arXiv:1311.0078v2 [math.OC] 24 Jun 2014 [email protected] [email protected] Taringoo), sscnb eue oteaayi faeulbi located equilibria on a defined of systems dynamical analysis many However, the origin. to the reduced at be stability Consequently, can origin. availabysis the transformations to equilibria coordinate shift to simplification to due significant analysis, to the leads proper in space spaces vector Euclidean attendant of the ties spaces. of fo- Euclidean application has on the evolving There, theory systems dynamical stability on of cused development the Traditionally, exampl for see literature, 23]. 18, the [16, in Lyapunov of analyzed sense extensively the been in th Stability characterizing equilibria. and of t analyzing stability is for theory tool mathematical stability Lyapunov core differen 34]. 33, dynamica or [16, equations see of differential inclusions, trajectories of solutions of as Th stability systems theory. the control in addresses topic theory important an is theory Stability many in found be [3–5]. can see systems settings, such mechanical of with Examples ge tools. 33], differential 3–5, ric of [1, application on the example requires evolve analysis for their naturally see that manifolds, dynamics Riemannian include systems Many Introduction 1 re with in functions systems dynamical Lyapunov for of theorems nonline converse properties for theorems characterizes Lyapunov and converse proposes paper This Abstract icvr rjc DP120101144. Project Discovery ⋆ e words: Key dy equili perturbed of of neighborhoods stability the normal obtain called we so functions, the Lyapunov to analysis our rpitsbitdt Automatica to submitted Preprint hswr a upre yteAsrla eerhCouncil Research addresses: Email Australian the by supported was work This oa hrceiaino ypnvfntoso Riemanni on functions Lyapunov of characterization local A [email protected] yaia ytm,Reana aiod,Goei curves Geodesic manifolds, Riemannian systems, Dynamical a lcrcladEetoi niern eatet h Uni The Department, Engineering Electronic and Electrical aznTaringoo Farzin [email protected] Yn Tan). (Ying Daa Neˇsi´c),(Dragan PtrM Dower), M. (Peter R a ee .Dower M. Peter , n odnmclsseseovn nReana aiod.Thi manifolds. Riemannian on evolving systems dynamical to (Farzin manifolds omet- aia ytm nReana manifolds. Riemannian on systems namical anal- pc oteReana itnefnto.W xedclassi extend We function. distance Riemannian the to spect has ruso imninmnfls yepoigtedrvdpr derived the employing By manifolds. Riemannian on briums tial rdnmclssesdfie nsot once Riemannian connected smooth on defined systems dynamical ar he le is e e s - l imnindsac ucin eepo h oinof stability on systems the notion dynamical apply for the and the results manifolds employ on Riemannian We on based geodesics function. manifolds dy- distance Riemannian of Riemannian Lyapunov stability on Lyapunov such systems define of we namical properties end, this local To some functions. man- prove Riemannian theo- on and Lyapunov evolving ifolds converse systems several dynamical for present rems we paper, this In Rieman- of [19]. notion see the function, employing distance by nian spaces consid- metric be can as manifolds spaces ered Riemannian metric general, compact In derived. on Lyapunois systems complete dynamical of for existence the functions [7,36], in particular, results In func- Recent Lyapunov [7,10,11,13,15,22,29,31,34,36,3 in of 26]. documented [24, are properties tions example and existence for the see concerning research, ne a of is manifolds area on evolving systems for analysis Stability inevit is theory stability the of framework traditional generalizatio the prop- a Consequently, space spaces. vector Euclidean the of erties possess necessarily not do manifolds h tblt hoypeetdi 1,1] ti hw that shown is It 16]. of [12, results in standard presented the theory of some stability equilibrium, invoke an the and of 32], space 19, to [14, tangent manifold the see on Riemannian system a dynamical on con- a equations to dynamical operator the lift vert a on evolving introduce We those to Eu- manifolds. extended Riemannian on are evolving [12,16,37] systems spaces dynamical clidean for results [2,5,8], see theory stability manifolds, the stability Riemannian on on of evolving version defined systems for a systems Using dynamical manifolds. for Riemannian functions Lyapunov of a . rgnNe Dragan , est fMlore itra 00 Australia 3010, Victoria, Melbourne, of versity ⋆ si ˇ c ´ a igTan Ying , R spromdb restricting by performed is s n ooti h existence the obtain to a 0Otbr2018 October 20 a Lyapunov cal an prisof operties manifolds able. of n 8]. w v in a normal neighborhood [19] of an equilibrium of a dy- Table 1 namical system, the constructed Lyapunov functions satisfy Symbols and Their Descriptions certain properties which can be used to analyze the stability Symbol Description and robustness of the underlying dynamical system. These results are extended and applied to study perturbed dynamic M Riemannian manifold systems. Geometric features of the normal neighborhoods, X(M) space of smooth time invariant such as existence of unique length minimizing geodesics and vector fields on M their local representations enable us to closely relate the sta- X R bility results obtained for dynamical systems in Rn to those (M × ) space of smooth time varying defined on Riemannian manifolds. vector fields on M
TxM tangent space at x ∈ M In terms of exposition, Section 2 presents some mathemati- ∗ cal preliminaries needed for the subsequent analysis. Section Tx M cotangent space at x ∈ M 3 presents the main results for the existence of Lyapunov T M tangent bundle of M functions for dynamical systems evolving on Riemannian T ∗M cotangent bundle of M manifolds. These results are employed in Section 4 to derive ∂ basis tangent vectors at x ∈ M the stability of perturbed dynamical systems on Riemannian ∂xi manifolds. The paper concludes with some closing remarks dxi basis cotangent vectors at x ∈ M in Section 5. f(x,t) time-varying vector fields on M
|| · ||g Riemannian norm 2 Preliminaries || · ||e Euclidean norm || · || induced norm In this section we provide the differential geometric material which is necessary for the analysis presented in the rest of g(·, ·) Riemannian metric on M the paper. We define some of the frequently used symbols d(·, ·) Riemannian distance on M of this paper in Table 1. Φf flow associated with f T F pushforward of F
Definition 1 Let M be a an n dimensional manifold. A co- TxF pushforward of F at x ordinate chart on M is (U, φ), where U is an open set in M R>0 (0, ∞) and φ is a homomorphism from U to φ(U) ⊂ Rn, see [21]. R≥0 [0, ∞) C∞(M) space of smooth functions on M 2.1 Riemannian manifolds ≃ isomorphism B(x,r) metric ball centered at x with radius r Definition 2 (see [21], Chapter 3) A Riemannianmanifold Br(0) Ball with radius r in tangent spaces (M,g) is a differentiable manifold M together with a Rie- mannian metric g, where g is defined for each x ∈ M via an inner product gx : TxM × TxM → R on the tan- can be connected via a path γ ∈ P(x, y), where gent space TxM (to M at x) such that the function defined by x 7→ gx(X(x), Y (x)) is smooth for any vector fields X, Y ∈ X(M). In addition, . γ piecewise smooth, P(x, y) = γ :[a,b] → M (2.1) (i) (M,g) is n dimensional if M is n dimensional; ( γ(a)= x, γ(b)= y )
(ii) (M,g) is connected if for any x, y ∈ M, there exists
a piecewise smooth curve that connects x to y. Theorem 1 ([19], P. 94) Suppose (M,g) is an n dimen- sional connected Riemannian manifold. Then, for any x, y ∈ M, there exists a piecewise smooth path γ ∈ P(x, y) that . Note that in the special case where M = Rn, the connects x to y. Riemannian metric g is defined everywhere by gx = n i,j=1 gij (x)dxi ⊗ dxj , where ⊗ is the tensor product on The existence of connecting paths (via Theorem 1) between ∗ ∗ Tx M × Tx M, see [21]. pairs of elementsof an n dimensional connected Riemannian P manifold (M,g) facilitates the definition of a corresponding As formalized in Definition 2, connected Riemannian man- Riemannian distance. In particular, the Riemannian distance ifolds possess the property that any pair of points x, y ∈ M d : M × M → R is defined by the infimal path length
2 between any two elements of M, with One may show, for a smooth vector field f, the integral flow Φf (s,t0, ·): M → M is a local diffeomorphism , . b see [21]. Here we assume that the vector field f is smooth d(x, y) = inf g (γ ˙ (t), γ˙ (t)) dt . (2.2) γ(t) and complete, i.e. Φf exists for all t ∈ (t0, ∞). γ∈P(x,y) a Z q . Note that in the special case where M = Rn, the Rieman- 2.3 Geodesic Curves nian distance (2.2) simplifies to d(x, y)= kx − yke. Geodesics are defined [14] as length minimizing curves on Using the definition of Riemannian distance d of (2.2), Riemannian manifolds which satisfy (M, d) defines a metric space as formalized by the follow- ing theorem. ∇γ˙ (t)γ˙ (t)=0, (2.9)
Theorem 2 ([19], P. 94) Any n dimensional connected Rie- where γ(·) is a geodesic curve on (M,g) and ∇ is the Levi- mannian manifold (M,g) defines a metric space (M, d) via Civita connection on M, see [19]. The solution of the Euler- the Riemannian distance d of (2.2). Furthermore, the in- Lagrange variational problem associated with the length duced topology of (M, d) is the same as the manifold topol- minimizing problem shows that all the geodesics on an n ogy of (M,g). dimensional Riemannian manifold (M,g) must satisfy the following system of ordinary differential equations: Next, the crucial pushforward operator is introduced. n i Definition 3 For a given smooth mapping F : M → N γ¨i(s)+ Γj,kγ˙j (s)γ ˙ k(s)=0, i =1, ..., n, (2.10) from manifold to manifold the pushforward is j,k=1 M N T F X defined as a generalization of the Jacobian of smooth maps between Euclidean spaces as follows: where n T F : TM → TN, (2.3) i 1 il ∂gjl Γj,k = g (gjl,k + gkl,j − gjk,l), gjl,k = , 2 ∂xk where l=1 X (2.11) T F : T M → T N, (2.4) x x F (x) where all the indexes i,j,k,l run from 1 up to n = dim(M) . and ij −1. Note that is the entity of the and [g ] = [gij ] gij (i, j) metric g. T F (X ) ◦ h = X (h ◦ F ), X ∈ T M,h ∈ C∞(N). x x x x x Definition 4 ([19], p. 72) The restricted exponential map is (2.5) defined by
expx : TxM → M, expx(v)= γv(1), v ∈ TxM, (2.12)
2.2 Dynamical systems on Riemannian manifolds where γv(1) is the unique maximal geodesic [19], P. 59, initiating from x with the velocity v up to one. This paper focuses on dynamical systems governed by dif- ferential equations on a connected n dimensional Rieman- Throughout, restricted exponential maps are referred to as nian manifold M. Locally these differential equations are exponential maps. An open ball of radius δ > 0 and cen- defined by (see [21]) tered at 0 ∈ T M in the tangent space at x is denoted by . x Bδ(0) = {v ∈ TxM | ||v||g < δ}. Similarly, the corre- X R x˙(t)= f(x(t),t), f ∈ (M × ), sponding closed ball is denoted by Bδ(0). Using the lo- x(0) = x0 ∈ M,t ∈ [t0,tf ]. (2.6) cal diffeomorphic property of exponential maps, the corre- sponding geodesic ball centered at x is dened as follows. The time dependent flow associated with a differentiable time dependent vector field f is a map Φf satisfying Lemma 1 ([19], Lemma 5.10) For any x ∈ M, there ex- ists a neighborhood Bδ(0) in TxM on which expx is a dif- Φf :[t0,tf ] × [t0,tf ] × M → M, feomorphism onto expx(Bδ(0)) ⊂ M. (s0,sf , x) 7→ Φf (sf ,s0, x) ∈ M, (2.7) Definition 5 ([19]) In a neighborhood of x ∈ M, where and expx is a local diffeomorphism (this neighborhood always exists by Lemma 1), a geodesic ball of radius δ > 0 is dΦf (s,s0, x) denoted by expx(Bδ(0)) ⊂ M. The corresponding closed = f(Φf (t,s0, x),t). (2.8) ds geodesic ball is denoted by exp (Bδ(0)). s=t x
3 Definition 6 For a vector space V , a star-shaped neighbor- hood of 0 ∈ V is any open set U such that if u ∈ U then αu ∈ U, α ∈ [0, 1].
Definition 7 ([19], p. 76) A normal neighborhood around x ∈ M is any open neighborhood of x which is a diffeo- morphic image of a star shaped neighborhood of 0 ∈ TxM under expx map.
Definition 8 The injectivity radius of M is . i(M) = inf i(x), (2.13) x∈M where
. R i(x) = sup{r ∈ ≥0| expx is diffeomorphic onto 1 1 expx(Br(0))}. Fig. 1. S and S \ {p} (2.14) unit circle S1 ⊂ R2 in Figure 1. A local coordinate system [20] for S1 is given by the local homeomorphism ψ : S1 → R (see also Figure 1) defined by Definition 9 The metric ball with respect to on is d (M,g) ψ defined by θ 7→ (sin(θ), cos(θ)) ∈ R2, θ ∈ (0, 2π) ⊂ R1. (3.17)
. 1 B(x, r) = {y ∈ M | d(x, y) < r}. (2.15) Inthe case of theremovalofa point p from S , the Euclidean distance between points converging in S1 \{p} to p ∈ S1 from either side converges to zero. However, at the same time, the Riemannian distance converges to 2π which is the 1 The following lemma reveals a relationship between normal largest distance on S between any pair of points. neighborhoods and metric balls on (M,g). We generalize the stability notion for dynamical systems on Riemannian manifolds as follows. Lemma 2 ([32], p. 122) Given any ǫ ∈ R>0 and x ∈ M, suppose that exp is a diffeomorphism on B (0) ⊂ T M, x ǫ x Definition 10 For the time-varying dynamical system x˙ = and B(x, r) ⊂ exp B (0) for some r ∈ R . Then x ǫ >0 f(x(t),t), f ∈ X(M × R), x¯ ∈ M is an equilibrium if exp Br(0) = B(x, r). (2.16) x Φf (t,t0, x¯)=¯x, t ∈ [t0, ∞), (3.18)
where Φf is the integral flow of f defined by (2.7).
We note that Bǫ(0) is the metric ball of radius ǫ with respect Definition 11 ([2,5,8,16]) For the dynamical system to the Riemannian metric g in TxM. x˙ = f(x(t),t), f ∈ X(M × R), an equilibrium x¯ ∈ M is
3 Lyapunov Analysis on Riemannian Manifolds (i) uniformly Lyapunov stable if for any neighborhood Ux¯ of x¯ ∈ M and any initial time t0 ∈ R, there exists a neigh- We extend the notion of stability to dynamical systems borhood Wx¯ of x¯, such that evolving on Riemannian manifolds. This problem has been addressed in [1,5,27] in a geometric framework. The main ∀x0 ∈Wx¯, Φf (t,t0, x0) ∈Ux¯, ∀t ∈ [t0, ∞). motivation here is to characterize the local properties of Lya- (3.19) punov functions based upon the Riemannian distance func- tion. These properties will be of great importance in analyz- ing a range of dynamical systems evolving on manifolds. (ii) uniformly locally asymptotically stable if it is Lyapunov stable and for any t0 ∈ R, there exists Ux¯ such that It is importantto note that, dependingon the geometry of the state space of a particular dynamical system, Riemannian ∀x0 ∈Ux¯, lim Φf (t,t0, x0)=¯x, i.e. distance might be significantly different than the Euclidean t→∞ lim d(Φf (t,t0, x0), x¯)=0, t ∈ [t0, ∞). (3.20) distance of embedded manifolds. As an example consider a t→∞
4 ∗ ∗R with dxχ ∈ Tx M as per (3.25), and dtχ ∈ Tt . (iii) uniformly globally asymptotically stable if it is Lya- R punov stable and for any t0 ∈ , Definition 13 ([1,5,16]) (Lyapunov Candidate Functions) A smooth function v : M × R → R is a Lyapunov function ∀x0 ∈ M, lim Φf (t,t0, x0)=¯x, t ∈ [t0, ∞). (3.21) X R t→∞ for the time-variant vector field f ∈ (M × ) if v is locally positive definite in a neighborhood of an equilibrium x¯ for (iv) uniformly locally exponentially stable if it is locally t ∈ [t0, ∞) and Lf v is locally negative semi-definite in a asymptotically stable and for any t0 ∈ R, there exist Ux¯ and neighborhood of x¯. K, λ ∈ R>0 such that Definition 14 The time-variant sublevel set Nb,t of a pos- ∀x ∈Ux, d(Φf (t,t , x ), x¯) ≤ itive semidefinite function v : M × R → R is defined as 0 ¯ 0 0 . Kd(x0, x¯) exp(−λ(t − t0)),t ∈ [t0, ∞). (3.22) Nb,t = {x ∈ M, v(x, t) ≤ b}. By Nb,t(¯x) we denote a connected sublevel set of M containing x¯ ∈ M. (v) globally exponentially stable if it is globally asymptot- ically stable and for any t0 ∈ R, there exist K, λ ∈ R>0, The following lemma shows that there exists a connected such that, compact sublevel set of an equilibrium point of a dynamical system on a Riemannian manifold. ∀x0 ∈ M, d(Φf (t,t0, x0), x¯) ≤ Kd(x0, x¯) exp(−λ(t − t0)),t ∈ [t0, ∞). (3.23) Lemma 3 Let x¯ ∈ M and v : M × R → R denote an equilibrium and a Lyapunov function respectively for system (2.6). Then, for any neighborhood Ux¯ of x¯ and any t ∈ R, there exists b ∈ R>0, such that Nb,t(¯x) is compact, We note that the convergenceon M is defined in the topology x¯ ∈ int(Nb,t(¯x)) and Nb,t(¯x) ⊂Ux¯, where int(·) gives the induced by d which is the same as the original topology of interior of a set. M by Theorem 2.
Definition 12 ([5,16]) A function χ : M → R is Proof. The proof is based on the proof given in [5], Lemma locally positive definite (positive semi-definite) in a R neighborhood of x¯ ∈ M if χ(¯x) = 0 and there 6.12. In this case we fix time t ∈ and consider v(·,t): M → R as a smooth time-invariant function. In this case, exists a neighborhood Ux¯ ⊂ M such that for all x ∈U \{x¯}, 0 <χ(x) (respectively 0 ≤ χ(x)). we apply the results of Lemma 6.12 in [5] to complete the x¯ proof. Given a smooth function χ : M → R, the Lie derivative of χ along a time invariant vector field f ∈ X(M) is defined by . To analyze the behavior of dynamical systems on manifolds Lf χ = dχ(f), (3.24) we employ the notion of comparison functions defined in [16]. where dχ : TM → R is the differential form of χ. In any neighbourhood of x ∈ M, dχ is given locally by Definition 15 ([16]) A continuous function α : [0,b) → R n ≥0 is of class K if it is strictly increasing and α(0) = 0, ∂χ ∗ and of class K∞ if b = ∞ and limr→∞ α(r)= ∞. dχ = dxi ∈ Tx M, (3.25) ∂xi i=1 X Definition 16 ([16]) A continuous function β : [0,b) × . ∗ R → R is of class KL if for each fixed s, β(·,s) ∈ K where n = dim(M) and Tx M is the cotangent space of M ≥0 ≥0 at x, see [21]. and for each fixed r ∈ [0,b), β(r, ·) is decreasing with lims→∞ β(r, s)=0. Remark 1 For time-varying dynamical systems evolving on M, the Lie derivative of a smooth time-varying function The following theorem provides K and KL comparison χ : M × R 7→ R is defined by function bounds for trajectories of uniformly stable dynam- ical systems evolving on Riemannian manifolds. . ∂ L χ = dχ ,f(x, t) , (3.26) f(x,t) ∂t Theorem 3 Any time-varying dynamical system of the form (2.6), evolving on a connected n dimensional Riemannian where manifold (M,g), satises the following properties:
∗ ∗R dχ = dxχ ⊕ dtχ ∈ Tx M ⊕ Tt , • If an equilibrium x¯ ∈ M is uniformly Lyapunov stable, (3.27) then there exists a class K function α and a neighborhood
5 Nx¯, such that have δ(r2) < δ(r1). Denote the associated neighborhoods r1 r2 of B(¯x, r1) and B(¯x, r2) by Wx¯ and Wx¯ respectively, see d(Φf (t,t0, x0), x¯) ≤ α(d(x0, x¯)), 3.30. Then δ(r2) <δ(r1) implies that x0 ∈ Nx¯,t ∈ [t0, ∞). (3.28) c c ∃x ∈ Wr1 , s.t. x ∈/ Wr2 , (3.32) • If x¯ is uniformly asymptotically stable then there exists a 0 x¯ 0 x¯ class KL function β and a neighborhood Nx¯, such that c c r1 where B(¯x, r1) ⊂ B(¯x, r2). However, x0 ∈ Wx¯ results in , which d(Φf (t,t0, x0), x¯) ≤ β(d(x0, x¯),t − t0), Φf (t,t0, x0) ∈ B(¯x, r1) ⊂ B(¯x, r2), t ∈ [t0, ∞) r2 x0 ∈ Nx¯,t ∈ [t0, ∞). (3.29) contradicts x0 ∈/ Wx¯ . Hence, δ(r1) ≤ δ(r2).c
Choose a ζ ∈ Kcsuch that ζ(r) ≤ δ(r), r ∈ R≥0 (this −1 Proof. Let us consider a neighborhoodUx¯ ⊂ expx¯ Bi(¯x)(0), is always possible since δ is non-decreasing), and ζ : R where i(¯x) is the injectivity radius at x¯ ∈ M and Bi(¯x)(0) ⊂ [0, supr∈[0,∞) ζ(r)) 7→ ≥0 is a K class function. Note that Tx¯M. Note that i(¯x) > 0, see Proposition 2.1.10 in [17]. In ζ is bounded by δ, hence, supr∈[0,∞) ζ(r) is bounded. Now order to prove the first assertion we note that the uniform choose . Nx¯ = expx¯ Bsupr∈[0,∞) ζ(r)(0) ⊂ expx¯ Bδ(i(¯x))(0) Lyapunov stability of x¯, implies that there exists Wx¯ ⊂ . −1 Then, r = ζ (d(x0, x¯)), x0 ∈ Nx¯, implies M, such that x0 ∈ Wx¯ results in Φf (t,t0, x0) ∈ Ux¯ for all t ∈ [t0, ∞). Hence, Wx¯ ⊆ Ux¯ ⊂ expx¯ Bi(¯x)(0) and d(x0, x¯)= ζ(r) ≤ δ(r), x0 ∈ Nx¯, (3.33) Φf (t,t0, x0) remains in a normal neighborhood of x¯.
Lemma 2 implies that exp B (0) = B(¯x, r), provided 0 < and hence, by (3.30) x¯ r . r ≤ i(¯x). Hence, for any Ux¯ = B(¯x, r), 0