1: \documentclass[letterpaper,10pt]{article}
2: \usepackage{amsmath,mathptmx,graphicx,eprint}
3:
4: \setreportheading{Masahiko Yoshioka}{PREPRINT}
5:
6: \begin{document}
7:
8: \title{Cluster synchronization in an ensemble of neurons interacting
9: through chemical synapses}
10: \author{Masahiko Yoshioka}
11: \date{August 22, 2003}
12: \author{Masahiko Yoshioka\thanks{Electric Address:
13: myosioka@brain.riken.go.jp} \\
14: {\normalsize\it Brain Science Institute, The Institute of Physical and Chemical
15: Research (RIKEN)}\\
16: {\normalsize\it Hirosawa 2-1, Wako-shi, Saitama, 351-0198, Japan}}
17: \date{August 22, 2003\\
18: {\normalsize (Revised on May 18, 2005)}}
19:
20: \maketitle
21:
22: \begin{abstract}
23: In networks of periodically firing spiking neurons that are
24: interconnected with chemical synapses, we analyze cluster state, where
25: an ensemble of neurons are subdivided into a few clusters, in each of
26: which neurons exhibit perfect synchronization. To clarify stability of
27: cluster state, we decompose linear stability of the solution into two
28: types of stabilities: stability of mean state and stabilities of
29: clusters. Computing Floquet matrices for these stabilities, we
30: clarify the total stability of cluster state for any types of neurons and any
31: strength of interactions even if the size of networks is infinitely
32: large. First, we apply this stability analysis to investigating
33: synchronization in the large ensemble of integrate-and-fire (IF)
34: neurons. In one-cluster state we find the
35: change of stability of a cluster, which elucidates that in-phase
36: synchronization of IF neurons occurs with only inhibitory synapses. Then, we
37: investigate entrainment of two clusters of IF neurons with
38: different excitability. IF neurons with fast decaying synapses show the
39: low entrainment capability, which is explained by a pitchfork
40: bifurcation appearing in two-cluster state with change of synapse decay
41: time constant. Second, we analyze one-cluster state of Hodgkin-Huxley
42: (HH) neurons and discuss the difference in synchronization properties
43: between IF neurons and HH neurons.
44: \end{abstract}
45:
46: \section{Introduction}
47:
48: It has been revealed that periodically firing interneurons exhibit
49: in-phase synchronization during
50: the gamma oscillations (20-80 Hz) and the sharp wave burst (100-200
51: Hz)\cite{buzsaki3}. Interneurons are found to be connected through
52: inhibitory chemical synapses. Therefore, a significant effort has been
53: devoted to understand a role of inhibitory chemical synapses in in-phase
54: synchronization in a large ensemble of neurons\cite{wang}. One major
55: analytical approach to investigate synchronization of neurons is the
56: phase reduction method, in which behavior of periodically firing neurons
57: are reduced to the simple phase
58: dynamics\cite{ermentrout,hansel,bressloff3,kuramoto}. This phase
59: reduction method is, however, applicable only to the case of weak
60: couplings. To understand a role of strong couplings in synchronization
61: of neurons we have to adopt different approach.
62:
63: One difficulty in investigating strongly coupled neurons is time delayed
64: interactions due to chemical synapses. Taking account of these time
65: delayed interactions Hansel {\it et al.} have computed Floquet matrix
66: and analyzed synchronization in a couple of strongly coupled
67: neurons\cite{hansel}. The size of this Floquet matrix, however,
68: increases as the size of neural networks increases. Therefore, it is
69: difficult to apply their approach to investigating the large size of
70: neural networks.
71:
72: Bressloff {\it et al.} have presented another scheme to deal with
73: chemical synapses, which allows us to analyze stability of networks of
74: integrate-and-fire (IF) neurons without computing the explicit form of
75: Floquet matrix\cite{bressloff3}. In some large size of neural networks
76: they have found the degeneracy of eigenvalues, which makes it easy to
77: analyze synchronization of many IF neurons. Actually, such degenerate
78: eigenvalues in stability analysis are found not only in IF neurons
79: but also in many synchronization phenomena induced by mean field
80: interactions. A most prominent example of this degeneracy is seen in
81: synchronization in an ensemble of chaotic oscillators such like Lorenz
82: equations and logistic maps\cite{fujisaka,kaneko,maistrenko,pikovsky}.
83: Just using the general properties of mean field
84: interactions we can decompose linear stability of synchronous state of
85: chaotic oscillators to two different components, which define the so
86: called tangential Lyapunov exponents and transversal Lyapunov exponents.
87: It must be noted that the result of this decomposition clearly indicates
88: the occurrence of degeneracy regarding transversal Lyapunov exponents.
89: Synchronization in many chaotic oscillators is thus characterized by
90: only a small number of exponents included in tangential and transversal
91: Lyapunov spectrum even if the system size is infinitely large.
92:
93: In the present paper we employ these sophisticated reduction techniques
94: in the chaos synchronization theory to investigate synchronization in
95: the large number of neurons. The target of the analysis is cluster
96: state, where an ensemble of neurons are subdivided into a few clusters,
97: in each of which neurons exhibit perfect synchronization. To evaluate
98: the degeneracy of eigenvalues we carry out the above-mentioned
99: decomposition of a linear stability and define stability of
100: mean state (tangential Floquet multipliers) and stabilities of clusters
101: (transversal Floquet multipliers). Stability of mean state elucidates
102: if cluster state is stable in the dynamics among clusters while
103: stabilities of clusters clarify whether small perturbations in each
104: cluster converge to vanish. We explicitly compute Floquet matrices of
105: these stabilities for arbitrary neuron dynamics. Therefore, we can
106: elucidate stability of cluster state for any types of neurons, even if
107: the size of networks is infinitely large and neurons are connected
108: through strong couplings.
109:
110: To give a concrete example of the present stability analysis we first
111: analyze networks of IF neurons interacting through uniform chemical
112: synapses. In this analysis, we find the change
113: of stability of a cluster, which elucidates that in-phase synchronization
114: of a large ensemble of IF neurons occurs with only inhibitory chemical
115: synapses. In addition, we investigate two clusters of neurons with
116: different excitability, and discuss the relationship of their
117: entrainment properties to the synapse decay time constant. Second, we
118: analyze one-cluster state of Hodgkin-Huxley (HH) neurons and discuss the
119: difference in synchronization condition between IF neurons and HH
120: neurons.
121:
122: The paper is organized as follows. In Sec.~\ref{sec:dynamics} we
123: present the dynamics of neural networks that include $Q$~clusters of
124: spiking neurons. In Sec.~\ref{sec:analysis}, we present the
125: analysis for cluster state of the neural networks. This analysis is
126: applied to networks of IF neurons in Sec.~\ref{sec:if}. Then, we analyze
127: synchronization of HH neurons in Sec.~\ref{sec:hodgkin}. Finally, in
128: Sec.~\ref{sec:discussion}, we give a brief summary and discuss the
129: future problems that can be solved by the present approach.
130:
131: \section{Networks of spiking neurons coupled with chemical synapses}\label{sec:dynamics}
132:
133: We consider a spiking neuron whose state is represented by
134: $n$-dimensional vector
135: \begin{equation}
136: {{\bf x}}={\left ( {v,w_{1},w_{2},\ldots,w_{n-1}} \right )}^{{\tiny\mbox{T}}},
137: \end{equation}
138: where $v$ represent the membrane potential and
139: ${\left\{ {w_l} \right \}}_{l=1,\ldots,n-1}$ describe gating of ion channels.
140: Typically, the dynamics of a spiking neuron is defined by Hodgkin-Huxley
141: (HH) equations, FitzHugh-Nagumo (FN) equations, and so on. We simply
142: represent these neuron dynamics by
143: \begin{equation}
144: \dot{{{\bf x}}}={\bf F}{\left ( {{{\bf x}}} \right )}.\label{eq:general}
145: \end{equation}
146: In the analysis in Sec.~\ref{sec:analysis}, we assume spiking neurons in
147: the form of Eq.~(\ref{eq:general}). Nevertheless, in Sec.~\ref{sec:if},
148: we will investigate integrate-and-fire (IF) neurons, which cannot be
149: expressed by Eq.~(\ref{eq:general}) since $v$ of IF neuron changes
150: discontinuously. This discontinuity of IF neuron requires a minor
151: corrections of the analysis in Sec.~\ref{sec:analysis}. We will discuss
152: this minor correction in Sec.~\ref{sec:if}.
153:
154: We assume that $N$ spiking neurons ${\left\{ {{{\bf x}}_i} \right \}}$ are interconnected
155: through chemical synapses. To describe the dynamics of synaptic
156: electric currents, we define spike timing by the time when membrane
157: potential $v_i=[{{\bf x}}_i]_1$ (the first element of vector ${{\bf x}}_i$) exceeds
158: the threshold value $\theta=0$. We represent $k$-th spike timing of
159: neurons~$i$ by $t_i(k)$, which satisfies
160: \begin{equation}
161: v_i[t_i(k)]={\left [ {{{\bf x}}_i[t_i(k)]} \right ]}_1=\theta\label{eq:spiketiming}
162: \end{equation}
163: and
164: \begin{equation}
165: \dot{v}_i[t_i(k)]={\left [ {\dot{{{\bf x}}}_i[t_i(k)]} \right ]}_1>0.
166: \end{equation}
167: Then, the dynamics of
168: networks of spiking neurons is expressed as
169: \begin{equation}
170: \dot{{{\bf x}}}_i={\bf F}_i({{\bf x}}_i)+(I_i,0,\ldots,0)^{{\tiny\mbox{T}}}, \label{eq:dynamics}
171: \end{equation}
172: where function~${\bf F}_i({{\bf x}}_i)$ represents the dynamics of
173: neuron~$i$. Variable $I_i$ represents a sum of synaptic electric currents, which is defined by
174: \begin{equation}
175: I_i=\sum_{j=1}^N J_{ij}\sum_{k=-\infty}^\infty S{\left [ {t-t_j(k)} \right ]}\label{cur},
176: \end{equation}
177: where $J_{ij}$
178: represents synaptic coupling from neuron~$j$ to neuron~$i$, and
179: function~$S(t)$ describes
180: time evolution of synaptic electric current.
181: We assume $S(t)$ taking the form
182: \begin{equation}
183: S(t)=\left\{\begin{array}{lc}
184: \displaystyle 0 & t<0,\\
185: \displaystyle \frac{1}{\tau_1-\tau_2}{\left ( {e^{-t/\tau_1}-e^{-t/\tau_2}} \right )} & 0\le t .
186: \end{array}
187: \right . \label{eq:synapse}
188: \end{equation}
189: where $0<\tau_2<\tau_1$.
190: Constants $\tau_1$ and $\tau_2$ are termed decay time and rise time,
191: respectively.
192:
193: \subsection{Neural networks composed of $Q$ clusters of neurons}
194:
195: In some problems, we have to consider neural networks including several
196: clusters of neurons, such like networks including both interneurons and
197: pyramidal neurons. Moreover, we will later study entrainment of two
198: clusters of IF neurons that have different excitability between
199: clusters. In the present study we analyze neural networks that are
200: composed $Q$ clusters of neurons. We assume that neurons share the same
201: dynamical properties within each cluster, that is, we assume
202: \begin{equation}
203: {\bf F}_i({{\bf x}})={\bf F}_q({{\bf x}}),\quad i\in U_q,\ 1\le q \le Q,\label{eq:generalf}
204: \end{equation}
205: where $U_q$ represents the set of neurons that belong to cluster~$q$.
206: In addition, we assume that synaptic couplings~$J_{ij}$ depend
207: only on
208: cluster indexes of pre and postsynaptic neurons, that is,
209: we assume synaptic coupling $J_{ij}$ of the form
210: \begin{equation}
211: J_{ij}=\tilde{J}_{qq'}/N,\quad i\in U_q,\ j\in U_{q'}.\label{generalj}
212: \end{equation}
213: Substituting Eqs.~(\ref{eq:generalf}) and (\ref{generalj}) into
214: Eqs.~(\ref{eq:dynamics}) and (\ref{cur}) we obtain the dynamics of
215: $Q$~clusters of neurons:
216: \begin{gather}
217: \dot{{{\bf x}}}_i={\bf F}_q({{\bf x}}_i)+(I_q,0,\ldots,0)^{{\tiny\mbox{T}}}, \label{eq:clusterx}\\
218: I_q=\frac{1}{N}\sum_{q'=1}^Q\sum_{j\in U_{q'}}
219: \tilde{J}_{qq'}\sum_k S{\left [ {t-t_j(k)} \right ]},\quad i\in U_q.\label{eq:clusteri}
220: \end{gather}
221: Note that synaptic electric current in Eq.~(\ref{eq:clusteri})
222: depends only on cluster
223: index~$q$ because of the assumption in Eq.~(\ref{generalj}).
224:
225:
226: \section{Analysis}\label{sec:analysis}
227:
228: \subsection{Cluster synchronization of periodically firing neurons}
229:
230: In the present analysis we focus on
231: cluster state, in which spike timing of neurons are written in the form
232: \begin{gather}
233: t_i^\ast(k)=t_q^\ast(k)=t_q^\ast+kT,\nonumber\\
234: 0\le t_q^\ast<T,\ i\in U_q,\ q=1,\ldots,Q, \label{focus}
235: \end{gather}
236: where asterisks indicates the quantity in stationary state.
237: In this state, neurons emit periodic spikes synchronously within each cluster.
238: We further assume that in cluster state not only spike timing but also
239: neuron states are synchronized within each
240: cluster (i.e., ${{\bf x}}_i^\ast={{\bf x}}_q^\ast\ (i\in U_q)$).
241: Substituting Eq.~(\ref{focus})
242: into Eqs.~(\ref{eq:clusterx}) and (\ref{eq:clusteri}),
243: we obtain the
244: dynamics of stationary state as
245: \begin{gather}
246: \dot{{{\bf x}}}_q^\ast={\bf F}_q({{\bf x}}_q^\ast)+(I_q^\ast,0,\ldots,0)^{{\tiny\mbox{T}}}, \label{sdynv}\\
247: I_q^\ast=\sum_{q'} \tilde{J}_{qq'}r_{q'}\tilde{S}{\left ( {t-t_{q'}^\ast} \right )},\label{scur}
248: \end{gather}
249: where $\tilde{S}(t)$ is defined by $\tilde{S}(t)=\sum_k S(t+kT)$ and $r_q=N_q/N$ represents the ratio of the number of neurons in
250: cluster~$q$ to the total number of neurons.
251:
252: To obtain the explicit form of cluster state we have to
253: calculate $T$ and $t_1^\ast,t_2^\ast,\ldots,t_q^\ast$ so as to
254: obtain $I_q^\ast$ and ${{\bf x}}_q^\ast$.
255: It is obvious that we can safely assume $t_{1}^\ast=0$, and we can calculate remaining $Q$~unknown
256: variables:~$T,t_{2}^\ast,t_{3}^\ast,\ldots,t_{Q}^\ast$
257: from {{{Eqs.}~{(\ref{sdynv})}}} and {{(\ref{scur})}} following the same scheme as our previous
258: study\cite{myosioka5,myosioka6}.
259: Note that we can compute these variables not only for IF neurons but
260: also for general neuron dynamics, as far as the stable
261: cluster state is concerned.
262:
263:
264: \subsection{Decomposition of linear stability}
265:
266: To investigate linear stability of cluster state we assume the
267: infinitesimal deviations of state of neurons:
268: \begin{equation}
269: {{\bf x}}_i={{\bf x}}_q^\ast+\delta{{\bf x}}_i,\quad i\in U_q\label{eq:dx}
270: \end{equation}
271: and infinitesimal deviations of spike timing:
272: \begin{equation}
273: t_i(k)=t_{q}^\ast(k)+\delta
274: t_i(k),\quad i\in U_{q}.\label{eq:dt}
275: \end{equation}
276: From Eq.~(\ref{eq:spiketiming}), we obtain
277: \begin{gather}
278: \delta t_i(k)=-\delta v_i[t_{q}^\ast(k)]/c_{q}=-{\left [ {\delta{{\bf x}}_i[t_{q}^\ast(k)]} \right ]}_1/c_{q},\nonumber\\
279: i\in U_{q}\label{eq:deltat}
280: \end{gather}
281: with
282: \begin{equation}
283: c_{q}=\dot{v}_{q}^\ast[t_{q}^\ast(k)]={\left [ {\dot{{{\bf x}}}_q^\ast[t_{q}^\ast(k)]} \right ]}_1.
284: \end{equation}
285: Note that constant $c_{q}$ is independent of $k$ because of the periodicity of the solution.
286: To obtain the relation in Eq.~(\ref{eq:deltat}), we must assume
287: continuous neuron dynamics such as HH neurons and FN neurons.
288: Note that we have to carry out the more
289: careful calculation in discontinuous dynamics like IF neuron as we will discuss in Sec.~\ref{sec:if}.
290: Expanding the dynamics in Eqs. (\ref{eq:clusterx}) and (\ref{eq:clusteri})
291: to the first order we obtain
292: the dynamics for the deviations:
293: \begin{gather}
294: \delta\dot{{{\bf x}}}_i={\bf F}^\prime_q({{\bf x}}_q^\ast) \delta{{\bf x}}_i+(\delta I_q,0,\ldots,0)^{{\tiny\mbox{T}}}\label{eq:individualx},\\
295: \delta I_q=-\frac{1}{N}\sum_{q'}\sum_{j\in
296: U_{q'}}\tilde{J}_{qq'}\sum_k S'{\left [ {t-t_{q'}^\ast(k)} \right ]}\delta t_j(k),\label{eq:individuali}
297: \end{gather}
298: where ${\bf F}^\prime_q({{\bf x}}_q^\ast)$ denotes Jacobi matrix.
299:
300: The naive evaluation of this $N\times n$-dimensional dynamics yields
301: an eigenvalue problem of the large size of matrix.
302: Therefore, for each cluster, we define mean state of neurons:
303: \begin{equation}
304: \overline{{{\bf x}}}_q=\frac{1}{N_q}\sum_{i\in U_q}{{\bf x}}_i
305: \end{equation}
306: and mean spike timing:
307: \begin{equation}
308: \overline{t}_q(k)=\frac{1}{N_q}\sum_{i\in U_q}t_i(k).
309: \end{equation}
310: Noting Eqs.~(\ref{eq:individualx}), (\ref{eq:individuali}), and (\ref{eq:deltat}), we can write the dynamics for
311: $\delta\overline{{{\bf x}}}_q$ and $\delta\overline{t}_q(k)$
312: in the closed form
313: \begin{gather}
314: \delta\dot{\overline{{{\bf x}}}}_q={\bf F}^\prime_q({{\bf x}}_q^\ast)\delta\overline{{{\bf x}}}_q+(\delta
315: I_q,0,\ldots,0)^{{\tiny\mbox{T}}},\label{eq:meanf}\\
316: \delta I_{q}=-\sum_{{q}'}\tilde{J}_{{q}{q}'}r_{{q}'}\sum_k S'{\left [ {t-t_{q'}^\ast(k)} \right ]}\delta \overline{t}_{q'}(k)\label{eq:meani}
317: \end{gather}
318: with
319: \begin{equation}
320: \delta \overline{t}_{q}(k)=-\delta\overline{v}_{q}[t_{q}^\ast(k)]/c_{q}=-{\left [ {\delta\overline{{{\bf x}}}[t_{q}^\ast(k)]} \right ]}_1/c_{q}.\label{eq:meant}
321: \end{equation}
322: Eqs.~(\ref{eq:meanf})-(\ref{eq:meant}) define the decomposed stability of the original $N$-body
323: stability.
324: We term this decomposed stability stability of mean state.
325: It must be noted that
326: stability of mean state in
327: Eqs.~(\ref{eq:meanf})-(\ref{eq:meant}) is effectively a problem in a
328: network of $Q$ neurons
329: with couplings $J_{{q}{q}'}r_{{q}'}$ since, to the first order,
330: Eqs.~(\ref{eq:meanf})-(\ref{eq:meant}) are equivalent to
331: \begin{gather}
332: \frac{d}{dt}{\left ( {{{\bf x}}_q^\ast+\delta\overline{{{\bf x}}}_q} \right )}={\bf F}_q{\left ( {{{\bf x}}_q^\ast+\delta\overline{{{\bf x}}}_q} \right )}+(I_q,0,\ldots,0)^{{\tiny\mbox{T}}},\label{eq:unperturbedmeanx}\\
333: I_q=\sum_{q'}\tilde{J}_{qq'}r_{q'}\sum_kS{\left [ {t-t_q^\ast(k)-\delta\overline{t}_{q'}(k)} \right ]}.\label{eq:unperturbedmeani}
334: \end{gather}
335:
336: Stability of mean state is a necessary condition for the
337: full stability, but not a sufficient condition.
338: To investigate synchronization of neurons in each cluster we
339: introduce deviations around the averaged state:
340: \begin{equation}
341: {{\bf x}}_i=\overline{{{\bf x}}}_q+\delta\tilde{{{\bf x}}}_i,\quad i\in U_q.
342: \end{equation}
343: Subtracting Eq.~(\ref{eq:meanf}) from Eq.~(\ref{eq:individualx}) we obtain
344: the dynamics of $\delta\tilde{{{\bf x}}}_i$ as
345: \begin{gather}
346: \delta\dot{\tilde{{{\bf x}}}}_i={\bf F}^\prime_q({{\bf x}}_q^\ast)\delta{\tilde{{{\bf x}}}}_i,\quad i\in U_q.\label{eq:tildex}
347: \end{gather}
348: Eq~(\ref{eq:tildex}) defines another decomposed
349: stability.
350: We term this decomposed stability stability of a cluster.
351: Stability of cluster~${q}$ is satisfied when
352: $N_q$ deviations $\delta\tilde{{{\bf x}}}_i\ (i\in
353: U_{q})$ converge into ${\bf 0}$.
354: These $N_q$ dynamics are, however, identical.
355: Therefore, it suffices to evaluate one set of
356: deviations to
357: determine the stability of one cluster.
358: Note that the determination
359: of the stability of a
360: cluster is effectively a problem of a single neuron dynamics
361: under the unperturbed synaptic electric current~$I_q^\ast$ since, to the
362: first order, Eq.~(\ref{eq:tildex}) is equivalent to
363: \begin{equation}
364: \frac{d}{dt}{\left ( {{{\bf x}}^\ast_q+\delta\tilde{{{\bf x}}}_i} \right )}={\bf F}_q({{\bf x}}_q^\ast+\delta\tilde{{{\bf x}}}_i)+(I_q^\ast,0,\ldots,0)^{{\tiny\mbox{T}}},\quad i\in U_q.\label{eq:efftildex}
365: \end{equation}
366:
367:
368: \subsection{Floquet matrices for stabilities of clusters}\label{sec:floquet-cluster}
369:
370: We can determine
371: stabilities of clusters following the ordinary procedure of Floquet theory.
372: Since the solution~${{\bf x}}^\ast_q$ is periodic,
373: ${\bf F}^\prime_q({{\bf x}}_q^\ast)$ is also periodic.
374: Therefore, a solution of Eq.~(\ref{eq:tildex}) is written in the form
375: \begin{equation}
376: \delta\tilde{{{\bf x}}}_i[t_q^\ast(k+1)]={\bf M}^{{{\tiny \perp}}}_q\delta\tilde{{{\bf x}}}_i[t_q^\ast(k)],\quad i\in U_q.
377: \end{equation}
378: Calculating Eq.~(\ref{eq:efftildex}) with small initial
379: perturbations
380: we can obtain every elements in matrix~${\bf M}_q^{{\tiny \perp}}$.
381: $n\times n$~matrix ${\bf M}_q^{{\tiny \perp}}$ has $n$~eigenvalues
382: ${\left\{ {\lambda_{ql}^{{\tiny \perp}}} \right \}}_{l=1,\ldots,n}$.
383: When cluster~$q$ is stable,
384: $\delta\tilde{{{\bf x}}}_i\ (i\in U_q)$ must converge to zero after a long time.
385: Therefore, the stability of cluster~$q$ is fulfilled when the largest
386: absolute eigenvalue~$|\lambda_{q1}^{{\tiny \perp}}|$ satisfies the condition
387: \begin{equation}
388: |\lambda_{q1}^{{\tiny \perp}}| < 1.\label{eq:stabilityofclusters}
389: \end{equation}
390:
391:
392: \subsection{Floquet matrix for stability of mean state}
393:
394: Determination of the stability of mean state is not an easy problem
395: since the calculation of $\delta I_{q}$ in {{{Eq.}~{(\ref{eq:meani})}}} requires long past
396: deviations of spike timing.
397: To solve this problem, following Hansel~{\it et al.}\cite{hansel}, we introduce
398: the variables:
399: \begin{eqnarray}
400: z_{{q}1}&=&\sum_{{q}'}\tilde{J}_{{q}{q}'}r_{{q}'}\sum_{t_{{q}'}^\ast(k')<t}S{\left [ {t-t_{{q}'}^\ast(k')-\delta
401: \overline{t}_{{q}'}(k')} \right ]}\label{eq:z1}\\
402: z_{{q}2}&=&\sum_{{q}'}\tilde{J}_{{q}{q}'}r_{{q}'}
403: \sum_{t_{{q}'}^\ast(k')<t}e^{-{\left [ {t-t_{{q}'}^\ast(k')-\delta\overline{t}_{{q}'}(k')} \right ]}/\tau_1}.\label{eq:z2}
404: \end{eqnarray}
405: By means of these variables we can exactly rewrite $I_{q}$ in Eq.~(\ref{eq:unperturbedmeani}) in the
406: truncated form
407: \begin{eqnarray}
408: I_{q}&=&\sum_{{q}'}\tilde{J}_{{q}{q}'}r_{{q}'}\sum_{t_{q}^\ast(k) \le
409: t_{{q}'}^\ast(k') }
410: S{\left [ {t-t_{{q}'}^\ast(k')-\delta\overline{t}_{{q}'}(k')} \right ]}\nonumber\\
411: &&+e^{-{\left [ {t-t_{q}^\ast(k)} \right ]}/\tau_2}z_{{q}1}[t_{q}^\ast(k)]\nonumber\\
412: &&+S{\left [ {t-t_{q}^\ast(k)} \right ]} z_{{q}2}[t_{q}^\ast(k)],\qquad t_{q}^\ast(k)< t.\label{eq:truncated}
413: \end{eqnarray}
414: Therefore, $\delta I_{q}$ is written as
415: \begin{eqnarray}
416: \delta I_{q}&=&-\sum_{{q}'}\tilde{J}_{{q}{q}'}r_{{q}'}\sum_{t_{q}^\ast(k) \le
417: t_{{q}'}^\ast(k') }
418: S'{\left [ {t-t_{{q}'}^\ast(k')} \right ]}\delta\overline{t}_{{q}'}(k')\nonumber\\
419: &&+e^{-{\left [ {t-t_{q}^\ast(k)} \right ]}/\tau_2}\delta
420: z_{{q}1}[t_{q}^\ast(k)]\nonumber\\
421: &&+S{\left [ {t-t_{q}^\ast(k)} \right ]}\delta z_{{q}2}[t_{q}^\ast(k)],\qquad t_{q}^\ast(k)< t.\label{eq:truncateddelta}
422: \end{eqnarray}
423: This means that once we know $\delta z_{{q}1}[t_{q}^\ast(k)]$ and
424: $\delta z_{{q}2}[t_{q}^\ast(k)]$, we can neglect past deviations of spike
425: timing $\delta\overline{t}_{q'}(k')$ that arose before $t=t_{q}^\ast(k)$.
426:
427: To take the advantage of $z_{{q} 1}$ and $z_{{q}2}$, we define the vector
428: \begin{equation}
429: {\bf y}_{q}={\left ( {[\overline{{{\bf x}}}_q]_1,\ldots,[\overline{{{\bf x}}}_q]_n,z_{q1},z_{q2}} \right )}^{{\tiny\mbox{T}}}.
430: \end{equation}
431: We safely assume $t_q^\ast\le t_{q+1}^\ast\ (q=1,\ldots,Q-1)$.
432: Then, since Eq.~(\ref{eq:meani}) is equivalent to
433: Eq.~(\ref{eq:truncateddelta}), from Eqs.~(\ref{eq:meanf}),
434: (\ref{eq:truncateddelta}), and (\ref{eq:meant}) we can show that
435: deviation~$\delta{\bf x}_q[t_q^\ast(k+1)]$ is determined from only
436: $\delta{\bf y}_q[t_q^\ast(k)]$ and $\delta\overline{t}_{q'}(k_{qq'})$
437: with
438: \begin{equation}
439: k_{qq'}=\left \{\begin{array}{cc}
440: k& q\le q',\\
441: k+1 & q'<q.
442: \end{array}
443: \right . \label{eq:defkprime}
444: \end{equation}
445: Moreover, deviations~$\delta
446: z_{q1}[t_q^\ast(k+1)]$ and $\delta z_{q2}[t_q^\ast(k+1)]$ are given as
447: \begin{eqnarray}
448: &&\delta z_{q1}[t_q^\ast(k+1)]\nonumber\\
449: &=&-\sum_{q'}\tilde{J}_{{q}{q}'}r_{{q}'}
450: S'{\left [ {t_q^\ast(k+1)-t_{{q}'}^\ast(k_{qq'})} \right ]}\delta\overline{t}_{{q}'}(k_{qq'})\nonumber\\
451: &&+e^{-T/\tau_2}\delta\label{eq:deltaz1}
452: z_{{q}1}[t_{q}^\ast(k)]+S{\left ( {T} \right )}\delta z_{{q}2}[t_{q}^\ast(k)],\\
453: &&\delta z_{q2}[t_q^\ast(k+1)]\nonumber\\
454: &=&\frac{1}{\tau_1}\sum_{q'}\tilde{J}_{{q}{q}'}r_{{q}'}
455: e^{-{\left [ {t_q^\ast(k+1)-t_{{q}'}^\ast(k_{qq'})} \right ]}/\tau_1}\delta\overline{t}_{{q}'}(k_{qq'})\nonumber\\
456: &&+e^{-T/\tau_1} \delta z_{q2}[t_q^\ast(k)].\label{eq:deltaz2}
457: \end{eqnarray}
458: Therefore, we can also determine $\delta z_{q1}[t_q^\ast(k+1)]$ and
459: $\delta z_{q2}[t_q^\ast(k+1)]$ from the above
460: mentioned variables: $\delta{\bf
461: y}_q[t_q^\ast(k)]$ and $\delta\overline{t}_{q'}(k_{qq'})$.
462: We can summarize these relationships in the form
463: \begin{eqnarray}
464: \delta{\bf y}_q[t_q^\ast(k+1)]
465: &=&\sum_{{q}'=1}^{Q}{\bf A}_{{q}{q}'}\delta{\bf
466: y}_{{q}'}[t_{q'}^\ast(k_{qq'})]
467: +{\bf B}_{q} \delta{\bf y}_{q}[t_{q'}^\ast(k)],\nonumber\\
468: &&\label{eq:relationship}
469: \end{eqnarray}
470: where
471: \begin{equation}
472: {\bf A}_{qq'}=
473: \left ( \begin{array}{cccc}
474: {\displaystyle \frac{\partial \delta{\bf y}_q[t_q^\ast(k+1)]}{\partial
475: \delta\overline{t}_{q'}(k_{qq'})}{\left ( {-\frac{1}{c_{q'}}} \right )}}&{\bf 0}&\ldots&{\bf 0}
476: \end{array}\right)
477: \end{equation}
478: and
479: \begin{equation}
480: {\bf B}_q=
481: \left ( \begin{array}{ccc}
482: {\displaystyle
483: \frac{\partial \delta{\bf y}_q[t_q^\ast(k+1)]}{\partial [\delta{\bf
484: y}_q[t_q^\ast(k)]]_1}}&
485: \ldots &
486: {\displaystyle \frac{\partial \delta{\bf y}_q[t_q^\ast(k+1)]}{\partial [\delta{\bf
487: y}_q[t_q^\ast(k)]]_{n+2}}}
488: \end{array}\right).
489: \end{equation}
490: In this equation, ${\bf A}_{qq'}\delta{\bf
491: y}_{q'}[t_{q'}^\ast(k_{qq'})]$ represents the contribution from $\delta\overline{t}_{q'}(k_{qq'})$.
492:
493: In Sec.~\ref{sec:floquet-cluster}, we obtain ${\bf M}_q^{{\tiny \perp}}$ by
494: calculating Eq.~(\ref{eq:efftildex}) with small perturbations. In the
495: similar manner, we can compute ${\bf A}_q$ and ${\bf B}_{qq'}$
496: explicitly for arbitrary neuron dynamics. We obtain $\partial
497: \delta{\bf x}_q[t_q^\ast(k+1)]/\partial
498: \delta\overline{t}_{q'}(k_{qq'})$ and $\partial \delta{\bf
499: x}_q[t_q^\ast(k+1)]/\partial [\delta{\bf y}_q[t_q^\ast(k)]]_l$ by
500: calculating Eqs.~(\ref{eq:unperturbedmeanx}) and (\ref{eq:truncated})
501: with small perturbations. Partial derivatives of $z_{q1}$ and $z_{q2}$
502: have been given in Eqs.~(\ref{eq:deltaz1}) and (\ref{eq:deltaz2}).
503: Therefore, we can compute every elements in matrices ${\bf A}_q$ and
504: ${\bf B}_{qq'}$. For the further details of calculation of ${\bf
505: A}_{q}$ and ${\bf B}_{qq'}$ see Ref.~\cite{myosioka6}(, though the
506: definitions of ${\bf A}_{q}$, ${\bf B}_{qq'}$, and so on in
507: Ref.~\cite{myosioka6} are slightly
508: different from the present ones.)
509:
510: We introduce vector
511: \begin{equation}
512: {\bf Y}(k)={\left ( {{\bf y}_1[t_1^\ast(k)]^{{\tiny\mbox{T}}} \ldots {\bf y}_Q[t_Q^\ast(k)]^{{\tiny\mbox{T}}}} \right )}^{{\tiny\mbox{T}}}.
513: \end{equation}
514: Then, we can rewrite the relationship in Eq.~(\ref{eq:relationship}) in the form
515: \begin{equation}
516: \delta{\bf Y}(k+1)={\bf M}^{{\tiny \parallel}} \delta{\bf Y}(k)
517: \end{equation}
518: with
519: \begin{equation}
520: {\bf M}^{{\tiny \parallel}}={\bf M}^{{\tiny \parallel}}_{Q}{\bf M}^{{\tiny \parallel}}_{{Q}-1}\ldots {\bf M}^{{\tiny \parallel}}_1, \label{eq:mta}
521: \end{equation}
522: where
523: \begin{equation}
524: {\bf M}^{{\tiny \parallel}}_q=\left(
525: \begin{array}{ccccccc}
526: {\bf E}& &{\bf 0}& & & & \\
527: &\ddots & & & &{\bf 0}& \\
528: {\bf 0}& &{\bf E}& & & & \\
529: {\bf A}_{q1} &\ldots&{\bf A}_{qq-1} & {\bf A}_{qq}+{\bf B}_q &{\bf A}_{qq+1}
530: &\ldots&{\bf A}_{qQ} \\
531: & & & &{\bf E}& & {\bf 0}\\
532: & {\bf 0}& & & &\ddots& \\
533: & & & &{\bf 0} & &{\bf E}
534: \end{array}
535: \right).\label{eq:mtaq}
536: \end{equation}
537: Matrix~${\bf M}_{q}^{{\tiny \parallel}}$ updates $\delta {\bf y}_{q}(k)$ to $\delta
538: {\bf y}_{q}(k+1)$, and hence matrix~${\bf M}^{{\tiny \parallel}}$ updates all the deviations.
539: $Q(n+2)\times Q(n+2)$ matrix~${\bf M}^{{\tiny \parallel}}$ has $Q(n+2)$ eigenvalues
540: ~${\left\{ {\lambda_{l}^{{\tiny \parallel}}} \right \}}_{l=1,\ldots,Q(n+2)}$, in which
541: a trivial
542: eigenvalue one is always included as in the case of ordinary Floquet matrix.
543: The stability of mean state is satisfied when all other eigenvalues are
544: less than one in absolute value, that is, the largest absolute
545: eigenvalue $|\lambda_1^{{\tiny \parallel}}|$ and the second largest absolute
546: eigenvalue $|\lambda_2^{{\tiny \parallel}}|$ satisfy
547: \begin{equation}
548: \left | \lambda_2^{{\tiny \parallel}} \right | <1= \lambda_1^{{\tiny \parallel}} .
549: \end{equation}
550:
551:
552: \section{Cluster synchronization in networks of integrate-and-fire (IF)
553: neurons}\label{sec:if}
554:
555: Let us apply the above analysis to networks of IF neurons
556: that are defined as
557: \begin{equation}
558: \dot{v_i}=-v_i+v_r+I_{{{\mbox{\footnotesize {ext}}}},q}+I_i,\qquad i\in U_q.\label{ifv}
559: \end{equation}
560: When $v_i$ exceeds the threshold value $\theta=0$, $v_i$ is reset to
561: $v_0=-1$. The resting potential $v_r$ is set to $v_r=1$, which leads
562: intrinsic firing of neurons.
563: We assume that these IF neurons are interconnected with
564: uniform couplings:
565: \begin{equation}
566: J_{ij}=\frac{g}{N}.\label{uniform}
567: \end{equation}
568: As we have mentioned, the discontinuity
569: of IF neurons require a minor correction of the
570: stability analysis in Sec.~\ref{sec:analysis}.
571: Since derivative $\dot{v}_i$ changes discontinuously at spike timing,
572: we define
573: $c_{{q}}^-=\dot{v}_{q}^\ast[t_{q}^\ast(k)-0]$
574: and $c_{{q}}^+=\dot{v}_{q}^\ast[t_{q}^\ast(k)+0]$.
575: To take account of discontinuity of $v_i$,
576: we extend the perturbed solution $v_q^\ast+\delta v_i$ before/after spike
577: timing~$t_i(k)=t_q^\ast(k)+\delta t_i(k)$ as
578: illustrated in Fig.~\ref{fig:schematic}, and then define
579: $\delta v_i^-[t_q^\ast(k)]$ and $\delta v_i^+[t_q^\ast(k)]$.
580: These deviations satisfy the condition,
581: \begin{equation}
582: \delta t_i(k)=-\delta v_i^-[t_q^\ast(k)]/c_q^-=-\delta v_i^+[t_q^\ast(k)]/c_q^+.\label{eq:cplusminus}
583: \end{equation}
584: We define two types of
585: mean state variables:
586: $\delta\overline{v}_q^-=(1/N)\sum_{i\in U_q}\delta v_i^-$ and
587: $\delta\overline{v}_q^+=(1/N)\sum_{i\in U_q}\delta v_i^+$, and two types
588: of deviations around the mean state:
589: $\delta v_i^-=\delta\overline{v}^-_i+\delta\tilde{v}_i^-$ and $\delta
590: v_i^+=\delta\overline{v}^+_i+\delta\tilde{v}_i^+$.
591: Neuron~$i$ behaves continuously in the time interval
592: $t_i(k)<t<t_i(k+1)$, during which
593: we can carry out the decomposition of
594: linear stability discussed in Sec.~\ref{sec:analysis}.
595: Therefore, noting Eqs.~(\ref{eq:tildex}), (\ref{ifv}),
596: and (\ref{eq:cplusminus}),
597: we obtain
598: \begin{equation}
599: \delta \tilde{v}_i^-[t_q^\ast(k+1)]=e^{-T}\delta \tilde{v}_i^+[t_q^\ast(k)]=\frac{c_q^+}{c_q^-}e^{-T}\delta \tilde{v}_i^-[t_q^\ast(k)].
600: \end{equation}
601: Hence, matrix~${\bf M}_q^{{\tiny \perp}}$ takes the form
602: \begin{equation}
603: {\bf M}_q^{{\tiny \perp}}={\left ( {\frac{c_q^+}{c_q^-}e^{-T}} \right )}.
604: \end{equation}
605: From Eq.~(\ref{eq:stabilityofclusters}), we obtain the condition for
606: stability of cluster~$q$:
607: \begin{equation}
608: \left |\lambda_{q1}^{{\tiny \perp}}\right|=\left |\frac{c_{{q}}^+}{c_{{q}}^-}e^{-T}\right |<1.
609: \end{equation}
610:
611: Following the similar scheme, we can derive matrices ${\bf A}_{{q}{q}'}$
612: and ${\bf B}_{{q}}$. Substituting ${\bf A}_{{q}{q}'}$ and ${\bf
613: B}_{{q}}$ into Eqs.~(\ref{eq:mta}) and (\ref{eq:mtaq}) yields ${\bf
614: M}^{{\tiny \parallel}}$, by which we can determine the stability of mean state.
615:
616: \subsection{One-cluster state of IF neurons ($Q=1$)}
617:
618: We begin with investigating one-cluster state $Q=1$. In this state, all
619: neurons take part in shaping one-cluster in-phase synchronization.
620: One-cluster solution of Eqs.~(\ref{sdynv}) and (\ref{scur}) is found
621: with only $g<1$ since too strong synaptic couplings with $g\ge 1$ leads
622: bursting of neurons. To elucidate the stability of the solution with
623: $g<1$, assuming $\tau_1=3.5$, $\tau_2=0.1\tau_1$, and
624: $I_{{{\mbox{\footnotesize {ext}}}},1}=0$, we calculate $|\lambda_{1}^{{\tiny \parallel}}|$,
625: $|\lambda_{2}^{{\tiny \parallel}}|$, and $|\lambda_{11}^{{\tiny \perp}}|$ as a function of
626: parameter $g$ as shown in Fig.~\ref{c1}{(a)}. Since the second largest
627: absolute eigenvalue of ${\bf M}^{{\tiny \parallel}}$ (i.e., $|\lambda_{2}^{{\tiny \parallel}}|$)
628: is always less than one, the stability of mean state is always
629: satisfied. However, the largest absolute eigenvalue of ${\bf
630: M}_1^{{\tiny \perp}}$ (i.e., $|\lambda_{11}^{{\tiny \perp}}|$) becomes grater than one
631: with excitatory coupling $g>0$. Therefore, the stability of a cluster
632: is satisfied with only inhibitory coupling $g<0$. These results imply
633: that while a self-coupled single neuron ($N=1$) can exhibit stable
634: periodic firing with both inhibitory and excitatory couplings, in-phase
635: synchronization of multiple neurons ($N>1$) can take place with only
636: inhibitory couplings $g<0$. Since the networks show the same
637: synchronization properties in all the decay time $\tau_1>0$,
638: $\tau_1-g$~phase diagram takes the simple form as described in
639: Fig.~\ref{c1}{(b)}. It turns out that in-phase synchronization of a
640: large number of IF neurons occurs with only inhibitory synapses in all
641: the value $\tau_1>0$.
642:
643: Figure~\ref{fig:simulation} shows the result of the numerical
644: simulations. While the networks with inhibitory couplings ($g=-0.5$)
645: exhibits the perfect in-phase synchronization, the network with
646: excitatory couplings ($g=0.5$) settles into the asynchronous state, in
647: which neurons fire periodically with uniformly distributed phase shifts.
648: Our stability analysis explains these numerically results well.
649:
650: \subsection{Two-cluster state of IF neurons ($Q=2$ and $I_{{{\mbox{\footnotesize {ext}}}},1}= I_{{{\mbox{\footnotesize {ext}}}},2}=0$)}
651:
652: We then investigate two-cluster state $Q=2$ for inhibitory coupling
653: $g<0$ assuming $r_1=r_2=0.5$ and $I_{{{\mbox{\footnotesize {ext}}}},1}=I_{{{\mbox{\footnotesize {ext}}}},2}=0$. It
654: has been shown that a couple of IF neurons exhibit a pitchfork
655: bifurcation with change of synapse decay time constant\cite{hansel}. We
656: now show that this pitchfork bifurcation occurs even in systems of two
657: clusters of neurons. Figure~\ref{bifurcation} shows $\tau_1-\varphi_2$
658: bifurcation diagram, where $\varphi_2$ denotes $t_2/T$. There are three
659: types of solutions: in-phase ($\varphi_2=0,1$), anti-phase
660: ($\varphi_2=0.5$), and out-of-phase solutions. Evaluating eigenvalues
661: of ${\bf M}^{{\tiny \parallel}}$, ${\bf M}_1^{{\tiny \perp}}$, and ${\bf M}_2^{{\tiny \perp}}$, we
662: find that the solutions denoted by thick lines satisfy the stability of
663: mean state and the stabilities of clusters.
664:
665: \subsection{Entrainment of two clusters of IF neurons with different excitability ($Q=2$ and
666: $I_{{{\mbox{\footnotesize {ext}}}},1}=0\ne I_{{{\mbox{\footnotesize {ext}}}},2}$)}
667:
668: We extend the above result to investigate the case when the excitability
669: of neurons are different between two clusters. Fixing
670: $I_{{{\mbox{\footnotesize {ext}}}},1}=0$, we investigate the behavior of $\varphi_2$ with
671: change of $I_{{{\mbox{\footnotesize {ext}}}},2}$ in {{{Fig.}~\ref{8}}}. With $I_{{{\mbox{\footnotesize {ext}}}},2}=0$, we
672: find three stable and two unstable solutions, which are consistent with
673: the preceding results in {{{Fig.}~\ref{bifurcation}}}. The in-phase solution
674: $\varphi_2=0,1$ is extended by the change of $I_{{{\mbox{\footnotesize {ext}}}},2}$ within the
675: interval $-0.019{{\raisebox{-1ex}{\mbox{$\stackrel{{<}}{\sim}$}}}} I_{{{\mbox{\footnotesize {ext}}}},2} {{\raisebox{-1ex}{\mbox{$\stackrel{{<}}{\sim}$}}}} 0.020$. In this interval
676: two clusters of neurons show synchronized firing with small phase
677: difference, that is, entrainment occurs. To examine the robustness of
678: this entrainment, we plot this range of $I_{{{\mbox{\footnotesize {ext}}}},2}$ as a function
679: of $\tau_1$ in {{{Fig.}~\ref{entrainment}}}. The remarkable feature of this
680: phase diagram is the narrow range of $I_{{{\mbox{\footnotesize {ext}}}},2}$ with short decay
681: time constant~$\tau_1$, and it is interesting that the pitchfork bifurcation described in
682: Fig.~\ref{bifurcation} explains this narrow range of $I_{{{\mbox{\footnotesize {ext}}}},2}$.
683: In this bifurcation diagram, the out-of-phase
684: solutions merge into the in-phase solutions at $\tau_1=0$. Therefore,
685: the entrained solution in {{{Fig.}~\ref{8}}} vanishes in the limit
686: $\tau_1\rightarrow 0$, and this vanishment explains the zero range of
687: $I_{{{\mbox{\footnotesize {ext}}}},2}$ at $\tau_1=0$ in {{{Fig.}~\ref{entrainment}}}.
688:
689: On the other hand, the out-of-phase solution ($\varphi_2=0.5$) is
690: considerably robust against the change of $I_{{{\mbox{\footnotesize {ext}}}},2}$, especially
691: with short $\tau_1$ ($-0.083{\raisebox{-1ex}{\mbox{$\stackrel{{<}}{\sim}$}}}I_{{{\mbox{\footnotesize {ext}}}},2}{\raisebox{-1ex}{\mbox{$\stackrel{{<}}{\sim}$}}}0.080$
692: with $\tau_1=1.5$.) Nevertheless, when we apply the external currents
693: to halves of neurons of both clusters (i.e., $Q=4,r_1=r_2=r_3=r_4=0.25,
694: I_{{{\mbox{\footnotesize {ext}}}},1}=I_{{{\mbox{\footnotesize {ext,3}}}}}=0,I_{{{\mbox{\footnotesize {ext}}}},2}=I_{{{\mbox{\footnotesize {ext}}}},4}\ne
695: 0,\varphi_1=0,\varphi_2\sim 0,\varphi_3\sim 0.5,\varphi_4\sim 0.5$), the
696: range for successful entrainment is found to be narrow
697: ($-0.016{\raisebox{-1ex}{\mbox{$\stackrel{{<}}{\sim}$}}}I_{{{\mbox{\footnotesize {ext}}}},2}=I_{{{\mbox{\footnotesize {ext}}}},4}{\raisebox{-1ex}{\mbox{$\stackrel{{<}}{\sim}$}}}0.017$ with
698: $\tau_1=1.5$). It seems that cluster synchronization easily breaks when
699: we apply heterogeneous external electric currents that cause splitting
700: of clusters.
701:
702: \section{One-cluster state ($Q=1$) in networks of Hodgkin-Huxley (HH) neurons}\label{sec:hodgkin}
703:
704: To explore the biological plausibility of synchronization in IF neurons
705: we study more realistic neuron model that is defined by HH equations.
706: A HH neuron, whose dynamics is described in
707: appendix~\ref{sec:appendix-hodgkin}, does not show intrinsic firing without external stimulus. Therefore, we apply constant external
708: electric current~$I_{\mbox{ext}}=10~(\mu A/cm^2)$ to all of HH neurons
709: and analyze synchronization in intrinsically firing homogeneous HH
710: neurons assuming the same synaptic couplings as Eq.~(\ref{uniform}).
711: Figure~\ref{fig:hodgkin} shows $\tau_1-g$~phase diagram, where the
712: condition for stable one-cluster state ($Q=1$ and $2\le N$) is
713: described. In the large area of inhibitory couplings ($g<0$) we find
714: stable in-phase synchronization. Beyond $\tau_1=7.0$ the behavior of
715: $|\lambda_{2}^{{\tiny \parallel}}|$ and $|\lambda_{11}^{{\tiny \perp}}|$ is similar to those
716: of IF neurons described in Fig.~\ref{c1}{(a)}, and the change of stability
717: occurs at $g=0$ because of $|\lambda_{11}^{{\tiny \perp}}|$. Below
718: $\tau_1=7.0$, however, synchronization with inhibitory couplings takes
719: place only below a certain negative value of $g$, and
720: excitatory couplings can induce synchronization in some conditions. $\tau_1-g$~phase
721: diagram of IF neurons (Fig.~\ref{c1}{(b)}) can explain synchronization
722: in HH neurons with slowly decaying synapses, though the synchronization
723: condition of HH neurons with fast decaying synapses is more complicated
724: than IF neurons.
725:
726: \section{Discussion}\label{sec:discussion}
727:
728: We have studied cluster state of networks of spiking neurons. We have
729: shown the analytical method that can deal with synchronization in the
730: large size of neural networks with arbitrary neuron dynamics and
731: arbitrary interactions. Employing this analysis we have investigated
732: networks of IF neurons interconnected through uniform chemical synapses. In the
733: analysis of one-cluster state, we have found the
734: change of stability of a cluster, which has elucidated that in-phase
735: synchronization of multiple IF neurons occurs only with inhibitory
736: couplings (Fig.~\ref{c1}). It must be noted that this analytical result well explains the
737: structure of interneurons in the real nervous system, where interneurons
738: are interconnected through inhibitory chemical synapses. In addition,
739: we have investigated the entrainment of two clusters of IF neurons with
740: different excitability (Fig.~\ref{entrainment}). We have explained the
741: narrow range of $I_{{{\mbox{\footnotesize {ext}}}},2}$ with short decay time constant $\tau_1$
742: in Fig.~\ref{entrainment} by the bifurcation diagram described in
743: Fig.~\ref{bifurcation}. Furthermore, we have investigated one-cluster
744: state of HH neurons. HH neurons show stable in-phase synchronization in
745: the large parameter region of inhibitory chemical synapses, though the
746: synchronization condition of HH neurons with fast decaying synapses is
747: more complicated than IF neurons (Fig.~\ref{fig:hodgkin}).
748:
749: Although van~Vreeswijk {\it et al.} have proposed another type of
750: stability criterion based on
751: function~$G(\phi)$\cite{vreeswijk1,vreeswijk2}, this stability criterion
752: is unsound in some conditions. One counterexample of their criterion is
753: a couple of IF neurons with couplings
754: $J_{11}=J_{22}=-J_{12}=-J_{21}=g/2$. With $\tau_1=3.5$ and
755: $\tau_2=0.1\tau_1$, in-phase synchronization of these neurons becomes
756: unstable beyond the critical point~$g=1.11$ as shown in
757: Fig.~\ref{fig:counterexample}. While our analysis based on linear
758: stability precisely yields this critical point, van~Vreeswijk's
759: criterion, namely, $G(\phi)=-g{\left ( {e^{-T}/2} \right )}\int_0^1 e^{T\theta}
760: {\left ( {\tilde{S}{\left [ {T{\left ( {\theta+\phi} \right )}} \right ]}-\tilde{S}{\left [ {T{\left ( {\theta-\phi} \right )}} \right ]}} \right )}d\theta$
761: with $T=\log{\left ( {v_0-v_r/v_0} \right )}$, fails to give the critical point.
762: Gerstner {\it et al.} have also investigated networks of IF
763: neurons\cite{gerstner}.
764: Their analysis, however, cannot treat the
765: realistic form of synaptic electric current~$S(t)$ that exerts the
766: long-time influence after activation within the finite size of matrix.
767:
768: The present decomposition of linear stability is simple enough to
769: investigate the general neuron dynamics including FN neurons and HH
770: neurons. Even when the behavior of neurons are chaotic\cite{feudel}, we
771: are still able to evaluate the stability of cluster state using tangential
772: Lyapunov exponents and transversal Lyapunov exponents\cite{myosioka8}, and such technique
773: may give a deeper understanding of the complicated behavior of
774: HH neurons around the arrow in Fig.~\ref{fig:hodgkin}. It is
775: interesting to apply the present analysis to networks including
776: pyramidal neurons as well as interneurons\cite{kopell}. The surface of
777: the neocortex is subdivided into numerous columnar organizations, each
778: of which is composed of several layers of neurons \cite{mountcastle}.
779: The internal and external dynamics of such columnar organizations would
780: also be the future target of the present analysis.
781:
782: \appendix
783:
784: \section{The Hodgkin-Huxley (HH) equations}\label{sec:appendix-hodgkin}
785:
786: The HH equations are the four-dimensional ordinary differential
787: equations, which describe the spike generation of the squid's giant
788: axon\cite{hodgkin}. The dynamics of a neuron state
789: vector~${{\bf x}}={\left ( {v,w_{1},w_{2},w_{3}} \right )}^{{\tiny\mbox{T}}}$ for a HH neuron is expressed as
790: \begin{eqnarray}
791: C_m\ \dot{v}&=&\overline{g}_{Na}w_2^3w_1{\left ( {v_{Na}-v} \right )}+\overline{g}_{K}w_3^4{\left ( {v_{K}-v} \right )}\nonumber\\
792: &&+\overline{g}_{L}{\left ( {v_{L}-v} \right )}+I_{\mbox{ext}},\\
793: \dot{w}_1&=&\alpha_1{\left ( {1-w_1} \right )}-\beta_1 w_1,\\
794: \dot{w}_2&=&\alpha_2{\left ( {1-w_2} \right )}-\beta_2 w_2,\\
795: \dot{w}_3&=&\alpha_3{\left ( {1-w_3} \right )}-\beta_3 w_3
796: \end{eqnarray}
797: with
798: \begin{eqnarray}
799: \alpha_1&=&0.01{\left ( {10-v} \right )}\left /{\left\{ {\exp{\left ( {\frac{10-v}{10}} \right )}-1} \right \}}\right . ,\\
800: \beta_1&=&0.125\exp{\left ( {-v/80} \right )},\\
801: \alpha_2&=&0.1{\left ( {25-v} \right )}\left /{\left\{ {\exp{\left ( {\frac{25-v}{10}} \right )}-1} \right \}}\right . ,\\
802: \beta_2&=&4\exp{\left ( {-v/18} \right )},\\
803: \alpha_3&=&0.07\exp{\left ( {-v/20} \right )},\\
804: \beta_3&=&1\left /{\left\{ {\exp{\left ( {\frac{30-v}{10}} \right )}-1} \right \}}\right .,
805: \end{eqnarray}
806: where $v_{Na}=50\mbox{ [mV]},\
807: v_K=-77\mbox{ [mV]},\ v_L=-54.4\mbox{ [mV]},\
808: \overline{g}_{Na}=120\mbox{ [mS/cm$^2$]},\ \overline{g}_K=36\mbox{
809: [mS/cm$^2$]},\ \overline{g}_L=0.3\mbox{ [mS/cm$^2$]},\ {\mbox{ and }} C_m=1\mbox{
810: [$\mu$F/cm$^2$]}.$
811: In the present study we set $I_{\mbox{ext}}=10~(\mu A/cm^2)$ to induce intrinsic firing of a HH neuron.
812:
813: \begin{thebibliography}{99}
814: \bibitem{buzsaki3} G.~Buzs{\'a}ki, Z.~Horv{\'a}th, R.~Urioste, J.~Hetke,
815: and K.~Wise, Science, 256, 1025 (1992).
816: \bibitem{wang} X.J.~Wang and G.~Buz{\'a}ki, J. Neurosci., 16, 6402 (1996).
817: \bibitem{ermentrout} G.B.~Ermentrout and N.~Kopell,
818: SIAM J. Math. Anal., 15, 215 (1984).
819: \bibitem{hansel} D.~Hansel, G.~Mato, and C.~Meunier, Neural Comp., 7,
820: 307 (1995).
821: \bibitem{bressloff3} P.C.~Bressloff and S.~Coombes, Neural Comp., 12, 91 (2000).
822: \bibitem{kuramoto} Y.~Kuramoto, Chemical oscillations, waves, and
823: turbulence (Springer-Verlag 1984).
824: \bibitem{fujisaka} H.~Fujisaka and T.~Yamada, Prog. Theor. Phys. 69, 32 (1983).
825: \bibitem{kaneko} K.~Kaneko, Physica D, 77, 456 (1994).
826: \bibitem{maistrenko} Y.~Maistrenko and T~Kapitaniak, Phys. Rev. E, 54, 3285
827: (1996).
828: \bibitem{pikovsky} A.~Pikovsky, O.~Popovych, and Yu.~Maistrenko, Phys. Rev. Lett., 87, 044102
829: (2001).
830: \bibitem{myosioka5} M.~Yoshioka, Phys. Rev. E, 65, 011903 (2002).
831: \bibitem{myosioka6} M.~Yoshioka, Phys. Rev. E, 66, 061913 (2002).
832: \bibitem{hodgkin} A.L.~Hodgkin and A.F.~Huxley, J. Physiol., 117, 500 (1952).
833: \bibitem{vreeswijk1} C.~van Vreeswijk, L.F.~Abbott, and G.B.~Ermentrout,
834: J. Comp. Neurosci, 1, 313 (1994).
835: \bibitem{vreeswijk2} C.~van~Vreeswijk, Phys. Rev. E 54, 5522 (1996).
836: \bibitem{gerstner} W.~Gerstner, J.L.~van~Hemmen, and J.D.~Cowan, Neural Comp., 8, 1653 (1996).
837: \bibitem{feudel} U.~Feudel, A.~Neiman, X.~Pei, W.~Wojtenek, H.~Braun,
838: M.~Huber, and F.~Moss, Chaos, 10, 231 (2000).
839: \bibitem{myosioka8} M.~Yoshioka, Phys. Rev. E, in press.
840: \bibitem{kopell} N.~Kopell, G.B.~Ermentrout, M.A.~Whittington, and
841: R.D.~Traub, Proc. Natl. Acad. Sci. USA, 97, 1867 (2000)
842: \bibitem{mountcastle} V.B.~Mountcastle, J. Neurophysiol. 20, 408 (1957)
843: \end{thebibliography}
844:
845: \begin{figure}[h]
846: {}{}{}
847: \begin{center}
848: \includegraphics[scale=1.5]{001.eps}
849: \end{center}
850: \caption{A schematic figure explaining the definition of
851: $\delta v_i^-[t_q^\ast(k)]$ and
852: $\delta v_i^+[t_q^\ast(k)]$.
853: Membrane potential of a IF neuron~$v_i$ changes
854: discontinuously at spike timing $t_q^\ast(k)+\delta t_i(k)$.
855: When $t_q^\ast(k)+\delta t_i(k)<t_q^\ast(k)$, we define $\delta v_i^-[t_q^\ast(k)]$ by
856: extending the solution as shown in the figure, and we define $\delta
857: v_i^+[t_q^\ast(k)]=\delta v_i[t_q^\ast(k)]$.
858: When $t_q^\ast(k)<t_q^\ast(k)+\delta t_i(k)$, we define these variables in
859: the opposite way.}\label{fig:schematic}
860: \end{figure}
861:
862: \begin{figure}
863: {}{}{}
864: {}{}{}
865: {}
866: \begin{center}
867: \begin{tabular}{ccc}
868: {(a)} & & \ \ \ \\
869: & \includegraphics[scale=1.5]{002.eps} \\
870: {(b)} \\
871: & \includegraphics[scale=1.5]{003.eps}
872: \end{tabular}
873: \end{center}
874:
875: \caption{{(a)}\ Absolute values of $\lambda^{{\tiny \parallel}}_1$,
876: $\lambda^{{\tiny \parallel}}_2$, and $\lambda^{{\tiny \perp}}_{11}$ for one-cluster state
877: ($Q=1$) of networks of IF neurons are plotted as a function of $g$ for
878: $I_{{\mbox{\footnotesize {ext,1}}}}=0$,$\tau_1=3.5$, and $\tau_2=0.1\tau_1$.
879: $\lambda^{{\tiny \parallel}}_1$ always takes one while $|\lambda^{{\tiny \parallel}}_2|$ is
880: always less than one. $|\lambda^{{\tiny \perp}}_{11}|$ is less than one only
881: when synapses are inhibitory ($g<0$). These eigenvalues behave in the same manner even with the other decay time $\tau_1>0$.
882: {(b)}\ $\tau_1$-$g$~phase diagram, where we fix $\tau_2=0.1\tau_1$.
883: A self-coupled single neuron ($N=1$) has the stable periodic solution
884: below $g=1$. However, synchronization of multiple neurons ($N>1$)
885: occurs with only inhibitory couplings $g<0$ since the stability
886: of a cluster is fulfilled with only inhibitory couplings $g<0$.
887: Beyond $g=1$, an excessive amount of positive synaptic electric
888: current leads bursting of neurons.}\label{c1}
889: \end{figure}
890:
891: \begin{figure}
892: {}{}{}
893: {}{}{}
894: {}
895: \begin{center}
896: \begin{tabular}{ccc}
897: {(a)} & & \ \ \ \\
898: & \includegraphics[scale=1.5]{004.eps} \\
899: {(b)} \\
900: & \includegraphics[scale=1.5]{005.eps}
901: \end{tabular}
902: \end{center}
903: \caption{The result of numerical simulations
904: with $N=100$, $\tau_1=3.5$, and $\tau_2=0.1\tau_1$.
905: Dots represent spike timing of neurons in a stationary state, which is
906: realized after a long run of simulation.
907: {(a)}\ With inhibitory synapses $g=-0.5$,
908: the perfect in-phase synchronization occurs.
909: {(b)}\ With excitatory synapses $g=0.5$, we observe asynchronous state, in
910: which neurons fire periodically with uniformly distributed phase shifts.}\label{fig:simulation}
911: \end{figure}
912:
913: \begin{figure}
914: {}{}{}
915: \begin{center}
916: \includegraphics[scale=1.5]{006.eps}
917: \end{center}
918: \caption{$\tau_1-\varphi_2$ bifurcation diagram for two-cluster state, where variable $\varphi_2$ denotes
919: $t_2/T$ ($Q=2$, $1\leq N_1=N_2$, $g=-3$, $\tau_2=0.1\tau_1$, and
920: $I_{{{\mbox{\footnotesize {ext}}}},1}=I_{{{\mbox{\footnotesize {ext}}}},2}=0$.)
921: The solutions represented by thick
922: lines satisfy the stability of mean state and stabilities of
923: clusters, while solutions represented by the dotted lines lack one or
924: both of stabilities.
925: The out-of-phase solutions plotted by the thin dotted
926: line ($\tau_1< 2.8$) is invalid since in these solutions
927: membrane potential $v_i$ crosses the
928: threshold $\theta$ multiple times.}\label{bifurcation}
929: \end{figure}
930:
931: \begin{figure}
932: {}{}{}
933: \begin{center}
934: \includegraphics[scale=1.5]{007.eps}
935: \end{center}
936: \caption{Entrainment of two clusters of neurons with different
937: excitability. The solution~$\varphi_2$ is plotted against
938: $I_{{{\mbox{\footnotesize {ext}}}},2}$ for the fixed value of $I_{{{\mbox{\footnotesize {ext}}}},1}=0$
939: ($Q=2,1\leq N_1=N_2,g=-3,\tau_1=3.5,$ and $\tau_2=0.1\tau_1$.)}\label{8}
940: \end{figure}
941:
942: \begin{figure}
943: {}{}{}
944: \begin{center}
945: \includegraphics[scale=1.5]{008.eps}
946: \end{center}
947: \caption{The upper and lower bounds of $I_{{{\mbox{\footnotesize {ext}}}},2}$ for
948: entrainment of two clusters of neurons are plotted against $\tau_1$
949: ($Q=2,1\leq N_1=N_2,\tau_2=0.1\tau_1,$ and $I_{{{\mbox{\footnotesize {ext}}}},1}=0$). The numbers in the
950: figure indicate the value of $g$.}\label{entrainment}
951: \end{figure}
952:
953: \begin{figure}
954:
955: {}{}{}
956: \begin{center}
957: \includegraphics[scale=1.5]{009.eps}
958: \end{center}
959: \caption{$\tau_1-g$~phase diagram for one-cluster state ($Q=1$) of
960: multiple Hodgkin-Huxley (HH) neurons ($2\le N$) under the condition
961: $\tau_2=0.1\tau_1$. ``s'' (``u'') in the figure indicates the region for
962: the stable (unstable) one-cluster state. Around the arrow we find a lot of
963: isolated regions for the stable one-cluster state. Note that we apply
964: constant external electric current~$I_{\mbox{ext}}=10~(\mu A/cm^2)$ to
965: all of HH neurons so as to induce intrinsic firing of neurons.}\label{fig:hodgkin}
966: \end{figure}
967:
968: \begin{figure}
969: {}{}{}
970: {}{}{}
971: {}{}{}
972: {}
973: \begin{center}
974: \begin{tabular}{ccc}
975: {(a)} & & \ \ \ \\
976: & \includegraphics[scale=1.3]{010.eps} \\
977: {(b)} \\
978: & \includegraphics[scale=1.3]{011.eps} \\
979: {(c)} \\
980: & \includegraphics[scale=1.3]{012.eps}
981: \end{tabular}
982: \end{center}
983: \caption{Stability of in-phase synchronization of a couple of IF neurons
984: interconnected with $J_{11}=J_{22}=-J_{12}=-J_{21}=g/2$ ($\tau_1=3.5$ and $\tau_2=0.1\tau_1$).
985: {(a)}\ Absolute values of $\lambda_1^{{\tiny \parallel}}$ and
986: $\lambda_2^{{\tiny \parallel}}$ are plotted as a function of $g$.
987: Beyond $g=1.11$, the in-phase synchronization becomes unstable.
988: {(b)}\ The result of numerical simulations with $g=1.0$.
989: A couple of neurons show in-phase synchronization.
990: {(c)}\ The result of numerical simulations with $g=1.2$.
991: Only a single neuron fires at high frequency in the winner-take-all fashion.}\label{fig:counterexample}
992: \end{figure}
993:
994: \end{document}
995: