q-bio0701052/rings.tex
1: \documentclass[aps,pre,twocolumn,showpacs,floatfix,superscriptaddress]{revtex4}
2: %\documentclass[aps,twocolumn]{revtex4}
3: %\usepackage{epsfig}
4: %\usepackage{pxfonts}
5: %\documentclass[aps,pre,showpacs,twocolumn,superscriptaddress]{revtex4}
6: %\usepackage[draft]{graphicx}
7: \usepackage{graphicx}
8: \usepackage{amsmath}
9: %\usepackage{amssymb}
10: \usepackage{epsfig}
11: %\bibstyle{apsrev.bib}
12: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
13: % this is how one defines new commands.
14: \newcommand{\on}{{\sc on}}
15: \newcommand{\off}{{\sc off}}
16: \newcommand{\bs}{\boldsymbol}
17: 
18: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
19: 
20: \begin{document}
21: 
22: \title{Attractors in continuous and Boolean networks} 
23: 
24: \author{Johannes Norrell}
25: \author{Bj{\"o}rn Samuelsson}
26: \author{Joshua E.\ S.\ Socolar}
27: \affiliation{
28: Physics Department and Center for Nonlinear and Complex Systems,
29: Duke University, Durham, North Carolina 27708}
30: 
31: \date{\today}
32: 
33: \begin{abstract}
34:   We study the stable attractors of a class of continuous dynamical
35:   systems that may be idealized as networks of Boolean elements, with
36:   the goal of determining which Boolean attractors, if any, are good
37:   approximations of the attractors of generic continuous systems.  We
38:   investigate the dynamics in simple rings and rings with one
39:   additional self-input.  An analysis of switching characteristics and
40:   pulse propagation explains the relation between attractors of the
41:   continuous systems and their Boolean approximations.  For simple rings,
42:   ``reliable'' Boolean attractors correspond to stable continuous
43:   attractors.  For networks with more complex logic, the qualitative
44:   features of continuous attractors are influenced by inherently
45:   non-Boolean characteristics of switching events.
46: \end{abstract}
47: 
48: \pacs{89.75.Hc, 05.45.--a, 02.30.Ks}
49: \maketitle
50: 
51: \section{Introduction}
52: Complex networks of interacting elements arise in many biological,
53: chemical, sociological, and physical contexts.  An important example
54: is the network of interactions among proteins and DNA in a cell.  The
55: binding of certain proteins to each other and to promoter regions of
56: DNA can create a combinatorially complex logic of gene expression.  It
57: is tempting to think of transcriptional networks and similar examples
58: as effectively Boolean in nature.  In such a picture, genes are turned
59: ``on'' and ``off'' in the presence of proteins produced when other
60: genes are turned on or off \cite{Davidson:06}, and many studies of the
61: fundamental principles governing networks of interacting genes have
62: focused Boolean models.  Recent work has highlighted distinctions in
63: the attractor structures as network architecture parameters are
64: varied, different distributions of Boolean functions are incorporated,
65: and/or different updating schemes are employed \cite{Aldana:03a,
66:   Aldana:03b, Greil:05, Moreira:05, Kesseli:06, Paul:06,
67:   Samuelsson:06}.
68: 
69: The Boolean models are generally understood to be idealized
70: representations of underlying continuous and perhaps stochastic
71: processes, so it is important to understand any artifacts introduced
72: in Boolean approximations.  Here we investigate some continuous,
73: deterministic systems motivated by models of transcriptional
74: interactions and designed to be good candidates for a Boolean
75: analysis.  The goal is to elucidate the most important effects that
76: may lead the continuous dynamics to differ qualitatively from
77: expectations based on the Boolean models.
78: 
79: We investigate the temporal structure and stability of the attractors
80: of continuous dynamical systems and those of the corresponding Boolean
81: models.  Two effects are found to be crucial for understanding the
82: structure of the continuous attractors: first, when an \on-\off\
83: symmetry of typical Boolean models is broken, the possibility of
84: stable pulse propagation down a chain can be lost; and second, the
85: memory of past inputs at a given node causes shifts in the temporal
86: spacing between multiple pulses on a feedback loop.  These effects
87: make it difficult to put the attractors of the continuous network into
88: correspondence with the attractors of their Boolean counterparts.
89: 
90: We study rings of $N$ elements (Fig.~\ref{fig:rings}) governed by
91: delay differential equations.  The delays are introduced to represent
92: intermediate steps in the process mediating the interactions between
93: elements.  We employ a form developed originally as a mean-field
94: description of the dynamics of transcription factor expression
95: \cite{Andrecut:06}, though for present purposes it simply provides
96: a generic model of elements with sigmoidal responses to their inputs:
97: \begin{align}
98: \label{eqn:continuous}
99:   {\dot x}_j(t) & = f_j\bs(x_{j-1}(t-\tau)\bs) - x_j(t)\,, \\
100: \label{eqn:f1in}
101:   f_j(x_i) & = 
102:     \eta_j\biggl(\frac{1+d_i^jx_i^{\nu}}{1+b_i^j x_i^{\nu}}\biggr)\,,
103: \end{align}
104: where $\eta_j$, $b_i^j$, and $d_i^j$ are constants and $\tau$ is a
105: time required for the signal produced by $x_{j-1}$ to reach $x_j$.
106: All subscripts are taken modulo $N$.  The parameters are chosen such
107: that the output of $f_j$ switches sharply between low and high values
108: as the input is smoothly varied.  We assume that all $x_j$ have the
109: same decay rate (chosen to be unity) and all arguments of $f_j$ have
110: the same time delay.  We also consider the effects of adding a
111: self-input to $x_1$, also with delay $\tau$:
112: \begin{align}
113: \label{eqn:selfinput}
114:   {\dot x}_1(t) & = f_1\bs(x_{0}(t-\tau),x_{1}(t-\tau)\bs) - x_1(t)\,, \\
115: \label{eqn:f2in}
116:   f_1(x_0,x_1) & = \eta_1\biggl(
117:       \frac{1+d_0^1 x_0^{\nu}+d_1^1 x_1^{\nu}+ d_{01}^1(x_0 x_1)^{\nu/2}}{\displaystyle
118:        1+b_0^1 x_0^{\nu}+b_1^1x_1^{\nu}+ b_{01}^1(x_0 x_1)^{\nu/2}}
119: \biggr)\,.
120: \end{align}
121: A Boolean idealization is obtained in the limit in which $f_j$ is a
122: step function and the decay term $-x_j(t)$ has an infinite
123: coefficient.
124: \begin{figure}[b]
125:   \includegraphics*[width=0.7\columnwidth]{rings.eps}
126: \caption{\label{fig:rings} 
127:   A simple ring and a self-input ring.  {\bf\textsf C} indicates that
128:   the element copies its input; {\bf\textsf F} indicates that the
129:   element is either a copier or an inverter on the simple ring, or
130:   that it performs one of the two-input logic functions on the
131:   self-input element.}
132: \end{figure}
133: 
134: We find that continuous systems can exhibit stable oscillations in
135: cases where Boolean reasoning would suggest otherwise and that in some
136: cases where Boolean reasoning predicts a stable attractor, the
137: corresponding continuous attractor does not have the expected
138: structure.
139: 
140: \section{Boolean systems}
141: We begin by summarizing the known behavior of Boolean systems where
142: each $x_j$ is taken to be a Boolean variable and its dependence on its
143: inputs is specified by a Boolean function $F_j$.  A common choice is
144: to update all the elements synchronously, setting each $x_j$ at each
145: discrete time step to the value $F_j$ returned just after the previous
146: step.  Synchronously updated systems are easy to simulate, but are not
147: generic.  To avoid artifacts of synchronicity, Klemm and Bornholdt
148: \cite{Klemm:05a} present a class of autonomous Boolean systems running
149: in continuous time.  Here the output of each $F_j$ is fed through a
150: filter that delays the signal by a fixed time (analogous to our
151: $\tau$, which they set equal to 1) and cuts out short pulses.  (See
152: Appendix~\ref{app:autonomous} for details.)  These autonomous
153: networks have state cycles that correspond to the attractors of a
154: synchronously updated Boolean network.  In contrast to synchronous
155: Boolean networks, however, autonomous networks permit the study of
156: infinitesimal fluctuations in the timing of switching events.  One can
157: externally impose a delay in one switching event and see whether the
158: sequence of time intervals between switching events is restored by the
159: dynamics.  If all possible small time delays evolve back to the same
160: sequence of switching times, the state cycle of the autonomous system
161: is stable.  If, on the other hand, a subset of the switching times in
162: the cycle remain delayed compared to the others for some perturbation,
163: the state cycle is marginally stable.  There are no unstable state
164: cycles, because there is no way for an infinitesimal perturbation to
165: get amplified.  As a consequence, autonomous networks have an infinite
166: set of marginally stable cycles.
167: 
168: Klemm and Bornholdt coined the terms ``reliable'' and ``unreliable''
169: to denote attractors in the {\em synchronous} system that correspond
170: to stable and marginal cycles, respectively, in the corresponding
171: autonomous system \cite{Klemm:05a}. In accordance with their
172: convention, we use the term ``attractor'' for any periodic state-cycle
173: in a synchronously updated network.  Unreliable attractors are not
174: expected to be observed in real systems because errors in timing can
175: accumulate and eventually cause a transition to a different attractor.
176: One of our goals is to determine whether the set of reliable
177: attractors is in one-to-one correspondence with the set of attractors
178: of a continuous system.
179: 
180: The dynamics of simple Boolean rings have been well characterized
181: \cite{Flyvbjerg:88, Klemm:05a, Greil:05}.  In a simple ring, each
182: element either copies or inverts the state of its input.  A ring with
183: an even (odd) number of inverters is dynamically equivalent to a ring
184: with zero (exactly one) inverters since a pair of inverters can be
185: transformed to copiers by redefining the meaning of {\sc on} and {\sc
186:   off} for all elements between them.  For a ring with no inverters,
187: there are two fixed points; the states with all elements \on\ or all
188: elements \off.  For a ring with one inverter there is no fixed point
189: because at all times there must be at least one element whose value is
190: not consistent with its input.  We refer to this local inconsistency
191: as a ``kink'' and we identify the kink as ``positive'' (``negative'')
192: when the element's input is \on\ (\off).  A single kink traveling
193: around the ring forms a stable cycle with the kink changing its sign
194: each time it passes the inverter.  For synchronous updates, the
195: separation between kinks cannot change, so every state lies on an
196: attractor; there are no transients.  For the autonomous system, any
197: cycle with more than one kink is marginally stable because there is no
198: restoring mechanism for a perturbation in the time lag between two
199: kinks.  The multi-kink synchronous attractors are therefore
200: unreliable.
201: 
202: For rings in which element $1$ has a self-input in addition to its
203: input from element $0$, the attractor structure is nontrivial.
204: Without loss of generality, we assume that the remaining elements are
205: all copiers.  There are ten possible choices of the Boolean function
206: $F_1$ at element $1$ for which both inputs are relevant.  These
207: comprise five pairs that are related by an \on--\off\ symmetry.  For
208: two of these pairs, the only attractors are fixed points, so there are
209: only three nontrivial cases; (I) $F_1 =$ {\sc nor}, (II) $F_1 =$ {\sc
210:   xor}, and (III) $F_1 = (x_0$ {\sc and not} $x_1)$.
211: 
212: For the present discussion, we restrict attention to the case $N=4$.
213: In Case I, there is a single attractor that is (surprisingly)
214: unreliable.  In Cases II and III, the all-\off\ state is a fixed
215: point.  Case II also has a reliable cycle containing all 15 other
216: states.  Case III has two cyclic attractors, both unreliable: a
217: 4-state cycle consisting of the state $1000$ and its cyclic
218: permutations; and a 2-state cycle consisting of the states $1010$ and
219: $0101$.
220: 
221: Our first important observation is that the definition of reliability
222: employed by Klemm and Bornholdt assumes the time delay for an element
223: is the same regardless of whether an input is turning \on\ or \off.
224: The breaking of this symmetry (see below) leads to different
225: propagation speeds for positive and negative kinks, which can destroy
226: or stabilize some marginally stable cycles.  For example, the
227: unreliable 4-cycle in Case III is stabilized if positive kinks move
228: faster than negative kinks.  Though the duration of the \on\ pulse
229: increases as it goes around the ring, it is cut back to $\tau$ each
230: time the pulse passes the self-input element.
231: 
232: \section{Analysis of continuous systems}
233: We now turn to the analysis of continuous systems described by
234: Eqs.~\eqref{eqn:continuous}--\eqref{eqn:f2in}.  We choose the Hill
235: coefficient (or cooperativity) $\nu=2$ corresponding to regulation
236: performed by dimers \cite{Andrecut:06}.  We fix $\eta$, $b$, and $d$
237: as follows in order to observe clearly identifiable \off\ and \on\
238: states.  For a single element that receives one input, we set
239: \begin{equation}
240: \label{eqn:singleinputparams}
241:   \begin{array}{llll}
242:   \eta=1, & b=0.001, & d=0.1 & {\textrm{\ for a copier}}; \\
243:   \eta=100, & b=0.1, & d=0 & {\textrm{\ for an inverter}}. 
244:   \end{array}
245: \end{equation}
246: With these choices, there are three fixed points for simple rings with
247: only {\sc copy} elements: the all-\off\ state $x=1.13$; the all-\on\ 
248: state $x=88.9$; and the unstable {\em switching value} $x=10.0$.  For
249: an {\sc invert} element, a steady \off\ input of $1.13$ produces
250: $x=88.7$ (close to the \on\ value); an \on\ input of $88.9$ produces
251: $x=0.13$ (close to the \off\ value); and $x=9.67$ is an unstable fixed
252: point.  For the self-input element, the values listed in
253: Table~\ref{tab:parameters} are chosen to represent the classes of
254: functions whose Boolean idealizations would be Cases I, II, and III
255: above.
256: \begin{table}[b] \label{tab:parameters}
257: \caption{Parameter values for self-input elements.}
258: \begin{tabular}{c|ccccccc}
259: \hline
260:  Case \rule[0pt]{0pt}{11pt} & $\eta$ & $b_0^1$ & $b_1^1$ & $b_{01}^1$ & $d_0^1$ & $d_1^1$ & $d_{01}^1$ \\
261: \hline
262:  I & $100$ & $0.1$ & $0.1$ & $0$ & $0$ & $0$ & $0$ \\
263:  II & 1 & $0.001$ & $0.001$ & $\;0.02\;$ & $\;0.1\;$ & $\;0.1\;$ & $\;0\;$ \\
264:  III & $1$ & $0.001$  & $0$ & $0.02$ & $0.1$ & $0$ & $0$ \\
265: \hline
266: \end{tabular}
267: \end{table}
268: 
269: To study the propagation of positive and negative kinks, we derive the
270: time it takes for a single kink to pass from one element to the next
271: in a chain of single-input elements.  The expression levels are
272: denoted by $x_0(t),x_1(t),x_2(t),\ldots$, where $x_0(t)$ serves as an
273: input and the other elements obey Eq.~\eqref{eqn:continuous}.  
274: For initial conditions, we
275: assume that $x_j(t)$ has the fixed value $\alpha_j$ for all $t\le0$,
276: with $\alpha_j=f_j(\alpha_{j-1})$ for $j=1,2,\ldots$.
277: 
278: If $x_0$ switches values, approaching a constant value $\beta_0$ at long times, each of
279: the $x_j$ will eventually approach a new value $\beta_j$.  For
280: mathematical convenience, we define rescaled quantities ${\hat x}_j(t)
281: \equiv [x_j(t) - \beta_j]/(\alpha_j - \beta_j)$ and ${\hat f}_j(x)
282: \equiv [f_j(x) - \beta_j]/(\alpha_j - \beta_j)$ with the properties
283: ${\hat x}_j(0) = 1$ and ${\hat x}_j(t)\rightarrow 0$ for large $t$.
284: We further define a specific time $s_j$ associated with the switch
285: from $\alpha_j$ to $\beta_j$ by the formula
286: \begin{equation}
287: \label{eqn:switchingdef}
288:   s_j \equiv \int_0^\infty dt\,{\hat x}_j(t).
289: \end{equation}
290: 
291: The formal solution of Eq.~\eqref{eqn:continuous} is
292: \begin{equation}
293: \label{eqn:dynamicsintegral}
294:   x_j(t) = \int_{-\infty}^tdu\,f_j\bs(x_{j-1}(u-\tau)\bs)\,e^{u-t}
295: \end{equation}
296: for all $j\neq 0$.  Substituting into Eq.~\eqref{eqn:switchingdef} and
297: subtracting $s_{j-1}$ from both sides yields an expression for the
298: time delay between the switching of elements $j-1$ and $j$:
299: \begin{equation}
300: \label{eqn:switchingdiff}
301:   s_j - s_{j-1} = \tau + 1
302:     + \int_0^\infty dt\,[{\hat f}_j\bs(x_{j-1}(t)\bs)-{\hat x}_{j-1}(t)].
303: \end{equation}
304: This delay is the sum of the explicit delay $\tau$, an intrinsic delay
305: of unity associated with the unit coefficient of the $-x_j$ term of
306: Eq.~\eqref{eqn:continuous}, and an additional term that depends on the
307: details of the input function ${\hat f}_j$.
308: 
309: For chains of identical copiers with the parameters
310: specified above, it is straightforward to show that the additional
311: delay is positive (negative) for negative (positive) kinks and
312: therefore that positive kinks will propagate faster, which implies
313: that an \on\ pulse will expand in width and an \off\ pulse will
314: shrink and disappear.  By changing the parameters $\eta$, $b$, and
315: $d$, it is possible to reverse this situation, but arranging for
316: precisely equal propagation speeds requires fine tuning.
317: 
318: In the numerical simulation of Fig.~\ref{fig:kinkspeeds} it is clear
319: that the kink moves faster in its positive form than in its negative
320: form.  (To bring out the asymmetry, this figure was made for $\tau =
321: 2$, a relatively short time.  Though the propagation time from one
322: site to the next depends on $\tau$, the difference between propagation
323: times for positive and negative kinks does not.)  The asymmetry is
324: also present in the continuous two-element rings in
325: Ref.~\cite{Klemm:05b} and in the repressilator simulation of
326: Ref.~\cite{Elowitz:00}.  The electronic model with step function
327: switching in Ref.~\cite{Glass:05}, however, does not break the
328: symmetry for the case studied, in which the switching level is halfway
329: between the \on\ and \off\ voltages.
330: \begin{figure}[t]
331:   \includegraphics*[width=\columnwidth]{sep.eps}
332: \caption{\label{fig:kinkspeeds}
333:   Propagation of a single kink around a ring of four elements with one
334:   inverter and $\tau=2$.  The thick, thin, dashed, and grey lines
335:   represent element 1, 2, 3, and 0, respectively.  Note that
336:   propagation is faster when the kink is positive and slower after the
337:   inverter converts it to a negative kink.  }
338: \end{figure}
339: 
340: Until now, we have neglected the fact that the propagation speed of a
341: kink through a given element is influenced by kinks that have
342: previously passed through.  
343: Consider now an \on\ pulse consisting of a positive kink followed by
344: a negative kink.
345: From Eq.~\eqref{eqn:dynamicsintegral}, we
346: see that $x_j(t)$ has an exponentially decaying memory of the events
347: in $x_{j-1}(t)$.  Because $f_j(x)$ is monotonic, the memory of low
348: values of $x_{j-1}(t)$ before the pulse in $x_{j-1}(t)$ will speed
349: up the effect of the trailing edge.  The interaction
350: shortens the pulse duration at $x_j(t)$, and monotonicity of $f_j$
351: ensures that the effect is stronger for shorter pulses.  Similar
352: reasoning shows that an \off\ pulse will also be shorter than the
353: time between kinks one would predict  
354: from Eq.~\eqref{eqn:switchingdiff} alone.
355: 
356: In a chain of copiers, different propagation speeds of
357: positive and negative kinks will lengthen (shorten) a traveling \on\
358: (\off) pulse.  In a chain that contains inverters arranged such that a
359: single pulse spends equal amounts of time in its \on\ and \off\
360: configurations, the asymmetry between positive and negative kinks
361: alone may not change the average pulse duration, but the pulse would
362: still be shortened by the memory effect.
363: 
364: Because the strength of the memory effect increases as two kinks
365: approach each other, a pulse or other sequence of kinks cannot
366: propagate stably on a chain of single-input elements --- only a single
367: kink can have a stable shape as it advances.  For simple rings, then,
368: stable attractors must have only zero or one kink.  These two
369: possibilities correspond precisely to the fixed points and single-kink cycles that
370: are the reliable attractors of the corresponding Boolean systems.
371: 
372: For rings with a self-input (Fig.~\ref{fig:rings}), the situation is
373: more complicated.  First, note that the analysis of pulse propagation
374: gives a quantitative measure of the asymmetry discussed above.  The
375: asymmetry enables some attractors classified as unreliable by the
376: (symmetric) Boolean analysis to be stable in the continuous system.
377: Second, memory effects associated with multiple inputs to a single
378: element can lead to repulsion between pulses and stabilization of new
379: attractors
380: related to synchronous Boolean ones but with shifts in the timing
381: between pulses.  
382: 
383: \section{Numerical results}
384: We studied the three cases numerically using an integration scheme
385: that takes advantage of the structure of
386: Eq.~\eqref{eqn:dynamicsintegral} as described in
387: Appendix~\ref{app:methods}.  In Case I, most initial conditions
388: lead first to a long transient that corresponds to the unreliable
389: attractor observed in the synchronous Boolean network.  The transient
390: eventually gives way to a stable attractor of
391: Fig.~\ref{fig:timeseries}(a), which has two pulses separated in time
392: by approximately $(5/2)(\tau + 1)$, an example of the stabilization of
393: an intermediate inter-pulse interval where the synchronous Boolean
394: attractor has pulses separated in time by alternating intervals of $2$
395: and $3$ time steps.  We note in passing that the path by which the
396: continuous system arrives at the attractor is rather robust: for a
397: wide range of initial conditions, the system goes to the Boolean-like
398: transient in a short time and then gradually shifts to the attractor.
399: In some cases, however, we observe the attractor shown in
400: Fig.~\ref{fig:timeseries}(b), which has three pulses with a time
401: separation of $(5/3)(\tau + 1)$ between each pair.  This attractor
402: corresponds to a marginally stable cycle of the autonomous Boolean
403: network, but the timing is incompatible with the synchronous updating
404: scheme.  
405: \begin{figure}[t]
406:   \includegraphics*[width=\columnwidth]{nor.eps} \\
407:   \includegraphics*[width=\columnwidth]{nor3.eps} \\
408:   \includegraphics*[width=\columnwidth]{xor.eps} \\
409:   \includegraphics*[width=\columnwidth]{singlePulse.eps} \\
410:   \includegraphics*[width=\columnwidth]{doublePulse.eps}
411: \caption{\label{fig:timeseries}
412:   Time series for Cases I (a and b), II (c), and III (d and e), with
413:   $\tau=10$.  The thick, thin, dashed, and grey lines represent
414:   element 1, 2, 3, and 0, respectively.  Case I shows (a)
415:   non-synchronous timing stabilization due to memory effects (with
416:   time axis labels shifted by $10^6$ to account for a long transient)
417:   and (b) a stable attractor that has no analogue in the corresponding
418:   synchronous Boolean
419:   network.  Case II shows (c) an inherently non-Boolean attractor.  Case
420:   III shows (d) single pulse and (e) double pulse attractors, both of
421:   which are unreliable in a Boolean network.}
422: \end{figure}
423: 
424: We note that the appearance of attractor periods having
425: $5(\tau+1)$ as an integer multiple can be understood analytically.
426: Kaufman and Drossel have analyzed the possible attractor periods in
427: synchronous Boolean networks in rings with a single cross-link
428: \cite{Kaufman:05}.  A straightforward extension of their method to the
429: autonomous networks of our ring with a {\sc nor} self-input (Case I)
430: reveals that all cycles in the autonomous model must have a period of
431: the form $5/m$, where $m$ is an integer and the time unit is the time
432: required for a kink to advance through one element.  The corresponding
433: period in the continuous model is therefore $5(\tau +1)/m$.
434: 
435: In Case II, two attractors are observed: the fixed point and an
436: oscillatory attractor that appears to be aperiodic.  The latter
437: behavior is sensitive to intermediate variable values and is thus
438: inherently non-Boolean.  In Case III, we observe one fixed point and
439: two limit cycles, corresponding to each of the three Boolean
440: attractors, with no long transients.  The limit cycles are {\em
441:   unreliable} in the Boolean case but {\em stable} in the continuous
442: case --- examples of stabilization due to asymmetry in kink
443: propagation speeds and a pulse-shortening self-input element.  
444: This is the effect alluded to above at the end of the 
445: Boolean systems section.  Though the pulse has broadened in traveling
446: around the ring, a new trailing edge is generated by the self-input,
447: which occurs one delay time after the arrival of the leading edge.
448: 
449: The
450: difference in pulse height in the two limit cycles is due to the
451: dependence of $f_1(x_0, x_1)$ on small variations in $x_1$ when $x_0$
452: is nominally \on\ and $x_1$ is nominally \off.  The variations in the
453: \off-state of $x_1$ are caused by the delayed self-input that
454: suppresses the \off-state during a time of approximately $\tau$
455: following each pulse.  In the limit cycle with two pulses, the
456: suppression is still active when the next pulse starts.  This is not
457: the case for the single-pulse attractor and the slightly higher
458: \off-state leads to a suppression of the pulse height.
459: 
460: The above observations remain valid for all sufficiently large values
461: of $\tau$.  For small $\tau$, the systems studied have only fixed
462: point attractors.  As $\tau$ is increased, the oscillatory attractors
463: are born by means of saddle-node bifurcations of cycles
464: \cite{Strogatz:00} in Cases II and III.  In Case I, a subcritical
465: bifurcation to a state cycle with period near $3\times(5/2)(\tau+1)$
466: is followed by a subcritical, symmetry-restoring bifurcation with
467: period near $(5/2)(\tau+1)$.  The cycle with period near
468: $3\times(5/2)(\tau+1)$ has no direct correspondence to a cycle in the
469: autonomous Boolean network.
470: 
471: \section{Discussion}
472: 
473: Our study of simple and single-self-input rings has elucidated some
474: important non-Boolean features of continuous systems: (i) the
475: asymmetry in the reaction time of an element when an input switches
476: \on\ or \off; (ii) deviations from nominal \on\ and \off\ values; and
477: (iii) memory effects due to the exponential decay of variables to
478: their steady state values for fixed inputs.  These features are
479: crucial for the stabilization and destabilization of oscillations in
480: the systems we have investigated.  In simple rings, they lead to a set
481: of stable attractors that coincides precisely with the reliable
482: attractors identified by Klemm and Bornholdt \cite{Klemm:05a}.  When
483: more complex logic is introduced, as illustrated in here by adding a
484: self-input to one element in a ring, the stable continuous attractors
485: do not correspond to the reliable attractors.
486: 
487: These observations are important for generalizing the well-developed
488: theory of large random Boolean networks to generic systems.  We would
489: like to know, for example, whether large, complex networks exhibit a
490: well-defined dynamical phase transition similar to the ``order-chaos''
491: transition in ensembles of Boolean networks
492: \cite{Kauffman:69,Derrida:86a,Samuelsson:06}.  An important feature of
493: the Boolean network transition is the rapid scaling of the number of
494: attractors with system size in the disordered regime, which suggests
495: that we should try to understand the set of attractors of large and
496: complex continuous systems.
497: 
498: The non-Boolean effects discussed in this paper may also be directly
499: relevant for understanding cell cycle oscillations in yeast, where
500: there is evidence for a transcriptional oscillator and recent
501: proposals for the genes involved suggest a fundamental ring of four
502: elements with multiple feed-forward and feedback links
503: \cite{pramila:06}.  We have seen, for example, that a pulse of
504: activity may propagate stably in such networks even where Boolean
505: reasoning suggests otherwise.  In systems where distinct elements have
506: significantly different time delays, we expect additional deviations
507: from the dynamics in synchronous or autonomous Boolean networks.
508: Future studies along these lines should elucidate the behavior of
509: larger rings and more complex network structures.
510: 
511: %% ------------------------------------------------------------------------ %%
512: %  ACKNOWLEDGMENTS
513: %% ------------------------------------------------------------------------ %%
514: 
515: \begin{acknowledgments}
516: We thank S.~Kauffman, A.~Ribeiro, and M.~Andrecut for stimulating
517: conversations and the U.~Calgary Inst.\ for
518: Biocomplexity and Informatics for hosting extended visits by
519: Samuelsson and Socolar.  This work was supported by the National
520: Science Foundation through Grant No.~PHY-0417372 and the Graduate
521: Research Fellowship Program.
522: \end{acknowledgments}
523: 
524: 
525: \appendix
526: 
527: \section{Autonomous dynamics}
528: \label{app:autonomous}
529: 
530: For completeness, we describe the implementation of autonomous Boolean
531: systems introduced by Klemm and Bornholdt \cite{Klemm:05a}.  The
532: important features are that switching times are determined by local
533: time delays and that repeated switching on times scales much shorter
534: than the delay time is suppressed.  There is no external clock and
535: no stochastic rule for determining update times.
536: 
537: As in Ref.~\cite{Klemm:05a}, we assume a time delay of $1$ unit
538: between the switching of a given node and the switching it induces in
539: nodes directly linked to it.  At each node, a low-pass filter is
540: assumed to suppress spikes of duration much shorter than $1$.  We let
541: $\epsilon$ be the minimum duration of an output pulse that passes
542: the filter (and require $0<\epsilon\ll1$).  If the inputs to a given
543: node switch at times that would lead to a pulse shorter than
544: $\epsilon$, the output from that node is assumed to stay constant
545: during that time.
546: 
547: For notational convenience, we write $F_j(t)$, suppressing its direct
548: dependence on its inputs.  We also let the Boolean values \off\ and
549: \on, respectively, correspond to the real values $0$ and $1$.  Then,
550: $x_j(t)$ is given by
551: \begin{align}\label{eqn:filter}
552:   x_j(t) &= \Theta\biggl[\frac{1}{2\epsilon}
553:             \int_{-\epsilon}^{\epsilon}d\delta\left(F_j(t-1+\delta) - \frac{1}{2}\right)
554:             \biggr]\,,
555: \end{align}
556: where $\Theta(x)$ is the Heaviside step function; $\Theta(x) = 0$ for
557: $x<0$ and $1$ for $x\ge0$.
558: 
559: When the timing of some switching events are perturbed
560: infinitesimally, $F_j$ may switch twice in rapid succession, which
561: would generate a positive or negative pulse of infinitesimal duration.
562: The filter of Eq.~\eqref{eqn:filter} suppresses such spikes, ensuring
563: that the number of switching events remains constant (for cycles in
564: which all spikes are longer than $\epsilon$ in duration) and
565: therefore that the stability of a state cycle is well defined with
566: respect to infinitesimal timing perturbations.
567: 
568: \section{Numerical integration method}
569: \label{app:methods}
570: 
571: Numerical integration of the time-delay equations
572: \eqref{eqn:continuous}--\eqref{eqn:f2in} was accomplished using a
573: fourth-order scheme based on the solution shown in
574: Eq.~\eqref{eqn:dynamicsintegral}.  To evolve the system through a time
575: step $h$, we define values of $x$ at the time points $t = nh$ for all
576: integer $n$ and define $m\equiv \tau/h$, where $\tau$ is the delay
577: time and $h$ is chosen such that $m$ is an integer.  To advance from
578: time step $n-1$ to $n$ we use the following formula (suppressing the
579: element index subscript $j$ for notational simplicity):
580: \begin{align}
581: \label{eqn:numerics}
582:   x_n & = 2h\left(\frac{1}{6}f_{n-m}+\frac{2}{3}e^{-h}f_{n-m-1}+\frac{1}{6}e^{-2h}f_{n-m-2}\right) \nonumber \\ 
583:      \  & \phantom{=}~ +\, e^{-2h}x_{n-2}\,,
584: \end{align}
585: where $f_n$ is the function defined by Eq.~\eqref{eqn:f1in}
586: [or~\eqref{eqn:f2in}] evaluated at time $t=nh$.  The integrator is
587: initialized with a desired value of $x_0$, where it is
588: assumed that $x_{n} = x_0$ for all $n<0$.  Integrations were carried
589: out using $h=0.1$.  Decreasing $h$ to $0.01$ had no noticeable effect
590: on the results.
591: 
592: Attractors were found by running from many (of order 50) different
593: initial conditions.  In some cases, such as the attractor shown in
594: Fig.~\ref{fig:timeseries}(b), it was necessary to arrange initial
595: conditions in which some nodes were artificially held at constant
596: values and released at different times.  The bifurcation structures as
597: a function of the time-delay parameter $\tau$ were determined by
598: performing integrations in which $h$ was increased or decreased very
599: slowly and observing transitions in the oscillation patterns of all
600: $x$'s.
601: 
602: \vfill
603: 
604: \bibliography{rings}
605: 
606: \end{document}
607: 
608: 
609: