131 73 17MB
English Pages 352 [343] Year 1972
Lecture Notes in Mathematics A collection of informal reports and seminars Edited by A. Dold, Heidelberg and B. Eckmann, Z(Jrich Series: Mathematics Institute, University of Warwick Adviser: D. B. A. Epstein
294
Stability of Stochastic Dynamical Systems Proceedings of the International Symposium Organized by "The Control Theory Centre", University of Warwick, July 10-14, 1972 Sponsored by the "International Union of Theoretical and Applied Mechanics"
Edited by Ruth F. Curtain University of Warwick, Coventry Warwickshire/England
Springer-Verlag Berlin.Heidelberg • New York 1 9 7 2
A M S Subject Classifications (1970): 34F05, 3 4 H 0 5 , 6 0 H 1 0 , 93-02, 9 3 D 9 9 , 93E05, 93E10, 93E15, 93E99
I S B N 3-540-06050-2 S p r i n g e r - V e r l a g B e r l i n • H e i d e l b e r g • N e w Y o r k I S B N 0-387-06050-2 S p r i n g e r - V e r l a g N e w Y o r k • H e i d e l b e r g • B e r l i n
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically those of translation, reprinting, re-use of illustrations, broadcasting, reproduction by photocopying machine or similar means, and storage in data banks. Under § 54 of the German Copyright Law where copies are made for other than private use, a fee is payable to the publisher, the amount of the fee to be determined by agreement with the publisher. © by Springer-Verlag Berlin • Heidelberg 1972. Library of Congress Catalog Card Number 72-91895. Printed in Germany. Offsetdruck: Julius Beltz, Hemsbach/Bergsm
INTRODUCTION
The symposium on the "Stability of Stochastic Dynamical Systems'~ held at Warwick University, July 10 - ]4th, 1972, was sponsored by the International Union of Theoretical and Applied Mechanics, support we appreciate.
(IUTAN), whose
Following IUTAM policy, participation was by
invitation and was limited to around 60 so as to encourage lively discussion. A full list of participants is given~
The main theme of the symposium was the stability and other properties of differential equations with stochastic coefficients.
Both the general
mathematical aspects and applications were discussed.
The contents have been arranged according to the daily sessions, when a "key-note" speaker gave a one hour address to set the daily "theme". Other lectures were of twenty minutes duration.
As each contributor was
asked to submit a complete typed version of his address for the symposium~ these proceedings are in general more detailed than the actual talks in no need of further introduction.
Ruth F. Curtain,
1972
and are
V International Scientific Committee
Prof. J.A. SH~CLIFF (U.K.) Chairman Prof. P. BLAQUIERE (France) Dr. P. BRUNOVSKY (Czechoslovakia) Prof. K. ITO (Japan) Prof. R.E K A ~
(U.S.A.)
Prof. L. MARKUS (U.K. and U.S.A.) Prof. H. ROS~m3Roc~< (U.K.) Prof. v.v. SOLODOVNIKOV (U.S.S.R.) Prof. H.R. SCHWARZ (W. Germany) (IFAC Representative) List of Part icipan% s
Prof. S.T. ARIARATNAM
C aaada
Dr. L. ARNOLD
West Germany (FDR)
Prof. K.J. ~STROM
Sweden
Dr. M.C. AUMASSON
France
Dr. J.F. BARRETT
U.K.
Dr. A. BENSOUSSAN
France
Prof. P. BLAQUIERE
France
Prof. R. BROCKETT
U.S.A.
Dr. P. BRUNOVSKY
Czechoslovakia
Dr. HELGA BUNKE
DDR
Dr. P. CAINES
U.K.
Dr. J.M.C. CLARK
U.K.
Dr. RUTH CURTAIN
U.K.
Dr. M. DAVIS
U.K.
Prof. J.L. DOUCE
U.K.
Prof. T. DUNCAN
U.S.A.
Dr. R. ELLIOTT
U.K.
Prof. A. FRIED.MAN
U.S.A.
Dr. A.T. FULLER
U.K.
Dr. C.J. HARRIS
U.K.
Prof. P.J. HARRISON
U.K.
Prof. U. HAUBSMANN
Canada
Mr. D.B. HERNANDEZ-CASTANO
U.K.
Dr. M. HUGHES
U.K.
Prof. K. ITO
U.S.A.
VI Dr. J.G. JAMES
U.K°
Prof. F. KOZIN
U.S.A.
Prof. H.J. KUSHNER
U.S.A.
Prof. J.A. LEPORE
U.S.A.
Dr. P. MANDL
Czechoslovakia
Prof. L. MARKUS
U.S .A./U.K.
Prof. D. ~[AYNE
U.K.
Prof. R. MONOPOLI
U.S.A.
Dr. T. MOROZAN
Rumania
Dr. L.A. MYSAK
Canada
Dr. T. NAKAMIZO
Japan
Dr. P.C. PARKS
U.K.
Dr. A. PRITCHARD
U.K.
Prof. P. SAGIROW
West Germany (FDR)
Prof. P. SETHNA
U.S.A.
Prof. J.A. SHERCLIFF
U.K.
Dr. Ing. W. WEDIG
West Germany
Prof. P. WHITTLE
U.K.
Prof. J.C. WILLEMS
U.S.A.
Prof. J.L. WILL.IS
Belgium
Dr. D. WISHART
U.K.
Dr. J.A. Z ~ A N
Austria
Local Organizing Committee (University of Warwick)
Prof. J.A. SHERCLIFF Prof. L. MARKUS Dr. RUTH CURTAIN Dr. P.C. PARKS
(Chairman)
CONTENTS
I. M A T H E M A T I C A L K.
FOUNDATIONS
ITO:
Stochastic
Differentials
of Continuous
Local
Quasi-Martingales.
.
P. MANDL: An Application
of It$'s F o r m u l a
to Stochastic
Control
Systems
. .
8
A. P R I E D M A N : ~ S t a b i l i t y and A n g u l a r B e h a v i o r of Solutions of Stochastic Differential Equations . . . . . . . . . . . . . . . . . . . . . .
14
T. NOROZAN: Boundedness
!I. K.J.
IDENTIFICATION
for Stochastic
Systems
. . . . . . . . . .
21
AND FILTERING
~STROM:~
System D.Q.
Properties
Identification
. . . . . . . . . . . . . . . . . . . . . .
35
NAYNE:
P a r a m e t r i z a t i o n snd I d e n t i f i c a t i o n of Linear M u l t i v a r i a b l e Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
A. ~ENSOUSSAN: O p t i m i z a t i o n of Sensor~' L o c a t i o n in a D i s t r i b u t e d F i l t e r i n g Prob]e~ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . T.E.
62
DUNCAN:
Some B a n a c h - V a l u e d
IIl.
STOCHASTIC
H.J.
KUSHNER:~
Stochastic
Processes
STABILITY ,,
with A p p l i c a t i o n s . . . . . . . . . .
I
Stability . . . . . . . . . . . . . . . . . . . . . . .
U.G. HAUSSNANN: S t a b J ] i z a t i o ~ of L i n e a r
85
Systems
with N u l t i p l i c a t i v e
Noise
97
. .125
VIII J.L.
WILLENS:
L y a p u n o v F u n c t i o n s and Global F r e q u e n c y D o m a i n S t a b i l i t y C r i t e r i a for a Clsss of S t o c h a s t i c F e e d b a c k Systems . . . . . . D.J.G.
139
JAMES:
Stability
of N o d e l - R e f e r e n c e
Systems
with Random
Inputs
....
147
W. WEDIG: R e g i o n s of I n s t a b i l i t y for a Linear S y s t e m w i t h R a n d o m Parametric Excitation . . . . . . . . . . . . . . . . . . . . . T. N A ~ M I Z O
and Y. SAWARAGI:
A n a l y t i c a l Study on n-th O r d e r L i n e a r S y s t e m w i t h S t o c h a s t i c Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . .
IV.
160
STOC}~STIC
STABILITY
173
!!
F. KOZIN: Stability Ch.J.
of the L i n e a r
Stochastic
System
. . . . . . . . . . .
P_&RRIS:
The F o k k e r - F l a n c k - K o l m o g o r o v E q u a t i o n in the A n a l y s i s of Nonl i n e s r F e e d b a c k S t o c h a s t i c Systems . . . . . . . . . . . . . . . J.A.
LEPORE
and R.A.
BROCKETT
Average P.R.
Value
230
STOLTZ:
S t a b i l i t y of L i n e a r C y l i n d r i c a l Shells S u b j e c t e d to S t o c h a s t i c Excitations . . . . . . . . . . . . . . . . . . . . . . . . . . R.W.
186
and J.C.
WILLENS:
Criteria
for S t o c h a s t i c
Stability
. . . . . . . .
239
252
SETHNA:
U l t i m a t e B e h a v i o u r of a Class of S t o c h a s t i c D i f f e r e n t i a l Systems D e p e n d e n t on a P a r a m e t e r . . . . . . . . . . . . . . . . . . . .
273
H. BUNK-E: Stable P e r i o d i c S o l u t i o n s of W e a k l y N o n l i n e a r S t o c h a s t i c Differential Equations . . . . . . . . . . . . . . . . . . . . .
283
V. A P P L I C A T I O N S S.T. A R I A R A T N A M : S t a b i l i t y of M e c h a n i c a l Systems u n d e r S t o c h a s t i c P a r a m e t r i c Excitation . . . . . . . . . . . . . . . . . . . . . . . . . . .
291
IX L.A. ~ S A K : Waves in a Rotating Stratified Fluid with LsteraJly Varying Random Inhomogeneities . . . . . . . . . . . . . . . . . . . . .
303
P.S. SAGIROW:
The Stability of a Satellite with Parametric Excitation by the Fluctuations of the Geomagnetic Field . . . . . . . . . . . . .
VI. OTHER CONTRIBUTIONS
(not presented
311
at the meeting)
V.G. KOLOMIETZ: Application of Averaging Principle in Nonlinear Oscillatory Stochastic Systems . . . . . . . . . . . . . . . . . . . . . . . .
317
V.V. SOLO DOVNIKOV and V.F. BIRIUKOV: Optimization of Multi-Dimensional Stochastic Systems and Stability of Solutions . . . . . . . . . . . . . . . . . . . . .
324
These were key-note
lectures
STOCHASTIC DIFFERENTIALS
OF CONTINUOUS LOCAL QUASI-MARTINGALES* A
KIYOSI ITO Cornell University The purpose of this paper is to unify the known results on stochastic differentials
in terms of differential quasi-martingales
so
that we can understand the intuitive meaning more easily. I. Let
[~t'
~-algebras. by
E t.
NOTATIONS AND DEFINITIONS
0 ~ t ~ i}
be a right continuous increasing family of
The conditional expectation relative to ~ t
is denoted
We consider the following classes of stochastic processes. is the class of all stochastic processes adapted to {~t}and
having continuous ~is
sample paths.
the class of all continuous
[ ~ t ] , namely all processes of martingales
X ~C~
local martingales
relative to
such that there exists a sequence
X n c g~ satisfying P(Xn(t ) : X(t) for every t) ~ i.
is the class of all
X ~
whose sample paths are of bounded
variation almost surely. is the class of all continuous all
X
local quasi-martingales,
such that X = M + L, M c ~ ,
L ~ g;
see D.L. Fisk [2], S. Orey [7] and K.M. dR
Rao [8].
is the class of all randominterval functions
(stochastic differentials)
induced by
dQ(s,t) Similarly for
d~,
d~
and
Q
= Q(t)-Q(s), d~.
*This work was supported by NSF GP-28109.
as follows: (s < t).
dQ
namely
2. Let us introduce DEFINITION
STOCHASTIC
DIFFERENTIALS
three basic operations
on stochastic
differentials.
dQ2 = edQ1 i f and only i f
i.
t
dQ2(s,t ) ~ Q2(t)
=f
-Q2(s)
dQl(e )
(s < t).
S
The integral is defined in the same way as stochastic Q1 = M1 + LI' such that both that
MI ~
and
LI c~
Ml(t )
and
EO dQl(O)=iAl~mo
~
S
i=l
A = {0 = e I < elO.
~ o, for ~ e R~, we have
~c(~(t,to,~))~ K(to,~)+
Zl b(r)P I where b(r)= inf c(x) J~l>.n.
Let Ar=Isup I~(t,to,~)l > r~ ,At,r=I Jx(t ~to,~)~ > r) = ~?~At, r Since F(Ar~:iim
P(Bm,r) , iim
b(r)P(Bm, r) ~K(to~)+ Theorem 2 is proved.
b(r)= ~,~ and , ~'or a l l
m >. to
and Bm,r=
25 Theorem 3o IZ If)
tim
in_f V(b,x) = be
(2) ~V(t+l,~[t+L,to,9))~_~EV(t,x(t,to,~))+F t oe where
z or all t e ~,
N, t ~ to, ~ e S o n ~
~ e(o,1), T¢[~ ~)
bhen the solubions of ( ~ ,N,~)
aze ablimatel,y bounded in probabill-
Proof . Let
t o ~ N, ~ > o,
~ C
S(~),qj~
From condition (2) it follows that
+ t-to
K(to,~)+
t-to-Z)
1-/~ for a l l t>~ t o , t 6. N.
ience
a(r)~ l
+ ....
~x(~,~o,~)l > r ]
o such that V(t,x) ~
for t 6 N
ixt >~ R o (2) EV(t+l,x(t+l,to,~))e~EV(b,~(t,to,~)) t >~ to, R r= S o D ~
then the solutions of ~
where , N,~)
for all teN,
toeN ,
~(o,i) aze equ.iul~.imatelF bounded in pro-
babilit~ Proof. Let to 6 N , c o, ~ 6 S(~.)G3~
26
We have $ ptl~(t,to,~)
I > ~ot~- E V ( t , ~ ( t , t o , ~ ) ) ~ " ~ t - t o
K(~o,~ ), for aZZ
teN, and thus, Theorem
#
t >. t o
is proved.
Oonsider the It6 system dx (t)=g( t ,x(t) )dt +B ( t ,~( t )dw( t )
(2)
where g: [o,~,)x R ~ --> R ~ is s continuous function, B(t,m) is a ~xpmatrix whose elements blm(t,x) are continuous functions and w(t) is the p dimensional process of Brownian Let W:[Os~)x
R~-~[o~)
motion,
be a function of class C 2, and
~heo~em 5. I_~ (~)
lira inf W(% ]~)= ~, &~
t >O
(2) ~W(t,x) ~ ~(t)
for all t ~ o, x e
!
where
~;[oj~)--~[oj~)
is a continuous function
with l ~ ( t ) d t ~
then the solutions of the sFstem (2) are strongly bounded in probaPro of.
I~et to>. o, ~ > o , ~ ~ S(~)EI~R ~ Let V(t,~)=W(t,~)+ •f~(s)ds,
a(r)=inf V(t,~), K(to,~)=su p V(~o,' t o ,
Ix(t,to,~)(~)l>
rJ and ~,~(t,~)=min{
By Ira's formula, we have
~,v(z~(t),~(z[~(t),to,~)}-V(to,~)~ J'~Av(s,~(S,to,~'))ds~_ o Hence
t,~.(~).}
27 :or all t >. t o But = lim P {Zo,~ ) ~ " 5 Thus
and Theorem 5 is proved. Gonsider the discrete system
~(t+l)=A(~)~(t)+~(t,~) ~(~,~,), t ~ N
(3)
where A(t) is a ~ - m a t r i x , G ( t , x )
is a @~-matrix,
whose elements
are continuous functions, and ~(t,~ ) is a p-dimensional random vector. Suppose that the trivial solution of the deterministic system :r(t+l)=A(t)y(t), t ~ :~ is exponentially stable, i.e. there exis~
ly(t,to,y)l~i~ W ~-t° Theorem 6. If
~o
, ~e(0,~)such tha~
~, for all t~N,to~N, t~t o, ~ R
sup l ~ ( t , x ) l ~ ; te N
~
sup E l ~ ( t ) } ~ ~eN
xe R ~ then the solutions of the s~ste m ~(~ are nltimat~ly
bounded in proba-
bility with ~espect to X. oo__._~f. Pr Let ~ e So, to~ N. Let x(t) be the solution of the system (3), with ~(to)(~ )= ~(~ and y(t,to,~) the solution of the system y(t+l)=~(t)y(t),
y(to,to,~): ~Let V(t,x)=sup(~y(t+s,t,x)|~ ~ )
t
e
It ms easy to see that ~x~V(t,~)_~
for t ~ N, x e
R~
N
:~£
with
28
v(u+i,y(u+~,to,~))-v(t,y(U,~o,~)) ~_ -(l-w)v(u,y(~,~o,yO), ~ ~o,teN,
Ye
Rf
; IV(~,XL)-V(~,~2) I ~_ ~ l ~ i - x 2 1 ,
~l,X2
for Rg
We may w£ite
v(t+i,~ (~+l))-v(t,~(~)) =v(u+l,y( u+i,y(u+~ ~-~,~(t)))-v(t,~(~)) + +v (t +i,~( t +l) )-W+l,y( ~+ L, u ,y(U))) Thus
v(~+].,y(u+].))-v(t,y(u))
__< - ( i - ~ ) v ( u , ~ ( u ) ) +
~I y ( u + l , ~ ( ~ ) ) -
-:~( t+ 1)l ~_ - ( 1-~r)v( u ,:~(~) ) +# I G@.,] (0)11 ,~( ~ )~ _z_ -(1.,¥) v ( t , X ( t ) ) + + ~ q I~(±)I where cl=su p iG(t,~)~ x~R ~ Hence
mV(U+i,](~+i)) z_~ L~(~,~(t))+ ~elc 2 c2
= sup ~ l?(t)i t~o
From Theorem 3 it 1"ollows that bhe solution of the system (5) are ultimately bounded in probability wibh respect
reX.
Theorem 7. If
(~)
sup IG(~,~)I~
M i~l'~
ten
(2) L - 7 - ~ z
M~>O
(3) Z_ ~. i ",/(L )1~ ~ ten ~he~ ~,hp ,enln~.innn nf b~e s,ystem (>) aft strongly bounded in probab ilib~ w i~h respce~
to X .
i)D 0 o f . Leb ~ o 6 N ,
~ £ S . Con~i(ier ~(~) and V(b,x) as in the proof of o
Theorem 6. We
have
v(t+l, :~(~+].))£ ~rv(u,:~(u))+#lm(~,x(~))~ I~(u)l_~v(u,:~(u)) +
29
Hence
_sv~.t+l,:,,(t+l)) --_ :~V(t,:~(t))-(L-~"
._ ~ ~)e~.x~,~.~l -, ~ ~rl. ~t~¢
Using Theorem 2, the theorem is proved. Consider the aiscrete system
:x('~+l)=A,'x(t)+a(t,',J)bJ;(~(t,~Z)~+Bg(|;,~ where A is a ~ - m a ~ r i x ,
B
),
6"(t)=c'x(t),
is a e,~-matrix,
teN,
b and c are
('q-) ~-dimens-
sional vectors, c' represents the transpose~ of the vector c, a(t,co), t e n
is a sequence of random variables,g(t,o~), t e N is a
seguence of p-dimensional random vectors and ~e ~& where ~ se~ of continuous functions f:RI--7 R1with
is the
o~-~'~)1
whereY(1)
~'=b-
+1
+
.4." O]+ o, a~ct (-.~.- -/~ g~)~-..~./-~>~ Let
%o>o such that
Let H=
A°
-~o(-L
,
/'
Z(cc,), ~ ( I ) ko
Obviously that M is a positive definite matrix . It is easy to see that
2h
- ~ b,Hb=Ao,A,HA_H=_qq, _ ~ ]
c where q = 4~
Let x ( t , ~ ) ,
as in the proof of the preceding theorem.
We have
4 x'(t,~) [A'HA-F~ ~(t,~ )+/~b'Hb fz(~t,~))+~'(t): = ~'(t,~) [A'~A-HJ~(t,~)+ i~(q(t,~)[/¢b'Hb-. + ]
where T(t)=IB'HBI
+
E Ig(t,~o)l ~
Hence
E
-(q,~(t,~))~ - ~o/~ ~(t,~ )1.2 ~o q'~(t,~)e(~(t~)) -~ol~(t,~)l~
where
~
= sup ~(t), tan
~i = ] ~~o
Using Theorem 3, we conclude that Theorem 9 is proved. We remark that it can be proved that the trivial solution of the
system (G), with g ( t , ~ ) ~ o a n d A
in mean square for all fE ~ Theorem 9 holds.
= o is e~ponentially stable
if and only if condition (2) of
33 Consider the following It6 system a~ : Ax + b ~ ( ~ ) d ~ ( t ) ,
where A is a ~-~-matrix,
(7)
~=c'~
w(t)
~d
b,c e R,
is t he pro-
tess of Brownian motion Theorem l o .
Th e triVl~laL ,So~,utilllO~
of t h e system ( 7 ) i s exlPonen-
tialLy stable in mean square fq~ all f e ~FA.
matrim
A is stable and
P~ OllO~i
~
if and only.if the
('/A-LA])'~I"~LA
~r
>0
(L=~-.~)
"
By ~ ( G )
-A'
will be denoted t~e matri~ which verifies
~(G)-
~(G)A=G.
wo
0
£#
J-~
Thus, by Theorem in [2] , it follows that if the trivial solution of (7) ~zor ~(~-)=k6- ,is exponentially stable in mean squarc~¢~ l l the matrix A is stable and (~c'(A-iZI) -I bi~dA >o j--Co
Suppose, now,that
~ 2~
h'A
Let
~ =b'~(cc')b,
and
(
the matrix A is stable and
.~,
i c'(~-i
~ I) -l
A---T- ÷ ~ ( i )
it is easy t o see t~at
= ,
>O
such that
- ~2
>o
' "h
Let R=
~~>~
~--b'~(T)b,
2 ..... ;o~o)L ~
bl ~ dA
"
-
and W(~)=x'~m. 2 - b ' HI3= ~ -'6---
c
A' l]+L~iii=-qq'- ~¢i,
~+b'~l~((c
) , ~=~'z
~lence
z - -F
f~ (~)]_~ _(q,
From this relation,
~- O}
Then it is not necessarily
Px{XT+t
x
be a Markov process whose
that w.p.l.
(3)
(Observe that
is slightly too broad a class of
by
x(~) = inf(t:
and suppose that
s ~ O, t ~ O.
Indeed,
(3) should
and we will restrict
to processes
where
our
(3) is true
T.
random variable
T
(defined on a set
~ C T
100
i.e.,
w
process
is a Markov time if we can tell whether or not x
s
up to time
Definition.
t
only, for each
T < t
by watching the
t > O.
If (3) holds for all Markov
times
T, then
xt
is said
to be a stron5 Markov process. Definition.
We will consider only strong Markov processes.
a Markov process for which the function of for each
t > O
x
given by
Exg(Xt)
and real valued continuous and bounded
If
xt
is
is continuous
g('), then
xt
is a
Feller process. A Feller process whose paths are continuous from the right is a strong Markov process.
([5], Theorem 3.10).
Definition.
Let
measurable functions on
B
denote the Banach space of real valued bounded
X, and
B0
the subset of
Exf(X t) + f(x),
weakly as
t + O.
exists (i.e., there is pointwise convergence, h + O) and is in
infinitesimal operator Suppose w.p.1., and
for which
f c B0
If the weak limit
Exf(X t) - f(x)
as
B
T
f E ~(A),
BO, we say that A, and write
g(-)
g(x), and the left hand side is bounded is in the domain
~(~)
of the weak
Af = g.
is a Markov time and
ExT < ~, and
xt
is right continuous
(the continuity conditions can be weakened;
see [5], P.
133) then we have the important relation (4), known as Dynkins formula ([5], P. 133), T (4)
Exf(X ~) - f(x) = E x I ~f(Xs)dS" 0
101
The operator
A
plays a role for Markov processes,
similar to the role
the differentiation operator plays for differentiable non-random real valued functions.
Equation (4) is an analog of the deterministic
integral - differential
relationship and, as such, will play an important role in the sequel.
The non-homo6eneous case. to apply (4) to functions follows.
Define
Redefine
B, B 0
t
f(x,t)
If
xt
is non-homogeneous,
of both state and time, then we can proceed as
to be a state of the process (replace
appropriately,
let
f a~(~)
Ex~tf(Xs,t+s)
and
s ÷ O, and
- f(x,t)
conditions on
T
Ex,tg(Xs,t+s) + g(x,t)
X
g = Af
by
if
X × [0,~)).
f(.,.) e B
and
g(x,t)
S
weakly as
or if we wish
weakly as
s + O.
Then under the
in (4), T
(4')
Ex,tf(xT,t+T)
- f(x,t) = Ex, t
Af(xs,t+s)ds" 0
In (4') we understand that time is measured from the origin of time
2.
x
in (4') is the value of the state
S
s
t; i.e., the value
units of time after the initial
t.
A Few Sources of Stochastic Stability Problems Stochastic stability problems occur in almost all phases of physics,
control theory, numerical analysis and economics where dynamical models subject to random disturbances appear, and the process is of interest over a long period of time.
0nly a few simple problem types will be mentioned here. Suppose that
tion
~t
Yt'
is a Markov process which drives the differential equa-
= f(y~,yt,a), where
~
is a parameter,
driving term, or random variations interested in the range of
a
!
Yt
may represent an external
in some parameter of the equation.
for which
mains bounded in some statistical sense.
Yt ~ 0
w.p.1., or for which
For
fixed at
~
We may be lytl
re-
a0, we may be inter-
102
ested in the range of initial conditions
Py,y,{
E.g., Yt
~>t
sup
y,y'
for which (for some
lYtl i X )
>0
i > ~ > 0)
! 6.
may represent a stress in a mechanical structure, and it may be of in-
terest to keep the stress less than ~ >
0.
Also, it may be desirable to know
whether
Py,y,{
sup >
as
y
or
y'
t
lytl ~ X} + 0 >
0
or both tend to zero (a type of stability of the origin w.p.1.).
The above stability properties are all properties of the paths of the processes.
There are many problems of interest concerning the asymptotic behavior
of the moments and of the measures of the process - and even in cases where the process is of interest for only a finite time [i], [3]. A large class of stability problems arise in tracking situations. example,
For
suppose that we are driving on a road and sample our instruments and
errors (e.g., distance from the center of the lane and from other cars) somewhat irregularly (as is usually done), then can we track the center of the lane within a certain error, etc.
Tracking problems arise in radar and machine tool systems.
Many types of stochastic convergence can be studied. w.p.l, convergence mainly.
Here we deal with
Other stability problems deal with (a) convergence
w.p.l, to a set, (b) recurrence - the process always returns to a bounded set w.p.l.,
(c) no finite escape time w.p.l.,
tain moments of the process, to an invariant measure.
(d) convergence or boundedness of cer-
(e) convergence of the distributions of the process
Types (d-e) are usually more difficult to treat than
w.p.l, convergence, but also are of considerable practical importance.
3.
A Brief Review of Deterministic
Stability
Some results in deterministic
stability are briefly reviewed because,
in a certain abstract sense, the stochastic results are analogies of the deter-
103
ministic results. Let Rr
to
Rr
Rr
denote Euclidean
r-space, f(.)
a continuous function from
and suppose that there is a continuous solution to the homogeneous
differential equation
~ = f(x).
Let
V(-)
denote a continuous, non-negative,
real valued, continuously differentiable function on noted by
Vx('))
satisfying
fined by
Qk £ {x: V(x) < ~}
V(O) = O, V(x) > O, for
(whose gradient is de-
IxI # 0.
Let the set de-
be bou4qded with the derivative of
positive along trajectories in
(5)
for
Rr
V(x t)
non-
QI; namely
V(x t) = Vi(xt)f(x t) ~ -k(x t) ~ 0
x t g QI.
Let
V(xt)
x0 = x
be in
QI.
is non-increasing.
The following statements can be made:
Then
x t g Qk
for all
t ~ O.
From
t
(6)
V(x) - V(xt) = I k(Xs)dS > 0 0 I
we have that
Ik(Xs)dS < ~.
This, and the uniform continuity of
k(x s)
[o,~)
on
0
imply that
k(x s ) + 0
as
Furthermore as
s ~ ~ , and
x s + {x: k(x) : O}NQ~
x + O, the maximum excursions of
~ K~. IxtI decrease to
zero. Define an invariant set of points Then there is a function with
x t, t e (-~,~)
x 0 = x, and furthermore
xt £ G
G
in
Rr
as follows.
which satisfies the equation
for all
t g (-~,~).
entire trajectories over the doubly infinite time interval Let the trajectory and assume (5).
xt
Let
be bounded.
Thus
G
x e G. ~ = f(x)
contains
(-~,~).
In particular, let
x 0 = x e Qk,
Then the invariance theorem [7] states that the path tends to the
largest invariant set contained in
K k.
The theorem is important since it is often used to show that the tend to a much smaller set than sets to which
xt
can tend.
K k.
xt
It gives a very nice characterization of the
104
Example.
Define the differential equation on
R 2,
iI = x2 x2 = -g(xl) - ax2 where t
I g(s)ds
as
÷
t ~ ~,
sg(s)
> 0
for
s # 0
0 g(0) = 0,
a > 0.
Define the Liapunov function
v(x) =
+ 2 i1 g(s)ds. 0
Then
V'(x)f(x) = -k(x) = -2axe. x We can conclude that
x2t ÷ O.
But what about
It is natural to expect that
Xlt?
Xlt ÷ 0
also, and indeed (although the
Liapunov function argument does not directly yield it) it can be proved using a limiting argument, using the facts that
V(x t)
is non-increasing and
Yet it would be much simpler to merely substitute
x2t ~ 0
x2t + 0.
in the differential
equation, and see what trajectories are possible; namely, put the limit of into the equation, directly. conclude that
Xlt + 0
x2t
The invariance theorem allows us to do this, and to
also.
In examples involving functional differential or
more complicated systems, the invariance theorems can save an enormous amount of work. In the sequel, we will develop stochastic counterparts of all the concepts which we just used.
~ile
homogeneity is required for the invariance
theorem, there are straightforward non-homogeneous function theorems.
extensions of the Liapunov
i05
4.
Stopped Processes The weak infinitesimal operator
~
and Dynkins formula (4) will be used
to replace (5)~ (6) for the stochastic problem.
The domain
to be a subset of a set of bounded functions on
X.
tions
V(.)
~(~)
was defined
However, the Liapunov func-
which are most likely to be used, and to which
~
is to be applied,
are usually unbounded (as is usual in the deterministic case).
Even if
V(')
were
bounded, the process may have a stability property only in a bounded or compact set
Q
in
X.
I.e., ~f(x)
may be non-positive only in some neighborhood
Q
of the origin. There is no loss of generality in studying the process only while the paths are in such a set
Q.
For we can often (always, if
find a sequence of sets
Qn t X, and,
X
is
G-compact)
if desired, study the behavior of
X
by
studying the "limits" of the behavior of the process up to, say, T , where n inf(t: x t ~ Qn }.
Thus, we can bound
V(x)
for
x
or we can define a new process by merely stopping
T
"s u f ficiently far" from xt
on first exit from
=
n Q,
Q.
The latter approach is much more convenient. Let
Q
be a set in
conditions under which following.
Let
xt
X.
Dynkin ([5], Chapter 4) gives various general
T = inf{x t ~ Q}
is a Markov time.
be right continuous w.p.l.
(a)
Q
is open and has compact closure.
(b)
Q
is open, X
compact. (c)
Q
We mention only the
(Lemma 4.1)
is a metric space (or metrizable)
and
xt
is continuous.
Define the stopped process either (a)-(c) above hold, and let
xt = X t A T ' xt
(p. iii) where
t ~ T = min(t,T).
be right continuous w.p.l.
a strong Markov process ([5], Theorem 10.2).
Then
Let xt
is
Unless otherwise mentioned, ~Q
be used to denote the weak infinitesimal operator of the process
x t.
Let
will xt
Then to apply Dynkins formula to an unbounded function
V('), we only need check that the restriction of xt
is
(Lemma 4.1)
is open, and
be continuous w.p.l.
X - Q
V(')
to
Q
is in ~(~Q).
is right continuous w.p.l., we need to check whether the restriction of
If V(-)
106 to the union over ~(~Q).
5.
x = x0
in
Such verification
Stochastic
Q
of the almost sure range
usually seems to be straightforward
Stability and Asymptotic
Unless otherwise mentioned, this section.
Xs, s ! T,
After the theorems
is in
in examples.
Stability
we will use the following assumptions
are proved,
in
extensions to more general cases
will be discussed. (AI)
X
is Euclidean
(A2)
V(')
(A3)
Define Let
r-space.
is a non-negative
real valued and continuous
Q~ = {x: V(x) < X}
xt
of exit from
(A4)
QX.
A~
sup xaQ k
(where the definition
for some
strong Markov pro-
x0 = x
of
in
xt = x t A T x "
V(.) Q
opera-
is assumed res-
of the almost sure
xt).
Px( sup
t > s > 0
Observe that,
is not empty.
for the weak infinitesimal
TX = inf(t: x t ~ Q~}~ and
tricted to the union over range of
R r.
X, defined until at least the first time
Write
xt, where
V(') g ~ ( A x )
(AS)
Q~
denote a right continuous homogeneous
cess on the state space
tor of
and assume that
function on
[Ixs-xlI >
a} ÷ 0
as
t + 0
for any a > 0.
i_f_f Y ~ Qk' but is in the almost sure ran6e of
x 0 = x g Qk' then
~kV(y)
~ -k(y) = 0.
Xs, s ~ T k
We will use this fact implicitly
in.the followin6 theorems. Theorem i. tion as x
AkV(x ) t + ~.
e
Assume
(A1)-(A4).
Let
AkV(x) ~ 0
is defined for the stopped process).
Hence
V(xt)
Px {
sup
conver~es
Then
(recall that the operaV(~t)
conver6es w.p.l.,
for almost all paths remaining
in
Qk"
For
Qk'
(7)
l_~f V(O) = 0
o~ > t > 0
and
V(xt) >_ ~}
V(x) # 0
for
=
Px {
sup V(xt) >_ ~}
oo > t > 0
x # O, then as
i V(x)/~.
Ixl + O, the probability
in (7)
107
goes to zero (a type of stability of the origin). Proof.
Applying (4)
gives t
(8)
t~]T~
~V(Xs)dS < O.
ExV(X t) - V(x) = E X I ~A~V(xs)ds = Ex I ~ 0
0
Thus, w.p.l., :u
E~X V(X t) < V(~ s) S
or, equivalently, since measures
xs
is Markov, (-@s
is the smallest
G-algebra which
Xr, r _< s) %
E[V(xt+s) I ~ s ] < V(X~s), w.p.l.
Thus of
(V(xt),_~t} V(xt).
is a non-negative super-martingale.
(7) i s t h e s u p e r - m a r t i n g a l e p r o b a b i l i t y i n e q u a l i t y .
statements are obvious.
x
and
t
The r e s t o f t h e
Q.E.D.
N on-homo6eneouscase. pends on
This gives the convergence
or that
xt
Suppose that the Liapunov function is non-homogeneous.
V(',')
de-
We state the following
Theorem 2, without proof. Theorem 2.
Let the real valued continuous functions (o_~n Rr, R r x [0,~),
R r, resp.) VI('), V(',.), V2(. )
satisfy, for some real
t O _> 0, and
% > 0,
Vl(X) i V(x,s) i V2(x)
for
s > tO
and
x c Q~ = {x: Vl(X) < ~}. Le t
xt
be a right continuous stron6
Markov process defined until at least the first exit time ~enote the weak infinitesimal operator of the process stopped on first exit from t > t O • Then, for
t >to,
Q~.
Suppose
TI
from
Q~.
Let
(xt, tNT%), which is
V(x,t) g°-~(A%) and
AiV(x,t) SUPs > 0VI(Xs) >- ~} < Px't0{~ 0> ssu p>
V(Xs,S+t 0) >_ I)
i V(x,to)/l"
Also
V(xt,t+t0)
conver~es for almost all paths for which
t ~ 0, where we use
x0 = x
V2(x) ÷ 0
a_~s Ixl ÷ 0.
Let also
i - V(x,t0)/~.
and
Vl(X) > 0
for
Ixl # 0; then for any E > 0
and any nei6hborhood of the ori6in
AI, there is a neighborhood
i__ff x ~ A2, the probability of
ever leavin~
-k(x) ~ 0 Thus
i_nn QI.
k(x t) ~ 0
leave leavin~
QI.
< ~
(Asymptotic Stability). Then
k(~t) ÷ 0
(and also
V(xt)
so that,
is no ~reater than
Assume (AI)-(AS).
in probability and
V(~ t)
Let
s.
AI V(x)
conver~es
w.p.l.
conver~es) for almost all paths which never l-V(x)/X
be uniformly continuous in
Q~.
for almost all paths which never leave
QI"
and
k(-)
V(x) ÷ ~
a_~s Ixl + ~, then
to the probability of never
s
are unbounded.
n {x: k(x) < ~} = K E>0
w.p.l.
The
If they are bounded, replace
E > O.
K
c
= {x: k(x) > s > O} N Q1 -
is finite
This follows from the inequality t
(lO)
% V(x) _> -ExV(~t) + V(x) = E x I k(Xs)dS _> ~Ex T' (t,g),
0 where
Tx(t,E)
Rr ,
The key to the proof is the fact that the total time which the
can spend outside of the set
w.p.l, for any
x t ÷ [ n {x:k(x) < g } ] N ~>0 If the hypotheses hold for all
{x: k(x) : 0}.
by
Proof. % x
xt ÷
Then
two s e n t e n c e s i s i n t h e t o p o l o g y f o r t h e c o m p a c t i f i e d
QI n {x: k(x) < ~}
N{k(x) < ~}
process
AI
A2 C AI
QI.)
convergence in the last
if the
xt
(E~uation (7) 6ives a lower bound
Let Q1 - P1
to; thus
a_~s Ixl ÷ 0; then the right hand side of (9) ~oes to zero VI(0) = 0
Theorem 3.
for all
for the initial condition at the initial time
there is convergence with at least probability Let
V(xt,t+t 0) ~ ~
is the total time that
follows from (i0).
k(x s) >__ s
in
[0,t].
That
P k(x~t) ÷ 0
109 We next prove the first statement of the second paragraph of the theorem. Let
T(t,a)
w.p.l,
denote the total time that
for any
E > 0
and
x a Q~.
(A5), t h e u n i f o r m s t o c h a s t i c Let
C
k(~ ) > g s --
in
[t,~).
Then
T(t,E) + 0
The rest of the proof combines this fact with
stability
assumption, to yield the w.p.1,
convergence,
denote the set
C E = {x: k(x) < a} ~ Qk.
Assume that
k(x) < 0
for some
by uniform continuity of between 6(a))
Qk N cCa for
and
k(-)
C /2
0 4 e < eO' and
x E Qk, for otherwise the theorem is trivial. in
(C~
is the complement
Qk O C cc
Define the Markov times lows.
(If
dn
or
O'n
Qk, there is some of
so that the distance
Ca) is positive
(say
_>
is not empty.
~n' q'n
is not defined at
q0' = inf{t: ~x t a Cc/2} ,
a0
Then,
(finite on sets
On' ~'n' resp.) as fol-
~, set it equal to
~
there.)
d O = 0,
(~l = inf(t: x~ t ~ Qk n C c~, t _> (~} ' ~'n = inf{t: x t a Ca/2,
t >__ (~n-i } ' gn = inf{t: ~x t E Q1 n cC~, t > ~ n }, etc.
(ii)
There is some p >
sup Px { sup JXs-Xl x~Qk p > s > 0
0
so that
! 6(a)/2} L 1/2.
Define
A = {~: xg c Cc n Q~, o < s < p, o < ~}. n +s a/2 n n If
~ a An
infinitely often, then the total time out of
for the c o r r e s p o n d i n g p a t h E IA
n
~ ~
w.p.l,
i = O,...,n-1,
xt(m).
if and only if
are in
~
Then (~
n
~ a A n measures
c Ca/2 U Qk
only finitely
is infinite
often w.p.1.
% Xs, s ~ ~n' thus all
But
Ai,
) n
Z Px{Anl~q
([5], P. 398-399),
n
} ÷ ~
w.p.l.
and, by the strong Markov property
(Px+{A} = 0
for
t = ~ )
110
(12)
Z Px{AnI~(~
n
} > Z Px(~ {
sup p>s>O
n
1 > _ ~ Z I{a
Thus
Gn < ~
iXs_Xl
The remaining
follow easily from what we have already proved.
(1) k(x) = 0
See the hypothesis
S1
is deleted,
and
ing, so there is a discontinuity
k(.)
(2)
If the hypotheses
Ixl ~ ~, then
x t ÷ {x: k(x) = 0}
(3)
If
be continuous
for
k(')
X = Rr x ~ S 1.
Q~
Q~
may be unbounded.
subset
S
of
X.
t ~ Ws, ~.
Thus
topology of
in
Q~, then
(A5) may not hold.
Suppose that (A5) holds if
TS, C < ~
Q~
x t + {x: k(x) = 0} U {~}
w.p.1., w.p.1,
S2
is absorb-
X. and
V(x) ÷ ~
as
xt
for any state process.
Furthermore
{x: k(x) < ~}g
is replaced by any compact For any compact so that
S
and
xt ~ S - C
for
in the one-point-compactification
R r.
subsidiary
conditions
can be used to eliminate the point
Refer to the next section for the definition
k(x t) ~ 0.
~
k(xh) ~ 0
Then, we can obtain the following.
Sometimes
"invariant
The set
w.p.1.
is unbounded,
> O, there is a random variable
Q~, nor even that
has a hole in it, i.e.,
of Theorem 3 hold for all
AxV(x) ~ -k(x) ~ 0
If
in
on the boundary of
space, provided that the Dynkins formula is valid for the (4)
of the theorem
and proof of [1], Theorem 2, Chapter 2.
k(x) ~ 1 of
statements
Q.E.D.
There are examples arising in control theory where a target set
< ~o}
and Extensions
It is not necessary that
anywhere.
n
< oo} •
n
only finitely often w.p.l.
Discussion
> 6(~)_} I{q
set".
Let the measures
Then
xt
set whose s u p p o r t
is
of the
tends in probability contained
probability to the union of
in c QX
[{x:
of the terms "weakly bounded"
process
xt
be weakly bounded,
(~}.
and and let
to the support of the largest invariant
k(x)
c = O} n Qx ] u QX.
and a subset of
Thus,
xt
{x: k(x) = 0} N QX"
tends
in
111
The remarks and results for unbounded problem for a process of the type are Markov processes.
The process
not converge in any sense. ponent
Yt
If
(A5), if we replace
# = f(u,y), where ut
(u t,yt ) = x t
We may be concerned with the convergence of the com-
X
is a metric space, the proof still goes through under Rr
in (A2)
will not usually be bounded.
then the path Ql
u t, and the pair
may serve as a time varying parameter, and
by a metric space
verify (A5) and the uniform continuity of
Then
are motivated by the stability
only, but the Liapunov function may depend on both components. (5)
Ql
Ql
k(')
X.
It may be difficult to
in this case, and the closure of
But it sometimes happens that if
Xs, Tl > s ~ O, is contained in a bounded subset of
is "effectively"
contained in a bounded subset, and if
x = x 0 g Ql, Ql k(.)
continuous and (A5) holds on this subset, then the proof goes through. for a specific example. guarantee that
xt
(A2)-
w.p.l. is uniformly See [6]
(A5) plays a crucial role in the proof (since we need to c CE/2
does not jump (w.p.l.) from
to
C c£
and back to
C ce/2
infinitely often in a total integrated time which is finite), and some form of uniform stochastic continuity condition is probably essential.
II.
INVARIANT SET THEOREMS AND APPLICATIONS TO STOCHASTIC DYNAMICAL SYSTEMS In this Section we will develop a stochastic theory of invariance analog-
ous to the deterministic theory in [7], [8]. given conditions, the measures of the process measures,
and that
xt
The main conclusion is that~ under xt
tend to an invariant set of
tends to the closure of the support set of this set of
measures in probability as
t ÷ ~.
Note that we are using the terms "invariance" and "invariant" according to their usage in the general theory of dynamical systems.
The term has
nothing to do with the stochastic notion of invariant measure.
In this Section
x
t
will be a homogeneous
strong Markov process.
opment in [2], with some changes and corrections.
We essentially follow the devel-
112 i.
Definitions Let
Let
@
xt, be a separable metric space.
denote the initial measure of the process;
m(t,¢,.) time
X, the state space of the process
i.e., P{x 0 e A} = ¢(A).
denote the measure induced on the Borel sets of
t, with initial measure
¢.
X
Let
by the process at
The semigroup property +
m(t+s,~,.)
= m(t,m(s,~),.)
holds. Let in
~
f(.)
~
denote the space of probability measures on
is said to converge weakly to in
bounded if, for each e.
~ W(~)
if
f[~n ] ÷ f[~].
A set
Define an
u-limit set
if there is a sequence
as a set
tn ÷ ~
K
C X
W(~)
so that
X.
M = {~a}
e > 0, there is a compact set ++
in A
ures
ME
~
m'(t,'),
m(t,m'(s,~),.)
is an invariant set if for each for
t ~ (-~,~)
= m'(t+s,.)
where
for any
xt
x e X
is in the support set
S(~)
N
x.
of
set
S(~)
Similarly
in ~
+
for which
~a(X-K ) < e
with the property:
f[m(tn,~)] + f[~]
(~
for all
is a f(.) e C X.
m'(0,') = ~(.), the initial measure, and t ~ 0
and
s ~ (-~,~). t e (-~,=)
and initial condition of
~
if
~(N) > 0
~.
Thus for each
Exf(Xt)
is continuous in ~
x
is written for
for
~(')
~ £ M,
and satisfying the Let
~
be in
-~.
for each neighborhood of
U S(~) is the support set of a set Q in ~ . ~gQ S(Q) is not necessarily closed. The process x t
Occasionally for simplicity or
is weakly
S(Q) =
is closed, but
Feller process if
for every
~ e M, there is a sequence of meas-
there is a trajectory of measures defined for all law of motion of the process
{~n )
We may abbreviate
weak limit of a sequence of measures taken along the trajectory) A set
A sequence
ff(X)~n(dX ) ÷ ff(x)~(dx)
CX, the space of continuous bounded functions on
the convergence relation as
for all
~
X.
t > 0 and
and
The is a
f(.) e Cx.
m(s,$)
for
m(s,~,.)
m(s,~(.),').
++It is important to keep in mind that the trajectory of measures.
~-limit set is an
~-limit set of a
113
Next, the main theorem and a useful corollary will be given.
Then the
conditions of the theorem will be replaced by more easily verifiable conditions.
2.
The Invariance Theorem Theorem 4.
Assume (BI)-(B3).
(BI)
The trajectory
{m(t,¢), t ~ 0}
(B2)
For each
i_n_n C X, f[m(t,¢)]
finite
f(') t
is continuous in
(t ~ O) interval, unif0rm!y in
weakly bounded (B3)
is weakly bounded.
f[m(t,@)]
@, for
¢
for each fixed
W
[l.e., a_~s @n ÷ @' f[m(t'@n)] + f[m(t,@)]
Then
W(@)
~(t)
f(') s CX, a_~s t Proof.
for each
t ~ 0. f(') s C X
weakly compact invariant set
i_n_n W(~), t ~ O, so that
f[m(t,#)]
for all
in an~
t > 0.]
is a non-empty, weakl~bounded~
and there is a sequence
¢
0n any
set.
is weakly continuous ~n
and each
t
- f[~(t)]
÷ 0
~.
According to Theorem !, Section I, Chapter 9 of [i0], a suffici-
ent condition for a sequence in -~ to have a weakly convergent subsequence is that it be weakly bounded. Let
{E.}
each
G.
I ~
W(@)
is not empty.
denote a real sequence which tends to zero.
i
are compact sets
Thus
G.l
so that
Gi+ I ~ Gi
and
there is a countable family ~ !
and dense in
i
CG .
Each element of ~ m
By (AI), there
m(t,@,X-G i) --< si' all
t _> 0.
For
of continuous functions, defined on
C i,
can be extended to a continuouS function
i
X
without increasing its norm (using the normality of the metric space and [9],
Theorem 1.5.3). U ~i" i
Write
Let
.°~ i
G = U G i. i
denote the countable family of such extensions and Observe that, for any
f(.) g C X,
-~=
oi
114
(*) also holds for
m(t,@,-)
replaced by an element in the weak closure of
(m(t~¢,')). Let
m(t ,¢,') n
Define the function
converge weakly to
Fn(.,.)
4(')
in the
u-limit set
W(¢).
by
Fn(t'f) = I f(x)m(tn+t'@'dx)"
If
tn-T _~ O, then
Fn(t,f) = J f(x)m(t+T,m(tn-T,@),dx).
is weakly bounded, (B2) uniformly in
implies that
n, for each
f(').
Fn(t,f)
is continuous in
[-T,T].
verges to a continuous function of
an
fg(')
in ~
t, F(t,f)
interval.
for which
and
for all
n
con-
and uniformly g ~ O, there is and
t h - t n,
f e CX. 4(t,.)
~(t,A) =
IA
[-T,T]
Fn(t,f)
f(') E ~ ,
f(') g C X
IFn(t,f s) - Fn(t,f) I < e
Define the set function
f(') g CX, and
on
t n) for which
for each
Since, for any
the asserted convergence is for all
where
t
By successive applications of the
diagonal procedure, we can extract a subsequence (of
[-T,T]
(m(tn-T,@))
Thus Ascoli's Theorem implies that there is a
uniformly convergent subsequence on
on any compact
Since
by
inf F(t,f) f h IA
is the indicator function of the Borel set
A
in
X.
The argument in [lO], pp. 441-444, can be used to prove that, for each t ~ (-~,~), 4(t,')
is a unique probability measure, 4(t,G) = 1
and
F(t,f) = I f(x)4(t,dx)
for each
f(') c Cx.
4(0,') = 4('). ported in in
Cx,
G.
Thus
W
m(t +t,¢ ") + ¢(t,') n
The weak closure of
~
{m(t,~,')}
for each
t g (-~,~)
where
is also weakly bounded and is sup-
Thus, by (B3), we can write, for any
t c (.~o ~), s ~ O, and
f(')
115
f[m(s,m(tn+t,¢))] ÷ f[m(s,~(t))] f[m(s,m(tn+t,¢)) ] = f[m(O,m(tn+t+s,¢)) ] f[m(O,@(t+s))] = f[~(t+s)],
which implies that
~(t+s,') = m(s,~(t),')
mine the measures uniquely. ~(t)
Thus
{~(t)}
since the continuous functions deterobeys the law of the process and each
is in an invariant set. Let
(thus
f[~n(')]
{~n }
f[~(')]
converge for each
are weakly bounded).
on
CX, and
~(G) = 1.
W
m(ti(n),¢,') ~ ~n (')
as
f(.) ~ CX, as
There is a measure
n ÷ =, where
~(')
We need to show that
for which
~(') E W(¢).
i ÷ =, for some real sequence
t.(n) ~ ~. i
~n (') e W(~) f[~n(')] +
For each
n,
Since
lim lim If(x)m(ti(n),¢,dx) = lim If(X)~n(dX) n i n :
for each for which
f g ~ , we can extract a subsequence
If(x)~(ax), (t }
of the double sequence
m(t ,~,.) ~ $(.). Only the last assertion of the theorem remains to be proved.
there is a sequence or
{ti(n)}
{t n}
so that for any subsequence
{t~}, and some
Suppose that f(')
in
CX,
(*)
By weak boundedness of to some
~(') E ~g .
lim sup
inf
n
~Ew(¢)
Jf[m(t~,¢)] - f[~(')]l >
o.
{m(t',¢,.)}, there is a subsequence which converges weakly n This
~(')
must also be in
W(¢), a contradiction to (*).
Q.E.D. Theorem 5. (i)
P
Assume (B1)-(B3) of Theorem 4. --
Then
x t ÷ S(W($)) ~ C, the closure of the support set of the invariant
116
set W(@)
i.e.,
P.{inf
'
@
(ii) X
and let
Let
Gn .
of an invariant
as
t ÷ ~, for an~
be a real valued, non-nesative Let
Then
G
xt
conver6es
in ~robability
X
that, for each
Let
NE(C)
denote an
m(t,¢,X-Gn)_
0,
(*)
lim P@{x t e X-N£(C)} t ÷ ~ ~
since (*) implies so that
for which
function on
to the lar6est support
n
(i)
and continuous
denote compact sets in
n
set whose s u p p o r t i s c o n t a i n e d i n
Proof.
e > 0.
--
k(')
P k(x t)_ + O.
en ÷ O, Gn+l D
Ixt-Y I > g} ÷ 0
yeC
P@{x t
(i).
Suppose
= 0,
(*) is violated.
e X-~e(C)} ~ e 0 > O.
Then there are
There is a function
t
n
f(') e C X
+ ~
and
£0 > 0
satisfying
n 0 ! f(x) ~ i, f(x) = 0 {t~}
of
on
{tn}, m(t~,@,')
f[~(')] ~ e 0 > O.
Thus
in the support set of
Ne/2(C),
f(x) = i
on
X-Ne(C).
converges weakly to a
~(')
X-Ne(C) , which is disjoint @('), a contradiction
in
from
For some subsequence W(@)
and
~, contains
to the definition
of
Under the conditions
of Theorem i, if
for the stopped process are weakly bounded, to the stopped process. each
~
is compact,
k(')
in Theorem
is not bounded, from
Ql
If the conditions
then
{m(t,~,.)}
(5) is the
k(')
Theorem 6.
Proof.
Let
Q
some point
Q,E.D.
of Theorem 4. is compact,
then the measures
and we can apply the invariance theorem of Theorem i hold for all
is weakly bounded.
of Theorem 3.
it may be that the measures
are weakly bounded.
(BI)-(B3)
+
C.
(ii) follows easily from (i), and the proof is omitted.
Discussion of the Conditions
f[m(t~,¢)]
Usually,
Furthermore,
for the process
I < ~, and the function
even if each
Ql
stopped on exit
See Example 2 in [6].
(B3) holds for a Feller process on any topological W
Cn (') + ¢(').
We must show that
state space.
117
I
(*)
f(x)m(t,¢n,dX) - Jf(x)m(t,¢,dx) ÷ 0
for all
f(') e Cx.
Write (*) as
I
[f(y)m(t,x,dy)](¢n(dX)
- @(dx))
= lht(x)[@n(dX) - ¢(dx)] m(t,x,.) is in
CX
denotes the measure with initial condition
x
O. and
ht(x) = Exf(Xt)
which
W
by the Feller property, and the convergence follows since
Cn(" ) + ¢(.).
Q.E.D. Remark. Theorem 6 implies that condition (B3) is not very restrictive. Theorem 7.
Let
(*) as
Px{[Xt-X] > g) ~ 0 t ~ 0, uniformly for
K c X, let the family Proof.
Let
x
in any compact set.
{m(t,x,'), x g K, t ~ T} {@a}
For each real
T > 0
be weakly bounded.
Then (B2) holds.
denote a weakly bounded set of measures.
second hypothesis implies that the family
{m(t,¢a,'), t ~ T, all
a}
bounded (we omit the proof, which is not hard). Write, for
and compactum
t > O, s > O, s+t < T,
IIf(x)[m(t+s,~G,dx ) - m(t,~,dx)] I IIExf(Xs ) - f(x) Im(t,~G,dx) = I IExf(Xs) - f(x) Im(t'¢a'dx) G' + I IExf(Xs) - f(x)Im(t'~a'dx)" X-G'
Then the is weakly
118
Choose compact
G'
to make the second term less than
Then, using the first hypothesis, choose for
s < sO
and
x E G'
sO > 0
~, for all
so that
IExf(Xs) - f(x) I ~
thus proving the right continuity of
To prove left continuity, write, for
a, t ~ T. E
Exf(Xt)
T > t-s > 0, s > O, - -
m
- -
IIf(x)[m(t,(~c~,dx) - m(t-s,@a,dx)]l IExf(Xs)-f(x)Im(t-s,¢a,dx).
~ I IExf(Xs)-f(x)Im(t-s,@a,dx) + I G' Choose compact
G'
and then choose
X-G' E ~ ~
so that the second term is
sO
so that
E IExf(X s) - f(x) I < ~
for
0 ~ t-s ~ T, and all s ~ sO
for
and all
x g G'.
Q.E.D.
III. Example i.
EXAMPLES
A relatively simple example is the diffusion process given
by the It$ equation
dx I = x2dt dx 2 = -g(Xl)dt - ax2dt - x2cdz where t I g(s)ds ÷ ~
as
t ÷ ~, sg(s) > O, s # 0, g(0) = 0,
0 and in
g(') R 2.
satisfies a local Lipschitz condition. Then
xt
Let
Q
be a bounded open set
can be defined up until the first exit time from
stopped process is a continuous Feller process and (BI)-(B3) hold.
V(x) = x~ + 2 il g(s)ds 0 is in ~(AQ)
and, for
x E Q,
~Qv(~I = x2(c 2 2 - 2a).
a,
Q, and the The function
119
Let
c
2
< 2a.
Then
Px { sup V(xt) _> l} _< V(x)/l ÷ 0 ~>t >0 and
xt
can be uniquely defined on
[0,~)
satisfy a global Lipschitz condition.
as ~ ÷
w.p.1., even though
g(.)
does not
It is a continuous Feller process and (B1)-
(BB) hold. Let
c 2 < 2a.
Then
x2t -~ 0
w.p.l, and by Theorem i, x t
tends in pro-
bability to the smallest invariant set whose support satisfies x2t = O, for all P Thus x t -~ O. This and the w.p.l, convergence of V(xt) implies t h a t x t + 0
t.
w.p.l.
Example 2.
For the second example, we take a problem arising in the
identification of the coefficients of a linear differential equation. The system to be identified is the scalar input, scalar output asymptotically stable, reduced form, system
n-1
(i)
di
di
m
(dtnd + i=0~ai dt i)y = (~0 bi 1---r)u' dt n > m,
where
u(t)
c.u.(t) i
i
'
is the input. where
~(t)
We wish to know the
ai,b i.
is a stationary Markov process.
The input
u(t)
is
The "equation error" method
of P. M. Lion ("Rapid Identification of Linear and Nonlinear Systems", Proc. 1966 Joint Automatic Control Systems Conference, University of Washington, Seattle) will be used.
For this method, Some estimate of the derivatives of a smoothed input and
output are needed. Let
H(s)
denote a transfer function the degree of whose denomenator ex-
ceeds the degree of the numerator by at least the "derivatives of the smoothed
u,y
n.
For any real number
as
Yk(S)
= H(s)(s+c)ky(s),
k = 0 .....
n
Uk(S)
= H(s)(s+c)ku(s),
k = 0 .....
m
c, define
120
and the equation error
c(t)
(2)
as
e(t) = Yn(t) +
where {~i,Bi }
are to be prescribed.
n-i m Z ~iYi(t) + Z Biui(t) 0 0
Let the system
(y,y(1),...,y(n-l))
state variablized by the minimal order, (with asymptotically stable
be
A y) i y =
AYx y + BYu, y = HYx y, and write
y(s)
N(s)
:DqiTu(s)
n-1Qi(s)
+
Z
D 7xi(O)
0
where the last term goes to zero exponentially. Let us impose the following conditions: (CI)
](t)
is a right continuous stationary Feller strong Markov pro-
cess with
Elu(t)l 2 = M O < ~.
formable paths w.p.l. all (C2)
~
w.p.l, for
0
sup ~>h>O
lu(h)-ul > e} + O u = u(O)
as
6 ÷ 0
uniformly for the initial
in any compact region.
E[7(t+T)7'(t)lT(s) , s ! 0] ÷ ~(T), the covariance of the cesses.
(C4)
In particular, f e-ktlu(t)Idt < ~
k > O.
condition (C3)
Thus, the paths are Laplace trans-
Let
E~(t) = 0.
pro-
(This condition is not essential.)
Su(~), the spectral density of
There are real numbers
7(t)
{a2,B~}
u(t), is nonzero over some interval.
so that
g(t) ~ 0
if all
x.(0)l = 0.
To see this write the Laplace transform of (2) where we have
e(s) = Yn(S) + = H(s)
if
n-I m o o ( s) + ~ ~iui(s) Z ~iYi 0 0 s+c)n N(s)
D--UV[+
n-i
o
Z ~i(s+c)
0
i N(s)
m
o
+
D-77Y + ~ Si(s c) 0
u(s) = 0
121 m
N(s) _
0
o i 8i(s+c) ~
~
n
o i Z a.(s+c) 1 0 For {e°,6°} 1
1
used in (2), g(t) ÷ 0
systems generating
Yi(t), ui(t)
exponentially.
o n
n
=i.
In fact, we suppose that the
are connected to their inputs at
that their initial conditions do not depend on the process non-random.
=
u(t).
The condition can be relaxed to allow for random
t = 0, and Then
e(t)
is
Yi(O), ui(O) , at
some extra complication in the analysis. The parameter adjustment procedure is
k 32 dj = - ~ 3---~= -k g yj
(B) k 892
Define the column vectors
o o . ,Bm_8~ ) z = (aO-&O,...,&n_ I - an_l,.. w = (yo,...,Yn_l, u 0 ..... Um). Then (4)
= -kwe = -kw{[Yoa O +...+ Yn_l~n_l +...+ Um~m]
÷ Yn ÷ lY0 0 ÷÷ (5)
where
Um °l
= -kww'z + ~t o 6t = -kW[Yn + YO~O +'''+ Um~ ]" We can assume that the
stable systems of the form Yk' Uk' y' u, z namely
- Ey0 o ÷÷
Yk(t), uk(t)
are the outputs for asymptotically
.Yk AYkxYk Yk Yk Yk x = + B y, Yk(t) = H x , etc.
Thus all
are state variabilized, and the composite state variabilization,
x(t), is a right continuous strong Markov process and Feller.
122
Furthermore, it is uniformly stochastically continuous in the sense of (C2). + Let of
E ~(0) I2 < ~.
£ = -kww'z.
Then
Let
¢(t,s)
l¢(t,s)l ! 1
denote the fundamental matrix solution
and co
Iz(t)I
u,B,Qp-IQBu
where etA'pe tA dt,
--co
then there is no stabilizin$
control.
Consider now the case
R
m
unstable modes of the corresponding Rm,
implies
that
C(u)
matrix A,
is sufficiently
Assume that
imaginary eisenvalues. stabilizin$
Restricting i.e.
(ii) of theorem 4.2 is satisfied.
satisfied provided
THEOREM 4.3.
# R n.
control.
If
DD'
large.
is positive
R m # R n,
the system (i.i) to the
to a certain complement With
P = I
(4.1) can be
The final theorem results.
definite and that
then for
of
C
sufficiently
A
has no purely larse there is no
130 References
[i]
U.G. Haussmann, Noise",
[2]
"Optimal Stationary Control with State and Control Dependent
SIAM. J. Control, 9(1971), pp. 184-198.
.....,
"Stability of Linear Systems with Control Dependent Noise",
Ibid., forthcoming.
[3]
W.M. Wonham,
"Optimal Stationary Control of a Linear System with State-
Dependent Noise", SIAM. J. Control, 5(1967), pp. 486-500.
[4]
,
Random Differential Equations in Control Theory, Probabilistic
Methods in Applied Mathematics, vol. II, A.T. Bharucha-Reid, ed., Academic Press, New York, 1969.
[5]
M. Zakai,
"A Lyapunov Criterion for the Existence of Sta%ionary Probability
Distributions for Systems Perturbed by Noise", SIAM. J. Control, 7(1969), pp. 390-397.
LYAPUNOV
FUN,CTIONS A N D
GLOBAL
A CLASS
FREQUENG, y , D O ~ A I N
OF S T O C H A S T I C
FEEDBACK
STABILITY
CRITERIA
FOR
SYSTENS
J A c Q ~ % L. WIL~MS University Gent,
of Gent
Belgium
ABSTRACT This paper deals with the stability of a particular class of stochastic systems;
feedback systems are considered which have a feedback gain with a
deterministic
gain which may be nonlinear and/or time-varying and a stochas-
tic component which is white noise. Lyapunov functions are constructed and criteria for global stability are derived similar to the results available for related deterministic
feedback systems,
erion, the Popov criterion,
such as the Routh-Hurwitz
crit-
and the circle criteria. I. INTRODUCTION
The difficult step in the application of Lyapunov theory to analyse the stability of deterministic systems
systems as well as the stability of stochastic
(Kushner 1967), is the construction of a suitable Lyapunov function.
There is indeed no general systematic procedure functions.
For deterministic
for generating Lyapunov
systems very interesting results,
Popov criterion and the circle conditions,
such as the
have been obtained for a partic-
ular class of feedback systems containing a linear time-invariant path element and either a nonlinear or a time-varying
forward
feedback element
(Zames 1966, J.C. Willems 1971a, J.L. Willems 1970). This is the motivation for considering a similar class of stochastic
feedback systems,
where the
gain of the feedback element has a stochastic white noise component. the purpose of this paper to investigate constructing Lyapunov functions
whether or not the procedures
for the deterministic
for the stability analysis in the stochastic the path integral method
(Brockett
for
case remain useful
case as well. In particular
1970) and the Kalman-Yacubovitch-Meyer
lemma (Kalman 1963) are used to generate useful Lyapunov functions stochastic
It is
for the
feedback system.
Section 2 deals with the case where the deterministic
part of the feed-
back element is a constant gain; a Lyapunov function is obtained which proves a necessary and sufficient
condition
for mean-square
stability;
the white
132
noise component of the feedback gain has a destabilizing effect on the meansquare stability properties. The criterion is also sufficient for Lyapunov stability with probability one, but not necessary. In Section 3 a nonlinear time-varying feedback element is considered;
stability criteria are obtained
which are closely related to the circle criterion for deterministic systems (Zames 1966), but which also show a destabilizing effect of the white noise gain component. A particularity of this analysis,
which does not appear in
deterministic stability theory, is that different solutions of the spectral factorization problem in the path integral method or of the Kalman-Yacubovitch-Meyer lemma yield non-equivalent stability criteria. Section 4 deals with the stability analysis when the deterministic
feedback gain component
is either nonlinear and time-invariant or linear and time-varying.
Criteria
similar to the well known Popov criterion are derived. Some possible extensions and generalizations are discussed in Section 5. 2.
LINEAR TIME-INVARIANT FEEDBACK
In this section the stability is considered of a system consisting of a
linear time-invariant forward path element with rational transfer function
H(s) and a multiplicative feedback whose gain is the sum of a deterministic constant k and a stochastic white noise component. This system is described by the stochastic It8 differential equation dx = (A x - kbcx)dt - b c x d 6
(I)
where matrices and vectors are denoted by underlined symbols. The triple
Ao,~,~
is a realization of the transfer function H(s) = c(Is-Ao)-1~, and
denotes a scalar Wiener process with independent increments and E((~(t)-~(~)) 2) = ~ I t - ~ l ,
E(d~(t)) = 0
(3)
Let G(s) be the transfer function of the deterministic closed loop system G(s) = p ~ )
= £(!S-Ao-kbc)-1~ = H(s)/(kH(s)+1)
Suppose that p(s) is strictly Hurwitz, and denute the polynomial p(s) by p(s) = s n + Pn_1 sn-1 + ... + Po ' q(s) = qn_1 sn-1 + ... + qo where, without loss of generality, p(s) is assumed a monic polynomial. Then (I) is equivalent to
133
p(D)y dt + q(D)y d ~ =
O
(4)
where D = d/dt. In the sequel either equation (I) or equation (4) is used, whichever is more convenient. Both the state vectors ~ and ~ = [ y
Dy °..
Dn-ly IT are used; they are related by a linear transformation, and they are identical, if A has the companion form (Brockett 1970), ~ = [qo qn_1 IT, and ~ = [O
0 ... 0
ql "'"
I] . For the stability analysis of a stochastic
equation, such as (I), instead of the derivative of the Lyapunov function in deterministic stability theory, the sign definiteness of LV(~) should be considered (Kushner 1967), where LV(x) is the sum of the derivative for zero noise and
?Yn
e
where Yi denotes the i
th
component of the vector ~. Hence, if quadratic
Lyapunov functions are used, it is the aim to find a Lyapu~ov function whose deterministic derivative contains a term proportional to (q(D)y) 2. This can be achieved by means of the path integral technique (Brockett 1970). Let the polynomial h(s) be the solution of the set of linear equations ½[h(s)p(-s) + h(-s)p(s)]~ q(s)q(-s)
for all s
(5)
A unique solution exists if p(s) has no zeros on the imaginary axis. Using the Lyapunov function Ft(~) V(~) = J t(O)[ p(D)z h(D)z -(q(D)z) 2] dt
(6)
m
(where the notation is as introduced by Brockett), we obtain LV(~) = -(q(O)y) 2 + ½ hn_ I ~2(q(D)y)2
(7)
Considering the path independence of the integral in (6) and evaluating this integral along solutions of p(D)z = O, the positive definiteness of V(~) is easily proved. By means of the Lyapunov theorems for stochastic systems (Kushner 1967) the following criterion is readily established : Criterion I
The null solution of system (I) is mean-square stable
in the
large and stable in the large with probability one, if p(s) is strictly Hurwitz, and ~'2hn_i/2 ~ I
(8)
134
To prove asymptotic stability pole-shifting techniques can be applied, or the following Lyapunov function is used
v(~) :
s
t(~)
[p(O)z(h(O)z+am(O)z) - (q(O)z)2 -a (z2~Dz)2+...
t (~)
+ (Dn'Iz)2)] dt
where m(D) is the solution of
½[p(D)m(-D)÷p(-O}m(D)~ = ~
(-~)i D2i
i=O Suppose ~ h n _ I / 2
~2qn_l/a = ~w(O)/2
(22)
E,t
where w(t) denotes the inverse Laplace
transform of G(s).
This criterion is proved by means of the following quadratic
Lyapunov
function v(~)
: ftt(~)[q(D)z (~)
where r(s) is the negative of p(s) and q(~)
p(D)z - (r(D)z) 2 ] dt
spectral
(23)
factor of the even part of the product
:
r(s) : ~ [p(s)q(-sl+p(-s)q(s)]T h i s Lyapunov f u n c t i o n y i e l d s
the desired
result
(24) since
LV(~) = -(r(D)y) 2 - (k(y,Dy .... ,t)-~qn_i/2)(q(D)y)2
140
The asymptotic stability is proved as in Section 2. Remark 6
The criterion reduces to a well known stability criterion for det-
erministic feedback systems if the stochastic component of the feedback gain is absent. Criterion 3 is stricter than its deterministic counterpart; indeed w(O) or qn-1 is positive if G(s) is positive real. The condition on the deterministic part of the feedback gain gets more restrictive as ~
increases.
The conditions of Criterion 3 require the passivity of the linear part of the open loop and a sufficient degree of passivity for the deterministic or average feedback element; the degree of passivity required depends on the statistics of the noise component of the feedback gain but also on the forward path element,
since w(O) appears in (22) and (23). In deterministic
theory the passivity conditions of the forward path and the feedback path are uncoupled. Remark 7
The criterion is equivalent to a corollary obtained by Willems and
Blankenship (1971, Criterion 4) for systems where the deterministic component of the feedback gain is constant.
I
(Note that a factor ~ was omitted in that
reference in the statements of corollaries 3 and 4). Here it is shown that the criterion is valid for a much larger class of stochastic systems. Remar~ 8
The Lyapunov function can also be used for the system dE = (A_~x - k(~,t)bcx) dt - f(~,t)bcx d ~
(25)
The stability conditions are : (i) G(s) positive real. (ii) For all x and t : ~2 k(x,t)~ ~--qn-1 (f(x,t))
(26)
For asymptotic stability it suffices that either of the conditions holds in the strict sense, as explained above. Remark ~
Consider the case where the mean is non-zero. For system
with E(d~) = m, the function
G(s)/(mG(s)+1)
real. For system (25), inequality k(x,t)~ ~2w(O)
(19)
is required to be positive
(26) should be replaced by
(f(x,t)) 2 - mf(x,t)
For k(~,t) : O the system described by (16) is obtained. Using the same technique for Lyapunov function construction as in the proof of Criterion 3 the following related results are readily derived :
141
Criterion 4
The stochastic
system (19) has a null solution which is mean
square stable in the large and stable in the large with probability one if (i) G(s) -I - a
is a positive real function,
and
(ii) for all x and t :
k(~,t)~ a + ~2w(0)/2 for some (positive or negative)
(27)
constant a. The stability properties are
asymptotic if either of both conditions are true in the strict sense. Criterion ~
The null solution of the stochastic system (19) is mean square
stable in the large and stable in the large with probability one if there exist constants k I and k 2 such that (i) The function kiG(s)+1 k2G(s)+1 is positive real. (ii) For all x and t k1+k 2 k1+k 2 ---~--. - b ~ k(~,t) ~ --~--- + b
with
[~i_k2 b =
(28)
J
]~
[---~-- - ~- (qn_1(k1+k2)+2Pn_1-2rn_1)
where rn_ 1 is the coefficient
(29)
of s n-1 in the polynomial r(s)
:
r(s) :~[(klq(s)÷p(s))(kaq(-sl+p(-s)l+(klq(-s)÷p(-s))(kaq(s) +p(s) ] + that is the positive spectral
(klq(s)+p(s))(k2q~s)+p(~)).
Remark
10
factor of the even part of the polynomial
The frequency domain conditions of Criteria 4 and 5 have an
interesting interpretation in terms of circle conditions. criterion 4, requires the frequency response G(jW) completely within the circle shown in Fig.
Condition
(i) of
for positive a to lie
I. For negative a, the frequency
response G(j~) should lie completely outside the circle shown in Fig. 2, and should encircle the disk as many times as G(s) has poles with positive real parts (unstable open loop poles). This corresponds to a criterion obtained by Blankenship and Willems
for a much smaller class of stochastic
tems. The geometric interpretation
of condition
sys-
(i) of Criterion 5 is shown
142
Fig.
Fi~. 2
I
Im
Re
J
Frequency Fig.
domain
condition
of Criterion
4.
I : a positive
Fig. 2 : a negative
Fig. 3 Frequency
domain
condition
of Criterion
~ ~ Q(jw / I
-~2
I
kl,k 2 positive
~e
5
G(jw) -1
kl
k I positive, k 2 negative
kl,k 2 negative
-I ~Re
k2
143
in Fig. 3. Remark 11
The condition on the feedback element is different depending on
whether the positive or negative spectral factor is used for generating the Lyapunov function. This is a particularity which does not appear for deterministic systems. This condition is least restrictive if the positive spectral factor is used. This same conclusion is obtained if the Kalman-Yacubovich-Meyer lemma is used for constructing the Lyapunov function as is shown below. The choice of the spectral factor has no effect on the stability condition if k I or k 2 are infinite,
as in the cases considered in Criteria 3
and 4. Remark 12
The same stability criteria and Lyapunov functions can also be
obtained by means of the Kalman-Yacubovich-Meyer lemma and its generalizations instead of the path integral technique. Only Criterion 5 is discussed here. The positive real character of (kiG+1)/(k2G+1) is sufficient
for the
existence of a negative definite symmetric matrix P satisfying the matrix inequality (Willems J.C°
1971b)
P_~A + ATp - (Pb+mcT)(~T~+m~) + klk2cTc ~ O
(3o)
where m = (k1+k2)/2 . The stability criteria are then obtained from the Lyapunov function V(~) = -xTpx_ __ . Here also the stability result depends on the choice of the solution ~ of the matrix inequality (30). Willems (1971b) has shown that the smallest(-~) is obtained if the inequality
(30) holds with
the equality sign and if that particular solution is chosen such that the eigenvalues of A-b(bTp+mc) have negative real parts; this corresponds to the choice of the positive spectral factor as the solution of the factorization problem in the path integral method. 4. THE STOCHASTIC POPOV CRITERION It is wellknown in deterministic stability theory that less restrictive stability results than the circle criteria can be obtained if the deterministic feedback gain is assumed to be either nonlinear and time-invariant or time-varying and linear. In this section it is shown that similar conclusions are obtained for the stochastic feedback system considered in Section 3. The case of a nonlinear time-invariant deterministic
feedback is considered first,
and a result similar to the Popov criterion is derived. Consider the system considered in Section 3, but suppose that the deterministic feedback component is nonlinear, memoryless and stationary. The system is described by the It8 equation
144
d~ : (Ax-f(cx)bcx)
dt - bcx d ~
(31)
or p(D)y dt + f(q(D)y)q(D)y Criterion
6
dt + q(D)y d ~
The null solution of the stochastic
(32)
= 0
system described
by (31) or
(32) is mean square stable in the large or stable in the large with probability one, if there exist a positive
constant
a such that
(i) the function (1+as)q(s)/p(s) is positive
real,
and
(ii) for all u :
f u)
2 (aqn-2Pn-1+qn-l-2rnrn-1 +a(f(u")+u dr(u), 2 ) du )qn-1
where r n and rn_ I are the coefficients
of D n and D n-1 in the positive
ral factor of the even part of the polynomial The stability is asymptotic
(33) spect-
(1+as)q(s)p(-s).
if either of the conditions
holds in the strict
sense. The criterion is proved by means of the Lyapunov
function
(t(y) V(~) = J
dt [(1+aD)q(D)y
p(D)y - (r(D)y) 2 + aq(D)yf(q(D)y)q(D)y]
Co) where r(s) = ~ 2 [(1+as)q(s)p(-s)+(1-as)q(-s)p(s)] Remark
I~
Condition
The criterion
reduces to the Popov criterion if ~2 vanishes.
(33) is much simpler if qn-1 or w(O) vanishes,
is of degree
(n-2). Then f(u) ~
Criterion
+
that is, if q(s)
(33) requires
aqn_2Pn_1~2/2
6 reduces to the conditions
of Criterion 3 if a = O.
Next we consider the case of a linear deterministic
feedback gain compon-
ent. Then the system equation is d~ = (A_~x-k(t)bcx) dt - bc_~x d~
(34)
145
and the following stability criterion is obtained : Criterion 7
The null solution of the stochastic system (34) is mean square
stable in the large and stable in the large with probability one, if there exists a positive constant a such that the function
(1+as)q(s)/p(s)
is
positive real, and the gain k(t) is nonnegative and satisfies (with the same notations as above) k(t) ~ 2 ~ ( t )
~2
+ ~- (aq
p +q -r r ) n-2 n-1 n-1 n n-1
(35)
for all t. It is asymptotically stable if either of both conditions is true in the strict sense. R e m a r k 14
Criterion 7 reduces to C r i t e r i o n 3 i r a
= O. Otherwise the cond-
ition on G(s) is less restrictive, but the condition (35) is stricter. means of the techniques used in the stability analysis of deterministic
feed-
back systems the criteria obtained in this section can be transformed and extended in various ways. Here Criteria 6 and 7 are given to illustrate the possible criteria one can obtain. 5. DISCUSSION In this paper some stability criteria are derived for systems having a stochastic element in a feedback structure. The analysis has shown that the techniques for generating Lyapunov functions developed for the stability analysis of deterministic feedback systems also yield interesting stability results in the stochastic case. The analysis of this paper could be extended and generalized in various ways. Without conceptual problems, multivariable feedback systems can be dealt with; the path integral method is not well suited for this case, but the procedures indicated in Remark 6 and Remark 12 apply to multivariable
feedback systems as well. Discrete systems could also
be considered (Willems and Blankenship 197~); the analysis is even more straightforward,
since the subtleties of ItScalculus disappear. It would be
interesting to consider other types of noise (Blankenship 1972), and to see how the frequency domain conditions are affected. The weak point of the analysis of this paper is that only quadratic Lyapunov functions are generated, and all conditions are sufficient
for mean square stability. It would
be interesting to consider other types of Lyapunov functions in order to obtain less conservative conditions for stability with probability one, which would not necessarily imply mean square stability (Kushner 1967).
146
REFERENCES Brockett, R.W., Finite Dimensional Linear Systems , New York : Wiley, 1970. Blankenship, G.L., Stability of Uncertain Systems, Ph. D.thesis, M.I.T., Report ESL-R-448, June 1971. Blankenship, G.L., Asymptotic properties of stochastic systems : a nonlinear integral equation, Technical Memo 24, Systems Research Centre, Case Western Reserve University, Cleveland, Ohio, 1972. Kalman, R.E., Proceedings of the Nat. Academy of Science of the U.S.A., 49, pp. 201-205, 1963. Kleinman, D.L., IEEE Trans. on Automatic Control, AC-14, 429-430, 1969. Kozin, F., Automatica , 5, 95-112, 1969. Ku~hner, H.J., Stochastic Stability and Control, New York : Academic Press, 1967. Rabotnikov, Iu. L., Prikl. Mat. Mekh., 28, 935-940, 1964. Willems, J.C., The Analysis of Feedback Systems, Cambridge, Mass. : M.I.T. Press, 1971(a). Willems, J.C., IEEE Trans. on Automatic Control, AC-16, 621-634, 1971(b). Willems, J.C., and Blankenship, G.L., IEEE Trans. on Automatic Control, AC-16, 292-299, 1971. Willems, J.L., Stability Theory of Dynamical Systems, London : Nelson and New York : Wiley Interscience,
1970.
Zames, G., IEEE Trans. on Automatic Control, AC-11, 228-238 and 465-477, 1966.
STABILITY OF MODEL-REFERENCE S Y S ~ M S WITH RANDOM INPUTS
D,J.G.
JAMES
Lanchester Polytechnic, Coventry, England
I.
Introduction In recent years model-reference adaptive control systems have proven to be one
of the most popular methods in the growing fie ld of adaptive control.
The input to
the system is also fed to a reference model, the output of which is proportional to the desired response; the outputs of the model and system are then differenced to form an error signal.
Since this error signal is to be zero when the system is in
its optimum state it is used as a demand signal for the adaptive loops which adjusts the variable parameters in the system to their desired values. Various methods of synthesizing the adaptive loops have been proposed but the one that has proven most popular is that developed by Whitaker et.al. (1961) and referred to as the sensitivity or MIT rule.
Here the performance criterion is taken
as the integral of error squared and this leads to a rule that a particular parameter be adjusted according to the rule Rate of change of parameter
=
- Gain x (error) x ~ (error) ....
(~rameter)
Although the MIT rule results in practically realizable,systems mathematical analysis of the adaptive loops, even for simple inputs, prove to be very difficult. A stability analysis for sinusoidal input has been previously considered by the author (James 1969, 1971); however, in practice, a more realistic input is a random one and the purpose of this paper is to investigate stability for such an input. Since the object is to pose the problem and illustrate the difficulties involved we shall limit our discussiGn to a first order MIT system with controllable gain.
2.
L i n e a r stochastic,,,,,,,,,s,ystems All the stability problems c o n s i d e r e d i n t h i s
paper reduce to one of
investigating the stability of a system of linear differential equations with random coefficients.
A vast amount of literature dealing with such systems has been
148 published in recent years and various types of stability have been proposed (Kozin 1966, 1969).
In this work we shall confine ourselves to three types of stability,
namely, stability in mean, stability in mean square and almost sure asymptotic stability (a.s.a.s.) and will investigate such stability u~der two types of random coefficient variations, viz: (i) Gaussian white noise processes and (ii) Gaussian no n - wh i t e p r o c e s s e s . 2.| ~auseian white noise processes We shall be concerned with a system of equations of the form
w~.r.A.,
is an ~ . n
constant . ~ , ~
an lqXtl
m a t r i x and /~ = ~/~:0{I:)~
= ~ m } ,'
=
, ~ ,j = I, 2, ....-
1, 2 .... ,rl
an ~
column vector of
Gaussian white noise processes having properties
e[
o,
--
Such a system has been studied extensively in the literature (Ariaratnam and Graeffe 1965, Caughey and Dienes 1962, Bogdanoff and Kozin 1962) and we shall confine ourselves here to a brief outline of the method of stability analysis. The response of system (I) is a continuous n-dimensional Markov process and such processes are completely described by the Fokker-Planck equation ,1
rl
appropriate
81:-.o In stability
~i:--..o
6t;
i n v e s t i g a t i o n a knowledge of t h e moments o f t h e system r e s p o n s e i s
usually sufficient
and a system o f f i r s t
the moments o f o r d e r
K , X~a.. . . . . "~nKM entire
JC
state space.
N
order differential
equations determining
are r e a d i l y o b t a i n e d from (2) by m u l t i p l y i n g throughout by
i h(,+~a÷
"-
-+Kn=~
, and i n t e g r a t i n g by p a r t s over the
Necessary and sufficient conditions for stability in the mean
and mean square are then readily obtained by applying the Routh-Hurwitz criteria to
149 the system of differential equations obtained in the cases ~ = I, 2 respectively. 2.2 Gaussian non-white processes If the coefficient variations are Gaussian but non-white then the response of the system no longer forms a Markov process.
However it is possible to construct
linear time invariant filters which, with Gauesian white noise as input, will have the required non-white Gaussian coefficient variations as output.
The response of
the total system, which includes these linear filters, will then form a Markov process.
Unfortunately, the additional state variables introduced by the filters
render the system equations non-linear with the result that the moment equations, obtained from the appropriate Fokker-Planck, can no longer be solved recursively and one cannot obtain criteria for stability in the mean square. We shall be concerned with linear systems of the form
] where
A~. is an
processes and
~
nxrl
constant stahility matrix,
"
V~ ~ L~o,-,)
(3)
~A(~) are stationary ergodic
constant matrices.
Sufficient conditions guaranteeing a.s.a.s, of (3) have been obtained by many authors and the most recent improvements in the stability criteria obtained appear to be those due to Infante (1968) and Man (1970).
By applying the results of the
extremal properties of the eigenvalues of pencils of quadratic form Infants showed that a sufficient condition for a.s.a.s, in the large of system (3) is that
i.~-#
where .~..]~)
denotes the far, st ei~nv~ue of the ~tri~
symmetric positive definite matrix.
~'} and ~
is a
By simultaneously reducing two quadratic forms
to diagonal from Man extended the development ~
Infante to obtain the sufficient
condition far a.s.a.s, in the form II
i
(5)
~,---o
where
and G
are positive definite constant matrices satisfying
150
3.
,GaSn adaption model reference s~stem Consider a model and system to be governed r e s p e c t i v e l y by the equations
(6) where the time constant "I" and model gain K
are constant and known, but the
process gain Ke is unknown and possibly time varying.
The Iroblem here is to
determine a suitable adaptive loop to control K c so that
kv ~c eventually equals
the model gain ~ . The MIT rule gives
where 4.
= @Kv/~
, and this leads to the scheme of f i g . 1 .
Random Input
4.1 Stability analysis If to the system of fig.1 a general random signal g,.(l:J,@~l~ are zero and
when but ~
where
KV~a~ K
~(~
is applied at time
t=O,
and if subsequently K¥ remains constant
is adjusted according to (7) then the system equations (6) become
~£B-
K - K vK~(t~.
Despite the recent progress in stochastic stability theory, methods of investigating the stability of (8), where the system is not asymptotically stable when the noise terms are equated to zero, are not forthcoming.
However, digital simulation
of the system, which will be discussed in section 4.2, suggests that stability boundaries exist fcr such an input. we shall assume that
.he..
~ >>T
In order to have a first look at the probl~,
N
a~ A . ( . = o, I, 2, .... N)
are r~dem v a r ~ b l s s drawn from an
amplitude probability distribution l~(~). Substituting (9) in (8) and solving within the time interval ~ ~ ~ ~ • ( ~ . ~ furnishes the following recurrence relationship for
3c (~" "~..'~
151 Fig.1
First order system - MIT ~ain adaption
(gin(t)
K I+Ts
>
>
E)i(t)
gs(t)l Kv I +Ts
,ZT which on successive application leads to &,o
~T
Thus in order to investigate the stability of the system we must examine, by letting T-~ ~
the convergence of the infinite product
~T
+'=o
By considerin~ a large number of terms and their distribution and by considering the logarithm of the product of these terms we are led to consider the integral
If
~ >0
then ~
to zero,
~(~-bO
diverges and the system is unstable; if ~ < O ~
~1 ~oozand the system is stable.
If
~
then -~ diverges has a Gaussian
distribution with zero mean and variance o-z then (10) leads to the stabili~ criterion
~e-~ 7"Vs
1.2
157
Fig.3
Stab ilit~ boundaries for random input
5-
b-
T
Runge-Kutto
7%7 7Y 6 =175
CrGnk-Nicolson
\
.
0
I
28
!
!
56
84
!
1
112
140 >
-7-~-6
168
CD 0 b.) 0
o
c) c~
CD CD-
0
0
cD I
o
r
I
c--
c~ c-3 I
I
o
"lJ
"1
-1"1 o x-
o
o
I
I
Z_ °
o
?
0
(.11 0 I
(aL.) C~ C) I
~7Y8
?
laJ
9
o o
r.
o
159
The digital simulation results for the case when the input is purely random suggests that an outstanding problem is that of obtaining stability criteria for the linear system
~(k~ ~ A~t~ ~ )
, where the time-varying elements of
~)
are
correlated Gaussian non-white processes and the system is not asymptotically stable when the time-varying elements of A~)
are made identically zero.
ACKNOWLEDGMENT The author would llke to express his gratitude to Dr P. C. Parks, University of Varwick, for his helpful discussions concerning the work described in this paper. REFERENCES ARIARATNAM, S.T., and GRAEFE, P.W.U., 1965, Int.J.Control, I, 239; 1965, Int.J.Control, 2, 161; 1965, Int.J.Control, 2, 205. BOGDANOFF, J.J., and KOZIN, F., 1962, J.Acoust.Soc.Am., 34, 1065. CAUGHEY, T.K., and DIENES, J.K., 1962, J.Math. and Phys., 41 , 300. INFANTE, E .F., 1 968, A.S.M.E.Jour.App.Mech.,
5, 7.
JAMES, D.J.G., 1969, Int.J.Control, 9, 311; 1971, Amer.Inst.Aero.Astro.,
9, 950;
1972, Int.J.Control, to be published. KOZIN, F., 1966, Paper 3A, ~rd IFAC Congr., London; 1969, Automatica, 5, 95. MAN, F.T., 1970, A.S.M.E.Jour.App.Mech.,
37, 541.
PARKS, P.C., 1 966, IEEE Trans .Aut .Cont., AC-I I , 362. WHITAKER, M.P., OSBURN, P.V., and KEZER, A., 1961, Inst.Aero.Sci., paper 61-39.
REGIONS
OF I N S T A B I L I T Y
WITH RANDOM
FOR A LINEAR
PARANETRIC
SYSTE~
EXCITATION
W. WEDIG
Universit~t
Karlsruhe
(TH),
Karlsruhe,
Germany
INTRODUCTION The
dynamic
loads has b e e n leads
stability
to a linear
ternal
so-called
efficient.
second
order
stability
for
subjected
properties
are w e l l - k n o w n
by the M a t h i e u
flJ we k n o w those
diagram
frequencies the
and amplitudes°
system
becomes
bility
regions
quency
of the
If the stability The
zero
square,
equation
all second
stability
siam white The
trated
noise
(see fig. with
old value. constant
!),
But,
in the
of the above,
remain
of Gaus-
excitations,
above
the
It is p h y s i c a l l y
frequencies
such s t a b i l i t y
of the s t a b i l i t y
specillus-
density
suitab-
Comparing
we find the
of w h i c h vary
coefficients
in m e a n
case
spectral system.
her.
noise
the
coefficient.
zero and a flat
mits
for n o n - w h i t e
fre-
manner,
may be g r a p h i c a l l y
and we find no i n s t a b i l i t y that
insta-
f2J.
limit
excitation
of
itself.
coordinates
stability
obvious,
excitation
the natural
a stochastic
the critical
to harmonic
corresponding
said to be stable
condition
map m e n t i o n e d
excitations
important,
twice
of its state
frequency
is a uni-
in a stochastic
the m e a n value
if we r e c o r d
in contrast
and most near
has been a p p l i e d
stability
co-
of e q u i l i b r i u m
frequency
contains
to infinity
w h i c h has
stability
two,
is then
order m o m e n t s
by the n a t u r a l
the
above
definition
f3J,
for the
fluctuates
system
tends
corresponding
ly m u l t i p l i e d figure
mentioned of the
to the
of values
situated
ex-
force.
According
the natural
excitation
if the time
This
are
the
for periodic
the p o s i t i o n
The first
diagram
and near
external
when
bounded,
trum.
system
position
for w h i c h
unstable.
of this
ranges
axial
study
dependent
problems
to a time v a r y i n g
equation.
this
in w h i c h
as a time
such stability
system
fluctuating
cases,
equation,
appears
of this
Strutt
under
In simple
differential
excitation
example
column,
structures
investigated.
parametric
A technical
form p i n - e n d e d The
of elastic
extensively
this
same thresh-
we have here regions,
in a similar limits
must
equation.
a
the liman-
exist
These
li-
161
mits
are to be d e r i v a t e d
If we select
in the following.
STATE
EQUATIONS
the i n i t i a l l y
mentioned
a simple
example
equation
m a y be taken
of stability
problems,
column under
the
end thrust
corresponding
as
stability
in the f o r m
y ÷ 2w, O 2 . w$[1+ cx(t)]y = O. The u n k n o w n transverse
time f u n c t i o n motion
of the
parameter
dimensionless
coefficient
For lightly
damping
nential
of the
damped
as well,
excited
(E, DLLT).
then
~t is its n a t u r a l
problems
and w e a k l y to one
equation
external
in the m a t e r i a l
such m e c h a n i c a l
in c o m p a r i s o n
of this
column,
dimensionless a viscous
y(t)
(1)
excitation
by means of the
describes
frequency,
the ¢ is a
Ex(t) and D is a
of which we are a s s u m i n g
column.
we can restrict systems,
our interest
to
so that $ and D are small
It is then possible
to make
the
expo-
substitution
y(t) = T(t) expGcu~DO for
y(t}
and %o convert
the original
(2) equation
i) into
the
shorter
form
T÷~211÷ ¢,x(t)]T = e,=¢/(1-09, containing of the state
the small
system.
This
=
parameter stability
¢7 and the damped equation
natural
can be w r i t t e n
frequency
as a set of
equations
T-z~÷z 2, t= iv,(z~-z2), (3a)
z; where
z 7 and
z2
[zr
x rz, . z/2].
are the g e n e r a l i z e d
(3b) coordinates
of the
system.
162
LOW-PASS
We
shall first
Such p r o c e s s e s through
having
of all discuss
can be o b t a i n e d
a low-pass
a low-pass
by p a s s i n g
x(t)
process
~(t). noise ~(t)
G a u s s i a n white
=
filter
an a r b i t r a r y
ter's r e s p o n s e
PROCESSES
limit
frequency
is s t a t i o n a r y
~g. A f t e r
and its spectral
a short
time,
density has
the fil-
the w e l l -
known form
Herein
the
coefficient
its a u t o - c o r r e l a t i o n Since the input
x(t)
response
form M a r k o v p of all
= ~(~
ker-Planck
function
~(t)
z 2 and
3
i=l OZi
the
equation
the e x p e c t a t i o n
the c o n d i t i o n a l
z,
and
z2
probability
z 3 is given by the
a i
there
its
of (3) density
corresponding
available. values
and
Fok-
For
E z,l,
b, , --limS E[~zj Azl, •J
bgj are the
indicated
4t~.0~¢
J
incremental
moments
and
in [4].
our purposes,
of all state
of s o l v i n g however,
coordinates
the F o k k e r a knowledge
is sufficient.
of
Accor-
of these m o m e n t s =
to m u l t i p l y
a,- lira "
is no g e n e r a l m e t h o d
M.,.~.k7 E[z~'zsk~z~"] state
2
by the m e t h o d
ding to the d e f i n i t i o n
of the
coordinates
z -(z,;z2,z3 )r,
coefficients
computed
At present,
we have
of
is G a u s s i a n white noise,
state
L ~j=J uLiua~ i
p..- p(z,t;'Zo,to),
Planck
as the
~(t), so that E[~(tl)~(t2) ] = So6(tT-t2).
spectrum
equation
C1~
in w h i c h
is g i v e n by
Therefore,
zl,
3
m a y be
constant
of the f i l t e r
as well
processes.
coordinates
So is the
both
coordinates
L#z~z22z;~ p d~dz2dzs, sides
of the F o k k e r - P l a n c k
and to integrate
e q u a t i o n by p o w e r s
it over the
entire
state
plane. As a result equations
of this procedure,
of the moments,
which
we then o b t a i n
are best w r i t t e n
the d i f f e r e n t i a l
in m a t r i x
form [5].
163
(~)
2z~ " = A kt~fk. Tlg=/ k(k_ l)M k_2* ¢~L~k., '
¥,-1Efz, z,#A, "
~,-
LE[z22z~]J
i, ,ool -kr~
They contain the d i m e n s i o n l e s s ance
~=
~ g ~ / 2 of the low-pass
are n o t e d above.
moments
of the system
limit frequency process.
0
.
-1
~g=~g/2~and
the vari-
The m a t r i c e s A~ and R of (4)
decision.
of the equation's
It becomes
of linear differential
evident,
is the exponential
If k is equal
vector are the second order
that these moments
general
are weakly
so that we actually have a
equations with constant
APPROXIMATE A sufficiently
process.
in-
(3), whose b e h a v i o u r we have to examine for the
coupled by E 7 with higher order moments, sequence
e-~
the power of the low-pass
the components
stability
,
The index k of the moment vector ~k is a positive
teger and indicates to zero,
0
0-i-k~J
coefficients.
SOLUTION
solution of the h o m o g e n e o u s
equations
(4)
fumction
l~k(t) = ~k exp(2~ 9~ with an amplitude multiplied equations
by
vector
2~.
(5)
~k and the u n k n o w n
W h e n we introduce
(4), we get an infinite
eigenvalue
9 adequately
this f u n c t i o n into the m o m e n t ~
system of h o m o g e n e o u s
algebraic
equa-
tions. ~C, '
2_~.E
For
~T equals zero,
~-~,.E
the latter can exactly be solved.
obtain three times the infinite magnitude
~- -k,7,,.~,,,
In this case we
of eigenvalues
164
and the
corresponding
c(O)
_2~ (k÷21)!
amplitude
vectors,
"n(n+l)/2]
as n o t e d
below.
~-',ik÷21.,=O, n-O,± 1, k~l--O~l,...
n(n-l)/2J The n u m b e r n has the roots tive
integer
plitude
proximate genvalue
The
£1 is not equal
solution
of its
can be solved It can easily
real
to zero,
part.
Thus,
the
the three for each posi-
comoonents
values
of
of the am-
k.
it is reasonable
to seek an ap-
method
the u n k n o w n
amplitude
expanding vector
of (5)
in power
eise-
El'
coefficients
results
(o(O)E.A JC(o + ~'.o-j)p(j)
¢[..
that
for i n c r e a s i n g
by a p e r t u r b a t i o n
small p a r a m e t e r
comparisom
which
are u n b o u n d e d
and determines
Ak, which we can obtain
9 and the c o r r e s p o n d i n g
of the
n =~I
values
matrices
k. It should be remarked,
vectors
In case
ries
three
of the diagonal
in the recurrence
2
~
formula
(,-~
step by step. be shown,
only
that
the first
the a p p r o x i m a t i o n s
eigenvalue
of this
has the greatest
eigenvalue
need to be
noted here.
9(')=[i In particular, vanishing These
the
solutions
and the others
results
contain
are c a l c u l a t e d
only positive
xp - l/(~÷i/p), are complex,
of an odd degree
conjugate
STABILITY As m e n t i o n e d
up to the fourth
real parts,
~p - I/?rt:i/p:
respectively
above,
because
are
approximation. the values
(p= 1,2)
complex.
IN M E A N
the first
of a p p r o x i m a t i o n
SQUARE
eigenvalue
has the greatest
real
of
165
part.
It therefore
order m o m e n t s with the
determines
of the
initially
state
the
increasing
equations
introduced
behaviour
of all secomd
(3) and it determines,
exponential
function
(2),
together
the b e h a v i o u r
of the moment
E[y2(t)]:E#z,+z2}~exp(-2u6DO of the
original
position
stability
of the system
equation
(i) will
(i) as well.
be stable
Consequently,
in m e a n
the
zero
square
fim=E[y2(t)]= k= const, if th@ value
of this root
is not
greater
than the
damping
coefficient
D.
2~9-2~D~0 In this
stability
solution
condition
of the first
we now have
eigenvalue.
to introduce
If we restrict
terms up to the order
~
tical v a r i a n c e
of the parametric
Upon density
(~I~} 2
computation
in a stability spectral
map
densities
point
system.
(fig.
This
excitation
we have
been
coefficients
The
value
to those
for the
cri-
e x(t) [6].
the critical
see that
however,
smaller
densities
of its limit the
near
twice
spectral
frequency of all
kmown
white
the
frequency
stability
of
con-
noise [3] and sto-
[7].
varying
spectral
demsities
do not
endanger
the stability
without
~o
critical
region,
the natural
is the already
the bandwiths,
may be,
envelope
of an instability
for G a u s s i a n
w i t h weakly
processes,
as much. spectral
derived
values
the limit
is s i t u a t e d
threshold
chastic
Low-pass
i). We
defines
of w h i c h
having
their
equation
2
it for the v a r i o u s
dition~
system
of this variance,
~ 2~(~,~
and we can plot
the
our solution
a quadratic
as well
~
lowest
, we obtaim
the approximate
the greater destabilizing
of the
the m a x i m u m
of
the
In
system.
166
I
0 1~
"
ol
4
Figure the case
of small
cess will
first
per p r o x i m i t y
limit
touch,
of twice
From
frequencies, and then
an i n c r e a s i n g
cross,
the natural
OF THE POWER
point
of view, which
For this
t e r m of the it since
ments's
the r e c u r r e n c e
to simplify
equations
= T and of the
purpose,
expansions.
formula
the real
dimensionless
add up two of these
Id
22
sum of the q u a d r a t i c
SERIES these
results
of course,
2
moments
are only en-
the power
series
(6)
to k n o w the
impossible
to cal-
(7) is too complicated. we must ~=
first
coordinates t/~
of the
so that we obtain
equation
of the proin the up-
of the system
conditions
state
velocity
equations,
limit
it would be n e c e s s a r y
It is,
this formula,
by t a k i n g
processes
spectrum
stability
frequency
under
In order
for the
the
the m a t h e m a t i c a l
are convergent. culate
m a p of low-pass
if we can determine
general
'
......
i: Stability
CONVERGENCE
sured,
~''---
transform
the mo-
of the p o s i t i o n system.
a single
We t h e n
differential
167
E[(~2 ÷J~2)x~ = Dk exp(2~ 9~ Ok and
the a m p l i t u d e
the
eigenvalue
9 of w h i c h we shall
investigate
in the following. The m i x e d moment, is smaller
than half
/E~
remaining
on the right
the absolute
side
of the
sum of the quadratic
If on the r i g h t - h a n d
side we n o w increase
by means
side decrease
eigenvalue,
simpler
(8),
moments.
~2)xk÷]//2 _~ E[~x~'~
and on the l e f t - h a n d the
equation
recurrence
by the
of this relation, we t h e n obtain
formula
=
for a m a j o r a n t
of the
t h a n zero.
coefficient
The
amplitudes,
the indices
Do and the
of which
eigenvalue
are greater
~ itself
are given
by
Applying
the p e r t u r b a t i o n
two terms
method,
we can exactly
~k " uk ;,(o)÷¢,u~ ;,(1)÷ E;T _2F,(2) u k +..., and must
then estimate
the h i g h e r
-(I) 2(1÷1), D;;÷,=(1,.2,1÷I)~ '/2~9 , -(0)
ed by c o m p a r i n g
and
the
we introduce
formula
terms
once
again. 1+
2'_1!,[i... j -I1, . _ .
1
21
/-I (2j-l)2j],
:n÷t) n
!
inequalities
of this
coefficients
of the same powers
this
and o b t a i n
~2,.J The most
second majorant
second majorant two inequations
important
part
n
2 _~ 2177 1..
H[1÷1+2i(2i.3)] - i=t F[[1÷2i(2i÷3)]-
i=1
D(°)..- 1, D.o(° 0 d-1,2...)
=(2n) n O~, =(1,2;1)1a; 21(a~/rle) 2nj;~[,
=(2n,I)
The r e s p e c t i v e purpose
the first
O;;.,=(l;2,1÷l)~(l"3)a,~(a'~"~, )2 j~.2[1÷2j(2j÷l")]" l..-O,l, J2~.., n..-1,2,3 ....
2/
O;,,(1,2,1)ax,
rence
calculate
of the' expansions
into the
I. For this
corresponding
for the amplitudes
of the first
12- .].
of
may be prov-
2t
2 .1
leads
inequality
.
.I .
l-
to
],
=" -=. 1"711 ÷ 2 i ( 2 i ÷ I ) ] , Z i=2
recur-
-'~-D~÷tJ
0,1,2..., n = 3/~,S...
(9)
168
and shall The result
be n r o v e d first
for
the
ln'7(n~l): In the (9)
of
are
to
p=n-land
easily
1"-2(n~-2):
p=n-3we
obtain
s u m on the
its two
able
{P can p=n-2in of
~-- O,
snd a double and
following.
coefficients
powers
1-I
case
glected we
two
in the
the
side.
on b o t h
Its
first
sum°
power
of
we
obtain
the
inequality
of
so t h a t
of the inequality.
1"-3(n~-3): E-'[1. 2i(2i.,. 3) ÷ 2j(2j .3)]- ~ (n-2) Z2i(2i+ n-, n-, 3) ~ ~.. 22i(2i+I) Similarly,
side
t e r m m a y be n e -
to a s i n g l e
sides
~j=l{i,q)
and
n-l~-n/2.
can be r e d u c e d
sums
calculated
sum on the r i g h t - h a n d
left-hand
last ~rms
compare
a single
be
i=l
Z 1=2
of the
coefficients
"
of a g e n e r a l
{.
n-I
n-1
In-p(n-~ p) : (n-p+ l) Z 2 i,(2 ~+3). . . Zip_fl2ip_2+3)_~ ~ 2 2i, . . . 2 ip.2(2ip_2+l).
,,~"J~".,.~p~--l ~-2J.,
It is o b v i o u s , smaller comes
that
n/2.
than
greater.
this
For
calculate
on the r i g h t - h a n d
(p-l)
side
finite
p-I
last
inequality
increasing
Consequently.
•
the
sums
side
is f u l f i l l e d ,
values
it m u s t
-
of p its
be p r o v e d
? 2ip_2(2ip .2+3) =- .
I ° 22i,(2~+3)... i7,...i~>.2=1 (J,,--J,-2) W h e n we
"-'~' f,i. .,p.2--2 .,,,.2:
on b o t h
a single
finite
left-hand
o n l y for p
is
side beequals
n.
:,;/).
~-;
sides
(n-p+1)
.....2,;
i,,. • JpL2=2
(g-
if
of t h i s
pro@uct
inequality , we o b t a i n
a n d on the
left-hand
products.
p-1
p-I
p-1
TT2i(2i.3)+ 17"2i(2i+3)+ ... + 172i:2i+3) ~ p TT2i(2i+l).
i=l(i*U
The
first
i=lfi.2)
two
terms
i=1 (i#p-1)
of t h i s
(2p÷1)/5 +(2p+1)/K The
other
inequality
By means is an u p p e r
may
of t h e s e
bound
+
sum
of the
suffice
...
we
= i--2
to v e r i f y ~-
be p r o v e d
results
Z
p/2,
the
inequality.
p..-3,,(,,...n.
in a s i m i l a r m a n n e r . arrive
amplitudes
at a c l o s e d
as w e l l
as of the
solution,
eigehvalue
self.
21÷I
oo I 2n n i (1,2;l+I)G, 0
the b o u n d n e s s
analytically.
different
-
w h e r e w(s)
n
XiqkLFk~(O)
t e r m of the c h a r a c t e r i s t i c
do is to c o m p u t e
Invoking
n
[
i n t e g r a l must v a n i s h
of the
system,
i.e.,
i ... + a2s+ a 1
for k+~=odd.
Frequency Domain Stability Criterion:
frozen
(25)
Here we a r r i v e
at:
Assume that the frozen system is stable.
Then the linear stochastic system of Eq. (1) is mean square stable if and only if n
i -
n
~ ~ (-I) m-k q k & K m > 0 k=l &=i
where Km =
(j~)m-l~(j~0
2d0a
-oo
ii.
There
is an a l t e r n a t i v e
to the N e v e l s o n minantal
calculations, A(0)
Aij(0)
After
criterion•
we can o b t a i n = 2n H
This will
w a y to c o m p u t e D(0).
& Khasminskii
performing
lead us
some d e t e r -
as*
(26)
n
= (_i) n-k 2 n-I Hij
where a n an_ 2 an_4
Hn
•..
i
an_ I an_ 3 ...
0
an
an_ 2 . • .
0
0
0
°.•
0
e n e n _ I en_ 2 ...
e1
0
I
an_ I an_ 3 ...
0
0
an
an_ 2 ...
0
0
0
0
aI
0
,
Hij
aI
=
...
in w h i c h *This is not difficult, but somewhat lengthy.
So it is omitted here (cf. [12,13])
179
em = It should for the
be noted frozen
its first
for
i + j : 2m
0
for
i + j # 2m
that
system,
H n is equal
fact
Algebraic Criterion:
order
H.. is r e s u l t e d ij (en,en_l,...,el).
Hurwitz determinant
from H
by r e p l a c i n g n Thus we can compute
Hij Hn
: (-I) n-k
gives
to the n-th
and that
row by a row vector ~ij(0)
The above
I
as
(27)
:
Assume that the frozen system is stable.
Then the linear
stochastic system of Eq. (I) is stable in the mean square sense, if and only if i n
(-1)n-k
n
(k+g=even) which is very relevant to Nevelson & Khasminskii criterion [14] Some other discussions on stability will be includes in appendices. 4. E Q U I V A L E N T 4.1
Derivation
of E q u i v a l e n t
the d i f f e r e n t i a l
equation
dnxe ~+ d tn where
b k ~ 0, and
In some Eq.(28)
instead
criterion
constant
of the
described
fl(t) is a Gaussian
white
it is of interest
form
covariance
selection
taken
here
e{X_e(t)}= {xe(t) , ~e(t)} that
noise
with E{B(t)fl(T)}= qS(t-T).
to be asked q should
here
of
is how the
be selected.
The
that
E{~(t)}
= coy
Eq.(29)
(28)
to use the r e p r e s e n t a t i o n
The q u e s t i o n
verified
by
dk-lxe
of Eq.(1).
coY
a system
coefficients
b k and the noise
for the
It is easily
Consider
k~I= b k dtk_-------i--: bo + B(t)
applications,
coefficients
System
with
n
SYSTEM
{[(t),
holds
(29)
~(t)}
for the
same
initial
condition
if we choose
bk = ak n q = qoo- 2 ~ qokmk(t)+ k=l It is also p r o v e d equivalent
that
in the sense
k = 0,i, n k=l
the constant that
...
, n
(30)
n ~ qk~[mk~( t)+ m k ( t ) m ~ ( t ) ] ~=I
it will
coefficient have
the
s y s t e m with same
Eq.(30)
correlation
is
matrix
180
and hence the same cross spectral densities this sense, the system defined by Eq.(28)
(in the steady state).
is said to be equivalent
In to
the original system defined by Eq.(1). Application to State Estimation
Briefly the problem is to estimate the
n state vector ~(t) of the stochastic
linear system described by Eq.(1),
provided that the m vector valued observation: d~(t) = H~(t)dt + Rdl(t) is available.
(31)
Here H and R are mxn and mxs matrices respectively,
and
~(t) is the s vector independent Wiener process. The equivalent
system to Eq.(1)
d~e(t) where
~ = (0,0,
is given by Eq.(28) with Eq.(30):
= A~e(t)dt + ~[aodt+ dB(t)]
... ,i)'
32
Similarly the equivalent
system to Eq.
31
can be expressed by dYe(t) = H~e(t)dt + Rd~(t) Thus the equivalent Eqs.(32)
33
Kalman filter can be immediately obtained based on
and (33) as dL(t)
= AL(t)dt+
~a odt+ K[d~e(t)- HX_e(t)dt]
(34
where K = PH'R -I
(35
= AP + PA'- PH'R-IHp + uu'q Since the actual observation process is given by Eq.(31), ~e(t)
in Eq.(34)
izable filter.
the process
should be replaced by ~(t) to obtain the physical realThus
d~(t) = [AS(t)+ ~ao]dt + K[d[(t)- H~(t)dt]
(36)
It is also not difficult to prove that the estimate x(t) governed by Eq.(36)
is the optimal linear estimate
The use of the equivalent
in the minimal variance
sense.
system provides with a simple method for
determining the optimal linear estimate of the stochastic
linear system.
In appendix A.3 extension will be made to the more genral case. 5. CONCLUSION Some analytical results have been presented here for the single input-single
output linear stochastic
system.
with the derivation of moment differential moment of output variables
is presented
The description begins
equation from which the second
in somewhat
compact form.
the correlation matrix can be expressed in a similar way.
~en
It turns out
181
that the output frozen
system:
spectral however
An explicit mean square
density
the magnitude
stability.
The resultant
for estabilishing
the stability computation
it is very doubtful
for the general stochastic
is increased
form of the condition
Thus the required However
has the same shape
vector
of the sigle
whether
case.
the equivalent
stochastic
system,
some simple
order
is presented.
be straightforwardly
extended
a simple method output
criterion
system.
does exist
to the feedback
et al [15].
system
is derived
as ones
to the state
Some of results to the more
the
conslderablly.
attempt
statistics
and its application
noisy observation
gives
Input-single
An alternative deterministic
that it has the same second
for quaranteeing
labour can be lessened
system has been made by Sawaragi
Finally
by the COnstant factor.
is obtained criterion
as that for the
general
in the sense
for the original estimation
presented case,
from the
here may
as indicated
in
appendix. APPENDICES A.1 Stochastic Liapunov Method i
Consider
the free system as
dx i = xi+idt n
(37)
dx n = -k~l[akdt+ An equiblium
solution
in question. probability tisfies
is evidently
speaking
the equiblium
LV(x)
if there
= - k(~)
Let the Liapunov V(£) where
of Eq.(37)
Roughly one,
exists
,
P is an nxn symmetric --
the first frozen
=
a Liapunov
system.
whose
solution
function
(Ax)'V
--
be of the quadratic
stability
is stable
V(~)
Assume
and
sa-
form
positive
definite
matrix.
Then
x + Pnn ~ ' Q ~
(38)
is related
that the frozen
to the stability
system is stable,
of the
then the matrix
so that
the covariance
from Eqs.(37)
with
> 0 which
(A~)'V x = - x'Qx
because
is
P = (Pij)
term on the RHS of which
P can be chosen
~=0,
< 0 (cf. Kushner[10]).
function
= x'Px
LV(x)
dak(t)]x k
matrix
(39)
Q is non-negative
definite.
It follows
(38) that
LV(~)
= (Pnn- l)x'Qx
from which the stability
condition
(40) is obtained
as
Pnn > I.
Thus the
182
problem is only to evaluate
the element Pnn"
From Eq.(39)
aiPjn + ajPin- Pi-l,j- Pi,j-i = qij
(41)
(i,j = 1,2,...,n) which is a set of equation putation,
for n(n+l)/2
unknown Pij"
By a direct
the Pnn can be solved as (cf. Nakanizo[12]) n n l)n- k 1 X [ (Pnn = 2 ~ n k= 1 ~=i qk~Hk~
Thus we arrive at the same algebraic should be noted that the stability lity with probability
criterion
(42) as presented
of second moments
in 2.
It
implies the stabi-
one.
Stability of r-th moments*
A.2
com-
It follows t
from Eq.(5)
that
~{V(x)} - ~{v(~o)} =I E{LV(x_)}dt O
If E{V(~)}
~ n~ (_l)m_kqkiKm k=l 4=1 (k+~=2m)
A.3 Application of EquivalentSystem to State Estimation system may in general be described by n dxi(t) =j=i ~ [aijdt+ dwij(t)]xJ+ dWio(t)
A linear sZochastic (47)
(i = 1,2,...,n) where aij are non-random coefficients, and wij(t)[i=l,...,n:j=0,1,...,n] are the standard Brownian motion processes. It is supposed that there is a set of noisy observations n
dYi(t ) = ~l[hijdt+ dvij(t)]xj+ dVio(t ) (48) j= (i = 1,2,...,m) where hij are constant, and vij(t)[i=l,...,n:j=0,1,...,m] are Brownian motion processes. Equations (47) and (48) can be written in vectormatrix form
d~(t) = A~(t)dt+ dW(t)~+ dWo(t)
(49)
d~(t) = H~(t)dt+ dV(t)~+ d~o (t)
(5O)
The incremental covariances of Brownian motion processes are given by ~{dwij (t)dWk~(t)) = qijk~dt (i,k=l,...,n: J,~=0,1,...,n) ~{dvij(t)dVk~(t)} = rijk~dt (i,k=l,...,m: j,~=0,1,...,m) and
184
~{dwij(t)dWk~(t)}
= Sijkzdt (i=l,...,n; J=0,1,...,n; k=l,...,m; ~=0,1,...,m) The present problem is to find the minimal variance linear estimate of the state x(t) provided that the process {y(T) t /to llx(t; Xo 'to ) II > c
Definition I n.
}
< c
(z. 3)
L zapunov Stability in the M e a n
The equilibrium solution is stable in the m e a n if the expectation exists and given
e > 0 , there exists
6(e, to) such that
llxo II< 5 implies
z ~t t s~p ll~t; x o, %)II } < ~ >~t ° Definition I
a. S,
c2.4)
A l m o s t Sure L y a p u n o v Stability
The equilibrium solution is said to be almost surely stable if lira
sup
P{ ilXo II ~o t >~t o llxlt; =o, to)II=
0 }= 1
12.5)
A l m o s t sure L y a p u n o v stability states that the equilibrium solution is stable for almost all sample systems.
This is the s a m e as saying the Definition
I holds with probability one. Asymptotic stability can be extended to the stochastic case in a similar way.
Definition IIp
Asymptotic. .Stability . . . . . .in . .Probability .
T h e equilibrium solution is said to be asymptotically stable in probability if ID holds and if there exists
5 > 0 such that
llx° II < 5 implies
191 lira T'*~ for any
{sup t~T
P
11 x ( t ; x o , t o )
II > ~
}
(z.6)
: 0
C > 0 .
Definition ~rn"
Asymptotic Stability in the M e & n
T h e equilibrium solution is said to be asymptotically stable in the
= e a n if ~
holds and if there e=sts lira E ~ sup
T~
T
D e f i n i t i o n TI
a.s°
IIx(t; % , % )
~ > 0 such that
II
llXo II < ~
~plies
: 0.
(Z.7)
A l m o s t Sure A s y m p t o t i c S ~ b i l i t 7
The e q u i l i b r i u m s o l u t i o n is said to be a l m o s t s u r e l y a s 3 a ~ p t o t i c a l l y stable i f
I
a. s.
holds and
P~.T-.~o
t>~T
llx(t;Xo'to)]]:°
~l
.
(z. 8)
T h e definitions above, as w e have stated before, are direct transitions to the stochastic setting of L y a p u n o v stability. These stability concepts are concerned with sample behavior on the half line. In the early stages of the development of this subject, m o s t studies w e r e concerned with the stability of various statistics of the solution process at a given time, rather than of the samples.
This is probably due to the fact
that it is easier to study statistical behavior than to study sample behavior. T w o typical stability concepts related to the m o m e n t s
are as follows;
[ 53, [15~ . Definition Ill. L yapunov Stability of the M.gan T h e equilibrium solution is said to possess stability of the m e a n if the expectation exists and lira ,,"Xo II ~
0 E t ll~It; x o, to~ I~ : 0
for allt >ito .
Iz. 9~
192 D e f i n i t i o n IV.
Exponential
The equilibrium
Stability of the M e a n
solution is said to possess
mean if the expectation
exists and if there
exponential
exists constants
s t a b i l i t y of t h e
a,
8,
5,
all
llx° II < 6 implies
greater than zero such that
E{ IIx(t;x o, to~ II} < ~ llXoII exp :- ~lt-t o) S for all t > t
iz 10~
o
Although the stability definitions IH and IV above do not appear to be as strong a restriction on the solution process as given in Ip-I a. s. ' IIp-IIa. s. ' there are significant implications in IIl and IV for s a m p l e stability behavior [ 16~, [ 17~ . H o w e v e r ,
stability of the m o m e n t s
alone does not always provide
a satisfactory intuitive basis u p o n w h i c h to judge the stability characteristics of the s y s t e m .
This can easily be illustrated by the simple first order linear
Ito differential equation, dx = axdt + ~x
w h e r e a,
dB
,
(2.11)
q are constamts, and the
process with
E{ Bzlt/} : t,
The solution process o b t a i n e d v i a t h e Ito c a l c u l u s
B -process is the zero m e a n
Wiener
~14~ . t o (2. 11) w i t h i n i t i a l c o n d i t i o n x ( o ) = x °
, is
as
x(t) = x ° exp [(a-qZ/z)t + v B(t) ] , with probability
one.
Furthermore,
E From
{
(2. IZ)
xn(t
t h e nt h m o m e n t s
,}
= x n exp o
E, o
of (2. 12) a r e e a s i l y s h o w n t o b e Z 2
/2)nt + ~
t
L . A
(2. 13)
(2.13), w e find that there is exponential stability of the n
a n d only i f
th
moment
if
2
a < ~
(n-n 2)
(2. 14)
Thus, for a < 0 , the first m o m e n t are unstable.
is exponentially stable, but higher m o m e n t s
F o r a < _(yZ , the first and s e c o n d m o m e n t s
stable, a n d higher m o m e n t s
are unstable, etc., etc.
are exponentially
It s e e m s
difficult to
associate a physical m e a n i n g to the s y s t e m behavior, k n o w i n g only that the
193
first m o m e n t N
moments
even m o r e
is stable but the s e c o n d m o m e n t are stable a n d all higher m o m e n t s
is unstable, or that the first are unstable.
To make
matters
interesting, it is well k n o w n that the s a m p l e functions of the B r o w n i a n
m o t i o n g r o w no faster than ~ t log log t , with probability one.
Therefore, the
stability of the s a m p l e solutions (Z. iZ) are d e t e r m i n e d by the algebraic sign of 2 ~2 a - Z , b u t (3. 1 5 ) , p o s s e s s e s
case is also illustrated development,
in
F i g . II.
dramatic
w i t h (3.15) since the
a horizontal
For further
asymptote.
details concerning
s e e [19] .
What has been accomplished that the sufficiency
b y (3. I6) i s a l s o c o m p a r e d
in this section?
conditions that have previously
Basically, appeared
we have seen
in the literature
for linear systems with stationary ergodic coefficients apparently are quite conservative w h e n c o m p a r e d with the yet to be determined, true stability region. Simply by adding slightly m o r e of the coefficient process, previous results. [iZ],[ 6 ] and
detail concerning the statistical structure
w e have achieved a rather dra/-natic advance over
The succession of stronger sufficiency conditions beginningwith
[ ii ] and n o w the conditions (3. ii), (3. 16) bring with t h e m the
question "where does it finally end? ".
W h a t is the true stability boundary for
the second order system (3. 6 ) with a stationary ergodic Oaussian coefficient? T o this question, unfortuantely, w e have no a n s w e r at the present time. But, let us return to the equality (3. 2 ) . U p o n dividing by t , w e obtain
log/Ix(t)lip log I~(oIIIp _
t
i = T
,~
t
J O
x'[(A+F(s))'P
+
(3.17)
P(A +F(s))] ds
x 'PX
If it can be established that the quotient on the left hand side of (3.17) remains negative as t approaches infinity, with probability one, then it m u s t follow that
199
liml itll 0
t'~c°
,
1318
with probability one, yielding the almost sure asymptotic stability of the equilibriu~ solution of (3. i), since (3.18) implies Lyapinov stability for linear systems. If the quotient remains positive as t approaches infinity, with probability one, the equilibrium solution is unstable. Hence, the algebraic sign of the limit, as t approaches infinity, of the integral on the right hand side of (3.17) b e c o m e s the necessary and sufficient condition for almost asymptotic stability. That is, if the limit is negative, w e have stability and if the limit is positive, w e have instability. O n e must, therefore, establish the existence of the limit with probability one, and then evaluate this limit. At this time, there does not appear to be an easy w a y to accomplish this for the systems that w e have been investigating in this section. However,
quite recently Khaz'minskii has recognized that such a limit
can be studied for linear Ito stochastic differential equations [ 20]. The next section is devot ed to applications of 14~az'minskii's results o
200 IV.
Linear It~ Differential Equations In this section w e shall be c o n c e r n e d
with linear stochastic Ito ~ differential
equations, w h i c h can be written in differential f o r m for i = 1. . . . .
~ , as
L b! n L dx. = Z x: d r + E Z ~ x. dB , z j=l I ~ r=l j=l ir j r where
b!1 ' ~ r
are constants a n d the
eesses f o r w h i c h
0
It is well k n o w n
B
(4.1)
are mutually independent W i e n e r
E
:
lt-sl
pro-
•
~ I ] that the~unique, solution process to the stochastic
s y s t e m (4. i) is a M a r k o v
diffusion process, with an associated generator < ,
defined by t Z ai~(x) i, j=l
u = (Bx, g r a d u) + where
&
aij{x) = E k, s =i
n rE= l
.2 x
(4. z) j
k s g i r ° j r x k Xs
and B = (b~) i.
U p o n applying the It~ differential f o r m u l a [22]
log llxll, where
x is the solution p=ooess to the system (4. iI and llxll is
the Euclidean n o r m
(x, x) I/2 , one obtains the expression in differentials,
d log IIx ll~(log IIx/I dt whero,
to t h e function
k = x/llxl!,
n
+ z
r=l
(o(r)L ~)dBr(t)
i
~ t r ) - - ( q r ) , i, j ; 1
.....
(4.3) ~,
a n % i s givenby (4. Z).
In particular,
~Y~
log llx II :
Q(~)
=
1
~
(B~, ~) + ~ - Z:l aij(~) - i,Ej:l
a i j (k) kik j
(4.4)
If w e substitute (4.4) into (4.3), integrate the resulting equation, a n d divide by
t , one obtains
201
log Ilk(t)II - log Ilx(o)U
f
1
J L(BX(s), k(s))+~i~=laij(X(s)
)
O
~
a
i, j=]
t + O
lj
(X(s))Xi(s)~j(s) as (4.5)
n
r=l
( a ( r ) k(s), k(s) dBr(S ) ,
which is the analogue to the formula (3. 17 ) of Section Ill. Thus it follows that the stability properties that w e are attempting to determine will be implied by the limit of the integrals on the right hand side of (4.5) as
t approaches infinity.
A/though, w e w e r e stopped at this point for stochastic systems of the type studied in Section Ill, K_haz'minskii has s h o w n the w a y through the d i l e m m a for systems of the type studied in the present section, chapter 6~. He recognized that the vector
[ see [ 20], and [23],
k = x/ llxll , w h e r e
process to (4. i), itself satisfies an ItAo differential equation.
x is the solution That is,
k
satisfies an equation of the form, dl = Al(k)dt + AZ(X)dB Hence, the sphere
(4.6)
k process is a M a r k o v process defined on the surface of the
IIX II = i.
ditions.
,
Furthermore,
this process is ergodic under certain con-
W e note, however, that the ergodic properties of the
determined by its singularities. determined by the nature of the larities.
In particular,
k - process are
the ergodic properties are
X - process in the neighborhood of its singu-
A singularity of a M a r k o v process is defined as a point at which
the diffusion component vanishes. singularities are the solutions to
For the X-process,
given by (4.6), the
AZ(X ) = 0 .
Since the second integral on the right hand side of (4.5) approaches
202 zero with probability one as t approaches infinity [Z3] it follows f r o m (4.4), (4.5) that
lira t-*~
log llx(t) II
- log t "
l[ x ( o ) Ii
= E ~Q(k)i"
(4.7)
-
with probability one, in the ergodic case. This is exactly the formula w e want since it yields the necessary and sufficient condition for almost sure sample asymptotic stability in terms of
the e x p e c t a t i o n E t Q ( K ) ) . If t h e e x p e c t a t i o n is n e g a t i v e , the d e s i r e d s t a b i l i t y property follows. probability one.
If the expectation is positive, the samples are unstable with W e m u s t evaluate the arithmetic sign of the expectation.
As a sim~e
example,
c onsider the first order linear ItAo equation
(2. II). F o r this case the generator is simply
~=ax
d ~
o"2 +-2--x
Z dE d-~x "
One easily finds f r o m (4.4), Z
Q(X) = (a - S Z - ) kz Recalling that l Z = 1 , w e immediately obtain the well k n o w n conditions >0
E~Q(X)} = ( a - 4 )
,
~2/2
a becomes
9t k'i"
dr: ÷
~,
Ki -~
i
~. ~'~.~j EL 9i.~-' i.,]
I"
K'i"
j
Nj
l" s and E2(t,t)__ = O.
Hence
b
E, ~c) = ¢p. e,cp { - C~-s~I - "/CF,~'~./~pct*- t'~.a,., = , Lz (t
Ez &l
-
_I
exp.(-zc~-s 0).
There is however an alternative input/output des-
cription which, although it has roots going back at least as far in time as do the concepts of transfer function and impulse response, has become particularly prevelant in the last half decade.
This descrlntlon gives the so-called
f!ankel matrix
of ZI defined by: cANB
CAB CAB
H ~-
: cANB
CA2B
...
:
.o° .°.
cAN+IB
.o•
cAN+IB
~.,IIO = [W(i+J-2)(0)] .
cA2N-IB
It turns out that many qualitative input/output properties of Zlare most easily described in terms of H. It is well-known that there exist many minimal realizations ~A,B,C}
of a given
G(s), W(t), or H, but that they all may be recovered from one of them by the transformation group {A,B,C} R nxn. the
~ {SAS-I,sB,CS -I} with S an arbitrary invertible e l ~ e n t of
The dimension of a minimal realization of a given transfer function is called
M~4illan degree. We will consider the following class of systems El:
Definition 2:
Z I is said to be
completely symmetric if m=p and*
The infinite matrix H is said to be
nonnegative definite (denoted by ~ 0) if all its
finite truncations are nonnegative definite, i.e. if N and for all sequences {zl}~'.
H = H' ~ 0.
N ~ z~ CAi+JBzj ! 0 for all i,j=0
256
The following lemma gives a very useful alternative characterization of completely symmetric systems.
Its proof, which is not germane to our purposes, is
an immediate consequence of some known facts in realization theory and is left to the reader. Lemma I: C(Is-A)-IB
I is completely symmetric if and only if its transfer function G(s) admits a realizatio~ {AI,BI,C I} with
=
A I = A~ and B 1 = C~.
Thus Z 1 is completely symmetric if and only if there exists a nonslngular matrix S such that SAS -I = (SAS-I) ' and SB - (CS-I) '
(nxn)
Completely svmmetrlc systems
have the property that the el~envalues of A are all real.
Thls Is in fact also the
case after applying symmetric feedback and it may be shown that E 1 is completely symmetric if G(s) = C'(s) and If A-BKC has real ei~envalues for all K = K'.
Note
also that Z Is completely svmmetrlc if and only if its transfer function admits the k Ri partial fraction expansion G(s) = i=l~ ~ with R i = R i' _> O. If m=p=l then ~lis completely symmetric if and only if the poles and the zeros of the transfer function G(s) are real and interlace, i.e. If ~i,%2 .... 'In are the Doles and If Zl,Z 2 ..... z r are the zeros of G(s), then r = n-l, h i and z i are real, and ~i > Zl > 12 > "'" > Zn-i > ~n"
Thls pole-zero pattern is illustrated in Figure 2. Im
•
n
Figure 2:
Zn-I
1
Re
i
Typical pole~zero pattern of a completely symmetric system,
Completely symmetric systems are a natural generalization of relaxation systems (see Wlllems
[1972]) which are completely svmmetrlc systems which satisfy the
additional stability requirement Re %[A] ~ O.
Thus E 1 is a relaxation system if and !
only if its transfer function admits a realization {AI,BI,C I} wlth A 1 = A I ~ O and B 1 = C I.
There are various other ways of defining a relaxation system.
It may be
The backgrou~materlal of realization theory used here may be found in Brockett [1970], Chapter 2, or Kalman [1969], Section i0.Ii.
257
shown that Z I defines a relaxation system if and only if H = H' ~ 0 end oH = ~ ' ~ O, where @H denotes the shifted Hankel matrix of Z I, i.e., ~ with the first block row (or column) deleted.
Alternatively,
Z1 defines a relaxation system if and only if
its impulse response W(t) = ceAtB is a completely monotonic ~unction on [0,~), i.e. dk W(t) ~ W'(t) and ( - l ) k - - W(t) > 0 for all t > 0 and k = 0,1,2, Relaxation dt k -_ ..- . systems play an important role in physics. They describe the response of various classes of systems
such as R-C and R-L electrical networks, viscoelastic materials
thermal systems, and chemical reactions. We now state the main result of this section. Theorem i:
Assume that Z I i8 completely 8y~ynetric and that K - K' almost surely.
Let ~max A
~{Xmax[A_BKC]}"
max
Then Z is almost surel~. asFmptoticall~ . . . stable if
m a x-- %1 with eaualitv holding if and only if K = 0 almost surely.
Note also that Theorem 2 is easily extended to the
case where K does not possess a density function. 4.
n-i qn-i s +'''+qo Let g(s) = n. n-i s *Pn_l s +"'+Po
and the zeros of g(s).
and let %1,...,%n and Zl,...,Zn_ I denote the poles
Thus %1 > Zl > %2 > "''>
% (K) denote the zeros of p(s)+Kq(s). n
Zn-i >
%n"
Let %max(K) - hi(K) > ...>
From root-locus considerations it is easily seen
260 that z i < li(K) < li(0) < li(-K) < zi_ 1 for K > 0 and i=1,2 .... ,n (where we have put n
ZO
~ and zn ~ -~).
n
Since -
ing uDDer bound for I
max
~ I i + Kgn_ I : - ~ li(K) we thus obtain the followi:l i-I (see Figure 3):
0
qn-i
which requires in particular that l I < 0. Examples:
i.
If K is uniformly distributed between the limits K_ and K+ then Z is
almost surely asymptotically stable if: Z+
Z_
+
g(z+) g(z_)
[Z_ /z+
The limiting behavior of I as K-~ +max
+ !
1 max where a =
0
This ine0uality is easily verified directly
where z+ ~ Xmax(K+) and z_ ~ Xmax(K_). 1 from the graph of f ( o ) g(~) . 2.
d~ -g(~) ->
is given by (see Figure 3):
for K + ~
Zl
for K + -~
i -Kan- i+~
n n-i qn-2 [ li - ~ zi = - - Pn-l" i=l i=l qn-I
Thus as K becomes more and more distributed
at large absolute values we see that almost sure asymptotic stability results if: n ziP+ + (i!iki
n-i -
[0
i~l= zi)P- - qn-I j_~
where P+ __A P(K > 0) and P_ A= P(K < 0).
K p (K)dE
For the uniformly distributed case studied in
Example i with K+ > 0 and K_ < 0 this condition requires n •
3.
n-I
K2
+
qn-1
Consider the equation studied by Infante [1968], p. ii:
. f (t)-8
n + Xc
< 0
;
B 6 = ~n-
},c
261 where 8, £, % > 0. problem.
This equation describes the kinetics of a simple nuclear reactor
It is easily seen that Theorem 2 applied to this case with 1
s+l ~
g(s) " £
s(s+ ~ +
~)
and k(t) - -f(t).
Thus almost sure asymptotic stability results if:
where p(.) denotes the density function of f. 1.2
A Frequency-D~ain
Stabillt X Crlterion
In this section we will derive another criterion for almost sure asymptotic stability of the system
Z.
We first recall the definition of a positive real
function: Deflni,,tion 3: s.
Let H(s) be a matrix of real rational functions of the complex variable
It is said to be positive real if H(s) + H'(~) ~ 0
for all Re s ~ 0, s ~ Doles
of N(s). There exist various equivalent conditions for positive realness. may be found in most books on electrical network synthesis [1957], Chapter i, or Newcomb
[1966]).
Such conditions
(se~ for example, ~uillemin
Positive real functions play a fundamental
role in the theory of passive systems, particularly in the analysis and synthesis of electrical networks.
They have recently also shown to be an essential tool for
obtaining frequency-domain stability criteria for feedback systems.
A time-domaln
condition for positive realness is given in the following lemma, the celebrated
Kalman-Yacubovich-Popov L e ~ a 2:
lemma:
Consider the minimal system: = Fz + Gv
and let ~ be a real number. there exists a solution
; w = Hz,
Then H(l(s-O)iF)-Ic is positive real if and only if
0 = O' > 0
to the relations:
F'0 + QF ~ -200 OC = H' For a proof of L e n a
2 we refer the reader to Willems
[19721.
262 The value of the above lemma in stability analvsls lles in the fact that the quadratic form induced by the matrix 0 yields a very suitable candidate for a LvaDunov function.
It Dlays a crucial role in the following theorem which is the
main result of this section: Theorem 3:
Let m = p.
Then E i8 almost surely a8ymptoticall~ stable i~ there exists
a constant (m~m) matrix A and a real n~mber ~ such that: (i)
(il) and
(iii)
Proof:
A + A'
> 0
;
F(s-O) ~ G(s-o)(l-A(s-o)G(s-~)) -I is positive real; ~{mln[~,~min[(K+K')(A+A')-l]]}
> O
o
We will ass~ne that (I-ACB) is invertlble and that the McMillan degree of
F(s) is n.
The general case may be resolved by a subsequent limiting argument which
is left to the reader• It is easily seen that F(s) is the transfer function of the system: = Az + B(v+AO)
w = Cz ,
;
or
= (A+B(I-ACB)-IACA)z+B(I-ACB)-Iv
; w = Cz .
This system is minimal since the McMillan degree of F(s) is assumed to be n. by condition
Thus
(ll) and Lemma 2 there exists a matrix 0 = O' > 0 such that [A+B(I-ACB)-IAcA]'Q+Q[A+B(I-ACB)-IAcA]
< -2~0
and OB(I-ACB) -I Let S be an invertible
=
C'
(nxn) matrix such that S'S = O and let x I = Sx.
The
equation for x I is given by: ~i
TM
(AI-BIK(t)CI)Xl
where A I = SAS -I, B I * SB, and C I = CS -1. (AI+C~ACIAI)' + (AI+C{ACIA1) ~ -2ol.
Moreover, B I = C~(I-ACIB I) and
Consider now the derivative of V(x I) = X~Xl+
y~Ayl, where Yl = ClXl' along solutions of the above differential equation. calculation using the above relations shows that: V(x I) ~ -2OXlXl-2YlK(t)Y 1
-2~V(Xl)+2y~(oA-K(t))y I .
A simple
263 Let %(t) = Imln[(K(t)+K'(t))(A+A')-i ] and let P be a nonsingular matrix such that P'P =A +A'
Since %(t) = %min[P-l(K(t)+K'(t))(P') -I] it thus follows that YiK(t)Yl
%(t)Yily I for all YI"
Hence
V(xI) ~-2OV(x I) + 2(o-l(t))y~Ay I • We now distinguish two cases:
(1) (li)
and
l(t) > C which imnlies V(x I) _< -2OV(x I) ; %(t) ~ 0 which, since V(x I) ~ Yi A YI' implies:
V(x 1) < -2SV(x 1) + 2 ( O - t ( t ) ) V ( x 1) = - 2 t ( t ) V ( x 1) • Hence
V(xI) ! - 2 m i n [ C , ~ ( t ) ] V ( x 1) and
V ( X l ( t ) ) _< V ( X l ( t o ) ) e x p ( -
2
f
t
t0
min[l,c(t)]dt)
.
By the ergodic hypothesis and condition (lii) this indeed imnlies that lim V(Xl(t))=O almost surely.
Thus lira Xl(t) = S l i m x(t) = 0 almost surely, which proves the t-~m t-~o
theorem. • Notes:
5.
If K + K' > £I > 0 almost surely and if G(s) is positive real then
Theorem 2 predicts almost sure asymptotic stability by considerin~ the limit o ÷ 0 and A ÷ 0.
In this sense Theorem 2 is thus a generalization of the circle criterion.
The advantage of the theorem is that it allows the gain K(t) to become negative provided however this is compensated by K(t) being sufficiently positive at some other time. One of the disadvantages of Theorem 3 is the inherent difficulty in verifying the average value condition from the distribution of K since %mln{[(K+K')(A+A')-I]} is a very nonlinear function of K.
In the scalar case however one may resolve the
various conditions in Theorem 3 much further•
Thus we arrive at the following
more explicit criterion for systems with a single stochastic parameter: qn_iSn-l+.•.+~ Theorem 4: Assu~e that m = ~ ~ 1 and let g(s) = C(Is-A)-IB = o sn+nn_Isn-l+...+po
denote the transfer function of ~I • Then Z is almost surely asymptotically stable if there exists a real constant 6 such that
(i) 8{min[B,K] } > C, ;
264
(ii) and
the pole8 of g(s)
(iii)
lie
i n Re s < - q n _ l 8 ;
the locus of G(J~-qn_lS) ,
-~ < ~ < m
does not encircle
or
intersect the closed disc centered on the negative real axis of i the complex plane and passing through the origin and the point - ~ . Proof:
By Theorem 3 it suffices to show that there exists a constant ~ > 0 such
that F(s-~) = g(s-~)(l-l(s-~)g(s-s)) -I is positive real and ~{min[u%,k]} > ~. that this implies u > 0. 1
g(s-O)
Now F(s-~) is Dositive real if and only if F-l(s-@)
Note -
r(s-~)
~(s-~) is ~ositive real.
Since F-l(s-O) = (___~i _ ~)(s-O) + q(s-~) with qn-i r(s) a polynomial of degree at most (n-l) it follows that X < ~ and that F-l(s-o) --
On_ I
i will be positive real for some % if and onlv if it is nositive real for ~ = ~ , an- I which is thus the optimal value of X to consider. The condition ~n-i > 0 follows from the frequency domain condition (ill) as a result of the behavior of g(Jm-o) for + ~.
Pick now s = 8qn_l.
In order to complete the proof of the theorem it suffices to show that F-I(s-G) = 1
1
g(S-qn-IB)
qn-I
hess
s + 8
is ~ositive real.
By one of the test of positive real-
this can be achieved by proving that Re F-l(s-O)]s=j~
~ 0 and (since F-l(s-o)
has no more zeros than Dole~ that the roots of q(s-o) lie in Re s < ~.
The real part
condition comes down to asking g(s-O)]s=j~ to have the non-intersection property stated in condition
(iii).
kq(s-O) lie in Re s < ~
By the non-encirclement
By letting k + m this implies that the roots of
for k > B.
q(s-O) lle indeed i n Re s < O.
condition the roots of p(s-o)+
By the non-lntersectlon property g(J0a-~) ~ 0 for
-co < ~ < ~ and we conclude that the roots of q(s-o) indeed lie in Re s < @ as desired. • Notes: 6.
It may be shown that conditions
(if) and (iii) of Theorem 4 will be veri-
fled for 8 ~ 81 if they are verified for 81 .
Thus the optimal 8 to consider is the
smallest number which satisfies condition (i) of the theorem. 7.
If K has density function p(K) then condition
h(S) A S
p(K)dK +
(i) of Theorem 4 requires that:
Kp(K)dK > 0 --co
dh(B)
N o w - - ~ - - - - ~ 0 , h ( 0 ) ~ 0 and h ( ~ ) = ~{K}. if
and o n l y i f
B > B*.
~{K} > 0, and i f
Thus Theorem 4 w i l l
Thus t h e r e e x i s t s
so, then there exists
predict
a B
a fl s u c h t h a t
such t h a t
almost sure asymptotic stability
h(B) > 0
h(8) > 0 f o r of I
265
if ~[K] > 0, if the poles of g(s) lie in Re s < qn_l ~* and if g(J~-qn_16*) satisfies the frequency domain condition of Theorem 4.
This procedure lends itself very nicely
to the graphical analysis illustrated in Figure 4.
h(fl)
Re
--- Im
/
-I/
Figure 4: E_xamples:
4.
qn-i s*)
I~lustrating the application of Theorem 4.
Assume that K is uniformly distributed between K
and K+ with K_ ~
and K+ + K_ ~ O.
Then B* = K+ -~IK~-K~ . Expressed in terms of the spread A K = K + - K K+-K_ (~and the mean M = 2 ....... this yields ~* = AK _ ¢~ )2 which in the range of interest AK * > M > 0 shows that ~ increases with AK for fixed M. 2 --
-
-
This again indicates the
destabilizing effect due to the uncertainty in K. 5.
Let Z 1 be a completely symmetric system as defined in Section i.i.
Then con-
ditions (ii) and (iii) of Theorem 4 will be satisfied as long as qn_l ~ < -%1 with %1 11 the largest pole of g(s). The stability condition then becomes ~{min[- - - ,K]} > 0 qn-I which is similar t~ but more conservative tha~ the condition obtained in Note 4. Thus Theorem 2 which only applies to completely symmetric systems gives a sharper stability estimate than Theorem 4 which applies to general systems. 2.
ANALYSIS OF THE MEAN AND THE COVARIANCE EOUATIONS
This last section of the paner is concerned with the stability analysis of the mean and the covariance of the state of ~ where K(t) is assumed to be a white stochastic process.
For simpliclt~ we will consider only the case in which the
266 process K(t) is scalar valued, but we will treat the non-stationary case.
If we
denote the mean of K(t) by k(t) and the variance by ~2(t) then Z is described by the stochastic differential equation: Z' : dx = (A-k(t)bc)x dt + q(t)bcx dR , where A ¢ R nxn, b ¢ R nxl , c ~ R Ixn, and ~ denotes a Wiener process with zero mean and unit covariance.
This stochastic differential equation is to be interpreted in
the sense of Ito and we will take it as the startin~ point of our analysis. It is well-known that if k(t) and q(t) are sufficiently smooth (e.g., locally integrable) then for all given X(to) there exists a unioue solution to Z' for t > t . --
o
Let ~(t) ~ ~{x(t)},
r(t) ~ ~{x(t)x'(t)}, and R(t) ~ ~{(x(t)-~(t))
(x(t)-~(t))'} denote respectively the mean, the second moment matrix, and the
covariance matrix of x(t).
These are governed by the equations:
= (A-~(t)bc)~
;
= (A-~(t)bc) r+r (A-~(t)bc) '+q 2(t)bcrc 'b ' and
;
R(t) = r(t)-~(t)~' (t),
with initial conditions ~(t o) = x(t o) and r(t o) = X(to)X'(tO)We will be concerned with the asymptotic properties of these variables.
The
relevant stochastic stability concepts are now defined: Definition 4:
square,
E' is said to be asymptotically stable in the mean, in the mean
or in the covariance if, respectively, lim ~(t) = 0,
lim r(t) = 0, or
llm R(t) = 0 for all given initial conditions X(to). t-~o It is easily seen from the relations r(t) ~ R(t)+~(t)~'(t) and R(t) = R'(t) ~ 0 that mean square asymptotic stability implies stability in the mean and in the covariance.
The stability of the mean is a standard deterministic stability
problem for which many criteria have been derived.
These criteria involve the trans-
fer function g(s) = c(Is-A)'~ and properties of k(t) as, for example, its bounds (e.g. in the circle criterion: see Brockett [1970], Section 35), bounds on its derlvative~ or its periodicity.
The stability of the differential equation which
expresses the evolution of the second moment matrix P(t) is much more intricate
267 to analyze and we will show how criteria like the multlvariable circle criterion may be used.
If q2(t) = 0 then its stability is equivalent to the stability of
the mean equation, whereas if q2(t) # 0 then more stringent conditions will have to be imposed. 2.1
Multilinear_System Theory It is easy to see that if x I and x 2 are vectors which satisfy the linear
equations: ;
Xl = Al(t)Xl and
;
x2 = A2(t)x2
XleR x2eR
nI n2
, ,
then the product XlX 2' satisfies also a linear equation, namely:
_dddtXlX2 " Al(t)xlx2 + XlX2A2(t)
"
By taking x I = x 2 we see that if x satisfies a linear equation, then so does XX
I .
This idea generalizes from auadratic forms to homogeneous p-th degree forms. These facts have been known at least since Lyapunov's thesis, but they have to the present time been used very little in system theory.
They may for example be
exploited in the minimization of homogeneous performance measures of degree p > 2 for linear dynamical systems. The above ideas may be used in setting up transfer functions for a class of bilinear systems. ~. M
We will make some use of the Kronecker product denoted here by
Thus the Kronecker product of M e R nxm and R £ R Dx~ is the element ~
R E R np:~nq defined by: mllR
M~ R~ ~2-~f-_ m22R mplR
mlqR
ml2R °.°
mp2R
m2qR
m
Pq
R
The main use of this notation is that if an (nxn) matrix 0 is written in lexo2 graphic notation as the n -vector
268
0v = c°l(qll' q12 . . . . . then (MQ) v ~ (I ~
qln . . . . .
qnl' qn2 . . . . ' qnn )
M)O v.
Consider now the following lemma: Lemma 3:
Let {A,b,c} be a minimal realization of the transfer funotion g(s) =
c(Is-A)-ib.
Then the differential equation: = AQ + 0A + bv' + vb'
defines a minimal realization on the
;
w = cO ,
dimensional space of symmetric (nxn)
n(n+l)
2
matrices of the transfer function: g[2l(s) = ( c ~ I + I ~ c ) ( I s - I ~ A - A ~
l)-l(b~
We will not give a detailed proof of this lemma.
l+l~b)
.
The proof exploits the fact
that the above matrix eauation describes the 5ilinear system d - - xx' = Axx' + xx'A' + bux' + xu'b' dt
;
yx t = CXX' where ~ = Ax + bu; y = cx. The dynamical system identified in the statement of Lemma 3 plays an important role in the analysis of the covarlance e~uation under cons~deratlon. from this lemma that controllability
and observabilltv will be preserved.
poles of g[2](s) are given by {%I(A)+lj(A)},
i,J=l ..... n.
convenient general formula for deriving g[2](s) from g(s). however, Example:
The
There appears to be no In a specific case
it is a relatively straightforward matter to calculate g[2](s). 6.
Let [A,b,c] be the standard controllable representation
[1970], n. 106) of Q(s) = 1 s 2+as+b
~[2] (s) = -
2.2
We know
Then "
i s3+3as2+(2a2+4b)s+4ab
s(s+2a)
The Circle Criterion for the Covariance Equation We now return to the covariance eouation: = (A-k(t)bc)F + F(A-k(t)bc)' + q2(t)bcrc'b'
?s
(see Brockett
269 which we model as the feedback system: E{ : Q = AO + QA'+bv'+vb'+bwb'
z~ : v=-~(t)y,
;
y - cO, z = c O c ' ,
w=~2(t)~
It follows from Lemma 3 that l I' is comnletelv controllable and completely observable. Let ~(s) =A ICll (s)
Gl2(S)]
LG21(s)
~22(s)j
,
where
y(s) = Gll(S)V(S ) + Cl2(S)W(S)
and
z(s) = G21(s)v(s) + G22(s)w(s) !
denote the transfer function of Z I.
G(s) = |
c~
c
,
It is easily calculated that G(s) is given by:
(Is-A@
I-I@A)
-1
[b ~ I + I ~ b
I b $ b]
L
Thus t h e s t a b i l i t y deterministic
of t h e c o v a r i a n c e
feedback system with
equation
is equivalent
to the stability
(n+l) feedback loons, with transfer
nj
of a
function
G(s) i n t h e f o r w a r d l o o o and R a i n m a t r i x k(t)I F(t) = 0
_~2 (t
in the feedback loop. The multivariable
circle criterion and its various generalizations
immediately applicable to this situation.
is thus
We will illustrate this only in the
simplest case.
Let If" If denote some norm on R n+l and let matrix norms be
induced norms.
The small loop gain theorem due to Zames [1966] thus leads to:
Theorem 5:
Assume that Re l[A] < 0.
Then ~' is asymptotically stable in the mean
square if:
( sup tl~(J~)ll)( -eo 0.
The qualitative similarity between
these conditions and those for harmonic excitation is worth noting.
Stability with probability one may be investigated by the Liapunov method as discussed in Kushner (1967).
For the system defined by Eq.(7), the
scalar function
VCa
n
o : 1
satisfies the requirements of a Liapunov function. stochastic Liapunov functions, i f ~ { V ( a ) } where~is
According to the theorem on
~ 0 then a ÷ [a:~{V(a)} = 0] w.p.l.,
the differential generator of the process a(t).
the necessary differentiation, we obtain
After performing
298 n n 2 [V(a)] = -¢ ~ ~ Aija. i=l j=l ) where the coefficients A.. are as defined by (8). 19 for asymptotic stability w.p.l are ~ Aij > O,
Hence, sufficient conditions
j = 1,2,...,n,
that is, k~. n k.. k.. k.. dil'- -!~! S ( 2 w i ) - 2 3"=i ~ 4wi 13( wI)i S~lj + ~ JISij) > O, i = 1 . .2. . 1 j~i
.,n.
(13)
For a single degree of freedom system, the result reduces to the condition for stability in the mean.
For w
O
= Wr ± Us, (13) leads to the pair of conditions
k k k rs ( rs + Sr)s(~r±~s) drr - 4~ " ~ - w
> 0
k k k sr sr rS)s(~ri~s) dss " 4-~- (~--- ± w
> 0
r
S
r
S
(i0)
s
(14)
r
Gzroscopic Szstems The class of systems considered has equations of motion of the form 2 2 ql + gSql - 29~2 + (~I -~ )ql + f(t)ql = 0
(1S) q2 + ~Bct2 + 2 ~ 1
2 2 + (w2-fi)q2
+ f(t)q2
= 0
where f(t) is a stationary stochastic process with zero mean.
These equations
represent the transverse flexural motion of a light elastic rotating shaft of unequal flexural rigidities carrying a single rigid mass at its mid-length and subjected to a randomly varying axial thrust, see for e.g. Dimentberg
(1961).
The coordinates ql' q2 denote the transverse displacement of the mass measured with respect to rotating principal axes. ~ is the angular velocity of the shaft and ~i,~2,
(~I < ~2 ) are the natural frequencies of transverse vibration;
represents the coefficient of internal damping, Fig.4.
8
299
{2
q2 ~
q
l
~l .... S'rJJi
Fig. 4
The eigenfrequencies ml, ~2 of the system (15) for ¢ = O, f(t) = 0 are given by the roots of the equation
. 2 2
2-2
(Wl+~2+2fl)~
+
(@a2)(@~2)
= O.
For real ~, fl should be outside the range ~i ~ ~
~ ~2
Considering only these values of ~ (for which the unloaded shaft is stable), Eqs. (IS) may be converted to standard form by the transformation
ql = al sin 91 + a 2 sin 9 2 q2 = ~lal cos ~I + ~2a2 cos 9 2 ql = alml cos 91 + a2~ 2 cos 9 2
q2
=
-elalJ1 sin ~1 - ~2a2~2 sin ~2'
where
91 = ~it+~1,
%
= ~2t+%,
--2 . 2 ~2, ~o1- (o~1-~)
--2
o~2 =
. 2 ~2. o~2- (COl-~ J 2~2
2~ 1
2f,~2 =
--2 . 2 2. a)2- (~2-~)
300 Proceeding now exactly as for non-gyroscopic systems and assuming the spectral density of f(t) to be eS(co), averaged equations may be set up for the squares of the amplitudes al, a 2 from which the following conditions for stability in the mean may be obtained: For co = O
2~.1
(i = 1,2):
2 2.2 col-co2J B > 16f~2~2 S(2~ i)
if ~ < col
i
2
2 2
(col-co2) B >
--2 --2 2 --2 S(2coi) 2 (col-co2+4~ )col
if ~ > co2"
For coo = col+co2: --2 2 2 > (colco2+Wl-~ ) 2 2 -4 (col-~) colco2 . • --
>
S(COI+CO2) if ~ < col
2 ~2.2~4
8 [colco2+col-~ ) ~ (~2_w~) [16~4_ .--2 --2.2~-(col-co2) Icolco2
S(COI+CO2) if ~ > co2
For coo = Wl-co2: . • --
>
2 ~2.2~4
8 Lcolco2-col+~ ) ~ (co~_~2) [ 1 6 ~ 4 _
--2 2 2 > (colco2-col+~) 2 2 -4 (~ -col)colco2
.--2 --2.2 ~-- -Lcol-co2J Jcolco2
S(COl-CO2)
if ~ < col
S(col-co2) if ~ > co2"
Sufficient conditions for asymptotic stability w.p.l may also be obtained as 22 before using the Liapunov function V(a) = al+a 2 but are not presented here. Again, the qualitative similarity between the stochastic and the deterministic results may be noted.
However, while in the case of sinusoidal excitation,
instability near coo = col+co2 is present only when ~ < col and that near coo=~i-$2 only when ~>co2' see Mettler
(1967), in the stochastic case there is
no such distinction; both forms of instability can occur in both ranges of values of ~.
301
CONCLUSIONS This survey has dealt with the stability of coupled linear oscillatory systems under stochastic excitation of small intensity.
The systems considered
are typically encountered in the study of the dynamic stability of elastic structural and mechanical systems subjected to randomly fluctuating loads. Certain similarities between the stability conditions for the case of stochastic excitation and those for deterministic sinusoidal excitation have been emphasized. The results presented here are obtained using the method of averaging which is valid to the first approximation only and hence correspond only to instabilities of the first order.
To obtain conditions corresponding to higher
order instabilities, higher approximations must be made in the solution of the governing equations in standard form.
It may then be possible to investigate
the influence, if any, of the values of the excitation spectrum at higher multiples and fractions of the natural frequencies and combination frequencies. Evidence that the spectrum at these frequencies may influence the stability is provided by some recent results of Bolotin (1971) for a single degree of freedom system excited parametrically by filtered Gaussian white noise.
For such a
system, differential equations governing the response moments were derived by Bogdanoff and Kozin (1962). Using a truncated linearised form of these equations, Bolotin has obtained approximate conditions for second moment stability and shown them graphically on a plot similar to the Strutt diagram for the Mathieu equation for various values of the damping parameter.
The stability boundaries reveal a
dip when the centre frequency of the excitation is close to the system natural frequency,
in addition to the one near twice the natural frequency.
While
these results do indicate a qualitative trend, the validity of the approximations made in obtaining them seem somewhat drastic and questionable.
It appears that
Wedig (1972) has also obtained similar results in his paper at this Symposium. Non-linear oscillatory systems under harmonic parametric excitation possess several complicated and interesting instability phenomena, Mettler (1965, 1967).
A corresponding general investigation of such systems under stochastic
excitation appears to be lacking. ACKNOWLEDGMENTS The research for this paper was supported (in part) by the National Research Council of Canada under Grant No. A-1815.
The assistance of Mr. D.S.F. Tam
and Mr. S.C. Oen in the preparation of the paper is gratefully acknowledged.
302
REFERENCES Ariaratnam, S.T. and Graefe, P.W.U., 1965, Int. J. Control, ~, 161. Ariaratnam, S.T., 1965, Proc. Int. Conf. on Dynamic Stability of Structures, Ed: G. Herrmann, Pergamon, 1966, 255. Ariaratnam, S.T., 1967, Proc. Can. Cong. Appl. Mech., Vol. III, 163. Ariaratnam, S.T., 1969, Proc. IUTAM Symp. on Instability of Continuous Systems, Ed: H. Leipholz, Springer-Verlag, 78. Bogdanoff, J.L. and Kozin, F., 1962, J. Acous. Soc. Am., 34, I063. Bolotin, V.V., 1964, The Dynamic Stability of Elastic Systems , Holden-Day, Inc. Bolotin, V.V., 1971, Study No. 6, Stability, Ed: H. Leipholz, Solid Mechanics Division, University of Waterloo, 385. Caughey, T.K. and Dienes, J.K., 1962, J. Math. Phys., 40-41, 300. Caughey, T.K. and Gray, A.H. Jr.,1965, J. Appl. Mech., 32, 365. Chelpanov, Dimentberg,
I.B., 1962, P ~ , 26, 762, English Translation,
1145.
F., 1961, Flexural Vibration of Rotating Shafts, Butterworths.
Graefe, P.W.U., 1966, Ing. Arch., 3_SS, 202. Gray, A.H. Jr.~ 1967, J. Appl. Mech., 34, 1017. Infante, E.T., 1968, J. ?ppl. Mech., 35. Kozin, F., 1963, J. Math. Phys., 42, 59. Kushner, H.J., 1967, Stochastic Stability and Control, Academic Press. Man, F.T., 1970, J. Appl. Mech., 37, 541. Mettler, E., 1965, Proc. Int. Conf. on Dynamic Stability of Structures, Ed: G. Herrmann, Pergamon, 1966, 169. Mettler, E., 1967, Proc. Fourth Conf. on Nonlinear Oscillations, Academia, 1968; 51.
Prague,
Samuels, J.C., 1960, J. Acoust. Soc. Am., 32, 594. Samuels, J.C., 1961, J. Acoust. Soc. Am., 33, 1782. Stratonovich, R.L., 1963, Topics in the Theory of Random Noise, Vol. I, Gordon and Breach. Stratonovich, R.L., and Romanovskii, Yu.M., 1958, Nauchnye doklady vysshei shkoly, Fiziko-mat. nauk., 3, 221, Reprinted in Non-Linear Transformations of Random Processes, Ed: Kuznetsov, P.I., R.L. Stratonovich and V.I. Tikhonov, Pergamon, 1965, 332. Wedig, W., 1972, Proc. IUTAM Symposium on Stability of Stochastic Dynamical Systems, Springer-Verlag. Weidenhammer,
F., 1964, Ing. Arch., 33, 404.
WAVES IN A ROTATING STRATIFIED FLUID WITH LATERALLY VARYING RANDOM INHOMOG~EITIES By Lawrence A. Mysak Department
of Mathematics and Institute of Oceanography University of British Columbia Vancouver, B.C.,Canada ABSTRACT
We discuss the propagation and stability of internal waves in a rotating stratified unbounded fluid with randomly varying buoyancy frequency, N • The first order smoothing approximation is used to derive the mean wave dispersion relation when ~ is of the f o r m N Z = No2(f÷~j~{) , where N o = constant~ O ~ £ z ~ and~ is a centered stationary random function of the horizontal direction X This form f o r . ~ represents a stochastic model of the lateral variations in the temperature and salinity microstructure in the ocean. From the complex dispersion relation, expressions are obtained for the phase speed change and spatial growth rate ( ~ 2); in particular, attention is focused on the asymptotic behaviour of these expressions for short and long correlation lengths ( ~ 3). i. INTRODUCTION It is now well established from observations that on the gross temperature and salinity depth profiles in the ocean, there are superimposed small step-like variations
(e.g., see Gregg and Cox, 1972, and their references).
That is, below
the surface mixed layer it is apparent that in many regions of the ocean, the density stratification consists of thin, sharply-defined homogeneous thicknesses varying from a few centimeters to several meters.
layers with
Consequently,
observed depth profiles of the Brunt-V~is~l~ or buoyancy frequency, which involves the vertical density gradient and which is of fundamental
importance in the theory
of internal waves, exhibit highly irregular fluctuations about their mean values (e.g., see Gregg and Cox, 1972; McGorman and Mysak,
1972,hereafter referrred to
as I). Although it is not yet fully understood how this so-called or 'microstructure'
arises in practice,
'fine-structure'
it is believed that both convective and
diffusive processes are involved in its formation (Gregg and Cox, 1972; J.S. Turner and C.F. Chert, private communication).
O~very
plausible mechanism is the so called
salt finger instability which is caused by the different rates of diffusivity of heat and salt (Turner and Stommel~1964).
If a warm salty body of water overlies
a cold, fresh and initially heavier body of water, then, because of the rapid diffusion of heat downward, the system becomes gravitationally unstable and long narrow convection cells (salt fingers) are formed. Under certain conditions the salt fingers become self limiting and a fairly regular step-like structure in the temperature and s~linity profiles is formed.
Fine-structure
suggestive of salt
fingering has been reported, for example, west of the Strait of Gibraltar in the depth range 1 2 0 0 - 1 5 0 0 ~ ~ which is just below the warm salty core of the Mediterranean
304
water (Tait and Howe, 1968).
The inverse situation, in which warm salty water
overlies lighter, cold fresh water, also appears to lead to layering.
In this case
sharp interfaces separating layers in turbulent convective motion are formed.
Such
layering has been observed recently in the Arctic ocean in the depth range 200500 ~
, which is just above the warm saline core of the Atlantic Intermediate
Water (Neshyba et aI)1971).
In an attempt to understand this inverse situation,
Drs. Chen and Turner have recently shown in the laboratory that horizontal gradients in the diffusing components can lead to a fairly well defined (though changeable) layer structure. A particularly interesting feature of the Chen~Turner experiments was the existence of horizontal variations in the microstructure.
This is to say, the
thickness of a particular layer varied considerably in the horizontal direction and often two layers merged into one.
Similar remarks also apply to the horizontal
behaviour of microstructure observations recently made in the Pacific off Vancouver Island (Nasmyth, 1973).
Thus, in view of the considerable observational,
experimental and theoretical interest in internal waves (see Turner, 1973), it appears worthwhile at this time to determine the effect of the horizontal variability in the microstructure on the propagation and stability of internal waves.
Below,
we shall discuss this problem via a study of the internal wave dispersion relation that is implied by a stochastic wave equation.
The starting point of our analysis
is the assumption that in the internal wave equation, the buoyancy frequency, ~ has the form
N where
% =
-
constant, O < f Z ~
----
d Z/ 4
(2.3)
0 and~Z={l~Iz-~z-{Z))O. The
the frequency passbands
~z~
O ' ~ v ~ / ~ (PBI) and
PBI is the usual oceanic situation.
Since#o~
kinetic energy is proportional to~o~Z
latter inequality holds for
~/o~
~Z
e - d ~ for N = ~
(PBII);
and the wave the factor e #~/z
( ~ = wave amplitude),
in (2.2) ensures that the kinetic energy is conserved. we do not set ~ = ~
~Z L
In the subsequent analysis
(the Boussinesq approlimation), so that (2.2), which we shall
call the deterministic solution, implies amplification for ~ > O propagating wave phase) and attenuation for ~ phase). In terms of the polar coordinates
(upward
(downward propagating wave
Ko)~ where(~=~o(dos~e))eq.(2.3)
takes the form
KO =
d / ~ 7 ( ~,.)Z g,o_gq,.-~ -- .f;~ 7--.,O)I/Z
The deterministic phase velocity thus has magnitude
C =ff/~ o
which is along the direction of energy propagation,
is given by
=
2-
(2.4)
and the group velocity,
q"
(2.5)
It is easy to show that the group velocity is always orthogonal to the local tangent of the O-= constant curves as given by (2.3); for relatively short waves (K~J/2~),
~C-o and~o D are essentially orthogonal.
306 For ~ 2
of the form (I.i), eq. (2.1) becomes a stochastic differential
equation for ~ ) ~ J
, where
¢ ~X) ~ j ~ ) = ~ X ~ ) e -,/- /A'~) ~.J = 0
:~
:
where
-i~-~ ~ - ~ O ) :
al
-
-
(2.6)
a e,
l
~ p C < ~ ( e ~- :2")-' "-
M
.[
(2.7)
are r e s p e c t i v e l y d e t e r m i n i s t i c and s t o c h a s t i c d i f f e r e n t i a l o p e r a t o r s , a n d < j ~ } = O l where < . >
denotes ensemble average.
According to Keller and Veronis (1969), (2.6) admits plane wave solutions of the form ( ~ > - - ~ ÷ ~ is correct to O ( f 2 )
provided that the following dispersion relation, which , is satisfied:
(2.8) where
Do
-~ e - ~
e ~
q, = : . : k x r . e ~ ) and ~ - 1
is the i n t e g r a l operator oO
de:
(z.~)
--o0 in which ~
is the following Green's function:
~ :×,~)
- g 1
310
From the above tables we thus draw the following geophysically relevant conclusion:
For PBI internal waves and for short correlation lengths (which is
the case in practice f o r ~ = / ~ )
(see I) and probably also the case forl/~
=~{X)j
although this point needs to be confirmed), vertical and lateral random fluctuations in
~
have the same average effect on the propagation and stability of the
waves.
The random fluctuations make the waves travel slower and enhance the
deterministic growth/decay behaviour. Acknowledsements This paper was partially supported by the National Research Council of Canada and was written while the author was a Senior Visitor during 1971-1972 in the Department
of Applied Mathematics and Theoretical Physics, University of
Cambridge.
REFERENCES Frisch, U. Probabilistic Methods in Applied Mathematics,
Vol i. (ed. A.T. Bharucha-
Reid), Academic Press, New York (1968) Gregg, M. C.jand Cox, C. S. Deep-Sea Research, 19, 355-376 (1972) Keller, J. B.~and Veronis, G. J" Geophys.
Research, 74, 1941-1951 (1969)
McGorman, R. E.~and Mysak, L. A. Geophys. Fluid Dynamics, Nasmyth, P. In proceedings University,
in press (1972)
of 'Fourth Colloquium on Ocean Hydrodynamics',
Liege
in press (1973)
Neshyba, S., Neal, V. T.~and Denner, W. J. Geophys. Research, 76, 8107-8120 (1971) Tait, R. I.~ and Howe, M. R. Deep-Sea Research, 15, 275-280 (1968) Turner, J. S. Buoyancy Effects in Fluids, Cambridge University Press, in press (1973) Turner, J. S.jand Stommel, H. Proc. Nat. Acad. Sci. USA, 52, 49-53 (1964)
THE STABILITY
OF A SATELLITE
BY THE FLUCTUATIONS
WITH PARAMETRIC
OF THE GEOMAGNETIC
EXCITATION FIELD
Peter S. Sagirow University
The motion
of a satellite
field by a satellite domains ferent
1. Satellite Satellites
stabilized
fixed magnetic
corresponding definitions
of Stuttgart,
to different
with magnetic
field in two different
ways
interaction duced
H
a damping
to
moment
with respect
depending
above are described
(1) ~
properties
k
and
As these
velocity
by the
to
in-
currents H~ .
are
It is
of the satellite
in a circular
equatorial
caused by the moments
orbit
mentioned
[6]
~
is the moment I
is caused
field.
is a positive
of the shell,
in the pitch
B
rod inter-
proportional
and the eddy currents
+ kH2Y ' + HI~inY
is the yaw angle,
magnetic moment
is proportional
of the satellite
B~"
and to dif-
by the geomagnetic
by the field.
on the angular
by equation
models
A second moment
of the earth
case of a satellite
the yaw oscillations
to the yaw axis,
a restoring
the second moment
In the most simple
Here,
causing
to the geomagnetic
The stability
are compared.
are influenced
of the satellite
H
integral
[1]. The stabilizing
the field
in the shell
proportional
linearization
of the field.
between
to the geomagnetic
Field
elements
acts with the field directly to the intensity
with respect
rod is considered.
of the stochastic
in the Geomasnetic
Germany
constant
of inertia with respect
depending
is the moment
axis of the satellite.
on the magnetic
of the magnetic
The intensity
H
rod fixed
can be described
by
(2)
H = Ho ( t * ~ )
where
H o = const
field.
Transforming
( 3)
where time.
~
denotes
the fluctuations
the time and introducing
of the geomagnetic
the constant
c = k ( FIo)'/~(B I ) v z
we obtain
(4)
and
the normed equation
Y +cC1+~)~ 9 +(I+~) si~ = 0 the dot denotes
the derivative
with respect
to the transformed
312 2. Linearization
of the Nonlinear Noise Term
Suppose the fluctuations The nonlinear term
ferent ways. Assuming v
by
~
q(t)
to be Gaussian with
v : (I+~)2
= 1+2~ .
~
: I+2q +~2
E~ : O, E ~ ~ : ~ .
can be linearized in dif-
to be small we can neglect
~z
and replace
A more sophisticated method is the statistical
earization proposed by Kazakov and determine the constants
Ev=Ev
[3]. We replace
~i • ~i
v
by
lin-
v i : ~i+ ~i~
either by the requirements
E(v - Evil= E(v-Ev)
,
or by the requirement I
E ( v - v a )~
In the first case we obtain case ~
~3 : I + ~ Z , ~3
= ( ~ z + ~3)/2
The coefficients
•
2 Thus,
"-
rain.
~, : I+~'
~z: K 4 + 2 ~ i
A further possibility
and in the second
[3] is
~
1+~ z
four linearized models are obtained:
~i , ~i
are given below:
mode I
'~i
~i
1
t
2
2
4+~ z
3
I + 6 '2
2
4
I ÷g2
4 • ¢FVFiT
3. Introduction of the White Noise The fluctuations ing frequencies
q (t)
of the geomagnetic
and amplitudes
field have rapidly
[2] and can be approximated
chang-
by white
noise
where
w(t)
is a normed Wiener process with
Then• equation
(6)
Ew(t)
(5) reads as
~ + c(~i+~i6w)6~ +(1+6~vlsinW " 0
= O, Ew2(t)
= t
3t3 or written correctly as a stochastic
(7 ) where
=
dx
( "
- sin x~ - c=i x=
x : (x~,x 2)'r
)
: (q~, ~)T
For small oscillations
the eqs.
system
df
+
and
T
(
°
)
-~'sinx~-~ex
z dw(~')
denotes transposition.
(6) and (7) can be linearized:
(8) or
(9)
=
4. Stability
(° -1
-c=i
° 1
×dw~
-~1~i¢
in the Mean Square
The mean square stability Khasminski
-0"
can be checked by the criterion of Nevel'son-
[5]. Interpreting
obtain the stability (10)
eq.
(8) in the sense of It8 we directly
condition 2c~i
>
c 2 ~ i = ~ z + 6,z
To the linearized models derived above correspond
the three explicit
stability conditions
~
0
I shows the stability domains in
the (c, ~z)_plane. Interpreting eq.
(8) in the sense of Stratonovich
to derive the equivalent
(11)
It8 equation.
[7], first we have
This equation is
~*('r..,(xi-~6'~'CZpi~'*gc~iw ) ~ + ('J-~i'62"Cl~i+~v)~
and leads to the Nevel'son-Khasminski
The stability domains
stability
condition
corresponding to the different
models are shown in Fig. 2.
= 0
linearization
314
3 0.5
2 1
0
m
0
1
2
3
Fig. I: Domains of stability in mean square for eq. interpreted in the sense of It8
4-
c
(8)
0 l,
0,5 3 4 2
.....
I
1
....
I
i
2
3
Fig. 2: Domains of stability in mean square for eq. interpreted in the sense of Stratonovich
,,,,
,i
4
(8)
c
315
5. Stabi!it[.in
Probability
The necessary and sufficient
condition for stability in probability
can be obtained introducing the polar coordinates = tan-4(x4/xz) and
~
and considering the processes
= inr(t)
r = (x~ +x ~ )4/2 , ~
= ~(t)
on
r : 1
This method is proposed in [4] and leads to a sta-
bility condition in integral ficulties evaluating
form. However,
the occuring integrals
due to the numerical difthis condition turns out
to be non effective. Sufficient
stability conditions
techniques.
can be obtained by stochastic Liapunov
As Liapunov function for the linear system (9) can serve
the quadratic
form
v(x) = Ax~ + Bx4x , + Cx~
(13) The stability
condition depends on the constants
A,B,C
and is de-
rived applying the L-operator to (15) and requiring the positive definiteness
of the functions
-Lv(x)
(~4)
v(x)
and
= ( 5 - C G i ) x~ + (Bc~ i - 2A+2C-2C~2c~i ) x , x z + * (2c~;
- cc~i~
However. the best possible choice of
~ - 5 ) x~
A.B.C
leads to the old con-
d i t i o n (I0). Considering the stability of the nonlinear system (7) and using the Liapunov function
(15) with
A-c~i/2 the stability
B=t
in probability
once more condition stability
,
,
C=O+cZ~)/2c=i,
D=~C(1-~2c~i)
can be shown to hold for
~
satisfying
(I0). This result follows also directly from the
of the linearized system.
6. Conclusions All results obtained agree qualitatively with the general insight in the considered motion. c ~&
and
c~
,
All stability bounds
e.g. for
k40
~2(c)
tend to zero for
(vanishing damping)
and for
I ~O
(vanishing magnetic moment of the stabilizing rod). The maximum of 62(c)
which must be expected between the mentioned minima occurs for
all models in the relatively small intervall and
k
the value
c = 0,4
0,3lllYJO aKTyaJI:&HOCT:b. qaCTO 3a;n,a~y Ol!TVlMll!3a~ll!lll MHOrOMepHOW CTOXaCTlll~SCKO~ C~CTeMhl ~eJI8C006pa3HO TpaKTO-
BaT:b KaK 3a;n,at~y OllTli!Mll!3a~Tt!l'l MHOrOMepHOM Jn!H8HHOiii H8CTa~oHapHOH CI1CT6Mbl CO CTOXaCTli!'I8CKli!MI1 B03,I\8WCTBl/lflMll!,
K
TaKOM 3a~aqe CB0~11TCH 3a-
,I\a~a OITT!IIMJ/!3al.\ll!lll MHOrOMepHOrO KOppeKTli!PYJOU18rO ycTpO~CTBa B
KJiacce
JIJ/!HetiHhiX HSC Tai(l/IOHapHHX ycTpOi1C TB, 6CJI!Il 061>eKT C TOXaC Tl'l ~UC Kvliii KOTOpb!e ~pyrllle aHanOrl/I~Hhle 3a,IIatlll!
1
1
!I! Re-
TaK~e OITTl/IMl/13aW1R MHOrOMepHOtl
a
CTOXaCTll!qeCKOtl Cl/ICT8Mhl 1 KOr~a Tpe6yeTCR, qTQOhl cpe~Hl/18 xapaKTepi'!CTlllKlll o6eCIT8l!Tt!BaJITI! HylltHOe Kat16C'l'J.JO CI1CT8Mhl. Hl'!llte paCCMaTpli!BaiOTCH HSKOTOpue BOITpOChl,
ec Tee TBSHHbiM o6pa30M B03HlID
liTO MHOrOMepHyiD J!Tt!H8hHYJO Cii!CTSMY MO~HO OITl/IChlBaTI> C ll0-
MaTpMt~aoti.
K( t,1:) ,
lct, A.)
(2)
OlK!d,ZJ;aH!dH, Tor,ZJ;a B
'f:..)
:Me
~0'
i
,1J;JIH
cymecTnyeT e~HHCTBeHHOe perneH!16
onepaTop
Bwny
cny-
ero jj~JiaeMh!M rrpeoopaaoBaH!1eM;
C ,li,8HHb!M ypenHeHHeM CBH38H orpaHW~6llHb!J/l B
llycn
(II)
~YHKL\~H Mel!t~Y-BXOf\HNM
B3aHMHa.H KOppeJIHL\HOHHaH cKanHpHaH
nyc n
E t. a~ oc ) ,
B>o ;
tlJYHKL\~.H nxo~Horo
~ai1Hh!M rrpoueccoM l1
Ktti-
R2 c.t) , t
ti.f~Lt>), 1.d,21... ~e~JCTBym~M
Ha
JIID6oti QJYHKL\HH
~:.. (t)' -
f"l"t,t)
, ~elilcTBYID!l\!1~
KOMIIOHeHTbl