Delay-induced blow-up in a planar oscillation model

In this paper we study a system of delay differential equations from the viewpoint of a finite time blow-up of the solution. We prove that the system admits a blow-up solution, no matter how small the length of the delay is. In the non-delay system every solution approaches to a stable unit circle in the plane, thus time delay induces blow-up of solutions, which we call"delay-induced blow-up"phenomenon. Furthermore, it is shown that the system has a family of infinitely many periodic solutions, while the non-delay system has only one stable limit cycle. The system studied in this paper is an example that arbitrary small delay can be responsible for a drastic change of the dynamics. We show numerical examples to illustrate our theoretical results.


Introduction
In various disciplines of the science, mathematical modelling offers a description of phenomena. In some phenomena, the history of the state, not only the current state, affects the change of the state, thus it is reasonable to consider the effect of time delay [7]. Up to now the theory of delay differential equations have been intensively developed [9,17].
In this paper we study a system of delay differential equations from the viewpoint of the blow-up solutions. Here we use the terminology "blow-up" as a finite time blow-up of the solutions, i.e., the solution diverges (in a suitable topology) in finite time. The blow-up phenomenon has been widely investigated in partial differential equations, see [2,11,19,18] and references therein. There are also extensive studies in Volterra integral equation, as an alternative formulation of partial differential equation of parabolic type with a point source term [3,18,21]. Compactification of the phase space is a method to study the blow-up solutions of ordinary differential equations [6,14]. Numerical analysis has been an unavoidable tool for understanding the blow-up phenomenon. Numerical method to compute the blow-up solutions for polynomial systems of ordinary differential equations has been proposed for ordinary differential equations with application to partial differential equations, see [11] and references therein. Recently, numerical validation for the existence of the blow-up solutions is proposed [20].
The blow-up phenomenon in delay differential equations has not been much studied, as far as we know, except for a few studies [8,13,22,23,10,1]. Perhaps the reason is that many examples of delay differential equations, which appear in population biology, control theory, etc, negative feedback condition is usually imposed which excludes blow-up solutions. One also sees that the following delay differential equation does not have a blow-up solution (at least if the initial function is continuous) for τ > 0 and is reduced to a famous example of the ordinary differential equation having a blow-up solution when τ = 0. Thus one may speculate that time delay inhibits the blow-up phenomenon in general (which is certainly not).
In the paper [8] the authors study the existence of the blow-up solutions for a class of delay differential equations. The authors are interested if adding (and multiplying) a delay term to an ordinary differential equation affects the solution behavior. Using the comparison principle, the authors obtain conditions that the delay term does not change the qualitative properties concerning the global existence and blow-up of the solution. In [23] the authors study the blow-up phenomenon of differential equations with piecewise constant arguments, in comparison with the corresponding ordinary differential equations. The blow-up phenomenon is studied in Volterra integro-differential equations. See [10,1] and also [13,22] as applications to a parabolic type partial differential equations to study of Volterra integro-differential equations.
It seems that the blow-up phenomenon that stems from the time delay has not been reported, to the best of our knowledge. Our motivations are to demonstrate whether the time delay itself induces blow-up of the solution, and to understand the mechanism of blow-up of solutions. In this paper, we propose an example model that time delay drastically changes the solution behavior and induces the blow-up solution together with a family of infinitely many periodic solutions, where most of solutions are shown to be unstable. In our equation, a blow-up solution exists no matter how small the length of the delay is, while non-delay equation does not have a blow-up solution. We thus call these phenomena "delay-induced blow-up". This paper is organized as follows. In Section 2, we introduce a planar system of delay differential equations which we study in this paper. In the absence of time delay, the system becomes a planar system of ordinary differential equations. It can be seen that every solution except the trivial solution approaches to the limit cycle, thus no solution blows up. Concerning the system of delay differential equations, we present our two main theorems that show blow-up of solutions is possible due to time delay and that the system admits infinitely many periodic solutions. In this section we use a special initial function in order to show blow-up of solutions for any τ > 0. We demonstrate numerical examples of blow-up solutions and global solutions. See Remark 3.1, Figure 1 and also Section 6. In Section 3, by a careful estimation of the solution we show that there is a blow-up solution for the system of delay differential equations and provide a proof of Theorem 1. In Section 4, we study existence of a periodic solution with constant radius and constant angular velocity in the plane. It is shown that there exist infinitely many periodic solutions which appear due to time delay. In Section 5, we study a characteristic equation which characterizes stability of the periodic solutions. It is shown that most of periodic solutions are not stable except for the only one periodic solution that is a continuation of the periodic solution of the non-delay model. In Section 6, we provide numerical examples which illustrate our theoretical results. In Section 7, we discuss our results.

A planar system of delay differential equations and main results
Let τ ≥ 0 be a parameter for time delay. In this paper we consider the following planar system of delay differential equations For the special case τ = 0, the system (1) is reduced to the following system of ordinary differential equations with a = 1: Here a ∈ R is a parameter. This system is a famous model for Hopf bifurcation at a = 0. (See [12], for instance.) From the elementary calculation, one can see that every solution of (2) except the trivial solution (x (t) , y (t)) ≡ (0, 0) tends to a periodic solution of minimal period 2π and satisfies lim t→∞ x 2 (t) + y 2 (t) = 1.
i.e., the limit cycle of (2) is the unit circle.
For the system (1) we prove that the system admits blow-up solutions due to the presence of time delay. We prove the following theorem in Section 3.
Theorem 1. For any τ > 0, there exist blow-up solutions for the planar system of delay differential equations (1).
Then in Sections 4 and 5, we further investigate the system (1) and show the existence of infinitely many periodic solutions. We also study the stability of the periodic solutions. The following theorem is proved in Sections 4 and 5.
Theorem 2. For any τ > 0, there exist infinitely many unstable periodic solutions of (1) with constant radius and angular velocity. Moreover, there exists a positive τ * such that the system (1) admits only one asymptotically stable periodic solution with constant radius and angular velocity for 0 < τ < τ * .
By the change of the variables, from (1), we obtain the following polar coordinate system Now we prove that there exists a solution such that r blows up in a finite time with θ → π 4 as r → ∞. Therefore, we can conclude that x and y blow up in a finite time. For (3), we consider the following initial condition where R > 0 and s(t) is a continuous function such that Remark 3.1. In this paper we consider a special initial condition (4) in order to show an existence of blow-up solutions for any positive τ . In this section we will prove the blow-up of solutions in t < τ for sufficiently large R > 0. For other initial data, there are no mathematical results on global existence of solutions except periodic solutions studied in Section 4 and blow-up of the solutions. In Figure 1  Since −τ ≤ t − τ < − 1 2 τ for t ∈ 0, 1 2 τ , from the initial condition (4), we have Then, for t ∈ 0, 1 2 τ , the solution of the system of delay differential equations (3) with the initial condition (4) is given by the following system of ordinary differential equations Step 1: Step 2: Lemmas 3.2 and 3.3 Step 3: Lemma 3.4 Blow-up process solution orbit with the initial condition We are going to prove that the solution of the system of ordinary differential equations (5) with the initial condition (6) blows up in t ≤ 1 2 τ . Figure 2 shows each step for blow-up of solutions. The process for blow-up of solutions is divided into the following 3 steps: Step 1: The angle of the solutions is monotonically increasing from −π/2 to −π/4 in t ∈ [0, T 1 ] for some T 1 > 0. In this region, the radius of the solutions may decrease, and thus we establish a decay estimate of the solutions in Lemma 3.2.
Step 2: Both radius and angle of the solutions monotonically increase in t ∈ [T 1 , T 2 ] for some T 2 > T 1 . The angle varies from −π/4 to 0 and the radius grows up beyond a threshold for an emergence of a nullcline of the angle.
For the exposition, we define Then the system (5) with the initial condition (6) becomes with the initial condition In the proof below we need to estimate the solution r. For the estimation we use the following equation which is obtained from (8).
Step 1: The decay estimate of the radius.
Lemma 3.2. There exists T 1 > 0 such that φ monotonically increases from π/4 to π/2 for t ∈ [0, T 1 ]. One also has Moreover, there exists R 1 > 0 such that Proof. The solution of the system of ordinary differential equation (8) with the initial condition (9) exists for sufficiently small t. We show that there exists T 1 > 0 such that the solution exists for t ∈ [0, T 1 ] satisfying that φ monotonically increases from π/4 to π/2 for 0 ≤ t ≤ T 1 . We first obtain an a priori estimate for r for φ ∈ 1 4 π, 1 2 π . From (10) it follows that One also obtains the following estimation Integrating the inequalities (12) and (13), we obtain the following estimation Using the a priori bound (14), from the equation (8b), we see that φ is an increasing function and that φ (t) ≥ 1, provided φ ∈ 1 4 π, 1 2 π . Therefore, there exists T 1 > 0 such that φ monotonically increases from π/4 to π/2 for t ∈ [0, T 1 ]. The inequality (11) holds from the estimation (14). Since, from the inequality (11), we have R exp(−π/4) ≤ r for t ∈ [0, T 1 ], the following estimation Step 2: Emergence of the nullcline of the angle. Next we have the following estimation.
In Figure 3, we plot the graph of (φ, φ ). We show that, for sufficiently large R, a nullcline for φ exists in π, 3 2 π , where the right-hand side of the φ-equation (8b) becomes 0. We see in Lemma 3.5 that φ has a upper bound for suitable large R and then r(t) blows up in a finite time. We now let R be a sufficiently large number such that R 2 exp(− π 4 ) > √ 2. Then, for r > R exp(− π 4 ), there is a φ-nullcline in π, 5 4 π that is given as It is easy to obtain the following elementary lemma.
Step 3: Blow-up of solutions. Finally we show the blow-up of solutions.
Thus we obtain the conclusion.

Existence of periodic solutions
We consider a periodic solution with a constant radius, i.e. r(t) ≡ ρ > 0, for the system (3). Then, θ (t) is also a constant from equations (3). Thus, we treat a periodic solution of the form where ω ∈ R is an angular velocity.
Lemma 4.2. For j ∈ Z \ {0, 1}, the equation (27a) with cos ωτ > 0 has exactly one root ω j on the following each interval Therefore, we obtain roots ω j for j ∈ Z if τ ≤ τ * and for j ∈ Z \ {0, 1} if τ > τ * of the equation (27a). In Figure 5(b) we plot the branches for ω as a function of τ . Once ω of the periodic solution (24) is given, ρ is determined from (27b). The radius ρ can be determined as ρ j given as, similar to (29), This implies that the periodic solutions of the form (24) has larger radius than the unit circle which is the trajectory of the periodic solution for τ = 0. We also note that there is a root ρ = 1 if τ = 2nπ, n ∈ N. In Figure 6 we plot the branches for the radius as a function of τ . Summarizing the above findings, we obtain the following result for the existence of the periodic solution (24) for the system (3). The result is not intuitive and not expected that the delay induces many periodic solutions from the non-delay system (1).
Theorem 3. For each τ > 0, the equation (25) has infinitely many roots, which are countable. Thus the system (3) has infinitely many periodic solutions, which are countable, of the form (24).

Analysis of the characteristic equation for the stability of the periodic solutions
To analyze stability of the periodic solution obtained in Section 3, we study a system of a delay differential equation and an integral equation, employing the principle of linearized stability for the coupled systems of renewal equations and delay differential equations [4].
of the system (3), given as in Section 3.
In the Appendix A, by linearization of the system (30) at the equilibrium (31), we obtain the following characteristic equation 2τ 0 e −λs ds, λ ∈ C, with ρ = ρ j , j ∈ N. Note that we have a family of the characteristic equations (32) which are indexed by the equilibrium about which we linearize. Define η := λτ . Then, from (32), we obtain the following equation where f : C → C defined as We study (33) to analyze the distribution of the roots of (32) with respect to the imaginary axis. First let us study the existence of real roots for the characteristic equation. Since f is a continuous function, if ρ 4 j τ < 1 then there is a negative real root and if ρ 4 j τ > 1 then there is a positive real root. We also see that if ρ 4 j τ = 1 then there exists a root 0 for the function f .
For the uniqueness of the root, we study the equation 0 = ηf (η). Here Let us define We consider an intersection of g 1 and g 2 . It is easy to see that g 1 (0) = g 2 (0) = τ 2 . We compute Therefore, one sees that g 1 is a downward-convex function (attaining minimum at η = log τ ) and g 2 is upward-convex function. Hence, the intersection of the functions g 1 and g 2 except 0 is unique. Thus we obtain the conclusion.
For each j ∈ N we obtain the estimation of ρ 4 j τ with respect to 1 as follows. Lemma 5.2. The following statements are true.
From Lemmas 5.1 and 5.2 and the principle of linearized stability [4], we obtain the following result concerning instability of the periodic solutions.
Theorem 4. The periodic solution of the form (24) for j = 1 is unstable for τ < τ * . The periodic solution of the form (24) for j ∈ Z \ {0, 1} is unstable.
Let us consider stability of the periodic solution of the form (24) for j = 0. For j = 0 the characteristic equation (33) may have an imaginary root with positive real part. Thus so far we cannot determine stability of the periodic solution from Lemma 5.1 and 5.2. Below we exclude this possibility to conclude that the periodic solution of the form (24) for j = 0 is asymptotically stable for τ < τ * .
First let us show that there is a compact region in the complex plane for the existence of a root with positive real part. Thus for any ε > 0 there exists δ > 0 such that τ < δ implies |η| < ε.
Therefore, from the application of the Rouche's theorem (see e.g. Lemma 2.8 in Chapter XI of [5]), we obtain the following conclusion.
Theorem 5. If τ < τ * , then the periodic solution of the form (24) with j = 0 is asymptotically stable and the periodic solution j = 0 is unstable. If τ > τ * then every periodic solution (24) is unstable.

Numerical simulations
In this section we demonstrate numerical solutions of the delay differential equation (1) with the special initial conditions x(t) = y(t) = δ, t ∈ [−τ, 0], where δ ∈ R.
First we fix τ = 0.392. For τ < τ * ≈ 0.398284 · · · , we show that the periodic solution of the form (24) with j = 0 is asymptotically stable (Theorem 5). From (27), we can compute the radius of the stable periodic solution as r = ρ 0 ≈ 1.18547 · · · . In Figure 7, a trajectory of the solution for δ = −36 is plotted in (x, y) plane, which shows that the periodic solution with the radius r = ρ 0 ≈ 1.18547 · · · attracts the solution. Observe that there is an unstable periodic solution in the vicinity of the stable periodic solution (in Figure 7, the trajectory of the asymptotically stable periodic solution and of the unstable periodic solution are illustrated as the dashed orange circle and as the dashed blue circle, respectively). In Figure 8, a trajectory of the solution with the initial condition δ = −37 is plotted. In this case, the solution winds around the unstable periodic solution (dashed blue circle) and then leaves and goes far away.
In Figure 9, we plot log r(t) for several initial conditions δ. The numerical experiment suggests that the solution exists globally for −36 < δ ≤ 0 and blows up for δ < −37: in short, the solution blows up for large |δ|. We can numerically observe many blow-up solutions, which blow up even after t = τ /2. In this paper, we prove the existence of blow-up solutions, which blow up in the time interval (0, τ /2). The numerical simulation suggest that many solutions blow up in finite times.
We also numerically compute the solution for τ > τ * . Figure 10 shows a transient behavior of a solution. The solution stays around the origin for a while and then it blows up in a finite time. Although for τ = 0.3985 > τ * , there does not exist   Figure 9. We plot the growth of log r(t) for several δ. Here τ is fixed (τ = 0.392). The numerical simulation suggests that larger |δ| causes the solution to blow up faster.
periodic solution for j = 0 and j = 1 around the origin, those periodic solutions may indirectly affect the transient behavior of the solution.

Discussion
In this paper we show an example of a blow-up phenomenon in a planar system of delay differential equations. In our system, the delay completely changes the system: many blow-up solutions and periodic solutions suddenly appear, no matter how small the length of the delay is. In neutral delay differential equations, it is known that arbitrary small delay can destabilize the system (see Chapter 1.7 in [9]). Here we find that arbitrary small delay can induce blow-up solutions in delay differential equations. Numerical simulations suggest that many solutions either tend to the stable periodic solution or blow up in a finite time. It is not obvious if more complicated solution behavior exists. The transient behavior observed in Figure 10 is interesting, as it looks that the solution tries to find a stable periodic solution which does not exist in this parameter setting.
Many results concerning the existence of the blow-up solutions are available in Volterra integral equations (see [2,3] and references therein). When delay differential equations can be formulated as Volterra integral equations, we may apply the blow-up results for Volterra integral equations to delay differential equations, see e.g. [1,21] for a relation between Volterra type integral equations and integrodifferential equations. However, we cannot apply the results for Volterra integral equations to a system of integral equations which is rewritten formally from our target system of delay differential equations. Since delay differential equations form an infinite dimensional dynamical system, it is also not straightforward to apply the results established in ordinary differential equations.
In the context of mathematical modelling, our example suggests that arbitrary small delay can be responsible for a drastic change of the dynamics, thus one should be careful when ignoring small delay. Other blow-up mechanisms in delay differential equations will be explored in our future work.