For continuous systems, Lyapunov functions provide a general technique to establish stability. For example, the simple system $x' = -x$, a Lyapunov function is $V(x) = \frac{1}{2}x^2$. It is easy to see that

  1. $V(0) = 0$
  2. $V(x) > 0$ for $x \neq 0$
  3. $V(x)' = \frac{dV(x)}{dx}\frac{dx}{dt} = -x^2 < 0$ for $x\neq 0$

Is there an analogous technique for sampled-data systems?

For example, suppose I take the simple continuous system $x' = -x$ and naively turn it into a sampled-data system by introducing a zero-order hold. In other words, every $\delta$ time units, $x'$ is set to $-x$ and held at that value for the next $\delta$ time units. More precisely,

$$\forall k \in N, ~~~~ \forall t \in [k*\delta, (k+1)*\delta], ~~~~ x'(t) = -x(k*\delta)$$

Looking at the solutions to this system shows that it is asymptotically stable for $\delta < 2$. However, looking at the solutions of a system is only possible for simple systems. Is there instead a way of establishing asymptotic stability of this system using some sort of Lyapunov function?


For linear systems, you can consider the corresponding discrete-time system. In your example since $$\dot{x}(t)=-x(k\delta)\quad\forall t\in[k\delta,(k+1)\delta)$$ if we integrate over $[k\delta,(k+1)\delta)$ we obtain $$x((k+1)\delta)=x(k\delta)-\delta x(k\delta)=(1-\delta)x(k\delta)$$ which is asymptotically stable iff $$|1-\delta|<1$$ i.e. if $\delta<2$. This approach can be generalized to arbitrary linear systems. Applying a sampled-data control $$u(t)=u(k\delta)\qquad \forall t\in[k\delta,(k+1)\delta)$$ to a linear system of the form $$\dot{x}=Ax+Bu$$ creates the discrete-time system $$x((k+1)\delta)=A_d x(k\delta)+B_du(k\delta)$$ with $$A_d:=e^{A\delta}\\B_d:=\int_0^{\delta}{e^{A(\delta-s)}ds}\cdot B$$ If the pair $(A_d,B_d)$ is controllable then there exists some $K$ such that the eigenvalues of the matrix $A_d+B_dK$ lie in the interior of the unit circle. Selecting then the sampled-data state-feedback control law $$u(k\delta)=K\cdot x(k\delta)$$ we obtain the stable closed-loop discrete-time system $$x((k+1)\delta)=(A_d+B_d K) x(k\delta)$$ with $\lim_{k\rightarrow\infty}x(k\delta)=0$. For the state vector within the sampled-data interval it holds true that $$x(t)=\left[ e^{A(t-k\delta)}+\int_{k\delta}^t{e^{A(t-s)}ds}\cdot BK\right]x(k\delta)\qquad \forall t\in[k\delta,(k+1)\delta)$$ From the above equation, asymptotic stability can be proved now for the continuous time system ($\lim_{t\rightarrow\infty}x(t)=0$) using the fact that there exist some $C_1,C_2>0$ such that $$\sup_{t\in[k\delta,(k+1)\delta)}\|e^{A(t-k\delta)}\|=\sup_{\tau\in[0,\delta)}\|e^{A\tau}\|\leq C_1\\ \sup_{t\in[k\delta,(k+1)\delta)}\|\int_{k\delta}^t{e^{A(t-s)}ds}\cdot BK\|=\sup_{\tau\in[0,\delta)}\|\int_{k\delta}^{\tau+k\delta}{e^{A(\tau+k\delta-s)}ds}\cdot BK\|\\=\sup_{\tau\in[0,\delta)}\|\int_{0}^{\tau}{e^{A(\tau-w)}dw}\cdot BK\|\leq C_2$$ and $\lim_{k\rightarrow\infty}x(k\delta)=0$.

Edit: For varying-length sample time intervals $\delta_i$ ($i=1,2,\cdots$) let us define the sample times $$t_k:=\sum_{i=1}^k{\delta_i}$$ and the sampled and hold control law $$u(t)=K_k x\left(t_k\right)\quad t\in\left[t_k,t_{k+1}\right)$$ Then, the matrices describing the discrete-time system $A_d,B_d$ are not constant i.e. $$x(t_{k+1})=A_{d,k}x(t_k)+B_{d,k}u(t_k)$$ with $$A_{d,i}=e^{A\delta_{i+1}}\\ B_{d,i}:=\int_0^{\delta_{i+1}}{e^{A(\delta_{i+1}-s)}ds}\cdot B$$ The closed-loop system takes the form $$x(t_{k+1})=(A_{d,k}+B_{d,k}K_k)x(t_k)$$ If $$\sup_{k\in\mathbb{N}}\|A_{d,k}+B_{d,k}K_k\|=\epsilon<1$$ then $$\|x(t_{k+1})\|=\left\|\prod_{i=1}^k{(A_{d,i}+B_{d,i}K_i)} x(0) \right\|\leq \prod_{i=1}^k\|A_{d,i}+B_{d,i}K_i\| \|x(0)\|\leq \epsilon^k \|x(0)\|$$ and therefore $$\lim_{k\rightarrow\infty}x(t_k)=0$$ For the state vector within the sampled-data interval it holds true that $$x(t)=\left[ e^{A(t-t_k)}+\int_{t_k}^t{e^{A(t-s)}ds}\cdot BK_k\right]x(t_k)\qquad \forall t\in[t_k,t_{k+1})$$ Assuming an upper bound on the allowed time span between two samples i.e. $\delta_k\leq \bar{\delta}$ $\forall k\in\mathbb{N}$ for some $\bar{\delta}>0$ then there exist constants $C'_1,C'_2>0$ such that $$\sup_{k\in\mathbb{N}}\sup_{t\in[t_k,t_{k+1})}\|e^{A(t-t_k)}\|=\sup_{k\in\mathbb{N}}\sup_{\tau\in[0,\delta_{k+1})}\|e^{A\tau}\|\leq\sup_{\tau\in[0,\bar{\delta})}\|e^{A\tau}\|\leq C'_1\\ \sup_{k\in\mathbb{N}}\sup_{t\in[t_k,t_{k+1})}\|\int_{t_k}^t{e^{A(t-s)}ds}\cdot BK_k\|=\sup_{k\in\mathbb{N}}\sup_{\tau\in[0,\delta_{k+1})}\|\int_{t_k}^{\tau+t_k}{e^{A(\tau+t_k-s)}ds}\cdot BK_k\|\\=\sup_{k\in\mathbb{N}}\sup_{\tau\in[0,\delta_{k+1})}\|\int_{0}^{\tau}{e^{A(\tau-w)}dw}\cdot BK_k\|\leq \sup_{\tau\in[0,\bar{\delta})}\|\int_{0}^{\tau}{e^{A(\tau-w)}dw}\|\cdot \sup_{k\in\mathbb{N}}\|BK_k\|\leq C'_2$$ Now we can prove similarly that $\lim_{t\rightarrow\infty}x(t)=0$.

For nonlinear systems the problem is more difficult and the results are typically based on discrete-time approximations of the system. See for example the papers

Sufficient conditions for stabilization of sampled-data nonlinear systems via discrete-time approximations, D. Nesic, A.R. Teel, P.V. Kokotovic, Systems and Control Letters, 38, pp. 259-270, 1999

L. Grune, D. Nesić, Optimization based stabilization of sampled-data nonlinear systems via their approximate discrete-time models, SIAM Journal on Control and Optimisation, 42 (2003), pp. 98–122

for further details.

  • $\begingroup$ Fantastic explanation. I have two follow-up questions. (1) If the sampling period is variable, but bounded by $\delta$, does the same approach work? If I understand things correctly, this would complicate the description of the discrete-time system, but ultimately, the same $C_1$ and $C_2$ will still bound the corresponding expressions. (2) Is there a general approach that works for non-linear systems? I suspect that computing the corresponding discrete-time system is not possible in general for non-linear systems. $\endgroup$ Dec 30 '15 at 8:50
  • $\begingroup$ @BmoreDaniel Thank you. I will edit my answer to reply to your questions. $\endgroup$
    – RTJ
    Dec 30 '15 at 9:09

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.