1 Introduction↩︎

Recent years have seen an increase in methods for stability analysis and control of partial differential equations (PDEs) without discretization. Often, these methods rely on Lyapunov functions to establish convergence of the system state, bound \(L_2\) gain, or design observers and controllers. Unlike ODEs, for which we have a well-established universal state-space representation, PDEs take a variety of forms – with no clear and consistent notion of state space. As a result, basic definitions of stability vary from PDE to PDE and, even for a given PDE, will vary with the set of imposed boundary conditions. Semigroup theory [1] provides definitions of state which apply to a large class of PDEs. However this approach simply shifts the problem into various definitions of inner product, domain, generator, et c.. The use of Lyapunov functions further complicates the problem. While it is accepted that a Lyapunov function should be positive, bounded, and non-increasing, the form taken by these upper and lower bounds is not consistent. As a result, even for a given PDE with a given set of boundary conditions, we may obtain a variety of non-equivalent stability results depending on how the Lyapunov function is defined and bounded.

To illustrate, consider a well-studied PDE: the wave equation \(\ddot{\mathbf{ u}}=\partial_s^2\mathbf{ u}\) with pinned boundary conditions \(\mathbf{ u}(t,0)= \mathbf{ u}(t,1)=0\). This PDE is not stable in the \(L_2\times L_2\) sense in that there exists no \(C\ge 0\) such that \(\lVert{\mathbf{ u}(t)}\rVert_{L_2}+\lVert{\dot{\mathbf{ u}}(t)}\rVert_{L_2}\le C (\lVert{\mathbf{ u}(0)}\rVert_{L_2}+\lVert{\dot{\mathbf{ u}}(0)}\rVert_{L_2})\) for all \(t \ge 0\). However, this PDE is stable in the \(H^1 \times L_2\) sense, in that for \(V(u)=\lVert{\dot{u}}\rVert_{L_2}+\lVert{\partial_s u}\rVert_{L_2}\), \(\dot{V}=0\) and hence \(\lVert{\partial_s\mathbf{ u}(t)}\rVert_{L_2}+\lVert{\dot{\mathbf{ u}}(t)}\rVert_{L_2} \le \lVert{\partial_s\mathbf{ u}(0)}\rVert_{L_2}+\lVert{\dot{\mathbf{ u}}(0)}\rVert_{L_2}\) ³. This illustration shows that obvious notions of stability often do not hold for PDEs (beam equations have the same problem). Nonetheless, it seems reasonable that we would want to be able to include wave equations in any Lyapunov framework for analysis and control of PDEs. Clearly, then, we must allow for some weaker notions of stability in the analysis and control of PDEs.

The need for a weaker notion of stability (and well-posedness) is well-known. Such notions for the wave equation and can be found in [2]. Other examples where stability depends on the choice of norm can be found in [3]. For a broader class of PDE, the framework developed in [4]–[6] for non-coercive Lyapunov functions weakens the positivity condition \(V(\mathbf{ u})\ge \epsilon \lVert{\mathbf{ u}}\rVert^2\) to \(V(\mathbf{ u})>0\) for all \(\mathbf{ u} \neq 0\). A strict negativity condition on the derivative of the Lyapunov function is then used to upper bound the solutions. However, the difficulty in proving stability of PDEs such as the wave equation lies not in positivity of the Lyapunov function, but in the upper bound. For example the Lyapunov function used for the wave equation discussed above is coercive in \(\lVert{\mathbf{ u}}\rVert\). However, it is not upper bounded by \(C\lVert{\mathbf{ u}}\rVert\) for any \(C>0\).

In this context, the contribution of this paper is primarily methodical organization and unification. We classify several properties of Lyapunov functions typically used in analysis of PDEs. These properties are divided into 3 categories: positivity (lower bounded), upper bounded, and negativity of the derivative. We then interpret those properties systematically using a hierarchy of notions of stability. The difficulty, as suggested above, is that PDEs, boundary conditions, and Lyapunov functions vary substantially from case to case. To address this problem, this paper is structured around an operator or integral formulation of the PDE referred to as a Partial Integral Equation (PIE) or fundamental state representation [7]. Specifically, if the boundary conditions of a PDE are suitably well-defined, there exists a bijection (Green’s function) from \(L_2\) to the set of sufficiently regular functions which satisfy the associated set of boundary conditions.

For example, consider the heat equation \(\partial_t\mathbf{ u}=\partial_s^2\mathbf{ u}\) where solutions are constrained to lie in some domain (similar to \(D(A)\) for semigroups), e.g. \(X:=\{\mathbf{ u} \in H^2\;:\; \mathbf{ u}(0)=\mathbf{ u}(1)=0\}\). Then if we define \(\mathcal{ T}\) (Green’s function) as \[(\mathcal{ T} \mathbf{ x}):=\int_0^1G(s,\theta)\mathbf{ x}(\theta)d\theta\quad G(s,\theta)=\begin{cases}\theta (s-1)&\theta\le s\\s (\theta-1) & s \le \theta, \end{cases}\]

we have that \(\mathcal{ T} : L_2\rightarrow X\), \(\mathcal{ T}\partial_s^2 \mathbf{ u}=\mathbf{ u}\) and \(\partial_s^2 \mathcal{ T} \mathbf{ x}=\mathbf{ x}\) for any \(\mathbf{ u} \in X\) and \(\mathbf{ x} \in L_2\). We refer to \(\mathbf{ x}:=\partial_s^2\mathbf{ u}\) as the fundamental state. Using this bijection, the dynamics of a classically well-posed PDE may be equivalently ⁴ written as \(\mathcal{ T}\dot{\mathbf{ x}}=\mathcal{ A} \mathbf{ x}\). Of course, for the heat equation, this is simply⁵ \(\mathcal{ T}\dot{\mathbf{ x}}=\mathbf{ x}\) and \(\mathcal{ T}=(\partial_s^2)^{-1}\) on the domain \(X\). An equation of the form \(\mathcal{ T}\dot{\mathbf{ x}}=\mathcal{ A}\mathbf{ x}\) where \(\mathcal{ T}\) and \(\mathcal{ A}\) have the structure given above is a Partial Integral Equations (PIEs). An operator \(\mathcal{ T}\) of this type is a Partial Integral (PI) operator and the set of PI operators define a linear algebra, with the set of PI operators with polynomial kernels being a sub-algebra. It has been shown that any PDE with suitably well-defined domain, \(X\), admits such a representation with polynomial kernels [7]–[9].

Our approach, then, is to allow for classification of bounds on a Lyapunov function in terms of both the PDE state and the fundamental (PIE) state. Next, we classify notions of stability in terms of these bounds – focusing on two cases of Lyapunov stability: PDE stability (\(\lVert{\mathbf{ u}(t)}\rVert\le C \lVert{\mathbf{ u}(0)}\rVert\)) and the weaker notion of PIE to PDE stability (\(\lVert{\mathbf{ u}(t)}\rVert\le C \lVert{\mathbf{ x}(0)}\rVert\)). These notions are extended to exponential and finite-energy stability. This classification system then allows for interpretation of almost any PDE stability result which is based on the use of Lyapunov functions. Finally, to illustrate the application of these results, we express sufficient conditions for these various stability notions in terms of operator inequalities. These operator inequalities are applied to representative PDEs.

2 Notation↩︎

\(L_2[\Omega]\) denotes the space of Lebesgue square integrable functions on \(\Omega\). Unless otherwise stated, \(\Omega=[0,1]\). We use the Sobolev space \(H^n:=\{\mathbf{ u}\;:\;\partial_s^i \mathbf{ u}\in L_2,\; i \le n\}\). For \(\mathbf{ u} \in H^n\), we use the shorthand \(\mathbf{ u}_{s}:=\partial_s \mathbf{ u}\), \(\mathbf{ u}_{ss}:=\partial_s^2 \mathbf{ u}\), et c. \(\mathcal{ L}(X)\) denotes the Banach space of bounded linear operators on Banach space \(X\) with operator norm. Bolded lowercase (e.g. \(\mathbf{ u}\)) indicates a function of a spatial variable. Calligraphic uppercase (e.g. \(\mathcal{ A}\)) indicates a bounded linear operator. For matrix \(N\), the \(i,j\)th element is denoted \([N]_{i,j}:=N_{i,j}\).

3 Fundamental (PIE) State and Bijection \(\mathcal{ T}\)↩︎

\(\Pi_3\) and \(\Pi_2\) are vector spaces and, when the \(R_i\) are square, are linear composition *-subalgebras of \(\mathcal{ L}(L_2)\). The subspaces of \(\Pi_3\) and \(\Pi_2\) with \(R_i\) polynomial are likewise *-subalgebras. Analytic formulae for composition and adjoint of \(\mathcal{ P}\) in terms of the parameters \(R_i\) can be found in [8], Lemmas 4 and 6. Furthermore, for \(\mathcal{ P} \in \Pi_3\) and \(\mathcal{ Q} \in \Pi_2\), we have \(\mathcal{ PQ},\mathcal{ QP}\in \Pi_2\). Extension to multivariate domains is the sum and composition of \(\Pi_2,\Pi_3\) operators on each spatial dimension. Extension to mixed spatial domains such as \(\mathcal{ L}(\mathbb{R}^n \times L_2[0,1] \times L_2[[0,1]^2])\) can be found in [7], [9].

PDEs are typically defined in terms of \(\dot{\mathbf{ u}}=A\mathbf{ u}\), where \(A:X\rightarrow L_2\) is a differential operator and \(\mathbf{ u}\) is constrained to lie in some subspace, \(X\), of a Sobolev space. We presume the PDE is well-posed in the classical sense (See Definitions in [7]), so that for every \(\mathbf{ u}_0\in X\), there exists \(\mathbf{ u}(t)\in X\) for which \(\dot{\mathbf{ u}}=A\mathbf{ u}\) and \(\mathbf{ u}(0)=\mathbf{ u}_0\). If the PDE is well-posed and \(X\) has the form in Eqn. 1 , then \(A:X \rightarrow L_2\) is invertible on \(X\) and \(\mathcal{ T}:=A^{-1} \in \Pi_3\). This results in an equation of the form \(\mathcal{ T} \dot{\mathbf{ x}}(t)=\mathbf{ x}(t)\) where we refer to \(\mathbf{ x}(t) \in L_2\) as the fundamental state. However, solving for \(A^{-1}\) explicitly requires solution of certain differential equations. In [8], a framework was proposed whereby, roughly speaking, only the highest derivative, \(D^\alpha\), in \(A\) is inverted⁶. This results in a Partial Integral Equation (PIE) of the form \(\mathcal{ T} \dot{\mathbf{ x}}(t)=\mathcal{ A} \mathbf{ x}(t)\) where \(\mathcal{ T}=(D^{\alpha})^{-1}:L_2 \rightarrow X\) and \(\mathcal{ A}=A\mathcal{ T} \in \Pi_3\). Operator \(\mathcal{ T}\) then defines a one-to-one map between solutions of the PIE and PDE – implying many problems of well-posedness [10] and stability [8] may be equivalently formulated in this framework. Significantly, analytic formulae can be obtained for the polynomial kernels, \(R_1,R_2\) in \(\mathcal{ T}\) – implying that these operators can be readily manipulated and optimized using software such as PIETOOLS [11]. This construction is illustrated in the following lemma which concisely states the result in[12] and where we use \(D^n =\partial_s^n\), \[\begin{align} X=&\Bigl\{g \in H^n[a,b] \;:\; \qquad \text{for } i=1,\cdots,n\label{eqn:gohberg95domain2}\\[-2mm] &\qquad \sum_{j=1}^n \alpha_{i,j}(D^{j-1}g)(a)+\sum_{j=1}^n \beta_{i,j}(D^{j-1}g)(b)=0\Bigr\}\notag \end{align}\tag{1}\] and \[W_n(t)=\begin{bmatrix}1 &t&\frac{1}{2}t^2& \cdots &\frac{t^{n-1}}{(n-1)!}\\0&1&t&\cdots&\frac{t^{n-2}}{(n-2)!}\\\vdots &\vdots&\vdots&\vdots&\vdots\\0 &0&0&\cdots&t\\0 &0&0&\cdots&1\end{bmatrix}.\]

We refer to \(\mathbf{ u}\) as the PDE state and \(\mathbf{ x}=D^\alpha \mathbf{ u}\) as the fundamental or PIE state. Stability notions and Lyapunov functions may then be defined in terms of the PIE state and, occasionally, interpreted in terms of the PDE state, \(\mathbf{ u}=\mathcal{ T}\mathbf{ x}\).

4 Positivity, Boundedness, and Negativity↩︎

For each notion of stability which will be introduced in Sec. 5, we will associate a Lyapunov stability condition. Each of these Lyapunov conditions includes three parts: a lower bound on the function (positivity), an upper bound on the function, and a bound on the negativity of the derivative of the function. Before introducing notions of stability, we characterize variations on these 3 properties. Specifically, we list possible properties of a Lyapunov function, \(V\), in terms of PIE state, \(\mathbf{ x}=D^{\alpha}\mathbf{ u}\) where the PDE state is obtained from the PIE state as \(\mathbf{ u}=\mathcal{ T}\mathbf{ x}\) for some PI operator, \(\mathcal{ T}\). This then implies that \(\lVert{\mathbf{ u}}\rVert_{L_2}\le \lVert{\mathcal{ T}}\rVert\lVert{\mathbf{ x}}\rVert_{L_2}\). We start with positivity properties, from weakest to strongest.

The positivity types in Defn. 2 are hierarchical⁷ – PIE positive implies PDE positive implies positive definite implies positive semidefinite. The first holds because \(V(\mathbf{ x})\ge \epsilon \lVert{\mathbf{ x}}\rVert^2\) and \(\lVert{\mathcal{ T} \mathbf{ x}}\rVert\le \lVert{\mathcal{ T}}\rVert\lVert{\mathbf{ x}}\rVert\) imply \(V(\mathbf{ x})\ge \epsilon \lVert{\mathbf{ x}}\rVert^2\ge \frac{\epsilon}{\lVert{\mathcal{ T}^2}\rVert} \lVert{\mathcal{ T}\mathbf{ x}}\rVert^2.\) However, PDE positivity does not imply PIE positivity, since there is no \(C\) such that \(\lVert{\mathbf{ x}}\rVert\le C \lVert{\mathcal{ T} \mathbf{ x}}\rVert\).

For upper bounds, from weak to strong, we have

Any Lyapunov function of the form \(V(\mathbf{ x})=\langle \mathbf{ x}, \mathcal{ P} \mathbf{ x} \rangle\) with PI operator \(\mathcal{ P}\in \Pi_3\) is PIE bounded. Furthermore, PDE bounded implies PIE bounded since \(\lVert{\mathcal{ T} \mathbf{ x}}\rVert \le \lVert{\mathcal{ T}}\rVert\lVert{\mathbf{ x}}\rVert\).

For negativity, we presume \(V(\mathbf{ x})\) is sufficiently regular so that, for a given \(\mathcal{ T} \dot{\mathbf{ x}}=\mathcal{ A} \mathbf{ x}\), we may define \(\dot{V}(\mathbf{ x})\).

Apart from “Lyapunov negative”, negativity conditions are ordered from weakest to strongest – i.e. PIE negative implies PDE negative implies negative semidefinite. The first holds since \(-\lVert{\mathbf{ x}}\rVert^2\le -\frac{1}{\lVert{\mathcal{ T}}\rVert^2}\lVert{\mathcal{ T}\mathbf{ x}}\rVert^2\). The strength of “Lyapunov negativity” relative to PIE and PDE negativity depends on positivity and boundedness of \(V\). Specifically, Lyapunov negativity implies PDE or PIE negativity if \(V\) is PDE or PIE positive, respectively. Likewise, PDE or PIE negativity implies Lyapunov negativity if \(V\) is PDE or PIE, respectively.

5 Stability Notions and Lyapunov Functions↩︎

In this section, we define 4 notions of stability and, for each notion of stability, associate a sufficient condition in terms of existence of a Lyapunov function which satisfies a set of bounds as defined in Sec. 4. For consistency, these notions of stability and Lyapunov conditions are all expressed in terms of a PIE \(\mathcal{ T} \dot{\mathbf{ x}}=\mathcal{ A} \mathbf{ x}\) with fundamental state, \(\mathbf{ x} \in L_2\). In all cases, we presume that the PIE is obtained from a classically well-posed PDE using the definitions and formulae in [7] (or Lemma 1). This ensures that any solution of the PIE, \(\mathbf{ x}(t)\) defines a solution of the PDE as \(\mathbf{ u}(t)=\mathcal{ T} \mathbf{ x}(t)\) and hence when a stability condition involves \(\mathcal{ T} \mathbf{ x}\), this is interpreted as the state of the underlying PDE from whence the PIE was obtained.

5.1 Notions of Lyapunov Stability↩︎

Let us begin with the notions of Lyapunov stability. These notions are particularly important for energy-conserving systems such as wave and beam equations.

Note that PDE to PIE stable never holds when the PIE state is a partial derivative of the PDE state (\(\mathbf{ x}=D^{\alpha}\mathbf{ u}\)) since at time \(t=0\), this requires \(\lVert{\mathbf{ x}(0)}\rVert\le C \lVert{\mathcal{ T}\mathbf{ x}(0)}\rVert\). However, such notions may hold for integro-differential equations [14]. In the following, we examine the relationships between the several proposed notions of Lyapunov stability.

Lemma 2 shows that PIE to PDE stable is the weakest notion, since it is implied by any of the others. However, PDE stable does not imply PIE stable and PIE stable does not imply PDE stable.

5.1.0.1 Conditions for Lyapunov Stability

Having defined several notions of Lyapunov stability, we now provide conditions for such stability properties in terms of properties of a candidate Lyapunov function.

5.2 Notions of Exponential Stability↩︎

For exponential stability, from weak to strong, we have

As in Lemma 2, Exp. PIE or PDE stable implies Exp. PIE to PDE stable and Exp. PDE to PIE stable implies all others.

Lyapunov Conditions for Exponential Stability To keep the conditions tight and concise, in Lemma 4 we use Lyapunov negativity instead of PIE or PDE negativity. This allows us to directly use \(V(\mathbf{ x}(t))\le -\alpha V(\mathbf{ x}(t))\) to conclude \(V(\mathbf{ x}(t))\le e^{-\alpha t}V(\mathbf{ x}(0))\). We will return to PIE and PDE negativity in Subsec. 5.3 and Sec. 6, while for now recalling only that the relation to Lyapunov negativity depends on PIE and PDE positivity and boundedness.

5.3 More Exotic Flavours: Finite Energy Stability↩︎

One might interpret “non-coercive” Lyapunov functions as \(V\) positive definite, but not PDE positive. However, we don’t have a consistent mechanism to enforce \(V\) positive definite and so we will relax this to \(V\) positive semidefinite. In order to ensure a non-zero Lyapunov function, we then enforce a strict negativity condition on the derivative. A consequence of this formulation is that the Lyapunov function does not enforce a pointwise-in-time decay rate in the same way that exponential stability does, but rather only imposes a bound on the energy of the solution as measured by its temporal \(L_2\) norm – i.e. \(\lVert{\mathbf{ x}}\rVert_{L_2} = \sqrt{\int_0^\infty \lVert{\mathbf{ x}(t)}\rVert^2 dt}\). We refer to this notion of energy as "Finite-Energy" Stability⁸.

Lyapunov Conditions for FE stability Unlike exponential stability, negativity of the derivative is critical in FE stability. Hence, we do not use Lyapunov negativity in Lemma 5.

6 Sufficient Conditions for Stability↩︎

In this section, we briefly provide a list of conditions for our stability notions in terms of feasibility of operator inequalities defined in terms of \(\mathcal{ T}\) and \(\mathcal{ A}\) and variable \(\mathcal{ P} \in \Pi_3\) or \(\mathcal{ Q} \in \Pi_3\). Here the variable label \(\mathcal{ P}\) will be used when the variable is constrained to be self-adjoint and \(\mathcal{ Q}\) otherwise. The operator inequality constraints are all interpreted as positivity on \(L_2\) and may be verified by semidefinite programming algorithms in⁹ software such as PIETOOLS [11].

Note that PIE bounded in satisfied for any \(\mathcal{ P} \in \Pi_3\).

Every sufficient condition proposed in this section relies on one of the following two canonical candidate Lyapunov functions. Specifically, given \(\mathcal{ T}\), we have \[V_1(\mathbf{ x}):=\langle \mathcal{ T}\mathbf{ x}, \mathcal{ P} \mathcal{ T} \mathbf{ x} \rangle\quad \text{and}\quad V_2(\mathbf{ x}):=\langle \mathbf{ x}, \mathcal{ Q} \mathcal{ T} \mathbf{ x} \rangle\] where \(V_1\) and \(V_2\) are parameterized by PI operators \(\mathcal{ P}\in \Pi_3\) and \(\mathcal{ Q}\in \Pi_3\), respectively. Here we constrain \(\mathcal{ P}=\mathcal{ P}^*\) and \(\mathcal{ Q} \mathcal{ T}=\mathcal{ T}^*\mathcal{ Q}^*\). See Notes 4 and 5 in [13] for further justification of \(V_1\) and \(V_2\), respectively.

6.1 Operator Inequalities for Lyapunov Stability↩︎

In the following, we combine \(V_1\) and \(V_2\) with Lemmas 3 and 6 to obtain operator inequalities for each category of Lyapunov stability – See Note 3 in [13] for a proof.

6.2 Operator Inequalities for Exponential Stability↩︎

We combine \(V_1\) and \(V_2\) with Lemmas 4 and 6 to obtain the following operator inequalities for each category of exponential stability¹⁰.

6.3 Operator Inequalities for Finite Energy Stability↩︎

Finally, we combine \(V_1\) and \(V_2\) with Lemmas 5 and 6 to obtain the following operator inequalities for each category of finite energy stability.

7 Summary↩︎

Table [tab:summary] summarizes the weakest Lyapunov conditions required for each notion of stability.

8 PDEs and Stability Analysis in PIETOOLS↩︎

We consider 4 illustrative PDEs and test each notion of stability indicated in Table [tab:summary] using PIETOOLS to enforce the associated LPI constraints as indicated in Lemmas 7, 8, and 9. In all cases, the PDE, BCs, and definition of PIE state, \(\mathbf{ x}\), are given. LPI parameters were chosen as \(\alpha=\epsilon=.001\). Table [tab:results] lists each notion of stability PIETOOLS was able to verify for each example¹¹.

Example 1: Hyperbolic System of Balance Laws [15] \[\dot{\mathbf{ u}}=-M\mathbf{ u}-L\mathbf{ u}_s\qquad K_1\mathbf{ u}(t,0)+K_2 \mathbf{ u}(t,1)=0 \qquad \mathbf{ x}:=\mathbf{ u}_s \] \[\begin{align} L&=\begin{bmatrix} -3.76& 0& 0\\ 0 & 0.5& 0\\ 0& 0& 5.25\end{bmatrix}\quad M=\begin{bmatrix} 1.03& .05 & .92\\ 1.03 & .05 & .92\\ 1.03 & .05& .92\end{bmatrix}\\ K_1&=\begin{bmatrix} 1 & -1.95& 34.5\\ 0& 0& 0\\ 0& 0& 0\end{bmatrix},\quad K_2=\begin{bmatrix} 0& 0& 0\\ -0.55& 1& 0\\ -1.75& 0 & 1\end{bmatrix}\\[-6mm] \end{align}\] Example 2: Reaction-Diffusion w/ spatial coefficients [16]. \[\begin{align} \dot{\mathbf{ u}}(t,s)&=a(s)\mathbf{ u}(t,s)+b(s)\mathbf{ u}_s(t,s)+c(s)\mathbf{ u}(s)\\ \mathbf{ u}(t,0)&=\mathbf{ u}_s(t,1)=0 \qquad \mathbf{ x}:=\mathbf{ u}_{ss}\qquad a(s)=s^3-s^2+2\\ b(s)&=3s^2 - 2s,\quad c(s)=-s^3/2 + 1.3s^2- 1.5s + 5.35\\[-7mm] \end{align}\]

Example 3: Euler-Bernoulli Beam Equation. Given \(\ddot{\mathbf{ u}}=-\mathbf{ u}_{ssss}\), use the first-order form defined by \(\mathbf{ u}_1=\dot{\mathbf{ u}}\) and \(\mathbf{ u}_2=\mathbf{ u}\), and we have \(\dot{\mathbf{ u}}_1=-\mathbf{ u}_{2,ssss}\), \(\dot{\mathbf{ u}}_2=\mathbf{ u}_1\) where \(\mathbf{ u}_2(t,0)=\mathbf{ u}_{2,ss}(t,1)=\mathbf{ u}_{2,s}(t,0)=\mathbf{ u}_{2,sss}(t,1)=0\). Fundamental state is defined as \(\mathbf{ x}:=\begin{bmatrix}\dot{\mathbf{ u}}& \mathbf{ u}_{ssss}\end{bmatrix}^T\).

Example 4: Tip-Damped Wave Equation [17]. Given \(\ddot{\mathbf{ u}}=\mathbf{ u}_{ss}-2 \dot{\mathbf{ u}}-\mathbf{ u}\), use the first-order form defined by \(\mathbf{ u}_1=\mathbf{ u}_t\) and \(\mathbf{ u}_2=\mathbf{ u}\), and we have \[\dot{\mathbf{ u}}=\begin{bmatrix}-2 & -1\\1&0\end{bmatrix}\mathbf{ u}+\begin{bmatrix}1&0\\0&0\end{bmatrix}\mathbf{ u}_{ss} \qquad \mathbf{ x}:=\begin{bmatrix}\dot{\mathbf{ u}}_s\\ \mathbf{ u}_{ss}\end{bmatrix}\] where \(\mathbf{ u}_1(t,0)=\mathbf{ u}_2(t,0)=0\) and \(\mathbf{ u}_s(t,1)=-\dot{\mathbf{ u}}(t,1)\).

9 Conclusion↩︎

We have proposed a hierarchical classification of Lyapunov, exponential, and finite-energy stability notions in terms of the fundamental state, \(\mathbf{ x}\), which is related to the PDE state as \(\mathbf{ x}=D^\alpha \mathbf{ u}\) and \(\mathbf{ u}=\mathcal{ T} \mathbf{ x}\) and which satisfies \(\mathcal{ T} \dot{\mathbf{ x}}=\mathcal{ A} \mathbf{ x}\) for some PI operators, \(\mathcal{ T},\mathcal{ A}\). This framework unifies Lyapunov stability conditions which may or may not include partial derivatives of the PDE state in lower bounds, upper bounds, and negativity conditions. Lyapunov conditions which establish stability, but not PDE stability often imply PIE to PDE stability. This framework allows for classification of PDEs such as the wave and beam equations. Furthermore, operator inequality conditions are provided for each notion of stability and may be verified using software such as PIETOOLS.

9.1 Notes and Commentary: Section 4↩︎

9.2 Notes and Commentary: Section 5↩︎

9.3 Notes and Commentary: Section 6↩︎

Note that the alternative condition for Exp. PIE to PDE stable may be considerably stronger than the condition for Exp. PIE to PDE stability in Lem. 8.

References↩︎

---

1. M. Peet is with the School for the Engineering of Matter, Transport and Energy, Arizona State University. e-mail: mpeet@asu.edu↩︎

2. This work was supported by the National Science Foundation under grants No. 2429973 and 2337751.↩︎

3. However, this tells us nothing about the actual PDE state, \(\mathbf{ u}(t)\). Fortunately, we know that \(\lVert{\mathbf{ u}}\rVert\le C\lVert{\partial_s\mathbf{ u}}\rVert\) where \(C\) is the Poincaré constant and hence \(\lVert{\mathbf{ u}(t)}\rVert_{L_2}+\lVert{\dot{\mathbf{ u}}(t)}\rVert_{L_2} \le C\lVert{\partial_s\mathbf{ u}(0)}\rVert_{L_2}+\lVert{\dot{\mathbf{ u}}(0)}\rVert_{L_2}\). However, as indicated, this bound cannot be strengthened to the first, more tradition notion of stability or well-posedness in \(\mathbf{ u}\).↩︎

4. For notational simplicity, we assume sufficient regularity for commutation of spatial and temporal derivatives. Otherwise we use \(\partial_t (\mathcal{ T}\mathbf{ x})=\mathcal{ A} \mathbf{ x}\).↩︎

5. More generally, we may always write the dynamics of a PDE \(\dot{u}=Au\) as \(\mathcal{ T} \dot{\mathbf{ x}}=\mathbf{ x}\) where \(\mathcal{ T}=A^{-1}\) on \(X\). However, in this case, the kernels of operator \(\mathcal{ T}\) may not be polynomial and require analytic solution of certain differential equations. If such an operator is known, however, the results of this paper can be applied using this \(\mathcal{ T}\) and \(\mathcal{ A}=I\).↩︎

6. If \(A\) is not a differential operator, boundary conditions still require differentiation in order to obtain a bijection from \(L_2\) to \(X\).↩︎

7. For an explanation why “Positive definite” was not included in this list, see Note 1 in [13]↩︎

8. See Note 2 in [13] to relate asymptotic and finite-energy stability.↩︎

9. Alternatively, these may be tested via discretization – i.e. late lumping.↩︎

10. If we wish to avoid bilinearity in decision variables \(\alpha\) and \(\mathcal{ Q}\) or \(\mathcal{ P}\), we may replace the Lyapunov negativity conditions in Lem. 4 with PIE or PDE negativity – See Note 8.↩︎

11. Note that a lack of verification does not necessarily imply the system is not stable in the given sense.↩︎