betaFeedback

WorksMathematics

Learning Morse Lyapunov functions through invertible neural networks

Abstract

Towards data-driven stability certificates of dynamical systems, we are witnessing renewed interest in studying and exploiting the structure of Lyapunov functions. Especially Lyapunov functions that are locally quadratic (i.e., Morse) gained popularity ever since it was understood that they can be written as the norm squared of a diffeomorphism. This topological representation alludes to parametrizations using invertible neural networks. In this note we discuss when and how such functions can be learned. In particular, using neural ODEs (NODEs), we characterize Sobolev approximation of these diffeomorphisms on compact sets and show how this naturally relates to stability. We also connect our results to PLNets and supply numerical experiments for illustration.

Introduction

After pioneering work by Massera [23], Kurzweil [18] and Wilson [33], there is now an extensive theory describing when the existence of a continuous or smooth Lyapunov function VV is equivalent to the existence of an asymptotically stable attractor under some dynamical system, e.g., see [10, 5, 3, 8]. Towards a data-driven stability analysis, however, one hopes to say more beyond regularity of those Lyapunov functions, e.g., to work with suitable parametric function spaces when searching for a Lyapunov function, see Manchester et al. [22] for a recent overview and Kolter and Manek [17] for work that propelled this field of study. In this article we focus solely on, structure of the Lyapunov function, but note that a lot of work aims to study or identify some unknown dynamical system together with such a stability certificate, e.g., see Harting et al. [11].

In a companion note (Jongeneel [14]), we discuss how the structure of Lyapunov functions helps in uncovering topological features of dynamical systems. We exploit those results below.

Setting

In this note we work with CrC^r-smooth dynamical systems of the form

x˙=f(x),x∈Rn,(1)\dot{x}=f(x), \quad x\in\mathbb{R}^{n}, \tag*{(1)}

for some r∈N≥1∪{∞}r\in\mathbb{N}_{\geq1}\cup\{\infty\}, and such that some equilibrium point x∗∈Rnx^{*}\in\mathbb{R}^{n} is globally asymptotically stable (GAS). To aid the reader, we briefly recall global asymptotic stability, (e.g., see Bacciotti and Rosier [3], p. 28). Consider (1), and let the corresponding semiflow be denoted by ϕ\phi, i.e., ddtϕt(x)∣τ=t=f(ϕt(x))\left.\frac{\mathrm{d}}{\mathrm{d}t}\phi^{t}(x)\right|_{\tau=t}=f(\phi^{t}(x)) for all (t,x)∈R≥0×Rn(t,x)\in\mathbb{R}_{\geq0}\times\mathbb{R}^{n}. We will not be concerned with existence of solutions. A point x∗∈Rnx^{*}\in\mathbb{R}^{n} is an equilibrium point of ff when f(x∗)=0f(x^{*})=0 or equivalently, x∗x^{*} is a fixed point of the time-tt map: ϕt(x∗)=x∗\phi^{t}(x^{*})=x^{*} for any t≥0t\geq0. Then, the equilibrium point x∗x^{*} is said to be globally asymptotically stable (GAS) (under ff) when: (i) x∗x^* is Lyapunov stable, that is, for any open neighbourhood Uε∋x∗U_\varepsilon\ni x^* there is an open set Uδ⊆UεU_\delta\subseteq U_\varepsilon such that ϕt(Uδ)⊆Uε\phi^t(U_\delta) \subseteq U_\varepsilon ∀t≥0\forall t \ge0; and (ii) x∗x^* is globally attractive, that is, lim⁡t→+∞ϕt(x0)=x∗\lim_{t \to+\infty} \phi^t(x_0) = x^* for all x0∈Rnx_0 \in\mathbb{R}^n.

It is hard to work directly with the definition of GAS. Luckily, x∗x^* being GAS is equivalent to the existence of a smooth (i.e., C∞C^\infty-smooth) function V:Rn→R≥0V : \mathbb{R}^n \to\mathbb{R}_{\ge0} that is: (i) coercive (i.e., all sublevel sets {x∈Rn:V(x)≤c}\{x \in\mathbb{R}^n : V(x) \le c\} are compact); (ii) such that {x∗}=V−1(0)\{x^*\} = V^{-1}(0); and (iii) away from x∗x^*, decreasing strictly under solutions of (1), i.e., LfV(x):=⟨∇V(x),f(x)⟩<0L_fV(x) := \langle\nabla V(x), f(x) \rangle< 0 for all x∈Rn∖{x∗}x \in\mathbb{R}^n \setminus\{x^*\}, e.g., see [3 Thm 2.4]. Functions VV satisfying properties (i), (ii) and (iii) will be simply called smooth Lyapunov functions (this instead of smooth, strict, proper Lyapunov functions with respect to equilibria). Without loss of generality, in the remainder of this note we focus on x∗=0x^* = 0. We also remark that we work on Rn\mathbb{R}^n for simplicity.

Regarding further notation and structure, following Kvalheim [20], let S0r(Rn)S^r_0(\mathbb{R}^n) be the set of CrC^r-smooth vector fields on Rn\mathbb{R}^n, for some r∈N≥1∪{∞}r \in\mathbb{N}_{\ge1} \cup\{\infty\}, such that x∗=0x^* = 0 is GAS.

Suppose now that we have a family of Lyapunov function candidates, parametrized through θ∈Θ⊆Rd\theta\in\Theta\subseteq\mathbb{R}^d via some neural network: {Vθ∈C∞(Rn;R≥0):θ∈Θ}\{V_\theta\in C^\infty(\mathbb{R}^n;\mathbb{R}_{\ge0}) : \theta\in\Theta\}. With the aforementioned terminology in place, the central question of this note is to understand precisely when and how these functions can be selected to be Morse. Better yet, we discuss why this is interesting.

We recall that a function v∈C2(Rn;R)v \in C^2(\mathbb{R}^n;\mathbb{R}) is a Morse function when all its critical points (i.e., x∗∈Rnx^* \in\mathbb{R}^n such that ∇v(x∗)=0\nabla v(x^*) = 0) are non-degenerate, that is, the Hessian ∇2v(x∗)\nabla^2v(x^*) is non-singular. Inspired by a rich set of topological results from Morse theory (e.g., see [24]), one may motivate the study, globally, of Lyapunov functions that are Morse, i.e., ⟨q,∇2V(x∗)q⟩>0\langle q,\nabla^2V(x^*)q\rangle> 0 for all q∈Rn∖{0}q \in\mathbb{R}^n \setminus\{0\} and some equilibrium x∗x^*. The fact that the Hessian is symmetric positive definite is written as ∇2V(x∗)≻0\nabla^2V(x^*) \succ0. Such a Lyapunov function will be called a Morse Lyapunov function.

Contribution: In our companion note, inspired by Boumal et al. [6], we studied topological ramifications of working with Lyapunov functions that are Morse [14]. In this work, being inspired by advances in invertible neural networks (e.g., Ishikawa et al. [13]) we detail effectiveness of employing invertible neural networks to learn Morse Lyapunov functions as stability certificates. In particular, we provide (i) an obstruction for the existence of Morse Lyapunov functions (Proposition 1); (ii) as the obstruction is mild, we can provide theory for when NODE-based INNs yield Morse Lyapunov functions on compact sets (Theorem 5, Corollary 7); and (iii) we link these results to PŁ Lyapunov functions, providing context for the distortion parameter τ\tau (Remark 8).

Simple Lyapunov functions do not suffice, in general

First, when (1) is linear, VV can be chosen to be quadratic (e.g., x↦⟨x,Px⟩x \mapsto\langle x,Px\rangle for P≻0P \succ0), globally, e.g., see [30 Thm. 18]. Under hyperbolicity, some arguments survive, locally, for nonlinear systems, e.g., see [21 §9] and [10 §28] for early results or [4] and [11] for recent learning-oriented results.

Globally, one may hope that more can be said beyond CrC^r-smoothness of VV, e.g., one may hope that any Lyapunov function can be chosen to be a polynomial. This hope was refuted by Ahmadi et al. [2]. In a similar vein, there are vector fields that render the origin GAS, but provably fail to admit a convex Lyapunov function [15 Ex. III.3]. Moreover, one can find examples where a smooth Lyapunov function exists, but not one that is locally real-analytic, see [3 Prop. 5.2 (b] and [1].

One may continue and hope that Lyapunov functions can always be chosen such that they are Morse, for instance. With the proposition below we illustrate that in general, not much more can be said beyond simple CrC^r-regularity. The example to have in mind is a planar ODE of the form:

x˙1=−εx13+x2x˙2=−x13−x23,(2)\begin{aligned} \dot{x}_1 &= -\varepsilon x_1^3+x_2\\ \dot{x}_2 &= -x_1^3-x_2^3, \tag*{(2)} \end{aligned}

for some ε>0\varepsilon>0. One can use the Lyapunov function R2∋(x1,x2)↦V(x1,x2):=x14/4+x22/2\mathbb{R}^2\ni(x_1,x_2)\mapsto V(x_1,x_2):=x_1^4/4+x_2^2/2 to check that x∗=0x^*=0 is indeed GAS. However, one can show that VV is not Morse, and in fact, no Morse Lyapunov function exists for (2) ([14 Prop. 3.5]. It seems reasonable that hyperbolicity plays a role here, however, the real problem is having a linearization where the imaginary part cannot be diagonalized ([14 Cor. 3.9]). We generalize this observation below, that is, not only does a locally quadratic Lyapunov function fail for (2), even a Lyapunov function that is sandwiched between positively homogeneous polynomials, in ∥x∥\lVert x\rVert, of equal degree (e.g., quadratic forms), fails to exist cf. the quadratic bounds in Kolter and Manek [17].

Proposition 1 (An obstruction to local homogeneity) Let f∈S0r(Rn)f\in S_0^r(\mathbb{R}^n), for some r∈N≥1∪{∞}r\in\mathbb{N}_{\geq1}\cup\{\infty\} and n∈N≥1n\in\mathbb{N}_{\geq1}. Suppose that Df(0)\mathrm{D}f(0) has eigenvalues on the imaginary axis that are not semisimple. Suppose that we pick some C0C^0 function V:Rn→R≥0V:\mathbb{R}^n\to\mathbb{R}_{\geq0} such that

a1∥x∥p≤V(x)≤a2∥x∥p∀x∈U,(3)a_1\lVert x\rVert^p\leq V(x)\leq a_2\lVert x\rVert^p\quad\forall x\in U, \tag*{(3)}

for some 0<a1≤a20<a_1\leq a_2, p∈N≥1p\in\mathbb{N}_{\geq1} and UU being some neighbourhood of 00. Then, VV cannot be a Lyapunov function for x∗=0x^*=0 under ff.

Proof. Since Df(0)\mathrm{D}f(0) has an eigenvalue that is not semisimple, there is some x0∈Ux_0\in U such that lim⁡t→+∞∥etDf(0)x0∥=+∞\lim_{t\to+\infty}\lVert e^{t\mathrm{D}f(0)}x_0\rVert=+\infty. This implies in particular that there is some T>0T>0 such that a1∥eTDf(0)x0∥p>a2∥x0∥pa_1\lVert e^{T\mathrm{D}f(0)}x_0\rVert^p>a_2\lVert x_0\rVert^p. However, then for a sufficiently small δ>0\delta>0 (i.e., to stay within UU) one has

V(ϕT(δx0))≥a1∥ϕT(δx0)∥p=a1∥δeTDf(0)x0+r(δ)∥p,V(\phi^T(\delta x_0))\geq a_1\lVert\phi^T(\delta x_0)\rVert^p=a_1\lVert\delta e^{T\mathrm{D}f(0)}x_0+r(\delta)\rVert^p,

where δ↦r(δ)\delta\mapsto r(\delta) is a remainder term such that lim⁡δ→0+r(δ)/δ=0\lim_{\delta\to0^+}r(\delta)/\delta=0. Continuing, after a further decrease in δ>0\delta>0, if necessary, one obtains:

a1δp∥eTDf(0)x0+r(δ)/δ∥p>a2δp∥x0∥p≥V(δx0).a_1\delta^p\lVert e^{T\mathrm{D}f(0)}x_0+r(\delta)/\delta\rVert^p>a_2\delta^p\lVert x_0\rVert^p\geq V(\delta x_0).

Hence, VV is increasing along a solution and so it cannot be a valid Lyapunov function.

Note, one may continue this line of arguments and generalize (3) to a subset of K∞K_\infty functions. We leave this, and in fact a complete characterization, for future work.

A typical non-hyperbolic example that does admit a Morse Lyapunov function is of the form

x˙1=x2−x1(x12+x22),x˙2=−x1−x2(x12+x22),(4)\begin{aligned} \dot{x}_1&=x_2-x_1(x_1^2+x_2^2),\\ \dot{x}_2&=-x_1-x_2(x_1^2+x_2^2), \tag*{(4)} \end{aligned}

e.g., see Sastry [28] and Jongeneel [14]. One can check that (x1,x2)↦V(x1,x2):=x12+x22(x_1,x_2)\mapsto V(x_1,x_2):=x_1^2+x_2^2 is a Morse Lyapunov function indeed. In particular, V(x)=−2∥x∥22V(x)=-2\lVert x\rVert_2^2.

Morse Lyapunov functions

Inspired by genericity results (e.g., see Peixoto [26], Hirsch, [12] or Palis and De Melo, [25]), one might expect that Morse Lyapunov functions still provide for a sufficiently rich class of functions. To that end, let HS0r(Rn)⊂S0r(Rn)\mathcal{H}S_{0}^{r}(\mathbb{R}^{n}) \subset S_{0}^{r}(\mathbb{R}^{n}) be the subset of vector fields such that 0∈Rn0 \in\mathbb{R}^{n} is a hyperbolic equilibrium point (i.e., ℜ{λi(Df(0))}≠0\Re\{\lambda_i(Df(0))\} \ne0 for i∈[n]i \in[n]). In our companion paper we have provided the following converse result, confirming part of what one might expect.

Theorem 2 (On the existence of Morse Lyapunov functions (Jongeneel, [14]). Let f∈HS0r(Rn)f \in\mathcal{H}S_{0}^{r}(\mathbb{R}^{n}), for some r∈N≥1∪{∞}r \in\mathbb{N}_{\ge1} \cup\{\infty\} and n∈N≥1n \in\mathbb{N}_{\ge1}, then, there is a smooth Morse Lyapunov function certifying that 00 is GAS.

A generalization of the Morse Lemma (Milnor, [24]) makes Theorem 2 interesting, that is, through Boumal et al. [6] it follows that Morse Lyapunov functions are elements of

V≔{V∈C∞(Rn;R≥0)∣∃ φ∈Diff⁡∞(Rn;Rn):V(φ−1(y))=inf⁡y′∈RnV(y′)+∥y∥222},\mathcal{V} \coloneqq\left\{V \in C^{\infty}(\mathbb{R}^{n};\mathbb{R}_{\ge0}) \mid\exists\ \varphi\in\operatorname{Diff}^{\infty}(\mathbb{R}^{n};\mathbb{R}^{n}) : V(\varphi^{-1}(y)) = \inf_{y' \in\mathbb{R}^{n}} V(y') + \frac{\lVert y\rVert_{2}^{2}}{2}\right\},

where Diff⁡r(Rn;Rn)\operatorname{Diff}^{r}(\mathbb{R}^{n};\mathbb{R}^{n}) denotes the group of CrC^{r}-smooth diffeomorphisms from Rn\mathbb{R}^{n} to itself (CrC^{r}-smooth bijective maps with a CrC^{r}-smooth inverse). In fact, one may pick φ∈Diff⁡∞,+(Rn;Rn)\varphi\in\operatorname{Diff}^{\infty,+}(\mathbb{R}^{n};\mathbb{R}^{n}) thanks to symmetry in y↦∥y∥22y \mapsto\lVert y\rVert_{2}^{2} (e.g., see Jongeneel, [14]), that is, the diffeomorphism can be selected to be orientation preserving (like time-tt maps x↦ϕt(x)x \mapsto\phi^{t}(x), for ϕ\phi a flow). One may spot that results of this form are closely related to the study of linearizing nonlinear dynamical systems. In general, a diffeomorphism does not suffice, one needs to sacrifice regularity at the equilibrium, see Grüne et al. [9]; Jongeneel [14]; Kvalheim and Sontag [20].

Remark 3 (Regions of attraction that are not Rn\mathbb{R}^{n}). We have focussed on global asymptotic stability on Rn\mathbb{R}^{n} to keep the exposition simple. However, consider some smooth nn-dimensional manifold MM and let p∗∈Mp^{*} \in M be asymptotically stable under some smooth vector field ff on MM, with B∗B^{*} being the corresponding region (basin) of attraction. A topological fact is that B∗B^{*} is diffeomorphic to Rn\mathbb{R}^{n} (Wilson Jr, 1967, Thm. 2.2). Hence, effectively anything we discuss carries over.

Then, inspired by Wilson Jr (1967), Cheng et al. [7] proposed to use a function gθg_{\theta}, parametrized via BiLipNet (or PLNet for that matter) e.g., see Wang et al. [32], and study Lyapunov functions of the form x↦Vθ(x)=∥gθ(x)∥22x \mapsto V_{\theta}(x) = \lVert g_{\theta}(x)\rVert_{2}^{2}, which we understand as a predecessor of using V\mathcal{V} (as global smooth PL (Lyapunov) functions, i.e., ∥∇Vθ(x)∥22≥μVθ(x)\lVert\nabla V_{\theta}(x)\rVert_{2}^{2} \ge\mu V_{\theta}(x), for some μ>0\mu> 0 and all x∈Rnx \in\mathbb{R}^{n}, are a subset of V\mathcal{V} (Boumal et al., [6]). They identified that such a parametrization directly aids in enforcing structure relevant for Lyapunov functions. Since gθg_{\theta} is bi-Lipschitz, the map is invertible, and in contrast to other work, VθV_{\theta} is not enforced to be convex. Convexity may be restrictive (Jongeneel and Schwan, [15]). Then, motivated by finding neural architectures that guarantee ∇Vθ(x)≠0\nabla V_{\theta}(x) \ne0 for all x≠x∗x \ne x^{*}, Zhong et al. [35] proposed PolarNet in a recent preprint. They effectively recover a subset of the results from Boumal et al. [6], cf. (Zhong et al., [35]). They work with x↦Vθ(x)=∥ψθ(x)∥22x \mapsto V_{\theta}(x) = \lVert\psi_{\theta}(x)\rVert_{2}^{2} for an appropriate map ψθ:Rn→Rn\psi_{\theta} : \mathbb{R}^{n} \to\mathbb{R}^{n} and claim that if ψθ\psi_{\theta} is an affine coupling flow architecture, then any C2C^{2} diffeomorphism can be approximated, up to arbitrary precision. The authors appeal to Teshima et al. [31] to substantiate this claim. However, that work provides a universal approximation property with respect to the LpL_{p} or sup-norm; and these norms are not of use in the context of Lyapunov inequalities. Then, somewhat paraphrasing, Zhong et al. [35] write that asymptotic stability implies that a Lyapunov function exists that can be accurately modelled using PolarNet. Below we formalize and sharpen this claim.

Topologies to certify stability

Merely universal approximation in some norm does not readily suffice in the context of Lyapunov functions. Suppose that VV is an unknown Lyapunov function with respect to some dynamical system (1), then any V~\widetilde{V} that is close, in the sup-norm, to VV is by no means a valid Lyapunov function. To get a grip on regularity, define the CkC^k-norm of a CkC^k-smooth map h:Rd→Rnh : \mathbb{R}^d \to\mathbb{R}^n, on some (bounded) domain Ω⊆Rd\Omega\subseteq\mathbb{R}^d, for simplicity, through

∥h∥Ck(Ω):=sup⁡i∈N[0,k]sup⁡x∈Ω∥Dih(x)∥.\|h\|_{C^k(\Omega)} := \sup_{i\in\mathbb{N}_{[0,k]}} \sup_{x\in\Omega} \|D^i h(x)\|.

The standard example to have in mind is as follows, given a function h∈C1(R;R)h \in C^1(\mathbb{R};\mathbb{R}), then a perturbation of the form R∋x↦e(x):=ϵsin⁡(x/ϵ2)\mathbb{R} \ni x \mapsto e(x) := \epsilon\sin(x/\epsilon^2) results in ∥h+e∥C0([0,1])≤∥h∥C0([0,1])+ϵ\|h+e\|_{C^0([0,1])} \leq\|h\|_{C^0([0,1])}+\epsilon. However, ∥h+e∥C1([0,1])≥h′(0)+1/ϵ\|h+e\|_{C^1([0,1])} \geq h'(0)+1/\epsilon. In the context of Lyapunov functions, merely C0C^0 control is futile, e.g., ∥V−V~∥C0(K)<ϵ\|V-\widetilde{V}\|_{C^0(K)} < \epsilon, for a small ϵ>0\epsilon> 0, does not imply that LfV~<0L_f\widetilde{V} < 0, despite LfV<0L_fV < 0.

Without going into measure theory, the LpL^p-norm of such a function hh is defined through ∥h∥Lp(Ω):=(∫Ω∥h∥p dμ)1/p\|h\|_{L_p(\Omega)} := \left(\int_\Omega\|h\|^p \,d\mu\right)^{1/p} for p∈N≥1p \in\mathbb{N}_{\geq1}. The sup-norm, on the other hand, is simply defined as ∥h∥Ω,∞:=sup⁡x∈Ω∥h(x)∥\|h\|_{\Omega,\infty} := \sup_{x\in\Omega}\|h(x)\|. It is now important to note that both the LpL_p-norm and the sup-norm do not control the CkC^k-norm for k>0k > 0. Hence, they are of little use in the context of learning Lyapunov functions. This is known, Harting et al. [11] consider C1C^1-control as well, although not directly by construction, but by appealing to standard universal approximation results.

Universal approximation properties of invertible neural networks

Zhong et al. [35] appeal to Teshima et al. [31] and claim that LpL_p or sup-norm control suffices to learn Lyapunov functions. We will appeal to more recent results by Ishikawa et al. [13] regarding invertible neural networks (INNs) and control of a Sobolev norm. To follow Ishikawa et al. [13], we need to work with n≥2n \geq2. We return to this in Remark 6.

Let G⊆Lip⁡r(Rn;Rn)\mathcal{G} \subseteq\operatorname{Lip}^r(\mathbb{R}^n;\mathbb{R}^n) (i.e., some CrC^r-smooth globally Lipschitz maps) and let ϕg\phi_g be the flow corresponding to y˙=g(y)\dot{y}=g(y) for any g∈Gg \in\mathcal{G}. Then define the set of diffeomorphisms Φ(G):={x↦ϕg1(x):g∈G}\Phi(\mathcal{G}) := \{x \mapsto\phi_g^1(x) : g \in\mathcal{G}\}. Now we may define a set of invertible neural networks (INNs) through

INN⁡Φ(G):={W1∘ϕ1∘⋯∘Wk∘ϕk:k∈N≥1,ϕi∈Φ(G),Wi∈Aff⁡(Rn)},\operatorname{INN}_{\Phi(\mathcal{G})}:=\{W_1 \circ\phi_1 \circ\cdots\circ W_k \circ\phi_k : k \in\mathbb{N}_{\geq1}, \phi_i \in\Phi(\mathcal{G}), W_i \in\operatorname{Aff}(\mathbb{R}^n)\},

where Aff⁡(Rn)\operatorname{Aff}(\mathbb{R}^n) denotes the affine group, i.e., x↦Wi(x)=Aix+bix \mapsto W_i(x)=A_i x+b_i for some Ai∈GL⁡(Rn)A_i \in\operatorname{GL}(\mathbb{R}^n) and bi∈Rnb_i \in\mathbb{R}^n. Importantly, INN⁡Φ(G)⊆Diff⁡r(Rn;Rn)\operatorname{INN}_{\Phi(\mathcal{G})} \subseteq\operatorname{Diff}^r(\mathbb{R}^n;\mathbb{R}^n). Moreover, by composing with some appropriate W0∈Aff⁡(Rn)W_0 \in\operatorname{Aff}(\mathbb{R}^n), one can always force the resulting diffeomorphism to be such that 00 maps to itself.

Then, we say that a set of maps G^⊆Cr(Rn;Rn)\widehat{\mathcal{G}} \subseteq C^r(\mathbb{R}^n;\mathbb{R}^n) is a Wr,∞W^{r,\infty}-universal approximator for G⊆Cr(Rn;Rn)\mathcal{G} \subseteq C^r(\mathbb{R}^n;\mathbb{R}^n) if for any compact set K⊆RnK \subseteq\mathbb{R}^n, any ϵ>0\epsilon> 0 and any g∈Gg \in\mathcal{G} there is a g^∈G^\widehat{g} \in\widehat{\mathcal{G}} such that

∥g−g^∥K,r,∞:=∑∣α∣≤rsup⁡x∈K∥Dαg(x)−Dαg^(x)∥<ϵ.\|g-\widehat{g}\|_{K,r,\infty}:=\sum_{|\alpha|\leq r}\sup_{x\in K}\|D^\alpha g(x)-D^\alpha\widehat{g}(x)\|<\epsilon.

The next theorem characterizes Wr,∞W^{r,\infty}-universality of neural ODEs (NODEs), that is, using NODEs, one has control over a Sobolev norm.

Theorem 4 ([13 Thm. 44]) Let r∈N≥0r \in\mathbb{N}_{\geq0} and suppose that G^⊆Lip⁡r(Rn;Rn)\widehat{\mathcal{G}} \subseteq\operatorname{Lip}^r(\mathbb{R}^n;\mathbb{R}^n) is a Wr,∞W^{r,\infty}-universal approximator for Lip⁡r(Rn;Rn)\operatorname{Lip}^r(\mathbb{R}^n;\mathbb{R}^n), then INN⁡Φ(G^)\operatorname{INN}_{\Phi(\widehat{\mathcal{G}})} is a Wr,∞W^{r,\infty}-universal approximator for Diff⁡max⁡{r,1}(Rn;Rn)\operatorname{Diff}^{\max\{r,1\}}(\mathbb{R}^n;\mathbb{R}^n).

Concretely, what Theorem 4 means is that if we fix some r≥0r \ge0, then for some appropriate set of globally Lipschitz vector fields G^\widehat{\mathcal{G}}, one may hand us any diffeomorphism φ∈Diff⁡max⁡{r,1}(Rn;Rn)\varphi\in\operatorname{Diff}^{\max\{r,1\}}(\mathbb{R}^{n};\mathbb{R}^{n}), a compact set K⊆RnK \subseteq\mathbb{R}^{n} and some ϵ>0\epsilon> 0, then we know that there is some k∈N≥1k \in\mathbb{N}_{\ge1} such that

∥φ−W1∘φ1∘⋯∘Wk∘φk∥K,r,∞<ϵ(5)\lVert\varphi- W_{1} \circ\varphi_{1} \circ\cdots\circ W_{k} \circ\varphi_{k} \rVert_{K,r,\infty} < \epsilon \tag*{(5)}

for φ1,…,φk∈Φ(G^)\varphi_{1},\ldots,\varphi_{k} \in\Phi(\widehat{\mathcal{G}}) and W1,…,Wk∈Aff⁡(Rn)W_{1},\ldots,W_{k} \in\operatorname{Aff}(\mathbb{R}^{n}). Appropriate choices of G^\widehat{\mathcal{G}} are discussed in [13 p. 19], e.g., MLPs with finite weights and tanh activation functions (since we typically consider r>0r > 0).

As mentioned before, one can always find an INN⁡Φ(G^)\operatorname{INN}_{\Phi(\widehat{\mathcal{G}})}-generated diffeomorphism φθ\varphi_{\theta} that maps 00 to 00. Regarding control, let KK be a compact neighbourhood of 00 and let φθ:=W1∘φ1∘⋯∘Wk∘φk\varphi_{\theta} := W_{1} \circ\varphi_{1} \circ\cdots\circ W_{k} \circ\varphi_{k} be such that (5) holds for φ\varphi generating some V∈VV \in\mathcal{V}. Let x↦W0(x):=x−b0x \mapsto W_{0}(x) := x - b_{0} for b0:=φθ(0)b_{0} := \varphi_{\theta}(0), then W0∘φθW_{0} \circ\varphi_{\theta} (e.g., adjust W1W_{1}) is a diffeomorphism and such that

∥φ−W0∘φθ∥K,r,∞=∥φ−W1∘φ1∘⋯∘Wk∘φk+b0∥K,r,∞<ϵ+∥b0∥K,r,∞<2ϵ\lVert\varphi- W_{0} \circ\varphi_{\theta} \rVert_{K,r,\infty} = \lVert\varphi- W_{1} \circ\varphi_{1} \circ\cdots\circ W_{k} \circ\varphi_{k} + b_{0} \rVert_{K,r,\infty} < \epsilon+ \lVert b_{0} \rVert_{K,r,\infty} < 2\epsilon

as (5) implies that ∥φ(0)−φθ(0)∥<ϵ\lVert\varphi(0) - \varphi_{\theta}(0) \rVert< \epsilon and φ(0)=0\varphi(0) = 0 as φ\varphi generates some V∈VV \in\mathcal{V}.

Approximation of stability certificates

Combining Theorem 2 and Theorem 4, we can detail approximation quality of NODE-based INNs (i.e., INN⁡Φ(G^)\operatorname{INN}_{\Phi(\widehat{\mathcal{G}})}) to construct Lyapunov functions. Regarding notation, we recall that ∥g∥K,∞:=∥g∥K,0,∞=sup⁡x∈K∥g(x)∥\lVert g \rVert_{K,\infty} := \lVert g \rVert_{K,0,\infty} = \sup_{x \in K} \lVert g(x) \rVert. Regarding the power of INNs, one should observe that below we effectively only need to be concerned with LfVθ<0L_{f}V_{\theta} < 0, other properties are direct, globally.

Theorem 5 (Universal approximation of Morse Lyapunov functions for hyperbolic equilibria) Let f∈HS0r(Rn)f \in\mathcal{H}\mathcal{S}_{0}^{r}(\mathbb{R}^{n}), for some r∈N≥1∪{∞}r \in\mathbb{N}_{\ge1} \cup\{\infty\} and n∈N≥2n \in\mathbb{N}_{\ge2}, then

(i) there is a Lyapunov function V∈VV \in\mathcal{V} (i.e., V=∥φ∥22V = \lVert\varphi\rVert_{2}^{2} for φ∈Diff⁡∞(Rn;Rn)\varphi\in\operatorname{Diff}^{\infty}(\mathbb{R}^{n};\mathbb{R}^{n})) such that PA+ATP≺0PA + A^{\mathsf{T}}P \prec0 for A:=Df(0)A := Df(0), P:=∇2V(0)P := \nabla^{2}V(0), certifying that x∗=0x^{*} = 0 is GAS;

(ii) for any compact neighbourhood KK of 00 there is an ϵK>0\epsilon_{K} > 0 such that if some diffeomorphism φθ∈Diff⁡1(Rn;Rn)\varphi_{\theta} \in\operatorname{Diff}^{1}(\mathbb{R}^{n};\mathbb{R}^{n}) satisfies ∥φθ−φ∥K,1,∞<ϵK\lVert\varphi_{\theta} - \varphi\rVert_{K,1,\infty} < \epsilon_{K}, then x↦Vθ(x):=∥φθ(x)−φθ(0)∥22x \mapsto V_{\theta}(x) := \lVert\varphi_{\theta}(x) - \varphi_{\theta}(0) \rVert_{2}^{2} is a Lyapunov function satisfying LfVθ<0L_{f}V_{\theta} < 0 on K∖{x∗}K \setminus\{x^{*}\}, thus certifying global asymptotic stability on {x∈K:Vθ(x)≤min⁡x∈∂KVθ(x)}\{x \in K : V_{\theta}(x) \le\min_{x \in\partial K} V_{\theta}(x)\}; and

(iii) such a φθ\varphi_{\theta} can be generated through an invertible neural network INN⁡Φ(G^)\operatorname{INN}_{\Phi(\widehat{\mathcal{G}})}, where G^⊆Lip⁡1(Rn;Rn)\widehat{\mathcal{G}} \subseteq\operatorname{Lip}^{1}(\mathbb{R}^{n};\mathbb{R}^{n}) is a W1,∞W^{1,\infty}-universal approximator of Lip⁡1(Rn;Rn)\operatorname{Lip}^{1}(\mathbb{R}^{n};\mathbb{R}^{n}).

Proof (i) Since f∈HS0r(Rn)f \in\mathcal{H}\mathcal{S}_{0}^{r}(\mathbb{R}^{n}), by Theorem 2 there is a Lyapunov function V∈VV \in\mathcal{V} certifying that x∗=0x^{*} = 0 is GAS, i.e., there is a φ∈Diff⁡∞,+(Rn;Rn)\varphi\in\operatorname{Diff}^{\infty,+}(\mathbb{R}^{n};\mathbb{R}^{n}) such that x↦V(x)=∥φ(x)∥22x \mapsto V(x) = \lVert\varphi(x) \rVert_{2}^{2}. The fact that VV can be chosen to satisfy a Lyapunov equation (or inequality PA+ATP≺0PA + A^{\mathsf{T}}P \prec0 for that matter), locally, follows from the proof of [14 Thm. 3.1].

(iii) By Theorem 4, we can find a φθ∈Diff⁡1(Rn;Rn)\varphi_{\theta} \in\operatorname{Diff}^{1}(\mathbb{R}^{n};\mathbb{R}^{n}), generated by some invertible neural network INN⁡Φ(G^)\operatorname{INN}_{\Phi(\widehat{\mathcal{G}})}, such that ∥φ−φθ∥K,1,∞<ϵK\lVert\varphi- \varphi_{\theta} \rVert_{K,1,\infty} < \epsilon_{K}.

(ii) We claim that there is a choice of ϵK\epsilon_{K} such that x↦∥φθ(x)∥22=:Vθ(x)x \mapsto\lVert\varphi_{\theta}(x) \rVert_{2}^{2} =: V_{\theta}(x) is a Lyapunov function on KK. As discussed in § 4, if needed, one can redefine the map φθ\varphi_{\theta} to be $x \mapsto φθ(x)−φθ(0)\varphi_\theta(x)-\varphi_\theta(0), i.e., via composition with some W0∈Aff⁡(Rn)W_0 \in\operatorname{Aff}(\mathbb{R}^n). Hence Vθ(x)=0⟺x=0V_\theta(x)=0 \Longleftrightarrow x=0. We remark, with some abuse of notation, that if ∥φ−φθ∥K,1,∞<12ϵK\|\varphi-\varphi_\theta\|_{K,1,\infty}<\frac{1}{2}\epsilon_K, then the shifted map satisfies ∥φ−φθ∥K,1,∞<ϵK\|\varphi-\varphi_\theta\|_{K,1,\infty}<\epsilon_K.

Secondly, we see that Vθ(x)>0V_\theta(x)>0, for all x∈K∖{0}x\in K\setminus\{0\} and in fact, all sublevel sets {x∈Rn:Vθ(x)≤c}\{x\in\mathbb{R}^n:V_\theta(x)\le c\}, for c>0c>0, are compact as those sets are diffeomorphisms of the compact balls Bc(0)={x∈Rn:∥x∥22≤c}B_{\sqrt{c}}(0)=\{x\in\mathbb{R}^n:\|x\|_2^2\le c\}.

The remaining property to check is that LfVθ<0L_fV_\theta<0 on K∖{0}K\setminus\{0\}. We know that LfV<0L_fV<0 on K∖{0}K\setminus\{0\}, plus that x↦LfVθ(x)=Lf∥φθ(x)∥22=2⟨φθ(x),Dφθ(x)f(x)⟩x\mapsto L_fV_\theta(x)=L_f\|\varphi_\theta(x)\|_2^2=2\langle\varphi_\theta(x),D\varphi_\theta(x)f(x)\rangle. Consider now some compact ball Bϱ(0)‾⊆int⁡(K)\overline{B_\varrho(0)}\subseteq\operatorname{int}(K), for some ϱ>0\varrho>0 that we select later, and study LfVθL_fV_\theta on the compact set W:=K∖Bϱ(0)W:=K\setminus B_\varrho(0). Let −2β:=max⁡x∈WLfV(x)<0-2\beta:=\max_{x\in W}L_fV(x)<0. Then, as LfVθ=LfV+(LfVθ−LfV)L_fV_\theta=L_fV+(L_fV_\theta-L_fV). We have that LfVθ<0L_fV_\theta<0 on WW if (LfVθ−LfV)≤β(L_fV_\theta-L_fV)\le\beta on WW. Now, as DφTφ−DφθTφθ=(Dφ−Dφθ)Tφ+DφθT(φ−φθ)D\varphi^{\mathsf T}\varphi-D\varphi_\theta^{\mathsf T}\varphi_\theta=(D\varphi-D\varphi_\theta)^{\mathsf T}\varphi+D\varphi_\theta^{\mathsf T}(\varphi-\varphi_\theta) and ∥Dφθ∥W,∞≤∥Dφ∥W,∞+∥Dφθ−Dφ∥W,∞\|D\varphi_\theta\|_{W,\infty}\le\|D\varphi\|_{W,\infty}+\|D\varphi_\theta-D\varphi\|_{W,\infty} see that

∣LfVθ−LfV∣=∣Lf(V−Vθ)∣≤∥f∥W,∞∥DV−DVθ∥W,∞≤2∥f∥W,∞∥DφTφ−DφθTφθ∥W,∞≤2∥f∥W,∞∥Dφ−Dφθ∥W,∞∥φ∥W,∞+2∥f∥W,∞∥φ−φθ∥W,∞(∥Dφ∥W,∞+∥Dφθ−Dφ∥W,∞).\begin{aligned} |L_fV_\theta-L_fV|&=|L_f(V-V_\theta)|\le\|f\|_{W,\infty}\|DV-DV_\theta\|_{W,\infty}\\ &\le2\|f\|_{W,\infty}\|D\varphi^{\mathsf T}\varphi-D\varphi_\theta^{\mathsf T}\varphi_\theta\|_{W,\infty}\\ &\le2\|f\|_{W,\infty}\|D\varphi-D\varphi_\theta\|_{W,\infty}\|\varphi\|_{W,\infty}\\ &\quad+2\|f\|_{W,\infty}\|\varphi-\varphi_\theta\|_{W,\infty}(\|D\varphi\|_{W,\infty}+\|D\varphi_\theta-D\varphi\|_{W,\infty}). \end{aligned}

Let κ1:=∥f∥W,∞\kappa_1:=\|f\|_{W,\infty}, κ2:=∥φ∥W,∞\kappa_2:=\|\varphi\|_{W,\infty} and κ3:=∥Dφ∥W,∞\kappa_3:=\|D\varphi\|_{W,\infty}. If ∥φ−φθ∥W,1,∞≤ϵ\|\varphi-\varphi_\theta\|_{W,1,\infty}\le\epsilon holds for

0<ϵ≤−(κ2+κ3)+(κ2+κ3)2+2β/κ12,0<\epsilon\le\frac{-(\kappa_2+\kappa_3)+\sqrt{(\kappa_2+\kappa_3)^2+2\beta/\kappa_1}}{2},

then ∣LfVθ−LfV∣≤β|L_fV_\theta-L_fV|\le\beta. Hence, max⁡x∈WLfVθ≤−β<0\max_{x\in W}L_fV_\theta\le-\beta<0.

Now we discuss what happens within Bϱ(0)B_\varrho(0), better yet, how to select ϱ>0\varrho>0. Observe that P=∇2V(x∗)=2Dφ(0)TDφ(0)P=\nabla^2V(x^*)=2D\varphi(0)^{\mathsf T}D\varphi(0), and this PP satisfies PDf(0)+Df(0)TP≺0PDf(0)+Df(0)^{\mathsf T}P\prec0. Following a standard argument (e.g., see the proof of Khalil, [16]Thm. 4.7), it can be shown that there is a ϱ>0\varrho>0 and κϱ>0\kappa_\varrho>0 such that LfV(x)<−2κϱ∥x∥22L_fV(x)<-2\kappa_\varrho\|x\|_2^2 on Bϱ(0)B_\varrho(0). Indeed, the reason that such a VV exists follows since f∈HS0r(Rn)f\in\mathcal{H}S_0^r(\mathbb{R}^n), see Jongeneel, [14]Thm. 3.1.

Then, as above, a useful decomposition is as follows:

12Lf(Vθ−V)=⟨φθ−φ,Dφf⟩+⟨φ,D(φθ−φ)f⟩+⟨φθ−φ,D(φθ−φ)f⟩.\frac{1}{2}L_f(V_\theta-V)=\langle\varphi_\theta-\varphi,D\varphi f\rangle+\langle\varphi,D(\varphi_\theta-\varphi)f\rangle+\langle\varphi_\theta-\varphi,D(\varphi_\theta-\varphi)f\rangle.

Akin to κϱ\kappa_\varrho, one can set κf:=sup⁡x∈Bϱ(0)∖{0}∥f(x)∥2/∥x∥2\kappa_f:=\sup_{x\in B_\varrho(0)\setminus\{0\}}\|f(x)\|_2/\|x\|_2, where one may shrink ϱ>0\varrho>0 if needed. In addition, set κ4:=∥Dφ∥Bϱ(0),∞\kappa_4:=\|D\varphi\|_{B_\varrho(0),\infty}. Now, if the Sobolev bound ∥φ−φθ∥Bϱ(0),1,∞≤ϵ′\|\varphi-\varphi_\theta\|_{B_\varrho(0),1,\infty}\le\epsilon' holds, one obtains, via the mean value theorem and convexity of Bϱ(0)B_\varrho(0) (e.g., ∥φθ(x)−φ(x)∥2≤ϵ′∥x∥2\|\varphi_\theta(x)-\varphi(x)\|_2\le\epsilon'\|x\|_2), that ∣Lf(Vθ−V)(x)∣≤2κfϵ′(2κ4+ϵ′)∥x∥22|L_f(V_\theta-V)(x)|\le2\kappa_f\epsilon'(2\kappa_4+\epsilon')\|x\|_2^2. Hence, if ϵ′>0\epsilon'>0 is such that

0<ϵ′≤−2κ4+4κ42+2κϱ/κf2,0<\epsilon'\le\frac{-2\kappa_4+\sqrt{4\kappa_4^2+2\kappa_\varrho/\kappa_f}}{2},

then ∣LfVθ(x)−LfV(x)∣≤κϱ∥x∥22|L_fV_\theta(x)-L_fV(x)|\le\kappa_\varrho\|x\|_2^2. In its turn, this implies that LfVθ(x)<−κϱ∥x∥22L_fV_\theta(x)<-\kappa_\varrho\|x\|_2^2 on Bϱ(0)B_\varrho(0).

At last, we conclude by selecting ϵK=12min⁡{ϵ,ϵ′}\epsilon_K=\frac{1}{2}\min\{\epsilon,\epsilon'\}.

Theorem 5 can only handle n≥2n\ge2 (as item (iii) relies on Ishikawa et al. [13]), however, we mention the following fact. As such, the case n=1n=1 is irrelevant from the learning perspective.

Remark 6 (The case n=1n = 1). Let f∈S0r(Rn)f \in S_{0}^{r}(\mathbb{R}^{n}), for some r∈N≥1∪{∞}r \in\mathbb{N}_{\ge1} \cup\{\infty\} and suppose that n=1n = 1. In this case, the trivial Morse Lyapunov function, i.e., x↦V(x)=x2x \mapsto V(x) = x^{2}, always works as xf(x)<0xf(x) < 0 on R∖{0}\mathbb{R} \setminus\{0\}. See Sontag and Sussmann [29] for further remarks.

The proof of Theorem 5 reveals that as long as asymptotic stability can be certified using a Morse Lyapunov function, one can still characterize a form of universality. However, on an annulus/shell.

Corollary 7 (Universal approximation of Morse Lyapunov functions). Let f∈S0r(Rn)f \in S_{0}^{r}(\mathbb{R}^{n}), for some r∈N≥1∪{∞}r \in\mathbb{N}_{\ge1} \cup\{\infty\} and n∈N≥2n \in\mathbb{N}_{\ge2}. Suppose, there is a Lyapunov function V∈VV \in\mathcal{V} (i.e., V=∥φ∥22V = \lVert\varphi\rVert_{2}^{2} for φ∈Diff⁡∞(Rn;Rn)\varphi\in\operatorname{Diff}^{\infty}(\mathbb{R}^{n};\mathbb{R}^{n})), certifying that x∗=0x^{\ast} = 0 is GAS. Then,

  • (i) for any compact neighbourhood KK of 00 and radius ϱ>0\varrho> 0 such that Bϱ(0)‾⊂int⁡(K)\overline{B_{\varrho}(0)} \subset\operatorname{int}(K), there is an ϵK,ϱ>0\epsilon_{K,\varrho} > 0 such that if some diffeomorphism φθ∈Diff⁡1(Rn;Rn)\varphi_{\theta} \in\operatorname{Diff}^{1}(\mathbb{R}^{n};\mathbb{R}^{n}) satisfies ∥φθ−φ∥K,1,∞<ϵK,ϱ\lVert\varphi_{\theta} - \varphi\rVert_{K,1,\infty} < \epsilon_{K,\varrho}, then x↦Vθ(x):=∥φθ(x)−φθ(0)∥22x \mapsto V_{\theta}(x) := \lVert\varphi_{\theta}(x) - \varphi_{\theta}(0)\rVert_{2}^{2} is a Lyapunov function satisfying LfVθ<0L_{f}V_{\theta} < 0 on K∖Bϱ(0)K \setminus B_{\varrho}(0); and

  • (ii) such a φθ\varphi_{\theta} can be generated through an invertible neural network INN⁡Φ(G^)\operatorname{INN}_{\Phi(\widehat{\mathcal{G}})}, where G^⊆Lip⁡1(Rn;Rn)\widehat{\mathcal{G}} \subseteq\operatorname{Lip}^{1}(\mathbb{R}^{n};\mathbb{R}^{n}) is a W1,∞W^{1,\infty}-universal approximator of Lip⁡1(Rn;Rn)\operatorname{Lip}^{1}(\mathbb{R}^{n};\mathbb{R}^{n}).

We interpret Corollary 7 as showing that one can learn a Morse Lyapunov function that certifies practical asymptotic stability, cf. [21 §25]. Arguably the most powerful ramification follows when f∈HS0r(Rn)f \in\mathcal{HS}_{0}^{r}(\mathbb{R}^{n}), as in that case, there is always a Lyapunov function V∈VV \in\mathcal{V}. To include Bϱ(0)B_{\varrho}(0), one needs to handle that LfVθL_{f}V_{\theta} vanishes at x∗=0x^{\ast} = 0. To that end, Theorem 5 (i) requires more control of VV, locally. Such a VV always exists (Jongeneel [14]), but we do not enforce this structure through INN⁡Φ(G^)\operatorname{INN}_{\Phi(\widehat{\mathcal{G}})}. Hence, Corollary 7 provides at the moment the most practically relevant guarantees when using vanilla NODEs and INN⁡Φ(G^)\operatorname{INN}_{\Phi(\widehat{\mathcal{G}})}. We are not aware of similar results for other contemporary architectures. However, the following remark illustrates that the above extends to PŁ-functions. Exploiting this is future work.

Remark 8 (Equivalence of Morse and PŁ Lyapunov functions on compact sets). It turns out that on compact sets, using smooth PŁ Lyapunov functions is equivalent to using Morse Lyapunov functions. First, C2C^{2}-smooth PŁ-functions, with a unique critical point, are Morse by [27 Cor. 2.17] and [6 Thm. 1.1]. The other direction also works, on compact sets. The local argument follows from [27 Cor. 2.11] and continuity allows for an extension to arbitrary compact neighbourhoods. To be somewhat complete, suppose VV is Morse. Then, around x∗=0x^{\ast} = 0 we have V(x)=12⟨x,Px⟩+o(∥x∥22)V(x) = \frac{1}{2}\langle x, Px\rangle+ o(\lVert x\rVert_{2}^{2}) for P>0P > 0 and ∇V(x)=Px+o(∥x∥2)\nabla V(x) = Px + o(\lVert x\rVert_{2}), but then there is a ϱ>0\varrho> 0 such that on Bϱ(0)B_{\varrho}(0) we have V(x)≤2λmax⁡(P)∥x∥22V(x) \le2\lambda_{\max}(P)\lVert x\rVert_{2}^{2} and ∥∇V(x)∥22≥12λmin⁡(P)∥x∥22\lVert\nabla V(x)\rVert_{2}^{2} \ge\frac{1}{2}\lambda_{\min}(P)\lVert x\rVert_{2}^{2}. Putting it together yields ∥∇V(x)∥22≥μϱV(x)\lVert\nabla V(x)\rVert_{2}^{2} \ge\mu_{\varrho}V(x) for μϱ:=(1/8)λmin⁡(P)2/λmax⁡(P)\mu_{\varrho} := (1/8)\lambda_{\min}(P)^{2}/\lambda_{\max}(P). Now, if we work on a compact neighbourhood KK of 00, we can define μ1:=max⁡x∈K∖Bϱ(0)V(x)>0\mu_{1} := \max_{x\in K\setminus B_{\varrho}(0)} V(x) > 0 and μ2:=min⁡x∈K∖Bϱ(0)∥∇V(x)∥22>0\mu_{2} := \min_{x\in K\setminus B_{\varrho}(0)} \lVert\nabla V(x)\rVert_{2}^{2} > 0 to obtain ∥∇V(x)∥22≥(μ2/μ1)V(x)\lVert\nabla V(x)\rVert_{2}^{2} \ge(\mu_{2}/\mu_{1})V(x) on K∖Bϱ(0)K\setminus B_{\varrho}(0). We can conclude by setting μ:=min⁡{μϱ,μ2/μ1}>0\mu:= \min\{\mu_{\varrho},\mu_{2}/\mu_{1}\} > 0.

Numerical experiments

The purpose of this section is to provide an illustration of the above, hence we consider systems that can be visualized (i.e., planar dynamical systems).

We designed the experiments such that all architectures have roughly the same number of parameters. We compare the NODE-INN (e.g., Ishikawa et al. [13] and this work) against Coupling-INN (e.g., Teshima et al. [31]; Zhong et al. [35]), PLNet (via BiLipNet) (e.g., Wang et al. [32]) and a standard MLP (e.g. using a tanh activation). The learning is implemented via Py-Torch, using a scale-free loss function of the form

θ↦ℓ(θ):=1N∑i=1Nmax⁡{0,LfVθ(xi)Vθ(xi)+m}\theta\mapsto\ell(\theta) := \frac{1}{N}\sum_{i=1}^{N} \max\left\{0,\frac{L_fV_\theta(x_i)}{V_\theta(x_i)}+m\right\}

for some margin m>0m > 0, samples {xi}i=1N\{x_i\}_{i=1}^{N} and θ↦Vθ(x):=∥φθ(x)−φθ(0)∥22\theta\mapsto V_\theta(x) := \lVert\varphi_\theta(x)-\varphi_\theta(0)\rVert_2^2, where the precise form of φθ\varphi_\theta depends on the architecture. We remark that for any VθV_\theta generated through NODE-INN, positive definiteness, properness and the absence of spurious critical points hold by construction. Only the condition LfVθ<0L_fV_\theta< 0 needs to generalize beyond the samples. In contrast, for MLP, all defining properties of a Lyapunov function need to generalize.

Experiment A

We start with learning a Lyapunov function for (4). We present some experimental results in Figure 1. One observes similar behaviour when comparing NODE-INN and PLNet, although PLNet is faster, both per iteration and with respect to convergence as a function of total iterations.

Figure 1. Experiment A. NODE-INN generates φθ\varphi_\theta via INN⁡Φ(G)\operatorname{INN}_{\Phi}(\mathcal{G}) with 3 layers and tanh activation functions generating Lipschitz vector fields. Numerical integration is done via RK4. Coupling-INN is effectively the same but uses affine coupling flows instead. The PLNet is implemented with ν1=1/2\nu_1 = 1/2 and ν2=2\nu_2 = 2, whereas MLP is a standard tanh network (2 hidden layers with 20 neurons). We train using 512 samples, m=10−3m = 10^{-3} and learning rate 10−210^{-2}. The purple levelsets correspond to a Lyapunov function for (4) and the thinner colour-coded curves are some levelsets of VθV_\theta found after 500 iterations of Adam. We show the outcomes of 100 independent runs. The thick curves visualize the first run. The training loss displays the mean and the 25th-75th percentile.

Experiment B

In the second experiment we consider a nonlinear system that is smoothly conjugate to a linear hyperbolic system. In particular, let x˙=Ax\dot{x} = Ax and the matrix JJ be defined through

A:=(−1k0−1),for some k≥0,andJ:=(0−110).A := \begin{pmatrix} -1 & k \\ 0 & -1 \end{pmatrix}, \quad\text{for some } k \ge0, \quad\text{and} \quad J := \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix}.

Now define y:=h(x)y := h(x) as the time-1 map of the flow under x˙=β∥x∥22Jx\dot{x} = \beta\lVert x\rVert_2^2 Jx, for some β>0\beta> 0. Let x↦V(x)=∥x∥22x \mapsto V(x) = \lVert x\rVert_2^2 and see that ⟨x,β∥x∥22Jx⟩=0\langle x, \beta\lVert x\rVert_2^2 Jx\rangle= 0. Hence, hh is a twist (as the angular velocity is variable). Now, as AA is Hurwitz, there is a quadratic form x↦VP(x):=⟨x,Px⟩x \mapsto V_P(x) := \langle x, Px\rangle that certifies that x∗=0x^* = 0 is GAS under x˙=Ax\dot{x} = Ax, but then VP∘h−1V_P \circ h^{-1} certifies that y∗=0y^* = 0 is GAS under y˙=Dh(h−1(y))Ah−1(y)\dot{y} = Dh(h^{-1}(y))Ah^{-1}(y). This setting fits well to NODE-INN, and under appropriate choices of β\beta and kk, the nonlinearities can be significant. We present some numerical experiments in Figure 2. NODE-INN seems to perform well, whereas other architectures appear to resist convergence. However, we recall that with respect to BiLipNet, one employs (ν1,ν2)(\nu_1,\nu_2)-Lipschitz maps F:Rn→Rn\mathcal{F} : \mathbb{R}^n \to\mathbb{R}^n, that is, ν1∥x−y∥≤∥F(x)−F(y)∥≤ν2∥x−y∥\nu_1\lVert x-y\rVert\le\lVert\mathcal{F}(x)-\mathcal{F}(y)\rVert\le\nu_2\lVert x-y\rVert for all x,y∈Rnx,y \in\mathbb{R}^n. To train the corresponding networks, one sets ν1\nu_1 and ν2\nu_2, with the distortion τ:=ν2/ν1\tau:= \nu_2/\nu_1 (or (ν2/ν1)2(\nu_2/\nu_1)^2 for VθV_\theta) controlling the conditioning, e.g., the condition number of ∇2Vθ(0)\nabla^2V_\theta(0) is bounded from above by τ2\tau^2. The results in Figure 2 are for some particular τ\tau, adapting this parameter may improve performance, as shown in Figure 3. This is interesting in the light of Remark 8; to work with PLNet, one needs to find an appropriate τ\tau, and this number may be large, cf. μϱ\mu_\varrho from Remark 8. At the same time, this restriction of the search space may aid convergence. A precise characterization is left for future work.

Figure 2. Experiment B. NODE-INN, Coupling-INN, PLNet and MLP generate their respective map φθ\varphi_\theta as in Experiment A. The purple levelsets correspond to a Lyapunov function for the nonlinear system from § 5.2, for k=6k=6 and β=1/4\beta=1/4, and the thinner colour-coded curves are some levelsets of VθV_\theta found after 1000 iterations of Adam. We show the outcomes of 30 independent runs. Visualization is as in Figure 1.

Figure 3. Experiment B, where the distortion τ\tau with respect to VθV_\theta is adapted.

Conclusion

We have shown that Morse Lyapunov functions are a class of Lyapunov functions with a rich structure that can be exploited both theoretically and algorithmically. Completing the converse theory, also for the other architectures, and providing a rigorous numerical analysis are left for future work.

Generative AI Use

Generative AI tools (Claude Fable 5.2) aided in comprehending some results in the neural networks literature and those tools were used for code and experimental implementation. The author reviewed and verified all AI-assisted content and takes responsibility for the final content of the paper.

References

  1. [1]Amir Ali Ahmadi and Bachir El Khadir. A globally asymptotically stable polynomial vector field with rational coefficients and no local polynomial Lyapunov function. Syst. Control Lett., 121: 50–53, 2018.arxiv.org/abs/1803.06087
  2. [2]Amir Ali Ahmadi, Miroslav Krstic, and Pablo A Parrilo. A globally asymptotically stable polynomial vector field with no polynomial Lyapunov function. In Proc. IEEE CDC, pages 7579–7580, 2011.DOI
  3. [3]Andrea Bacciotti and Lionel Rosier. Lyapunov Functions and Stability in Control Theory. Springer Science & Business Media, Berlin, 2005.
  4. [4]Matthieu Barreau and Nicola Bastianello. Learning maximal regions of attraction on compact sets using Taylor Lyapunov networks. arXiv preprint arXiv:2408.17246v3, 2024.arxiv.org/abs/2408.17246
  5. [5]Nam Parshad Bhatia and Giorgio P Szegö. Stability Theory of Dynamical Systems. Springer, Berlin, 1970.DOI
  6. [6]Nicolas Boumal, Christopher Criscitiello, and Quentin Rebjock. Smooth, globally Polyak–Łojasiewicz functions are nonlinear least-squares. arXiv preprint arXiv:2604.07972, 2026.arxiv.org/abs/2604.07972
  7. [7]Jing Cheng, Ruigang Wang, and Ian R Manchester. Learning stable and passive neural differential equations. In Proc. IEEE CDC, pages 7859–7864, 2024.arxiv.org/abs/2404.12554
  8. [8]Albert Fathi and Pierre Pageault. Smoothing Lyapunov functions. T. Am. Math. Soc., 371(3):1677–1700, 2019.
  9. [9]Lars Grüne, Eduardo D Sontag, and Fabian R Wirth. Asymptotic stability equals exponential stability, and ISS equals finite energy gain—if you twist your eyes. Syst. Control Lett., 38(2):127–134, 1999.
  10. [10]Wolfgang Hahn. Stability of Motion. Springer, Berlin, 1967.DOI
  11. [11]Alice Harting, Karl Henrik Johansson, Sophie Tarbouriech, and Matthieu Barreau. Locally stable neural ODEs with characterized region of attraction. arXiv preprint arXiv:2606.19109, 2026.
  12. [12]Morris W Hirsch. Differential Topology. Springer, New York, 1976.DOI
  13. [13]Isao Ishikawa, Takeshi Teshima, Koichi Tojo, Kenta Oono, Masahiro Ikeda, and Masashi Sugiyama. Universal approximation property of invertible neural networks. J. Mach. Learn. Res., 24(287): 1–68, 2023.arxiv.org/abs/2204.07415
  14. [14]Wouter Jongeneel. On global Lyapunov functions being Morse. arXiv preprint arXiv:2610.05626, 2026b.arxiv.org/abs/2610.05626
  15. [15]Wouter Jongeneel and Roland Schwan. On continuation and convex Lyapunov functions. IEEE Trans. Automat. Contr., 69(10):6895–6906, 2024.DOI
  16. [16]Hassan K Khalil. Nonlinear Systems. Prentice Hall, Upper Saddle River, 2002.
  17. [17]J Zico Kolter and Gaurav Manek. Learning stable deep dynamics models. In Proc. NeurIPS, volume 32, pages 11128–11136, 2019.arxiv.org/abs/2001.06116
  18. [18]Jaroslav Kurzweil. On the inversion of Ljapunov’s second theorem on stability of motion. AMS Transl. Ser. 2, 24:19–77, 1963.
  19. [20]Matthew D Kvalheim and Eduardo D Sontag. Global linearization of asymptotically stable systems without hyperbolicity. Syst. Control Lett., 203:106163, 2025.arxiv.org/abs/2502.07708
  20. [21]Joseph La Salle and Solomon Lefschetz. Stability by Liapunov’s Direct Method with Applications. Academic Press, New York, 1961.
  21. [22]Ian R Manchester, Ruigang Wang, and Nicholas H Barbara. Neural networks in the loop: Learning with stability and robustness guarantees. Annu. Rev. Control Robot. Auton. Syst., 9, 2026.DOI
  22. [23]Jose L Massera. Contributions to stability theory. Ann. Math., 64(1):182–206, 1956.DOI
  23. [24]John Milnor. Morse Theory. Princeton University Press, Princeton, 1963.
  24. [25]Jacob Jr Palis and Wellington De Melo. Geometric Theory of Dynamical systems: an Introduction. Springer, New York, 1982.
  25. [26]Mauricio M Peixoto. On an approximation theorem of Kupka and Smale. J. Differ. Equ., 3(2):214–227, 1967.DOI
  26. [27]Quentin Rebjock and Nicolas Boumal. Fast convergence to non-isolated minima: four equivalent conditions for C^2 functions. Mathematical Programming, 213(1):151–199, 2025.
  27. [28]Shankar Sastry. Nonlinear Systems. Springer, New York, 1999.
  28. [29]Eduardo Sontag and Hector Sussmann. Remarks on continuous feedback. In Proc. IEEE CDC, pages 916–921, 1980.DOI
  29. [30]Eduardo D Sontag. Mathematical Control Theory: Deterministic Finite Dimensional Systems. Springer, New York, 1998.
  30. [31]Takeshi Teshima, Isao Ishikawa, Koichi Tojo, Kenta Oono, Masahiro Ikeda, and Masashi Sugiyama. Coupling-based invertible neural networks are universal diffeomorphism approximators. In Proc. NeurIPS, pages 3362–3373, 2020.arxiv.org/abs/2006.11469
  31. [32]Ruigang Wang, Krishnamurthy Dj Dvijotham, and Ian Manchester. Monotone, bi-Lipschitz, and Polyak-Lojasiewicz networks. In Proc. ICML, pages 50379–50399. PMLR, 2024.arxiv.org/abs/2402.01344
  32. [33]F Wesley Wilson. Smoothing derivatives of functions and applications. T. Am. Math. Soc., 139: 413–428, 1969.DOI
  33. [35]Yuan Zhong, Jiaxin Cheng, Hefu Ye, and Yicong Zhou. Polarnet: Single-minima neural network for modeling Lyapunov functions. arXiv preprint arXiv:2605.09546, 2026.arxiv.org/abs/2605.09546

Paper details

vlnc:W_b8sx89rk
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Claude Fable 5.2

View statement
Looking for related papers…
Contents