You are on page 1of 23

Fundamental limitations for control of unstable S ISO plants

K. Havre and S. Skogestad


N-7034 Trondheim, Norway.
 y

Chemical Engineering, Norwegian University of Science and Technology,

Submitted 10 Oct. 1997 for possible publication in Automatica.

Abstract This paper examines the fundamental limitations on closed-loop performance imposed by instability in the plant (Right Half Plane (R HP) poles). The main limitation is that instability requires active use of plant inputs, and we quantify this is terms of tight lower bounds on the input magnitudes required for disturbance and measurement noise rejection. These new bounds involve the H1 -norm, which has direct engineering signicance. The output performance in terms of disturbance rejection or reference tracking is only limited if the plant has R HP-zeros, and for a one degree-of-freedom controller the presence of R HP-poles further deteriorate the response, whereas there is no additional penalty for having R HP-poles if we use a two degrees-of-freedom controller. It is important to stress that the derived bounds are controller independent and that they are tight, meaning that there exists controllers which achieve the lower bounds.

Introduction

An unstable plant can only be stabilized by use of feedback control which implies active use of the plant inputs. If measurement noise and/or disturbances are present (which is always the case in practical control), then the input usage may become unacceptable. In this paper, the above statements are quantied by deriving tight lower bounds on the H1 -norm of the closed-loop transfer functions SV and TV , where S and T are the sensitivity and complementary sensitivity functions. The transfer function V can be viewed as a generalized weight, which for our purpose should be independent of the feedback controller K . Some reasons for deriving such bounds are: 1) The lower bounds provide direct insights to the limitations imposed by R HP zeros and poles in S ISO systems. 2) The lower bounds derived are independent of the controller, so they can be used as controllability measures.
 Also afliated with: Institute for energy technology, P.O.Box 40, N-2007 Kjeller, Norway, E-mail: kjetil@ife.no. y Author to whom correspondence should be addressed, Fax: (+47) 73 59 40 80, E-mail: skoge@kjemi.unit.no.

3) In some cases we can show that the bounds are tight. This implies that we in these cases can nd a controller K , analytically, which achieves an H1 -norm of the closed-loop transfer function equal to the lower bound. 4) We can quantify, in terms of the H1 -norm, the best achievable closed-loop effect of the worst case disturbance, measurement noise and references both at the input and at the output of the plant. One important application is that we can quantify the minimum input usage for stabilization in the presence of worst case measurement noise and disturbances. It appears that even for S ISO systems this has been a difcult task, which has not been solved analytically until now. To give the reader some appreciation of the basis of the bounds and their usefulness, we consider as a motivating example an unstable plant with a R HP-pole p. We want to obtain a lower bound on the H1 -norm of the closed-loop transfer function KS from measurement noise n to plant input u. We rst rewrite KS = G,1 T , which is on the form TV with V = G,1 . The basis of our bound is the use of the maximum modulus principle and the interpolation constraint T p = 1, which must apply to achieve internal stability. We obtain (see Theorem 1 for details)

1 kKS sk1 = kG,1T s k1  jG, mspj


where Gms is the stable and minimum phase version of G (if G also has a R HP-zero z we get the z+pj 1 additional penalty j jz,pj ). As an example, consider the plant Gs = s,10 , which has an unstable pole 1 . For any linear feedback controller K , we nd that the lower p = 10. We obtain Gmss = s+10 bound 1 kKS sk1  jG, ms pj = 2p = 20 must be satised. Thus, if we require that the plant inputs are bounded with kuk1  1, then we cannot allow the magnitude of measurement noise to exceed knk1 = 1=20 = 0:05. The basis for our results is the important work by Zames (1981), who made use of the interpolation constraint S z  = 1 and the maximum modulus theorem to derive bounds on the H1 -norm of S for plants with one R HP-zero. Subsequently, these results were extended to unstable plants with one R HP-pole and then to plants with combined RHP zeros and poles, e.g. (Doyle, Francis and Tannenbaum, 1992, pp. 9395) and (Skogestad and Postlethwaite, 1996). However, these generalizations to unstable plants did not consider the input usage which involves the closed-loop transfer function KS . An important contribution of this paper is therefore to use the trick KS = G,1 T , which enable us to derive lower bounds on input usage, by using the general lower bound on kTV sk1 with V = G,1 . But when G is unstable (with R HP-pole p), then V = G,1 has R HP-zeros for s = p. A second important contribution compared to earlier work is the ability to include R HP zeros and poles in the weight V (under the assumption that SV and TV are stable). A third important contribution is that we show that the lower bounds are tight. That is, we give analytical expressions for stable controllers which achieves an H1 -norm of the closed-loop transfer function which is equal to the lower bound. Several authors, among them Kwakernaak (1995), have noted the symmetries between sensitivity and complementary sensitivity and the roles of R HP zeros and poles of G. In this paper, the symmetries are also reected in where performance is measured. We nd that RHP-zeros of G pose limitations on performance measured at the output of the plant whereas R HP-poles of G pose limitations on performance measured at the input of the plant (input usage). The bounds on kS sk1 for plants with R HP-zero derived by Zames (1981) are also valid for multivariable systems. It is important to note that all the results given in this paper have been generalized to multivariable systems (Havre and Skogestad, 1997b). However, the notation becomes complicated 2

in the multivariable case, with the result that it is difcult to understand the implications of the bounds. In the S ISO case, the bounds may easily be derived by hand for a particular plant. However, in the multivariable case, we must in general evaluate the bounds numerically. The paper is organized as follows: First we introduce the notation and present some basics from linear control theory. In Section 3 we derive the general lower bounds on kSV sk1 and kTV sk1, and in Section 4 we prove the tightness of the lower bounds. In Section 5 we show some applications and implications of the lower bounds both on output performance, input usage, peaks in sensitivity and complementary sensitivity, and we give some simple examples to illustrate the applications and the implications. In Section 6 we discuss briey the relationship to stabilization with input constraints. In Section 7 we derive a lower bound applicable to two degrees-of-freedom (2-DOF) control. The proofs of the results which are not given in the main text are given in Appendix A.

Basics from linear control theory

We consider linear time invariant transfer function models on the form

ys = Gsus + Gdsds (1) where u is the manipulated input, d is the disturbance, y is the output, G is the S ISO plant model and Gd is the S ISO disturbance plant model. The measured output is ym = y + n where n is the
measurement noise. The H1 -norm of a stable rational transfer function magnitude jM j! j over all frequencies.

M s is dened as the peak value in the


(2)

kM sk1 , sup jM j!j


!

2.1 Zeros and poles


In a rational transfer function M the zeros and poles are the roots of the numerator and denominator polynomials. That is, the zeros zj and the poles pi are the solutions to the following equations

M zj  = 0

and

M ,1 pi = 0

(3)

When we refer to zeros and poles we mean the zeros and poles of the plant explicitly stated.

G unless otherwise

2.2 Factorizations of R HP zeros and poles


A rational transfer function M s with zeros zj and poles pi in the open R HP, factorized in Blaschke products as follows1

fzj ; pig 2 C + , can be


(4) (5) (6)

M s = Bz M  Mm s ,1 M  M s M s = Bp s , 1 M s = Bz M  Bp M  Mms s


where
1

Note that the notation on the all-pass factorizations of R HP zeros and poles used in this paper is reversed compared to the notation used in (Skogestad and Postlethwaite, 1996; Havre and Skogestad, 1997a). The reason to this change of notation is to get consistent with what the literature generally denes as an all-pass lter.

m

m

s,z

s+z

s,p
s

1 s+p

s,z s,p

 ms
s

 - s+? z s+p

(a) on R HP-zeros z

(b) on R HP-poles p

(c) on R HP zeros and poles

Figure 1: Operations on R HP zeros and poles for scalar transfer functions

Minimum phase (subscript m) version of M with the R HP-zeros mirrored across the imaginary axis. Ms Stable (subscript s) version of M with the RHP-poles mirrored across the imaginary axis. Mms Minimum phase, stable (subscript ms) version of M with the RHP zeros and poles mirrored across the imaginary axis. Bz M  Stable all-pass rational transfer function jBz j!j = 1; 8! containing the RHP-zeros (subscript z ) of M . Bp M  Stable all-pass rational transfer function jBpj!j = 1; 8! containing the RHP-poles (subscript p) of M as R HP-zeros. The all-pass lters are

Mm

Bz M s = BpM s =

Nz Y s

j j =1 s + z Np Y s , pi i i=1 s + p

, zj

(7)

(8)

where Nz is the number of R HP-zeros zj 2 C + and Np is the number of R HP-poles pi 2 C + in M . In most cases M = G and to simplify the notation we often omit to show that the all-pass lters are dependent on G, i.e. we write Bp s and Bz s in the meaning of Bp Gs and Bz Gs. Figure 1 demonstrates the operations, m for R HP-zeros, s for R HP-poles and the combined operator ms of scalar transfer functions. The order of the two operations m and s in the combined operator ms is arbitrary. It also follows that

1 G,1ms = Gms,1 = G, ms
And we note that

(9)

kM sk1 = kMm sk1 = kMmssk1 (10) The rst identity follows since jBz M j! j = 1; 8! , and the latter identity follows since M is stable, i.e. Mms = Mm and Bp Mm  = Bp M  = 1.
To prove the main results in this paper we make use of the following Lemma. L EMMA 1. Consider a stable S ISO transfer function AB which can be expressed by the product of the S ISO transfer functions A and B , where both A and B may be unstable. Then

kAB k1 = kAB m k1 = kAmsBms k1


4

(11)

~= d d

r ~cR r+6 - K y n ~-c m N n


q

- wu e - + wP 6 ? Gd + y +? u -G c c
q

z2 z1 -

+ +

Figure 2: One degree-of-freedom control conguration

2.3 Closing the loop


A typical control problem is shown in Figure 2. In the gure possible performance weights are given in dashed lines. Mainly for simplicity, but also because it is most practically relevant, we assume that the performance weights wP and wu are stable and minimum phase. If intergrators (poles at s = 0) are present in wP and wu then we need the same number of integrators in L = GK , to have a stable closed-loop transfer function. In Figure 2 we have included both the reference r and the measurement noise n, in addition to disturbances d as external inputs. The transfer functions, Gd , R and N can be ~, r ~ and n ~ are normalized in magnitude. Normally, viewed as weights on the inputs, and the inputs: d N is the inverse of signal to noise ratio. For most practical purposes, we can assume that R and N are stable. However from the technical point of view it sufce that the unstable modes in N or R can be stabilized through the input u. We apply negative feedback control

u = K r , ym = K r , y , n
The closed-loop transfer function F from

(12)

r ~ ~5 v = 4d n ~
is


2 3

to

wP y , r 1 z= z z2 = wuu

where the sensitivity S and the complementary sensitivity T are dened as

wP SR wP SGd ,wP TN F s = , wuSKR ,wuKSGd ,wuKSN

(13)

S , 1 + GK ,1 = 1 +1GK (14) GK T , 1,S = 1+ (15) GK To have good control performance (keep z1 small) with a small input usage (keep z2 small) we need to have kF sk1 small. That is we want all the S ISO transfer functions in (13) small. In addition, there are robustness issues. For example, we wish to have kwunc T sk1 small, where wunc is the
magnitude of the relative plant uncertainty. The rst requirement for being able to satisfying all these objectives (e.g. having all seven transfer functions mentioned above small), is that the weights wP , wu and wunc are such that the objectives 5

can be achieved. For example, since S + T = 1 we cannot have wP R and wunc large at the same frequency if we want to have kwP SRsk1 (tight control of setpoint changes) and kwunc T sk1 (the closed-loop response is insensitive to plant uncertainty) small. However, the presence of R HP zeros and poles in the plant G provide additional limitations, which is the focus of this paper.

2.4 Interpolation constraints


If G has a R HP-zero z or a R HP-pole p then for internal stability of the feedback system the following interpolation constraints must apply (e.g. Skogestad and Postlethwaite, 1996):

T z = 0; S p = 0;

S z = 1 T p = 1

(16) (17)

Lower bounds on the tions

H1-norm of closed-loop transfer func-

In this section we will give the main results, which are lower bounds on the H1 -norm of closedloop transfer functions which can be written on the form TV or SV . The generalized weight V is assumed to be independent of the feedback controller K . V may be unstable but TV and SV must be stable. That is, it must be possible to stabilize all transfer functions by controlling the output y using the input u (this implies that all unstable modes of V also are modes of G). Some examples. Consider the six transfer functions in (13). The rst two can be written on the form SV by selecting V11 = wP R and V12 = wP Gd . The remaining four can be written on the form TV by selecting V13 = wP N , V21 = wuG,1R, V22 = wuG,1Gd and V23 = wuG,1N . From this we see that the weight V may be unstable (if one or both of Gd and G,1 are unstable) and may contain R HP-zeros (if one or both of Gd and G,1 contain R HP-poles). In the rst result, which is the lower bound on kTV sk1 , we consider any number of R HP-zeros in the plant G and one R HP-pole at a time. Then by maximizing over all R HP-poles in the plant G we nd the largest lower bound on kTV sk1 which takes into account one R HP-pole and all R HP-zeros. T HEOREM 1 (L OWER BOUND ON kTV sk1 ). Consider the S ISO plant G with Nz  0 R HP-zeros zj 2 C + and Np  1 RHP-poles pi. Let V be a rational transfer function, and assume that TV is (internally) stable. Then the following lower bound on kTV sk1 applies:

kTV sk1  R max jBz,1pij  jVmspij -poles, pi


HP

(18)

The lower bound (18) is independent of the controller K , if the weight V is independent of K . The ,1 pi j takes into account the interactions between all the RHP-zeros zj 2 C + and the single factor jBz R HP-pole pi of G. As we shall see this factor can be quite large if G contains one or more RHP-zeros close to the R HP-pole pi .
Proof of Theorem 1.

R EMARK 1. With jBz pi j we mean jBz Gsj evaluated at s = pi . R EMARK 2. The assumption that TV is internally stable, means that zero/pole cancellations between G and K .

TV

is stable and we have no R HP

1) Factor out R HP zeros and poles in

T and V . Lemma 1 gives fs T V s fs

2) Introduce the stable scalar function   = m ms  . 3) Apply the maximum modulus theorem to   at the R HP-poles

kTV sk1 = kTms Vmssk1 = kTm Vmssk1 where the last equality holds since T is stable, i.e. Tms = Tm .

pi of G.

4)

,1 pi j  jVms pi j jf pij = jBz (19) Note that f pi  is independent of the controller K if V is independent of K . 2 Since these steps holds for all R HP-poles pi , Theorem 1 follows. In the next result, which is the lower bound on kSV sk1, we consider any number of R HP-poles in the plant G and one R HP-zero at a time. Then by maximizing over all R HP-zeros in the plant G we nd the largest lower bound on kSV sk1 which takes into account one R HP-zero and all R HP-poles.
T HEOREM 2 (L OWER BOUND ON kSV sk1 ). Consider the S ISO plant G with Nz  1 R HP-zeros zj and Np  0 RHP-poles pi 2 C + . Let V be a rational transfer function, and assume that SV is (internally) stable. Then the following lower bound on kSV sk1 applies:

5) 6) Evaluate the lower bound.

kf sk1  jf pij ,1 pi  to get Resubstitute the factorization of R HP-zeros in T , i.e. use Tm pi  = T pi Bz ,1 pi Vms pi  f pi = Tmpi Vmspi  = T piBp Use the interpolation constraint (17) for R HP-poles pi in G, i.e. use T pi  = 1.

,1 z j  jV z j kSV sk1  R max jBp j ms j -zeros, zj


HP

(20)

Remarks on Theorems 2 and 1: 1) The lower bounds on kTV sk1 and kSV sk1 involve Vms . Thus, we get the same result if the weight V is replaced by its stable minimum phase counterpart with the same magnitude Vms. Note that for V = V 1 V 2 we have

1 V2 k kTV sk1 = kTm Vms ms 1

(21)

Which means that we can treat the different factors of V independently. ,1 pi j  1 2) The bound on kTV sk1 is caused by the R HP-poles pi in G, and the term jBz gives an additional penalty for plants which also have RHP-zeros. For the case when G has no ,1 pi  = 1. R HP-zeros, then Bz ,1 zj j  1 3) The bound on kSV sk1 is caused by the R HP-zeros zj in G, and the term jBp gives an additional penalty for plants which also have RHP-poles. For the case when G has no ,1 zj  = 1. R HP-poles, then Bp 4) In all the lower bounds which follows from Theorems 2 and 1, one of the following two factors appears

Nz Q

jB,1p j
z i

=
7

j =1 Nz Q j =1

jpi + z j j jpi , zj j

1

(22)

Np Q

jB,1z
p

j j

i=1 Np Q i=1

jzj + p ij jzj , pij

1

(23)

,1 pi j is a measure for the interactions between all RHP-zeros zj 2 C + of G The factor jBz and the single R HP-pole pi of G. If one or more R HP-zeros are close to the RHP-pole pi , then ,1 zj j combines all R HP-poles pi 2 C + jBz,1 pij is much larger than one. In a similar way jBp ,1 zj j is much larger than one if of G together with the single R HP-zero zj of G. Clearly, jBp one or more R HP-poles are located close to the R HP-zero zj .

Tightness of lower bounds

Theorems 1 and 2 provide lower bounds on kTV sk1 and kSV sk1 . The question is whether these bounds are tight, meaning that there actually exist controllers which achieve the bounds? The answer is yes if there is only one R HP-zero or one R HP-pole. Specically, we nd that the bound kTV sk1 is tight if the plant G has one RHP-pole and any number of RHP-zeros, and that the bound on kSV sk1 is tight if the plant G has one R HP-zero and any number of R HP-poles. We prove tightness of the lower bounds by constructing controllers which achieve the bounds. First, we consider the controller which minimizes kTV sk1 . T HEOREM 3 (K WHICH M INIMIZE kTV sk1 ). Consider the S ISO plant G with one R HP-pole p and Nz  0 R HP-zeros zj 2 C + . Then the feedback controller K which minimize kTV sk1 is given by 1 ,1 K s = G, (24) ms Ko s; Ko s = PQ s where

,1 p V p V ,1 s P s = Bz ms ms ,1 s 1 , B s P s Qs = 1 , Bz s P sm = Bp z
With this controller we have

(25) (26) (27)

kTV sk1 = jBz,1pj  jVmspj

which shows that the bound given in Theorem 1 is tight when the plant has one R HP-pole. We stress that the bound given in Theorem 1 is generally not tight if the plant has more than one R HP-pole. The controller in Theorem 3 yields a constant (at) frequency response jTV j! j for all !. We note that no properness restriction has been imposed on the controller, so the controller given in Theorem 3 may be improper. Also note that the controller K s in Theorem 3 is always stable and minimum phase. This may seem surprising since it is known that some plants with RHP zeros and poles require an unstable controller (Youla, Bongiorno and Lu, 1974) to achieve closed-loop stability. However, these results assume that the loop transfer function GK is proper or strictly proper, and does therefore not apply in our case where K may be improper. In practice, controllers are often made proper by adding high-frequency dynamics, e.g. by multiplying with 1="s + 1m where " is small, and m is some integer. This works in most cases. However, it will not work for plants which needs R HP zeros or poles in the controller to make the closed-loop transfer function TV stable, and we therefore conclude that our lower bound on TV may not be tight in such cases. Remark: We have written may not be tight since numerical results using standard state-space H1 -synthesis in 8

M ATLAB, have shown that the bounds are tight even for plants which result in an unstable state-space H1-controller. Thus for some plants, in addition to the stable improper controller given by (24), there may exist an unstable proper controller which yields the same minimum value of kTV sk1 . Next, we consider the controller which minimizes kSV sk1 . T HEOREM 4 (K WHICH M INIMIZE kSV sk1 ). Consider the S ISO plant G with one R HP-zero z and Np  0 R HP-poles pi 2 C + . Then the feedback controller K which minimize kSV sk1 is given by 1 ,1 K s = G, (28) ms Ko s; Ko s = PQ s where

,1 z  V z  V ,1 s Qs = Bp ms ms ,1 s 1 , B s Qs P s = 1 , Bp s Qsm = Bz p


With this controller we have

(29) (30)

,1 z j  jV z j kSV sk1 = jBp ms

(31)

which shows that the bound given in Theorem 2 is tight when the plant has one R HP-zero. The comments following Theorem 3 also apply to the bound in Theorem 2 and to the controller given in Theorem 4.

Applications of lower bounds

The lower bounds on kTV sk1 and kSV sk1 in Theorems 1 and 2 can be used to derive a large number of interesting and useful bounds.

5.1 Bounds on important closed-loop transfer functions


Consider again the six transfer functions in (13), and the weighted complementary sensitivity function wunc T . For simplicity we assume that wP , wu, wunc , R and N are all stable minimum phase (or have been replaced by the stable minimum phase counterparts with same magnitude). From Theorems 1 and 2 we obtain: Output performance, reference tracking:

,1 z j  jRz j kwP SRsk1  R max jwP zj j  jBp j j -zeros, zj


HP

(32)

Output performance, disturbance rejection:

,1 z j  jG  j jwP zj j  jBp kwP SGdsk1  R max j d ms s=zj -zeros, zj


HP

(33)

Output performance, measurement noise rejection:

kwP TN sk1  R max jwP pij  jBz,1pi j  jN pij -poles, pi


HP

(34)

Input usage, reference tracking:

1 jwupi j  jBz,1 pij  jG, kwuKSRsk1 = kwuTG,1Rsk1  R max ms Rpi j -poles, pi


HP

(35)

Input usage, disturbance rejection:

,1 p j  jG,1 G  j j w  p  j  jB kwuKSGd sk1 = kwuTG,1Gdsk1  R max u i i z ms d ms s=pi -poles, pi


HP

(36)

Input usage, measurement noise rejection:

,1 p j  jG,1 N p j kwuKSN sk1 = kwuTG,1N sk1  R max j w  p  j  jB u i i i z ms -poles, pi


HP

(37)

Closed-loop sensitivity to plant uncertainty:

jwunc pij  jBz,1 pij kwuncT sk1  R max -poles, pi


HP

(38)

Note that we mainly have inherent limitations on (output) performance when the plant has R HP-zeros. The exception is for measurement noise, where the requirement of stabilizing an unstable pole may give poor performance. On the other hand, all the bounds on input usage are caused by the presence of R HP-poles. This is reasonable since we need active use of the input in order to stabilize the plant. This is considered in more detail in the next section.

5.2 Implications for stabilization with bounded inputs


Our bounds involve the H1 -norm, and their large engineering usefulness may not be immediate. In the following we will concentrate on the bounds involving input usage and we will use the lower bounds to derive and quantify the conclusion: Bounded inputs combined with disturbances and noise may make stabilization impossible. The input signal for a one degree-of-freedom (1-D OF) controller due to disturbance d, measurement noise n of magnitude N and reference r of magnitude R is

u = KS Rr ~ , Gd d , N n ~

(39)

Measurement noise. The transfer function from normalized measurement noise n ~ to the input u is KSN . Then from (37) with wu = 1

1 jBz,1pij  jG, kuk1 = kKSN sk1  R max ms pi N pi j -poles, pi


HP

(40)

Thus, to have kuk1

 1 for kn ~ k1 = 1, we must require jGmspij  jBz,1pij  jN pij for the worst case pole pi

(41)

(we have here assumed that N is minimum phase). That is: To keep the input magnitude less than one kuk1  1 we must require that the plant gain is larger than the measurement noise at frequencies corresponding to the unstable poles. 10

To better understand this statement, we will make use of the interpretation of the H1 -norm in terms of steady-state sinusoids. Consider the case when kn ~ k1 = 1 and assume that the lower bound in terms of kuk1 = kKSN sk1 in (40) is larger than one (i.e. (41) is not satised). In this case, no matter what linear controller we design, there will always be a sinusoidal noise signal

nt = nmax sin!0t; nmax = jN j!0 j ut = umax sin!0t + '

such that the resulting input signal has umax 1 (the value of ' is not of interest here). For a given controller K , the worst case frequency !0 may be chosen as the frequency ! where jKSN j!j has its peak value, i.e. jKSN j!0j = kKSN sk1. Disturbances. Similar results as those for measurement noise apply to disturbances if we replace N by Gd . From (36) with wu = 1 we obtain ,1 p j  jG,1 G  j jB kuk1 = kKSGdsk1  RHP max (42) i z ms d ms s=pi -poles, pi

To have kuk1

 1 for kdk1 = 1 we must require jGmspij  Bz,1 pi  jGd msjs=pi

for the worst case pole pi

(43)

That is: To keep the input magnitude less than one kuk1  1 we must require that the plant gain is larger than the gain of the disturbance plant at frequencies corresponding to the unstable poles. References. For reference changes with kr ~k1 = 1, we nd the same bound (42), but with Gd replaced by R. However, the implications are less severe since we may choose not to follow the references (e.g. set R = 0). Also, in the case of reference changes we may use a 2-D OF controller, such that the burden on the feedback part of the controller K is less. This is discussed in Section 7.

5.3 Combined R HP zeros and poles


It is well known that the combination of R HP zeros and poles imply peaks in the sensitivity and complementary sensitivity for S ISO systems which are larger than one. This has previously been quantied by Freudenberg and Looze (1988) in terms of sensitivity integral relations, but not directly in terms of kS sk1 and kT sk1 . However, from (20) and (18) we obtain ,1 z j  1 kS sk1  RHP max jBp j -zeros, zj kT sk1  RHP max jBz,1 pij  1 -poles, pi which are large if jzj , pi j is small. If all R HP zeros and poles are different by a factor of 10 or more, then the interaction between them are small. For a plant with one R HP-zero z and one R HP-pole p we obtain ,1 z j = jB,1 pj = jz + pj Mpz , jBp z jz , pj Peaks in S and T less than Mpz are thus unavoidable. For example if z=p = 10 gives Mpz = 1:22, whereas z=p = 1:5 gives Mpz = 5. By using the lower bounds we have derived and quantied the conclusion: Closely located R HP zeros and poles imply large sensitivity peaks. 11

5.4 Examples
E XAMPLE 1. The intention with this example is to show the engineering application of the lower bound on kKSN sk1 and to demonstrate the use of Theorem 3 to nd the feedback controller K which minimize kKSN sk1. We consider the unstable plant
1 Gs = s , p;

with R HP-pole at p. From (40) we have the following lower bound on the H1 -norm of the transfer function from normalized measurement noise n ~ to input u (we assume that N is minimum phase)

1 kKSN sk1  jG, ms pj  jN pj 1 ,1 In our case G,1 = s , p, G, ms s = s + p, Gms p = 2p, and the lower bound becomes kKSN sk1  2p  jN pj (44) The controller which minimizes k TV sk1 and achieves the bound (44) is given in Theorem 3. Rewriting KSN = TG,1N and by using V = G,1N we obtain Vmss = s + pN s, where we have assumed N to s,p . Thus, from Theorem 3 we obtain be stable minimum phase. Furthermore, Bz s = 1, Bp s = s +p

p  N p P s = s + p  N s


2

and

2p  N p s +p Qs = s , p  1 , s + p  N s

which gives

Remark: It seems like this controller has a R HP-zero for s = p, but this is not the case for its minimal realization since

p  N p s , p K s = s +2p N s , 2p  N p

s p  N sjs=p , 2p  N p = 0 For the special case where N s is a constant N s = N we get the proportional feedback controller ps , p = 2p K s = s2+ p , 2p As a numerical example, let p = 10, then 1 Gs = s , 10 and we must have for any stabilizing feedback controller K
 + 

kKSN sk1  20 jN pj Thus with k n ~ k1 = 1 we will need excessive inputs k uk1 1 if jN pj  jGms pj = 0:05. Assume that N s = N p = 0:05, then K s = 2p = 20. This controller gives a at frequency response, i.e. jKSN j!j = 20; 8!. Thus, at any frequency !0 the closed-loop response in u due to nt = 0:05 sin!0t; is ut = sin!0t + ' 8! So, the input ut oscillates between 1. The response in u and y due to nt = 0:05 sin4t is shown in
Figure 3.

12

0.5

ut nt y t

0.5

1 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

Time [s]

Figure 3: Closed-loop response at input with K = 20

u and output y of the plant G, due to nt = 0:05 sin4t (dashed),

E XAMPLE 2. In this example we consider disturbance rejection for a plant with one R HP zero and pole. Let

We see that the disturbance is of magnitude kd and enters at the input of the plant. Note that
 +z Gdms = k zkd s s+p

s,z Gs = k z s , p;

Gds = kd Gs

with

z = 2; p = 1

and

The factors involving the interactions between the R HP zero z and pole p become

1 zs+p G, ms = k s + z

and we nd that peaks in the sensitivity since

S and the complementary sensitivity T less than 3 are unavoidable + pj jz + pj = 3 kS sk1  jjz = 3 and k T sk1  z , pj jz , pj Since G has a R HP-zero, we have a bound on the H1 -norm of the closed-loop transfer function from disturbance d to output e = y , r ,1 z j  jGd ms j = jz + pj  2 jk  kd j = 2 jk  kd j kSGdsk1  jBp s=z jz , pj jz + pj and for k dk1 = 1, the output e will be unacceptable k ek1 1 for jk  kd j 0:5. Similarly, since G has a R HP-pole p we have a bound on the H1 -norm of the closed-loop transfer function from disturbance d to input u jz+pj ,1 pj  jG, 1 kKSGdsk1  jBz mspj  jGd ms js=p = jz,pj  jkd j = 3 jkd j
and for k dk1
=1

,1 z j = jB,1 pj = jz + pj = 3 jBp z jz , pj

the input usage will be unacceptable k uk1

1

when jkd j

1 3.

E XAMPLE 3. In this example we look at the effect of a R HP zero and pole in Gd . Let the plant be
5 Gs = 10s + 1 s , 1

where Bz s = 1 since there is no R HP-zeros in G. We consider the three disturbances

kd Gd1s = s , 10 :2s + 1 ;

kd Gd2 s = s + 10 :2s + 1

13

and

For disturbance d1 we must assume that the unstable pole at p = 1 is the same as the one in the plant G, such that it can be stabilized using feedback control. There is no R HP-zero in G, so we have no lower bound on kSGdk sk1. However, since G has a RHP-pole p there is a bound on kKSGdk sk1, and we nd that the same lower bound applies to all three disturbances k 2 f1; 2; 3g, since


kd s , 2 Gd3s = s + 10 :2s + 1s + 2

kd Gd1ms = Gd2 ms = Gd3ms = s + 10 :2s + 1


10 + 1 + 1 5

We obtain

1 kKSGdk sk1  jG, ms Gdk ms js=p =

 + 10 2 + 1

kd :s

s=1

11 6

 jkd j

6 Thus, for kdk1 = 1 and if we require k uk1  1 we need to have jkd j  11 6  0:55. encounter excessive plant inputs (for all controllers) if jkd j 11

 0:55. In other words, we may

Stabilization with input saturation

Our results provide tight lower bounds for the required input signals for an unstable plant. Can these bounds be used to say anything about the possibility of stabilizing a plant with constrained inputs (e.g. jutj  1; 8t)? Assume that we have found, from one of these bounds, that we need kuk1 1. That is, at some frequency !0 we need ut = umax sin!0 t, with umax 1. Will the system become unstable in the case where input is constrained such that jutj  1 8t? Unfortunately, all our results are for linear systems, and we have not derived any results for this nonlinear effect of input saturation. Nevertheless, for simple low order systems we nd as expected very good agreement between our lower bounds and the actual stability limit in systems with input saturation. Intuitively, this agreement should be good if the input remains saturated for a time which is longer than about 1=p, where p is the R HP-pole.

6.1 Examples
E XAMPLE 1
CONTINUED.

Consider again the plant


1 Gs = s , 10

K = 20 which minimizes kKSN sk1 when N is constant. jKS j!j = 20; 8!, from which we know that sinusoidal measurement noise
with the controller

With this controller we get

nt = n0 sin!0t ut = 20n0 sin!0 t + ' for any frequency !0 . Thus, for n0 = f  0:05 we have that ut = f sin!0 t + ', and for f 1 the plant input will exceed 1 in magnitude. The question is: what happens if the inputs are constrained to be within 1? Will the stability be maintained? We will investigate this numerically by considering three frequencies; !0 = 1 [rad/s], !0 = 10 [rad/s] and !0 = 100 [rad/s]. First, Figure 4 shows the response to nt = 1:01  0:05 sint !0 = 1 [rad/s], f = 1:01. We see that the plant becomes unstable due to the input saturation. Next, we consider !0 = 10 [rad/s]. In this case we do not
14
cause the input to become

ut
Constrained Unconstrained

0.5

y t y t
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

0.5

1 0

Time [s]

Figure 4: Closed-loop response at input u and output y of the plant G, due nt = 1:01  0:05 sint
1 0.5 0 0.5 1 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

ut
Constrained Unconstrained

y t y t

Time [s]

Figure 5: Closed-loop response at input u and output y of the plant G, due nt = 1:29  0:05 sin10t get instability with f = 1:01 and !0 = 10 [rad/s]. We nd numerically that we need to increase the magnitude of the sinusoidal noise to about f = 1:29 to get instability for this particular frequency. Figure 5 shows the response to nt = 1:29  0:05 sin10t ! = 10 [rad/s] and f = 1:29. Finally, as shown in Figure 6 we get instability with nt = 1:6  0:05 sin100t ! = 100 [rad/s] and f = 1:6. We experience that we have to increase the magnitude of the noise somewhat to get instability for sinusoidal measurement noise with frequency around the bandwidth and higher. However, we are still within a factor of two for a large frequency range for this particular plant. Measurement noise usually contain a large range of frequencies, which makes it even more probable that one loose stability of the plant if the lower bounds exceeds the allowable input range. Note that the control system designer seldom wants the input to saturate when stabilizing an unstable plant due to the possibility of loosing stability. So our engineering bounds are really applicable in practical controller design. As a nal simulation, Figure 7 shows the closed-loop response due to a step of size 1:01  0:05 in n. (1% increase relative to the limit which cause u to exceed 1). This input signal can be viewed as consisting of innite number of frequencies with decreasing magnitude, where the steady-state effect is the most important and can be viewed as a slowly varying sinusoid with !0 = 0 [rad/s] and amplitude 1:01  0:05. As can be seen from the gure, the unconstrained input exceeds 1 slightly. When the input is constrained to be within 1, stability of the plant is lost. E XAMPLE 3
CONTINUED.

Consider again the plant


5 Gs = 10s + 1 s , 1

15

1 0.5

ut
Unc. y t

0 0.5 1 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

y t

Time [s]

Figure 6: Closed-loop response at input (unconstrained input not shown)


1

u and output y of the plant G, due nt

= 16

:  0:05 sin100t,

ut
Constrained Unconstrained

0.5

y t

0.5

y t
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

1 0

Time [s]

Figure 7: Closed-loop response at input

ns = 1:01 

0:05 s

u and output y of the plant G, due to step in measurement noise,


=

In the simulations shown in this example, we have used the disturbance plant Gd

However, it does not really matter which Gdk one uses, except that the initial responses may be different. By using Theorem 3 with V = G,1 Gd , we obtain:

kd s , 2 Gds = s + 10 :2s + 1s + 2

Gd3

10s + 1 11 55 0:2s + 1 Vmss = k5d 0 :2s + 1 ; Vmsp = 6  kd; P s = 6 10s + 1 The H1 -optimal controller minimizing k KSGd sk1 becomes

, 1 ; G s = 5 Bz s = 1; Bps = s ; ms s+1 s + 110s + 1

kd Gdms s = s + 10 :2s + 1 ;


and

s+1 Qs = 49 6 10s + 1

6 the controller K1 results in k K1 SG sk = 1, and when k = 0:55 which is not proper. For kd = 11 d d 1 6 (0:55 is the value of k used in the simulations) kK1 SG sk = 1:008. We note that the specter of d d 1 11 K1SGdj! is at (constant). To get a realizable (proper) controller we add second order dynamics at high frequency to obtain the H1 -suboptimal controller
e 1 s = 49 0:2s + 110s + 1 K 11 0:01s + 12

K1s = 49 0:2s + 110s + 1 11

(45)

16

e 1 SGd with kd The H1 -norm of the closed-loop transfer function K e 1 SGd sk = 1:027; kK 1

= 0 55

is

for !

= 1 35

[rad/s].

To compare with a more traditional controller, which emphasize tight control at low frequencies, we also consider controlling the plant G using the feedback controller
4  10s + 1 K s = 0:s 0:1s + 12

2
= 0 55

(46)

With this K the H1 -norm of the closed-loop transfer function KSGd for kd

becomes

kKSGdsk1 = 2:845; for ! = 2:056 [rad/s]. e 1 SGd for K e 1 given by (45) is shown in Figure 8 The magnitude of the closed-loop transfer functions K together with the magnitude of KSGd for K given in (46). From the gure we see that forcing jKSGd j! j
10
1

Magnitude

10

10

10

10

10

10 Frequency [rad/s]

10

10

e 1 SGd (solid) Figure 8: Closed-loop transfer functions KSGd (dashed) and K

to be small at low frequencies, results in a peak in the medium frequency range (compare jKSGd j! j with e 1 SGd j! j in Figure 8). jK The non-linear constrained and the linear unconstrained responses to the unit step in disturbance d using e 1 given by (45) and the controller K given by (46), are shown in Figures 9 the suboptimal H1 -controller K and 10. From the simulations we see that the input saturates (it may be difcult to separate the unconstrained input from the constrained input in Figure 9, since the unconstrained input only slightly exceeds ,1), with the consequence that we loose stability of the plant for both controllers.

Two degrees-of-freedom control


u = K1 r , K2 y + n wP SGK1 , 1R
(47)

In this section we consider the 2-D OF controller where

(the 1-D OF considered above follows by setting K1 = K2 = K ). For a 2-D OF controller the closedloop transfer function from references r ~ to outputs z1 = wP y , r becomes (48)

We then have the following special lower bound on this transfer function. 17

4 3 2 1 0 1 0 2 4 6 8 10 12
Constrained Unconstrained

y t

ut
14 16 18 20

Time [s]

Figure 9: Responses in y and e 1 given by (45) input with K


4 3 2 1 0 1 0 2

u due to unit step in disturbance d for constrained (juj  1) and unconstrained

Constrained Unconstrained

ut

y t

10

12

14

16

18

20

Time [s]

Figure 10: Responses in y and u due to unit step in disturbance input with K given by (46)

d for constrained (juj  1) and unconstrained

T HEOREM 5. Consider the S ISO plant G with Nz  1 R HP-zeros zj and Np  0 R HP-poles pi 2 C + . Let the performance weight wP be stable and minimum phase, and let the closed-loop transfer function wP SGK1 , 1R be stable. Then for a two degrees-of-freedom controller the following lower bound applies kwP SGK1 , 1Rk1  RHP max (49) jwP zj j  jRms zj j -zeros, z Furthermore, the bound (49) is tight if the plant has one R HP-zero z , and the controllers K1 and K2 which achieve the lower bound (49) are given by 1 z   ,1 , w ,1sR,1 sw z R z  K1 = Bpz G, (50) P ms ms ms P m K2 = The controller given in Theorem 4, minimizing kSGsk1. (51)
R EMARK 1. The bound (49) is clearly a lower bound (both for 1-D OF and 2-D OF controllers). The important fact is that (49) provides a tight lower bound for a plant with one RHP-zero and with the 2-D OF controller given in Theorem 5. ,1 is stable and R,1 is stable. From Theorem 4 it follows that R EMARK 2. It follows that K1 is stable since wP ms K2 is stable.
j

The bound in (49) should be compared to the corresponding bound for 1-DOF controller (32): ,1 z j  jR z j jwP zj j  jBp kwP SRsk1  RHPmax j ms j -zeros, zj 18

(52)

The fact that the lower bound (49) is tight when the plant has one R HP-zero and 2-D OF is applied makes it possible to conclude that only the R HP-zero pose limitations in this case. Thus, with a 2-D OF controller there is no additional penalty for having R HP-poles in G when performance is measured as z1 = wP y , r . However, from (52) we see that the penalty for having both a R HP-zero zj and ,1 zj j  1 for a 1-D OF controller. R HP-poles is jBp

Discussion

Form the lower bounds on input usage (see Section 5.2) we can easily quantify how much measurement noise and the magnitude of disturbance we can tolerate to avoid that the input exceeds some prespecied limits. We nd this quantication appealing, and it should be useful for control engineers doing practical control design. We therefore used the term engineering bounds for this application of the lower bounds in the second part of Example 1. Here we will only stress that these bounds are of fundamental theoretical importance, and they are (in many cases) tight for the best possible controller. So the bounds are exact, i.e. these bounds are not rules of thumb. In the H1 -controller design procedure, the H1 -norm of some weighted closed-loop transfer function is minimized. It has been shown that the resulting minimization problem is a convex problem, which can be solved numerically for example by introducing Linear Matrix Inequalities (LMI) or using -iteration. In this paper we have looked at single closed-loop transfer functions which can be written as TV or SV . Practical H1 -controller designs are usually set up as a stacked transfer function consisting of several closed-loop transfer functions. Usually the sensitivity appears as a factor in one or more of the closed-loop transfer functions, which is the origin to the name mixed sensitivity. The controller designed will then reect a trade-off between the different requirements expressed in each of the closed-loop transfer functions. For example, it is common to put weight on both the output performance and input usage. This can be expressed as in the mixed S=KS H1 controller design where  w S P the problem is to nd the controller K such that the H1 -norm of wu KS is minimized, i.e.

min K maxfkwP S sk1 ; kwuKS sk1g 




wP S s wuKS s




Lower and upper bounds on the H1 -norm of the mixed S=KS sensitivity are

wP S s wuKS s

p  2 maxfkwP S sk1 ; kwuKS sk1g

which shows that our individual lower bounds on kwP S sk1 and kwu KS sk1 provide useful information also for practical H1 -controller designs. In the -iteration the H1 -minimization over the controller K is transformed to a convex minimization problem in the free variable , which is the H1 -norm of the closed-loop transfer function2. Most packages3 perform the -iteration using the bisection method. That is, given a high and a low value of (upper and lower bound) and a stabilizing controller, the bisection method is used to iterate on the value of . This modern controller synthesis shows one application of lower and upper bounds on the H1 -norm of general closed-loop transfer functions. The lower bounds derived in this
2 3

In M ATLAB Robust Control Toolbox is the inverse of the H1 -norm of the closed-loop transfer function. See M ATLAB, -tools or Robust Control Toolbox.

19

paper can be used as the low value of supplied to the -iteration. This follows since the largest singular value of matrix is larger than the largest element in the matrix. So, the largest lower bound on the H1 -norm of a S ISO transfer function in a larger multivariable stacked transfer function matrix still is a lower bound on the H1 -norm of the stacked closed-loop transfer function in question.

Conclusion
We have derived tight lower bounds on closed-loop transfer functions. The bounds are independent of the controller, and therefore reects the controllability of the plant. The bounds extend and generalizes the S ISO results by Zames (1981), Doyle et al. (1992) and Skogestad and Postlethwaite (1996) to also handle non-minimum phase and unstable weights. This allow us to derive new lower bounds on input usage due to disturbances, measurement noise and reference changes. The new lower bounds on input usage make it possible to quantify the minimum input usage for stabilization of unstable plants in the presence of worst case disturbances, measurement noise and reference changes. It is proved that the lower bounds are tight, by deriving analytical expressions for stable controllers which achieves an H1 -norm of the closed-loop transfer functions equal to the lower bound for large classes of systems. Theorem 5 express the benet of applying a 2-D OF controller compared to a 1-D OF controller when the plant is unstable and has a RHP-zero. The application of the lower bounds have been illustrated and the implications have studied in several examples. Nonlinear simulations have been used to nd the amount of noise and disturbances which in combination with input constraints, cause loss of stability for unstable plants. The results show good agreement between this amount of noise and disturbances and the corresponding values predicted by the lower bounds, in the examples studied.

References
Doyle, J. C., Francis, B. and Tannenbaum, A. (1992). Feedback Control Theory, Macmillan Publishing Company. Freudenberg, J. S. and Looze, D. P. (1988). Frequency Domain Properties of Scalar and Multivariable Feedback Systems, Vol. 104 of Lecture Notes in Control and Information Sciences, Springer-Verlag, Berlin. Havre, K. and Skogestad, S. (1997a). Effect of R HP zeros and poles on performance in multivariable systems, Accepted for publication in Journal of Process Control . Havre, K. and Skogestad, S. (1997b). Limitations imposed by R HP zeros/poles in multivariable systems, Proc. from ECC97, Brussels, Belgium. Kwakernaak, H. (1995). Symmetries in control system design, Trends in Control - A European Perspective, Alberto Isidori (ed.) pp. 1751. Skogestad, S. and Postlethwaite, I. (1996). Multivariable Feedback Control, Analysis and Design, John Wiley & Sons, Chichester. Youla, D. C., Bongiorno, J. J. and Lu, C. N. (1974). Single-loop feedback stabilization of linear multivariable dynamical plants, Automatica 10: 159173.

20

Zames, G. (1981). Feedback and optimal sensitivity: model reference transformations, multiplicative seminorms, and approximate inverses, IEEE Transactions on Automatic Control AC-26(2): 301320.

A Proofs of the results


Proof of Lemma 1. The rst identity in (11) follows since extracting R HP-zeros in the product AB in terms of the all-pass lter Bz AB , does not change the H1 -norm. The reason is of course that Bz AB  is all-pass for s = j!. To prove the latter identity, assume A has R HP-zeros which does not appear in the product AB , then ,1B . Similarly, if B has B has R HP-poles for those R HP-zeros, and these R HP-poles can be factorized as Bp R HP-zeros which does not appear in the product AB , then A has R HP-poles for those R HP-zeros, and these ,1A. We obtain R HP-poles can be factorized as Bp

,1 A Bz A Ams B,1 B  Bz B  Bms = B,1 A Bz A B,1 B  Bz B  Ams Bms AB = Bp p p p | z
|

=Bz AB

=ABm

Since, AB is stable then AB m

AmsBms , and it follows that

,1A Bz A B,1 B  Bz B  = Bz AB  Bp p ,1A Bz B  are the R HP-zeros of B which are not R HP-poles in A and Bz A Bp ,1 B  are the Note that, Bp R HP-zeros of A which are not R HP-poles in B . 2
Proof of Theorem 2. 1) Factor out R HP zeros and poles in

S and V . Lemma 1 gives fs S V s fs

2) Introduce the stable scalar function   = m ms  . 3) Apply the maximum modulus theorem to   at the R HP-zeros

kSV sk1 = kSmsVms sk1 = kSmVmssk1 where the last equality holds since S is stable, i.e. Sms = Sm .

zj of G.

4)

,1 zj j  jVms zj j jf zj j = jBp Note that f zj  is independent of the controller K if V is independent of K . Since these steps holds for all R HP-zeros zj , Theorem 2 follows.

5) 6) Evaluate the lower bound.

kf sk1  jf zj j ,1 zj  to get Resubstitute the factorization of R HP-zeros in S , i.e. use Sm zj  = S zj Bp ,1 zj Vms zj  f zj  = Smzj Vmszj  = S zj Bp Use the interpolation constraint (16) for R HP-zeros zj in G, i.e. use S zj  = 1.
(53)

,1 s is stable and the remaining matrices Proof of Theorem 3. The transfer function P is stable, since Vms ,1 p and Vms p are nite constant matrices. Consider 1 , Bz sP s which has a R HP-zero for s = p Bz
1

, Bz sP s = Bps 1 , Bz sP sm

or

1

,1s 1 , Bz sP s , Bz sP sm = Bp

We obtain

1 ,1 s Bz s Ko = M s = ,1 + Ls = GK s = Bp 1 , M s 1 , M s

21

where Since, S ,1
=1+

L = 1,1M we have
= 1

,1 p Vms p V ,1 s M s = Bz s Bz ms

S s

,1 p Vms p V ,1 s = 1 , Bz s P s , M s = 1 , Bz s B ms |z z


=P s

Consider S at the complex value p

,1 p Vms p V ,1 p = 0 S p = 1 , Bz p Bz ms


We can therefore write

,1 s1 , Bz s P s S s = Bps Qs , Qs = Bp 1 where Q is stable. Since, P , Q and G, ms are all stable we have that K is stable. Furthermore, ,1 s Bz s P sQ,1 sBp sQs = Bz sP s T = Bp
and we get

TV s
The H1 -norm of TV is

= =

=V s | ,1 p Vms p V ,1 s Bz V s B,1 V s Vms s Bz s Bz ms p , 1 , 1 Bz p Vmsp Bz s Bz V s Bp V s
z

,1 pj  jVms pj kTV sk1 = jBz (54) ,1 V s is all-pass for s = j!. Since the value of kTV sk in (27) is the same as since Bz s Bz V s Bp 1 the lower bound (18), this is the controller which minimize k TV sk1 . 2 ,1 s is stable and the remaining matrices Proof of Theorem 4. The transfer function Q is stable, since Vms ,1 z  and Vms z  are nite constant matrices. Consider 1 , BpsQs which has a R HP-zero for s = z Bp ,1 1 , Bp sQs = Bz s 1 , Bp sQsm or 1 , Bp sQsm = Bz s 1 , Bp sQs
We obtain
1+

Ls

= = =

,1 s Bz s Gms G,1 Ko = 1 + B,1s Bz s Ko GK s = 1 + Bp ms , p  , 1 , 1 , 1 ,1 1 + Bp s 1 , Bp s Qs Q s = 1 + Bp s Q s , 1 ,1 s Q,1 s Bp
1+

Which implies that

S s = Bps Qs Since both Qs and Bp s are stable, it follows that S s is stable. At the complex value z , we have ,1 z  and S z  = Bp z Qz  = 1 Qz = Bp It then follows that T = 1 , S is stable and has a R HP-zero for s = z T z = 1 , S z = 1 , 1 = 0 Since T s = 1 , S s = 1 , Bp sQs, we obtain ,1 s 1 , Bp s Qs T s = Bz s P s , P s = Bz
22

1 where P is stable. Since, Q, P and G, ms are all stable we have that K is stable. We get

SV s
The H1 -norm of SV is

= =

=V s | ,1 z  Vms z  V ,1 s Bz V s B,1 V s Vms s Bps Bp p ms , 1 , 1 Bp z Vms z Bp s Bz V s Bp V s
z

(55)

,1 z j  jVms z j kSV sk1 = jBp (56) ,1 V s is all-pass for s = j!. Since the value of kSV sk in (31) is the same as since Bp s Bz V s Bp 1 the lower bound (20), this is the controller which minimize k SV sk1 . 2
Proof of Theorem 5. We rst prove the lower bound (49). From Lemma 1 we have

kwP SGK1 , 1Rsk1 = kwP SGK1 , 1ms Rms sk1 since wP is stable and minimum phase. Consider the scalar function f s = wP SGK1 , 1ms Rms which is
analytic (stable) in R HP since the closed-loop system is stable. By applying the maximum modulus theorem to f s we get

kwP SGK1 , 1ms Rms sk1 = kf sk1  jf zj j

We get

jf zj j = jwP SGK1 , 1ms Rms js=zj = jwP zj ,1Rms zj j = jwP zj j  jRms zj j The second equality follows since SGK1 must have R HP-zeros for s = zj , since G has R HP-zeros for s = zj , and S and K1 must be stable (no R HP-poles in S or K1 to cancel the R HP-zeros in G). It then follows that SGK1 , 1 has no R HP-zeros for s = zj . We next prove that the controllers K1 and K2 given in Theorem 5, achieves this lower bound for the case when the plant has one R HP-zero z . From equation (55) in the proof of Theorem 4 we nd that SG with K = K2 (minimizing kSGsk1) and V = G becomes ,1z Gms z Bz s SGs = Bp
We obtain

SGK1s , 1

= = =

,1 sR,1 swP z Rms z m , 1 Bz s1 , wP ms , ,1 sR,1 swP z Rms z  , 1 , 1 Bz sBz s 1 , wP ms , 1 , 1 ,w sR swP zRms z
P ms

R HP-zero for s = z
|

which gives

,1 R wP z Rms z  wP SGK1 s , 1Rs = ,Bz R Bp ,1 R wP z Rms z  and R HP-poles in R may only Since wP SGK1 s , 1Rs is stable, so is Bz R Bp cancel against R HP-zeros in SGK1 , 1. It follows that

kwP SGK1s , 1Rsk1 = jwP zj  jRmszj and the controllers K1 and K2 given in (50) and (51) minimizes the H1 -norm of wp SGK1 , 1Rs.

23

You might also like