Neural Excitability and Singular Bifurcations
 Peter De Maesschalck^{1} and
 Martin Wechselberger^{2}Email author
https://doi.org/10.1186/s1340801500292
© De Maesschalck and Wechselberger 2015
Received: 29 May 2015
Accepted: 23 July 2015
Published: 6 August 2015
Abstract
We discuss the notion of excitability in 2D slow/fast neural models from a geometric singular perturbation theory point of view. We focus on the inherent singular nature of slow/fast neural models and define excitability via singular bifurcations. In particular, we show that type I excitability is associated with a novel singular Bogdanov–Takens/SNIC bifurcation while type II excitability is associated with a singular Andronov–Hopf bifurcation. In both cases, canards play an important role in the understanding of the unfolding of these singular bifurcation structures. We also explain the transition between the two excitability types and highlight all bifurcations involved, thus providing a complete analysis of excitability based on geometric singular perturbation theory.
Keywords
AMS Subject Classification
1 Excitable Systems
Most neurons are excitable, i.e. they are typically silent but can fire an action potential or produce a firing pattern in response to certain forms of stimulation. The fact that equivalent stimulation can elicit qualitative different spiking patterns in different neurons demonstrates that intrinsic coding properties differ significantly from one neuron to the next.

Type I: The stable equilibrium (resting state) disappears via a saddlenode on invariant circle (SNIC) bifurcation.

Type II: The stable equilibrium (resting state) loses stability via an Andronov–Hopf bifurcation (HB), either sub or supercritical.
Remark 1
In the canonical model (1), the nonlinear nature of \(G(v)\) is essential to guarantee relaxation type behaviour as observed in type I neurons which distinguishes this polynomial model from the classic FitzHugh–Nagumo model [7, 8] which can only produce type II (and type III) behaviour. In more biophysically inspired twodimensional systems, the Morris–Lecar model [9] presents a prime example which is able to produce all three excitability types; see also [4–6, 10].
1.1 Slow–Fast Excitable Systems
An important feature of most neural systems is that they evolve on multiple time scales; see, e.g., [11]. It is the interplay of the dynamics on different temporal scales that creates complicated rhythms. Multiple (or slow–fast) timescale problems are usually modelled by singularly perturbed systems such as (1) where the timescale separation of the ‘fast’ variable v (voltage) and the ‘slow’ variable w (recovery variable) is explicitly identified through the singular perturbation parameter \(\varepsilon \ll1\). The interest in such slow–fast systems goes towards the presence of socalled relaxation oscillations. Along an orbit of relaxation oscillation type, parts where the velocity of the phase variables is small (the slow parts) are alternated with high velocity peaks on short time intervals (the fast parts). During the slow parts, the phase state is \(O(\varepsilon )\)close to the critical set \(f(w,v,I)=0\) (because then \(\Vert (w',v')\Vert = O(\varepsilon )\)), whereas during the fast part the phase state is at an \(O(1)\)distance from this critical set. Tonic firing as observed in (1) is exactly of relaxation type.
While this GSPT approach to explain relaxation type behaviour in neural systems is well known to the mathematical and computational neuroscience community, see e.g. [5, 11], it is not often or consequently used to explain the underlying bifurcation structure in such singularly perturbed systems. This is the focus of Sects. 3–6 and we study singular bifurcations and their unfoldings in a normal form introduced in Sect. 3.1.
A closer look at the bifurcation diagram of type II excitability in Fig. 1 reveals that shortly after the Andronov–Hopf bifurcation the amplitude of the bifurcating limit cycles explodes dramatically under a very tiny (an exponentially small) parameter change. This is known as a canard explosion [14, 15] and indicates that the singular perturbation nature of the neural model is also reflected in the bifurcation structure. Note also a similar dramatic change in frequency near this singular Andronov–Hopf bifurcation. We will review this (wellknown) bifurcation phenomenon in Sect. 4.
Additional bifurcation structure is necessary to explain the transition from type II to type I excitability. This is covered in Sect. 5 where an incomplete canard explosion is identified. This lesserknown phenomenon refers to a premature termination of a canard explosion in a homoclinic bifurcation.
Similarly to the singular Andronov–Hopf bifurcation, one has to expect that the SNIC bifurcation associated with type I excitability shown in Fig. 1 must have a singular nature. We identify a singular Bogdanov–Takens/SNIC bifurcation point as the organising centre for type I excitability. Unfolding this type I singular bifurcation structure is the main focus of Sect. 6 which is based on the blowup method, a desingularisation technique for nilpotent singularities that has been successfully implemented in geometric singular perturbation problems with loss of normal hyperbolicity [14–16].
Finally, we summarise our results in Sect. 7 and discuss its implications for possible numerical observations in slow–fast neural models.
2 The Setup for Slow–Fast Excitable Systems
We start with introducing basic assumptions on the singularly perturbed system (3), respectively, (4).
Assumption 1
Assumption 2
This assumption implies that the wnullcline is a graph \(\{w=\psi _{\lambda}(v)\}\), and \(\psi_{\lambda}\) is a monotonically increasing function. While mathematically not necessary, this reflects the property of a typical neural model where \(g=0\) is given as a graph of a sigmoidal function over v.
Assumption 3

there can be no more than one equilibrium on the lower attracting branch \(S_{a}^{}\);

the existence of exactly two equilibria indicates a saddlenode bifurcation either on \(S_{r}\) or at \(F^{}\).
Remark 2
Assumption 3 sets the scene for the transition from an excitable to an oscillatory state. It excludes the possibility of a transition from an oscillatory state to a(nother) steady state (known as depolarisation block in neuroscience), by restricting the parameter space from above. This is not necessary but allows us to focus on the onset of oscillations, not the termination.
Assumption 4
The first condition in (12) is equivalent to imposing the requirement that \(\varphi _{I}'(v)=0\), \(\varphi _{I}''(v)<0\); in other words, the fold \(F^{+}\) is a regular local maximum of \(\varphi _{I}\).^{3} Together with the second condition, this determines the direction of the reduced flow near \(F^{+}\) and qualifies this fold point as a jump point: all orbits that come in along the upper branch \(S_{a}^{+}\) jump off the fold \(F^{+}\) and follow the fast fibre towards q; see Fig. 2 (right). Note that no equilibrium on \(S_{r}\) can approach the upper fold \(F^{+}\) while varying parameters (see also Remark 2).
These basic assumptions hold for many twodimensional neuronal models including the canonical system (1). This model has a cubicshaped critical manifold S (Assumption 1). The function \(G(v)\) is monotonically increasing (Assumption 2). It is the nonlinear nature of G imposed by Assumption 3 that allows us to explore the cases of different numbers of equilibria (restricted to \(S_{a}^{}\) and \(S_{r}\) only). If \(0< I_{1}< I_{F^{+}}\) and \(v_{\mathrm{th}}< v^{+}\), where \(I_{F^{+}}\) is defined implicitly by \(\varphi _{I_{F^{+}}}(v^{+})=G(v^{+})\), then Assumptions 3 and 4 are fulfilled and \(F^{+}\) is a regular jump point for all \(I\in[I_{0},I_{1}]\) and \(c \in[0,c_{1}]\).
3 Singular (or Slow–Fast) Bifurcations
Remark 3
System (1) models type III excitability if \(I_{\mathrm {bif}}\) lies outside of the interval \([I_{0},I_{1}]\), i.e. \(I_{\mathrm {bif}}>I_{1}\). In general, type III excitability can be characterised by the property that \(I_{\mathrm{bif}}\) lies outside of the interval \([I_{0},I_{1}]\).
3.1 Normal Form Near the Singular Fold Point
While in general one has to do a bit more work, the normal form for system (3) near the singular fold \(F^{}\) shows similarity with the normal form (14) of the canonical model. The local shape of the vector field (3) near the fold \(F^{}\) is described in the following proposition.
Proposition 1
Proof
Remark 4
These two local singular bifurcation points are associated with the two different neural excitability types I and II. Note that a Bogdanov–Takens bifurcation is a codimension2 bifurcation that includes a codimension1 Andronov–Hopf bifurcation in its unfolding. So, we also expect to find a connection between these two bifurcations as c tends to zero.
Remark 5
Although type III neurons are not associated with any bifurcation for fixed current input, these slope detectors play an important role in identifying dynamic changes and producing transient responses. We refer to [2] for details and [10] where type III neurons and excitability are discussed in the context of GSPT.
4 Type II Excitability: Singular Andronov–Hopf Bifurcation and Canard Explosion
Assumption 5
Remark 6
System (1) satisfies Assumption 5 for a range of λ values, along a parameter curve \(I=I_{\mathrm {bif}}^{\mathrm{II}}(\lambda)\).
Note in Fig. 1 that the \(O(\sqrt{\varepsilon })\) branch of the Andronov–Hopf bifurcation suddenly changes dramatically near \(I=I_{c}\). This almost vertical branch marks the unfolding of canard cycles within an exponentially small parameter interval of the bifurcation parameter I. This is often referred to as a canard explosion [14, 15]; it provides the necessary continuous connection between the small Andronov–Hopf limit cycles and the large relaxation cycles as shown in Fig. 1. In the singular limit, canard cycles can be identified as follows: Note that the stability switch of the equilibrium in the reduced problem (11) is due to the singular nature of system (11) at \(F^{}\); a stability switch of a single equilibrium without interacting with another equilibrium in a onedimensional regular perturbation problem is otherwise not possible. In fact, for \(I=I_{\mathrm{bif}}\) there exists no equilibrium in the reduced problem (11) due to a cancellation of a simple zero. Hence, a trajectory is able to cross from \(S_{a}^{}\) to \(S_{r}\) with nonzero speed which is a hallmark of a singular canard. One can construct singular canard cycles that are formed through concatenations of slow canard segments and fast fibres as shown in Fig. 4(d). Note that these singular canard cycles have \(O(1)\) amplitude and have a frequency \(O(1)\) on the order of the slow time scale. These singular canard cycles will unfold to the above mentioned canard cycles in a canard explosion. The following summarises these observations.
Theorem 1
Proof
We refer to [15], where the main part of the statement is shown. Here we just restrict to computing \(H_{1a}\) and \(K_{1a}\) in the normal form (15).
Let us start with \(H_{1a}\), given (21). A simple asymptotic analysis reveals that a singularity is located at \((x,y) = (\varepsilon ^{2}\frac{H_{1a}}{c^{2}} + O(\varepsilon ^{3}),\varepsilon \frac {H_{1a}}{c}+O(\varepsilon ^{2}) )\), about which the linearisation of the vector field has a trace given by \((\sigma+ \frac{2H_{1a}}{c})+O(\varepsilon ^{2})\). The Hopf bifurcation hence occurs at \(H_{1a}=\frac{c}{2}\sigma\).
Next we focus on the canard value \(H_{1a}+K_{1a}\). Since the singular Hopf point is of generic nature, the parameter value at which canards are present are the same parameter values for which there exists a smooth asymptotic expansion \(x = \varphi (y) + \varphi _{1}(y)\varepsilon + \varphi _{2}(y)\varepsilon ^{2} + O(\varepsilon ^{3})\) representing an invariant graph. Expressing the invariance by plugging the series in the differential equations yields expressions for \(\varphi _{1}\) and \(\varphi _{2}\), given \(a_{c} = (H_{1a}+K_{1a})\varepsilon + O(\varepsilon ^{2})\). Then imposing the requirement that \(\varphi _{2}\) should not have a pole at \(y=0\) yields a condition on \(H_{1a}+K_{1a}\) that leads to the required result. □
Remark 7
By actively using the singular nature of canards, the above calculation also presents an alternative way to find the first Lyapunov coefficient \(K_{1a}\) to determine the criticality of the singular Andronov–Hopf bifurcation.
In the singular limit, we have \(I_{h}=I_{c}=I_{\mathrm{bif}}\) indicating the singular nature of the bifurcation. Note that the classic definition of type II excitability refers to the slow \(O(\varepsilon )\) frequency band of the large relaxation oscillations which does not vary much (and not to the actual intermediate \(O(\varepsilon ^{1/2})\) singular Andronov–Hopf bifurcation frequency).
Lemma 1
For fixed \(0<\varepsilon \ll1\), the normal form (15) has for \(c\approx c_{\mathrm{bautin}}\) a codimension2 Bautin (generalised Andronov–Hopf) bifurcation point.
The Bautin bifurcation [17] indicates for fixed \(0< c< c_{\mathrm {bautin}}\) that we are dealing with a subcritical singular Andronov–Hopf bifurcation (under the variation of the parameter a) which is accompanied by another codimension1 bifurcation, a saddlenode of periodic orbits (SNPO) bifurcation that has branched of the codimension2 Bautin point. Due to the singular nature of our problem, this SNPO bifurcation is a bifurcation of canard cycles. Thus it happens exponentially close to the canard parameter value \(a_{c}\) defined in Theorem 1.
5 Type I/II Excitability Transition Regime: Incomplete Canard Explosion
Recall from (17) that for \(\varepsilon =0\) we observe a codimension2 cusp bifurcation at \((a,c)=(a_{\mathrm{cusp}},c_{\mathrm {cusp}})\), which persists along a parameter curve \(\{(a_{\mathrm{cusp}}(\varepsilon ),c_{\mathrm {cusp}}(\varepsilon ),\varepsilon ) : \varepsilon \in[0,\varepsilon _{0}]\}\) in \((a,c,\varepsilon )\) parameter space:
Lemma 2
For fixed \(0<\varepsilon \ll1\), the normal form (15) has for \(c\approx c_{\mathrm{cusp}}\) a codimension2 cusp bifurcation point.
This lemma indicates that the cusp has no singular nature^{8} with respect to the limit \(\varepsilon \to0\). For fixed \(0< c< c_{\mathrm{cusp}}\), we have three equilibrium states, at least two of which are located on \(S_{r}\) (see Assumption 3). This changes the global bifurcation structure of our problem. While we still observe a singular Andronov–Hopf bifurcation with respect to the parameter a, the growth of the limit cycles is, however, bounded as one approaches a homoclinic connection towards one of these additional equilibria (which is of saddle type). The next theorem discusses this scenario, which describes a first transition from type II excitability towards the type I limiting situation. We formulate the results concerning the main system, but we will make the description according to the parameters introduced for the normal form (15) in Sect. 3.1. For the canonical model (1), the relation between a and I is trivial, while in general the relation between \((I,\lambda)\) and \((a,c)\) may be complicated, though Sect. 3.1 entails a procedure on how to compute the change of parameters.
Remark 8
We denote by \(c_{\mathrm{sn}}^{}>0\) the ccoordinate where one of the SN branches intersects the caxis (‘SNIC of canard type’ in Fig. 5). In the normal form (15) excluding the higher order (bigoh) terms, we know that \(0< c_{\mathrm{sn}}^{}< c_{\mathrm {cusp}}\). We assume that this is also the case with the bigoh terms included.
Theorem 2
 1.
For \(a_{\mathrm{sn}}^{+}< a\), the fold \(F^{}\) is of regular jump type and a large stable relaxation cycle exists.
 2.
At \(a=a_{\mathrm{sn}}^{+}\), a saddlenode bifurcation of singular points on the middle branch \(S_{r,\varepsilon }\) in an \(O(c)\)neighbourhood of \(F^{}\); the large relaxation cycle persists.
 3.
For \(a_{h} < a < a_{\mathrm{sn}}^{+}\), the system has a saddle \(p_{+}\) and an unstable focus/node \(p_{}\) on the middle branch \(S_{r,\varepsilon }\) surrounded by the large relaxation cycle. The unstable focus/node \(p_{}\) is closer to the fold \(F^{}\).
 4.
At \(a=a_{h}\), \(p_{}\) changes stability and a subcritical singular Andronov–Hopf bifurcation takes place; the large relaxation cycle persists.
 5.
For \(a_{c}< a< a_{h}\), repelling smallamplitude limit cycles appear around the stable focus \(p_{}\); the large relaxation cycle persists.
 6.
For \(a_{s}< a< a_{c}\), small jumpback canard cycles appear that rapidly grow in amplitude (canard explosion); the large relaxation cycle perturbs to a largeamplitude jumpforward canard cycle.
 7.
At \(a=a_{s}\), a small jumpback homoclinic loop of canard type, issued from the saddle \(p_{+}\), appears together with a stable largeamplitude canard cycle.
 8.
For \(a_{\ell} < a < a_{s}\), the small homoclinic loop breaks and only the stable largeamplitude canard cycle persists.
 9.
At \(a=a_{\ell}\), a largeamplitude homoclinic loop of canard type, issued from the saddle \(p_{+}\), appears together with the outer largeamplitude cycle.
 10.
As a decreases from \(a_{\ell}\), largeamplitude canard cycles appear that grow in amplitude until it disappears in a saddlenode bifurcation of limit cycles at \(a=a_{\mathrm{snpo}}\).
Proof
Before proving this theorem, we would like to mention that the distance between \(a_{h}\) and \(a_{c}\) is \(O(\varepsilon )\), just like in the previous singular Andronov–Hopf case, while the values \(a_{c}\), \(a_{s}\), \(a_{\ell}\) and \(a_{\mathrm{snpo}}\) are all exponentially close to each other.
For \(a>a_{h}\), the location of the singularities implies that the fold \(F^{}\) is a jump point and a stable relaxation cycle is present. This shows parts (1)–(3).
Remark 9
Note, there exists also a third equilibrium on the middle branch, an unstable node/focus denoted by n, bounded away from the fold \(F^{}\). This third equilibrium bifurcates with \(p_{+}\) along the second saddlenode branch (see Fig. 5).
Remark 10
The canard parameter value \(a_{c}\) is not strictly defined, as also a “smallamplitude limit cycle” is not strictly defined. We choose a δneighbourhood of the fold and the moment where the canard cycles grow out of this δneighbourhood along the canard explosion, we define the parameter value \(a=a_{c}(\varepsilon )\). (In other words, \(a_{c}\) lies beyond the “birth of canards”.)
Incomplete canard explosion and homoclinic saddle loops. The presence of the saddle \(p_{+}\) shows that the canard cycles cannot grow unlimitedly during the canard explosion. We give here a short overview of the proof of the presence of small canard cycles because we will use elements of the proof to show the existence of canard homoclinics.
We can further rely on the results in [20], where it is shown that the maps (26) are smooth up to and including (at its extension) the boundary \(x=x_{S}^{+}\). This implies that the canard value \(a_{s}=\varepsilon A_{\mathrm{canard}}(x_{S}^{+},\sqrt{\varepsilon })\) obtained above is actually a parameter curve along which the vector field has a homoclinic saddle loop of canard type (of ‘jumpback’ type). This proves part (7) of the theorem.
Remark 11
Around the unstable canard cycle (or homoclinic a bit later on), there appears a big relaxation oscillation of canard type. It lies close to the full relaxation oscillation, but travels an \(O(c)\)distance along the middle repelling branch (see Fig. 6). The exact distance travelled along the middle branch can be computed using slowdivergence integrals and exit–entry relations; we refer to the literature [21].
By introducing an alternative section \(\tilde{S}\) between the middle and upper branch instead of S, we can treat homoclinic saddle loops of the ‘jumpaway’ type in a completely similar way. The only thing that changes is that points of \(\tilde{S}\) undergo a largeamplitude oscillation in their way to T in positive time (travelling along \(S_{a}^{+}\) towards the jump point \(F^{+}\), jumping off towards \(S_{a}^{}\)). The smoothness of the transition maps and the application of the implicit function theorem is analogous. This defines \(a=a_{\ell}< a_{s}\) and proves part (9) of the theorem.
For \(a< a_{\ell}(\varepsilon )\), the homoclinic connection breaks into a repelling largeamplitude cycle. While a proceeds to the outside of an \(O(\varepsilon )\)neighbourhood of 0, it encounters the relaxationlike attracting cycle that surrounded all repelling cycles. They disappear in a saddlenode bifurcation of limit cycles at \(a=a_{\mathrm{snpo}}< a_{\ell}\). This proves part (10) of the theorem.
5.1 Termination of Homoclinic Saddle Loops: SNICs
In the canonical form (1), the singular Andronov–Hopf curve intersects the saddlenode bifurcation curve along which the saddle \(p_{+}\) collides with a third singularity, a node n on the middle branch \(S_{r}\). Expressing (1) in the local coordinates (14), this singular bifurcation point is marked in blue in Fig. 5 and has coordinates \((a_{\mathrm{sn}}^{},c_{\mathrm{sn}}^{})=(0,\frac{1}{4\beta})\). In this section, we discuss how this codimension2 singular bifurcation point perturbs to \(\varepsilon >0\).
The SNcurve perturbs regularly for positive values of ε to a curve \(c = c_{\mathrm{sn}}(a,\varepsilon )\). Observe that the part of the AH curve between the origin and this codimension2 bifurcation point perturbs to a wedge of canard curves, corresponding to the incomplete canard explosion discussed in Theorem 2.
As in the case of the saddle homoclinics, denote by \(x_{S}^{+}\) the intersection of the section S with the unstable separatrix of the singularity \(p_{+}\), up to the point where it becomes a saddlenode. Then the part \(\{x< x_{S}^{+}\}\) of S is the part for which the backward map \(B\colon S\to T\) is well defined. Applying the results in [20], we know that the map B has a \(C^{k}\)smooth extension (for any k) to the boundary of its definition domain. It implies that we can define \(B(x_{S}^{+},\varepsilon ,a,c_{\mathrm{sn}}(a,\varepsilon ))\) and also \(F(x_{S}^{+},\varepsilon ,a,c_{\mathrm{sn}}(a,\varepsilon ))\), like in (26) but where we made the dependence of c explicit. Therefore, solving \(FB=0\) using the implicit function theorem with respect to the rescaled parameter A (recall \(a=\sqrt{\varepsilon }A\)) allows us to prove the presence of saddlenode homoclinics of (‘jumpback’) canard type.
As before, replacing the section S with the alternative \(\tilde{S}\), we can apply the same reasoning to prove the presence of a canard value along which there is a saddlenode homoclinic of (‘jumpaway’) canard type. Since the canard values \(A_{\mathrm{canard}}\) are smooth in terms of \(x_{S}\), we clearly see that the homoclinic loop curves (of jumpback and jumpaway type) terminate at corresponding saddlenode homoclinic at the saddlenode bifurcation curve.
The exponential wedge on the SNcurve between the two terminal points are terminal points of SNIC canard curves that correspond to heteroclinic canards as shown in Fig. 7. Visually, it is clear how in Fig. 7, a heteroclinic connection tends towards a SNIC as n and \(p_{+}\) approach each other in a SN bifurcation. The method of proof is similar to the one exposed before.
Remark 12
In the case \(c_{\mathrm{sn}}^{}< c< c_{\mathrm{cusp}}\) fixed, the second SNbifurcation is located ahead of the singular (subcritical) AHbifurcation and we observe a complete canard explosion (including a SNPO bifurcation of canard cycles).
In the case \(c_{\mathrm{cusp}}< c< c_{\mathrm{bautin}}\) fixed, there is only the singular (subcritical) AHbifurcation and we also observe a complete canard explosion (including a SNPO bifurcation of canard cycles).
6 Type I Excitability: Singular Bogdanov–Takens/SNIC Bifurcation
Assumption 6
Remark 13
Under these conditions it is well known that in εdependent rescaled coordinates, a regular Bogdanov–Takens bifurcation takes place; see [16]. As a consequence, the presence of smallamplitude homoclinics is clear in some parameter subset. Furthermore, as (singular) Andronov–Hopf bifurcations form part of the bifurcation diagram, canardtype orbits are present. Indeed, the double singularity in the slow dynamics at \(x=0\) may unfold in a way that the fold point becomes a canard point and an extra saddlesingularity in the slow dynamics on the middle branch may appear. In that way, an incomplete canard explosion can be observed that terminates in a canardtype saddle homoclinic (‘jumpback’, without ‘head’). In fact, these are phenomena that appear locally near the Bogdanov–Takens fold point.
Besides the smallamplitude phenomena near the Bogdanov–Takens point, we consider orbits that are close to the singular saddlenode homoclinic loop Γ shown in Fig. 4. We expect the existence of largeamplitude saddlenode homoclinics (SNICs) and as in the previous section, we also expect largeamplitude saddle homoclinics as well as relaxation oscillations.
6.1 Blowup of the Singular Fold
As is usual in geometric desingularisation, we study the flow on the halfsphere in different (coordinate) charts. Two charts are important: the chart \(K_{1}\) (or the phasedirectional rescaling chart), and the chart \(K_{2}\) (or the family rescaling chart). The \(K_{1}\) chart is used to extend the orbits along the slow manifold (which are directed towards the fold) to a neighbourhood that is at distance \(O(\varepsilon )\) from the origin. While this transition is the most technical and least obvious for the reader who is not accustomed to the blowup method, it fortunately is that part where the study of system (30) agrees with the results in studies of a classical regular jump point. Hence, we do not present detailed computations in the chart \(K_{1}\), but focus on presenting important facts and refer to the literature [14, 15] for a detailed analysis.
Calculations in \(K_{1}\) reveal two hyperbolic saddle singularities, \(p_{s}\) and \(p_{n}\), and two semihyperbolic singularities \(p_{a}\) and \(p_{r}\) along the equator. Combining information from the slow–fast dynamics near \(F^{}\) with information obtained in chart \(K_{1}\) allows one to reconstruct the dynamics near the circle shown in Fig. 9.
6.2 Local and Global Codimension2 Bifurcations
System (33) describes the flow in the interior of the sphere as shown in, e.g., Fig. 9, and it can be analysed by means of classic bifurcation analysis. Together with the information obtained from the global return mechanism, we are able to describe all observed local and global bifurcations in \((A,C)\) parameter space.
Bogdanov–Takens bifurcation.
Lemma 3
For \(\varepsilon =0\), system (33) undergoes a subcritical Andronov–Hopf bifurcation when \(A=A_{h}(C)= \frac{1}{4} +\frac{1}{2} C\), \(C>1\), and a saddlenode bifurcation of singularities when \(A=A_{\mathrm{sn}}^{+}(C)=C^{2}/4\). Both bifurcation curves meet in a Bogdanov–Takens bifurcation point at \((A,C)=(1/4,1)\). Both bifurcations persist for \(\varepsilon > 0\).
Proof
There are two singular points on \(X=Y^{2}\), located at \(Y=Y_{\pm} := \frac{1}{2}(C \pm\sqrt{C^{2}4A})\). The singularity at \(Y=Y_{+}\), denoted \(p_{+}\), is always a saddle. The singularity at \(Y=Y_{}\), denoted \(p_{}\) is of focus/node type for \(C>1\) and undergoes a change in the sign of the trace along \(A_{h}=\frac{1}{4}+\frac{1}{2}C\), which indicates an Andronov–Hopf bifurcation. Along this parameter line, \(p_{}\) is weakly unstable; a basic calculation shows that the first Lyapunov coefficient is positive. Hence the Andronov–Hopf bifurcation is subcritical.
The two singular points \(p_{+}\) and \(p_{}\) collide along \(A_{\mathrm{sn}}^{+}=C^{2}/4\) indicating a saddlenode bifurcation at \(p_{\pm}\). □
Remark 14
Saddlenode homoclinic bifurcation. The following proposition states some properties of system (33) for \(\varepsilon =0\). As mentioned before, we focus on the interaction of local dynamics with the global return mechanism (31). In particular, we want to understand the dynamics of the centre separatrix of \(p_{a}\).
Proposition 2
 1.
When \(C<\frac{1}{2}\), the separatrix coming from \(p_{a}\) connects to a centrestable separatrix of the saddlenode singularity \(p_{\pm}\). The unique unstable centre separatrix of \(p_{\pm}\) connects to \(p_{n}\).
 2.
When \(C=\frac{1}{2}\), the separatrix coming from \(p_{a}\) connects to the hyperbolic attracting separatrix of the saddlenode singularity \(p_{\pm}\). The unique unstable centre separatrix of \(p_{\pm}\) connects to \(p_{n}\).
 3.
When \(C > \frac{1}{2}\), the separatrix coming from \(p_{a}\) connects directly to \(p_{n}\) along a regular orbit. This is the jump scenario. In particular, the BT point \(p_{\pm}\) (\(C=1\)) is not connected to the separatrix.
 4.
The attracting separatrix of the saddlenode point \(p_{\pm}\) and the separatrix coming from \(p_{a}\) break regularly with respect to the parameter C.
Proof
The proof uses basics from planar theory of vector fields (e.g. invariant curves, isoclines, positive invariant sets). It requires some computations, but as it concerns basic properties we have left the details for the reader.
For \(C<1/2\), we define \(W = XY^{2} + (C1)Y  \frac{1}{2}(C1)C\). Notice that the saddlenode singularity \(p_{\pm}\) is a point on the parabola \(W=0\) and that \(\dot{W}\vert _{W=0} = (12C)(Y\frac{C}{2})^{2}\) which is positive except at the SN point \(p_{\pm}\). Using information from infinity (i.e. from chart \(K_{1}\)), we see that the separatrix from \(p_{a}\) enters the region \(\{W>0\}\) which is positively invariant. Hence the ωlimit set of the separatrix has to be the vertex of the parabola. Finally, the hyperbolic separatrix of the saddlenode singularity \(p_{\pm}\) is tangent to \(\partial\{W=0\}\) which implies it lies outside the positive invariant set \(\{W>0\}\). This proves part (1).
For \(C=\frac{1}{2}\), the singularity \(p_{\pm}\) lies on the invariant parabola from Lemma 3, which then coincides with the separatrix coming from \(p_{a}\). It is not hard to verify that the hyperbolic eigenspace of the saddlenode singularity \(p_{\pm}\) coincides with the tangent space of the parabola. This proves part (2).
For \(C>\frac{1}{2}\), we define \(V= XY^{2}\frac{1}{2}Y(\frac{1}{8}\frac{C}{2})\). One can verify that \(\dot{V}\vert _{V=0} = \frac{1}{16}(2C1)^{2}<0\), so that \(\{ V\leq0\}\) is a positive invariant set. It is a symbolic computation to verify that the separatrix coming from \(p_{a}\) enters this invariant set, and hence cannot leave.^{11} On the other hand, V computed at the saddlenode point \(p_{\pm}=(\frac{1}{4} C^{2},\frac{1}{2} C)\) yields \(\frac{1}{4}(C\frac{1}{2})>0\). We conclude that the separatrix from \(p_{a}\) cannot reach \(p_{\pm}\). Since \(p_{n}\) is the only other option for a ωlimit, it shows part (3).
As for the regular breaking of the connection in part (4): we compute the stable separatrix of the saddlenode \(p_{\pm}\) and compare it with the separatrix coming from \(p_{a}\). For the comparison we choose an arbitrary section crossing \(\{V=0\}\) and parameterise it by the levels of V. It is not hard to see that the separatrix coming from \(p_{a}\) intersects any such section at Vvalues that are \(O(C\frac{1}{2})^{2}\). On the other hand, a variational computation of the stable separatrix of \(p_{\pm}\) reveals that it is given by \(V = \frac{1}{4}e^{14Y} (C\frac{1}{2}) + O(C\frac{1}{2})^{2}\). Since \(\frac{1}{4}e^{14Y}\neq0\), it explains the transversality. This finishes the proof of the theorem. □
Clearly, the saddlenode curve \(A_{\mathrm{sn}}^{+}(C)=C^{2}/4\) persists within a manifold \(A_{\mathrm{sn}}^{+}(C,\varepsilon )=C^{2}/4 + O(\varepsilon )\). The regular breaking property formulated in the proposition ensures that the results persist for \(\varepsilon >0\).
Theorem 3
There exists a parameter surface \(A_{\mathrm{sn}}^{+}(C,\varepsilon )=C^{2}/4 + O(\varepsilon )\) along which a saddlenode singularity \(p_{\pm}\) exists. On this surface, there exists a curve \(C=\frac{1}{2} + O(\varepsilon )\) along which a saddlenode homoclinic (\(\mathit{SN}\mbox{}\mathit{HOM}_{\ell}\)) connection appears containing the hyperbolic separatrix of the saddlenode. For \(C<\frac{1}{2} + O(\varepsilon )\) on this parameter surface, there is a SNIC connection containing a centre separatrix of the saddlenode. For \(C>\frac{1}{2} + O(\varepsilon )\), there is no SNIC connection.
Proof
Restrict to the saddlenode surface. Let \(\gamma_{C,\varepsilon }\) be the unstable separatrix of the saddlenode \(p_{\pm}\) that connects to \(p_{n}\). It smoothly intersects in a point \(P_{C,\varepsilon }\) the section \(\varSigma_{n}\). The global return mechanism (31) takes this point to a point \(Q_{C,\varepsilon }\) on \(\varSigma_{a}\), where \(Q_{C,0}\) lies on the centre separatrix. On the other hand, let \(\nu_{C,\varepsilon }\) be the hyperbolic stable separatrix of the saddlenode \(p_{\pm}\) that intersects \(\varSigma_{a}\) at a point \(R_{C,\varepsilon }\). From Proposition 2, we know that \(Q_{\frac{1}{2},0}=R_{\frac{1}{2},0}\), and, parameterizing the section \(\varSigma _{a}\) by a regular coordinate θ, we also know that \(\frac{\partial }{\partial C}(Q_{C,0}R_{C,0})\neq0\) at \(C=\frac{1}{2}\). Hence, we can apply the implicit function theorem to prove the presence of a curve \(C=\frac{1}{2}+O(\varepsilon )\) along which both points coincide and a saddlenode homoclinic connection appears. The rest of the statements follow easily from the properties at the singular limit. □
Resonant homoclinic bifurcation.
Proposition 3
 1.
When \(C<\frac{1}{2}\), the centre separatrix of \(p_{a}\) connects to the node \(p_{}\).
 2.
When \(C=\frac{1}{2}\), a SNbifurcation takes place (see Proposition 2).
 3.
When \(C>\frac{1}{2}\), a centre separatrix of \(p_{a}\) connects to the hyperbolic saddle \(p_{+}\), and one of unstable separatrices of the saddle connects to \(p_{n}\). The ratio of eigenvalues is given by \(\rho(C) := 24C < 0\), and the saddle is strongly resonant at \(C=\frac{3}{4}\).
 4.
For any given \(C>\frac{1}{2}\), the saddle connection breaks regularly with respect to the parameter A as one moves away from \(A_{\ell}(C)=\frac{1}{16}+\frac{C}{4}\).
Proof
Recalling V from the proof of Proposition 2, we see that along \(A_{\ell}=\frac{1}{16}+\frac{C}{4}\), \(\{V=0\}\) is invariant and hence contains the centre separatrix of \(p_{a}\). It is easy to verify that when \(C<\frac{1}{2}\), only the node \(p_{}\) lies on \(\{V=0\}\), and when \(C>\frac{1}{2}\) only the saddle \(p_{+}\) lies there. In the second case, the node \(p_{}\) is found to lie in \(\{V>0\}\). So, the unstable separatrix of the saddle in \(\{V<0\}\) can only connect to \(p_{n}\). The computation of the eigenvalues is direct.
Theorem 4
Let \(C_{\min}>\frac{1}{2}\). There exists a parameter surface \(A_{\ell}(C,\varepsilon )=\frac{1}{16}+\frac{C}{4} + O(\varepsilon )\), \(C > C_{\min}\) along which a largeamplitude saddle homoclinic (\(\mathit{HOM}_{\ell}\)) connection exists. On this surface, there exists a curve \(C=\frac{3}{4} + O(\varepsilon )\) along which the homoclinic changes stability (resonant \(\mathit{HOM}_{\ell}\)): for lower values of C, the homoclinic is stable, for larger values it is unstable. From this curve emerges a surface \(A=A_{\mathrm{snpo}}(C,\varepsilon )\) along which a SNPO bifurcation takes place. The surfaces \(A_{\mathrm {snpo}}(C,\varepsilon )\) and \(A_{\ell}(C,\varepsilon )\) are exponentially close.
Proof
The presence of the homoclinic surface follows from a reasoning completely analogous to the one in the proof of Theorem 3. The change of stability is simply an eigenvalue computation: the equation \(\rho(C)=1\) is perturbed regularly under the εperturbation.
The emergence of an SNPO branch from the resonant saddle homoclinic is standard (see [24]), and based on three features: (i) the ratio of eigenvalues is perturbed regularly upon variation of a parameter (C), (ii) the separatrix connection breaks regularly upon variation of another parameter (A), and (iii) the divergence integral along the homoclinic loop is nonzero. Properties (i) and (ii) follow directly from the singular limit analysis in Proposition 3. Property (iii) follows from the slow–fast nature of the global return mechanism: the divergence integral computation is dominated by the passages along the slow branches \(S_{a}^{\pm}\), which are both attracting and yield a contribution of the order \(K/\varepsilon \), for some \(K>0\), while the fast parts and the parts near the folds yield an \(O(1)\) contribution. While this argument does not prove that the SNPO branch is uniformly defined up to the limit, the proof of such a result is based on combining the local Dulac map of the saddle with a return mechanism. Since all properties are uniform and since the global return mechanism is sufficiently smooth up to and including the singular limit, the method for showing SNPO branches is valid uniformly in ε. □
Figure 10 summarises all observed codimension2 bifurcations and the bifurcating codimension1 branches.
Remark 15
There are no additional bifurcations (proof omitted).
Remark 16
The homoclinic surface \(A_{\ell}(C,\varepsilon )\) defined in Theorem 4 can be extended to \(C=\frac{1}{2}\), up to and including its intersection with the SNsurface \(A_{\mathrm{sn}}^{+}(C,\varepsilon )\) from Theorem 3. At the singular limit, this is seen in Fig. 10, but a proof is needed for \(\varepsilon >0\). In such a proof, one would need to blow up the vector field once more at the saddlenode singularity \((x,y) = (\frac{1}{16},\frac{1}{4})\) and at the parameter value \((a,c)=(\frac{1}{16},\frac{1}{2})\), using a family blowup, in order to uniformly separate the saddle from the node. The technical issues involved in such a construction go beyond the scope of what we intend to expose in this paper.
Remark 17
The bifurcation curves AH, SNPO, \(\mathrm{HOM}_{\ell}\) and \(\mathrm{HOM}_{s}\) shown in Fig. 8 and Fig. 10 are the same. To rigorously prove this, we would need to include the parameters \((a,c)\) in the blowup analysis, i.e. we would have to blow up the origin \((x,y,\varepsilon ,a,c)=(0,0,0,0,0)\). Again, the technicalities involved in such a construction go beyond the scope of what we intend to expose in this paper.
7 Discussion
Excitability is an important subject area in neuroscience and its modern treatment dates back to Alan Hodgkin’s seminal work [1]. His distinction of three neural excitability classes based on injected steps of currents and observed corresponding distinct frequency–current (f–I) curves still forms the basis in understanding bifurcation mechanisms of neural excitability. FitzHugh was the first to use dynamical systems techniques for the qualitative description of action potential generation and threshold phenomena [7, 8, 25]. Rinzel and Ermentrout [3, 4] provided then a mathematical framework based on bifurcation theory to distinguish between these excitability types: SNIC bifurcation for type I and Andronov–Hopf bifurcation for type II.
This dynamical systems approach pioneered by Rinzel and Ermentrout is also used to explain more complicated neural activity such as bursting patterns. Here, the inherent multiple timescale structure of neural models given through inherent slow and fast cell membrane processes is actively used to explain the bursting pattern. The bifurcation structure found in the fast subsystem provides a possible key to understanding the genesis of bursting patterns; see, e.g., [5]. Interestingly enough, the spiking pattern itself within a burst is also often a result of a multiple timescale structure. This relaxation type behaviour is typically ignored in the bursting literature, because it would mean to consider models with (at least) three time scales—fast, intermediate and slow—which has become only recently a new research focus [26–28].
On the other hand, the literature on neural excitability (see, e.g. [5]) clearly uses slow–fast decomposition to explain action potential generation for tonic spiking models, although the accompanying bifurcation analysis of such a system often ignores or just inconsequently uses the given slow/fast structure. This article actively explores the singular nature of (twodimensional) neural models and identifies a novel singular bifurcation based on the slow–fast structure—the singular Bogdanov–Takens/SNIC bifurcation—that is key to understanding type I and (part of) type II excitability. Using readily available tools and results from geometric singular perturbation theory [14–16, 19–22] and bifurcation theory [17, 24] we are able to unfold this singular bifurcation and identify important codimension2 bifurcation points—Bautin, Bogdanov–Takens, resonant homoclinic and saddlenode homoclinic—which organise the bifurcation landscape for \(\varepsilon >0\) and help to explain the transitions between type I and type II excitability. For example, based on the position of the Bautin point in parameter space we identify a supercritical Andronov–Hopf bifurcation as a clear indicator of type II excitability while a subcritical Andronov–Hopf bifurcation does not necessarily guarantee a frequency band (significantly) bounded away from zero, i.e. the model could be close to type I and thus close to a homoclinic. Another important indicator for this proximity to type I is the cusp bifurcation and its corresponding saddlenode branches. Within the cusp region there are three equilibria including a saddle which is necessary to form a homoclinic loop. In this type I–II transition regime, properties of model neurons that are considered type II might show behaviour usually associated with type I and vice versa. Care has to be taken when inferring properties from such a simple excitability classification; see also [5], where many of these bifurcations and observations have been highlighted.
Our analytical bifurcation results provide important information for the computational neuroscience community. We show that a SNIC bifurcation associated with type I excitability only exists in a small parameter regime. Thus it is more likely to observe a (large) saddle homoclinic in one parameter continuation of neural models, although it might be very close to a saddlenode and, hence, be mistaken for a SNIC. Similarly, if one observes a subcritical Andronov–Hopf bifurcation close to a saddlenode, especially where the corresponding periodic orbits terminate nearby in a (small) homoclinic, then the other observed (large) homoclinic that terminates close to a saddlenode cannot be a SNIC.
A main obstacle in a numerical bifurcation analysis is not only the stiffness of the underlying problem but also the close proximity of different bifurcation branches. Our analytical bifurcation results should be seen as a helpful guide for numerical continuation. For example, the numerical bifurcation diagrams presented, e.g., for the Morris–Lecar neural model in [29], Fig. 6, or for a 2D sodium spiking model in [30], Figs. 1–2, are incomplete since the exponentially close branches \(\mathrm{HOM}_{s}\), \(\mathrm{HOM}_{\ell}\), SNPO and SNIC (Fig. 8) are hard to distinguish numerically and the identification of certain codimension2 points (Fig. 10)—Bogdanov–Takens, resonant \(\mathrm{HOM}_{\ell}\) and \(\mathrm{SN}\mbox{}\mathrm{HOM}_{\ell}\)—is a very difficult task.
This first condition in (12) is in fact already a consequence from condition (10) together with (8).
Note that the cusp is not a ‘singular’ bifurcation since it persists in the singular limit as a bifurcation of the reduced problem, i.e. the cusp forms a regular bifurcation structure within this singularly perturbed system.
We refrain from writing the subindex 2 on \((X,Y)\) in this chart to distinguish the coordinates from \(K_{1}\).
The separatrix of \(p_{a}\) has an expansion \(Y = \sqrt{X}\frac{1}{4} + \frac{18C}{32}X^{1/2} \frac {(2C1)^{2}}{64}X^{1} + O(X^{3/2})\) as \(X\to\infty\) and \(V = \frac {(2C1)^{2}}{32}X^{1/2} + O(X^{1})\) along this separatrix.
Declarations
Acknowledgements
MW was supported by ARC Future Fellowship grant FT120100309. PDM was supported by the FWO grant G093910.
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
Authors’ Affiliations
References
 Hodgkin AI. The local electric changes associated with repetitive action in a nonmedullated axon. J Physiol. 1948;107(2):165–81. View ArticleGoogle Scholar
 Meng X, Huguet G, Rinzel J. Type III excitability, slope sensitivity and coincidence detection. Discrete Contin Dyn Syst. 2012;32(8):2729–57. MathSciNetView ArticleMATHGoogle Scholar
 Rinzel J. Excitation dynamics: insights from simplified membrane models. Fed Proc. 1985;44:2944–6. Google Scholar
 Rinzel J, Ermentrout G. Analysis of neural excitability and oscillations. In: Koch C, Segev I, editors. Neuronal modeling: from synapses to networks. Cambridge: MIT Press; 1998. Google Scholar
 Izhikevich EM. Dynamical systems in neuroscience: the geometry of excitability and bursting. Cambridge: MIT Press; 2007. Google Scholar
 Prescott SA, De Konick Y, Sejnowski TJ. Biophysical basis for three distinct dynamical mechanism of action potential initiation. PLoS Comput Biol. 2008;4(10):e1000198. View ArticleGoogle Scholar
 FitzHugh R. Thresholds and plateaus in the Hodgkin–Huxley nerve equations. J Gen Physiol. 1960;43:867–96. View ArticleGoogle Scholar
 FitzHugh R. Impulses and physiological states in theoretical models of nerve membrane. Biophys J. 1961;1:455–66. View ArticleGoogle Scholar
 Morris C, Lecar H. Voltage oscillations in the barnacle giant musclefiber. Biophys J. 1981;35(1):193–213. View ArticleGoogle Scholar
 Wechselberger M, Mitry J, Rinzel J. Canard theory and excitability. In: Nonautonomous dynamical systems in the life sciences. Cham: Springer; 2013. p. 89–132. View ArticleGoogle Scholar
 Rubin JE, Terman D. Geometric singular perturbation analysis of neuronal dynamics. In: Handbook of dynamical systems. Vol. 2. Amsterdam: Elsevier; 2002. p. 93–146. Google Scholar
 Fenichel N. Geometric singular perturbation theory for ordinary differential equations. J Differ Equ. 1979;31(1):53–98. MathSciNetView ArticleMATHGoogle Scholar
 Jones CKRT. Geometric singular perturbation theory. In: Dynamical systems (Montecatini Terme, 1994). Heidelberg: Springer; 1995. p. 44–118. (Lecture notes in mathematics; vol. 1609). View ArticleGoogle Scholar
 Dumortier F, Roussarie R. Canard cycles and center manifolds. Providence: Am Math Soc; 1996. (Memoirs of the American Mathematical Society; vol. 577). Google Scholar
 Krupa M, Szmolyan P. Relaxation oscillation and canard explosion. J Differ Equ. 2001;174(2):312–68. MathSciNetView ArticleMATHGoogle Scholar
 De Maesschalck P, Dumortier F. Slowfast Bogdanov–Takens bifurcations. J Differ Equ. 2011;250(2):1000–25. View ArticleMATHGoogle Scholar
 Kuznetsov YA. Elements of applied bifurcation theory. New York: Springer; 1995. (Applied mathematical sciences; vol. 112). MATHGoogle Scholar
 Eckhaus W. Relaxation oscillations including a standard chase on French ducks. In: Asymptotic analysis, II. Berlin: Springer; 1983. p. 449–94. (Lecture notes in mathematics; vol. 985). View ArticleGoogle Scholar
 De Maesschalck P, Dumortier F. Canard solutions at nongeneric turning points. Trans Am Math Soc. 2006;358(5):2291–334 (electronic). View ArticleMATHGoogle Scholar
 De Maesschalck P, Dumortier F. Canard cycles in the presence of slow dynamics with singularities. Proc R Soc Edinb, Sect A. 2008;138(2):265–99. View ArticleMATHGoogle Scholar
 De Maesschalck P, Dumortier F. Time analysis and entry–exit relation near planar turning points. J Differ Equ. 2005;215(2):225–67. View ArticleMATHGoogle Scholar
 Boutat M. L’étude à l’infini de la bifurcation de Takens–Bogdanov. C R Acad Sci Paris Sér I Math. 1993;316(2):183–6. MathSciNetMATHGoogle Scholar
 Wechselberger M. Extending Melnikovtheory to invariant manifolds on noncompact domains. Dyn Syst. 2002;17:215–33. MathSciNetView ArticleMATHGoogle Scholar
 Homburg AJ, Sandstede B. Homoclinic and heteroclinic bifurcations in vector fields. In: Handbook of dynamical systems. Vol. 3. Amsterdam: Elsevier; 2010. p. 379–524. Google Scholar
 FitzHugh R. Mathematical models of threshold phenomena in the nerve membrane. Bull Math Biophys. 1955;17:257–78. View ArticleGoogle Scholar
 Krupa M, Popović N, Kopell N. Mixedmode oscillations in three timescale systems: a prototypical example. SIAM J Appl Dyn Syst. 2008;7(2):361–420. MathSciNetView ArticleMATHGoogle Scholar
 Vo T, Bertram R, Wechselberger M. Multiple geometric viewpoints of mixed mode dynamics associated with pseudoplateau bursting. SIAM J Appl Dyn Syst. 2013;12(2):789–830. MathSciNetView ArticleMATHGoogle Scholar
 Nan P, Wang Y, Kirk V, Rubin J. Understanding and distinguishing three time scale oscillations: case study in a coupled Morris–Lecar system. SIAM J Appl Dyn Syst. 2015. doi:10.1137/140985494 Google Scholar
 Tsumoto K, Kitajima H, Yoshinaga T, Aihara K, Kawakami H. Bifurcations in Morris–Lecar neuron model. Neurocomputing. 2006;69:293–316. View ArticleGoogle Scholar
 Guckenheimer J, Tien JH, Willms AR. Bifurcations in the fast dynamics of neurons: implications for bursting. In: Coombes S, Bressloff PC, editors. Bursting: the genesis of rhythm in the nervous system. Singapore: World Scientific; 2005. p. 89–122. View ArticleGoogle Scholar