arxiv: v1 [math.oc] 1 Apr 2014

Transcription

1 On the NP-completeness of the Constrained Minimal Structural Controllability/Observability Problem Sérgio Pequito, Soummya Kar A. Pedro Aguiar, 1 arxiv: v1 [math.oc] 1 Apr 2014 Abstract This paper studies the problem of, given the structure of a linear time invariant system and a set of possible inputs (respectively outputs), finding the smallest subset of input (respectively output) vectors that ensures system s structural controllability (respectively observability). We refer to this problem as the constrained minimal input/output selection (CMIOS) problem, since the selection has to be performed on an initial given set of possible inputs/outputs. We show that the CMIOS is NP-complete, which addresses a recent open question of whether there exist polynomial algorithms (in the size of the system plant matrices) that solve the CMIOS problem. Hence, we show that such algorithms are unlikely to exist (for arbitrarily given input/output sets). Further, we provide a polynomial reduction from the CMIOS to the set covering problem under some additional assumptions on the input/output set instances, applicable in the context of certain multi-agent networked control problems such as leader-follower type control. Finally, we provide an example that illustrates our findings. I. INTRODUCTION Research on large-scale control systems has grown considerably over the last few years, triggered by technological advances in sensing and actuation infrastructures and relatively low cost of deployment. Such pervasive sensing and actuation present tremendous opportunities for enhanced system control, although, at the cost of handling and processing enormous amounts of sensor data for system state inference and subsequently coordinating generated control signals among the actuators distributed throughout the system. Thus, it is of importance to understand which subsets of sensors and actuators (hence the smallest amount of data that need to be processed and coordination required) are crucial for achieving desirable system monitoring (observability) and control (controllability) performance. These and related questions form the Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, PA Institute for System and Robotics, Instituto Superior Técnico, Technical University of Lisbon, Lisbon, Portugal Department of Electrical and Computer Engineering, Faculty of Engineering, University of Porto, Porto, Portugal

2 2 core of the input/output selection problems [5], [12], [13] in large-scale control systems. In this paper, we focus on the problem of, given the structure of a possibly large scale linear time invariant system and a set of possible inputs (respectively outputs), finding the smallest subset of input (respectively output) vectors that ensures system s structural controllability (respectively observability). We refer to this problem as the constrained minimal input/output selection (CMIOS) problem. Formally, consider the system ẋ(t) = Ax(t)+Bu(t) (1) y(t) = Cx(t), t R + 0 (2) where x R n is the state, u R p and y R m denote the input and output vectors, respectively. Additionally, let Ā {0, } n n denote the zero/non-zero or structural pattern of the system matrix A, whereas B {0, } n p and C {0, } m n the structural pattern of the input and output matrices respectively. Notice that the structural matrices defined above determine the autonomous coupling between the system state variables and the locations of actuators (inputs) and sensors (outputs) deployed in the system. Such structural matrices are the object of study in structural systems theory [4], where the pair (Ā, B) is said to be structurally controllable if there exists a numerical realization (A,B) in (1) with the same structure as (Ā, B) that is controllable. Similarly, the pair (Ā, C) is said to be structurally observable if there exists an observable numerical realization (A,C) in (1)-(2) with the same structure as (Ā, C). In fact, a stronger characterization holds and it can be shown that the set of non-controllable numerical realizations (A,B) of a structurally controllable pair (Ā, B) has zero Lebesgue measure in the product space R n n R n p, i.e., in other words, almost all numerical realizations of a structurally controllable pair are controllable [4]. A similar characterization of structural observability holds. Hereafter, we restrict attention to structural system theoretic properties and consider the CMIOS design problem as noted above. More specifically, given the structural matrix and possible input/output configurations, the input selection version of the CMIOS problem, referred to as the constrained minimal input selections (CMIS) problem, consists of identifying the smallest subset of inputs that ensure structural controllability and may be formally posed as follows

3 3 P 1 Given Ā {0, }n n and B {0, } n p, determine J = arg min J {1,...,p} J (3) s.t. (Ā, B J ) is structurally controllable, where J is a subset of indices associated with the inputs and B J corresponds to the subset of columns in B with index in J. The results that we obtain for the CMIS problem P 1 readily extend to the corresponding output selection problem by the duality between observability and controllability in linear systems, and, hence, in what follows, we focus on the CMIS only. Finally, note that the current setup considers continuous time systems, however, all our results apply to the discrete time setting as well due to similar controllability/observability criteria. Problem P 1 has been previously explored by several authors, see [1] and references therein. In fact, [1] provided the motivation for the present technical note, where we address the following question posed by the authors: Is there a polynomial solution to P 1? As pointed out in [1] given Ā and a structural input matrix B, verifying, if a specific subset of inputs B J with J {1,...,p} is such that the pair (Ā, B J ) is structurally controllable, can be performed using polynomial complexity algorithms in the size of the state space and the input matrix. Hence, from the computational complexity viewpoint, it follows that the CMIS problem is in NP, see [6] for an introduction to the topic. The main contribution of the present paper consists in showing that the CMIS (hence the CMIOS) is in fact an NP-complete problem. This is achieved by reducing a well known NPcomplete problem [2], the set covering problem, to the CMIS. Hence, in particular, polynomial complexity algorithms that solve general instances of the CMIS (and CMIOS) are unlikely to exist. The NP-completeness of the CMIS further implies that the CMIS may be polynomially reduced to other (more standard) NP-complete problems. Practically, such reduction may lead to efficient (polynomial complexity) approximation schemes for solving the CMIS with guaranteed suboptimality bounds. While we do not provide such reductions from general CMIS instances to other NP-complete problems, for a certain restricted subclass of CMIS problems (with some additional conditions on the dynamic matrix structure) we explicitly construct a reduction to the set covering problem. Nonetheless, this restricted class is practically relevant and, as shown

4 4 later, subsumes important applications in multi-agent control such as leader-follower problems [8], [9]; as a demonstration, we show how our reduction can be used to solve the leader-selection problem and a more general variant of it, which we refer to as the constrained leader-selection problem. Some aspects of such reduction was also presented in [1], whereas, in this paper we explicitly construct a reduction to the set covering problem and demonstrate its applicability in the context of leader-selection problems. Finally, note that, although the CMIS is shown to be NP-complete, there could be subclasses of the CMIS (with restricted input instances) that admit polynomial complexity algorithmic solutions. In particular, we show that a practically relevant subclass of CMIS problems, in which the B matrix is restricted to be structurally similar to the n n identity matrix 1 (but Ā is arbitrary), falls under the purview of a related structural control design problem addressed in [10] and admits polynomial complexity algorithmic solutions. The rest of this paper is organized as follows: Section 2 introduces some preliminaries on computational complexity theory, associated complexity classes and polynomial reductions between problems. Additionally, in Section 2.1 we review some concepts and results in structural systems theory to be used in the sequel. Section 3 presents the main result of this paper, i.e., the CMIOS is NP-complete. Additionally, a polynomial reduction from the CMIS to the set covering problem is provided in Section 4, under certain restricted assumptions on the CMIS instances. Finally, an illustrative example is provided in Section 5. II. PRELIMINARIES AND TERMINOLOGY In this section, we review the set covering problem [3], and some necessary and sufficient conditions that ensure system s structural controllability. A (computational) problem is said to be reducible in polynomial time to another if there exists a procedure to transform the former to the latter using a polynomial number of operations on the size of its inputs. Such reduction is useful in determining the qualitative complexity class [6] a particular problem belongs to. For instance, recall that a problem P in NP (i.e., the class of non-deterministic polynomial algorithms) is said to be NP-complete if all other NP problems 1 A structural input matrix B that is structurally similar to the n n identity matrix is referred to as a dedicated input configuration, in that, each input can actuate or is connected to at most a single state variable (see Section 3 for details). Such dedicated input configurations are common in several large-scale multi-agent networked control systems such as the power system, see [7], for example.

5 5 can be polynomially reduced to P [6]. The set of NP-complete problems is referred to as the NP-complete class. The following result may be used to check for NP-completeness of a given problem. Lemma 1 ([6]): If a problem A is NP-complete, B is in NP and A is reducible in polynomial time to B, then B is NP-complete. In this paper, we will often use the following well known NP-complete problem as a means to obtaining desired reductions. Definition 1 ([3]): (Set Covering Problem) Given a set of m elements U = {1,2,...,m} and n a set of n sets S = {S 1,...,S n } such that S i U, with i {1,,n}, and S i = U, the set covering problem consists of finding a set of indices I {1,2,...,n} corresponding to the minimum number of sets covering U, i.e., i=1 I = argmin I {1,2,...,n} I subject to: U = i IS i. A. Structural Systems Structural systems provide an efficient representation of a linear time invariant system as a directed graph (digraph). A digraph consists of a set of vertices V and a set of directed edges E V,V of the form (v i,v j ) where v i,v j V. If a vertex v belongs to the endpoints of an edge e E V,V, we say that the edge e is incident to v. We represent the state digraph by D(Ā) = (X,E X,X), i.e., the digraph that comprises only the state variables as vertices denoted by X = {x 1,,x n } and a set of directed edges between the state vertices denoted by E X,X = { (x i,x j ) X X : Ā j,i 0 }. Similarly, we represent the system digraph by D(Ā, B) = (X U,E X,X E U,X ), where U = {u 1,,u p } corresponds to the input vertices and E U,X = { (u i,x j ) U X : Bi,j 0 } the edges identifying which state variables are actuated by which inputs. Further, we say that an input u i is assigned to a state variable x j if B i,j 0. A directed path between the verticesv 1 andv k is a sequence of edges{(v 1,v 2 ),(v 2,v 3 ),...,(v k 1,v k )}. If all the vertices in a directed path are different, then the path is said to be an elementary path.

6 6 A cycle is a directed path such that v 1 = v k and all remaining vertices in the direct path are distinct. We also require the following graph theoretic notions [3]: A digraph D is strongly connected if there exists a directed path between any two vertices. A strongly connected component (SCC) is a maximal subgraph D S = (V S,E S ) of D such that for every u,v V S there exist paths from u to v and from v to u. By visualizing each SCC as a virtual node, we can build a directed acyclic graph (DAG) representation, in which a directed edge exists between vertices belonging to two SCCs if and only if there exists a directed edge connecting the corresponding SCCs in the original digraph D = (V,E). The construction of the DAG associated with D(Ā) can be performed efficiently in O( V + E ) [3]. The SCCs in the DAG may be categorized as follows. Definition 2 ([10]): An SCC is said to be linked if it has at least one incoming/outgoing edge from another SCC. In particular, an SCC is non-top linked if it has no incoming edges to its vertices from the vertices of another SCC. GivenD = (V,E), we can construct a bipartite graph B(S 1,S 2,E S1,S 2 ) [10], where S 1,S 2 V and the edge set E S1,S 2 = {(s 1,s 2 ) E : s 1 S 1,s 2 S 2 }. Such bipartite graphs will be used throughout in connection with the CMIS and we provide some elementary concepts associated with bipartite graphs. Given B(S 1,S 2,E S1,S 2 ), a matching M corresponds to a subset of edges in E S1,S 2 that do not share vertices, i.e., given edges e = (s 1,s 2 ) and e = (s 1,s 2 ) with s 1,s 1 S 1 and s 2,s 2 S 2, e,e M only if s 1 s 1 and s 2 s 2. A maximum matching M is a matching M with the largest number of edges among all possible matchings. Note that, in general, a maximum matching may not be unique. A maximum matching can be computed efficiently in O( S 1 S 2 E S1,S 2 ) using, for instance, the Hopcroft-Karp algorithm [3]. Given a matching M, an edge is said to be matched with respect to (w.r.t.) M, if belongs to M. In addition, we say that a vertex v V 1 V 2 is matched if it is incident to some matched edge in M, otherwise we say that the vertex is free w.r.t. M. Incident and free vertices can be further characterized as follows: a vertex in S 2 is a right-matched vertex if it is incident to an edge in M, otherwise, it is an right-unmatched vertex. A maximum matching in which there are no free vertices (or equivalently, either left/right-unmatched vertices) is called a perfect matching.

7 7 Given a state digraph D(Ā) = (X,E X,X), a particular bipartite graph of interest is its bipartite representation, given by B B(X,X,E X,X ) and we refer to it as the state bipartite graph. The state bipartite graph may be used to characterize all possible structurally controllable pairs (Ā, B) as given in the following result. Proposition 1 ([10]): The pair (Ā, B) is structurally controllable if and only if there exists a maximum matching of B(Ā) with a set of right-unmatched vertices U R such that B has, at least, m non-zero entries, where m is the number of right-unmatched vertices, one in each of the rows corresponding to the different state variables in U R and located at different columns, and, at least, β non-zero entries, where β denotes the number of non-top linked SCCs, each belonging to a row (state variable) corresponding to a distinct non-top linked SCC of D(Ā) and located in arbitrary columns. Remark 1: A configuration of inputs B verifying the conditions in Proposition 1 is referred to a feasible input configuration. To obtain the minimum number of non-zero entries in B, we have to maximize the intersection between the set of right-unmatched vertices and the non-top linked SCCs. Observe that the set of β non-zero entries may share columns with the first set of m non-zero entries. Thus, informally speaking, in order to obtain the input configuration B with the minimum number of non-zero entries (the sparsest input configuration) ensuring structural controllability, we need to compute a maximum matching of B(Ā) that has the maximum number of right-unmatched vertices belonging to the different non-top linked SCCs of D(Ā). In [10] a polynomial complexity algorithm was provided to obtain such a maximum matching and, in particular, it was shown that the sparsest input configuration design problem admits a polynomial complexity solution. We note in this context that the CMIS is in general different from the sparsest input configuration design problem as noted above, however, certain restricted instances of the CMIS (notably, the ones with dedicated inputs, see Section 3) can be polynomially reduced to the sparsest input design problem and hence, may be shown to admit polynomial complexity solutions. Furthermore, as a direct consequence of the construction in Proposition 1, it follows that any B such that (Ā, B) is structurally controllable must have at least m distinct non-zero columns, i.e., m different inputs. From Proposition 1 we immediately obtain the following. Corollary 1: Given D(Ā) = (X,E X,X) and its DAG representation, constituted by k SCCs, denoted by {N i } k i=1, where N i = (X i,e Xi,X i ), let N i1,...n im be the non-top linked SCCs in the

8 8 DAG representation with {i 1,...,i m } {1,...,k} and B(Ā) the state bipartite graph. If B(Ā) has a perfect matching, then (Ā, B) is structurally controllable if and only if for each non-top linked SCC there exists an input (corresponding to a column in B) assigned to at least one of its state variables. Notice that, since B(Ā) a perfect matching, the set of right-unmatched vertices associated with any maximum matching of B(Ā) is empty and hence, only the condition that is related to the non-top linked SCCs needs to be satisfied in Proposition 1 to ensure structural controllability. III. MAIN RESULTS In this section we show that the CMIS presented in P 1 is NP-complete. Since CMIS is in NP 2 it suffices to show that the set covering problem (an NP-complete problem) can be polynomially reduced to P 1, which is achieved in the following. Theorem 1: The constrained minimum input selection problem presented inp 1 is NP-complete. Proof: The proof follows by using Lemma 1. First, note that there exist polynomial algorithms to verify if (Ā, B(J)), for somej {1,...,p} is structurally controllable [1] and hence, it follows that the CMIS is in NP. To show that CMIS is NP-complete, by Lemma 1, it then suffices to show that an NP-complete problem (the set covering problem) can be polynomially reduced to the CMIS. To this end, consider a general set covering problem instance with sets {S i } i I, the index set I = {1,...,p} and universe U = S i, where U = n. Subsequently, construct Ā {0, } n n i I to be a diagonal matrix with non-zero entries, i.e.,, in its diagonal. Additionally, select B {0, } n p, such that its (i,j ) entry is given as follows: for i {1,...,n} and j {1,...,p}. B i,j =, if i S j 0, otherwise, Now, note that D(Ā) = (X,E X,X) consists of n non-top linked SCCs and the associated state bipartite graph B(Ā) has a perfect matching. Now, recall that, by Corollary 1, (Ā, B(J)), for 2 The CMIS is an NP problem since for a given k N there exists a polynomial enumeration of all possible subsets of k columns of B and each such subset can be verified to be feasible or not using a polynomial algorithm (in the size of the system plant matrices), see [4].

9 9 some J {1,...,n}, is structurally controllable if and only if each non-top linked SCC of D(Ā) contains a state variable that is connected from an input (corresponding to a non-zero column in B(J)). We first show that a feasible solution to the CMIS leads to a feasible solution of the set covering problem, and secondly, a minimal solution leads to a minimal solution of the set covering problem. To show feasibility, let B(J), for some J, be a feasible solution to the CMIS, i.e., (Ā, B(J)) is structurally controllable. It then follows that there exists edges from the inputs associated with indices in J to all the state variables (corresponding to the non-top linked SCCs in D(Ā)), which implies by the construction of B that the family of subsets{s j } j J cover U. To obtain minimality, suppose, on the contrary, that J constitutes a minimal solution to the CMIS, but the family {S j } j J is not a minimum covering of U. Then, there exists J {1,...,p} with J < J such that the family {S j } j J covers U. This, in turn, by the construction of B and Corollary 1 implies that the pair (Ā, B(J )) is structurally controllable. Since J < J, we conclude that B(J ) is not a solution to the CMIS, a contradiction. From Theorem 1 we obtain one of the main results of this paper. Corollary 2: The constrained minimum input/output selection (CMIOS) problem is NP-complete. Proof: The result readily follows from Theorem 1 and noting that the results for input selection in Theorem 1 readily extend to the corresponding constrained output selection by the duality between controllability and observability in linear time invariant systems. Discussion of the Results Previously, we have concluded that the CMIOS is NP-complete, which, however, does not rule out the possibility that there exist subclasses of the CMIOS (with restricted input instances) that admit polynomial complexity algorithmic solutions (in the size of the system plant matrices). In fact, a particularly interesting subclass of the CMIOS is one in which the collection of inputs (respectively outputs) initially given consist of all possible dedicated inputs (respectively dedicated outputs), i.e., the matrix B (respectively C) consists ofninputs (respectivelynoutputs) each of which is assigned to a single distinct state variable. Formally, the above restricted subclass of the CMIS, referred to as the minimum dedicated input selection problem, can be posed as

10 10 follows: Given Ā {0, }n n and the input matrix B =I n, i.e., the n n diagonal matrix with non-zero diagonal entries, determine J such that J = arg min J {1,...,N} J (4) s.t. (Ā,I n(j)) is structurally controllable, where I n (J) corresponds to the columns of I n with indices in J. Similarly, we can formulate the dual minimum dedicated output selection problem. In [10], the minimum dedicated input selection problem, posed in (4), is shown to be equivalent to the following sparsest input design problem: Given Ā {0, }n n, determine B such that B = arg min B 0 (5) B {0, } n n s.t. (Ā, B) is structurally controllable B.,j 0 1, j = 1,...,n, where B.,j corresponds to the j-th columns of B and M 0 counts the number of non-zero entries in the matrix M {0, } n 1 n 2. The problem (5) (in fact, a more general variant of (5)) was shown to be polynomially solvable in [10], from which we readily conclude that the minimum dedicated input selection (and output selection, by duality) is polynomially solvable. Finally, we note that the sparsity minimization objective (as in (5)) is not generally equivalent to the CMIOS, which is consistent with the fact that the CMIOS general instance is NP-complete, whereas, the sparsest input/output design problems addressed in [10] are in P, i.e., polynomially solvable. IV. PARTIAL POLYNOMIAL REDUCTION OF THE CMIS TO THE SET COVERING PROBLEM In Section 3 we have showed that the CMIOS is an NP-complete problem without explicitly deriving a polynomial reduction from the CMIOS to another (standard or known) NP-complete problem. In this section we provide a partial polynomial reduction from the CMIS (consequently, by duality, from the CMIOS) to the set covering problem (see Theorem 2 below). By partial reduction we mean that it is only valid if the state digraph satisfies certain additional properties,

11 11 to be made precise in Assumption 1. Notably, the set of state digraphs satisfying Assumption 1 for which the proposed reduction holds, include dynamical systems commonly encountered in multi-agent networked control applications (see Section 4.1 for details). Further, in Section 4.2 we show how the polynomial reduction obtained in Section 4.1 can be used to solve leaderselection problems in multi-agent networks. A. Polynomial Reduction of the CMIS to the Set Covering Problem Throughout this section, we assume that the system dynamic matrices, i.e., the Ā matrices in the CMIS, satisfy the following condition. Assumption 1 The structural dynamic matrix Ā is such that the state bipartite graph B(Ā) = B(X,X,E X,X ) associated with Ā has a perfect matching. In other words, the set of rightunmatched vertices associated with any maximum matching of B(Ā) is empty. We now provide a polynomial reduction from the CMIS to the set covering problem under Assumption 1. Theorem 2: Consider the CMIS problem with system matrix instance Ā {0, }n n and input matrix B {0, } n p, where Ā satisfies Assumption 1. Denote by Ni, i = 1,...,k, the k non-top linked SCCs of D(Ā). The CMIS problem can then be polynomially reduced to the set covering problem with universe U = {1,...,k} and sets {S j } j=1,...,p, where S j = {i U : B r,j =, x r N i }. Proof: The proof requires two steps: 1) to show that the stated reduction to the set covering problem can be achieved by performing a polynomial number of operations with respect to the size of Ā and B; and 2) to prove the correctness of the reduction, i.e., to show that, under Assumption 1, the solution to the CMIS can be readily determined from the minimal solution of the set covering problem. The proposed reduction is polynomial since the non-top linked SCCs of D(Ā) can be determined polynomially, for instance, by computing the DAG associated with D(Ā) (see Section 2.1). Subsequently, the sets S j and the universe U, constituting the set covering problem, can be constructed with linear complexity in the number of state variables in D(Ā). To show correctness, suppose, on the contrary, we have J {1,...,p} such that {S j } j J is a (minimal) solution to the set covering problem, and B(J ) is not a (minimal) solution to the CMIS. Hence, there exists J {1,...,p}, with J < J, such that B(J ) is a solution

12 12 to CMIS. Now note that since Ā satisfies Assumption 1, the bipartite graph B(Ā) consists of a perfect matching, and hence, by Corollary 1, for each non-top linked SCC N i of D(Ā), there exists an input corresponding to an index in J that is assigned to a state variable in N i. Thus, by construction of the set covering problem, the family {S l } l J covers U = {1,...,k}. Since J < J, it follows that the family {S j } j J is not a minimal set covering of U, a contradiction. In the next section we introduce a class of multi-agent networked control problems, referred to as leader-selection problems. Further, we explain how the reduction obtained in Theorem 2 can be used to solve these leader-selection problems. B. Leader Selection Problems We introduce two (structural) variants of leader-selection problems [11], namely, i) structural (unconstrained) leader-selection, and ii) structural constrained leader-selection, as presented next in L 1 and L 2 respectively. 1) Structural (Unconstrained) Leader Selection Problem: The structural (unconstrained) leaderselection problem can be posed as follows: Consider a multi-agent network consisting of N agents and let G denote the (directed) inter-agent communication graph, where each agent i has the ability to transmit scalar data to its neighbors, denoted by N i = {j : (i,j) G}. Let the (distributed) inter-agent interaction dynamics be given as follows: ẋ(t) = Wx(t)+D(J)u(t), (6) where x R N denotes the collection of the agents states, with the i-th entry x i being the state of agent i. The weight matrix W has sparsity induced by G, i.e., W j,i R if i = j or (i,j) G, otherwise W j,i = 0. Additionally, u R J represents the collection of inputs assigned to a subset of J agents (state variables), designated as leaders. Each such input is dedicated, in that, it is assigned to a single agent (state variable). Hence, the interaction dynamics in (6) is given by a linear time invariant dynamical system as in (1), with system matrix W and input matrix D(J), where D is the diagonal (with zero-free diagonal) matrix and D(J) denotes denoting the collection of columns of D with indices in J. Let W {0, } N N denote the sparsity of W induced by G and I N the sparsity of the diagonal matrix D, corresponding to non-zero in the

13 13 diagonals and zero elsewhere. The structural (unconstrained) leader-selection problem may then be posed as follows: L 1 Determine J where J = arg min J {1,...,N} J (7) s.t. ( W,I N (J)) is structurally controllable. Remark 2: Note that problem (7) corresponds to the dedicated input selection problem posed in (4), which can be mapped to the sparsest input design problem (5) (for general W ) and hence solvable using polynomial complexity algorithms. 2) Structural Constrained Leader Selection Problem: We now consider a more challenging variant of leader-selection, the structural constrained leader-selection problem. Recall that the leader agents in Section essentially play the role of input agents that are coupled to the rest of the network by (dedicated) links and are responsible for driving the overall network state to a desired goal. The formulation in Section 4.2.1, from the design or leader-selection perspective, assumes that any agent can play the role of a leader (input agent) with no constraint whatsoever. In practice, the ability of functioning as leaders may be limited to a subset of agents only, possibly determined by various socio-economic-geographic factors and varying from one application to the other. To this end, we propose a generalized leader-selection problem, in which the set of agents with potential leader functionality is constrained; moreover, as dictated by the application, such potential leader agents are not necessarily dedicated, in that, they may actuate or connect to multiple network agents. To formalize, consider a set of N follower agents and let L = {1,,L}, be the given set of L potential leader agents whose goal is to control the collection of N followers. Furthermore, denote by B R N L the input matrix representing the actuation exercised by the potential leader agents, i.e., the entry B f,l indicates how leader l L actuates the follower f {1,,N}. Finally, for a susbet J L denote by B(J) the collection of columns in B corresponding to indices in J. Supposing, a set J of leaders are active or selected, the dynamics of the follower agents is given by ẋ (t) = W x (t)+b(j)u(t), (8)

14 14 where W represents the interaction between the N follower agents, and x R N denotes the collection of the agents states. Given the interaction pattern W, the set L of potential leaders and their actuation pattern B, the structural constrained leader-selection problem may then be posed as follows. L 2 Determine J where J = argmin J L J (9) s.t. ( W, B(J)) is structurally controllable. Remark 3: The non-structural version of the leader-selection problems which aims to design the numerical parameters in the structural matrices ( W,I N (J)) (or ( W, B(J))) in addition to minimizing the set of leader agents J can also be addressed within the current framework. For instance, the non-structural version L 2 of the constrained leader-selection problem L 2 may be posed as: Given W (conforming to a specified follower interaction graph) and B (denoting possible leader configurations), determine a numerical matrix pair (W,B ) with the smallest number of non-zero columns in B, such that (W,B ) is controllable and the structure of (W,B ) conforms to ( W, B) (i.e., an entry in W or B must be zero if the corresponding entry in W or B is zero). Since controllability is a generic property, it is readily seen that the non-structural problem L 2 can be addressed by first solving the structural problem L 2, and subsequently, almost all numerical instances (W,B) of the structural pair ( W, B(J )), where J is the minimal subset of leaders obtained in L 2, are solutions of L 2. A similar relation holds between the structural unconstrained leader-selection problem L 1 and its non-structural counterpart L 1. 3) Solution to the Structural Leader Selection Problems: Subsequently, we show that L 1, L 2 can be solved using set covering problems by employing the reduction developed in Theorem 2. To make the connection between the reduction in Theorem 2 and the leader-selection problems in (6) and (8), we start by showing that the respective dynamics matrices satisfy Assumption 1. Proposition 1: The structural dynamics matrices W, W {0, } N N associated with the leader-selection problems L 1, L2 satisfy Assumption 1. Proof: Let Ā {0, } N N denote the structural matrix W or W (depending on which problem we consider). The proof follows by noticing that D(Ā) consists of self-loops on all the

15 15 state vertices, corresponding to the non-zero diagonal entries in Ā. Consequently, the matching M = {(x i,x i ), i = 1,...,n} is a maximum matching associated with the state bipartite graph B(Ā), which is a perfect matching. In other words, the set of right-unmatched vertices of B(Ā) is empty, and hence Assumption 1 holds. Because Assumption 1 holds for the problems L 1 and L 2, by invoking Theorem 2, we can solve the structural leader-selection problems using a set covering problem. This is formally stated as follows. Corollary 3: The problems L 1, L 2 can be polynomially reduced to set covering problems as given in Theorem 2. V. ILLUSTRATIVE EXAMPLE Consider the system state digraphs depicted in Figure 1. The agent states are depicted by black vertices and the inter-agent dynamical coupling by the black directed edges. Furthermore, consider potential input vertices depicted by blue vertices, where: in Figure 1 a) we pose the structural unconstrained leader-selection problem, whereas, in Figure 1 b), we consider a structural constrained leader-selection problem, in which the blue directed edges represent which leaders can actuate which agents. Hereafter, we illustrate how, both the structural leader-selection problems can be solved using the reduction developed in Theorem 2 (see also Corollary 3). Structural (Unconstrained) Leader Selection Problem: The goal is to solve the leader-selection problem L 1 as formulated in (7) with the structure of the dynamics matrix induced by the state digraph represented by the black vertices and edges as depicted in Figure 1 a). To this end, note that, by Proposition 1 and Corollary 3, L1 can be reduced to a set covering problem (see Theorem 2). From Theorem 2, to set up the set covering problem, we obtain S l = for l {1,...,9} since none of the (potential) inputs u 1,...,u 9, i.e. the dedicated inputs assigned to agents 1 to 9 respectively, are assigned to variables in non-top linked SCCs. In addition, S 10 = {1}, S 11 = {2}, S 12 = {3}, S 13 = {4}, where each set comprises the index of the non-top linked SCC it belongs to, and subsequently the universe U = {1,2,3,4}. It is readily seen that the solution to the set covering problem is unique and comprises the sets S l, with l {10,11,12,13}. Hence, from the viewpoint of leader-selection, agents 10 to 13 should be designated as leaders, which uniquely solves the leader-selection problem. Thus, an input

16 16 must be assigned to the state variables x l (l {10,11,12,13}), as depicted in Figure 1 a) by the blue vertices. It is important to note that in general the set covering problems resulting from structural unconstrained leader-selection problems have the characteristic that the sets S l s comprise at most a single state variable. It is readily seen that such instances of the set covering problem may be solved using polynomial complexity algorithms (recall the set covering problem is NP-complete in general); in fact, to cover the universe, we only need to consider a set for each of the elements in the universe. This in accordance with the fact that (7) can be solved using a polynomial complexity algorithm (see Remark 2). Structural Constrained Leader Selection Problem: Now consider the constrained leader-selection problem L 2 as formulated in (9), with the state digraph induced by the structural dynamics matrix given by the black vertices and edges as depicted in Figure 1 b) and the set of potential leaders depicted by the blue vertices. Additionally, the set of followers actuated by the potential leaders is depicted by the blue edges, i.e., B {0, } 9 4 with all entries equal to zero except: B 1,1 = B 2,1 = corresponding to input u 1 assigned to state variables x 1 and x 2 respectively and, similarly, B2,2 = B 3,2 =, B3,3 = B 4,3 =, B7,4 = B 8,4 =. Now note that, by Proposition 1 and Corollary 3, L2 can be reduced to a set covering problem (see Theorem 2). From Theorem 2, to set up the set covering problem, we obtain S 1 = {1}, S 2 = {1,2}, S 3 = {2} and S 4 =. In other words, agent 1 can only actuate followers from the non-top linked SCC N 1, agent 2 can actuate followers from the non-top linked SCCs N 1,N 2 and so on. Additionally, the universe is U = {1,2} and in this particular example (note that in general the set covering problem in NP-complete), it is straightforward to see that the solution of the set covering problem consists of the set S 2 only. Thus agent 2 should be designated as the leader, which is the solution to the structural constrained leader-selection problem. VI. CONCLUSIONS AND FURTHER RESEARCH In this paper, we have showed that constrained minimal input/output selection (CMIOS) problem is NP-complete, which implies that in general efficient (polynomial complexity) solution procedures to the CMIOS are unlikely to exist. Although the general instance of the CMIOS is NP-complete, there may exist important subclasses or restricted instances of the CMIOS which may admit polynomial complexity algorithmic solutions; in this paper, we have identified one such subclass, namely, CMIOS instances with dedicated inputs/outputs, which can

17 17 Fig. 1. The non-top linked SCCs are depicted by gray dashed boxes and all agents have self-loops (not drawn to keep the illustration simple). In a) we depict the inter-agent communication graph (the agents are depicted by black vertices with associated states as labels) given by the black edges. In addition potential leaders (depicted by blue vertices) are shown to which a dedicated input may be assigned. Alternatively, in b) we depict a communication graph and possible locations for leaders (depicted in blue). be solved polynomially. The NP-completeness of the CMIOS further implies that the CMIOS is polynomially reducible to other NP-complete problems. For a restricted subclass of CMIOS problems, which subsumes practically relevant multi-agent networked control applications such as leader-selection problems, we have explicitly constructed a reduction from the CMIOS to the set covering problem. As future research, it may be worthwhile to obtain reductions from more general instances of the CMIOS to other standard NP-complete problems, notably the ones with good approximation guarantees, such as the SAT [6]. REFERENCES [1] Christian Commault and Jean-Michel Dion. Input addition and leader selection for the controllability of graph-based systems. Automatica, 49(11): , [2] Stephen A. Cook. The complexity of theorem-proving procedures. In Proceedings of the third annual ACM symposium on Theory of computing, STOC 71, pages , New York, NY, USA, ACM. [3] Thomas H. Cormen, Clifford Stein, Ronald L. Rivest, and Charles E. Leiserson. Introduction to Algorithms. McGraw-Hill Higher Education, 2nd edition, [4] Jean-Michel Dion, Christian Commault, and Jacob Van der Woude. Generic properties and control of linear structured systems: a survey. Automatica, pages , [5] Magnus Egerstedt. Complex networks: Degrees of control. Nature, 473(7346): , May [6] Michael R. Garey and David S. Johnson. Computers and Intractability: A Guide to the Theory of NP-Completeness. W. H. Freeman & Co., New York, NY, USA, [7] M.D. Ilic, L. Xie, U.A. Khan, and J.M.F. Moura. Modeling of future cyber physical energy systems for distributed sensing and control. IEEE Transactions onsystems, Man and Cybernetics, Part A: Systems and Humans, 40(4): , July [8] Mehran Mesbahi and Magnus Egerstedt. Graph theoretic methods in multiagent networks. Princeton University Press, 2010.

18 18 [9] Reza Olfati-Saber, J. Alex Fax, and Richard M. Murray. Consensus and Cooperation in Networked Multi-Agent Systems. Proceedings of the IEEE, 95(1): , January [10] Sergio Pequito, Soummya Kar, and A. Pedro Aguiar. A framework for structural input/output and control configuration selection of large-scale systems. Submitted to IEEE Trans. on Automatic Control, Available in [11] Amirreza Rahmani, Meng Ji, Mehran Mesbahi, and Magnus Egerstedt. Controllability of multi-agent systems from a graph-theoretic perspective. SIAM J. Control Optim., 48(1): , February [12] Dragoslav D. Siljak. Large-Scale Dynamic Systems: Stability and Structure. Dover Publications, [13] Sigurd Skogestad. Control structure design for complete chemical plants. Computers and Chemical Engineering, 28(1-2): , 2004.