Hopf-Like Bifurcations in Planar Piecewise Linear Systems
Abstract
Continuous planar piecewise linear systems with two linear zones are considered. Due to their low differentiability specific techniques of analysis must be developed. Several bifurcations giving rise to limit cycles are pointed out.
Full text
Publicacions Matem`atiques, Vol 41 (1997), 135–148. HOPF-LIKE BIFURCATIONS IN PLANAR PIECEWISE LINEAR SYSTEMS Emilio Freire, Enrique Ponce and Francisco Torres Abstract Continuous planar piecewise linear systems with two linear zones are considered. Due to their low differentiability specific techniques of analysis must be developed. Several bifurcations giving rise to limit cycles are pointed out. 1. Introduction and preliminary results We are interested in nonlinear planar oscillators of the form (1) ˙x=f(x)−y, ˙y=x−by, where b∈Rand the function fgives account of the nonlinearity. System (1) exhibits a structure very common in the oscillator’s field: the x-dynamics could be viewed, under an adequate selection of f(x), as a bistable system with a negative feedback represented by the y-state, which in turn has a linear first order dynamics. Examples of systems modelled by equation (1) appear in different domains: van der Pol- Duffing oscillators in mechanics and electronics, Fitzhugh [1]-Nagumo [2] equations of nerve conduction, Boissonade-De Kepper [3] equations of chemical oscillating reactions, among others. In the study of such oscillators, one finds some situations where the nonlinearities involved are correctly modelled by piecewise linear functions with three linear pieces [4], [5]. To understand the corresponding dynamics, it is important to first consider the two linear pieces case, because some phenomena encountered in this configuration are inheritated by the more general case of three pieces [6]. Thus, we are naturally led to consider continuous planar piecewise linear systems with two linear zones, hereafter denoted 2CPL systems.
136 E. Freire, E. Ponce, F. Torres In this situation, by means of rigid motions, we can consider without loss of generality that the boundary between the two zones coincides with the vertical axis. So we write ˙x =F(x)=ALx+cL,if x1≤0, ARx+cR,if x1>0, where x=(x1,x 2)∈R2,AL,ARare 2 ×2 matrices and cL,cR∈ R2. Here the dot denotes derivatives with respect to τ. The continuity requirement implies that cL=cR=cand the matrices AL,ARmust have the same second column, i.e. our systems become (2) ˙x =F(x)=ALx+c,if x1≤0, ARx+c,if x1>0, AL=l11 a12 l21 a22 , AR=r11 a12 r21 a22 . By using scalar equations, we have equivalently (3) ˙x1=ϕ1(x1)+a12x2+c1, ˙x2=ϕ2(x1)+a22x2+c2, with ϕi(x1)=li1x1,if x1≤0, ri1x1,if x1>0,and i=1,2. For system (2), we are mainly interested in the bifurcations giving rise to limit cycles. If we define the parameters tL= trace(AL) and tR= trace(AR) and denote by int(.) the interior region of a closed Jordan curve in the plane, then a first result concerning periodic orbits is the following. It can be easily deduced from Green’s formula. Proposition 1. If Γis a closed orbit of system (2) then intL(Γ) tLdx1dx2+intR(Γ) tRdx1dx2=tL·SL+tR·SR=0, where intL(Γ) = int(Γ) ∩{x1<0},intR(Γ) = int(Γ) ∩{x1>0}, SL= area(intL(Γ)),SR= area(intR(Γ)). From Proposition 1, the following result about necessary conditions for the existence of isolated closed orbits, that is limit cycles, can be stated.
Bifurcations in piecewise linear systems 137 Proposition 2. If Γis a limit cycle of system (2) then Γlives in the two zones, and the conditions tL·tR<0and a12 =0hold. Proof: Within each zone, system (2) is linear and linear systems cannot have limit cycles, therefore Γ lives in both zones. If tL·tR>0,from Proposition 1, system (2) cannot have periodic orbits (this is an analogous of Bendixson’s criterion). As Γ lives in the two zones and so SL·SR>0, if one of the traces is zero, then the other trace value is also zero. This leads to a conservative system where no limit cycles are possible, getting a contradiction. As is well known [7], every periodic orbit in a continuous planar vector field must contain at least one equilibrium in its interior. Then, if a12 = 0, system (2) is uncoupled and the equilibrium in the interior of Γ should have an associated invariant vertical straight line, which precludes the existence of periodic orbits. We now give a canonical form to which all 2CPL systems capable of presenting limit cycles can be reduced. Proposition 3. If a 2CPL system has a limit cycle, then the system can be transformed into the following canonical form (4) ˙x=f(x)−y, ˙y=x−by −a, where f(x)=µLx, if x≤0, µRx, if x>0, and the dot now denotes derivatives respect to a new variable s. Proof: From Proposition 2, we can assume that tL=tRand a12 =0. Now, if we do in (3) a change of variables of the form u=x1, v=αx1−a12x2−c1, then the vertical boundary is preserved. In the new variables, we have (5) ˙u=ϕ1(u)+αu −v, ˙v=αϕ1(u)−a12ϕ2(u)+(α−a22)(αu −v)+a22c1−a12c2. Considering the second equation piecewise linear term, αϕ1(u)−a12ϕ2(u)=(αl11 −a12l21)u, if u≤0, (αr11 −a12r21)u, if u>0,
138 E. Freire, E. Ponce, F. Torres if we select α=l21 −r21 l11 −r11 a12 =l21 −r21 tL−tR a12, the second equation in (5) is linear. We get a system of the form (6) ˙u=g(u)−v, ˙v=ρu +σv +ξ, where g(u)=m·u, if u≤0, n·u, if u>0, and m−n=tR−tL.Ifρ≤0, a simple computation shows that the possible equilibria have real eigenvalues and, consequently, the system has invariant straight lines and no periodic orbits. Thus, we can assume ρ>0 in (6) and from the change, x=√ρu, y =v, τ =s √ρ,a=−ξ √ρ,b=−σ √ρ,µ L=m √ρ,µ R=n √ρ, the conclusion follows. We remark that canonical form (4) is, modulo a translation of the variables, a system of form (1) which constitutes the main motivation of this work. Thus, system (1) includes all the systems with two-zones piecewise linear nonlinearities which can exhibit limit cycle behaviour. In this paper, we aim to study for the canonical form (4) its Hopflike bifurcations, that is bifurcations that involve an equilibrium and a limit cycle. To do so, we introduce the parameters for the trace and the determinant in each zone, tΛ=µΛ−b, dΛ=1−bµΛ, where Λ = L, R, standing for left or right, respectively. We will assume dL,d R>0, which implies that system (4) has only one equilibrium point. Also, we will restrict ourselves to the case where DΛ=dΛ−t2 Λ 4>0,for Λ = L, R, that is, we have linear focus dynamics in both zones. This case will be named the focus-focus case. Other cases will appear elsewhere. In the next section, we include the main results about Hopf-like bifurcations of codimension one in a certain three parameter space, showing four surfaces associated to three different kind of such bifurcations. These codimension one bifurcations surfaces are arranged around four
Bifurcations in piecewise linear systems 139 codimension two curves emanating from one codimension three bifurcation point. The study of periodic oscillations in piecewise linear planar system traces back to the Andronov’s school. In fact, part of this work closely follows Andronov et al. [8], but we fill some gaps there, and we give, under the previous assumptions, the general bifurcation set for two-zones piecewise linear oscillators. 2. Hopf-like bifurcations In this section, we give for the focus-focus case the main results of the paper about the bifurcations which are associated to changes of stability of an equilibrium point along with the appearance of a limit cycle. We call these phenomena Hopf-like bifurcations because they have the qualitative ingredients of differentiable Hopf bifurcations but, as it will be seen, there arise some dissimilarities. We need to write the solutions for each zone, under our assumptions of having only one equilibrium point and linear focus dynamics in both zones. Namely, (7) x(s)−a dΛ y(s)−µΛ a dΛ =estΛ 2C(s) x(0) −a dΛ y(0) −µΛ a dΛ , where the matrix C(s)is (8) C(s)= cos(ωΛs)+(b+µΛ) sin(ωΛs) 2ωΛ−sin(ωΛs) ωΛ sin(ωΛs) ωΛ cos(ωΛs)−(b+µΛ) sin(ωΛs) 2ωΛ and ωΛ=√DΛ, for Λ = L, R, standing for left or right zone, respectively. We also introduce two parameters which are crucial for the following analysis, namely the weighted traces (9) γL=µL−b 2ωL =tL 2ωL , γR=µR−b 2ωR =tR 2ωR . If a= 0, then the equilibrium is at the origin and so in the boundary between the two linear zones. Its topological type is determined in the next proposition.
140 E. Freire, E. Ponce, F. Torres Proposition 4. Considering system (4) under the assumptions DL,D R>0,and a=0, the following statements hold. (a) If γR+γL>0, the origin is a globally repulsive nonlinear focus. (b) If γR+γL=0, the origin is a global nonlinear center. (c) If γR+γL<0, the origin is a globally atractive nonlinear focus. Proof: It suffices to compute a Poincar´e return map Pon the halfstraight line S+={x=0,y≥0}. The map Pis built by a composition of the mapping PL, which applies points of S+into points of S−= {x=0,y≤0}using the orbits in the region x<0, and the mapping PR:R+→R+, which applies points of S−into points of S+using the orbits in the region x>0, see Figure 1a. Taking in (7) and (8) the values a=0,x(0) = 0, y(0) = p, and working in the left zone, for sL=π/ωLwe obtain q=−y(sL)=PL(p)= pexp(γLπ). In an analogous way, putting a=0,x(0) = 0, y(0) = −qand working in the right zone, it turns out for sR=π/ωRthat r=PR(q)= qexp(γRπ). Hence, P(p)=e(γL+γR)πp, and statements (a)-(c) follow easily. When a= 0 in (4) the equilibrium point is located in the interior of one zone. The flow in this zone is linear, but the global flow is nonlinear and, as we will see now, a limit cycle can appear when tL·tR<0. The next result closely follows the ideas used in Andronov et al. [8], which also appeared in Ye [9]. However, in both references there are some little gaps and so we include its complete proof. r=PR(q)=P(p)y p ˆp=P−1 L(0) x q=PL(p) ϕ −2π−ππ ˆτ 2π τ (a) (b) Figure 1. (a) Poincar´e maps used through the paper. (b) Graph of function ϕfor γ>0.
Bifurcations in piecewise linear systems 141 Theorem 5. Consider system (4) with DL,D R>0,a=0, and tL·tR<0. Then there is only one equilibrium point and the following statements hold. (a) If γL+γR>0and sign(a) = sign(γL), then the equilibrium is asymptotically stable, and it is surrounded by a unique unstable limit cycle. (b) If γL+γR≤0and sign(a) = sign(γL), then the equilibrium is globally asymptotically stable and no limit cycles exist. (c) If γL+γR≥0and sign(a) = sign(γR), then the equilibrium is unstable and no limit cycles exist. (d) If γL+γR<0and sign(a) = sign(γR), then the equilibrium is unstable and it is surrounded by a unique stable limit cycle. Moreover, when one limit cycle exists, its amplitude depends linearly on the parameter “a” and its period does not depend at all on that parameter. Proof: Note that the equilibrium is in the left zone for a<0 and when a>0 it is in the right zone. We shall begin the proof by choosing a<0 and considering only assertions (a) and (b). Thus we have that γL<0. So tL<0, and the equilibrium is an asymptotically stable focus located in the left zone. To study the existence of limit cycles we now compute a Poincar´e map P, which will be obtained in parametric form by means of certain phase angles. To this end let us introduce the auxiliary function, ϕγ(τ)=1−eγτ (cos τ−γ·sin τ) with the symmetry properties ϕ−γ(−τ)=ϕγ(τ), ϕ−γ(τ)=ϕγ(−τ),∀γ,τ ∈R and the graph of Figure 1b. Note that for γ>0, the function ϕγhas a first positive zero for a certain value ˆτ=ˆτ(γ)∈(π,2π), and relative maxima at τ=−π,τ=π. Taking now (x(0),y(0)) = (0,−q), with q>0 (that is, a point of S−, as defined before) as initial point of an orbit, it is concluded that the trajectory evolves in the rigth zone until it eventually reaches the left zone. Note that the flow on S−(respectively, S+) is pointing to the right (respectively, to the left) and so the trajectory leaves the right zone in apoint(0,r)ofS+, after a time sR. From (7) and (8) for s=sRand
142 E. Freire, E. Ponce, F. Torres solving for qand r, we can write, (10) q(τR)=−aωRe−γRτRϕγR(τR) dR·sin τR r(τR)=−aωReγRτRϕ−γR(τR) dR·sin τR , where τR=ωRsR, and τR∈(0,π). When τRvaries from 0 to π, the functions q(τR) and r(τR) increase monotonically from 0 to ∞. From (10) we can implicitely define the half-return map on the rigth zone as the C∞map PR:R+→R+, with PR(q)=r, and PR(0)=0, see Figure 1a. Integrating backwards from the point (0,−q)onS−, we analogously obtain a first intersection point (0,p)∈S+after a time −sL. Again from (7) and (8) and s=−sL, we can solve for pand qand write, (11) p(τL)=aωLe−γLτLϕγL(τL) dL·sin τL q(τL)=aωLeγLτLϕ−γL(τL) dL·sin τL , where τL=ωLsL∈(π, ˆτ]. Note that ϕ−γL(ˆτ) = 0, and so q(ˆτ)=0. We will denote ˆp=p(ˆτ). When τLdecreases from ˆτto π, the function p(τL) increases monotonically from ˆpto ∞and the function q(τL) from 0 to ∞. For representing the half-return map with time reversal on the left zone, we define the C∞map P−1 L:R+→R+, where P−1 L(q)=p>ˆp and P−1 L(0) = ˆp, see Figure 1a. Elementary calculations show that the function p=P−1 L(q) has the following derivatives, (12) dp dq =ϕ−γL(τL) ϕγL(τL), d2p dq2=−2dL1+(γL)2sin3τL aωL[ϕγL(τL)]3[sinh(γLτL)−γLsin τL]. Now using that τL∈(π, ˆτ] and the properties of the function ϕγ, the function p(q) is increasing with limq→∞p(q)=e−γLπand p(q)>0,for q≥0.
Bifurcations in piecewise linear systems 143 So we have that 0 ≤p(q)<e −γLπfor q≥0 and consequently, the inverse map PLis well defined and differentiable in the interval (ˆp, ∞). The expressions of the derivatives of the function r=r(q) are identical to (12) if we change all subscripts from Rto L. So taking into account that τR∈(0,π), the function r(q) turns out to be increasing with limq→0r(q)=1,limq→∞r(q)=eγRπand r(q)>0,for q≥0. We deduce that 1 <r (q)<e γRπ. We can define the Poincar´e map r=P(p)=PR(PL(p)), where Pis defined from [ˆp, ∞)to[0,∞), with P(ˆp) = 0. Accordingly we obtain by using (10)-(12), (13) dP dp =ϕγL(τL) ϕ−γL(τL) ϕγR(τR) ϕ−γR(τR)=p P(p)e2(γLτL+γRτR)=p P(p)e2ν(q), where ν(q)=γLτL+γRτR, and τL,τRare the phase angles of half-return maps through q. Note that when p→ˆp+, then τL→ˆτand τR→0. If p→∞, then τL→πand τR→π. Therefore from (13) we have (14) limp→ˆp+ dP dp =∞, limp→∞ dP dp =e(γR+γL)π. We remark that if ¯pis a fixed point of the map P, then (13) reduces to dP dp p=¯p =e2ν(¯q),where ¯q=PL(¯p). Moreover, as the functions r(q) and p(q) verify r(0) = 0 <ˆp=p(0), by continuity we have r(q)<p(q) for q<¯q. It can be deduced that (15) 1 ≤r(¯q) p(¯q)=dP dp p=¯p =e2υ(¯q) and so ν(¯q)≥0. In our working case, the function ν(q) is strictly increasing because to a larger qcorresponds a larger τRand a smaller τL. Therefore, this function can vanish only in one fixed point, at most. As a consequence, a continuum of fixed points cannot happen and the fixed points, if any, are isolated.