Ultimate Boundedness for Switched Systems with Multiple Equilibria Under Disturbances

$\mathbb{R}$ and $\mathbb{Z}$ denote the real and integer numbers, and $\mathbb{R}_{+}$, $\mathbb{Z}_{+}$ the non-negative reals and integers, respectively. The Euclidean norm is denoted by $\|\cdot\|$ and $B_{\delta}(x)\subset\mathbb{R}^{n}$ denotes an open-ball (Euclidean) of radius $\delta>0$ centered at $x\in\mathbb{R}^{n}$. Let $\mathcal{A}\subseteq\mathbb{R}^{n}$, then $\accentset{\circ}{\mathcal{A}}$ denotes the interior while $\overline{\mathcal{A}}$ denotes the closure of $\mathcal{A}$, respectively. The index $k\in\mathbb{Z}_{+}$ represents discrete time. The discrete-time disturbance $d:\mathbb{Z}_{+}\to\mathbb{R}^{m}$ is a sequence $\{d_{k}\}_{k\in\mathbb{Z}_{+}}$ with $d_{k}\in\mathbb{R}^{m}$ for $k\in\mathbb{Z}_{+}$. The norm of $d$ is $\|d\|_{\infty}:=\sup_{k\in\mathbb{Z}_{+}}\|d_{k}\|$. Let $t\in\mathbb{R}_{+}$ represent the continuous time. The disturbance $d:\mathbb{R}_{+}\to\mathbb{R}^{m}$ that acts in continuous time is assumed to be a piecewise continuous signal with norm $\|d\|_{\infty}:=\sup_{t\geq 0}\|d(t)\|$. Abusing notation, we use $d$, $\|\cdot\|_{\infty}$ and $\mathcal{D}$ for both discrete- and continuous-time disturbances. No ambiguity arises as it will always be clear from context whether the signal is discrete or continuous. Finally, a function $\alpha:\mathbb{R}_{+}\to\mathbb{R}_{+}$ is of class $\mathcal{K}_{\infty}$ if it is continuous, strictly increasing, $\alpha(0)=0$, and $\lim_{s\to\infty}\alpha(s)=\infty$. A function $\beta:\mathbb{R}_{+}\times\mathbb{R}_{+}\to\mathbb{R}_{+}$ is of class $\mathcal{KL}$ if it is continuous, $\beta(\cdot,t)$ is of class $\mathcal{K}_{\infty}$ for any fixed $t\geq 0$, $\beta(s,\cdot)$ is strictly decreasing, and $\lim_{t\to\infty}\beta(s,t)=0$ for any fixed $s\geq 0$; see [36] for class $\mathcal{K}_{\infty}$ and $\mathcal{KL}$ functions.

Let $\mathcal{P}$ be a finite index set and consider the family of discrete-time systems

where $x\in\mathbb{R}^{n}$ is the state of the system and $d_{k}\in\mathbb{R}^{m}$ is the value at time $k$ of the discrete disturbance signal $d$, which belongs to the set of bounded disturbances $\mathcal{D}:=\{d:\mathbb{Z}_{+}\to\mathbb{R}^{m}~|~\|d\|_{\infty}<\infty\}$. It is assumed that, for each $p\in\mathcal{P}$, the mapping $f_{p}:\mathbb{R}^{n}\times\mathbb{R}^{m}\to\mathbb{R}^{n}$ is continuous in its arguments, and that there exists a unique $x_{p}^{*}\in\mathbb{R}^{n}$ which satisfies $x_{p}^{*}=f_{p}(x_{p}^{*},0)$. Note that the vast majority of the relevant literature assumes that all subsystems $f_{p}$ share a common equilibrium point; here, we relax this assumption, and allow for $x_{p}^{*}\neq x_{q}^{*}$ when $p\neq q$.

To state the main result, we will require each system in the family (1) to be input-to-state stable, as defined below.

Let $\sigma:\mathbb{Z}_{+}\to\mathcal{P}$ be a switching signal, mapping the discrete time $k$ to the index $\sigma(k)\in\mathcal{P}$ of the subsystem that is active at $k$. This gives rise to a discrete switched system of the form

We are interested in establishing boundedness and ultimate boundedness of the solutions of (3) under bounded disturbances, provided that the switching signal is sufficiently “slow on average”. Definition 2 below makes this notion precise.

We can now state the main result of this section for discrete switched systems.

Let us now briefly discuss some aspects of Theorem 1. First, note that by [37, Definition 3.2], conditions (5)-(6) imply that $V_{p}$ is an ISS-Lyapunov function for the $p$-th subsystem, which by [37, Lemma 3.5], entails that $f_{p}$ is ISS. Since we are interested in switching among systems from the family (1), condition (7) is added to ensure that the ratio of the Lyapunov functions corresponding to the subsystems involved in switching is bounded. With conditions (5)-(6) and (7) in place, Theorem 1 states that if each system in the family (1) is ISS, the solution of (3) is uniformly²² 2 The term “uniformly” refers to uniformity over the set of switching signals that satisfy (4) for $N_{0}\geq 1$ and $N_{\rm a}\geq\overline{N}_{\rm a}$, as required by (8) of Theorem 1. bounded and uniformly ultimately bounded within the compact set $\mathcal{M}(\bar{\omega})$ characterized by (10). Furthermore, (10) indicates that the “size” of $\mathcal{M}(\bar{\omega})$ reduces proportionally with the norm $\|d\|_{\infty}$ of the disturbance. Note, however, that $\mathcal{M}(\bar{\omega})$ does not collapse to a point when the disturbance signal $d$ vanishes; indeed, if $d=0$, (11) implies that $\bar{\omega}(0)=c>0$ and the solutions of (3) are ultimately bounded to the zero-input compact set $\mathcal{M}(c)$ that contains the equilibria $x^{*}_{p}\in\mathcal{M}(c),~\forall p\in\mathcal{P}$.

It should be noted that Theorem 1 does not establish ISS for (3), since the compact set $\mathcal{M}(c)$ is not positively invariant³³ 3 In the terminology of [38], $\mathcal{M}(\bar{\omega})$ is not a zero-invariant set for (3). under the zero-input dynamics of (3). In fact, for suitable switching signals satisfying the requirements of the theorem, solutions of (3) can be found that start within $\mathcal{M}(c)$ and—while evolving in the absence of the disturbance—escape from $\mathcal{M}(c)$ before they return to $\mathcal{M}(c)$ and be trapped forever in it; see also Remark 2 for how this behavior can emerge. Note also that although the estimate (9) is reminiscent of (2) in Definition 1, it extends only up to a finite integer $K$, and it does not represent point-to-set distance from $\mathcal{M}(c)$ as establishing set-ISS for (3) would require. However, when all the subsystems in the family (1) share the same equilibrium, then ISS can be recovered, as the following corollary shows. Corollary 1 provides the counterpart of [8, Theorem 3.1] for discrete switched systems.

While the set $\mathcal{M}(\bar{\omega})$ in Theorem 1 is not positively invariant, one can identify a (compact) subset of initial conditions in $\mathcal{M}(c)$ such that the corresponding solutions never leave $\mathcal{M}(\bar{\omega})$. This property corresponds to the notion of practical stability in the terminology of [19], and is important in certain motion planning applications; e.g., see [11, 12] for planning motions with limit-cycle walking bipedal robots. This is made precise by the following corollary.

As in Section II-A, let $\mathcal{P}$ be a finite index set and consider the family of continuous-time systems

where $x\in\mathbb{R}^{n}$ is the state of the system and $d(t)\in\mathbb{R}^{m}$ is the value of the continuous-time disturbance signal $d$ at time $t$ which belongs to the set of bounded disturbances $\mathcal{D}:=\{d:\mathbb{R}_{+}\to\mathbb{R}^{m}~|~\|d\|_{\infty}<\infty,~d~{\rm piecewise~continuous}\}$. It is assumed that, for each $p\in\mathcal{P}$, the vector field $f_{p}:\mathbb{R}^{n}\times\mathbb{R}^{m}\to\mathbb{R}^{n}$ is locally Lipschitz in its arguments, and that there exists a unique $x_{p}^{*}\in\mathbb{R}^{n}$ with $0=f_{p}(x_{p}^{*},0)$. As in Section II-A, we allow for $x_{p}^{*}\neq x_{q}^{*}$ when $p\neq q$.

Analogous to Section II-A, we will require each system in the family (13) to be input-to-state stable, as defined below.

Let $\sigma:\mathbb{R}_{+}\to\mathcal{P}$ be a switching signal mapping the time instant $t$ to the index $\sigma(t)\in\mathcal{P}$ of the subsystem that is active at $t$. It is assumed that $\sigma(t)$ is right-continuous. The switching signal gives rise to the continuous-time switched system

The solution $x(t):=\phi(t,x(0),\sigma(t),d(t))$ of (15) is a sequential concatenation of each subsystem’s solution as governed by the switching signal. Let $\{t_{n}\}_{n=1}^{\infty}$ with $t_{n}\in\mathbb{R}_{+}$ be a strictly monotonically increasing sequence of switching times. Clearly, continuity of $f_{p}(x,d)$ and piecewise continuity of $d(t)$ imply that $x(t)$ is continuous over $(t_{n},t_{n+1})$, i.e. between subsequent switches. Furthermore, for any $t_{n}$, the subsystem $f_{\sigma(t_{n})}$ that is switched in and is active over $[t_{n},t_{n+1})$ is initialized by $x(t_{n})=\lim_{t\nearrow t_{n}}x(t)$ ensuring that $x(t)$ is continuous at $t_{n}$. Hence, $x(t)$ is continuous for all $t\geq 0$.

As in the discrete-time case, the main result of this section is stated for switching signals $\sigma$ with sufficiently slow switching on average; the following definition formalizes this notion.

We are now in a position to state the main result of this section for continuous switched systems.

Suppose that $V_{p}$ is a function satisfying (5)-(6) in the case of discrete or (17)-(18) in the case of continuous switched systems. The $\kappa$-sublevel set of $V_{p}$ is defined as

and the union of the sublevel sets over $\mathcal{P}$ is denoted as

Next, we define a positive constant

which is well defined since $\mathcal{M}(\kappa)$ is compact for any $\kappa>0$ and $\mathcal{P}$ is finite. Intuitively, the definition of $\omega(\kappa)$ by (26) enlarges each sublevel set $\mathcal{M}_{p}(\kappa)$ so that the resulting enlarged set $\mathcal{M}_{p}(\omega(\kappa))$ includes the sets $\mathcal{M}_{q}(\kappa)$ for all $q\in\mathcal{P}$. An illustration of this construction can be seen in Fig. 1 and the following remark makes this intuition precise.

We now establish a relationship between the values $V_{p}(x)$ and $V_{q}(x)$ of any pair $V_{p},V_{q}$ of ISS-Lyapunov functions at a given point $x\in\mathbb{R}^{n}$ as the system switches between the corresponding two subsystems $p,q\in\mathcal{P}$, $p\neq q$. Consider the ratio $V_{q}(x)/V_{p}(x)$, and let

which is bounded as shown in the following arguments. Let $a_{1}:=\limsup_{\|x-x_{p}^{*}\|\to\infty}V_{q}(x)/V_{p}(x)$ which is bounded by (7) and (19). As a consequence of [39, Theorem 3.17] it follows that there exists a $r>0$ such that for any $x$ with $\|x-x_{p}^{*}\|>r$, we have $V_{q}(x)/V_{p}(x)\leq a_{1}+1$. Expand $r$ if necessary to ensure that $\mathcal{M}(\kappa)\subset B_{r}(x_{p}^{*})$. Note that $V_{q}(x)/V_{p}(x)$ is continuous on $\mathbb{R}^{n}\setminus\{x_{p}^{*}\}$ hence it is also continuous on $\overline{B}_{r}(x_{p}^{*})\setminus\accentset{\circ}{\mathcal{M}}_{p}(\kappa)\subset\mathbb{R}^{n}\setminus\{x_{p}^{*}\}$ which is compact. Then there exists $a_{2}>0$ such that $V_{q}(x)/V_{p}(x)<a_{2}$ for any $x\in\overline{B}_{r}(x_{p}^{*})\setminus\accentset{\circ}{\mathcal{M}}_{p}(\kappa)$. Therefore $V_{q}(x)/V_{p}(x)<\max\{a_{1}+1,a_{2}\}$ for all $x\not\in\accentset{\circ}{\mathcal{M}}_{p}(\kappa)$ ensuring the boundedness of $\mu_{p}(\kappa)$ in (27).

This constant provides a bound on how much the value of the Lyapunov function can change on switching. Clearly,

To make this bound independent of $p$, let

which implies that

Due to the interchangeability of the indices $p$ and $q$, it also holds that $V_{p}(x)\leq\mu(\kappa)V_{q}(x)$ as long as $x\notin\accentset{\circ}{\mathcal{M}}_{q}(\kappa)$. Hence, when $x\notin\accentset{\circ}{\mathcal{M}}_{p}(\kappa)\cup\accentset{\circ}{\mathcal{M}}_{q}(\kappa)$, we can write $V_{q}(x)\leq\mu(\kappa)^{2}V_{q}(x)$, from which it follows that

since $V_{q}$ is positive definite for $x\notin\accentset{\circ}{\mathcal{M}}_{p}(\kappa)\cup\accentset{\circ}{\mathcal{M}}_{q}(\kappa)$. Finally, in the context of the switched system (3), it is worth noting that (29) holds for a switching instant $k_{n}$ even if $x_{k_{n}}\in\accentset{\circ}{\mathcal{M}}(\kappa)$, as long as $x_{k_{n}}\notin\accentset{\circ}{\mathcal{M}}_{\sigma(k_{n}-1)}(\kappa)$. A similar statement can be made for the continuous switched system (15).

The aforementioned set constructions allow us to provide explicit expressions for the average dwell-time bound $\overline{N}_{\rm a}$ in (8) and (20), and for the characterization of the sets $\mathcal{M}(\bar{\omega})$ in (10) and (22) of Theorems 1 and 2, respectively. Although these expressions are derived in the proofs of Section IV below, we provide them here to ease their use in applications.

For the sake of notational convenience in (6), let $\lambda:=\max_{p\in\mathcal{P}}\lambda_{p}$ and $\hat{\alpha}(\|d\|_{\infty}):=\max_{p\in\mathcal{P}}\{\alpha_{p}(\|d\|_{\infty})\}$. Then, for any $p\in\mathcal{P}$ we can write (6) as

where $\lambda\in(0,1)$ and $\hat{\alpha}$ are independent of $p$.

Let $\delta$ be any constant that lies within $(\lambda,1)$, then it will be shown in the proof of Theorem 1 that the lower bound on the average dwell-time, i.e. $\overline{N}_{\rm a}$ in (8), is

where the constant $\mu(\kappa)\geq 1$ is defined by (28). Furthermore, the compact set $\mathcal{M}(\bar{\omega})$ in Theorem 1(ii) is characterized by

from which the constant $c>0$ and the class-$\mathcal{K}_{\infty}$ function $\tilde{\alpha}$ participating in (11) can be readily identified.

It is remarked that the constants $\delta$ and $\kappa$ are design parameters available for tuning the frequency of the switching signal. The choice of $\delta\in(\lambda,1)$ provides a tradeoff between robustness of the switched system (3) and the switching frequency. In more detail, if $\delta$ is chosen close to $\lambda$, the lower bound on the average dwell-time (32) becomes large, thus limiting the number of switches $N_{\sigma}(k,\underline{k})$ in any time interval $[\underline{k},k)$, as (4) implies. On the other hand, if $\delta$ is chosen close to $1$, the size of the compact set $\mathcal{M}({\bar{\omega}})$ to which the solutions ultimately converge increases as $\bar{\omega}$ in (33) increases. As a result, the solutions of (3) are permitted to wander in a larger set, indicating low robustness of (3) to disturbances. Hence, slower switching signals result in tighter trapping regions.

The effect of $\kappa$ is more involved. Picking smaller values of $\kappa$ results in inner sublevel level sets of $V_{p}$, thus reducing the size of $\mathcal{M}(\kappa)$ in (25) and causing $\mu_{p}(\kappa)$ to increase as the supremum in (27) increases. Hence, smaller values of $\kappa$ result in larger values of $\mu(\kappa)$ by (28), leading to slower switching frequencies as well as larger compact attractive sets $\mathcal{M}({\bar{\omega}})$; observe the role of $\mu(\kappa)$ in (32) and (33). On the other hand, picking a larger $\kappa$ will result in smaller $\mu(\kappa)$ allowing for faster switches, however, $\omega(\kappa)$ also increases making its effect on $\mathcal{\bar{\omega}}$ unclear; observe the $\mu^{N_{0}}\omega(\kappa)$ term in (33).

To simplify notation in (18), we define $\lambda:=\min_{p\in\mathcal{P}}\lambda_{p}>0$ and $\hat{\alpha}(\|d\|_{\infty}):=\max_{p\in\mathcal{P}}\{\alpha_{p}(\|d\|_{\infty})\}$. Then, for any $p\in\mathcal{P}$,

where $\lambda>0$ and $\hat{\alpha}(\|d\|_{\infty})$ are independent of $p$.

Let $\delta$ be any constant in the open interval $(0,\lambda)$, then the lower bound on the average dwell time $\overline{N}_{\rm a}$ in Theorem 2 is

where $\mu(\kappa)\geq 1$ is defined by (28).The compact set $\mathcal{M}(\bar{\omega})$ in Theorem 2(ii), within which solutions of (15) ultimately become trapped corresponds to

from which the constant $c>0$ and the class-$\mathcal{K}_{\infty}$ function $\tilde{\alpha}$ in (23) can be easily recognized.

As in the case of the discrete switched systems, the constant $\delta\in(0,\lambda)$ presents a tradeoff between the robustness of the system and the switching frequency; setting $\delta$ close to $0$ increases the disturbance term in (36) while setting $\delta$ close to $\lambda$ increases $\overline{N}_{\rm a}$ in (35). Regarding the effect of $\kappa$, the discussion is identical to that in Section III-B.

We begin by establishing an important estimate in the following lemma that will be used in the proof of Theorem 1.

The proof of the lemma can be found in Appendix A. Now we are ready to present the proof of Theorem 1.