You are on page 1of 8

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2020.3032496, IEEE
Transactions on Automatic Control
1

Cooperative Output Regulation of Heterogeneous


Multiagent Systems: A Global Distributed
Control Synthesis Approach
Ahmet Taha Koru1 , Selahattin Burak Sarsılmaz2 , Tansel Yucelen3 , and Eric Norman Johnson1

Abstract—Agent-wise local design methods to synthesize dis- structure. Consequently, the gain synthesis for a distributed
tributed control gains focus on the individual dynamics of each control law is known to be NP-hard since forcing some entries
agent to guarantee the overall stability of the system. They are of the control gain matrix to be zero yields a nonconvex
powerful tools due to their scalability. However, the agent-wise
local design methods are incapable of maximizing the overall feasible set [9]. Some research results restrict the Lyapunov
system performance through, for example, decay rate assignment. functions to be in a structured form to make the problem
On the other hand, design methods, which are predicated on tractable, e.g., [10] and [11]. As a result, the design becomes
a global condition, leads to non-convex optimization problems. an optimization over a convex subset of the nonconvex feasible
This paper considers a global design of an internal model- set. One of the main questions that arises is the existence of
based distributed dynamic state feedback control law for the
linear cooperative output regulation problem. We present a a solution to the structured Lyapunov inequality [12].
convex formulation of this global design problem based on a The linear algebra literature has attempted to address this
structured Lyapunov inequality. Then, the existence of solutions question for the last several decades (see, e.g., [13]–[16] and
to the structured Lyapunov inequality is investigated. Specifically, references therein). For a given Hurwitz matrix A, there exist
we analytically show that the solutions exist for the systems symmetric positive definite matrices P such that AT P + P A
satisfying the agent-wise local sufficient condition. Finally, we
compare the proposed method with the agent-wise local design is symmetric negative definite. Such matrices P are so-called
method through numerical examples in terms of conservatism, stability factors for A [16]. References [13]–[15] investigate
performance maximization, graph dependency, and scalability. the existence of diagonal stability factors. The necessary
and sufficient conditions in [16] further this investigation by
considering block diagonal matrices. There are more control-
I. I NTRODUCTION theoretic results recently. Reference [17] presents a sufficient
A. Literature Review H∞ condition for the existence of solutions to block diag-
onally structured Lyapunov inequality. The authors of [18]
Distributed control approaches to the cooperative output
study the diagonal stability of a class of systems with a
regulation of heterogeneous linear time-invariant (LTI) mul-
cyclic interconnection structure. The main result of [19] is
tiagent systems draws the attention of the researchers in the
a necessary and sufficient condition for a diagonal solution to
last decade [1]–[8]. In particular, the authors of [3], [5], and
the ARE.
[8] study the internal model-based distributed dynamic state
feedback control law. They present global and agent-wise
local sufficient conditions on the existence of a control law B. Contribution
that solves the linear cooperative output regulation problem. In this paper, we propose a global design method to synthe-
The control gains are synthesized using the algebraic Riccati size gains of the internal model-based dynamic state feedback
equation (ARE) in [3] and using linear matrix inequalities control law for the heterogeneous LTI multiagent systems
(LMIs) in [5]. Both methods meet the agent-wise local suffi- based on a structured Lyapunov inequality. The resulting
cient condition. However, the existing literature lacks a global feasible set is convex in decision variables and solvable by
design approach. polynomial-time algorithms. We also show that the agent-wise
Global design methods consider the overall closed-loop local sufficient condition implies the existence of solutions
dynamics of the multiagent system. When all matrices de- to the structured Lyapunov inequality. Hence, for a given
termining the agent dynamics are represented in a compact fixed (i.e., time-invariant) augmented directed graph, the set of
form, the resulting overall control gain matrix has a particular stabilizing controllers satisfying the agent-wise local sufficient
* This research was supported in part by the Dynamics, Control, and
condition is a subset of the one whose elements are the results
Systems Diagnostics Program of the National Science Foundation under of the proposed global method.
Grants CMMI-1657637. In what follows, we concisely compare the agent-wise local
1 A. T. Koru and E. N. Johnson are with the Department of Aerospace
Engineering at Pennsylvania State University, University Park, PA, USA
and global design in four categories. Section V illustrates each
(emails: ahtakoru@gmail.com, eric.johnson@psu.edu). item through numerical examples.
2 S. B. Sarsılmaz is with the Department of Aeronautics and Astronautics
• Conservatism: The global design presented in this paper
at the University of Washington, Seattle, WA, USA (email: burak@uw.edu).
3 T. Yucelen is with the Department of Mechanical Engineering at the is less conservative than the agent-wise local design. The
University of South Florida, Tampa, FL, USA (email: yucelen@usf.edu). reason is twofold. On one hand, the agent-wise local

0018-9286 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Auckland University of Technology. Downloaded on November 01,2020 at 13:55:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2020.3032496, IEEE
Transactions on Automatic Control
2

sufficient condition implies the existence of solutions the follower agents. The notation IN stands for the index set
to the structured Lyapunov inequality for a given fixed {1, 2, . . . , N }. The augmented graph Ḡ is the extension of the
augmented directed graph. On the other hand, the global graph G by including the leader node. Finally, the following
design can synthesize control gains that may violate the definition tailors the internal model concept, which is given in
agent-wise local sufficient condition (see the first example Definition 1.22 and Remark 1.24 of [20], for the purpose of
in Section V). this paper.
• Global Performance: The global design can optimize
Definition 1. Given any square matrix A0 , a pair of matrices
the distributed control gains considering a broad set of
(G1 , G2 ) is said to incorporate a p-copy internal model of the
global performance criteria besides local ones. Examples
matrix A0 if G1 and G2 are given by
include decay rate assignment (once again, we refer to
the first example in Section V), pole placement in LMI G1 = diag(βl ), G2 = diag(σl ),
regions, and norm minimization. The drawback of the h×h h
local design is the lack of information about the global where βl ∈ R and σl ∈ R , for l = 1, . . . , p, satisfy the
performance. following conditions:
• Graph Dependency: The agent-wise local design re- i) The pair (βl , σl ) is controllable.
quires graph-wise global information. However, this in- ii) The minimal polynomial of A0 is equal to the character-
formation is not the exact structure of the graph but the istic polynomial of βl .
spectral radius of a matrix depending on the graph. There-
fore, the agent-wise local design ensures the stability of II. P ROBLEM F ORMULATION
the system over a family of graphs that share similar Consider the cooperative output regulation of a group of N
properties. Dissimilarly, the global design depends on follower agents over a fixed directed graph G with dynamics
the exact structure of the graph. Additional optimization given by
constraints regarding a family of graphs overcome this
disadvantage by trading-off with more computational cost ẋi (t) = Ai xi (t) + Bi ui (t) + Ei w(t),
(see the second example in Section V). xi (0) = xi0 , t ≥ 0, (1a)
• Scalability: The main disadvantage of the global design ei (t) = Ci xi (t) − F w(t), (1b)
is its computational complexity as it synthesizes all
control gains simultaneously. On the contrary, the agent- where xi (t) ∈ Rni is the state, ui (t) ∈ Rmi is the input,
wise local design breaks the problem into smaller sub- ei (t) ∈ Rp is the tracking error, and w(t) ∈ Rq is the
problems that consider the agents one by one. Therefore, exogenous signal that yields the reference input to be tracked
it depends less on the system scale (see the third example (i.e., F w(t)) and the external disturbance to be rejected (i.e.,
in Section V). Ei w(t)). The exosystem with the following dynamics
ẇ(t) = A0 w(t), w(0) = w0 , t≥0 (2)
C. Notation and Definitions generates the exogenous signal.
The notation of this paper is standard. Specifically, R, Rn , The tracking error ei (t) is assumed to be available to a
R n×m
respectively stand for the sets of all real numbers, n- nonempty proper subset of all agents. Each agent has also
dimensional real column vectors, and n × m real matrices; access to the relative output error; that is, ei (t) − ej (t) for all
Sn+ denotes the set of all n × n symmetric positive definite j ∈ Ni . Based on the available information to each agent, a
matrices. In addition, we use X  0 (, ≺, ) to denote local virtual regulating error is employed
a symmetric positive definite (positive semidefinite, negative  
definite, negative semidefinite) matrix. The element-wise order 1 X
evi (t) , aij (ei (t) − ej (t)) + gi ei (t)
Y > 0 (≥) means that the all entries of the vector/matrix di + gi
j∈Ni
Y are positive (nonnegative), and “,” denotes equality by
definition. We also write In for the n × n identity matrix, as in [8]. The nonnegative real number gi represents whether
0n×m for the n × m matrix with zero entries, 1n for the n × 1 the agent i observes the leader node (gi > 0) or not (gi = 0).
vector of all ones, ρ(·) for the spectral radius of a square A positive gi corresponds to the weight of an edge from the
matrix, ⊗ for Kronecker product, and diag(Xi ) to denote the leader node to the agent i in the fixed augmented directed
block diagonal matrix whose elements on the main diagonal graph Ḡ. Throughout the paper, Ḡ also consists of the weights
are Xi ’s for all i in the corresponding index set. For matrix gi and aij .
X = [xij ], the absolute value notation denotes |X| = [|xij |]. For each agent, consider now the dynamic state feedback
The notation O(·), so called “big O notation” in computer control law given by
science, characterizes the functions according to their growth żi (t) = G1i zi (t) + G2i evi (t), zi (0) = zi0 , t ≥ 0, (3a)
rate. For the graph definitions, we only mention the important
ui (t) = K1i xi (t) + K2i zi (t), (3b)
ones here, where we follow [8] for the rest. In particular, we
write A = [aij ] ∈ RN ×N to denote the adjacency matrix, D = where zi (t) ∈ Rnz , K1i ∈ Rmi ×ni , K2i ∈ Rmi ×nz , G1i ∈
PN
diag(di ) with di = j=1 aij to denote in-degree matrix of Rnz ×nz , and G2i ∈ Rnz ×p . Next, we state the definition of
the graph G, which models the information exchange between the linear cooperative output regulation problem from [3].

0018-9286 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Auckland University of Technology. Downloaded on November 01,2020 at 13:55:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2020.3032496, IEEE
Transactions on Automatic Control
3

Definition 2. Given the system in (1), together with the gfi (s) = Cfi (sI − Afi )−1 Bfi .
exosystem in (2), and the fixed augmented directed graph Ḡ,
The following lemma states that the property (a) of Defini-
find a distributed control law in (3) such that:
tion 2 implies the property (b) under mild conditions.
a) The closed-loop system matrix due to (1) and (3) is
Hurwitz. Lemma 1 (Theorem 4.1 in [8], global condition). If As-
b) For any initial condition of the closed-loop system and sumptions 1-3 hold and Ag is Hurwitz, then the dynamic state
the exosystem, limt→∞ ei (t) = 0 for all i ∈ IN . feedback control law in (3) solves the problem in Definition
2.
This paper makes the following assumptions:
Remark 1. As in the linear output regulation problem,
Assumption 1. The leader system matrix A0 has no eigen- the internal model-based approach is known to be robust
value with negative real part. against plant uncertainties in decentralized and distributed
Assumption 2. The pair (G1i , G2i ) contains a p-copy in- output regulation problems (see, e.g., [22]–[25]).
ternal model of A0 for all i ∈ IN . We now state the agent-wise local sufficient condition,
Assumption 3. The fixed augmented directed graph Ḡ has which is presented in [3]. Later, in [8], the condition of Afi
a spanning tree. being Hurwitz is included in the theorem to fill a gap.
Assumptions 1 and 2 are standard in linear output regulation Lemma 2 (Theorem 2 in [3], agent-wise local sufficient
theory (see Chapter 1 in [20]); and, so is Assumption 3 in condition). Let Assumption 3 hold, and let γ , 1/ρ(FA)
multiagent systems literature (see Remark 3.2 in [21]). when ρ(FA) 6= 0; γ , ∞ otherwise. If Afi is Hurwitz and
In this paper, we formulate the control gain synthesis of the kgfi k∞ < γ (6)
distributed dynamic state feedback control law considering the
global condition (i.e., the property (a) of Definition 2) as a for all i ∈ IN , where kgfi k∞ is the H∞ norm of gfi (s), then
convex optimization problem through a structured Lyapunov the matrix Ag is Hurwitz.
inequality. Moreover, we investigate the existence of solutions
The existence of a P  0 to the Lyapunov inequality ATg P +
to the structured Lyapunov inequality.
P Ag ≺ 0 is necessary and sufficient for the property (a) of
Definition 2. Yet, the feasible set is non-convex due to the
III. G LOBAL D ISTRIBUTED C ONTROL S YNTHESIS structure of K in Ag for the decision variables P and K. On
Now, we briefly overview the related results of [3] and [8]. the other hand, the dual of the structured Lyapunov inequality
The details of the derivation of the overall closed-loop system
and the exosystem-free local agent dynamics can be found in
ATg P + P Ag ≺ 0, (7a)
[8]. The overall closed-loop system in the compact form is  
diag(P1i ) diag(Poi )
η̇(t) = Ag η(t) + Bg wa (t), (4a) P ,  0, (7b)
∗ diag(P2i )
e(t) = Cg η(t) + Dg wa (t), (4b)
yields convex formulations of the control gain synthesis prob-
where Ag , A + BK, and lem, where P1i ∈ Sn+i and P2i ∈ Sn+z necessarily, and
T Poi ∈ Rni ×nz for all i ∈ IN . Now, we present the main
η(t) , xT1 (t), . . . , xTN (t), z1T (t), . . . , zN
T

(t) , results of this section.
T
e(t) , eT1 (t), . . . , eTN (t) , wa (t) , 1N ⊗ w(t),

Lemma 3. Consider a fixed augmented directed graph Ḡ as
in Assumption 3. There exist matrices P in the form of (7b)
with the following matrices
  and K in the form of (5c) such that (7a) holds if and only if
diag(Ai ) 0 there exist matrices
A, , (5a)
diag(G2i )Wdiag(Ci ) diag(G1i ) 
diag(Q1i ) diag(Qoi )

    Q,  0, (8)
diag(Bi ) diag(Ei ) ∗ diag(Q2i )
B, , Bg , , (5b)
0 −diag(G2i )WFa 
Y , diag(Y1i ) diag(Y2i ) ,

(9)
 
K , diag(K1i ) diag(K2i ) , (5c)
  where Q1i ∈ Sn+i and Q2i ∈ Sn+z necessarily, Qoi ∈ Rni ×nz ,
Cg , diag(Ci ) 0 , Dg , −Fa , Fa , IN ⊗ F, (5d) Y1i ∈ Rmi ×ni , and Y2i ∈ Rmi ×nz for all i ∈ IN , such that
F , diag (di + gi )−1 , W , (IN − F A) ⊗ Ip .

(5e)
QAT + AQ + Y T B T + BY ≺ 0. (10)
The matrix F is well-defined, since di + gi > 0 for all i ∈ IN
Proof: The “if” part: Since Q  0, so is Q−1 . By
owing to Assumption 3. The matrices, which define the local
Sylvester’s law of inertia, we have (A + BY Q−1 )T Q−1 +
agent dynamics, are given by
  Q−1 (A + BY Q−1 ) ≺ 0 from congruence transformation of
Ai + Bi K1i Bi K2i (10) with Q−1 . The matrix Q−1 has the form of (7b), which
Afi = ,
G2i Ci G1i can be verified by Theorem 2.1 of [26]. Moreover, Y Q−1
takes the form of (5c). Hence, (7a) holds with P , Q−1 and
 
0  
Bfi = , Cfi , Ci 0 , K , Y Q−1 . The proof of the “only if” part is similar.
−G2i

0018-9286 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Auckland University of Technology. Downloaded on November 01,2020 at 13:55:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2020.3032496, IEEE
Transactions on Automatic Control
4

Theorem 1. Consider a fixed augmented directed graph Ḡ IV. O N THE E XISTENCE OF S OLUTIONS TO THE
as in Assumption 3. If there exist matrices Q in (8) and Y in S TRUCTURED LYAPUNOV I NEQUALITY
(9) such that the LMI in (10) holds and if the control gains Although Section III presents a convex formulation of the
are recovered as global design problem, there may not exist K satisfying (7). In
h i−1 fact, it does exist for all control gains satisfying the agent-wise
Ki , K1i K2i = Y1i Y2i Q∗1i Q
   oi
Q2i (11) local sufficient condition. Let Assumption 3 hold throughout
this section. Let the sets
for all i ∈ IN , then Ag is Hurwitz. If, in addition, Assumptions
1 and 2 hold, then the dynamic state feedback control law in KL , {K | Afi is Hurwitz and (6) is feasible ∀i ∈ IN } ,
(3) solves the problem in Definition 2. KS , {K | (7) is feasible} ,
Proof: Let T be a permutation matrix such that denote all stabilizing controllers in the feasible set of the
     agent-wise local sufficient condition and in the structured
Ini 0ni ×nzi
T , diag diag , (12) Lyapunov inequality, respectively. Mathematically speaking,
0nzi ×ni Inzi
we prove that K ∈ KL implies K ∈ KS ; hence KL ⊆ KS
h i for a fixed augmented directed graph Ḡ. Example 1 shows
with nzi = nz for all i ∈ IN . The matrix Q∗1i Q Q2i
oi
 0 that the converse is not true by presenting a K ∈ KS but
h i
Q1i Qoi K 6∈ KL . Therefore, the method presented in this manuscript
in (11) for all i ∈ IN since diag ∗ Q2i = T QT T is
congruent to Q. By (11), Y1i = K1i Q1i + K2i QToi and Y2i = is less conservative than the agent-wise local design. However,
K1i Qoi + K2i Q2i for all i ∈ IN . Hence, Y = KQ, where K KS = KL if G contains no loop (i.e., acyclic), see Remark 7.
has the form of (5c). Thus, (7) holds with P , Q−1 and K, We derive the result by three steps:
by the proof of Lemma 3. This proves the first statement of 1) Deriving an LMI involving ρ(FA) and FA.
the theorem. The second statement follows from Lemma 1. 2) Presenting an inequality from the literature.
3) Proving that whenever K ∈ KL , there exists a P  0 in
Remark 2. We restrict Q to be in form of (8) which is the form of (7b) such that (7a) holds; hence, K ∈ KS .
a proper subset of the set of all symmetric positive definite
matrices. This leads to conservatism in control synthesis. How- Step 1: The agent-wise local sufficient condition considers
ever, (11) recovers the control gains owing to this restriction. the spectral radius of the matrix FA, whereas the matrix
Consequently, the distributed control gain synthesis condition itself appears in the overall closed-loop dynamics. Therefore,
is an LMI instead of a bilinear matrix inequality (BMI). the relationship between ρ(FA) and FA determines the rela-
Hence, the problem becomes tractable with polynomial-time tionship between the agent-wise local sufficient condition and
algorithms such as SeDuMi [27]. the structured Lyapunov inequality. Corollary 1 presents this
relationship in terms of an LMI. We start with the definition
The LMI (10) is in a standard form. It is straightforward of an M-matrix.
to apply well-known optimal control design methods, e.g.,
H∞ / H2 minimization, and pole placement in LMI regions Definition 3 (Definition 1.2 in Chapter 6 of [31]). Any
[28]. Those methods can be applied to a subset of agents matrix Ξ of the form Ξ = pI − B is called an M-matrix,
i ∈ IN as local performance criteria by considering the where B ≥ 0 and p ≥ ρ(B).
quintuples (Q1i , Qoi , Q2i , Y1i , Y2i ) or to overall system as a Lemma 4 (Theorem 4 in [14]). If I − |B| is a nonsin-
global performance criterion by considering Q and Y . As gular M-matrix1 , there exists a diagonal R  0 such that
an example, the decay rate λ of the overall system can be R − B T RB  0.
maximized by modifying the condition (10) to
Corollary 1. For any positive real number c < γ, there
QAT + AQ + Y T B T + BY + 2λQ ≺ 0, (13) exists a diagonal R  0 such that
1
and using a bisection algorithm over λ. The iterations will Ψ, R − F ARAT F T  0. (14)
c2
converge to the global maximum λ∗S in the convex set defined
by (8) and (13) with (9). The optimal control gains are K ∗ = Proof: The matrix cFA ≥ 0; thus, |cAT F T | = cAT F T .
Y ∗ Q∗−1 , where (13) is feasible with Q∗ , Y ∗ and λ∗S . The matrix IN − cAT F T is a nonsingular M-matrix since
Another approach to synthesize distributed control gains is ρ(cAT F T ) = cρ(FA) = c/γ < 1. We conclude from Lemma
to utilize nonlinear solvers. These solvers linearize the concave 4 that there exists a diagonal R  0 satisfying (14).
parts of the constraint set at a feasible K [29]. Although The existence of a diagonal R  0 satisfying (14) suffices to
nonlinear solvers do not require a structured Lyapunov func- derive the main result of this section. However, if needed, LMI
tion, finding an initial K for a BMI problem is difficult [30]. solvers can be used to construct such a matrix. Furthermore,
Furthermore, effectiveness of nonlinear solvers highly depends an analytical construction is possible for a special case (see
on the initial guess as they optimize in its neighborhood. The Appendix).
decay rate λ∗S is the global maximum in the aforementioned 1 In [14], M-matrices are nonsingular by definition. Thus, we add “non-
convex set, which is a subset of the overall non-convex set. singular” term to the theorem which is not mentioned in the original script.
Therefore, K ∗ is a good initial guess for such algorithms. Note that Ξ in Definition 3 is nonsingular if and only if p > ρ(B).

0018-9286 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Auckland University of Technology. Downloaded on November 01,2020 at 13:55:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2020.3032496, IEEE
Transactions on Automatic Control
5

Step 2: Solvability of the ARE characterizes the H∞ norm P B(FARA


b T T b T P + CgT (R−1 ⊗ Ip )Cg .
F ⊗ Ip )B (18)
of an LTI system and the stability of its system matrix [32].
We utilize the following lemma, which is widely used in LMI We now conclude from (16), (17), and (18) that
applications, to show that the ARE is an upper bound over the (A
b + B(FA
b ⊗ Ip )Cg )T P + P (A
b + B(FA
b ⊗ Ip )Cg )  M,
structured Lyapunov inequality with a proper construction of
the stability factor P . hence, (7a) is feasible.
Lemma 5 (See Lemma 6.2 in [33]). For any real matrices Remark 3. Theorem 2 constructs a function such that
B, C, and any symmetric positive definite matrix R, with " #
diag(ri−1 Pb1i ) diag(ri−1 Pboi )
compatible dimensions, P = fR (P1 , . . . , PN ) =
b b ,
∗ diag(ri−1 Pb2i )
BC + C T B T  BRB T + C T R−1 C.
where P is a stability factor of Ag . We employ this function
Step 3: Now, we prove the existence of solutions to the to prove the existence of solutions to the structured Lyapunov
structured Lyapunov inequality. inequality. In future research, this function may lead novel
Theorem 2. KL ⊆ KS for a fixed augmented directed graph agent-wise local design methods which contain information
Ḡ. about the stability/performance of the overall system with
more general dynamics, e.g., networked control systems [35],
Proof: Fix K ∈ KL . Hence, the pair (K1i , K2i ) satisfies switching networks [36], and nonlinear agent dynamics [37].
the hypotheses of Lemma 2 for each agent. There exists a
positive real number γ
b such that Remark 4. A similar function can be constructed concern-
ing the dual problem. Let Q b i  0 such that Qb i AT + Af i Q bi +
fi
γ̄ , max kgfi k∞ < γ
b < γ. 1 T T
i∈IN b2 Bf i Bf i + Qi Cf i Cf i Qi ≺ 0. Inequality (10) is feasible for
b b
γ
As per Lemma 2.2 in [32], which is also known as bounded the matrix
" #
real lemma, there exists Pbi  0 such that the LMI represen- diag(ri Q
b 1i ) diag(ri Q
b oi )
¯
Q = fR (Q1 , . . . , QN ) =
tation of the ARE b 2i ) ,
b b
∗ diag(ri Q
1 T b
ATfi Pbi + Pbi Afi + 2 Pbi Bfi Bfi T
Pi + Cfi Cfi ≺ 0 (15) along with Y = KQ.
γ
b
is feasible for all i ∈ IN . For any positive real number ri , Remark 5. Theorem 2 can be considered as an alternative
ri 1 T proof of Lemma 2 without employing small-gain theorems.
T
Mi , ATfi Pi + Pi Afi + 2 Pi Bfi Bfi Pi + Cfi Cfi ≺ 0
γ
b ri Remark 6. An open problem here is whether the structured
holds as it is the multiplication of (15) by 1/ri , where Pi , Lyapunov inequality in (7) implies Afi is Hurwitz for all i ∈
ri−1 Pbi . For the following partition IN or not. We know that Ag being Hurwitz does not imply
  that Afi is Hurwitz for all i ∈ IN ; see the detailed discussion
P1i Poi in Appendix A of [8].
Pi = , P1i ∈ Sn+i , P2i ∈ Sn+z , Poi ∈ Rni ×nz
∗ P2i
Remark 7. The sets KS and KL are equal if G contains no
and the permutation matrix T in (12), the stability factor P , loop. To show this, we begin with KL ⊆ KS ⊆ KG where
T T diag(Pi )T  0 takes the form of (7b). Similarly, M , KG , {K | Ag is Hurwitz}. Lemma 2 does not require (6)
T T diag(Mi )T yields in this case, see Remark 4.5 in [8]. Thus, KL = KG from

1
 Lemma 1 in [38]. Consequently, KL = KS = KG .
T bT P
M = A P + PA + PB
b b b R ⊗ Ip B
b2
γ
+CgT (R−1 ⊗ Ip )Cg ≺ 0 (16) V. I LLUSTRATIVE N UMERICAL E XAMPLES
In this section, we compare the agent-wise local design and
in a compact form, where
  the presented method through three numerical examples. We
b , diag(Ai + Bi K1i ) diag(Bi K2i ) ,
A
set each nonzero aij and gi to one.
diag(G2i Ci ) diag(G1i )
 
0 Example 1: Conservatism and Maximizing a Global
B
b, , R , diag(ri ).
−diag(G2i ) Performance Criterion
Note that Ag = A b + B(FA
b ⊗ Ip )Cg . Let R be the one in
Corollary 1 for c = γ
b. Then, inequality (14) yields 1 3 4
P B(FARA
b T T
F ⊗ Ip )Bb T P  1 P B(Rb ⊗ Ip )B b T P, (17)
b2
γ
0 2 5
from the spectrum property of the Kronecker product (see
Proposition 7.1.10 in [34]). Furthermore, as per Lemma 5,
Fig. 1. Ḡ of Example 1 and Ḡ1 of Example 2.
P B(FA
b ⊗ Ip )Cg + CgT (FA ⊗ Ip )T B
bT P 

0018-9286 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Auckland University of Technology. Downloaded on November 01,2020 at 13:55:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2020.3032496, IEEE
Transactions on Automatic Control
6

λ = 0.05 λ = 0.5 λ=2


1.0 0.5 0.5
0 0 0
ei1 (t)

ei1 (t)

ei1 (t)
−1.0 −0.5
−0.5
−2.0 −1.0
−3.0 −1.0 −1.5
−4.0 −1.5 −2.0
0 20 40 60 80 0 2 4 6 8 0 0.5 1 1.5 2
λ = 0.05 λ = 0.5 λ=2
1.8 0.4 0.5
1.2 0 0
ei2 (t)

ei2 (t)

ei2 (t)
0.6 −0.5
−0.4
0 −1.0
−0.6 −0.8 −1.5
−1.2 −1.2 −2.0
0 20 40 60 80 0 2 4 6 8 0 0.5 1 1.5 2
Time (sec) Time (sec) Time (sec)

Fig. 2. The tracking error ei (t) = [ei1 (t), ei2 (t)]T of Example 1 for control gains designed with various decay rates λ in (13).

This example is borrowed from [5] and modified. Specifi- Example 2: Graph-Dependent Design
cally, consider a group of five agents with the matrices
 0.3 −2
  1.8 −0.8  1 3 4
Ai = −0.2 , Bi = , i = 1, 3, 5,
 0.1
−0.1 1.2
 0.9 1.6
0 1],
Ci = 0.4 1.4 , Ei = [ 0.5 i = 1, 3, 5,
h 6 0 1i
0 2 5
h0 1 0 i
Ai = 00 1 , Bi = 01 , i = 2, 4,
0 0 −2
h 01 00 i
Ci = [ 10 01 00 ] , Ei = 01 , i = 2, 4 Fig. 3. Ḡ2 of Example 2.
10

and a leader with the matrices Both approaches are graph-wise global. The structured
Lyapunov inequality in (7) depends on the whole structure of
the graph, whereas the agent-wise local sufficient condition
A0 = [ 00 10 ] , F = [ 10 01 ] .
depends on ρ(FA) as in (6). Thus, the agent-wise local
design is more tolerant of graph variations. To be specific, the
The pair (G1i , G2i ) given by
resulting control gains solve the cooperative output regulation
0 1 0 0 0 0 problem for a family of graphs. As an example, consider
G1i = 0000 , G2i = 10 ,
0001 00 the agent and leader matrices as well as its 2-copy internal
0000 01
model in Example 1, and the fixed augmented directed graphs
contains a 2-copy internal model of A0 for all i ∈ IN . For the Ḡ1 and Ḡ2 seen in Figures 1 and 3, respectively. Let Fj
fixed augmented directed graph seen in Figure 1, γ = 1.11. and Aj describe the structure of Ḡj , for j = 1, 2. We note
From Theorem 1, we calculate the distributed control gains as here γ1 = 1.11 and γ2 = 1.10, where γj , 1/ρ(Fj Aj ).
Control gains for which Afi is Hurwitz and kgfi k∞ < 1.10
for all i ∈ IN ensure the solvability of the cooperative output
K1 =
 −0.77 −8.47 −1.27 −9.11 −1.17 −8.02
 regulation problem not only for Ḡ1 and Ḡ2 but also for any
4.76 1.05 −0.97 −11.23 1.30 12.23 , (19a)
 −0.47 −1.50 0.10 −0.18 −0.88 −0.39 −2.31
 fixed augmented directed graph whose γ ≥ 1.10. The agent-
K2 = 9.72 −6.52 0.04 2.11 16.25 −1.23 −11.98 , (19b) wise local design does not change the size of the optimization
 −0.94 −9.98 −1.80 −11.31 −1.71 −10.60

K3 = 5.93 1.36 −1.34 −14.07 1.89 15.69 , (19c) problem.
 −0.51 −0.89 0.11 −0.21 −0.94 −0.31 −1.15 
K4 = , (19d) On the other hand, the control gains in (19) only guarantees
 4.91 −4.91 −0.10 1.46 7.30 −1.45 −8.81
the solvability of the problem over Ḡ1 . If the resulting gains
K5 = −0.99 −9.65 −1.78 −10.53 −1.74 −10.51

5.41 1.26 −1.27 −12.31 1.80 13.79 . (19e)
are required to be used in different setups, we may employ
the design with a common Lyapunov function. Consider the
With these matrices, kgf4 k∞ = 1.123 ≥ γ which violates the
matrices given by
agent-wise local sufficient condition. Hence, we here conclude  
that KL is a proper subset of KS . diag(Ai ) 0
A(j) = ,
The agent-wise local design does not contain any informa- diag(G2i )Wj diag(Ci ) diag(G1i )
tion about the global decay rate. Designing the control gains where Wj = (IN − Fj Aj ) ⊗ Ip , j = 1, 2. Searching for
with Theorem 6 of [5] ends up with a decay rate λ = 0.078. a common Lyapunov function, QA(j)T + A(j) Q + Y T B T +
On the other hand, the bisection method maximizes the decay BY ≺ 0, j = 1, 2, yields the control gains given by
rate of the overall system up to λ = 2.12 when it is applied
K1 = −4.87 −0.59 0.76 13.19 −1.02 −15.22
 
to the LMI in (13). Figure 2 shows the simulation results for 0.29 −10.28 −1.23 −15.26 −0.83 −7.59 ,
K2 = −0.70 −2.49 0.11 −0.21 −1.36 −0.53 −4.52
 
different decay rates. 16.42 −11.05 0.11 2.96 31.35 −1.83 −21.93 ,

0018-9286 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Auckland University of Technology. Downloaded on November 01,2020 at 13:55:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2020.3032496, IEEE
Transactions on Automatic Control
7

 −6.49 −0.70 0.99 17.46 −1.24 −19.21



K3 = 0.36 −13.07 −1.46 −18.85 −1.00 −9.62 , whereas it takes about 0.15 seconds with the agent-wise local
K4 =
 −0.99 −1.64 0.11 −0.31 −1.53 −0.44 −2.12

, design for the same system scale.
10.90 −10.77 0.06 2.26 14.75 −2.33 −15.86
 −2.73 −0.51 0.41 5.56 −0.75 −8.02

K5 = 0.37 −6.01 −0.91 −8.58 −0.56 −3.61 ,
The Global Design The Agent-wise Local Design
which solves the cooperative output regulation problem over 600 0.8
both fixed augmented directed graphs Ḡ1 and Ḡ2 . However,

Time (sec)
450 0.6
the number of constraints is doubled. 300 0.4
We synthesize the control gains to be used in different fixed 150 0.2
graphs. For the agent-wise local design, the existence of a
0 0
common Lyapunov function should be discussed to guarantee 0 25 50 75 100 0 250 500 750 1000
the stability under arbitrary switching. For the global approach Number of Agents N Number of Agents N
presented within the context of this example, the solvability of
the cooperative output regulation problem may require addi- Fig. 5. Computation time with respect to the system scale for the global
tional conditions despite the existence of a common Lyapunov design and the agent-wise local design.
function. Therefore, the cooperative output regulation problem
under switching graphs needs further investigation in both
cases. VI. C ONCLUSION
We showed that the global distributed control synthesis
Example 3: Scalability problem is convex in decision variables when the Lyapunov
inequality is structured. We proved that satisfaction of the
agent-wise local sufficient condition implies the existence
0 1 2 3 ··· N
of solutions to the structured Lyapunov inequality. (i) The
proposed method was less conservative than the agent-wise
Fig. 4. The path graph for N agents. local design for a fixed augmented directed graph. Besides,
it could maximize overall system performance. Therefore, the
This example compares the scalability properties of the proposed method was very efficient when high performance
proposed method and the agent-wise local design; in par- was a demand. Moreover, nonlinear solvers could utilize the
ticular, it investigates the relationship between computational results of this global approach as an initial guess for better
complexity and the number of agents N . For this purpose, solutions by removing the restrictions on the stability factor
consider N agents with homogeneous dynamics over the P . In terms of scalability, however, it took long for this
fixed augmented directed graph seen in Figure 4. We use global approach to solve a given large scale problem, whereas
the quadruple (A1 , B1 , C1 , E1 ) and the pair (G11 , G21 ) of it took much shorter for the agent-wise local design. (ii)
Example 1 for the agents, and the pair (A0 , F ) for the leader. Future research can focus on extensions of the global design
Since γ depends on N , it is the function γ(N ) in this example. to more general dynamics, e.g., networked control systems,
For any N ≥ 2, the agent-wise local design is to calculate switching networks, and nonlinear agent dynamics. (iii) The
the control gains such that Af1 is Hurwitz and kgf1 k∞ < proof technique in Theorem 2 may also help to develop agent-
γ(N ). Since the agents have homogeneous dynamics, the wise local design methods for more general dynamics with
computational complexity of the agent-wise local design is respect to the discussion in Remarks 3, 4 and 5.
independent of the system scale, and in O(1), once γ(N )
is calculated. However, the calculation of γ(N ) has a com-
putational complexity O(N 2 ) with recent algorithms; see, A PPENDIX
e.g., [39]. Therefore, the overall design is in O(N 2 ). It is Lemma 6 (Theorem 2.3 in Chapter 6 of [31], items (D16)
possible to make the method independent of the system scale and (I27)). Let Ξ = [ξij ] ∈ RN ×N such that ξij ≤ 0 when
by employing the condition kgf1 k∞ ≤ 1 as Remark 4.5 in [8] i 6= j. Then, the following statements are equivalent:
suggests. However, synthesizing such control gains may not • Ξ is a nonsingular M-matrix.
be possible since it is a very conservative condition. • Every real eigenvalue of Ξ is positive.
On the other hand, the number of decision variables (#v ) • There exists z > 0 with Ξz > 0.
and the total number of rows (#r ) in the LMI conditions
of Theorem 1 increase linearly with respect to N . The Lemma 7 (Lemma 4.1 in Chapter 6 of [31]). Let Ξ =
computational complexity is in O(#2v #2.5 3.5
r + #r ) for the [ξij ] ∈ RN ×N such that ξij ≤ 0 when i 6= j. The matrix Ξ is
semidefinite programming solver toolbox SeDuMi [40]. As an M-matrix if and only if Ξ + I is a nonsingular M-matrix
a result, the computational complexity grows at the order for any positive real number .
of 4.5 with respect to N , and is in O(N 4.5 ). Therefore, it The matrix FA is nonnegative; thus, ρ(FA) (i.e., 1/γ)
is more expensive to design large scale systems with the is an eigenvalue of FA, and corresponding left- and right-
proposed method. The computation time with an average eigenvectors are nonnegative (see Theorem 1.1 in Chapter 2
laptop computer for various system scales can be seen in of [31]). The following proposition considers the case where
Figure 5. We see that it takes approximately ten minutes the eigenvectors are positive.
to solve the problem for N = 100 with the global design,

0018-9286 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Auckland University of Technology. Downloaded on November 01,2020 at 13:55:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2020.3032496, IEEE
Transactions on Automatic Control
8

Proposition 1 (Analytic construction of R). Let Assump- [16] D. Carlson, D. Hershkowitz, and D. Shasha, “Block diagonal semista-
T bility factors and Lyapunov semistability of block triangular matrices,”
tion 3 hold. Suppose the right- vr , [vr1 , . . . , vrN ] and
T Linear Algebra and its Applications, vol. 172, pp. 1–25, 1992.
the left-eigenvector vl , [vl1 , . . . , vlN ] of FA correspond- [17] A. Sootla and J. Anderson, “On existence of solutions to structured
ing to 1/γ are positive. Let c be a positive real number, Lyapunov inequalities,” in American Control Conference (ACC), 2016,
R , diag(vri /vli ), and Ψ be the matrix defined in (14). Then, 2016, pp. 7013–7018.
[18] M. Arcak and E. D. Sontag, “Diagonal stability of a class of cyclic
Ψ  0 when c < γ, and Ψ  0 when c = γ. systems and its connection with the secant criterion,” Automatica,
1 1 vol. 42, no. 9, pp. 1531–1537, 2006.
Proof: Clearly, vr = Rvl . Let Ξ , R− 2 ΨR− 2 = [19] A. Aleksandrov and O. Mason, “Diagonal Riccati stability and applica-
T
1
c2 IN − R−1/2 FARAT F T R−1/2 , and z , [z1 , . . . , zN ] tions,” Linear Algebra and its Applications, vol. 492, pp. 38–51, 2016.
√ [20] J. Huang, Nonlinear output regulation: Theory and applications.
with zi = vri vli > 0. Then, we have the identities
1 1 Philedelphia, PA, USA: SIAM, 2004.
R− 2 z= vl and R− 2 vr = z. After algebraic manipulations, [21] F. L. Lewis, H. Zhang, K. Hengster-Movric, and A. Das, Cooperative
Ξz = c12 − γ12 z. control of multi-agent systems: Optimal and adaptive design approaches.
London, England: Springer-Verlag, 2014.
Case (i) c < γ. The vectors z > 0 and Ξz > 0. By [22] E. Davison, “The robust decentralized control of a general servomecha-
Lemma 6, Ξ is a nonsingular M-matrix. Note that Ξ is nism problem,” IEEE Transactions on Automatic Control, vol. 21, no. 1,
symmetric. Therefore, Ξ  0 by Lemma 6. From congruence pp. 14–24, 1976.
1 [23] X. Wang, Y. Hong, J. Huang, and Z.-P. Jiang, “A distributed control
transformation by R 2 , Ψ  0. approach to a robust output regulation problem for multi-agent linear
Case (ii) c = γ. Let  > 0 be given. Ξz = 0 which yields systems,” IEEE Transactions on Automatic Control, vol. 55, no. 12, pp.
(Ξ + I)z = z > 0. From Lemma 6, we conclude that Ξ + I 2891–2895, 2010.
[24] Y. Su, Y. Hong, and J. Huang, “A general result on the robust cooper-
is a nonsingular M-matrix; therefore, Ξ is an M-matrix from ative output regulation for linear uncertain multi-agent systems,” IEEE
Lemma 7 since  > 0 is arbitrary. Then, Ξ  0 by item Transactions on Automatic Control, vol. 58, no. 5, pp. 1275–1279, 2012.
(C8) of Theorem 4.6 in Chapter 6 of [31]. From congruence [25] X. Wang, Y. Su, and D. Xu, “Cooperative robust output regulation
1 of linear uncertain multiple multivariable systems with performance
transformation by R 2 , Ψ  0. constraint,” Automatica, vol. 95, pp. 137–145, 2018.
[26] T.-T. Lu and S.-H. Shiou, “Inverses of 2× 2 block matrices,” Computers
R EFERENCES & Mathematics with Applications, vol. 43, no. 1-2, pp. 119–129, 2002.
[27] J. F. Sturm, “Using SeDuMi 1.02, a MATLAB toolbox for optimization
[1] Y. Su and J. Huang, “Cooperative output regulation of linear multi-agent over symmetric cones,” Optimization methods and software, vol. 11, no.
systems,” IEEE Transactions on Automatic Control, vol. 57, no. 4, pp. 1-4, pp. 625–653, 1999.
1062–1066, 2012. [28] M. Chilali and P. Gahinet, “H∞ design with pole placement constraints:
[2] ——, “Cooperative output regulation of linear multi-agent systems by an LMI approach,” IEEE Transactions on automatic control, vol. 41,
output feedback,” Systems & Control Letters, vol. 61, no. 12, pp. 1248– no. 3, pp. 358–367, 1996.
1253, 2012. [29] Q. T. Dinh, S. Gumussoy, W. Michiels, and M. Diehl, “Combining
[3] C. Huang and X. Ye, “Cooperative output regulation of heterogeneous convex–concave decompositions and linearization approaches for solv-
multi-agent systems: An H∞ criterion,” IEEE Transactions on Auto- ing BMIs, with application to static output feedback,” IEEE Transactions
matic Control, vol. 59, no. 1, pp. 267–273, 2014. on Automatic Control, vol. 57, no. 6, pp. 1377–1390, 2011.
[4] Y. Li, X. Wang, J. Xiang, and W. Wei, “Synchronised output regu- [30] Y. Wang, R. Rajamani, and A. Zemouche, “Sequential LMI approach for
lation of leader-following heterogeneous networked systems via error the design of a BMI-based robust observer state feedback controller with
feedback,” International Journal of Systems Science, vol. 47, no. 4, pp. nonlinear uncertainties,” International journal of robust and nonlinear
755–764, 2016. control, vol. 28, no. 4, pp. 1246–1260, 2018.
[5] F. A. Yaghmaie, F. L. Lewis, and R. Su, “Output regulation of linear [31] A. Berman and R. J. Plemmons, Nonnegative matrices in the mathemat-
heterogeneous multi-agent systems via output and state feedback,” ical sciences. Philedelphia, PA, USA: SIAM, 1994, vol. 9.
Automatica, vol. 67, pp. 157–164, 2016. [32] K. Zhou and P. P. Khargonekar, “An algebraic Riccati equation approach
[6] H. Cai, F. L. Lewis, G. Hu, and J. Huang, “The adaptive distributed to H∞ optimization,” Systems & Control Letters, vol. 11, no. 2, pp. 85–
observer approach to the cooperative output regulation of linear multi- 91, 1988.
agent systems,” Automatica, vol. 75, pp. 299–305, 2017. [33] K. Gu, V. L. Kharitonov, and J. Chen, Stability of time-delay systems.
[7] M. Lu and L. Liu, “Cooperative output regulation of linear multi- Boston, MA, USA: Birkhäuser, 2003.
agent systems by a novel distributed dynamic compensator,” IEEE [34] D. S. Bernstein, Matrix Mathematics, 2nd ed. Princeton, NJ, USA:
Transactions on Automatic Control, vol. 62, no. 12, pp. 6481–6488, Princeton University Press, 2008.
2017. [35] A. T. Koru, E. N. Johnson, T. Yucelen, and S. B. Sarsilmaz, “On
[8] S. B. Sarsilmaz and T. Yucelen, “A distributed control approach for het- cooperative control of linear multiagent systems over networks with
erogeneous linear multiagent systems,” International Journal of Control, limited bandwidth,” in 2020 American Control Conference (ACC), 2020,
available online. pp. 1659–1664.
[9] V. Blondel and J. N. Tsitsiklis, “NP-hardness of some linear control [36] Y. Su and J. Huang, “Cooperative output regulation with application to
design problems,” SIAM Journal on Control and Optimization, vol. 35, multi-agent consensus under switching network,” IEEE Transactions on
no. 6, pp. 2118–2127, 1997. Systems, Man, and Cybernetics, Part B (Cybernetics), vol. 42, no. 3, pp.
[10] D. Stipanovic and D. D. Šiljak, “Robust stability and stabilization 864–875, 2012.
of discrete-time non-linear systems: the LMI approach,” International [37] A. Isidori, L. Marconi, and G. Casadei, “Robust output synchronization
Journal of Control, vol. 74, no. 9, pp. 873–879, 2001. of a network of heterogeneous nonlinear agents via nonlinear regulation
[11] D. D. Šiljak and A. Zečević, “Control of large-scale systems: Beyond theory,” IEEE Transactions on Automatic Control, vol. 59, no. 10, pp.
decentralized feedback,” Annual Reviews in Control, vol. 29, no. 2, pp. 2680–2691, 2014.
169–179, 2005. [38] F. A. Yaghmaie, F. L. Lewis, and R. Su, “Output regulation of hetero-
[12] A. I. Zecevic and D. D. Šiljak, “Global low-rank enhancement of geneous linear multi-agent systems with differential graphical game,”
decentralized control for large-scale systems,” IEEE Transactions on International Journal of Robust and Nonlinear Control, vol. 26, no. 10,
Automatic Control, vol. 50, no. 5, pp. 740–744, 2005. pp. 2256–2278, 2016.
[13] M. Araki, “Application of M-matrices to the stability problems of [39] D. Bini and V. Pan, “Practical improvement of the divide-and-conquer
composite dynamical systems,” Journal of Mathematical Analysis and eigenvalue algorithms,” Computing, vol. 48, no. 1, pp. 109–123, 1992.
Applications, vol. 52, no. 2, pp. 309–321, 1975. [40] D. Peaucelle, D. Henrion, Y. Labit, and K. Taitz, “User’s guide for
[14] P. J. Moylan, “Matrices with positive principal minors,” Linear Algebra SeDuMi interface 1.04,” LAAS-CNRS, Toulouse, 2002.
and its Applications, vol. 17, pp. 53–58, 1977.
[15] D. Hershkowitz, “Recent directions in matrix stability,” Linear Algebra
and its Applications, vol. 171, pp. 161–186, 1992.

0018-9286 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Auckland University of Technology. Downloaded on November 01,2020 at 13:55:16 UTC from IEEE Xplore. Restrictions apply.

You might also like