Copyright © IFAC Design Methods of Control Systems.

Zurich. Switzerland. 1991

THE GENERALIZED H2 CONTROL PROBLEM

M. A. Rotea

School of Aeronautics and AstronaUlics, Purdue Universiry,

West Lafayette, IN 47907, USA

Abstract. In this paper we consider the problem of finding a feedback controller such that
the controlled or regulated signals have a guaranteed maximum peak \'alue in response to ar-
bitrary (but bounded) energy exogenous inputs. \Iore specifically. we gi\'e a complete solution
to the problem of finding a stabilizing controller such that the closed loop gain from L 2 [0. x)
to L",[O. x) is below any specified level. \Ve consider both state-feedback and output feedback
problems. In the state-feedback case it is shown that if this synthcsis problem is soh·able. then a
solution can be chosen to be a constant state-feedback gain. \ecessary and sufficient conditions
for the existence of solutions as well as a formula for a state-fecdback gain that soh'cs this control
problem are obtained in terms of a finite dimensional com'ex feasibility program. The output
feedback case is reduced to a state-feedback problem by llsing a novel separation property.

Keywords. Multivariable control theory. linear optimal control. state-space methods. convex
programming.
The 11= and the 1.1 optimization theories also find ap-
INTRODUCTION
plication in robust stabilization problems. Indeed, since
Consider the causal finite dimensional linear time-invariant the I/." and the 1.1 norms represent the gain of a system
feedback system shown in the following figure where Q de- (i.e. they are induced operator norms). if a controller is de-
notes the plant and C the controller. l-iE-re, w dcnotes an signed to keep this gain is small, the resulting system will
remain stable in the presence of certain classes of modeling
uncertainty.
Recently. Wilson (1989) has introduced a number of
interesting "system gains". From the results of Wilson it
follows that, if the feedback system is internally stable and
there is no direct feedthrough from IV to :: then, T.w is
exogenous input, while z denotes a controlled signal. Let a bounded operator from L 2 [0"x) to 1.,,[0,(0), and its
T.w denote the closed loop map from the exogenous input induced norm is given by
tu to the controlled output z.
Many important control problems can be formulated as (1 )
the problem of finding a dynamic controller C such that
the feedback system is internally stable and the closed loop where T,w denotes tlJ(' corresponding closed loop transfer
map T.w is small in a suitable defined sense. Theories such matrix. The function f(.) is eit her t he maximum eigenvalue
as H2 , Hoc, and L1 optimal control are some of the con- or the maximum diagonal ent ry, depending on the "spatial"
crete and well-known examples of this point of view. The norm used on the cont rolled signal:: (I) (a precise definition
theory of H2 optimization (Wiener-IIopf or LQG theory) is of spatial norms is gi\·enlatcr). \ote that when z is a scalar
typically used when the exogenous input has known spC'c- signal. (I) reduc('s to t he familiar 112 norm of T,,, .. On the
tral characteristics (i.e. white noise) and one is interested ot her hand. if z is a \'cctor-\'alued signal t his induced norm
in reducing the power of the regulated variable z. Sf'e. for is no longer t he standard 112 norm. Furthermore, due to
example, Youla et aL (1976a 1976b). In many practical t he presence of t he maximum eigenvalue or t he maximum
situations thC' exogenous input does not haw fixed spectral diagonal entry. 11.11 need not 1)(' difff'rf'ntiahle.
characteristics, but instead its power spectrum is bounded. Let I denote a positiv(' constant. The synthesis problem
In this case, it s('('ms more meaningful to reduce the max- considered in this paper is th(' following:
imum (or worst-case) power amplification that can occur. GC1!(m/ized 112 COII/ro! p1'Obhm. "Find (if pos-
This point of \'iew has led to the theory of Hoo optimiza- sible) an admissible controller C such that the
tion introduced by Zames (1981). For survey papers in this closed loop system is internally stable and the
area the reader is referred to Francis and Doyle (1987) and induced norm (1) (t hp gain from L2 to L",) sat-
Khargonekar (1990). It should be noted that the Hoo frame- isfies 111-;" 11 < ,)."
work may also be used when one is interested in reducing
the worst-case energy of the output when the input is u n- The importance of this synthesis problem should be clear.
known but has bounded energy. Finally, the theory of L1 It guarantees t hat whenever t Iif' exogenous input w has
optimization arose in an attempt to deal with time-domain bounded energy. the controlled output z has a specified
specifications such as keeping the maximum (or worst-case) maximum peak \"alue. As in the standard H2 control prob-
excursion of the controlled variable small when the exoge- lem, the performance measure (I) is a function of the in-
nous input is unknown but has bounded amplitude. This tegral of the "square" of t he closed loop transfer function.
problem was introduced by Vidyasagar (1986) and Dahleh This motivates us to call the abm"e synthesis problem the
and Pearson (1987a 1987b) obtained an elegant solution for gfneralized 112 con/1'01 ]lmb!em.
it.

77
 Previous research on this class of synthesis problems extension into the right half plan('. The norm on this 'pac('
includes the work of Wilson, Grimble, and Zhu and Skel- is defined in the usual way and it is denot('.! by 11·lll'
ton. Wilson (1990) showed that the stationary Kalman This conference paper is a sumIllary of a re('ent 1"'1)('1'
filter is optimal when filtering performance is measmed by by the author (Rotea. 1991). \\'here flll'tlwr clptails ma\' 1)('
the system gain (I). In this case T,w should be interpreted found.
as the Illap frolll noise to estimation eITor. 1'11(' "control"
interpretation of this result is that the Kalman filter gain TilE GENERALIZED //2 CO:\THOL 1'1l01l1.E\,1
is opt imal wit h respect to the cost (I). when "full out-
put injection" to the state equation is a\'ailable. Grimble In this section wc define the !!,('n('raliz(,d J/ , cost. This
is a matrix-valued performan(,c measure whcN' maxilllulll
(1990) established a connection between the standard LQG
eigenvalue or maximum diagonal cnt r!' is t II<' ill("l<'c<l lIorlll
problem and a "dual" version of the generalized Ih con-
introduced in (I). As it lI'ill becolllc c1car. t IJ(' generalized
trol problem. In om context, the results of Grimble show
H2 cost is most useful for stating our r('sldts. Tltis s(,ction
that an LQG controller is also optimal for the generalized
concludes with a pr('cise formulation of tlte g('I1<'raliz('d J/ ,
fl2 control problem for a plant with "larger weighting" in
control problem.
the regulated variables. Zhu and Skelton (1990) also ap-
proached the genC'l'alized H2 problem via the solution to Analysis
an LQG problem. They proposed an iterative scheme for
Consider a finite-dimensional lill('M til11<'-ill\'ilri"lIt ,,'st('111
finding "LQG weights" so that a solution to the LQG prob-
lem also soh'es the generalized 1/2 problem. There are no described by the state-span' modcl:
guarantees that this scheme will conwrge. lloth Grimble T ._ { .i· = F.r + (;11' . .1'(0) = O. I:!)
and Zhu and Skelton considered only the case in which the .- 2 = H.r + JtI'.
maximum C'igen\'alue is used in tll(' definition of the induced
where the matrices F, (,'.11, and J aI'(' rcal '111<1 of CUIIlI',tl-
nonll.
ible dirm'IlsioIls. Suppose that T is int('1'1lally ,Ial>l". i.e. ,..
In this paper wc give a complete solution to the gen-
is a stability matrix. Lp( T,,,, d('notc tllc trallsf('r Illatri~
eralized fl2 control problem. first. \\'C' show that when
from U' to ::. Recall that IIT,,,,lh < x if and Old." if.1 = n.
the state of if (or the state and the exogenous input) is
and i1l this case we Illay define'
available for feedback. sialic slalc-fecdback gains offer the
best possible performance for the generalized 1/ 2 synthe- (:l)
sis problem. That is, in til(' state-f('edback case, dynamic
compensators buy nothing extra as far as the generalized
The gcncralized H2 pc/fol'mann 111((1,,'11'( (01' co,,1) fM tl)('
1/1 synthesis problem concerns. further, necessary and suf-
linear time-invariant system T is (\C-filwd a, folloll's:
ficient conditions for the solvability of the generalized 112
synt hesis problem with state-feedback, as lI'ell as a formula (I)
for a solution. arc obtained \'ia a finile-dimensional con- where for 7' = 2, F 2 (S) = .'I; and if /' = x. 1",,(8) =
t'Lr frasibilily program. In the case of output feedback, it.
diag(S). Note that Jr(L,.) 2: 0, and that if .1,(T,") is
is shown that generalized 112 controllers can be chosen to finite, i.e. Amar(Jr(T,w)) < 00, thell wc Illust ha\'l' .J = 0
be a combination of a standard /,'almal1 ]illfl' and a slalc- in (2). As in the definition of the III nonll. J,.(T",.) is
feedback gain for the generalized J1 2 synthesis problem of a defined in terms of the integral (3), this mot i\'ates the ni\llH'
suitahly constructed auxiliary plant. "generalized H2 cost."
While the approach taken here is somewhat similar to Wilson (1989) showed that tile mat rix-\'aluf'd perfor-
the approach of Bord et al. (1988) in that th",y may also mance measure defined in (4) giws risc to thc folloll'ing
reduce the generalized Il2 synthesis problem to a convex system gains:
optimizat ion probkm, there are significant differences be-
tween O\ll' results and those of Boyd et al. (1988). In par- VAmax(Jr(T,w)) =
ticular, we reduce t he generalized 112 cont roller synthesis
sup {11::11."".r/111I'112 : /I' E {2[D.xl./I· 'I o}. (.i)
problem to a con\'('x optimization problem over a subset of
q x 11 and 11 x 11 symmetric real matrices, where q and 11 where
are respecti\'ely the control input and the state vector di- IIZlloo.r := sup{II::(t)II, : I E [0. x)},
mensions. By comparison, the results of Boyd et al. (1988)
and the r-spatial norIll is gi\'f'n by
applied to the present problem would reduce it to a convex
optimization probleIll over an infinite-dimensional space of if/' =:!
stable transfer functions.
if ,. = x.
The notation used in this paper is filirly standard. For a
given matrix A, A' denotes its transpose and A- its conju- where s := dim(::). The subscript ",." in (I) st ands for t II<'
gate transpose. If.-l is square, diag(A) denotes the diagonal spatial norm used on the output 2(t).
mat rix formed lI'ith the diagonal entries of A. If A = A' When Jr(T,w) is finite (i.I'" in (2). J = 0) it call 1)('
is real, its maximum eigenvalue and diagonal entry are de- easily computed in time domain as follo\\'s. Let C d(,l1ot(,
noted by Ama,(A) and dmar(A), respectively. If A and B the controllability gramian of the pail' (F. (;). That is. L-
are real symmetric matrices, A 2: B (resp., A> B) denotes is the unique solution to
A - B positive semidefinite (resp., definite). Given a square FLc + LeF' + CC' = o.
matrix A, the matrix-\'alued function
Then
if r = 2
if 7' = 00
The following result provides an altel'llati\'e characteriza-
is extensively used throughout this paper. Linear time-
tion for the generalized H2 performance measure J,.(T,".)
in\'ariant systems are described by state-space models and
that will become useful later. for a proof of this result seC'
they are denoted by "script" symbols while the cOl'1'espond-
ing transfer matrices are denoted by "print" symbols. For Rotea (1991).
instance, if denote a system with transfer matrix C. The Lemma 1 Conside7' Ihc syslem T dffinfll in (.3) and Id
I-lardy space H2 consists of matrix-valued functions that T zw denote the Im1!sffl' mah·i.r f7'Om tt' 102. I,d Q = Q' > 0
are square integrable on the imagiuary axis with analytic be given. Then, T is inlernally slable and J,(T,,,,) < F,(Q)

78
 = 0 and Ihc1'c
if and only if J c:risls }., = }" > 0 sllch Ihal denote the generalized H2 cost cOlTesponding to the feed-
back system in Fig. 1.
FY + Y F' + CC' < O. (6)
F,(HY H' - Q) < O. (7)
STATE-FEEDBACK / FULL INFOIUIr\TION
Synthesis In this section we consider the case in which the plant to
Consider the finite-dimensional linear time-ill\'ariant feed - be controlled is gil'en by a state-spiICe model where either

w®z
back system depicted in Fig. I. where the plant 9 and the

u y

C
y
the state vector or the state and the exoge nou s input arc
measured (full information structure) . We show that the
generalized H2 synthesis problem wit h dynamic full infor-
mation feedback has a solution if and only if the gener-
a.lized H2 synthesis problem with mC1110ryless (consta nt)
state-feedback has a solut ion. lIence. for full information
and/OI' state-feedback problems. generalized 112 controllers
Fig. I. The synt hesis fram(,lI'ork. can be chosen to be a cOllslanl sta lc-fadback gain. This re-
sult is analogous to the cOlTcsponding r('suits for the LQR
controller Care gil'en by some stale-space mod('ls. The sig- problem (Kalman, 1960), the standard 11,,,,, control prob-
nal w denotes the exogenous input I·CctOr. II'llile z denotes lem (Khargonekar et ai., 1988; and Doyl", et ai., 1989).
the controlled output \·ector. The signals 11 and y denote and more recently for a class of mixed lid 11 '" problems
the control input and the measured out put. respect il·('ly. (I~hargonekar and Rotea. 1990), This section concludes
The transfer matrices of the plant and the controller arc with a convex optimization approach to the constant state-
denoted by C and C , respectil·"ly. \\'e denote the closed feedback problem.
loop transfer matrix by T,u" A cont rollcr C is called ad-
missible (for the plant g)
if C inl<'rnally stabili z('s the plant \Ve consider the generalized 112 sy nthesis problem for
the following plants (see also Fig. I): the "state- feedback
9. The set of all admissible cont rollers C for t hc plant 9 is
denoted by A(Q). Here "A" stands for ""dlni ss ible'·. Note plant", gil'en by the state-space model
that A(9) '" 0 if "nd only if 9 is stahil iz<lblc from 11 and X = A,.T + B1w + B2ll
detectable from y . 9,j:= z = C.r
{ y = x,
+ Du (8)
Let Q denote a posilil'e definite real ,,'nlnH'tric matrix.
The generalized 112 controlkr synt hesis ,;1'01>1"111 is defined
as follows: and the "full information plant" gil'en by

" Find (if poss ible) a controll er C E A(9 ) such x=Ax+BIW+B2U

that J,(T,w) < F,(Q) (I' =:1 or x) ." 9j,:= z=Cx+Du (9)
{ y = [lJ w']' .
Although no explicit frcqucncy d('pendent II'cights lI'erc in -
troduced in the abol'e synt llf'sis probl('nl. it is assumed that All the matrices introduced in (8) and (9) are real and
all weighting function s hal'(' been absorbed in th(' general- of compatible dimensions . As before, we let C'j and C j ,
ized plant 9. Note also that C E A(9) soh'es thi s sm thesis denote the transfer matrices of (8) and (9), respectively.
problem if and only if the closed loop syst(,1ll is il;ternally The subscripts "sf" and "fi" denote "state-feedback" and
stable and it satisfies the di st urbanc(' at t('nllation con dition "full information" structure, respectil'ely. Note that the
state-feedback and full information plants hal'e the same
dynamical equation. The only difference between them is
for some constant Q < " el'('I"y 1/' E 1,2[0, (0), and zero in the measurement equation. Note also that there is no
initial conditions. This is a simple consequ('nce of t he signal feed through term from the exogenous signal w to the con-
interpretation of the generalized //2 cost J"(T,,,,) gil'en in trolled output z. Although it is possible to include this
(5). Even though t\1(, constant lI'('ight F,,( Q) cou ld hal'e term, we have chosen not to do so in order to keep the
been absorbed in t he generalized plant 9 11'(' hill'e chosen presentation as simple as poss ible.
not to do so since this notation is most appropriate for Theorem 1 Conside,' Ih e syslcms 9 ,j and 9 j, defined
stating our results. in (8) and (9), ,·especlivc/y. LcI Q = Q' > 0 be given
In the next section wc will be interested in memory- and consider Ihe sets of admissible cantrollcrs A(Qj,) and
less (i .e. static) controllers for solving the abol'e syn thesi s Am(9,j). Then, Ih e following slatemenls a,'e (quit' alenl :
problem. In such a case, the generalized 1-/ 2 synthesis prob-
lem is defined in the following way. Define first the set of
/. There exists C E A(Qj,) such Ihal J,(Cj"C) < F,(Q).
memory less stabilizing controllers A m(Q) := {C E A(Q) : 2, There exisls JC E AmW ,j) sltch Ihal J,,(C,j, II) <
C E HqX P } , where q = dim(u) and p = dim(y) (sce Fig. I). Fr(Q).
Then , the memory less synthesis problem is:
"Find (if possible) a controller le E A m(9) suc h Due to the structure of 9 j, and 9 ,j it immediately
that J,(T, w) < Fr(Q) (r = 2 or (0)." follows that second statement in Theorem I implies the
first one. For if a memory less state-feedback controller JC
In Am(.), the subscript "m" stands for memoryless con-
satisfies Condition 2, then the full informat ion controller
trollers. (In this case it is certainly possible to identify C
C := [JC 0] satisfies Condition I. Theorem I shows that
with C and think of Am(9) as a set of matrices. Howel'er ,
the converse is also true.
for consistency of notation. we will use C when we \I'ant to
Proof Suppose that Condition 1 hold s, and let the full
think of the static gain as a system and C when we want
it to be interpreted as a matrix! ) information controller C be given by
Remarks on notation. Consider the feedback sl'ste m shown ~ = A c~ + BelT + B c2 w
in Fig. 1. Given a plant 9 and an internally st~b ilizing con- {U = Cc~ + Del X + Dc2 w .
troller C, the generalized JI 2 cost of t he closed loop system
The closed loop system cOlTesponding to the interconnec-
is a function of the transfer matrix T,u' only, Since T,w
tion of 9 j, and C is given by
depends only on the individual transfer matrices C and C.
in the sequel we let ~ = Ft/J + Cw
{ z = Ht!, + Jw,
J,(G, C) := J,(T,w(G, C))

79
where F, G, H, J are easily compuled rrom the s tale-s pace such Ihal J,(G',j, l\ ) < F,( Q ) if alld only if lI, c,'e e,risls
models or the plant and th e controller. ( 1\', Y) E HQxn X Hnxn sTleh Ihal Ih e fol/oll'ing eondilioll is
Since J,(Gj;,C) < F,(Q) , usi ng Lemma I 11'1" may nOli' salisfied:
conclude that J = D De2 = 0, and that there exists l" = Y = Y' > 0
Y' > 0 such that L(II', y) < 0 (1 7)
{ F,(,\I ( \\', l ')) < F,(Q),
L := FY + l" F' + GG' < 0, (10)
F,(HY H' ) < F, (Q), ( 11 ) ,l/o,.coI'CI', if Ihis condilion holds, Ihe slal e-feedback gain
Using this matrix Y we construcllhe rollowing memory- 1\0 := 1\')"-' salisfies " '0 E A ,,(t;;,j) and J ,(C',j ,l,o) <
less controller K. ror (;,j, Lel n = dim (,r) and /l e = dim(O, F,(Q),
and partition Y and L accordin g to plant and controller Due to space limitation s we will not gi,'(' a proor or this
dimensions, That is, theorem, The intuiti o n behind Theorem 2 is to repl ace th('
sea rch OW l' state- re('dbac k gains I,' E A ", (9, j) by a search
0\'1"1' the set of matrices defined by condition ( 17 ), This is
done by using Lemma I in combination lI'ith the chang(' or
where dim(l 'tl = l1xll,dim(12) = llxll e,dim(l3) = ll eX /l c' "iuiables '" = II'l '- ', \\'hf'l'c l ' is a ,;ollltion to the lin car
and similarly ror L. :':ote that l', > 0, }:j - l 'l l',-'l i > 0,
matrix ineqllality th at enro rces intcl'II ill stilbi lit y (cf. (6)),
and L, < 0, Define th e st ate- reedbi\ck g~in A sim ilar change or variilblcs wa s lI,;cd by l3e rtll' ssou et a l.
1\:=De, +(CY;+Dc1U;)l ',- I, ( 12) ( 1989) in a diffcr('nt contcxt. For rllrthf'l' d('tails, sce Rotra
(199 1),
Th e closed loop system resulting rroll' the intcrconnection
At first sight. Theorem 2 does not S(,('1ll ,'cry attracti\'c,
or (;,j and K,' is g i\'f'n by
Indeed, \\'hile thc problem or deciding whrt her t 11('1'(' ('xists
i = Fm'\' + [J,Il ' ,,- E A.,,(t;;,j) suc h that J"(C',j, /,') < I'~,(Q) can be tackled
{ z = llm.r , "ia non linear programming with on ly th(' reillmatrix I,' as
thc decision \'ariablc , the probkm or decid in g wll('t 11('1' (17)
where Fm := A + [J11,' , 11", := C + 0 /,', and the gain I,'
is gi\'en by ( 12 ), Simple algebraic lI1anipulat ion, ,how that has a rea l soluti o n ill\'oh'es a search m'c r a larger SpiKe, ;\ot
on ly lI'e mus t find the ,n a t rix I I ' (whos(' dilll('nsion ('qllill,
the "( I, I )" block or t he mat rix /, in (10) sat isfi('s
that or 1,') but also thc posit ive definitc matrix l" ,
+ lJ';', + 13 1J; + IJ lnclO~IJ]; ,
L, = Fml ", 1 (1:3) The next result shows that this o"e rparam et ri za ti o n is
Using the ract that J = DD" = 0, it is easy to "rr iry th~t most use rul since the se t de fin ed by (17) is CO ll\'ex, ( On
the ot her hand, th e set of admissible stati c slate-reedback
Hm y,I!~, = IIl ' II' - DC(}:) - l '; l ',- 'l 'I)C; J)', (1 ,1)
ga in s, A m (9 ,j), is no t necessari ly con\'ex,) Therefore, find-
Since Y3 - l ; l ',- 'l i > 0 and I" < 0, lI'e conclude 1'1'0111 ing a pail' or rea l matrices ( W, Y) that satisfies the condit ion
(11), (13), and ( 11 ) that t he> posit i,'(' ddillitr lI,atrix l ', in Theorem 2 is equivalent to a conl'e,r feasibilily p,.oblu1!,
satisfies This m ea ns that there are e fficient numerical algorithms
to solve it. A numb e r or good a lgor it hm s to address t hi s
Fm}" + Y, F~" + 13,13; < 0,
problem may be found in 130yd and Yilng (1989),
F, (If m Y,II;,,):S F,,( lIl ' I!' ) < F,,(Q),
Lemma 2 ConsidCl' Ih e plant (; ,j defined in (8), LcI L(,)
From these inequ a lit ics and LCll1ll1 a I , it rollows that F", is a
and .1/(,) denole Ih e mal,.ir-t'allledfunclion_' d(fined in ( 15)
stability matrix (i,r, '" E A ", (t;;,j)) and that J , (G'j' I,') <
F,(Q ), which completes thc proof. _
and ( 16), I'cspeclircly , Lcl Q = Q' be giro/. LcI <I> denole
Ih e subsc/ of HQxn X H,,,n of al/ pil i,., (I I', l ') slIeh Ihal (17)
Although Th eore>m I pro"ides a nicr st rllct mal illror ma -
holds, Th en, <I> is COIII'er,
tion , in that it tell s us that ror rllll inro rlllatioll rcedbac k or
sta te-feedback , ge> ne>rali zed 112 cant roll ers ca ll 1)(' chosrn to The proof of thi s lemma is rath er lo ng and it will be
be a constant state- reed back ga in, this t h('orcll1 dol'S 1I0t say ommitted , The key idea s are ilS rollo\\' s, First not(' that
whe th er such a gain exists and/or how to compute it, \Ye the map L(,) defined in ( 1,5) is ilffi ne, Ilrn ce, thc srt
will de\'elop a cOl1\'ex programmill g approach ror soh'ing t hc {(I\', l") : Y = l " and L( II'. l') < O,} is ob\' iously con-
stat ic state-reed ba ck proble lll. \I orc sprcifically, gi,'rn t hc \'ex, To establish Lemmil 2 it would suffice to show thilt
plant (; ,j defilled in (8) , lI'e gi,'r a nrcrssiI\'" and suffi cie nt the "Q-I evel set" or the map ,\1 (,) definrd in ( 16 ), i,f', lhr
condit ion ror the existence or an admissible stat e- reedb ac k set {(II ', Y) : Y = y' > 0 and " ~( '\/ ( lI'.l ')) < F,(Q)} is
gain J\ such that J,(G,j, J\ ) < F,(Q), I' = 2 or 00, This COll\'ex , The author has show n that this is thr case, For
condition is gi\'en in term s or a set or matrix inrqualities, rurther details see Rotea ( 1991),
We also show that this set or matri x inequalities defines a To illu st rate ho\\' the resu lt s or this spct ion may be us('d
convex subset of a finit e dimensional space of real matri- to sol\'e the generalized J-[2 sy nth esis problem ror the statc-
ces , Hence, the pro blem or chec king whether this spt or reedbilck plant g,j, le t us consider the CilSf' l' = 2 ( i,f' , thf'
inequalities has a solu tion is readily so h'ed by us ing con\'ex spat ial norm on .;(t) is the eucl id ean norm), Gi\'en a real
programming techniqu es, Furth e rmore, the construction of positi\'e definite mat rix Q, from Theorems 1 and 2, we kn o\\'
a real matr ix /{ such that J,(G,j' /{ ) < F,( Q ) is straight- that there exists C E A(t;;,j) (poss ibly dynamic) such that
forward from our necessa ry and sufficient cond it ion, J 2 (G',j, C) < Q ir and on ly ir the matrix in equalit ies
With reference to the plant (; ,j defined in (8), let n =
AY + YA' + 8 2 1\' + II " B~ + B,B; < 0
dim(x) and q = dim(Il), Let ( II'. Y) E RQXn X Rnxn be
(C Y + OIVP'-' (C Y + OIl ')' < Q ( 18)
given and define { Y = }", > 0,
L(II', Y) := AY + YA' + B211' + IV'B; + B,B;, ( 15 )
ha\'e a real solution ( 11 ', }"), In this case, the real malrix
If Y is nonsingular, we define also /\' := IVy-I is a solu tion to the synt hes is problem ,

M(\II, Y) := (C Y + DW)y-' (C Y + 011' )', (16) A little bit of work using Schur comp lements shows that
( 18) may be writte n as
Theorem 2 Considel' Ih e syslem (;,j defined in (8) and
AY + Y A' + B21V + IV' B~ + B, B; < 0
let A m(9, j ) denote Ih e sd of admissible memol'yless COIl-
+ C I\"D' + OIVC' - Q OIV]
/1'ollel's fol' this pial>!, LcI G',j denol e Ih e lran sf cl' mall'ix of
(;,j , Lel Q = Q' > 0 be git'ell. Tituc c,risls K E A m(t;;,j) l [
CYC'
IV'O' _y < 0,

80
:\ote that the I('ft hilnd side of tllis system of inequa lities is t he ort hogona li ty assumption can be madc with no Im" vf
amne in (II'. )'), Iloyd and Yang (198!l) ga\'(' a Illllnber of genera li ty, In deed. a prel iminary feedback transforma ti on,
nu m('rica I algori t h ms a ppropria t e for si III i la I' COl1\'ex feasi- wi ll en force D2[ B; D;] = [0 1].
bility problems, The applicat ion of this algorit hills to our It is well-known Ihat . under thr abo\'c assllmptions.
synt hrsis problem is ClIlTent Iy und('r ,t udy and tll(' rrsults there ex ists a (unique) real symmetric matrix ) ' sllch that
of this investigations will 1)(' r('port('d in the future, (20)
\ rhen thr maxinlllln diagonal rntry is usrd in thr drf-
initioll of the gelleralized I II performan('e measure (i,e, and A - Y C~C2 is stable. FurtllPrmor('. ) . 2: 0 and if th('
I' = 00). it may b(' shown that th(' g('nrralized II I control
pa ir ( A . Bd is controllable then ) ' is definit('. In order to
problem is equi\'al('nt to a multicritC'rion II I control prob- gi\'c the ma in result of this srction \1'(' defill(, the following
lem, This simplr follows from the fact that in this case the aux il ia ry system:
performance requir('mrnt is to kerp thr 1/ 2 norm of e\'e ry .i· = / 1.1' + ) '(';1' + iJl/l
single row of F,(Q)-'/2T, u' below onc. Boyd and BalTat
9 'f()') := { ~: ~.'l' + D,ll (21 )
(1990) ga\'e an algorithm that can not fail in drtcrmin ing
whether thp re'lllting multicriterion III problpm is feasib le,
where the matrices A.C,. D , . 13 1 arr as in (19). Let G'f()')
\rhen the problpm is fpa5ible. this algorithm also finds an
denote the transfer matr ix from (I'. /I) to (' . y) in (21) . lI erc
LQH controllrr gain that 501\'es the I!;eneralized 112 control
thc notat ion 9'f() ') nwans that this allxiliary plant de-
problem. ,\ppli('d to tll<' prc'spnt problem. the algorithm
pends on ) ' and has state-f('cdback stru('t\ll'e. The main
proposed by Bo.\'(1 and Barrat sparchrs 0\'('1' a ,pitCC whose
result of this section is gi\'cn ncxt.
dimcnsion is 011<' kss than the IIllmb('r of ('omponents of
thr rpglllated Olltput ::. In this sensr. our n'sults for the Th e o re m 3 Considu the f((dbacl.· intel'connrction of
case I' = 00 apprar to br romplltationally nlon' intrnsin' I,
Fig. 1. whel'e the plant 9 I" gll'( by (1.9). Supposc that
than those of Boyd and IlalTill. ,\n '1<I\'an\;lg(' of our ap- Asslllll]>tions ;-\'-:\ :1 hold. Th(n. th(l'( uI"t.< a Ul/Iquc (I'eal
proach is that no ,lsslInlplioll' 011 probl('1I1 dilta (i.(' , 011 thc symllletric) matl'i.r ) ' that snti.<jir.< till .. \ IIF (JO) and sllch
state-space' data 1I1trodllc<'d in (~)) <lr(' re'<[lIirrd, tha t A - ) 'C~Cl is stable. ,1101'(0/'(1'. ) . 2: O. l,cI Q = Q' >
:\pxl, il will 1)(' shownthill thc IlIixed II I synllH'sis proh- o be 9il'('n and cOl/sidcl' Ih( .<rls of admi""tbh conll'o//rl'''
Icm in thr casc of of Olltpllt f('('dback call be reduccd to the A W) and A ..,W,f() ')). Then. Ihe fo//oll'in!J .<Iallll/rl/Is al'r
st atr-ferdhack prohl('m cOII,id('rt'd ilbo\'(' , Thus. t he ap- cqui/'alcnt :
proach de\'elop('d in this ,,'ction is dire'clly applicahlr to
thr morc gcncral Olltpllt f('edhilck CilS('.
I. There crists C E AW) ,<nch Ihal .1,.((" . C) < I~(Q).

) There cris is K E Am(9.~J( } '·)) 8/1('11 Iha/ J,·(("sj. /\') <

I'~(Q - C, )"C;).
01 "ll'l ' 'I' FI-:I ': \)IL\( '1\
Further, gi1'(,1I any I·rallllol/·i./, 1\' Ihal "alisjir., Ihis lasl COI/-
III this s('ct iOIl W(' cOllsidcr tl,,· g('II('rilli/('<I III coni 1'01 prob-
dittoll. Ih( dYlla1l1ic oulpul fr(dbark ('Qnll'O//(1'
Icm for thc cas(' of 0111 pili i'<-edback, Thl.maillreslllt of
this srelioll (T'IH'orem :\) sh,)w, Ih,d i1 grlH'ralizrd II I out- C.= { ~ = (A - \ 'C;C 2 + /32 /\')~ + H';!! (22)
p"t fredback cOlltroll,'1'. if it exisls, call be cl,oS('n to br . 11 = 1\'(
i1 combill<ltion of a stilndard l\allll'"1 filt(,r, and a stat('-
sa lisfies C E A (9 ) and .1,(G.C) < f ;.(Q).
frrdback gain for the grlleraliZl'd III synthesis problem of
an auxiliary plant. T heorem 3 shows t h at thc g('nrra lizcc\ 112 s)' nt hesis
SIIPPOSP that the planl ill Fig. I is gi\'ell hy I he' following p roblem in tll(' output feedback case has a solution if and
stat('-spacc modcl: on ly if the genera li7.ed 112 synthesis pl'Obkm for the (state-
feedback) p lant 9 'f{)') has a solutioll. If a sollltion to th is
.i·=,~.r+lj,,,,+ Ul/l
state-feedback problem exists. the'n a gcnrralized 112 con-
9:= ::=(,.1'+/),/1 ( \9) troller for the p lant 9 can be chosen to 1)(' a combination of
{
y = C/.I· + /)111'. the Kalman est imator (of thc state of thr plant gi\'en the
where all the matric,'s ill (11)) aI'<' I'<'itlmiltriccs of compal- measurements y), and a state-fcedback gain for the gene r-
ihle dimensions. ,\s hrfol'('. \I'P I<'t (; <I('not(' thr transf('r a li zed H2 synthesis problem COIT('sponding to t he aux ili ary
matrix of (19). :\otr thal Ihrrr is no f('<,dlhrollgh tprm plant 9 'f( Y )' lI ence. this thcorclll. along \\'ith the resu lts
from thp exogpnous sigllal It' to thc controllrd Olltput z. in t he p rev io us sect ion . provi(\P a ('ompk t (' solut ion for t hi s
Similarly. thcr(' is no dir("t fr('c\through from Ihe cOlltro l c lass o f synt hesis problems. This is suml1larizrd in t he fo l-
input 11 to thr m('aslll','d out pili !/ . .\!though it is possiblr lowi ng concept ua l a lgo ril hm:
to include tl1('s(' trrms. \I'r ha\'r choS('n nol to do so in ol'<iPr I . Find the stabilizing solut iOIl to t he A H E (20) . ?\ote I hal
to kpcp tll(' prcsentation as simpl,' as possib!r'. In additioll. this solution exists .
wc will makc till' following as,,"nptions: 2. l:sing Theorem 2. determine \\'hct h('1' I hr grnrralized 112
,\ I) The Iriplr (Cl, A. 13 1 ) is stabilizabk and (ktpct ablp. state- feedback problem for the auxiliar~' plant 9'f() ') is
sol\'ab le. If a solution 1\' cxists. t hc controller (22) solves
A2 ) D2 has full row rank and D2[B; D;] = [0 ! ].
the general ized H2 synthesis pl'Ohlrm for thr plant 9, Oth-
A 3) The pair (A, B I ) has no uncontrollable modes on the e r wise, t he generalized 112 synt hrsis problem for the plant
imaginary axis. 9 does not ha\'e a solution.
Clearly AI is necessary. otherwise thr sct of admissib le con- It shoul d a lso be obsen'cd t hat although the abO\'e
trollers is empty. Assumption A2 is a standard assumption resu lt resemb les t he separation st ructlll'e of t hc standa rd
in fi lteri ng theory. It ensures that there is full mcasurement LQ G p roblem, t here is a differencE' . In Theorem 3, the
no ise and that there is no correlation between proccss a nd state- feedback gain !\' of Condition 2 cou ld depend on the
measurement noise. Finally. in the \·icw of A 2 • Assumpt ion stab ili zing solution }' to the filtering A H E (20).
A3 together with (C 2 , A) detectable constitutes a necessary A detali ed argument, leading to a proof of Theorem 3,
and sufficient condition to guarantee that the l\ alman filter, may be fou nd in R otea ( 1991). lI ere. \\'e just gi\'e the cen-
that estimates the state of 9. is stable. These are common tral ideas behind th is result. First note that . under our
assumptions in the standard H2 and !loo filtering t heories, ass u mpt io ns, the stabi lizing solution ) ' to the A HE (20)
Note that the full row rank property of D2 implies that exists an d satisfies Y 2: O. ~o\\'. define the following auxi l-

81
iary system: forced by the H oc constraint.

.i; = Ax + Y C~r + 8 2 11
H(Y):= v = Clx + Dill REFEREi\CES
{
Y = C2 x + ". l3ernussou J., P. L. D. Peres. and J, C. Geromcl ( 1989). An
Let H(Y) denote the transfer matrix frol11 (1'.11) to (1', y) LP oriented procedure for quadratic stabilization of un-
in (23). certain systems. Sysl. COIII,·. Dell ., 11. 3, 65-72.
Let C E A(9) denote any admissible cont roller. and 130yd S. P. , V. l3alakri shan , C. H. l3arratt , N. ~!. Khraishi,
consider the feedback systems shown below. Obserye that X. Li , D. G. ~leyer, and S. A. i\orman (1988). A new
CAD method and associated architectures for linear con -
trollers. IEEE Tmlls. Oll.-t utomal. CO'III'.. 11. 3. 268-283.
Doyd S.P. and C. H. Darratt (1990) . em Linear controller
design: limit s of performa nce. Prent ice- Hall. New Jersey.
Doyd S. P. and Q. Yang (1989). Structured and simultaneous
Lyapuno\' functions for system stabilit~· problems. Int.
JOll. Contml.:ill. 6. 2215-22 I0.
c

H(Y) and 9 share the same (A, 8 2 , C 2 ) triple. lI ence, C E Dahleh 111. A. and J. 13. Pearson ( 19R7a). I1 optimal feed-
A(H(Y)). Further, in Rotea ( 199 I) it is shown that the back controllers for lI1I~10 discrNe-time syste ms. IEEE
integral of the "square" of the closed loop transfer matrices Tmns. 011 Aulomal. COllI,· .. ;11, ·1. 31 -1-322.
T,w and Tvr in the figure abo\'e arc re lat ed by Dahleh ~1. A. and J. 13. Pearson (1987h). I'I optimal com-
pensators for continuous-time systems. IEEE Tmlls. 011
1
-2
7r
/00
- 00
T, w(jw)T;w(j..,.') d..,.. = Aulomnt. COIlI .... 31. 10. R89-89'i.
Doyle J. C .. I,. GIO\·er. P. P. ['hargonekar. and n. A. Franci s
Cl yc; +
27r - 00
~ /00
Tvr(j",,)r.:,. (j~·) d..:. (24) (1989). State-space solutions to standard 112 and 1100
and control problems . IEEE Tl'Olls. on .-\"Iomal. Conlr"
This separation prope rty of the gen('ralized 112 cost of li, 8, 831-841.
the closed loop (which hold s for any admissible controllrr C) Francis 13. A. and J. C. Doylc (19R7) . Linear control theory
is the key to esta blish Theore m :3 . Indeed. from (24), it fol- with an 1100 optimalit)' criterion. SIIl.II Jou, Contml and
lows t,hat a controller C E A(9) sat isfies J,((,',C') < F,.(Q) Oplimi:alion. £,1, 815-8 l·1.
c

Grimble M. J. (1990). Relation ship between the trace and the

if and only if the same controller satisfies C E A(H(Y)) and
maximum eigenvalue norms for linear quadratic control
Jr(H(Y) ,C) < Fr(Q - CIYC;). Theorem :.1 is finally ob-
design . IEEE Tmns. Oil Aulo mnl. Conlm/,~, 10, lt76-
tained by solving the generalized 112 synt hesis problem for
11 8 !.
the auxiliary plant HP "). It is important to note that, eyen
l\alman H. E. (1960). Contributions to th(' Iheol)' of optimal
though this generalized Ilz problem for the plant HP '-) is control. lJol. Soc ..lInl. Mu-ira.1. 102.
still an output feedback problem. it is a nluch simpler one.
I,hargonekar P. P. (1990) . State-space 1100 control theory.
In fact, since the open loop transfer matrix from ,. to !J in
Report no. CGR-27, EECS, The Uniyersity of Mich igan,
H(Y) has a stable inverse (recall that A - )'C~Cl is stable), Khargonekar P. P., I. R. Petersen, and lIl. A. Rotea (1988).
it is possible to construct an output fe('dback controller C H oc-optimal control with state-feedback. IEEE Trons.
that achieves exactly the same pe rformance proyided by a on Aulomal. COlllm[, 11, 8,786-788.
full information controller acting on HP "). Or, by Theo- Khargonekar P. P. and M. A. Rotea (1990). lIlixed H2/ Hoc
rem I , a state-feedback controller acting on g,jP·'). For control: A convex optimization approach. To appear in
further details on this argument see Rotea (1991 ). IEEE Tmns. on Automat. Contm/.
Rotea M. A. (1991). The generalized 112 control problem .
Submitted.
CONCLUSIONS Vidyasagar Iv!. (1986). Optimal rejection of persistent boun-
In this paper we haye considered a (s ub-optima l ) general- ded disturbances . IEEE Trans. Aulomat. Contr.,;U, 527-
ized Hz control problem. This synt hesis problcm is well 535,
'vVilso n D. A. (1989). Convolution and Ifankel operator norms
motivated since it re presents a problcm of disturbance at-
for linear systems. IEEE Tmns. on Automat. Control,
tenuation, as measured by the peak-yalue of t he controlled
M , 1,94-9 7.
outputs, in the face of unknown energy exogeIlous input s . Wilson D. A. (1990). Extended optimalit)' properties of the
For these class of synthesis problem s wc ha\'e shown that: linear quadratic regulator and stat ionary [,alman fi lter.
1. When either the state of the plant. or the s t ate and IEEE Trons. on A lItomal . Cont1'Ol, ;t2, .) , 583-585.
You]a D. C., J. J. Dongiorno . and Il.A. Jabr (1976a) . Mod-
the exogenous input. is available for feedback, sialic slnlf-
ern Wiener-][opf design of optimal controllers. Part I:
feedback gains offer the best possible performance for the The single-input-output case. IEEE Tmlls. on Automatic
generalized H2 synthesis problem. Control. 21, 3-13.
2. In the case of state-feedback, necessary and sufficient Youla D. C .. J. J. Dongiorno. and H. A. Jabr (1976b). Modern
conditions, as well as a formula for a solution, may be ob- Wiener-Hopf design of optimal controllers. Part ll : The
tained via a finite-dimensional convex feasib ilily program. multivariable case. IEEE Tmns. on Automalic Contml,
21, 319-338.
3, In the case of output feedback , generalized Hz controllers Zhu G. and R . E . Skelton (1990). ~Iixed L2 and Loo prob-
can be chosen to be a combination of a standard A'alman lems by weight selection in quadrati c optimal control. To
filter and a state-feedback gain for the generalized Hz syn- appear in /nlemalional JOll. of Conl,·o/.
thesis problem of a suitably constructed auxiliary plant. Zames G . (1981). Feedback and optimal sens itivity: Model
reference transformations, multiplicative seminorms, and
Finally, the results in this paper may be combined with approximate inverses . IEEE T1YI1ls. on Automatic Con-
t hose in Khargonekar and Rotea (1990) in order to solye trol, 26, 301-320.
the control problem of finding a controller such that an
"upper bound" for the generalized H2 cost on one closed
loop transfer matrix and the Hoc norm of some other closed
loop transfer matrix are below desirable levels. This mu 1-
ticriterion control problem is extremely important since it
represents a problem of nominal performance , as measured
by the genera lized Hz cost, with robust stability, as ell-

82