betaFeedback

WorksMathematics

On necessary conditions for feedback stabilizability when the control space is a fiber bundle

Abstract

Within control theory, a question of great interest is to understand when a stabilizing feedback exists. Building upon Krasnosel'ski\uı and co-workers, necessary conditions for the existence of a continuous stabilizing feedback have been pioneered by Brockett and Coron, providing so-called adversarial and homological conditions, respectively. Kvalheim recently generalized their conditions from equilibria to compact attractors and—in line with earlier work by Coron—showed that when a control system can be identified with a vector bundle, the homology condition is stronger, that is, the homology condition implies the adversary condition. Kvalheim also showed that this comparison breaks when extending the case of vector bundles to fiber bundles. In this work, we propose and characterize an appropriate adversarial condition for control systems represented by fiber bundles. The homology condition, on the other hand, is already well-defined, showcasing its general applicability. With the generalized conditions at hand, we show that by appropriately relaxing the regularity of the feedback under consideration, that is, by allowing for mild discontinuities, the homology condition implies the adversarial condition for fiber bundles. Thereby, in this work, we extend these earlier comparative results.

Introduction

Within the broad field of automatic control, it is of great importance to understand, a priori, if a stabilization problem admits a solution. For instance, suppose we are given a control system x˙=u\dot{x}=u, with x∈Rx \in\mathbb{R} the state and u∈Ru \in\mathbb{R} the input. As uu is unconstrained, for all states, the so-called control space (see §1.1) can be identified with the trivial vector bundle π1:X×U→X\pi_1 : X \times U \to X, defined through (x,u)↦π1(x,u):=x(x,u) \mapsto\pi_1(x,u) := x, where X:=RX := \mathbb{R}, U:=RU := \mathbb{R} and for any x∈Xx \in X the fiber is simply π1−1(x)=U\pi_1^{-1}(x) = U, see §1.2 for more on bundles. The fiber captures the space of available inputs, given a state xx. Now, suppose that the goal is to stabilize (see §2) the origin using continuous feedback, that is, finding a continuous map μ1:R→R\mu_1 : \mathbb{R} \to\mathbb{R} (more precisely, a section X→X×UX \to X \times U, see §1.1) such that the origin is stable under x˙=μ1(x)\dot{x}=\mu_1(x). Then, as the control system is linear and controllable, such a feedback is guaranteed to exist. One may simply select x↦μ1(x):=−xx \mapsto\mu_1(x) := -x. What if, in addition, the constraint ∣u∣=1|u|=1 must hold? Satisfying this constraint using linear feedback is impossible; the control space fails to be a vector bundle and is effectively the fiber bundle π2:X×S0→X\pi_2 : X \times S^0 \to X, with non-contractible fibers π2−1(x)=S0≅{−1,1}\pi_2^{-1}(x)=S^0 \cong\{-1,1\}. Here, we employ standard notation and recall that “≅\cong” denotes a context-dependent isomorphism. To continue, the feedback x↦μ2(x):=−sgn⁡(x)x \mapsto\mu_2(x) := -\operatorname{sgn}(x) does lead to stabilization of the origin. Compared to utilizing μ1\mu_1, the key difference is that the closed-loop vector field is not continuous anymore. On the other hand, the solutions of the corresponding differential equation, in the sense of Carathéodory, Krasovskii and Filippov, are continuous, forward in time. Can all of this—especially when working with more intricate control systems—be understood in a principled manner? That is the central theme of this note.

What we are looking for is a better understanding of structural features of a control system that enable or preclude stabilization by feedback of a certain regularity. With the focus on uncovering structural features, we study necessary conditions for solving feedback stabilization problems.

To be more explicit, we illustrate the conceptual framework by highlighting the most influential results in this area. First, Brockett’s seminal result [6 Thm. 1.(iii)]. Consider a control system x˙=f(x,u)\dot{x}=f(x,u), with f:Rn×Rm→Rnf:\mathbb{R}^{n}\times\mathbb{R}^{m}\to\mathbb{R}^{n} being continuously differentiable1^{1} (C1C^{1}) and f(0,0)=0f(0,0)=0. Then, Brockett’s condition states that there is a C1C^{1} feedback x↦μ(x)x\mapsto\mu(x) such that 0 is locally asymptotically stable under the closed-loop vector field x↦f(x,μ(x))x\mapsto f(x,\mu(x)), only if (x,u)↦f(x,u)(x,u)\mapsto f(x,u) maps a neighbourhood of (0,0)(0,0) onto a neighbourhood of 0. Topological necessary conditions of this form have been highly influential as one can draw relatively strong conclusions based on generally available information, e.g., see Example 3.4 for a system that is controllable, but cannot be stabilized using continuous feedback, as follows from applying Brockett’s condition. This should be contrasted with linear controllable systems where controllability implies the existence of a globally stabilizing continuous (linear) feedback [45 Ch. 2].

Another influential necessary condition is due to Coron [11 Thm. 2]. Here, we consider the same control systems as Brockett. Let Σ:=dom⁡(f)∖f−1(0)\Sigma:=\operatorname{dom}(f)\setminus f^{-1}(0), that is, Σ\Sigma is the subset of the control space that cannot be mapped to 0 under ff. Then, Coron’s condition states that2^{2}, for n≥2n\ge2, there is a C0C^{0} feedback x↦μ(x)x\mapsto\mu(x) such that 0 is locally asymptotically stable under the closed-loop vector field x↦f(x,μ(x))x\mapsto f(x,\mu(x)), only if

f∗Hn−1(Σ)=Hn−1(Rn∖{0})≅Z.(1)f_{*}H_{n-1}(\Sigma)=H_{n-1}(\mathbb{R}^{n}\setminus\{0\})\cong\mathbb{Z}. \tag*{(1)}

Here, Hk(⋅)H_{k}(\cdot) denotes the kk-th singular homology group with integer coefficients and f∗f_{*} denotes the corresponding induced homomorphism, e.g., consult [5 Ch. IV] or [18 Ch. 2] for more on the relevant algebraic topology. Coron’s condition can be shown to be stronger than Brockett’s condition [11 p. 229], and as with Brockett, this condition is implied by the index theoretic results stated in [24 §52], see also [4, 25]. On the other hand, Brockett’s condition is arguably easier to work with.

Besides relaxing differentiability to continuity, compared to original work by Brockett, generalizations of Brockett’s condition pertain to stabilization of rather generic compact sets (not just equilibrium points) where the control system is defined using structures reminiscent of fiber bundles (not just an Euclidean product space) [30]. The same is true for generalizations of Coron’s condition [27, 28].

Then, recently, Kvalheim showed, in great generality, how well-known necessary conditions for continuous feedback stabilization compare [28]. To be precise, one may focus on (i) the adversary condition, that is, Brockett’s condition [6] and its generalizations [42, 41, 7, 30]; and (ii) the homology condition, that is, Coron’s condition [11] and its generalizations [32, 33, 27, 34]. See also [44, 43, 24, 46, 9, 12, 38, 1, 29] for related work in this space. Here, it is interesting to note that most of these results essentially hinge on the homotopy invariance of the topological degree of a map (e.g., see [17 p. 81, 108]), that is, one typically computes some invariant for a canonical system and then relates this invariant to a homotopy class of the canonical system, e.g., see [22 Ex. 3.4]. Now, regarding the comparison itself, Kvalheim shows that when a control space (i.e., the domain of ff) can be understood as a vector bundle with oriented base space, then, the homology condition is stronger than the adversary condition, that is, if the homology condition holds, so does the adversary condition [28 Thm. 4]; the result of Coron extends. Surprisingly, the situation for fiber bundles is less clear, e.g., when working with sphere bundles, there are counterexamples by Kvalheim [28 Ex. 4] and de Sa and Ames [13 Ex. 4] that show cases where the homology condition holds but the adversarial condition does not (see also Example 1.1 below). One might ask the following.

Question A: “What do these counterexamples convey if one is solely interested in stabilization? Is the homology condition truly occasionally weaker than the adversarial condition?”

An affirmative answer to the latter would be surprising given prior results by Coron and Kvalheim. The goal of this work is to resolve this conundrum. More details regarding Question A are provided below, in particular regarding the meaning of the word “truly” as those counterexamples are evidently correct within the context of those papers.

Contribution We show in this note that the aforementioned counterexamples are not truly counterexamples if one is solely interested in stabilization. The reason that for these examples the homology condition holds, while the adversarial condition does not, is simply the result of the homology condition being necessary for a larger class of feedback controllers; feedback that is possibly discontinuous. Indeed, those counterexamples admit a stabilizing feedback, just not a continuous one.

We first show that the case of fiber bundles can be better understood by relaxing the regularity of the feedback. In particular: (i) we construct a new adversarial condition (see Theorem 3.3); (ii) show that this condition clarifies the fiber bundle counterexample by Kvalheim (see Examples 1.1-1.2); and (iii) we show that the homology condition remains stronger (see Corollary 3.6 and Theorem 3.7). Hence, the contribution of this note is to provide further motivation for using the homology condition. We have summarized the above in Figure 1. The goal of this work is not to be as general as possible (e.g., we do not work with generic fiber bundles). Lastly, we point to open problems and conclude in § 4.

Summary diagram of the context and contribution of this work

Figure 1. A summary of the context and contribution of this work. (1a) When the control system (see § 1.1) is such that U→XU \to X is a vector bundle, with XX oriented, then, the homology condition (ii) is stronger than the adversary condition (i). (1b) When U→XU \to X is a fiber bundle (e.g., a sphere bundle), there are examples showcasing that this ordering fails (see Example 1.1). Nevertheless, the system studied in Example 1.1 does admit a stabilizing controller, however, one that is merely continuous on N∖AN \setminus A (not on all of NN). (2) Indeed, we show that the homology condition (ii) holds true for stabilizing control systems with mild discontinuities on AA, hence one could argue that (1b) displays shortcomings of the adversary condition, not of the homology condition. (1c) To overcome these shortcomings, we develop the spherical adversary condition (iii); this adversarial condition accommodates mild discontinuities, handles Example 1.1 and is again implied by the homology condition.

Control systems

To study further necessary conditions for feedback stabilization, we start by defining a (continuous) control system via the triple C:=(X,U,f)\mathcal{C}:=(X,\mathcal{U},f). Here, XX is the state space (an nn-dimensional smooth manifold), U\mathcal{U} a topological space together with a continuous surjective map πU:U→X\pi_{\mathcal{U}}:\mathcal{U}\to X (e.g., a fiber bundle, see §1.2) known as the control space and lastly, f:U→TXf:\mathcal{U}\to TX is a continuous fiber-preserving map defined on the control space, that is, given a continuous section σ:X→U\sigma:X\to\mathcal{U} (i.e., πU∘σ=id⁡X\pi_{\mathcal{U}}\circ\sigma=\operatorname{id}_X), f∘σ:X→TXf\circ\sigma:X\to TX is a vector field. As σ\sigma can be understood as feedback, f∘σf\circ\sigma should be understood as a closed-loop vector field, e.g., consider the following commutative diagram:

U→fTXσ↓↓πXX X(2)\begin{CD} U @>{f}>> TX \\ @V{\sigma}VV @VV{\pi_X}V \\ X @. X \tag*{(2)} \end{CD}

Representing a control system through (1.2) reveals how state-dependent input constraints come to play when synthesizing (stabilizing) feedback controllers. Studying control systems through such a diagram is commonly attributed to Brockett (e.g., see [40 §13.5] and the accompanying notes). Note that (1.2) generalizes the case of x˙=f(x,u)\dot{x}=f(x,u) with f:X×U→TXf:X\times U\to TX, that is, for such a control system we work with a trivial vector bundle πu:U=X×U→X\pi_u:U=X\times U\to X. This remark is important as one might be accustomed to the notation f(x,u)f(x,u) and not f(u)f(u) (or f(ux)f(u_x)). Also note that one typically3 studies local sections, that is, a map σ:N⊆X→U\sigma:N\subseteq X\to U such that πu∘σ=id⁡N\pi_u\circ\sigma=\operatorname{id}_N.

Next, we provide both (classes of) necessary conditions, for a particular feedback stabilization problem. Here, 0TX:X→TX0_{TX}:X\to TX denotes the zero section of the tangent bundle TXTX, that is, x↦0TX(x)=(x,0)x\mapsto0_{TX}(x)=(x,0), locally. With some abuse, we will also call the image 0TX(X)0_{TX}(X) the zero section.

Necessary conditions for feedback stabilization

1.2. Necessary conditions for feedback stabilization. Suppose we work with a control system C=(X,U,f)\mathcal{C}=(X,U,f). Then, given a sufficiently regular (e.g., such that an Euler characteristic4 χ(⋅)\chi(\cdot) is well-defined) compact subset A⊆XA\subseteq X one may ask if there is a (local) section σ\sigma such that AA is (locally) asymptotically stable under the closed-loop vector field f∘σf\circ\sigma. This is a question of stabilization through continuous feedback (see § 2 for more on stability theory). A necessary condition that generalizes Brockett’s condition [6 Thm. 1.(iii)], as proposed by Kvalheim and Koditschek [30 Thm. 3.2], is as follows. Note, one needs χ(A)≠0\chi(A)\ne0.

(i) The adversary condition For any neighbourhood N⊆XN\subseteq X of AA, there is a neighbourhood N0⊆TXN_0\subseteq TX of the zero section 0TX0_{TX}, such that for any N0N_0-valued continuous vector field EE over NN, E(x)=f(ux)E(x)=f(u_x) for some x∈Nx\in N and ux∈πu−1(N)u_x\in\pi_u^{-1}(N).

Observe that, like Brockett’s original necessary condition, the adversary condition hinges on continuity, that is, around the attractor AA, the closed-loop vector field can be made arbitrarily small in magnitude. As we argue below, this is not necessary when one is concerned with stability and continuity of solutions (e.g., semiflows) alone; continuity of the closed-loop vector field is not necessary.

Next, Coron’s condition has been recently generalized by Kvalheim [27 Thm. 2]. Essentially, by generalizing the homotopy argument as exploited by Krasnosel’skii and Zabreiko [24 Lem. 52.1], cf. [27 Thm. 1]. Regarding notation, given a subset Y⊆XY\subseteq X, let ΣY:=πu−1(Y)∖f−1(0TX)\Sigma_Y:=\pi_u^{-1}(Y)\setminus f^{-1}(0_{TX}), that is, ΣY⊆U\Sigma_Y\subseteq U is the subset of the control space not mapped to the zero section of TYX:=⋃y∈YTyXT_YX:=\bigcup_{y\in Y}T_yX under ff, cf. [11].

(ii) The homology condition Let GG be any uniquely integrable continuous vector field on XX such that AA is asymptotically stable under GG. Then, for any sufficiently small neighbourhood NN of AA we have

G∗H∙(N∖A)⊆f∗H∙(ΣN∖A)inH∙(TN∖AX∖0TX).(3)G_*H_\bullet(N\setminus A)\subseteq f_*H_\bullet(\Sigma_{N\setminus A})\quad\text{in}\quad H_\bullet(T_{N\setminus A}X\setminus0_{TX}). \tag*{(3)}

Here, H∙(X)H_\bullet(X) denotes graded homology (i.e., H∙(X)=⨁kHk(X)H_\bullet(X)=\bigoplus_k H_k(X)). In this note, we employ singular homology with integer coefficients for simplicity of exposition. One may consider Čech cohomology instead to capture shape equivalence of attractors and their basin of attraction. For more details regarding this generalization, see [23, 16, 30].

When it comes to comparing these conditions, Kvalheim shows the following. When the control system has the structure of a vector bundle, that is, πu:U→X\pi_u : U \to X is a vector bundle, and XX is orientable, then, the homology condition is a sharper necessary condition (i.e., (ii) ⇒\Rightarrow (i): when the homological condition holds, so does the adversarial condition) [28 Thm. 4].

Before continuing, we recall some terminology. Let EE and BB be topological spaces, then, a tuple (E,B,π)(E,B,\pi), such that π:E→B\pi: E \to B is a continuous surjective map, with a vector space structure on each fiber π−1(b)\pi^{-1}(b), is a kk-dimensional vector bundle when for any b∈Bb \in B, there is an open neighbourhood U∋bU \ni b such that π−1(U)\pi^{-1}(U) is homeomorphic to U×RkU \times\mathbb{R}^{k}, with this homeomorphism restricting to a vector space isomorphism when taking π−1(b)\pi^{-1}(b) to {b}×Rk\{b\} \times\mathbb{R}^{k}. We point the reader to [19] for more on vector bundles. Relaxing the vector space structure provides us with a fiber bundle. That is, let EE, BB and FF be topological spaces, then the tuple (F,E,B,π)(F,E,B,\pi) is a fiber bundle when π:E→B\pi: E \to B is a continuous surjective map and for any b∈Bb \in B there is an open neighbourhood U∋bU \ni b such that π−1(U)\pi^{-1}(U) is homeomorphic to U×FU \times F, and this homeomorphism is fiber-preserving.

Now, let π:E→B\pi: E \to B be an nn-dimensional vector bundle (e.g., πx:TB→B\pi_x : TB \to B), then, there are several methods to define the so-called sphere bundle Sn−1↪E→BS^{n-1} \hookrightarrow E \to B. One may simply demand that all fibers are topological spheres, that is, π−1(b)≅Sn−1\pi^{-1}(b) \cong S^{n-1} for all b∈Bb \in B. Indeed, a sphere bundle is a fiber bundle that is not a vector bundle. As vector bundles, with BB being paracompact and Hausdorff, have an inner product [19 Prop. 1.2], one may also define the sphere bundle, induced by EE, as S(E):={e∈E:⟨e,e⟩=1,e∈π−1(b),b∈B}S(E) := \{e \in E : \langle e,e\rangle= 1, e \in\pi^{-1}(b), b \in B\}. We will use the latter. Note that if BB is a smooth manifold with tangent bundle TBTB, we may endow BB with a Riemannian metric gg, and thus, S(TB)S(TB) will be defined using this metric. We will only comment on the employed metric if relevant. The norm induced by this metric will be denoted by ∥⋅∥g\lVert\cdot\rVert_g or simply ∥⋅∥\lVert\cdot\rVert. In general, one could define the zero section 0E:B→E0_E : B \to E and study a quotient space of E∖0EE \setminus0_E, e.g., use Rn∖{0}≅Sn−1\mathbb{R}^{n} \setminus\{0\} \cong S^{n-1}. For instance, regarding (1.3) we have that

TN∖AX∖0TX≅S(TN∖AX).(4)T_{N\setminus A}X \setminus0_{TX} \cong S(T_{N\setminus A}X). \tag*{(4)}

Then, as stated before, up and including (oriented) vector bundles, it is known that the homology condition implies the adversary condition. One may wonder if the comparison stops there. To that end, in the case of fiber bundles, Kvalheim provides the following counterexample [28 Ex.4]. Counter in the sense that this example shows that the homology condition (ii) does not always imply the adversary condition (i) when looking at fiber bundles that are not vector bundles.

Example 1.1 ((ii) ⇒\Rightarrow (i): a counterexample using sphere bundles [28 Ex.4]). Let X:=RnX := \mathbb{R}^{n}, πu:U→X\pi_u : \mathcal{U} \to X be the unit-sphere bundle, that is, U:=S(TX)⊂TX\mathcal{U} := S(TX) \subset TX, and let f:U→TXf : \mathcal{U} \to TX be the inclusion map. Now, consider A:={0}A := \{0\}. The homology condition (ii) readily holds. To see this, first note that (any) GG is homotopy equivalent to G/∥G∥G/\lVert G\rVert, away from AA. Then πu∘G=G/∥G∥\pi_u \circ G = G/\lVert G\rVert and as G=f∘GG = f \circ G so conclude on set inclusion. The adversary condition (i), on the other hand, fails as there is no input available to render the vector field sufficiently small, that is, for any vector field EE such that ∥E(x)∥∈[0,1)\lVert E(x)\rVert\in[0,1) for all x∈Xx \in X, there is no admissible ux∈πu−1(U)u_x \in\pi_u^{-1}(U) as ∥f(ux)∥=1∉(0,1)\lVert f(u_x)\rVert= 1 \notin(0,1).

Now consider the following example.

Example 1.2 (Finite-time stability and sphere bundles). Consider a control system on Rn\mathbb{R}^{n}, where x˙=u\dot{x}=u with u∈Sn−1u \in S^{n-1}, that is, we are in the setting as in

Example 1.1 above. Now define the feedback μ\mu as follows. On Rn∖{0}\mathbb{R}^{n}\setminus\{0\}

x↦μ(x):=−x∥x∥2,x \mapsto\mu(x):=-\frac{x}{\lVert x\rVert_{2}},

whereas at x=0x=0, we let x↦μ(x)=Sn−1x\mapsto\mu(x)=\mathbb{S}^{n-1}, that is, the map is set-valued. Now, we can study x˙∈F[μ](x)\dot{x}\in\mathcal{F}[\mu](x), for F\mathcal{F} the Filippov operator (see § 2). Indeed, the origin is (finite-time) asymptotically stable under this dynamical system5^{5}, despite failing to satisfy the adversary condition (i). ◦

The example above shows that the counterexample (Ex. 1.1) only holds due to a lack of continuity of any closed-loop vector field around 00. Stability of the origin under these vector fields, on the other hand, is not obstructed (see § 4 for a similar example in the context of CBFs). Neither is continuity of the solutions. Still, is the fact that the homology condition does hold for Example 1.1 a coincidence or a feature? Better yet, can we alter the adversarial condition and make it suit Example 1.1? With this in mind, we ask the following.

Question B: “Is there an adversarial condition, suitable for Example 1.1-1.2; a condition that is applicable in a sufficiently general setting; a condition that is again implied by the homology condition?”

An affirmative answer would resolve Question A as motivated and stated in the above—the homology condition should be understood as being simply more general than the adversarial condition. Indeed, it turns out that a mild adjustment of the adversary condition solves this question, as formalized in this note. It is clear that for such a generalization, the magnitude of adversaries should not play an important role, their “direction” is of importance. Indeed, in Example 1.2, the control system is such that for any neighbourhood NN of the point 00, one can find for any z∈Sn−1z\in\mathbb{S}^{n-1} some ux∈π−1(N∖{0})u_x\in\pi^{-1}(N\setminus\{0\}) such that z=f(ux)/∥f(ux)∥2z=f(u_x)/\lVert f(u_x)\rVert_{2}.

Control systems revisited

With Example 1.2 in mind, we introduce a slight modification into the definition of a control system C=(X,U,f)\mathcal{C}=(X,U,f). Let D⊆XD\subseteq X be a set where closed-loop vector fields may be discontinuous, but remain measurable and locally essentially bounded. If D≠∅D\ne\varnothing, we interpret the resulting discontinuous differential equation in the sense of Filippov, that is, we study set-valued maps F[f∘σ]:X⇒TX\mathcal{F}[f\circ\sigma]:X\Rightarrow TX (see § 2). Note, the discontinuity may arise due to σ\sigma, ff or both. For simplicity, however, we will assume that ff and σ\sigma are only allowed to be discontinuous on ΣD\Sigma_D and DD, respectively (and thus, those maps are continuous on ΣX∖D\Sigma_{X\setminus D} and X∖DX\setminus D, respectively). Overloading notation, we denote the resulting tuple by C=(X,U,f,D,F)\mathcal{C}=(X,U,f,D,\mathcal{F}) (i.e., if D=∅D=\varnothing, we recover (X,U,f)(X,U,f) as in that case F[f∘σ]=f∘σ\mathcal{F}[f\circ\sigma]=f\circ\sigma).

We remark that this definition is inspired by Example 1.1-1.2 and finite-time stabilization. One may study a more general setting.

Lyapunov stability

In this section we follow Mayhew and Teel [35] and highlight a converse Lyapunov theorem applicable to some discontinuous dynamical systems, see also [10]. A reader familiar with stability theory may safely skip this section.

Let XX be an nn-dimensional smooth, Hausdorff and second-countable manifold (see [31 Ch. 1]). Then, there is a smooth embedding X↪R2n+1X \hookrightarrow\mathbb{R}^{2n+1} (XX is diffeomorphic to some closed submanifold of R2n+1\mathbb{R}^{2n+1} [31 Cor. 6.16]). Hence, we will assume in what follows that XX is an embedded submanifold of some Euclidean space.

Now, we will study stability under set-valued vector fields F:X⇉TXF : X \rightrightarrows TX, that is, F(x)⊆TxXF(x) \subseteq T_xX. More concretely, we study stability under the differential inclusion (5).

x˙∈F(x),x∈X.(5)\dot{x} \in F(x), \qquad x \in X. \tag*{(5)}

To handle the inclusion (5), it is convenient to assume that the map F:X⇉TXF : X \rightrightarrows TX satisfies the so-called basic conditions [35 Def. 5]. These conditions are satisfied precisely when

(i) FF is locally bounded, that is, for any compact set K1⊆XK_1 \subseteq X there is a compact set K2⊆TXK_2 \subseteq TX such that F(K1)⊆K2F(K_1) \subseteq K_2;

(ii) FF is outer semicontinuous, that is, graph⁡(F)\operatorname{graph}(F) is closed; and

(iii) F(x)F(x) is nonempty, compact and convex, for any x∈Xx \in X.

Recall, if F:X→TXF : X \to TX is not set-valued, but still locally bounded, then graph⁡(F)\operatorname{graph}(F) being closed implies that FF is continuous.

Consider now a possibly discontinuous ODE x˙=f(x)\dot{x} = f(x), x∈Xx \in X, with ff measurable and locally essentially bounded. Then, the Filippov operator FF [15 p. 85] is defined pointwise through

F[f](x):=⋂r>0⋂N:μ(N)=0conv⁡‾f(Br(x)∖N).(6)\mathcal{F}[f](x) := \bigcap_{r>0}\bigcap_{\substack{N:\mu(N)=0}} \overline{\operatorname{conv}} f\bigl(B_r(x) \setminus N\bigr). \tag*{(6)}

where μ\mu denotes the appropriate Lebesgue measure and conv⁡‾(B)\overline{\operatorname{conv}}(B) the closure of the convex hull of a set BB. Applying the Filippov operator to such an ff leads to a set-valued vector field that satisfies the basic conditions. The rationale is to pass from a discontinuous differential equation to a more regular differential inclusion.

Now, a solution to an inclusion (5) is an absolutely continuous curve [0,T]∋t↦ξ(t)∈X[0,T] \ni t \mapsto\xi(t) \in X that satisfies

ddtξ(t)∣t=s∈F(ξ(s))for a.e. s∈[0,T].\left.\frac{\mathrm{d}}{\mathrm{d}t}\xi(t)\right|_{t=s} \in F(\xi(s)) \quad\text{for a.e. }s \in[0,T].

Then, an inclusion (5) is said to be forward complete when all solutions are defined for all t≥0t \geq0.

Regarding stability, a compact set A⊆XA \subseteq X is said to be (locally) asymptotically stable under (a forward complete) inclusion (5) when

(i) AA is stable, that is, for any open neighbourhood U∞U_\infty of AA, there is an open neighbourhood U0U_0 of AA such that solutions t↦ξ(t)t \mapsto\xi(t) under (5) that start in U0U_0 stay in U∞U_\infty for all time (i.e., ξ(0)∈U0⟹ξ(t)∈U∞ ∀t≥0\xi(0) \in U_0 \Longrightarrow\xi(t) \in U_\infty\ \forall t \geq0); and

(ii) AA is attractive, that is, there is an open neighbourhood VV of AA such that for any open neighbourhood WW of AA there is T≥0T \geq0 such that solutions t↦ξ(t)t \mapsto\xi(t) under (5) starting in VV enter WW in time TT, and stay there (i.e., ξ(0)∈V⟹ξ(t)∈W ∀t≥T\xi(0) \in V \Longrightarrow\xi(t) \in W\ \forall t \geq T).

Let AA be asymptotically stable, then the basin of attraction, denoted B(A)B(A), is the union of all neighbourhoods VV, such that AA is attractive from VV. As AA is compact, without loss of generality, one may define stability through metric neighbourhoods as well [3 Ch. V].

Then, based on [8], [35] provides the following converse Lyapunov theorem.

Theorem 2.1 (A converse Lyapunov theorem for inclusions [35 Cor. 13]). Let the compact subset AA, of some embedded submanifold X⊆RmX \subseteq\mathbb{R}^{m}, be asymptotically stable under x˙∈F(x)\dot{x} \in F(x), with F:X⇉TXF : X \rightrightarrows TX satisfying the basic conditions, then, there is a smooth function V:B(A)→R≥0V : B(A) \to\mathbb{R}_{\geq0} such that

(i) α1(ω(x))≤V(x)≤α2(ω(x))\alpha_{1}(\omega(x)) \leq V(x) \leq\alpha_{2}(\omega(x)), where α1,α2∈K∞\alpha_{1}, \alpha_{2} \in\mathcal{K}_{\infty} and ω\omega is a any proper continuous indicator function for AA; and

(ii) max⁡v∈F(x)⟨∇V(x),v⟩≤−V(x)\max_{v \in F(x)} \langle\nabla V(x),v\rangle\leq-V(x).

In the next section we exploit Item (ii) from Theorem 2.1. It allows for capturing the “direction” of a discontinuous vector field via properties of a smooth object (a gradient flow generated by VV).

Homotopy to a smooth gradient flow Let FF be a vector field on a smooth manifold XX such that AA is asymptotically stable with basin of attraction B(A)⊆XB(A) \subseteq X. Let FF be continuous on B(A)∖AB(A) \setminus A and suppose there is a smooth Lyapunov function VV such that LFV<0L_{F}V < 0 on B(A)∖AB(A) \setminus A (here, LFVL_{F}V denotes the Lie derivative of VV along FF [31 p. 227]). Since XX is smooth, XX admits a Riemannian metric gg [31 Prop. 3.13]. Then, we can define the Riemannian gradient of VV through the relation Df(x)[v]=⟨grad⁡V(x),v⟩gDf(x)[v] = \langle\operatorname{grad} V(x),v\rangle_{g}; this relation must hold for any smooth function f:B(A)→Rf : B(A) \to\mathbb{R} and any v∈TxXv \in T_{x}X. It follows that the set AA is also asymptotically stable under the vector field −grad⁡V-\operatorname{grad} V (i.e., check that −grad⁡V(V)<0-\operatorname{grad} V(V) < 0 on B(A)∖AB(A) \setminus A). Now, define for any s∈[0,1]s \in[0,1] the vector field Hs:=(1−s)F−sgrad⁡VH_{s} := (1-s)F-s\operatorname{grad} V. It follows that HsH_{s} is continuous on B(A)∖AB(A) \setminus A and that LHsV<0L_{H_{s}}V < 0 on B(A)∖AB(A) \setminus A, as such, [0,1]∋s↦Hs[0,1] \ni s \mapsto H_{s} yields a homotopy, on B(A)∖AB(A) \setminus A, from FF to a smooth gradient flow. In particular, it follows that HsH_{s} is nonzero for any ss, that is, we work with the continuous map

Hs:B(A)∖A→TB(A)∖AX∖0TXH_{s} : B(A) \setminus A \to T_{B(A)\setminus A}X \setminus0_{TX}

As we will work in a setting where the existence of such a Lyapunov function can be guaranteed (e.g., via [35, 14]), we can and will use the existence of this homotopy in the remainder of this note. For simplicity, we will write ∇V\nabla V instead of grad⁡V\operatorname{grad} V.

A spherical adversary condition

The first result is a Brockett-like necessary condition for stabilization. It will allow for correctly interpreting Examples 1.1 and 1.2 in relation to the homology condition.

The result is inspired by [30], but with the focus on a less regular closed-loop vector field. This is possible by simply focussing on the smooth gradient flow generated by a Lyapunov function. For simplicity, our compact attractors are slightly less general than the ones in [30 §2] (e.g., Kvalheim and Koditschek consider shape equivalence of an attractor AA and its basin of attraction B(A)B(A), we consider homotopy equivalence), as captured by the following assumption.

Assumption 3.1 (On the pair (A,X)(A,X), with A⊆XA \subseteq X). Let XX be a smooth, orientable, second-countable, Hausdorff manifold and let AA be a compact subset of XX such that ι:A↪X\iota: A \hookrightarrow X is cofibration (e.g., AA is an embedded submanifold).

Note that the pair (A,X)(A,X) from Example 1.2 satisfies Assumption 3.1. We point the reader to [5 Ch. VII] for more comments on cofibrations, and to [20, 21] for more on the relation with dynamical systems. Now, given a set AA, we define its Euler characteristic through (see [18 Thm. 2.44]):

χ(A):=∑k=0∞(−1)krank⁡Hk(A).(7)\chi(A) := \sum_{k=0}^{\infty}(-1)^{k}\operatorname{rank} H_{k}(A). \tag*{(7)}

When the pair (A,X)(A, X) satisfies Assumption 3.1 and AA is asymptotically stable under some sufficiently regular differential equation, we know that (3.1) is well-defined, e.g., since AA is homotopy equivalent to a smooth manifold, see [20].

As will be clarified, it is useful to momentarily impose another, stronger assumption. This assumption is akin to [32, 33] and helps in making sense of “directions”.

Assumption 3.2 (On XX). Let XX be an open subset of Rn\mathbb{R}^{n}.

The most important ramification of Assumption 3.2 is that TX≅X×RnTX \cong X \times\mathbb{R}^{n}. Hence, (closed-loop) vector fields F:X→TXF:X \to TX, or other fiber preserving maps into TXTX, can be identified with maps F′:X→RnF':X \to\mathbb{R}^{n}. This is not the most general setting we will consider; we comment on appropriate generalizations later (e.g., see Corollary 3.6 and §3.1).

Theorem 3.3 (A spherical adversary condition). Let (A,X)(A, X) satisfy Assumptions 3.1-3.2 and consider a control system C=(X,U,f,A,F)\mathcal{C}=(X,U,f,A,\mathcal{F}) in the sense of §1.3. Let N⊆XN \subseteq X be an open neighbourhood of AA and suppose there is a section σ:N→U\sigma:N \to U such that AA is asymptotically stable under f∘σf \circ\sigma, in the sense of Filippov. Then,

(i) χ(A)\chi(A) is well-defined; and

(ii) if χ(A)≠0\chi(A) \ne0, then, for any open neighbourhood N′⊆NN' \subseteq N of AA

ΣN′∖A∋ux↦f(ux)∥f(ux)∥2∈Sn−1\Sigma_{N'\setminus A} \ni u_x \mapsto\frac{f(u_x)}{\lVert f(u_x)\rVert_2} \in S^{n-1}

is surjective, with ff understood as f:U→Rnf:U \to\mathbb{R}^{n}.

Proof. By Assumption 3.1, we may assume that XX is an embedded manifold of some Euclidean space.

Then, as AA is asymptotically stable under a closed-loop vector field F[f∘σ]:N⊆X⇉TNX\mathcal{F}[f\circ\sigma]:N \subseteq X \rightrightarrows T_NX that satisfies the basic conditions (by the definition of C\mathcal{C}), Theorem 2.1 (see [35 Cor. 13]) implies there exists a C∞C^\infty Lyapunov function V:B(A)⊆N→R≥0V:B(A) \subseteq N \to\mathbb{R}_{\ge0}. In particular, AA is asymptotically stable under x˙=−∇V(x)\dot{x}=-\nabla V(x). Let Mc:=V−1([0,c])⊆B(A)M_c:=V^{-1}([0,c]) \subseteq B(A) for some c>0c>0.

As XX is orientable, AA is compact and ∇V\nabla V is of constant rank on ∂Mc\partial M_c, the set ∂Mc\partial M_c is a compact, orientable, (n−1)(n-1)-dimensional embedded submanifold of XX [31 Thm. 5.12]. It may happen that ∂Mc\partial M_c is not connected (e.g., let X=R3X=\mathbb{R}^{3} and A=S2A=S^{2}, see also [32]). However, by compactness, ∂Mc\partial M_c consists out of finite components, say k∈N>0k \in\mathbb{N}_{>0} components.

Then, under our standing assumptions (i.e., AA is asymptotically stable and ι:A↪X\iota:A \hookrightarrow X is a cofibration), we have that χ(Mc)\chi(M_c) is well-defined6 and χ(Mc)=:χ(A)\chi(M_c)=:\chi(A), as B(A)B(A), AA and McM_c are all homotopy equivalent [20].

As TX≅X×RnTX \cong X \times\mathbb{R}^{n}, one can identify sections into the sphere bundle S(TX)S(TX) with maps into Sn−1S^{n-1}. Now consider the maps g,h:∂Mc→Sn−1g,h:\partial M_c \to S^{n-1} as defined through x↦g(x):=f(σ(x))/∥f(σ(x))∥2x \mapsto g(x):=f(\sigma(x))/\lVert f(\sigma(x))\rVert_2 and x↦h(x):=−∇V(x)/∥∇V(x)∥2x \mapsto h(x):=-\nabla V(x)/\lVert\nabla V(x)\rVert_2. Note that hh is smooth, but gg is only known to be continuous. Then, as Lf∘σV<0L_{f\circ\sigma}V<0 on x∈B(A)∖Ax\in B(A)\setminus A (and thus on ∂Mc\partial M_c), we can construct a (single) straight-line (continuous) homotopy from gg to hh (recall the last paragraph of §2). This means that deg⁡(h)=deg⁡(g)\deg(h)=\deg(g) [17 p. 108]. Then, as χ(A)≠0\chi(A)\ne0, the Poincaré-Hopf theorem [36 p. 35-36] implies

(−1)ndeg⁡(h)=deg⁡(−h)=χ(A)≠0,(-1)^n\deg(h)=\deg(-h)=\chi(A)\ne0,

and thus we know that deg⁡(g)=deg⁡(h)≠0\deg(g) = \deg(h) \ne0. Hence, gg must be surjective over ∂Mc\partial M_c, where our choice of c>0c > 0 was arbitrary. Then, to conclude, given any open neighbourhood N′⊆NN' \subseteq N of AA, we can always find a sufficiently small c′>0c' > 0 such that Mc′⊆N′M_{c'} \subseteq N' (e.g., as XX is smooth, endow XX with a Riemannian metric and appeal to standard results regarding metric neighbourhoods of compact sets [39 §27]).

As discussed in §4(b), χ(A)≠0\chi(A) \ne0 is not an artifact of the proof, this condition cannot be removed. We also emphasize that Theorem 3.3 is not equivalent to showing that the Gauss map from some smooth levelset is surjective, as pointed out in [30 Footnote 4], this is always the case. Instead, we focus on the degree. The crux is that surjectivity is not homotopy invariant, but the degree of a map is.

Now, given Theorem 3.3, we study the following version of the adversary condition.

(iii) The spherical adversary condition For any open neighbourhood N⊂XN \subset X of AA and any z∈Sn−1z \in S^{n-1}, there is a ux∈πu−1(N∖A)u_x \in\pi_u^{-1}(N \setminus A) such that z=f(ux)/∥f(ux)∥2z = f(u_x)/\|f(u_x)\|_2, with ff understood as f:U→Rnf : \mathcal{U} \to\mathbb{R}^n.

Indeed, the spherical adversary condition applies to Examples 1.1-1.2 whereas the standard adversary condition does not.

To continue, we show that the spherical adversary condition complies with what is known regarding the nonholonomic integrator, e.g., see [6, 46, 42].

Example 3.4 (The nonholonomic integrator). Consider the classical nonholonomic integrator

x˙1=u1x˙2=u2x˙3=x1u2−x2u1,(8)\begin{aligned} \dot{x}_1 &= u_1 \\ \dot{x}_2 &= u_2 \\ \dot{x}_3 &= x_1u_2 - x_2u_1, \tag*{(8)} \end{aligned}

compactly written as x˙=f(x,u):=∑igi(x)ui\dot{x} = f(x,u) := \sum_i g_i(x)u_i. Interestingly, this system is controllable (e.g., see (8) is driftless and check that rank⁡(g1,g2,[g1,g2])=3\operatorname{rank}(g_1,g_2,[g_1,g_2]) = 3 [40 Ch. 3]). Now consider stabilizing the origin, it readily follows that Brockett’s condition fails as there is no pair (xε,uε)(x_\varepsilon,u_\varepsilon) such that f(xε,uε)=(0,0,ε)f(x_\varepsilon,u_\varepsilon) = (0,0,\varepsilon), for some ε>0\varepsilon> 0. Hence, there is no differentiable (or even continuous) feedback such that the origin is stabilized. Regarding Coron’s condition, one may check that f∗Hn−1(Σ)≅f∗H2(S1)≇Zf_*H_{n-1}(\Sigma) \cong f_*H_2(S^1) \not\cong\mathbb{Z}. Clearly, there is also no pair (xs,us)(x_s,u_s) such that f(xs,us)/∥f(xs,us)∥2=(0,0,1)f(x_s,u_s)/\|f(x_s,u_s)\|_2 = (0,0,1). Thus, the spherical adversary condition fails to hold, that is, introducing mild discontinuities does not resolve the question of stabilization. However, we may stabilize a set AA, in finite-time, with χ(A)=0\chi(A) = 0, as shown in §4(b).

The spherical adversary condition (iii) is all about the “directions” a control system can generate around some set AA, and the intention of proposing precisely this condition is that it is directly of use (e.g., there is no need to first perform any form of regularization via Filippov operators and the like).

A sidenote here is that in some cases, an application of the Filippov operator fills in precisely the gap in the tangent space that would otherwise prevent the standard adversary condition from being true (e.g., going from a sphere- to disk bundle).

Remark 3.5 (Comparison to [42]). Ryan studies a setting akin to Brockett (i.e., for x˙=f(x,u)\dot{x} = f(x,u), with ff continuous, find a feedback to stabilize 0), but Ryan allows for mildly discontinuous feedback. Here, the additional assumption is that for KK convex, f(x,K)f(x,K) is convex. Moreover, feedback controllers are upper semicontinuous maps x↦x \mapsto

μ(x)\mu(x), such that μ(x)\mu(x) is nonempty, compact and convex, with 0∈μ(0)0 \in\mu(0). A Brockett-like necessary condition can be recovered [42 Cor. 1]. This construction is reminiscent of applying the Filippov operator to a discontinuous vector field, however, it is more stringent (i.e., compare applying the Filippov operator to either a discontinuous vector field or only to the controller, see [9]). A simple example that cannot be handled using the machinery from [42] would be x˙=sin⁡(u)\dot{x} = \sin(u) with u∈S0u \in S^0. Nonetheless, x˙=sin⁡(−sgn⁡(x))\dot{x} = \sin(-\operatorname{sgn}(x)) renders the origin asymptotically stable, in the sense of Filippov.

Next, before comparing the spherical adversary condition (iii) to the homology condition (ii), we ensure that the homology condition is still meaningful (i.e., still necessary) in our setting, that is, in the context of § 1.3. This readily follows from Kvalheim’s homotopy theorem [28 Thm. 2], [27 Thm. 1] and the structure as exploited in the first part of the proof of Theorem 3.3. Note, we do not rely on Assumption 3.2 or χ(A)≠0\chi(A) \ne0.

Corollary 3.6 (The homology condition under discontinuous feedback). Let (A,X)(A, X) satisfy Assumption 3.1 and consider a control system C=(X,U,f,A,F)\mathcal{C} = (X, \mathcal{U}, f, A, \mathcal{F}) in the sense of §1.3. Let N⊆XN \subseteq X be an open neighbourhood of AA and suppose there is a section σ:N→U\sigma: N \to\mathcal{U} such that AA is asymptotically stable under f∘σf \circ\sigma, in the sense of Filippov, then the homology condition holds.

Proof. Let AA be asymptotically stable on some sufficiently small neighbourhood N′⊆NN' \subseteq N of AA. First, one constructs a (straight-line) homotopy from f∘σf \circ\sigma to −∇V-\nabla V, without introducing any zeros, on N′∖AN' \setminus A. Then, by Kvalheim’s homotopy theorem [27 Thm. 1], there is a homotopy from −∇V-\nabla V to GG, on W∖AW \setminus A, where W⊆N′W \subseteq N' is a possibly smaller neighbourhood of AA, again, without introducing any zeros. Now we may conclude as in the proof of [28 Thm. 3], that is

G∗H∙(W∖A)=−∇V∗H∙(W∖A)=(f∘σ)∗H∙(W∖A)⊆f∗H∙(ΣW∖A),G_*H_\bullet(W \setminus A) = -\nabla V_*H_\bullet(W \setminus A) = (f \circ\sigma)_*H_\bullet(W \setminus A) \subseteq f_*H_\bullet(\Sigma_{W\setminus A}),

in H∙(TW∖A∖0TX)H_\bullet(T_{W\setminus A} \setminus0_{TX}). Note, here we do exploit that both ff and σ\sigma are continuous on ΣW∖A\Sigma_{W\setminus A} and W∖AW \setminus A, respectively. For the spherical adversary condition this was not needed (only continuity of f∘σf \circ\sigma on W∖AW \setminus A was used, and only indirectly to appeal to smooth Lyapunov functions and the topological degree).

Theorem 3.7 ((ii) ⇒\Rightarrow (iii): the homology condition is always stronger than the spherical adversary condition). Let the pair (A,X)(A, X) satisfy Assumptions 3.1–3.2 and consider a control system C=(X,U,f,A,F)\mathcal{C} = (X, \mathcal{U}, f, A, \mathcal{F}) in the sense of §1.3 with χ(A)≠0\chi(A) \ne0. Then, if the homology condition holds, so does the spherical adversary condition.

Proof. As in [28], we show that if the spherical adversary fails to hold, so does the homology condition. Note, we use again the identification TX≅X×RnTX \cong X \times\mathbb{R}^n, e.g., H∙(W×Rn)≅H∙(W)H_\bullet(W \times\mathbb{R}^n) \cong H_\bullet(W).

With the above in mind, consider GG as in the homology condition. By the properties of GG, we know there is a smooth Lyapunov function VV (with respect to AA being asymptotically stable under GG) [14]. This means that we can, for instance, simply pick G:=−∇V:Mc⊆X→TMcXG := -\nabla V : M_c \subseteq X \to T_{M_c}X. Since ∂Mc\partial M_c is a compact, orientiable, embedded submanifold, we have Hn−1(∂Mc)≅ZkH_{n-1}(\partial M_c) \cong\mathbb{Z}^k with k∈N>0k \in\mathbb{N}_{>0} being the finite number of disjoint components (e.g., consider H0(∂Mc)H^0(\partial M_c) and Poincaré duality, noting that dim⁡(∂Mc)=n−1\dim(\partial M_c) = n - 1).

Then, for n≥2n \ge2 we have

−∇V∗Hn−1(Mc∖A)≅−∇V∗Hn−1(∂Mc)=−(∇V/∥∇V∥2)∗Hn−1(∂Mc)≅deg⁡(h)⋅Z,\begin{aligned} -\nabla V_*H_{n-1}(M_c \setminus A) &\cong-\nabla V_*H_{n-1}(\partial M_c) \\ &= -(\nabla V/\lVert\nabla V\rVert_2)_*H_{n-1}(\partial M_c) \cong\deg(h)\cdot\mathbb{Z}, \end{aligned}

where ∂Mc∋x↦h(x):=−(∇V(x)/∥∇V(x)∥2)∈Sn−1\partial M_c \ni x \mapsto h(x) := -(\nabla V(x)/\|\nabla V(x)\|_2) \in S^{n-1} and deg⁡(h)=∑i=1kdeg⁡(hi)\deg(h) = \sum_{i=1}^{k} \deg(h_i), for hih_i acting on the ii-th component of ∂Mc\partial M_c. Since χ(A)≠0\chi(A) \ne0, we know from the Poincaré-Hopf theorem [36 p. 35-36] that deg⁡(h)≠0\deg(h) \ne0.

For n=1n = 1, one has H0(S0)≅Z2H_0(S^0) \cong\mathbb{Z}^2 and hence the degree construction does not readily extend. However, since −(∇V/∥∇V∥2)-(\nabla V/\|\nabla V\|_2) is surjective, one can conclude that −∇V∗Hn−1(Mc∖A)≅Z2-\nabla V_*H_{n-1}(M_c \setminus A) \cong\mathbb{Z}^2.

Now, as the spherical adversary condition fails to hold, there is a z∈Sn−1z \in S^{n-1} such that z≠f(ux)/∥f(ux)∥2z \ne f(u_x)/\|f(u_x)\|_2 for all ux∈πu−1(N∖A)u_x \in\pi_u^{-1}(N \setminus A), for NN being some sufficiently small open neighbourhood of AA.

In what follows we assume that c>0c > 0 is selected so that Mc⊆NM_c \subseteq N. This is always possible by compactness of AA (and XX admitting a metric).

Now consider the second part of the homology condition. It follows that

f∗Hn−1(ΣMc∖A)=(f/∥f∥2)∗Hn−1(ΣMc∖A),f_*H_{n-1}(\Sigma_{M_c\setminus A}) = (f/\|f\|_2)_*H_{n-1}(\Sigma_{M_c\setminus A}),

where the normalization is well-defined as f−1(0)∩ΣMc∖A=∅f^{-1}(0) \cap\Sigma_{M_c\setminus A} = \varnothing. Then, as the normalized map to the sphere is not surjective, consider the factorization

(f/∥f∥2):ΣMc∖A→f~Sn−1∖{∗}↪ιSn−1,(f/\|f\|_2) : \Sigma_{M_c\setminus A} \xrightarrow{\widetilde{f}} S^{n-1} \setminus\{*\} \xhookrightarrow{\iota} S^{n-1},

i.e., f:=ι∘f~f := \iota\circ\widetilde{f}, and note that Sn−1∖{∗}≅{∗}S^{n-1} \setminus\{*\} \cong\{*\}. However, this means that (f/∥f∥2)∗Hn−1(ΣMc∖A)(f/\|f\|_2)_*H_{n-1}(\Sigma_{M_c\setminus A}) factors through Hn−1({∗})≅0H_{n-1}(\{*\}) \cong0 for n≥2n \ge2 and through H0({∗})≅ZH_0(\{*\}) \cong\mathbb{Z} for $n = 1. Hence, the homology condition fails.

We also observe that Theorem 3.7, [28 Thm. 4] and the original result by Coron [11 (B)], show that only the top homology is enough to imply the adversary condition, that is, for dim⁡(X)=n\dim(X) = n, one may consider the “top homology condition”

G∗Hn−1(N∖A)⊆f∗Hn−1(ΣN∖A)in Hn−1(TN∖AX∖0TX),(9)G_*H_{n-1}(N \setminus A) \subseteq f_*H_{n-1}(\Sigma_{N\setminus A}) \quad\text{in } H_{n-1}(TN\setminus AX \setminus0_{TX}), \tag*{(9)}

for some sufficiently small neighbourhood NN of AA.

Generalizations

(iiib) The sphere bundle adversary condition For any sufficiently small neighbourhood N⊆XN \subseteq X of AA there is local continuous section σ:N∖A→U\sigma: N \setminus A \to\mathcal{U} such that

f∘σ∥f∘σ∥g:N∖A→S(TN∖AX)(10)\frac{f \circ\sigma}{\|f \circ\sigma\|_g} : N \setminus A \to S(T_{N\setminus A}X) \tag*{(10)}

is well-defined, but cannot be extended to all of NN.

Note, this is not precisely a generalization, in Theorem 3.3 we can exploit the underlying Euclidean structure and go one step further by exploiting the topological degree7. What is more, the sense of “direction” is lost and the explicit dependency on a continuous section σ\sigma makes (10) impractical. This is similar to employing the index condition by Krasnosel’skiĭ directly8. Nevertheless, as the spherical adversary condition (iii), one may use the sphere bundle adversary condition (iiib) to show, for example, that there is no S1\mathbb{S}^{1}-valued feedback that stabilizes the origin in (8).

Arguably more practical would be a homological condition that does not rely on searching for sections. As alluded to before, following a path that led to adversarial conditions will provide us with a condition on top homology. As only the top degree appears, this shows again the advantage of the homology condition. To be more explicit, as in the proof of Theorem 3.3, given a (local) section σ:N→U\sigma:N\to U, which may be discontinuous on AA, pick c>0c>0 and set V−1([0,c))=:Mc⊆NV^{-1}([0,c))=:M_c\subseteq N. Since −∇V:∂Mc→T∂McX-\nabla V:\partial M_c\to T_{\partial M_c}X is a vector field, we have that πx∘−∇V=id⁡∂Mc\pi_x\circ-\nabla V=\operatorname{id}_{\partial M_c}. Hence,

(π∘−∇V)∗Hn−1(∂Mc)=Hn−1(∂Mc)≅Zk,(\pi\circ-\nabla V)_*H_{n-1}(\partial M_c)=H_{n-1}(\partial M_c)\cong\mathbb{Z}^{k},

for some k>0k>0. Thereby, −∇V∗Hn−1(∂Mc)≅−∇V∗Hn−1(Mc∖A)-\nabla V_*H_{n-1}(\partial M_c)\cong-\nabla V_*H_{n-1}(M_c\setminus A) cannot be trivial. Now, we know that −∇V-\nabla V is homotopic to f∘σf\circ\sigma on Mc∖AM_c\setminus A, without introducing new zeros. Hence

−∇V∗Hn−1(N∖A)=(f∘σ)∗Hn−1(N∖A)⊆f∗Hn−1(ΣN∖A)-\nabla V_*H_{n-1}(N\setminus A)=(f\circ\sigma)_*H_{n-1}(N\setminus A)\subseteq f_*H_{n-1}(\Sigma_{N\setminus A})

cannot be trivial, as a subgroup of Hn−1(TN∖AX∖0TX)H_{n-1}(T_{N\setminus A}X\setminus0_{TX}). Combining the above with Kvalheim’s homotopy theorem [27 Thm. 1], we get that

0≇G∗Hn−1(N∖A)⊆f∗Hn−1(ΣN∖A)inHn−1(TN∖AX∖0TX).0\not\cong G_*H_{n-1}(N\setminus A)\subseteq f_*H_{n-1}(\Sigma_{N\setminus A})\quad\text{in}\quad H_{n-1}(T_{N\setminus A}X\setminus0_{TX}).

Note, deriving this condition on top homology, relies only on the existence of a section σ\sigma and the non-trivial top homology of N∖AN\setminus A. Evidently, under mild assumptions, such a condition is implied by the homology condition (ii). Note that the above does not rely on χ(A)≠0\chi(A)\ne0. Hence, when it comes to generality, Corollary 3.6 provides us with the most general necessary condition. However, as adversarial conditions are typically easier to employ, when it comes to applicability, one might hope to improve upon (10).

Remark 3.8 (Connection to the work by Mansouri). In [32, 33], Mansouri provides necessary conditions similar to a top homology condition. For X=RnX=\mathbb{R}^{n}, AA some mm-dimensional compact embedded submanifold of XX and f:Ω⊆Rn×Rm→Rnf:\Omega\subseteq\mathbb{R}^{n}\times\mathbb{R}^{m}\to\mathbb{R}^{n} a C1C^{1} map representing the control system, i.e., x˙=f(x,u)\dot{x}=f(x,u), Mansouri studies necessary condition for the existence of a C1C^{1} feedback such that AA is stabilized. Comments on continuous feedback are provided. Specifically, let Σf,ϵ:={(x,u)∈Ω:f(x,u)≠0, d(x,A)<ϵ}\Sigma_{f,\epsilon}:=\{(x,u)\in\Omega:f(x,u)\ne0,\ d(x,A)<\epsilon\}, then:

(a) when m=n−1m=n-1 (and n≥2n\ge2), Mansouri shows that χ(MA)⋅Z⊆f∗Hn−1(Σf,ϵ)\chi(M_A)\cdot\mathbb{Z}\subseteq f_*H_{n-1}(\Sigma_{f,\epsilon}), for all ϵ>0\epsilon>0, where MAM_A is the manifold enclosed by AA; on the other hand

(b) when m≤n−2m\le n-2, χ(A)⋅Z⊆f∗Hn−1(Σf,ϵ)\chi(A)\cdot\mathbb{Z}\subseteq f_*H_{n-1}(\Sigma_{f,\epsilon}), for all ϵ>0\epsilon>0.

Compared to this work, we observe a difference in the state space, regularity of the attractor and the regularity of the vector fields. Otherwise, one should observe that if one constrains the control systems, sharper necessary conditions can be derived, e.g., we can speak of explicit subgroups instead of only stating that they are non-trivial. Also note the difference between Σf,ϵ\Sigma_{f,\epsilon} and ΣN∖A\Sigma_{N\setminus A}. Comparing to (ii) and with the above in mind, we believe it is beneficial to consider all degrees.

Discussion and conclusion

(a) (On sufficiency): Clearly, these necessary conditions are not sufficient. As the index condition by Krasnosel’skiĭ and Zabreiko [22 p. 82], all conditions above fail to capture that x˙=x\dot{x}=x (sometimes denoted as ∂x\partial x), with x∈R2nx\in\mathbb{R}^{2n}, cannot be stabilized (i.e., cannot render 0∈R2n0\in\mathbb{R}^{2n} asymptotically stable).

(b) (On χ(A)≠0\chi(A)\ne0): The standard adversarial condition (i) cannot be extended to the case of χ(A)=0\chi(A)=0. The same is true for the spherical adversarial condition, as proposed in this note (iii). Nevertheless, a top-homological version of (iii) can be applied, although, trivially. To emphasize the importance of filling this gap, we elaborate on a recurring example by Mansouri, see [32, 33]. Consider the control system

x˙1=u1x˙2=u2x˙3=(x2u1−x1u2)ex32,(11)\begin{aligned} \dot{x}_1 &= u_1\\ \dot{x}_2 &= u_2\\ \dot{x}_3 &= (x_2u_1-x_1u_2)e^{x_3^2}, \tag*{(11)} \end{aligned}

with (x,u)∈R3×R2(x,u)\in\mathbb{R}^3\times\mathbb{R}^2 (i.e., this control system can be identified with a vector bundle). Note, the system is a minor modification of the one presented in [32]. Now, replacing uu by the feedback x↦(μ1(x),μ2(x))x\mapsto(\mu_1(x),\mu_2(x)) defined through

μ1(x):=−x2x3−x1(x12+x22−1)μ2(x):=x1x3−x2(x12+x22−1)(12)\begin{aligned} \mu_1(x)&:=-x_2x_3-x_1(x_1^2+x_2^2-1)\\ \mu_2(x)&:=x_1x_3-x_2(x_1^2+x_2^2-1) \tag*{(12)} \end{aligned}

results in a closed-loop system with A={(x1,x2,0)∈R3:x12+x22=1}A=\{(x_1,x_2,0)\in\mathbb{R}^3:x_1^2+x_2^2=1\} being locally asymptotically stable. See that (A,X)(A,X) satisfies Assumption 3.1 and χ(A)=0\chi(A)=0. One may check this using the Lyapunov function x↦V(x):=(x12+x22−1)2+x32x\mapsto V(x):=(x_1^2+x_2^2-1)^2+x_3^2. In fact, this Lyapunov function asserts almost global asymptotic stability (i.e., the system is asymptotically stable on R3∖{(0,0,z)∈R3:z∈R}).\mathbb{R}^3\setminus\{(0,0,z)\in\mathbb{R}^3:z\in\mathbb{R}\}). Control system (4.1) is used in conjunction with feedback (4.2) to show that the adversary condition cannot be readily extended to the case of χ(A)=0\chi(A)=0 (i.e., the example above admits a stabilizing feedback, but the condition fails). We argue now that the same is true for the spherical adversary condition.

Indeed, there is no (x,u)(x,u) such that f(x,u)=(0,0,ε)f(x,u)=(0,0,\varepsilon), for any ε>0\varepsilon>0, so the spherical adversary condition fails in general, for any compact attractor AA such that χ(A)≠0\chi(A)\ne0. Nonetheless, we can find a stabilizing feedback, where the input is (S1↪R2)(S^1\hookrightarrow\mathbb{R}^2)-valued and as such, the control system cannot be identified with a vector bundle. In particular, we will project (4.2) onto S1S^1. Here, a simple computation shows that ∥μ(x)∥22=(x12+x22)V(x)\|\mu(x)\|_2^2=(x_1^2+x_2^2)V(x), and this expression is only zero on A∪{(0,0,z)∈R3:z∈R}A\cup\{(0,0,z)\in\mathbb{R}^3:z\in\mathbb{R}\}. Using (4.1) and the Lyapunov function VV from above, one observes that normalizing μ\mu, preserves asymptotic stability (now understood in the sense of Filippov). Moreover, under the normalized feedback

⟨∇V(x),x˙⟩=−4∥μ(x)∥22(x12+x22)(x12+x22−1)−2∥μ(x)∥22(x12+x22)x32ex32≤−4∥μ(x)∥22(x12+x22)(x12+x22−1)−2∥μ(x)∥22(x12+x22)x32,\begin{aligned} \langle\nabla V(x),\dot{x}\rangle &=-\frac{4}{\|\mu(x)\|_2^2}(x_1^2+x_2^2)(x_1^2+x_2^2-1)-\frac{2}{\|\mu(x)\|_2^2}(x_1^2+x_2^2)x_3^2e^{x_3^2}\\ &\le-\frac{4}{\|\mu(x)\|_2^2}(x_1^2+x_2^2)(x_1^2+x_2^2-1)-\frac{2}{\|\mu(x)\|_2^2}(x_1^2+x_2^2)x_3^2, \end{aligned}

as inf⁡z∈R{ez2}=1\inf_{z\in\mathbb{R}}\{e^{z^2}\}=1. Then, recall that for any c>0c>0, Mc=V−1([0,c])M_c=V^{-1}([0,c]) is compact and forward invariant. Now, pick any c∈(0,1)c\in(0,1) and see that on McM_c we have (x2+y2−1)2<1(x^2+y^2-1)^2<1 and thus (x2+y2−1)2<(x2+y2−1)(x^2+y^2-1)^2<(x^2+y^2-1). What is more, for such a c>0c>0, see that min⁡x∈Mc{x12+x22}=1−c\min_{x\in M_c}\{x_1^2+x_2^2\}=1-\sqrt{c}. Hence, we can write

⟨∇V(x),x˙⟩∣Mc≤−4∥μ(x)∥22(x12+x22)(x12+x22−1)2−2∥μ(x)∥22(x12+x22)x32≤−1∥μ(x)∥22(x12+x22)V(x)=−(x12+x22)V(x)≤−1−cV(x)=:−αV(x),\begin{aligned} \left\langle\nabla V(x),\dot{x}\right\rangle\big|_{M_c} &\leq-\frac{4}{\|\mu(x)\|_2^2}(x_1^2+x_2^2)(x_1^2+x_2^2-1)^2-\frac{2}{\|\mu(x)\|_2^2}(x_1^2+x_2^2)x_3^2\\ &\leq-\frac{1}{\|\mu(x)\|_2^2}(x_1^2+x_2^2)V(x)\\ &=-\sqrt{(x_1^2+x_2^2)}\sqrt{V(x)}\\ &\leq-\sqrt{1-\sqrt{c}}\sqrt{V(x)}=:-\alpha\sqrt{V(x)}, \end{aligned}

for α∈(0,1)\alpha\in(0,1). This means, however, that AA is locally finite-time asymptotically stable (e.g., on McM_c) under (11) and the feedback x↦μ(x)/∥μ(x)∥2x\mapsto\mu(x)/\|\mu(x)\|_2 (with μ(x):=0\mu(x):=0 on AA), e.g., see [2 Thm. 4.2]. This, despite not satisfying the spherical adversary condition (iii).

A top-homological generalization, and in particular the homology condition (ii), do not rely on χ(A)=0\chi(A)=0 and must be true. However, as in Remark 3.8, several conditions become trivial. We leave the study of non-trivial necessary conditions for circular attractor for future work, see also [27 p. 538].

(c) (On Control Barrier Functions (CBFs)): Consider the control system x˙=x+u\dot{x}=x+u, with x∈R2x\in\mathbb{R}^2 and u∈S1↪R2u\in S^1\hookrightarrow\mathbb{R}^2. In [13], we find Brockett-like necessary condition for the existence of a continuous controller such that some set SS is safe (forward invariant). They consider S:={x∈R2:∥x∥22≤1}S:=\{x\in\mathbb{R}^2:\|x\|_2^2\leq1\} and show that there is no continuous feedback [13 Ex. 4], employing effectively the same arguments as in Example 1.1 ([28 Ex. 4]). However, one may select the controller κ:R2→S1\kappa:\mathbb{R}^2\to S^1, as defined through x↦κ(x):=−x/∥x∥22x\mapsto\kappa(x):=-x/\|x\|_2^2 for x∈R2∖{0}x\in\mathbb{R}^2\setminus\{0\} while κ(x):=0\kappa(x):=0 otherwise. Indeed, as the closed-loop system becomes

x˙={(∥x∥22−1∥x∥22)x,for x∈R2∖{0}0otherwise,\dot{x}= \begin{cases} \left(\dfrac{\|x\|_2^2-1}{\|x\|_2^2}\right)x, & \text{for }x\in\mathbb{R}^2\setminus\{0\}\\ 0 & \text{otherwise,} \end{cases}

we render SS safe (and 0∈S0\in S is asymptotically stable in the sense of Filippov).

(d) (On discontinuous feedback): We emphasize that the type of discontinuities we consider are mild (see §1.3) and are not intended to overcome topological obstructions. Indeed, typically, if there is no continuous feedback, then, there is no robust discontinuous feedback either [42, 9, 35]. For us, discontinuities emerge due to the possibility of working with fiber bundles with non-contractible fibers cf. Example 1.1-1.2 and the robustness enters due to application of the Filippov operator. Handling discontinuities outside of AA is left for future work, see for instance [9 Ex. 3] and [37].

In this work, the study of certain control systems is enabled by the existence of a smooth Lyapunov function. In a line of work pioneered by Lyapunov, Morse, Conley and many others, this note may be interpreted as another example where classification of dynamical (control) systems through the existence of these smooth functions is fruitful. One should compare this with classification of systems based on the regularity of the vector field itself.

Considering more general control systems, e.g., allowing more general discontinuities and fiber bundles, is left for future work.

Acknowledgments

The first author would like to thank Matthew D. Kvalheim for helpful discussions.

References

References

  1. [1]M.-A. Belabbas and J. Ko, Dynamic global feedback stabilization: why do the twist?, in Geometry, Topology and Control System Design: Proceedings of a Banff International Research Station Workshop, M. A. Belabbas, ed., American Institute of Mathematical Sciences, Springfield, 2025, pp. 47–60.arxiv.org/abs/2404.18380
  2. [2]S. P. Bhat and D. S. Bernstein, Finite-time stability of continuous autonomous systems, SIAM J. Contr. Optim., 38 (2000), pp. 751–766.DOI
  3. [3]N. P. Bhatia and G. P. Szegö, Stability theory of dynamical systems, Springer, Berlin, 1970.DOI
  4. [4]N. Bobylev and M. Krasnosel’skiĭ, Deformation of a system into an asymptotically stable system, Autom. Remote Control, 35 (1974), pp. 1041–1044.
  5. [5]G. E. Bredon, Topology and geometry, Springer, New York, 1993.
  6. [6]R. W. Brockett, Asymptotic stability and feedback stabilization, in Differential Geometric Control Theory, Boston, 1983, Birkhäuser, pp. 181–191.
  7. [7]C. I. Byrnes, On Brockett’s necessary condition for stabilizability and the topology of Liapunov functions on ℝⁿ, Commun. Inf. Syst., 8 (2008), pp. 333–352.DOI
  8. [8]C. Cai, A. R. Teel, and R. Goebel, Smooth Lyapunov functions for hybrid systems part II: (pre) asymptotically stable compact sets, IEEE T. Automat. Control, 53 (2008), pp. 734–748.DOI
  9. [9]F. Ceragioli, Some remarks on stabilization by means of discontinuous feedbacks, Syst. Control Lett., 45 (2002), pp. 271–281.DOI
  10. [10]F. H. Clarke, Y. S. Ledyaev, and R. J. Stern, Asymptotic stability and smooth Lyapunov functions, J. Differ. Equ., 149 (1998), pp. 69–114.DOI
  11. [11]J.-M. Coron, A necessary condition for feedback stabilization, Syst. Control Lett., 14 (1990), pp. 227–232.DOI
  12. [12]J.-M. Coron, Control and Nonlinearity, American Mathematical Society, Providence, 2007.
  13. [13]M. De Sa and A. D. Ames, Topological obstructions to the existence of control barrier functions, arXiv e-prints, (2026), https://arxiv.org/abs/2603.18422.
  14. [14]A. Fathi and P. Pageault, Smoothing Lyapunov functions, T. Am. Math. Soc., 371 (2019), pp. 1677–1700.
  15. [15]A. F. Filippov, Differential Equations with Discontinuous Righthand Sides, Springer Science & Business Media, Dordrecht, 1988.
  16. [16]M. Gobbino, Topological properties of attractors for dynamical systems, Topology, 40 (2001), pp. 279–298.DOI
  17. [17]V. Guillemin and A. Pollack, Differential Topology, American Mathematical Society, Providence, 2010.
  18. [18]A. Hatcher, Algebraic Topology, Cambridge University Press, Cambridge, 2002.
  19. [19]A. Hatcher, Vector bundles, 2003, http://www.math.cornell.edu/~hatcher.
  20. [20]W. Jongeneel, On topological properties of compact attractors on Hausdorff spaces, in 2024 European Control Conference (ECC), 2024, pp. 186–191.DOI
  21. [21]W. Jongeneel, On topological properties of closed attractors, (2025), https://arxiv.org/abs/2511.10429.
  22. [22]W. Jongeneel and E. Moulay, Topological Obstructions to Stability and Stabilization: History, Recent Advances and Open Problems, Springer Nature, Cham, 2023.DOI
  23. [23]L. Kapitanski and I. Rodnianski, Shape and Morse theory of attractors, Comm. Pure Appl. Math., 53 (2000), pp. 218–242.doi.org/10.1002/(sici)1097-0312(200002)53:2<218::aid-cpa2>3.0.co;2-w
  24. [24]A. Krasnosel’skiĭ and P. P. Zabreĭko, Geometrical methods of nonlinear analysis, Springer, Berlin, 1984.
  25. [25]M. A. Krasnosel’skiĭ, The operator of translation along the trajectories of differential equations, American Mathematical Society, Providence, 1968.
  26. [26]M. D. Kvalheim, A generalization of the Hopf degree theorem, Proc. Am. Math. Soc., 151 (2023), pp. 453–454.
  27. [27]M. D. Kvalheim, Obstructions to asymptotic stabilization, SIAM J. Contr. Optim., 61 (2023), pp. 536–542.
  28. [28]M. D. Kvalheim, Relationships between necessary conditions for feedback stabilizability, in Geometry, Topology and Control System Design: Proceedings of a Banff International Research Station Workshop, M. A. Belabbas, ed., American Institute of Mathematical Sciences, Springfield, 2025, pp. 167–179.
  29. [29]M. D. Kvalheim, Stabilizability of first-order dynamics in second-order systems, (2026), https://arxiv.org/abs/2604.10757.
  30. [30]M. D. Kvalheim and E. D. Koditschek, Necessary conditions for feedback stabilization and safety, J. Geom. Mech., 14 (2022), pp. 659–693.arxiv.org/abs/2106.00215
  31. [31]M. Lee, Introduction to Smooth Manifolds, Springer, New York, 2012.
  32. [32]A.-R. Mansouri, Local asymptotic feedback stabilization to a submanifold: Topological conditions, Syst. Control Lett., 56 (2007), pp. 525–528.DOI
  33. [33]A.-R. Mansouri, Topological obstructions to submanifold stabilization, IEEE T. Automat. Contr., 55 (2010), pp. 1701–1703.DOI
  34. [34]A.-R. Mansouri, On Coron’s necessary condition for local asymptotic stabilization, in Geometry, Topology and Control System Design: Proceedings of a Banff International Research Station Workshop, M. A. Belabbas, ed., American Institute of Mathematical Sciences, Springfield, 2025, pp. 337–340.
  35. [35]C. G. Mayhew and A. R. Teel, On the topological structure of attraction basins for differential inclusions, Syst. Control Lett., 60 (2011), pp. 1045–1050.DOI
  36. [36]J. Milnor, Topology from the Differential Viewpoint, Princeton University Press, Princeton, 1965.
  37. [37]R. Montgomery and R. Sanfelice, Hybrid control, Morse theory and Ivan Kupka, in Ivan Kupka Legacy: A Tour Through Controlled Dynamics, B. Bonnard, M. Chyba, D. Holcman, and E. Trélat, eds., American Institute of Mathematical Sciences, Springfield, 2024, pp. 135–153.
  38. [38]E. Moulay and Q. Hui, Conley index condition for asymptotic stability, Nonlinear Anal.-Theor., 74 (2011), pp. 4503–4510.DOI
  39. [39]J. Munkres, Topology, Pearson, Harlow, 2014.
  40. [40]H. Nijmeijer and A. van der Schaft, Nonlinear Dynamical Control Systems, Springer, New York, 1990.DOI
  41. [41]R. Orsi, L. Praly, and I. Mareels, Necessary conditions for stability and attractivity of continuous systems, Int. J. Control, 76 (2003), pp. 1070–1077.DOI
  42. [42]E. P. Ryan, On Brockett’s condition for smooth stabilizability and its necessity in a context of nonsmooth feedback, SIAM J. Contr. Optim., 32 (1994), pp. 1597–1604.DOI
  43. [43]E. Sontag and H. Sussmann, Remarks on continuous feedback, in Proc. IEEE Conf. Decis. Control, 1980, pp. 916–921.DOI
  44. [44]F. W. Wilson Jr, The structure of the level surfaces of a Lyapunov function, J. Differ. Equ., 3 (1967), pp. 323–329.DOI
  45. [45]W. M. Wonham, Linear multivariable control, Springer, New York, 1979.
  46. [46]J. Zabczyk, Some comments on stabilizability, Appl. Math. Opt., 19 (1989), pp. 1–9.DOI

Paper details

vlnc:W_xub5zx6a
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

Not recorded for this paper

View statement
Looking for related papers…
Contents