On Common Linear/Quadratic Lyapunov Functions for Switched Linear Systems

Chapter 25 On Common Linear/Quadratic Lyapunov Functions for Switched Linear Systems

Melania M. Moldovan and M. Seetharama Gowda

Nonlinear Analysis and Variational Problems (pp. 415–429) eds.: P. Pardalos, Th. M. Rassias, and A. A. Khan;c 2009 Springer

Dedicated to the memory of Professor George Isac

Abstract Using duality and complementarity ideas, and Z-transformations, in this article, we discuss equivalent ways of describing the existence of common linear/quadratic Lyapunov functions for switched linear systems. In particular, we extend a recent result of Mason-Shorten on positive switched system with two con- stituent linear time-invariant systems to an arbitrary ﬁnite system.

25.1 Introduction

n×n Given a ﬁnite set of matrices {A1,A2,...,Am} in R , the dynamical system

x˙+ Aσ x = 0, σ ∈{1,2,...,m}, (25.1) where the switching signal σ is a piecewise constant function from [0,∞) to {1,2,...,m}, is called a switched (continuous) linear system. Because of numerous applications of such systems, these and their variants have been well studied in the literature, see e.g., the monograph [16] and recent survey article [27]. Given a signal σ, a solution to (25.1) is a continuous and piecewise continuously differ- entiable function x(t) that satisﬁesx ˙ + Aσ x(t)=0 for all t ∈ [0,∞) except at the switching instances of σ. For (uniform exponential) asymptotic stability of (25.1)

Melania M. Moldovan Department of Mathematics & Statistics, University of Maryland, Baltimore County, Baltimore, Maryland 21250, USA, e-mail: [email protected] M. Seetharama Gowda Department of Mathematics & Statistics, University of Maryland, Baltimore County, Baltimore, Maryland 21250, USA, e-mail: [email protected]

415 416 Melania M. Moldovan and M. Seetharama Gowda corresponding to arbitrary signals, which means that there exist numbers M ≥ 1 and β > 0 such that −β ||x(t)|| ≤ Me t ||x(0)|| for all t ≥ 0, for all solutions x(t), and all signals, a sufﬁcient condition is the existence of a common quadratic Lyapunov function (CQLF) V(x) := xT Px, where P is a symmetric matrix satisfying the conditions

T + = , ,..., . P 0 and Ai P PAi 0 for all i 1 2 m (25.2)

The existence of such a P has been studied by numerous authors under various sufﬁcient conditions, such as commutativity [22], simultaneous triangulariza- tion [21], Lie algebraic conditions [17], etc. See also, [25], [26], [15], [24]. Similar to the continuous case, there is a switched (discrete) linear system given by x(k + 1)+Aσ x(k)=0, σ ∈{1,2,...,m}, k = 1,2,.... (25.3) These systems have also been well studied, see e.g., [19]. As in the continuous case, the stability can be studied by constructing a CQLF V(x) := xT Px, where P is a symmetric matrix satisfying the conditions

− T = , ,..., . P 0 and P Ai PAi 0 for all i 1 2 m (25.4)

n If we restrict the dynamics of the system (25.1) to the nonnegative orthant R+ of Rn, we get a positive switched system [7]. Here we require any trajectory of (25.1) n n with a starting point in R+ to remain in R+. This condition requires all matrices Ai to be Z-matrices (which are matrices with nonpositive off-diagonal entries). In this setting, a common linear Lyapunov function V(x)=xT d can be constructed, where d is a vector in Rn satisfying the conditions

> T > = , ,..., . d 0 and Ai d 0 for all i 1 2 m (25.5) Alternatively, for the same positive switched system, a common quadratic copositive Lyapunov function V(x) := xT Px can be constructed, where P is a symmetric matrix satisfying the conditions

T + n = , ,..., . P and Ai P PAi are strictly copositive on R+ for all i 1 2 m (25.6)

n T (Here, a matrix A is strictly copositive on R+ means that x Ax > 0 for all 0 = x ∈ n R+.) By considering the cone of (symmetric) positive semideﬁnite matrices and transformation L(X)=AT X + XA, various authors have used theorems of alternative (or duality theory) to formulate conditions equivalent to (25.2). For example, a result of Kamenetskiy and Pyatnit- skiy [14] says that condition (25.2) holds if and only if there do not exist positive semideﬁnite matrices Yi (i = 1,2,...,m) with at least one Yi nonzero such that 25 On Common Linear/Quadratic Lyapunov Functions for Switched Linear Systems 417

m ( + T )= . ∑ AiYi YiAi 0 i=1 It is possible to write similar equivalent conditions for (25.4), (25.5), and (25.6). Our first objective in this paper is to present a unified result that covers all these equivalent conditions. This result was motivated by the observation that the underlying transformations in (25.2), (25.4), (25.5), and (25.6), namely, the Lyapunov transformation LA defined on the cone of positive semidefinite matrices, the Stein transformation SA on the cone of positive semidefinite matrices, the matrix A on n the cone of nonnegative orthant in R , and the Lyapunov transformation LA (corresponding to a Z-matrix) on the cone of symmetric copositive matrices have the following common property:

∗ x ∈ K, y ∈ K , and, x,y = 0 ⇒L(x),y ≤0, where K denotes a cone with dual K∗. Transformations satisfying the above property are called Z-transformations. By using the properties of Z-transformations [12], we formulate our unifying condition, see Theorem 25.8. Our second objective is to demonstrate the relevance of complementarity problems in the study of common linear/quadratic Lyapunov functions. Specifically, we show how a complementarity result of Song, Gowda, and Ravindran [28] can be used to extend a recent result of Mason-Shorten on positive switched systems [20]. Here is an outline of this paper. In Section 25.2, we recall necessary matrix theory concepts and describe essential properties of Z-transformations. Section 25.3 deals with complementarity ideas. Section 25.4 deals with duality ideas and our unifying result covering the existence of linear/quadratic Lyapunov functions. Finally, in Section 25.5 we present our extension of the Mason-Shorten result. A word about our notation. Throughout this paper, we will use standard matrix theory and complementarity theory notation. For example, we will use Z-matrices and positive stable matrices (instead of Metzler matrices and Hurwitz matrices), work with positive definite matrices (instead of negative definite matrices), etc. Also, our (dynamical) systems are written in the formx ˙+ Ax = 0 instead ofx ˙ = Ax.

25.2 Preliminaries

25.2.1 Matrix Theory Concepts

The space Rn carries the usual inner product which is written as either x,y or xT y. n n n ◦ In R , we denote the nonnegative orthant by R+ and its interior by (R+) . We write n n ◦ d ≥ 0 when d ∈ R+ and d > 0 when d ∈ (R+) . A matrix A ∈ Rn×n is said to be 418 Melania M. Moldovan and M. Seetharama Gowda

• a Z-matrix if all its off-diagonal entries are nonpositive. The negative of a Z- matrix is called a Metzler matrix; • a P-matrix if all its principal minors are positive; • a copositive matrix (strictly copositive matrix) if xT Ax ≥ 0(> 0) for all 0 = x ≥ 0; • an S-matrix if there is a d > 0 such that Ad > 0; • positive stable if the real part of any eigenvalue of A is positive; • Schur stable if the absolute value of any eigenvalue of A is less than one; • completely positive if there is a nonnegative (rectangular) matrix B with A = BBT , or equivalently, A is a sum of matrices of the form xxT with x ≥ 0. Consider a nonempty set C of matrices in Rn×n. A matrix A ∈ Rn×n is called a column representative of C if for every j = 1,2,...,n, the jth column of A is the jth column of some matrix in C . Similarly, row representatives of C are deﬁned. n×n For A1,A2,...,Am,D1,...,Dm ∈ R , where each Di is a diagonal matrix, the following holds [29]: m det(∑AiDi)=∑det(K)det(E), (25.7) 1 where K is a column representative of {A1,A2,...,Am}, E is a column representative of {D1,...,Dm}, and the same indexed columns are selected to form K and E. Here the summation is over all column representatives of {A1,...,Am}.

25.2.2 Z-Transformations

Throughout this paper, H denotes a ﬁnite dimensional real Hilbert space. We reserve the symbol K for a proper cone in H, that is, K is a closed convex cone in H such that K ∩ (−K)={0} and K − K = H. We denote the dual of K by K∗ and the interior by K◦. A linear transformation L : H → H is said to be a Z-transformation (or said to have the Z-property) on K if

∗ x ∈ K, y ∈ K , x,y = 0 ⇒L(x),y ≤0. (25.8)

n n When H = R and K = R+,aZ-transformation is nothing but a Z-matrix. We remark that a Z-transformation is the negative of a “cross-positive” transformation introduced in [23]. It is known [5] that L has the Z-property on K if and only if any trajectory of the dynamical systemx ˙+ L(x)=0 which starts in K stays in K. The above concept is illustrated in the following examples. Examples 25.1 and 25.2 given below appear in [12]. Example 25.3 is new.

Example 25.1. Let H = S n, the space of all n × n real symmetric matrices with inner product given by X,Y = trace(XY), where the trace of a matrix is the sum 25 On Common Linear/Quadratic Lyapunov Functions for Switched Linear Systems 419

n of its diagonal elements (or the sum of its eigenvalues). Let K = S+, the cone of positive semideﬁnite matrices in S n. We use the notation n n ◦ X 0 when X ∈ S+ and X 0 when X ∈ (S+) . For any A ∈ Rn×n, let 1 L (X)= (AX + XAT ) and S (X)=X − AXAT A 2 A denote the Lyapunov and Stein transformations corresponding to A. We verify that n n LA and SA are Z-transformations on S+: Since S+ is self-dual and

X 0, Y 0, and X,Y = 0 ⇒ XY = 0,

T LA(X),Y = trace(AXY)=0&SA(X),Y = trace(XY) −trace(AXA Y) ≤ 0, (25.9) where the last inequality comes from the facts that AXAT 0 (when X 0) and the n inner product of any two elements in S+ is nonnegative. Thus, both LA and SA are n Z-transformations on S+. Example 25.2. As in Example 25.1, let H = S n with X,Y = trace(XY). Let

n n K = {X ∈ S : X is copositive on R+}.

Then K is a proper with dual

∗ K = {X ∈ S n : X is completely positive},

∈ = ∑N T ∈ ∗ ∈ n see [3], Thm. 2.3. Given X K and Y i=1 yiyi K (where yi R+ for all i) and X,Y = 0, we claim that XY is a nonnegative matrix with zero diagonal. To see this, ∑N T = ∑N ( T )= , = . ﬁrst observe that i yi Xyi i trace Xyiyi X Y 0 As X is copositive, we T = ≥ get yi Xyi 0 for all i; because X is copositive and symmetric, we have Xyi 0. 1 ( + )T ( + ) ≥ ≥ (This follows from limt↓0 t tx yi X tx yi 0 for all x 0.) From this, we see = ∑N T ( )= that XY i Xyiyi is a nonnegative matrix. As trace XY 0, XY must have zero diagonal. Now suppose A ∈ Rn×n. We claim that

LA is a Z-transformation on K if and only if A is a Z-matrix.

To see this, suppose that LA is a Z-transformation on K. We show that the (1,2) and (2,1) entries of A are nonpositive. Let e1,e2,...,en denote the standard unit coordinate vectors in Rn. Let

=[ ], = T , = T , X xij Y e1e1 and W e2e2 ∗ where all entries of X are zero except x12 = x21 = 1. We see that X ∈ K, Y,W ∈ K , and X,Y = 0 = X,W . A simple computation shows that

trace(LA(X)Y)=a12 and trace(LA(X)W)=a21. 420 Melania M. Moldovan and M. Seetharama Gowda

By the Z-property of LA, we must have a12 ≤ 0 and a21 ≤ 0. A similar argument will prove that other off-diagonal entries are also nonpositive. Hence A is a Z-matrix. Now conversely, suppose that A is a Z-matrix. Let X ∈ K, Y ∈ K∗, with X,Y = 0. Then XY and YX are nonnegative matrices with zero diagonals. Then

trace(LA(X)Y)=trace(AXY) ≤ 0.

This proves that when A is a Z-matrix, LA has the Z-property on K. Example 25.3. Let H be a Euclidean Jordan algebra with inner product ·,· and Jordan product x◦y [6]. Let K denote the cone of squares {x◦x : x ∈ H}. Then K is a self-dual (proper) cone. Examples of Euclidean Jordan algebras include Rn with the usual inner product and componentwise product (as the Jordan product), the Jordan spin algebra L n (n > 1) whose underlying space is R × R(n−1) with the usual inner n product and (x0,x) ◦ (y0,y)=(x0y0 + x,y ,x0y + y0x), and matrix algebras S of all n×n real symmetric matrices, H n of all n ×n complex Hermitian matrices, Qn of all n × n quaternion Hermitian matrices, and O3 of all 3 × 3 octonion Hermitian matrices. In the matrix algebras, the Jordan and inner product are given respectively ◦ = 1 ( + ) , = ( ) by X Y : 2 XY YX and X Y : Retrace XY . In any Euclidean Jordan algebra, for any element a, the so-called Lyapunov transformation La is a Z-transformation on the cone of squares, where,

La(x)=a ◦ x.

We now recall some properties of Z-transformations [12].

Theorem 25.4. Suppose L is a Z-transformation on a proper cone K in a Hilbert space H. Then the following are equivalent: (1) There exists a d ∈ K◦ such that L(d) ∈ K◦. (2) L is invertible with L−1(K◦) ⊆ K◦. (3) L is positive stable, that is, the real part of any eigenvalue of L is positive. (4) L +tI is invertible for all t ∈ [0,∞). (5) All real eigenvalues of L are positive. (6) There is an e ∈ (K∗)◦ such that LT (e) ∈ (K∗)◦. n n Moreover, when H = R and K = R+, the above properties (for a Z-matrix) are further equivalent to (7) LisaP-matrix.

Remark 25.5. When A ∈ Rn×n is a Z-matrix, the positive stability of A can be de- scribed in more than 50 equivalent ways, see [2]. In particular, we have the equivalence of the following for a Z-matrix: (1) A is positive stable. (2) A is a P-matrix. (3) There exists a d > 0 such that Ad > 0 (or AT d > 0). (4) There exists a (diagonal) D 0inS n such that AD + DAT 0. 25 On Common Linear/Quadratic Lyapunov Functions for Switched Linear Systems 421

(5) LA is positive stable. 1 (λ + μ) (Regarding item (5), we note that the eigenvalues of LA are of the form 2 , where λ and μ are eigenvalues of A.) We now add one more condition to this list: A Z-matrix A is positive stable if and only if there exists a strictly copositive matrix C such that AC +CAT is strictly copositive. This follows from the above theorem applied to L = LA on the cone of symmetric copositive matrices.

25.3 Complementarity Ideas

Suppose K is a self-dual cone in H. For any two elements x,y ∈ H, let

x y = x − ΠK(x − y), where ΠK denotes the (orthogonal) projection of H onto K. It is known, see [11], that is a commutative, but (in general) non-associative binary operation on H.In n n the case of H = R and K = R+, we use the standard notation x ∧y in place of x y; we observe that x ∧ y = min{x,y} and that ∧ is associative. Given a matrix-tuple A =(A1,A2,...,Am) of matrices Ai n×n n in R and vector-tuple q =(q1,q2,...,qm) of vectors qi in R , the vertical linear complementarity problem [10] VLCP(A,q) is to ﬁnd a vector x ∈ Rn such that

x ∧ (A1x + q1) ∧···∧(Amx + qm)=0.

n ◦ If −qi ∈ (R+) , then a solution to the above problem satisﬁes x ≥ 0 and Aix ≥−qi > 0 for all i; hence a small perturbation of such an x produces a vector d > 0 such that Aid > 0 for all i.

Theorem 25.6. ([10], Section 6.3) The VLCP(A,q) has a unique solution for all q if and only if every row representative of the set {A1,A2,...,Am} is a P-matrix.

When m = 1, the VLCP reduces to the well known linear complementarity problem [4] that has been extensively studied in the optimization literature. Now coming to the general case of a self-dual K, it can be easily verified that x y ∈ K ⇒ x,y ∈ K. Thus, for a finite set of linear transformations Li : H → H and ◦ vectors −ei ∈ K ,ifanx ∈ H satisfies the complementarity problem

x f (x)=0, where f (x) is a combination of Li(x) − ei, i = 1,2,...,m under the binary operation ◦ ◦ in some order, then we can assert the existence of a d ∈ K such that Li(d) ∈ K . n n In particular, when H = S and K = S+ and Li = L T , we can relate the existence Ai 422 Melania M. Moldovan and M. Seetharama Gowda of a common quadratic Lyapunov function to a solution of a complementarity problem. We refer the reader to [9] and [12] for results highlighting this connection. See also, Section 25.5.

25.4 Duality Ideas

In this section, we present our result that generalizes the result of Kamenetskiy and Pyatnitskiy [14] and at the same time uniﬁes several similar results. We begin with a theorem of alternative.

Theorem 25.7. (Theorem of Alternative [2], Page 9) Let H1 and H2 be Hilbert spaces with proper cones K1 ⊆ H1 and K2 ⊆ H2. For a linear transformation L : H1 → H2, consider the following systems: ◦ ◦ (i) L(x) ∈ (K2) , x ∈ (K1) , T ( ) ∈ ∗, = ∈− ∗. (ii) L y K1 0 y K2 (iii) L(x) ∈ K2, 0 = x ∈ K1, T ( ) ∈ ( ∗)◦, ∈−( ∗)◦. (iv)L y K1 y K2 Then, exactly one of the systems (i) and (ii) is consistent and exactly one of the systems (iii) and (iv) is consistent. Theorem 25.8. Let H be a ﬁnite dimensional real Hilbert space and K be a proper cone in H. Let Li : H → H be linear for i = 1,2,...,m, where L1 is positive stable and has the Z-property on K. Consider the following statements: ◦ ◦ (1) There exists a d ∈ K such that Li(d) ∈ K for all i = 1,...,m. ∈ ∗, = ,..., ∑m T ( )= (2) There do not exist xi K i 1 m with some xi nonzero such that 1 Li xi 0. (3) There exists a nonzero d ∈ K such that Li(d) ∈ K for all i = 1,...,m. ∈ ( ∗)◦, = ,..., ∑m T ( )= . (4) There do not exist xi K i 1 m such that 1 Li xi 0 Then (1) ⇔ (2) and (3) ⇔ (4). ∗ Proof. Assume that (1) holds but not (2). Then there are x1,x2,...,xm in K with at ∑m T ( )= least one xi nonzero such that 1 Li xi 0. Then

m m = T ( ), = , ( ) > 0 ∑Li xi d ∑ xi Li d 0 1 1 as xi,Li(d) ≥0 for all i = 1,2,3,...,m and xi,Li(d) > 0 when xi = 0. This contra- diction proves that (1) ⇒ (2). Now assume the negation of (1). Since L1 is positive stable and has the Z-property on K, it is invertible, see Theorem 25.4; hence (as m = 1) we can deﬁne L : H → Hm−1 by ( ) =( −1( ),..., −1( )). L x : L2L1 x LmL1 x 25 On Common Linear/Quadratic Lyapunov Functions for Switched Linear Systems 423

We now apply Theorem 25.7 with K1 := K in H and K2 := K ×...×K (m−1 times) in Hm−1. We claim that condition (i) of Theorem 25.7 cannot hold. Assuming the ◦ ◦ ◦ ◦ contrary, there is an x ∈ K such that L(x) ∈ (K2) = K ×···×K . This implies ( −1( )) ∈ ◦ = , ,..., that Li L1 x K for all i 2 3 m. Let = −1( ). d : L1 x −1( ◦) ⊆ ◦ ∈ ◦ ( )= ∈ ◦. Since L1 K K (by Theorem 25.4), d K and L1 d x K From ( −1( )) ∈ ◦ ( ) ∈ ◦ = , ,..., Li L1 x K , we get Li d K for i 2 3 m. This cannot happen as we have assumed the negation of (1). So condition (i) of Theorem 25.7 fails. Therefore, by condition (ii) of Theorem 25.7, we get a y such that

∗ ∗ ∗ 0 = y ∈−(K ×···×K ) and LT (y) ∈ K .

Now let ∗ ∗ −y =(x2,...,xm) ∈ K × ... × K . Then for any u ∈ H we have , T ( ) = ( ), = − ( −1( ),..., −1( )),( ,..., ) u L y L u y L2L1 u LmL1 u x2 xm m m m = − −1( ), = − , −T T ( ) = ,− −T T ( ) . ∑ LiL1 u xi ∑ u L1 Li xi u ∑ L1 Li xi i=2 i=2 i=2 Hence m = T ( )=− −T T ( ). x1 : L y ∑ L1 Li xi i=2 T Applying L1 to both sides of the above equation, we get T ( )+ T ( )+...+ T ( )= . L1 x1 L2 x2 Lm xm 0 ∗ As x1 ∈ K and (x2,x3,...,xm) is nonzero, we get the negation of (2). Hence ∼ (1) ⇒∼ (2),or(2) ⇒ (1). This completes the proof of the equivalence (1) ⇔ (2). The proof of (3) ⇔ (4) is similar, but one has to use conditions (iii) and (iv) of The- orem 25.7.

We illustrate the theorem by the following examples.

n n Example 25.9. Let H = S , K = S+, and

1 L (X)= (A X + XAT )(X ∈ S n), Ai 2 i i ∈ n×n ( = , ,..., ) ( )T = where Ai R i 1 2 m are positive stable. As LA and LA LAT are both n n×n Z-transformations on S+ for any A ∈ R , the above theorem gives the equivalence of the following statements [14]: 424 Melania M. Moldovan and M. Seetharama Gowda

n T (i) There exists D 0inS such that A D + DAi = 2L T (D) 0 for all i = i Ai 1,2,...,m. (ii) There do not exist Yi 0 (i = 1,2,...,m) with at least one nonzero Yi such that ∑m ( + T )= . i=1 AiYi YiAi 0 We note that the matrix D in item (i) produces a common quadratic Lyapunov function V(x) := xT Dx for the switched linear systemx ˙+ Aσ x = 0.

= S n = S n ( )= − T ( ∈ S n), Example 25.10. Let H , K +, and SAi X X AiXAi X ∈ n×n ( = , ,..., ) ( )T = where Ai R i 1 2 m are Schur stable. Now, SA and SA SAT are n n×n Z-transformations on S+ for any A ∈ R , see Example 25.1. Also, the Schur stability of Ai implies that SAi is positive stable for each i, see e.g., Theorem 11 in [8] and Theorem 25.4. Hence the above theorem gives the equivalence of the following statements: S n − T = , ,..., (i) There exists D 0in such that D Ai DAi 0 for all i 1 2 m. (ii) There do not exist Yi 0 (i = 1,2,...,m) with at least one nonzero Yi such that ∑m ( − T )= . i=1 Yi AiYiAi 0 The matrix D in item (i) produces a common quadratic Lyapunov function V(x) := xT Dx for the discrete switched system x(k + 1)+Aσ x(k)=0.

n n n×n Example 25.11. Let H = R , K = R+, and Ai ∈ R (i = 1,2,...,m) be positive stable Z-matrices. Then the following are equivalent: > n T > = , ,..., (i) There exists d 0inR such that Ai d 0 for all i 1 2 m. (ii) There do not exist vectors yi ≥ 0 (i = 1,2,...,m) with at least one nonzero yi ∑m = . such that i=1 Aiyi 0 In this setting, the vector d in item (i) produces a common linear Lyapunov function v(x) := xT d for the positive switched systemx ˙+ Aσ x = 0.

Example 25.12. Let H = S n, K be the cone of symmetric copositive matrices and n×n Ai ∈ R for i = 1,2,...,m be positive stable Z-matrices. Then the following are equivalent: S n T + (i) There exists a strictly copositive matrix C in such that Ai C CAi is strictly copositive for all i = 1,2,...,m. (ii) There do not exist completely positive matrices Yi (i = 1,2,...,m) with at least ∑m ( + T )= . one nonzero Yi such that i=1 AiYi YiAi 0 Here, the matrix C in item (i) produces a common copositive quadratic Lyapunov function v(x) := xTCx for the positive switched systemx ˙+ Aσ x = 0.

We remark that if condition (i) of Example 25.11 holds, then C := ddT satisﬁes condition (i) of Example 25.12. However, the converse is false, see Remark 25.20 in Section 25.5. 25 On Common Linear/Quadratic Lyapunov Functions for Switched Linear Systems 425 25.5 Positive Switched Linear Systems

In a recent paper [20], Mason and Shorten prove that for two positive stable Z- matrices A1 and A2, the following are equivalent: > n T > T > (i) There exists a vector d 0inR such that A1 d 0 and A2 d 0. (ii) Every column representative of {A1,A2} is positive stable. (iii) Every column representative of {A1,A2} has positive determinant. In an earlier work, based on complementarity ideas, Song, Gowda, and Ravindran [28] proved the equivalence of the following for any compact set A of positive stable Z-matrices in Rn×n: (1) There exists a d > 0 such that AT d > 0 for all A ∈ A . (2) Every column representative of A is a P-matrix. In this section, we extend the above Mason-Shorten result to any compact set of matrices and at the same time improve the result of Song, Gowda, and Ravindran by requiring that only the determinant be positive for every column representative of A . We recall that for a Z-matrix (see Remark 25.5), P, S, and positive stability properties are equivalent. We begin with some lemmas. n×n Lemma 25.13. Suppose A = {A1,A2,...,Am}⊆R with det(C) > 0 for any column representative C of A . Then for any set of positive diagonal matrices {D1,...,Dm} we have det(A1D1 + ... + AmDm) > 0. Proof. The result follows from the identity (25.7).

Lemma 25.14. Suppose A1,A2,...,Am are positive stable Z-matrices and det(C) > 0 for any column representative C of {A1,A2,...,Am}. Then there exists 0 = u ≥ 0 such T ≥ = ,..., . that Ai u 0 for all i 1 m Proof. If the stated conclusion fails, then by the equivalence of items (3) and (4) in Theorem 25.8 (see Example 25.11), there exist yi > 0(i = 1,...,m) such that

A1y1 + A2y2 + ...+ Amym = 0.

If Di denotes the diagonal matrix with diagonal yi, then

(A1D1 + A2D2 + ...+ AmDm)e = 0, where e is the transpose of the vector (1,1,...,1) in Rn. This means that

det(A1D1 + A2D2 + ...+ AmDm)=0.

However, Lemma 25.13 together with the hypothesis shows that this is not possible. 426 Melania M. Moldovan and M. Seetharama Gowda

Lemma 25.15. Suppose that A1,A2,...,Am are positive stable Z-matrices and det(C) > 0 for any column representative C of {A1,A2,...,Am}. Then there exists d > 0 such T > = ,..., . that Ai d 0 for all i 1 m Proof. Let E be the matrix in Rn×n with every entry 1. We can choose a small positive ε such that Ai − εE is a P ∩ Z-matrix for all i and any column representative of {A1 −εE,A2 −εE,...,Am −εE} has positive determinant. An application of the ≥ ( T − ε ) ≥ previous lemma produces a nonzero u 0 such that Ai E u 0 for all i. This = ≥ T > > implies that 0 u 0 and Ai u 0 for all i; by continuity, there exists a d 0 such T > that Ai d 0 for all i.

n×n For a set A = {A1,A2,...,Am} in R , let A # := set of all column representatives of A and m m A := ∑AiDi : Di is a nonnegative diagonal matrix and ∑Di = I . 1 1

We note that A ⊆ A # ⊆ A and that Ais convex.

Theorem 25.16. Let A = {A1,A2,...,Am} be a set of positive stable Z-matrices in Rn×n. Then the following are equivalent: > T > = , ,..., (1) There exists a d 0 such that Ai d 0 for all i 1 2 m. (2) There exists a d > 0 such that CT d > 0 for all C ∈ A . (3) Every matrix in A is positive stable (equivalently, a P-matrix). (4) Every matrix in A # is positive stable (equivalently, a P-matrix). (5) Every matrix in A # has positive determinant. (6) Every matrix in A has positive determinant.

= ∑m ∈ A T = ∑m T > Proof. If (1) holds, then for any matrix C 1 AiDi , C d 1 DiAi d 0by the imposed conditions on Di. Thus condition (2) holds. Now each matrix C in A is a Z-matrix. The implication (2) ⇒ (3) follows from Remark 25.5. The implications (3) ⇒ (4) ⇒ (5) are obvious. The implication (5) ⇒ (6) follows from the identity (25.7). Finally, the implication (6) ⇒ (5) is obvious and the implication (5) ⇒ (1) follows from Lemma 25.15. This completes the proof of the theorem.

Several remarks are in order. 25 On Common Linear/Quadratic Lyapunov Functions for Switched Linear Systems 427

Remark 25.17. Thanks to Theorem 2, we can relate item (4) above to the uniqueness of solution in the vertical linear complementarity problem VLCP(AT ,q) for any q AT =( T , T ,..., T ). vector-tuple . Here, A1 A2 Am Item (5) is related to the so-called column W -property and horizontal linear complementarity problems, see [29].

Remark 25.18. A conjecture due to Mason and Shorten [18] and D. Angeli says that for a set A = {A1,A2,...,Am} of positive stable Z-matrices, the following are equivalent: (a) Every matrix in the convex hull of A is positive stable. (b) The switched systemx ˙+ Aσ x = 0 is (uniformly) asymptotically stable for arbitrary signals.

In [13], Gurvits, Shorten, and Mason disprove this conjecture by constructing a 2×2 set {A1,A2} of positive stable Z-matrices in R for which item (a) holds but not (b). We note here that if we replace (a) by (a)Every matrix in A # is positive stable (or equivalently, every matrix in the convex set Ais positive stable), then (a) ⇒ (b). This raises the question whether (b) ⇒ (a). Based on the work of Akar et al [1], we show that this is false.

Example 25.19. Consider the following positive stable Z-matrices

1 −1 − 1 = = 1 2 . A1 − 1 and A2 − 2 1 11

It is easily veriﬁed that every convex combination of A1 and A2 is positive stable. In this context (because every diagonal entry of A1 and A2 is one), by a result of Akar et al [1], the positive systemx ˙+ Aσ x = 0 is (uniformly) asymptotically stable for arbitrary signals. Thus condition (b) above holds. However, the column representative formed by the ﬁrst column of A2 and the second column of A1 is not a P-matrix (that is, it is not positive stable). Hence (a) fails to hold.

Remark 25.20. It is well known that a Z-matrix A is positive stable if and only if there is a diagonal matrix D 0 such that AD + DAT 0. Now for a ﬁnite set A = {A1,A2,...,Am} of Z-matrices, consider the following statements: (i) Every row and column representative of A is positive stable. + T = , ..., . (ii) There exists a diagonal matrix D 0 such AiD DAi 0 for all i 1 2 m We claim that (i)⇒ (ii) but not conversely. To see this, assume that (i) holds, in which case, by the above theorem, there exist vectors u > 0 and v > 0inRn such that > T > ( = , ,..., ). Aiu 0 and Ai v 0 i 1 2 m Let D be a diagonal matrix with Dv = u. Then D has positive diagonal entries and

( + T ) = + ( T ) > AiD DAi v Aiu D Ai v 0 428 Melania M. Moldovan and M. Seetharama Gowda

Z + T P for all i. This means that for each i, the symmetric -matrix AiD DAi is a - + T matrix, hence AiD DAi 0. Thus item (ii) holds. To see that (ii) need not imply (i), we merely construct two 2 × 2 symmetric positive deﬁnite Z-matrices {A1,A2} such that a column representative of {A1,A2} is not a P-matrix. Then with D = I, item (ii) holds but not (i).

Theorem 25.21. Let A be a compact set of positive stable Z-matrices in Rn×n. Then the following are equivalent: (1) There exists a d > 0 such that AT d > 0 for all A ∈ A . (2) Every column representative of A is a P-matrix. (3) Every column representative of A has positive determinant.

Proof. The equivalence between (1) and (2) was proved by Song, Gowda, and Ravindran in [28]. The implication (2)⇒(3) is obvious and (3)⇒(2) follows from the previous theorem applied to a ﬁnite set of matrices.

References

1. M. Akar, A. Paul, M.G. Safonov, and U. Mitra, Conditions on the stability of a class of second-order switched systems, IEEE Transactions on Automatic Control, 51 (2006) 338- 340. 2. A. Berman and R.J. Plemmons, Nonnegative matrices in the mathematical sciences, SIAM, Philadelphia, PA, 1994. 3. A. Berman and N. Shaked-Monderer, Completely positive matrices, World Scientiﬁc, New Jersey, 2003. 4. R.W. Cottle, J.S. Pang, and R.E. Stone, The linear complementarity problem, Academic Press, Boston, 1992. 5. L. Elsner, Quasimonotonie und Ungleischungen in halbgeordneten Raumen,¨ Linear Algebra and its Applications, 8 (1974) 249-261. 6. J. Faraut and A. Koranyi, Analysis on symmetric cones, Oxford University Press, Oxford 1994. 7. L. Farina and S. Rinaldi, Positive linear systems: Theory and applications, Wiley, New York 2000. 8. M.S. Gowda and T. Parthasarathy, Complementarity forms of theorems of Lyapunov and Stein, and related results, Linear Algebra and its Applications, 320 (2000) 131-144. 9. M.S. Gowda and Y. Song, Some new results for the semideﬁnite linear complementarity problems, SIAM Journal on Matrix Analysis, 24 (2002) 25-39. 10. M.S. Gowda and R. Sznajder, The generalized order linear complementarity problem, SIAM Journal on Matrix Analysis, 15 (1994) 779-795. 11. M.S. Gowda, R. Sznajder, and J. Tao, Some P-properties for linear transformations on Eu- clidean Jordan algebras, Linear Algebra and its Applications, 393 (2004) 203-232. 12. M.S. Gowda and J. Tao, Z-transformations on proper and symmetric cones, Mathematical Programming, Series B, 117 (2009) 195-221. 13. L Gurvits, R. Shorten, and O. Mason, On the stability of switched positive linear systems, IEEE Transactions on Automatic Control, 52 (2007) 1099-1103. 14. V.A. Kamenetskiy and Ye.S. Pyatnitskiy, An iterative method of Lyapunov function con- struction for differential inclusions, Systems and Control Letters, 8 (1987) 445-451. 25 On Common Linear/Quadratic Lyapunov Functions for Switched Linear Systems 429

15. C. King and M. Nathanson, On the existence of a common quadratic Lyapunov function for a rank one difference, Linear Algebra and its Applications, 419 (2006) 400-416. 16. D. Liberzon, Switching in systems and control, Birkhauser,¨ Boston, 2003. 17. D. Liberzon, J.P. Hespanha, and A.S. Morse, Stability of switched systems: a Lie-algebraic condition, Systems and Control Letters, 37 (1999) 117-122. 18. M. Mason and R. Shorten, A conjecture on the existence of common quadratic Lyapunov functions for positive linear systems, Proceedings of American Control Conference, 2003. 19. M. Mason and R. Shorten, On common quadratic Lyapunov functions for stable discrete-time LTI systems, IMA Journal of Applied Mathematics, 69 (2004) 271-283. 20. M. Mason and R. Shorten, On linear copositive Lyapunov functions and the stability of positive switched linear systems, IEEE Transactions on Automatic Control, 52 (2007) 1346-1349. 21. Y. Mori, T. Mori, and Y. Kuroe, A solution to the common Lyapunov function problem for continuous-time systems, Decision and Control, 4 (1997) 3530-3531. 22. K.S. Narendra and J. Balakrishnan, A common Lyapunov function for stable LTI systems with commuting A-matrices, Automatic Control, IEEE Transactions, 39 (1994) 2469-2471. 23. H. Schneider and M. Vidyasagar, Cross-positive matrices, SIAM Journal on Numerical Anal- ysis, 7 (1970) 508-519. 24. R. Shorten, O. Mason, and C. King, An alternative proof of the Barker, Berman, Plemmons (BBP) result on diagonal stability and extensions, Linear Algebra and its Applications, 430 (2009) 34-40. 25. R.N. Shorten and K.S. Narendra, On the stability and existence of common Lyapunov functions for stable linear switching systems, Decision and Control, 4 (1998) 3723-3724. 26. R.N. Shorten and K.S. Narendra, Necessary and sufﬁcient conditions for the existence of a common quadratic Lyapunov function for M stable second order linear time-invariant systems, American Control Conference, Chicago, 2000. 27. R. Shorten, F. Wirth, O. Mason, K. Wulff, and C. King, Stability criteria for switched and hybrid systems, SIAM Review, 49 (2007) 545-592. 28. Y. Song, M.S. Gowda, and G. Ravindran, On some properties of P-matrix sets, Linear Alge- bra and its Applications, 290 (1999) 237-246. 29. R. Sznajder and M.S. Gowda, Generalizations of P0- and P-properties; extended vertical and horizontal linear complementarity problems, Linear Algebra and its Applications, 223/224 (1995) 695-715.