Notes on homoclinic bifurcation theory
Ale Jan Homburg KdV Institute for Mathematics University of Amsterdam Plantage Muidergracht 24 1018 TV Amsterdam The Netherlands [email protected]
Version 20051027
1 Introduction
These notes contain a proof of the disappearance of an isolated periodic orbit in a homoclinic bifurcation with a homoclinic loop to a hyperbolic saddle equilibrium. The notes include a treatment of the implicit function and the stable manifold theorem.
2 Implicit function theorem
At several places we will find the need to solve equations on Banach spaces. We include the statement of the implicit function theorem allowing us to do so. Implicit function theorems on Banach spaces were developed by Hildebrandt and Graves [12]. We follow the treatment in [7]. See [25] for an account of the history of implicit functions theorems.
2.1 Contraction theorem
The Banach contraction theorem says that a contraction on a complete metric space has a unique fixed point. Let S be a metric space with metric d. Recall that T : S → S is a contraction if for some θ ∈ [0, 1),
d(T (x), T (y)) ≤ θd(x, y). (2.1)
The constant θ is called a contraction constant.
1 Note that in an equivalent metric d˜, this gives d˜(T (x), T (y)) ≤ Cθnd˜(x, y) Theorem 2.1. If (S,d) is a complete metric space and T : S → S is a contraction, then n there is a unique fixed point x¯ of T in S. Also, for any x0 ∈ S, T x0 → x¯ as n →∞.
2.2 Uniform contraction theorem
For a contraction depending on a parameter, also the fixed point will depend on a parameter. A contraction is called uniform if the contraction constant can be chosen independent of the parameter. Theorem 2.2. Let U, V be open sets on Banach spaces X,Y , let U¯ be the closure of U, T : U¯ × V → U¯ a uniform contraction on U¯ and let g(y) be the unique fixed point of T (·,y) in U¯. If T ∈ Ck(U¯ × V,X), 0 ≤ k < ∞ then g(·) ∈ Ck(V,X).
Proof. We will show that g is continuous. This follows easily using the fixed point formulas and the definition of contraction. Compute |g(y + h) − g(y)| = |T (g(y + h),y + h) − T (g(y),y)| ≤ |T (g(y + h),y + h) − T (g(y),y + h)| + |T (g(y),y + h) − T (g(y),y)| ≤ θ|g(y + h) − g(y)| + |T (g(y),y + h) − T (g(y),y)|. Therefore 1 |g(y + h) − g(y)| ≤ |T (g(y),y + h) − T (g(y),y)|. 1 − θ This implies continuity of g. Derivatives of g are treated as follows. Start at the fixed point equation T (g(y),y)= g(y), differentiate to obtain equations for derivatives, solve these equations and prove that the solution indeed equals the derivative of g. Formally differentiating T (g(y),y)= g(y) with respect to y gives
DxT (g(y),y)Dg(y)+ DyT (g(y),y)= Dg(y). So Dg(y) should be a solution M(y) to
DxT (g(y),y)M(y)+ DyT (g(y),y)= M(y). (2.2)
This equation can be solved since DxT (g(y),y) gives a contraction. To establish that M(y) equals the derivative of g, we must show that |g(y + h) − g(y) − M(y)h| = o(|h|) as h → 0. Write γ(y)= g(y + h) − g(y). Now γ(y) = T (g(y)+ γ(y),y + h) − T (g(y),y)
= DxT (g(y),y)γ(y)+ DyT (g(y),y)h + R, (2.3)
2 where R = T (g(y)+ γ(y),y + h) − T (g(y),y) − DxT (g(y),y)γ(y) − DyT (g(y),y)h. Since T is differentiable, for any ǫ > 0 there is δ > 0 with |R| < ǫ(|γ(y)| + |h|) if |γ(y)|, |h| < δ. Since γ is continuous in h we may further restrict δ so that this estimate holds for |h| < δ. −1 From (2.3) we get γ(y)=(I − DxT (g(y),y)) (DyT (g(y),y)h + R), so that |γ(y)| < C|h| for some C > 0, if |h| < δ. This implies |R| < ǫ(1 + C)|h| for |h| < δ.
Equation (2.2) shows that (I − DxT (g(y),y))M(y) = DyT (g(y),y). With (2.3) this yields (I − DxT (g(y),y))(γ(y) − M(y)h)= R. We obtain |γ(y) − M(y)h| < Kǫ|h| for some K > 0, if |h| < δ. This proves that M(y) equals the derivative Dg(y). Higher order derivatives are treated by induction. Assume that T is of class Ck and g has been shown to be Cj, j 2.3 Implicit function theorem The implicit function theorem is a corollary of the uniform contraction theorem, by combining it with a Newton type algorithm. Theorem 2.3. Suppose X,Y,Z are Banach spaces, U ⊂ X,V ⊂ Y are open sets, F : U × V → Z is continuously differentiable, (x0,y0) ∈ U × V , F (x0,y0)=0 and DxF (x0,y0) has a bounded inverse. Then there is a neighborhood U1 × V1 ∈ U × V of (x0,y0) and a function f : V1 → U1, f(y0) = x0 such that F (x, y)=0 for (x, y) ∈ U1 × V1 if and only if k k x = f(y). If F ∈ C (U × V,Z), k ≥ 1, then f ∈ C (V1,X). −1 Proof. If L = [DxF (x0,y0)] , G(x, y) = x − LF (x, y), then the fixed points of G are the solutions of F = 0. The map G has the same smoothness properties as F , G(x0,y0) = 0, DxG(x0,y0) = 0. Therefore, |DxG(x, y)| ≤ θ < 1 for all x, y is a neighborhood U1 × V1 of (x0,y0). It is easy to choose the neighborhood so that G : U1 × V1 → U1. The result now follows from Theorem 2.2. Remark 2.4. Note that the above reduction to an application of the uniform contraction theorem works if y 7→ F (x, y) is continuous, as long as x 7→ F (x, y) is continuously differ- entiable and DxF (x0,y0) has bounded inverse. See [5]. 3 Invariant manifolds Consider a flow φt for a differential equationx ˙ = f(x) on a compact manifold M. A submanifold W ⊂M is invariant if φt(W )= W , t ∈ R. It is positive invariant if this holds for t > 0. It is locally invariant if f(x) ∈ TxW for x ∈ W . Invariant manifolds play an important role in the qualitative study of flows. There exists a substantial theory on robust invariant manifolds, for which perturbed flows have perturbed invariant manifolds. 3 We will see a glimpse of the general theory by studying dynamically defined invariant man- ifolds near equilibria. 3.1 Stable manifold theorem Consider a differential equation x˙ = f(x) (3.1) on a smooth compact manifold M. We assume that X is a smooth, that is C∞, map. Write (t, x0) 7→ φt(x0) for the flow of X giving the solution to (3.1) equal to x0 at time t = 0. Standard theory of ordinary differential equations shows that φt(x0) depends smoothly on (t, x0). Suppose (3.1) has an equilibrium p ∈M, f(p) = 0. Suppose further that the origin is a hyperbolic equilibrium, i.e. the spectrum of Df(p) is disjoint from the imaginary axis. s u s u There is a Df(p) invariant splitting E ⊕ E of TpM where Df(p) restricted to E (E ) has spectrum with negative (positive) real part. We refer to Es as the stable subspace of Df(p) and to Eu as the unstable subspace of Df(p). Define the stable set s W (p)= {x ∈M| lim φt(x)= p}. t→∞ u The unstable set W (p) is defined as the stable set for φ−t. There are several proofs of the stable manifold theorem (for a hyperbolic equilibrium) stated below. We give a variant of Perron’s proof [29], see also [19, 20, 21, 32, 16]. A more geometric approach to stable manifolds is with graph transforms. This approach originates with [11]. An extensive use of graph transforms is in [13]. We quote D.S. Anosov [2]: Every five years or so, if not more often, someone discovers the theorem by Hadamard and Perron, proving it by Hadamard’s method of proof or by Perron’s. s s s Proposition 3.1. W (p) is an injectively immersed smooth manifold with TpW (p)= E . Proof. We construct a local stable manifold defined as s Rn Wloc(p)= {x ∈ | d(x, p) < σ, lim φt(x)= p} t→∞ s s for a sufficiently small σ > 0. Then W (p) = ∪t≤0Wloc(p). Using a local coordinate system we may assume that p equals the origin in Rn and that we are working with a differential equationx ˙ = f(x) on Rn. Using a test function we may assume that X is globally close to Df(0). Indeed, let ψ : Rn → R be a test function, i.e. a smooth nonnegative function such that ψ = 0 outside a neighborhood of 0 and ψ = 1 near 0. Consider the rescaled version ψǫ(x) = ψ(x/ǫ). Then ψǫX + (1 − ψǫ)Df(0) converges together with first order derivatives to Df(0) as ǫ → 0. 4 Consider the time 1 map F = φ1. This is a smooth map. We will construct the local stable manifold for F and show that this is also invariant for φt. Write B for the space of bounded sequences N → Rn, equipped with the supnorm. Note that B is a Banach space. s u n n i Take coordinates x =(xs, xu) ∈ E × E on R . Let πi : R → E , i = s,u, be the canonical coordinate projections. Define Γ : Es × B → B by (π F (ν(i − 1)), π F −1(ν(i + 1))), i> 0, Γ(a , ν)(i)= s u s (a , π F −1(ν(1))), i =0. s u It is straightforward to show that Γ is smooth, see [32]. Here we note that a map G : B → B, G(ν)(i) = g(ν(i)), for a given smooth map g : Rn → Rn, has derivative DG(ν)ζ(i) = Dg(ν(i))ζ(i). −1 By a linear coordinate change, we may assume that Df(0)| Es and DX (0)| Eu are con- tractions. From this it follows that for fixed as, ν → Γ(as, ν) is a contraction on B. for X sufficiently close to Df(0) (assured by choosing ǫ small enough). This map possesses a unique fixed point Γ(as,ζ)= ζ. Since Γ is a smooth map, the fixed point depends smoothly on as by the uniform contraction theorem. Write ζas for the fixed point of Γ(as, ·). Note that π1(ζas (0)) = as. We claim that s Wloc = ∪as∈Es ζas (0) (3.2) We will first show that the right hand side equals the local stable manifold of f. The fixed point ζas satisfies πsζas (1) = πsf(ζas (0)), −1 πuζas (0) = πuf (ζas (1)). From this one concludes that ζas (1) = F (ζas (0)). Continuing one sees that ζas (i +1) = F (ζas (i)), i ≥ 0. Thus ζas is a positive orbit of f. It is easily seen that ζas (i) converges to 0 as i →∞. Namely, since ζas (i) lies on a smooth curve, π1ζas (i + 1) lies at least a fixed factor λ for some λ < 1 closer to 0 than π1ζas (i) (this is trivial for the linear map x 7→ DF (0)x and therefore holds for F as well, if F is close enough to x 7→ DF (0)x). This shows that the right hand side of (3.2) is indeed the local stable manifold of f. As φtζas (i +1)= φt(F (ζas (i))) = F (φt(ζas (i))), N {φtζas (i)}, i ∈ , lies on the local stable manifold of f as well. Uniqueness implies that the right hand side of (3.2) is the local stable manifold of the flow φt. We sketch an alternative construction directly for the flow φt. Write A = Df(0) and f(x)= Ax + g(x). Let πs and πu be coordinate projections as before. Use coordinates x =(xs, xu) in Es × Eu. An orbit {z(t)}, t ≥ 0 in W s(0) satisfies t ∞ At A(t−v) A(t−v) z(t)= e z(0) + e (πsg(z(v)), 0)dv − e (0, πug(z(v)))dv. Z0 Zt 5 0 Rn The right hand side gives a contraction on the space Cb ([0, ∞), ) of bounded continuous orbits. Its fixed point, for t = 0, gives a point on the stable manifold with prescribed s coordinate πsz(0) on E . 3.2 Center stable manifolds In the bifurcation study below we will assume that stable and unstable manifolds of a hyper- bolic equilibrium contain orbits moving with different exponential speeds to the equilibrium as time goes to plus or minus infinity. The general setup is as follows. Suppose that the sta- ble subspace Es of Df(p) admits a Df(p) invariant splitting Ews ⊕ Ess, where the spectrum of Df(p)| Ess has real parts lying to the left of the real parts of the spectrum of Df(p)| Ews . We call Ews principal stable directions (also weak stable directions) and Ess strong stable directions. There are invariant manifolds, invariant under the flow of X, associated to these Df(p) invariant directions. In this context we distinguish s s s • the stable manifold W (p) with TpW (p)= E , ss ss ss • the strong stable manifold W (p) with TpW (p)= E , ws,u ws,u ws u • the center unstable manifold W (p) with TpW (p)= E ⊕ E , Unstable manifolds, strong unstable manifolds and center stable manifolds, assuming appro- priate Df(p) invariant splittings, have analogous definitions. We will briefly indicate the constructions of these manifolds, using variants of Perron’s method of proof for stable manifolds. Several details are omitted. The stable manifold has been dealt with in Proposition 3.1. Graph transform techniques, which we do not consider here, form perhaps a more consistent approach to the construction of invariant manifolds of different kinds. ss ss ss Proposition 3.2. W (p) is an injectively immersed smooth manifold with TpW (p)= E . Proof. We indicate the difference with the construction of W ss,s(p) in Proposition 3.1. ss s,u,uu n n i Take coordinates x = (xss, xs,u,uu) ∈ E × E on R . Let πi : R → E , i = ss or i = s,u,uu, be the canonical coordinate projections. Define Γ : Ess × B → B by (π f(ν(i − 1)), π f −1(ν(i + 1))), i> 0, Γ(a , ν)(i)= ss s,u,uu ss (a , π f −1(ν(1))), i =0. ss s,u,uu Let the real number r be such that the line of complex numbers with real part r divides the r spectra of Df(p)| Ess and Df(p)| Ews,u . Note that λ < 0. Write λ = e , which is smaller than 1. Define A on B by Aν(i)= λ−iν(i). 6 N Rn −i Let Bλ = {ν : → | supi kν(i)kλ < ∞}. Then Bλ equipped with the norm kνk = −i ¯ ss supi kν(i)kλ . is a Banach space. Define Γ : E × Bλ → Bλ by −1 Γ(¯ ass, ν)= AΓ(ass, A ν). This is a smooth map for λ < 1. For fixed ass, ν → Γ(¯ ass, ν) is a contraction on Bλ. The −i i−1 i unique fixed point ζass satisfies ζass (i) = λ f(λ ζass (i − 1)), so that λ ζass (i) is an orbit. ws,u 1 ws,u Proposition 3.3. W (p) is an injectively immersed C manifold with TpW (p) = Ews,u. It is not uniquely defined, but its tangent bundle along W u(p) is smooth and uniquely defined. Proof. By reversing the direction of time, we need to construct a center stable manifold W s,wu(0). The construction in Proposition 3.2 works for center stable manifolds. Note that now λ > 1. We remark that center stable manifolds are continuously differentiable, but in general not smooth. The difficulty being that the map Γ¯ is not smooth, see [20, 22, 10]. We do not include a proof of differentiability here. Uniqueness and smoothness of the tangent bundle T W ws,u(p) along W u(p) is shown as fol- n n lows. Let Gws,u(R ) be the Grassmannian manifold of linear subspaces in R with dimension ws,u n n dim E . Lift the time one map F to a map F¯ on R × Gws,u(R ): F¯(x, α)=(F (x), DF (x)α). Observe that this is a skew product map. Clearly, (p, Ews,u) is a fixed point of F¯. A direct computation shows that (p, Ews,u) is a hyperbolic fixed point and the unstable directions n include the tangent space of the fiber {0} × TEws,u Gws,u(R ). The computation can be done using local coordinates in Rn ×L(Ews,u, Ess) near (p, Ews,u) In these local coordinates, F¯ takes the form F¯(x, v)=(F (x),w), with graph w = DF (x)graph v. To compute eigenvalues in the fiber over the origin, it suffices to consider the map v 7→ w, graph w = DF (0)graph v. See [14]. The tangent bundle T W ws,u(p) along W u(p) is the unstable manifold of the fixed point (p, Ews,u) of F¯. It is therefore unique and smooth. 3.3 Families of differential equations Consider a family x˙ = fµ(x) (3.3) of differential equations with (µ, x) 7→ fµ(x) smooth. If for µ = µ0, (3.3) has a hyperbolic equilibrium p, then it has a hyperbolic equilibrium pµ that continuous p for µ near µ0. Also µ 7→ pµ is smooth. 7 By incorporating the parameter µ in the construction of stable manifolds above, one sees s that W (pµ) depends smoothly on µ. The same holds for strong stable manifolds. Center stable manifolds depend continuously differentiable on µ. Note that can consider an ex- tended differential equationx ˙ = fµ(x), µ˙ = 0 and construct a center stable manifold of the equilibrium (pµ0 ,µ0). 4 Homoclinic bifurcations For a smooth compact manifold M, let X (M) be the set of C∞ vector fields on M equipped with the C∞ topology. Consider a system of differential equationsx ˙ = f(x), X ∈ X (M) such that X has an equilibrium p: f(p) = 0. We will assume that p is hyperbolic. Note that, by the implicit function theorem, p persists perturbations of X. There is a Df(p) invariant 1 2 1 2 splitting E ⊕ E of TpM where Df(p) restricted to E (E ) has spectrum with negative (positive) real part. We distinguish the following cases. • p is of saddle-saddle type: the eigenvalues of Df(p)| E1 and Df(p)| E2 closest to the imaginary axis are real and simple. • p is of saddle-focus type: the eigenvalues of Df(p)| E1 closest to the imaginary axis are complex conjugate, the eigenvalue of Df(p)| E2 closest to the imaginary axis is real, • p is of bifocus type: the eigenvalues of Df(p)| E1 and Df(p)| E2 closest to the imaginary axis are complex conjugate. The eigenvalues nearest to the imaginary axis are called the principal eigenvalues or weak eigenvalues. 4.1 Blue sky catastrophe A blue sky catastrophe is a bifurcation in which a periodic orbit developes infinite period as the parameter moves to the bifurcation value, after which the periodic orbit disappears [28]. Homoclinic bifurcations, treated in this section, lead to this phenomenon. There are other bifurcations, not involving an equilibrium, that show this phenomenon as well, see e.g. [18]. Some authors reserve the phrase blue sky catastrophe for bifurcations not involving an equilibrium. A homoclinic orbit (to p) is a solution ofx ˙ = f(x) converging to p for t → ±∞. We will consider an equilibrium p of saddle-saddle type. There is thus a Df(p) invariant splitting ss ws wu uu E ⊕ E ⊕ E ⊕ E of TpM in strong stable, one dimensional weak stable, one dimen- sional weak unstable and strong unstable directions. Notation for invariant manifolds is as introduced above. Let Γ = {γ(t) | t ∈ R} be a homoclinic orbit to p. A number of open assumptions will be made. 8 xss Γ xu xs Figure 1: The orbit flip. (H1) T W s(p) intersects T W u(p) along Γ only in Rγ˙ , (H2) Γ ∈ W s(p)\W ss(p) ∩ W u\W uu, (H3) Γ ∈ W s,wu(p) ⋔ W u(p) and Γ ∈ W s(p) ⋔ W ws,u(p), (H4) λs + λu =6 0. Here A ⋔ B stands for a transverse intersection of manifolds A and B. Violation of (H 2) is called an orbit flip [30], of (H 3) an inclination flip [15]. A homoclinic orbit with λs + λu =0 is called a resonant homoclinic orbit [6]. Denote by XH the set of vector fields f with a homoclinic orbit to a saddle-saddle type equilibrium as above. In particular the open and dense conditions (H1), (H2), (H3), (H4) are assumed to hold. Consider a family of differential equationsx ˙ = fµ(x), depending on a real parameter µ, with f0 ∈ XH . For µ near µ0, fµ possesses an equilibrium pµ near p = p0. Invariant manifolds, such as the stable manifold W s(p), have a continuation in µ. We will s write W (pµ) and similarly for the other invariant manifolds. Take a local cross section Σ transverse to the flow of f0 at a point in the homoclinic orbit ss,s u,uu Γ. Write d(µ) for the signed distance between the intersections of W (pµ) and W (pµ) with Σ. Then d is a smooth function of µ and d(0) = 0. Recall condition (H 4). By changing the direction of time, if necessary, we may assume that λs + λu < 0 at µ = 0. The following theorem is due to Shil’nikov [31]. A two dimensional version is in [1]. Theorem 4.1. Let x˙ = fµ(x) be a family of differential equations, depending on a real ∂ parameter µ, with f0 = f as above. Assume that ∂µ d(µ) =6 0 at µ = 0. The family has a 9 W s ss x Γ xu xs W s ss x Γ xu xs Figure 2: The inclination flip. The geometry of the convergence of the stable manifold to the equilibrium, for negative times, depends on eigenvalues. 10 hyperbolic periodic orbit for µ > 0 that converges in the Hausdorff topology to Γ as µ → 0. Its period converges to infinity as µ → 0. 5 Planar vector fields We will present a proof of Theorem 4.1 for families of differential equations in the plane. Lemma 5.1. In smooth local coordinates x =(xs, xu) near p, and after a multiplication by a smooth positive function, the vector field X is given by the set of differential equations s s x˙ s = λ xs + f (x), u x˙ u = λ xu, where f s(x) satisfies s 2 f (x) = O(|xs| ), as kxsk→ 0. Proof. We may take smooth coordinates x =(xs, xu) near p, so that s W (p) = {xu =0}, (5.1) u W (p) = {xs =0}, (5.2) In such coordinates we have s s x˙ s = λ xs + f (x), u u x˙ u = λ xu + f (x), s u s u for quadratic and higher order terms f , f with f (x)= O(|xs|kxk) as xs → 0 and f (x)= u u O(|xu|kxk) as xu → 0. Dividing the vector field by the positive function λ + f (x)/xu gives u s an expression with f = 0 and still f (x)= O(|xs|kxk). 2 Write f s(x , x )= kx x + gs(x , x ) with ∂ gs(x , x )=0in(x , x )=(0, 0). A direct s u s u s u ∂xs∂xu s u s u computation shows that a polynomial coordinate change removes the term kxsxu from the differential equation for xs. Indeed, let ys = xs + axsxu and yu = xu. Theny ˙u = λuyu and y˙s = λsys +(k + aλu)ysyu + ... 2 where the dots stand for terms O(|ys| ) and terms of order three and higher. As λu =6 0, we can choose a such that k + aλu = 0. This does the trick. s 2 2 We may now assume f (xs, xu) = O(|xs| )+ O(|xs|kxk ). To finish the proof, we must remove terms O(xsxu) from the equation for xs. We follow the reasoning from [26]. Consider a coordinate change ys = xs + p(xu)xs, yu = xu 11 for a function xu 7→ p(xu) that vanishes at xu = 0. Write the differential equations in the new coordinates as y˙s = λsys + P (ys,yu)ys, y˙u = λuyu. Treating p as a variable, compute s s y˙s =x ˙ s +px ˙ s + px˙s = λsxs + f (xs, xu) +px ˙ s + pλsxs + pf (xs, xu)= λsys +py ˙ s + h.o.t., where h.o.t. are terms of quadratic and higher order involvingp,p,y ˙ s,yu. Along the unstable manifold ys = 0, P (0,yu) =p ˙+h.o.t., where the higher oder terms are of quadratic and higher order in yu,p. We look for a function p so that P (0,yu) vanishes. This yields differential equations p˙ = h.o.t. y˙u = λuyu We obtain p as the one dimensional unstable manifold for this system of differential equations. Remark 5.2. One can smoothly linearize the vector field restricted to the stable manifold s λst W (0). This is done through the inverse of the map e 7→ xs(t), or 1 λ ln z s s z 7→ z 1+ f (xs(v), 0)dv . Z0 ! One can estimate that this is a smooth map. This brings s 2 f (x)= O(|xs| |xu|). Take a cross section Σ transverse to the homoclinic orbit. in out Proposition 5.3. Let Ψ : Σ → Σ be the first return map given by the flow φt of X. Write β = −λs/λu. In coordinates near p as in Lemma 5.1, There exists φ =6 0, ω > 0, so that as xu → 0, xu 7→ Ψ(xu) satisfies dk Ψ(x ) − φxβ ≤ C |x |β+ω−k, dxk u u k+l u u for positive constants Ck+ l. 12 in out Proof. Let Σ ⊂{xs = δ}, Σ ⊂{xu = δ} be cross-sections close to p. By a rescaling we may assume δ = 1. Note that in the normal form s s x˙ s = λ xs + f (x), u x˙ u = λ xu, s 2 this leaves us with f (x)= δO(|xs| ). It suffices to consider the first hit map Sin : Σin → Σout as the first hit map Sout : Σout → Σin is a local diffeomorphism by the flow box theorem [27]. Write β = −λs/λu. We will show that Sin satisfies k in in in β in β+ω−k D in S (x ) − φ(x ) ≤ C |x | , xu u u k+l u