LIMITATIONS IMPOSED BY RHP ZEROS/POLES IN

MULTIVARIABLE SYSTEMS
K. Havre and S. Skogestad
Chemical Engineering, Norwegian University of Science and Technology, N-7034 Trondheim, Norway
 Fax: (+47) 73 59 40 80 and e-mail: skoge@kjemi.unit.no

H
Keywords: System theory, Achievable 1 performance,
- Wu -z3
Multivariable systems, R HP-zeros and poles, Stabilization.

-- +e- WP -z1
c
Abstract d~ = d 6
This paper examines the limitations imposed by Right Half Plane ?
(R HP) zeros and poles in multivariable feedback systems. The main Gd
k k
result is to provide lower bounds on WXV (s) 1 where X is S , r~ - R r- u- G -+ ?+ y- WT -z2
ym - K
q
c
SI , T or TI (sensitivity and complementary sensitivity). Previously q q

H
derived lower bounds on the 1 -norm of S and T are thus general- n~ - N n-+ c
ized to the case with matrix-valued weights, including bounds for ref- +
erence tracking and disturbance rejection. Furthermore, new bounds
which quantify the minimum input usage for stabilization in the pres- Figure 1: Two degrees-of-freedom control configuration
ence of measurement noise and disturbances are derived.

largest singular value of M (j! )

1 Introduction
It is well known that the presence of R HP-zeros and poles pose
fundamental limitations on the achievable control performance.
This was quantified for S ISO systems by Bode (1945) more
than 50 years ago, and most control engineers have an intu-
itive feeling on the limitations for scalar systems. Rosenbrock
(1966; 1970) pointed out that multivariable R HP-zeros pose
similar limitations.
In this paper we extend the work of Zames (1981), Skogestad
and Postlethwaite (1996) and Havre and Skogestad (1996b),
and quantify the fundamental limitations imposed by R HP-
zeros and poles in terms of lower bounds on the peaks (H1 -
norm) of important closed-loop transfer functions.
Notation. We consider linear time invariant transfer function
models on the form
kM (s)k1 , sup  (M (j!)) (2)
!
Closing the loop. In this paper we consider the general two
degrees-of-freedom (2-D OF) control configuration shown in
Figure 1. In the figure the performance weights are given in
dashed lines. We have included both references r and measure-
ment noise n in addition to disturbances d as external inputs.
The three matrices Gd , R and N can be viewed as weights on
the inputs, and the inputs d~, r~ and n
~ are normalized in magni-
tude. Normally N is diagonal and [N ]ii is the inverse of signal
to noise ratio.
The controller can be divided into a positive feedback part
from y (K2 ) and a feed forward part from r (K1 )
u = K1r + K2ym = K1 r + K2 (y + n) (3)

y(s) = G(s)u(s) + Gd (s)d(s) (1) The closed-loop transfer function F from [ r~T d~T n~ T ] to z
becomes
where u is the vector of manipulated inputs, d is the vector of 2
WP (SGK1 , I )R WP SGd ,WP TN
3
disturbances and y is the vector of outputs. We often omit to F (s) = 4 WT SGK1R WT SGd ,WT TN 5 (4)
show the dependence on the complex variable s for transfer Wu SI K1 R Wu K2 SGd WuK2 SN
functions. When we refer to zeros and poles and their direc-
tions we mean the zeros and poles of the plant G unless other- where the sensitivity S , the complementary sensitivity T and
wise explicitly stated. In order to be able to stabilize the pair the input sensitivity SI are defined by
(G, Gd ) we must require that all unstable modes in Gd is also
a mode of G.
S , (I , GK2 ),1 (5)
The H1 -norm of a stable rational transfer function matrix T , I , S = ,GK2 (I , GK2 ),1 (6)
M (s) is defined as the peak value overall frequencies of the SI , (I , K2 G),1 (7)
with positive feedback controller K2 . We also define the input where xR 2 C n and xL 2 C n are the eigenvectors correspond-
complementary sensitivity ing to the two eigenvalue problems

TI , I , SI = ,K2 G(I , K2 G),1 (8) AxR = pxR ; xHL A = xHL p

By setting K1 = ,K2 in the above equations, the one degree-
of-freedom (1-D OF) control configuration can be analyzed.
2 Zeros and poles in multivariable systems
2.1 Zero directions
D EFINITION 1 (Z EROS, MacFarlane and Karcanias, 1976).
zi 2 C is a zero of G(s) if the rank of G(zi ) is less than the
normal rank of G(s). The zero polynomial is defined as z (s) =
QN
i=1 (s , zi ) where Nz is the number of finite zeros of G(s).
z apply
Note, that the pole directions are normalized, i.e. kyp k = 1 and
kupk = 1. We will only consider distinct poles in this paper,
for the case where the pole p is not distinct refer to (Havre and
Skogestad, 1996c).
2.3 Interpolation constraints on , S T , SI and TI
C ONSTRAINT 1 (R HP - ZERO ). If G(s) has a R HP-zero at s =
z with output zero direction yz , then for internal stability of the
feedback system the following interpolation constraints must

The normal rank of G(s) is defined as the rank of G(s) at all s yzH T (z ) = 0; yzH S (z ) = yzH (12)
except a finite number of singularities (which are the zeros). TI (z )uz = 0; SI (z )uz = uz (13)

D EFINITION 2 (Z ERO D IRECTIONS ). If G(s) has a zero for In words, (12) says that T must have a R HP-zero in the same
s = z 2 C then there exist non-zero vectors, denoted the output output direction as G and that S (z ) has an eigenvalue of 1
zero direction yz 2 C l and the input zero direction uz 2 C m , with corresponding left eigenvector yz . In a similar way TI
such that yzH yz = 1, uH has R HP-zero with the same input direction as G.
z uz = 1 and
yzH G(z ) = 0; G(z )uz = 0 (9) C ONSTRAINT 2 (R HP - POLE ). If G(s) has a R HP-pole at
  s = p with output direction yp , then for internal stability of the
For a system G(s) with state-space realization C A B , the feedback system the following interpolation constraints must
D
zeros z of the system, the zero input directions uz and the state apply
input zero directions xzi 2 C n (n is the number of states) can
all be computed from the generalized eigenvalue problem
S (p)yp = 0; T (p)yp = yp (14)
H
up SI (p) = 0; uHp TI (p) = uHp (15)
A , sI B xzi = 0
    

C D uz 0 (10) The constraint (12) was first proved by Zames (1981), the proof
of (14) is given in (Skogestad and Postlethwaite, 1996). The
Similarly one can compute the zeros z and the output zero di- proofs of (13) and (15) follows similarly.
rections yz from GT .

3 Factorizations of R HP-zeros and poles

2.2 Pole directions
Bode (1945) states that the poles are the singular points at A rational transfer function matrix M (s) with R HP-poles pi
which the transfer function fails to be analytic. This means can be factorized in Blaschke products either at the output (sub-
that the system G evaluated1 at s = p, G(p), is infinite in some script o) or at the input (subscript i) of M (s) as follows
directions at the input and the output. This is the basis for the
following definition of input and output pole directions.
M (s) = Bpo (M )Mso (s); M (s) = Msi Bpi (M (s)) (16)

Mso , Msi – Stable (subscript s) versions of M (s) with the

D EFINITION 3 (P OLE D IRECTIONS ). If s = p 2 C is a dis- R HP-poles mirrored across the imaginary axis.
tinct pole of G(s) then there exist one output direction yp 2 C l Bpo (M ), Bpi (M ) – All-pass2 rational transfer function matri-
and one input direction up 2 C m with infinite gain for s = p. ces (all singular values are equal to 1 for s = j! ) containing
A B the   the R HP-poles (subscript p) of M (s).
For a system G(s) with state-space realization C D The all-pass filters are (Havre and Skogestad, 1996a)
pole directions yp and up for a distinct pole p can be computed
from (Havre and Skogestad, 1996c) Bpo(M (s)) = QNp (I + 2Re(pi ) y^ y^H )
i=1 s , pi pi pi (17)
Bpi (M (s)) = QNp 2Re( pi ) H
yp = CxR =kCxR k; up = B H xL =kB H xL k (11) i=1 (I + s , pi u^pi u^pi ) (18)
1 Strictly speaking, the transfer function G(s) can not be evaluated at s = p, B
2 Strictly speaking, our definitions of
po (M ) etc. give the inverse of what
since G(s) is not analytic in the point s = p. is usually called the all-pass transfer function.
Bpo(M ) is obtained by factorizing at the output one RHP-pole 4 Main result
at a time, starting with M = Bp1 o (M )Mp1 o where
In this section we derive general lower bounds on the H1 -
Bp1 o (M (s)) = I + 2Re( p1 ) H
s , p1 y^p1 y^p1 norm of closed-loop transfer functions when the plant G has
and y^p1 = yp1 is the output pole direction of M for p1 . This one or more R HP-zeros and/or poles, by combined use of the
procedure may be continued to factor out p2 from Mp1 o where interpolation constraints and the maximum modulus theorem.
y^p2 is the output pole direction of Mp1 o (which need not coin- The bounds are applicable to closed-loop transfer functions on
cide with yp2 , the pole direction3 of M and so on. A similar the form
procedure may be used to factorize the poles at the input of W (s)X (s)V (s) (28)
M (s). where X may be S , T , SI or TI . The idea is to derive lower
In a similar sequential manner, the R HP-zeros can be factor- bounds on k WXV (s) k1 which are independent of the con-
ized either at the output or at the input of M (s) troller K2 . In general, we assume that WXV is stable. The
,1(M )Mmo (s); M (s) = Mmi B,1(M (s)) (19) “weights” W and V must be independent of K2 . They may be
M (s) = Bzo zi unstable but it must be possible to stabilize all transfer func-
Mmo, Mmi – Minimum phase versions of M (s) with the tions through the outputs. This implies that the unstable modes
R HP-zeros mirrored across the imaginary axis. of W and V must also appear in G. Otherwise, the system is
Bzo(M ), Bzi (M ) – All-pass rational transfer function matri- not stabilizable.
ces (all singular values are equal to 1 for s = j! ) containing
T HEOREM 1 (L OWER B OUNDS ON k WXV (s) k1 ). Consider
the R HP-zeros (subscript z ) as R HP-poles.
a plant G with Nz R HP-zeros zj with input directions uzj , out-
We get (Havre and Skogestad, 1996a) put directions yzj and Np R HP-poles pi with input directions
Bzo (M (s)) = QNz (I + 2Re(zj ) y^z y^H ) upi and output directions ypi . Let the closed-loop transfer
j =1 s , zj j zj (20)
functions WSV , WTV , WSI V and WTI V be stable. For
Bzi (M (s)) = QN 2Re(zj ) H each R HP-zero zj the following bounds apply
j =1 (I + s , zj u^zj u^zj )
z (21)
,1 (G) V ) j k
k WSV (s) k1  k Wmo js=zj yzj k2
k yzHj V Bzi (Bpo s=zj 2
3.1 Some useful properties (29)
Assume that M ,1 exists, then k WSI V (s) k1 kBzo (W Bpi,1 (G))W js=zj uzj k2
k uHzj Vmi js=zjk2
(30)
Bzi (M ) = Bpo (M ,1 ); Mmi = (M ,1 ),so1 (22) and for each R HP-pole pi the following bounds apply
Bzo (M ) = Bpi (M ,1 ); Mmo = (M ,1 ),si1 (23) k WTV (s) k1  kBzo (W Bzo ,1 (G)) W j yp k
k ypH Vmi j k
s=pi i 2 i s=pi 2
Proof of (22) and (23). (31)
k WTI V (s) k1  k Wmo js=pi upi k2
k uHpi V Bzi (Bzi,1 (G)V ) js=pik2
M = Mmi Bzi,1 (M ) M ,1 = Bpo (M ,1 ) (M ,1 )so
and (32)
invert the latter to obtain M = (M ,1 ), ,1 (M ,1 ), and (22) fol-
so1 Bpo Proof : See Appendix A.
lows from inspection. (23) follows similarly from the output factor- R EMARK. The somewhat messy notation can easily be interpreted.
ization of R HP-zeros and input factorization of poles for M ,1 . 2 As an example take the last factor of (29): Factorize the R HP-poles at
B
the output of G into an all-pass filter po (G), take its inverse (yields
Repeated factorization: R HP-zeros) and multiply on the right with V (may add R HP-zeros if V
Bpi (Bpi (M )) = Bpi (M ) Bpo (Bpo (M )) = Bpo (M ) (24) is non-minimum phase), then factorize at the input the R HP-zeros of

Bzi (Bzi,1(M )) = Bzi (M ) Bzo(Bzo,1 (M )) = Bzo (M ) (25)

the product into an all-pass transfer function, take its inverse, multiply
on the left with yzj V and finally evaluate the result for s = zj .
Bzi (Bpo,1(M )) = Bpo (M ) Bzo (Bpi,1(M )) = Bpi (M ) (26)
Bpi (Bzo (M )) = Bzo (M ) Bpo(Bzi (M )) = Bzi (M ) (27) 4.1 Some important special cases
Proof of (24)–(27). The two first are obvious (repeating a factorization G has no RHP-poles. ,1(G)
Then Bpo = I and Bpi,1(G) = I ,
of the same type gives back the same all-pass filter). If we assume that and (29) and (30) become
M ,1 exist then for zi ( po
B B,1 (M )) in (26) we have
k WSV (s) k1  k Wmo js=zj yzj k2
k yzHj Vmi js=zjk2 (33)
Bzi (B,1 (M )) = Bzi (B,1 (M ,1 )) = Bzi (M ,1 ) = Bpo (M )
po zi k WSI V (s) k1  k Wmo js=zj uzj k2
k uHzj Vmi js=zjk2 (34)
where (22) is used twice with M set to M ,1 , the second part of (26)
and (27) follows similarly by also using (23). When M is not invert- The bounds (31) and (32) do not apply in this case.
ible the proofs are more complicated (need to go into the state-space
realizations of the factorizations to prove these relations). 2 G has no RHP-zeros. ,1(G)
Then Bzo = I and Bzi,1(G) = I ,
and (31) and (32) become
fact: y^p2 = pH1 o (M ) s=p2 yp2 . Here B s=s0 means the rational
B j j
k WTV (s) k1  k Wmo js=pi ypi k2
k ypHi Vmi js=pik2
3 In

transfer function matrix B (s) evaluated at the complex number s = s0 . Thus, (35)
it provides an alternative to B (s0 ), and it will mainly be used to avoid double k WTI V (s) k1  k Wmo js=pi upi k2
k uHpi Vmi js=pik2 (36)
parenthesis. It will also be used in combination with the absolute value for
j j j
S ISO transfer functions, i.e. m s=s0 = m(s0 ) . j The bounds (29) and (30) do not apply in this case.
Scalar stable weights. For the case with a scalar stable input 4. Closed-loop transfer function from references of
weight V (s) = v (s)
I , then the last terms in (29) and (32) magnitude R to weighted outputs for 1-D OF controller,
become WP (,SGK2 , I )R = ,WP SR. Select W = WP , X = S and
vm
z }| {
V = R, and apply the bound (29).
,
V Bzi (Bpo (G) V ) = vBzi (v) Bzi (Bpo
1 ,1 (G)) = vm Bpo (G) (37) Note that we can also look at the combined effect of the two
V Bzi (Bzi,1 (G) V ) = vBzi (v) Bzi (Bzi,1 (G)) = vm Bzi (G) (38) above by selecting V = [ Gd R ].
| {z }
vm 5. Closed-loop transfer function from measurement noise
of magnitude N to weighted outputs, ,WP TN . Select
Since v (s) is stable, R HP-pole/zero cancellation between v and W = WP , X = T and V = N , and apply the bound (31).
Bpo,1(G) and Bzi,1(G) are avoided.
Similarly, with a scalar stable output weight W = w(s)
I ,
the two first terms of (30) and (31) become 5.2 Input usage
wm The above provide generalizations of previous results, but we
z }| {
Bzo(W Bpi,1 (G)) W = Bzo(w) w Bzo(Bpi,1 (G)) = wm Bpi (G) (39) can also from Theorem 1 derive some new bounds in terms
,1 (G)) W = Bzo (w) w Bzo (Bzo
Bzo(W Bzo ,1 (G)) = vm Bzo (G) (40) of input usage. These new bounds provide very interesting in-
| {z } sights, for example, into the possibility of stabilizing an unsta-
wm ble plant with inputs of bounded magnitude.
Thus, for the case where both weights are scalar and stable, the The basis of these new bounds is to note that the transfer
function from the outputs to the inputs, K2 S , can be rewritten
as ,K2 S = TI G,1 . When G is unstable, G,1 has one or
bounds (29)–(32) become
k wSv(s) k1  jwm (zj )j
jvm (zj )j
k yzHj Bpo (G) js=zjk2 (41) more R HP-zeros, it is important that the bounds in Theorem 1
k wSI v(s) k1  jwm (zj )j
jvm (zj )j
kBpi (G) js=zj uzj k2 can handle the case when V = G,1 has R HP-zeros. Other-
wise, G,1 evaluated at the pole of G would be zero in a certain
(42)
k wTv(s) k1  jwm (pi)j
jvm (pi )j
kBzo (G) js=pi ypi k2 (43) direction, and we would not derive any useful bounds.
k wTI v(s) k1  jwm (pi)j
jvm (pi )j
k uHpi Bzi (G) js=pik2 (44) Some examples of bounds involving the input magnitude are
(see Figure 1):
S ISO-case. For the S ISO-case where all transfer functions are
scalars, some very nice and simple bounds can be derived, for 1. Closed-loop transfer function from disturbances d
further details see Section 6. to weighted inputs u, Wu K2 SGd = ,Wu TI G,1 Gd . Select
W = Wu, X = TI and V = G,1 Gd , and apply the bound (32).
5 Applications of lower bounds 2. Closed-loop transfer function from references r of mag-
nitude R to weighted inputs u, Wu K2 SR = ,Wu TI G,1 R.
The lower bounds on k WXV (s) k1 in Theorem 1 can be used Select W = Wu , X = TI and V = G,1 R, and apply (32).
to derive a large number of interesting and useful bounds. 3. Closed-loop transfer function from measure-
ment noise n of magnitude N to weighted inputs u,
5.1 Output performance WuK2 SN = ,WuTI G,1 N . Select W = Wu, X = TI
The previously derived bounds in terms of the H1 -norms of and V = G,1 N , and apply the bound (32).
S and T given in (Zames, 1981; Skogestad and Postlethwaite, Note that we can also look at the combined effect of the three
1996; Havre and Skogestad, 1996b) follow easily, and further above by selecting V = [ Gd R N ].
generalizations involving output performance can be derived. For further implications of these bounds see the S ISO case
1. Weighted sensitivity, WP S . Select W = WP , X = S discussed in Section 6.
and V = I . Note, this generalizes the previously found bounds
to the case with a matrix valued weight. For the special case 5.3 Tightness of bounds
WP (s) = I we derive from (29)
Numerical results indicate that the bounds in Theorem 1 are
kS (s)k1  R max
-zeros, zj
HP
kyzHj Bpo (G)js=zjk2  1 (45) tight, that is, in most cases there exists a controller which
achieves the lower bound under consideration. For example,
2. Weighted complementary sensitivity, WT T . Select consider the problem:
W = WT , X = Tand V = I . For WT (s) = I we derive from  Does there exists a controller which can stabilize a given
(31) unstable plant in the presence of disturbances, with k u k2 
kT (s)k1  R max kBzo (G)js=pi ypi k2  1 (46) 1 for all kdk2  1 ?
-poles, pi
HP The answer is “yes” if and only if there exists a controller for
3. Closed-loop transfer function from disturbances to which kK2 SGd k1 < 1. To test this bound, we can solve the
weighted outputs, WP SGd (see Figure 1). Select W = WP , H1 -problem
X = S and V = Gd , and apply the bound (29). min
K 2
kK2 SGd(s)k1
Alternatively, we can use 6.2 R HP-poles and stabilization

minK2 k K2 SGd (s) k1 = minK2 k TI G,1 Gd (s) k1
 maxpi k uHpi G,1 Gd Bzi (Bzi,1 (G)G,1 Gd ) js=pik2
Numerical results indicate that the inequality is always an
equality when the number of disturbances are fewer or equal
to the number of outputs.
5.4 2-D OF controller
 Input saturation combined with disturbances and noise may
make stabilization impossible.
The input signal for a 1-D OF controller is (positive feedback)
u = K2 S (Gd d + N n~ , Rr~) (50)
Disturbances. The input magnitude in response to a “worst-
case” sinusoidal disturbance jdj = 1, has the following lower
bound
jumaxj = kK2SGd (s)k1  (G,1 Gd )ms s=p
Mpz
 
(51)
For a 2-D OF controller the closed-loop transfer function from
references to outputs is bounded as follows where “ms” means that both the R HP-poles and zeros are mir-
rored into the left half plane.
k WP (SGK1 , I )R k1  k (WP )mo js=zj yzj k2
k yzHj Rmi (zj ) k2 Proof of (51). Follows directly from (44) with v = G,1 Gd and w = 1
when v is stable. For the case with v unstable (i.e. when G has R HP-
(47)
Proof : See Appendix A; note that this bound does not follow zeros) some more careful analysis is needed. 2
directly from Theorem 1. The bound in (47) should be com-
pared to the following bound for a 1-D OF controller (which E XAMPLE 1 Let
follows from Theorem 1) G(s) = (10s+1)(
5 Gd (s) = (s+1)(0kd:(2ss,+1)(
s,1) and 2)
s+2)
k WP SR(s) k1  k (WP )mo js=zj yzjk2
, 1 10 s +1 k
Then (G Gd )ms = 5 0:2s+1 and we get jumax j
d 
H , j(G,1 Gd )ms js=1 = 11 k ju j  1
k yzj RBzi (Bpo (G) R) js=zjk2 (48) have k  6  0:556 . In other words, we may encounter problems
1 d . If we require then we must
d 11
The bounds in (47) and (48) apply to all R HP-zeros zj in G. stabilizing the plant if kd > 0:55. Indeed, this is confirmed by
a particular nonlinear simulation example (see, Skogestad and
We see that for the 2-D OF controller only the R HP-zeros pose
Postlethwaite, 1996, page 192).
limitations.
References. For reference changes of magnitude R we find
6 Implications of the bounds (S ISO-case) the same bound (51), but with Gd (s) replaced by R(s). How-
ever, the implications for stabilization are less since we always
Some of the important implications of the bounds are most eas- may choose not to follow the references (e.g. set R = 0).
ily understood by studying the S ISO case. We consider a plant Measurement noise. More importantly, the same bound (51),
with a single R HP-pole p and/or a single R HP-zero z . Then but with Gd (s) replaced by N (s), applies to measurement
 noise of magnitude N (s), since u = K2 SN n~. More specif-
if G has a
Bp (G) , Bpi (G) = Bpo (G) = 1 + s2,pp = ss+,pp R HP-pole
ically, we have
jumaxj = kK2 SN (s)k1  (G,1 N )ms s=p
Mpz (52)
 

if G has a
Bz (G) , Bzi (G) = Bzo(G) = 1 + s,z = ss,+zz 2z
R HP-zero Thus, if we require juj  1, then we can expect problems
with stabilization5 if jGms (p)j < jN (p)j
Mpz (we have here as-
Also note4 that jBp (G) js=z = jBz (G) js=p = jjzz+ pj
,pj . We define sumed that N (s) is minimum-phase). That is:
(
jz+pj
 We expect problems if the plant gain is less than the mea-
def jz,pj if G has both a R HP-zero and pole
Mpz = surement noise at frequencies corresponding to the unstable
1 otherwise pole.

E XAMPLE 2 Consider the plant G(s) = 1=(s , 10) with the unstable
6.1 Combined R HP-zeros and poles mode p = 10. With measurement noise we have (Mpz = 1)
 jumax j = k K2 SN (s) k1  (G,1 N )ms s=p =
 
Closely located R HP-poles and zeros imply large sensitivity
 
jN (p)j
 (p+10)
peak.
1  = 20
jN (p)j


From (45) and (46) we get

kS (s)k1  Mpz and kT (s)k1  Mpz (49)
which is large if jz , pj is small. If the R HP-zero and pole are
different by a factor of 10 or more, then the interaction between
them are small, for example, p=z = 10 gives jjzz, +pj = 1:22.
pj
Thus, if we require juj  1, then we can expect problem with stabi-
lization if the noise is such that jN (p)j > 0:05. Simulations with a
proportional controller K2 = ,20 (positive feedback) and input sat-
uration 1 confirms this bound. For a higher controller gain the noise
may need to be increased by a factor of two to get instabilities.
5 We cannot rigorously determine nonlinear stability from a linear analy-

sis, but if the required linear inputs exceed the saturation bounds around the
4 jBp (G) js=zj , jBp (G) js=zj j. frequency of the R HP -pole, then nonlinear instability is highly likely.
6.3 R HP-zeros and output performance A Proofs of the results
From (41) with w = 1 and v = Gd we find for a plant with Proof of Theorem 1. Due to limited space we will only prove (29),
R HP-zero z the proofs of (30)–(32) follows similarly. To prove (29), factorize
B
R HP-zeros in W at the output, i.e. W = zo ,1 (W )Wmo. Define6
jemaxj = kSGd(s)k1  j(Gd )ms js=z
Mpz (53)
, , , 1 B
S1 T (GsoK2 ) , it then follows that S = S1 po ,1 (G). We get7
and we must require j(Gd )ms js=z
Mpz < 1 to guarantee , 1 , 1
jej < 1 for jdj = 1. Similarly, we have from (48) for refer- WSV = Bzo (W ) WmoS1 Bpo (G) V
,1 (W ) Wmo S1 (B,1 (G) V )m;i B,1 (B,1 (G) V )
= Bzo
ence changes with a 1-D OF controller |
po
{z }
zi po
jemaxj = kSR(s)k1  jR(z )j
Mpz (54) ,(WS1V )m
where we have assumed that R is stable and minimum-phase. Introduce the scalar transfer function f (s) = xH 1 (WS1 V )m x2 ,
The factor Mpz does not apply to the 2-D OF control configura- which is analytic (stable) in R HP. We want to choose x1 and x2 so that
j j j j
f (s) obtains maximum J (s) = maxkx1 k=1; kx2 k=1 f (s) . Then
tion, since (47) gives
2
the following relation holds for all zj R HP
jemaxj = k(SGK1 , 1)R(s)k1  jR(z )j (55)
k WSV (s) k1 = k (WS1 V (s))m k1  k J (s) k1  J (zj )
i.e. only the R HP-zeros pose fundamental limitations. Thus,
The first equality follows since (WS1 V )m only differ from WSV
the effect of the R HP-pole may be counteracted with a 2-D OF
with all-pass filters on the input and output. The first inequality fol-
controller, at least if we neglect the presence of model uncer-
lows since the largest singular value measures the maximum gain
tainty. of a matrix independent of direction, i.e.   (A)  k k
Aw 2 and
 (A) wH A 2 for any vector w with w 2 = 1. The second
k k k k
References inequality follows from the maximum modulus theorem. We get

Bode, H. W. (1945). Network Analysis and Feedback Amplifier De-

J (zj ) = kx k=1max xH W S (B,1 (G) V )mi js=zj x2
; kx2 k=1 1 mo 1 po
1
sign, D. Van Nostrand Co., New York.  kx k=1max xH W y yH S (B,1 (G) V )mi js=zj x2
; kx k=1 1 mo zj zj 1 po
Boyd, S. and Desoer, C. A. (1985). Subharmonic functions and per- 1 2
,1 (G) V )mi j k
= k Wmo js=zj yzj k2
k yzHj (Bpo
formance bounds in linear time-invariant feedback systems, IMA s=zj 2
J. Math. Contr. and Info. 2: 153–170.
Chen, J. (1995). Sensitivity integral relations and design trade-offs
2
in linear multivariable feedback systems, IEEE Transactions on Proof of (47). Factorize R HP-zeros of R and WP at the input and the
Automatic Control AC-40(10): 1700–1716. B
output, i.e. R = Rmi zi ,1 (R) and WP = zoB,1 (WP ) (WP )mo . Then
Havre, K. and Skogestad, S. (1996a). Directions and factorizations of
zeros and poles in multivariable systems. Internal report, Nor-
kWP (SGK1 , I )Rk1 = k(WP )mo (SGK1 , I )Rmi k1
wegian University of Science and Technology, Trondheim. Introduce the scalar function f (s) = xH
1 WP (SGK1 , I )Rx2 which
Havre, K. and Skogestad, S. (1996b). Effect of R HP zeros and poles is analytic (stable) in R HP since the closed-loop system is stable. We
on performance in multivariable systems, Proc. from Control’96, j j
want to choose x1 and x2 so that f (s) obtains maximum J (s) =
University of Exeter, UK, pp. 930–935. Also submitted to Jour- j j
maxkx1 k=1; kx2 k=1 f (s) . We get
nal of Process Control.
k (WP )mo (SGK1 , I )Rmi k1  k J (s) k1  J (zj )
Havre, K. and Skogestad, S. (1996c). Selection of variables for reg-
ulatory control using pole directions. AIChE Annual meeting, The first inequality follows since the largest singular value measures
Chicago 10–14 Nov., 1996, Paper 45f. the maximum gain of a matrix independent of direction, i.e.   (A) 
MacFarlane, A. G. J. and Karcanias, N. (1976). Poles and zeros of Aw 2 and  (A) wH A 2 for any vector w with w 2 = 1. The
k k k k k k
linear multivariable systems: A survey of algebraic, geometric second inequality follows from the maximum modulus theorem. We
and complex variable theory, International Journal of Control get
24: 33–74. J (zj ) = kx k=1max xH W (SGK1 , I )R(zj )x2
; kx2 k=1 1 P
Rosenbrock, H. H. (1966). On the design of linear multivariable sys- 1

tems, Third IFAC World Congress. Paper 1A.  kx k=1max xH W y yH (SGK1 , I )R(zj )x2
; kx k=1 1 P zj zj
1 2

= kWP (zj )yzj k2
k(yzHj G(zj ) K1 (zj ) , yzHj )R(zj )k2

Rosenbrock, H. H. (1970). State-space and Multivariable Theory,
Nelson, London.
| {z }
Skogestad, S. and Postlethwaite, I. (1996). Multivariable Feedback 0
Control, Analysis and Design, John Wiley & Sons, Chichester. = kWP (zj )yzj k2
kyzHj R(zj )k2
Zames, G. (1981). Feedback and optimal sensitivity: model ref-
erence transformations, multiplicative seminorms, and approx- 2
imate inverses, IEEE Transactions on Automatic Control AC- 6 Note that generally S = S
1 6
mi since only R HP-zeros in S due to R HP-
poles in G has been factorized in S1 , S may have additional R HP -zeros due to
26(2): 301–320.
R HP -poles in K2
7 The notation (WS V )
1 m means the minimum phase version of WS1 V
Internal reports and M ATLAB-software to compute the
bounds in Theorem 1 are available on the internet, see with the R HP -zeros of W factorized on the output and the R HP -zeros of S1 V
http://www.kjemi.unit.no/~skoge. on the input.