292 TRANSACTIONS
IEEE ON AUTOMATIC CONTROL, JUNE 1968
161 H. J. Kusbner. 'On the differential eauations satisfied by conditional prob-
ability densities of Markov processes. with applications." RIAS. Baltimore, bid.
Tech. Rept. T R 6 4 4 , March 1964.
171 R. S. Bucy. "Konlinear filtering,DI E E E Trans. A u h a t i c ControZ (Corre-
spondence), vol. AC-IOLp. 198. April 1,965.
161 J. K. Lubbock. The optmllatlon of a class of nonlinear filters." Proc. I E E
(London), vol. 107, pt. C! pp. 60-84, March 1960.
191 N. K. Sinha, "Optmum nonlmear filters for random signals." I R E Inlmraf'L
Cono. Rec.. vol. 10, pt. 2. pp. 88-91. March 1962.
1101M. Lal. 'A note on optimization of a class of nonlinear systems.* IEEE
Trans.Automatic Control (Correspondence). v?1. AC-11, pp. 137-138. January 1966.
In1 A. H. Nuttall. 'Theory and appllcatlon of the separable class of random
ProcesSes."M.I.T. Electronics Research Lab.. Cambridge, Mass. Tech Rept. TR 343,
hlay 1958.
1191 J. L. Broxn. Jr.. 'On a cross-correlation property for stationary random
processes,' I R E Tram. Ixfonnafiun Theory. vol. IT-3, pp. 28-31. March 1957.
A Simplified Viewpoint of Hyperstability
BRIAN D. 0. ANDERSON, MEMBER, IEEE
Abstracf-The dehition of a hyperstable system according to
Popov is given a network theoretic interpretation, and proofs are
presented outlining the connection between passive and hyperstable
systems.
I. INTRODUCTION
In 1963, Popov introduced the notion of hyperstability for linear
single-input single-output time-invariant systemsI1l and subse-
quently extended his ideas.PI The key feature of his results is the
existence of a n equivalence between hyperstable systems and sys-
tems which can be described by a transfer function that is positive
real, Le., the transfer function is like that which maps the input into
the output of a passive system.
Using well-known ideas from the theory of passive systems and
networktheory,Popov's statementsareinterpreted in terms of
energy flow and excitation of energy storage elements and are pre-
sented in what are believed to be more transparent proofs of his
theorem. These results are also presented in multivariable format
for greater generality.
As Popov himself points out, stabilityquestions associated with a
system containing memoryless nonlinear feedback can be interpreted
as questions of hyperstability, thus emphasizing the tieup between
positive real functions, or matrices, and regulating systems withnon-
linear feedback.[']
11. DEFINITION
AKD NOTATIOK
Consider systems described by state space equations of the form
*=Fx+Gh
y=H'x+Ju
where throughout the paper i t is assumed that the pair [ F , G] is
completely controllable[~land [F,H ] completely observable.[31 I t
is also assumed that thevectors u and y have the same dimension.
Hyperstability is a property of the system which requires the
state vector to remain bounded if the inputs u are restricted to be-
longing to a subset of the set of all possible inputs. The subset of in-
puts u is defined by those u ( - )which satisfy for all T
Here 6 is a positive constant depending on the initial state x ( 0 ) of the
system but independent of the timeT and thesymbol 11 denotes the 11
Euclidean norm.
The system (1) is termed "hyperstable" if for any u ( . ) in the
subset defined by (2) the following inequality holds for some positive
constant K and all t :
Manuscript received May 9, 1967; revised October 24. 1967, and Febru?ry 23.
1968. Tbis work, ma: supported by the Aust.dian R?~ear$~ GranJs Cpmmlttee.
The author IS \mth theDept. of Electncal Engmeenng. Unlvershty of New-
castle. Kew South Wales, Australia 2308.
The system is termed"asymptoticallyhyperstable" if for any
u ( . ) in the subset defined by (2) which is also bounded the inequality
(3) holds together with
lim z(t) = 0. (4)
f+m
An important point is that the definitions of hyperstability and
asymptotic hyperstability are "coordinate free" as map easily be
checked, that is, if a new basis is chosen fo' the state space, causing
replacement of x by e= Tx, and of F by F= T F P ' , etc., then hy-
perstability in the first coordinate system implies hyperstability in
the second and vice versa. The same holds, of course, for asymptotic
hyperstability. Consequently, hyperstability and asymptotic hyper-
stability are properties of the transfer-function matrix
Z(S)= J + H ' ( d - F)-'G (3
a s distinct from being properties of a particular minimal realization
of Z(s). A minimal realization of Z ( s ) is a quadruple { F, G, H , J }
such that ( 5 ) holds with [F,G] completely controllable and [F, H ]
completely observable.
A rational transfer-function matrix as in (5) is termed "positive
real" if the following three conditions are satisfiedL41:
1) Z ( s ) has real elements for real s,
2) The elements of Z(s) have no poles in Re[s] >O and poles on
the ju a x i s are simple, and such that the associated residue
matrix is non-negative definite Hermitian,
3) For any real value of w such that no element of ZGw) has a
pole for this value, Z(jw)+Z'*(jw) is non-negative definite
Hermitian.
For real rational Z(s), 1) and 3) may be replaced byP1:
2) Z(s)+Z'*(s) is non-negative definite Hermitian in Re[s]>O.
111. THERESULTS OF POPOV
Popov's two results, generalized to the multiple-input situation,
are asfollows.
Theorem 1
A necessary and sufficient condition for the transfer-function
matrix Z(s) of ( 5 ) t o define a hyperstable system is that Z(s) be
positive real.
Theorem 2
A necessary and sufficient condition for the transfer-function
matrix Z(s) of ( 5 ) to define an asymptotically hyperstable system is
that Z ( s ) be strictly positive real, in the sense that Z ( s ) is positive
real, all poles of Z(s) should lie in the half-plane Re[s]<Oand
Z ( j w ) +Z'*(j w ) should be positive definite Hermitian for all real w.
These two theorems have interesting interpretations in terms of
energy constraints in passive systems. 4 s is well known, appropriate
choice of input variable u and outputvariable y of a passive system
leads to a transfer-function matrix for the system which is positive
real with JoTu'(t)y(t)dt having the interpretation of the energy de-
livered into the system from the external world over the timeinterval
[0, T].141\Vith a correct choice of coordinate basis, it is also possible
to interpret x ( t ) as having entries which are measures of the excita-
tion of the energy storage elements of the system.
Of the set of all possible inputs, the inputs in the subset defined
by (2) are those for which the energy flow into thesystem up till time
T is bounded by the maximum excitation over [0, T ]of the energy
storageelementswithin the system. Hyperstability requires, [see
(3)1, that for such inputs themaximum excitations should be bounded
in terms of the initial excitations. The preceding interpretation in
fact will enable us to provide proofs of Theorems 1 and 2.
IV. PROOFS
OF THE RESULTS
Positive Realness Implies Hyperstabdity
Since Z(s) is positive real, one minimal reahation { F,G, H, J } of
Z(s) is provided through the medium of an interconnection of pas-
SHORT PAPERS 293

siveinductors, passive resistors, idealtransformers, andgyrators
which constitute a network synthesis ofZ(s).[.‘l For it is basic t o
network theory that associated witha positive real Z(s) there is
always a network made up of such passive components, the port
impedance of which is precisely Z(s). Capacitors may be eliminated
by replacing them with a gyrator andinductor. As is also well known,
a suitable state-variable vector x for Z(s) is provided by the instan-
taneousinductorcurrents. Moreover, as canalways be assumed,
this state vector is of minimum dimension if the number of reactive
elements in the synthesis of Z(s) is minimal.
If all inductors have value unity, which can always be arranged
using transformer normalizations, the stored energy in the network
a t time t is fr’(t)x(t),and passivity requires that for all T
JOTZ~’(t)Y(i)ar + +X’(O)X(O) 1 +X’(T)S(T).
In other words, the sum of the initial stored energy of the network
and the energy directed into the network terminals over [0, T ] must
be no less than the stored energy a t time T. The “greater than” sign
in (6) will hold if the network dissipates energy during [0, TI.
If allowable inputs are those satisfying ( 2 ) , then, for all T ,
w ( T ) ~ ( T )I +x‘(o)x(o) + a l l j ~ ( o ) l l i ~ ~ ~ ~ ~ ~ ~ ( ~the
matrices P = P‘>O, L,and W O such thatP1
From (7) it follows that Ilx(t)l/ is bounded, for if not, for a sequence of
-.
values of T, T = T I ,Tz, , the left-hand side of (7) would increase
faster than the right-hand side, so that for sufficiently large ~ ~ x ( T ~ ) ~ ~
the inequality would fail. Let Af denote supost+ Ilx(t)ll. Then (7)
yields
term is zero and thefirst term alone predominates; then the inequal-
ity ( 1 2 ) means that the integral on the left of (15) diverges to - m.
Consequently the input u ( t ) of (13) obeys the inequality ( 2 ) . At the
same time, it is clear that Ilx(t)ll is unbounded, increasing more or
less as e”O*. Thus lack of hyperstability is demonstrated.
Strict Posdive Realness Implies Asymptotic Hyperstability
Note first thatztrict positive realness implies that for asufficiently
small positive u, Z(s)=Z(s-u) is positive real: condition 1) of the
positive real definition is satisfied for i ( s ) , for arbitrary real u. Con-
dition 2) is satisfied for sufficiently small u because all poles of
Z(s) lie in Re s<O. ToA= that-condition 3) is satisfied, observe that
the non-negativity of ZGw) +Z’*(ju) is equivalent to the non-nega-
(6)
tivity for all w of a series of polynomials which are the minors of the
matrix d * d [ i ( i w ) +.?*(jw)], d being the lowest common denominator
of the elements of 80‘0).As u+O, these minors which approach the
corresponding minors associated withZ(jw) are strictly positive by as-
sumption. So for sufficiently small u, _the minors associated with
ZGw) are non-negative. Observing that Z ( s ) is given by
i(s) = J + H ‘ [ d - ( F + r1)I-G (16)
(f) ) l l positive
. realness of i ( s ) implies the existence of real constant
P(F uI) (F’ + + +
uI)P = - LL’ (17 4
PG H - LWo (17b)
W O ’ W O =J J’. +
(lfc)

w
I +IIX(O)llZ dll40)lll.1f+ (8)
I t will now be shown that V=x‘Px is a Liapunov function for the
system ( 1 ) with any bounded input 21 satisfying (2). Of course V is
or positive definite; also,
{if - ~ [ l l ~ ~ o ) 5
l l IlX(O)ll2
ll~ + dll40)lll~ (9) P = 2x’Pi = d(PF + F’P)x + 2x‘PGu
I fll.co)ll +dlld0)lll)~ = - 2adPx - (X’L - u’Wd)(L’x - WOU) + 2u‘y. (18)
whence Thus
x I Il~C0)ll+ 2~[ll.(O)lll (10) vI - 2av + 2u’y. (19)
or for all t
Setting W=pbV(X)dX and integrating (19) yields
Ilx(0ll I2(11~(0,ll + 6) (11)
which is a special case of (3). 05w < - 2uw + V ( t 0 ) + 2 JtIdydt.
0
Hyperstability Implies Positwe Realness
Now ( 1 ) is certainly hyperstable, thus (2) and (3) yield
I t will actually be proved that the negative statement, lack of
positive realness, implies lack of hyperstability. In particular, assume
that thepositive real definition of condition 2 ) does not hold, i.e., that
Z(s)+Z’*(s) fails to be non-negative definite Hermitian for some
value of s, for example, uo+jwo with u0>0. By continuity of the ele- for some constant RI and all t . Hence
ments of Z(.) in Re[s]>O, wg is assumed to be nonzero. Then there 0IJv 5 - 2 u w + V ( f 0 )+ R1. (22)
exists a vector 240 such that
The definition of 14’ and (22) imply W is bounded above, and that w
ue’*[zGo +jwo) + z’*(uO +jwo)]uo < 0. (12) is non-negative and bounded above. Hence @EL1,the space of func-
Moreover, there exists an initial state SOsuch that theresponse to the tions which are absolutely integrable on [0, a).The boundedness of
input x , following from hyperstability as distinct from asymptotic hy-
perstability which has yet to beproven, and the assumed bounded-
ness of u guarantee that ir in (18) is bounded. Hence f17 is uniformly
continuous. Since also TqEL’, it follows that m + O as t+ m, i.e.,
limb, x ( t ) = 0.
is given for t > O by
y(t) = Re Z(UO+jwo)u,@rejwof. (14) Asymptotic Hypedability Implies Stvid Positive Realness
Explicit calculation set out for example in Newcomb ([.‘I ch. 4) yields I t is a simple matter to modify the earlier argument that hyper-
stability implies positive realness. The strict positive realness can
fail in two ways, with positive realness still being maintained.
Either Z(s) can have a pole on t h e j u axis, or there exists a vector
up and frequency w o for which
+- +
uo wo
Z{+~Z’(~~O
u0)uoe + zd*[z(jw0) + zr*(jwo)juo= 0. (23)
where a is a constant, independent of time (the details of the calcula- I n the former case, there exists an initial state which is Liapunov
tion of (15) are of no concern here). For suitablychosen T the second stable but not asymptotically stable in the presence of zero input.
294 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, JUNE 1968

Since asymptotic stability is a requirement for asymptotic hyper-
stability, it is seen that failure of Z(s) t o be strictly positive real for
the first reason results in a lack of asymptotic hyperstability.
In the latter case, arguments like those used immediately after
(8) lead to the conclusion that there exists a n initial state xo and
sinusoidally varying (wa#O) or constant (w0=0) input u(t) for which
J,,Tu‘(t)y(t)dt is bounded; moreover x ( t ) does not decay to zero. Thus
with inequality (2) holding, the hyperstability is not asymptotic.
v. coNcLusIoNs
The interpretation of the hyperstabilityconcept in terms of
properties of passive systems as given in Section I11 is straightfor-
ward but does not seem to have been given before. In particular it
allows the writing down of (6), essentially by physical arguments.
This equation, whose simple deduction is not to be found in Popov,[’l
allows a quicker derivation of the result that positive realness im-
plies hyperstability.
The stability of systems with nonlinear memoryless feedback can
be studied with the aid of the hyperstability inequality (2). Further
stability results could possibly be deduced by searching for other
forms of feedback also allowing satisfaction of (2).
REFEREKCES
M. Popov. =Thesolution of a new stabilityproblem for controlled systems.”
111 V.
81 -.
A%fomatMnaa& Remote Conirol. vol. 24, pp. 1-23, January 1963.
Hyperstability and optimality of automatic systems with several
control functions,” Rm. Roumanian Sci. Tech.-Eledrotech. and Engrg.. vol. 9. no.
4, pp. 629490. 1964.
131 L.A. Zadeh and C. A. Desoer. Linear System Themy. New York: McGraw-
Hill, 1963.
141 R. W. Nemcomb. Linear Mdtiport Synthesis. New York: hfcGraw-Hill.
1966.
151 B. D. 0. Anderson. ’A system theory criterion for positive real matrices,”
S I A M J . Confrol.E-. A, vol. 5 , pp. 171-182, May 1967.

Correspondence
Comments on “Interpolation and where m
Extrapolation Schemes in Dynamic x = n-dimensional state vector
&<=quantization increment in ith
Programming” state variable
In the above paper, Bard* proposes a n u = control vector
alternative procedure for carrying out the t = stage variable
interpolations and extrapolationsrequired 6t = increment in stagevariable
by the state increment dynamic program- over which control is applied

ming procedure. The classical dynamic
programming computational procedure2
often requires excessive high-speed storage
in medium- t o large-size problems. State in-
crementdynamic programming3*4 provides
a method forreducing the high-speed storage
requirement without losing theother de-
sirable properties of the classical procedure.
In this correspondence, Bard’s methods are
compared to those given in Larson.3.4
The basic philosophy behind the state
increment dynamic programming procedure
is to compute the minimum cost at the next
state on the basis of values of the minimum
cost function in asmall neighborhood about
the present state and stage. To accomplish
this 6t, the increment in the stage variable t
over which each control u is applied, is
chosen4 as
,At 1 (1)
Manuscript received December 21. 1967.
1 Y. Bard, I E E E Trans. Arrtaatic Cmfrol(Shorf
Papers). vol. AC-12. pp. 97-99. February 1967.
2 R. E. BelIman, Dynamic Programming. Priice-
ton, h’.J.: Princeton University Press, 1961.
s R. E. Larson. “Dynamic pro,mmming with re-
duced computational requirements.’ IEEETram.
Aulomafic Control. vol. AC-10. pp. 135-143,April
1965.
4-. State I n c r m J DynamicProgramming.
hTmYork: American Elsevier Publ. Co., 1968.
Ai = increment in stage variable be-
tween successive computations
of optimal control
f.(x, 21, t ) = @xi/&) = rate of change of ith
state variable as determined
by system evaluation.
This choice of St ensures that the change in
any statevariable xi is a t most Axi, and that
the change in t is a t most At. The next state
thus lies on the surface of a hypercube in
(n+l)-dimensional x - t space, with dimen-
sion 2Axi along the y axis and At along the t
axis. The minimum cost a t this next state
can be computed from values at the quan-
tized states and stages on this surface. -4
complete computational procedure based on
this concept is described in Larson3 and dis-
cussed in more detail in Larson.‘
The variousinterpolation andextrap-
olation schemes are most conveniently dis-
cussed in terms of a one-dimensional ex-
ample. The computations at a given state
x* and stage t* are illustrated in Fig. 1. -4s
discussed in Larson,’ it is generally true that
minimum cost and optimal control have al-
ready been computed at either x*+& or
x * - A x for stage t*; for this example it is
assumed that the computations a t X”-LY,
t* have already been made.
There are threeranges of 6t that must be
considered. In region I, %<At andf(x*, ZL, t )
-
t r t*+ At
@ = POINT AT WHC
I H COUPUTATII*US
ARE CURRENTLY TAKIXG PLACE
0 = POINTS AT WICH MlNlLlUM COST
HAS ALREADY BEEN COMPUTED
Fig. 1. Ranges of dt over which interpolations and
extrapolations are required.
<O. The next state is thus x*-&, and the
corresponding value of the stagevariable
falls in the range t*<t<t*+At. In this region
all the procedures considered here determine
the minimum cost at the next state byinter-
polation in t. In the case of n state variables,
this valueisobtained by interpolationin
(x- 1) state variables and t4.
In region 11, 6 t = M For this case, all
procedures determine the minimum cost by
interpolation in x . In the n-dimensional case,
the interpolation is in n state variables, ex-
actly asin the classical procedure.
In region 111, t<At and f ( x * , u, t*)>O.
I t is in this region that there are differences
between the methods proposed by Larsons>*
and Bard.’ I t should be noted that, as
pointed out in Larson,3,4 the concept of the
preferred direction of motion can be used to