Block-Diagonal Solutions to Lyapunov Inequalities and Generalisations of Diagonal Dominance

Transcription

1 Block-Diagonal Solutions to Lyapunov Inequalities and Generalisations of Diagonal Dominance Aivar Sootla, Yang Zheng and Antonis Papachristodoulou Abstract Diagonally dominant matrices have many applications in systems and control theory Linear dynamical systems with scaled diagonally dominant drift matrices, which include stable positive systems, allow for scalable stability analysis algorithms For example, it is known that Lyapunov inequalities for this class of systems admit diagonal solutions In this paper, we present an extension of scaled diagonally dominance to block partitioned matrices We show that our definition describes matrices admitting block-diagonal solutions to Lyapunov inequalities and that these solutions can be computed using linear algebraic tools We also show how in some cases the Lyapunov inequalities can be decouplento a set of lower dimensional linear matrix inequalities, thus leading to improved scalability We conclude by illustrating some advantages and limitations of our results on numerical examples I INTRODUCTION Systems admitting diagonal matrix solutions to Lyapunov inequalities are of particular interest in control theory, since they allow for a lower computational complexity of stability analysis Necessary and sufficient conditions for the existence of diagonal solutions were deriven [1], which are, however, hard to check On the other hand, it is wellknown that stable linear systems invariant on the positive orthant (or positive systems) admit diagonal matrix solutions to Lyapunov inequalities [2] Therefore, generalisations of positivity attracted some attention as well, eg eventual positivity [3], [4], which inherits some properties of positivity However, in the context of Lyapunov inequalities, perhaps, a more relevant generalisation is based on (scaled) diagonally dominant matrices These are defined through linear constraints on the absolute values of the individual entries of the matrix Under some conditions scaled diagonally dominant drift matrices admit diagonal solutions to Lyapunov inequalities [5], which can be computed using linear programming [6] Instead of considering individual entries of the matrix, one can partition the matrix into blocks and apply the diagonal dominance conditions to some norms of these blocks as in [7] Although some authors considered block versions of scaled diagonal dominance [6], [8], [9], construction of block-diagonal solutions to Lyapunov inequalities was not fully addressed In this paper, we present a novel block generalisation of scaled diagonal dominance In comparison to generalisations in [7], [9], our definition appears to be more suitable for stability analysis, since it includes a control theoretic concept of the norm The authors are with the Department of Engineering Science, University of Oxford, Parks Road, Oxford, OX1 3PJ, UK aivarsootla, yangzheng, antonis}@engoxacuk The authors would like to thank Dr James Anderson for valuable discussions It is fairly computationally cheap to check if a matrix satisfies our definition of scaled diagonal dominance, which is a useful property for stability analysis of large-scale systems We also show that the introduced class of matrices admits block-diagonal solutions to Lyapunov inequalities Furthermore, we decouple the Lyapunov inequality into a number of linear matrix inequalities of smaller dimensions, which naturally leads to reduced memory requirements and computational complexity To summarise, we can check in advance (by performing a set of cheap computations), if a matrix admits a block-diagonal solution to a Lyapunov inequality Furthermore, we can find these solutions using linear algebra or a set of smaller linear matrix inequality constraints The rest of the paper is organised as follows We cover the preliminaries of positive systems theory, scaled diagonal dominance and some facts from systems theory in Section II We introduce our extension of scaled diagonal dominance to block partitioned matrices in Subsection III-A We show how block-diagonal solutions to Lyapunov inequalities are constructen Subsection III-B, and present decoupled stability tests based on our results in Subsection III-C We illustrate our results on numerical examples in Section IV and conclude in Section V The proofs of some auxiliary results are founn the Appendix Notation Let S k + (respectively, S k ++) denote the set of k k positive semidefinite (respectively, positive definite) matrices We also write A 0 if A S k +, and A 0 if A S k ++ We denote the positive orthant R n >0, that is, the set of all vectors x with positive entries The operator denotes a matrix transpose We denote the maximal singular value of a matrix A as σ(a), while the minimal as σ(a) The norm of an asymptotically stable transfer function G(s) is computed as G H = max w R G(ıω) 2, where ı is the imaginary unit and A 2 is the induced matrix norm equal to σ(a) Finally, let diag A 1,, A n } denote a block-diagonal matrix with matrices A i on the block-diagonal II PRELIMINARIES AND PROBLEM FORMULATION Consider the linear time invariant system ẋ(t) = Ax(t) + Bu(t), y(t) = Cx(t), with the transfer function G(s) = C(sI A) 1 B, where x(0) = x 0, x(t) R n, u(t) R m, y(t) R k It can be shown that system (1) is stable [10] with u(t) = 0 for all t if and only if there exists P 0 such that (1) P A + A P 0 (2)

2 The linear matrix inequality (LMI) (2) is called a Lyapunov inequality, ants solution defines a Lyapunov function of the form V (x) = x(t) P x(t) for system (1) with u(t) = 0 We will also use the following result from control theory called the Bounded Real Lemma [10] Proposition 1: For system (1), there exists γ such that γ > G H if and only if there exists P 0 such that P A + A P + P BB P + C Cγ 2 0 (3) In some cases, we can guarantee the existence of a diagonal matrix P satisfying (2) One of such cases is the class of dynamical systems with Metzler drift matrices (or positive systems) Definition 1: A matrix A R n n is said to be Metzler if all the off-diagonal elements are nonnegative Analysis of positive systems is computationally and conceptually simpler than analysis of general types of systems For example, the following result (which is a combination of results in [11], [12], [13]), allows to replace semidefinite constraints in analysis and design methods with linear ones, which leads to scalable algorithms Proposition 2: Consider a system ẋ = Ax with a Metzler drift matrix A Then the following statements are equivalent i) There exists d R n >0 such that Ad R n >0; ii) There exists e R n >0 such that e A R n >0; iii) A is Hurwitz (has eigenvalues with negative real parts) iv) There exists a diagonal P such that P A + A P 0 The points (i) and (ii) in Proposition 2 imply that Hurwitz Metzler matrices belong to another well-known class of matrices Definition 2: A matrix A R n n with entries a ij is called strictly row scaled diagonally dominant if there exist positive scalars d 1,, d n such that: a ii > d j a ij i = 1,, n (4) The matrix A is called strictly column scaled diagonally dominant if there exist positive scalars e 1,, e n such that: e i a ii > e j a ji i = 1,, n (5) The matrix A is strictly row (respectively, column) diagonally dominant if = 1 (respectively, e i = 1) for all i In order to illustrate the connection between diagonal dominance and positivity we introduce the following concept Definition 3: Let the matrix M(A), with entries M ij (A) defined as max aii, 0} if i = j, M ij (A) = (6) a ij otherwise, be called the comparison matrix We slightly modified the definition of the comparison matrix compared to the classic one (cf [5]) in order to streamline the stability analysis For example, if A is Metzler or A is lower triangular (a ij = 0, for all i < j, and a ii < 0) then A is Hurwitz if and only if M(A) is Hurwitz More generally, if M(A) is Hurwitz, then A is Hurwitz and A admits a diagonal solution to (2) [5], which can be constructed using linear algebra, linear or second order cone programming [6] In the proof of these results Proposition 2 is applied to a Hurwitz Metzler matrix M(A), which leads to existence of positive, e i such that (4) and (5) hold, that is A is strictly row and column scaled diagonally dominant We, finally, note that the matrices with Hurwitz M(A) belong to a well-studied class of matrices called H matrices We will not discuss in detail this class of matrices, but refer the reader to [5], [8] for details In this paper, we discuss a generalisation of scaled diagonal dominance to block partitioned matrices We say that a matrix A R N N has α = k 1,, k n }-partitioning with N = n k i, if the matrix A is written as follows i=1 A 11 A 12 A 1n A 21 A 22 A 2n A = A n1 A n2 A nn where A ij R ki kj We say that A is α-diagonal if it is α-partitioned and A ij = 0 if i j We aim at characterising α-diagonally stable matrices A R N N, which are such that there exists an α-diagonal positive definite X R N N satisfying (2) If the partitioning is trivial, ie α = 1,, 1} = 1, we will not mention α and say that an α-diagonal (respectively, α-diagonally stable) matrix A is diagonal (respectively, diagonally stable) We will also use a version of the Gershgorin circle theorem for the α-partitioned matrices Proposition 3 ([7]): For an α-partitioned matrix A R N N, where α = k 1,, k n } and N = n i=1 k i, every eigenvalue of A satisfies (λi A ii ) A ij 2 for at least one i where i = 1,, n III GENERALISATIONS OF DIAGONAL DOMINANCE A α-comparison Matrix ants Properties We start by introducing a novel generalisation of the comparison matrix M(A) to the block partitioned case Definition 4: Given an α-partitioned matrix A, we define the matrix M α (A) as follows M α (si Aii ) ij(a) = 1 1 if i = j, (7) A ij 2 otherwise If A ii is not Hurwitz, then we can continuously extend the function so that = 0 Therefore, Definition 4 is well-posed In [7], [9], a similar definition of M α (A) is used, but is 2 and for stability analysis it is required replaced by A 1 ii 1 that A ii are Metzler and Hurwitz Since (si A ii ) 1 H = A 1 2 for Hurwitz Metzler matrices and we do not have any restrictions on A ii besides stability, our definition appears to be better suited for stability analysis It is tempting

3 to call the set of matrices such that the comparison matrices M α (A) are Hurwitz as block scaled diagonally dominant or block-h matrix similarly to [7], [9] There are, however, several definitions of block-h matrices ann order to minimise confusion we will resist of introducing new nomenclature Now we will discuss the properties of our extension If M α (A) is Hurwitz then according to Proposition 2 there exist positive scalars, e i such that for all i = 1,, n: > e i > σ(a ij )d j, (8) σ(a ji )e j (9) In the trivial partitioning case, ie, α = 1,, 1} = 1, we have (si a ii ) 1 1 = max a ii, 0} and a ii is Hurwitz if and only if it is negative Therefore, the α-partitioned generalisation of the comparison matrix reduces to our previous definition Furthermore, stability of the matrix M 1 (A) ensures stability of the matrix A In the α-partitioned case, a similar statement can be made Proposition 4: An α-partitioned matrix A is Hurwitz, if M α (A) is Hurwitz The proof can be founn the Appendix In what follows we will show that stability of M α (A) implies a stronger property of A, namely, α-diagonal stability, a result which carries over from the case α = 1 However, some of the properties of scaled diagonally dominant matrices are not preserven our generalisation We have the following result with the proof in the Appendix Proposition 5: There exists a matrix A and a partitioning α such that M α (A) is a Hurwitz matrix, however, for any α-diagonal positive definite P the matrix M α (A P + P A) is not a Hurwitz matrix Proposition 5 shows the contrast to the trivial partitioning case (ie, α = 1) If M 1 (A) is Hurwitz, then there exists a diagonal P such that M 1 (P A + A P ) is Hurwitz This seems as a minor obstacle, however, in the case α = 1 this property was used to compute diagonal solutions to the Lyapunov inequalities Therefore, in what follows we need to find another way to compute α-diagonal P B Computation of α-diagonal Lyapunov Matrices We start by considering the following auxiliary result Proposition 6: Let M α (A) be a Hurwitz matrix, then there exist γ ij R n 0, W ij S ki +, P i S ki ++ such that P i A ii + A iip i + γ ii I ki + W ii 0, Wij P i A ij A ij P 0, i γ ij I kj γ ii > γ ji, W ii W ij (10) Proof The proof is constructive and we find explicitly P i, W ij and γ ij satisfying LMIs (10) Let A be α-partitioned and let M α (A) be a Hurwitz matrix, which implies that there exists positive scalars, e i such that (8), (9) hold Let also A i, i = ( A i,1 A i,i 1 A i,i+1 A i,n ) Γ i = diag ( γ i,1 I γ i,i 1 I γ i,i+1 I γ i,n I )} H where γ ij = e i / if i = j σ(a ij )e i /d j otherwise (11) The scalars γ ij are equal to zero, if A ij is Therefore, we introduce matrices Γ i and Ãi, i, which are obtained by removing all zero blocks from Γ i and A i, i Let J i = j [1,, n] γ ij 0, j i} We have that Ãi, i Γ 1/2 i j J i σ 2 2 = σ A ij A ij/γ ij j Ji ( Aij A ij) /γij = j J i (σ(a ij )) 2 /γ ij = σ(a ij )d j /e i < /e i (12) Using inequality (12) we can obtain further bounds (si A ii ) 1 à i, i Γ 1/2 i 2 (si A ii ) 1 2 Ãi, i Γ 1/2 i 2 2 < (si A ii ) 1 H /e i = γ 1 ii This according to Proposition 1 implies that for some P i 0 P i A ii + A iip i + γ ii I + P i à i, i Γ 1 i à i, ip i 0 By noticing that γ ii > n γ ji anntroducing new variables W ij P i A ij A ij P i/γ ij, we obtain (10) It is noticeable, that there is a certain asymmetry in the seemingly related variables W ij, γ ij in (10), in the sense that one is a matrix, while the other is a scalar Actually, if the main goal in mins stability analysis, we can relax the conditions (10) and consider the following LMIs with P i S ki ++, W ij, V ji S ki + P i A ii + A iip i + V ii + W ii 0, Wij P i A ij A ij P 0, i V ij V ii V ji, W ii W ij, (13a) (13b) (13c) which leads to the main theoretical result of the paper Theorem 1: Let an α-partitioned A satisfy (13) for some P i S ki ++, W ij, V ji S ki +, then the matrix A is α- diagonally stable Furthermore, P A + A P 0 with P = diag P 1,, P n } 0 Proof Let R ij = R ii R jj for i j, where Rii R ki N partitionento blocks of the size k i k j for all j = 1,, n All the blocks are zero matrices, except for the k i s block entry, which is an identity matrix We have

4 the following decomposition: P A + A P = ( R ii (P i A ii + A iip i + V ii + W ii )Rii i=1 R ii (V ii + W ii i=1 (W ij + V ji ))R ii ) Wij P R i A ij ij A ij P i V ij R ij n It is straightforward to show that i=1 R ii( V ii + n V ji)rii, n i=1 R ii( W ii + n W ij)rii are negative definite, while other sums are negative semidefinite, therefore P A + A P 0 This result also implies that a matrix A is α-diagonally stable provided that M α (A) is Hurwitz The class of matrices satisfying LMIs (13) can be seen as a generalisation of matrices with a stable M(A) in their own right Let α = 1 and A = a ij } n i,j=1, then constraints (13) are simplified to a ii w ii + v ii wij v ij, a ij, 2p i p i (14) w ii > w ij, v ii > v ji Proposition 7: The matrix A satisfies (14) if and only if M(A) is Hurwitz Proof If M(A) is Hurwitz then according to Proposition 2 it is diagonally stable and as a result a ii are negative Furthermore, there exist positive e i, such that a ii > n a ij d j and a ii e i > n a ji e j Now we can set w ij = a ij e i d j /d 2 i, v ij = a ij e i /d j, p i = e i / and verify that A satisfies (14) Now let (14) be fulfilled Consider a strictly column diagonally dominant (and hence Hurwitz) Metzler matrix v ii i = j V = v ij i j This implies that there exists positive scalars such that v ii d 2 i > n v ijd 2 j according to Proposition 2 For any positive scalars x, y we have xy (x + y)/2, therefore 1 2p i a ij d j / vij (d j / ) 2 w ij v ij (d j / ) 2 + w ij < 1 2p i (w ii + v ii ) a ii, which shows that M(A) is strictly row scaled diagonally dominant and hence Hurwitz C Decoupled Stability Tests According to Theorem 1, we can test stability of an α = k 1,, k n }-partitioned matrix A using LMIs (13) These LMIs provide only a sufficient condition for stability, but they are decentralisen the sense that the semidefinite constraints p i are of orders k i, and we do not need to impose a semidefinite constraint of order n i=1 k i We can fully decouple the stability tests by setting, for example, V ij = γ ij I kj for some fixed γ ij, while eliminating W ij using the Schur complement formula We get P i A ii + A iip i + P i A ij A ij/γ ij P i + I ki (ε i + γ ji ) = 0, i = 1,, n (15) where ε i are positive predefined scalars The choice of the gains γ ij is essential and we present a few ad-hoc choices Test A The equations (15) have solutions P i 0 with γ ij = σ(a ij ) Test B The equations (15) have solutions P i 0 with 1 σ(a ij ) > 0, γ ij = 0 σ(a ij ) = 0 Test C The matrix M α (A) is Hurwitz, that is there exist positive scalars e i, such that (8), (9) hold, which implies that the equations (15) have solutions P i 0 with γ ij = σ(a ij )e i /d j The sets of matrices satisfying these stability tests intersect, but none of them includes the other It is possible to find matrices, which satisfy only one of tests and fail two others This cannot be done in the trivial partitioning case (ie, α = 1,, 1}), but in the two block case with α = 2, 2} we present such examples in what follows IV NUMERICAL EXAMPLES Example 1 First, we present an example verifying the result of Proposition 5 Let δ = 163, Q = blkdiagq 1, Q 1 } B δi A =, B =, Q δi B = 7 11 With α = 2, 2}, the M α (A) is Hurwitz, and the matrix QA + A Q is negative definite We can verify if there exists P = blkdiagp 1, P 2 } 0 such that M α (P A + A P ) is a Hurwitz matrix using LMIs In particular, it can be shown that (si P A A P ) 1 1 = σ(p A + A P ), hence we have the following matrix inequalities: P i A ii + A iip i γ ii I ki, i = 1, 2 ( ) γ 12 I k1 P 1 A 12 + A 21P 2 P 2 A 21 + A 0, 12P 1 γ 12 I k2 γ11 γ B = 12 0, γ 12 γ 22 where B 0 is equivalent to B being Hurwitz for symmetric matrix B Numerical computations show that there exists no P = blkdiagp 1, P 2 } 0 such that M α (P A + A P ) is a Hurwitz matrix But if we set δ = 16, then it is straightforward to check that M α (QA + A Q) is a Hurwitz matrix

5 A = B = C = (16) Example 2 Consider matrices A, B and C in (16) It can be verified that the matrix A satisfies Test A and fails Tests B,C, the matrix B satisfies Test B and fails Tests A,C and finally the matrix C satisfies Test C, while fails Tests A,B Test B uses binary information about the interconnections (if they exist or not), while Test A uses also the information on the gains of the interconnections Therefore, it may seem counter-intuitive that Test A sometimes fails when Test B prevails, since in Test A we use more information about the system than in Test B In our examples, if the gain σ(a 12 ) is larger than one, then we make the Riccati equation for i = 1 less conservative by normalising A 12 However, we make the other Riccati equation (with i = 2) more conservative by increasing the term γ 12 I, which requires to make the norm of the system (si A 22 ) 1 A 21 smaller Therefore, for large gains A ij either of the tests can perform better depending on the drift matrices A ii The main conservatism of Test C is in the transition to the Riccati equations We use the Cauchy-Schwartz inequality, which may be conservative if the eigenspaces of the matrices A ii, A ij are not aligned In control-theoretic language, this corresponds to the mode of A ii closest to the imaginary axis being poorly controllable using the input matrix A ij On the other hand, we also scale the gains σ(a ij ), which provides extra freedom In our examples, these limitations of the tests are not apparent, which indicates that even the slightest changes in the gains σ(a ij ) and the eigenspaces of A ij can result in the failure of one of the completely decoupled tests Example 3 We proceed with a rather theoretical observation It is well-known that an α-triangular matrix A is Hurwitz if and only if the blocks on α-diagonal are Hurwitz, which also implies that it is α-diagonally stable We will consider this class of matrices through our generalisation of scaled diagonal dominance on a specific example Let A = While setting α = 2, 3, 1}, compute the comparison matrix M α (A) = , which is Hurwitz as long as elements on the diagonal are negative It can be verified that the generalisation of the scaled diagonal dominant matrices from [7], [6] will not yield conclusive results on stability analysis of the matrix A Our definition, however, confirms that stability of the blocks A ii (and hence stability of M α (A)) is necessary and sufficient for α-diagonal stability of an α-triangular matrix A Example 4 Now we consider another class of matrices called border block diagonal [14] Consider the following α-partitioned matrix A 11 A 12 A 13 A 1n A 21 A A = A A n1 0 0 A nn If the matrix A satisfies (13) then the conditions are simplified since A ij = 0 unless i = j, i = 1 or j = 1 We can set directly V ij = 0, W ij = 0 if A ij = 0, furthermore, we sum LMIs in (13b) containing W 1i and W i1 for every i, after rearranging the LMIs to fit the dimensions After these operations, it can be shown that conditions (13) imply the following LMIs Q 1 A 11 + A 11Q 1 + j>1 Y j 0 Q j A jj + A jjq j + Z j 0 ( Y j Q 1 A 1j A j1 P ) j A 1j Q 0, 1 Q j A j1 Z j (17) where we set Y j = W 1j + V j1, Z j = W j1 + V 1j These conditions can be obtained directly from the LMI QA + A Q 0 provided that Q = diag Q 1,, Q n } 0 using standard decomposition techniques (cf [15]) Therefore, conditions (17) are necessary and sufficient for α-diagonal stability of border block diagonal matrices Conditions (17) are less restrictive than stability of M α (A) and conditions (13) applied to the matrix A, at the same time one can view conditions (17) as conditions (13) applied to QA + A Q with P i = I, and W ij = V ji V CONCLUSION In this paper, we presented a generalisation of scaled diagonal dominance for block partitioned matrices Our main goal was to provide conditions on the drift matrix, which facilitate the stability analysis of large-scale systems, in the spirit of positive systems theory In particular, we derived sufficient conditions for existence of block-diagonal solutions to Lyapunov inequalities We also decoupled the Lyapunov inequality into a set of linear matrix inequalities of lower dimensions, which can potentially be used for distributed stability analysis We note that our results are related to work on contraction theory (cf [16]), input-tostate stability of networked systems (cf [17]) or rather its linear predecessors [14], and chordal sparsity for Lyapunov inequalities [18] We will consider these connections in future work Furthermore, our results can be applied to decentralised control problems as indicaten [19]

6 REFERENCES [1] D Carlson, D Hershkowitz, and D Shasha, Block diagonal semistability factors and lyapunov semistability of block triangular matrices, Linear Algebra ants Applications, vol 172, pp 1 25, 1992 [2] A Berman and R J Plemmons, Nonnegative Matrices in the Mathematical Sciences SIAM, 1994, vol 9 [3] C Altafini, Representing externally positive systems through minimal eventually positive realizations, in Proc IEEE Conf Decision Control, 2015, pp [4] A Sootla, Properties of Eventually Positive Linear Input-Output Systems, ArXiv e-prints arxiv: , Sept 2015 [5] D Hershkowitz and H Schneider, Lyapunov diagonal semistability of real H-matrices, Linear algebra ants applications, vol 71, pp , 1985 [6] A Sootla and J Anderson, On existence of solutions to structured Lyapunov inequalities, in Proc of Am Control Conf IEEE, 2016, pp [7] D G Feingold, R S Varga, et al, Block diagonally dominant matrices and generalizations of the gerschgorin circle theorem, Pacific J Math, vol 12, no 4, pp , 1962 [8] B Polman, Incomplete blockwise factorizations of (block) H- matrices, Linear Algebra Appl, vol 90, pp , 1987 [9] S-h Xiang and Z-y You, Weak block diagonally dominant matrices, weak block H-matrix and their applications, Linear Algebra Appl, vol 282, no 1, pp , 1998 [10] K Zhou, J C Doyle, and K Glover, Robust and optimal control Prentice Hall New Jersey, 1996 [11] K Fan, Topological proofs for certain theorems on matrices with non-negative elements, Monatshefte für Mathematik, vol 62, no 3, pp , 1958 [12] R S Varga, On recurring theorems on diagonal dominance, Linear Algebra ants Applications, vol 13, no 1, pp 1 9, 1976 [13] A Rantzer, Scalable control of positive systems, European Journal of Control, vol 24, pp 72 80, 2015 [14] D D Šiljak, Large-scale dynamic systems: stability and structure North Holland, 1978, vol 2 [15] Y Zheng, G Fantuzzi, A Papachristodoulou, P Goulart, and A Wynn, Fast ADMM for semidefinite programs with chordal sparsity, arxiv preprint arxiv: , 2016 [16] G Russo, M Di Bernardo, and E D Sontag, A contraction approach to the hierarchical analysis and design of networked systems, IEEE Trans Autom Control, vol 58, no 5, pp , 2013 [17] S N Dashkovskiy, B S Rüffer, and F R Wirth, Small gain theorems for large scale systems and construction of iss lyapunov functions, SIAM J Control Optim, vol 48, no 6, pp , 2010 [18] R P Mason and A Papachristodoulou, Chordal sparsity, decomposing sdps and the lyapunov equation, in Proc of Am Control Conf IEEE, 2014, pp [19] Y Zheng, M Kamgarpour, A Sootla, and A Papachristodoulou, Convex design of structured controllers using block-diagonal Lyapunov functions In submission to Automatica, 2017 APPENDIX Proof of Proposition 4: We note that the proof of the following result employs the technique usen [7] We prove the result by contradiction Let A have eigenvalues with a nonnegative real part and let M α (A) be Hurwitz, which implies there exists positive scalars such that (8) holds for every i Since A has eigenvalues with a nonnegative real part, then so does the matrix D 1 AD is with D = diag d 1 I k1,, d n I kn } Let λ be the eigenvalue of D 1 AD with a nonnegative real part By Proposition 3 there exists an index i such that (λi A ii ) A d j d j ij = A ij 2 2 (18) However, (λi A ii ) 1 2 (si A ii ) 1 H for any λ such that Re(λ) 0 Combining (8) and (18) gives: A ij 2 d j (λi A ii ) > A ij 2 d j We arrive at the contradiction and complete the proof Proof of Proposition 5: Consider the matrix A with B δi A = δi B with δ > 0, Hurwitz matrix B R k k such that (a) σ(b) 1 (b) σ(x 0 )δ 1/2, where X 0 is a solution to X 0 B + BX 0 + I = 0 (c) the matrix M α (A) is Hurwitz with α = k, k} Such a matrix ( exists if B) R k k with k 2 For 8 8 example, let B = and δ = We will show that under assumptions (a)-(c) there does not exist an α-diagonal matrix X 0 such that the matrix M α (A X + XA) is Hurwitz The matrix M α (A X + XA) is Hurwitz if and only if there exist X 1 0, X 2 0, γ such that B X 1 + X 1 B 0, σ(b X 1 + X 1 B) > σ(x 1 + X 2 )γδ, B X 2 + X 2 B 0, σ(b X 2 + X 2 B)γ > σ(x 1 + X 2 )δ, which implies that B X 1 + X 1 B + σ(x 1 + X 2 )Iδγ 0 B X 2 + X 2 B + σ(x 1 + X 2 )Iδ/γ 0, It can be verified that P = σ(x 1 +X 2 )X 0 (γ+1/γ)δ satisfies B P + P B + σ(x 1 + X 2 )I(γ + 1/γ)δ = 0, and X 1 + X 2 P It follows that σ(x 1 + X 2 ) > σ(x 0 )σ(x 1 + X 2 )(γ + 1/γ)δ, which cannot be fulfilled since σ(x 0 )(γ + 1/γ)δ 1 for all γ > 0 due to b) Indeed, γ 2 γ/(δσ(x 0 )) (γ 1/(2δσ(X 0 ))) /(2δσ(X 0 )) /(2δσ(X 0 )) 2 0 δσ(x 0 ) 1/2 This completes the proof Remark 1: The proof of Proposition 5 holds when B R k k with k 2 Indeed, if b is a positive scalar then b = σ(b), X 0 = ( 1/(2b) ) therefore we need to pick δ b b δ However, A = has a positive eigenvalue if δ > b δ b and has an eigenvalue at the origin if δ = b Hence, no two by two matrix can satisfy the conditions in the proof of Proposition 5