Analysis of feedback systems with structured

uncertainties
John Doyle
Indexing terms; Control theory, Feedback, Robustness, Sensitivity

Abstract: The paper introduces a general approach for analysing linear systems with structured uncertainty
based on a new generalised spectral theory for matrices. The results of the paper naturally extend techniques
based on singular values and eliminate their most serious difficulties.

1 Introduction blem leads to the definition of a function n that provides

The last several years has seen something of a revolution in necessary and sufficient conditions for structured matrix
multivariable control theory. Central to this revolution is a perturbation problems. Section 4 examines the properties of
renewed focus on the fundamental issue in feedback design: the function ju and expresses it in familiar matrix algebraic
providing performance in the face of uncertainty. This develop- terms. Section 5 develops the tools necessary to find gradients
ment has been supported by a renewed interest in the frequency for singular values, which are used in Section 6 to provide
domain. An indication of the growing acceptance of these techniques to compute n in important special cases. Section 7
points of view can be found in the recent Special Issue on has some examples and a discussion of recent experience in
'Linear multivariable control systems' in the February 1981 computing ii, and Section 8 gives a summary and conclusions.
IEEE Transactions on Automatic Control, where five of the
first six articles [1—5] deal with uncertainty, robustness and 2 Importance of the block-diagonal perturbation problem
sensitivity using frequency-domain tools. Analysis methods based on singular values have been successful
There are many results of both practical and theoretical in providing a framework within which to develop multiloop
importance which have come out of this research. As a conse- generalisations of classical single-loop techniques. While this
quence, it is now well known that multivariable sensitivity has been an important development, singular-value methods
and robustness cannot be reliably evaluated one loop at a have serious limitations. Consider, for example, the problem
time [1, 6 ] . Singular-value methods have proved useful in of analysing a linear multivariable feedback system with two
providing both a reliable multivariable measure of sensitivity multiplicative perturbations appearing simultaneously at the
and robustness [1, 3, 7] and, more generally, a framework inputs and outputs [8—10].
within which to develop reliable multiloop versions of the There are essentially two direct applications of singular-
main tools of classical control theory [1—9]. value techniques to this problem. First, the system may be
While these results are reasonably well understood within rewritten to isolate the two perturbations as a single perturba-
the control theory community, many of the more subtle tion with a two-block-diagonal structure. This structure can
aspects of multivariable feedback systems are less well under- then be ignored by replacing the block-diagonal perturbation
stood. Multiloop systems have important properties that have with one full matrix perturbation. In this configuration,
no analogues in the single-loop case, For example, there are standard singular-value analysis can be applied. Of course, the
problems caused by the fact that the signals and responses in results may be, in general, arbitrarily conservative. A second
multiloop systems vary not only with frequency, but also approach is simply to treat the two perturbations one at a
with direction. While some published results [8—10] have time, leading possibly to arbitrarily optimistic answers.
given preliminary indications of this, the issue of directionality If these first two approaches to the simultaneous pertur-
in multiloop systems is largely unexplored. bation problem happen to give the same results, then the
The most serious limitation associated with the methods 'true' answer is obviously determined. In general, this will
based on singular values is that they treat a limited, albeit not be the case. It should be emphasised again that these two
important, class of uncertainties. Roughly speaking, most of approaches can yield arbitrarily bad estimates; i.e., depending
these results give, at best, exact characterisation (i.e. involving on the problem, the gap between the upper and lower bounds
necessary and sufficient conditions) of the feedback proper- may be arbitrarily large. No other information can be obtained
ties of a system model relative to perturbations which are from these bounds other than that the 'true' answer lies
norm bounded but otherwise unconstrained. There is no between them. This will become evident in the course of this
systematic mechanism for exploiting information about the paper.
structure of a perturbation.
An extension of the direct application of singular values is
The aim of this paper is to develop techniques for analys-
to analyse the differential sensitivity of the singular values
ing systems with arbitrary constraints on their structure.
evaluated at one point relative to perturbations at the other
Section 2 discusses some particularly important examples of
[10]. Reference 10 is extremely well written and would pro-
structured uncertainties in an attempt to motivate the general
vide excellent background for the present paper. However,
results in later Sections.
this approach suffers from the fundamental difficulty associ-
Sections 3—5 present the main theoretical results. In ated with differential sensitivity techniques in not applying to
Section 3, the block-diagonal perturbation problem is formu- large perturbations. In addition, the results in Reference 10
lated. This problem is quite general since any norm-bounded provided only directional sensitivity information (i.e. direc-
perturbation problem, regardless of structure, can be trivially tional derivatives) when singular values are not clustered. If
rewritten as a block-diagonal perturbation problem. This pro- differential sensitivity methods are to be effective, it is often
important that worst-case directions (as with gradients) be
Paper 2205D, first received 4th May and in revised form 14th September obtained. The results of this paper can be modified for com-
1982 puting worst-case differential directions for clustered singular
The author is with Honeywell Inc., MN 17—2375, 2600 Ridgway values. Although not developed further here, this may make it
Parkway, PO Box 312, Minneapolis, MN 55440, USA
242 0143-7054/82/060242 + 09 $01.50/0 IEEPROC, Vol. 129, Pt. D, No. 6, NOVEMBER 1982
possible to complete the results in Reference 10 for clustered Let
singular values.
What is really desired, of course, is an exact analysis of the
original simultaneous independent perturbation problem,
which applies to large perturbations. One application of the
u
7=
results in this paper is just such an analysis method. This will
be illustrated in Section 7 by a simple 'textbook' example.
The problem of evaluating robustness with respect to simul- Thus Xb is the set of block-diagonal matrices with structure
taneous input and output perturbations is a special case of the determined by Jf whose norm is not greater than 5, and
more general problem of evaluating robustness with respect Xoo is all such matrices with no restriction on the norm.
to perturbations with arbitrary constraints on their structure. What is desired is a function (depending, of course, on
Consider the general problem of an interconnected linear
system with multiple independent norm-bounded perturbations
occurring throughout. By rearranging the system, it is always
possible to isolate the perturbations as a single large block-
diagonal perturbation. with the property that for
Stability analysis, for example, then essentially boils down
to ensuring that / + MA remains nonsingular at all frequencies
and for all A under consideration. Here, A is a block-diagonal det (/ + MA) =£ 0 for all A G X8 (1)
perturbation and Mis the transfer function from the collective
outputs of the perturbations to their inputs. Readers unfam-
if, and only if,
iliar with manipulating systems to isolate specific components
as a diagonal matrix are encouraged to try some examples and
convince themselves that it can be done. This notion is essential
to understanding how the results in this paper may be applied
to analyse control systems. This could be taken as a definition of fi. Alternatively, //
The key problem then involves the matrix problem of could be defined as
determining necessary and sufficient conditions such that
det (/ +MA)¥: 0. All norm-bounded linear perturbation
problems reduce almost trivially to this. The main results in
this paper are a partial solution to the general block-diagonal 0 if no A e Xm solves det (/ + MA) = 0
perturbation problem and a complete solution for the case tx(M) =
of three or fewer blocks. mm (amax(A) | d e t ( / + MA) = 0

3 Definitions and problem formulation otherwise

The following notation will be used throughout:
^f(k) = alegbra of complex k x k matrices
^/{k) = unitary matrices in Jf(k)
This definition of ju shows that a well defined function satisfies
°max(M) = maximum singular value of M
p(M) eqn. 1. It probably has little additional value since the opti-
= spectral radius = magnitude of largest eigenvalue
= least (most negative) eigenvalue for Hermitian H misation problem involved does not appear to have any useful
3
M " = conjugate transpose of M properties.
diag M2, M 3 , . . . . , Mn) = block-diagonal matrix with In order to proceed further, some additional definitions are
Mj (not necessarily square) on the diagonal needed. Again, assume the 2«-tuple ^"is given.
Then let
To provide a description of block-diagonal perturbations, let
JT= ( m i , m2,. . . , mn, kx, k2,..., kn) be a 2«-tuple of
positive integers. All the definitions that follow depend on %= U{K)C\XX (block-diagonal unitary matrices)

= {diag (tf i 4 , , d2lki ,...,dmilki,dm .,dmIkm)\dteR+ = (0,oo)}

= {6i*&(dl,d2,...,dK)\d,e[0,~)}

= {diag (i>,, vx , v2,.. . , v2, v3 vn.x,vn,...,vn)\ VjECkJ*l,v;Vj = 1,/ = 1,2,...,/!}

JT, but, to simplify notation, this dependency will not be Note that these definitions all depend on J2f, but to explicitly
explicitly represented. Unless specifically noted otherwise, represent the dependency (for example, \xx, X^.s, ^ e t c . )
^ w i l l be assumed to be an arbitrary but fixed 2«-tuple of would be unnecessarily cumbersome. Using these definitions,
positive integers. Let several useful properties of fi are stated in the following
Section, leading up to the first main result of the paper.
K = and
The problem formulation taken here is certainly not the
most general possible. For example, the entire development in
For each 5 ER, 8 > 0, let X8 <ZJ?{K) be

X8 = { d i a g ( A 1 , A 1 , . . . , A 1 , A 2 , A 2 , . . . , A 2 , A 3 , A n _,, A n , A n ,. . . , A J
and amax(Ay) < 5 for each/ = 1, 2 , . . . , n}
IEEPROC, Vol. 129, Pt. D, No. 6, NOVEMBER 1982 243
this paper could be done with nonsquare perturbations with norm on C h. Let z G Cfe be such that p(z) = 0 and \z\ =
no conceptual change in the results or proofs. The notation, min (l|z|Llp(z) = 0); i.e. z is a solution of p(z) = 0 with
however, would become even more cumbersome than it zG<C fe
already is. The particular approach taken is the simplest that minimum || • IU norm.
allows both p and a max: to emerge as special cases of yi. Hope- Lemma 1
fully, this will aid the reader by maintaining continuity with There exists x i C fe such that p(x) = 0 and | x / | = | | i
standard linear algebra as well as with some of the more for ally = 1 , . . . k.
popular existing methods in multivariable control.
Proof
Let Xj = 2j for each z;- such that |z;-1 = | | z | L . If \zj\-
4 Properties of 11
HzIL for all / , then x=z satisfies the lemma. So suppose
The following properties of/i are easily proven from eqn. 1. that one of the components of z, say for convenience zk,
As always, JT= {m\, . . . , mn, kx, . . . , kn) is an arbitrary has \zk | < | | z | U . Then p ( z ) = S p , . ^ , . . . ,zk_x)zk, where
2«-tuple for some n G / + , unless specified otherwise, and each pr is a polynomial in A: — 1 variables. Suppose that for
some r, pr{zx,. . . , zk_l)i^Q. This will lead to a contra-
n diction. By continuity of the roots of a polynomial, for any
m k
K = Z ii e > 0 there exists 5 > 0 such that, for each y GCh~l such
that || y — ( z ! , . . . , ffe_ j ) II < 5, there exists a w G C such
(a) vipcM) = \a\fi(M) for all MGjf(K) that \w — zk\<e and Xpr(y)wr = 0. Then, by choosing e
sufficiently small, there exists a / G C f e ~ 1 and w G C such
(b) M(/) = 1 that P(j>i,y2,...,yk-i, vv) = 0 and IICPi,. . . , yk.l, w)
(c) fji(AB)<omax(AMB) fox all A, I L < | | z | U . This contradicts the minimising property of z.
(A) forallAGX 6 Thus pr(zx, . . . , zfe _!) = 0 for all r, 0 < r < q, and p(z x ,
(d) = a . . . , zk _ ! , w) = 0 independent of w. Let xk =\\z IL. This
If n = 1 and m t = 1, then argument may be repeated for each j = 1, 2,. .. , k so that
either Xj = 2j or Xj = \\z |U and p(x) = 0 for x = ( x 2 , . . . ,
li{M) = amax(M) for jcfe). This completes the proof.
Hn = l,kl = l,then This lemma may seem a bit obscure at first glance. What it
K = mx basically says is that if HzIL is to be minimised subject to
the constraint that p(z) = 0, then at least one minimising
x6 = {x/|xe c , i x i < 5 } solution x lies on the poly disc where each \Xj \ — | | J C | L . This
will allow a great simplification in the characterisation of i±.
ju(M) = p{M) for all J?(K)
First though, note that the proof of lemma 1 immediately
implies the following two additional lemmas.

(h) For all A € I K a n d for all Lemma 2

DAD- 1
= A If z is real [Im(zy) = 0 for all / ] , then there exists xGUk
such that p(x) = 0 and \xj \ = \\z IL for all / = 1, 2 , . . . , k.
(0 For all U G ^ and M G J?(K),
Lemma 3
If z is real and non-negative [Im(zy) = 0 and R e ( z ; ) > 0 for
a l l / ] , then there exists x GUk non-negative such t h a t p ( x ) = 0
(/) For all D G ^ a n d M GJ?(K), andx ; - = ||z||c=.
Theorem 1
l For a
(k) max p ( UM) < /i(A/) < . inf amaje (DMD~ )
H(M) = max p(MU)
for
Proof
(/) For all A G Xoo, there exist U, V G ^ a n d Suppose MGJ?(K). If /i(M) = 0, then the result follows
such that immediately from property (k), and so assume n(M) = 1/8
> 0 . Then there exists A G I 6 (not necessarily unique) such
A = UH V*
that omax(A) = 8 and det(/ + MA) = 0. By property (0,
(m) n is continuous in the usual metric topology on there exist U, KG ^ a n d 2 G ^ s u c h that
A = UEV*
Property (/) is just a block version of the singular-value decom-
position (SVD). Properties (e) and (/) show that JU has as special
Then
cases both the spectral radius and maximum singular value or
spectral norm. Property (i) means that pi is ^-invariant. The det(/ + MA) = 0
main theorem of this Section will show that the left-hand
inequality in property (k) is actually an equality and thus if, and only if,
that ix is, in some sense, the fundamental ^-invariant. In order
det (I+ MUXV*) =0
to prove this theorem, some simple lemmas involving poly-
nomial equations are needed. This last equation may be viewed as a polynomial in the diag-
Suppose p: Cfe-*• C is a polynomial in k complex variables onal elements of 2 . By assumption, E is a minimum norm
of degree no more that q G 2 for each variable. For z = (now viewed as || • II- on the diagonal elements) solution to
( z l 5 . . . , z f e )GC f e , let ||z|U = max \zj\ be the usual || • IL this polynomial. This satisfies the conditions of lemma 3 and
244 IEEPROC, Vol. 129, Pt. D, No. 6, NOVEMBER 1982
therefore 2 may be replaced by a scalar 5/. Thus tive alternative. The first step needed for solving infa majc
(DMD~l) is a technique for computing descent directions.
det (I+ MUXV*) = 0 This problem is taken up in the following Section.
implies
det (/ + 8MUV*) = 0 5 Differentiability properties of singular values

which implies This Section will develop the necessary tools for computing
'gradients' for singular values. Although singular values are not
p(MUV*) > 1/6 = fi(M) in general differentiable functions of the matrix elements, it
By property (k), the reverse inequality also holds, and so is always possible to compute a generalised gradient which
serves the same purpose as would a gradient. This Section will
max p(MU) = ii(M) be brief and not at all self-contained and only original results
f will be proved. An excellent background for this Section is
Corollary Reference 10, as well as standard texts [11—13].
Suppose M: SI -+J?(k) is a matrix-valued function whose
= max p(V*MW) elements are real analytic in some neighbourhood SI of 0 G U Q.
V, Denote by omax(x) [for a majc (Af(jt)), x G SI] the maximum
Proof singular value of AT as a function of x G£2 C UQ.
max p(MU) = \=> there exist £/G %f and x G C f e such The directional derivatives of omax(x) at JC = 0 depend
only on M{x) to first order, and so without loss of generality
that Aft/* = XJC. This implies there exist assume
K, W G ^ s u c h t h a t M W K * * = XJC
M{x) = Mo + X
which implies
p(MWV*) = p{V*MW) > X for x = (*!, x2 , . . . , xq). Using the SVD, M(0) may be
written as
which in turn implies
Af(0) = MO = o w ( O ) ^ V\ + U2 2 2 V$
max p(V*MW) > X
v, Here Ux, Vl,GCkxr with r the multiplicity of a majc (0) and

Similarly, [UlU2],[VlV2]EUik)
max p(V*MW) = X Suppose amax (0) > 0 [i.e. M(0) * 0 ] . Then, letting (a m a j c ) 0 =
V, a
max (°)> °max (*) m a y b e written as
implies = \(omax)0 + j) + 0(*) I (2)
max p(MU) > X where

Thus Hj = Hf = ReiUfMjVi) = hiUfMjVx + V\MfU{)

H(M) = max p(V*MW) for all Af G ^ ( X ) Define Vt C R q by

Vi = {x£UQ, xj = v*HjV\v*v = 1 and v is an eigen-
9
This theorem and corollary express /i in terms of familiar linear vector for some 0
algebraic quantities. Note that it is now possible to use the and denote by coV! the convex hull of V x . Denote by x = min
corollary to define a general matrix decomposition that would (coVi) the unique point x G coV! that minimises
have both the singular-value decomposition and Jordan form
as special cases. While such a unification may prove to be of min || x || 2
a great theoretical interest, it is not essential to the aims of x 6: coy,
this paper and will not be pursued further here. A well known property of convex sets that will be used in the
Instead, the remainder of this paper will be devoted to following theorem is that (x, x) > || x || 2 for all x G coVx [14].
developing methods to compute ii{M). Unfortunately, the The significance of Vt is expressed in the following theorem.
optimisation problem expressed in the corollary cannot be
Theorem 2
solved in general by simple gradient or local techniques.
Examples have been generated which have multiple local If JC =5^ 0, then there exists e 0 > 0 such that
maxima. While recent computational experience with some °max ( - ex) < (omax)0 for all 0 < e < e0
alternative algorithms is very promising, no efficient algorithms
have been developed which have proven guaranteed conver- Proof
gence to the global maximum. Let v be a unit eigenvector associated with the eigenvalue
An alternative approach will now be taken by considering XminiXXjHj) and let ^ G V , be defined by ys = v*HjV for
the right-hand side inequality of property (k). In important 7 = 1 , 2 , . . ..tf.T
special cases it will be shown that this inequality is actually
an equality, and that inf a^,^ (DMD~l) has no local minimum j) = V*(?XjHj)V = Cx,^» 0
which is not global. In fact, a tedious but straightforward cal- Then from eqn. 2 there exists e0 > 0 such that amax ( — ex)
culation shows that amajc(Z)MZ)~1) is convex inD. The proof < iPmax)o for all 0 < e < e 0 .
of this is omitted since it is not actually needed for any of the
main results, but it makes this optimisation problem an attrac- This theorem implies that, although omax{x) is not necessarily
IEEPROC, Vol. 129, Pt. D, No. 6, NOVEMBER 1982 245
differentiable, x = min (coVx) serve one of the purposes of a Proof
gradient in providing a descent direction. If x =£ 0, then a Suppose
direction (for example, —x) can be found which reduces all
the singular values in the cluster. In the special case where
omax(0) has multiplicity one, then Vx trivially reduces to the
ordinary gradient. It may be thought of as the direct limit of
the sets of possible gradients in neighbourhoods of 0. Unfor- for some /. Then
tunately, Vx is a rather awkward set to work with since it
depends, in part, on singular vectors of M{x) for x =£ 0. a b
As an alternative, consider the set sin0(cosi// — /sini//)] | _
b
v2 = {xeuq,xj = v*Hjv\veCq,v*v = 1}
COS0
Clearly, Vx C V 2 . While V2 has a much simpler description
than V x , it serves the same purpose, as is shown by the [sin 6 (cos i// + /sini//)
following lemma and theorem.
= acos •d +2Re(6cos0 sin
Lemma 4
min (coVO = min (coV2) (l/2)(a-c)cos20
Proof + sin26 Re(Z>(cosi// + /sini//))
Let y E V2 and x = min ). Then y}- = v*HjV for some v
such that v*v = 1. Then = 0/2) (a+ c)
(x,y) = 2XJ(V*HJV)
cos20
Re(6) cosi//
= (x,x) for some Jt
since x = sin20 sini// .

Therefore \\x\\< || min (coV2) || and since j C V2, x = min Let gi :R3 -*• U be the affine map defined by the scalar and
(coV2). vector in the last equation so that/i(P 2 ) =gi(S2) as shown.
Define gj for / = 2, 3,. . . , q similarly so that the resultant
3 2 22
Theorem 3 g:UJ ^U is affine. Then/(P ) =g(S ) as desired.
i = coV2 Corollary
For /:C 2 defined above, if q = 1 or 2, then f{P2) is
convex.
Proof
It suffices to prove that dist {x, coVj) = dist (x, coV2) for all Proof
xe Uq. Fix jf = ( x 1 > . . . ,xg)e Uq. Define Vt and V2 by If q = 1 or 2, then for any affine map g:U3
replacing H}- by Hj—Xjl in the definitions of V! and V2, convex. Apply the theorem.
respectively. Then, by lemma 4,
Corollary
dist (x,coVi) = || min (coVi) II2 = II min (coV2)||2 Let f:&^Mq be defined as above for arbitrary rEI+. If
q = 1 or 2, then f(Pr) is convex.
= dist (x,coV2)
Proof
This holds for any , and so the proof is complete. Let x, y E C r , x ±y and || * || 2 = \\y \\ 2 = 1. It suffices to
In order to make effective use of V2 in, say, an optimisation show that z{t) = tf(x) + (1 - 0 / 0 0 e/(P r ) for all t E
algorithm or a sensitivity calculation, there must be some [0,1]. Let
reliable algorithm for computing min (coV2) [or maybe
max(coV2) for sensitivity] given the {Mj}. The Appendix
briefly describes an approach to this. The rest of this Section H, = y]
is concerned with properties of V2 that will prove useful in
the following Section.
Suppose that Hj = Hj* E ^ ( r ) , / = 1, 2,. . . , q, and let for all; = 1, 2,. . ., q, and let/:C 2 -> Uq be defined as before
/ : C r -» UQ be defined by fj(x) = x*HjX f o r * E C r . Let in terms of the Hj. Then f(x), f(y)€f(P2)Cf(Pr) and
f(P2) is convex by the previous corollary. Thus z(t)Ef(P2)
Pn = = 1}CC" f(Pr) for all t E [0, 1 ], and so f(Pr) is convex.
and This last corollary implies immediately that, )
for fiERa and q = 1 or 2, then the V2 defined as before is
Sn = {xGUn +l
\xTx = convex. This will prove quite useful in the following Section.
Note that the results of this Section and Appendix could
Note that this notation is not standard. be used to study the sensitivity of clustered singular values to
infinitesimal parameter variations. The aim of this paper,
Theorem 4 however, is to develop techniques for handling large perturba-
3
For r = 2, there exists an affine map g: -> Uq such that tions, and so the following Section will make use of these
f(P2)=g(S2). results to compute the function \L.
246 IEEPROC, Vol 129, Pt. D, No. 6, NOVEMBER 1982
6 Computation of fi These two theorems have important implications for comput-
ing (i. Suppose that, conceptually, a gradient search is to be
This Section will combine the results of the preceding two
used to compute
Sections and concentrate on the right-hand side inequality in
property (k) of Section 4. Suppose in this Section that in the inf am
2/i-tuple ^ rrij = 1 for each / (there are no repeated blocks).
Then reduce Jf to an «-tuple 3F = {kx,. . . , kn) and let with min(coV2) serving as a gradient when there are clustered
K=Xkj. Define DMn ^J?(K) by singular values. This search can proceed until a local minimum
D{x) = diag (exp O i ) / , exp ( x 2 ) / , . . . , exp (xn)f) is found, i.e. 0GcoV 2 . [As was mentioned earlier, amax
(DMD'1) is convex in D, and so such a local minimum would
ForM o e^?(K), define M: Rn -+^{K) by be global.] If, in fact, 0 G V2 at this point, then, by theorem 5,
Ii has also been found. Unfortunately, it is not generally
M{x) = D(x)M0(D(x)yl = D(x)M0D(-x) true that 0 G coV2 implies 0 G V2 •
Theorem 6 implies that, for the case n < 3 (three or fewer
and let amflJC (x), Ui,Vi,Vlt V2 etc. be defined as in Section 4 blocks), V2 = coV2, and so such a gradient search would (at
for M(x). Let A = f/j and £ = £/2 and assume that M is least conceptually) always yield n. This result is rather remark-
normalised so that amajc (0) = 1. Then write able since the optimisation involves only n — 1 variables,
regardless of the size of the blocks. Furthermore, in view of
theorem 3 and its corollaries, the computation of the general-
ised gradients for these cases should be quite straightforward.
A = a2 . . . c r ] = This provides the desired technique for exact analysis of
multivariable systems with simultaneous input/output pertur-
bations (i.e. two blocks), as well as more general situations.

and similarly for 5 in terms of {bf} and {fy}. Here a / G C K x l , 7 Examples and computational experience
a / G C f e / x r (similarly for 5 ) . It is then easy to verify that Two numerical examples and a discussion of computational
experience are contained in this Section. The examples are
intended to be purely illustrative, and no particular signifi-
Theorem 5 cance should be attributed to them.
= 1 if, and only if, 0 G V2(M). The first example uses a 3 x 3 matrix for which n will be
computed for five different structured perturbations. The
Proof nominal matrix is
0GV 2 (M)if, and only if:
(i) there exists.y G p r such that 1 1.
\-2i -10
2
y*Hjy = 0 for all/
if, and only if, M = 24 + 6/ 3/ 60 - 8 0 /
(ii) WffyW = l|a,VH for all/ 6 1
2 + 2/
if, and only if, 5 5 V
(iii) for all /, there exists Uj G U(kt) such that and the five uncertainty structures are shown in Table 1. The
perturbation in case 1 is block diagonal, and so ji can be
computed directly from M. In cases 2—5, however, the matrix
if, and only if,
must be rearranged to make the perturbation block diagonal.
(iv) there exists f/G ^such that By = UAy This can always be done by pre- and postmultiplying M by
(U = dtog(UltU2,...,Un)) appropriate matrices, denoted here by L and R, respectively.
if, and only if, The last two columns in Table 1 give JTfor that case and the
(v) B*UAy = y corresponding /i (LMR).
if, and only if, Cases 1—3 have three or fewer blocks, and n was computed
(vi) there exists 0 =£ x G C K such that using the approach suggested in Section 6. Cases 4 and 5 have
more than three blocks, and n was computed using an algo-
AB*Ux = x rithm based on theorems 1 and 2. The global maximum was
if, and only if, found for case 4, as was verified by computing infc mojc
(vii) MUx = x (since omax (M) = 1) (DLMRD~X) and obtaining agreement to within the accuracy
if, and only if, of the algorithms' termination conditions (in this case, four
(viii) p(MU) = 1 digits for single precision).
Therefore, using property (k) from Section 4, n(M) = 1 if, and This example illustrates that fi depends heavily on the
only if, 0GV 2 (M). assumed structure of the uncertainty. This should not be
interpreted as a limitation on its usefulness. On the contrary,
Theorem 6 it emphasises the importance of having a method for analysing
For n < 3, a ^ ( M ) = ii(M) if, and only if, 0 G coV2. matrix perturbation problems that maintains structure. The
Proof next example illustrates how a particular uncertainty structure
For any «, V2 depends only on « — 1 variables, since V2 could naturally arise in a feedback system.
lies in the (n — l)-dimensional subspace of R n orthogonal to The following example was originally constructed by the
(1, 1,. . . , 1); i.e. D(x)M0D{ — x) is constant along any line author to illustrate that loop-at-a-time analysis was inadequate
parallel to the line through 0 and (1, 1,. . . , 1). For n < 3, by for studying simultaneous variations in multiloop systems and
the second corollary to theorem 3, V2 = coV2, and the result to introduce the use of singular values. An example was pub-
then follows from Theorem 1. lished by the author [6]. The actual design analysed here was
IEEPROC, Vol. 129, Pt. D, No. 6, NOVEMBER 1982 247
proposed by Sain et al. [15]. A thorough and lucid discussion bation of norm 1 and that all smaller perturbations can be
of this example, including Sain's design, can be found in tolerated without instability. Unfortunately, this provides very
Reference 10. little information (other than the upper bound) on the toler-
Consider the feedback configuration in Fig. 1, where the able level for simultaneous perturbations; i.e. these maximum
nominal plant P has a transfer function singular values are a lower bound for n for simultaneous
perturbations.
9 -10 In order to analyse the system for simultaneous variations,
s+ 1 s+ 1 it is rearranged to isolate the A,s as a block-diagonal pertur-
Pis) = bation in standard feedback configuration, as in Fig. 3. For
-8 9 this example, one choice is
s+2

and M =
-(I+PK)~lP (I + PK)~lPK
9(s+l) 10 (s + 2)
1 with
K(s) =
0.0159s
8 (s + 1) 9 0 + 2) A! 0
A =
0 A2
The feedback compensation K is the product of the two
elements of the compensator in References 15 and 10. Only
the product K affects the system's feedback properties. M
r
p 1* A ,,

1 K A

Fig. 3 Standard feedback configuration

Fig. 1 Block diagram for example design
If the block structure were ignored at this point, then amax
For this example, the nominal plant is considered to (M) (the upper curve in Fig. 2) would provide a tight bound
have simultaneous multiplicative perturbations [1] at the on tolerable perturbations A. Unfortunately, amax{M) pro-
input and output. In applications, such perturbations might vides only a conservative bound for the block-diagonal case.
arise from actuators, sensors and unmodelled dynamics. The For the structure as given, the corresponding M is plotted
perturbations could then be weighted to reflect the fact in Fig. 4. This plot has several interpretations. The simplest
that the level of uncertainty varies in frequency and direction. is that the system can be stabilised by simultaneous pertur-
This example has no physical motivation, and so for simplicity bations with norms approximately equal to 0.1, and smaller
this system will be evaluated for robustness with respect to perturbations may be tolerated without instability. The
unweighted size of the norm-bounded perturbations. It is bounds in Fig. 2 were almost useless for this example. More
desired to compute the smallest norm-bounded perturbation generally, it is quite easy to construct examples where the gap
(as a function of frequency) which produces instability. for both bounds is arbitrarily large.
If only one perturbation is considered at a time, then ii The relatively large value of n in this analysis should not be
is simply the maximum singular value of the transfer function considered as an indictment of this design. It merely indicates
(I + PK)-lPK or (I + KPylKP, for an output or input that the design has poor margins relative to this particular
perturbation, respectively. These two transfer functions uncertainty structure; with respect to another structure they
happen to be the same for this example, and their maximum may be much better (for example, the margins for uncertainty
singular value is plotted as the lower curve in Fig. 2. This on just one side may be good). A design that tolerated
implies, for example, that there exists a destabilising pertur- larger simultaneous variations of the type considered here

10

10 10

9 i

-2
10 10"2 Id 2 - 2
10"1 1 10 X)2 K)3 10 ID"' 1 10 10 10 10
frequency, r a d / s frequency, r a d / s
Fig. 2 Singular-value bounds Fig. 4 Plot of M for example design
248 IEEPROC, Vol. 129, Pt. D, No. 6, NOVEMBER 1982
 Table 1 : Uncertainty structures

Case

(1.1.1. 10.24

1.1.D

A, 0 0 1 0 0 1 0 (1.1. 4.65

0 0 0 0 1 0 1 1.1)

0 0 0 p 0_

1 0 0 1 0 0 0 (1,1.1. 102.8

1 0 0 0 1 1 0 1,1.2)

0 1 0 0 0 0 1

0 0 1

A, 0 0 (1,1. 1, 1, 10.8

A
2 A, 0 1, 1,1. D

0 0 A

A, 0 0 (1,1,2, 10.6

A, A3 0 1,1.1)

0 0 A

would have to sacrifice some other aspect of performance. This is suggestive but, of course, not conclusive. As expected,
There was no physical basis for the uncertainty: it was chosen for three dimensions it is always 1 to within reasonable nu-
to be illustrative. Thus the results should not be interpreted merical error.
too broadly. Because the computer programs are experimental, i.e. con-
This example reiterates the point that vastly different taining many diagnostics and obviously inefficient code, it is
numbers are obtained, depending on the assumed structure of impossible to draw any meaningful conclusions about com-
the problem. For practical problems, the structure is dictated putational speed. As a single data point, it took approximately
to a large degree by physical reality and engineering con- three times longer to compute the curve in Fig. 4 as it did
straints. The extent to which a design engineer can capture and to compute the bounds in Fig. 2.
handle the natural structure determines the extent to which
the conclusions based on any analysis are relevant to the
practical problem. It is hoped that the ideas introduced in this 8 Summary and conclusions
paper and illustrated in the examples will make a contribution This paper has introduced a general approach for analysing
towards better techniques for handling structured uncertainty linear systems with structured uncertainties based on a new
in feedback systems. generalised spectral theory for matrices. This basic theory
An important consideration in the application of n is a addresses the norm-bounded perturbation problem with
numerical software. In as much as the ideas in this paper are arbitrary structure. The strongest results are for perturbations
relatively new and have not been published before, it may be with three or fewer blocks, for which (conceptual) algorithms
some time before reliable numerical software (as in Unpack with guaranteed convergence were proposed. One application
or Eispack) is available to compute n. The computational of these results is a generalisation of standard singular-value
experience to date is most encouraging, however. Programs analysis techniques for multivariable feedback systems to treat
have been developed to compute both bounds in property simultaneous input/output uncertainty (i.e. two blocks).
(k) of Section 4. Recall that the lower bound is always an These results are merely a beginning, and much more work
equality (see Section 4), but the global maximum may be remains to be done. For example, existing multivariable
difficult to find. The optimisation problem in the upper bound control methods provide little more than minor extensions of
is convex, but it is only guaranteed to yield n for three or SISO techniques. Initial study indicates that consideration
fewer blocks (see Section 6). of multiple simultaneous perturbations leads to wholly new
The ratio between the lower and upper bounds produced phenomena, the explanation of which will provide a far deeper
by these new programs has been computed for over 50 000 understanding of multiloop feedback systems. Linear multi-
pseudo-randomly generated matrices of dimension three to variable control theory will need to be thoroughly re-examined
ten, mostly with scalar blocks. The worst-case ratio was in this new light. It is hoped that the more general results in
approximately 0.95, although examples have been constructed this paper could also provide the beginning of a nontrivial
analytically where the ratio was ^ 0 . 8 5 . It is interesting to theory of decentralised control and/or large-scale systems,
note that the ratio seems not to decrease after four dimensions. where use of structural information is essential.
IEEPROC, Vol. 129, Pt. D, No. 6, NOVEMBER 1982 249
9 Acknowledgments 12 WILKENSON, J.H.: 'The algebraic eigenvalue problem' (Clarendon
Press, 1965)
Many people contributed to this paper, but I would par- 13 RELLICH, F.: 'Perturbation theory of eigenvalue problems' (Gordon
ticularly like to thank Dr. Joe Wall, of Honeywell Systems and & Breach, 1969)
Research Center, for his help throughout the research and 14 LUENBERGER, D.G.: 'Introduction to linear and nonlinear pro-
writing. Jim Freudenberg of the University of Illinois and gramming' (Addison-Wesley, 1973)
15 SAIN, M.K., MA, A., and PERKINS, D.: 'Sensitivity issues in
Honeywell SRC made a major contribution to the implemen- decoupled control system design'. Proceedings of Southeast sym-
tation of the algorithms used in the examples and offered posium on system theory
many useful suggestions regarding the paper.
This work has been supported by Honeywell internal 11 Appendix
research and development funding, the US Office of Naval
Reserach under ONR research grant N00014-82-C-0157, Let {Hj}f=1,/, andP r be defined as in the paragraph preceding
and the US Air Force Office of Scientific Research grant theorem 3, and let V2 =f(Pr). To compute min(coV2),
F49620-82-C-0090. consider the following algorithm to produce a sequence
This work is in the public domain in the USA.
(a) Pick any vx e Pr and let xx = f{vx).
(b) Inductively definex n = min(co[jcn_j ,f(vn-i)]).
10 References
(c) Then vn = any unit eigenvector for X m {
1 DOYLE, J.C., and STEIN, G.: 'Multivariable feedback design:
Concepts for a classical/modern synthesis', IEEE Trans., 1981, This produces a sequence {xn} that has nonincreasing norm,
AC-26, pp. 4-16 and thus {II*JI} converges. Therefore there exists a sub-
2 POSTELTHWAITE, I., EDMUNDS, J.M., and MacFARLANE, sequence {xnit} such that both {xnk} and {vnk} converge to,
A.G.J.: 'Principal gains and principal phases in the analysis of say, JC0 and v0, respectively. Note that y =f(vn) minimises
linear multivariable feedback systems', ibid., 1981, AC-26, pp. 32-46
the min<jcn,.y> and (xnJ(vn)) = \mirl(2xjnHj). Then, by
3 SAFONOV, M.G., LAUB, A.J., and HARTMANN, 'Feedback 1
properties of multivariable systems: The role and use of the return taking limits, (x0f(v 0)) = \min (?:x H
0 j) and y=f(v0)
difference matrix', ibid., 1981, AC-26, pp. 47-65 minimises min<x0 ,y).
4 CRUZ, J.B., FREUDENBERG, J.S., and LOOZE, D.P.: 'A re- Now suppose x0 ¥= min (coV2) (leading to a contradiction).
lationship between sensitivity and stability of multivariable feedback Then by definition of the convex hull, <x0 ,f(v0)) <\\x0 II2 and
systems', ibid., 1981, AC-26, pp. 66-74
5 LEHTOMAKI, N.A., SANDELL, N.R., Jr., and ATHANS, M.: ||x 0 ||-I|min(co[x 0 ,/(Wo])li>efor s o m e e > ° - B v dist-
'Robustness results in linear-quadratic Gaussian based multivariable ance of a convergent subsequence, there exists TV sufficiently
control designs', ibid., 1981, AC-26, pp. 75-92 large such that \\xN-xo\\<e/2 and \\f(vN)-f(v0)\\<
6 DOYLE, J.C.: 'Robustness of multiloop linear feedback systems' e/2. Then
17th IEEE conference on decision and control, San Diego, USA,
Jan. 1979 ||xoll>e+||min(co[x 0 ,/(wo)])ll
7 DOYLE, J.C.: 'Multivariable design techniques based on singular
value generalisation of classical control' AGARD lecture series >e+ [mm(co[xN,f(vN)])-e/2]
117 on multivariable analysis and design techniques. Sept. 1981
8 WALL, J.E., DOYLE, J.C., and HARVEY, C.A.: 'Tradeoffs in the = e/2+ 1 1 * ^ ||
design of multivariable feedback systems'. Proceedings of 18th
Allerton conference on communication control and computing, This contradicts the fact that {||x n ||} is nonincreasing, and
Oct. 1980, pp. 715-725
9 DOYLE, J.C.: 'Limitations on achievable performance of multi- thus {xn} converges to min (coV2).
variable feedback systems'. AGARD lecture series 117 on multi- In this algorithm f(vn) is an intersection point of V2 and
variable analysis and design techniques, September. 1981 the support hyperplane perpendicular to xn. There are,
10 FREUDENBERG, J.S., LOOZE, D.P., and CRUZ, J.B.: 'Robustness of course, a number of other ways that this point could be
analysis using singular value sensitivities', Int. J. Control, 1982,
pp.95-116 used to obtain a convergent algorithm. The particular algo-
11 KATO, T.: 'Perturbation theory for linear operators' (Springer- rithm outlined here is rather crude in not taking full advantage
Verlag, 1976) of the structure of V2.

250 IEEPROC, Vol 129, Pt. D, No. 6,.NOVEMBER 1982