1: \section{Main discussion}
2:
3: In nonlinear contraction theory, the analysis of dynamical systems is
4: greatly simplified by studying stability and nominal motion
5: separately. We propose a similar point of view for analyzing
6: synchronization in networks of dynamical systems. In section
7: \ref{sec:analysis}, we study specific conditions on the coupling
8: structure which guarantee exponential convergence to a linear
9: subspace. In section \ref{sec:syminv}, we examine how symmetries
10: and/or diffusion-like couplings can give rise to specific
11: flow-invariant subspaces corresponding to concurrent synchronized
12: states.
13:
14: \subsection{Some coupling structures and conditions for exponential
15: synchronization}
16:
17: \label{sec:analysis}
18:
19: \subsubsection{Balanced diffusive networks}
20:
21: \label{sec:balanced}
22:
23: A balanced network \cite{OlfMur} is a directed diffusive network which
24: verifies the following equality for each node~$i$ (see figure
25: \ref{fig:balanced} for an example)
26: \[
27: \sum_{j \neq i} \bfK_{ij}=
28: \sum_{j \neq i} \bfK_{ji}
29: \]
30:
31: Because of this property, the symmetric part of the Laplacian matrix
32: of the network is itself the Laplacian matrix of the underlying
33: undirected graph to the network\,\footnote{In fact, it is easy to see
34: that the symmetric part of the Laplacian matrix of a directed graph
35: is the Laplacian matrix of some undirected graph \emph{if and only
36: if} the directed graph is balanced.}. Thus, the positive
37: definiteness of $\bfV\bfL\bfV^\top$ for a balanced network is
38: equivalent to the connectedness of some well-defined undirected graph.
39:
40: \begin{figure}[ht]
41: \begin{minipage}[ht]{0.5\textwidth}
42: \centering
43: \includegraphics[scale=0.7]{fig/balanced.eps}
44: \end{minipage}
45: \begin{minipage}[ht]{0.5\textwidth}
46: \flushleft
47: \caption[]{A balanced network with Laplacian matrix \\
48: $\bfL=\left(
49: \begin{array}{rrrr}
50: 4&-2&0&-2\\
51: -3&3&0&0\\
52: -1&-1&2&0\\
53: 0&0&-2&2
54: \end{array}\right)$}
55: \label{fig:balanced}
56: \end{minipage} \\[10pt]
57: \begin{minipage}[ht]{0.5\textwidth}
58: \centering
59: \includegraphics[scale=0.7]{fig/balanced2.eps}
60: \end{minipage}
61: \begin{minipage}[ht]{0.5\textwidth}
62: \flushleft {Its underlying undirected graph,
63: with Laplacian matrix \\
64: $\bfL_s=\frac{\bfL+\bfL^\top}{2}=\left(
65: \begin{array}{rrrr}
66: 4&-2.5&-0.5&-1\\
67: -2.5&3&-0.5&0\\
68: -0.5&-0.5&2&-1\\
69: -1&0&-1&2
70: \end{array}\right)$}
71: \end{minipage}
72: \end{figure}
73:
74: For general directed diffusive networks, finding a simple condition
75: implying the positive definiteness of $\bfV\bfL\bfV^\top$ (such as the
76: connectivity condition in the case of undirected networks) still
77: remains an open problem. However, given a particular example, one can
78: compute $\bfV\bfL\bfV^\top$ and determine directly whether it is
79: positive definite.
80:
81: \subsubsection{Extension of diffusive connections}
82:
83: In some applications \cite{WangSlo2}, one might encounter the
84: following dynamics
85: \[
86: \left\{ \begin{array}{l}
87: \dot{\bfx}_1=\bff_1(\bfx_1,t) + k\bfA^\top(\bfB\bfx_2-\bfA\bfx_1)\\
88: \dot{\bfx}_2=\bff_2(\bfx_2,t) + k\bfB^\top(\bfA\bfx_1-\bfB\bfx_2)
89: \end{array} \right.
90: \]
91:
92: Here $\bfx_1$ and $\bfx_2$ can be of different dimensions, say $d_1$
93: and $d_2$. $\bfA$ and $\bfB$ are constant matrices of appropriate
94: dimensions. The Jacobian matrix of the overall system is
95: \[
96: \bfJ=\left(
97: \begin{array}{cc}
98: \frac{\partial\bff_1}{\partial\bfx_1}& \\
99: & \frac{\partial\bff_2}{\partial\bfx_2}
100: \end{array}
101: \right) - k\bfL, \quad \mathrm{where~} \bfL=
102: \left(
103: \begin{array}{cc}
104: \bfA^\top\bfA & -\bfA^\top\bfB \\
105: -\bfB^\top\bfA & \bfB^\top\bfB
106: \end{array}
107: \right)
108: \]
109:
110: Note that $\bfL$ is symmetric positive semi-definite. Indeed, one
111: immediately verifies that
112:
113: \[
114: \forall \bfx_1, \bfx_2 :
115: \left(
116: \begin{array}{cc}
117: \bfx_1 & \bfx_2
118: \end{array}
119: \right) \bfL \left(
120: \begin{array}{c}
121: \bfx_1 \\
122: \bfx_2
123: \end{array}
124: \right) =
125: (\bfA\bfx_1-\bfB\bfx_2)^\top (\bfA\bfx_1-\bfB\bfx_2) \geq 0
126: \]
127:
128: Consider now the linear subspace of
129: $\mathbb{R}^{d_1}\times\mathbb{R}^{d_2}$ defined by
130: \[
131: \sM=\left\{ \left(
132: \begin{array}{c}
133: \bfx_1 \\
134: \bfx_2
135: \end{array}
136: \right) \in \mathbb{R}^{d_1}\times\mathbb{R}^{d_2} :
137: \bfA\bfx_1-\bfB\bfx_2=\zeros \right\}
138: \]
139: and use as before the orthonormal projection $\bfV$ on $\sM^\perp$, so
140: that $\bfV\bfL\bfV^\top$ is positive definite. Assume furthermore that
141: $\sM$ is flow-invariant, i.e.
142: \[
143: \forall (\bfx_1,\bfx_2)\in \mathbb{R}^{d_1}\times\mathbb{R}^{d_2},
144: \left[\bfA\bfx_1=\bfB\bfx_2\right] \Rightarrow
145: \left[\bfA\bff_1(\bfx_1)=\bfB\bff_2(\bfx_2)\right]
146: \]
147: and that the Jacobian matrices of the individual dynamics are
148: upper-bounded. Then large enough $k$, i.e. for example
149: \[
150: k\lambda_\mathrm{min}(\bfV\bfL\bfV^\top)> \max_{i=1,2}\left(
151: \sup_{\bfa_i,t}\lambda_\mathrm{max}
152: \frac{\partial\bff_i}{\partial\bfx_i}(\bfa_i,t)\right)
153: \]
154: ensures exponential convergence to the subspace $\sM$.
155:
156: The state corresponding to $\sM$ can be viewed as an extension of
157: synchronization states to systems of different dimensions. Indeed, in
158: the case where $\bfx_1$ and $\bfx_2$ have the same dimension and where
159: $\bfA=\bfB$ are non singular, we are in the presence of classical
160: diffusive connections, which leads us back to the discussion of
161: section \ref{sec:global}.
162:
163: As in the case of diffusive connections, one can consider networks of
164: so-connected elements, for example :
165: \[
166: \left\{ \begin{array}{l} \dot{\bfx}_1=\bff_1(\bfx_1,t) +
167: \bfA_B^\top(\bfB_A\bfx_2-\bfA_B\bfx_1)
168: + \bfA_C^\top(\bfC_A\bfx_3-\bfA_C\bfx_1)\\
169: \dot{\bfx}_2=\bff_2(\bfx_2,t) +
170: \bfB_C^\top(\bfC_B\bfx_3-\bfB_C\bfx_2)
171: + \bfB_A^\top(\bfA_B\bfx_1-\bfB_A\bfx_2) \\
172: \dot{\bfx}_3=\bff_3(\bfx_2,t) +
173: \bfC_A^\top(\bfA_C\bfx_1-\bfC_A\bfx_3) +
174: \bfC_B^\top(\bfB_C\bfx_2-\bfC_B\bfx_3)
175: \end{array} \right.
176: \]
177: leads to a positive semi-definite Laplacian matrix
178: \[
179: \left(
180: \begin{array}{rrr}
181: \bfA_B^\top\bfA_B & -\bfA_B^\top\bfB_A & \zeros\\
182: -\bfB_A^\top\bfA_B & \bfB_A^\top\bfB_A & \zeros\\
183: \zeros & \zeros & \zeros
184: \end{array}
185: \right)+ \left(
186: \begin{array}{rrr}
187: \zeros & \zeros & \zeros\\
188: \zeros & \bfB_C^\top\bfB_C & -\bfB_C^\top\bfC_B \\
189: \zeros & -\bfC_B^\top\bfB_C & \bfC_B^\top\bfC_B \\
190: \end{array}
191: \right)+ \left(
192: \begin{array}{rrr}
193: \bfA_C^\top\bfA_C & \zeros & -\bfA_C^\top\bfC_A\\
194: \zeros & \zeros & \zeros \\
195: -\bfC_A^\top\bfA_C & \zeros & \bfC_A^\top\bfC_A
196: \end{array}
197: \right)
198: \]
199: and potentially a flow-invariant subspace
200: \[
201: \sM=\{\bfA_B\bfx_1=\bfB_A\bfx_2\} \cap \{\bfB_C\bfx_2=\bfC_B\bfx_3\}
202: \cap \{\bfC_A\bfx_3=\bfA_C\bfx_1\}
203: \]
204:
205: The above coupling structures can be implemented in nonlinear versions
206: of the predictive hierarchies used in image processing
207: (e.g.~\cite{LueWil,DayHin,RaoBal,Korner,GeoHaw,Rao}).
208: %They allow information flow between layers to be minimized while
209: %maintining predictive capabilities in case of temporary interruption
210: %of input ${\bf I}(t)$.
211:
212: \subsubsection{Excitatory-only networks}
213:
214: \label{sec:excit}
215:
216: One can also address the case of networks with excitatory-only
217: connections. Consider for instance the following system and its
218: Jacobian matrix\,\footnote{For the sake of clarity, the elements are
219: assumed to be 1-dimensional. However, the same reasoning applies for
220: the multidimensional case as well: instead of $\mathrm{span}{\tiny
221: \left\{ \left(\begin{array}{c}
222: 1\\
223: 1
224: \end{array}\right)
225: \right\}}$, one considers
226: $\mathrm{span}\{\bfeg_1,\dots,\bfeg_d\}$ as in section
227: \ref{sec:global}.}
228: \[
229: \left\{\begin{array}{l}
230: \dot{x}_1=f(x_1,t)+k x_2 \\
231: \dot{x}_2=f(x_2,t)+k x_1
232: \end{array}\right.
233: \qquad \bfJ=\left(\begin{array}{cc}
234: \frac{\partial{f}}{\partial{x}}(x_1,t)&0\\
235: 0&\frac{\partial{f}}{\partial{x}}(x_2,t)
236: \end{array}\right) +
237: k\left(\begin{array}{cc}
238: 0&1\\
239: 1&0
240: \end{array}\right)
241: \]
242:
243: Clearly, $\mathrm{span}\{(1,1)\}$ is flow-invariant. Applying the
244: methodology described above, we choose
245: $\bfV=\frac{1}{\sqrt{2}}(1,-1)$, so that the projected Jacobian matrix
246: is $\frac{1}{2}\left(\frac{\partial{f}}{\partial{x}}(x_1,t)+
247: \frac{\partial{f}}{\partial{x}}(x_2,t)\right)-k$. Thus, for $k >
248: \sup_{a,t}\frac{\partial{f}}{\partial{x}}(a,t)$, the two elements
249: synchronize exponentially.
250:
251: In the case of diffusive connections, once the elements are
252: synchronized, the coupling terms disappear, so that each individual
253: element exhibits its natural, uncoupled behavior. This is not the case
254: with excitatory-only connections. This is illustrated in figure
255: \ref{fig:fn} using FitzHugh-Nagumo oscillator models (see appendix
256: \ref{sec:oscillators} for the contraction analysis of coupled
257: FitzHugh-Nagumo oscillators).
258:
259: \begin{figure}[ht]
260: \begin{minipage}[ht]{0.32\textwidth}
261: \centering
262: \includegraphics[scale=0.35]{fig/fn-single.eps}
263: \end{minipage}
264: \begin{minipage}[ht]{0.32\textwidth}
265: \centering
266: \includegraphics[scale=0.35]{fig/fn-diff.eps}
267: \end{minipage}
268: \begin{minipage}[ht]{0.32\textwidth}
269: \centering
270: \includegraphics[scale=0.35]{fig/fn-excit.eps}
271: \end{minipage}
272: \caption[]{From left to right : a single oscillator, two
273: oscillators coupled through diffusive connections, two oscillators
274: coupled through excitatory-only connections.}
275: \label{fig:fn}
276: \end{figure}
277:
278: \subsubsection{Rate models for neuronal populations}
279:
280: \label{sec:rate}
281:
282: In computational neuroscience, one often uses the following simplified
283: equations to model the dynamics of neuronal populations
284: \[
285: \tau \dot{\bfx}_i= -\bfx_i+ \Phi\left(\sum_{j\neq
286: i}k_{ij}\bfx_j(t)\right)+ \bfu_i(t)
287: \]
288:
289: Assume that the external inputs $\bfu_i(t)$ are all equal, and that
290: the synaptic connections $k_{ij}$ verify $ \exists c, \forall i,\
291: \sum_{j\neq i}k_{ij} = c $ (i.e., that they induce input-equivalence, see
292: section~\ref{sec:syminv}). Then the synchronization subspace
293: \mbox{$\{\bfx_1=\dots=\bfx_n\}$} is flow-invariant. Furthermore, since
294: each element, taken in isolation, is \emph{contracting} with
295: contraction rate $1/\tau $, synchronization should occur when the
296: coupling is not too strong (see remark (ii) in
297: section~\ref{sec:global}).
298:
299: Specifically, consider first the case where $\Phi$ is a linear
300: function : $\Phi(\bfx)=\mu\bfx$. The Jacobian matrix of the global
301: system is then $-\bfI_n+\mu\bfK$, where $\bfK$ is the matrix of
302: $k_{ij}$. Using the result of remark (ii) in section
303: \ref{sec:global}, a sufficient condition for the system to be
304: contracting (and thus synchronizing) is that the couplings are
305: \emph{weak} enough (or more precisely, such that
306: $\mu\lambda_\mathrm{max}(\bfK) < 1$).
307:
308: The same condition is obtained if $\Phi$ is now e.g. a
309: multidimentional \emph{sigmoid} of maximum slope $\mu$ (see remark
310: (iii) in section \ref{sec:global}).
311:
312: Besides the synchronization behavior of these models, their natural
313: \emph{contraction} property for weak enough couplings of any sign is
314: interesting in its own right. Indeed, given a set of (not necessarily
315: equal) external inputs $\bfu_i(t)$, all trajectories of the global
316: system will converge to a unique trajectory, independently of initial
317: conditions.
318:
319: \subsection{Symmetries, diffusion-like couplings, flow-invariant
320: subspaces and concurrent synchronization}
321:
322: \label{sec:syminv}
323:
324: Synchronized states can be created in at least two ways : by
325: architectural and internal\footnote{Internal symmetries can easily
326: be analyzed within our framework as leading to flow-invariant
327: subspaces, and we shall use this property in section \ref{sec:cpg} for
328: building central pattern generators. However, they will not be
329: discussed in detail in this article. The interested reader can consult
330: \cite{DioGolSte}.} symmetries
331: \cite{GolSte,GolSteTor,DioGolSte,PogSanNij} or by diffusion-like
332: couplings \cite{WangSlo,JadMotBar,OlfMur,LinBroFra,Belykh1}. Actually,
333: we shall see that both, together or separately, can create
334: flow-invariant subspaces corresponding to concurrently synchronized
335: states.
336:
337: \subsubsection{Symmetries and input-equivalence}
338:
339: In section \ref{sec:global}, we argued that, in the case of coupled
340: \emph{identical} elements, the global synchronization subspace $\sM$
341: represents a flow-invariant linear subspace of the global state space.
342: However, several previous works have pointed out that larger (less
343: restrictive) flow-invariant subspaces may exist if the network
344: exhibits symmetries~\cite{Zhang,Belykh1,PogSanNij}, even when the
345: systems are {\it not} identical~\cite{GolSte}.
346:
347: The main idea behind these works can be summarized as follows. Assume
348: that the network is divided into $k$ aspiring synchronized groups
349: $S_1,\dots,S_k$\,\footnote{Some groups may contain a single element,
350: see section \ref{sec:single}.}. The flow-invariant subspace
351: corresponding to this regime (in the sequel, we shall call such a
352: subspace a \emph{concurrent synchronization subspace}), namely
353: \[
354: \{(\bfx_1;\dots;\bfx_n) : \forall 1\leq m \leq k,
355: \forall i,j \in S_m : \bfx_i=\bfx_j\}
356: \]
357: is flow-invariant if, for each $S_m$, the following conditions are
358: true :
359:
360: \begin{enumerate}
361: \item if $i,j\in S_m$, then they have a same individual (uncoupled)
362: dynamics
363: \item if $i,j\in S_m$, and if they receive their input from elements
364: $i'$ and $j'$ respectively, then $i'$ and $j'$ must be in a same
365: group $S_{m'}$, and the coupling functions (the synapses)
366: $i'\rightarrow i$ and $j'\rightarrow j$ must be identical. If $i$
367: and $j$ have more than one input, they must have the same number of
368: inputs, and the above conditions must be true for each input. In
369: this case, we say that $i$ and $j$ are input-symmetric, or more
370: precisely, {\it input-equivalent} (since formally ``symmetry'' implies
371: the action of a group).
372: \end{enumerate}
373:
374: One can see here that symmetry, or more generally
375: input-equivalence, plays a key role in concurrent
376: synchronization. For a more detailed discussion, the reader is
377: referred to \cite{GolSte,GolSteTor}.
378:
379: \textbf{Remark :} One can thus turn on/off a specific symmetry by
380: turning on/off a single connection. This has similarities to the fact
381: that a single inhibitory connection can turn on/off an entire network
382: of synchronized identical oscillators~\cite{WangSlo}.
383:
384:
385: \subsubsection{Diffusion-like couplings}
386:
387: The condition of input-equivalence can be relaxed when some
388: connections \emph{within a group} are null when the connected elements
389: are in the same state. Such connections are pervasive in the
390: literature : diffusive connections (in a neuronal context, they
391: correspond to electrical synapses mediated by gap junctions
392: \cite{SheRin,fukuda06}, in an automatic control context, they
393: correspond to poursuit or velocity matching strategies
394: \cite{OlfMur,LinBroFra}, \dots), connections in the Kuramoto model
395: \cite{IzhiKura,JadMotBar,Strogatz} (i.e. in the form
396: $\dot{x_i}=f(x_i,t)+\sum_{j}k_{ij}\sin(x_j-x_i)$), etc.
397:
398: Indeed, consider for instance diffusive connections and assume that
399: \begin{itemize}
400: \item $i,i',j,j' \in S_m$
401: \item $i'\rightarrow i$ has the form $\bfK_1(\bfx_{i'}-\bfx_i)$
402: \item $j'\rightarrow j$ has the form $\bfK_2(\bfx_{j'}-\bfx_j)$ with
403: possibly $\bfK_1\neq \bfK_2$
404: \end{itemize}
405: Here, $i$ and $j$ are not input-equivalent in the sense we defined
406: above, but the subspace $\{\bfx_i=\bfx_j=\bfx_{i'}=\bfx_{j'}\}$ is
407: still flow-invariant. Indeed, once the system is on this
408: synchronization subspace, we have $\bfx_i=\bfx_{i'}$,
409: $\bfx_j=\bfx_{j'}$, so that the diffusive couplings $i'\rightarrow i$
410: and $j'\rightarrow j$ vanish.
411:
412: One can also view the network as a directed graph $G$, where the
413: elements are represented by nodes, and connections $i\to j$ by
414: directed arcs $i\to j$. Then, the above remark can be reformulated as
415:
416: \begin{description}
417: \item[1 :] for all $m$, color the nodes of $S_m$ with a color $m$,
418: \item[2 :] for all $m$, erase the arcs representing diffusion-like
419: connections and joining two nodes in $S_m$,
420: \item[3 :] check whether the initial coloring is balanced (in the
421: sense of \cite{GolSte}) with respect to the so-obtained graph.
422: \end{description}
423:
424: It should be clear by now that our framework is particularly suited to
425: analyze concurrent synchronization. Indeed, a general methodology to
426: show global exponential convergence to a concurrent synchronization
427: regime consists in the following two steps
428:
429: \begin{itemize}
430: \item First, find an flow-invariant linear subspace by taking advantage of
431: potential symmetries in the network and/or diffusion-like connections.
432: \item Second, compute the projected Jacobian matrix on the orthogonal
433: subspace and show that it is uniformly negative definite (by
434: explicitly computing its eigenvalues or by using results regarding
435: the form of the network, e.g. remark (i) in section \ref{sec:global}
436: or section \ref{sec:analysis}).
437: \end{itemize}
438:
439: \subsection{Illustrative examples}
440:
441: %Let us illustrate our statements with the examples in figure
442: %\ref{fig:examples}.
443: \begin{figure}[ht]
444: \begin{minipage}[ht]{0.31\textwidth}
445: \centering
446: \includegraphics[scale=0.4]{fig/square-a.eps}
447: \end{minipage}
448: \begin{minipage}[ht]{0.31\textwidth}
449: \centering
450: \includegraphics[scale=0.4]{fig/square-b.eps}
451: \end{minipage}
452: \begin{minipage}[ht]{0.33\textwidth}
453: \centering
454: \includegraphics[scale=0.45]{fig/toy.eps}
455: \end{minipage}
456: \caption[]{Three example networks}
457: \label{fig:examples}
458: \end{figure}
459:
460: \begin{enumerate}
461: \item The first network has three non-trivial flow-invariant subspaces
462: other than the global sync subspace, namely $\sM_1=\{\bfx_1=\bfx_2,
463: \bfx_3=\bfx_4\}$, $\sM_2=\{\bfx_1=\bfx_3, \bfx_2=\bfx_4\}$, and
464: $\sM_3=\{\bfx_1=\bfx_4, \bfx_2=\bfx_3\}$. Any of these subspaces is
465: a strict superset of the global sync subspace, and therefore one
466: should expect that the convergence to any of the concurrent sync
467: state is ``easier'' than the convergence to the global sync
468: state~\cite{Zhang,Belykh1,PogSanNij}. This can be quantified from
469: (\ref{equ:synccond}), by noticing that
470: \begin{equation} \label{percolation}
471: \sM_A \supset \sM_B
472: \Rightarrow\sM_A^\perp \subset \sM_B^\perp \ \Rightarrow
473: \lambda_\mathrm{min}(\bfV_A\bfL\bfV_A^\top) \ge
474: \lambda_\mathrm{min}(\bfV_B\bfL\bfV_B^\top) \ \ \
475: \end{equation}
476: While in the case of identical systems and relatively uniform
477: topologies, this ``percolation'' effect may often be too fast to
478: observe, (\ref{percolation}) applies to the general concurrent
479: synchronization case and quantifies the associated and possibly very
480: distinct time-scales.
481: \item The second network has only one non-trivial flow-invariant
482: subspace $\{\bfx_1=\bfx_2, \bfx_3=\bfx_4\}$.
483: \item If the dashed blue arrows represent diffusive connections then
484: the third network will have one non-trivial flow-invariant subspace
485: $\{\bfx_2 =\bfx_3=\bfx_4, \bfx_5=\bfx_6=\bfx_7\}$, even if these
486: extra diffusive connections \emph{obviously break the symmetry}.
487: \end{enumerate}
488:
489:
490: Let's study in more detail this third network, in which the
491: connections between the round element and the square ones are modelled
492: by trigonometric functions (we shall see in section \ref{sec:single}
493: that their exact form has no actual influence on the convergence
494: rate).
495: \[
496: \left\{ \begin{array}{l}
497: \dot{v}_1=f(v_1)+a_1\cos(v_2)+a_2\sin(v_3)\\
498: \dot{v}_2=g(v_2)+a_4\sin(v_1)+c_1v_6\\
499: \dot{v}_3=g(v_3)+a_4\sin(v_1)+b_1(v_2-v_3)+b_2(v_4-v_3)+c_1v_5\\
500: \dot{v}_4=g(v_4)+a_4\sin(v_1)+b_3(v_3-v_4)+c_1v_7 \\
501: \dot{v}_5=h(v_5)+c_2v_2+(d_2v_7-d_1v_5)\\
502: \dot{v}_6=h(v_6)+c_2v_3+(d_2v_5-d_1v_6)\\
503: \dot{v}_7=h(v_7)+c_2v_4+(d_2v_6-d_1v_7)\\
504: \end{array} \right. \]
505: The Jacobian matrix of the couplings is
506: \[
507: \bfL=\left( \begin{array}{rrrrrrr}
508: 0&a_1\dot{v}_2\sin(v_2)&-a_2\dot{v}_3\cos(v_3)&0&0&0&0 \\
509: -a_4\dot{v}_1\cos(v_1)&0&0&0&0&-c_1&0 \\
510: -a_4\dot{v}_1\cos(v_1)&-b_1&b_1+b_2&-b_2&-c_1&0&0 \\
511: -a_4\dot{v}_1\cos(v_1)&0&-b_3&b_3&0&0&-c_1 \\
512: 0&-c_2&0&0&d_1&0&-d_2 \\
513: 0&0&-c_2&0&-d_2&d_1&0 \\
514: 0&0&0&-c_2&0&-d_2&d_1 \end{array}\right)
515: \]
516: As we remarked previously, the concurrent synchronization regime
517: $\{v_2=v_3=v_4,v_5=v_6=v_7\}$ is possible. Bases of the linear
518: subspaces $\sM$ and $\sM^\perp$ corresponding to this regime are \[
519: \left(\begin{array}{r} 1\\0\\0\\0\\0\\0\\0 \end{array} \right),
520: \left(\begin{array}{r} 0\\1\\1\\1\\0\\0\\0 \end{array} \right),
521: \left(\begin{array}{r} 0\\0\\0\\0\\1\\1\\1 \end{array} \right) \quad
522: \textrm{for } \sM,\textrm{ and}\quad \left(\begin{array}{r}
523: 0\\\frac{\sqrt6}{3}\\\frac{-\sqrt6}{6}\\\frac{-\sqrt6}{6}\\0\\0\\0
524: \end{array} \right), \left(\begin{array}{r}
525: 0\\0\\\frac{-\sqrt2}{2}\\\frac{\sqrt2}{2}\\0\\0\\0 \end{array}
526: \right), \left(\begin{array}{r}
527: 0\\0\\0\\0\\\frac{\sqrt6}{3}\\\frac{-\sqrt6}{6}\\\frac{-\sqrt6}{6}
528: \end{array} \right), \left(\begin{array}{r}
529: 0\\0\\0\\0\\0\\\frac{-\sqrt2}{2}\\\frac{\sqrt2}{2} \end{array}
530: \right) \quad \textrm{for }\sM^\perp. \] Group together the vectors
531: of the basis of $\sM^\perp$ into a matrix $\bfV$ and compute
532: \[
533: \bfV\bfL_s\bfV^\top=\left(\begin{array}{rrrr}
534: \frac{b_1}{2}&-\frac{\sqrt{3}(2b_1+b_2-b_3)}{6}&\frac{c_1-2c_2}{4}&-\frac{c_1\sqrt{3}}{4}\\
535: -\frac{\sqrt{3}(2b_1+b_2-b_3)}{6}&\frac{b_1+2(b_2+b_3)}{2}&-\frac{c_1\sqrt{3}}{4}&-\frac{c_1+2c_2}{4}\\
536: \frac{c_1-2c_2}{4}&-\frac{c_1\sqrt{3}}{4}&\frac{2d_1+d_2}{2}&0\\
537: -\frac{c_1\sqrt{3}}{4}&-\frac{c_1+2c_2}{4}&0&\frac{2d_1+d_2}{2}
538: \end{array}\right) \] As a numerical example, let $b_1=3\alpha$,
539: $b_2=4\alpha$, $b_3=5\alpha$, $c_1=\alpha$, $c_2=2\alpha$,
540: $d_1=3\alpha$, $d_2=4\alpha$ and evaluate the eigenvalues of
541: $\bfV\bfL_s\bfV^\top$. We obtain approximately $1.0077\alpha$ for
542: the smallest eigenvalue. Using again FitzHugh-Nagumo oscillators
543: and based on their contraction analysis in appendix
544: \ref{sec:oscillators}, concurrent synchronization should occur for
545: $\alpha>10.25$. A simulation is shown in figure \ref{fig:exfig}.
546: One can see clearly that, after a transient period, oscillators 2,
547: 3, 4 are in perfect sync, as well as oscillators 5, 6, 7, but that
548: the two groups are not in sync with each other.
549:
550: \begin{figure}[ht]
551: \centering
552: \includegraphics[scale=0.5]{fig/fn-ex.eps}
553: \caption{Simulation result for network 3.}
554: \label{fig:exfig}
555: \end{figure}
556:
557: \subsection{Robustness of synchronization}
558:
559: So far, we have been considering \emph{exact} synchronization of
560: \emph{identical} elements. However this assumption may seem
561: unrealistic, since real systems are never absolutely identical. We
562: use here the robustness result for contracting systems (see
563: theorem~\ref{theorem:robust}) to guarantee approximate synchronization
564: even when the elements are not identical.
565:
566: Consider, as in section \ref{sec:global}, a network of
567: $n$ dynamical elements
568: \begin{equation}
569: \label{equ:robustsync}
570: \dot{\bfx}_i = \bff_i(\bfx_i,t) + \sum_{j \neq i} \bfK_{ij}(\bfx_j -
571: \bfx_i) \qquad i=1,\ldots,n
572: \end{equation}
573: with now possibly $\bff_i\neq\bff_j$ for $i\neq j$. This can be
574: rewritten as
575: \begin{equation}
576: \label{equ:robust}
577: \left(\begin{array}{c}
578: \dot{\bfx}_1\\
579: \vdots\\
580: \dot{\bfx}_n
581: \end{array}\right) =
582: \left(\begin{array}{c}
583: \bfc(\bfx_1,t)\\
584: \vdots\\
585: \bfc(\bfx_n,t)
586: \end{array}\right) -
587: \bfL
588: \left(\begin{array}{c}
589: \bfx_1\\
590: \vdots\\
591: \bfx_n
592: \end{array}\right)+
593: \left(\begin{array}{c}
594: \bff_1(\bfx_1,t)-\bfc(\bfx_1,t)\\
595: \vdots\\
596: \bff_n(\bfx_n,t)-\bfc(\bfx_n,t)
597: \end{array}\right)
598: \end{equation}
599: where $\bfc$ is some function to be defined later. Keeping the
600: notations introduced in section \ref{sec:global}, one has
601: \[
602: \dot{\bfxg}=\bfcg(\bfxg,t)-\bfL\bfxg+\bfd(\bfxg,t)
603: \]
604: where $\bfd(\bfxg,t)$ stands for the last term of equation
605: (\ref{equ:robust}).
606:
607: Consider now the projected auxiliary system on $\sM^\perp$
608: \begin{equation}
609: \label{equ:projecteddisturbed}
610: \dot{\bfy}=\bfV\bfcg(\bfV^\top\bfy+\bfU\bfU^\top\bfxg,t)-
611: \bfV\bfL\bfV^\top\bfy+\bfV\bfd(\bfV^\top\bfy+\bfU\bfU^\top\bfxg,t)
612: \end{equation}
613:
614: %We show in
615: %section \ref{sec:robustcontraction} that, for strong enough coupling
616: %strength, all trajectories of the system will exponentially converge
617: %to a boundary layer of thickness $D/\lambda$ around the sync subspace
618: %$\sM$, where $\lambda$ is the contraction rate of the auxiliary system
619: %and $D$ is a measure of the dissimilarity of the elements (see
620: %equation (\ref{equ:D}).
621:
622: Assume that the connections represented by $\bfL$ are strong enough
623: (in the sense of equation (\ref{equ:synccond})), so that the
624: undisturbed version of (\ref{equ:projecteddisturbed}) is contracting
625: with rate $\lambda>0$. Let $D=\sup_{\bfxg,t}\|\bfV\bfd(\bfxg,t)\| $,
626: where $D$ can be viewed as a measure of the dissimilarity of the
627: elements. Since $\bfy=\zeros$ is a particular solution of the
628: undisturbed system, theorem \ref{theorem:robust} implies that the
629: distance $R(t)$ between any trajectory of
630: (\ref{equ:projecteddisturbed}) and $\zeros$ verifies, after a
631: transient period, $R(t)\leq D/\lambda$. In the $\bfx$-space, it means
632: that \emph{any trajectory will eventually be contained in a boundary
633: layer of thickness $D/\lambda$ around the synchronization subspace
634: $\sM$}.
635:
636: The choice of $\bfc$ can now be specified so as to minimize
637: $D/\lambda$. Neglecting for simplicity the variation of $\lambda$, a
638: possible choice for $\bfc(\bfx,t)$ is then the center of the ball of
639: smallest radius containing $\bff_1(\bfx,t),\dots,\bff_n(\bfx,t)$, with
640: $D$ being the radius of that ball.
641:
642: Consider for instance, the following system (similar to the model used
643: for coincidence detection in \cite{WangSlo} and section
644: \ref{sec:manies})
645: \[
646: \dot{x}_i=f(x_i)+ I_i + k(x_0-x_i) \qquad {\rm where}\ I_\mathrm{min}
647: \le I_i \le I_\mathrm{max},\ \forall i
648: \]
649: In this case, choosing
650: $c(x)=f(x)+\frac{I_\mathrm{max}+I_\mathrm{min}}{2}$, one can achieve
651: the bound $D/\lambda$, where $\lambda$ is the contraction rate of $f$
652: and $D=\frac{I_\mathrm{max}-I_\mathrm{min}}{2}$.
653:
654: \textbf{Remark :} Assume that two spiking neurons are approximately
655: synchronized, as just discussed. Then, since spiking induces large
656: abrupt variations, the neurons must spike approximately at the same
657: time. More specifically, if the bound on their trajectory discrepancy
658: guaranteed by the above robustness result is significantly smaller
659: than spike size, then this bound will automatically imply that the two
660: neurons spike approximately at the same time.
661:
662: %%% Local Variables:
663: %%% mode: latex
664: %%% TeX-master: "neuro"
665: %%% End:
666: