InTechOpen uses cookies to offer you the best online experience. By continuing to use our site, you agree to our Privacy Policy.

Computer and Information Science » Numerical Analysis and Scientific Computing » "Advances in Discrete Time Systems", book edited by Magdi S. Mahmoud, ISBN 978-953-51-0875-7, Published: December 5, 2012 under CC BY 3.0 license. © The Author(s).

# Stochastic Mixed LQR/H∞ Control for Linear Discrete-Time Systems

By Xiaojie Xu
DOI: 10.5772/51019

Article top

## Overview

Figure 1. The trajectories of mean values of states of resulting system in Example 5.1.

Figure 2. The trajectories of mean values of states of resulting system in Example 5.2.

# Stochastic Mixed LQR/H∞ Control for Linear Discrete-Time Systems

Xiaojie Xu1

## 1. Introduction

Mixed H2/H control has received much attention in the past two decades, see Bernstein & Haddad (1989), Doyle et al. (1989b), Haddad et al. (1991), Khargonekar & Rotea (1991), Doyle et al. (1994), Limebeer et al. (1994), Chen & Zhou (2001) and references therein. The mixed H2/H control problem involves the following linear continuous-time systems

 x˙(t)=Ax(t)+B0w0(t)+B1w(t)+B2u(t),x(0)=x0z(t)=C1x(t)+D12u(t)y(t)=C2x(t)+D20w0(t)+D21w(t) (1)

where, x(t)Rn is the state, u(t)Rm is the control input, w0(t)Rq1 is one disturbance input, w(t)Rq2 is another disturbance input that belongs to L2[0,) , y(t)Rr is the measured output.

Bernstein & Haddad (1989) presented a combined LQG/ H control problem. This problem is defined as follows: Given the stabilizable and detectable plant (1) with w0(t)=0 and the expected cost function

 J(Ac,Bc,Cc)=limt→∞E{xT(t)Qx(t)+uT(t)Ru(t)} (2)

determine an n th order dynamic compensator

 x˙c(t)=Acx(t)+Bcy(t)u(t)=Ccxc(t) (3)

which satisfies the following design criteria: (i) the closed-loop system (1) (3) is stable; (ii) the closed-loop transfer matrix Tzw from the disturbance input w to the controlled output z satisfies Tzw<γ ; (iii) the expected cost function J(Ac,Bc,Cc) is minimized; where, the disturbance input w is assumed to be a Gaussian white noise. Bernstein & Haddad (1989) considered merely the combined LQG/ H control problem in the special case of Q=C1TC1 and R=D12TD12 and C1TD12=0 . Since the expected cost function J(Ac,Bc,Cc) equals the square of the H2 -norm of the closed-loop transfer matrix Tzw in this case, the combined LQG/ H problem by Bernstein & Haddad (1989) has been recognized to be a mixed H2/H problem. In Bernstein & Haddad (1989), they considered the minimization of an “upper bound” of Tzw22 subject to Tzw<γ , and solved this problem by using Lagrange multiplier techniques. Doyle et al. (1989b) considered a related output feedback mixed H2/H problem (also see Doyle et al. 1994). The two approaches have been shown in Yeh et al. (1992) to be duals of one another in some sense. Haddad et al. (1991) gave sufficient conditions for the exstence of discrete-time static output feedback mixed H2/H controllers in terms of coupled Riccati equations. In Khargonekar & Rotea (1991), they presented a convex optimisation approach to solve output feedback mixed H2/H problem. In Limebeer et al. (1994), they proposed a Nash game approach to the state feedback mixed H2/H problem, and gave necessary and sufficient conditions for the existence of a solution of this problem. Chen & Zhou (2001) generalized the method of Limebeer et al. (1994) to output feedback multiobjective H2/H problem. However, up till now, no approach has involved the combined LQG/ H control problem (so called stochastic mixed LQR/ H control problem) for linear continuous-time systems (1) with the expected cost function (2), where, Q0 and R>0 are the weighting matrices, w0(t) is a Gaussian white noise, and w(t) is a disturbance input that belongs to L2[0,) .

In this chapter, we consider state feedback stochastic mixed LQR/ H control problem for linear discrete-time systems. The deterministic problem corresponding to this problem (so called mixed LQR/ H control problem) was first considered by Xu (2006). In Xu (2006), an algebraic Riccati equation approach to state feedback mixed quadratic guaranteed cost and H control problem (so called state feedback mixed QGC/ H control problem) for linear discrete-time systems with uncertainty was presented. When the parameter uncertainty equals zero, the discrete-time state feedback mixed QGC/ H control problem reduces to the discrete-time state feedback mixed LQR/ H control problem. Xu (2011) presented respectively a state space approach and an algebraic Riccati equation approach to discrete-time state feedback mixed LQR/ H control problem, and gave a sufficient condition for the existence of an admissible state feedback controller solving this problem.

On the other hand, Geromel & Peres (1985) showed a new stabilizability property of the Riccati equation solution, and proposed, based on this new property, a numerical procedure to design static output feedback suboptimal LQR controllers for linear continuous-time systems. Geromel et al. (1989) extended the results of Geromel & Peres (1985) to linear discrete-time systems. In the fact, comparing this new stabilizability property of the Riccati equation solution with the existing results (de Souza & Xie 1992, Kucera & de Souza 1995, Gadewadikar et al. 2007, Xu 2008), we can show easily that the former involves sufficient conditions for the existence of all state feedback suboptimal LQR controllers. Untill now, the technique of finding all state feedback controllers by Geromel & Peres (1985) has been extended to various control problems, such as, static output feedback stabilizability (Kucera & de Souza 1995), H control problem for linear discrete-time systems (de Souza & Xie 1992), H control problem for linear continuous-time systems (Gadewadikar et al. 2007), mixed LQR/ H control problem for linear continuous-time systems (Xu 2008).

The objective of this chapter is to solve discrete-time state feedback stochastic mixed LQR/ H control problem by combining the techniques of Xu (2008 and 2011) with the well known LQG theory. There are three motivations for developing this problem. First, Xu (2011) parametrized a central controller solving the discrete-time state feedback mixed LQR/ H control problem in terms of an algebraic Riccati equation. However, no stochastic interpretation was provided. This paper thus presents a central solution to the discrete-time state feedback stochastic mixed LQR/ H control problem. This result may be recognied to be a stochastic interpretation of the discrete-time state feedback mixed LQR/ H control problem considered by Xu (2011). The second motivation for our paper is to present a characterization of all admissible state feedback controllers for solving discrete-time stochastic mixed LQR/ H control problem for linear continuous-time systems in terms of a single algebraic Riccati equation with a free parameter matrix, plus two constrained conditions: One is a free parameter matrix constrained condition on the form of the gain matrix, another is an assumption that the free parameter matrix is a free admissible controller error. The third motivation for our paper is to use the above results to solve the discrete-time static output feedback stochastic mixed LQR/ H control problem.

This chapter is organized as follows: Section 2 introduces several preliminary results. In Section 3, first,we define the state feedback stochastic mixed LQR/ H control problem for linear discrete-time systems. Secondly, we give sufficient conditions for the existence of all admissible state feedback controllers solving the discrete-time stochastic mixed LQR/ H control problem. In the rest of this section, first, we parametrize a central discrete-time state feedback stochastic mixed LQR/ H controller, and show that this result may be recognied to be a stochastic interpretation of discrete-time state feedback mixed LQR/ H control problem considered by Xu (2011). Secondly, we propose a numerical algorithm for calclulating a kind of discrete-time state feedback stochastic mixed LQR/ H controllers. Also, we compare our main result with the related well known results. As a special case, Section 5 gives sufficient conditions for the existence of all admissible static output feedback controllers solving the discrete-time stochastic mixed LQR/ H control problem, and proposes a numerical algorithm for calculating a discrete-time static output feedback stochastic mixed LQR/ H controller. In Section 6, we give two examples to illustrate the design procedures and their effectiveness. Section 7 is conclusion.

## 2. Preliminaries

In this section, we will review several preliminary results. First, we introduce the new stabilizability property of Riccati equation solutions for linear discrete-time systems which was presented by Geromel et al. (1989). This new stabilizability property involves the following linear discrete-time systems

 x(k+1)=Ax(k)+Bu(k);x(0)=x0y(k)=Cx(k) (4)

J2:=k=0{xT(k)Qx(k)+uT(k)Ru(k)}

under the influence of state feedback of the form

 u(k)=Kx(k) (5)

where, x(k)Rn is the state, u(k)Rm is the control input, y(k)Rr is the measured output, Q=QT0 and R=RT>0 . We make the following assumptions

Assumption 2.1 (A,B) is controllable.

Assumption 2.2 (A,Q12) is observable.

Define a discrete-time Riccati equation as follows:

 ATSA−A−ATSB(R+BTSB)−1BTSA+Q=0 (6)

For simplicity the discrete-time Riccati equation (6) can be rewritten as

 Πd(S)=Q (7)

Geromel & Peres (1985) showed a new stabilizability property of the Riccati equation solution, and proposed, based on this new property, a numerical procedure to design static output feedback suboptimal LQR controllers for linear continuous-time systems. Geromel et al. (1989) extended this new stabilizability property displayed in Geromel & Peres (1985) to linear discrete-time systems. This resut is given by the following theorem.

Theorem 2.1 (Geromel et al. 1989) For the matrix LRm×n such that

 K=−(R+BTSB)−1BTSA+L (8)

holds, SRn×n is a positive definite solution of the modified discrete-time Riccati equation

 Πd(S)=Q+LT(R+BTSB)L (9)

Then the matrix (A+BK) is stable.

When these conditions are met, the quadratic cost function J2 is given by

J2=xT(0)Sx(0)

Second, we introduce the well known discrete-time bounded real lemma (see Zhou et al., 1996; Iglesias & Glover, 1991; de Souza & Xie, 1992).

Lemma 2.1 (Discrete Time Bounded Real Lemma)

Suppose that γ>0 , M(z)=[ABCD]RH , then the following two statements are equivalent:

i. M(z)<γ .

ii. There exists a stabilizing solution X0 ( X>0 if (C,A) is observable ) to the discrete-time Riccati equation

ATXAX+γ2(ATXB+CTD)U11(BTXA+DTC)+CTC=0

such that U1=Iγ2(DTD+BTXB)>0 .

Next, we will consider the following linear discrete-time systems

 x(k+1)=Ax(k)+B1w(k)+B2u(k)z(k)=C1x(k)+D12u(k) (10)

under the influence of state feedback of the form

 u(k)=Kx(k) (11)

where, x(k)Rn is the state, u(k)Rm is the control input, w(k)Rq is the disturbance input that belongs to L2[0,) , z(k)Rp is the controlled output. Let x(0)=x0 .

The associated with this systems is the quadratic performance index

 J2:=∑k=0∞{xT(k)Qx(k)+uT(k)Ru(k)} (12)

where, Q=QT0 and R=RT>0 .

The closed-loop transfer matrix from the disturbance input w to the controlled output z is

Tzw(z)=[AKBKCK0]:=CK(zIAK)1BK

where, AK:=A+B2K , BK:=B1 , CK:=C1+D12K .

The following lemma is an extension of the discrete-time bounded real lemma ( see Xu 2011).

Lemma 2.2 Given the system (10) under the influence of the state feedback (11), and suppose that γ>0 , Tzw(z)RH ; then there exists an admissible controller K such that Tzw(z)<γ if there exists a stabilizing solution X0 to the discrete time Riccati equation

 AKTX∞AK−X∞+γ−2AKTX∞BKU1−1BKTX∞AK+CKTCK+Q+KTRK=0 (13)

such that U1=Iγ2BKTXBK>0 .

Proof: See the proof of Lemma 2.2 of Xu (2011). Q.E.D.

Finally, we review the result of discrete-time state feedback mixed LQR/ H control problem. Xu (2011) has defined this problem as follows: Given the linear discrete-time systems (10)(11) with w L2[0,) and x(0)=x0 , for a given number γ >0, determine an admissible controller that achieves

supwL2+infK{J2} subject to Tzw(z)<γ .

If this controller K exists, it is said to be a discrete-time state feedback mixed LQR/ H controller.

The following assumptions are imposed on the system

Assumption 2.3 (C1,A) is detectable.

Assumption 2.4 (A,B2) is stabilizable.

Assumption 2.5 D12T[C1D12]=[0I] .

The solution to the problem defined in the above involves the discrete-time Riccati equation

 ATX∞A−X∞−ATX∞B^(B^TX∞B^+R^)−1B^TX∞A+C1TC1+Q=0 (14)

where, B^=[γ1B1B2] , R^=[I00R+I] .

Xu (2011) has provided a solution to discrete-time state feedback mixed LQR/ H control problem, this result is given by the following theorem.

Theorem 2.2 There exists a discrete-time state feedback mixed LQR/ H controller if the discrete-time Riccati equation (14) has a stabilizing solution X and U1= Iγ2B1TXB1>0 .

Moreover, this discrete-time state feedback mixed LQR/ H controller is given by

K=U21B2TU3A

where, U2=R+I+B2TU3B2 , and U3=X+γ2XB1U11B1TX .

In this case, the discrete-time state feedback mixed LQR/ H controller will achieve

supwL2+infK{J2}=x0T(X+γ2XwXz)x0 subject to Tzw<γ .

where, A^K=AK+γ2BKU11BKTXAK , Xw=k=0{(A^Kk)TAKTXBKU12BKTXAKA^Kk} , and Xz=k=0{(A^Kk)TCKTCKA^Kk} .

## 3. State Feedback

In this section, we consider the following linear discrete-time systems

 x(k+1)=Ax(k)+B0w0(k)+B1w(k)+B2u(k)z(k)=C1x(k)+D12u(k)y(k)=C2x(k) (15)

with state feedback of the form

 u(k)=Kx(k) (16)

where, x(k)Rn is the state, u(k)Rm is the control input, w0(k)Rq1 is one disturbance input, w(k)Rq2 is another disturbance that belongs to L2[0,) , z(k)Rp is the controlled output, y(k)Rr is the measured output.

It is assumed that x(0) is Gaussian with mean and covariance given by

E{x(0)}=x¯0cov{x(0),x(0)}:=E{(x(0)x¯0)(x(0)x¯0)T}=R0

The noise process w0(k) is a Gaussain white noise signal with properties

E{w0(k)}=0,E{w0(k)w0T(τ)}=R1(k)δ(kτ)

Furthermore, x(0) and w0(k) are assumed to be independent, w0(k) and w(k) are also assumed to be independent, where, E[] denotes expected value.

Also, we make the following assumptions:

Assumption 3.1 (C1,A) is detectable.

Assumption 3.2 (A,B2) is stabilizable.

Assumption 3.3 D12T[C1D12]=[0I] .

The expected cost function corresponding to this problem is defined as follows:

 JE:=limT→∞1TE{∑k=0T(xT(k)Qx(k)+uT(k)Ru(k)−γ2‖w‖2)} (17)

where, Q=QT0 , R=RT>0 , and γ>0 is a given number.

As is well known, a given controller K is called admissible (for the plant G ) if K is real-rational proper, and the minimal realization of K internally stabilizes the state space realization (15) of G .

Recall that the discrete-time state feedback optimal LQG problem is to find an admissible controller that minimizes the expected quadratic cost function (17) subject to the systems (15) (16) with w(k)=0 , while the discrete-time state feedback H control problem is to find an admissible controller such that Tzw<γ subject to the systems (15) (16) for a given number γ>0 . While we combine the two problems for the systems (15) (16) with w L2[0,) , the expected cost function (17) is a function of the control input u(k) and disturbance input w(k) in the case of γ being fixed and x(0) being Gaussian with known statistics and w0(k) being a Gaussain white noise with known statistics. Thus it is not possible to pose a discrete-time state feedback stochastic mixed LQR/ H control problem that achieves the minimization of the expected cost function (17) subject to Tzw<γ for the systems (15) (16) with wL2[0,) because the expected cost function (17) is an uncertain function depending on disturbance input w(k) . In order to eliminate this difficulty, the design criteria of discrete-time state feedback stochastic mixed LQR/ H control problem should be replaced by the following design criteria:

supwL2+infK{JE} subject to Tzw<γ

because for all wL2[0,) , the following inequality always exists.

infK{JE}supwL2+infK{JE}

Based on this, we define the discrete-time state feedback stochastic mixed LQR/ H control problem as follows:

Discrete-time state feedback stochastic mixed LQR/ H control problem: Given the linear discrete-time systems (15) (16) satisfying Assumption 3.1-3.3 with w(k)L2[0,) and the expected cost functions (17), for a given number γ>0 , find all admissible state feedback controllers K such that

supwL2+{JE} subject to Tzw<γ

where, Tzw(z) is the closed loop transfer matrix from the disturbance input w to the controlled output z .

If all these admissible controllers exist, then one of them K=K will achieve the design criteria

supwL2+infK{JE} subject to Tzw<γ

and it is said to be a central discrete-time state feedback stochastic mixed LQR/ H controller.

Remark 3.1 The discrete-time state feedback stochastic mixed LQR/ H control problem defined in the above is also said to be a discrete-time state feedback combined LQG/ H control problem in general case. When the disturbance input w(k)=0 , this problem reduces to a discrete-time state feedback combined LQG/ H control problem arisen from Bernstein & Haddad (1989) and Haddad et al. (1991).

Remark 3.2 In the case of w(k)=0 , it is easy to show (see Bernstein & Haddad 1989, Haddad et al. 1991) that JE in (17) is equivalent to the expected cost function

JE=limkE{xT(k)Qx(k)+uT(k)Ru(k)}

Define Q=C1TC1 and R=D12TD12 and suppose that C1TD12=0 , then JE may be rewritten as

JE=limkE{xT(k)Qx(k)+uT(k)Ru(k)}=limkE{xT(k)C1TC1x(k)+uT(k)D12TD12u(k)}=limkE{zT(k)z(k)}

Also, the controlled output z may be expressed as

 z=Tzw0(z)w0 (18)

where, Tzw0(z)=[AKB0CK0] . If w0 is white noise with indensity matrix I and the closed-loop systems is stable then

JE=limkE{zT(k)z(k)}=Tzw022

This implies that the discrete-time state feedback combined LQG/ H control problem in the special case of Q=C1TC1 and R=D12TD12 and C1TD12=0 arisen from Bernstein & Haddad (1989) and Haddad et al. (1991) is a mixed H2 / H control problem.

Based on the above definition, we give sufficient conditions for the existence of all admissible state feedback controllers solving the discrete-time stochastic mixed LQR/ H control problem by combining the techniques of Xu (2008 and 2011) with the well known LQG theory. This result is given by the following theorem.

Theorem 3.1 There exists a discrete-time state feedback stochastic mixed LQR/ H controller if the following two conditions hold:

i. There exists a matrix ΔK such that

 ΔK=K+U2−1B2TU3A (19)

and X is a symmetric non-negative definite solution of the following discrete-time Riccati equation

 ATX∞A−X∞−ATX∞B^(B^TX∞B^+R^)−1B^TX∞A+C1TC1+Q+ΔKTU2ΔK=0 (20)

and A^c=AB^(B^TXB^+R^)1B^TXA is stable and U1=Iγ2B1TXB1>0 ;

where, B^=[γ1B1B2] , R^=[I00R+I] , U3=X+γ2XB1U11B1TX , U2=R+I+ B2TU3B2 .

ii. ΔK is an admissible controller error.

In this case, the discrete-time state feedback stochastic mixed LQR/ H controller will achieve

supwL2+{JE}=limT1Tk=0Ttr(B0TXB0R1(k)) subject to Tzw<γ

Remark 3.3 In Theorem 3.1, the controller error is defined to be the state feedback controller K minus the suboptimal controller K=U21B2TU3A , where, X0 satisfies the discrete-time Riccati equation (20), that is,

ΔK=KK

where, ΔK is the controller error, K is the state feedback controller and K is the suboptimal controller. Suppose that there exists a suboptimal controller K such that AK=A+B2K is stable, then K and ΔK is respectively said to be an admissible controller and an admissible controller error if it belongs to the set

Ω:={ΔK:AK+B2ΔK  is stable }

Remark 3.4 The discrete-time state feedback stochastic mixed LQR/ H controller satisfying the conditions i-ii displayed in Theorem 3.1 is not unique. All admissible state feedback controllers satisfying these two conditions lead to all discrete-time state feedback stochastic mixed LQR/ H controllers.

Astrom (1971) has given the mean value of a quadratic form of normal stochastic variables. This result is given by the following lemma.

Lemma 3.1 Let x be normal with mean m and covariance R . Then

E{xTSx}=mTSm+trSR

For convenience, let AK=A+B2K , BK=B1 , CK=C1+D12K , AK=A+B2K , BK=B1 , CK=C1+D12K , and K=U21B2TU3A , where, X0 satisfies the discrete-time Riccati equation (20); then we have the following lemma.

Lemma 3.2 Suppose that the conditions i-ii of Theorem 3.1 hold, then the both AK and AK are stable.

Proof: Suppose that the conditions i-ii of Theorem 3.1 hold, then it can be easily shown by using the similar standard matrix manipulations as in the proof of Theorem 3.1 in de Souza & Xie (1992) that

(B^TXB^+R^)1=[U11+U11B^1U21B^1TU11U11B^1U21U21B^1TU11U21]

where, B^1=γ1B1TXB2 .Thus we have

ATXB^(B^TXB^+R^)1B^TXA=γ2ATXB1U11B1TXA+ATU3B2U21B2TU3A

Rearranging the discrete-time Riccati equation (20), we get

X=ATXA+γ2ATXB1U11B1TXAATU3B2U21B2TU3A+C1TC1+Q+ΔKTU2ΔK=AKTXAK+γ2AKTXBKU11BKTXAK+CKTCK+Q+KTRK+ΔKTU2ΔK

that is,

 AK∗TX∞AK∗−X∞+γ−2AK∗TX∞BK∗U1−1BK∗TX∞AK∗+CK∗TCK∗+Q+K∗TRK∗+ΔKTU2ΔK=0 (21)

Since the discrete-time Riccati equation (20) has a symmetric non-negative definite solution X and A^c=AB^(B^TXB^+R^)1B^TXA is stable, and we can show that A^c=AK+ γ2BKU11BKTXAK , the discrete-time Riccati equation (21) also has a symmetric non-negative definite solution X and AK+γ2BKU11BKTXAK also is stable. Hence, (U11BKTXAK,AK) is detectable. Based on this, it follows from standard results on Lyapunov equations (see Lemma 2.7 a), Iglesias & Glover 1991) that AK is stable. Also, note that ΔK is an admissible controller error, so AK=AK+B2ΔK is stable. Q. E. D.

Proof of Theorem 3.1: Suppose that the conditions i-ii hold, then it follows from Lemma 3.2 that the both AK and AK are stable. This implies that Tzw(z)RH .

Define V(x(k))=xT(k)Xx(k) , where, X is the solution to the discrete-time Riccati equation (20), then taking the difference ΔV(x(k)) , we get

 ΔV(x(k))=xT(k+1)X∞x(k+1)−xT(k)X∞x(k)=xT(k)(AKTX∞AK−X∞)x(k)+2wT(k)BKTX∞AKx(k)+wT(k)BKTX∞BKw(k)+2w0T(k)B0TX∞AKx(k)+2w0T(k)B0TX∞B1w(k)+w0T(k)B0TX∞B0w0(k) (22)

On the other hand, we can rewrite the discrete-time Riccati equation (20) by using the same standard matrix manipulations as in the proof of Lemma 3.2 as follows:

ATXAX+γ2ATXB1U11B1TXAATU3B2U21B2TU3A+C1TC1+Q+ΔKTU2ΔK=0

or equivalently

 AKTX∞AK−X∞+γ−2AKTX∞BKU1−1BKTX∞AK+CKTCK+Q+KTRK=0 (23)

It follows from Lemma 2.2 that Tzw<γ .Completing the squares for (22) and substituting (23) in (22), we get

ΔV(x(k))=z2+γ2w2γ2U112(wγ2U11BKTXAKx)2+xT(k)(AKTXAKX+γ2AKTXBKU11BKTXAK+CKTCK)x(k)+2w0T(k)B0TXAKx(k)+2w0T(k)B0TXB1w(k)+w0T(k)B0TXB0w0(k)=z2+γ2w2γ2U112(wγ2U11BKTXAKx)2xT(k)(Q+KTRK)x(k)+2w0T(k)B0TXAKx(k)+2w0T(k)B0TXB1w(k)+w0T(k)B0TXB0w0(k)

Thus, we have

JE=limT1TE{k=0T(xT(k)Qx(k)+uT(k)Ru(k)γ2w2)}=limT1TE{k=0T(ΔV(x(k))z2γ2U112(wγ2U11BKTXAKx)2+2w0T(k)B0TXAKx(k)+2w0T(k)B0TXB1w(k)+w0T(k)B0TXB0w0(k))}limT1TE{k=0T(ΔV(x(k))+2w0T(k)B0TXAKx(k)+2w0T(k)B0TXB1w(k)+w0T(k)B0TXB0w0(k))}

Note that x()=limTx(T)=0 and

x(k)=AKkx0+i=0k1AKki1B0w0(i)+i=0k1AKki1B1w(i)w0T(k)B0TXAKx(k)=w0T(k)B0TXAKAKkx0+w0T(k)B0TXAKi=0k1AKki1B0w0(i)+w0T(k)B0TXAKi=0k1AKki1B0w(i)

we have

E{k=0Tw0T(k)B0TXAKx(k)}=E{k=0T(w0T(k)B0TXAKi=0k1AKki1B0w0(i))}=k=0Ti=0k1tr{B0TXAKkiB0E(w0(i)w0T(k))}=0

Based on the above, it follows from Lemma 3.1 that

supwL2+{JE}=limT1TE{xT(0)Xx(0)+k=0Tw0T(k)B0TXB0w0(k)}=limT1T{x¯0TXx¯0+tr(XR0)+k=0Ttr(B0TXB0R1(k))}=limT1Tk=0Ttr(B0TXB0R1(k))

Thus, we conclude that

supwL2+{JE}=limT1Tk=0Ttr(B0TXB0R1(k)) subject to Tzw<γ Q.E.D.

In the rest of this section, we give several discussions.

A. A Central Discrete-Time State Feedback Stochastic Mixed LQR/ H Controller

We are to find a central solution to the discrete-time state feedback stochastic mixed LQR/ H control problem.This central solution involves the discrete-time Riccati equation

 ATX∞A−X∞−ATX∞B^(B^TX∞B^+R^)−1B^TX∞A+C1TC1+Q=0 (24)

where, B^=[γ1B1B2] , R^=[I00R+I] . Using the similar argument as in the proof of Theorem 3.1 in Xu (2011), the expected cost function JE can be rewritten as:

 JE=limT→∞1TE{∑k=0T(xT(k)Qx(k)+uT(k)Ru(k)−γ2‖w‖2)}=limT→∞1TE{∑k=0T[−ΔV(x(k))−‖z‖2−γ2‖U112(w−γ−2U1−1BKTX∞AKx)‖2+xT(AKTX∞AK−X∞+γ−2AKTX∞BKU1−1BKTX∞AK+CKTCK+Q+KTRK)x+2w0T(k)B0TX∞AKx(k)+2w0T(k)B0TX∞B1w(k)+w0T(k)B0TX∞B0w0(k)]} (25)

Note that

 AKTX∞AK−X∞+γ−2AKTX∞BKU1−1BKTX∞AK+CKTCK+Q+KTRK=ATX∞A−X∞−ATX∞B^(B^TX∞B^+R^)−1B^TX∞A+C1TC1+Q+(K+U2−1B2TU3A)TU2(K+U2−1B2TU3A) (26)

It follows from (25) and (26) that

 JE=limT→∞1TE{∑k=0T(xT(k)Qx(k)+uT(k)Ru(k))}=limT→∞1TE{∑k=0T[−ΔV(x(k))−‖z‖2+γ2‖w‖2−γ2‖U112(w−γ−2U1−1BKTX∞AKx)‖2+‖U212(K+U2−1B2TU3A)x‖2+2w0T(k)B0TX∞AKx(k)+w0T(k)B0TX∞B0w0(k)]} (27)

If K=U21B2TU3A , then we get that

supwL2+infK{JE}=limT1T{x¯0TXx¯0+tr(XR0)+k=0Ttr(B0TXB0R1(k))}=limT1Tk=0Ttr(B0TXB0R1(k))

by using Lemma 3.1 and the similar argument as in the proof of Theorem 3.1. Thus, we have the following theorem:

Theorem 3.2 There exists a central discrete-time state feedback stochastic mixed LQR/ H controller if the discrete-time Riccati equation (24) has a stabilizing solution X0 and U1=Iγ2B1TXB1>0 .

Moreover, if this condition is met, the central discrete-time state feedback stochastic mixed LQR/ H controller is given by

K=U21B2TU3A

where, U3=X+γ2XB1U11B1TX , U2=R+I+B2TU3B2 .

In this case, the central discrete-time state feedback stochastic mixed LQR/ H controller will achieve

supwL2+infK{JE}=limT1Tk=0Ttr(B0TXB0R1(k)) subject to Tzw<γ

Remark 3.5 When ΔK=0 , Theorem 3.1 reduces to Theorem 3.2.

Remark 3.6 Notice that the condition displayed in Theorem 3.2 is the same as one displayed in Theroem 2.2. This implies that the result given by Theorem 3.2 may be recognied to be a stochastic interpretation of the discrete-time state feedback mixed LQR/ H control problem considered by Xu (2011).

B. Numerical Algorithm

In order to calculate a kind of discrete-time state feedback stochastic mixed LQR/ H controllers, we propose the following numerical algorithm.

Algorithm 3.1

Step 1: Fix the two weighting matrices Q and R , set i=0 , ΔKi=0 , U2(i)=0 and a small scalar δ , and a matrix M which is not zero matrix of appropriate dimensions.

Step 2: Solve the discrete-time Riccati equation

ATXiAXiATXiB^(B^TXiB^+R^)1B^TXiA+C1TC1+Q+ΔKiTU2(i)ΔKi=0

for Xi symmetric non-negative definite such that

A^ci=AB^(B^TXiB^+R^)1B^TXiA is stable and U1(i)=Iγ2B1TXiB1>0 .

Step 3: Calculate U3(i) , U2(i) and Ki by using the following formulas

 U3(i)=Xi+γ−2XiB1U1(i)−1B1TXiU2(i)=R+I+B2TU3(i)B2Ki=−U2(i)−1B2TU3(i)A+ΔKi (28)

Step 4: Let ΔKi+1=ΔKi+δM (or ΔKi+1=ΔKiδM ) and U2(i+1)=U2(i) .

Step 5: If Ai=A+B2Ki is stable, that is, ΔKi is an admissible controller error, then increase i by 1 , goto Step 2; otherwise stop.

Using the above algorithm, we obtain a kind of discrete-time state feedback stochastic mixed LQR/ H controllers as follows:

Ki=U2(i)1B2TU3(i)A±iδM(i=0,1,2,,n,)

C. Comparison with Related Well Known Results

Comparing the result displayed in Theorem 3.1 with the earlier results, such as, Geromel & Peres (1985), Geromel et al. (1989), de Souza & Xie (1992), Kucera & de Souza (1995) and Gadewadikar et al. (2007); we know easily that all these earlier results are given in terms of a single algebraic Riccati equation with a free parameter matrix, plus a free parameter constrained condition on the form of the gain matrix. Although the result displayed in Theorem 3.1 is also given in terms of a single algebraic Riccati equation with a free parameter matrix, plus a free parameter constrained condition on the form of the gain matrix; but the free parameter matrix is also constrained to be an admissible controller error. In order to give some interpretation for this fact, we provided the following result of discrete-time state feedback stochastic mixed LQR/ H control problem by combining directly the proof of Theorem 3.1, and the technique of finding all admissible state feedback controllers by Geromel & Peres (1985) ( also see Geromel et al. 1989, de Souza & Xie 1992, Kucera & de Souza 1995).

Theorem 3.3 There exists a state feedback stochastic mixed LQR/ H controller if there exists a matrix L such that

 L=K+U2−1B2TU3A (29)

and X is a symmetric non-negative definite solution of the following discrete-time Riccati equation

 ATX∞A−X∞−ATX∞B^(B^TX∞B^+R^)−1B^TX∞A+C1TC1+Q+LTU2L=0 (30)

and AK+γ2BKU11BKTXAK is stable and U1=Iγ2B1TXB1>0 .

Where, B^=[γ1B1B2] , R^=[I00R+I] , U3=X+γ2XB1U11B1TX , U2=R+I+ B2TU3B2 .

Note that AK=AK+B2ΔK , K=K+ΔK and

A^c=AB^(B^TXB^+R^)1B^TXA=AK+γ2BKU11BKTXAK

This implies that AK+γ2BKU11BKTXAK is stable if A^c is stable and ΔK is an admissible controller error. Thus we show easily that in the case of ΔK=L , there exists a matrix L such that (29) holds, where, X is a symmetric non-negative definite solution of discrete-time Riccati equation (30) and AK+γ2BKU11BKTXAK is stable if the conditions i-ii of Theorem 3.1 hold.

At the same time, we can show also that if ΔK=L is an admissble controller error, then the calculation of the algotithm 3.1 will become easilier. For an example, for a given admissible controller error ΔKi , the step 2 of algorithm 3.1 is to solve the discrete-time Riccati equation

ATXiAXiATXiB^(B^TXiB^+R^)1B^TXiA+Q^=0

for Xi being a stabilizing solution, where, Q^=C1TC1+Q+ΔKiTU2(i)ΔKi . Since A^ci=A B^(B^TXiB^+R^)1B^TXiA is stable and ΔKi is an admissible controller error, so AK+γ2BKU1(i)1BKTXiAK is stable. This implies the condition ii displayed in Theorem 3.1 makes the calculation of the algorithm 3.1 become easier.

## 4. Static Output Feedback

This section consider discrete-time static output feedback stochastic mixed LQR/ H control problem. This problem is defined as follows:

Discrete-time static ouput feedback stochastic mixed LQR/ H control problem: Consider the system (15) under the influence of static output feedback of the form

u(k)=Fy(k)

with wL2[0,) , for a given number γ>0 , determine an admissible static output feedback controller F such that

supwL2+{JE} subject to Tzw<γ

If this admissible controller exists, it is said to be a discrete-time static output feedback stochastic mixed LQR/ H controller. As is well known, the discrete-time static output feedback stochastic mixed LQR/ H control problem is equivalent to the discrete-time state feedback stochastic mixed LQR/ H control problem for the systems (15) (16), where, K is constrained to have the form of K=FC2 . This problem is also said to be a structural constrained state feedback stochastic mixed LQR/ H control problem.Based the above, we can obtain all solution to discrete-time static output feedback stochastic mixed LQR/ H control problem by using the result of Theorem 3.1 as follows:

Theorem 4.1 There exists a discrete-time static output feedback stochastic mixed LQR/ H controller if the following two conditions hold:

i.There exists a matrix ΔK such that

 ΔK=FC2+U2−1B2TU3A (31)

and X is a symmetric non-negative definite solution of the following discrete-time Riccati equation

 ATX∞A−X∞−ATX∞B^(B^TX∞B^+R^)−1B^TX∞A+C1TC1+Q+ΔKTU2ΔK=0 (32)

and A^c=AB^(B^TXB^+R^)1B^TXA is stable and U1=Iγ2B1TXB1>0 .

Where, B^=[γ1B1B2] , R^=[I00R+I] , U3=X+γ2XB1U11B1TX , U2=R+I+ B2TU3B2 .

ii. ΔK is an admissible controller error.

In this case, the discrete-time static output feedback stochastic mixed LQR/ H controller will achieve

supwL2+{JE}=limT1Tk=0Ttr(B0TXB0R1(k)) subject to Tzw<γ

Remark 4.1 In Theorem 4.1, define a suboptimal controller as K=U21B2TU3A , then ΔK=FC2K . As is discussed in Remark 3.1, suppose that there exists a suboptimal controller K such that AK=A+B2K is stable, then ΔK is an admissible controller error if it belongs to the set:

Ω:= { ΔK:A+B2FC2 is stable}

It should be noted that Theorem 4.1 does not tell us how to calculate a discrete-time static output feedback stochastic mixed LQR/ H controller F . In order to do this, we present, based on the algorithms proposed by Geromel & Peres (1985) and Kucera & de Souza (1995), a numerical algorithm for computing a discrete-time static output feedback stochastic mixed LQR/ H controller F and a solution X to discrete-time Riccati equation (32). This numerical algorithm is given as follows:

Algorithm 4.1

Step 1: Fix the two weighting matrices Q and R , set i=0 , ΔKi=0 , and U2(i)=0 .

Step 2: Solve the discrete-time Riccati equation

ATXiAXiATXiB^(B^TXiB^+R^)1B^TXiA+C1TC1+Q+ΔKiTU2(i)ΔKi=0

for Xi symmetric non-negative definite such that

A^ci=AB^(B^TXiB^+R^)1B^TXiA is stable and U1(i)=Iγ2B1TXiB1>0 .

Step 3: Calculate U3(i+1) , U2(i+1) and ΔKi+1 by using the following formulas

U3(i+1)=Xi+γ2XiB1U1(i)1B1TXiU2(i+1)=R+I+B2TU3(i+1)B2ΔKi+1=U2(i+1)1B2TU3(i+1)A(C2T(C2C2T)1C2I)

Step 4: If ΔKi+1 is an admissible controller error, then increase i by 1 , and goto Step 2; otherwise stop.

If the four sequences X0,X1,,Xi, , U1(1),U1(2),,U1(i), , U2(1),U2(2),,U2(i), , and U3(1),U3(2),,U3(i), converges, say to X , U1 , U2 and U3 , respectively; then the both two conditions displayed in Theorem 4.1 are met. In this case, a discrete-time static output feedback stochastic mixed LQR/ H controllers is parameterized as follows:

F=U21B2TU3AC2T(C2C2T)1

In this chapter, we will not prove the convergence of the above algorithm. This will is another subject.

## 5. Numerical Examples

In this section, we present two examples to illustrate the design methods displayed in Section 3 and 4 respectively.

Example 5.1 Consider the following linear discrete-time system (15) under the influence of state feedback of the form u(k)=Kx(k) , its parameter matrices are

A=[0240.2],B0=[0.10.2],B1=[0.50.3],B2=[10]C1=[1000],C2=[1001],D12=[01]

The above system satisfies Assumption 3.1-3.3, and the open-loop poles of this system are p1= 2.7302 , p2=2.9302 ; thus it is open-loop unstable.

Let R=1 , Q=[1001] , γ=9.5 , δ=0.01 , M=[0.041.2] ; by using algorithm 3.1, we solve the discrete-time Riccati equation (20) to get Xi , U1(i) , Ki ( i=0,1,2,,10) and the corresponding closed-loop poles. The calculating results of algorithm 3.1 are listed in Table 1.

It is shown in Table 1 that when the iteration index i=10 , X100 and U1(10)=0.2927 0 , thus the discrete-time state feedback stochastic mixed LQR/ H controller does not exist in this case. Of course, Table 1 does not list all discrete-time state feedback stochastic mixed LQR/ H controllers because we do not calculating all these controllers by using Algorithm in this example. In order to illustrate further the results, we give the trajectories of state of the system (15) with the state feedback of the form u(k)=Kx(k) for the resulting discrete-time state feedback stochastic mixed LQR/ H controller K= [0.30712.0901] . The resulting closed-loop system is

x(k+1)=(A+B2K)x(k)+B0w0(k)+B1w(k)z(k)=(C1+D12K)x(k)

where, A+B2K=[0.30710.09014.00000.2000] , C1+D12K=[1.000000.30712.0901] .

 IterationIndex i Solution of DARE Xi AdditionalCondition U1(i) State FeedbackController Ki The Closed-Loop Poles 0 X0>0 0.5571 [−0.2886−1.9992] p1=−0.2951, p2=0.2065 1 X1>0 0.5553 [−0.2892−2.0113] p1=−0.1653, p2=0.0761 2 X2>0 0.5496 [−0.2903−2.0237] p1,2=−0.0451 ±j0.1862 3 X3>0 0.5398 [−0.2918−2.0363] p1,2=−0.0459 ±j0.2912 4 X4>0 0.5250 [−0.2940−2.0492] p1,2=−0.0470 ±j0.3686 5 X5>0 0.5040 [−0.2969−2.0624] p1,2=−0.0485 ±j0.4335 6 X6>0 0.4739 [−0.3010−2.0760] p1,2=−0.0505 ±j0.4911 7 X7>0 0.4295 [−0.3071−2.0901] p1,2=−0.0535 ±j0.5440 8 X8>0 0.3578 [−0.3167−2.1049] p1,2=−0.0584 ±j0.5939 9 X9>0 0.2166 [−0.3360−2.1212] p1,2=−0.0680 ±j0.6427 10 X10>0 −0.2927

#### Table 1.

The calculating results of algorithm 3.1.

To determine the mean value function, we take mathematical expectation of the both hand of the above two equations to get

x¯(k+1)=(A+B2K)x¯(k)+B1w(k)z¯(k)=(C1+D12K)x¯(k)

where, E{x(k)}=x¯(k) , E{z(k)}=z¯(k) , E{x(0)}=x¯0 .

Let w(k)=γ2U11B1TX(A+B2K)x¯(k) , then the trajectories of mean values of states of resulting closed-loop system with x¯0=[32]T are given in Fig. 1.

#### Figure 1.

The trajectories of mean values of states of resulting system in Example 5.1.

Example 5.2 Consider the following linear discrete-time system (15) with static output feedback of the form u(k)=Fy(k) , its parameter matrices are as same as Example 5.1.

When C2 is quare and invertible, that is, all state variable are measurable, we may assume without loss of generality that C2=I ; let γ=6.5 , R=1 and Q=[1001] , by solving the discrete-time Riccati equation (24), we get that the central discrete-time state feedback stochastic mixed LQR/ H controller displayed in Theorem 3.2 is

K=[0.37192.0176]

and the poles of resulting closed-loop system are p1=0.1923 , p2=0.0205 .

When C2=[15.4125] , let γ=6.5 , R=1 , Q=[1001] , by using Algorithm 4.1, we solve the discrete-time Riccati equation (32) to get

X=[148.90068.83168.83169.5122]>0,U1=0.0360

Thus the discrete-time static output feedback stochastic mixed LQR/ H controller displayed in Theorem 4.1 is F=0.3727 . The resulting closed-loop system is

x(k+1)=(A+B2FC2)x(k)+B0w0(k)+B1w(k)z(k)=(C1+D12FC2)x(k)

where, A+B2FC2=[0.37270.01744.00000.2000] , C1+D12FC2=[1.000000.37272.0174] .

Taking mathematical expectation of the both hand of the above two equations to get

x¯(k+1)=(A+B2FC2)x¯(k)+B1w(k)z(k)=(C1+D12FC2)x¯(k)

where, E{x(k)}=x¯(k) , E{z(k)}=z¯(k) , E{x(0)}=x¯0 .

Let w(k)=γ2U11B1TX(A+B2FC2)x¯(k) , then the trajectories of mean values of states of resulting closed-loop system with x¯0=[12]T are given in Fig. 2.

#### Figure 2.

The trajectories of mean values of states of resulting system in Example 5.2.

## 6. Conclusion

In this chapter, we provide a characterization of all state feedback controllers for solving the discrete-time stochastic mixed LQR/ H control problem for linear discrete-time systems by the technique of Xu (2008 and 2011) with the well known LQG theory. Sufficient conditions for the existence of all state feedback controllers solving the discrete-time stochastic mixed LQR/ H control problem are given in terms of a single algebraic Riccati equation with a free parameter matrix, plus two constrained conditions: One is a free parameter matrix constrained condition on the form of the gain matrix, another is an assumption that the free parameter matrix is a free admissible controller error. Also, a numerical algorithm for calculating a kind of discrete-time state feedback stochastic mixed LQR/ H controllers are proposed. As one special case, the central discrete-time state feedback stochastic mixed LQR./ H controller is given in terms of an algebraic Riccati equation. This provides an interpretation of discrete-time state feedback mixed LQR/ H control problem. As another special case, sufficient conditions for the existence of all static output feedback controllers solving the discrete-time stochastic mixed LQR/ H control problem are given. A numerical algorithm for calculating a static output feedback stochastic mixed LQR/ H controller is also presented.

## References

1 - K. J. Astrom, 1970 Introduction to stochastic control theory Academic Press INC
2 - M. Athans, 1971 The role and use of thr stochastic linear-quadratic-Gaussian problem in control system design IEEE Trans. Aut. Control 16 6 529 552
3 - T. Basar, P. Bernhard, 1991H-optimal control and related minmax design problems: a dynamic approach, Boston, MA: Birkhauser.
4 - D. S. Bernstein, W. M. Haddad, 1989 LQG control with an H performance bound: A Riccati equation approach IEEE Trans. Aut. Control 34 3 293 305
5 - X. Chen, K. Zhou, 2001 Multiobjective H2/H control design SIAM J. Control Optim. 40 2 628 660
6 - C. E. de Souza, L. Xie, 1992 On the discrete-time bounded real lemma with application in the characterization of static state feedback H controllers Systems & Control Letters 18 61 71
7 - J. C. Doyle, K. Glover, P. P. Khargonekar, B. A. Francis, 1989a State-space solutions to standard H2 and H control problems IEEE Trans. Aut. Control 34 8 831 847
8 - J. C. Doyle, K. Zhou, B. Bodenheimer, 1989b Optimal control with mixed H2 and H performance objectives Proceedings of 1989 American Control Conference Pittsb-urh, PA 2065 2070
9 - J. C. Doyle, K. Zhou, K. Glover, B. Bodenheimer, 1994 Mixed H2 and H perfor-mance objectives II: optimal control IEEE Trans. Aut. Control 39 8 1575 1587
10 - K. Furata, S. Phoojaruenchanachai, 1990 An algebraic approach to discrete-time H control problems Proceedings of 1990 American Control Conference San Diego 2067 3072
11 - J. Gadewadikar, F. L. Lewis, L. Xie, V. Kucera, M. Abu-Khalaf, 2007 Parameterization of all stabilizing H static state-feedback gains: application to output-feedback design Automatica 43 1597 1604
12 - J. C. Geromel, P. L. D. Peres, 1985 Decentrailised load-frequency control IEE Proceedings 132 5 225 230
13 - J. C. Geromel, A. Yamakami, V. A. Armentano, 1989 Structrual constrained controllers for discrete-time linear systems Journal of Optimization and Applications 61 1 73 94
14 - W. M. Haddad, D. S. Bernstein, D. Mustafa, 1991 Mixed-norm H2/H regulation and estimation: the discrete-time case Systems & Control Letters 16 235 247
15 - P. A. Iglesias, K. Glover, 1991 State-space approach to discrete-time H control INT. J. Control 54 5 1031 1073
16 - P. P. Khargonekar, M. A. Rotea, 1991 Mixed H2/H control: A convex optimization approach IEEE Trans. Aut. Control 36 7 824 837
17 - V. Kucera, C. E. de Souza, 1995 A necessary and sufficient condition for output feedback stabilizability Automatica 31 9 1357 1359
18 - H. Kwakernaak, 2002 H2-optimization-theory and application to robust control design Annual Reviews in Control 26 45 56
19 - D. J. N. Limebeer, B. D. O. Anderson, P. P. Khargonekar, M. Green, 1992 A game theoretic approach to H control for time-varying systems SIAM J. Control and Optimization 30 2 262 283
20 - D. J. N. Limebeer, B. D. O. Anderson, B. Hendel, 1994 A Nash game approach to mixed H2/H control IEEE Trans. Aut. Control 39 1 69 82
21 - E. Tse, 1971 On the optimal control of stochastic linear systems IEEE Trans. Aut. Control 16 6 776 785
22 - X. Xu, 1996 A study on robust control for discrete-time systems with uncertainty A Master Thesis of 1995 Kobe university, Kobe, Japan, January, 1996
23 - X. Xu, 2008 Characterization of all static state feedback mixed LQR/H controllers for linear continuous-time systems Proceedings of the 27th Chinese Control Conference Kunming, Yunnan, China 678 682 July 16-18, 2008
24 - X. Xu, 2011 Discrete time mixed LQR/H control problems Discrete Time Systems Mario Alberto Jordan (Ed.) 978-9-53307-200-5 InTech, Available from http://www.intechopen.com/
25 - H. Yeh, S. S. Banda, B. C. Chang, 1992 Necessary and sufficient conditions for mixed H2 and H optimal control IEEE Trans. Aut. Control 37 3 355 358
26 - K. Zhou, J. C. Doyle, K. Glover, 1996 Robust and optimal control Prentice-Hall, INC.