IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 63, NO. 4, APRIL 2018 1059 Stochastic Consentability of Linear Systems With Time Delays and Multiplicative Noises Xiaofeng Zong, Tao Li , Senior Member, IEEE, George Yin , Fellow, IEEE, Le Yi Wang , Fellow, IEEE, and Ji-Feng Zhang , Fellow, IEEE
Abstract—This paper develops stochastic consentability Index Terms—Consentability, consensus, multiagent of linear multiagent systems with time delays and multi- system, multiplicative noise, time delay. plicative noises. First, the stochastic stability for stochastic differential delay equations driven by multiplicative noises is examined, and the existence of the positive definite so- I. INTRODUCTION lution for a class of generalized algebraic Riccati equations (GAREs) is established. Then, sufficient conditions are de- ULTIAGENT systems have attracted much attention in duced for the mean square and almost sure consentability recent years. The research progress in such systems has and stabilization based on the developed stochastic stabil- M ity and GAREs. Consensus protocols are designed for linear provided valuable insights and benefits in understanding, de- multiagent systems with undirected and leader-following signing, and implementing distributed controllers for such sys- topologies. It is revealed that multiagent consentability de- tems; see [1]Ð[3]. To date, the consensus problem of perfect pends on certain characterizing system parameters, in- models, which assume that each agent can obtain its neighbor in- cluding linear system dynamics, communication graph, formation timely and precisely, has reached a reasonable degree channel uncertainties, and time delay of the deterministic term. It is shown that a second-order integrator multiagent of maturity; see [4] for example. However, networked systems in system is unconditionally mean square and almost surely practical applications often operate in uncertain communication consentable for any given noise intensities and time delay, environments and are inevitably subjected to communication and that the mean square and almost sure consensus can latency and measurement noises [5]Ð[7]. Hence, time delays be achieved by carefully choosing the control gain accord- and measurement noises should be taken into consideration for ing to certain explicit conditions. examining multiagent consensus problems. Manuscript received February 20, 2017; accepted July 14, 2017. Date In the literature, Olfati-Saber and Murray [7] gave optimal of publication July 27, 2017; date of current version March 27, 2018. delay bounds for consensus under undirected graphs. Bliman This work was supported in part by the National Natural Science Foun- and Ferrari-Trecate [8] studied average consensus problems for dation of China under Grant 61522310, in part by Shanghai Rising-Star Program under Grant 15QA1402000, in part by the National Key Basic undirected networks with constant, time varying and nonuni- Research Program of China (973 Program) under Grant 2014CB845301, form time delays, and presented some sufficient conditions for in part by the Fundamental Research Funds for the Central Universities, average consensus. Different types of time delays (communi- China University of Geosciences (Wuhan) under Grant CUG170610, and in part by the U.S. Air Force Office of Scientific Research under FA9550- cation delay, identical self-delay and different self-delay) were 15-1-0131. This paper was presented in part at 12th IEEE Conference investigated in [9] for output consensus. These works have fo- on Control and Automation, Kathmandu, Nepal, June 1Ð3, 2016. Rec- cused on first-order multiagent systems. For second-order mul- ommended by Associate Editor E. Zhou. (Corresponding author: George Yin.) tiagent systems, Yu et al. [10] gave necessary and sufficient X. Zong is with the School of Automation, China University of Geo- consensus conditions related to time delays. Zhou and Lin [11] sciences, Wuhan 430074 China, and with Hubei Key Laboratory of Ad- showed that consensus problems of linear multiagent systems vanced Control and Intelligent Automation for Complex Systems, Wuhan 430074, China (e-mail: [email protected]). with input and communication delays could also be solved by T. Li was with the School of Mechatronic Engineering and Automation, truncated predictor feedback protocols. Shanghai University, Shanghai 200072 China. He is now with the Shang- For multiplicative noises (or state-dependent noise), Ni and hai Key Laboratory of Pure Mathematics and Mathematical Practice, Department of Mathematics, East China Normal University, Shanghai Li [12] studied consensus problems of continuous-time systems 200241, China (e-mail: [email protected]). in which the noise intensity functions are the absolute values of G. Yin is with the Department of Mathematics, Wayne State University, relative states. Li et al. [13] revealed that multiplicative noises Detroit, MI 48202 USA (e-mail: [email protected]). L. Y. Wang is with the Department of Electrical and Computer En- may enhance the almost sure consensus, but may have damaging gineering, Wayne State University, Detroit, MI 48202 USA (e-mail: effect on the mean square consensus, indicating a distinct fea- [email protected]). ture from additive noises which are always destabilizing factors. J.-F.Zhang is with the Key Laboratory of Systems and Control, Institute of Systems Science, Academy of Mathematics and Systems Science, These results have been established for first-order multiagent Chinese Academy of Sciences, Beijing 100190 China, and also with systems. For linear discrete-time systems, Li and Chen [14] gave School of Mathematical Sciences, University of Chinese Academy of the mean square consensus analysis by solving a modified alge- Sciences, Beijing 100149, China (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available braic Riccati equation. When the time delay and multiplicative online at http://ieeexplore.ieee.org. noise coexist, stochastic consensus conditions were examined Digital Object Identifier 10.1109/TAC.2017.2732823 in [15] for second-order discrete-time models. However, little is
0018-9286 © 2017 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. 1060 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 63, NO. 4, APRIL 2018 known about the consensus for general continuous-time linear designing control protocols under time delays and measurement multiagent systems with time delays and multiplicative noises. noises. The contribution of this paper is detailed as follows. All the aforementioned works have focused on the issue 1) Two new techniques (Theorems II.1 and II.2) are devel- of convergence to consensus, which is concerned with con- oped to study the multiagent consentability and consensus ditions on the given control protocols under which the agents under measurement noises and time delays. We first ex- can achieve certain agreements asymptotically. However, they ploit the degenerate Lyapunov functional [25] to establish do not address the issue of consentability. Consentability is the mean square and almost sure exponential stability cri- concerned with conditions on system parameters under which a teria of stochastic differential delay equations (SDDEs). consensus protocol exists, and is of great importance, both theo- We show that the stability is independent of the time delay retically and practically, for cooperative control protocol devel- with stochastic influence. Then, the existence of positive opment such as flocking behavior, agent rendezvous, and robot definite solutions to GARE is obtained. The GARE plays coordination. Note that Ma and Zhang [16] considered general an important role in stochastic linear-quadratic control linear continuous-time multiagent systems and obtained nec- problems [26]Ð[28]. Here, both sufficient and necessary essary and sufficient conditions on system parameters for con- conditions are presented for guaranteeing the existence of sentability. This problem was then generalized in [17] to the case a positive definite solution to the GARE. As a byproduct, with input constraints and uncertain initial conditions in terms necessary and sufficient conditions are given for the ex- of linear matrix inequalities (LMIs). For discrete-time models, istence of stochastic feedback stabilization control laws. Zhang and Tian [18] investigated second-order multiagent sys- That is, it is sufficient that the open-loop dynamics (A, B) tems with Markov-switching topologies, and proved that such is controllable and the product of the sum of real parts of networked systems are mean square consentable under linear unstable open-loop poles and the square of noise inten- consensus protocols if and only if the union of the graphs in sity is less than 1/2, while it is necessary that (A, B) is the switching topology has globally reachable nodes. For linear stabilizable and the product of the maximal real part of multiagent systems, You and Xie [19] obtained necessary and the open-loop poles and the square of noise intensity is sufficient conditions for single-input multiagent consentability less than 1/2. and revealed how the agent dynamic, network topology, and 2) The necessary and sufficient conditions for multiagent communication data rate affect consentability. Gu et al. [20] consentability and consensus are revealed. We first de- employed a properly designed dynamic filter into local con- rive the mean square and almost sure consentability con- trol protocols to relax consensusability conditions. For the issue ditions for leader-free linear multiagent systems. a) It is of structural controllability of multiagent systems, we refer the proved that if the agent dynamics (A, B) is controllable, λu N −1 2 λ ∗ readers to [21] and references therein. At present, consentability 4 0 N σ¯ < 2 and the time delay τ1 <τ1 , then the lin- for multiagent systems with multiplicative noises, even for the ear multiagent system is mean square and almost surely λu delay-free case, remains an open problem, due to the lack of consentable, where 0 denotes the sum of the real parts suitable techniques to treat intrinsic complications from general of the unstable eigenvalues of A, N is the number of linear systems. agents, λ2 is the algebraic connectivity of the graph, σ¯ ∗ This paper aims to establish the mean square and almost sure is the bound of the noise intensities, and τ1 is a bound consentability and develop the consensus protocol design of of the time delay in the deterministic term. Especially, linear multiagent systems with time delays and multiplicative some special cases are given to show that the stabiliz- noises. Our results accommodate different time delays in deter- ability of (A, B) and the restriction on noise intensity ministic (drift) and stochastic (diffusion) terms. Departing from are necessary for the mean square consentability. b) For the feedback structure of [22], [23], where each agent’s state second-order integrator multiagent systems under undi- feedback was used, our consensus protocol requires only lo- rected graphs, we show that it is mean square and almost cal relative-state measurements. In this case, each agent’s state surely consentable regardless of noises and time delays. is not locally stabilized, leading to a more difficult consensus Some necessary conditions on the mean square consensus analysis problem. It can be verified that a consensus problem is are also obtained for the delay-free case. Then, the con- equivalent to the stability problem of the corresponding closed- sentability and consensus results are extended to leader- loop system. Hence, for consensus analysis under measurement following multiagent systems which include stochastic noises, the key is to establish stochastic stability of the corre- stabilization by delayed noisy feedback as the special sponding stochastic equations driven by additive or multiplica- case. tive noises, see [12]Ð[15], [24]. However, for linear multiagent The rest of the paper is structured in five parts. Section II systems of orders higher than one, multiplicative noises are of- develops the main tools such as stochastic stability and GARE ten degenerate and the existing stochastic stability theory does for examining the consentability of linear multiagent systems. not involve the issues for such stochastic differential equations. Section III investigates consentability conditions and pursues Time delays add further difficulty in deriving consentability consensus protocols of leader-free linear multiagent systems conditions and consensus protocols of linear multiagent sys- under undirected graphs. Section IV extends our investigation tems with multiplicative noises. New ideas and techniques are to the leader-following case and stochastic stabilization with needed to resolve these difficulties. In this paper, stochastic delayed noisy feedback. Section V presents simulation results stability and a generalized algebraic Riccati equation (GARE) to verify the theoretical analysis. Section VI concludes the paper are established for obtaining multiagent consentability and with some remarks that outline possible directions for future ZONG et al.: STOCHASTIC CONSENTABILITY OF LINEAR SYSTEMS WITH TIME DELAYS 1061 research in this field. For clarity of presentation, the proofs of Theorem II.1: Suppose Assumption II.1 and condition (2) most technical results are relegated to the appendix. hold. If there exists a matrix P>0 such that 1 Notation: The symbol N denotes the N-dimensional col- ¯T ¯ ¯T ¯ T A P + P A +(A P A + A1 PA1 )τ1 + DP < 0 (3) umn vector with all ones; ηN,i denotes the N-dimensional col- umn vector with the ith element being 1 and others being zero; where A¯ = A0 + A1 , then there exist C0 ,γ0 > 0 such that J = 1 1 1T I N N N N N ; N denotes the -dimensional identity matrix. 1 γ T 2 −γ0 t 0 For a given matrix or vector A, its transpose is denoted by A , its Ey(t) ≤ C0 e , lim sup log y(t) < − , a.s. (4) →∞ t 2 Euclidean norm is denoted by A, and its maximum real part of t the eigenvalues is denoted by max(Re(λ(A))). For a real sym- That is, the trivial solution to SDDE (1) is mean square and metric matrix A, λmax(A) and λmin(A) denote its maximum and almost surely exponentially stable. − minimum eigenvalues, respectively. For two matrices A and B, In applications, the diffusion fi (y(s τ2 ))dwi (s) may A ⊗ B denotes their Kronecker product. For symmetric matri- have various forms (see Sections III and IV). Note that d t T − − ces X and Y , the notation X ≥ Y (respectively, X>Y) means M,P0 M (t)= i=1 0 fi (y(s τ2 ))P0 fi (y(s τ2 ))ds. that the matrix X − Y is positive semidefinite (respectively, So we can use the following condition to replace (2): positive definite). For a, b ∈ R, a ∨ b represents max{a, b} and T d M,P0 M (t) ≤ y (t − τ2 )DP y(t − τ2 )dt (5) a ∧ b denotes min{a, b}.Let(Ω, F, P) be a complete proba- 0 bility space with a filtration {Ft }t≥0 satisfying the usual condi- which is a direct consequence of (2) that does not involve the tions. For a given random variable or vector X, the mathematical concrete form of diffusion. In fact, the proof of Theorem II.1 expectation of X is denoted by EX. For continuous martin- in Appendix is based on (5). Theorem II.1 also leads to the gales M(t) and N(t), their quadratic variation is denoted by following corollaries. M,N (t) (see [29]). We also denote M (t):= M,M (t). Corollary II.1: Suppose Assumption II.1 and condition (2) For the fixed τ>0,weuseC([−τ,0], Rn ) to denote the space hold. If there exists a positive definite matrix P such that n ¯T ¯ of all continuous R -valued functions ϕ defined on [−τ,0] with TP := A P + P A + DP < 0, then for any τ1 ∈ [0,τ1 (P )), the norm ϕ C =supt∈[−τ,0] ϕ(t) . the trivial solution to SDDE (1) is mean square and almost λmin(−TP ) surely exponentially stable, where τ1 (P ):= ¯T ¯ T . A P A+A 1 PA1 II. STOCHASTIC STABILITY AND GARE Corollary II.2: Suppose that A¯ is symmetric, τ1 =0, and When the control of multiagent systems with the delayed fi(y)=iy, i =1,...,d. Then, the trivial solution to SDDE and noisy measurements is studied, the consensus problem of (1) is mean square exponentially stable if and only if 2A¯ + d 2 the closed-loop systems is transformed into the stability prob- i=1 i In < 0. lem of SDDEs. Hence, the stochastic stability theorem (see Remark II.1: Concerning Theorem II.1, first it can yield Theorem II.1) should be first established. After the stability the- an explicit delay bound (see Corollary II.1). Second, it pro- orem, we need to fix how to design the feedback control based vides a delay dominated stability criterion and improves the on the delayed and noisy measurements. The solution will resort LMI stability theorems in [30], which takes the SDDE dy(t)= − − to a GARE (see Theorem II.2). This section is to establish the y(t τ1 )dt + y(t)dw(t) for example, and gave delay bound ∗ two tools. τ1 < 0.1339. Our Theorem II.1 yields a better delay bound ∗ τ1 < 0.5. Third, it shows that the stability of stochastic delay A. Stochastic Stability of SDDEs systems does not necessarily depend on the time-delay in diffu- sion. In fact, the necessary condition of the mean square stability This section aims to establish the mean square and almost is independent of such delay in Corollary II.2. sure exponential stability of the following SDDE: Remark II.2: Corollary II.1 shows that the mean square sta- ble linear time-invariant SDEs can tolerate a certain time delay, − dy(t)=[A0 y(t)+A1 y(t τ1 )]dt + dM(t) (1) which depends on the original system parameters. Especially, for the scalar SDDE: dy(t)=(a0 y(t)+a1 y(t − τ1 ))dt + σy(t − × d t ∈ Rn n − ∗ 2(a +a )+σ 2 where A0 ,A1 , M(t)= i=1 0 fi(y(s τ2 ))dwi − 0 1 τ2 )dw(t), we have a time delay bound τ1 = (a +a )2 +a2 (s), τ ,τ ≥ 0, f : Rn → Rn , d>0, {w (t)}d are indepen- 0 1 1 1 2 i i i=1 for mean square stability if the delay-free SDE dy(t)=(a + dent Brownian motions. The functions {f (x)}d satisfy the 0 i i=1 a )y(t)dt + σy(t)dw(t) is mean square stable, which is equiv- following assumption. 1 alent to 2(a + a )+σ2 < 0. Assumption II.1: f (0) = 0, i =1,...,d, and there exist 0 1 i Remark II.3: Note that the sufficient conditions for the mean positive constants { }d such that for any y ,y ∈ Rn , i i=1 1 2 square and almost sure exponential stability in Theorem II.1 f (y ) − f (y )≤ y − y , i =1,...,d. i 1 i 2 i 1 2 do not involve time delay τ in the diffusion term. Corollary We also give the initial data y(t)=ϕ(t) for t ∈ [−τ,0], τ = 2 II.2 also shows that the mean square stability is independent τ ∨ τ , and ϕ ∈ C([−τ,0], Rn ). Here, we assume that for each 1 2 of time delay τ when τ =0, which is consistent with the P > 0, there exists a D ≥ 0 such that 2 1 0 P0 scalar case in [31]. Theorem II.1 and Corollary II.2 also give ¯ d that if A is symmetric and fi (y)=i y, i =1,...,d, then the T T ≤ trivial solution to SDDE (1) is mean square and almost surely fi (y)P0 fi (y) y DP0 y. (2) λ (−(2A¯+ d 2 I )) ≤ min i =1 i n i=1 exponentially stable if τ1 ¯T ¯ T . Compared A A+A 1 A 1 1062 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 63, NO. 4, APRIL 2018 with [32, Th. 2.1], not only does it improve the stability criterion, dependent noise: but also avoids solving a complex quadratic equation. dx(t)=[Ax(t)+Bu(t)]dt + σB0 u(t)dw(t) (9) B. GARE where u(t)=Kx(t) is the input control. The authors proved In this section, we develop the existence and uniqueness of that the stochastic system (9) can be stabilized for any given the solution P>0 to the following GARE: noise intensity σ>0 if and only if (A, B) is stabilizable and the columns of input coefficient matrix B (Rank(B )
In the following sections, we apply Theorems II.1 and II.2 for t ∈ [−τ,0], and ϕ ∈ C([−τ,0]; RNn) is deterministic. Let to establish multiagent consentability, consensus, and stochastic f·(·) denotes the family of noise intensity functions {flji(·), stabilization. i =1, 2, ..., N, j ∈ Ni , l =1, 2, ..., d}. The collection of all admissible distributed protocols with different control gain III. LEADER-FREE LINEAR MULTIAGENT SYSTEMS matrices is denoted by ⎧ A. Linear Multiagent Systems ⎨ U · | ≥ We consider a system consisting of N (N ≥ 2) agents where (τ1 ,τ2 ,f·( )) = ⎩u(t) ui (t)=K zji(t),t 0, the agents are indexed by 1, 2,...,N, respectively. The dy- j=N i namics of the N agents are described by the continuous-time systems: K ∈ Rm ×n ,i=1,...,N. . (15)
x˙ i(t)=Axi(t)+Bui(t),t∈ R+ (12) Under the measurement noise, the consensus definitions are ∈ Rn ∈ Rn×n ∈ Rn×m where i =1, 2, ..., N, xi(t) , A , B , diversified due to the different asymptotic behaviors in the prob- ∈ Rm and ui(t) is the input control of the ith agent, ability sense, where the mean square and the almost sure consen- T T T respectively. Denote x(t)=[x1 (t), ..., xN (t)] and u(t)= sus are two important topics. Here, we give the definitions on the T T T [u1 (t), ..., uN (t)] . Here, the input control u(t) is to be de- mean square and the almost sure consentability and consensus. signed. The information flow structures among different agents Definition III.1: We say that the linear systems (12) are mean are modeled as a connected undirected graph G = {V, E, A}, square (or almost surely) consentable w.r.t. U(τ1 ,τ2 ,f·(·)),if where V = {1, 2, ..., N} is the set of nodes with i represent- there exists a protocol u ∈U(τ ,τ ,f·(·)) solving the mean E A 1 2 ing the ith agent, denotes the set of directed edges and = square (or almost sure) consensus, that is, it makes the agents ∈RN ×N G [aij ] is the adjacency matrix of with element aij =1 have the property that for any initial data ϕ ∈ C([−τ,0]; RNn) or 0 indicating whether or not there is an information flow from 2 and all distinct i, j ∈V, limt→∞ Exi(t) − xj (t) =0 (or agent j to agent i directly. Also, N denotes the set of the node i lim →∞ x (t) − x (t) =0, a.s.). ∈ N t i j i’s neighbors, that is, for j Ni , aij =1, and degi = j=1 aij Remark III.1: The delayed measurement (f (·) ≡ 0)orthe G lji is called the degree of i. The Laplacian matrix of is defined as noisy measurement without delay (τ = τ =0) was investi- L D−A D 1 2 = , where = diag(deg1 , ..., degN ). It is obvious that gated in [7] and [13], respectively. In fact, in the complicated L is a symmetric matrix and admits a zero eigenvalue, denoted environment, the information communication is often subject to λ λ ≤ ≤ λ by 1 ; other eigenvalues 0 < 2 ... N are positive due to time-delays and measurement noises simultaneously. The gen- G the connectivity of . eral measurement (14) is to describe this phenomenon and has We consider the distributed protocol in the following form: been examined in [15] for the first-order multiagent consensus. It is worth noting that the current results are still new even for ui(t)=K zji(t) (13) the delay-free case (τ = τ =0). j∈N i 1 2 Remark III.2: Multiagent consensus and consentability ∈ Rm ×n where K is the feedback gain matrix to be designed based on precise (noise and delay free) relative state measure-
zji(t)=xj (t − τ1 ) − xi(t − τ1 ) ments were well studied in [7] and [18], [40]. Consensus based on relative state measurements require less information than that d based on both absolute and relative state measurements, and for + f (x (t − τ ) − x (t − τ ))ξ (t),j∈ N (14) lji j 2 i 2 lji i many realistic cases, each agent may not have the ability to l=1 get the absolute state information due to the limited perception is the state measurement of the agent i from its neighbor agent j, capacity and the lack of global coordinates. Multiagent con- T ∈ Rd τ1 ,τ2 are the time delays, ξji(t)=(ξ1ji(t),...,ξdj i (t)) sentability has not been well studied in the delayed and/or noisy · Rn → Rn is the measurement noise, flji( ): is the noise environment. Here, the admissible protocol (15) can be viewed − − − intensity function. In this paper, xj ( t τ1 ) xi(t τ1 ) is as a natural extension from the precise relative state measure- d − − called the deterministic term and l=1 flji(xj (t τ2 ) ments to the delayed noisy version. The skills developed in this xi(t − τ2 ))ξlji(t) is called the stochastic term. We assume that paper can also be used for the case with absolute state feedback. the noises are independent Gaussian white noises. To be exact, Applying Theorems II.1 and II.2 produces the follow- they satisfy the following assumption. ing theorem, which gives the consentability conditions and Assumption III.1: The noise process ξji(t)=(ξ1ji(t),..., the consensus protocol design for linear systems (12) under T ∈ Rd t ≥ λ ≥ ξdj i (t)) satisfies 0 ξji(s)ds = wji(t), t 0,i=1, 2, max(Re( (A))) 0. ...,N,j ∈ Ni , where {wji(t),i=1, 2,...,N,j ∈ Ni } are in- Theorem III.1: Suppose that Assumption III.1 holds, flji(x) dependent d-dimensional Brownian motions defined on the = σljix, σlji ≥ 0, and max(Re(λ(A))) ≥ 0. Then, linear sys- complete probability space (Ω, F, P). tems (12) are mean square and almost surely consentable w.r.t. Note that the update of state x(t) depends on the past U(τ1 ,τ2 ,f·(·)) if states x(s), s ∈ [t − τ,t) with τ = τ1 ∨ τ2 . We need to de- 1) (A, B) is controllable; − λu N −1 2 λ fine the initial function on [ τ,0]. We assume that x(t)=ϕ(t) 2) 4 0 N σ¯ < 2 ; 1064 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 63, NO. 4, APRIL 2018
∈ ∗ T 3) τ1 [0,τ1 ), (0,σjix2 ) , σji > 0. Then, linear systems (12) are mean square − ∗ λ −4 N 1 σ¯ 2 (λu + ) N −1 2 1 ∧ 2 N 0 consentable w.r.t. U(τ1 ,τ2 ,f·(·)) only if λN > 4β σ , where τ1 = 2A2 P 6(λu + )λ , P>0 is the N 0 N N ∈ λu λu ∈ where σ = min σji. solution to the GARE (11) with α ( 0 , 0 + ), i,j=1 λ − N −1 2 λu 2 4 N σ¯ 0 ) 2 d N 2 (0, − ), σ¯ = max σ . In particular, 4 N 1 σ¯ 2 l=1 i,j=1 lji N B. Second-Order Integrator Multiagent Systems T −1 T the protocol (13) with K = k(Im + B PB) B P , k ∈ (k, k¯), solves the mean square and almost sure consensus, In view of Theorems III.1 and III.4, we can see that if λ − λ2 − ¯ λ λ2 − max(Re(λ(A))) > 0, the mean square consentability of linear where k =[ 2 2 2αρ]/ρ, k =[ 2 + 2 2αρ]/ρ, ρ =(2N −1 σ¯2 +3λ τ )λ . systems (12) might be destroyed by the large noise intensities. If N N 1 2 λ Remark III.3: If A =0, then condition 1) in Theorem III.1 max(Re( (A))) = 0, then the consentability of linear systems (12) holds for any given noise intensities if τ1 =0. However, we implies that B is invertible. In this case, for any given τ1 , τ2 , and need to solve the GARE (11) in order for finding the consensus σlji, i, j =1,...,N,l =1,...,d, the linear multiagent sys- protocol. Moreover, we do not know whether the restriction on tems (12) are mean square consentable w.r.t. U(τ1 ,τ2 ,f·(·)). −1 1 the time delay can be removed if max(Re(λ(A))) = 0. These In fact, K = −kB with k ∈ (0, − ) can be used to λ τ + N 1 σ¯ 2 N 1 N motivate us to give further investigation. achieve the stochastic consensus. λ In this section, we consider the second-order integrator sys- If A is stable (max(Re( (A))) < 0), then deterministic con- T 2 tems, that is, xi(t)=[yi (t),vi (t)] ∈ R with the dynamics sensus of x˙ i = Axi follows because each xi(t) tends to zero. Moreover, we can prove the following theorem, which shows y˙i(t)=vi(t), v˙i(t)=ui (t),i=1, 2, ..., N (16) that small control gain does not affect the consensus of the which can be considered as the special case of the linear system original system x˙ i = Axi . (12) with A =(01), and B =(0, 1)T . Here, y (t) and v (t) can Theorem III.2: Suppose that Assumption III.1 holds, flji(x) 00 i i be considered as the position and the velocity of the ith agent, = σljix, σlji ≥ 0, and max(Re(λ(A))) < 0. Then, the linear systems (12) are mean square and almost surely consentable respectively. It is easy to see that (A, B) is controllable and λ(A)=0. We will show that second-order integrator systems w.r.t. U(τ1 ,τ2 ,f·(·)) for any given τ1 ,τ2 ,σlji > 0. Especially, the choice K = kBT P can guarantee the mean square and al- (16) must be mean square and almost surely consentable w.r.t. U · most sure consensus of the linear system (12), where k ∈ R (τ1 ,τ2 ,f·( )) for any given system parameters. Here, we focus | |λ T 2 N −1 T | | on the choice of the protocol for solving stochastic consensus. satisfies k N PBB P (2 +σ ¯ N BB P k ) < 1, P = ∞ A T t At For this case, we consider the different uncertainties in each 0 e e dt. When τ vanishes, we can obtain the following corollary, position and velocity communication. Then, the protocol (13) 1 can be rewritten as follows: which is a direct consequence of Theorem III.1. Corollary III.1: Suppose that Assumption III.1 holds, ui(t)=k1 z1ji(t)+k2 z2ji(t) (17) f (x)=σ x, σ ≥ 0, τ =0 and max(Re(λ(A))) ≥ 0. lji lji lji 1 j∈N i j∈N i Then, linear systems (12) are mean square and almost surely ∈ R1×2 − − − consentable w.r.t. U(τ1 ,τ2 ,f·(·)) if conditions 1) and 2) in where K =[k1 ,k2 ] ,z1ji(t)=yj (t τ1 ) yi(t τ1 ) − − − − − Theorem III.1 hold. Moreover, the protocol (13) with K = + f1ji(yj (t τ2 ) yi(t τ2 ))ξ1ji(t),z2ji(t)=vj (t τ1 ) T −1 T ¯ − − − − k(Im + B PB) B P , k ∈ (k, k), solves the mean square vi(t τ1 )+f2ji(vj (t τ2 ) vi(t τ2 ))ξ2ji(t) are the and almost sure consensus, where P> 0 is the solution to position and the velocity measurements of the agent i from λ2 T ∈ λu 2 λ − λ2 − ¯ its neighbor j ∈ Ni , ξji(t)=[ξ1ji,ξ2ji] is the 2-D Gaussian GARE (11) with α ( 0 , ), k =[ 2 2 2αρ]/ρ, k = 2ρ white noise, that is, ξ and ξ are the scalar independent [λ + λ2 − 2αρ]/ρ, ρ =2N −1 σ¯2 λ . 1ji 2ji 2 2 N 2 Gaussian white noise, and f (·) and f (·) satisfy the Theorem III.1 and Corollary III.1 focus on the suf- 1ji 2ji following assumption. ficient conditions on stochastic consentability. We now Assumption III.2: For each (j, i), f (0) = f (0) = 0 examine the necessary conditions for the mean square 1ji 2ji and there exist nonnegative constants σ¯ and σ¯ consentability. 1ji 2ji such that f (y ) − f (y )≤σ¯ y − y , f (y ) − Theorem III.3: Suppose that Assumption III.1 holds and 1ji 1 1ji 2 1ji 1 2 2ji 1 f (y )≤σ¯ y − y for all y ,y ∈ R. τ =0. Then, linear systems (12) are mean square consentable 2ji 2 2ji 1 2 1 2 1 By using Theorem II.1 and choosing appropriate P in (3), we w.r.t. U(τ ,τ ,f·(·)) only if (A, B) is stabilizable. 1 2 can obtain the following consensus conditions. The following example gives the necessity of certain restric- Theorem III.5: Suppose that Assumptions III.1 and III.2 tion on the noise intensities for the mean square consentability hold, and d =2. Then, the protocol (17) solves the mean square of the two-dimensional dynamics. and almost sure consensus of the multiagent system (16) if Theorem III.4: Consider the multiagent systems (12) with k1 > 0,k2 > 0, and − 01 2 N 1 n =2,τ1 = τ2 =0,A= k 2λ τ +¯σ (2 + k λ τ )+k λ τ 0 β 1 N 1 1 N 1 N 1 1 2 1
N − 1 T λ − 2 − 2 λ2 and B =(0, 1) . Suppose that Assumption III.1 holds < 2 2 k2 1 k2 σ¯2 4k2 N τ1 (18) T N with d =2, β>0 and f1ji(x)=(σjix1 , 0) ,f2ji(x)= ZONG et al.: STOCHASTIC CONSENTABILITY OF LINEAR SYSTEMS WITH TIME DELAYS 1065
N N where σ¯1 = maxi,j=1 σ¯1ji and σ¯2 = maxi,j=1 σ¯2ji solve the stochastic leader-following consentability and con- Remark III.4: Theorem III.5 shows that for any time delays sensus problems, where we only need the subgraphs formed by { }N τ1 ,τ2 and σ1ji,σ2ji i,j=1, we can find the appropriate control the followers to be undirected. This is the following section’s gains k1 ,k2 to guarantee the mean square and almost sure con- intention. { }N sensus. In fact, for any time delays τ1 ,τ2 and σ1ji,σ2ji i,j=1, λ2 we first choose k2 such that k2 < λ2 λ 2 N −1 .Let IV. LEADER-FOLLOWING MULTIAGENT SYSTEMS 2 N τ1 + 2 σ¯ 2 N λ − 2 N −1 − k2 be fixed. Then, we have p1 := 2 2 k2 (1 k2 σ¯2 N ) A. Leader-Following Linear Multiagent Systems 2 λ2 4k2 N τ1 > 0. Based on the given k2 , we choose k1 such that k p (k )
0 − ≥ − ≥ such that σ¯1ji y1 y2 f1ji(y1 ) f1ji(y2 ) σ1ji y1 For the ith follower, the dynamics is described as (12) − − ≥ − ≥ − y2 , σ¯2ji y1 y2 f2ji(y1 ) f2ji(y2 ) σ2ji y1 with ui(t) defined by (13). Note that this is different from ∈ R y2 for all y1 ,y2 , Section III-A since for each agent i, its neighbors Ni may con- By Theorem III.5 and stochastic stability theorem of stochas- tain the leader 0. Considering the information flow from the tic ordinary differential equations, we can prove the following leader to the followers, we denote the topology by G = {V , A} necessary conditions and sufficient conditions. with V = {0, 1, 2,...,N} and Theorem III.6: Suppose that Assumptions III.1 and III.3 hold, d =2, and τ1 = τ2 =0. Then, the protocol (17) solves 00N ×N (N +1)×(N +1) the mean square consensus if A = ∈ R , a0 A − − 2 N 1 λ − 2 2 N 1λ k1 > 0,k1 σ¯1