IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 37, NO.

9, SEPTEMBER 1992 1405

CONCLUSIONS and
Random and nonrandom consultation schemes are examined
and different mathematical formulations of the decision making
problem in the presence of consultation cost are analyzed. The REFERENCES
problem of consulting sensors is cast in a general framework [ll R. R. Tenney and N. R. Sandell, Jr., “Detection with distributed
suggested for sensor integration that satisfies design criteria that sensors,” ZEEE Trans. Aerospace Electron. Syst., vol. AES-17, pp.
guarantee the benefits of data fusion 1111. The analysis and the 501-510, July 1981.
numerical results indicate that the optimal solutions to the [21 S. C. A. Thomopoulos, R. Viswanathan, and D. K. Bougoulias,
“Optimal decision fusion in multiple sensor systems,” ZEEE Trans.
different schemes introduced in this note satisfy the three data Aerospace, Electron. Syst., vol. AES-23, no. 5 , Sept. 1987.
fusion design criteria which we advocate to be essential for the [31 S. C. A. Thomopoulos, D. K. Bougoulias, and L. Zhang, “Optimal
design of any practical decision making system, namely mono- and suboptimal distributed decision fusion,”SPZE’s 1988 Tech.
tonicity with respect to fused information, monotonicity with Symp. Optics, Electro-Optics, Sensors, Orlando, FL, Apr. 4-8, 1988.
R. Srinivasan, “Distributed radar detection theory,” ZEE Proceed-
respect to the cost associated with aquiring the information, and ings, Part F , vol. 133, no. 1, pp. 55-60, Feb. 1986.
robustness with respect to a priori uncertainty. Comparison be- S. C. A. Thomopoulos, R. Viswanathan, and D. K. Bougoulias,
tween the random and nonrandom consultation schemes demon- “Optimal distributed decision fusion,” ZEEE Trans.Aerospace Elec-
strates that nonrandom consultation considerably reduces the tron. Syst., vol. 25, no. 5, pp. 761-764, Sept. 1989.
communication requirements for achieving a desired perfor- R. Viswanathan, S. C. A. Thomopoulos, and R. Tumuluri, “Opti-
mal serial distributed decision fusion,” ZEEE Trans. Aerospace
mance level compared to the communication requirements for Electron. Syst., vol. 24, no. 4, July 1988.
achieving the same performance level with random consultation. J. D. Papastavrou and M. Athans, “A distributed hypothesis-test-
Additional analytical and numerical results from different for- ing team decision with communications cost,” in Proc. 25th Conf:
mulations of the problem can be found in [12]. Decision Contr., Athens, Greece, Dec. 1986.
J. D. Papastavrou, “Distributed detection with selective communi-
APPENDIX cations,” Ph.D. dissertation, Dept. Elect. Eng., M.I.T., Cambridge,
MA, 1986.
To derive the ROC of the serial combination of S1 and S2, H. L. Van Trees, Detection, Estimation and Modulation Theory, Part
we consider a system of two sensors S1 and S2 in which the Z. New York Wiley, 1968.
J. V. Di Franco and W. L. Rubin, Radar Detection. MA: Artech
decision u 2 of sensor S2 is transmitted to sensor S1 and is then House, 1980.
used together with the raw data Z,available to S1 to arrive at a S. C. A. Thomopoulos, “Sensor integration and data fusion,” J.
final decision ul. To that extent, we follow an analysis similar to Robotic Syst. (Special Issue), vol. 7, no. 3, 1990.
) p(rllHl),
[61. Denoting the distribution of rl as p ( r l ~ H oand S. C. A. Thomopoulos and N. N. Okello, “Distributed detection
the likelihood ratio at sensor S1 becomes with consulting sensors and communication cost: analysis and
numerical results,” IPIS Lab., Dept. Elect. Eng., SIU, Carbondale,
IL. Tech. Rep. TR-SIU-EE-10, 1988.
L(r19 U 2 1 H l ) -
- P(rllHl)[PD, S(u2 - 1) + (1 - P D 2 ) Vu,)]
-, “Distributed detection with consulting sensors and communi-
L(rlY ‘2IHO) P(r11HO)[PF2 2‘(‘ - l) + ( l - P F 2 ) S(u2)1 cation cost,” Decision Contr. Syst. Lab., Dep. Elect. Comput. Eng.,
The Pennsylvania State Univ., University Park, PA, Tech. Rep.,
(‘4.1) Oct. 1990.
where
P D ~ Pr(U2 = 1IHl) and P F =
~ Pr(U2 = IIHo) (A.2)
are the detection and false alarm probabilities at S2, respec-
Cyclomonotonicity and Stabilizability Properties of
tively, u 2 = k implies that sensor S2 decides Hkrk = 0, 1, and
S ( x ) is Kronecker’s delta Solutions of the Difference Periodic Riccati
Equation
S(x)= x=o
[o X Z O Giuseppe De Nicolao
Hence, if t is the threshold at sensor S1, the test at S1 ,4b&~~--This note deals with the Kalman filter associated with a
reduces to discrete-timelinear T-periodic system. The problem considered is that of

P(T1IHl)PDZ ;f if u2 =
selecting an initial covariance matrix such that the periodic filter based
on the first T values of the Kalman gain is stabilizing. Sufficient
conditions are given that hinge on the cyclomonotonicity of the solution
p(rllHO)PF2 H, of the periodic Riccati equation. Potential applications are found in
filter design, quasi-linearization techniques for the periodic Riccati
P(rllffl)(1 - PD2)
p(rllHO)(l - P F 2 ) H,
‘ if ‘2 = O. (A.3)
equation, and the design of receding-horizon control strategies for
periodic and multirate systems. When specialized to time-invariant
systems, our results give rise to new sufficient conditions for the cy-
Alternatively domonotonicity of the solutions of the time-invariant Riccati equation
and the existence of periodic stabilizing feedbacks.
tl,l if u2 = 1
A(rl)zo tlp if u2 = o (*e4) Manuscript received January 4, 1991; revised July 26, 1991. This work
was supported by the MURST-Project, “Models, Identification, Systems
where Control, Signal Processing,” and the Centro di Teoria dei Sistemi of the
CNR, Milan, Italy.
p(r1 IHI ) The author is with the Centro di Teoria dei Sistemi-C.N.R., Diparti-
A(rl) = ~

mento di Elettronica, Politecnico di Milano, Milan, Italy.

P (r2 IHo) IEEE Log Number 9201322.

0018-9286/92$03.00 0 1992 IEEE

1406 IEEE TRANSACTIONSON AUTOMATICCONTROL,VOL. 37, NO. 9, SEPTEMBER 1992

I. INTRODUCTION Since multirate sampled-data systems can be modeled by means

The solution of the optimal filtering probleh for discrete-time of periodic systems, see e.g., [3], adaptive control of multirate
linear periodic systems calls for the difference periodic Riccati systems is a potential area of application for the results of this
equation (PRE) note.
In the context of time-invariant systems, a problem analogous
P(t + 1) = A ( t ) P ( t ) A ( t ) ' + Q ( t ) - A ( t ) P ( t ) C ( t ) ' to the one considered herein was dealt with in a number of
notes [4]-[9]. There, the time-invariant difference Riccati equa-
* [C(t ) P ( t ) C ( t ) ' + I ] - ' C ( t ) P ( t ) A ( t ) ' , tion (RE) was considered
P(T)=P, (1)
P( t + 1 ) = A P ( t ) A ' +Q - A P ( t ) C ' [ CP( t ) C ' + I ] -'CP( t ) A ' ,
where A(.):Z + R"'", C(.): Z --* Rp'", e(.):
Z + R"'", Q ( t )
P(0) ' P o (2)
2 0, V t are real periodic matrix functions and T is the (integer)
period.' The symmetric nonnegative definite matrix P, is the together with the Kalman gain K ( t ) = A P ( t ) C ' [ C P ( t ) C ' + I]-'.
initial state covariance. The celebrated Kalman gain is given by The problem consisted in selecting an initial condition Po such
that the matrices A - K(t)C have all their eigenvalues strictly
K(t) = A(t)P(t)C(t)'[C(t)P(t)C(t)'+ I ] -'.
inside the unit disk for all t 2 0. In other words, conditions were
At present, a consistent corpus of results is available for linear searched under which the "frozen" Kalman filter is asymptoti-
periodic systems. In particular, stability, controllability, observ- cally stable. For this purpose, some strict connections between
ability, stabilizability, and detectability together with their con- such stability property and the (nonincreasing) monotonicity of
nections with the periodic solutions of the PRE, have been the solutions of the R E were established [4]. With suitable
extensively dealt with (see Section I1 for a concise survey). monotonicity results at hand, the analysis was carried out by
In general, unless the initial condition P,. is suitably chosen, resorting to the so-called fake algebraic Riccati equation
the Kalman gain K ( - )is not periodic. On the other hand, there (FARE). In this approach, the original RE is transformed into
is an obvious interest for the design of periodic stable predictors. an FARE, i.e., an ad hoc algebraic Riccati equation (ARE)
In this nete, we try to answer the following question: for what constructed in such a way as to admit P ( t ) as a solution. Then,
values of the initial covariance P, is the periodic gain standard stability results for the ARE are employed. As shown
in [6] and [9], the requirement of nonincreasing monotonicity
+
~ , ( t iT) 4K ( t ) , t = T,T + I,..., + T - I , i Ez limits the analysis to initial covariances greater than or equal to
the strong solution. Hence, in [9] the FARE techniques were
such that the periodic closed-loop matrix A ( . ) - K,(.)C(-) is
extended to include the monotonicity of the first difference (first
(asymptotically) stable? In other words, when does it happen
derivative, in continuous-time [6]) of the solution, therefore
that, by periodicizing the first T values of a (not necessarily
allowing for a variety of initial conditions.
periodic) Kalman gain, a periodic stabilizing gain is obtained?
In this note, fake Riccati techniques are extended to periodic
This question is widely motivated. First, let us consider the
discrete-time systems in order to derive periodic stabilizing fil-
filtering of detectable and nonstabilizable periodic systems. It is
ters from the time varying solution of the PRE. Our approach
well known that in the presence of unreachable characteristic
parallels the one adopted in the time-invariant case and the fake
multipliers on the unit circle, there does not exist a stabilizing
periodic Riccati equation (FPRE) is introduced. One main dif-
periodic solution of the PRE [l].Anyway, under certain condi-
ference, however, is that cyclomonotonicity (see Definition 4)
tions, the time-uarying filter is stable and converges to a periodic
replaces monotonicity as a key tool in the analysis. Analogously,
equilibrium, although the limiting periodic filter is not stable. If
the monotonicity of the first difference is replaced by second-
sufficient conditions for the above problem were available, we
order monotonicity (see Definition 5).
could stop updating the PRE (1) at a certain point and the last T
Since time-invariant systems can be seen as periodic systems
gains would provide a suboptimal periodic and stable filter. It is
of arbitrary period, our results apply to the time-invariant RE as
worth remarking that unreachable multipliers arise, for example,
well. In particular, in some time-invariant filtering problems,
in the modeling of constant or sinusoidal deterministic disturb-
FARE techniques may not work, whereas the FPRE approach
ances of unknown amplitude. ~r
yields a periodic stabilizing filter (see the Example in Section
A second application is concerned with the initialization of
IV).
quasi-linearization techniques for the computation of the peri-
odic solution of the PRE [2]. As the quasi-linearization proce- 11. PRELIMINARIES
dure has to be initialized by means of a stabilizing periodic
feedback, in [2] it -was heuristically suggested to integrate the Periodic systems: Stability and structural properties: Consider the
PRE over a "sufficiently long" range and then use the last T discrete-time periodic system
values of the Kalman gains to build a (hopefully stabilizing) X(t + 1) = A ( t ) x ( t ) + B ( t ) u ( t )
periodic feedback. It is apparent that solving the problem con-
y(t) = Cjt)x(t).
sidered in this note would give a theoretical basis to such a
heuristic procedure. The transition matrix associated with A ( . )will be denoted by
Finally, we turn to the dual of Kalman filtering, i.e., optimal q A ( t ,T ) . In the literature on periodic systems, matrix QA(7) =
control. If the original problem is translated into an optimal WA(7 + T, T ) is named the monodromy matrix at T . In general,
control setting, it can be seen that we are actually concerned the monodromy matrix QA(7)depends on T . However, its eigen-
with guaranteeing the stabilizability properties of receding- values, the so-called charactektic multipliers of A ( . ) , together
horizon control strategies. In the design of adaptive controllers, with their multiplicities, are time-independent; see, e.g., [lo]. It
this would allow us to skip the cumbersome computation at each is well known that matrix A ( . ) is (asymptotically) stable if and
step of the periodic solution of the Riccati equation without only if all its characteristic multipliers belong to the open-unit
sacrificing the asymptotic stability of the closed-loop system. disk.
, -

IEEE TRANSACTIONS ON AUTOMATIC CONTROL VOL 3-. SO. 9. SEPTEMBER 19Y2 1401

We now turn to the structural properties of the pairs ( .4( .I. B ( .1 where
and A(.),C(.>>.
Definition 1 [ l ] : A characteristic multiplier A of A ( . ) is said to
+ z]-'c(t)
A,(?)= A ( t ) - A(t)Pl(t)C(t)'[C(t)P1(f)C(t)'
be ( A ( . ) C(-))-unobservable
, at T if there exists w f 0 such that (3.a)
k(t)= C(t)P,(t)C(t)' + I (3 .b)
QA('r)W = hw
D(t> Q 2 ( t ) - Qi(t>.
Proof The proof can be carried out by following step by
step the proof of the analogous result for the difference Riccati
Definition 2 [ I ] : A characteristic multiplier A of A ( . )is said to equation with time-invariant coefficients [9], and is therefore
be ( A ( . ) ,B(.))-unreachable at T if there exists U # 0 such that omitted.
Lemma 3 [I]: Suppose that ( A ( . ) , C ( . ) is ) detectable and let
a A ( 7 ) ' u = ALJ P+(.)be a strong solution of the PRE (1). Furthermore, let Pl(.)
be a symmetric T-periodic solution of the following PRE:
B(f - l)'qA(T + T , t ) ' U = 0, 7 <t I 7 + T. P ( t + 1) = A ( t ) P ( t ) A ( t ) ' + Q , ( t ) - A ( t ) P ( t ) C ( t ) '

Remark: IL can be seen that, if A # 0 is an ( A ( . ) , B(.))- .[ C ( t ) P (t)C( t ) ' + I ] -IC( t ) P ( t ) A ( t ) ! .

unreachable characteristic multiplier at T , then it is ( A ( . ) ,B(.))- If Q ( t > 2 Q,(t>,V t , then P'(t) 2 P,(t>, Vt.
I

unreachable at any time point, i.e., the unreachability notion is Lemma 4 [14]: Let A with IAl = 1 be an ( A ( . ) , e(.))-
time-independent for the nonzero characteristic multipliers [l], unreachable characteristic multiplier of A ( . ) and consider any
[Ill. symmetric periodic solution P(.)of the PRE. Then, @,&)'U = ALJ
The following result can be easily proven by resorting to the implies P ( t ) u = 0.
definition of the unreachable multiplier. The proof of the analo- Theorem 1 [2]: Assume that ( A ( . ) , C ( . ) )is detectable and
gous continuous-time result can be found in [12]. (A(.), e(.)) is stabilizable. Then, the PRE (1) admits a unique
Lemma 1 [131: Consider any T-periodic gain matrix L ( . ) : nonnegative definite periodic solution. Moreover, such a solu-
Z + R n x p . If A is an ( A ( . ) - L ( . ) C ( . ) , e(.) +
L(.)L(.)')- tion is stabilizing.
unreachable characteristic multiplier of A ( . ) - L ( . ) C ( . ) ,then A
111. CYCLOMONOTONICITY
AND STABILIZABILITY
is an ( A ( . ) ,Q(.))-unreachable characteristic multiplier of A(.).
Furthermore, if U is an eigenvector of @ A - L c ( ~ ) ' with eigen- Our analysis of the stabilkability properties of the solutions of
value A, it is also an eigenvector of QA(7)' with eigenvalue A. the PRE will call for the following notion of cyclomonotonicity.
The pair ( A ( . ) B(.))
, is reachable (observable) if all the char- Definition 4: The symmetric matrix function P ( t ) = P ( t ) ' ,
acteristic multipliers A of A ( . ) are ( A ( . ) ,B(.))-reachable t 2 T is said to be a cyclomonotonic nonincreasing sequence
[ ( A ( . ) C(-))-observablel.
, The pair ( A ( - )C(.))
, is said to be de- + +
of period T , if P(T + kT T ) I P(T kT), k 2 0, for some
tectable (stabilizable) if all the characteristic multipliers A of fixed T .
4 . 1 such that (AI 2 1 are (4.1, C(.))-observable [ ( A ( . ) lN.1)- , One could also introduce a notion of strong monotonicity by
reachable]. For further details on the theory of linear periodic +
requiring that P ( t T ) I P ( t ) , t 2 T . In other words, a se-
systems, the interested reader is referred to [lo]. quence is strongly cyclomonotonic of period T if every subse-
quence starting from t with T I t IT + T - 1 is cyclomono-
Periodic Riccati equations: Below we recall some definitions tonic. The following lemma provides a sufficient condition for
and known results concerning the PRE. the cyclomonotonicity of the solution of the PRE. Furthermore,
In correspondence of a T-periodic solution of the PRE (l), it shows that as far as the solutions of the PRE are concerned,
one can define the closed-loop matrix cyclomonotonicityentails strong cyclomonotonicity.
Lemma 5: Let P ( - ) be the solution of the PRE (1). If
F(t) =A(t) - A(t)P(t)C(t)'[C(t)P(t)C(t)' + z]-'c(t). +
P(T + T ) I P ( T ) , for some T , then P ( t T ) I P ( t ) for all
t 2 7.
Definition 3 [ l ] : A T-periodic symmetric positive semidefinite +
Proof Let P , ( t ) = P ( t T ) , P 2 ( t )= P ( t ) , and P ( t > = P ( t )
solution P'(.) of the PRE is called strong if all the characteris- - P ( t + T ) . Then, by using Lemma 2, it follows that
tic multipliers of the corresponding closed-loop matrix belong to
the closed-unit disk.
P ( t + 1) = A , ( t ) P ( t ) A , ( t ) ' -A,(t)P(t)C(t)'
Now, consider a periodic matrix function X ( . ) (not necessarily +
. [ C ( t ) P ( t ) C ( t ) ' k ( t ) ]- l C ( t ) P ( t ) x l ( t ) ' (4)
a solution of the PRE), with X ( t + T ) = X ( t ) , and let r(t)= where xl(t) and I?;(t)are a s i n (3). Since (4)is a (time-varying)
A(t)X(t)C(t)[C(t)X(t)C(t)' + Z1-I be the corresponding gain. difference Riccati equation, P ( T ) 2 0 implies P ( t ) 2 0, t 2 T.
Then, X ( . ) will be said to be stabilizing if the matrix A ( . ) - Before attacking the problem stated in the Introduction, a
T(.)C(.) is asymptotically stable. couple of preliminary definitions are still needed.
Lemma 2: Let P I ( . )and P2(.) be the solutions of two PRE's Consider a generally time-varying matrix function X ( . ) . The
(1) with the same A(.),C(.>matrices but possibly different e(-) symbol X , ( . )will denote a T-periodic matrix function such that
matrices e,(.)
and Q2(.) and with initial conditions P , ( T ) = P,
t I s I t T - 1, +
2 0 and P2(7) = P2 2 0, respectively. Then, the matrix P ( t ) = X,(S) = X ( s ) ,
P2(t) - P,(t) satisfies the following Riccati equation: X*(s + T ) = X,(S),VS.
Moreover, give4 a (generally time-varying)solution P(.)of the
P(t + 1) = L q ( t ) P ( t ) L i , ( t ) ' + Q ( t ) - L q ( t ) P ( t ) C ( t ) ' PRE (l), define e,(.) as

.[ C ( t ) P ( t ) C ( t ) ! + I?( t ) ]-IC( t ) P ( t)A,(t ) ' DAs)= (:::; + P(t) - P(t + T )

s f t + T - 1
+
s = t kT - 1
1408 IEEE TRANSACI'IONSON AUTOMATIC CONTROL, VOL. 37, NO. 9, SEPTEMBER 1992

for k 2 0. Our approach to the study of the stabilizing proper- 2)A(T + T - 1 ) ' =~ 0. This procedure can be iterated to yield
ties of a time-varying solution P(.) of the PRE consists of Q ( t - 1)PA(7+ T , t ) ' ~
= 0, T <t IT + T
viewing P,(.)as a periodic solution of a suitable PRE. Indeed,
from (1) evaluated over the interval [ T , T + TI, it follows that i.e., A is ( A ( . ) e(.))-unreachable,
, and furthermore
P,(.)is a periodic solution of the following PRE: V(T)@~(T)'U
= 0.
P,(t + 1) = A ( t ) P , ( t ) A ( t ) ' + i S , ( t ) - A ( t ) P , ( t ) C ( t ) ' Recalling that (PA(7)'u= Au, A # 0, it follows that V(T)U= 0,
i.e., u*P(T)u = 0. Since u*P(T + T)u = 0 and u*P(T)u= 0, by
-[C ( t ) P , ( t ) C ( t ) ' + Z]-'C(t)P,(t)A( t ) ' . ( 5 )
the very definition of D,(-)
it follows that
Equation (5) is an ad hoc PRE that has been built starting
from a predefined periodic solution P,(-).For this reason, in GT(t- ~ ) W , ( T + T , t)'u = 0, T <t < T + T
analop with the fake algebraic Riccati equation (FARE) [6]( 5 ) i.e., A is ( A ( . ) @,(.))-unreachable
, also.
will be named fake periodic Rccati equation (FPRE). Theorem 3: Let P(.)be the solution of the PRE (1) with
Now, we are in a position to exploit the stability theory initial condition P ( T ) = P, 2 0 and assume that
relative to the periodic solutions of the PRE. For this, the
aonnegative definiteness of @,(.), as well as the stabilizability of 1) ( A ( . ) C(.))
, is detectable
the pair (A(.),QT(.)) are to be assumed. 2) (A(.), @
,
(
e
) is stabilizable
Theorem 2: Let P ( . ) be the solution of the PRE (1) with 3) P(T 4- T ) IP ( T ) .
initial condition P ( T ) = P, 2 0 and assume that Then, the periodic matrix functions PT+kT(.)are stabilizing for
1) ( A ( . ) C(.))
, is detectable each k 2 0.
2) (A(.),@J.)) is stabilizable Proofi First note that due to assumption 3) and Lemma 5,
3) Q(T + T - 1) + P ( T ) - P(T + T ) 2 0. aT+kT(t) 2 0, k 2 0. The periodic matrix function PT+kT(.) sat-
isfies the FPRE
Then, the periodic matrix function PJ.1 is stabilizing.
Progt: The periodic matrix function P,(.)is a nonnegative P7+kT(t + l)
'
periodic solution of the FPRE (5). By the assumptions, =A( kT( A( ' + GT+
kT ( ) -A( )&
' 7' ( t IC(
( A ( . ) ,C(.)) is detectable and ( A ( - ) , a,(.))
is stabilizable. Then,
'T+

Theorem 1 entails that the FPRE ( 5 ) admits a unique nonneg- *[ ( ' ) ' + 1 -Ic(
k T ( )( 'T+ +k T ( A ( ' ( 6 ) )e '
ative definite periodic solution, which necessardy coincides which may be rewritten as a Lyapunov equation
with PJ.). The stabilizability property of P,(.) follows by ob-
serving that under the stated assumptions, the unique nonnega- 'T+kT(l + '1 Fk(t)Pr+kT(t)Fk(t)'

tive definite periodic solution of the FPRE (5) is stabilizing + @T+k7'(l) + Kk(t)Kk(t)' (7)
(Theorem 1).
where
In analogy with the time-invariant case [4], it is possible to
strengthen the above result by showing that under slightly dif- ~k (t ) = A (t )PT+ k T (t )c( ) C( t >pT+k T (t + 1I - '
ferent assumptions, the stabilizing property of the solution is
preserved in forward time, i.e., PT+kT(.)is stabilizing, for all
Fk(t)= A ( t ) - Kk(t)C(t).
k 2 0. In order to prove this result, we first need the following: From (7) it follows that
Lemma 6: Let U be an eigenvector of mA(7)' with eigenvalue
A. Consider a nonnegative defifiite solution P ( . ) of the PRE (1). 'T+kT(' +T,
T+ T
Then
1) u*P(T + T)u = 0 and
= mF,(T>P,+kT(T>~F,(.>' + c
f=T+1
*Fk(T + T ?t )

2)A#O '[@,+kT(t - 1) +Kk(f - I ) K k ( t - l ) ' ] q ~ ~ ( TT+y t ) ' .

imply that A is both (A(.),Q(2) and ( A ( - )@J.))-unreachable
, (8)
and u*P(T)u = 0 (where U* denotes the conjugate transpose
of U ) . Let U be an eigenvector of mFJ7)' associated with the eigen-
Proofi Let P ( t ) = V(t)'V(t).As shown in [4, Lemma 51, the value A. Then, (8) yields
PRE can be given the alternative expression T+ T

+ 1) =A(t)V(t)'H(t)V(t)A(t)' + Q ( t )
P(t (1 - IAI*)v*P,+~~(T)v = v*qFk(7+ T , t ) [ Q T + k T ( t- 1)
f=T+1

where H ( t ) = [Z + V(t)C(t)'C(t)V(t)']-'> 0. Now, consider + K k ( t - 1)K,(t - l)']q~,(T + T , t ) ' ~ (9)

the above expression for t = T + T - 1. By multiplying on the
left by U* and on the right by U,u*P(T+ T ) u = 0 implies V(T+ Now, two cases are possible.
T - ~ ) A (+T T - 1)'u = 0 and Q(T + T - 1)u = 0. In turn, by Case 1: u * P , + ~ ~ ( T#) u0. Equation (9) entails [AI I 1. Note
multiplying the expression that if JAl < 1, the thesis follows. Hence, assume by contradic-
tion that IAl = 1. By expression (9)
V ( T+ T - ~ ) ' V (+T T - 1)
+ T - ~ ) ' H ( T+ T - 2 )
= A ( T + T - 2)V(T [GT+kT(t - '>
+ Kk(t - l)Kk(t - ')'I
.V(T+ T - 2 ) A ( T + T - 2)' + Q(T + T - 2 ) WF,(~+T,t)'u=0, T<t<'T+T
on the left by u*A(T+ T - 1) and on the right by A(T + T - so that A is (Fk(-),@T+kT(.) + K,(.)K,(.)')-unreachable. This,
1)'u, from V(T+ T - ~ ) A ( +T T - 1)'u = 0 it follows that V(T in view of Lemma 1, implies that is also an ( A ( * )QT+kT(.))-
,
+ T - 2)A(T + T - 2)'A(T + T - 1 ) ' =~ 0 and Q(T + T - unreachable characteristic multiplier and U is such that QA(7)'u
 -
IEEE TRANSACTIONS ON AUTOMATTC CONTROL, VOL. 37, NO. 9, SEPTEMBER 1992 1409

= Au. Then, by Lemma 4,u * P , + ~ ~ ( T=) u0 and we fall into Case Proofi Observe that
2.
Case 2: U * P , + ~ ~ ( T=) 0,
U A # 0. As in Case 1, A turns out to DT+kT+T(' + - '1 - (Z,+kT(' + - ')
be an eigenvalue of aA(7)' with eigenvector U . Then, Lemma 6 =- P ( T + kT + 2 T ) + ~ P ( +T kT + T ) - P ( T + k T ) .
can be recursively applied to show that u*P,+~~(T)u = 0 and A is Then,
- condition 4) of the theorem and Lemma 7 imply
( A ( - )~,,,,(-))-unreachable,
, for i = k - 1,k - 2 , . . - ,0. In par- 2 D,+,,Q), V t , k 2 0. Since, by condition 31, (Z,(t)
D,(.))
ticular, since ( A ( . ) , =
is stabilizable, it follows that [ A ( < 1. 2 0, it follows that e,+&) 2 O,(t>2 O,yt, k 2 0. Then, by
It is worth observing that a necessary condition for Theorem 3 recalling assumption 2), the pairs ( A ( - ) ,Q,, k T ( . ) ) are easily
to hold is that the initial covariance P, is greater than or equal shown to be stabilizable for all k 2 0. Finally, the application of
to the strong solution P + ( T ) of (1) evaluated at T . In fact, Theorem 2 yields the thesis.
assumption 3) implies D T ( t )2 Q(t), V t . Then, as P,(.) is the
periodic stabilizing (and hence strong) solution of the FPRE (6), I v . CYCLOMONOTONICITYAND THE TIME-INVARIANT
P J T ) = P, 2 P + ( T ) (Lemma 3). An analogous limitation was RICCATIEQUATION
also encountered in the time-invariant FARE analysis and was The fact that a time-invariant system can be viewed as a
overcome by working out some monotonicity results for the first periodic system of arbitrary period allows us to draw some
difference of the solution of the Riccati equation [9]. In the interesting consequences from the results of the previous sec-
periodic case, in order to extend the FPRE analysis to initial tion.
conditions not necessarily above the strong solution, we now We begin with the issue of cyclomonotonic behavior of solu-
introduce the notion of second-order cyclomonotonicity. tions of the time-invariant RE. The observation that in general,
Dejnition 5: The symmetric matrix function P ( t ) = P ( t ) ' , t 2 the solution P ( . ) of the RE, besides being not monotonic, need
T , is said to be a second-order cyclomonotonic nonincreasing be neither increasing nor decreasing cyclomonotonic traces back
sequence of period T , if to [15]. Examples of failure of cyclomonotonicity can also be
P ( T + kT + 2 T ) - ~ P ( +T kT + T ) + P ( T + k T ) I0, found in [4]. Here, by means of Lemma 5 we are able to provide
a sufficient condition for cyclomonotonicity as stated in the
k2O. following.
A sufficient condition for second-order cyclomonotonicity of Corollary 1: Assume that the time-invariant pair ( A , C ) is
the solution of the PRE is given below. Again, for the solutions detectable and let P ( . ) be the solution of the time-invariant R E
of the PRE, second-order cyclomonotonicity entails strong sec- (2) with initial condition P(0) = Po. Assume that Po > 0 and
+ +
ond-ordercyclornonotonicity, i.e., P ( t 2 T ) - 2 P ( t T ) + P ( t ) Po > P + , where P + denotes the (unique) strong solution of the
I0, t 2 7 . ARE
Lemma 7: Let P ( - ) be the solution of the PRE (1). If +
P = APA' Q -APC'[CPC' I]-'CPA'. + (10)
P(T + 2 T ) - 2P(T + T ) + P ( T ) I0 Then, there exists an integer T 2 0 such that P ( . ) is strongly
cyclomonotonic (nonincreasing) of period T , i.e., P ( t + T ) I
for some T, then
H t ) , t 2 0.
P(t + 2 T ) - 2 P ( t + T ) + P ( t ) I0 Proofi Under the stated assumptions, the solution P ( . ) of
the R E asymptotically converges to the strong solution P + of
for all t 2 T . the ARE [16, theorem 4.21. Now, letting A(t) = P ( t ) - P + , and
Prmfi Let us define for t 2 T
x E R"

+T)
P l ( t ) iP ( t x ' [ P 0 - P ( t ) ] x =.'[Po - P + ] x - x ' A ( t ) x
2 A,,ll~Il~ - IlA(t)ll2ll~ll~
P2(t) P(t + 2T)
where A,, > 0 is the minimum eigenvalue of Po - P I . Since
P,(t) P(t) - Pl(t) lim A ( t ) = 0
t+m

P2(t) 4P 2 ( t ) - P l ( t ) there exists a time point T such that Po - P ( T ) 2 0, i.e.,

P ( T ) IPo. Then, the cyclomonotonic behavior of the solution
S ( t ) 4P 2 ( t ) + PI@). P ( . ) straightforwardly follows from Lemma 5 .
Another interesting issue concerns the possibility of deriving a
Then, by following step by step the proof of lemma 3.1 in [9], periodic stabilizing gain from the time-varying solution of the
one can show that S ( t ) I0, t 2 T. time-invariant RE. The following result is a straightforward
We end this section with a stability result which hinges on consequence of Theorem 3.
second-order cyclomonotonicityand does handle a wide range of Corollary 2: Let P ( . ) be the solution of the time-invariant R E
initial conditions. (2) with initial condition P(0) = Po 2 0 and assume that ( A , C )
Theorem 4: Let P ( . ) be the solution of the PRE (1) with is detectable. If there exists an integer T such that
initial condition P ( T ) = P, 2 0 and assume that
i) ( A ( . ) ,~ o ( . ) is
) stabilizable
1) ( A ( . ) C(-))
, is detectable ii) P ( T ) IP(0)
cT(.))+
2) ( A ( - ) , is stabilizable
3) Q(T+ T - 1) P ( T )- P(T T ) 2 0. + then, the T-periodic matrix functions PkT(.) are stabilizing for
each k 2 0.
+ +
4) P(T + 2 T ) - 2P(T T ) P ( T ) 5 0.
Note that, provided that ( A , Q ) is stabilizable, ii) implies i) so
Then, the periodic matrix functions PT+kT(.)
are stabilizing for that under stabilizability, one has only to look for cyclomono-
each k 2 0. tonicity in order to obtain a periodic stabilizing feedback.
1410 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 37, NO. 9, SEPTEMBER 1992

Finally, by putting together Corollaries 1 and 2, it is possible is such that the characteristic multipliers of the periodic closed-
to guarantee the existence of a periodic stable filter under fairly loop matrix A - Ko(-)C are
mild conditions.
Corollary 3: Assume that ( A ,C ) is detectable and ( A , Q) is A, = 0.5, h2 = 0.75.
stabilizable. Let P ( . ) be the solution of the time-invariant R E
(2) with initial condition P(0) = Po > 0 and Po > P’, where P’ CONCLUDING
REMARKS
denotes the strong (and stabilizing) solution of the ARE (10).
Then, there exists an integer T such that P ( T ) IP(0). More- In this note, conditions have been derived that guarantee the
over, the periodic matrix functions PkT(.) are stabilizing for each cyclomonotonicity and second-order cyclomonotonicity of solu-
k 2 0. rn tions of the difference periodic Riccati equation. By means of
Note that Corollary 3 provides an effective constructive proce- these results, it has been shown how to devise periodic stable
dure. Indeed, one only needs to integrate the RE and store the filters using the finite-horizon solutions of the periodic Riccati
Kalman gains until a time point T is reached such that P ( T ) I equation. Cyclomonotonicity analysis and periodic stabilization
P(0). As shown in the example below, there may be situations in techniques were also extended to the time-invariant Riccati
which cyclomonotonicity is more easily achievable than mono- equation. Potential applications range from design of stable
tonicity. Therefore, in case the fake algebraic Riccati equation filters to the implementation of stable receding horizon control
cannot be applied, fake periodic Riccati techniques provide an strategies for periodic and multirate control systems.
additional tool to devise stabilizing feedbacks for time-invariant
systems. REFERENCES
Example: Let C. E. de Souza, “Existence conditions and properties for the
maximal periodic solution of periodic Riccati difference equations,”
Int. J. Cont., vol. 50, pp. 731-742, 1989.
S. Bittanti, P. Colaneri, and G. De Nicolao, “The difference
periodic Riccati equation for the periodic prediction problem,”
with a > 0 and b > 0. It can be seen that the strong solution of IEEE Trans.Automat. Contr., vol. 33, pp. 706-712, 1988.
the ARE (10) is given by P. Colaneri, R. Scattolini, and N. Schiavoni, “Stabilization of
multirate sampled-data systems,” Automatica, vol. 26, pp. 377-380,
P’= [; ;] 1990.
R. R. Bitmead, M. R. Gevers, I. R. Petersen, and R. J. Kaye,
“Monotonicity and stabilizability properties of solutions of the
which is not stabilizing. The solution of the difference R E (2) is Riccati difference equation: Propositions, lemmas, theorems, falla-
cious conjectures and counterexamples,” Syst. Contr. Lett., vol. 5,
pp. 309-315, 1985.
M. R. Gevers, R. R. Bitmead, I. R. Petersen, and R. J. Kaye,
“When is the solution of the Riccati equation stabilizing at every
instant?,” Frequency Domain and State Space Methods for Linear
Systems, C. I. Bymes and A. Lindquist, Eds. Amsterdam, The
Netherlands: North-Holland, 1986, pp. 531-540.

+ 1) = ib +1 0
1 M. A. Poubelle, R. E. Bitmead, and M. R. Gevers, “Fake algebraic
Riccati techniques and stability,” IEEE Trans. Automat. Control,
P(2i

lo (i
a

+ 1)a + 1 1‘
Since ( A , C ) is detectable and Po > 0, the solution P ( . )
vol. 33, pp. 379-381, 1988.
M. R. Gevers, R. R. Bitmead, and M. A. Poubelle, “Monotonicity
and its consequences for stability of receding horizon LQ-control,”
in Proc. Workshop on Riccati Equation Con&, Syst. Sign., Como,
Italy, 1989, pp. 34-37.
asymptoticallyconverges to the strong solution [16, theorem 4.21. C. E. de Souza, “Monotonicity and stabilizability results for the
As the strong solution is not stabilizing, it is of particular solutions of the Riccati difference equation,” in hoc. Workshop
Riccati Equation Contr. Syst., Sign., Como, Italy, 1989, pp. 38-41.
interest to find an easy way to compute a stabilizing filter. Now, -, “On stabilizing properties of solutions of the Riccati differ-
let a = 1 and b = 1/3. Then, there exists no integer value t ence equation,” IEEE Trans. Automat. Contr., vol. 34, pp.
such that P ( t + 1) IP ( t ) and the results of [4] and [9] cannot 1313-1316, 1989.
be applied. However, S. Bittanti, “Deterministic and stochastic linear periodic systems,”

[7 ,”. ] [::]
Time Series, Linear Syst., S. Bittanti, Ed. Berlin: Springer-Verlag,
a2 1 1986, pp. 141-182.
S. Bittanti, P. Colaneri, and G. De Nicolao, “An algebraic Riccati
equation for the discrete-time periodic prediction problem,” Syst.
P(0) - P ( 2 ) = = > 0. Conk Lett., vol. 14, pp. 71-78, 1990.
- - S. Bittanti and P. Colaneri, “Lyapunov and Riccati equations:
b+l 12 Periodic inertia theorems,” IEEE Trans. Automat. Contr., vol.
AC-31, pp. 659-661, 1986.
Hence, by Lemma 5, the solution P ( . ) is cyclomonotonic of C. E. de Souza and G. C. Goodwin, “Periodic solutions of matrix
period 2. Since Do(.) = P(0) - P(2) > 0, the assumptions of Riccati equation in optimal filtering of nonstabilizable periodic
Theorem 3 are satisfied and the periodic gains systems,” in Proc. 10th IFAC World Congress, Munich, Germany,
1987, pp. 249-254.
+ z]-’,
KZi(t) =AP(t)C’[CP(t)C’ G. De Nicolao, “On the convergence to the strong solution of
periodic Riccati equations,” Int. J. Contr., 1992, to be published.
2i It I2i 1, +K Z i ( t+ 2) = K , , ( t ) , V t P. E. Caines and D. Q. Mayne, “On the discrete-time matrix
Riccati equation of optimal control,” Int. J. Contr., vol. 12, pp.
are stabilizing for all i 2 0. For example, the periodic gain 785-794, 1970.
C. E. de Souza, M. R. Gevers, and G. C. Goodwin, “Riccati
equations in optimal filtering of nonstabilizable systems having
singular state transition matrices,” ZEEE Trans. Automat. Contr.,
vol. AC-31, pp. 831-838, 1986.