24 shows plots of the bound 1/|V| together with plots of the perturbed sensitivity functions S for the four extreme combinations of parameter values. In particular, the bound is violated when the value of the gain g drops to too small values.
20

|S | [dB]

bound
0 -20 -40 -60 -1 10

10

10

10

angular frequency [rad/s]

Figure 6.24: Magnitude plot of S for different parameter value combinations and the bound 1/ V

6.9.3 Setting up the problem


To set up the -synthesis problem we rst dene the perturbation model. As seen in Example 5.8.4 the loop gain has proportional perturbations
P (s) =

P(s) P0 (s) = P0 (s)

gg0 g0

1 + s
P ( j)|

(6.180)

We bound the perturbations as | W0 (s) = + s0 . 1 + s0

|W0 ( j)| for all , with (6.181)

294

Chapter 6. H -Optimization and -Synthesis

The number , with < 1, is a bound |g g0 |/g0 for the relative perturbations in g, and 0 is the largest possible value for the parasitic time constant . Numerically we let = 0.5 and 0 = 0.05. Figure 6.25 shows plots of the perturbation | P | for various combinations of values of the uncertain parameters together with a magnitude plot of W0 . Figure 6.26 shows
20

| P | [dB]

bound
0

-20 -40 -60 -1 10

10

10

10

angular frequency [rad/s]

Figure 6.25: Magnitude plots of the perturbations P for different parameter values and of the bound W0 the corresponding scaled perturbation model. It also includes the scaling function V used to normalize performance.

w Wo y u + p P + q uo V + Po + z

Figure 6.26: Perturbation and performance model Exercise 6.9.1 (Reduction to mixed sensitivity problem). In 6.8.2 (p. 289) it is argued that robust performance such that SV < 1 under proportional plant perturbations bounded by P W0 is achieved iff sup (|S( j)V ( j)| + |T ( j)W0 ( j)|) < 1.

(6.182)

A sufcient condition for this is that sup (|S( j)V ( j)|2 + |T ( j)W0 ( j)|2 ) <

1 2

2.

(6.183)

To see whether the latter condition may be satised we consider the mixed sensitivity problem consisting of the minimization of H , with H= SV . T W0 (6.184)

6.9. An application of -synthesis

295

(a) Show that the mixed sensitivity problem at hand may be solved as a standard H problem with generalized plant V 0 G= V P0 W0 V 1 P0 . P0

(6.185)

(b) Find a (minimal) state representation of G. (Hint: Compare (6.190) and (6.191).) Check that condition (b) of 6.5.5 (p. 274) needed for the state space solution of the H problem is not satised. To remedy this, modify the problem by adding a third component z1 = u to the error signal z, with small. (c) Solve the H -optimization problem for a small value of , say, = 106 or even smaller. Use the numerical values introduced earlier: g0 = 1, = 0.25, 0 = 1, = 0.5, = 0.5, and 0 = 0.05. Check that H cannot be made less than about 0.8, and, hence, the condition (6.183) cannot be met. (d) Verify that the solution of the mixed sensitivity problem does not satisfy (6.182). Also check that the solution does not achieve robust performance for the real parameter variations (6.178). The fact that a robust design cannot be obtained this way does not mean that a robust design does not exist. The next step in preparing for DK iteration is to compute the interconnection matrix G as in Fig. 6.22 from the interconnection equations q p z = G w . u y Inspection of the block diagram of Fig. 6.26 shows that p z y = = = W0 u, V w + P0 (q + u), V w P0 (q + u), q W0 P0 w . u P0 (6.187)

(6.186)

so that p 0 z = P0 P0 y

0 W1 W1 G

(6.188)

To apply the state space algorithm for the solution of the H problem we need a state space representation of G. For the output z we have from the block diagram of Fig. 6.26 z = V w + P0 u0 = 1 s2 + 20 s + 2 g0 0 w + 2 u0 . 1+ s2 s (6.189)

296

Chapter 6. H -Optimization and -Synthesis

This transfer function representation may be realized by the two-dimensional state space system x1 0 1 = 0 0 x2 z= 1 0 x1 + x2
20 1+ 2 0 1+

0 g0

w , u0

(6.190)

1 x1 w. + x2 1+

The weighting lter W0 may be realized as x3 = 1 1 x3 + u, 0 0 u + x3 .

p =

(6.191)

Exercise 6.9.2 (State space representation). Derive or prove the state space representations (6.190) and (6.191). Using the interconnection equation u0 = q + u we obtain from (6.1906.191) the overall state differential and output equations 0 0 x + g0 0 10 0 0 0 0 p 0 0 1 1 z = 1 0 0 x + 0 1+ y 1 0 0 0 1 0 0 x= 0 1 0 0
20 1+ 2 0 1+

q g0 w , u 1 0 1 q 0 w . u 0
0

(6.192)

1+

To initialize the DK iteration we select the compensator that was obtained in Example 5.2.2 (p. 199) with transfer function K0 (s) = k + sTd . 1 + sT0 (6.193)

This compensator has the state space representation x u = = 1 1 x + y, T0 T0 Td Td (k )x + y. T0 T0

(6.194)

6.9.4

DK iteration

The DK iteration procedure is initialized by dening a frequency axis, calculating the frequency response of the initial closed-loop system H0 on this axis, and computing and plotting the lower and upper bounds of the structured singular value of H0 on this frequency axis14 . The calculation of the upper bound also yields the D-scales.
14 The calculations were done with the M ATLAB -tools toolbox. The manual (Balas, Doyle, Glover, Packard, and Smith 1991) offers a step-by-step explanation of the iterative process.

6.9. An application of -synthesis

297

1.4

1.2

0.8 -2 10

10

-1

10

10

10

10

angular frequency [rad/s]

Figure 6.27: The structured singular value of the initial design The plot of Fig. 6.27 shows that the structured singular value peaks to a value of about 1.3. Note that the computed lower and upper bounds coincide to the extent that they are indistinguishable in the plot. Since the peak value exceeds 1, the closed-loop system does not have robust performance, as we already know. The next step is to t rational functions to the D-scales. A quite good-looking t may be obtained on the rst diagonal entry of D with a transfer function of order 2. The second diagonal entry is normalized to be equal to 1, and, hence, does not need to be tted. Figure 6.28 shows the calculated and tted scales for the rst diagonal entry. Because we have 1 1 perturbation
10
2

|D |
10
0

10

-2

fit

10 -2 10

-4

D-scale
10
-1

10

10

10

10

angular frequency [rad/s]

Figure 6.28: Calculated and tted D-scales blocks only, the left and right scales are equal. The following step in the -synthesis procedure is to perform an H optimization on the scaled system G1 = DGD1 . The result is a sub-optimal solution corresponding to the bound 1 = 1.01. In the search procedure for the H -optimal solution as described in 6.4.6 (p. 269) the search is terminated when the optimal value of the search parameter has been reached within a prespecied tolerance. This tolerance should be chosen with care. Taking it too small not only prolongs the computation but more importantly results in a solution that is close to but not equal to the actual H optimal solution. As seen in 6.6 (p. 221), solutions that are close to optimal often have large coefcients. These large coefcients make the subsequent calculations, in particular that of the frequency response of the optimal closed-loop system, ill-conditioned, and easily lead to erroneous results. This difculty is a weak point of algorithms for the solution of of the H problem that do not provide the actual optimal solution. Figure 6.29 shows the structured singular value of the closed-loop system that corresponds to the compensator that has been obtained after the rst iteration. The peak value of the structured

298

Chapter 6. H -Optimization and -Synthesis

1.4

1.2

initial solution

first iteration
1

0.8 -2 10

second iteration
10
-1

10

10

10

10

angular frequency [rad/s]

Figure 6.29: Structured singular values for the three successive designs singular value is about 0.97. This means that the design achieves robust stability. To increase the robustness margin another DK iteration may be performed. Again the Dscale may be tted with a second-order rational function. The H optimization results in a suboptimal solution with level 2 = 0.91. Figure 6.29 shows the structured singular value plots of the three designs we now have. For the third design the structured value has a peak value of about 0.9, well below the critical value 1. The plot for the structured singular value for the nal design is quite at. This appears to be typical for minimum- designs (Lin, Postlethwaite, and Gu 1993).

6.9.5 Assessment of the solution


The compensator K2 achieves a peak structured singular value of 0.9. It therefore has robust performance with a good margin. The margin may be improved by further DK iterations. We pause to analyze the solution that has been obtained. Figure 6.30 shows plots of the sensitivity function of the closed-loop system with the compensator K2 for the four extreme combinations of the values of the uncertain parameters g and . The plots conrm that the system has robust performance.
50

|S | [dB]

bound
0

-50

-100 -2 10

10

-1

10

10

10

10

angular frequency [rad/s]

Figure 6.30: Perturbed sensitivity functions for K2 and the bound 1/ V Figure 6.31 gives plots of the nominal system functions S and U. The input sensitivity function U increases to quite a large value, and has no high-frequency roll-off, at least not in the frequency region shown. The plot shows that robust performance is obtained at the cost of a large controller gain-bandwidth product. The reason for this is that the design procedure has no explicit or implicit provision that restrains the bandwidth.

6.9. An application of -synthesis

299

magnitude [dB]

100 50 0 -50 -100 -2 10

U S

10

-1

10

10

10

10

angular frequency [rad/s]

Figure 6.31: Nominal sensitivity S and input sensitivity U for the compensator K2
magnitude [dB]
100

50

1/ W2

UV
-50 -2 10 10
-1

10

10

10

10

angular frequency [rad/s]

Figure 6.32: Magnitude plots of UV and the bound 1/ W2

6.9.6 Bandwidth limitation


We revise the problem formulation to limit the bandwidth explicitly. One way of doing this is to impose bounds on the high-frequency behavior of the input sensitivity function U. This, in turn, may be done by bounding the high-frequency behavior of the weighted input sensitivity UV . Figure 6.32 shows a magnitude plot of UV for the compensator K2 . We wish to modify the design so that UV is bounded as |U( j)V ( j)| where 1 = . W2 (s) s + 1 (6.196) 1 , |W2 ( j)| , (6.195)

Numerically, we tentatively choose = 100 and 1 = 1. Figure 6.32 includes the corresponding bound 1/|W2 | on |U|. To include this extra performance specication we modify the block diagram of Fig. 6.26 to that of Fig. 6.33. The diagram includes an additional output z2 while z has been renamed to z1 . The closed-loop transfer function from w to z2 is W2 UV. To handle the performance specications SV 1 and W2 UV 1 jointly we impose the performance specication W1 SV W2 UV 1, (6.197)

where W1 = 1. This is the familiar performance specication of the mixed sensitivity problem.

300

Chapter 6. H -Optimization and -Synthesis


z2 p P + + w

W2 y

W0

q uo

V + P o + z1

Figure 6.33: Modied block diagram for extra performance specication

By inspection of Fig. 6.33 we see that the structured perturbation model used to design for robust performance now is p 0 z1 , 2 z2

q = P w 0

0 1

(6.198)

with P , 1 and 2 independent scalar complex perturbations such that

1.

Because W2 is a nonproper rational function we modify the diagram to the equivalent diagram of Fig. 6.34 (compare 6.7, p. 279). The compensator K and the lter W2 are combined to a

w z2 p P + x4 + q uo

Wo

V + Po + z1

y K ~ K W2 ~ u

W21

Figure 6.34: Revised modied block diagram

new compensator K = K W2 , and any nonproper transfer functions now have been eliminated. It is easy to check that the open-loop interconnection system according to Fig. 6.34 may be

6.9. An application of -synthesis

301

represented as 0 0 x= 0 0 p 0 z1 1 = z2 0 y 1 1 0 0 0 0 0 0 0 0 0 1 0 0 0 q g0 g0 0 w , x+ 1 0 0 u 0 0 1 0 0 0 0 0 1 1 0 0 q 0 1+ w . x+ 0 0 0 1 u 0 1 0 1+ 0 0 0
20 1+ 2 0 1+

0 10

(6.199)

To let the structured perturbation have square diagonal blocks we rename the external input w to w1 and expand the interconnection system with a void additional external input w2 : 0 20 0 0 0 1 0 0 1+ q 2 0 g0 0 0 0 g0 0 0 w 1 1+ , x+ x= 1 1 0 0 0 w2 0 0 0 0 0 u 0 0 0 1 0 0 0 (6.200) 0 0 0 0 0 0 1 1 p q 1 0 z1 1 0 0 0 0 0 w 1 1+ . = z2 0 0 0 0 x + 0 0 0 1 w 2 y 1 0 0 0 u 1 0 1+ 0 0 The perturbation model now is q = P 0 w 0
0

p . z

(6.201)

with p a scalar block and 0 a full 2 2 block. To obtain an initial design we do an H -optimal design on the system with q and p removed, that is, on the nominal, unperturbed system. This amounts to the solution of a mixed sensitivity problem. Nine iterations are needed to nd a suboptimal solution according to the level 0 = 0.97. The latter number is less than 1, which means that nominal performance is achieved. The peak value of the structured singular value turns out to be 1.78, however, so that we do not have robust performance. One DK iteration leads to a design with a reduced peak value of the structured singular value of 1.32 (see Fig. 6.35). Further reduction does not seem possible. This means that the bandwidth constraint is too severe. To relax the bandwidth constraint we change the value of in the bound 1/ W2 as given by (6.196) from 100 to 1000. Starting with H -optimal initial design one D-K iteration leads to a design with a peak structured singular value of 1.1. Again, robust performance is not feasible. Finally, after choosing = 10000 the same procedure results after one D-K iteration in a compensator with a peak structured singular value of 1.02. This compensator very nearly has robust performance with the required bandwidth. Figure 6.35 shows the structured singular value plots for the three compensators that are successively obtained for = 100, = 1000 and = 10000.

302

Chapter 6. H -Optimization and -Synthesis

1.4

1.2 1 0.8 0.6 -2 10

= 100

= 1000
= 10000

10

-1

10

10

10

10

angular frequency [rad/s]

Figure 6.35: Structured singular value plots for three successive designs

6.9.7 Order reduction of the nal design


The limited bandwidth robust compensator has order 9. This number may be explained as follows. The generalized plant (6.200) has order 4. In the (single) DK iteration that yields the compensator the D-scale is tted by a rational function of order 2. Premultiplying the plant by D and postmultiplying by its inverse increases the plant order to 4 + 2 + 2 = 8. Central subop timal compensators K for this plant hence also have order 8. This means that the corresponding W2 for the conguration of Fig. 6.33 have order 9. compensators K = K 1 It is typical for the D-K algorithm that compensators of high order are obtained. This is caused by the process of tting rational scales. Often the order of the compensator may considerably be decreased without affecting performance or robustness. Figure 6.36(a) shows the Bode magnitude and phase plots of the limited bandwidth robust
10
4

magnitude

10 10 10 10

(b),(c)

-2

(a)
-4

10 100

-2

10

-1

10

10

10

10

10

10

phase [deg]

(b),(c)
-100

(a)
-200 -2 10 10
-1

10

10

10

10

10

10

angular frequency [rad/s]

Figure 6.36: Exact (a) and approximate (b, c) frequency responses of the compensator compensator. The compensator is minimum-phase and has pole excess 2. Its Bode plot has break points near the frequencies 1, 60 and 1000 rad/s. We construct a simplied compensator in two steps:

6.9. An application of -synthesis

303

1. The break point at 1000 rad/s is caused by a large pole. We remove this pole by omitting the leading term s9 from the denominator of the compensator. This large pole corresponds to a pole at of the H -optimal compensator. Figure 6.36(b) is the Bode plot of the resulting simplied compensator. 2. The Bode plot that is now found shows that the compensator has pole excess 1. It has break points at 1 rad/s (corresponding to a single zero) and at 60 rad/s (corresponding to a double pole). We therefore attempt to approximate the compensator transfer function by a rational function with a single zero and two poles. A suitable numerical routine from M ATLAB yields the approximation K(s) = 6420 s + 0.6234 . (s + 22.43)2 + 45.312 (6.202)

Figure 6.36(c) shows that the approximation is quite good. Figure 6.37 gives the magnitudes of the perturbed sensitivity functions corresponding to the four extreme combinations of parameter values that are obtained for this reduced-order compensator. The plots conrm that performance is very nearly robust. Figure 6.38 displays the nominal performance. Comparison of the plot of the input sensitivity U with that of Fig. 6.31 conrms that the bandwidth has been reduced.
magnitude [dB]
50

bound
0

-50

-100 -2 10

10

-1

10

10

10

10

angular frequency [rad/s]

Figure 6.37: Perturbed sensitivity functions of the reduced-order limitedbandwidth design

magnitude [dB]

50

U
0

S
-50

-100 -2 10

10

-1

10

10

10

10

angular frequency [rad/s]

Figure 6.38: Nominal sensitivity S and input sensitivity U of the reducedorder limited- bandwidth design Exercise 6.9.3 (Peak in the structured singular value plot). Figure 6.35 shows that the singular value plot for the limited-bandwidth design has a small peak of magnitude slightly greater

304

Chapter 6. H -Optimization and -Synthesis

than 1 near the frequency 1 rad/s. Inspection of Fig. 6.37, however, reveals no violation of the performance specication on S near this frequency. What is the explanation for this peak?

6.9.8 Conclusion
The example illustrates that -synthesis makes it possible to design for robust performance while explicitly and quantitatively accounting for plant uncertainty. The method cannot be used navely, though, and considerable insight is needed. A severe shortcoming is that the design method inherently leads to compensators of high order, often unnecessarily high. At this time only ad hoc methods are available to decrease the order. The usual approach is to apply an order reduction algorithm to the compensator. Exercise 6.9.4 (Comparison with other designs). The compensator (6.202) is not very different from the compensators found in Example 5.2.2 (p. 199) by the root locus method, by H2 optimization in 3.6.2 (p. 133). and by solution of a mixed sensitivity problem in 6.2.5 (p. 254). 1. Compare the four designs by computing their closed-loop poles and plotting the sensitivity and complementary sensitivity functions in one graph. 2. Compute the stability regions for the four designs and plot them in one graph as in Fig. 6.5. 3. Discuss the comparative robustness and performance of the four designs.

6.10 Appendix: Proofs


6.10.1 Introduction
In this Appendix to Chapter 6 we present several proofs for 6.4 (p. 264) and 6.5 (p. 272).

6.10.2 Proofs for 6.4


We rst prove the result of Summary 6.4.1 (p. 264). Inequality for the -norm. By the denition of the -norm the inequality H is equivalent to the condition that ( H ( j)H( j)) for , with the largest singular value. This in turn is the same as the condition that i ( H H) 2 on the imaginary axis for all i, with i the ith largest eigenvalue. This nally amounts to the condition that H H 2 I on the imaginary axis, which is part (a) of Summary 6.4.1. The proof of (b) is similar. Preliminary to the proof of the basic inequality for H optimal control of Summary 6.4.2 (p. 250) we establish the following lemma. The lemma presents, in polished form, a result originally obtained by Meinsma (Kwakernaak 1991). Summary 6.10.1 (Meinsmas lemma). Let that admits a factorization = Z J Z with J= In 0 0 Im be a nonsingular para-Hermitian rational matrix function

(6.203)

and Z square rational. Suppose that V is an (n + m) n rational matrix of full normal column rank15 and W an m (n + m) rational matrix of full normal row rank such that W V = 0. Then
15 A rational matrix has full normal column rank if it has full column rank everywhere in the complex plane except at a nite number of points. A similar denition holds for full normal row rank.

6.10. Appendix: Proofs

305

(a) V V 0 on the imaginary axis if and only if W (b) V

1 1

W 0 on the imaginary axis. W < 0 on the imaginary axis.

V > 0 on the imaginary axis if and only if W

Proof of Meinsmas lemma. (a) V V 0 on the imaginary axis is equivalent to V = Z 1 A , B (6.204)

with A an n n nonsingular rational matrix and B an n m rational matrix such that A A B B on the imaginary axis. Dene an m m nonsingular rational matrix A and an m n rational matrix B such that A = 0. [ B A] B (6.205)

It follows that B A = A B, and, hence, A1 B = BA1. From A A B B on the imaginary axis it follows 1 1 B 1, and, hence, A A B B on the imaginary axis. Since that BA 1, so that also A W V = W Z 1 A = 0, B (6.206)

it follows from (6.205) that we have W Z 1 = U[ B A], with U some square nonsingular rational matrix. It follows that W = U[ B A]Z, and, hence, W
1

(6.207)

W = U( B B A A )U .
1

(6.208)

This proves that W

W 0 on the imaginary axis. The proof of (b) is similar.

We next consider the proof of the basic inequality of Summary 6.4.2 (p. 265) for the standard H optimal control problem. Proof of Basic inequality. By Summary 6.4.1(b) (p. 264) the inequality H 2 I on the imaginary axis. Substitution of H = G11 + G12 K(I G22 K )1 G21 , L leads to
G11 G11 + G11 G21 L + LG21 G11 + LG21 G21 L 2 I

is equivalent to H H (6.209)

(6.210)

on the imaginary axis. This, in turn, is equivalent to [I L]


G11 G11 2 I G21 G11 G11 G21 G21 G21

I 0 L

(6.211)

on the imaginary axis. Consider


1 1

= = I 0

11 21 12

12 22 1 22 11

(6.212)
12 1 22 21

0
22

I
1 22 21

0 . I

(6.213)

306

Chapter 6. H -Optimization and -Synthesis

For the suboptimal solution to exist we need that


12 1 22 21

11

= G11 G21 (G21 G21 )1 G21 G11 G11 G11 + 2 I 0

(6.214)

on the imaginary axis. This is certainly the case for sufciently large. Also, obviously 22 = G21 G21 0 on the imaginary axis. Then by a well-known theorem (see for instance Wilson (1978)) there exist square rational matrices V1 and V2 such that 12 1 22 21

11

= V1 V1 ,

22

= V2 V2 .

(6.215)

Hence, we may write = I 0


12 1 22

I Z

V1 0

0 V2

I 0

0 I

V1 0

0 V2

I
1 22 21

0 , I

(6.216)

Z is nonsingular. Taking (6.217) with ) yields (6.218)

which shows that V= I , L

admits a factorization. It may be proved that the matrix W= L , I

application of Summary 6.10.1(a) (p. 304) to (6.211) (replacing [ L I]


1

I 0 L

on the imaginary axis. Since G12 K(I G22 K )1 L 0 = = I I I we have K I 0 G12 I G22
G11 G11 2 I G21 G11 G11 G21 G21 G21 1

G12 G22

I (I G22 K )1 K

(6.219)

0 I

G12 G22

I 0. K

(6.220)

This is what we set out to prove.

6.10.3 Proofs for 6.5


Our derivation of the factorization of of 6.5.3 (p. 272) for the state space version of the standard problem is based on the Kalman-Jakubovi -Popov equality of Summary 3.7.1 (p. 143) in 3.7.2. This c equality establishes the connection between factorizations and algebraic Riccati equations. Proof (Factorization of = 0
G12 ).

We are looking for a factorization of


G11 G11 2 I G21 G11 G11 G21 G21 G21 1

I
G22

0 I

G12 G22

(6.221)

First factorization. As a rst step, we determine a factorization


G11 G11 2 I G21 G11 G11 G21 G21 G21 = Z1 J1 Z1 ,

(6.222)

1 such that Z1 and Z1 have all their poles in the closed left-half complex plane. Such a factorization is sometimes called a spectral co factorization. A spectral cofactor of a rational matrix may be obtained by

6.10. Appendix: Proofs

307

rst nding a spectral factorization of at hand, it is easily found that


G11 (s)G11 (s) 2 I G21 (s)G11 (s)

and then transposing the resulting spectral factor. For the problem

G11 (s)G21 (s) G21 (s)G21 (s)

H(sI A)1 F F T (sI AT )1 H T 2 I 0 C(sI A)1 F F T (sI AT )1 H T

0 2 I 0

H(sI A)1 F F T (sI AT )1 C T 0

C(sI A)1 F F T (sI AT )1 C T + I (6.223)

The transpose of the right-hand side may be written as R1 + H1 (s)W1 H1 (s), where 2 0 0 I 2 I 0 , H1 (s) = F T (sI AT )1 [H T 0 C T ], W1 = I. R1 = 0 0 0 I

(6.224)

Appending the distinguishing subscript 1 to all matrices we apply the Kalman-Jakubovi -Popov equality of c Summary 3.7.1 with A1 = AT , B1 = [H T 0 C T ], C1 = F T , D1 = 0, and R1 and W1 as displayed above. This leads to the algebraic Riccati equation 0 = A X1 + X1 AT + F F T X1 (C T C 1 T H H)X1 . 2 (6.225)

The corresponding gain F1 and the factor V1 are 1 1 2 H 2 H F1 = 0 X1 , V1 (s) = I + 0 X1 (sI AT )1 [H T 0 C T ]. C C

(6.226)

We need a symmetric solution X1 of the Riccati equation (6.225) such that AT + ( 12 H T H C T C)X1, or, equivalently, 1 A1 = A + X1 ( 2 H T H C T C) is a stability matrix. By transposing the factor V1 we obtain the cofactorization (6.222), with H T Z1 (s) = V1 (s) = I + 0 (sI A)1 X1 12 F 0 C , C 2 0 0 I 2 I 0 . J1 = R1 = 0 0 0 I (6.227)

(6.228)

(6.229)

Second factorization. The next step in completing the factorization of is to show, by invoking some algebra, that H(sI A1 )1 C T H(sI A1 )1 B 0 G12 1 = 0 I Z1 I G22 1 )1 X1 C T C(sI A1 )1 B I C(sI A = 0 0 I 0 H I + 0 (sI A1 )1 X1 C T 0 C B .

(6.230)

308

Chapter 6. H -Optimization and -Synthesis

To obtain the desired factorization of guishing subscript 2)

we again resort to the KJP equality, with (appending the distin(6.231) 0 12 I 0 0 0 . I (6.232)

B2 = [X1 C T B], R2 = 0, A2 = A1 , 1 2 I H 0 0 1 C2 = 0 , D2 = 0 I , W2 = J1 = 0 C I 0 0 This leads to the algebraic Riccati equation (in X2 ) 1 1 0 = AT X2 + X2 A1 2 H T H + C T C ( X2 X1 I )C T C( X1 X2 I ) + 2 X2 B BT X2 . Substitution of A1 from (6.227) and rearrangement yields 0 = AT X2 + X2 A ( X2 X1 I ) + X2 X1 ( 1 T H H( X1 X2 I ) 2

(6.233)

1 T H H C T C)X1 X2 + 2 X2 B BT X2 . 2

(6.234)

Substitution of X1 ( 12 H T H C T C)X1 from the Riccati equation (6.225) and further rearrangement result in 0 = ( X2 X1 I ) AT X2 + X2 A( X1 X2 I ) 1 T H H( X1 X2 I ) X2 ( 2 B BT F F T ) X2 . (6.235) 2 Postmultiplication of this expression by ( X1 X2 I )1 , assuming that the inverse exists, and premultiplication by the transpose yields the algebraic Riccati equation + ( X2 X1 I ) 0 = AT X2 + X2 A + H T H X2 ( B BT 1 F F T )X2 , 2 (6.236)

where X2 = 2 X2 ( X1 X2 I )1 . After solving (6.236) for X2 we obtain X2 as X2 = X2 ( X1 X2 2 I )1 . The gain F2 and factor V2 for the spectral factorization of are
T T F2 = L1 ( B2 X2 + D2 W2 C2 ) = 2

C( X1 X2 I ) , 2 BT X2

(6.237)

V2 (s) = I + F2 (sI A2 )1 B2 . (6.238) We need to select a solution X2 of the Riccati equation (6.233) such that A2 B2 F2 is a stability matrix. It is proved in the next subitem that A2 B2 F2 is a stability matrix iff 1 A2 = A BF2 = A + ( 2 F F T B BT )X2 is a stability matrix, where F2 = ( B BT equation (6.236). Since L2 = I 0 0 , 12 I
1 2

(6.239)

F F T )X2 is the gain corresponding to the algebraic Riccati

(6.240) is (6.241)

the desired spectral factor Z of Z (s) = I 0


1

0 V (s). I 2

6.10. Appendix: Proofs

309

Invoking an amount of algebra it may be found that the inverse M of the spectral factor Z may be expressed as M (s) = I 0 C 1 0 (sI A2 )1 (I 2 X1 X2 )1 [X1 C T B]. + T I B X2 (6.242)

Choice of the solution of the algebraic Riccati equation. We require a solution to the algebraic Riccati equation (6.233) such that A2 B2 F2 is a stability matrix. Consider A2 B2 F2 = A1 X1 C T C( X1 X2 I ) + 2 B BT X2 . Substitution of A1 = A +
1 2

(6.243)

X1 H T H X1 C T C and X = X2 ( X1 X2 2 I )1 results in 1 X1 H T H X1 X2 X1 H T H 2 (6.244)

A2 B2 F2 = ( A X1 X2 2 A +

X1 C T C X1 X2 + 2 B BT X2 )( X1 X2 2 I )1 . Substitution of X1 ( 12 H T H C T C)X1 from (6.225) results in A2 B2 F2 = ( 2 A X1 H T H + 2 B BT X2 X1 AT X2 F F T X2 )( X1 X2 2 I )1 . Further substitution of H T H + AT X2 from (6.236) and simplication yields 1 A2 B2 F2 = ( X1 X2 2 I )( A B BT X2 + 2 F F T X2 )( X1 X2 2 I )1 . Inspection shows that A2 B2 F2 has the same eigenvalues as A2 = A B BT X2 + A2 B2 F2 is a stability matrix iff A2 is a stability matrix.
1 2

(6.245)

(6.246) F F T X2 . Hence,

310

Chapter 6. H -Optimization and -Synthesis

A Matrices
This appendix lists several matrix denitions, formulae and results that are used in the lecture notes. In what follows capital letters denote matrices, lower case letters denote column or row vectors or scalars. The element in the ith row and jth column of a matrix A is denoted by Ai j. Whenever sums A + B and products AB etcetera are used then it is assumed that the dimensions of the matrices are compatible.

A.1 Basic matrix results


Eigenvalues and eigenvectors A column vector v n is an eigenvector of a square matrix A nn if v = 0 and Av = v for some . In that case is referred to as an eigenvalue of A. Often i ( A) is used to denote the ith eigenvalue of A (which assumes an ordering of the eigenvalues, an ordering that should be clear from the context in which it is used). The eigenvalues are the zeros of the characteristic polynomial A () = det(In A), ( , )

where In denotes the n n identity matrix or unit matrix. An eigenvalue decomposition of a square matrix A is a decomposition of A of the form A = V DV 1 , where V and D are square and D is diagonal.

In this case the diagonal entries of D are the eigenvalues of A and the columns of V are the corresponding eigenvectors. Not every square matrix A has an eigenvalue decomposition. The eigenvalues of a square A and of T AT 1 are the same for any nonsingular T. In particular A = T AT 1 . Rank, trace, determinant, singular and nonsingular matrices The trace, tr( A) of a square matrix A that
n nn

is dened as tr( A) =

n i=1

Aii . It may be shown

tr( A) =
i=1

i ( A). 311

312

Appendix A. Matrices

The rank of a (possibly nonsquare) matrix A is the maximal number of linearly independent rows (or, equivalently, columns) in A. It also equals the rank of the square matrix AT A which in turn equals the number of nonzero eigenvalues of AT A. The determinant of a square matrix A nn is usually dened (but not calculated) recursively by det( A) =
n j+1 A1 j det( Aminor ) j=1 (1) 1j

if n > 1 . if n = 1

Here Aminor is the (n 1) (n 1)-matrix obtained from A by removing its ith row and jth ij column. The determinant of a matrix equals the product of its eigenvalues, det( A) = n i ( A). i=1 A square matrix is singular if det( A) = 0 and is regular or nonsingular if det( A) = 0. For square matrices A and B of the same dimension we have det( A B) = det( A) det(B). Symmetric, Hermitian and positive denite matrices, the transpose and unitary matrices A matrix A nn is (real) symmetric if AT = A. Here AT is the transpose of A is dened elementwise as ( AT )i j = A ji , (i, j = 1, . . . , n). A matrix A nn is Hermitian if AH = A. Here AH is the complex conjugate transpose of A dened as ( AH )i j = A ji (i, j = 1, . . . , n). Overbars x + jy of a complex number x + jy denote the complex conjugate: x + jy = x jy. Every real-symmetric and Hermitian matrix A has an eigenvalue decomposition A = V DV 1 and they have the special property that the matrix V may be chosen unitary which is that the columns of V have unit length and are mutually orthogonal: V H V = I. A symmetric or Hermitian matrix A is said to be nonnegative denite or positive semi-denite if xH Ax 0 for all column vectors x. We denote this by A 0. A symmetric or Hermitian matrix A is said to be positive denite if xH Ax > 0 for all nonzero column vectors x. We denote this by A > 0. For Hermitian matrices A and B the inequality A B is dened to mean that A B 0. Lemma A.1.1 (Nonnegative denite matrices). Let A 1. All eigenvalues of A are real valued, 2. A 0 i ( A) 0 3. A > 0 i ( A) > 0 ( i = 1, . . . , n), ( i = 1, . . . , n),
nn

be a Hermitian matrix. Then

4. If T is nonsingular then A 0 if and only T H AT 0.

A.2. Three matrix lemmas

313

A.2 Three matrix lemmas


Lemma A.2.1 (Eigenvalues of matrix products). Suppose A and BH are matrices of the same dimension n m. Then for any there holds det(In A B) = nm det(Im BA). Proof. One the one hand we have Im A B In Im 0
1 Im B = A In

(A.1)

0 1 In AB

and on the other hand Im A B In Im A 0 Im BA = In 0 B . In

Taking determinants of both of these equations shows that m det(In 1 Im A B) = det A B = det(Im BA). In

So the nonzero eigenvalues of A B and BA are the same. This gives the two very useful identities: 1. det(In A B) = det(Im B A), 2. tr( A B) =
i i ( A B)

j j (BA)

= tr(BA).

Lemma A.2.2 (Sherman-Morrison-Woodburry & rank-one update). ( A + UV H )1 = A1 A1 U(I + V H A1 V H ) A1 . This formula is used mostly if U = u and V = v are column vectors. Then UV H = uvH has rank one, and it shows that a rank-one update of A corresponds to a rank-one update of its inverse, ( A + uvH )1 = A1 1 1 + vH A1 u ( A1 u)(vH A1 ) .

rank-one

Lemma A.2.3 (Schur complement). Suppose a Hermitian matrix A is partitioned as A= P QH Q R

with P and R square. Then A > 0 P is invertible, P > 0 and R QH P1 Q > 0. The matrix R QH P1 Q is referred to as the Schur complement of P (in A).

314

Appendix A. Matrices

References
Abramowitz, M. and I. A. Stegun (1965). Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables. Dover Publications, Inc., New York. Anderson, B. D. O. and E. I. Jury (1976). A note on the Youla-Bongiorno-Lu condition. Automatica 12, 387388. Anderson, B. D. O. and J. B. Moore (1971). Linear Optimal Control. Prentice-Hall, Englewood Cliffs, NJ. Anderson, B. D. O. and J. B. Moore (1990). Optimal Control: Linear Quadratic Methods. Prentice Hall, Englewood Cliffs, NJ. Aplevich, J. D. (1991). Implicit Linear Systems, Volume 152 of Lecture Notes in Control and Information Sciences. Springer-Verlag, Berlin. ISBN 0-387-53537-3. Bagchi, A. (1993). Optimal control of stochastic systems. International Series in Systems and Control Engineering. Prentice Hall International, Hempstead, UK. Balas, G. J., J. C. Doyle, K. Glover, A. Packard, and R. Smith (1991). Users Guide, Analysis and Synthesis Toolbox. The MathWorks, Natick, Mass. Barmish, B. R. and H. I. Kang (1993). A survey of extreme point results for robustness of control systems. Automatica 29, 1335. Bartlett, A. C., C. V. Hollot, and H. Lin (1988). Root locations of an entire polytope of polynomials: It sufces to check the edges. Math. Control Signals Systems 1, 6171. Basile, G. and G. Marro (1992). Controlled and Conditioned Invariance in Linear System Theory. Prentice Hall, Inc., Englewood Cliffs, NJ. Berman, G. and R. G. Stanton (1963). The asymptotes of the root locus. SIAM Review 5, 209218. Biaas, S. (1985). A necessary and sufcient condition for the stability of convex combinations of polynomials or matrices. Bulletin Polish Acad. of Sciences 33, 473480. Black, H. S. (1934). Stabilized feedback ampliers. Bell System Technical Journal 13, 118. Blondel, V. (1994). Simultaneous Stabilization of Linear Systems, Volume 191 of Lecture Notes in Control and Information Sciences. Springer-Verlag, Berlin etc. Bode, H. W. (1940). Relations between attenuation and phase in feedback amplier design. Bell System Technical Journal 19, 421454. Bode, H. W. (1945). Network Analysis and Feedback Amplier Design. Van Nostrand, New York. Boyd, S. P. and C. H. Barratt (1991). Linear Controller Design: Limits of Performance. Prentice Hall, Englewood Cliffs, NJ. 315

316

References

Bristol, E. H. (1966). On a new measure of interaction for multivariable process control. IEEE Transactions on Automatic Control 11, 133134. Brockett, R. W. and M. D. Mesarovic (1965). The reproducibility of multivariable systems. Journal of Mathematical Analysis and Applications 11, 548563. Campo, P. J. and M. Morari (1994). Achievable closed-loop properties of systems under decentralized control: conditions involving the steady-state gain. IEEE Transactions on Automatic Control 39, 932943. Chen, C.-T. (1970). Introduction to Linear Systems Theory. Holt, Rinehart and Winston, New York. Chiang, R. Y. and M. G. Safonov (1988). Users Guide, Robust-Control Toolbox. The MathWorks, South Natick, Mass. Chiang, R. Y. and M. G. Safonov (1992). Users Guide, Robust Control Toolbox. The MathWorks, Natick, Mass., USA. Choi, S.-G. and M. A. Johnson (1996). The root loci of H -optimal control: A polynomial approach. Optimal Control Applications and Methods 17, 79105. Control Toolbox (1990). Users Guide. The MathWorks, Natick, Mass., USA. Dahleh, M. A. and Y. Ohda (1988). A necessary and sufcient condition for robust BIBO stability. Systems & Control Letters 11, 271275. Davison, E. J. (1996). Robust servomechanism problem. In The Control Handbook, pp. 731 747. CRC press. Davison, E. J. and S. H. Wang (1974). Properties and calculation of transmission zeros of linear multivariable systems. Automatica 10, 643658. Davison, E. J. and S. H. Wang (1978). An algorithm for the calculation of transmission zeros of a the system (C,A,B,D) using high gain output feedback. IEEE Transactions on Automatic Control 23, 738741. DAzzo, J. J. and C. H. Houpis (1988). Linear Control System Analysis and Design: Conventional and Modern. McGraw-Hill Book Co., New York, NY. Third Edition. Desoer, C. A. and J. D. Schulman (1974). Zeros and poles of matrix transfer functions and their dynamical interpretations. IEEE Transactions on Circuits and Systems 21, 38. Desoer, C. A. and M. Vidyasagar (1975). Feedback Systems: Input-Output Properties. Academic Press, New York. Desoer, C. A. and Y. T. Wang (1980). Linear time-invariant robust servomechanism problem: a self-contained exposition. In C. T. Leondes (Ed.), Control and Dynamic Systems, Volume 16, pp. 81129. Academic Press, New York, NY. Dorf, R. C. (1992). Modern Control Systems. Addison-Wesley Publishing Co., Reading, MA. Sixth Edition. Doyle, J. C. (1979). Robustness of multiloop linear feedback systems. In Proc. 17th IEEE Conf. Decision & Control, pp. 1218. Doyle, J. C. (1982). Analysis of feedback systems with structured uncertainties. IEE Proc., Part D 129, 242250. Doyle, J. C. (1984). Lecture Notes, ONR/Honeywell Workshop on Advances in Multivariable Control, Minneapolis, Minn.

References

317

Doyle, J. C., B. A. Francis, and A. R. Tannenbaum (1992). Feedback Control Theory. Macmillan, New York. Doyle, J. C., K. Glover, P. P. Khargonekar, and B. A. Francis (1989). State-space solutions to standard H2 and H control problems. IEEE Trans. Aut. Control 34, 831847. Doyle, J. C. and G. Stein (1981). Multivariable feedback design: concepts for a classical/modern synthesis. IEEE Transactions on Automatic Control 26, 416. Engell, S. (1988). Optimale lineare Regelung: Grenzen der erreichbaren Regelg te in liu nearen zeitinvarianten Regelkreisen, Volume 18 of Fachberichte MessenSteuernRegeln. Springer-Verlag, Berlin, etc. Evans, W. R. (1948). Graphical analysis of control systems. Trans. Amer. Institute of Electrical Engineers 67, 547551. Evans, W. R. (1950). Control system synthesis by root locus method. Trans. Amer. Institute of Electrical Engineers 69, 6669. Evans, W. R. (1954). Control System Dynamics. McGraw-Hill Book Co., New York, NY. F llinger, O. (1958). Uber die Bestimmung der Wurzelortskurve. Regelungstechnik 6, 442 o 446. Francis, B. A. and W. M. Wonham (1975). The role of transmission zeros in linear multivariable regulators. International Journal of Control 22, 657681. Franklin, G. F., J. D. Powell, and A. Emami-Naeini (1986). Feedback Control of Dynamic Systems. Addison-Wesley Publ. Co., Reading, MA. Franklin, G. F., J. D. Powell, and A. Emami-Naeini (1991). Feedback Control of Dynamic Systems. Addison-Wesley Publ. Co., Reading, MA. Second Edition. Freudenberg, J. S. and D. P. Looze (1985). Right half plane poles and zeros and design tradeoffs in feedback systems. IEEE Trans. Automatic Control 30, 555565. Freudenberg, J. S. and D. P. Looze (1988). Frequency Domain Properties of Scalar and Multivariable Feedback Systems, Volume 104 of Lecture Notes in Control and Information Sciences. Springer-Verlag, Berlin, etc. Gantmacher, F. R. (1964). The Theory of Matrices. Chelsea Publishing Company, New York, NY. Reprinted. Glover, K. and J. C. Doyle (1988). State-space formulae for all stabilizing controllers that satisfy an H -norm bound and relations to risk sensitivity. Systems & Control Letters 11, 167172. Glover, K. and J. C. Doyle (1989). A state space approach to H optimal control. In H. Nijmeijer and J. M. Schumacher (Eds.), Three Decades of Mathematical System Theory, Volume 135 of Lecture Notes in Control and Information Sciences. Springer-Verlag, Heidelberg, etc. Glover, K., D. J. N. Limebeer, J. C. Doyle, E. M. Kasenally, and M. G. Safonov (1991). A characterization of all solutions to the four block general distance problem. SIAM J. Control and Optimization 29, 283324. Gohberg, I., P. Lancaster, and L. Rodman (1982). Matrix polynomials. Academic Press, New York, NY. Golub, G. M. and C. Van Loan (1983). Matrix Computations. The Johns Hopkins University Press, Baltimore, Maryland.

318

References

Graham, D. and R. C. Lathrop (1953). The synthesis of optimum transient response: Criteria and standard forms. Trans. AIEE Part II (Applications and Industry) 72, 273288. Green, M. and D. J. N. Limebeer (1995). Linear Robust Control. Prentice Hall, Englewood Cliffs. Grimble, M. J. (1994). Two and a half degrees of freedom LQG controller and application to wind turbines. IEEE Trans. Aut. Control 39(1), 122127. Grosdidier, P. and M. Morari (1987). A computer aided methodology for the design of decentralized controllers. Computers and Chemical Engineering 11, 423433. Grosdidier, P., M. Morari, and B. R. Holt (1985). Closed-loop properties from steady-state gain information. Industrial and Engineering Chemistry. Fundamentals 24, 221235. Hagander, P. and B. Bernhardsson (1992). A simple test example for H optimal control. In Preprints, 1992 American Control Conference, Chicago, Illinois. Henrici, P. (1974). Applied and Computational Complex Analysis, Vol. 1. Wiley-Interscience, New York. Hinrichsen, D. and A. J. Pritchard (1986). Stability radii of linear systems. Systems & Control Letters 7, 110. Horowitz, I. (1982, November). Quantitative feedback theory. IEE Proc. Pt. D 129, 215226. Horowitz, I. and A. Gera (1979). Blending of uncertain nonminimum-phase plants for elimination or reduction of nonminimum-phase property. International Journal of Systems Science 10, 10071024. Horowitz, I. and M. Sidi (1972). Synthesis of feedback systems with large plant ignorance for prescribed time domain tolerances. Int. J. Control 126, 287309. Horowitz, I. M. (1963). Synthesis of of Feedback Systems. Academic Press, New York. Hovd, M. and S. Skogestad (1992). Simple frequency-dependent tools for control system analysis, structure selection and design. Automatica 28, 989996. Hovd, M. and S. Skogestad (1994). Pairing criteria for decentralized control of unstable plants. Industrial and Engineering Chemistry Research 33, 21342139. Hsu, C. H. and C. T. Chen (1968). A proof of the stability of multivariable feedback systems. Proceedings of the IEEE 56, 20612062. James, H. M., N. B. Nichols, and R. S. Philips (1947). Theory of Servomechanisms. McGrawHill, New York. Kailath, T. (1980). Linear Systems. Prentice Hall, Englewood Cliffs, N. J. Kalman, R. E. and R. S. Bucy (1961). New results in linear ltering and prediction theory. J. Basic Engineering, Trans. ASME Series D 83, 95108. Karcanias, N. and C. Giannakopoulos (1989). Necessary and sufcient conditions for zero assignment by constant squaring down. Linear Algebra and its Applications 122/123/124, 415446. Kharitonov, V. L. (1978a). Asymptotic stability of an equilibrium position of a family of systems of linear differential equations. Differentialnye Uraveniya 14, 14831485. Kharitonov, V. L. (1978b). On a generalization of a stability criterion. Akademii Nauk Kahskoi SSR, Fiziko-matematicheskaia 1, 5357. Korn, U. and H. H. Wilfert (1982). Mehrgr ssenregelungen. Moderne Entwurfsprinzipien im o Zeit- und Frequenzbereich. Verlag Technik, Berlin.

References

319

Krall, A. M. (1961). An extension and proof of the root-locus method. Journal Soc. Industr. Applied Mathematatics 9, 644653. Krall, A. M. (1963). A closed expression for the root locus method. Journal Soc. Industr. Applied Mathematics 11, 700704. Krall, A. M. (1970). The root-locus method: A survey. SIAM Review 12, 6472. Krall, A. M. and R. Fornaro (1967). An algorithm for generating root locus diagrams. Communications of the ACM 10, 186188. Krause, J. M. (1992). Comments on Grimbles comments on Steins comments on rolloff of H optimal controllers. IEEE Trans. Auto. Control 37, 702. Kwakernaak, H. (1976). Asymptotic root loci of optimal linear regulators. IEEE Trans. Aut. Control 21(3), 378382. Kwakernaak, H. (1983). Robustness optimization of linear feedback systems. In Preprints, 22nd IEEE Conference of Decision and Control, San Antonio, Texas, USA. Kwakernaak, H. (1985). Minimax frequency domain performance and robustness optimization of linear feedback systems. IEEE Trans. Auto. Control 30, 9941004. Kwakernaak, H. (1986). A polynomial approach to minimax frequency domain optimization of multivariable systems. Int. J. Control 44, 117156. Kwakernaak, H. (1991). The polynomial approach to H -optimal regulation. In E. Mosca and L. Pandol (Eds.), H -Control Theory, Volume 1496 of Lecture Notes in Mathematics, pp. 141221. Springer-Verlag, Heidelberg, etc. Kwakernaak, H. (1993). Robust control and H -optimization. Automatica 29, 255273. Kwakernaak, H. (1995). Symmetries in control system design. In A. Isidori (Ed.), Trends in Control A European Perspective, pp. 1751. Springer, Heidelberg, etc. Kwakernaak, H. (1996). Frequency domain solution of the standard H problem. In M. J. Grimble and V. Ku era (Eds.), Polynomial Methods for Control Systems Design. Springerc Verlag. Kwakernaak, H. and R. Sivan (1972). Linear Optimal Control Systems. Wiley, New York. Kwakernaak, H. and R. Sivan (1991). Modern Signals and Systems. Prentice Hall, Englewood Cliffs, NJ. Kwakernaak, H. and J. H. Westdijk (1985). Regulability of a multiple inverted pendulum system. Control Theory and Advanced Technology 1, 19. Landau, I. D., F. Rolland, C. Cyrot, and A. Voda (1993). R gulation num rique robuste: Le e e placement de p les avec calibrage de la fonction de sensibilit . In A. Oustaloup (Ed.), o e Summer School on Automatic Control of Grenoble. Hermes. To be published. Laub, A. J. and B. C. Moore (1978). Calculation of transmission zeros using QZ techniques. Automatica 14, 557566. Le, D. K., O. D. I. Nwokah, and A. E. Frazho (1991). Multivariable decentralized integral controllability. International Journal of Control 54, 481496. Le, V. X. and M. G. Safonov (1992). Rational matrix GCDs and the design of squaringdown compensators- a state-space theory. IEEE Transactions on Automatic Control 37, 384392. Lin, J.-L., I. Postlethwaite, and D.-W. Gu (1993). K Iteration: A new algorithm for synthesis. Automatica 29, 219224.

320

References

Luenberger, D. G. (1969). Optimization by Vector Space Methods. John Wiley, New York. Lunze, J. (1985). Determination of robust multivariable I-controllers by means of experiments and simulation. Systems Analysis Modelling Simulation 2, 227249. Lunze, J. (1988). Robust Multivariable Feedback Control. Prentice Hall International Ltd, London, UK. Lunze, J. (1989). Robust Multivariable Feedback Control. International Series in Systems and Control Engineering. Prentice Hall International, Hempstead, UK. MacDuffee, C. C. (1956). The Theory of Matrices. Chelsea Publishing Company, New York, N.Y., USA. MacFarlane, A. G. J. and N. Karcanias (1976). Poles and zeros of linear multivariable systems: a survey of the algebraic, geometric and complex-variable theory. International Journal of Control 24, 3374. Maciejowski, J. M. (1989). Multivariable Feedback Design. Addison-Wesley Publishing Company, Wokingham, UK. McAvoy, T. J. (1983). Interaction Analysis. Principles and Applications. Instrument Soc. Amer., Research Triangle Park, NC. McFarlane, D. C. and K. Glover (1990). Robust Controller Design Using Normalized Coprime Factor Plant Descriptions, Volume 146. Springer-Verlag, Berlin, etc. Middleton, R. H. (1991). Trade-offs in linear control system design. Automatica 27, 281292. Minnichelli, R. J., J. J. Anagnost, and C. A. Desoer (1989). An elementary proof of Kharitonovs stability theorem with extensions. IEEE Trans. Aut. Control 34, 9951001. Morari, M. (1985). Robust stability of systems with integral control. IEEE Transactions on Automatic Control 30, 574577. Morari, M. and E. Zariou (1989). Robust Process Control. Prentice Hall International Ltd, London, UK. Nett, C. N. and V. Manousiouthakis (1987). Euclidean condition and block relative gain: connections, conjectures, and clarications. IEEE Transactions on Automatic Control 32, 405407. Noble, B. (1969). Applied Linear Algebra. Prentice Hall, Englewood Cliffs, NJ. Nwokah, O. D. I., A. E. Frazho, and D. K. Le (1993). A note on decentralized integral controllability. International Journal of Control 57, 485494. Nyquist, H. (1932). Regeneration theory. Bell System Technical Journal 11, 126147. Ogata, K. (1970). Modern Control Engineering. Prentice Hall, Englewood Cliffs, NJ. Packard, A. and J. C. Doyle (1993). The complex structured singular value. Automatica 29, 71109. Postlethwaite, I., J. M. Edmunds, and A. G. J. MacFarlane (1981). Principal gains and principal phases in the analysis of linear multivariable feedback systems. IEEE Transactions on Automatic Control 26, 3246. Postlethwaite, I. and A. G. J. MacFarlane (1979). A Complex Variable Approach to the Analysis of Linear Multivariable Feedback Systems, Volume 12 of Lecture Notes in Control and Information Sciences. Springer-Verlag, Berlin, etc. Postlethwaite, I., M. C. Tsai, and D.-W. Gu (1990). Weighting function selection in H design. In Proceedings, 11th IFAC World Congress, Tallinn. Pergamon Press, Oxford, UK.

References

321

Qiu, L., B. Bernhardsson, A. Rantzer, E. J. Davison, P. M. Young, and J. C. Doyle (1995). A formula for computation of the real stability radius. Automatica 31, 879890. Raisch, J. (1993). Mehrgr ssenregelung im Frequenzbereich. R.Oldenbourg Verlag, M nchen, o u BRD. Rosenbrock, H. H. (1970). State-space and Multivariable Theory. Nelson, London. Rosenbrock, H. H. (1973). The zeros of a system. International Journal of Control 18, 297 299. Rosenbrock, H. H. (1974a). Computer-aided Control System Design. Academic Press, London. Rosenbrock, H. H. (1974b). Correction on The zeros of a system. International Journal of Control 20, 525527. Saberi, A., B. M. Chen, and P. Sannuti (1993). Loop transfer recovery: Analysis and design. Prentice Hall, Englewood Cliffs, N. J. Saberi, A., P. Sannuti, and B. M. Chen (1995). H2 optimal control. Prentice Hall, Englewood Cliffs, N. J. ISBN 0-13-489782-X. Safonov, M. G. and M. Athans (1981). A multiloop generalization of the circle criterion for stability margin analysis. IEEE Trams. Automatic Control 26, 415422. Sain, M. K. and C. B. Schrader (1990). The role of zeros in the performance of multiinput,multioutput feedback systems. IEEE Transactions on Education 33, 244257. Savant, C. J. (1958). Basic Feedback Control System Design. McGraw-Hill Book Co., New York, NY. Schrader, C. B. and M. K. Sain (1989). Research on system zeros: A survey. International Journal of Control 50, 14071433. Sebakhy, O. A., M. ElSingaby, and I. F. ElArabawy (1986). Zero placement and squaring problem: a state space approach. International Journal of Systems Science 17, 17411750. Shaked, U. (1976). Design techniques for high feedback gain stability. International Journal of Control 24, 137144. Silverman, L. M. (1970). Decoupling with state feedback and precompensation. IEEE Trans. Automatic Control AC-15, 487489. Sima, V. (1996). Algorithms for linear-quadratic optimization. Marcel Dekker, New YorkBasel. ISBN 0-8247-9612-8. Skogestad, S., M. Hovd, and P. Lundstr (1991). Simple frequency-dependent tools for om analysis of inherent control limitations. Modeling Identication and Control 12, 159177. Skogestad, S. and M. Morari (1987). Implications of large RGA elements on control performance. Industrial and Engineering Chemistry Research 26, 23232330. Skogestad, S. and M. Morari (1992). Variable selection for decentralized control. Modeling Identication and Control 13, 113125. Skogestad, S. and I. Postlethwaite (1995). Multivariable Feedback Control. Analysis and Design. John Wiley and Sons Ltd., Chichester, Sussex, UK. Stoorvogel, A. A. (1992). The H-Innity Control Problem: A State Space Approach. Prentice Hall, Englewood Cliffs, NJ. Stoorvogel, A. A. and J. H. A. Ludlage (1994). Squaring down and the problems of almost zeros for continuous time systems. Systems and Control Letters 23, 381388.

322

References

Svaricek, F. (1995). Zuverl ssige numerische Analyse linearer Regelungssysteme. a B.G.Teubner Verlagsgesellschaft, Stuttgart, BRD. Thaler, G. J. (1973). Design of Feedback Systems. Dowden, Hutchinson, and Ross, Stroudsburg, PA. Tolle, H. (1983). Mehrgr ssen-Regelkreissynthese. Band I. Grundlagen und Frequenzbereo ichsverfahren. R.Oldenbourg Verlag, M nchen, BRD. u Tolle, H. (1985). Mehrgr ssen-Regelkreissynthese. Band II: Entwurf im Zustandsraum. o R.Oldenbourg Verlag, M unchen, BRD. Trentelman, H. L. and A. A. Stoorvogel (1993, November). Control theory for linear systems. Vakgroep Wiskunde, RUG. Course for the Dutch Graduate Network on Systems and Control, Winter Term 19931994. Truxal, J. G. (1955). Automatic Feedback Control System Synthesis. McGraw-Hill Book Co., New York, NY. Van de Vegte, J. (1990). Feedback control systems. Prentice-Hall, Englewood Cliffs, N. J. Second Edition. Vardulakis, A. I. G. (1991). Linear Multivariable Control: Algebraic Analysis and Synthesis Methods. John Wiley and Sons Ltd., Chichester, Sussex, UK. Verma, M. and E. Jonckheere (1984). L -compensation with mixed sensitivity as a broadband matching problem. Systems & Control Letters 14, 295306. Vidyasagar, M., H. Schneider, and B. A. Francis (1982). Algebraic and topological aspects of feedback stabilization. IEEE Trans. Aut. Control 27, 880894. Vidysagar, M. (1985). Control Systems SynthesisA Factorization Approach. MIT Press, Cambridge, MA. Williams, T. and P. J. Antsaklis (1996). Decoupling. In The Control Handbook, pp. 795804. CRC press. Williams, T. W. C. and P. J. Antsaklis (1986). A unifying approach to the decoupling of linear multivariable systems. Internat. J. Control 44(1), 181201. Wilson, G. T. (1978). A convergence theorem for spectral factorization. J. Multivariate Anal. 8, 222232. Wolovich, W. A. (1974). Linear Multivariable Systems. Springer Verlag, New York, NY. Wong, E. (1983). Introduction to random processes. Springer-Verlag, New York. Wonham, W. M. (1979). Linear Multivariable Control: A Geometric Approach. SpringerVerlag, New York, etc. Youla, D. C., J. J. Bongiorno, and C. N. Lu (1974). Single-loop feedback-stabilization of linear multivariable dynamical plants. Automatica 10, 159173. Young, P. M., M. P. Newlin, and J. C. Doyle (1995a). Computing bounds for the mixed problem. Int. J. of Robust and Nonlinear Control 5, 573590. Young, P. M., M. P. Newlin, and J. C. Doyle (1995b). Lets get real. In B. A. Francis and P. P. Khargonekar (Eds.), Robust Control Theory, Volume 66 of IMA Volumes in Mathematics and its Applications, pp. 143173. Springer-Verlag. Zames, G. (1981). Feedback and optimal sensitivity: Model reference transformations, multiplicative seminorms, and approximate inverses. IEEE Trans. Aut. Control 26, 301320.

References

323

Zhou, K., J. C. Doyle, and K. Glover (1996). Robust and Optimal Control. Prentice Hall, Englewood Cliffs. ISBN 0-13-456567-3. Ziegler, J. G. and N. B. Nichols (1942). Optimum settings for automatic controllers. Trans. ASME, J. Dyn. Meas. and Control 64, 759768.

324

References

Index
, 124 1 1 2 -degrees-of-freedom, 51 2-degrees-of-freedom, 11 acceleration constant, 63 action, 173 additive perturbation model, 209 adjoint of rational matrix, 264 algebraic Riccati equation, 105 amplitude, 169 ARE, 105 solution, 114 asymptotic stability, 12 B zout equation, 16 e bandwidth, 28, 80 improvement, 8 basic perturbation model, 208 BIBO stability, 13 biproper, 268 Blaschke product, 40 Bode diagram, 70 gain-phase relationship, 35 magnitude plot, 70 phase plot, 70 sensitivity integral, 37 Butterworth polynomial, 90 canonical factorization, 268 central solution (of H problem), 267 certainty equivalence, 121 characteristic polynomial, 14, 311 classical control theory, 59 closed-loop characteristic polynomial, 15 system matrix, 107 transfer matrix, 33 complementary sensitivity function, 23 325 sensitivity matrix, 33 complex conjugate transpose, 312 constant disturbance model, 280 constant disturbance rejection, 65 constant error suppression, 280 contraction, 211 control decentralized-, 185 controllability gramian, 174 crossover region, 32 decentralized control, 156, 185 integral action, 188 decoupling control, 156 indices, 180 delay margin, 21 time, 81 derivative time, 66 Descartes rule of sign, 201 descriptor representation, 276 determinant, 311 DIC, 188 disturbance condition number, 185 D-K iteration, 291 D-scaling upper bound, 238 edge theorem, 205 eigenvalue, 311 decomposition, 311 eigenvector, 311 elementary column operation, 159 elementary row operation, 159 energy, 169 equalizing property, 275 error covariance matrix, 118 error signal, 5 Euclidean norm, 168

326

Index

feedback equation, 6 xed point theorem, 211 forward compensator, 5 Freudenberg-looze equality, 40 full perturbation, 218 functional controllability, 178 reproducibility, 178 FVE, 81 gain, 88 at a frequency, 69 crossover frequency, 80 margin, 21, 80 generalized Nyquist criterion, 19 generalized plant, 126, 258 gridding, 203 Guillemin-Truxal, 89

interconnection matrix, 208 internal model principle, 65 internal stability, 13, 175 ITAE, 91 J-unitary, 266 Kalman-Jakubovi -Popov equality, 108 c Kalman lter, 118 Kharitonovs theorem, 204 KJP equality, 108 inequality, 108

L2 -norm, 168
lag compensation, 84 leading coefcient, 16 lead compensation, 82 linearity improvement, 8 linear regulator problem stochastic, 116 L -norm, 169 loop gain, 18 gain matrix, 18 IO map, 6, 10 shaping, 34, 82 transfer matrix, 15 transfer recovery, 123 LQ, 104 LQG, 115 LTR, 123 Lyapunov equation, 118 Lyapunov matrix differential equation, 142 margin delay-, 21 gain-, 21, 80 modulus-, 21 multivariable robustness-, 234 phase-, 21, 80 matrix, 311 closed loop system-, 107 complementary sensitivity-, 33 determinant, 311 error covariance-, 118 Hamiltonian, 114 Hermitian, 312 input sensitivity-, 33 intensity-, 116

H2 -norm, 124, 173 H2 -optimization, 125


Hadamard product, 186 Hankel norm, 173 Hermitian, 312 high-gain, 6 H -norm, 172, 213 H -problem, 258 central solution, 267 equalizing property, 275 Hurwitz matrix, 207 polynomial, 200 strictly-, 181 tableau, 201 induced norm, 171 inferential control, 128 -norm of transfer matrix, 213 of vector, 168 input-output pairing, 186 input sensitivity function, 30 matrix, 33 integral control, 66 integrating action, 64 integrator in the loop, 132 intensity matrix, 116 interaction, 156

Index

327

loop transfer-, 15 nonnegative denite-, 312 nonsingular, 311 observer gain-, 117 positive denite-, 312 positive semi-denite-, 312 proper rational-, 18 rank, 311 sensitivity-, 33 singular, 311 state feedback gain-, 105 symmetric, 312 system-, 14 trace, 311 unitary, 312 M-circle, 76 measurement noise, 31 MIMO, 11, 33, 153 minimal realization, 163 minimum phase, 37 mixed sensitivity problem, 251 type k control, 252 modulus margin, 21 monic, 109 , 234 -synthesis, 289 H , 236 multiloop control, 186 multiplicative perturbation model, 209 multivariable dynamic perturbation, 233 multivariable robustness margin, 234 N-circle, 76 Nehari problem, 260 Nichols chart, 77 plot, 77 nominal, 32 stability, 214 nonnegative denite, 312 nonsingular matrix, 311 norm, 168 Euclidean, 168 Hankel, 173 induced, 171 -, 213 L p -, 168 matrix, 171 on n , 168

p-, 168 spectral-, 170 submultiplicative property, 171 normal rank polynomial matrix, 159 rational matrix, 160 notch lter, 84 Nyquist plot, 18, 74 stability criterion, 18 observability gramian, 174 observer, 117 gain matrix, 117 poles, 122 open-loop characteristic polynomial, 15 pole, 88 stability, 20 zero, 88 optimal linear regulator problem, 104 order of realization, 163 output controllability, 178 functional reproducible, 178 principal directions, 184 overshoot, 81 para-Hermitean, 265 parametric perturbation, 198 parasitic effect, 32 parity interlacing property, 20 partial pole placement, 254 peak value, 169 perturbation full, 218 multivariable dynamic, 233 parametric, 198 repeated real scalar, 233 repeated scalar dynamic, 233 unstructured, 208 perturbation model additive, 209 basic, 208 multiplicative, 209 proportional, 209 phase crossover frequency, 80

328

Index

margin, 21, 80 minimum-, 37 shift, 69 PI control, 66 PID control, 66 plant capacity, 30 p-norm, 168 of signal, 168 of vector, 168 pole assignment, 16 excess, 252 of MIMO system, 161 placement, 16 pole-zero excess, 74 polynomial characteristic-, 14 invariant-, 159 matrix, Smith form, 159 matrix fraction representation, 226 monic, 109 unimodular matrix, 159 position error constant, 62 positive denite, 312 semi denite, 312 principal direction input, 184 output, 184 principal gain, 184 proper, 18 rational function, 15 strictly-, 16 proportional control, 66 feedback, 7 perturbation model, 209 pure integral control, 66 QFT, 93 quantitative feedback theory, 93 rank, 311 rational matrix fractional representation, 226 Smith McMillan form, 160 realization minimal, 163 order of, 163

regulator poles, 122 system, 2 relative degree, 74, 252 relative gain array, 186 repeated real scalar perturbation, 233 repeated scalar dynamic perturbation, 233 reset time, 66 resonance frequency, 80 peak, 80 return compensator, 5 return difference, 108 equality, 108, 144 inequality, 108 RGA, 186 rise time, 81 robustness, 7 robust performance, 242 roll-off, 32, 87 LQG, 132 mixed sensitivity, 252 root loci, 88 Schur complement, 114, 313 sector bounded function, 220 sensitivity complementary-, 23 function, 23 input-, 30 matrix, 33 matrix, complementary-, 33 matrix, input-, 33 separation principle, 121 servocompensator, 189 servomechanism, 189 servo system, 2 settling time, 81 shaping loop-, 82 Sherman-Morrison-Woodbury, 313 , , 170 signal amplitude, 169 energy, 169 L2 -norm, 168 L -norm, 169 peak value, 169 signature matrix, 265

Index

329

single-degree-of-freedom, 12 singular matrix, 311 singular value, 169 decomposition, 169 SISO, 11 small gain theorem, 212 Smith-McMillan form, 160 Smith form, 159 spectral cofactorizaton, 306 factor, 265 spectral norm, 170 square down, 162 stability, 11 asymptotic-, 12 BIBO-, 13 internal, 175 internal-, 13 matrix, 174 nominal, 214 Nyquist (generalized), 19 Nyquist criterion, 18 of polynomial, 206 of transfer matrix, 175 open-loop, 20 radius, 220 radius, complex, 220 radius, real, 220 region, 201 standard H problem, 258 standard H2 problem, 126 state estimation error, 117 feedback gain matrix, 105 state-space realization, 163 static output feedback, 164 steady-state, 60 error, 60 step function unit, 60 strictly proper, 16 structured singular value, 234 D-scaling upper bound, 238 lower bound, 236, 238 of transfer matrix, 236 scaling property, 236 upper bound, 236 submultiplicative property, 171 suboptimal solution (of H problem), 264

SVD, 169 Sylvester equation, 16 symmetric, 312 system matrix, 14 trace, 119, 311 tracking system, 2 transmission zero rational matrix, 160 triangle inequality, 168 two-degrees-of-freedom, 11 type k control mixed sensitivity, 252 type k system, 61 unimodular, 159 unit feedback, 5 step function, 60 unitary, 312 unstructured perturbations, 208 vector space normed, 168 velocity constant, 63 well-posedness, 176, 263 zero open loop, 88 polynomial matrix, 159