Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php SIAM J. MATH. ANAL. Vol. 22, No. 3, pp. 653-679, May 1991 (C) 1991 Society for Industrial and Applied Mathematics 006 THE BIFURCATIONS OF COUNTABLE CONNECTIONS FROM A TWISTED HETEROCLINIC LOOP* BO DENG? Abstract. Codimension-two bifurcation phenomena associated with nondegenerate heteroclinic loops are studied. The bifurcation curves of homoclinic orbits in the parameter space are characterized by the twist structure of the heteroclinic loops at the bifurcation points. Among other things, it is shown that heteroclinic orbits with any given winding number around a doubly twisted heteroclinic loop must bifurcate. Applications of these bifurcation phenomena are also discussed. Key words, twisted heteroclinic orbit, homoclinic orbit, periodic orbit, k-heteroclinic orbit, Sil’nikov’s variables, exponential expansions, strong A-lemmas, entrance sets, exit sets, bifurcation equations AMS(MOS) subject classifications. 34A34, 34C28, 34C99 1. Introduction. A heteroclinic loop takes place for a vector field F when there exist two heteroclinic orbits z*(t) and z*(t), with z* connecting an equilibrium point a l, to another one a2, and z2* connecting a2 to a l. To be precise, z/*(t) -> ai as --> -o and z/*(t) --> a as --> + for i,j= 1,2 and i#j. Figures 1.1 and 1.2 heuristically illustrate what could happen to two structurally different loops when a planar vector field F is perturbed slightly. In Fig. 1.1, either a homoclinic orbit or a periodic orbit would possibly bifurcate from the loop, while in Fig. 1.2 a heteroclinic orbit winding around the original loop for any finite times before reaching its destinations in both backward and forward evolutions would also be possible under perturbation. The very structure distinguishing the second loop from the first one is that a given heteroclinic orbit arises from and tends to the equilibria from different "sides" of the other heteroclinic orbit. The purpose of this paper is to study the bifurcations of a generic two-parameter family of vector fields in Ea, d _-> 2 which exhibit the above heteroclinic phenomena. The first obvious generalization is to assume that the equilibria ai of the equation :-- F(z) (1.1) FIG. 1.1 FIG. 1.2 Received by the editors April 26, 1988" accepted for publication (in revised form) March 23, 1990. Mathematics and Statistics, University of Nebraska-Lincoln, Lincoln, Nebraska ? Department of 68588-0323. 653 Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php 654 o DENG have the same dimension, m->_ l, for the stable manifolds W and the same unstable 1 and 2. Moreover, dimension, n d m _-> l, for the unstable manifolds W’ for both the ais are simple saddlepoints in the sense that (1.2) There exist principal eigenvalues Ai <0< for the linearization DzF(a) and constants A <0</2 such that for any other eigenvalue , of DzF(ai) either Reu<<Ai or Re,>/2i>/,for i=l and 2. Not as a generalization but as a generic restriction to all the cases, we assume that both equilibria are relatively contractive" (1.3) +/<0 for i=l and 2. That is, the principal attraction of a dominates the principal repelling. Concerning the structure on the intersection of the unstable manifold W’ of a and the stable manifold W] of aj which must be nontransverse along the heteroclinic orbit Fi := {z*(t)" 6 R}, we assume that they are in general position" (1.4) codim where and Tp W means the tangent space of a given manifold W at a base point p W. Also, motivated by the strong h-lemma from Deng (1989), the following strong inclination property, as another assumption, is also generic: lim t--) (1.5) lim t-)Too T,, W/ + Ta, WSi Tz(t T()= T,W’’+ T,W. , Here, W and W are the strong stable and strong unstable manifolds of a, respectively. See Fig. 1.3. If the vector field P is C then W and W are C as well (see, instead (see e.g., Shub (1987)). But, in general, W and W are proved to be C Deng (1989)). Moreover, W and W are (m-1)- and (n-1)-dimensional, respectively, characterized by the fact that the limits lim (1.6) lim z(t)-a 0 for z(0) - WU\ W uu, z(t)-a ,-+oollz(t)-all exist and are equal to unit eigenvectors for the principal unstable eigenvalue and principal stable eigenvalue, respectively. The strong inclination property is a generic property provided F is C with r >-7. See Deng (1989) for the proof. The last structural assumption reads W/\ W’/") ("1 W\ Ws) for i,j= 1, 2 and j. That is, by virtue of (1.6) this hypothesis says that the heteroclinic orbits arise from (1.7) F, and tend to the equilibria along principal eigendirections. It is certainly a generic condition. The assumptions (1.4), (1.5), and (1.7) together are referred to as nondegeneracy. They lead to our classifications of heteroclinic orbits into twisted and nontwisted as follows. THE BIFURCATIONS FROM A TWISTED HETEROCLINIC LOOP Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php W 655 e- e W’’ e; a2 W (a) (b) (c) FIG. 1.3. (a) a nontwisted loop, (b) a single twisted loop, (c) a double twisted loop. Let e-= lim (z*,(t)-a)/llz*,(t)-all, (1.8) e-= lim (z*(t)-a)/llz*i(t)-all. By (1.6) and (1.7), they are unit principal eigenvectors. See Fig. 1.3. Choose pi G and q F W]o sufficiently close to the other equilibrium a. Let p z(0) and q z(T) for a large T 0. Because of the strong inclination propey (1.5) and the principal asymptotic tangency (1.8), choosing p and qj close enough to a and a, respectively, implies F WTo sufficiently close to the equilibrium a (1.9) e[ Tp, ef Tq, d= Tp, +span (el), Nd Tq +span (el). Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php 656 o DEYG Since Tz(,), 0 <= <- T, defines a homotopy from Tp, to Tqj, the following definition is justified (see Fig. 1.3). DEFNa’ION 1.1. Let Fi be a nondegenerate heteroclinic orbit connecting two and ef point to opposite sides of Tp, and simple saddles. Fi is said to be twisted if it is nontwisted. respectively. Otherwise, Tqj, For the heteroclinic loop, cl (F1 (-.J F2) {a, a2} (-J 171 (-J F2, it is called double twisted if both F and F2 are twisted, single twisted if and only if one of them is twisted, and nontwisted if otherwise. As the last assumption we assume e (1.10) F’E a x2---, d is a generic C vector field with two parameters a (a l, a2) E2, having a nondegenerate heteroclinic loop at a 0. Here, the regularity r >_- 8. By genericity we mean that our results will hold for a residual subset of C (d X 2, d in the weak Whitney topology of Ck-convergence (see, e.g., Hirsch (1976)). To be more precise, we include the following as equivalent conditions for (1.10): (1.10a) The continuation of Fi: Given the fact that the heteroclinic orbit is a codimension 1 object by (1.4), we assume for every (a 1,0) there exists a heteroclinic orbit F2(al) {z*2(t, al)" } from a2 to al such that z2* is a C two-dimensional surface in the phase space. Similarly, z*(t, a2) forms a C surface in d as (t, a2) takes all values from 2 and each of the t-curves is a heteroclinic orbit from a to a2; (1.10b) The transverse crossing of the stable and unstable manifolds along F: d, (1, 0) d(0, lim--0 and lim--0, where dl(al, a2) denotes the continuously varying distance of W’;(a)fE and W(a)Z with d(0,0)=0 and E is an arbitrarily chosen Poincar6 cross section to F. A similar description applies for d2. Note that since the Poincar6 mapping introduced between any pair of two cross sections is diffeomorphism, the nonzero limiting property in (1.10b) above is independent of the choice of the cross section Z. Finally, to state our main theorems we need a few more terms. Let be a small tubular neighborhood of the heteroclinic loop cl (F1 t_J F). A k-periodic (k-per) orbit is a periodic orbit which is contained in OR and has winding number k in OR. Similarly, the closure of a k-homoclinic (k-hom) orbit has winding number k in OR. Accordingly, a k-heteroclinic (k-het) orbit F from a to a2 is such a heteroclinic orbit that cl (F t_J Fz) has winding number k + 1. Similarly, we define a k-heteroclinic (k-het2) orbit from a to al. Thus, F and F2 themselves are zero-heteroclinic orbits. Note that as long as R is chosen small enough, the above definition is independent of any particular choice of OR. Also, the terminology extends canonically to small perturbations of the vector field F(., 0). The first theorem, except for the directions of bifurcation and the k-heteroclinic orbits, is taken from Chow, Deng, and Terman (1990). THEOREM A. Suppose F is a generic two-parameter family of vector fields having a nondegenerate heteroclinic loop connecting two relatively contractive and simple saddle equilibria at a O, i.e., (1.2)-(1.5), (1.7), and (1.10a, b) are satisfied. Then there exists a small tubular neighborhood OR of the heteroclinic loop cl (F kJ F2) and a neighborhood THE BIFURCATIONS FROM A TWISTED HETEROCLINIC LOOP 657 of the bifurcation point a 0 in the parameter space such that up to a nonsingular and differentiable change of the parameters, which leaves the axes invariant as sets, the following is satisfied (cf. the bifurcation diagram Fig. 1.4): (i) There exists a C r-7 curve a2 hOml (al) with al > 0 in f such that there exists a homoclinic orbit to a in if hOm l. Moreover, hom is asymptotically if and only + and the direction of the bifurcation is determined the as to a a 0 tangent positive 1-axis by the twist of the heteroclinic orbit F2 as follows: >0 if F is twisted, hOml Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php f <0 otherwise. to al for k >- 1 if F2 is not twisted and there exists at least one 1-heteroclinic orbit from a to al on a C r-7 curve; a= 1-het (al) for a >0 otherwise. Moreover, 1-hete is asymptotically tangent to the 0 a -axis as a (iii) Analogous statements hold for homoclinic orbits to ae and k-heteroclinic orbits from a to a2. (iv) Let A={(al,a2):a2>hOml(al) if al>O or a>hom2(a2) if a2>O}. Then there exists a periodic orbit in 71 if and only if a A. (v) The homoclinic and periodic orbits do not coexist in all for a given parameter. They are all unique and are 1-hom and 1-per orbits, respectively. (ii) There does not exist any k-heteroclinic orbit from a2 +. Our main result is as follows. THEOREM B. In addition to the hypotheses of Theorem A, suppose the stable manifolds of the equilibria a and a are all one-dimensional; then the following is satisfied (cf. Fig. 1.4): (i) If F2 is twisted but F1 is not, then the 1-heteroclinic orbit from a to al is the Moreover, unique k-heteroclinic orbit for all k >= 1 and a . 1-het (a,) >hom, (al) for al > O. (ii) If the loop cl (F U F) is double twisted, then there exist two sequences of C r-7 curves a2=k-het(a) with a>O and al=k-hetl (ae) with a>O in f, respectively, satisfying 0 _--< k-het2 < (k + 1 )-het < hOml all k >= 0 such that there exists a k-heteroclinic orbit from a to a if and only if k-het2. Moreover, it is a unique heteroclinic orbit in 71 with respect to the parameter and k-het is asymptotically tangent to the a -axis as a 0 +. Furthermore, the homoclinic bifurcation curve hOml is inaccessible from below in the sense that for every al for a - k-het (a,) horn, (a,) An analogous statement also holds for the k-hetl - as k- +. curves. Theorem A provides us with a useful clue to the twist of a given heteroclinic loop: the two zero-heteroclinic continuation curves (which are the parameter axes in our theorems) divide the neighborhood f into four sectors. The 1-homoclinic bifurcation curves hom and hom lie in one sector for a double twisted loop, or in two adjacent sectors for a single twisted loop, or in two opposite sectors for a nontwisted loop. Keeping this fact in mind, let us examine the following bifurcation diagram Fig. 1.5 for traveling waves of the FitzHugh-Nagumo equation vt=Vxx+f(v)-w, w,=e(v-yw), e,y_-->O, where f(v)=-v+H(v-a) with H to be the Heaviside step function and O<a<1/2. Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php 658 loop. o ozy 0 hom (a) (b) FIG. 1.4. The bifurcation diagrams for (a) a nontwisted loop, (b) a single twisted loop, (c) a double twisted Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php THE BIFURCATIONS FROM A TWISTED HETEROCLINIC LOOP 659 tl-t t/2-het,//" -////--hom, / (c) FIG. 1.4.--continued A traveling wave solution (v, w)(x, t) is a function (v, w)(z) of z=x+ct, c>=O. Let u(z)= v’(z), then vc (v, u, w)(z) satisfies a first-order system of ODE (1.11) v’=u, u’=cu-f(v)+w, w’=e-(v-yw). c For fixed 0 < a < 1/2 and 0 < e << 1, numerical as well as rigorous arguments from Rinzel and Terman (1982) show that the front curve Or, on which there exits a front wave connecting the rest steady state ff to the exitable state g as shown in Fig. 1.6, crosses transversely the back curve 0B, on which there is a back wave from g to ft. Thus, at the intersection point 0* there exists a front wave and a back wave traveling at the same speed. This gives rise to a heteroclinic loop. Their numerical simulation also shows that the impulse curve 0e, homoclinic to if, and the g’-impulse curve 0, homoclinic to g, also bifurcate from the loop at 0". Note that their bifurcation directions of asymptotic tangency are exactly opposite our bifurcation diagram Fig. 1.4 for Theorems A and B. This is due to the fact that the steady states ff and g are relatively contractive simple saddles having one-dimensional stable manifolds only for the time reversed (z--z) system (1.11). What is most remarkable about these two curves is that both of them lie in the same sector in the parameter space. In fact, this has been rigorously proved (see (3.8) and (3.10) from Rinzel and Terman (1982)). Unfortunately, however, we can only speculate that Theorem A suggests the double twist for the heteroclinic loop. Indeed, we are facing a tantalizing dilemma here: either it is feasible to check the transverse crossing condition (1.10b) and the double twist of the loop due to the piecewise linearity of f but the vector field is not smooth enough, or it becomes a fairly open problem to do so for a smooth vector field, e.g., the usual cubic function f= v(v-1)(a-v). Nevertheless, the implication is interesting: for given Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php 660 o DENG OF 73 7 3’2 7 (a) (b) FIG. 1.5. (a) A heuristic bifurcation diagram produced from Rinzel and Terman (1982); (b) The conjectured complete diagram. w FIG. 1.6. The conjectured twist for the front-back wave loop. parameters e, 0 < a < 1/2 and 0 < Y3 Y << 1 there would be infinitely many fronts traveling at different speeds. The more "humps" a front were to carry the slower it would travel. If the humps were "too many" (infinity) the traveling wave arising from the rest state but would return to itself after a would never be able to reach the exitable state long excursion. Slowing down a little, it would become a traveling train, or periodic orbit for the ODE. On the other hand, push y slightly to the right of Y3, the above scenario would repeat for back waves and q-impulses. By their numerical evidence on the stability of the primary front and back waves with respect to the PDE, as well as other authors’ results on somewhat related stability 661 Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php THE BIFURCATIONS FROM A TWISTED HETEROCLINIC LOOP problems, it has been demonstrated that the stability of a given impulse is closely related to the direction at which the stable and unstable manifolds cross transversely as the speed parameter c varies (see, e.g., Evans (1972), Jones (1984), Kokubu, Nishiura, and Oka (1988)). Thus, we find the second implication is most interesting: there would be infinitely many stable transition waves connecting two stable patterns. However, all of these phenomena do not appear in the "chaotic" parameter region discussed by Evans, Fenichel, and Feroe (1982) and Hastings (1982), where for a given speed there are infinitely many impulses and traveling trains due to the Sil’nikov saddle-focus homoclinic explosion for (1.11) (see also Sil’nikov (1967)), whereas there would be a unique traveling front, or back, or impulse, or traveling train, except at the bifurcation point 0* in our case. Indeed, as long as there are two bistable steady states as shown in Fig. 1.6, the equilibria are not saddle-focus. Nevertheless, we would probably not be too surprised by the enormous, stable, yet not "chaotical" transporting capability that a nerve axion would inherit if our conjectures were true. In contrast to our conjectures above, we will discuss the existence of nontwisted heteroclinic orbits and thus the limited number of connections between two equilibrium states for another type of reaction diffusion systems in 7. We will also discuss in that section some ways newly discovered by other authors to check all the nondegenerate and generic conditions (1.4), (1.5), (1.7), and (1.10a, b) for their examples to which our theory is immediately applicable. 2. Preliminaries. This section is devoted to introducing the Sil’nikov variables for a Poincar6 map around the loop. Let 0< 6o be a small number and B(6o) --{Z: Z--Z ,’’’, z(d)), ]z(i (0} be the 6o-box of the origin. Let the coordinate be locally normalized near the equilibria so that (x, y)=0 corresponds to z ai and the local stable, unstable manifolds are given by the x-axis and y-axis in B(6o), respectively; i.e., Woc {y 0} B(8o) and W’oc {x=O}B(6o). In addition, the directions of the first x-component x (1) and and e-, the first y-component y(1) are chosen to be the unit principal eigenvectors respectively, as in (1.8). Let the points p and q from (1.9) in the definition of twist (1) be specifically given as Pi (xi, O) and q (0, y). We can assume xi 8o and Yi(1) 8o e- because of assumption (1.7) for the asymptotic tangency of F along the principal eigendirections. Let E and E’ be two small cross sections, or (d- 1)-dimensional boxes B(81) and e-, respectively with 0 81 80, centered at p and qi and perpendicular to and Poincar sections are cross provided 8o (see Fig. 2.1). They 81 are sufficiently be of whose initial the subset those small. Let trajectories in points (Xo, yo)eE This at the time exit cross section E’ at (xl, Yl) corresponz(x0, Yo). B(8o) first hit dence gives rise to the local Poincar map I-I" E’ by (Xo, Yo)-* (xl, Yl). Similarly, by the continuous dependence on initial data and parameters we can define a global Poincar map IIi E’ -* E. Here, without loss of generality Es is taken to be the domain is a proper subset of E not containing any point from of definition for I-Ii, whereas the stable manifold Woc. Let (, y)e a-1 and (x, )e a-1 be the normalized local coordinates on E and E’ so that (0,0) corresponds to the center points Pi and qi, respectively. Indeed, ,y(")) (see s=)-)i, and /=3-3i., where )---(X(2), ,X (m)) and )=(y(Z) -() the is snear time e Sil’nikov where be the Let principal 2.1). a, Fig. i(a) unstable eigenvalue for DzF(a, a). Then the Sil’nikov variable for the local map is (s, :, 7) and the Sil’nikov domain is e r r- r ..., A:= {(s, , o 662 DENG Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php r //" q constant -r .><> Q.:.-.y.r.._ r] Y(/[ constant e cross sections and te corresponding Poinear map en m FIG. 2.1. E2 ip2 const ant 1, n 2. where So(a)= e -,(")o for some large but fixed to. Note that the dependence on a is suppressed from rT, o’’, and ki. It has been proved by Sil’nikov (1967) that for the initial point (Xo, Yo)e and the end point (x,, y)e with r-- r(Xo, Yo) time units apart, the initial Yo and the end x, components are functions of the Sil’nikov variable: yo := Y/(s,:, ’9, a) and Xl:=Xi(s,, "9, a). Moreover, it has been observed by Deng (1989) that the maps p’A- o-, with (s, :, "9)with (s, :, ’9)- (X(s, :, "9, a), "9) are actually diffeo(, Y(s, sc, "9, a)) and p’" Ai morphisms of class C r. Thus, p gives rise to a smooth change of variables for the local Poincar6 map 1-I, which in turn is p’ under the new Sil’nikov variable. See Fig. 2.2. More important, we have the following exponential expansion result. PROPOSiTiON 2.1 (Deng (1988), (1989)). Let the strong stable manifold and the strong unstable manifold also be normalized such that Ws= {x (1)= 0, y 0} and W {x=0, y() 0} locally. Let vi(a) hi(a)/tx(a)-1 and (a) and i(a) be as in (1.2). Then for sufficiently small there exist C r-7 functions p(sc, a), ff(sc, "9, a), Nil(S, "9, o), and R2(s, "9, a) over k such that o- ’ r , X, s, "r], ce) qgi(, "r], ol s + + R Y(s, sc, "9, a)= ,(, "9, a)s+ R,2(s, (2.1a) with q and (2.1c) No(s, , for all s, "9) , , S, "9, "r], ol), a) satisfying (2.1b) and Rij , ,, o , (i(, ’9, a)= emao+ o(([l / Inl / IlJi "9, Ol e.6o + O((Iscl +1"91 + 8o)ao) ’9, ) satisfying k g ,1[= O(s l+v.+.’), ID, R,11= O(s IDe.) , ) R,21 O(SI+"), ]O(e,, A ,, 0 <--_ k <-_ r ]DsR,2[ O(sO’), 7, where e,--(1, O,.. ",o)T m, and e,, =(1, O,..., 0) r --F ), Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php THE BIFURCATIONS FROM A TWISTED HETEROCLINIC LOOP FIG. 2.2. The Sil’nikov change of variables for the local map when rn n 663 2. and i>O is a constant not greater than rain {li(a)/txi(a), (li(OI.)--i(O))f ld(O)} for all [a[ -< 60. Equation (2.1a) is referred to as the exponential expansion, and i the expansion coecient functions, and Rg, the remainders. From this proposition we immediately have the following proposition. PROPOSITION 2.2. For sufficiently small but fixed So, the domain of the local map contained is y) the on sector vertical in < boundary point Moreover, every H for (, y) (, (So, )), y) 6oSo/2. (See Fig. 2.2.) It is also easy to see from (2.1c) that the functions X and thus p and p, can be C extended to s 0. From now on let us use the same notation for the extended for the extended domain of A. functions, but Let us conclude this section with two lemmas which will be frequently used later. LEMMA 2.3. Let the global map Hi be expressed as P(x, a), y Qi(x, a), under the new coordinates for and Z. Let . , __ , , , DnP(O, O, O) 0 and M D,Q(O, O, O) e DnQ(O, O, O) (a-2)(a-2) where e (0, O, 0), or e. en both Mi and i are nonsingular for sufficiently small 60. M Proof Note first that all the column vectors of M except the middle one (0, e) span the linear subspace Tp W7 E) + Tp W} E}), which has dimension d 2 by the hypothesis (1.4) and the choices of E, which are transverse to the flow. On the other hand, the remaining column vector is approximately parallel to the principal unstable eigenvector e; by the exponential expansion property (2.1b). Thus, M achieves its maximal rank by (1.9). Moreover, the strong inclination propey also implies DnQ(0,0) is a diffeomorphism, thus the truncated square submatrix Mi achieves its maximal rank d- 2 as well. 664 o DENG Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php LEMMA 2.4. Let Mi be the same as in Lemma 2.3 above with e Ni with f--(0, 1,’’’, O)W, (2.2a) (0, 0, 0) and LDnQi(O, 0, 0) f (d-1)(d-1) (0, 0,’’’, 1). Then there is a constant mo so that lirn 6o->0 [det Mil> > 0 mo 0 and (2.2b) lim det N/= 0. 6o->0 o - 0 and the strong inclination Proof (2.2a) is true because of p(0, 0)/6-* e, as property (1.5) and (1.9). Since (0,f) is contained in T,jW.u, (2.2b) is also true for the [3 same reason. 3. Entrance and exit sets and their extensions. In this section we only consider the heteroclinic connections from a2 to al. Analogous analysis and result can be immediately extended to the a to a2 connections. Again, the parameter ce is suppressed from the text if no confusions arise. Let Ex := Wloc fq Z denote the intersection of the local unstable manifold of as with the exit cross section Z given in the previous section. is referred to as the initial exit set of W. It is obvious that there is a heteroclinic orbit (from a2 to a) if and only if there is a solution of the initial exit set which also lies on the local stable manifold of the other equilibrium a. Thus, we need to closely follow the images of the initial exit set under those successive local and global Poincar6 maps. To be precise, if we set the image of an empty set under a given map to be empty, then all the following sets are well defined: Ex En2k := H2(Exk) Ex/k:=Hi(En/k0p(A)), i--1,2, k=l,2,..., En k1: II(Ex 2k- ), r where p" Ais the Sil’nikov change of variables and p(Ai) is contained in the domain of Hi. En/k and Exk are referred to as the kth entrance set and the kth exit set of W near a, respectively. They might be empty except for the initial exit set Ex2 and the first entrance set En near a. Nevertheless, we have the following. PROPOSITION 3.1. There exists a (k-1)-heteroclinic orbit from a to al sufficiently close to the loop F[’2 if and only if En,J-, Ex,J- for l<=j<-k-1, i-1,2, and r En k Woc . By definition, for every point (:, y) Enlk with z,J. there exist (0, r/) Ex2 and z o- (,J., y), 1 -<_j -<_ k- 1 and i-- 1, 2 from the orbit of (0, r/) such that H21(0, r/) H_ H2(zk-) (, y). Using the "pull back, we have a z, H2 H(z) z,. (s, unique r/i)A satisfying z i= Ps(r)- Thus, the following proposition is valid. ’ , PROPOSrrION 3.2. (3.2a) The kth entrance set to a, Enlk, is nonempty if and only the if following system of l (2k- 1)(d 1) equations has solutions for the l + n 1 unknown variables q, 1, y with sr/=(s{, i, TJi) Ai satisfying the con<-_ and straints si > 0 for all 2 1 <-j k- 1" 1, , n,(o, ) p() 1-I,(p’(’-’)) (, y); Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php THE BIFURCATIONS FROM A TWISTED HETEROCLINIC LOOP 665 (3.2b) The kth exit set from a l, Ex k is nonempty if and only if after replacing (, y) k with k A the same statement of (3.2a) holds true" by Pl(’l) (3.2c) The kth entrance set to a2, En2k, is nonempty if and only if the following system of 12 2k(d- 1) equations has solutions for the 12 4-n- 1 unknown variables /, with i (s ji, rl rlk :, Y srll, Ai satisfying the constraints s > 0 for all 1, 2 and 1 <-j < k: I’[21(0, y) p,(’) nl(p’()) (, y); (3.2d) The kth exit set from a2, Ex2k is nonempty if and only if after replacing (, y) by p(k) with k2 A2 the same statement of (3.2c) holds true. Solving equations (3.2a-d) is more difficult with the constraints s > 0 than without them when those maps p and p’ are considered as the extended maps on the extended domain zi introduced in the previous section. Let us now study the extended equations and leave the consideration of the constraints to the next section. Note that each system of the extended equations (3.2a-d) might locally define an (n 1)-dimensional manifold near the origin of the/-dimensional Euclidean space NI, where li + n- 1. In fact, we have the following. LEMMA 3.3. Suppose the equilibria are simple saddle, relatively contractive, and the heteroclinic loop is nondegenerate. Then there exists a small constant 6 > 0 independent of k such that in the 6-box B(6) of the origin in Nl each of the systems of the extended equations (3.2a-d) defines an (n 1)-dimensional manifold J/[ in B (6) which contains the origin and can be written as the graph of a C vector-valued function of the last n 1 components, i.e., either 33 (y(2,..., y() or rh, where 11 + n 1, or 1 + n 1, accordingly. Proof. We prove the lemma by the implicit function theorem for equation (3.2a) only since the other cases are identical. By using the notation II 0 (Pi, Q) p (i, Y) and p’ (X, r/i) from the last section, we see that solving equation (3.2a) is equivalent to solving the zero of the following equation: r, a,(t’, 3) o, where -l+P(O, ) YI + Q2(O, --2 d- PI(X1, ’1) (?,g)= - -Y2+Q!(X, hi) + P.(x"k-1 -y + Q2(X2k-1 X= X(s, , Y= k-1 k-1 q2 k2-1 k2-1 k-1 y(, Y(s, so{, r/{). It is obvious that the existence of the rl{) and heteroclinic loop FlU F2 (at a =0!) implies (0, 0)=0. A simple calculation yields that the ll x ll square Jacobian matrix 0/0ff at (sr, fi) (0, 0) has the following diagonal property det _--7 (0, O) Idet diag (M2, M1 ,’’’, Ml, M=)I, Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php 666 o DENG where the diagonal blocks have the forms of Lemma 2.3 with all the blocks Mi except the last M2 taking e=-Oj(0, 0, 0). Therefore, the Jacobian matrix is nonsingular. Hence, by the implicit function theorem there exists a 3 > 0 such that (’, 33) 0 defines in B(3) an (n- 1)-dimensional manifold (containing the origin at a 0) which is the graph of a C function of the variable 33. Moreover, because of the diagonal block structure it is not difficult to see that 3 can be chosen to be independent of the number 13 of the equations 11 (2k- 1)(d 1). Let P be the canonical projection from E=Itl-d-lE d-1 onto the last d-1 components. Then Lemma 3.3 implies that the projection PJ// of the manifold J/ is also the graph of a C function of the last n- 1 coordinates. Therefore, we have the following. DEFINITION 3.4. n/k:=P and lxk:=p’(P) are called the extended kth entrance set and the extended kth exit set (of W) near ai, respectively, according to whether is taken to be the manifold defined by the extended equations (3.2a-d) for the entrance sets or the exit sets in Lemma 3.3. Since all the extended entrance and exit sets exist in some small but fixed 3-box B(3) of the center points on the entrance and exit cross sections, respectively, we can of the heteroclinic loop cl (F1 F2) such easily construct a tubular neighborhood that the intersections of with the entrance and exit cross sections (E and E’) are exactly those 3-boxes. Thus, we only need to consider the real entrance and exit sets (of W) in B(3) and rename Eft k :-" Eft k B(3) for simplicity of notation, where/ n or x. Now we are read,y to compare these sets with their extensions. Because the kth extended entrance set Enk near al is a graph over the last n- 1 coordinates fi on E and the nonemptiness of its intersection with the local stable manifold W of a forces 33 0, we have proved the following. COROLLARY 3.5. If a (k-1)-heteroclinic orbit from a2 to al exists in 11, then it must be unique (for the corresponding parameter). k COROLLARY 3.6. Let =graph G, H) with G(fi) and yl)= H(fi). If dim W dim W 1, IH(33)1 <_- 3oSo/4 and the derivative Ion(fi)l <-- 1/2for all Ifil < 3, then nk O tri f if and only if 0< H(0), where 3o and So are as in Proposition 2.2. Proof Since dim W dim W 1, G(y) 30, the x-component of the center of the entrance section E T. Since the boundary point (30, Yo)=(30, Y(so, O))6Otr satisfies >- 3oSo/2 > max HI by Proposition 2.2 and our assumption, the two boundary points (30, Yo) and (30, 0) must be in different sides of k if 0< H(0). The path connectedness of implies there must be a point (30, Y(sl, 30, 0)) lnk. This shows 0< H(0) is sufficient. To show that it also necessary, suppose it false, i.e., H(0)_-<0. Since nk graph (H), then 1)= n(0)-<_0. Thus, yl)-yol<=lDHIl[<-1fi[/2 by our assumption. Let Yl 6 n/kf-)tr7 # ; then Ifill<y 1 holds true by Proposition 2.2. It [-1 follows that Ifil-yo<y])-yo)<-ll/2 and 0-< 13311/2<y01), a contradiction. Our main result of this section is as follows. THEOREM 3.7. Let lnk graph G, H) with G() and y 1 H(). If the derivak < tire of H satisfies DHI k, where for all k, i= 1, 2 and Eft is nonempty, then E k = n : yo n o- yo .t ni or Xi. Proof. Suppose it is false; then there exists a first E k such that E k k. We claim first fl x. If fl ni then there exists a point Po n/k_ Enk. Hence, there exists < 0, where and the a point (, 3) with Po P(, 3), and (, 3) has at least one s j= projection P are as in Definition 3.4. To be precise, say s =< 0. Let us first note the following: Denote /3 k according to whether it is obtained by the extended equations for the kth entrance set when/3 n or the exit set when /3 xi. Now it is not difficult to see that if p /3 k then the point q, whose components since q satisfies consist of the first (2j 1)(d 1) + n 1 components of p, belongs to x 667 Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php THE BIFURCATIONS FROM A TWISTED HETEROCLINIC LOOP the first (2j-1)(d- 1) equations (cf. Proposition 3.2). Similarly, if q is obtained by keeping the first 2j(d- 1)+ n- 1 components of p, then q is in gx. Now, resume our assumption s-<0 and let qo be such a truncated point of Po which belongs to gx. Then P(qo) /x since _-< 0. This contradicts our assumption for Hence, the claim holds true. Since we have it follows that there exist (ss, r/s) zi with fq for the and p(sr2) j 1, 2, s, =< 0, and s2 > 0 such that p(srl) same reason as above on the truncated point q. Since is path connected, being a graph over the path connected set [931 < 6, there exists a (So, sCo, r/o) 6 ,i with So 0. Hence, P(’o) (sCo, Y(0, so, r/o)) (sCo, 0) 0r f’] Wloc. That is, (sCo, 0) graph (G, H). This implies ly)l _-< I1--< 11/2 by our assumption for all (s y) fq It follows that since by Proposition 2.2 ]93] < y() for all (sc, y) This [3 contradicts p,*.(’2) En f) c ln f) r,*.. Ex n. Ex x, s n- Enn sro n o- ,r IDHI ,En o, . nn. 4. Bifurcation equations. From now on, we shall spell out the parameter explicitly wherever it is necessary. In this section we consider the constraints s{ > 0 in terms of their sign changes with the parameter, in particular the sign changes of y(1)= H(0). Here (G, H) gives the graph of the extended entrance set. For this reason, we consider the following equations" 1-I2(0, r/, c)= (, y), II21(0, r/, a)--P(’l, a), II21(P(2 a), ce)= (, y), 172,(p(sr, c), a)= p(’,, a), II12(P(, ce), c)= (sc, y), 17112(P(’1, ce), a)= p(sr2, a), (4.1a) (4.1b) (4.1c) (4.1d) (4.1c’) (4.1d’) which introduce every new s{ or y(1) into our recursive construction of the entrance and exit sets in the last section, where (&, sc, r/). Note that each of the systems above defines a system of d-1 equations with l=(d- 1) + (n- 1)+2 unknown variables for (4.1a-b) or/2- 2(d- 1)+2 variables for (4.1c-d’), including the parameters a and a2. Thus, presumably, each of them defines an (l (d 1))-dimensional manifold in R6 accordingly. Indeed, we have the following lemma. LEMMA 4.1. Suppose the conditions of Theorem A are satisfied. Then there exists a small constant 6 > 0 such that in the &box B(6) of the origin in t, each of the extended equations (4.1 a-d) defines an (l (d 1)) -dimensional differential manifold in B (6) which can be written as the graph of a C function of the last n- 1)+ 2 variables (, a) or r/, a) when li l or the first m and the last n 1) + 2 variables si, i, a) or (&, r/, a) when li 12, accordingly. Moreover, up to only one nonsingular and differentiable change of the parameter for all the equations (4.1a-d’) considered, the following bifurcation equations are satisfied for solutions to (4.1a-d’) with the corresponding , , alphabetical order: (4.2a) (4.2b) (4.2c) (4.2d) (4.2c’) m,y (1)+ l + m,y (")= a+ Og2 -II- mly (I)+ q-" m.y (")= a2 + ’r2(a )s +u2 + 0(([/:721 -iI- [$21 <)Is=l l/t"2 + 13312), )S12 +v2 -’1- O((l2l -’[- ]S2I ff2)lS2] 1+"2 + (In,I + ISll t/0/lY (’>+’’" +tfi,Y (" O1-71(O)S1l+Vl /O((I,I/Isll 1 )IS, 1+ "’/11 2) S1 2 "It- T2( 668 BO DENG Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php (4.2d’) where vi, f’i are as in Proposition 2.1. Here mi mi(t) are differentiable functions of a satisfying 1/(26o) < ml <2/60, mi/ml o(1) as 6o-O fori # 1 and the analogousproperties also hold for ri. Moreover, the scalar functions, ’i % (a), called twist functions, are nonzero, differentiable, and satisfy (4.3) ’i(0) <0 >0 if F is twisted, otherwise. Furthermore, the change of the parameters leaves the parameter axes, as sets, invariant, but may reverse their directions. The basic framework for the proof of this lemma, in particular, the derivation of the bifurcation equations through a modified Lyapunov-Schmidt reduction, has much in common with the spirit of Chow, Deng, and Terman (1990). Thus, we will prove it in the Appendix with necessary modifications given to the twist terms ’(a)s +’ and the order estimates on the higher-order terms. The following corollaries concern the conditions of Corollary 3.6 and Theorem 3.7 when the parameter is taken into consideration. COROLLARY 4.2. Let n/k= graph (G(., a), H(., a)) with G(;, a) and y(= H(, a). Then 6o and 6 can be chosen sufficiently small but fixed such that IH(, a)l <= 6oSo/4 and IDH(f, a)[<=1/2 for all ]1, ]al<6 and k>=l, where D is the differentiation operator in y. Proof Using (4.2c) or (4.2c’), we have = [max IDHI t.[maxil lY<’>[ <= and (6)]/m1() + 0(6%)]/m1(o)=o(1)+26o0(6"), lmi(a)’6 / Imi(a)l (1)6 / where Vo min {/21, /’*2, 1, 2}" Choosing 6o and 6 so small but fixed implies the desired estimates. COROLLARY 4.3. 6 can be chosen small so that if ln/k= graph (G(., a), H(., a)) crosses the stable manifold Wloc f’l E, then it does so transversely in aj in the sense that 0H(0, a)/0aj>0 for I1<, In other words, if H(O, a)=0 then for the fixed ith component o a i, H(O, ce > 0 if and only if aj > o0 Proof. Using (4.2c) or (4.2c’) again, we have ml(a)H(0, a)=% + o(11/o) with 2}. Thus Vo min { v, v2, , ( [ Om----A 0H(0, c)/oc_-> 1 H(O, a) < by an appropriately chosen small COROLLARY 4.4. The first entrance set Enll to al intersects the domain of the local map nonempty if and only if 02 > O. Proof By (4.2b), s=2/o(InlllSll/lsl/,), implying s>0 if and only for cr if a>0. 5. Proof of Theorem A. As we mentioned earlier, Theorem A has been proved in Chow, Deng, and Terman (1990), except for the directions of homoclinic bifurcations and the k-heteroclinic orbits. Thus, we are going to outline the proof from that paper and provide the necessary details for the other part of the proof. 669 Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php THE BIFURCATIONS FROM A TWISTED HETEROCLINIC LOOP Consider the Poincar6 map II1 rI21 II2 I-[2 II from a subset of 0" into the entrance section E near al. Using the Sil’nikov changes of variables, we can similarly reduce the problem of finding periodic points of II1 into solving a system of equations for the unknown Sil’nikov variables with the constraints si > 0. The conditions of the nondegenerate heteroclinic loop and the relative contraction for the simple saddle equilibria imply that the extended system has a unique solution parametrized by a by the implicit function theorem. This uniqueness allows us to consider the simple homoclinic, periodic orbits only. Thus, by the implicit function theorem, we solve (’*, ’2*)(a)= (s*, :*, 7*, s*, :*, /*)(a) as the solution for the extended equations I-I2(p’(’l, a), a)= p(’2, a) and Hl(p’(’, a), a)= p(sr, a), where ri (s, :, /i) into the bifurcation equation (4.2d’) with I’*(a)[ O([a[). Now, substituting sr* and and (4.2d), we have s*2= a +O(la[ +o) and s*=a+O(la[+). It follows that the map a- (s*, s2*) is a diffeomorphism. Thus, the sector A for the periodic orbits is given by s* > 0 and s2* > 0 and the curve hom for 1-homoclinic orbits from a to a is given as a piece of the boundary 0A with s*- 0 but s2* > 0. Substituting s* 0 and s2* > 0 into (4.2d’) and (4.2d), again we have s*2 a + O([a[)[s*[ +o, implying al >0, ’* and o + ()*’+ + o([ [)[*1 ’+, az=hom (a)=[-Zz(a)+O(lal)]]s*2[ +2. Therefore, (.) the bifurcation direcimplying tions in (i) hold true because of (4.3). Finally, to complete the proof, we only need to prove (ii). Let us consider the 3(d- 1) equations 1-I2(0, rt, a)= p(sr, a), II,(p’(’, a), a) p(’2, a), and 1-I2,(p’(’2, a)) including the two parameters al and following 3(d 1) 1 equations first: a2, (:, 0) for 3(d-1)+l variables, and assume F2 is twisted first. We solve the - -, + P2(0, ’q, a) --rl+ Q2(0, ’q,a) P(X1, r/l, a) =0, (, )= --2 + Y2 QI(X,, "rl,, (5.2) 02 with where "- (Q(), - -+ n(x, n, ) , Q(f)), and solve the leftover equation Q(I)(X2, (r/, s, :, , s, :, r/, r/:, a)=0 later, :). The existence of the heteroclinic loop implies (0, 0)=0 and a simple calculation shows the Jacobian square matrix satisfies det (0, 0) [det diag (M2, M,, M)[, where Mi and M are the same as in Lemma 2.3 with e being (0, 0) in M. Hence, it follows from Lemma 2.3 that the Jacobian 0/0sr(0, 0) is nonsingular and sr can be solved as a C function sr* of a satisfying ’*(0)=0 by the implicit function theorem. Substituting sr=sr*(a) into the remaining equation Q(I)(X*, r/*, a)=0, we find it equivalent to solving a from (5.3) s*=P2(X2* O= Q:(X* ’0* a), n* a). Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php 670 o Dn6 Notice that this equation has the form of the connecting equation (4.2c) and the corresponding bifurcation equation (4.2c) of Lemma 4.1 applies Thus, it is equivalent to Since s2*, o(lal), this equation always has a unique solution a2: 1-het (al) for every a by the implicit function theorem Since F2 is twisted, then r:(a)< 0. This implies 1-het>0. Also 1-het= O(lall’/). To see if the constraint s*(a)>0 and s*(a) > 0 are satisfied at a 1-het we need to consider the other two bifurcation equations (4.2b) and (4.2d’) corresponding to the first two connections: *= (5.5a) S*l o=+ o(In* lls* l+ls* l/,), From (5.5a) and a2 1-het (a)> 0 it is obvious to see that s* >0 is automatically satisfied. Moreover, s* O(a2)= O([a11+2). Substituting this order for s* into (5.5b) yields s2* > 0 if and only if a > 0 since a is the leading term in the right-hand side when s* and a are of order O(la,l+). Let 1-het2 := 1-het I,>o be the desired curve To show the nonexistence of k-heteroclinic orbits for k>-1 under the nontwist assumption for F2, let us solve a system of equations similar to (5.2). Analogously, it is equivalent to solving those a from the following bifurcation equations so that j* si (a) > 0 for all i,j: 1" Ol1-1"- ’,Is,1" S2 (5.6) 0 = / 1+ ’+ o((l’,*l+lsl*ll)lsll*ll+l+(lnJ*l+ls’*lOls*l) =ls*l /= / o((1*1 / Is*l =)lsg*l /=). Since r > 0, the last equation implies a2 < 0. This forces This completes the proof. s1" < 0 from the first equation. 6. Proof of Theorem B. To prove Theorem B we need the following three lemmas. When dim W 1, two given graphs graph (/-/) over 33 on the entrance set E are denoted by E =< E if H(33) <- H2(33), or E1 < E2 if H(33) < H2(33) for all common y. A point p=(6o, y)E is said to satisfy p-<_ (<)E graph (H) if y()-< (<)H(33). In what follows, let be the corresponding first entrance set of W’ocf3 E’ and W to a. Their definitions are analogous to Ex2 and End, respectively. Now we have Lemma 6.1 (cf. Fig. 6.1). to ai LEMMA 6.1. If dim W1--dim W 1 and all the entrance sets Fni and up to a given number k >-j are nonempty graphs over the fi-axis, then E Fx: Fn En (6.1a) If F: (6.1b) If both F1 Proof. is twisted but F1 is not, then En < En/< Fnl < Enl for 3 <=j <- k; Fe are twisted, then Fn < Enk < En ki-1 < and < En < En. The proof is based on the following two simple observations: (1) The range o-’ of the local map Hi contains all the exit sets and lies to one side of the corresponding initial exit graph (FXl or Ex). The images of the exit sets under the global map Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php THE BIFURCATIONS FROM A TWISTED HETEROCLINIC LOOP 671 lie below (above), i.e., <(>), the corresponding first entrance graph to the other equilibrium aj if the connection Fi from ai to aj is twisted (nontwisted); (2) For a given pair of entrance graphs near a given equilibrium ai ordered by <, the ordering (<) for the consecutive entrance graphs near the other equilibrium aj is (not) to be reversed if the connection F from a to aj is (not) twisted. To show (6.1(a)), we have Exc G, thus En+l < for all j_-> 1 by (1). By (2), the single twist and En{ < End1 imply En2 < En] +1 for all j -> 2. Since < En1/2 for j => 1 +1 < Fn2 by (2) (see Fig. 6.1(a)). To show (6.1(b)), we have by (1), we have < +1 + for all j>_- 1 by (1). The twist of F2 and and < < Fnl imply Fnl < 2 for all j>_-0 by (2). The double twists and En < En imply En] < En2 by (2). Last, a +1 for all j-> 1 (see Fig. 6.1(b)). < simple inductive argument shows LEMMA 6.2. If dim W dim W 1 and [a2[ < al, then the kth exit set Exk from a is nonempty if all the previous k- exit sets from a are nonempty and the kth extended entrance set near a has nonempty intersection with the domain tr of the local map H exit set Ex2k-1 from a2 is nonempty. Since Proof We first claim that the (k-1)st kExk-, the (k-1)st entrance set En2 to a2 is nonempty. By [DH(.,a)[<-1/2 of Corollary 4.2 and Theorem 3.7, we have Enk-l= =graph (G(., a), H(., a)). Using the bifurcation equation (4.2c’) we have En En En En Fn En En En Ex .nk lnzk- H(O, G) [Gl + O({o[l+%)]/ml(o) > O, (a) ,(b) FIG. 6.1. (a) F is twisted but F is not twisted; (b) Both F and F are twisted. En Fn En Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php 672 BO DENG .. since al is dominating by our assumption o > 121. It follows from of Corollary 4.2 and Corollary 3.6 that Hence, n2-1 o- Ex2-I II2(En2- IH(Y, c)l -< 6oSo/4 r) # En The claim is proved. Obviously, this claim implies the kth entrance set to al is by Corollary 4.2 and Theorem 3.7. Thus, the nonempty and equal to its extension 13 condition nfq # implies the kth exit set near al is nonempty. Combining Corollary 3.6, Theorem 3.7, and Corollaries 4.2, 4.3 above we have the following important result. LEMMA 6.3. If dim W dim W 1 and Pi Wiloc(a)f"l Ei ,nki at some a o o nifqr if and a (a o1, a) with k >= 1, then for the fixed ith component ai ai, is nonempty then the same statement where < Moreover, if only if a > holds true for En/. Proof Let o graph (H(., a)). Corollary 4.3 implies H(0, a) > 0 for ai a if and only if aj > aj. Since Corollary 4.2 implies the conditions < and ]HI _-< oSo/4 of Corollary 3.6, we conclude from Corollary 3.6 that f3 r7 # for a a if and By the condition DHI < of Corollary 4.2 and Theorem 3.7, we have only if r a, n Ex En n n => 7. IDHI . En/ lni Proof of Theorem B. (i) Corollary 4.2 implies the conditions of Corollary 3.6. The Hence, by nonemptiness of the kth entrance set En k with k >= 3 implies En2 f3 o-1 Corollary 3.6, 0<H(0), where En2=graph(H), implying the center point Wloc(a) f3E lies below the second entrance set En2. It follows from (6.1a) that the existence of k-heteroclinic orbits from a2 to al is impossible for k-> 2. This together with Theorem A(ii) proves the uniqueness of the 1-heteroclinic orbit from a2 to Let 1-het2 be as in Theorem A. To show 1-het2 (a)> homl (a) for al>0, we notice that Fn11 is nonempty in the spirit of Corollary 4.4. By (6.1a) we have Enl2 < Fn En 1. It follows that with a2 increasing, the homoclinic connection Pl Fnll takes place before the single heteroclinic connection Pl Enl2 does by the transverse crossing property of Corollary 4.3 and Lemma 6.3. (ii) First we zoom in the region of the parameter where (k 1)-heteroclinic orbits (from a2 to al) can take place. We first claim there exists a (k-1)-heteroclinic orbit only if 0<=ce2<hOml (tel) with al>0, where homl is the bifurcation curve for the 1-homoclinic orbit at a l. If a < 0 then En2 < < p2 for all k >= 1 by Corollary 4.4 for k >_- 1. and the twist of F1. Thus, Enk C? o-2 by Corollary 3.6. Hence, tnlk+l If a2> hOml (al) then Pl < Fnl by Corollary 4.3 By Lemma 6 1 Pl < Fnl < Enlk for all k_-> 1. This proves our claim. The existence of these heteroclinic bifurcation curves (k-1)-het2 in the region 0--<aa<homl (al) is an immediate consequence of the following claim: in 0_-<a2 <k homl (al) there exists a unique a2 (k- 1)-het2 (al) for every k _-> 1 such that Pl < Enl if and only if a2> (k- 1)-het2 (al). We proceed by induction. When k 1, it is trivial by the existence of the primary heteroclinic orbits and the transverse crossing property of Corollary 4.4. Suppose the claim holds true for k- 1. Then, by Lemma 6.3 we have for 0/2 < (k 2)-het2 (al); hence, (k 1)-heteroclinic orbits do not exist. En -1 f-I for (k-2)-het2 (al) % a2hom Again, by Lemma 6.2 and 6.3 we have Enk cel. On the other hand, Corollary 4.4 implies Fn2 f’l or2 # G5 and thus Fn] # G5 for these parameters. The transverse crossing properties of Lemma 6.3 and Corollary 3.6 imply It follows from the conFn l<pl for a2<homl (al). Lemma 6.1 implies Fnl < tinuity of En] on the parameter that when 0<a-(k-2)-het2(al)<< 1, Enk-1 is sufficiently close to the stable manifold W, and Ex k-ll is close to FXl Therefore, holds. Since Pl Fnl < Enlk at a2 hOml (al), there must be an Fnl < Fn r En. . Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php THE BIFURCATIONS FROM A TWISTED HETEROCLINIC LOOP 673 such that Pl E Enlk at (a, ak-). Because Lemma 6.3 implies this crossing is transverse, k-1 Let ak- := (k- 1)-bet2 (a). This completes the claim. Pl < Enkl if and only if a2> a Furthermore, by the property of transverse crossing, the function (k-1)-bet2 is also differentiable by the implicit function theorem. Finally, to show the inaccessibility of the homoclinic bifurcation curve hom from below, we suppose to the contrary that (k- 1)-het2 (a l)- <hOml (eel) for some a for all k. Let limk_, En k E Then E ->_ p > Fn at Then, at a o (a a ) 2 En k 0 a Thus, Pl E because otherwise < Enk <p for sufficiently large k for En kl would be sufficiently close to the stable manifold W and tnk/l would be empty. Therefore, Pl < E But, in this case, by moving a2 down a little, i.e., 0< a-a2 << 1, p < E=< Enk would still hold true for all k and a =(a a2). Thus, there would exist a k such that a2<(k-1)-het2(al). This would imply Enk<p by our second claim [3 above. This is a contradiction. ce o, . Fn . . , 7. Remarks. (a) It seems that the bifurcation equations (5.6) are solvable for and z2 are all negative. Incidentally, by neglecting the higher-order terms, the truncated recursive formulas do give rise to a monotone increasing set s_i* for i=l,...,k-1 with se1" =a+ra1+,.’>0 and amonotone decreasing set si* for i= 1,..., k with slk* := a+ z2[sk-I*I+2=0. Furthermore, the (k1)-het curve is then defined by the recursive formulas. Unfortunately, this argument fails when those fuzzy error terms are taken into consideration. A similar situation of our losing control over the full system appears in the homoclinic bifurcations with resonant principal eigenvalues (i.e., h(0)-/Zl(0) in our notation) studied by Chow, Deng, and Fiedler (1990). This is the reason we impose the condition dim Ws= m 1 and approach our problem topologically by considering the entrance and exit sets and their extensions. We feel that this restriction may not be merely technical, since without it the position of Fnl relative to Enk may behave in an unpredictable way. (b) It can be easily seen from the bifurcation equation (5.1) that the asymptotic tangency of the homoclinic curve homi is completely determined by the sign hj +/xj for #j. That is, homi is asymptotically tangent to the ai-axis at a -0 for a relatively contractive aj, or the a-axis for a relatively repelling ay, or tangent to none of them for an a with principal resonant eigenvalues, i.e., hj +/x =0 at a =0. In all cases, however, k-heteroclinic orbits are expected to bifurcate at a twisted heteroclinic loop. In particular, as in the homoclinic doubling bifurcation for a twisted homoclinic orbit (see, e.g., Chow, Deng, and Fiedler (1990)), a double homoclinic bifurcation will also probably take place at a single twisted loop with the resonant eigenvalues, i.e., (1 + q)(1 + ,)- 1 at ce-0, or in the case where the heteroclinic loop is degenerate (see Yanagida (1986)). Relaxing the equal dimensionality dim W dim W assumption will also lead to countable k-heteroclinic connections which, in contrast to Theorem B, take place in an open set of the parameter space. This was observed by Deng (1989). Also, as the principal eigenvalue (either stable or unstable) becomes a pair of complex, the system itself at a --0 becomes rather chaotically complicated (see Tresser (1984) for the case where dim W dim W, and Bykov (1980) where dim W dim W). (c) The heteroclinic loop gives us another new bifurcation point which an oriented homoclinic path can hit globally in the parameter space (cf. Fig. 7.1). The orientation of a given homoclinic path is determined by the nonzero orbit index of the periodic orbits nearby. See Mallet-Paret and Yorke (1982), Fiedler (1985), and Chow, Deng, and Fiedler (1990) for more details on the orientation relative to the orbit index. Let us suppose the homoclinic paths hom and hom_ are oriented as shown in the figure. s j* (a)> 0 if the twist functions ’ 674 o DENG Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php k-het, / hom i The twisted heteroclinic FIG. 7.1. is the orbit index. loop_O-heq If it is assumed to be k-het O-het then the region A is on the right of homi curves. W -yw+O=O =f(v) FIG. 7.2 Following hom,, it will hit and terminate at the heteroclinic loop bifurcation point a 0. But, right at this point, the homoclinic orbit to a, trades itself to a homoclinic orbit to a2 and the curve hom2 arises from a 0 as if it is the continuation of hom. For this reason, we may also call our heteroclinic loop bifurcation the homoclinic trading bifurcation. However, if we "follow" (actually we do not know how at this moment) a heteroclinic path in the double twisted case, we will find doubly infinite heteroclinic trading partners at the homoclinic trading place a =0. Certainly, this immediately complicates any "global heteroclinic path following" attempt. But it also gives us one more hope that a global homoclinic path following result seems on its way (see a detailed discussion from Chow, Deng, and Fiedler (1990)). (d) While writing these remarks, I received a preprint by Kokubu, Nishiura, and Oka (1988). I found that our notion of twisted heteroclinic loop has been propagating faster than I could finish writing this paper. Their work demonstrates that the nondegenerate conditions (1.4), (1.5), (1.7), and (1.10a, b) are verifiable. Indeed, motivated by the idea for the Mel’nikov function and the method of singular limit eigenvalue problem developed by Nishiura and Fujii (1987) and Nishiura (1989), they derive not only an analogous function to detect the transverse crossing of the stable and unstable manifolds (also see, e.g., Kokubu (1988)), but also a computable twist function to detect the strong inclination property and the twist of a given heteroclinic orbit at the same time for the system of ODE for the traveling waves, (v, w)(x, t) (v, w)(x + ct). The reaction diffusion system they consider is as follows: ev, e2 v + f v w, (7.1) Wxx -Jl- I) llW -Jl" O, where f =-v3+ v. Starting at a standing front wave and a standing back wave which forms a heteroclinic loop (i.e., at c 0), they manage to obtain the local codimensionthree bifurcation unfoldings with c, O, y being the relevant parameters and globally W Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php THE BIFURCATIONS FROM A TWISTED HETEROCLINIC LOOP 675 extend the local bifurcation diagram. Among the most interesting is the nontwistedness of all the resulting heteroclinic loops involved due to various symmetries exhibited by the system. According to Theorem A, this implies that there are no multiple heteroclinic connections other than the persistence of the zero-heteroclinic orbits from the loops. They also show that some of the finite connections are actually unstable. Perhaps some comparisons between their system and the FitzHugh-Nagumo equation considered in the Introduction are worthwhile. First of all they model systems of different worlds--chemical reactions, predator and prey populations for the former while nerve impulses for the latter. Theoretically speaking, however, they are the same system but at different values of the diffusion parameter for the w dynamics. Indeed, if we move the origin to the left equilibrium state in Fig. 7.2, the parameter 0 is the same as the parameter a in the FitzHugh-Nagumo equation (1.11). Rescaling the time and the space variables in (7.1) yields wt 6Wx + e(v yw), VXX +f(v) w, where we renamed := z/e and e := ez. Thus, it is the same FitzHugh-Nagumo equation except for a large diffusion coefficient 6 for w. Since both systems have the same symmetries, I think the appearance of the second diffusion simply "untwists" the twisting structure somewhere. Thus, it is natural to ask whether it happens at some 6o> 0 or just at 0 =0. Indeed, there are two types of bifurcations involved. When 6o 0 the system of the ODE is singularly perturbed. When > 0, however, to untwist a heteroclinic loop a heteroclinic orbit must be degenerate in general and, in particular, it must violate the strong inclination property. None of these bifurcations problems has ever been fully investigated. Nevertheless, the idea developed in this paper offers more hope for solving the bifurcation of twists than the problem of singular perturbation. V o Appendix. Proof of Lemma 4.1. The proof for the first half part of the lemma is identical to the proof of Lemma 3.3 in 3. Thus, we omit it here. Using the notation from 2, we write equations (4.1a-4.1d) in the following equivalent forms: (L z, )=0, where with the subindex a, b, c, d in correspondence with the alphabets in the equation numbers. Here (r/2, :1) and z varies with equations as follows. (I)a(, Z, 0)-- " Q2(O, r/, - z--yl, , g, , )(C, z, )= (e(x_, ,(’, z, (o, = (Sl, 1) 0(x, n, )-y a(, z, a) , Q(X2, , a)- Y, Y(s, r12, z z , (s, :, y), (s, , s, rl), where X X2(s, rl, a) and Y rl, a). Delete the rnth component of and let =((a), (m-),(m+), (a-)7". We solve c=0 first by the implicit function theorem for r=sr*(z, a) and then solve the reduced remaining equation (") (’*, z, a) 0 later. Because of the existence of the heteroclinic loop we have (0, 0, 0) 0. Moreover, the square Jacobian 0/0sr(0, 0, 0)- M is nonsingular by Lemma 2.3. (Note that this Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php 676 BO DENG also implies 002/0T/(0 0, 0) nonsingular.) Hence, by the implicit function theorem there exists a differentiable function of Izl, [a[ < 6 satisfying st*(0, 0) 0 and [sr*[ < 6 such that (’, z, a)=0 if and only if sr*(z, a). To solve the reduced equation (m)(sr*, z, )=0, we need some important facts about ’*. Since usually are not the same for different indexes a, b, e, or d, we denote it by ’*, accordingly. Since and do not depend on the first y-component yl), and are functions of )1 and a only. Moreover, since X2(0, 2, r/, a)=0 and Y(0, s, r/l, a)=0 by the 0 property (2.1a) of exponential expansions, it is easy to see that when equations for all are restricted to sl =0, s2=0, and =0 (whichever applies) they are all reduced to the same equations as follows" ’* a ’= c sr. ’* ’* r/, )--1 =0, (r/2, ,). Q2(0, r/2, a) 0, Thus, the solution depends only on a. Therefore, the following functions of restrictions " (A.0) " sr*[;,=o, ’*[s,=o, sr*]s2=0,;,=o, and ’*[,=s2=o are in fact equal to the same function of a, say (u, v)(a), which is the solution to (A.0). Note also that [’*- (u, v)r[ O([;l[ +[s2[) or O([sl[ + [s2[), accordingly. As another preparation, we need the following procedures one way or another. Expand (P, Qz)(X, rt, a) at (x, rt)= (0, u), (P)(x, ( P2) + a)( ) x 8(P2, Q2) (O, u, Q2, Q o(x, n) "q u where (P2, O2)= (P2, Q2)(0, u, a). Expand q(s r/, a) at (:, r/)= (0, u) and q,(s at (s r/) (v, 0), respectively: (A.2) X(s2, 2, T]2, og) (2as+v2-[- O[(l2[[-lT]2- ul)ls211+v2[-]s211+v2+’2], Y(s1, 1, T/l, )-- ,aS1 + O[(ISl- vl / Ir/ll)lXll / ISll’+v’], (A.3) where o2 o(O, u, a) and /la (V, O, 1). Let (A.1) rt, a) , , , r/, a) oP (o, u, oQ2 (0, u, a) (d-1)x(d --2) and P (o, u, a) -I M2(o IOn (O, u, ) 0 -- a and Lemma 2.3 we may assume that, without loss of generality, Lz(a) has the maximal rank d-2 which is attained by the submatrix //(a). Note that this also implies 0O2/0r/(0, u, a) nonsingular. Also, up to rn-1 permutations we may still call Then, by the continuity of u on (= (’ and we have 4, .-(.-1 ]det M2(a)[ > mo6o for [a] < 6 by Lemma 2.3 and Lemma 2.4. 677 Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php THE BIFURCATIONS FROM A TWISTED HETEROCLINIC LOOP (0,. Now we are ready to solve (’)(’*, z, a)=O. Since (m), 0," ", O)(Sr*, Z, a). This implies ., det [L:(a), O(sr*, z, a)] (’*, z, a)-=O, (’*, z, a)= (-1) d-’)+" det /l:(a)om)( *, z, a). Hence, (I)(’(sr*, z, a)=0 is equivalent to det [L2(a), O(’*, z, a)] =0. (A.4) Since the simplifications for these equations are all identical we will only treat two a and c here, with emphasis on how the nonsingular change of typical cases parameters and the functions mi and ’i are obtained for all the bifurcation equations. When a, substituting (x, r/) (0, r/z*) into the Taylor expansion (A.1) and using [r/- U] 0(I;11) we have (*,z,)= Q + 0(P, Q) Or/ (0, u,,)(n2m-u)+O(I.,12) yl Since the second and the fourth terms all belong to the range of L2(a), they will disappear in (A.4). This implies det [L2(a), (y0)]=det[L2 (P2 2). Dividing this equation by det M2(a) and expressing the left-hand side in terms of a homogeneous linear combination in y]i) we have m,(a)yl(1)qt- m,(a)y] ") c2(a, where mi [ ( ) ] /det O det L2( a )’ M2( a and (A.5) c2 det L2(a), Q2 det M2(a), and ei En has zero components except for the ith component of 1. This has the form of (4.2a). Let us show that the functions rni satisfy the required properties and postpone the discussion of c2 until later. Since (0, q,j(0, 0))/60-*(0, el) as 60-*0 by the exponential expansion property (2.1b), det M2(a) is approximately the product of 6o and the numerator for ml as 6o-* 0. Hence, for small but fixed 60 we have 1/26o < ml < 2/60. Also, it follows from Lemma 2.4 that mi_det[L2(a),|O][/det[L2(a),|ol[r[\]/[[\’} =o(1) ml k \ / J/ Before we check the properties for k 2, \ el / as 60-*0. let us first obtain the bifurcation equation (4.2d)o Substitute (x, r/)=(X*, r/*) with X*z=X(s, 2, r/*, a) into (A.1) and use IX2*[ O([s2[ 1+"2) and Jr/z* u[ O([sl] + Is2[). Then substitute the exponential expansion (A.2) for X*. Finally, substitute the obtained (A.1) and (A.3) with sol :* into the 678 Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php function Bo OENG d (’*, Z, a). We have ,, * z, o Q + + o o(P2Q2) (o, u, o (o, , s , )(n*- u)- (0) + 0(([2[+[$2[2)[$2[1+v2+([1[+[$1[)[$1[). Similarly, the third and fouh terms belong to the range of Lz(a ); hence, they disappear in equation (A.4). This yields ( +det [ L()’O(P’Ox Q) (0, u, )] s+ + (the same form of higher order). o we obtain the desired form Dividing both sides by det M() det L(), l)], s c+ rs + (the same form of higher order), where the Nnction c of is the same as (A.5), and l+p "=det[ L()’O(P’Ox (O,u,)p/detM(). Q) Now we show c(, 0)=0 and Oca/O(O, 0) 0 and (4.3). Recall that Q Thus, from (A.5) we obtain det 0Q/0(0, u, ) det M( C2(ffl, if2) (__l)(m_l) 0. 2 Because oQ/o(O, u, ) is nonsingular it suces to show n(l=0 when =( 0) and 0.(/0 0 at (0, 0) by the product rule of differentiation. It is trivial to check (=0 at =(1,0) because of the existence of the primary heteroclinic connections from a to a on the -axis. Also, since (P, Q)r=(p, Q)r(0 u, is on the unstable manifold W() for (0, u)e Woc() by (1.10b) for the and Q 0 we have distance between W and W on , 0 < d2(l, 2) < () ](P2, Q2)-(, 0)] ]2. min (#,0) ilo (1) This implies Q2 at (0, 2) has a constant sign for 2 > 0, say >0, since 0 < d2(O, by our assumptions. Therefore, d(0, 2) () 2 2 0< for=(0,2). )/ > 0 at Passing the limit 2 0+ above, we have 02.2 (0, 0) by (1.10b). This completes the proof for c. To show (4.3), we notice first that u(0)=0 and the set of all the column vectors and of M2(0) forms a base for TZ ox o(x, n Downloaded 01/17/16 to 129.93.180.108. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php THE BIFURCATIONS FROM A TWISTED HETEROCLINIC LOOP 679 Project this vector onto the one-dimensional linear space span {(0, $1o)r}, which is complementary to the range of L2(0); namely, span {Tp,(W’f’lE), Tp,(WocVIE)}. We obtain O(P2, Q2)/ox(O, O, 0)O2o "72(0, qlO) r + h with h range L2(0). Hence, det [L.(O), O(P2, Q2)/ox(O, O, O)q2o] and -72 r2(0) follows. Of course, z2(0)> 0 if and [-I Definition 1.1. This completes the proof. "72 det M2(O) only if 12 is not twisted by our Acknowledgments. The author is indebted to the reviewers for their many useful suggestions. He also has benefited from many conversations with S. N. Chow, J. K. Hale, and J. Mallet-Paret. Special thanks go to D. Terman, who corrected the author’s misunderstanding of his work with J. Rinzel, which was the key motivation for beginning this work. REFERENCES V. V. BYKOV, Bifurcation of dynamical systems close to systems with a separatrix contour containing a saddle-focus, Methods of qualitative theory of differential euqations, Gor’kov, Gos. University, Gorki, Soviet Union, 1980, pp. 44-72. S.-N. CHOW, B. DENG, AND B. FIEDLER, Homoclinic bifurcations at resonant eigenvalues, J. Dynamical Systems and Differential Equations, 2 (1990), pp. 177-244. S.-N. CHOW, B. DENG, AND D. TERMAN, The bifurcation of homoclinic and periodic orbits from two heteroclinic orbits, SIAM J. Math. Anal., 21 (1990), pp. 179-204. B. DENG, The Sil’nikov problem, exponential expansion, strong A-lemma, C l-linearization and homoclinic bifurcation, J. Differential Equations, 79 (1989), pp. 189-231. Exponential expansion with principal eigenvalues, preprint, 1988. J. W. EVANS, Nerve axon equations. III. stability of the nerve impulse, Indiana Univ. Math. J., 22 (1972), pp. 577-594. J. W. EVANS, N. FENICHEL, AND J. A. FEROE, Double impulse solutions in nerve axon equations, SIAM J. Appl. Math., 42 (1982), pp. 219-234. B. FIEDLER, An index for global Hopf bifurcation in parabolic systems, J. Reine Angew. Math., 359 (1985), pp. 1-36. S. P. HASTINGS, Single and multiple pulse waves for the FitzHugh-Nagumo equations, SIAM J. Appl. Math., 42 (1982), pp. 247-260. M. W. HIRSCH, Differential Topology, Springer-Verlag, New York, 1976. K. R. T. JONES, Stability of the traveling wave solution of the FitzHugh-Nagumo system, Trans. Amer. Math. Soc., 286 (1984), pp. 439-469. H. KOKUBU, Homoclinic and heteroclinic bifurcation ofvectorfields, Japan J. Appl. Math., 5 (1988), pp. 455-501. H. KOKUBU, Y. NISHIURA, AND H. OKA, Heteroclinic and homoclinic bifurcation in bistable reaction diffusion systems, preprint KSU/ICS 88-08, 1988. J. MALLET-PARET AND J. A. YORKE, Snakes, oriented families of periodic orbits, their sources, sinks, and continuation, J. Differential Equations 43 (1982), pp. 419-450. Y. NISHIURA, Singular limit approach to stability and bifurcation for bistable reaction diffusion systems, to appear in Proc. Workshop on Nonlinear PDE’s, March 1987, Provo, Utah, P. Bates and P. Fife, eds., Springer-Verlag, New York, 1989. Y. NISHIURA AND H. FUJII, Stability ofsingularly perturbed solutions to systems of reaction-diffusion equations, SIAM J. Math. Anal. 18 (1987), pp. 1726-1770. J. RINZEL AND D. TERMAN, Propagation phenomena in a bistable reaction-diffusion system, SIAM J. Appl. Math., 42 (1982), pp. 1111-1137. M. SHUB, Global Stability of Dynamical Systems, Springer-Verlag, New York, 1987. L. P. SIL’NIKOV, The existence of a countable set of periodic motions in the neighborhood of a homoclinic curve, Soviet Math. Dokl., 8 (1967), pp. 102-106. C. TRESSER, About some theorems by L. P. Sil’nikov, Ann. Inst. H. Poincar6, 40 (1984), pp. 440-461. E. YANAGIDA, Branching of double pulse solutions from single pulse solutions in nerve axon equations, J. Differential Equations, 66 (1986), pp. 243-262.