SYMBOLIC MODELS FOR NONLINEAR CONTROL SYSTEMS WITHOUT STABILITY ASSUMPTIONS

arXiv:1002.0822v3 [math.OC] 18 Oct 2010

MAJID ZAMANI1 , GIORDANO POLA2 , MANUEL MAZO JR.3 , AND PAULO TABUADA1

Abstract. Finite-state models of control systems were proposed by several researchers as a convenient mechanism to synthesize controllers enforcing complex specifications. Most techniques for the construction of such symbolic models have two main drawbacks: either they can only be applied to restrictive classes of systems, or they require the exact computation of reachable sets. In this paper, we propose a new abstraction technique that is applicable to any smooth control system as long as we are only interested in its behavior in a compact set. Moreover, the exact computation of reachable sets is not required.

1. Introduction In the past years several different abstraction techniques have been developed to assist in the synthesis of controllers enforcing complex specifications. This paper is concerned with symbolic abstractions resulting from replacing aggregates or collections of states of a control system by symbols. When a symbolic abstraction with a finite number of states or symbols is available, the synthesis of the controllers can be reduced to a fixed-point computation over the finite-state abstraction [1]. Moreover, by leveraging computational tools developed for discrete-event systems [2, 3] and games on automata [4, 5, 6], one can synthesize controllers satisfying specifications difficult to enforce with conventional control design methods. Examples of such specification classes include logic specifications expressed in linear temporal logic or automata on infinite strings. The quest for symbolic abstractions has a long history including results on timed automata [7], rectangular hybrid automata [8], and o-minimal hybrid systems [9, 10]. Early results for classes of control systems were based on dynamical consistency properties [11], natural invariants of the control system [12], l-complete approximations [13], and quantized inputs and states [14, 15]. Recent results include work on piecewise-affine and multi-affine systems [16, 17], abstractions based on an elegant use of convexity of reachable sets for sufficiently small time [18], and the use of incremental input-to-state stability [19, 20, 21, 22]. Our results improve upon most of the existing techniques in two directions: i) by being applicable to larger classes of control systems; ii) by not requiring the exact computation of reachable sets which is an hard task in general. In the first direction, our technique improves upon the results in [15, 16, 17] by being applicable to systems not restricted to nonholonomic chained-form, piecewise-affine, and multi-affine systems, respectively, and upon the results in [19, 20, 21, 22] by not requiring any stability or stabilizability assumptions. In the second direction, our technique improves upon the results in [13, 14] by not requiring the exact computation of reachable sets. In [18] a different abstraction technique is proposed that is also applicable to a wide class of control systems and does not require the exact computation of reachable sets. Such technique is based on convexity of reachable sets which requires very small sampling times. In contrast, our technique imposes no restrictions on the sampling time. In this paper, we show that symbolic models exist if the control systems satisfy an incremental forward completeness assumption which is an incremental version of forward completeness. The main contribution of this paper is to establish that: This work has been partially supported by the National Science Foundation award 0717188, 0820061 and the Center of Excellence for Research DEWS, University of LAquila, Italy. 1

2

MAJID ZAMANI, GIORDANO POLA, MANUEL MAZO JR., AND PAULO TABUADA

For every nonlinear control system satisfying the incremental forward completeness assumption, one can construct a symbolic model that is alternatingly approximately simulated by the control system and that approximately simulates the control system. Furthermore, since a weaker version of incremental forward completeness holds on compacts for every smooth control system, a symbolic model describing the behavior on compacts can always be constructed. These relationships are weaker than the approximate bisimulation relationships, established in [19, 20, 21, 22], in the sense that they are only sufficient but not necessary to guarantee that any controller synthesized for the symbolic model can be refined to a controller enforcing the desired specifications on the original control system. In other words, any controller synthesized for the abstraction can be converted into a controller enforcing the specification on the original control system. However, failing to find a controller enforcing the specification on the symbolic model does not prevent the existence of a controller for the original control system. Hence, control designers are confronted with the choice between the following two alternatives when using approximate abstractions: (1) design a controller rendering the original control system incrementally input-to-state stable and then apply the existing abstraction techniques in [19, 20, 21, 22]; (2) or construct an abstraction using the results presented in this paper. Since most of the existing controller design techniques provide controllers enforcing stability rather than incremental stability, the second alternative provides a concrete approach to symbolic control design for unstable control systems. 2. Control Systems and Incremental Forward Completeness 2.1. Notation. The identity map on a set A is denoted by 1A . If A is a subset of B we denote by ıA : A ,→ B or simply by ı the natural inclusion map taking any a ∈ A to ı(a) = a ∈ B. The symbols N, Z, R, R+ and R+ 0 denote the set of natural, integer, real, positive, and nonnegative real numbers, respectively. The symbol Im denotes the identity matrix on Rm . Given a vector x ∈ Rn , we denote by xi the i–th element of x, by kxk the infinity norm p of x, and by kxk2 the Euclidean norm of x; we recall that kxk = max{|x1 |, |x2 |, ..., |xn |}, and kxk2 = x21 + x22 + · · · + x2n , where |xi | denotes Pnthe absolute value of xi . Given a matrix M = {mij } ∈ Rn×m , the infinity norm of M is kM k = max1≤i≤m j=1 |mij |. The closed ball centered at x ∈ Rn with radius ε is defined by Bε (x) = {y ∈ Rn | kx − yk ≤ ε}. For any A ⊆ Rn and µ ∈ R+ define [A]µ = {a ∈ A | ai = ki µ, ki ∈ Z, i = 1, · · · , n}. The set [A]µ will be used as an approximation of the set + A with precision collection S µ. Geometrically, for any µ ∈ R and λ ≥ µ the S of sets {Bλ (p)}p∈[A]µ is a covering n of A, i.e. A ⊆ p∈[A]µ Bλ (p). In the special case where A = R , the set p∈[A]µ Bλ (p) remains a cover even n n if we reduce λ so as to satify λ ≥ µ2 . A function d : Rn × Rn → R+ 0 is a metric on R if for any x, y, z ∈ R , the following three conditions are satisfied: i) d(x, y) = 0 if and only if x = y; ii) d(x, y) = d(y, x); and iii) n d(x, z) ≤ d(x, y) + d(y, z). Given a measurable function f : R+ 0 → R , the (essential) supremum (sup norm) + of f is denoted by kf k∞ ; we recall that kf k∞ = (ess) sup {kf (t)k, t ≥ 0}. A continuous function γ : R+ 0 → R0 , is said to belong to class K if it is strictly increasing and γ(0) = 0; γ is said to belong to class K∞ if γ ∈ K and γ(r) → ∞ as r → ∞. 2.2. Control Systems. The class of control systems that we consider in this paper is formalized in the following definition. Definition 2.1. A control system is a quadruple Σ = (Rn , U, U, f ), where: • Rn is the state space; • U ⊆ Rm is the input set which is compact and convex; • U is the set of all piecewise continuous functions of time from intervals of the form ]a, b[⊆ R to U with a < 0 and b > 0;

SYMBOLIC MODELS FOR NONLINEAR CONTROL SYSTEMS WITHOUT STABILITY ASSUMPTIONS

3

• f : Rn × U → Rn is a continuous map satisfying the following Lipschitz assumption: for every compact set Q ⊂ Rn , there exists a constant L ∈ R+ such that for all x, y ∈ Q and all u ∈ U, we have: kf (x, u) − f (y, u)k ≤ Lkx − yk. A curve ξ :]a, b[→ Rn is said to be a trajectory of Σ if there exists υ ∈ U satisfying: ˙ = f (ξ(t), υ(t)) , (2.1) ξ(t) for almost all t ∈ ]a, b[. We also write ξxυ (τ ) to denote the point reached at time τ under the input υ from initial condition x = ξxυ (0); this point is uniquely determined, since the assumptions on f ensure existence and uniqueness of trajectories [23]. Although we have defined trajectories over open domains, we shall refer to trajectories ξxυ :[0, τ ] → Rn and input curves υ : [0, τ ] → U defined on closed domains [0, τ ], τ ∈ R+ , with 0 n the understanding of the existence of a trajectory ξxυ and input curve υ 0 :]a, b[→ U such that 0 :]a, b[→ R 0 0 ξxυ = ξxυ0 |[0,τ ] and υ = υ |[0,τ ] . A control system Σ is said to be forward complete if every trajectory is defined on an interval of the form ]a, ∞[. Sufficient and necessary conditions for a system to be forward complete can be found in [24]. A control system Σ is said to be smooth if f is an infinitely differentiable function of its arguments. 2.3. Incremental forward completeness. The results presented in this paper require a certain assumption that we introduce in this section. Definition 2.2. A control system Σ is incrementally forward complete (δ-FC) if there exist continuos functions + + + + + + β : R+ 0 × R0 → R0 and γ : R0 × R0 → R0 such that for every s ∈ R0 , the functions β(·, s) and γ(·, s) belong 0 n + 0 to class K∞ , and for any x, x ∈ R , any τ ∈ R , and any υ, υ ∈ U, where υ, υ 0 : [0, τ ] → U, the following condition is satisfied for all t ∈ [0, τ ]: kξxυ (t) − ξx0 υ0 (t)k ≤ β(kx − x0 k, t) + γ(kυ − υ 0 k∞ , t).

(2.2)

Incremental forward completeness1 requires the distance between two arbitrary trajectories to be bounded by the sum of two terms capturing the mismatch between the initial conditions and the mismatch between the inputs as shown in (2.2). As an example, for a linear control system: ξ˙ = Aξ + Bυ, ξ(t) ∈ Rn , υ(t) ∈ U ⊆ Rm , the functions β and γ can be chosen as:

β(r, t) = eAt r; γ(r, t) =

(2.3)

Z

t



As

e B ds r,

0 At

At

where ke k denotes the infinity norm of e . Whenever the origin is an equilibrium point for Σ, the choice x0 = 0, υ 0 = 0 results in the estimate kξxυ (t)k ≤ β(kxk , t) + γ(kυk∞ , t) which is shown in [24] to be equivalent to forward completeness of Σ when it holds for all t ∈ R+ 0 . Hence, the systems satisfying (2.2) are termed incrementally forward complete. Descriptions of δ-FC in terms of Lyapunov-like functions and expansion metrics are reported in Section 5. 3. Symbolic Models and Approximate Equivalence Notions 3.1. Systems and control systems. We use systems to describe both control systems as well as their symbolic models. A more detailed exposition of the notion of system that we now introduce can be found in [1]. Definition 3.1. [1] A system S is a quintuple S = (X, U, −→, Y, H) consisting of: 1We note that δ-FC implies uniform continuity of the map φ : Rn × U → Rn defined by φ (x, υ) = ξ (t) for any fixed t ∈ R+ . t t xυ 0 Here, uniform continuity is understood with respect to the topology induced by the infinity norm on Rn , the sup norm on U , and n the product topology on R × U .

4

MAJID ZAMANI, GIORDANO POLA, MANUEL MAZO JR., AND PAULO TABUADA

• • • • •

A set of states X; A set of inputs U ; A transition relation −→⊆ X × U × X; An output set Y ; An output function H : X → Y .

System S is said to be: • metric, if the output set Y is equipped with a metric d : Y × Y → R+ 0; • countable, if X is a countable set; • finite, if X is a finite set. u u A transition (x, u, x0 ) ∈−→ is denoted by x - x0 . For a transition x - x0 , state x0 is called a u-successor, or simply successor, of state x. Since −→⊆ X × U × X is a relation, for any given state and input u ∈ U there may be: no u-successors, one u-successor, or many u-successors. We denote the set of u-successors of a state x by Postu (x) and by U (x) the set of inputs u ∈ U for which Postu (x) is nonempty. A system is deterministic if for any state x ∈ X and any input u, there exists at most one u-successor (there may be none). A system is called nondeterministic if it is not deterministic. Hence, for a nondeterministic system it is possible for a state to have two (or possibly more) distinct u-successors.

The results in this Section and in Section 4 rely on additional assumptions on U and U that we now describe. Such assumptions are not required for the results in Sections Qm 2 and 5. We restrict attention to control systems Σ = (Rn , U, U, f ) with input sets U of the form U = i=1 [ai , bi ] ⊆ Rm with bi > ai . For such input sets we define the constant µ b by µ b = min{|b1 − a1 |, · · · , |bm − am |}. We further restrict attention to trajectories generated by piece-wise constant input curves, by requiring U to contain only constant curves of duration τ ∈ R+ : U = {υ : [0, τ ] → U | υ(t) = υ(0), t ∈ [0, τ ]}. Given a control system Σ = (Rn , U, U, f ), and a time discretization parameter τ ∈ R+ , consider the system - , Yτ , Hτ ) consisting of: Sτ (Σ) = (Xτ , Uτ , τ

• • • • •

n

Xτ = R ; Uτ = U; υτ - x0τ if there exists a trajectory ξxτ υτ : [0, τ ] → Rn of Σ satisfying ξxτ υτ (τ ) = x0τ ; xτ τ Yτ = Rn ; Hτ = 1Rn .

The above system can be thought of as the time discretization of the control system Σ. 3.2. System relations. We first consider approximate simulation relations, introduced in [25], that are useful when analyzing or synthesizing controllers for deterministic systems. Definition 3.2. Let Sa = (Xa , Ua ,

- , Ya , Ha ) and Sb = (Xb , Ub ,

a

- , Yb , Hb ) be metric systems with

b

the same output sets Ya = Yb and metric d, and consider a precision ε ∈ R+ . A relation R ⊆ Xa × Xb is said to be an ε-approximate simulation relation from Sa to Sb , if the following three conditions are satisfied: (i) for every xa ∈ Xa , there exists xb ∈ Xb with (xa , xb ) ∈ R; (ii) for every (xa , xb ) ∈ R we have d(Ha (xa ), Hb (xb )) ≤ ε; (iii) for every (xa , xb ) ∈ R we have that : ua ub xa - x0a in Sa implies the existence of xb - x0b in Sb satisfying (x0a , x0b ) ∈ R. a

b

System Sa is ε-approximately simulated by Sb or Sb ε-approximately simulates Sa , denoted by Sa εS Sb , if there exists an ε-approximate simulation relation from Sa to Sb .

SYMBOLIC MODELS FOR NONLINEAR CONTROL SYSTEMS WITHOUT STABILITY ASSUMPTIONS

5

For nondeterministic systems we need to consider relationships that explicitly capture the adversarial nature of nondeterminism. The notion of alternating approximate simulation relation is shown in [20] to be appropriate to this effect. Definition 3.3. Let Sa and Sb be metric systems with the same output sets Ya = Yb and metric d, and consider a precision ε ∈ R+ . A relation R ⊆ Xa × Xb is said to be an ε-approximate alternating simulation relation from Sa to Sb if the following three conditions are satisfied: (i) for every xa ∈ Xa , there exists xb ∈ Xb with (xa , xb ) ∈ R; (ii) for every (xa , xb ) ∈ R we have d(Ha (xa ), Hb (xb )) ≤ ε; (iii) for every (xa , xb ) ∈ R and for every ua ∈ Ua (xa ) there exists ub ∈ Ub (xb ) such that for every x0b ∈ Postub (xb ) there exists x0a ∈ Postua (xa ) satisfying (x0b , x0a ) ∈ R. System Sa is alternatingly ε-approximately simulated by Sb or Sb alternatingly ε-approximately simulates Sa , denoted by Sa εAS Sb , if there exists an alternating ε-approximate simulation relation from Sa to Sb . It is readily seen from the above definitions that the notions of approximate simulation and of alternating approximate simulation coincide when the systems involved are deterministic. The importance of the preceding notions lies in enabling the transfer of controllers designed for a symbolic model to controllers acting on the original control system. More details about these notions and how the refinement of controllers can be performed are reported in [1]. 4. Symbolic Models for δ-FC Control Systems This section contains the main contribution of the paper. We show that the time discretization of a δ-FC control system, suitably restricted to a compact set, admits a finite abstraction. We consider a δ-FC control system Σ = (Rn , U, U, f ), and a quadruple q = (τ, η, µ, θ) of quantization parameters, where τ ∈ R+ is the time quantization, η ∈ R+ is the state space quantization, µ ∈ R+ is the input space quantization, and θ ∈ R+ is a design parameter. Given Σ and q, define the system: - , Yq , Hq ), (4.1) Sq (Σ) = (Xq , Uq , q

consisting of: • Xq = [Rn ]η ; • Uq = [U]µ ; uq • xq - x0q if kξxq uq (τ ) − x0q k ≤ β(θ, τ ) + γ (µ, τ ) + η2 ; q • Yq = Rn ; • Hq = ı : Xq ,→ Yq , where β and γ are the functions appearing in (2.2). In the definition of the transition relation, and in the remainder of the paper, we abuse notation by identifying uq with the constant input curve with domain [0, τ ] and value uq . The transition relation of Sq (Σ) is well defined in the sense that for every xq ∈ Xq and every uq ∈ Uq there uq - x0q . This can be seen by noting that by definition of Xq , for any always exists x0q ∈ Xq such that xq q

x ∈ Rn there always exists a state x0q ∈ Xq such that kx − x0q k ≤ η/2. Hence, for x = ξxq uq (τ ) there always exists a state x0q ∈ Xq satisfying kξxq uq (τ ) − x0q k ≤ η2 ≤ β(θ, τ ) + γ (µ, τ ) + η2 . We can now state the main result of the paper which relates δ-FC to existence of symbolic models. Theorem 4.1. Let Σ be a δ-FC control system. For any desired precision ε ∈ R+ , and any quadruple q = (τ, η, µ, θ) of quantization parameters satisfying µ ≤ µ b and η ≤ 2ε ≤ 2θ, we have Sq (Σ) εAS Sτ (Σ) εS Sq (Σ).

6

MAJID ZAMANI, GIORDANO POLA, MANUEL MAZO JR., AND PAULO TABUADA

Proof of Theorem 4.1. We start by proving Sτ (Σ) εS Sq (Σ). Consider the S relation R ⊆ Xτ × Xq defined by (xτ , xq ) ∈ R if and only if kHτ (xτ ) − Hq (xq )k = kxτ − xq k ≤ ε. Since Xτ ⊆ p∈[Rn ]η B η2 (p), for every xτ ∈ Xτ there exists xq ∈ Xq such that: kxτ − xq k ≤

(4.2)

η ≤ ε. 2

Hence, (xτ , xq ) ∈ R and condition (i) in Definition 3.2 is satisfied. Now consider any (xτ , xq ) ∈ R. Condition (ii) in Definition 3.2 is satisfied by the definition of R. Let us now show that condition (iii) in Definition 3.2 holds. Consider any υτ ∈ Uτ . Choose an input uq ∈ Uq satisfying: kυτ − uq k∞ = kυτ (0) − uq (0)k ≤ µ.

(4.3)

Note that the existence of such uq is guaranteed by the special form of U, described in Section 3, and by the inS υτ - x0τ = ξxτ υτ (τ ) equality µ ≤ µ b which guarantees that U ⊆ Bµ (p). Consider the unique transition xτ p∈[U]µ

τ

in Sτ (Σ). It follows from the δ-FC assumption that the distance between x0τ and ξxq uq (τ ) is bounded as: kx0τ − ξxq uq (τ )k ≤ β(ε, τ ) + γ (µ, τ ) .

(4.4) Since Xτ ⊆

S

p∈[Rn ]η

B η2 (p), there exists x0q ∈ Xq such that: kx0τ − x0q k ≤

(4.5)

η . 2

Using the inequalities ε ≤ θ, (4.4), and (4.5), we obtain: kξxq uq (τ ) − x0q k = kξxq uq (τ ) − x0τ + x0τ − x0q k ≤ kξxq uq (τ ) − x0τ k + kx0τ − x0q k η η ≤ β(ε, τ ) + γ (µ, τ ) + ≤ β(θ, τ ) + γ (µ, τ ) + , 2 2 which implies the existence of xq and

η 2

uq

- x0q in Sq (Σ) by the definition of Sq (Σ). Therefore, from inequality (4.5)

q

≤ ε, we conclude (x0τ , x0q ) ∈ R and condition (iii) in Definition 3.2 holds.

Now we prove Sq (Σ) εAS Sτ (Σ). Consider the relation R ⊆ Xτ × Xq . For every xq ∈ Xq , by choosing xτ = xq , we have (xτ, xq ) ∈ R and condition (i) in Definition 3.3 is satisfied. Now consider any (xτ , xq ) ∈ R. Condition (ii) in Definition 3.3 is satisfied by the definition of R. Let us now show that condition (iii) in Definition 3.3 holds. Consider any uq ∈ Uq . Choose the input υτ = uq and consider the unique x0τ = ξxτ υτ (τ ) ∈ Postυτ (xτ ) in Sτ (Σ). From the δ-FC assumption, the distance between x0τ and ξxq uq (τ ) is bounded as: kx0τ − ξxq uq (τ )k ≤ β(ε, τ ).

(4.6) Since Xτ ⊆

S

p∈[Rn ]η

B η2 (p), there exists x0q ∈ Xq such that: kx0τ − x0q k ≤

(4.7)

η . 2

Using the inequalities, ε ≤ θ, (4.6), and (4.7), we obtain: kξxq uq (τ ) − x0q k = kξxq uq (τ ) − x0τ + x0τ − x0q k ≤ kξxq uq (τ ) − x0τ k + kx0τ − x0q k ≤ β(ε, τ ) + η ≤ β(θ, τ ) + γ (µ, τ ) + , 2 which implies the existence of xq and

η 2

η 2

uq

- x0q in Sq (Σ) by definition of Sq (Σ). Therefore, from inequality (4.7)

q

≤ ε, we can conclude that (x0τ , x0q ) ∈ R and condition (iii) in Definition 3.2 holds.



SYMBOLIC MODELS FOR NONLINEAR CONTROL SYSTEMS WITHOUT STABILITY ASSUMPTIONS

7

The symbolic model Sq (Σ) has a countably infinite state set. In order to construct a finite symbolic model we note that in practical applications the physical variables are restricted to a compact set. Velocities, temperatures, pressures, and other physical quantities cannot become arbitrarily large without violating the operational envelop defined by the control problem being solved. By making use of this fact, we can directly compute a finite abstraction SqD (Σ) of Sτ (Σ) capturing the behavior of Sτ (Σ) within a given compact set D ⊂ Rn describing the valid range for the physical variables. The system SqD = (XqD , UqD , - , YqD , HqD ) qD is defined by: • XqD = [D]η ; • UqD = [U]µ ; uqD • xqD - x0q if every x0q ∈ [Rn ]η satisfying kξxqD uqD (τ ) − x0q k ≤ β(θ, τ ) + γ (µ, τ ) + qD • YqD = Rn ; • HqD = ı : XqD ,→ YqD .

η 2

belongs to XqD ;

Note that SqD (Σ) is a finite system since D is a compact set. Moreover, the relation R ⊆ XqD × Xq defined by (xqD , xq ) ∈ R if xqD = xq is a 0-approximate alternating simulation relation from SqD (Σ) to Sq (Σ). By combining SqD (Σ) 0AS Sq (Σ) with Sq (Σ) εAS Sτ (Σ) we conclude2 SqD (Σ) εAS Sτ (Σ). Hence, any controller synthesized for the finite model SqD (Σ) can be refined to a controller enforcing the same specification on Sτ (Σ). Moreover, in order to compute SqD (Σ) we only need inequality (2.2) in the definition of δ-FC to hold for times t for which ξ(t) ∈ D. As will be discussed in Section 5, this weaker version of inequality (2.2) holds for any smooth control system on a compact set D. We refer the interested readers to Appendix I for a numerical example showing the effectiveness of the proposed results. In the example, a controller is synthesized for an inverted pendulum subject to a schedulability constraint defined by a finite system. 5. Descriptions of incremental forward completeness This section contains the description of δ-FC in terms of Lyapunov-like functions and expansion metrics. We start by introducing the following definition which was inspired by the notion of incremental input-to-state stability (δ-ISS) Lyapunov function presented in [26]. Definition 5.1. Consider a control system Σ and a smooth function V : Rn × Rn → R+ 0 . Function V is called a δ-FC Lyapunov function for Σ, if there exist K∞ functions α, α, σ, and κ ∈ R such that: (i) for any x, x0 ∈ Rn α(kx − x0 k) ≤ V (x, x0 ) ≤ α(kx − x0 k); (ii) for any x, x0 ∈ Rn and for any u, u0 ∈ U ∂V ∂V 3 0 0 0 0 ∂x f (x, u) + ∂x0 f (x , u ) ≤ κV (x, x ) + σ(ku − u k). The following theorem describes δ-FC in terms of the existence of a δ-FC Lyapunov function. Theorem 5.2. A control system Σ is δ-FC if it admits a δ-FC Lyapunov function. Moreover, the functions β and γ in (2.2) are given by:  κt 
e −1 (5.1) β(r, t) = α−1 2eκt α(r) , γ(r, t) = α−1 2 σ(r) . κ The proof of the preceding result is reported in Appendix II and was inspired by the work in [24]. 2It is shown in [1] that the composition of two alternating simulation relations is still an alternating simulation relation. 3 Condition (ii) of Definition 5.1 can be replaced by ∂V f (x, u) + ∂V f (x0 , u0 ) ≤ ρ(kx − x0 k) + σ(ku − u0 k), where ρ is a K ∞ ∂x ∂x0

function. It is known that there is no loss of generality in considering ρ(kx − x0 k) = κV (x, x0 ), by appropriately modifying the δ-FC Lyapunov function V (see Lemma 11 in [27]).

8

MAJID ZAMANI, GIORDANO POLA, MANUEL MAZO JR., AND PAULO TABUADA

Remark 5.3. It can be easily checked that the quadratic Lyapunov function V (x, x0 ) = kx − x0 k22 satisfies conditions (i) and (ii) in Definition 5.1 if we restrict x and x0 to range in a compact subset of Rn . The arguments in the proof of Theorem 5.2 show that any such control system satisfies inequality (2.2), in the definition of δ-FC, for all t for which the trajectories ξ remain within the aforementioned compact set. This weaker version of inequality (2.2) is sufficient for the construction of the finite symbolic model SqD (Σ) introduced in Section 4. In addition to Lyapunov-like functions, the δ-FC condition can be described by resorting to expansion metrics. The notion of contraction metric was introduced in control theory in [28] as a tool for the study of stability properties of nonlinear systems. The interested reader may also wish to consult [29] where it is shown that contraction metrics were investigated more than 40 years before being used in control theory. In this paper we consider systems that are not necessarily stable and thus consider also expansion metrics. The variational system associated with a smooth control system Σ, when we have variations of the state and input, is given by the differential equation: ∂f ∂f d δξ + δυ (δξ) = (5.2) dt ∂x x=ξ ∂u x=ξ u=υ

u=υ

4

where δξ and δυ are variations of the state and input, respectively. A Riemannian metric G : Rn → Rn×n on Rn is a smooth map such that, for any x ∈ Rn , G(x) is a symmetric positive definite matrix [30]. For any x ∈ Rn and smooth functions Z, W : Rn → Rn , one can define the scalar function hZ, W iG as Z T (x)G(x)W (x). We will still use the notation hZ, W iG to denote Z T GW even if G does not represent any Riemannian metric. In the following, we introduce the notion of exponential expansion which was inspired by the notion of exponential contraction in [31, 28]. Note that in [31, 28], the notion of exponential contraction does not capture the variations on the input. In contrast, the following notion of exponential expansion does contain variations of the state and input. Definition 5.4. Let Σ = (Rn , U, U, f ) be a smooth control system on the manifold Rn equipped with a Riemannian metric G. Control system Σ is said to be an exponential expansion with respect to a metric G, if there exists some λ ∈ R and α ∈ R+ 0 such that:   1 1 ∂f 2 (5.3) hZ, ZiF + 2 ≤ λhZ, ZiG + αhZ, ZiG W, Z hW, W iI2m ∂u G  T ∂G n m for F (x, u) = ∂f G(x) + G(x) ∂f ∂x ∂x + ∂x f (x, u), any Z ∈ R , and any W ∈ R , or equivalently:     T T ∂f ∂f ∂G T ∂f Z + 2W (5.4) ZT G(x) + G(x) + f (x, u) G(x)Z ≤ ∂x ∂x ∂x ∂u 1

1

λZ T G(x)Z + α(Z T G(x)Z) 2 (W T W ) 2 , where the constant λ is called expansion rate. Note that the inequality (5.3) or (5.4) implies: 1 1 d 2 hδξ, δξiG ≤ λhδξ, δξiG + αhδξ, δξiG hδυ, δυiI2m , dt where δξ and δυ are variations of a state and an input trajectory of the control system Σ. The following result describes δ-FC in terms of the existence of an expansion metric.

(5.5)

4The variations δξ and δυ can be formally defined by considering a family of trajectories ξ (t, ) and inputs υ(t, ) parametrized xυ

by  ∈ R. The variations of the state and input are then δξ =

∂ξ ∂

and δυ =

∂υ . ∂

SYMBOLIC MODELS FOR NONLINEAR CONTROL SYSTEMS WITHOUT STABILITY ASSUMPTIONS

9

Theorem 5.5. If Σ = (Rn , U, U, f ) is an exponential expansion with respect to a Riemannian metric G on Rn , satisfying ωkyk22 ≤ y T G(x)y ≤ ωkyk22 for any x, y ∈ Rn and ω, ω ∈ R+ , then Σ is a δ-FC control system. Moreover, the functions β and γ in (2.2) are given by: r r  ωn λ t m α  λt e 2 r, γ(r, t) = e 2 − 1 r, (5.6) β(r, t) = ω ω λ where n and m denote the dimension of the state and input space, respectively. The proof of the preceding result can be found in Appendix II and was inspired by the work in [31]. 6. Discussion In this paper we showed that any smooth control system, suitably restricted to a compact subset of states, admits a finite symbolic model. Our results improve upon the existing work by being applicable to a large class of control systems and by not requiring the exact computation of reachable sets or the convexity of reachable sets. The symbolic models constructed according to the results presented in this paper can be used to synthesize controllers enforcing complex specifications given in several different formalisms such as temporal logics or automata on infinite strings. The synthesis of such controllers is well understood and can be performed using simple fixed-point computations as described in [1]. The current limitation of this design methodology is the size of the computed abstractions. The authors are currently investigating several different techniques to address this limitation such as integrating the design of controllers with the construction of symbolic models [32]. Efforts by other researchers include the use of non-uniform quantization [33]. Acknowledgements The authors would like to thank C. Manolescu for enlightening discussions. References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15]

P. Tabuada, Verification and Control of Hybrid Systems, A symbolic approach. Springer US, 2009. R. Kumar and V. Garg, Modeling and Control of Logical Discrete Event Systems. Kluwer Academic Publishers, 1995. C. Cassandras and S. Lafortune, Introduction to discrete event systems. Boston, MA: Kluwer Academic Publishers, 1999. L. de Alfaro, T. A. Henzinger, and R. Majumdar, “Symbolic algorithms for infinite-state games,” in CONCUR 01: Concurrency Theory, 12th International Conference, ser. Lecture Notes in Computer Science, no. 2154, 2001. P. Madhusudan, W. Nam, and R. Alur, “Symbolic computational techniques for solving games,” Electronic Notes in Theoretical Computer Science, vol. 89, no. 4, 2003. A. Arnold, A. Vincent, and I. Walukiewicz, “Games for synthesis of controllers with partial observation,” Theoretical Computer Science, vol. 28, no. 1, pp. 7–34, 2003. R. Alur and D. L. Dill, Automata, Languages and Programming, ser. Lecture Notes in Computer Science. Berlin: Springer, April 1990, vol. 443, ch. Automata for modeling real-time systems, pp. 322–335. T. Henzinger, P. W. Kopke, A. Puri, and P. Varaiya, “What’s decidable about hybrid automata?” Journal of Computer and System Sciences, vol. 57, pp. 94–124, 1998. G. Lafferriere, G. J. Pappas, and S. Sastry, “O-minimal hybrid systems,” Math. Control Signal Systems, vol. 13, pp. 1–21, 2000. T. Brihaye and C. Michaux, “On the expressiveness and decidability of o-minimal hybrid systems,” Journal of Complexity, vol. 21, no. 4, pp. 447–478, 2005. P. E. Caines and Y. J. Wei, “Hierarchical hybrid control systems: A lattice-theoretic formulation,” Special Issue on Hybrid Systems, IEEE Transaction on Automatic Control, vol. 43, no. 4, pp. 501–508, April 1998. X. D. Koutsoukos, P. J. Antsaklis, J. A. Stiver, and M. D. Lemmon, “Supervisory control of hybrid systems,” Proceedings of the IEEE, vol. 88, no. 7, pp. 1026–1049, July 2000. T. Moor, J. Raisch, and S. D. O’Young, “Discrete supervisory control of hybrid systems based on l-complete approximations,” Journal of Discrete Event Dynamic Systems, vol. 12, pp. 83–107, 2002. D. Forstner, M. Jung, and J. Lunze, “A discrete-event model of asynchronous quantised systems,” Automatica, vol. 38, pp. 1277–1286, 2002. A. Bicchi, A. Marigo, and B. Piccoli, “On the reachability of quantized control systems,” IEEE Transactions on Automatic Control, vol. 47, no. 4, pp. 546–563, 2002.

10

MAJID ZAMANI, GIORDANO POLA, MANUEL MAZO JR., AND PAULO TABUADA

[16] L. Habets, P. Collins, and J. V. Schuppen, “Reachability and control synthesis for piecewise-affine hybrid systems on simplices,” IEEE Transactions on Automatic Control, vol. 51, no. 6, pp. 938–948, 2006. [17] C. Belta and L. Habets, “Controlling a class of nonlinear systems on rectangles,” IEEE Transactions on Automatic Control, vol. 51, no. 11, pp. 1749–1759, 2006. [18] G. Reißig, “Computation of discrete abstractions of arbitrary memory span for nonlinear sampled systems,” in Proc. of 12th Int. Conf. Hybrid Systems: Computation and Control (HSCC), vol. 5469, pp. 306–320, April 2009. [19] G. Pola, A. Girard, and P. Tabuada, “Approximately bisimilar symbolic models for nonlinear control systems,” Automatica, vol. 44, no. 10, pp. 2508–2516, 2008. [20] G. Pola and P. Tabuada, “Symbolic models for nonlinear control systems: alternating approximate bisimulations,” SIAM Journal on Control and Optimization, vol. 48, no. 2, pp. 719–733, February 2009. [21] G. Pola, P. Pepe, M. Di Benedetto, and P. Tabuada, “Symbolic models for nonlinear time-delay systems using approximate bisimulations,” Systems and Control Letters, vol. 59, pp. 365–373, 2010. [22] A. Girard, G. Pola, and P. Tabuada, “Approximately bisimilar symbolic models for incrementally stable switched systems,” IEEE Transactions on Automatic Control, vol. 55, no. 1, pp. 116–126, January 2009. [23] E. D. Sontag, Mathematical control theory, 2nd ed. New York: Springer-Verlag, 1998, vol. 6. [24] D. Angeli and E. D. Sontag, “Forward completeness, unboundedness observability, and their lyapunov characterizations,” Systems and Control Letters, vol. 38, pp. 209–217, 1999. [25] A. Girard and G. J. Pappas, “Approximation metrics for discrete and continuous systems,” IEEE Transactions on Automatic Control, vol. 25, no. 5, pp. 782–798, 2007. [26] D. Angeli, “A lyapunov approach to incremental stability properties,” IEEE Transactions on Automatic Control, vol. 47, no. 3, pp. 410–21, 2002. [27] L. Praly and Y. Wang, “Stabilization in spite of matched unmodeled dynamics and an equivalent definition of input-to-state stability,” Math. Control Signal Systems, vol. 9, pp. 1–33, 1996. [28] W. Lohmiller and J. J. Slotine, “On contraction analysis for non-linear systems,” Automatica, vol. 34, no. 6, pp. 683–696, 1998. [29] J. Jouffroy, “Some ancestors of contraction analysis,” Proceedings of the 44th IEEE Conference on Decision and Control, pp. 5450–5455, December 2005. [30] J. M. Lee, Introduction to Smooth Manifolds. Springer-Verlag, 2003. [31] N. Aghannan and P. Rouchon, “An intrinsic observer for a class of lagrangian systems,” IEEE Transactions on Automatic Control, vol. 48, no. 6, pp. 936–945, 2003. [32] A. Borri, G. Pola, and M. Di Benedetto, “An integrated approach to the symbolic control design of nonlinear systems with infinite states specifications,” in Proceedings of 49th IEEE Conference on Decision and Control, Atlanta, USA, December 2010. [33] Y. Tazaki and J. Imura, “Discrete-state abstractions of nonlinear systems using multi-resolution quantizer,” in Proc. of 12th Int. Conf. Hybrid Systems: Computation and Control (HSCC), vol. 5469, pp. 351–365, April 2009. [34] P. Tabuada, “An approximate simulation approach to symbolic control,” IEEE Transactions on Automatic Control, vol. 53, no. 6, pp. 1406–1418, July 2008. [35] A. Girard, “Approximately bisimilar finite abstractions of stable linear systems,” in Hybrid Systems: Computation and Control, ser. Lecture Notes in Computer Science. Berlin: Springer, May 2007, vol. 4416, pp. 231–244. [36] S. Prajna, A. Papachristodoulou, P. Seiler, and P. A. Parrilo, “Sostools: Control applications and new developments,” in Proceedings of IEEE International Symposium on Computer Aided Control Systems Design, pp. 315–320, 2004. [37] Pessoa, “Electronically available at: http://www.cyphylab.ee.ucla.edu/pessoa,” October 2009. [38] H. K. Khalil, Nonlinear systems, 2nd ed. New Jersey: Prentice-Hall, Inc., 1996. [39] P. Petersen, Riemannian Geometry. Springer, 1997.

7. Appendix I Example 7.1. We illustrate the results in Theorem 4.1 on an inverted pendulum shown in Figure 1. We have the following model for the pendulum:  x˙ 1 = x2 , (7.1) Σ: h 1 x˙ 2 = gl sin(x1 ) − ml 2 x2 + ml cos(x1 )u, where x1 is the angular position, x2 is the angular velocity of the point mass, u is the applied force (control input), g = 9.8 is gravity’s acceleration, l = 0.5 is the length of the rod, m = 0.5 is the mass, and h = 2 is the rotational friction coefficient. All constants and variables are expressed in the International System. The eigenvalues of the linearized system around the equilibrium point (0, 0) are λ1 = 1.1433 and λ2 = −17.1433 showing that the original nonlinear system is unstable at (0, 0). Hence, the results in [34, 19, 20, 21, 35] do not apply to this system. We assume that u ∈ U = [−6, 6]. We work on the subset X = [− π2 , π2 ] × [−1, 1] of the

SYMBOLIC MODELS FOR NONLINEAR CONTROL SYSTEMS WITHOUT STABILITY ASSUMPTIONS

11

m x1 l

u

Figure 1. Inverted pendulum mounted on a cart. Table 1. The constants ahijk defining the δ-FC Lyapunov function V for the inverted pendulum. ahijk -0.0639 -0.0639 -0.20708 -0.14087 1.138 1.138 2.3092 2.3092 -2.2813 -4.6848 2.2001 2.2001 -2.1876 -2.2813 0.00228 -0.01408 0.01634 0.02707 -0.09059 0.01094 -0.01807 -0.00637

h 0 3 0 0 0 0 2 0 1 1 0 1 0 0 1 4 2 1 1 0 1 1

i 0 1 1 0 2 0 0 0 1 0 1 0 1 0 0 0 1 2 1 0 0 0

j 3 0 3 4 0 0 0 2 0 1 1 0 0 1 3 0 1 1 2 2 0 0

k 1 0 0 0 0 2 0 0 0 0 0 1 1 1 0 0 0 0 0 2 3 3

ahijk 0.02482 0.03899 -0.05880 -0.01634 0.02482 -0.0588 0.02707 0.03899 0.01094 -0.00636 -0.01807 0.001563 -0.02071 0.00228 -0.17639 0.05159 0.01501 0.05159 0.08251 -0.09059 0.01501

h 2 1 0 1 0 1 1 0 2 1 0 2 3 3 1 0 0 2 0 2 0

i 1 2 2 0 1 1 0 1 2 3 3 0 0 0 1 2 3 0 2 0 1

j 0 0 1 2 2 0 1 1 0 0 1 2 0 1 1 2 0 0 0 1 0

k 1 1 1 1 1 2 2 2 0 0 0 0 1 0 1 0 1 2 2 1 3

state space of Σ. In order to construct a symbolic abstraction for the preceding model, we need to find functions β and γ satisfying the incremental forward completeness property in (2.2). We use SOS programming [36] to P 0k find a δ-FC Lyapunov function of the form V ((x1 , x2 ), (x01 , x02 )) = ahijk xh1 xi2 x0j 1 x2 , 0 ≤ h + i + j + k ≤ 4, for the inverted pendulum described above. The resulting V is defined by the constants ahijk provided in Table 1. Function V satisfies conditions (i) and (ii) in Definition 5.1 with K∞ functions: α(r) = 0.5r2 , α(r) = 3r2 , 2. Using the results from Theorem 5.2, the functions β and γ are given σ(r) = 5r2 and √ positive constant κ√= √ by β(r, t) = 2 3et r and γ(r, t) = 2 5 e2t − 1r. Our objective is to design a controller forcing the trajectories of the system to reach the target set W = [−0.25, 0.25] × [−1, 1] and to remain indefinitely inside W . We furthermore assume that the controller is implemented on a microprocessor that is executing other tasks in addition to the control task. We consider a periodic schedule with epochs of three time slots where the first two time slots are allocated to the control tasks and the third time slot to another task. The expression time slot refers to a time interval of the form [kτ, (k + 1)τ [ with k ∈ N and where τ is the time quantization parameter. Therefore, the microprocessor schedule is given by: |aau|aau|aau|aau|aau|aau|aau|aau|aau|aau| · · ·

12

MAJID ZAMANI, GIORDANO POLA, MANUEL MAZO JR., AND PAULO TABUADA

q1 a

q2 a

q3 u

Figure 2. Ssc describing the schedulability constraints. The lower part of the states are labeled with the outputs a and u denoting availability and unavailability of the microprocessor, respectively.

Figure 3. Upper and central panels: evolution of x1 and x2 with initial condition (0.7, 0). Lower panel: input signal. where a denotes a slot available for the control task and u denotes a slot allotted to a different task. The symbol | separates each epoch of three time slots. The schedulability constraint on the microprocessor can be represented by the finite system Ssc in Figure 2. When Ssc is in state q1 or q2 the microprocessor computes the control input for the inverted pendulum. On the other hand, when Ssc is in state q3 , the microprocessor computes another task. Although we can easily consider more complex schedules, the constraints described by Ssc in Figure 2 already illustrate the computational constraints imposed by implementing control laws on shared microprocessors. For a precision ε = 0.01, we construct a symbolic model Sq (Σ) by choosing θ = 0.01, η = 0.02, τ = 0.5, and µ = 0.4 so that assumptions of Theorem 4.1 are satisfied. The computation of the abstraction Sq (Σ) was performed in the tool5 Pessoa [37]. A controller enforcing the specification is found by performing simple fixedpoint computations on Sq (Σ) using standard algorithms from game theory [1]. We solved a reachability game and a safety game, both implemented in Pessoa, to reach and stay indefinitely in the target set, respectively. In Figure 3, we show the closed-loop trajectory stemming from the initial condition (0.7, 0) and the evolution of the input signal. The domain of the synthesized controller for different initial states of the finite system modeling the schedulability of the microprocessor is shown in Figure 4. 5Pessoa can be freely downloaded from http://www.cyphylab.ee.ucla.edu/pessoa. All the files necessary to recreate this example are also available on Pessoa’s website.

SYMBOLIC MODELS FOR NONLINEAR CONTROL SYSTEMS WITHOUT STABILITY ASSUMPTIONS

1 0.8 0.6 0.4

x2

0.2 0 ï0.2 ï0.4 ï0.6 ï0.8 ï1 ï0.8

ï0.6

ï0.4

ï0.2

0 x1

0.2

0.4

0.6

0.8

0.2

0.4

0.6

0.8

0.2

0.4

0.6

0.8

(a) 1 0.8 0.6 0.4

x2

0.2 0 ï0.2 ï0.4 ï0.6 ï0.8 ï1 ï0.8

ï0.6

ï0.4

ï0.2

0 x1

(b) 1 0.8 0.6 0.4

x2

0.2 0 ï0.2 ï0.4 ï0.6 ï0.8 ï1 ï0.8

ï0.6

ï0.4

ï0.2

0 x1

(c) Figure 4. Domain of the controller forcing the inverted pendulum to reach and remain in [−0.25, 0.25] × [−1, 1] under the constraint described by Ssc in Figure 2. The different figures in (a), (b), and (c) correspond to Ssc initialized from states q1 , q2 , or q3 , respectively.

13

14

MAJID ZAMANI, GIORDANO POLA, MANUEL MAZO JR., AND PAULO TABUADA

8. Appendix II Proof of Theorem 5.2. The proof is inspired by the work in [24]. By using property (i) in Definition 5.1, we obtain: (8.1)

kξxυ (t) − ξx0 υ0 (t)k ≤ α−1 (V (ξxυ (t), ξx0 υ0 (t))) .

By using the property (ii) and the comparison lemma [38], one gets: (8.2)

V (ξxυ (t), ξx0 υ0 (t)) ≤ eκt V (ξxυ (0), ξx0 υ0 (0)) + eκt ∗ σ(kυ(t) − υ 0 (t)k)

where ∗ denotes the convolution integral6. By combining inequalities (8.1) and (8.2), one gets:
kξxυ (t) − ξx0 υ0 (t)k ≤ α−1 eκt V (x, x0 ) + eκt ∗ σ(kυ − υ 0 k)   eκt − 1 ≤ α−1 eκt V (x, x0 ) + σ(kυ − υ 0 k∞ ) = γ(ρ, φ) κ where γ(ρ, φ) = α−1 (ρ + φ), ρ = eκt V (x, x0 ), and φ = variable, we have: κt

eκt −1 κ σ (kυ

0


kξxυ (t) − ξx0 υ0 (t)k ≤ h e V (x, x ) + h



− υ 0 k∞ ). Since γ is nondecreasing in each

eκt − 1 σ (kυ − υ 0 k∞ ) κ



+ 0 0 where h(r) = γ(r, r) = α−1 (2r) and h : R+ 0 → R0 is a K∞ function. Moreover, using V (x, x ) ≤ α(kx − x k), one obtains:   κt
e −1 0 −1 κt 0 −1 σ(kυ − υ k∞ ) . kξxυ (t) − ξx0 υ0 (t)k ≤ α 2e α(kx − x k) + α 2 κ

Therefore, by defining functions β and γ as β(kx − x0 k, t) γ(kυ − υ 0 k∞ , t)


= α−1 2eκt α (kx − x0 k)  κt  e −1 −1 0 2 = α σ(kυ − υ k∞ ) , κ

the condition (2.2) is satisfied. Hence, the system Σ is δ-FC.



Proof of Theorem 5.5. The proof of Theorem 5.5 requires the following preliminary result. Lemma 8.1. The Riemannian manifold Rn equipped with a Riemannian metric G, satisfying7 ωkyk22 ≤ y T G(x)y for any x, y ∈ Rn and for some positive constant ω, is complete as a metric space, with respect to the Riemannian distance determined by G. Proof. The proof was suggested to us by C. Manolescu. First, for each pair of points x, y ∈ Rn we define the path space: Ω(x, y) = {χ : [0, 1] → Rn | χ is piecewise smooth, χ(0) = x, and χ(1) = y}. Recall that a function χ : [a, b] → Rn is piecewise smooth if χ is continuous and if there exists a partition a = a1 < a2 < · · · < ak = b of [a, b] such that χ|[ai ,ai+1 ] is smooth for i = 1, · · · , k − 1. We can then define the Riemannian distance dG (x, y) between points x, y ∈ Rn as s T Z 1  dχ(s) dχ(s) dG (x, y) = inf G(χ(s)) ds. ds ds χ∈Ω(x,y) 0 6eκt ∗ σ(kυ(t) − υ 0 (t)k) = R t eκ(t−τ ) σ(kυ(τ ) − υ 0 (τ )k)dτ . 0

7This condition is nothing more than uniform positive definitness of G.

SYMBOLIC MODELS FOR NONLINEAR CONTROL SYSTEMS WITHOUT STABILITY ASSUMPTIONS

15

It follows immediately that dG is a metric on Rn . The Riemannian manifold Rn is a complete metric space, equipped with the metric dG , if every Cauchy sequence8 of points in Rn has a limit in Rn . Assume {xn }∞ n=1 is a Cauchy sequence in Rn , equipped with the metric dG . By using the assumption on G, we have s T Z 1  dχ(s) dχ(s) dG (xn , xm ) = inf G(χ(s)) (8.3) ds ds ds χ∈Ω(xn ,xm ) 0 s T Z 1  dχ(s) dχ(s) √ √ ≥ ω inf ds = ωkxn − xm k2 , ds ds χ∈Ω(xn ,xm ) 0 n It is readily seen from the inequality (8.3) that the sequence {xn }∞ n=1 is also a Cauchy sequence in R with n respect to the Euclidean metric. Since the Riemannian manifold R with respect to the Euclidean metric is ∗ n a complete metric space, the sequence {xn }∞ n=1 converges to a point, named x , in R . By picking a convex n ∗ compact subset D ⊂ R , containing x , and using Lemma 8.18 in [30], we have: ωkyk22 ≥ y T G(x)y for any ∗ y ∈ Rn , x ∈ D, and some positive constant ω. Since the sequence {xn }∞ n=1 converges to x ∈ D, there exists ∞ some integer N such that the sequence {xn }n=N remains forever inside D. Hence, we have: √ √ (8.4) ωkxn − x∗ k2 ≤ dG (xn , x∗ ) ≤ ωkxn − x∗ k2 , ∗ n for n > N . Therefore, the sequence {xn }∞ n=1 converges to x ∈ R , equipped with the metric dG . Therefore, n R with respect to the metric dG is a complete metric space.

 By using the assumption on the metric and Lemma 8.1, we know that Rn is a complete metric space with respect to the Riemannian distance determined by G. Using the Hopf-Rinow theorem [39], we conclude that Rn with respect to the metric G is geodesically complete. The rest of proof is inspired by the proof of Theorem 2 in [31]. Consider two points x and x0 in Rn and a geodesic χ : [0, 1] → Rn joining x = χ(0) and x0 = χ(1). The geodesic distance between the points x and x0 is given by: s T Z 1  dχ(s) dχ(s) (8.5) dG (x, x0 ) = G(χ(s)) ds. ds ds 0 0 Consider the straight line χ bt (s) = (1 − s)υ(t) + sυ 0 (t), for fixed t ∈ R+ 0 , fixed υ, υ ∈ U, and for any s ∈ [0, 1]. The curve χ bt is a geodesic, with respect to the Euclidean metric, on the subset U ⊆ Rm joining υ(t) = χ bt (0) 0 and υ (t) = χ bt (1). Consider also the input curve υs defined by υs (t) = χ bt (s). Let l(t) be the length of the curve ξχ(s)υs (t) parametrized by s and with respect to the metric G, i.e.:

Z (8.6)

l(t) = 0

1

q

δξ T G(ξχ(s)υs (t))δξds, s.t. δξ =

∂ ξχ(s)υs (t). ∂s

In the rest of the proof, we drop the argument of the metric G for simplicity. By taking the derivative of (8.6) with respect to time, we obtain:
Z 1 d T d dt δξ Gδξ p l(t) = ds dt 2 δξ T Gδξ 0     T T ∂f ∂f ∂G T ∂f Z 1 δξ T G + f + G δξ + 2δυ Gδξ ∂x ∂x ∂x ∂u ∂ p = ds, s.t. δυ = υs (t). T ∂s 2 δξ Gδξ 0 8A sequence {x }∞ in a metric space X, equipped with a metric d, is a Cauchy sequence if lim n n=1 n,m→∞ d(xn , xm ) = 0.

16

MAJID ZAMANI, GIORDANO POLA, MANUEL MAZO JR., AND PAULO TABUADA

Since Σ is an exponential expansion with λ and α the constants introduced in Definition 5.4, the following inequality holds: Z d λ α 1√ T (8.7) l(t) ≤ l(t) + δυ δυds dt 2 2 0 λ α = l(t) + kυ(t) − υ 0 (t)k2 . 2 2 √ Using the inequality kυ(t)k2 ≤ mkυ(t)k, in which m denotes the dimension of the input space, one gets: √ λ mα λ t l(t) ≤ e 2 t l(0) + e 2 ∗ kυ(t) − υ 0 (t)k 2 √  λ mα  λ t t 2 e 2 − 1 kυ − υ 0 k∞ , ≤ e l(0) + λ where ∗ denotes the convolution integral. From (8.5) and (8.6), it can be seen that l(0) = dG (x, x0 ). However, for t ∈ R+ , l(t) is not necessarily the Riemannian distance, determined by G, because ξχυs (t) is not necessarily a geodesic. Nevertheless, dG (ξxυ (t), ξx0 υ0 (t)) ≤ l(t), and one has: √  λ mα  λ t e 2 − 1 kυ − υ 0 k∞ . (8.8) dG (ξxυ (t), ξx0 υ0 (t)) ≤ e 2 t dG (x, x0 ) + λ Using the assumptions on the metric, it can be readily checked that: √ ωkξxυ (t) − ξx0 υ0 (t)k ≤ dG (ξxυ (t), ξx0 υ0 (t)) , √ ωnkx − x0 k, dG (x, x0 ) ≤ where n denotes the dimension of the state space. Hence, the condition (8.8) reduces to: r r  ωn λ t m α  λt 0 2 (8.9) kξxυ (t) − ξx0 υ0 (t)k ≤ e 2 − 1 kυ − υ 0 k∞ , e kx − x k + ω ω λ which is the δ-FC condition in (2.2).

1 Department



of Electrical Engineering, University of California at Los Angeles, Los Angeles, CA 90095

E-mail address: {zamani, tabuada}@ee.ucla.edu URL: http://www.ee.ucla.edu/~zamani URL: http://www.ee.ucla.edu/~tabuada 2 Department

of Electrical and Information Engineering, Center of Excellence DEWS, University of LAquila, Poggio di Roio, 67040 LAquila, Italy E-mail address: [email protected] URL: http://www.diel.univaq.it/people/pola 3 INCAS3 , Dr. Nassaulaan 9, 9401 HJ Assen, The Netherlands and the Faculty of Mathematics and Natural Sciences, ITM, University of Groningen, Groningen, 9747AG, The Netherlands

E-mail address: [email protected] URL: http://www.rug.nl/staff/m.mazo/index