We’re sorry, something doesn't seem to be working properly.
Please try refreshing the page. If that doesn't work, please contact us so we can address the problem.
Stability of the splay state in networks of pulse-coupled neurons
© Olmi et al.; licensee Springer 2012
- Received: 31 January 2012
- Accepted: 9 October 2012
- Published: 22 November 2012
We analytically investigate the stability of splay states in the networks of N globally pulse-coupled phase-like models of neurons. We develop a perturbative technique which allows determining the Floquet exponents for a generic velocity field and implement the method for a given pulse shape. We find that in the case of discontinuous velocity fields, the Floquet spectrum scales as and the stability is determined by the sign of the jump at the discontinuity. Altogether, the form of the spectrum depends on the pulse shape, but it is independent of the velocity field.
PACS:05.45.Xt, 84.35.+i, 87.19.lj.
- pulse-coupled neural networks
- Floquet spectra
- splay states
The first objective of (neural) network theory is the identification of asymptotic regimes. Previous research activity led to the discovery of fully- and partially-synchronised states, clusters and splay or asynchronous states in pulse-coupled networks [1–4]. It has also been made clear that ingredients such as disorder (the diversity of neurons and the structure of connections) are very important in determining the asymptotic behaviour, as well as the possible presence of delayed interactions and plasticity [5, 6]. However, even if one restricts the analysis to identical, globally-coupled oscillators, there are very few theoretical results and they mostly concern fully-synchronised regime or specific types of neurons (e.g. the leaky integrate-and-fire model) [4, 7, 8].
In this paper, we develop a perturbative analysis for the stability of splay states (also known as antiphase states , ‘ponies on a merry-go-round’ , or rotating waves ) in ensembles of N globally pulse-coupled identical neurons. In a splay state, all the neurons follow the same periodic dynamics and their phases are evenly shifted. Accordingly, the phase, and potential, separation is of order . Splay states have been identified in experimental measurements performed on electronic circuits  and on multimode lasers . Theoretical studies have been devoted to splay states in fully-coupled Ginzburg-Landau equations , Josephson arrays [14, 15], laser models , traffic models , unidirectionally delay-coupled Stuart-Landau oscillators  and pulse-coupled neuronal networks . In the latter context, splay states have been mainly investigated in leaky-integrate-and-fire (LIF) neurons [2, 3, 19–21], but some studies have been also devoted to the θ-neurons  and to more realistic neuronal models . Finally, splay states are important in that they provide the simplest instance of asynchronous behaviour and can be thereby used as a testing ground for the stability of a more general class of dynamical regimes.
Our model neurons are characterised by a membrane potential u that is continuously driven by the velocity field from the resetting value toward the threshold (see the next section for a more precise definition). As threshold and resetting value can be identified with one another and thereby u interpreted as a phase, it will be customary to refer to the case as to that of a discontinuous velocity field. Additionally, we assume that the post-synaptic potential (PSP) has a stereotyped shape, the so-called α-pulse, that is characterised by identical rise and decay time . The linear stability analysis reveals that the eigenvectors are characterised by different spatial frequencies (when moving from the neuron with the smallest to that one with the largest membrane potential). It is therefore convenient to use the frequency ϕ (scaled to the average phase separation ) to parametrise the Floquet spectrum. As already discussed in , there exist two components, namely short (SWs) and long (LWs) wavelengths. SWs vary on ‘microscopic’ scales, i.e. correspond to finite ϕ values: they are typically marginally stable in the thermodynamic limit (). LWs vary on scales of order , i.e. correspond to vanishing frequencies: they have been studied in the context of mean-field theory, i.e. by analysing a suitable functional equation for the probability distribution of the membrane potential u [2, 24]. By developing an approach that is valid for arbitrary coupling strength and is perturbative in the inverse system-size , here we prove that the Floquet spectrum scales as and is proportional to . We are also able to determine the spectral shape and find it to be independent of the structure of the velocity field. The transition from SWs to LWs is signalled by the occurrence of a singularity in the spectrum for the frequency . In the crossover region (where k is large but small compared to N), we show that the exponents remain finite and coincide with those determined in the weak coupling limit by Abbott and Van Vreeswijk  with their mean field approach. This result is non-trivial, since it is not a priori obvious that the ‘macroscopic’ description discussed in  is fully contained in the ‘microscopic’ description derived in this paper, as they refer to two different levels of description.
More specifically, we first build a suitable event-driven map and expand it in powers of (a posteriori, we have verified that it is necessary to reach the fourth order). Afterwards, an expression of the splay state is determined: this task corresponds to finding a fixed point of the event-driven map in a suitably moving reference frame - analogously to what has previously been done in specific contexts [21, 25, 26]. In practise this task is carried out by first taking the continuum limit for various orders and then obtaining suitable differential equations. The solutions of such equations show that all finite-size corrections for both the period T and the membrane potential vanish up to the third order. Next, the stability analysis is carried out to determine the leading term of the Floquet spectrum. This task involves the introduction of a suitable Ansatz to decompose each eigenvector into the linear superposition of a slow and rapidly oscillating component. The following continuum limit shows that the two components satisfy an ordinary and a differential equation, respectively.
Altogether, the proof of our main result requires determining all terms up to the third order in the expansion of the splay state solution, while some third-order terms are not necessary for the tangent space analysis. Going beyond discontinuous fields would require extending our analysis to account for higher-order terms and this might not even be sufficient to characterise analytic velocity fields. In fact, previous numerical simulations  suggest that the Floquet exponents scale with higher powers of depends on which derivatives of are eventually discontinuous. Moreover, it is worth recalling that in the case of a strictly sinusoidal field, the theorem proved by Watanabe and Strogatz  implies that Floquet exponents ( for a splay solution) vanish exactly for any value of N.
From the analysis of the SW spectra, one can conclude that the splay state is stable in excitatory (inhibitory) networks whenever (). These conditions can be extended also to the crossover region, where our results coincide with those obtained in  (in the limit of a small coupling). Our analytical studies cannot, however, predict the behaviour of the LW component that may be responsible for the emergence of new collective solutions in excitatory networks [3, 28]. The overall scenario is partially reminiscent of the stability of synchronous and clustered regimes that is determined by the sign of the first derivative of the velocity-field averaged on the interval (the latter problem has been investigated in excitatory pulse-coupled integrate-and-fire oscillators subject to δ-pulses [1, 29]).
Section 2 is devoted to the introduction of the model and to a brief presentation of the main results, including an expression for the leading correction to the period for the LIF model, to provide evidence that it is typically of the fourth order. A general perturbative expression for the map is derived in Section 3, while Section 4 is devoted to deriving the splay state solution up to the third order in . The main result of the paper is discussed in Section 5, where the Floquet spectra are finally obtained. Section 6 contains some general remarks and a discussion of open problems. The technical details of some lengthy calculations have been confined in the appendices: Appendix A is devoted to the derivation of the splay state solution; Appendix B contains the derivation of the leading term (of order four) of the period T for the LIF model; Appendix C is concerned with the linear stability analysis.
where the sum in the rhs represents the source term due to the spikes emitted at times .
where is the interspike time interval and, for the sake of simplicity, we have introduced the new variable .
In this paper we focus on a specific solution of the network dynamics, namely on splay states, which are asynchronous states, where all neurons fire periodically with the period T and two successive spike emissions occur at regular intervals . A typical ‘raster plot’ for this state is reported in Figure 1(b).
In the large-N limit, it is natural to consider as a smallness parameter and thereby to expand the evolution equations in powers of . In order to perform this expansion, the unique condition to require is the differentiability of the velocity field in the definition interval . The only exception is represented by the boundaries of the interval where discontinuity is allowed.
where encodes the information on the pulse dynamics (see Eq. (71)). We did not dare to estimate the quartic contribution for generic velocity fields, not only because the algebra would be utterly complicated, but also since our main motivation is to determine the leading contributions in the stability analysis, and it turns out that it is sufficient to determine the splay state up to the third order.
where represents a zeroth-order phase, while and are the real and imaginary parts of the Floquet exponents, respectively.
Notice that since the total number of exponents is (the zero exponent has been removed by taking the Poincaré section), we are going to miss one of them. Furthermore, as shown in , the N th and th exponents are associated to the field evolution and they will be not considered in our analysis.
i.e. for discontinuous velocity fields, the real part of the spectrum scales as , while the imaginary part is of even higher order.
For continuous fields, it has been numerically observed that the scaling of the spectrum is at least . In other words, the shape of the spectrum is universal, apart from a multiplicative factor that vanishes if and only if , i.e. for true phase rotators where coincides with .
This expression, which holds in the limit of (), can be compared with the results obtained in , in the small-coupling limit (), for sufficiently large k. The equivalence of the methods has two implications: (i) the crossover component is ‘universal’ as it is valid also for large coupling constants; (ii) the ‘macroscopic’ stability is fully contained in our ‘microscopic’ analysis. In fact, numerical studies reveal a perfect correspondence also for the LW component that is not amenable to an analytic treatment [25, 30].
From Eqs. (7) and (8), it follows that the stability of the splay state can be inferred, for arbitrary coupling strength, from the sign of : in excitatory (inhibitory) networks, the state is stable whenever (). The same result was previously reported in the weak coupling limit in .1 It is, however, necessary to point out that such condition(s) do not account for instabilities that can arise in the LW component. This is, e.g., the case of the onset of partial synchronisation via a supercritical Hopf bifurcation [3, 28].
where . The time at which the st spike is emitted can be determined implicitly from Eq. (10) by setting since, by definition of the model, . In this reference frame the splay state corresponds to a fixed point of the map.
where one can further eliminate with the help of Eq. (3).
where we have introduced the short-hand notation for (and analogously for ℱ).
By introducing Eq. (16) in Eq. (10) and eliminating the n dependence, we obtain a recursive equation for the variable . The fixed-point solution corresponds to the ‘trajectory’ that, starting from , ends in and can be found by tuning the ‘parameter’ . The existence of one or more solutions is related to the dependence on T. Simple calculations reveal that , i.e. it is independent of T; moreover, since is the integral from time 0 to time of a positive defined function (), it is a monotonically increasing function of T which vanishes in zero. Accordingly, for any function F, the minimal value of is g (obtained for ), while the maximal value is unbounded from above. Therefore, there exists one and only one solution provided that .
where the variables are defined in Appendix A. Notice the dependence on variables is hidden in terms.
It is important to stress that the event-driven neuronal evolution in the comoving frame implies that , i.e. the first neuron will fire at the next step, and , i.e. the membrane potential of the last neuron has been just reset to zero. This implies that and , while for any .
By inserting this expansion into Eq. (19), we obtain an equation that can be effectively split into terms of different order that will be analysed separately. Notice that by retaining terms of order h, it is possible to determine the original variables at order .
4.1 Zeroth-order approximation
and where, for the sake of simplicity, the prime denotes derivative with respect to the variable and the dependence of F and on has been dropped.
4.2 First-order approximation
where is defined by Eq. (28). The further condition to be satisfied, , implies and thereby we have , i.e. first-order corrections vanish both for the period and the membrane potential.
4.3 Second-order approximation
which has the same structure as Eq. (30). Since one has also to impose the same boundary conditions as for the first order, namely , we can conclude that and, consequently, . Therefore, second-order corrections are absent too.
4.4 Third-order approximation
Therefore, we can safely conclude that third-order terms vanish too.
The LIF model can be solved exactly for any value of N, starting from the asymptotic value (). As shown in Appendix B, it turns out that the leading corrections are of the fourth order for both the period T and the membrane potential.
The fixed-point analysis has revealed that the finite-size corrections to the stationary solutions are of order . Since such deviations do not affect the leading terms of the linear stability analysis (as it can be verified a posteriori) they will be simply neglected. Therefore, for the sake of simplicity, from now on and will be simply referred to as T and .
where the variables are polynomials of τ defined in Appendix C.1.
where the auxiliary variables are defined in Appendix C.1.
where is, in principle, a complex number and, for the sake of simplicity, we have dropped its dependence on k. Finally, as already shown at the zeroth order, the eigenvalues correspond to a pure rotation (specified by ) with no expansion or contraction, i.e. .
that is the object of our investigation. The overline means that the function is evaluated in , corresponding to the infinite N limit.
5.1 Continuum limit
Similarly to the splay state estimation, it is convenient to take the continuum limit. However, at variance with the previous case, now one should take into account also the presence of fast scales associated to the ‘spatial’ dependence of .
are slowly varying variables. On the one hand, the existence of the slow component follows from the analogy with the real-space evolution. On the other hand, the presence of the rapidly oscillating terms , first noticed in Ref.  in the uncoupled limit, suggests the presence of the second slow field, namely . Anyway, the correctness and uniqueness of Ansatz (42) is ensured a posteriori by the consistency of the equations that are obtained for the various perturbation orders.
where . This allows expanding around , similarly to what has been done in Eq. (21). At variance with the computation of the fixed point, now there are also terms like and , whose computation requires a similar expansion, but around . By incorporating all the expansion terms within Eq. (41), we have finally an equation, where terms of different orders are naturally separated from one another. The calculations are summarised in Appendix C, and the final equation is (86). By separately treating the different orders, we obtain differential and ordinary equations for the Θ and Π variables. It turns out that it is necessary to consider in parallel different orders in the fast and slow terms to obtain Θ and Π to the same order. As a consequence, we will see that it is sufficient to expand up to order .
5.2 Zeroth-order approximation
where we made use of the definition (28) and is a suitable integration constant.
i.e. the eigenvectors are independent of the phase and equal to one another. In other words, the degeneracy has not been lifted.
5.3 First-order approximation
where is an integration constant associated with the solution of the previous equation.
Altogether, we can conclude that the second-order correction to the Floquet exponent vanishes as well, and one cannot lift the degeneracy among the eigenvectors.
5.4 Second-order approximation
where is an integration constant associated with the solution of the previous differential equation.
Accordingly, is real and depends on the difference between and , confirming the numerical findings in . Therefore, the imaginary terms are smaller than .
We have derived analytically the short-wavelength component of the Floquet spectrum of the splay solution in a fully-coupled network composed of generic suprathreshold pulse-coupled phase-like neurons in the large-N limit. Our analysis has revealed that, for discontinuous velocity fields, the spectrum scales as and the stability is controlled by the sign of the difference between the velocity at reset and at threshold. The shape of the spectrum is otherwise independent of the velocity field. It would be interesting to investigate the role of the pulse shape as well. As long it follows from a linear evolution equation, such as Eq. (3), it is indeed possible to replicate the analysis carried out in this paper. Numerical studies suggest that the scaling behaviour is truly universal, while the shape of the Floquet spectrum depends on the pulse shape . It would be interesting to discover whether and which pulse shapes may give rise to SW instabilities.
Networks of LIF neurons coupled via δ-like pulses are characterised by a finite (in)stability of the whole spectrum . The difference with the case of α-pulse is so strong that it cannot be reconciled even by taking the limit (zero pulse-width), indicating that the limits and zero pulse-width do not commute . This reveals that even the development of general stability theory of the simple splay states requires some further progress.
Finally, notice that although our analytical approach is able to cover the entire SW component and the crossover region, it does not cover the truly long wavelengths which require going beyond a perturbative approach.
where we have reported also the expansion of that is necessary to pass from expression (14) to (19). Please notice that while the membrane potentials and the period are expanded up to , as in (18) and (17), here we limit the expansion to terms, since the field variables appearing in the event-driven map are integrated over an interspike-interval (see (9)).
where the overline means that the function is computed in , which corresponds to the infinite N limit.
while identifies a term of order that is multiplied by . Since, while proceeding from lower- to higher-order terms, we find that (for ), it is not necessary to give the explicit expression of the functions as they do not contribute at all.
while we do not provide explicit expressions for as they turn out to be irrelevant.
Now we are in the position to analyse the different orders.
B.1 Zeroth order
which coincides with Eq. (23) with .
B.2 From first to third order
which thereby implies that first-, second- and third-order corrections vanish also for the membrane potential.
B.3 Fourth order
where the dependence of τ on n has been dropped, since we are considering a linearisation around the splay state.
where we exploited the equality , which follows from the fact that in the thermodynamic limit .
By now substituting the expansion (40) and retaining the leading terms, we obtain Eq. (41).
We thank Mathias Wolfrum for illuminating discussions in the early stages of this study. This research project is part of the activity of the Joint Italian-Israeli Laboratory on Neuroscience funded by the Italian Ministry of Foreign Affairs. SO and AT are grateful to the Department of Physics and Astronomy of the University of Aarhus for the hospitality during the final write up of this manuscript and AT acknowledges the Villum Foundation for the support received, under the VELUX Visiting Professor Programme 2011/12, for his stay at the University of Aarhus.
- Mirollo RE, Strogatz SH: Synchronization of pulse-coupled biological oscillators. SIAM J Appl Math 1990, 50: 1645–1662. 10.1137/0150098MathSciNetView ArticleGoogle Scholar
- Abbott LF, van Vreeswijk C: Asynchronous states in networks of pulse-coupled oscillators. Phys Rev E 1993, 48: 1483–1490.View ArticleGoogle Scholar
- van Vreeswijk C: Partial synchronization in populations of pulse-coupled oscillators. Phys Rev E 1996, 54: 5522–5537. 10.1103/PhysRevE.54.5522View ArticleGoogle Scholar
- Golomb D, Hansel D, Mato G: Theory of synchrony of neuronal activity. In Handbook of Biological Physics. Edited by: Gielen S, Moss F. Elsevier, Amsterdam; 2001.Google Scholar
- Gerstner W, Kistler WM: Spiking Neuron Models: Single Neurons, Populations, Plasticity. Cambridge University Press, Cambridge; 2002.View ArticleGoogle Scholar
- Dayan P, Abbott LF: Theoretical Neuroscience. MIT press, Cambridge; 2001.Google Scholar
- Brunel N: Dynamics of sparsely connected networks of excitatory and inhibitory spiking neurons. J Comput Neurosci 2000, 8: 183–208. 10.1023/A:1008925309027View ArticleGoogle Scholar
- Goel P, Ermentrout B: Synchrony, stability, and firing patterns in pulse-coupled oscillators. Physica D 2002, 163: 191–216. 10.1016/S0167-2789(01)00374-8MathSciNetView ArticleGoogle Scholar
- Hadley P, Beasley MR: Dynamical states and stability of linear arrays of Josephson junctions. Appl Phys Lett 1987,50(10):621–623. 10.1063/1.98100View ArticleGoogle Scholar
- Aronson DG, Golubitsky M, Krupa M:Coupled arrays of Josephson junctions and bifurcation of maps with SN symmetry. Nonlinearity 1991, 4: 861–902. 10.1088/0951-7715/4/3/013MathSciNetView ArticleGoogle Scholar
- Ashwin P, King GP, Swift JW: Three identical oscillators with symmetric coupling. Nonlinearity 1990, 3: 585–601. 10.1088/0951-7715/3/3/003MathSciNetView ArticleGoogle Scholar
- Wiesenfeld K, Bracikowski C, James G, Roy R: Observation of antiphase states in a multimode laser. Phys Rev Lett 1990, 65: 1749–1752. 10.1103/PhysRevLett.65.1749View ArticleGoogle Scholar
- Hakim V, Rappel WJ: Dynamics of the globally coupled complex Ginzburg-Landau equation. Phys Rev A 1992, 46: R7347-R7350. 10.1103/PhysRevA.46.R7347View ArticleGoogle Scholar
- Nichols S, Wiesenfeld K: Ubiquitous neutral stability of splay-phase states. Phys Rev A 1992, 45: 8430–8435. 10.1103/PhysRevA.45.8430View ArticleGoogle Scholar
- Strogatz SH, Mirollo RE: Splay states in globally coupled Josephson arrays: analytical prediction of Floquet multipliers. Phys Rev E 1993, 47: 220–227. 10.1103/PhysRevE.47.220View ArticleGoogle Scholar
- Rappel WJ: Dynamics of a globally coupled laser model. Phys Rev E 1994, 49: 2750–2755. 10.1103/PhysRevE.49.2750View ArticleGoogle Scholar
- Seidel T, Werner B: Breaking the symmetry in a car-following model. Proc Appl Math Mech 2006, 6: 657–658. 10.1002/pamm.200610309View ArticleGoogle Scholar
- Choe C-U, Dahms T, Hövel P, Schöll E: Controlling synchrony by delay coupling in networks: From in-phase to splay and cluster states. Phys Rev E 2010., 81: Article ID 025205(R) Article ID 025205(R)Google Scholar
- Bressloff PC: Mean-field theory of globally coupled integrate-and-fire neural oscillators with dynamic synapses. Phys Rev E 1999, 60: 2160–2170. 10.1103/PhysRevE.60.2160MathSciNetView ArticleGoogle Scholar
- Chow CC, Kopell N: Dynamics of spiking neurons with electrical coupling. Neural Comput 2000, 12: 1643–1678. 10.1162/089976600300015295View ArticleGoogle Scholar
- Zillmer R, Livi R, Politi A, Torcini A: Stability of the splay state in pulse-coupled networks. Phys Rev E 2007., 76: Article ID 046102 Article ID 046102Google Scholar
- Dipoppa M, Krupa M, Torcini A, Gutkin BS: Splay states in finite pulse-coupled networks of excitable neurons. SIAM J Appl Dyn Syst 2012, 11: 864–894. 10.1137/110859683MathSciNetView ArticleGoogle Scholar
- Brunel N, Hansel D: How noise affects the synchronization properties of recurrent networks of inhibitory neurons. Neural Comput 2006, 18: 1066–1110. 10.1162/neco.2006.18.5.1066MathSciNetView ArticleGoogle Scholar
- Treves A: Mean-field analysis of neuronal spike dynamics. Netw Comput Neural Syst 1993, 4: 259–284. 10.1088/0954-898X/4/3/002MathSciNetView ArticleGoogle Scholar
- Calamai M, Politi A, Torcini A: Stability of splay states in globally coupled rotators. Phys Rev E 2009., 80: Article ID 036209 Article ID 036209Google Scholar
- Olmi S, Livi R, Politi A, Torcini A: Collective oscillations in disordered neural networks. Phys Rev E 2010., 81: Article ID 046119 Article ID 046119Google Scholar
- Watanabe S, Strogatz SH: Constants of motion for superconducting Josephson arrays. Physica D 1994, 74: 197–253. 10.1016/0167-2789(94)90196-1View ArticleGoogle Scholar
- Mohanty PK, Politi A: A new approach to partial synchronization in globally coupled rotators. J Phys A 2006, 39: L415-L421. 10.1088/0305-4470/39/26/L01MathSciNetGoogle Scholar
- Mauroy A, Sepulchre R: Clustering behaviors in networks of integrate-and-fire oscillators. Chaos 2008., 18: Article ID 037122 Article ID 037122Google Scholar
- Olmi S, Politi A, Torcini A: In preparation, 2012. Olmi S, Politi A, Torcini A: In preparation, 2012.Google Scholar
- Zillmer R, Livi R, Politi A, Torcini A: Desynchronization in diluted neural networks. Phys Rev E 2006., 74: Article ID 036203 Article ID 036203Google Scholar
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.