betaFeedback

WorksPhysics and Astronomy

Noise-induced transitions between coexisting limit cycles beyond saddle-node bifurcations

Abstract

Nonlinear multistable systems with limit cycles may exhibit noise-induced transitions between oscillatory states, even in parameter regimes beyond the saddle-node bifurcations that determine the existence of these limit cycles. This behavior is associated with the ghost of a saddle-node bifurcation, a well-known feature of nonlinear systems near such bifurcation points. To investigate this phenomenon, we first consider a stochastic minimal model that supports two coexisting limit cycles within a parameter interval bounded by two saddle-node bifurcations. We numerically and theoretically estimate the probability of remaining near each stable attractor as a function of both the noise intensity and the distance to the ghost structure. The model is then extended by introducing spatially localized, non-homogeneous noise with enhanced intensity in a prescribed region of phase space. This modification is intended to emulate the asymmetric structure of the Hindmarsh-Rose model, in which bursting limit cycles approach each other closely in specific regions of phase space, thereby facilitating noise-induced transitions between them.

Highlights

  • A bi-variate, axisymmetric model with homogeneous noise is proposed and studied.

  • The birhythmicity exhibited by the model is extended by the noise.

  • The model is numerically and analytically characterized.

  • The model is amended by incorporating non-homogeneous noise.

  • The model reproduces behaviors observed in neurological nonlinear systems.

Introduction

Multistability is a widespread phenomenon in nonlinear systems [1]. It refers to the coexistence of multiple attractors, including fixed points, limit cycles, quasiperiodic regimes, and chaotic attractors, among others. The attractor eventually reached by the system depends on the initial conditions [2, 3]. A particularly relevant instance of multistability is birhythmicity, which denotes the coexistence of two periodic solutions with distinct amplitudes and frequencies [4].

Understanding the effects of noise on multistable systems is relevant across several areas of science, including physics, chemistry, and biology [5, 6, 7]. Noise plays a fundamental role in neuronal dynamics, both at the level of individual neurons and in the collective behavior of neuronal networks [8, 9]. When noise is introduced into a multistable system, random transitions between coexisting attractors may occur, as stochastic fluctuations can drive trajectories across the separatrices that delimit their basins of attraction [4]. For example, noise-induced chaos beyond the deterministic region of existence of a chaotic attractor has been reported in a neuronal discrete map [10]. In epilepsy research, noise has also been shown to trigger transitions that lead to epileptic seizures [11].

Multistable systems exhibiting one or more stable limit cycles in parameter regimes structured by saddle-node bifurcations are of particular interest, since transitions associated with these cycles can occur even beyond the corresponding bifurcation point. This behavior is associated with the ghost of a saddle-node bifurcation, a characteristic feature of nonlinear systems near these bifurcations [3]. Beyond the bifurcation, neither the stable nor the unstable limit cycle exists. However, if the initial condition lies close to the region of phase space where the annihilated limit cycles used to exist, the trajectory may remain in this region for a finite time before escaping from it. Consequently, in the presence of noise, transitions mediated by this ghost transient may occur and are qualitatively similar to the stochastic transitions observed inside the coexistence region, where an actual stable limit-cycle attractor is present.

The Hindmarsh–Rose model provides a relevant framework for investigating the effects of noise on birhythmicity. It was proposed in 1984 to reproduce phenomena commonly observed in experiments on single neurons, such as bursting, slow-fast dynamics, and diverse firing patterns [12, 13, 4]. As the parameter associated with the fast ion channels is varied, the model exhibits a branch of bursting limit cycles that undergoes successive folds through a sequence of saddle-node bifurcations, with an increasing number of spikes per burst. Regions of birhythmicity arise between consecutive bifurcations. When noise is introduced, transitions occur between the coexisting bursting limit cycles. Owing to the ghost phenomenon, such transitions can also occur outside, but close to, the birhythmicity regions. In Ref. [14], the phenomenon of stochastic birhythmicity in the Hindmarsh–Rose model was investigated in detail, and a minimal model was proposed to provide a qualitative description of the underlying mechanism. However, that model captures the phenomenon only partially, as it exhibits a single limit cycle coexisting with a fixed-point attractor rather than two coexisting limit cycles, and a single saddle-node bifurcation of limit cycles rather than a pair of such bifurcations. Consequently, it does not display genuine birhythmicity, although it does exhibit the noise-induced emergence of periodic oscillations beyond the saddle-node bifurcation.

To gain a deeper understanding of the mechanisms underlying birhythmicity in systems with a structure similar to that of the Hindmarsh–Rose model, as well as of the role played by the ghost phenomenon, we propose here a stochastic, axially symmetric birhythmic model in which two coexisting limit cycles arise within a parameter interval bounded by two saddle-node bifurcations. The simplicity of the model allows for a tractable probabilistic description based on the associated Fokker–Planck equation [15]. Within this framework, stochastic transitions between the coexisting limit cycles may occur throughout the phase space, in contrast to the strongly asymmetric Hindmarsh–Rose model, where such transitions are largely confined to regions in which the limit-cycle trajectories approach each other closely.

To capture this effect, we extend the simplified model by introducing non-homogeneous noise with enhanced intensity in a prescribed region of phase space. As a consequence, stochastic transitions become more likely to occur within that region. Overall, both the homogeneous- and non-homogeneous-noise models qualitatively reproduce the same hourglass-like structure observed in the parameter space of the Hindmarsh–Rose model, suggesting that this structure may be a robust feature of nonlinear systems in which two limit-cycle attractors coexist within a region bounded by saddle-node bifurcations. This hourglass-like structure emerges from long-time observations throughout parameter space. However, in the low-noise regime, even very long observation times may be insufficient for the system to reach stochastic equilibrium, since the corresponding relaxation times become exceedingly large and practically inaccessible. Consequently, a theoretical approach based on stationary probability distributions is proposed to describe the behavior of the system when the noise is high, while in the low-noise case, a complementary approach, inspired by Kramers’ escape problem [16, 17], is proposed.

The article is organized as follows. In § 3, we introduce the axially symmetric model and characterize stochastic birhythmicity in terms of the noise intensity and a control parameter. In § 4, we incorporate non-homogeneous noise into the model and analyze the resulting stochastic birhythmicity. Finally, § 5 summarizes the main conclusions.

Double saddle-node bifurcation model with homogeneous noise

Deterministic model

We introduce a simple model designed to emulate the birhythmicity observed in the Hindmarsh–Rose model, where the coexisting attractors remain relatively far from the origin but come into close proximity in a specific region of phase space. This model exhibits two coexisting stable limit cycles bounded by two saddle-node bifurcations. The model is a bivariate system of stochastic differential equations written in vector form as

drdt=F(r)+εξ(t),(1)\frac{d\mathbf{r}}{dt} = \mathbf{F}(\mathbf{r}) + \varepsilon\boldsymbol{\xi}(t), \tag*{(1)}

where r=(x(t),y(t))\mathbf{r}=(x(t),y(t)) collects the two time-dependent variables. The term F(r)\mathbf{F}(\mathbf{r}) represents the deterministic dynamics, while εξ(t)\varepsilon\boldsymbol{\xi}(t) accounts for intrinsic noise. In biological neurons, such noise reflects the intrinsic complexity of the system and the large number of underlying components and variables involved. The parameter ε\varepsilon controls the noise intensity, and ξ(t)\boldsymbol{\xi}(t) is a bivariate, uncorrelated white-noise process [15], defined by

⟨ξi(t)⟩=0,\langle\xi_i(t)\rangle= 0,
⟨ξi(t)ξj(t′)⟩=δijδ(t′−t),(2)\langle\xi_i(t)\xi_j(t')\rangle= \delta_{ij}\delta(t'-t), \tag*{(2)}

where δij\delta_{ij} represents the Kronecker delta coefficients and δ(t′−t)\delta(t'-t) is the Dirac delta function. The deterministic component is given by an autonomous vector field, which in polar coordinates naturally separates into radial and angular parts,

F(r)=f(r)r^+rωθ^,(3)\mathbf{F}(\mathbf{r}) = f(r)\hat{\mathbf{r}} + r\omega\hat{\boldsymbol{\theta}}, \tag*{(3)}

where r=∣r∣r = \lvert\mathbf{r}\rvert is the distance to the origin, while r^\hat{\mathbf{r}} and θ^\hat{\boldsymbol{\theta}} are the polar unit vectors. The parameter ω\omega represents the angular velocity around the origin, which throughout this section is fixed at ω=1\omega= 1. This choice is arbitrary and does not affect the long-time behavior of the system. The radial component is defined by the scalar function f(r)f(r),

f(r)=−r(N(r2)−b),(4)f(r) = -r\left(N(r^2) - b\right), \tag*{(4)}

where bb is a control parameter and N(r2)N(r^2) is a scalar function with an N-shaped profile,

N(r2)=10r2(r2−2.252)(r2−2.502)(r2−2.752).(5)N(r^2) = \frac{10}{r^2}\left(r^2 - 2.25^2\right)\left(r^2 - 2.50^2\right)\left(r^2 - 2.75^2\right). \tag*{(5)}

The specific values 2.25, 2.50, and 2.75 used in Eq. (5) are chosen so that the distance between the two coexisting limit cycles remains relatively small compared to their radii. We have also tested other choices of these values, as well as different N-shaped profiles, and verified that no significant qualitative changes arise in the results presented in this work.

The function N(r2)N(r^2) vanishes at r=2.25r = 2.25, r=2.50r = 2.50, and r=2.75r = 2.75, diverges to −∞-\infty as r→0r \to0, and diverges to ∞\infty as r→∞r \to\infty. It possesses a local maximum at (r,N)=(2.354,1.238)(r, N) = (2.354, 1.238) and a local minimum at (r,N)=(2.643,−1.168)(r, N) = (2.643, -1.168).

The deterministic system defined by Eq. (1) with ε=0\varepsilon= 0 has a singular point at r=0r = 0, where f(r)f(r) diverges. This point is always repelling, regardless of the values of the parameters bb and ω\omega. The equilibrium solutions correspond to limit cycles centered at the origin and rotating with angular velocity ω\omega. Their radii are given by the nonzero solutions of f(r)=0f(r) = 0, or equivalently by the solutions of N(r2)=bN(r^2) = b. Owing to the N-shaped profile of N(r2)N(r^2), this equation admits a single non-negative solution for b<−1.168b < -1.168 or b>1.238b > 1.238, and three non-negative solutions for −1.168<b<1.238-1.168 < b < 1.238. Consequently, the system possesses lower, middle, and upper branches of limit-cycle solutions, as illustrated in Figure 1. The lower and upper branches are stable, whereas the middle branch is unstable. Their stability is determined by the sign of the derivative of f(r)f(r). The lower and middle branches coalesce in a saddle-node bifurcation of limit cycles at (b,r)≈(1.238,2.354)(b, r) \approx(1.238, 2.354), whereas the middle and upper branches coalesce in a second saddle-node bifurcation at (b,r)≈(−1.168,2.643)(b, r) \approx(-1.168, 2.643). These bifurcations are indicated by dots in Figure 1. Therefore, the system exhibits birhythmicity for −1.168<b<1.238-1.168 < b < 1.238. As b→−∞b \to-\infty, the radius of the lower limit-cycle branch tends asymptotically to zero.

Figure 1. Bifurcation diagram of the deterministic simple model, showing the limit cycle radius rr as a function of the control parameter bb. The solid blue lines denote the stable limit cycle branches and dashed blue line denotes the unstable limit cycle branch. The diagram also marks the limit cycle saddle-node bifurcations, and the intersections of the branches with the b=0b = 0 axis.

Some simulations of the deterministic system for different initial conditions and values of bb are shown in Figure 2, where the distance to the origin, r(t)r(t), is plotted as a function of time. Within the birhythmicity region (Figure 2(b)), the system converges to one of the two stable limit cycles depending on whether the initial distance r(0)r(0) is smaller or larger than the radius of the unstable limit cycle located between them. In other words, the unstable cycle acts as a separatrix between the basins of attraction of the two stable cycles. Outside the birhythmicity region, only a single stable limit cycle exists, and the system converges to it regardless of the initial condition. However, near the boundaries of the birhythmicity region, the ghost phenomenon emerges [3]. In Figure 2(a), bb is set slightly below −1.168-1.168, corresponding to the left saddle-node bifurcation shown in Figure 1. In this case, the stable limit cycle has a radius of approximately 2.21. If r(0)<2.64r(0) < 2.64, corresponding to the radius of the colliding cycles at the bifurcation, the system converges rapidly to the stable limit cycle. By contrast, if r(0)>2.64r(0) > 2.64, the trajectory slows down near r=2.64r = 2.64 and remains in its vicinity for a considerable time before eventually converging to the stable limit cycle. A similar effect is observed in Figure 2(c), where bb is set slightly above 1.238, corresponding to the right saddle-node bifurcation in Figure 1. In this case, the stable limit cycle has a radius of approximately 2.79. If r(0)>2.35r(0) > 2.35, corresponding to the radius at the saddle-node bifurcation, convergence is rapid. However, if r(0)<2.35r(0) < 2.35, the trajectory slows down near r=2.35r = 2.35 before eventually converging to the stable limit cycle. Thus, the ghost phenomenon acts as a precursor of the stable limit cycle that disappears at each saddle-node bifurcation. This effect becomes increasingly pronounced as the system approaches the birhythmicity region: as bb approaches either saddle-node bifurcation, the time during which r(t)r(t) remains near the corresponding saddle-node radius diverges.

Figure 2. Simulations of Eq. (1) for ϵ=0\epsilon= 0 and different values of bb: (a) b=−1.2b = -1.2, (b) b=0b = 0, and (c) b=1.3b = 1.3. The initial conditions are uniformly distributed between r=2r = 2 and r=3r = 3. Time series of r(t)r(t) are shown as solid blue curves. The radii of the stable and unstable limit cycles are indicated by solid and dashed red curves, respectively. In panels (a) and (c), the radius of the nearby saddle-node bifurcation is also shown as a dash-dotted red line. Due to the ghost phenomenon, r(t)r(t) remains close to this radius for some time before approaching the stable limit cycle.

Stochastic model

When noise is introduced into the model (i.e., ϵ>0\epsilon> 0), the solutions r(t)r(t) of Eqs. (1),(2) become stochastic processes. Their time evolution can be described statistically by a probability distribution ρ(r,t)\rho(\mathbf{r}, t), governed by the associated Fokker-Planck equation [15, 18, 19, 20],

∂ρ∂t=−∇⋅(F(r)ρ)+ϵ22∇2ρ,(6)\frac{\partial\rho}{\partial t} = -\nabla\cdot\left(\mathbf{F}(\mathbf{r})\rho\right) + \frac{\epsilon^2}{2}\nabla^2 \rho, \tag*{(6)}

where ∇≡(∂/∂x,∂/∂y)\nabla\equiv(\partial/\partial x,\partial/\partial y) is the nabla operator. Equation (6) admits a stationary solution, namely a probability distribution to which the system converges after a sufficiently long time, independently of its initial state. Unlike the Hindmarsh--Rose model, which is strongly asymmetric, the present model is axially symmetric, i.e., statistically invariant under rotations. This symmetry, together with the existence of a stationary regime, ensures that the system is ergodic, so that expected values can be estimated as time averages [15]. It is therefore natural to assume that the stationary distribution depends only on the radial distance to the origin, ρs(r)=ρs(r)\rho_s(\mathbf{r})=\rho_s(r). After removing the time dependence and using the identity ∇2ρ=∇⋅(∇ρ)\nabla^2\rho=\nabla\cdot(\nabla\rho), Eq. (6) yields

∇⋅(−F(r)ρ+ε22∇ρ)=0.(7)\nabla\cdot\left(-\mathbf{F}(r)\rho+\frac{\varepsilon^2}{2}\nabla\rho\right)=0. \tag*{(7)}

In polar coordinates, and provided that there is no dependence on θ\theta, it turns out that ∇ρ=∂ρ∂rr^\nabla\rho=\frac{\partial\rho}{\partial r}\hat{\mathbf{r}} and ∇⋅H=1r∂∂r(rHr)\nabla\cdot\mathbf{H}=\frac{1}{r}\frac{\partial}{\partial r}(rH_r). Thus, Eq. (7) yields

1rddr(r(−f(r)ρs+ε22dρsdr))=0.(8)\frac{1}{r}\frac{d}{dr}\left(r\left(-f(r)\rho_s+\frac{\varepsilon^2}{2}\frac{d\rho_s}{dr}\right)\right)=0. \tag*{(8)}

The radial component f(r)f(r), defined in Eq. (4), can also be written as the negative gradient of a scalar potential, f(r)=−du/drf(r)=-du/dr, with

u(r)=16u1r6−14u2r4+12(u3−b)r2−u4ln⁡(r),(9)u(r)=\frac{1}{6}u_1r^6-\frac{1}{4}u_2r^4+\frac{1}{2}(u_3-b)r^2-u_4\ln(r), \tag*{(9)}

where the coefficients take the values u1=10u_1=10, u2=188.75u_2=188.75, u3≈1171.91u_3\approx1171.91, and u4≈2392.82u_4\approx2392.82. By applying Eq. (9) and imposing the normalization condition that the integral of ρs\rho_s over the entire plane equals unity, Eq. (8) reduces to

dudrρs+ε22dρsdr=0.(10)\frac{du}{dr}\rho_s+\frac{\varepsilon^2}{2}\frac{d\rho_s}{dr}=0. \tag*{(10)}

The solution of Eq. (10) assuming zero stationary radial probability current is

ρs(r)=Ce−2u(r)/ε2,(11)\rho_s(r)=Ce^{-2u(r)/\varepsilon^2}, \tag*{(11)}

where CC is a normalization constant. This function gives the probability density of finding the system in a vector state r\mathbf{r} such that ∣r∣=r|\mathbf{r}|=r. The probability density of the scalar radial coordinate rr is obtained by integrating ρs(r)\rho_s(r) along the circumference of radius rr centered at the origin,

ps(r)=2πCre−2u(r)/ε2.(12)p_s(r)=2\pi Cre^{-2u(r)/\varepsilon^2}. \tag*{(12)}

Simulations of Eqs. (1),(2) for different parameter values (b,ε)(b,\varepsilon) are shown in Figure 3, together with the corresponding histograms. Within the birhythmicity region (Figure 3(b)), bounded by the saddle-node bifurcations, the stochastic system undergoes random transitions between the two limit-cycle attractors. This behavior is reflected in the histogram of the radial distance rr, which exhibits two peaks located at the radii of the deterministic stable cycles, separated by a minimum at the radius of the unstable cycle. Outside the birhythmicity region (Figure 3(a) and Figure 3(c)), the system possesses a single stable limit-cycle attractor. When noise is introduced, the trajectory fluctuates around this attractor, but it may occasionally visit the ghost region associated with the annihilated cycle. This behavior is reflected in the corresponding histograms, which display a single peak centered near the radius of the deterministic attractor, together with a tail extending beyond the radius of the nearest saddle-node bifurcation.

Figure 3. Numerical simulations of Eqs. (1),(2) for ε=1.2\varepsilon= 1.2 and different values of bb: (a) b=−1.3b = -1.3, (b) b=0b = 0, and (c) b=1.3b = 1.3. In each case, the time series r(t)r(t) from a single simulation is shown after a transient, once the system has reached a steady state, together with the normalized histogram of rr obtained from 100 simulations of duration Δt=100\Delta t = 100 with initial conditions uniformly distributed between r=2r = 2 and r=3r = 3. Each histogram consists of 200 bins, shown in yellow, and is compared with the stationary probability distribution ps(r)p_s(r), shown as a solid blue curve. The radii of the stable and unstable deterministic limit cycles are indicated by solid and dashed red lines, respectively. In panels (a) and (c), the radius of the nearest saddle-node bifurcation is also shown as a dash-dotted red line, acting as a de facto separatrix between the stable limit cycle and the ghost.

This tail arises from the ghost phenomenon. Near a saddle-node bifurcation, noise may drive the system into the ghost region where the annihilated cycle would have existed, causing the trajectory to remain there for a relatively long time before eventually returning to the stable limit cycle. The numerically obtained histograms are in excellent agreement with the theoretically predicted stationary distribution ps(r)p_s(r). This agreement can be quantified by the Hellinger distance, which is of the order of 10−410^{-4} [21]. As shown below, however, this agreement deteriorates in the low-noise regime because the system requires exceedingly long times to reach stochastic equilibrium.

Although, in the simple model studied here, the peaks and valleys of the state distributions coincide with the deterministic stable and unstable solutions, respectively, such a correspondence need not hold in other stochastic systems as the noise intensity increases [22]. In the present case, however, this effect is barely noticeable, since the deterministic stable branches lie very close to one another within the birhythmicity region.

Overall, Figure 3 shows that, regardless of the value of bb, the system may undergo stochastic transitions between two stochastic states: a large-radius state and a small-radius state. When bb belongs to the deterministic birhythmicity region (Figure 3(b)), these states fluctuate around the two deterministic stable limit cycles. In this case, the deterministic unstable limit cycle naturally serves as the separatrix between them. When bb lies outside the birhythmicity region (Figure 3(a) and Figure 3(c)), one state fluctuates around the single stable limit cycle, whereas the other is associated with the ghost of the annihilated limit cycle. The radius of this ghost state is approximated by the radius of the corresponding saddle-node limit cycle. In the deterministic system, depending on which side of the ghost the initial condition is located, trajectories either approach the ghost transiently or move away from it immediately. Thus, the ghost acts as an effective (de facto) separatrix, making it a natural criterion for distinguishing between the large- and small-radius

stochastic states.

Following the discussion above, we can define a probability PLSP_{\mathrm{LS}} of observing the system evolving in the large-radius state at a given time. A numerical estimation of PLSP_{\mathrm{LS}} over the (b,ε)(b,\varepsilon) parameter plane, based on simulations of Eqs. (1),(2), is shown in Figure 4(a). For each pair (b,ε)(b,\varepsilon), several simulations of duration Δt=1000\Delta t=1000 are performed using dispersed initial conditions, and the fraction of time steps for which r(t)r(t) exceeds the separatrix radius is computed. For low noise levels (ε<0.03\varepsilon<0.03), PLSP_{\mathrm{LS}} is nearly zero on the left side of the deterministic birhythmicity region, nearly one on the right side, and ranges between 0.36 and 0.65 inside it. The latter occurs because, under weak noise, the system rapidly approaches one or the other stable limit cycle depending on whether r(0)r(0) lies below or above the separatrix, with little probability of crossing it during the subsequent evolution.

Figure 4. (a) Numerical estimation of the probability PLSP_{\mathrm{LS}} of observing the largest limit cycle, or its ghost, in the stochastic system, Eqs. (1),(2), as a function of (b,ε)(b,\varepsilon). For each parameter set, 100 simulations of duration Δt=1000\Delta t = 1000 are performed, with initial conditions uniformly distributed between r=2r = 2 and r=3r = 3. After t=500t = 500, the fraction of time points for which r(t)r(t) exceeds the radius of the separatrix is computed. The dashed black lines indicate the boundaries of the deterministic birhythmicity region, while the solid red lines mark the estimated boundaries of the de facto stochastic birhythmicity region, defined as the region in which both states occur with probability at least 2%. (b) Numerical computation of the theoretical probability PLSP_{\mathrm{LS}}, given by Eq. (13).

As the noise intensity increases, the region in which PLSP_{\mathrm{LS}} is neither one nor zero becomes narrower. Inside the deterministic birhythmicity region, but close to either saddle-node bifurcation, the stable limit cycle created at the bifurcation is less stable than its counterpart. In terms of the scalar potential u(r)u(r), this means that the well associated with its radius is shallower than that corresponding to the other attractor. This behavior is illustrated in Figure 5. As the parameter bb crosses either saddle-node bifurcation, a new well appears in the profile of u(r)u(r). Initially, this well is shallow, but it becomes progressively deeper as bb moves further into the deterministic birhythmicity region. Consequently, moderate noise can readily drive the system from the shallow well to the deeper one, whereas the reverse transition is much less likely. As a result, one state dominates the probability distribution, and both states are observed with comparable likelihood only near the center of the deterministic birhythmicity region. This defines a de facto stochastic birhythmicity region. More specifically, this region is defined here as the set of points (b,ε)(b,\varepsilon) for which both states have an observation probability of at least 2%, although this threshold is admittedly arbitrary.

Figure 5. Profiles of the scalar potential u(r)u(r) around the saddle-node bifurcations. (a) Left saddle-node bifurcation for b=−1.17b=-1.17 (solid line), b=−0.77b=-0.77 (dashed line), and b=−0.37b=-0.37 (dotted line). (b) Right saddle-node bifurcation for b=1.24b=1.24 (solid line), b=0.84b=0.84 (dashed line), and b=0.44b=0.44 (dotted line).

The stochastic birhythmicity region reaches its minimum width around ε≈4×10−1\varepsilon\approx4 \times10^{-1}. For higher noise intensities, the region broadens again, as fluctuations become sufficiently large to induce transitions between the two wells of u(r)u(r) regardless of their relative depths. At ε≈9×10−1\varepsilon\approx9 \times10^{-1}, the stochastic birhythmicity region becomes wider than its deterministic counterpart. In this regime, even outside the deterministic birhythmicity region, the system can switch between the stable limit cycle and the ghost of the annihilated cycle (see Figure 3). Overall, the simple model qualitatively reproduces the hourglass-shaped stochastic birhythmicity region observed in the stochastic Hindmarsh–Rose model [14]. This suggests that the hourglass-shaped structure may be a robust feature of systems exhibiting birhythmicity bounded by two saddle-node bifurcations of limit cycles.

The probability PLSP_{\mathrm{LS}} can also be theoretically estimated as the integral of the stationary distribution ps(r)p_{\mathrm{s}}(r) from the separatrix radius discussed above to infinity,

PLS=∫rsep∞ps(r) dr,(13)P_{\mathrm{LS}} = \int_{r_{\mathrm{sep}}}^{\infty} p_{\mathrm{s}}(r)\,\mathrm{d}r, \tag*{(13)}

This integral does not have a straightforward anti-derivative, but it is still relatively simple to compute numerically. The only point requiring special care is the choice of rsepr_{\mathrm{sep}}, the radius of the de facto separatrix. As discussed above, in the deterministic monorhythmic regions this radius is identified with that of the annihilated saddle-node cycle, whereas in the deterministic birhythmic region it corresponds to the radius of the unstable limit cycle. The latter is given by the middle nonzero solution of N(r2)=bN(r^2)=b. Although this equation is nontrivial, the unstable branch can be accurately fitted by a simple expression. Thus, rsepr_{\mathrm{sep}} can be written as the following piecewise function of bb,

rsep={2.643,b≤−1.168,−0.0932b+1.168+0.08851.238−b+2.5024,−1.168<b<1.238,2.354,b≥1.238.(14)r_{\mathrm{sep}} = \begin{cases} 2.643, & b \le-1.168, \\ -0.0932\sqrt{b+1.168} + 0.0885\sqrt{1.238-b} + 2.5024, & -1.168 < b < 1.238, \\ 2.354, & b \ge1.238. \end{cases} \tag*{(14)}

The numbers −0.0932-0.0932, 0.08850.0885, and 2.50242.5024 are obtained by fitting the function to the unstable branch, with coefficient of determination R2=0.9998R^2 = 0.9998. The other numbers are the coordinates of the saddle-node bifurcations (see Figure 1).

The color map of the theoretical integral PLSP_{\mathrm{LS}}, Eq. (13), is shown in Figure 4(b). The upper part of the hourglass-like structure is identical to that in Figure 4(a). However, the lower part is absent. Instead, for ε<0.2\varepsilon< 0.2, the probability PLSP_{\mathrm{LS}} is nearly zero on the left side and nearly one on the right side, with an abrupt transition around b=−0.021b = -0.021, where the two wells of the potential u(r)u(r) have equal depth. This discrepancy occurs because, in the lower central region of Figure 4(a), the system has not yet reached stochastic equilibrium. Since the noise intensity is very low, there barely is any flow through the separatrix over the simulation time Δt=1000\Delta t = 1000. The color map in Figure 4(b) indicates that, under birhythmicity and very low noise, the system is much more likely to occupy the deeper potential well, even when the depth difference between the two wells is small. If the system is initially in the basin of attraction of the shallower well, it will first approach this well, but noise will eventually drive it into the deeper well, although this transition may require an extremely long time. This conclusion may seem counterintuitive, but it is consistent with Eq. (12). For a potential u(r)u(r) with two wells of similar depth, −u(r)-u(r) has two peaks of similar height. However, since −u(r)-u(r) is multiplied by 2/ε22/\varepsilon^2, which is large for small ε\varepsilon, this small height difference is strongly amplified in the profile of ps(r)p_{\mathrm{s}}(r). Therefore, the simulations in Figure 4(a) would need to be run for much longer times in order to produce a map similar to that in Figure 4(b). Presumably, the same mechanism occurs in the Hindmarsh–Rose model: the hourglass-shaped map obtained in Ref. [14] reflects the fact that the system requires very long times to reach stochastic equilibrium when the noise intensity is low.

To further illustrate that, in the low-noise regime, the system requires an extremely long time to reach stochastic equilibrium, simulations of Eqs. (1),(2) are performed with b=−0.05b = -0.05. For this value of bb, the system is birhythmic and the potential u(r)u(r) is nearly symmetric, with the two wells having almost the same depth, as shown in Figure 6(a). The noise intensity is set to ε=0.18\varepsilon= 0.18, and uniformly distributed initial conditions are used. These simulations have a very long duration, Δt=106\Delta t=10^{6}, but only the last 0.05% of the data is used to construct the histogram shown in Figure 6(b). The histogram exhibits two sharp peaks around the deterministic attractors. This is intuitively expected, since the initial conditions are uniformly distributed, the attraction basins have similar sizes, and the noise intensity is small. However, the histogram does not match the stationary probability distribution ps(r)p_s(r). Indeed, ps(r)p_s(r) also has two sharp peaks around the deterministic attractors, but the left peak is much higher than the right one. This mismatch occurs because, even after such a long simulation time, the system has not yet reached stochastic equilibrium, which would require enough noise-induced transitions between attractors. Kramers’ theory provides an order of magnitude of the expected escape time from a potential well, given by Δtescape∼eΔu/ε2\Delta t_{\mathrm{escape}}\sim e^{\Delta u/\varepsilon^{2}}, where Δu\Delta u is the well depth, i.e., the difference between the local maximum and the local minimum [16]. For ε=0.18\varepsilon=0.18 and Δu=0.5\Delta u=0.5 (see Figure 6(a)), the escape time is of the order of 5×1065\times10^{6}.

Figure 6. (a) Profile of u(r)u(r) for b=−0.05b=-0.05. (b) Normalized histogram of rr obtained from 100 simulations of Eqs. (1),(2) for b=−0.05b=-0.05 and ε=0.18\varepsilon=0.18. The simulations have duration Δt=106\Delta t=10^{6} and use initial conditions uniformly distributed between r=2r=2 and r=3r=3. The histogram consists of 200 bins, shown in orange, and is constructed using only the last 0.05% of the simulation data. The stationary probability distribution ps(r)p_s(r) is also shown as a solid blue curve.

Theoretical estimation of PLSP_{\mathrm{LS}} in the low-noise regime

As illustrated in Figure 4, the colormap of the large state probability, PLSP_{\mathrm{LS}}, estimated from numerical simulations, is well reproduced by the integral formula given in Eq. (13), except in the deterministic birhythmicity region when the noise is low (namely, ε<0.3\varepsilon<0.3). Therefore, a theoretical expression that reproduces the simulational estimation of PLSP_{\mathrm{LS}} in this region of parameter space is yet to be found. To this end, we propose to carry out a procedure analogous to Kramers’ escape rate problem [16, 17].

We start from the Fokker–Planck equation, Eq. (6). Owing to the axial symmetry of the system, we neglect the dependence on θ\theta. However, because we are now considering a regime in which the system has not yet reached stochastic equilibrium, the time dependence must be retained. This reduction is valid, for instance, when the initial probability distribution is axially symmetric. Thus, Eq. (6) yields

∂ρ∂t=1r∂∂r(r(dudrρ+ε22∂ρ∂r)).(15)\frac{\partial\rho}{\partial t} = \frac{1}{r}\frac{\partial}{\partial r}\left(r\left(\frac{du}{dr}\rho+ \frac{\varepsilon^2}{2}\frac{\partial\rho}{\partial r}\right)\right). \tag*{(15)}

The probability density of the scalar coordinate rr is given by integrating ρ(r,t)\rho(r,t) along the circumference of radius rr centered at the origin. Thus, p(r,t)=2πC(t)ρ(r,t)p(r,t)=2\pi C(t)\rho(r,t), where C(t)C(t) is a normalization coefficient. We also define an effective potential, ueff(r)=u(r)+12ε2ln⁡(r)u_{\mathrm{eff}}(r)=u(r)+\frac{1}{2}\varepsilon^2\ln(r). Multiplying Eq. (15) by 2πC(t)2\pi C(t) leads to

∂p∂t=−∂Jr∂r,(16)\frac{\partial p}{\partial t} = -\frac{\partial J_r}{\partial r}, \tag*{(16)}

where JrJ_r is the radial probability flux given by

Jr=−(dueffdrp+ε22∂p∂r).(17)J_r = -\left(\frac{du_{\mathrm{eff}}}{dr}p+\frac{\varepsilon^2}{2}\frac{\partial p}{\partial r}\right). \tag*{(17)}

Notice that Eq. (16) is formally equivalent to the Fokker–Planck equation of a one-dimensional system evolving in the effective potential ueffu_{\mathrm{eff}} and subjected to white noise of intensity ε\varepsilon.

In the deterministic birhythmicity region, the system has three equilibrium radii: the radius of the small stable limit cycle, the radius of the large stable limit cycle, and, between them, the radius of the unstable limit cycle. We denote these three radii by rs1r_{s1}, rs2r_{s2}, and rur_u, respectively. As discussed above, rur_u acts as the separatrix between the basins of attraction of the two stable limit cycles. Therefore, the probability of observing the system in the large-radius state is obtained by integrating the probability distribution from rur_u to infinity,

PLS=∫ru∞p(r,t) dr.(18)P_{\mathrm{LS}} = \int_{r_u}^{\infty} p(r,t)\,dr. \tag*{(18)}

Subsequently, integrating Eq. (16) in the same range leads to

dPLSdt=−Jr(ru,t).(19)\frac{dP_{\mathrm{LS}}}{dt}=-J_r(r_u,t). \tag*{(19)}

We now aim to write Eq. (19) as an ordinary differential equation for PLSP_{\mathrm{LS}}. To this end, we use the following identity satisfied by JrJ_r:

∂∂r(pe2ε2ueff(r))=−2ε2Jre2ε2ueff(r).(20)\frac{\partial}{\partial r}\left(p e^{\frac{2}{\varepsilon^2}u_{\mathrm{eff}}(r)}\right)=-\frac{2}{\varepsilon^2}J_r e^{\frac{2}{\varepsilon^2}u_{\mathrm{eff}}(r)}. \tag*{(20)}

As discussed above, in the low-noise regime the system may require extremely long times to reach stochastic equilibrium. The evolution of the probability density is therefore slow, so that ∂p/∂t\partial p/\partial t can be approximated as zero over the timescale considered. According to Eq. (16), this implies that the radial probability flux JrJ_r depends only weakly on rr. Hence, over the interval [rs1,rs2][r_{s1},r_{s2}], JrJ_r can be approximated by its value at the separatrix, Jr(ru,t)J_r(r_u,t). Under this approximation, we integrate Eq. (20) over [rs1,rs2][r_{s1},r_{s2}],

p(rs2,t)e2ε2ueff(rs2)−p(rs1,t)e2ε2ueff(rs1)=−2ε2Jr(ru,t)∫rs1rs2e2ε2ueff(r) dr.(21)p(r_{s2},t)e^{\frac{2}{\varepsilon^2}u_{\mathrm{eff}}(r_{s2})}-p(r_{s1},t)e^{\frac{2}{\varepsilon^2}u_{\mathrm{eff}}(r_{s1})}=-\frac{2}{\varepsilon^2}J_r(r_u,t)\int_{r_{s1}}^{r_{s2}}e^{\frac{2}{\varepsilon^2}u_{\mathrm{eff}}(r)}\,\mathrm{d}r. \tag*{(21)}

We now consider an initial probability density consistent with the initial conditions of the simulations, namely a uniform distribution over the interval 2≤r≤32 \le r \le3. In the low-noise regime, the probability density rapidly concentrates around the deterministic attractors, forming two sharp peaks centered at rs1r_{s1} and rs2r_{s2} (see Figure 6(b)). The areas under these peaks are initially determined by the amount of probability lying on each side of the separatrix rur_u. Probability is then slowly transferred between the two attractors, so that the system evolves gradually toward stochastic equilibrium through changes in the peak heights. Since this transfer occurs on a much longer timescale than the local relaxation around each attractor, the density near each peak can be approximated by a quasi-equilibrium profile with a time-dependent amplitude. In particular, near the large-radius state rs2r_{s2}, we approximate the probability density as

p(r,t)=p(rs2,t)e−2ε2(ueff(r)−ueff(rs2)).(22)p(r,t)=p(r_{s2},t)e^{-\frac{2}{\varepsilon^2}\left(u_{\mathrm{eff}}(r)-u_{\mathrm{eff}}(r_{s2})\right)}. \tag*{(22)}

Next, we expand ueff(r)u_{\mathrm{eff}}(r) to second order around rs2r_{s2}. Since ueff′(rs2)u'_{\mathrm{eff}}(r_{s2}) is approximated as zero, this gives

p(r,t)=p(rs2,t)e1ε2ueff′′(rs2)(r−rs2)2.(23)p(r,t)=p(r_{s2},t)e^{\frac{1}{\varepsilon^2}u''_{\mathrm{eff}}(r_{s2})(r-r_{s2})^2}. \tag*{(23)}

Here, primes denote derivatives with respect to rr. The approximation ueff′(rs2)≈0u'_{\mathrm{eff}}(r_{s2}) \approx0 is justified because, for ε<0.3\varepsilon< 0.3, the critical points of ueff(r)u_{\mathrm{eff}}(r) are very close to those of u(r)u(r). Thus, the solutions of ueff′(r)=0u'_{\mathrm{eff}}(r)=0 are well approximated by the deterministic radii rs1r_{s1}, rur_u, and rs2r_{s2}, which satisfy u′(r)=0u'(r)=0.

We substitute Eq. (23) into Eq. (18). Assuming that the probability peak around rs2r_{s2} is narrow and well contained within the integration domain, the integral can be evaluated explicitly, yielding

PLS=πε2ueff′′(rs2)p(rs2,t).(24)P_{\mathrm{LS}}=\sqrt{\frac{\pi\varepsilon^2}{u''_{\mathrm{eff}}(r_{s2})}}p(r_{s2},t). \tag*{(24)}

Therefore, p(rs2,t)p(r_{s2},t) can be expressed in terms of PLSP_{\mathrm{LS}} as

p(rs2,t)=ueff′′(rs2)πε2PLS.(25)p(r_{s2},t)=\sqrt{\frac{u''_{\mathrm{eff}}(r_{s2})}{\pi\varepsilon^2}}P_{\mathrm{LS}}. \tag*{(25)}

Analogously, the value of the probability density at rs1r_{s1} can be expressed in terms of PSSP_{\mathrm{SS}}, the probability of observing the system in the small-radius state at a given time:

p(rs1,t)=ueff′′(rs1)πε2PSS.(26)p(r_{s1},t)=\sqrt{\frac{u''_{\mathrm{eff}}(r_{s1})}{\pi\varepsilon^2}}P_{\mathrm{SS}}. \tag*{(26)}

Since rs1r_{s1} and rs2r_{s2} are the radii of stable limit cycles, u′′(rs1)u''(r_{s1}) and u′′(rs2)u''(r_{s2}) are positive. In the low-noise regime, it is therefore reasonable to assume that ueff′′(rs1)u''_{\mathrm{eff}}(r_{s1}) and ueff′′(rs2)u''_{\mathrm{eff}}(r_{s2}) are also positive.

We now evaluate the integral on the right-hand side of Eq. (21). Since rur_u is the radius of the unstable limit cycle, we have u′(ru)=0u'(r_u)=0 and u′′(ru)<0u''(r_u)<0. In the low-noise regime, the integrand can be assumed to have a single dominant peak near r=rur=r_u, well contained within the integration domain. Moreover, since ueff(r)u_{\mathrm{eff}}(r) remains close to u(r)u(r) for sufficiently small ε\varepsilon, we can approximate ueff′(ru)≈0u'_{\mathrm{eff}}(r_u)\approx0 and assume ueff′′(ru)<0u''_{\mathrm{eff}}(r_u)<0. Expanding ueff(r)u_{\mathrm{eff}}(r) to second order around rur_u, the integral is approximated as

∫rs1rs2e2ε2ueff(r) dr≈πε2ueff′′(ru)e2ε2ueff(ru).(27)\int_{r_{s1}}^{r_{s2}}e^{\frac{2}{\varepsilon^2}u_{\mathrm{eff}}(r)}\,\mathrm{d}r\approx\sqrt{\frac{\pi\varepsilon^2}{u''_{\mathrm{eff}}(r_u)}}e^{\frac{2}{\varepsilon^2}u_{\mathrm{eff}}(r_u)}. \tag*{(27)}

Substituting Eqs. (25), (26), and (27) into Eq. (21), we obtain an expression for the radial flux Jr(ru,t)J_r(r_u,t) in terms of PLSP_{\mathrm{LS}} and PSSP_{\mathrm{SS}}:

Jr(ru,t)=R1PSS−R2PLS,(28)J_r(r_u,t)=R_1P_{\mathrm{SS}}-R_2P_{\mathrm{LS}}, \tag*{(28)}

where R1R_1 and R2R_2 are given by

R1=12π−ueff′′(rs1)ueff′′(ru)e−2ϵ2(ueff(ru)−ueff(rs1)),(29)R_1 = \frac{1}{2\pi}\sqrt{-u_{\mathrm{eff}}^{\prime\prime}(r_{s1})u_{\mathrm{eff}}^{\prime\prime}(r_u)}e^{-\frac{2}{\epsilon^2}(u_{\mathrm{eff}}(r_u)-u_{\mathrm{eff}}(r_{s1}))}, \tag*{(29)}
R2=12π−ueff′′(rs2)ueff′′(ru)e−2ϵ2(ueff(ru)−ueff(rs2)).(30)R_2 = \frac{1}{2\pi}\sqrt{-u_{\mathrm{eff}}^{\prime\prime}(r_{s2})u_{\mathrm{eff}}^{\prime\prime}(r_u)}e^{-\frac{2}{\epsilon^2}(u_{\mathrm{eff}}(r_u)-u_{\mathrm{eff}}(r_{s2}))}. \tag*{(30)}

Substituting Eq. (28) into Eq. (19) and using PSS=1−PLSP_{\mathrm{SS}}=1-P_{\mathrm{LS}}, we obtain the following differential equation for the time evolution of PLSP_{\mathrm{LS}}:

dPLSdt=−(R1+R2)PLS+R1.(31)\frac{dP_{\mathrm{LS}}}{dt}=-(R_1+R_2)P_{\mathrm{LS}}+R_1. \tag*{(31)}

The solution of this equation is

PLS=R1R1+R2+(PLS0−R1R1+R2)e−(R1+R2)t,(32)P_{\mathrm{LS}}=\frac{R_1}{R_1+R_2}+\left(P_{\mathrm{LS}}^0-\frac{R_1}{R_1+R_2}\right)e^{-(R_1+R_2)t}, \tag*{(32)}

where PLS0P_{\mathrm{LS}}^0 denotes the initial value of PLSP_{\mathrm{LS}}. For an initial probability density uniformly distributed over the interval 2≤r≤32\leq r\leq3, this initial value is given by PLS0=(3−ru)/(3−2)P_{\mathrm{LS}}^0=(3-r_u)/(3-2). (32) depends on the three radii rs1r_{s1}, rur_u, and rs2r_{s2}, which are the nonzero solutions of N(r2)=bN(r^2)=b. Since this equation is not trivial to solve explicitly, these radii can be approximated by fitting simple functions to the corresponding branches, yielding

rs1=−0.0938(1.238−b)0.4528+2.354,(33)r_{s1}=-0.0938(1.238-b)^{0.4528}+2.354, \tag*{(33)}
ru=−0.0932b+1.168+0.08851.238−b+2.5024,(34)r_u=-0.0932\sqrt{b}+1.168+0.0885\sqrt{1.238-b}+2.5024, \tag*{(34)}
rs2=0.0989(b+1.168)0.4592+2.643.(35)r_{s2}=0.0989(b+1.168)^{0.4592}+2.643. \tag*{(35)}

All three fits have a coefficient of determination of approximately R2=0.9998R^2=0.9998.

A colormap of the estimate of PLSP_{\mathrm{LS}} obtained from Eq. (32) is shown in Figure 7(a). Here, the time is set to t=1000t=1000, matching the duration of the numerical simulations. Equation (32) reproduces the numerical estimate of PLSP_{\mathrm{LS}} in the deterministic birhythmicity region for low noise, and the resulting map closely resembles that shown in Figure 4(a). For very low noise, for example ϵ≤0.01\epsilon\leq0.01, both R1R_1 and R2R_2 are negligible, so the exponential term in Eq. (32) can be approximated by 1. Therefore, PLSP_{\mathrm{LS}} remains close to its initial value PLS0P_{\mathrm{LS}}^0, which, as discussed above, is equal to 3−ru3-r_u. Along the deterministic birhythmicity interval, this quantity ranges from 0.36 to 0.65, in agreement with Figure 4(a) and Figure 7(a). For moderately low noise, namely 0.01<ε<0.30.01 < \varepsilon< 0.3, the exponential term in Eq. (32) can no longer be approximated by 1. In this case, both PLS0P_{\mathrm{LS}}^{0} and R1/(R1+R2)R_1/(R_1+R_2) contribute significantly to PLSP_{\mathrm{LS}}. In terms of accuracy, the difference between the value of PLSP_{\mathrm{LS}} predicted by Eq. (32) and that obtained from numerical simulations is below 0.05 for most values of (b,ε)(b,\varepsilon), and rarely exceeds 0.08.

Figure 7. Theoretical estimation of the probability PLSP_{\mathrm{LS}} of observing the largest limit cycle, given by to Eq. (32), as a function of (b,ε)(b,\varepsilon), for (a) t=1000t=1000 and (b) t=∞t=\infty.

The estimate of PLSP_{\mathrm{LS}} given by Eq. (32) also captures the behavior expected at stochastic equilibrium, as illustrated in Figure 7(b). The result is very similar to that obtained from the stationary probability distribution in Figure 4(b). As t→∞t \to\infty, the exponential term in Eq. (32) vanishes, and PLSP_{\mathrm{LS}} converges to R1/(R1+R2)R_1/(R_1+R_2), which can be written as 1/(1+R2/R1)1/(1+R_2/R_1). From Eqs. (29) and (30), we obtain

R2R1=ueff′′(rs2)ueff′′(rs1)e2ε2(ueff(rs2)−ueff(rs1)).(36)\frac{R_2}{R_1}=\sqrt{\frac{u_{\mathrm{eff}}^{\prime\prime}(r_{s2})}{u_{\mathrm{eff}}^{\prime\prime}(r_{s1})}e^{\frac{2}{\varepsilon^2}(u_{\mathrm{eff}}(r_{s2})-u_{\mathrm{eff}}(r_{s1}))}}. \tag*{(36)}

As discussed above, there is a threshold value of the control parameter, approximately b=−0.021b=-0.021, at which the two attractors have the same potential depth, i.e., u(rs1)=u(rs2)u(r_{s1})=u(r_{s2}). The limiting value of PLSP_{\mathrm{LS}} as ε→0\varepsilon\to0 depends on whether bb lies below or above this threshold. If b<−0.021b<-0.021, then u(rs1)<u(rs2)u(r_{s1})<u(r_{s2}) (see Figure 5(a)). Since ueff(r)→u(r)u_{\mathrm{eff}}(r)\to u(r) as ε→0\varepsilon\to0, the quantity 2ε2(ueff(rs2)−ueff(rs1))\frac{2}{\varepsilon^2}(u_{\mathrm{eff}}(r_{s2})-u_{\mathrm{eff}}(r_{s1})) diverges to +∞+\infty. Consequently, R2/R1R_2/R_1 also diverges, and therefore 1/(1+R2/R1)1/(1+R_2/R_1) converges to 0. On the other hand, if b>−0.021b>-0.021, then u(rs1)>u(rs2)u(r_{s1})>u(r_{s2}) (see Figure 5(b)). In this case, 2ε2(ueff(rs2)−ueff(rs1))\frac{2}{\varepsilon^2}(u_{\mathrm{eff}}(r_{s2})-u_{\mathrm{eff}}(r_{s1})) di verges to −∞-\infty as ε→0\varepsilon\to0. Hence, R2/R1R_2/R_1 converges to 0, and 1/(1+R2/R1)1/(1 + R_2/R_1) converges to 1. This explains the step-like structure of the colormap shown in Figure 7(b).

Double saddle-node bifurcation model with localized noise

Here, the model described in § 3 is modified to emulate the asymmetric behavior of the Hindmarsh–Rose model, in which noise is likely to induce transitions between coexisting bursting attractors, or their ghosts, only in regions where the attractors lie close to one another. Rather than breaking the axial symmetry of the deterministic attractors in the simple model, we break the axial symmetry of the noise by restricting it to a specific angular sector, namely −π/4<θ<π/4-\pi/4 < \theta< \pi/4, corresponding to the yellow region in Figure 8. This region is relatively small compared with the rest of phase space, in analogy with the Hindmarsh–Rose model, where the coexisting limit-cycle trajectories remain close to one another in the bursting region, which is also relatively small compared with the rest of the trajectories. The model equation reads

Figure 8. Numerical simulations of Eqs. (37), (38), (39) for ε=1.5\varepsilon= 1.5 and different values of bb: (a) b=−1.25b = -1.25, (b) b=0b = 0, and (c) b=1.25b = 1.25. Results from 100 simulations with dispersed initial conditions are shown in phase space, together with the active region, shown in yellow. The radial distance rr is computed each time the trajectory crosses the line θ=−π/2\theta= -\pi/2, shown as a solid green line, and the corresponding histogram is constructed. Each histogram also shows the radii of the stable and unstable deterministic limit cycles, indicated by solid and dashed red lines, respectively. In panels (a) and (c), the radius of the nearest saddle-node bifurcation is also shown as a dash-dotted red line.

drdt=F(r)+ϵχ(θ)ξ(t),(37)\frac{d\mathbf{r}}{dt} = \mathbf{F}(\mathbf{r}) + \epsilon\chi(\theta)\boldsymbol{\xi}(t), \tag*{(37)}

where F(r)\mathbf{F}(\mathbf{r}) is the deterministic field defined in Eqs. (3),(4),(5). In this case, the angular velocity is fixed at ω=50\omega= 50 so that more rotations around the origin are performed and more data can be collected in each simulation. This choice is nevertheless arbitrary. The bivariate, uncorrelated white-noise process ξ(t)\boldsymbol{\xi}(t) is the same as in Eq. (2),

⟨ξi(t)⟩=0,⟨ξi(t)ξj(t′)⟩=δijδ(t′−t),(38)\begin{aligned} \langle\xi_i(t) \rangle&= 0,\\ \langle\xi_i(t)\xi_j(t') \rangle&= \delta_{ij}\delta(t' - t), \tag*{(38)} \end{aligned}

Finally, χ(θ)\chi(\theta) is a step function,

χ(θ)={1if −π4<θ<π4,0.01otherwise.(39)\chi(\theta) = \begin{cases} 1 & \text{if } -\frac{\pi}{4} < \theta< \frac{\pi}{4},\\ 0.01 & \text{otherwise}. \end{cases} \tag*{(39)}

Thus, noise can trigger transitions between the circumferential stable limit cycles, or their ghosts, only within the region θ∈(−π/4,π/4)\theta\in(-\pi/4,\pi/4). We will refer to this sector as the active region and the remainder of phase space as the quiescent region. Ultimately, the system approaches either an attractor or a ghost in the quiescent region, depending on its state upon leaving the active region. This behavior is illustrated in Figure 8, which shows numerical simulations of Eqs. (37), (38), (39). To determine which state the system approaches in the quiescent region, the radial distance rr is recorded each time the trajectory crosses the negative part of the yy-axis, θ=−π/2\theta= -\pi/2, and a histogram is constructed. Inside the deterministic birhythmicity region (Figure 8(b)), the system may converge to either stable limit cycle, as reflected by the two histogram peaks located at their respective radii. Outside, but close to, the deterministic birhythmicity region (Figure 8(a) and Figure 8(c)), only one stable limit cycle exists; nevertheless, due to the ghost phenomenon, the system can also approach the precursor of the other cycle that is about to emerge at the saddle-node bifurcation. This is again reflected in the histograms, which show one peak at the radius of the stable cycle and another at the radius corresponding to the nearby saddle-node bifurcation.

The statistical behavior of the model defined by Eqs. (37), (38), (39) is summarized in Figure 9. For different values of (b,ϵ)(b,\epsilon), the radial distance rr, measured at the end of the quiescent region, is represented using scatter plots. In the low-noise regime, as shown in Figure 9(a), the points closely follow the deterministic bifurcation diagram. As the noise intensity increases (Figure 9(b) and (Figure 9c)), the scatter broadens, and the deterministic stable branches are no longer fully populated near the saddle-node bifurcations. This behavior arises because, in these parameter regions, the system is bistable, but one of the stable branches is significantly less stable than the other, corresponding to a shallower potential well (see Figure 5). Consequently, trajectories tend to remain near the deeper well, while noise-induced transitions to the less stable branch become unlikely during the active phase.

Figure 9. Stochastic bifurcation diagrams of Eqs. (37),(38),(39) for different noise levels: (a) ε=0.05\varepsilon= 0.05, (b) ε=0.2\varepsilon= 0.2, (c) ε=0.9\varepsilon= 0.9, and (d) ε=2.5\varepsilon= 2.5. For each pair (b,ε)(b,\varepsilon), 100 simulations are performed. Each time the trajectory crosses the lower part of the yy-axis, corresponding to θ=−π/2\theta= -\pi/2, the distance rr to the origin is recorded and plotted as a single point. The color of the points encodes their concentration, with blue indicating low density and green indicating high density. The deterministic stable and unstable limit-cycle branches are also shown as solid and dashed red lines, respectively.

This behavior closely parallels that observed in the symmetric version of the model discussed in § 3 (see Figure 4(a)). At still higher noise levels, as shown in Figure 9(d), the scatter becomes more broadly distributed. However, both stable branches become populated again, since the noise is now sufficiently strong to induce transitions in both directions during the active phase. The ghost phenomenon also becomes evident in this regime, as the scatter extends beyond the deterministic stable branches and the saddle-node bifurcations, thereby producing an enlarged stochastic birhythmicity region.

A more complete summary of the behavior of the model with localized noise over the (b,ε)(b,\varepsilon) parameter plane is shown in Figure 10. As discussed above, the system may approach two possible states during the quiescent region: either two stable limit cycles, or one stable limit cycle and a ghost, depending on whether bb lies inside or outside the birhythmicity region. For each pair (b,ε)(b,\varepsilon), simulations of Eqs. (37), (38), (39) are performed over a time interval Δt=1000\Delta t = 1000, using initial values of rr uniformly distributed between r=2r = 2 and r=3r = 3. The values of rr are recorded whenever the trajectory crosses the line θ=−π/2\theta= -\pi/2, and the fraction of crossings for which rr is closer to the radius of the larger state is computed. For this calculation, the boundary between the two states is chosen as the mean of their radii, an admittedly arbitrary choice. The results qualitatively reproduce the hourglass-shaped profile observed in both the symmetric model of § 3 and the Hindmarsh–Rose model.

Figure 10. Numerical estimation of the probability of observing the largest limit cycle, or its ghost, in the stochastic system, Eqs. (37), (38), (39), as a function of (b,ε)(b,\varepsilon). For each parameter set, 100 simulations with uniformly distributed initial conditions and duration Δt=1000\Delta t = 1000 are performed. After t=500t = 500, the values of rr are recorded each time the trajectory crosses the line θ=−π/2\theta= -\pi/2, and the fraction of crossings for which rr lies above the mean radius of the stable limit cycles, or their ghosts, is computed. The dashed black lines indicate the boundaries of the deterministic birhythmicity region, while the solid red lines mark the estimated boundaries of the de facto stochastic birhythmicity region, defined as the region in which both states occur with probability at least 2%.

As before, we define the de facto stochastic birhythmicity region as the set of (b,ε)(b,\varepsilon) values for which both states have a probability of observation of at least 2% (although this cutoff is arbitrary). For ε<10−1\varepsilon< 10^{-1}, the stochastic birhythmicity region coincides with its deterministic counterpart, since the noise is too weak to induce transitions and the outcome depends primarily on the initial conditions. As ε\varepsilon increases, the stochastic birhythmicity region narrows, reaching a bottleneck around ε≈6×10−1\varepsilon\approx6 \times10^{-1}. For larger noise intensities, the region broadens again, becoming wider than the deterministic region at ε∼100\varepsilon\sim10^{0}, where fluctuations are strong enough to induce transitions between the stable limit cycles or their ghosts. Finally, outside the stochastic birhythmicity region, only the smaller state on the left, or the larger state on the right, is likely to be observed. It is important to note that, as in the homogeneous-noise case, the hourglass-like structure described above reflects finite-time behavior rather than stochastic equilibrium in the low-noise regime. Reaching stochastic equilibrium, and thus obtaining a cup-shaped map similar to that in Figure 4(b), would require exceedingly long simulations. Moreover, such a map cannot be obtained from a simple radial integral such as Eq. (13), since the localized-noise model is not axially symmetric.

Conclusions

In this paper, we have studied the effects of noise on birhythmic systems in which two limit cycles coexist within a region bounded by two saddle-node bifurcations. The Hindmarsh–Rose model, which exhibits multiple such regions where bursting periodic regimes coexist, is of particular interest. When noise is introduced into this model, noise-induced transitions between bursting regimes are observed even beyond the saddle-node bifurcations. In other words, birhythmic behavior may arise in parameter regions where it is not expected deterministically.

To provide an accessible description of this stochastic birhythmicity phenomenon, we first proposed a simple, axially symmetric model with homogeneous noise, in which two circumferential limit cycles coexist within a region delimited by two saddle-node bifurcations. The effect of noise intensity on birhythmicity was analyzed numerically by estimating, over the (b,ε)(b,\varepsilon) parameter plane, the probability of observing the system near the largest limit cycle or its ghost. This probability displays an hourglass-shaped region of stochastic birhythmicity embedded between two regions of monorhythmicity. For sufficiently intense noise, this stochastic birhythmicity region extends beyond its deterministic counterpart. However, the hourglass-shaped profile is a finite-time effect: at low noise intensity, the system does not reach stochastic equilibrium within the simulation time, since the corresponding transition times become exceedingly long. A theoretical estimate based on the stationary probability distribution accounts for the high-noise regime, where stochastic equilibrium is reached within the simulation time, whereas the low-noise finite-time behavior is captured by an approximate approach inspired by Kramers’ escape problem. This suggests that similar hourglass-shaped profiles may also arise in finite-time simulations of the Hindmarsh–Rose model and in other systems exhibiting birhythmicity bounded by saddle-node bifurcations of limit cycles. Since the phenomenon is captured by both a simple model and a complex neuronal model, it may also be observable in experimental setups with similar characteristics.

To account for the lack of symmetry in the Hindmarsh–Rose model, the simplified model was then modified by introducing non-homogeneous noise with enhanced intensity in a specific region of phase space. As a result, noise-induced transitions between the two limit cycles occur preferentially within this region, reproducing the fact that, in the Hindmarsh–Rose model, transitions between attractors take place mainly when the system evolves in the vicinity of both limit cycles. For this asymmetric model, a similar numerical analysis was performed, and an hourglass-shaped stochastic birhythmicity region was again observed. This supports the idea that such a structure is a recurrent feature of finite-time numerical simulations and experimental observations of systems displaying birhythmicity bounded by saddle-node bifurcations [23, 24]. Moreover, the existence of a noise intensity that minimizes the probability of transitions between the two limit cycles, corresponding to the bottleneck of the hourglass, appears to be a robust characteristic of this class of systems.

Acknowledgments

This work has been supported by the Spanish State Research Agency (AEI) and the European Regional Development Fund (ERDF, EU) under Project No. PID2023-148160NB-I00 (MCIN/AEI/10.13039/501100011033).

Data Availability Statement

The data that support the findings of this study are available from the corresponding author, upon reasonable request.

References

  1. [1]Alvar Daza, Alexandre Wagemakers, and Miguel A. F. Sanjuán. Multistability and unpredictability. Phys. Today, 77(11):44–50, November 2024.DOI
  2. [2]Steven H. Strogatz. Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry, and Engineering. Westview Press, Boulder, CO, second edition, 2015.
  3. [3]Eugene M. Izhikevich. Dynamical Systems in Neuroscience: The Geometry of Excitability and Bursting. MIT Press, Cambridge, MA, 2007.DOI
  4. [4]Evdokiia Slepushkina, Irina Bashkirtseva, Philipp Kügler, and Lev Ryashko. Noise-driven bursting birhythmicity in the Hindmarsh–Rose neuron model. Chaos, 33(3):033106, 2023.
  5. [5]Irina Bashkirtseva, Lev Ryashko, and Svetlana Zaitseva. Stochastic sensitivity analysis of noise-induced transitions in a biochemical model with birhythmicity. J. Phys. A: Math. Theor., 53(26):265601, 2020.DOI
  6. [6]Yu. M. Shukrinov, I. R. Rahmonov, K. V. Kulikov, A. E. Botha, A. Plecenik, P. Seidel, and W. Nawrocki. Modeling of LC-shunted intrinsic Josephson junctions in high-Tc superconductors. Supercond. Sci. Technol., 30(2):024006, 2017.DOI
  7. [7]R. Arumugam, T. Banerjee, and P. S. Dutta. Rhythmogenesis, birhythmicity and chaos in a metapopulation model. Eur. Phys. J. Spec. Top., 226(9):2145–2156, 2017.DOI
  8. [8]Javier Used, Jesús M. Seoane, Irina Bashkirtseva, Lev Ryashko, and Miguel A. F. Sanjuán. Synchronization of two non-identical Chialvo neurons. Chaos Solitons Fractals, 183:114888, 2024.arxiv.org/abs/2404.07783
  9. [9]Joel Zirkle and Leonid L. Rubchinsky. Noise effect on the temporal patterns of neural synchrony. Neural Netw., 141:30–39, 2021.DOI
  10. [10]Lev Ryashko and Irina Bashkirtseva. Noise-induced chaos and generation of phantom attractors in a birhythmic neuron model. Chaos Solitons Fractals, 191:115841, 2025.DOI
  11. [11]Zhuan Shen, Honghui Zhang, Lin Du, Zichen Deng, and Jürgen Kurths. Initiation and termination of epilepsy induced by Lévy noise: A view from the cortical neural mass model. Chaos Solitons Fractals, 167:113038, 2023.DOI
  12. [12]J. L. Hindmarsh, R. M. Rose, and Andrew Fielding Huxley. A model of neuronal bursting using three coupled first-order differential equations. Proc. R. Soc. Lond. B Biol. Sci., 221(1222):87–102, 1984.pubmed.ncbi.nlm.nih.gov/6144106
  13. [13]Roberto Barrio, Santiago Ibáñez, and Lucía Pérez. Homoclinic organization in the Hindmarsh–Rose model: A three-parameter study. Chaos, 30(5):053132, 2020.DOI
  14. [14]Ignacio Ortega-Piwonka, Javier Used, Jesús M. Seoane, and Miguel A. F. Sanjuán. Stochastic birhythmicity outside the coexistence region in the Hindmarsh–Rose model. Phys. Rev. E, 111:044217, 2025.
  15. [15]Crispin W. Gardiner. Handbook of Stochastic Methods for Physics, Chemistry and the Natural Sciences, volume 13 of Springer Series in Synergetics. Springer-Verlag, Berlin, third edition, 2004.
  16. [16]H. A. Kramers. Brownian motion in a field of force and the diffusion model of chemical reactions. Physica, 7(4):284–304, 1940.DOI
  17. [17]S. Chandrasekhar. Stochastic problems in physics and astronomy. Rev. Mod. Phys., 15:1–89, 1943.
  18. [18]Adriaan D. Fokker. Die mittlere Energie rotierender elektrischer Dipole im Strahlungsfeld. Ann. Phys. (Leipzig), 348(5):810–820, 1914.DOI
  19. [19]Max Planck. Über einen Satz der statistischen Dynamik und seine Erweiterung in der Quantentheorie. Sitzungsber. K. Preuss. Akad. Wiss. Berlin, 1917.
  20. [20]Andrey Kolmogorov. Über die analytischen Methoden in der Wahrscheinlichkeitsrechnung. Math. Ann., 104(1):415–458, 1931.
  21. [21]Alison L. Gibbs and Francis E. Su. On choosing and bounding probability metrics. Int. Stat. Rev., 70(3):419–435, 2002.
  22. [22]Ignacio Ortega, Marcel G. Clerc, Claudio Falcón, and Nicolás Mujica. Subharmonic wave transition in a quasi-one-dimensional noisy fluidized shallow granular bed. Phys. Rev. E, 81:046208, 2010.pubmed.ncbi.nlm.nih.gov/20481808
  23. [23]Zhihui Wang, Yanying Yang, and Lixia Duan. Dynamic mechanism of epileptic seizures induced by excitatory pyramidal neuronal population. Electron. Res. Arch., 31(8):4427–4442, 2023.
  24. [24]L. B. Tagne Nkounga, F. M. Moukam Kakmeni, and R. Yamapi. Birhythmic oscillations and global stability analysis of a conductance-based neuronal model under ion-channel fluctuations. Chaos Solitons Fractals, 159:112126, 2022.

Paper details

vlnc:W_7jx7x629
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents