Dynamical systems analysis of spikeadding mechanisms in transient bursts
 Jakub Nowacki^{1},
 Hinke M Osinga^{2}Email author and
 Krasimira TsanevaAtanasova^{1}
DOI: 10.1186/2190856727
© Nowacki et al.; licensee Springer 2012
Received: 24 August 2011
Accepted: 13 February 2012
Published: 24 April 2012
Abstract
Transient bursting behaviour of excitable cells, such as neurons, is a common feature observed experimentally, but theoretically, it is not well understood. We analyse a fivedimensional simplified model of afterdepolarisation that exhibits transient bursting behaviour when perturbed with a short current injection. Using oneparameter continuation of the perturbed orbit segment formulated as a wellposed boundary value problem, we show that the spikeadding mechanism is a canardlike transition that has a different character from known mechanisms for periodic burst solutions. The biophysical basis of the model gives a natural timescale separation, which allows us to explain the spikeadding mechanism using geometric singular perturbation theory, but it does not involve actual bifurcations as for periodic bursts. We show that unstable sheets of the critical manifold, formed by saddle equilibria of the system that only exist in a singular limit, are responsible for the spikeadding transition; the transition is organised by the slow flow on the critical manifold near folds of this manifold. Our analysis shows that the orbit segment during the spikeadding transition includes a fast transition between two unstable sheets of the slow manifold that are of saddle type. We also discuss a different parameter regime where the presence of additional saddle equilibria of the full system alters the spikeadding mechanism.
Keywords
burst spike adding transient behaviour dynamical systems geometric singular perturbation theory1 Introduction
How a single spike or a burst of spikes is generated and regulated for neuron cells is one of the most fundamental questions in neuroscience [1]. Spike generation is closely related to neuronal excitability, which is the ability of the cell’s membrane potential to undergo a large excursion, called an action potential or a spike, when subjected to a sufficiently strong stimulus [1–3]. An excitable cell either responds in full to such a stimulus or not at all, which allows for a reliable transmission of information. Therefore, the different mechanisms for excitability and bursting have been widely studied; we refer to Izhikevich [4] for a comprehensive overview of mechanisms for neurons. Excitable behaviour has also been reported to occur in many other types of cells [1–3] as well as in physical systems, such as lasers [5, 6], electronic circuits [7–10] and chemical reactions [11]. Neuronal excitability can be very sensitive to even relatively small changes in, for example, the biophysical properties of a neuron [12–14] or its morphology [15]. This parameter sensitivity indicates that dynamical systems theory is particularly suited for explaining the rich dynamics found in excitable systems.
The classification of different bursting mechanisms was pioneered by Rinzel [16], who used a system decomposition into slow and fast subsystems. He showed that the burst can be divided into active (spiking) and silent phases, which follow different types of attractors of the fast subsystem. Hence, a classification of the bursting oscillators is provided by the structure of the bifurcation diagram of the fast subsystem. Rinzel’s classification was extended in Izhikevich’s studies [1, 4]. As an alternative approach, Golubitsky et al. [17] used singularity theory to classify the bifurcation diagrams of the fast subsystem and, hence, the different bursting mechanisms; see also [18]. These studies are primarily for systems with one slow variable; see Smolen et al. [19] for an extension to two slow variables and Ermentrout and Terman [3] for a summary of these ideas along with new results.
A classification of bursting mechanisms, however, does not answer questions about the number of spikes in a particular burst of the same type nor does it explain possible transitions between spiking and bursting. While the latter has received considerable attention over the years, the former has hardly been addressed. Terman [20] analysed transitions between bursting and tonic (continuous) spiking in a pancreatic βcell model. He recognised the importance of connecting classical slowfast analysis with full system bifurcation analysis and identified bifurcations of the periodic bursting solutions that organise the transitions between different parameter regimes. He further studied chaotic spiking that can arise in between such transitions [21]. Recently, Benes et al. [22] and Kramer et al. [23] reported that a new type of torus canard can play a role in the transition from spiking to bursting in a model of Purkinje cells. Other recent studies particularly focus on spikeadding in a periodic bursting oscillator; for example, see Govaerts and Dhooge [24], Guckenheimer and Kuehn [25], TsanevaAtanasova et al. [26] and Linaro et al. [27]. These studies show that the spikeadding mechanism is formed by a pair of saddlenode bifurcations of periodic orbits of the full system; bursts with different numbers of spikes are, in fact, different periodic attractors of the full system that may coexist only if the number of spikes differs by one [20, 26]. The recent study by Teka et al. [28] explains how one may predict the precise number of spikes in a burst; here, the use of two slow variables is essential. Methods to regulate the number of spikes are reported by Ghigliazza and Holmes [29], where a minimal HodgkinHuxley model of bursting is proposed to analyse spikeadding and transitions between bursting and tonic spiking in a more general context.
As discussed by Izhikevich [4], the mechanisms for generating spikes do not depend on whether the neuron exhibits spiking only as a transient phenomenon when subjected to a strong enough stimulus or whether it is spiking or bursting continuously. This is indeed the case if transient bursting is organised by the applied stimulus. Applying a stimulus has the effect of changing the righthand side of the underlying system of ODEs. The bifurcation diagram of the corresponding fast subsystem typically no longer has a stable equilibrium that corresponds to the resting potential, and a spike or burst arises from new attractors that exist only when the stimulus is on. As soon as the stimulus is switched off, the system relaxes back to the resting potential. Therefore, the mechanism is due to a change in the structure of the bifurcation diagram, which depends on the strength of the applied stimulus. Studies of this nature, where the type of burst is studied in dependence on the strength or duration of the stimulus, have been done, for example, by Tran et al. [30], Kim et al. [31] and Stern et al. [32].
In this paper, we investigate spikeadding in a transient burst in the model of hippocampal pyramidal neurons from Nowacki et al. [13]. In contrast to the above studies, the spikeadding occurs after the applied stimulus has been switched off. Hence, the bursting behaviour is governed by the underlying bifurcation structure of the original system, for which a stable equilibrium exists. The strength of the applied stimulus must be such that a first action potential is generated, but the stimulus is not responsible for generating any additional spikes. Our investigation can be compared to excitability in laser systems [5, 6] where the response after an applied stimulus is explained by the existence of a nearby homoclinic bifurcation with respect to a parameter. This mechanism is due to the presence of a saddle equilibrium that coexists with the stable equilibrium (i.e., the resting potential). The interesting aspect about our model is that there exists only a stable equilibrium, and we have been unable to identify any nearby saddletype invariant object in the parameter region of interest that could organise homoclinic bifurcations. We explain the transient behaviour following the ideas of Geometric Singular Perturbation Theory (GSPT) [33–36].
Another important difference with existing studies is that we do not rely on simulations to study the effects of an applied stimulus in numerical experiments. Instead, we use a continuationbased approach and show the process of spikeadding transitions during a transient burst in unprecedented detail, which is not possible using bruteforce simulation. Our numerical method is based on the continuation of orbit segments as solutions of a twopoint boundary value problem; this approach has already been applied to the bifurcation analysis of periodic orbits, including homoclinic or heteroclinic bifurcations [37], and more recently for the computation of invariant manifolds [38] and socalled slow manifolds in systems with multiple time scales [39, 40]. We divide the system into two separate orbit segments, with and without current injection, which are coupled only by the boundary conditions. This allows us to continue the orbit segments in a chosen parameter and analyse the precise nature of their continuous deformation even over exponentially small parameter variations. We complement this continuation with the computation of the twodimensional critical manifold of the fast subsystem, which comprises all the equilibria of the fast subsystem parametrised by two slow variables. The critical manifold is folded and consists of a number of stable and unstable sheets. In our setting, the unstable sheets are of saddle type, and we will refer to them as saddleunstable sheets. The fast subsystem also has families of periodic orbits that emanate from Hopf bifurcations on the critical manifold, which give rise to the spiking behaviour during a burst. Our analysis shows that, during the spikeadding transition, orbit segments trace saddleunstable slow manifolds that lie very close to corresponding saddleunstable sheets of the critical manifold; the distance between these two manifolds is of the same order as the ratio between the contraction/expansion rates towards and on the manifold, which is organised by the difference between the slow and fast time scales [33, 41]. This canardlike behaviour is very similar to behaviour during a spikeadding transition of a periodic burst [25], but it does not involve bifurcations, and coexistence of bursts with different numbers of spikes is not possible here. Thus, our analysis indicates that it is the presence of canardlike behaviour that organises the spike adding.
As for periodic bursting, the spikeadding transition in our model occurs over an exponentially small parameter interval [42–44]. Within this exponentially small parameter interval, we find an even smaller parameter interval during which the canardlike orbit segment includes a fast transition from a saddleunstable to another saddleunstable slow manifold. This phenomenon is similar to the socalled foldinitiated canards that have been observed for periodic orbits [45]. To understand this behaviour, we study the associated slow flow on the critical manifold and identify the effect of folds and folded singularities on the behaviour of the orbit segment. Our findings complement the study for planar systems by Guckenheimer et al. [45] and confirm that such foldinitiated canardlike behaviour occurs robustly during a spikeadding transition due to continuity of the vector field with respect to the parameter.
Our study concerns the analysis of a spikeadding mechanism where the full system has a unique stable equilibrium that does not undergo any bifurcations. We find that, for different values of model parameters, the system can have additional unstable equilibria that alter the nature of the spikeadding mechanism. More precisely, the appearance of two saddle equilibria on the critical manifold suppresses the foldinitiated transition between saddleunstable sheets and changes the behaviour of the orbit segment. The presence of the saddle equilibria give rise to a mechanism that is reminiscent of the excitability studies by Wieczorek et al. [5] and Krauskopf et al. [6]; a new spike is added via a heteroclinic connection of the perturbed state to a saddle equilibrium of the full system. However, this saddle equilibrium lies outside a neighbourhood of the resting potential, and we observe a canardlike transition before and after the heteroclinic connection that is similar to the canardlike transition that occurs when no additional equilibria are present. We compute the critical manifold for this situation and study the associated slow flow to explain this phenomenon; we find that the saddle equilibrium point lies in the middle of a saddleunstable sheet of the critical manifold, and it is an attractor with respect to the slow flow on the critical manifold.
This paper is organised as follows: In the next section we present our model of the study. In Section 3, we numerically identify the mechanism of spikeadding via continuation of the orbit segment. Next, in Section 4, we calculate the critical manifold of the fast subsystem along with the slow flow to explain the transition to bursting. Here, we also investigate the transition between two unstable slow manifolds of the saddle type and the changes in the spikeadding mechanism when additional equilibria of the full system are present. We end with a discussion in Section 5.
2 The model
We apply our analysis of transient bursts to pyramidal neuron cells from the CA1 and CA3 regions of the hippocampus. A detailed model of such neurons in HodgkinHuxley formalism was presented by Nowacki et al. [13], but for the purpose of this paper, we study a reduced version of this model. The simplified model consists of four ionic currents, namely, fast and slow inward currents, denoted as ${I}_{\mathrm{FI}}$ and ${I}_{\mathrm{SI}}$, respectively, and fast and slow outward currents, denoted as ${I}_{\mathrm{FO}}$ and ${I}_{\mathrm{SO}}$, respectively. Inward currents are responsible for the depolarisation or increase of the membrane potential, whereas outward currents hyperpolarise or decrease the membrane potential and return the cell back to its resting state (a stable equilibrium) [2, 46]. The fast inward current ${I}_{\mathrm{FI}}$ represents the fastest class of spikegenerating ${\mathrm{Na}}^{+}$ and ${\mathrm{Ca}}^{2+}$currents. The rates of change of these currents are usually similar to that of the membrane potential. Therefore, we assume that the gating of ${I}_{\mathrm{FI}}$ is instantaneous [13, 46–48]. The slow inward current ${I}_{\mathrm{SI}}$ mainly corresponds to the transient Ttype ${\mathrm{Ca}}^{2+}$current [13, 49–51] and represents the lowvoltage activated currents responsible for shaping the subthreshold behaviour of the model. The fast outward current ${I}_{\mathrm{FO}}$ represents highvoltage activated fast ${\mathrm{K}}^{+}$currents that we base on the delayed rectifier ${\mathrm{K}}^{+}$current [13, 46, 47]. Finally, ${I}_{\mathrm{SO}}$ represents muscarinicsensitive ${\mathrm{K}}^{+}$current [50, 52, 53], which has an activation rate of the same order as that of ${I}_{\mathrm{SI}}$.
Parameter values for the simplified model (1) as defined in (2)(4)

System (1) defined by Equations (2)(4) evolves on multiple time scales because ${C}_{\mathrm{m}}/{g}_{\mathrm{FO}}$ (as an approximation of the time scale for V) and ${\tau}_{x}$ with $x\in \{{m}_{\mathrm{SI}},{m}_{\mathrm{FO}},{m}_{\mathrm{SO}},{h}_{\mathrm{SI}}\}$ have different orders of magnitude. As indicated in Table 1${m}_{\mathrm{SO}}$ and ${h}_{\mathrm{SI}}$ are slow variables that vary on a time scale that is (roughly) 10 times slower than ${m}_{\mathrm{SI}}$ and ${m}_{\mathrm{FO}}$ and 100 times slower than V. In particular, this means that our model is capable of firing an arbitrarily large number of spikes during the ADP. More precisely, an increase in ${g}_{\mathrm{SI}}$, as in Figure 1 and throughout this paper, has the net effect that the slow variable ${h}_{\mathrm{SI}}$ becomes even slower so that more spikes can be fired during the time it takes for ${h}_{\mathrm{SI}}$ to relax back to its equilibrium value. In this paper, we are not interested in the exact nature of this process, but we mention here that a large number of spikes will also be accompanied by a noticeable reduction in oscillation amplitudes. As ${h}_{\mathrm{SI}}$ slows down, the dynamics will resemble more and more the behaviour organised by slow passage through a Hopf bifurcation [26, 54]. We focus on the process of spike adding and take advantage of the difference in time scales in Section 3, where it suffices to consider the timescale separation between the three fast variables V${m}_{\mathrm{SI}}$ and ${m}_{\mathrm{FO}}$, and the two slow variables ${m}_{\mathrm{SO}}$ and ${h}_{\mathrm{SI}}$.
The gating variables express the fractions of channels in a given state and naturally range over the interval $[0,1]$. The natural range of the membrane potential V is bounded by the two Nernst potentials [2, 46], i.e., ${E}_{\mathrm{O}}\le V\le {E}_{\mathrm{I}}$, where ${E}_{\mathrm{O}}=80.0\text{mV}$ and ${E}_{\mathrm{I}}=80.0\text{mV}$. It is beneficial for the numerical analysis if all variables vary over a similar range. Therefore, the computations are done using the scaled membrane potential $V/{k}_{v}$, where ${k}_{v}=100\text{mV}$. For our numerical investigations, we used the continuation package Auto[55, 56] for solving the boundary value problems. All numerical simulations were done with XPP [57] using the frontend package XPPy [58] in Python [59], and visualisations were done in Python using Matplotlib [60] and Mayavi [61].
3 Identifying the spikeadding mechanism
Spike adding happens after a current injection, that is, in the regime where ${I}_{\mathrm{app}}=0$. Hence, any numerical investigation of the transient behaviour must take into account a discontinuous jump from ${I}_{\mathrm{app}}>0$ to ${I}_{\mathrm{app}}=0$ on the righthand side of Equation (2). We view the orbit as a concatenation of two orbit segments that are the solution of two boundary value problems and define appropriate boundary conditions to account for the discontinuity in ${I}_{\mathrm{app}}$.
Figure 3b shows our starting solution, i.e., a single spike followed by ADP. Along the first plateau of the solution branch up to the first downward peak, all orbit segments are qualitatively like Figure 3b; in particular, the ADP is a small hump. As we follow the solutions into the downward peak, the hump of the ADP for the orbit segments stretches out as shown in Figure 3c, which lies at the bottom of the downward peak. Interestingly, as we follow the solution back up along the downward peak, the orbit segment generates a double step in the ADP, as shown in Figure 3d; we selected the orbit segment with the longest double step (with respect to time). As we continue to trace the solution up along the downward peak, the small spike at the end of the orbit segment grows into a fully developed spike, while the stretched double step retracts; the orbit segment shown in Figure 3e is representative of such a solution, and all orbit segments along the second plateau in Figure 3a are qualitatively like Figure 3e. Figure 3f represents a solution along the next plateau, which exhibits three spikes that are created via the same process as explained above for the twospike burst. In fact, the same spikeadding process takes place for all spikeadding transitions via the downward peaks in Figure 3a. We emphasise here that the stretched single and doublestep ADPs, as shown in Figure 3c,d, only exist along the downward peaks in Figure 3a, that is, in the exponentially small parameter interval during which a spikeadding transition occurs. Hence, such solutions are unlikely to be observed in actual experiments, and they are also very difficult to find in numerical experiments that use initialvalue integration methods. The fact that a spikeadding transition happens over an exponentially small parameter interval, during which the solution measure changes rapidly, suggests that it is organised by the difference in time scales present in system (1). Therefore, in order to obtain a better understanding, we proceed by using GSPT [34–36, 41, 62, 63].
4 Spikeadding organised by the critical manifold
Furthermore, because we are interested in spikeadding phenomena after the brief current injection, we set ${I}_{\mathrm{app}}=0$.
The important objects of study in the singular limit are equilibria and periodic orbits. Since ${m}_{\mathrm{SO}}$ and ${h}_{\mathrm{SI}}$ are parameters, these invariant objects occur in twoparameter families. The $({m}_{\mathrm{SO}},{h}_{\mathrm{SI}})$dependent families of equilibria are known as the critical manifold, which we denote by S. The equilibria on S can be stable or unstable, determined with respect to the threedimensional fast subsystem, and are typically separated by curves of fold or Hopf bifurcations. Similarly, we can expect the existence of $({m}_{\mathrm{SO}},{h}_{\mathrm{SI}})$dependent families of periodic orbits that emanate from a curve of Hopf bifurcations on the critical manifold; these periodic orbits can again be stable or unstable with respect to the threedimensional fast subsystem. Typically, the stable periodic orbits of the fast subsystem organise the spiking phase of the bursting oscillators [16, 19, 26].
The critical manifold S, when considered in the full fivedimensional phase space of system (1), is a twodimensional surface, or collection of surfaces, and the associated families of periodic orbits form a threedimensional manifold, or collection of manifolds, that we denote by P. Together, these collections of manifolds organise the behaviour of solutions of (1). If ${m}_{\mathrm{SO}}$ and ${h}_{\mathrm{SI}}$ vary slowly enough, then GSPT guarantees that a solution of (1) (with ${I}_{\mathrm{app}}=0$) will trace attracting sheets of S or P that correspond to the $({m}_{\mathrm{SO}},{h}_{\mathrm{SI}})$dependent families of attractors of the fast subsystem [41]. For example, the transient spikes of system (1) trace the manifold ${P}^{a}$ that corresponds to the family of attracting periodic orbits of the fast subsystem, while ${m}_{\mathrm{SO}}$ and ${h}_{\mathrm{SI}}$ are slowly varying [16, 64]. More precisely, solutions of (1) lie on socalled slow manifolds that are perturbations of the different sheets of S and P from the singular limit [41]. Solutions of (1) are characterised by fast transitions between, followed by exponential contraction onto the slow manifolds. The essential difference in behaviour during a spikeadding transition is the fact that the solution of (1) contains a segment that traces a slow manifold associated with a sheet of S that is unstable (of saddle type) rather than attracting; see also [25, 34, 36, 63]. While technically the solutions of (1) trace slow manifolds, we will abuse notation and write ‘sheet of S’ where we mean ‘slow manifold corresponding to the sheet of S.’
Stability properties of the critical manifold S and the manifold P of periodic orbits
Twodimensional critical manifold S  Threedimensional manifold P  

${S}_{1}^{a}$  ${S}_{1}^{r}$  ${S}_{2}^{a}$  ${S}_{2}^{r}$  ${S}_{3}^{r}$  ${S}_{3}^{a}$  ${P}^{r}$  ${P}^{a}$  
Dimension stable manifold  5  4  5  4  3  5  4  5 
Dimension unstable manifold  N/A  3  N/A  3  4  N/A  4  N/A 
The maxima and minima of the families of periodic orbits originating from H are shown in Figure 4c,d, using the same two viewpoints as in panels (a) and (b), respectively. The Hopf bifurcation is subcritical along the entire curve so that the emanating family of periodic orbits is unstable (of saddle type), with fourdimensional stable and unstable manifolds; we coloured this family magenta and labelled it ${P}^{r}$. The family of periodic orbits becomes stable via a fold of periodic orbits, after which it is coloured blue and labelled ${P}^{a}$, and ends in a homoclinic bifurcation involving equilibria on the sheet ${S}_{1}^{r}$. We refer again to Table 2 for an overview of the different families and their stability properties.
The spikeadding process occurs along the downward peak in Figure 3a, during which the parameter ${g}_{\mathrm{SI}}$ remains almost fixed, but the orbit segments of system (5)(8) change dramatically. We observe the formation of a stretched ADP, which initially gets increasingly longer and shortens again as we follow the orbit segments along the downward peak in Figure 3a. This transition is initiated by the fact that, at the special value ${g}_{\mathrm{SI}}\approx 0.5615{\text{mS/cm}}^{2}$, the injected current perturbs the orbit segment such that ${\mathbf{u}}_{\mathrm{ON}}(1)={\mathbf{u}}_{\mathrm{OFF}}(0)$ lies almost on the fourdimensional stable manifold of the saddleunstable sheet ${S}_{1}^{r}$; more precisely, ${S}_{1}^{r}$ has a corresponding slow manifold with a corresponding finitetime stable manifold that ${\mathbf{u}}_{\mathrm{ON}}(1)={\mathbf{u}}_{\mathrm{OFF}}(0)$ comes close to. The closer ${\mathbf{u}}_{\mathrm{ON}}(1)={\mathbf{u}}_{\mathrm{OFF}}(0)$ lies to this fourdimensional manifold, the closer the corresponding orbit segment comes to the slow manifold associated with ${S}_{1}^{r}$ and the longer it will trace this slow manifold. We approximate the slow manifold by the critical manifold ${S}_{1}^{r}$, and Figure 5a shows the orbit segment from Figure 3c, which traces ${S}_{1}^{r}$ almost up to the fold ${F}_{1}$ before it drops down to ${S}_{1}^{a}$ and converges to the resting potential.
Figure 5b,c,d,e,f illustrates orbit segments for the second upward part of the downward peak in Figure 3a. Interestingly, a doublestep ADP is created via a transition from ${S}_{1}^{r}$ to ${S}_{2}^{r}$ as shown. The orbit that was previously shown in Figure 3d traces both saddleunstable sheets ${S}_{1}^{r}$ to ${S}_{2}^{r}$ of S all the way up to the fold ${F}_{3}$. This transition from ${S}_{1}^{r}$ to ${S}_{2}^{r}$ is, in fact, a robust part of the spikeadding process that is a continuous (parameterdependent) variation of the case where the folds ${F}_{1}$ and ${F}_{2}$ are absent, that is, where ${S}_{1}^{r}$ connects directly to ${S}_{3}^{r}$; see Section 4.2 for more details. After reaching the top fold ${F}_{3}$, the membrane potential V initially increases instead of immediately decreasing down to the stable sheet ${S}_{1}^{a}$, and a small spike is created. As we continue to follow the solution up along the downward peak, the spike part of the orbit segment grows and moves back towards the attracting periodic orbit family ${P}^{a}$, as illustrated in Figure 5d,e. Finally, as shown in Figure 5f, the orbit segment traces ${S}_{1}^{r}$ for only a very short time before the second spike occurs; this orbit segment is selected almost at the end of the downward peak, after which orbit segments stop tracing ${S}_{1}^{r}$ altogether, and the transition from a one to twospike transient burst ends. We refer to Govaerts and Dhooge [24], Guckenheimer and Kuehn [25], and Osinga et al. [65] for similar transitions of periodic bursting oscillators.
We remark here that the manner of eventual convergence to the resting potential depends on the nature of the liftoff from the slow manifolds that correspond to the sheet ${S}_{1}^{r}$ or ${S}_{2}^{r}$. Recall that both sheets ${S}_{1}^{r}$ and ${S}_{2}^{r}$ of the critical manifold S have threedimensional stable manifolds; see Table 2. This means that the associated slow manifolds have a onedimensional repelling fast component [41], and orbit segments that trace these saddleunstable slow manifolds can leave it only along a single fast direction. We can see this in Figure 5 as a liftoff from ${S}_{1}^{r}$ ‘down’ in V, shown in Figure 5a, or a liftoff from ${S}_{1}^{r}$ ‘up’ in V, shown in Figure 5f; this uniquely defined change in direction along the onedimensional repelling fast component is real and not just due to the projection onto $({h}_{\mathrm{SI}},{m}_{\mathrm{SO}},V)$space. The same holds for the sheet ${S}_{2}^{r}$, for which Figure 5b,e is a good example that also shows the required liftoff up from ${S}_{1}^{r}$ in order to reach ${S}_{2}^{r}$. In what follows, the notions up and down are with respect to this uniquelydefined change in direction.
The behaviour of the orbit segment of system (1) in relation to the critical manifold S of the fast subsystem that corresponds to the first downward peak in Figure 3a is representative for what happens along the other downward peaks in Figure 3a. Each time ${g}_{\mathrm{SI}}$ reaches a special value such that the orbit segment comes close enough to the fourdimensional stable manifold of ${S}_{1}^{r}$, it gets trapped onto ${S}_{1}^{r}$ (or, more precisely, the corresponding saddleunstable slow manifold) for increasingly longer times, and the next spikeadding transition begins. For the parameters of Table 1, we found that this process always includes a transition between two saddleunstable sheets, which organises the doublestep ADP solutions. As mentioned before, the two sheets ${S}_{1}^{r}$ and ${S}_{2}^{r}$ together are perturbations of a single sheet connected to ${F}_{3}$ that can be obtained continuously via a small parameter variation (using a suitable parameter from Table 1), such that the fold curves ${F}_{1}$ and ${F}_{2}$ disappear in a curve of cusp bifurcations. Therefore, in the next section, we first explain the jump at the end of the canardlike behaviour, that is, the behaviour near the fold ${F}_{3}$ that separates the two saddleunstable sheets ${S}_{2}^{r}$ and ${S}_{3}^{r}$. We then discuss the transition between ${S}_{1}^{r}$ and ${S}_{2}^{r}$ in Section 4.2. Section 4.3 illustrates how the spikeadding mechanism can change when additional equilibria are present.
4.1 Slow flow on the critical manifold near ${F}_{3}$
that is, the slow flow on the twodimensional critical manifold S is defined by two ordinary differential equations for ${m}_{\mathrm{SO}}$ and ${h}_{\mathrm{SI}}$ and a single algebraic constraint ${f}_{1}^{\ast}(V,{m}_{\mathrm{SO}},{h}_{\mathrm{SI}},\lambda )=0$. Unfortunately, S is folded with respect to V so that ${m}_{\mathrm{SO}}$ and ${h}_{\mathrm{SI}}$ do not uniquely define V; however, the algebraic constraint does uniquely define ${m}_{\mathrm{SO}}$ or ${h}_{\mathrm{SI}}$ from given pair $(V,{m}_{\mathrm{SO}})$ or $(V,{h}_{\mathrm{SI}})$, respectively; compare also Figure 4a,b. Hence, it is advantageous to express the slow flow in terms of only one of the slow variables, ${m}_{\mathrm{SO}}$ or ${h}_{\mathrm{SI}}$, together with the fast variable V.
Figure 7a,b shows a projection onto $({h}_{\mathrm{SI}},{m}_{\mathrm{SO}},V)$coordinates of how the orbit segment follows the slow flow on ${S}_{2}^{r}$ as it approaches ${F}_{3}$. The first part of the orbit segment lies slightly above (relative to this projection) ${S}_{1}^{r}$, and after a jump, another part of this orbit segment lies slightly below ${S}_{2}^{r}$. In a neighbourhood of the folded focus on the fold curve ${F}_{3}$, the slow flow has the form of large semicycles that cause the orbit segment to trace ${S}_{2}^{r}$ laterally and, at the same time, push it toward ${F}_{3}$. Since the flow on the top sheet ${S}_{3}^{r}$ also points towards ${F}_{3}$, as shown in Figure 7c,d, the orbit segment cannot pass ${F}_{3}$ and reaches a socalled jump point; compare also with Figure 6b. At the jump point, the fast directions of the flow take over, which causes the formation of a small spike as the orbit segment leaves S; see also Figure 5c. Let us emphasise here that the behaviour of the orbit segments near ${F}_{3}$ does not involve interactions with the slow flow on ${S}_{3}^{r}$; the small spike and subsequent drop down to ${S}_{1}^{a}$ do not intersect the surfaces ${S}_{3}^{r}$ and ${S}_{2}^{r}$. As mentioned earlier in this section, the actual spike formation develops as soon as an orbit segment has reached ${F}_{3}$. After reaching ${F}_{3}$, the orbit segment will lie slightly above (relative to this projection) ${S}_{2}^{r}$ and experience a liftoff up from ${S}_{2}^{r}$ (or later only from ${S}_{1}^{r}$ as shown in Figure 5f). The spikeformation takes place on the fast time scale, and any perceived intersections with ${S}_{3}^{r}$ and ${S}_{2}^{r}$ are due to the projection onto $({h}_{\mathrm{SI}},{m}_{\mathrm{SO}},V)$coordinates.
4.2 Slow flow of the critical manifold near the folds ${F}_{1}$ and ${F}_{2}$
The formation of a new welldeveloped spike occurs over an exponentially small parameter interval ${g}_{\mathrm{SI}}\approx 0.5615{\text{mS/cm}}^{2}$ for which the effect of the injected current is precisely such that the orbit segment comes close to the fourdimensional stable manifold of ${S}_{1}^{r}$. The behaviour of the orbit segment near the top fold ${F}_{3}$ corresponds to the onset of such a new spike, but the process of reaching ${F}_{3}$, as illustrated in Figure 5a,b, as well as the further development of the spike, as illustrated in Figure 5d,e,f, involves the creation of a doublestep ADP; this behaviour is organised by a (fast) jump from ${S}_{1}^{r}$ to another saddleunstable sheet ${S}_{2}^{r}$. Such a jump, which is actually a jump between the two corresponding saddleunstable slow manifolds, is a phenomenon that occurs robustly as part of the spikeadding mechanism and has previously been observed for periodic orbits in planar systems; it was reported as a new type of canard called foldinitiated canards in a study by Guckenheimer et al. [45], and a slightly different version termed ducks with relaxation is discussed by Arnol’d [66], Ch.4, Sec.5.4]. In fact, the behaviour we observe in our model is essentially planar and very similar to the example discussed by Guckenheimer et al. [45]. Indeed, there are no folded singularities on the folds ${F}_{1}$ and ${F}_{2}$, which means that each fold point has the same effect on the dynamics, and the slow flow is essentially one dimensional. Furthermore, the repelling fast component of the slow manifolds associated with ${S}_{1}^{r}$ and ${S}_{2}^{r}$ is one dimensional as well.
Let us consider this continuous oneparameter family of orbit segments as identified by the moment of liftoff (first down and then up) from ${S}_{1}^{r}$. At the start of the spikeadding process, orbit segments trace only the saddleunstable sheet ${S}_{1}^{r}$ before a liftoff down to ${S}_{1}^{a}$ returns the system to its resting potential; the example in Figure 8a shows an orbit segment that almost reaches ${F}_{1}$. As ${g}_{\mathrm{SI}}\approx 0.5615{\text{mS/cm}}^{2}$ increases continuously (but only exponentially small), the orbit segments come increasingly closer to ${S}_{1}^{r}$ until one actually reaches ${F}_{1}$; these orbit segments grow increasingly longer stretched ADPs.
Using the analysis via the desingularised slow flow (11) as derived in Section 4.1, we can decide what happens when an orbit segment reaches ${F}_{1}$. We find that the desingularised slow flow (11) does not have any equilibria in the neighbourhood of the two folds ${F}_{1}$ and ${F}_{2}$, which means that there are no folded singularities on either ${F}_{1}$ or ${F}_{2}$; the fold curve ${F}_{1}$ is uniformly attracting, which we indicated by a darkgreen colour, and ${F}_{2}$ is uniformly repelling, indicated by a lightgreen colour. Hence, upon reaching ${F}_{1}$, the orbit segment simply jumps down toward the resting potential, and subsequent orbit segments exhibit a liftoff up from ${S}_{1}^{r}$. Since the sheet ${S}_{2}^{a}$ on the other side of ${F}_{1}$ is attracting, the fast directions will push these orbit segments toward ${S}_{2}^{a}$, provided that the liftoff up from ${S}_{1}^{r}$ occurs not too far away from ${F}_{1}$. Note that the slow flow on ${S}_{2}^{a}$ points back to ${F}_{1}$, so orbit segments that reach ${S}_{2}^{a}$ will flow to ${F}_{1}$ and again drop down to ${S}_{1}^{a}$; an example is shown in Figure 8b.
As we continue to increase ${g}_{\mathrm{SI}}\approx 0.5615{\text{mS/cm}}^{2}$ ever so slightly, orbit segments will converge to ${S}_{2}^{a}$ closer and closer near ${F}_{2}$, until the liftoff up from ${S}_{1}^{r}$ happens earlier, far enough from ${F}_{1}$, such that they may reach ${F}_{2}$. In order to determine what happens when an orbit segment reaches ${F}_{2}$, we must remember that system (1) and, hence, the slow flow on S depend continuously on ${g}_{\mathrm{SI}}$. This means that the family of orbit segments is also continuous, and the orbit segment that reaches ${F}_{2}$ is a continuous variation of an orbit segment that experiences a liftoff up from ${S}_{1}^{r}$ later so that it still converges to ${S}_{2}^{a}$ and flows back to ${F}_{1}$ as well as an orbit segment that experiences a liftoff up from ${S}_{1}^{r}$ earlier so that it misses ${F}_{2}$ altogether and forms a welldeveloped spike as illustrated in Figure 5f. Therefore, the family of orbit segments must include a subfamily of orbit segments that experience a liftoff from ${S}_{2}^{r}$. Just as for ${S}_{1}^{r}$, the liftoff will first be down as a continuous variation from the orbit segments that flow along ${S}_{2}^{a}$ and then up to transform into an orbit segment that misses ${F}_{2}$ altogether. This subfamily exists in a (doubly) exponentially small parameter regime of foldinitiated canard behaviour [45] and consists of orbit segments that exhibit a doublestep ADP.
Another way to understand this phenomenon is in terms of singularity theory, which suggests that ${S}_{1}^{r}$ and ${S}_{2}^{r}$ are part of the same surface that unfolds a cusp singularity. Recall that the critical manifold S is a collection of families of equilibria of the fast subsystem of (1); the folds ${F}_{1}$ and ${F}_{2}$ are curves of saddlenode bifurcations that exist robustly in the $({m}_{\mathrm{SO}},{h}_{\mathrm{SI}})$parameter plane organised by the two slow variables. Since ${F}_{1}$ and ${F}_{2}$ are typically not parallel, they will meet; note that ${F}_{1}$ and ${F}_{2}$ may need to be extended into an unphysical parameter regime of the $({m}_{\mathrm{SO}},{h}_{\mathrm{SI}})$plane. Singularity theory tells us that the two fold curves ${F}_{1}$ and ${F}_{2}$ typically meet at a cusp point and end there, which means that the two sheets ${S}_{1}^{r}$ and ${S}_{2}^{r}$ merge into one in this region of the $({m}_{\mathrm{SO}},{h}_{\mathrm{SI}})$plane. The existence of a doublestep ADP then merely depends on the location relative to the cusp point of the interaction between the orbit segments and the critical manifold S. We remark that a small change in one or more of the parameters given in Table 1 may move the cusp point into the physical regime or such that ${F}_{1}$ and ${F}_{2}$ no longer exist for physiologically realistic values of ${m}_{\mathrm{SO}}$ and ${h}_{\mathrm{SI}}$; in the latter case, the spikeadding transition will not feature a doublestep ADP.
4.3 Spikeadding when additional equilibria are present
It turns out that the spikeadding mechanism organised by canardlike behaviour during the downward peaks of Figure 3a always features a doublestep ADP stage involving a jump between ${S}_{1}^{r}$ and ${S}_{2}^{r}$. Hence, each downward peak in Figure 3a corresponds to a qualitatively similar transition as discussed for the first one at ${g}_{\mathrm{SI}}\approx 0.5615{\text{mS/cm}}^{2}$. If we increase ${g}_{\mathrm{FO}}$ from the fixed value ${g}_{\mathrm{FO}}=9.5{\text{mS/cm}}^{2}$ that was used in Figure 3 to the new value ${g}_{\mathrm{FO}}=9.6{\text{mS/cm}}^{2}$, then the nature of spike adding changes due to the presence of additional unstable equilibria of system (1). If we again continue the twopoint boundary value problem (5)(8) as before with $\lambda ={g}_{\mathrm{SI}}$ but ${g}_{\mathrm{FO}}=9.6{\text{mS/cm}}^{2}$ set to its new value, we get a bifurcation diagram similar to the one for ${g}_{\mathrm{FO}}=9.5{\text{mS/cm}}^{2}$ shown in Figure 3. In fact, the spikeadding mechanism for the first four additional spikes involves a doublestep ADP stage as we have seen in the previous section. However, for ${g}_{\mathrm{SI}}\approx 0.7672{\text{mS/cm}}^{2}$, that is, just before the transition from five to six spikes, a saddlenode bifurcation occurs on the saddleunstable sheet ${S}_{1}^{r}$. This creation of two new (unstable) equilibria prevents a doublestep ADP stage; the spikeadding mechanism only involves orbit segments exhibiting a stretched ADP with a singe step, and there is no longer a jump between saddleunstable slow manifolds.
The value ${g}_{\mathrm{SI}}\approx 0.7842{\text{mS/cm}}^{2}$ for this case with ${g}_{\mathrm{FO}}=9.6{\text{mS/cm}}^{2}$ is again special because at the end of the oscillations, when the orbit segment reaches the family of homoclinic orbits where ${P}^{a}$ ends, it lies extremely close to the fourdimensional stable manifold of ${S}_{1}^{r}$ so that it drops down and traces the saddleunstable sheet ${S}_{1}^{r}$ of S. The difference with the spikeadding mechanism illustrated in Figure 5 is that the behaviour of the orbit segment on ${S}_{1}^{r}$ is affected by the presence of the equilibria ${s}_{1}$ and ${s}_{2}$. With respect to the twodimensional slow flow on ${S}_{1}^{r}$, the equilibrium ${s}_{1}$ is an attractor, and all orbit segments on ${S}_{1}^{r}$ converge to ${s}_{1}$ provided that they lie in its basin of attraction, which is bounded by the onedimensional stable manifold of the saddle ${s}_{2}$. In terms of the full fivedimensional flow, ${S}_{1}^{r}$ is obviously unstable, and orbit segments that come close enough to ${S}_{1}^{r}$ will behave as dictated by the slow flow for only a finite amount of time; this means that convergence to ${s}_{1}$ will eventually be followed by a fast repulsion away from ${S}_{1}^{r}$. The orbit segment in Figure 9a,c enters a close enough neighbourhood of ${S}_{1}^{r}$ in the region of the basin of attraction of ${s}_{1}$; hence, during the time that it is following the slow flow, it converges to ${s}_{1}$, but we can clearly see in Figure 9c that the fast directions take over before it reaches ${s}_{1}$. Since this orbit segment was selected from the falling slope of the downward peak of the spikeadding mechanism, the orbit segment jumps straight down toward ${S}_{1}^{a}$, where it converges to the resting potential. Orbit segments on this slope that lie closer to the minimum of the downward peak would come closer to ${s}_{1}$ but still jump down toward ${S}_{1}^{a}$ when the fast directions take over. On the other hand, orbit segments from the rising slope of the downward peak eventually experience a liftoff ‘up’ from ${S}_{1}^{r}$ so that a large action potential occurs before converging back to the resting potential; this change of direction corresponds to the onset of a new spike, which is more dramatic and abrupt than the gradual increase in V followed by a smallamplitude spike as illustrated in Figure 5.
Continuity of the vector field (1) implies that there exists an orbit segment that actually converges to the saddle ${s}_{1}$ and never relaxes back to the resting potential. This happens when ${g}_{\mathrm{SI}}\approx 0.7842{\text{mS/cm}}^{2}$ is exactly at the value where the perturbed trajectory lies on the fourdimensional stable manifold ${W}^{s}({s}_{1})$ of the equilibrium ${s}_{1}$. In contrast to the fourdimensional stable manifold of ${S}_{1}^{r}$, the manifold ${W}^{s}({s}_{1})$ is invariant under the flow of (1), and this heteroclinic connection is a welldefined bifurcation for system (1).
We remark here that the presence of additional equilibria, such as ${s}_{1}$ and ${s}_{2}$ in the example discussed, only affects the spikeadding mechanism if the orbit segments that trace ${S}_{1}^{r}$ enter the basin of attraction of ${s}_{1}$. If such orbit segments trace ${S}_{1}^{r}$ on the other side of the stable manifold of ${s}_{2}$, then a doublestep ADP stage would occur. We know from our further model analysis (not shown) that the unstable equilibria persists for higher values of ${g}_{\mathrm{SI}}$ as well as ${g}_{\mathrm{FO}}$, and in all cases that we investigated, these additional equilibria on ${S}_{1}^{r}$ affect the spike generation in the way described above.
5 Discussion
In this paper, we performed a detailed analysis of the mechanisms of spike generation and spikeadding in a transient burst. Based on a reduction of our previous model [13], we identify these mechanisms using numerical continuation of orbit segments that are solutions to a wellposed boundary value problem. In our analysis, we utilised the separation of time scales in system (1). We calculated the twodimensional critical manifold S of the fast subsystem, which organises the behaviour of the system. The spikegeneration process is characterised by the fact that orbit segments trace saddleunstable slow manifolds that correspond to saddleunstable sheets of S. More precisely, there are two saddleunstable sheets, ${S}_{1}^{r}$ and ${S}_{2}^{r}$, with fourdimensional stable and threedimensional unstable manifolds; this means that the liftoff from the associated slow manifolds is characterised by a uniquely defined direction. The changes in sign of this direction mark the different phases of the spikeadding transition.
By considering the slow flow on S, we were able to explain the onset of a spike as well as the doublestep stretched ADP that leads up to it. For the value of ${g}_{\mathrm{FO}}=9.5{\text{mS/cm}}^{2}$ that we considered, the onset of a spike is organised by the top fold ${F}_{3}$ of S. This fold contains a foldedfocus singularity, but it is not accessible and the spikes are due to (regular) jump points. The folds ${F}_{1}$ and ${F}_{2}$ that are involved in the doublestep ADP do not contain any folded singularities, and they are uniformly attracting and repelling, respectively. Therefore, the first step in the stretched ADP ends at a regular jump point. The second step occurs due to a type of foldinitiated canard because the slow flow points away from ${F}_{2}$; the foldinitiated canardlike behaviour forms a robust part of the spikeadding mechanism that has also been observed for periodic orbits [45]. The actual spikeadding transition occurs in an exponentially small parameter regime that is very difficult, if not impossible, to find with bruteforce integration routines; therefore, it is also highly unlikely to observe anything like a stretched ADP or doublestep ADP in experiments.
We found that the nature of the spikeadding mechanism may change if ${g}_{\mathrm{FO}}$ increases slightly. For higher values of ${g}_{\mathrm{FO}}$, an increase in ${g}_{\mathrm{SI}}$ causes the appearance of two equilibria, ${s}_{1}$ and ${s}_{2}$, on ${S}_{1}^{r}$ that form a saddlenode pair with respect to the slow flow. As it turns out, the presence of these equilibria prevents the doublestep ADP. Instead, orbit segments that come close to ${S}_{1}^{r}$ during the spikegeneration process flow towards the attracting equilibrium ${s}_{1}$ before a liftoff in the fast direction. This means that the onset of a spike is now organised by ${s}_{1}$ rather than the fold ${F}_{3}$, and the stretched ADP involves only a single step. A spike generated by ${s}_{1}$ is dramatically different from one generated by ${F}_{3}$. While both spike generations happen in an exponentially small parameter interval, the increase in amplitude of a spike generated by ${F}_{3}$ is gradual and should be viewed as a variation of the orbit segment that depends continuously on ${g}_{\mathrm{SI}}$. On the other hand, a spike generated by ${s}_{1}$ is not a continuous variation, and a largeamplitude spike appears abruptly as ${g}_{\mathrm{SI}}$ is increased (on an exponentially small scale). The two families, one with and one without the additional spike, are separated by a heteroclinic connection (via a current injection) from the resting potential to the saddle ${s}_{1}$; our numerical method for continuation of the family only gets past this discontinuity because we do not impose relaxation back to the resting potential but keep ${T}_{\mathrm{OFF}}$ fixed instead. Similar spikeadding mechanisms can be organised by the presence of a saddle periodic orbit or other saddletype invariant object, but we did not observe this in our model.
In theory, it should be possible to have a doublestep ADP as part of the spikeadding transition even when additional equilibria are present. The occurrence of a doublestep ADP in this case only depends on whether tracing of ${S}_{1}^{r}$ commences in the basin of attraction of ${s}_{1}$ (restricted to ${S}_{1}^{r}$) or not, which is determined by the stable manifold of the saddle equilibrium ${s}_{2}$. In our numerical explorations, the orbit segments always commence tracing ${S}_{1}^{r}$ in the basin of attraction of ${s}_{1}$. Hence, we may conclude that ${g}_{\mathrm{FO}}$ and ${g}_{\mathrm{SI}}$ do not have a profound influence on the relative location where orbit segments begin to trace ${S}_{1}^{r}$ during the spikeadding process. However, other parameters of the system may alter this relative location and provoke a doublestep ADP even in the presence of additional equilibria. While this observation indicates a challenge for a precise definition of spikeonset in our context, the two different mathematical notions seem to have the same biological effect.
We believe that the canardlike transition involving saddleunstable sheets of the critical manifold lies at the heart of any spikeadding mechanism when no invariant saddletype objects are present. The different phases during the transition, however, could be organised by features other than regular jump points and foldinitiated canards. For example, it should be expected that other folded singularities may appear due to variations in the slow flow on S. An investigation of all possibilities remains an interesting and challenging project for future work.
Declarations
Acknowledgements
The authors thank Pablo Aguirre, Mathieu Desroches, Bernd Krauskopf, John Guckenheimer and Peter De Maesschalck for their helpful discussions. JN was supported by the grant EP/E032249/1 from the Engineering and Physical Sciences Research Council (EPSRC). HMO was supported by an EPSRC Advanced Research Fellowship grant, and KTA was supported by the EPSRC grant EP/I018638/1.
Authors’ Affiliations
References
 Izhikevich EM: Dynamical Systems in Neuroscience: The Geometry of Excitability and Bursting. MIT Press, Cambridge; 2006.Google Scholar
 Keener JP, Sneyd J: Mathematical Physiology: Cellular Physiology. Springer, New York; 2008.Google Scholar
 Ermentrout GB, Terman DH: Mathematical Foundations of Neuroscience. Springer, New York; 2010.View ArticleGoogle Scholar
 Izhikevich EM: Neural excitability, spiking and bursting. Int J Bifurc Chaos Appl Sci Eng 2000,10(6):1171–1266. 10.1142/S0218127400000840MathSciNetView ArticleGoogle Scholar
 Wieczorek S, Krauskopf B, Lenstra D: Multipulse excitability in a semiconductor laser with optical injection. Phys Rev Lett 2002,88(6):1–4.View ArticleGoogle Scholar
 Krauskopf B, Schneider K, Sieber J, Wieczorek S, Wolfrum M: Excitability and selfpulsations near homoclinic bifurcations in semiconductor laser systems. Opt Commun 2003,215(4–6):367–379. 10.1016/S00304018(02)022393View ArticleGoogle Scholar
 Nagumo J, Arimoto S: An active pulse transmission line simulating nerve axon. Proc IRE 1962,50(10):2061–2070.View ArticleGoogle Scholar
 Douglas R, Mahowald M, Mead C: Neuromorphic analogue VLSI. Annu Rev Neurosci 1995, 18: 255–281. 10.1146/annurev.ne.18.030195.001351View ArticleGoogle Scholar
 Guckenheimer JM, Hoffman K, Weckesser W: The forced van der Pol equation I: the slow flow and its bifurcations. SIAM J Appl Dyn Syst 2003,2(1):1–35. 10.1137/S1111111102404738MathSciNetView ArticleGoogle Scholar
 Indiveri G, LinaresBarranco B, Hamilton TJ, Van Schaik A, EtienneCummings R, Delbruck T, Liu SC, Dudek P, Häfliger P, Renaud S, Schemmel J, Cauwenberghs G, Arthur J, Hynna K, Folowosele F, Saighi S, SerranoGotarredona T, Wijekoon J, Wang Y, Boahen K: Neuromorphic silicon neuron circuits. Front Neurosci 2011, 5: 73.Google Scholar
 Brøns M, BarEli K: Canard explosion and excitation in a model of the BelousovZhabotinskii reaction. J Phys Chem 1991,95(22):8706–8713. 10.1021/j100175a053View ArticleGoogle Scholar
 Brown JT, Randall AD: Activitydependent depression of the spike afterdepolarization generates longlasting intrinsic plasticity in hippocampal CA3 pyramidal neurons. J Physiol 2009,587(6):1265–1281. 10.1113/jphysiol.2008.167007View ArticleGoogle Scholar
 Nowacki J, Osinga HM, Brown JT, Randall AD, TsanevaAtanasova K: A unified model of CA1/3 pyramidal cells: an investigation into excitability. Prog Biophys Mol Biol 2011,105(1–2):34–48. 10.1016/j.pbiomolbio.2010.09.020View ArticleGoogle Scholar
 Brown JT, Chin J, Leiser SC, Pangalos MN, Randall AD: Altered intrinsic neuronal excitability and reduced Na(+) currents in a mouse model of Alzheimer’s disease. Neurobiol Aging 2011, 32: 2109.e1–2109.e14. 10.1016/j.neurobiolaging.2011.05.025View ArticleGoogle Scholar
 Van Elburg RAJ, Van Ooyen A: Impact of dendritic size and dendritic topology on burst firing in pyramidal cells. PLoS Comput Biol 2010.,6(5):
 Rinzel J: A formal classification of bursting mechanisms in excitable systems. In Mathematical Topics in Population Biology, Morphogenesis, and Neurosciences. Edited by: Teramoto E, Yamaguti M. Springer, Berlin; 1987:267–281.View ArticleGoogle Scholar
 Golubitsky M, Josić K, Kaper T: An unfolding theory approach to bursting in fastslow systems. In Global Analysis of Dynamical Systems: Festschrift Dedicated to Floris Takens for His 60th Birthday. Edited by: Broer H, Krauskopf B, Vegter G. Institute of Physics Publishing, Bristol; 2001:277–308.Google Scholar
 Osinga HM, Sherman A, TsanevaAtanasova K: Crosscurrents between biology and mathematics: The codimension of pseudoplateau bursting. Discrete Contin Dyn Syst, Ser A 2012,32(8):2853–2877.MathSciNetView ArticleGoogle Scholar
 Smolen P, Terman DH, Rinzel J: Properties of a bursting model with two slow inhibitory variables. SIAM J Appl Math 1993,53(3):861–892. 10.1137/0153042MathSciNetView ArticleGoogle Scholar
 Terman DH: The transition from bursting to continuous spiking in excitable membrane models. J Nonlinear Sci 1992,2(2):135–182. 10.1007/BF02429854MathSciNetView ArticleGoogle Scholar
 Terman DH: Chaotic spikes arising from a model of bursting in excitable membranes. SIAM J Appl Math 1991,51(5):1418–1450. 10.1137/0151071MathSciNetView ArticleGoogle Scholar
 Benes GN, Barry AM, Kaper TJ, Kramer Ma, Burke J: An elementary model of torus canards. Chaos 2011.,21(2):
 Kramer M, Traub R, Kopell N: New dynamics in cerebellar purkinje cells: torus canards. Phys Rev Lett 2008,101(6):68103.View ArticleGoogle Scholar
 Govaerts W, Dhooge A: Bifurcation, bursting and spike generation in a neural model. Int J Bifurc Chaos Appl Sci Eng 2002,12(8):1731–1741. 10.1142/S021812740200542XMathSciNetView ArticleGoogle Scholar
 Guckenheimer J, Kuehn C: Computing slow manifolds of saddle type. SIAM J Appl Dyn Syst 2009,8(3):854–879. 10.1137/080741999MathSciNetView ArticleGoogle Scholar
 TsanevaAtanasova K, Osinga HM, Rieß T, Sherman A: Full system bifurcation analysis of endocrine bursting models. J Theor Biol 2010,264(4):1133–1146. 10.1016/j.jtbi.2010.03.030View ArticleGoogle Scholar
 Linaro D, Champneys A, Desroches M, Storace M: Codimensiontwo homoclinic bifurcations underlying spike adding in the HindmarshRose burster. SIAM J Appl Dyn Syst, in press. arXiv; 2011. [http://arxiv.org/abs/1109.5689] Linaro D, Champneys A, Desroches M, Storace M: Codimensiontwo homoclinic bifurcations underlying spike adding in the HindmarshRose burster. SIAM J Appl Dyn Syst, in press. arXiv; 2011. [http://arxiv.org/abs/1109.5689]
 Teka W, Tabak J, Vo T, Wechselberger M, Bertram R: The dynamics underlying pseudoplateau bursting in a pituitary cell model. J Math Neurosci 2011., 1:Google Scholar
 Ghigliazza RM, Holmes PJ: Minimal models of bursting neurons: how multiple currents, conductances, and timescales affect bifurcation diagrams. SIAM J Appl Dyn Syst 2004,3(4):636–670. 10.1137/030602307MathSciNetView ArticleGoogle Scholar
 Tran D, Sato D, Yochelis A, Weiss J, Garfinkel A, Qu Z: Bifurcation and chaos in a model of cardiac early afterdepolarizations. Phys Rev Lett 2009,102(25):1–4.View ArticleGoogle Scholar
 Kim MY, Aguilar M, Hodge A, Vigmond E, Shrier A, Glass L: Stochastic and spatial influences on druginduced bifurcations in cardiac tissue culture. Phys Rev Lett 2009,103(5):1–4.View ArticleGoogle Scholar
 Stern JV, Osinga HM, LeBeau A, Sherman A: Resetting behavior in a model of bursting in secretory pituitary cells: distinguishing plateaus from pseudoplateaus. Bull Math Biol 2008,70(1):68–88. 10.1007/s115380079241xMathSciNetView ArticleGoogle Scholar
 Jones CKRT: Geometric singular perturbation theory. In Dynamical Systems. Springer, Heidelberg; 1995:44–118.View ArticleGoogle Scholar
 Dumortier F: Techniques in the theory of local bifurcations: blowup, normal forms, nilpotent bifurcations, singular perturbations. In Bifurcations and Periodic Orbits of Vector Fields (Montreal, PQ, 1992). Kluwer Academic, Dordrecht; 1993:19–74.View ArticleGoogle Scholar
 Szmolyan P, Wechselberger M:Canards in ${\mathbb{R}}^{3}$. J Differ Equ 2001,177(2):419–453. 10.1006/jdeq.2001.4001MathSciNetView ArticleGoogle Scholar
 Wechselberger M:Existence and bifurcation of canards in ${\mathbb{R}}^{3}$ in the case of a folded node. SIAM J Appl Dyn Syst 2005,4(1):101–139. 10.1137/030601995MathSciNetView ArticleGoogle Scholar
 Champneys AR, Kuznetsov YA, Sandstede B: A numerical toolbox for homoclinic bifurcation analysis. Int J Bifurc Chaos Appl Sci Eng 1996,6(5):867–887. 10.1142/S0218127496000485MathSciNetView ArticleGoogle Scholar
 Krauskopf B, Osinga HM: Computing invariant manifolds via the continuation of orbit segments. In Numerical Continuation Methods for Dynamical Systems: Path Following and Boundary Value Problems. Edited by: Krauskopf B, Osinga HM, GalánVioque J. Springer, Dordrecht; 2007:117–154.View ArticleGoogle Scholar
 Desroches M, Krauskopf B, Osinga HM: Mixedmode oscillations and slow manifolds in the selfcoupled FitzHugh Nagumo system. Chaos 2008.,18(1):
 Desroches M, Krauskopf B, Osinga HM: The geometry of slow manifolds near a folded node. SIAM J Appl Dyn Syst 2008,7(4):1131–1162. 10.1137/070708810MathSciNetView ArticleGoogle Scholar
 Fenichel N: Geometric singular perturbation theory for ordinary differential equations. J Differ Equ 1979, 31: 53–98. 10.1016/00220396(79)901529MathSciNetView ArticleGoogle Scholar
 Benoît É, Callot JL, Diener F, Diener M: Chasse au canard. Collect Math 1981,31–32(1–3):37–119.Google Scholar
 Dumortier F, Roussarie R: Canard cycles and center manifolds. Mem Am Math Soc 1996,121(577):1–101. [With an appendix by Cheng Zhi Li.] [With an appendix by Cheng Zhi Li.]MathSciNetGoogle Scholar
 Lee E, Terman D: Uniqueness and stability of periodic bursting solutions. J Differ Equ 1999, 158: 48–78. 10.1016/S00220396(99)800187MathSciNetView ArticleGoogle Scholar
 Guckenheimer JM, Hoffman K, Weckesser W: Numerical computation of canards. Int J Bifurc Chaos Appl Sci Eng 2000,10(12):2669–2687. 10.1142/S0218127400001742MathSciNetView ArticleGoogle Scholar
 Hodgkin AL, Huxley AF: A quantitive description of membrane current and its application to conduction and excitation in nerve. J Physiol 1952,105(117):500–544.View ArticleGoogle Scholar
 Golomb D, Yue C, Yaari Y:Contribution of persistent ${\mathrm{Na}}^{+}$ current and Mtype ${\mathrm{K}}^{+}$ current to somatic bursting in CA1 pyramidal cells: combined experimental and modeling study. J Neurophysiol 2006,96(4):1912–1926. 10.1152/jn.00205.2006View ArticleGoogle Scholar
 Yue C, Remy S, Su H, Beck H, Yaari Y:Proximal persistent ${\mathrm{Na}}^{+}$ channels drive spike afterdepolarizations and associated bursting in adult CA1 pyramidal cells. J Neurosci 2005,25(42):9704. 10.1523/JNEUROSCI.162105.2005View ArticleGoogle Scholar
 Jaffe DB, Ross WN, Lisman JE, LasserRoss N, Miyakawa H, Johnston D:A model for dendritic ${\mathrm{Ca}}^{2+}$ accumulation in hippocampal pyramidal neurons based on fluorescence imaging measurements. J Neurophysiol 1994,71(3):1065–1077.Google Scholar
 Yaari Y, Yue C, Su H:Recruitment of apical dendritic Ttype ${\mathrm{Ca}}^{2+}$ channels by backpropagating spikes underlies de novo intrinsic bursting in hippocampal epileptogenesis. J Physiol 2007,580(2):435–450.View ArticleGoogle Scholar
 Blackmer T, Kuo SP, Bender KJ, Apostolides PF, Trussell LO: Dendritic calcium channels and their activation by synaptic signals in auditory coincidence detector neurons. J Neurophysiol 2009,102(2):1218–1226. 10.1152/jn.90513.2008View ArticleGoogle Scholar
 Yue C, Yaari Y: KCNQ/M channels control spike afterdepolarization and burst generation in hippocampal neurons. J Neurosci 2004,24(19):4614–4624. 10.1523/JNEUROSCI.076504.2004View ArticleGoogle Scholar
 Yue C, Yaari Y: Axosomatic and apical dendritic Kv7/M channels differentially regulate the intrinsic excitability of adult rat CA1 pyramidal cells. J Neurophysiol 2006,95(6):3480–3495. 10.1152/jn.01333.2005View ArticleGoogle Scholar
 Baer SM, Erneux T, Rinzel J: The slow passage through a Hopf bifurcation: delay, memory effects, and resonance. SIAM J Appl Math 1989, 49: 55–71. 10.1137/0149003MathSciNetView ArticleGoogle Scholar
 Doedel EJ: AUTO: a program for the automatic bifurcation analysis of autonomous systems. Congr Numer 1981, 30: 265–284.MathSciNetGoogle Scholar
 Doedel EJ, Oldeman BE: AUTO07P: Continuation and Bifurcation Software for Ordinary Differential Equations; 2007. [cmvl.cs.concordia.ca/auto/] Doedel EJ, Oldeman BE: AUTO07P: Continuation and Bifurcation Software for Ordinary Differential Equations; 2007. [cmvl.cs.concordia.ca/auto/]
 Ermentrout GB: Simulating, Analyzing, and Animating Dynamical Systems: A Guide to XPPAUT for Researchers and Students. SIAM, Philadelphia; 2002.View ArticleGoogle Scholar
 Nowacki J: XPPy; 2011. [http://seis.bris.ac.uk/~enxjn/xppy] Nowacki J: XPPy; 2011. [http://seis.bris.ac.uk/~enxjn/xppy]
 Oliphant T: Python for scientific computing. Comput Sci Eng 2007,9(3):10–20.View ArticleGoogle Scholar
 Hunter J: Matplotlib: a 2D graphics environment. Comput Sci Eng 2007,9(3):90–95.View ArticleGoogle Scholar
 Varoquaux G, Ramachandran P: Mayavi: making 3D data visualization reusable. In Proceedings of the 7th Python in Science Conference. SciPy, Pasadena; 2008:51–56.Google Scholar
 Hek G: Geometric singular perturbation theory in biological practice. J Math Biol 2010,60(3):347–386. 10.1007/s0028500902667MathSciNetView ArticleGoogle Scholar
 Desroches M, Guckenheimer JM, Krauskopf B, Kuehn C, Osinga HM, Wechselberger M: Mixedmode oscillations with multiple time scales. SIAM Rev 2012,54(2):211–288. 10.1137/100791233MathSciNetView ArticleGoogle Scholar
 Rinzel J, Ermentrout GB: Analysis of neural excitability and oscillations. In Methods in Neuronal Modelling. Edited by: Koch C, Sagev I. MIT Press, Cambridge; 1998:251–292.Google Scholar
 Osinga HM, TsanevaAtanasova KT: Dynamics of plateau bursting depending on the location of its equilibrium. J Neuroendocrinol 2010,22(12):1301–1314. 10.1111/j.13652826.2010.02083.xView ArticleGoogle Scholar
 Arnol’d VI: Dynamical systems V. In Encyclopaedia of Mathematical Sciences. Springer, Berlin; 1994.Google Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.