674 IEEE TRANSACTIONS ON .IUTOMIATIC CONTROL, DECEMBER 1973 the interval (O,t] or, equivalently, the triple ( A , B , C ) is controllable E C, the rank of Q( A,B,C)equals n where if, for each + Q(A,B,C) = The center matrix in (4)will be positive semidefinite if and only if conditions a), b), and c) are satisfied [ l l ] . I n addition, ( 3 ) will be a.ymptotically stable if 6-l(xt)# 0 along any nontrivial solution of (3) [9]. To show asymptotic stability, note that the solution of ( 3 )possesses the representat.ion [Q1',...,Q1n,Q*2,...,Q~n,.'',QnnI and 911 = C = AQP + BQ,-'& QF forj = Oorj> k = 0, (Le., cont,rollable in the sense of Kirillova and Churakova [6] ). The algebraic crit.erion for conlplete conbrollability given in Definition 1is equivalent t.o [7] X(t,o,+) X ( t - s)CC'X'(t - s) ds rank = n DeJninition 2: System (1) is point,wise complete on the interval (O,t] or, equivalent.ly, the pair ( A , B ) is pointwise complete if the rank of R(A , B ) equals n where X'(t S_O, X(t - s - r)BB'X'(t - s - T ) = 1' (5) exp (- AT)CC' exp (- A'T) T > 0, u ( x ( t ) ) = -C'W-'x(t) where 2x'(t)W'-'x(t) + I:, Improved Conditions for the L?-Stability of Nonstationary Feedback Systems CC' W-1. A straightforward calculation shows that verifying that V ( x ) L' Lyapunov functional for t.he closed-loop system + BX(t - is equivalent to verifying that t-,(xt) = 2x'(t)Wx(f) + T ) S (2) ll~ AI'x(t) + B'X(t ~ 11. I(. SUNDA4RESHASA X D 11. A . L.THATH.4CHriR Abstracf-Concerning the Lrstability of feedback systems containinga linear time-varying operator,some of the stringent restrictions imposed on the multiplier a s well as the linear partof the system, in the criteria presented earlier, are relaxed. x'(s)Qx(s) ds The authon' paper [ l ] presents new criteria for the Lrstabiliry of linear and nonlinear systems containing a linear time-varying operator G in L2, in an otherwise timeinvariant negative feedback loop. is a Lyapunov functional for i(t) = (7) x'(s)W-'QW-'x(s) ds is t.he Lyapunov functional associated with the st.abilized syst.em. (See [9] or [ l o ] for detailsregarding Lyapunov functionals and their derivatives.) .Prooj: Since ( A , C ) is controllable, W > 0. Let A^ A A - X([) = A^X(t) T ) B $ 0. REFEREXCES + CC' exp (- As)CC' exp ( - A ' s ) ds, -S- (6 1 111 T u . E. Osipor."Stabilization of controlsystemswithdelays." Di-feerentid'nve L-rarneniyn. rol. 1 . no. 5 , pp. 605-618. 1965. 121 D. It-,, Ross. "Optimal control ofs,-stems describedby dicerential-difference equations. Ph.D. dissertation.Dep.Elec.Eng..StanfordCniv..Stanford. Calif..1967. linear differentialdelay [3] G . J. Kazaroff."Stabilityandstahilizationof . systems." I E E E Trans. Automat.Contr. (Tech.XotesandCorresp.),vol. AC-18. pp. 317318, June 1973. F u n k . Ekrac., vol. 11, [A] D. L. Lukes. "Stahilizability and optimal control." pp. 39-50. 1968. [5] D . L. Kleinman. ";\n easy u.ey t o stabilize a linear constantsystem." I E E E Trans. dufomat. Contr. (Corresp.). vol. XC-15. p. 692. Dec. 1970. 161 F. XI. Kirillova and S. V. Churakova. "The problem of controllabilityof linear systems with after efiect." Diflerential'nye 1.rarnrniya. vol. 3. no. 3. pp. 436-446. 196i. Kith time-delays." in [7] D. H. Chyungand E. E. Lee."Optimalsystems Proc. 1966 IF.4C C o p g r . [ S ] T . 000.J. I.amssskx. and T . Sawarapi. "On t h e controllahility of systems with a time-var>-ing delay."Int. J . Contr.. 1.01. 14. no. 5 . pp. 975-987. 1 9 i l . [9] X. X. Krasox-skii. Stability of Motion. Stanford.Calif.:StanfordUniv. Press.1963. [ l o ] J . Hale. Functiond Di-ferentif Equations. Xew X-ork: Springer.1971. [ l l ] E. Irrein$er and A . Jameson. Conditions for nonexativeness of partitioned matrices I E E E T r a n s . A7domaf. Cordr. (Tech. Sores andCorresp.).vol. AC-17. pp. 14i-148. Feb. 1972. then (1) is stabilized by the cont.ro1law V(Xt) = d8. Hence the asymptotic stabilityof (3) is etablished R-hich, in turn, establishes the asymptotic stabilityof (2). This complet.esthe proof. * den0t.a the Penrose-Moore pseudoinverse. ' Theorem: If ( A , C ) is controllable and ( A , B ) is pointwise complete, and the conditions a ) Q 2 0,b) B'Q*Q, and c) Q - 4 WB'Q* B W 2 0 are satisfied where = T)B'Q(S) R ( i , B ) = 12. ds = n f o r O < t < 03. In the following theorem, W - X'(t - s) is the fundamental solution associated with (2). Assume that (6) is not satisfied. Then the controllability ofAt.he triple ( i , B , C ) is contra$cted since ( A , C ) controllable =) ( A , C ) controllable =) rank Q ( A , B , C ) = R . Assume that ( i )is not satisfied. Then the pointwise completeness of ( 2 , B ) is cont.radicted since ( A , B ) pointKise complete =) rank (i.e., pointwise completein the sense of Ono, Tamasaki,and Sawaragi [8] ). The algebraic criterion for point.wise completeness given in Definition 2 is equivalent. t o [SI and S X'!t - s)C $ 0 R ( A , B ) = [B,AB,...,A"-'B] Q X(t - Substit,ut.ion of ( 5 ) into (4) clearly implies p,(xt) # 0 along a nontrivial solution provided f o r o < [ < 03. rank x(f)+(o)+ - 7). i3 ) 3Ianuseript received April 27. 1973. Differentiating Vl(xr)along the solutions of (3)yields the quadratic The authors are withtheDeoartment form India. 560012,Institute~of~Science. Bangalore of ElectricalEngineering.Indian 675 TECHNICAL NOTES A N D CORRESPONDENCE These criteria permit a new class of linear causal operat.ors X in Lfe with nonst.ationary kernels to be used as multipliers. However, the criteria in [I] require G, M, and X - 1 t.o belong t o a class $9~4of linear causal operators in L e defined by the following condition: if H E Q m , then Hz(f) = ~ ( ~ , T ) x d( rT )tC x(-)E Lpe, R7it.h the ) 0 +f T > t and kernel sat.isfying the twin condit.ions h ( t , ~ = I h ( t ? ~ dt ) ) d~r < a.I t has of late been noticed that t.he latter int,egral condition renders the class EA to be of limited use in practical situations, since it excludes even t.he simple convolutions. The purpose of this correspondence is to remove this stringent requirement from t.he criteria of [ I ] . A careful study of the st,abilitycriteria of [ l ] reveals that t.he condition G, 31, and 31-1 E $EA is imposed only to ensure that these are operat.ors in Ln with finite gain. Let us now define cert,ain new classes QEB and QBC of linear causal operators H in L,, as follows. 1) if H E *RE, then .fr b Hz(t) = where F = F(t), and w = E[w(t)]= 0, E[w(t)wT(r)] = Qs(t - dj rj+l > t, = wj+l = Q(f). < .. . < t . 5 ~ T by + vi (2 1 Ntj+lJj)Tj -I; + wj+l *(tj+l,g)w(n)dg (7) where f w j ) is a discrete white-noise sequence with covariance f A,G)), (1) '$ T AjXj = where zj is an mj vector, xj = r(tj),and { v,) is a zero-mean whitenoise sequence t,hat, isstatisCically independent of w(f) and r(@. The problem of interest in this correxpondence is that of obtaining filtered estimates of x. By introducing the state t,ransit,ion matrix corresponding to F ( t ) , we can replace ( 1 ) by ~ ( ~ , T ) zd~ ( T+f ) X(.) E Lz,, h(f,T) = 0 = Q T), The model is observed at. discrete times ti, lo < tl .fr I= w ( t )is a whit.e-noiseprocess with A&' (21 + 1) = Jy" ~(tj+l,u)Q(u)OT(tj+l,u) du. (3) Evaluation of the @ and A , matrices generally adds computational complexit,gto thefiltering problem. Motivation for conversiy from the continuous formulation ( 1 ) t.o 2) if H E QEC, then H i s defined by ( 1 ) and (2) as earlier, with the thk discrete formulat.ion (1) is that t.he latt.er problem may be solved additiona1,condition using square-root filtering techniques (cf. [SI). These techniques are generally more accurate than is the discret.e Kalman filt,er. In this correspondence it is shown that t,he Dyer-MeReynolds (D-XI) square-root filter can be applied directly to the cont.inuous-discrete system (1) and (2). By writing diflerential equations for t.he D-3f It may be noted that H E QEB implies I:H~(t)ll6 KIII exp ( - k ~ t ) data array matrix, one circumvents the problem of computing @ ~ ( t ) '(where : denote the &norm and * denotes convolution), and A U + which are used by thediscrete square-root filter. and hence H is an operator in Lz wit.h finite gain. Further, H E QEC implie's (see Dunford and Schwartz [2]) ! , ; H z ( - ) 6 ' / ( S l L \ T ? ) : /I%(-):], T H E COKTISCOVE-DI~CRETE D T E R - ~ f c R E n o L D s FILTER and hence H is an operator in Lo with finite gain. With t.his, it is now The D-1.1 filter [ 11 consists of t.he data array (R,d)where RTR = simple to realize that the stabilitycriteria [ I , theorems 1 and 21 hold (COY (a))-l and d = Rr*, X* being the minimum variance estimate of in t0t.o with the operaton G, ~11,and ~1j-lpernlitt.ed to be members x (time arguments are suppressed for notat.iona1 simplicity). This of the class 8 9 ~U ~Q E B U QEC. filter is most useful for problems that. involve state estimat.es a t REFEREXES relatively fen t,imes, for example, after a quantity of nleasurements [ l ] hI. K. Sundareshan and M . -1.L. Thathachar, "Timedomain criteria for the have been processed. Except at. these times, it is possible t o deal only Lystahility of nonstationaryfeedbacksystems," IEKt.' Trans. Aulomaf. mith t.he data array (4)and (5). Each time an estimate is required, Contr., x.01. AC-18. pp. 8C-81. Feb. 1973. [Z] N. Dunfordand J. T. Schwarts. Linear O p e r a t o r s , part 1. New Tork: it is necesary to solve a matrix equat.ion Rx = d, and each t,ime t,he Interscience, 1958, p. 518. covariance of t.he estimate isrequired, one must compute P = R-lR-T.l When estimatesand/or covariances are required at numerouspoints, the computationalburden becomes excessive. The algorithm follows. A Square-Root Data Array Solution of the Propagalion t < t 5 tj+,: and Ih(t,r)l and KZ, 6 K I - exp (-%(t for some positive constants KI (3) 7)) .I, Continuous-Discrete Filtering Problem d - [Rjd] = [ -RF di GERALD J. BIERMXN Abstract-The Dyer-McReynolds [ 11 discrete square-root filtering algorithm is extended to accommodatecontinuous dynamics. Differential equations are given to represent the time evolution of the iilter data array. These equations are nonlinear, but it is shown that the nonlinearities act to enhance thestability of the solution. where r = -4RQRT. The differ_ential equation [ X l d ] = [Rj/dj]. is + rR!rd] init.ialized a.t times ti by t = tj L7pdating (at time tjll): Choose an orthonormal t.ransformation Tj+lsuch t.hat INTRODUCTION In this correspondence we consider the linear dynamic model d --5 = clt F-5 +w + + l ) R y - T ( j 1) and the mat,rix on the where cov ~ j += ~Rp-l(j right is upper triangular. Comput.ation of the triangularization (5) is not discussed here. References [ l ] ,[ 3 ] - [ 5 ] explain in detail how the computation is performed.