physics0501117/H1.tex
1: \documentclass[twocolumn,superscriptaddress,showpacs,preprintnumbers,amsmath,amssymb,aps,pre]{revtex4}
2: %\documentclass[preprint,superscriptaddress,showpacs,preprintnumbers,amsmath,amssymb,aps,pre]{revtex4}
3: \usepackage{graphicx}% Include figure files
4: \usepackage{dcolumn}% Align table columns on decimal point
5: \usepackage{bm}% bold math
6: \begin{document}
7: \title{Entropy in the natural time-domain\footnote{Published in Physical Review E {\bf70}, 011106 (2004).}}
8: \author{P. A. Varotsos}
9: \email{pvaro@otenet.gr}
10: \affiliation{Solid State Section, Physics Department, University of Athens, Panepistimiopolis, Zografos 157 84, Athens, Greece}
11: \affiliation{Solid Earth Physics Institute, Physics Department, University of Athens, Panepistimiopolis, Zografos 157 84, Athens, Greece}
12: \author{N. V. Sarlis}
13: \affiliation{Solid State Section, Physics Department, University of Athens, Panepistimiopolis, Zografos 157 84, Athens, Greece}
14: \author{E. S. Skordas}
15: \affiliation{Solid Earth Physics Institute, Physics Department, University of Athens, Panepistimiopolis, Zografos 157 84, Athens, Greece}
16: \author{M. S. Lazaridou}
17: \affiliation{Solid State Section, Physics Department, University of Athens, Panepistimiopolis, Zografos 157 84, Athens, Greece}
18: \begin{abstract}
19:  A surrogate data analysis is presented, which is based on the fluctuations of the ``entropy'' $S$
20: defined in the natural time-domain [Phys. Rev. E {\bf 68}, 031106, 2003]. This entropy is not a static one as, for example, the Shannon entropy. The analysis is applied to three 
21: types of time-series, i.e., seismic electric signals, ``artificial'' noises and 
22: electrocardiograms, and ``recognizes'' the non-Markovianity in all these signals. 
23: Furthermore, it differentiates the electrocardiograms 
24: of healthy humans from those of the sudden cardiac  death ones. If $\delta S$ and $\delta S_{shuf}$
25: denote the standard deviation when calculating the entropy by means of a time-window sweeping 
26: through the original data and the ``shuffled'' (randomized) data, respectively, it seems that the ratio 
27: $\delta S_{shuf}/\delta S$ plays a key-role. The physical meaning of $\delta S_{shuf}$ is
28: investigated.
29: \end{abstract}
30: \pacs{05.40.-a, 87.17.-d}
31: \maketitle
32: 
33: \section {introduction}\label{sec1}
34: 
35: In an electric signal consisting of N pulses, the natural time was introduced\cite{var01,var02} 
36: by ascribing to the $k$-th pulse the value $\chi_{k} = k/N$. The analysis is then made in terms of 
37: the couple ($\chi_{k}$, $ Q_{k}$) where $ Q_{k}$ stands for the duration of the $k$-th pulse.
38: The entropy $S$, defined\cite{var01, var03b}  as 
39: $S \equiv \langle \chi \ln \chi \rangle -\langle \chi \rangle 
40: \ln \langle \chi \rangle$,
41: where $\langle \chi \rangle = \sum_{k=1}^N p_k \chi_k$, $p_k$=$Q_{k}/\sum_{n=1}^N Q_n$ and 
42: $\langle \chi \ln \chi \rangle = \sum_{k=1}^N p_k \chi_k \ln \chi_k$,
43: was found\cite{var03b} to distinguish Seismic Electric Signals (SES) 
44: activities from artificial noises (AN), where the latter terminology stands for 
45: electrical disturbances which are recorded at a measuring site due to 
46: nearby man-made electric sources. More precisely, SES activities 
47: and AN have $S$-values smaller and larger than that
48:  ($S_{u}$) of a ``uniform'' (u)
49: distribution, respectively (as the latter was defined in 
50: Refs. \cite{var01,var03,var03b}). Furthermore, ion 
51: current fluctuations in membrane channels  (ICFMC) have $S$ very close to $S_{u}$ \cite{var03b}. 
52: 
53: The fact that a system contains nonlinear components does not necessarily reflect that a 
54: specific signal we measure from the system also exhibits nonlinear features. Thus, before 
55: analyzing this signal by applying nonlinear techniques, we must first clarify if the use of such 
56: techniques is justified by the data available. The method of surrogate data has been 
57: extensively used to serve such a purpose (see Ref. \cite{sch00} for a review). Surrogate data refer 
58: to data that preserve certain linear statistic properties of the experimental data, but 
59: are random otherwise\cite{cha95,siw01}. These data are prepared by various procedures; 
60: for example, Siwy et al.\cite{siw01} in order to study the nature of dwell-time series in ICFMC, among other methods, also used 
61: surrogate data which have been obtained by three  different procedures. 
62:  The present paper aims, in general, 
63: at presenting a kind of surrogate data analysis using the entropy fluctuations in the natural 
64: time-domain (see below) as discriminating statistics. Throughout the paper, the surrogate data 
65: are obtained by  shuffling the $Q_k$ randomly and hence their 
66: distribution is conserved. Applying such a procedure, we do the following: consider the
67:  null hypothesis that the data consist of 
68: {\em independent} draws from a fixed probability distribution of the dwell times;
69:  if we find significantly 
70: different serial correlations in the data and their shuffles, we can reject the hypothesis of 
71: {\em independence}, see paragraph 3.1 
72: of Ref. \cite{sch00}. In other words, the tested null hypothesis is that $Q_k$ 
73: are independent and identically distributed (iid) random variables, i.e., that there are no 
74: correlations between the lengths of consecutive intervals. If the original (continuous) time series
75: is Markovian then the null hypothesis for the $Q_k$ should hold, i.e., the $Q_k$ are iid. We
76: emphasize that the terminology ``Markovian'' throughout this paper always refers to the 
77: original time series.
78: 
79: 
80: 
81: 
82: Here, as a measure of the natural time entropy fluctuations we consider 
83: the standard deviation $\delta S$ 
84: when we calculate the value of $S$ for a number of consecutive pulses 
85: and study how $S$ varies when 
86: sweeping this time-window through the whole time-series. 
87: We use the following three data sets: Two of them  
88: are those treated in Ref. \cite{var03b}, i.e., 
89: SES activities and AN. As a third one, we preferred to use, 
90: instead of ICFMC, the case of electrocardiograms (ECG), for several reasons, 
91: chief among of which are: (a) They are publicly accessible \cite{GOL00}. 
92: (b) Instead of the single ICFMC example, a large variety of ECG are available 
93: (i.e., 105 individuals are employed here, 10 healthy and 95 patients). 
94: (c) The case of ECG is similar to ICFMC, in the sense that the $S$-value in ECG 
95: results very close to $S_{u}$ as in ICFMC investigated in \cite{var03b}. 
96: Note, however, that the intervals
97:  between heart beats fluctuate widely, e.g., \cite{chia02}.
98: 
99: 
100:  
101: A general agreement about whether or not normal heart dynamics are chaotic 
102: or not chaotic is still lacking (e.g., see Ref.\cite{p2} and references therein). 
103:  The most commonly used non-linear complexity measures are fractal dimensions of
104:  various kinds (e.g., correlation dimension, Renyi dimensions). Each of them measures 
105: different aspects of the {\em statistics} on the attractor. On the other hand, 
106: Liapunov exponents and the Kolmogorov-Sinai entropy (K-S entropy) and entropy rates 
107: are measures of the {\em dynamics} on an attractor. Except for the K-S entropy and entropy rates,
108:  the other categories of complexity measures assume a purely 
109: deterministic system (e.g., see Ref.\cite{p11}). Since a physiological time series may be due to a mixed process, 
110: stochastic and deterministic, the use of fractal dimensions in physiological time 
111: series has been occasionally criticized\cite{p11}. On the other hand, entropy is a 
112: concept equally applicable to deterministic as well as stochastic processes.
113: This is why we preferred to use the entropy in natural time (more
114: precisely its fluctuations $\delta S$ ) as  discriminating statistics. 
115: The following point, 
116: however, should be stressed. Complexity measures based on {\em static} entropy 
117: (e.g., Shannon entropy) quantify {\em statistical} order in the time series. 
118: The underlying key-property of these complexity measures is the probability
119:  distribution of the (dwell times in the) data analyzed; thus, the result of such computations should 
120: be independent of permutations performed on the (sequence of the dwell times in the) 
121: time series as in a surrogate (randomized)
122:  data set obtained by data shuffling. On the other hand, the entropy in 
123: natural time (and the relevant measures) considers, from its definition, the {\em sequential} 
124: order (of beats); in other words, $S$ is a {\em dynamic} entropy, i.e., it captures 
125: characteristics of the dynamics in a system. Additional comments on the importance 
126: of the fluctuations of $S$ in ECG will be forwarded in Section \ref{sec5}. 
127: 
128:  
129: 
130: In all examples, we use a sliding window of length three to ten pulses, 
131: except otherwise stated. Concerning the symbols: We reserve $\delta S$ {\em only} 
132: for the case when the calculation is made by a {\em single} time-window,
133:  e.g., 5 pulses. The symbol $\overline{ \delta S}$ denotes 
134: the average of the $\delta S$-values calculated for a sequence of single windows, 
135: e.g., 3, 4 and 5 pulses. Finally, $\langle \delta S \rangle$ stands for the  
136: $\delta S$-values averaged over a group of individuals, e.g., the 10 healthy subjects. 
137: 
138:  The present paper is organized as follows: In section \ref{sec2}, we investigate whether 
139: a distinction between SES activities and AN can be achieved by the 
140: $\delta S$-value alone. 
141: Furthermore, we examine if $\delta S$ can recognize the non-Markovianity in all the signals investigated. 
142: In section \ref{sec3}, we attempt to shed light on the quantity $\delta S_{shuf}$ 
143: calculated in a surrogate (randomized) data set obtained by data ``shuffling''. 
144: We find that $\delta S_{shuf}$ in ECG is a measure of 
145: $\sigma / \mu$ (where $\mu$ and $\sigma$ stand for the mean value and
146:  the standard deviation 
147: of the corresponding intervals, see below). Section \ref{sec4} shows that the $\delta S_{shuf}$-value 
148: differs from $\delta S$, as expected 
149: (cf. the entropy $S$  is {\em not} 
150: static entropy,as mentioned above). 
151: The prominent role of the ratio $\delta S_{shuf}/\delta S$ in distinguishing ECG of healthy humans 
152: from those suffered from sudden cardiac death is shown in Section \ref{sec5}.
153: The conclusions are summarized in Section \ref{sec6}. 
154: Finally, an Appendix is reserved to derive an exact 
155: relationship between $\delta S_{shuf}$ and $\sigma / \mu$   when $Q_k$ are iid.
156: 
157: 
158: 
159: \section{the possibility of employing $\delta S$ to ``recognize'' the non-Markovianity} \label{sec2}
160: We start by examining whether the $\delta S$-values alone can distinguish 
161: SES activities from AN as well as ``recognize'' their non-Markovianity. Recall\cite{var02,var03b},
162:  that SES and AN are time-series of dichotomous nature which are non Markovian. 
163: In a {\em dichotomous} Markovian time-series, the 
164: dwell times ($Q_k$) are exponentially distributed; for such a series we plot,  
165: in Fig. \ref{fig1}(a),
166:   the $\delta S$-value versus the time-window length. (Since in the 
167: calculation of $S$  
168: only ratios of  $Q_k$ are involved the result does not depend on the transition 
169: rates of the Markovian process.)  
170: The error shown in this case is on the average 7\%. (The calculation 
171: was made for a total number of $10^2$ pulses, see below. Note that this error 
172: decreases upon increasing the number of pulses, i.e., it becomes $\approx$ 2\% for $10^3$ pulses, 
173: which will be used later). 
174: In the same figure, we insert the $\delta S$-values calculated for the four SES 
175: activities (labelled K1, K2, A, U) and the six AN (labelled n1 to n6) depicted in Fig. 1 
176: of Ref. \cite{var03b}. An inspection of Fig.\ref{fig1}(a) reveals the following conclusions. 
177: First, {\em no} distinction between SES activities and AN (both of which have estimation 
178: errors comparable to the aforementioned error of the Markovian) is obvious. 
179: An inspection of Table I of \cite{EPAPS03},
180: reveals that the number of pulses in three (out of the four) 
181: SES activities is around $10^2$, for K2, U 
182: and A (while for K1, is $\approx 310$) and this is why we calculated here the 
183: Markovian case for $10^2$ pulses. Second, concerning the possibility of 
184: ``recognizing'' the non-Markovianity (as discussed and shown in 
185: Refs. \cite{var02,var03,var03b} 
186: by independent procedures):
187:  This could be possibly supported,
188:  {\em only} for the shorter 
189: time-windows  (i.e., 3, 4 and possibly 5 pulses) for all SES 
190: activities as well as for most 
191: AN (i.e., n6, n4, n3, n2, possibly n1, but {\em not} for n5), 
192: see Fig.\ref{fig1}(a). 
193: 
194: 
195: 
196: We now investigate if the $\delta S$-values alone can ``recognize'' 
197: the non-Markovianity in ECG. In a single sinus (normal) cycle of an ECG, 
198: the turning points are labelled with the letters P, Q, R, S and T. We used here the 
199: QT database from physiobank \cite{GOL00}( see also \cite{LAG97}), which consists 
200: of 105 fifteen-minute excerpts of Holter recordings as follows: 10 from MIT-BIH 
201: Normal Sinus Rhythm Database (i.e., healthy subjects, hereafter labelled H), 15 from MIT-BIH Arrhythmia Database (MIT), 
202: 13 from MIT-BIH Supraventricular Arrhythmia Database (MSV), 6 from MIT-BIH ST 
203: Change Database (MST), 33 from the European ST-T Database (EST), 4 from MIT-BIH 
204: Long-Term ECG Database (LT) and 24 from sudden death patients from BIH(SD). 
205: (cf. BIH denotes the Beth Israel Hospital). In Fig. \ref{fig2}, we plot, 
206: for the 
207: QRS-interval time-series, the $\delta S$-value averaged over each of the aforementioned 
208: seven groups versus the time-window length. Since all time-series 
209: of these seven groups have 
210: $\approx 10^3$ intervals, we insert in the same figure the results calculated 
211: for a Markovian case (cf. with the procedure mentioned in the previous 
212: paragraph) of comparable length $\approx 10^3$.  
213: We see that the Markovian case exhibits $\delta S$-values that are roughly one 
214: order of magnitude larger than those of the seven groups of humans, which clearly 
215: points to the non-Markovianity of {\em all} the signals in these groups.  We emphasize that the same 
216: conclusions are drawn if we 
217: consider, instead of QRS-, the series of QT-intervals, or the  beat-to-beat 
218: intervals (RR). 
219: In summary, the $\delta S$-value alone can well recognize
220:  the non-Markovianity in ECG. 
221: 
222: \begin{figure}
223: \includegraphics{Fig1}% Here is how to import EPS art
224: \caption{\label{fig1} (Color) (a) The $\delta S$-values for 
225: each SES-activity and artificial 
226: noise versus the time-window length. The corresponding 
227: values for a Markovian 
228: time-series ($10^2$ pulses) are also plotted (green). 
229: (b) $\overline{ \delta S}$ versus $\overline{ \delta S}_{shuf}$(time-window range 3-5) 
230: for all the SES 
231: activities and AN in (a). The straight line corresponds to 
232: $\overline{ \delta S}_{shuf}$ = $\overline{ \delta S}$.}
233: \end{figure} 
234: 
235: \begin{figure}
236: \includegraphics{Fig2}% Here is how to import EPS art
237: \caption{\label{fig2}  (Color) The $\langle \delta S \rangle$-values for the QRS-intervals (see the text) of the seven groups 
238: of humans versus the time-window length. The corresponding values for a 
239: Markovian time-series ($10^3$ pulses, labelled M) are also plotted. }
240: \end{figure} 
241: 
242: \begin{figure}[b]
243: \includegraphics{Fig3}% Here is how to import EPS art
244: \caption{\label{fig3}The $\sigma / \mu$-value, for each of the 105 individuals, versus 
245: the corresponding $\overline{ \delta S}_{shuf}$-value 
246: for the (a) RR-, (b) QT- and (c) QRS-intervals. The identity of the individual 
247: associated with each point can be found in Ref.\cite{EPAPS}.
248: }
249: \end{figure}
250: 
251: \begin{figure}[b]
252: \includegraphics{Fig4}% Here is how to import EPS art
253: \caption{\label{fig4} (Color) The $\sigma/\mu$-value for  RR-, QT- and 
254: QRS-intervals of the ten H 
255: versus the corresponding 
256: $\overline{ \delta S}_{shuf}$-value (time-window range 3-10 beats). 
257: The straight 
258: line results from a least squares fit of all the thirty points. 
259: For the identity of the individual associated with each point 
260: see Ref. \cite{EPAPS}. }
261: \end{figure}
262: 
263: \begin{figure}[b]
264: \includegraphics{Fig5}% Here is how to import EPS art
265: \caption{\label{fig5}The $\overline{ \delta S}$-value, 
266: for each of the 105 individuals versus the corresponding $\overline{ \delta S}_{shuf}$-value
267: for (a) RR-, (b) QT- and (c) QRS-intervals. The straight line, drawn in each case,
268: corresponds to $\overline{ \delta S}_{shuf}$ = $\overline{ \delta S}$.
269: For the identity of the individual associated with each point 
270: see Ref. \cite{EPAPS}.}
271: \end{figure}
272: 
273: \begin{figure}
274: \includegraphics{Figure6}% Here is how to import EPS art
275: \caption{\label{fig6} (Color)
276: The $\overline{ \delta S}$-value,
277: in each of the 10 H (black) and 24 SD (red), for the QT-intervals versus $\overline{ \delta S}_{shuf}$ 
278: (time-window range: 3-10 beats). Note that the values of the ordinates are
279:  appreciably smaller than the $\delta S$-value ($\approx 2\times 10^{-2}$) of the Markovian time-series 
280: ($10^3$ events) depicted in Fig. \ref{fig2}.}
281: \end{figure}
282: 
283: 
284: 
285: \section{the physical meaning of $\delta S_{shuf}$} \label{sec3}
286: 
287: In Fig. \ref{fig3}(a) we plot, for each of the 105 individuals, the value 
288: of $\sigma / \mu$ versus the corresponding value of $\overline{ \delta S}_{shuf}$ 
289: (time-window range 3-10 beats) for the RR-intervals. 
290: The same is repeated in Figs. \ref{fig3}(b) and \ref{fig3}(c) for the QT- and QRS-intervals, 
291: respectively. All these three plots, can be described by linear behavior 
292: and a least squares fitting to a straight line passing through the origin 
293: leads to the following slopes: 
294: 38.6 $\pm$ 0.6, 36.8 $\pm$ 0.2 and 40.1 $\pm$ 0.4, for the RR-, QT- and QRS-intervals, 
295: respectively.
296:  This points to the conclusion that 
297: $\delta S_{shuf}$ provides, as intuitively expected, a measure of $\sigma/\mu$. 
298: (This, however, {\em cannot}  be supported with certainty for 
299: the SES activities and AN.) 
300: Note that, although these three slopes are more or less comparable, 
301: they differ by 
302: amounts lying outside their standard error. 
303: Furthermore, it may be worthwhile to 
304: mention, that if we study {\em altogether} the RR-, QT- and QRS-intervals,
305:   for 
306: the 10 healthy humans {\em only} (Fig. \ref{fig4}),
307:  a good linearity of $\sigma/\mu$ 
308: versus $\overline{ \delta S}_{shuf}$ results with a slope 37.5 $\pm$ 0.4. 
309: (cf. if we study each of the three intervals separately, we find slopes that agree 
310: within the error margins, i.e., 37.5 $\pm$ 0.4, 37.1 $\pm$ 0.7 and 37.8 $\pm$ 0.1 for 
311: the RR-, QT- and QRS-intervals, respectively). The origin of this {\em common} 
312: behavior merits further investigation. 
313: 
314: 
315: One could argue that $Q_k$ may become iid upon their shuffling.
316:  In the Appendix, we show that,  when 
317: $Q_k$ are iid, $\delta S$ is  actually  proportional to $\sigma/\mu$; 
318: %(cf. this is the null hypothesis of the shuffling procedure employed), 
319: the following relationship is obtained:
320: \begin{eqnarray}
321: \delta S_{shuf} &=& \frac{\sigma}{\mu} \frac{1}{\sqrt{N-1}}  [  
322: \sum_{k=1}^N \left(\frac{k}{N}\ln \frac{k}{e N  \overline{\chi}  } \right)^2 \frac{1}{N} \nonumber \\
323: & -& \left( \sum_{k=1}^N \frac{k}{N^2} \ln \frac{k}{e N \overline{\chi}}   \right)^2 
324: ]
325: \label{eq1}
326: \end{eqnarray}
327: where 
328: \begin{equation}
329: \label{eq2}
330: \overline{\chi}  = \sum_{k=1}^N \frac{k}{N^2}   = \frac {1} {2}+ \frac {1} {2N}
331: \end{equation}
332: and $e$ denotes, as usually, the base of the natural logarithms.
333: The relation (\ref{eq1})  reveals  that $\delta S_{shuf}$ versus $\sigma/\mu$ must be 
334: a straight line with a slope  ranging from 34.2 to 40.4, for a time-window length 3 to 10. 
335: This result is comparable with the slopes determined above from the analysis of the ECG data.  
336: 
337: 
338: \section{on the difference between $\delta S$ and $\delta S_{shuf}$} \label{sec4}
339: 
340: We first comment on the difference between $\delta S$ and $\delta S_{shuf}$ in the SES activities and AN. In Fig. \ref{fig1}(b), the value of 
341: $\overline{ \delta S}$ versus the corresponding $\overline{ \delta S}_{shuf}$ was plotted, 
342: for each of the ten signals discussed in Fig. \ref{fig1}(a). The average values 
343: in Fig. \ref{fig1}(b) have been calculated over the three time-windows 
344: of 3, 4 and 5 pulses, 
345: since we mentioned in Section \ref{sec2}  that the ``recognition'' of the non-Markovianity  in all 
346: SES activities becomes possible  in this time-window range. 
347: If we disregard n6, 
348: and despite the errors of around 5\% (for the time-window range 3-5), we may say 
349: that there is a systematic tendency pointing to a value of 
350: $\overline{ \delta S}_{shuf}/\overline{ \delta S}$ larger than unity (cf. 
351: the same 
352: conclusion is drawn, if we take the averages over the time-window 
353: range 3-10). 
354: This is consistent with the non-Markovianity   of all these signals, 
355: because for a Markovian case we expect 
356: $\overline{ \delta S}_{shuf}$=$\overline{ \delta S}$. (Since, by definition, 
357: $\delta S_{shuf}$ corresponds to the entropy fluctuations upon {\em random} mixing 
358: of $Q_k$, see Section \ref{sec1}, it is naturally expected that in 
359: a Markovian case the two quantities 
360: $\delta S$ and $\delta S_{shuf}$ {\em should} coincide). Note that the reverse is {\em not} 
361: always true (thus the equality $\overline{ \delta S}_{shuf}$ =  $\overline{ \delta S}$ may 
362: also hold for {\em non}-Markovian time series) as  will be demonstrated  below 
363: with precise examples. 
364: 
365: We now proceed to compare $\delta S_{shuf}$ with $\delta S$ in ECG. 
366: Figure \ref{fig5}(a) depicts the $\overline{ \delta S}$-values, calculated for each 
367: of the 105 individuals, versus the corresponding $\overline{ \delta S}_{shuf}$ for 
368: the RR-intervals (time-window range 3-10 beats).
369: The same is repeated in Figs. \ref{fig5}(b) and \ref{fig5}(c) for 
370: the QT- and QRS-intervals, respectively. In each case, we also plot the straight 
371: line $\overline{ \delta S}_{shuf}$=$\overline{ \delta S}$ to visualize that the vast 
372: majority of points fall below this line. The non-equality of $\overline{ \delta S}_{shuf}$ 
373: and $\overline{ \delta S}$ has been also verified by applying the Wilcoxon paired signed-rank 
374: test recommended\cite{mot95} to be followed for non-Gaussian paired data. The tested null hypothesis is that the means of $\overline{ \delta S}_{shuf}$
375: and $\overline{ \delta S}$ are the same and is rejected at a level of significance well below 0.01,
376:  since the data of Figs. 5(a),(b) and (c) lead to normally distributed
377:  variables $z=-8.29, -6.81$ and $-6.32$,
378:  respectively (cf. the corresponding one-tailed asymptotic significance is given by $P(Z<z)$, 
379: i.e.,  the probability to obtain a normally distributed variable which is smaller than $z$).
380:  Note that a least squares fit to a straight line 
381: passing through the origin, results  in the following expressions: 
382: $\delta S=(0.76\pm 0.03)\delta S_{shuf}$, $\delta S=(0.85\pm 0.02)\delta S_{shuf}$, 
383: $\delta S=(0.94\pm 0.02)\delta S_{shuf}$ for the Figs. \ref{fig5}(a),\ref{fig5}(b),
384: \ref{fig5}(c), respectively. The sampling rate $F_s$ in ECG is 250 Hz; thus, if we take 
385: as an example the RR-intervals,  the experimental error in their allocation is around $1/F_s$. 
386: The latter reflects in the calculation of $\delta S$ and $\delta S_{shuf}$ errors 
387: which are drastically smaller than those required to eventually justify a compatibility of 
388: the expression $\overline{ \delta S} = (0.76\pm 0.03)\overline{ \delta S}_{shuf}$, 
389: found from Fig.\ref{fig5}(a), with a straight line of slope equal to unity,
390:  i.e., $\overline{ \delta S} = \overline{ \delta S}_{shuf}$.
391: 
392: The difference between $\delta S$ and 
393: $\delta S_{shuf}$ in ECG could be understood in the context that the former depends on 
394: the {\em sequential} order (of beats), as mentioned in Section \ref{sec1}, while the latter does not. Since 
395: short- (and long-) range correlations is a usual 
396: feature( see Ref. \cite{PNAS02} and references
397:  therein)
398:  in heartbeat dynamics, which are possibly destroyed (or become
399: weaker) upon randomizing the data, more
400: ``disorder'' is intuitively expected to appear after randomization, thus reflecting
401: $\delta S_{shuf} > \delta S$. 
402: Furthermore, note that in {\em all} the three plots of Fig. \ref{fig5} there are some 
403: drastic deviations from the straight line $\overline{ \delta S} = \overline{ \delta S}_{shuf}$. 
404: The origin of these deviations is currently investigated in detail. 
405: 
406: 
407: 
408: 
409: Finally, we further clarify the aforementioned point
410:  that the equality $\overline{ \delta S} = \overline{ \delta S}_{shuf}$ does {\em not} 
411: necessarily reflect Markovianity. In Fig.\ref{fig6}, we plot, for the QT-intervals, 
412: $\overline{ \delta S}_{shuf}$ versus $\overline{ \delta S}$ (for time-window range 3-10 beats) for SD and H. 
413: We see that there are several individuals (mainly SD, see also next section) of which 
414: their points lie practically (i.e., within the error margins) 
415: on the straight line $\overline{ \delta S} = \overline{ \delta S}_{shuf}$. 
416: If we plot their $\delta S$- (or $\delta S_{shuf}$-) values versus the time-window 
417: (in a similar fashion as in Fig. \ref{fig2}), we find that these values are distinctly smaller
418:  than those of the Markovian case, thus making clear that these
419:  individuals cannot be characterized as exhibiting Markovian behavior. 
420: (This non-Markovianity holds for {\em all} H and {\em all} SD.)
421: 
422: 
423: \section{the use of $\overline{ \delta S}_{shuf} / \overline{ \delta S} $ to 
424: distinguish ECG of healthy humans from the sudden cardiac death ones} \label{sec5}
425: 
426: Here we focus only on two groups of ECG, namely H and SD, and examine whether 
427: they can be distinguished by means of the ratio $\overline{ \delta S}_{shuf} / \overline{ \delta S} $. 
428: We calculate this ratio, for each 
429: type of interval, at two ranges: (i) a short (s) range 3-4 beats (cf. consider that the {\em smallest} 
430: number allowed for the natural time-domain analysis is 3 beats) and (ii) a longer (L) range 50-70 beats. 
431: For the sake of convenience, we define $\nu \equiv \overline{ \delta S}_{shuf} / \overline{ \delta S} $, 
432: and hence the following 
433: ratios are investigated: $\nu_s(\tau)$ and $\nu_L(\tau)$, where $\tau$ denotes the type of 
434: interval (i.e., $\tau$=RR, QRS or QT) and $s, L$ refer to the range studied (i.e., $s=$ 3-4 
435: beats and $L$=50-70 beats).
436: 
437: The calculated values for $\nu_s(\tau)$ and $\nu_L(\tau)$ for the three types of intervals are given, for all 
438: H and SD, in Table \ref{tab1}. The minima $\min_H[\nu_\kappa (\tau)]$ and maxima $\max_H[\nu_\kappa (\tau)]$ (where 
439: $\kappa$ denotes either the short, $\kappa=s$, or the longer, $\kappa=L$, range) among the healthy subjects are also inserted in 
440: two separate rows, for each type of interval and each range studied. These minima and maxima 
441: are labeled $H_{min}$ and $H_{max}$, respectively. The cases of SD which have smaller and larger values 
442: than $H_{min}$ and $H_{max}$ (reported in each column) are marked with superscripts ``a'' and ``b'' respectively.
443: 
444: A careful inspection of Table \ref{tab1} leads to the following main conclusion: {\em All} SD violate one or more H-limits 
445: (i.e., they have values that are smaller than $H_{min}$ or larger than  $H_{max}$). We intentionally emphasize that this 
446: conclusion is also drawn  {\em even} when disregarding the results for the QT-intervals. (Concerning the latter intervals: 
447: Only 5 SD out of 24 violate the H-limits; however,
448:  in {\em all} SD, their $\delta S$-values themselves are larger than those in H,
449:  see also Fig. \ref{fig6}. The usefulness of this difference will be discussed in 
450: detail elsewhere). In other words, when focusing our investigation solely on the RR-
451:  and QRS-intervals, {\em all} SD violate one or more of the four H-limits 
452: related to $\nu_s(RR)$, $\nu_L(RR)$, $\nu_s(QRS)$ and $\nu_L(QRS)$. This is important
453:  from practical point of view, because the RR- and QRS-intervals can be 
454: detected more easily (and accurately) than the QT- by means of an automatic
455:  threshold based detector (e.g., see Ref. \cite{jan97}, which evaluated the results 
456: of a detector that has been forwarded in Refs. \cite{lag90} and \cite{lag94} to 
457: determine automatically the waveform limits in Holter ECG).
458: 
459: 
460: \squeezetable
461: \begin{table*} 
462: \caption{ \label{tab1}The values of the ratios $\overline{\delta
463: S}_{shuf}/ \overline{{\delta S}}$ in the short ($s$) range 3-4 ($\nu_s$) or in the longer ($L$) range 
464: 50-70 beats ($\nu_L$) in H (sel16265 to sel17453) and SD (sel30 to sel17152) for the RR-, 
465: QRS- and QT-intervals} 
466: \begin{ruledtabular}
467: \begin{tabular}{c ccc ccc}
468: 
469:   individual & 3-4 beats ($\nu_s$) & & & 50-70 beats ($\nu_L$) & 
470: \\
471: 
472:  \cline{2-4}
473:  \cline{5-7}
474: \\
475:   & \vspace{0.05cm} RR & QRS & QT & RR & QRS & QT \\
476:  \hline 
477:  sel16265 & 1.82 & 1.00 & 1.24 & 0.48 & 1.02 & 0.76 \\
478:  sel16272 & 1.74 & 0.99 & 0.98 & 0.77 & 1.08 & 1.11 \\
479:  sel16273 & 2.21 & 1.00 & 1.48 & 0.50 & 0.88 & 0.71 \\
480:  sel16420 & 1.55 & 0.98 & 1.08 & 0.53 & 1.09 & 0.90 \\
481:  sel16483 & 2.25 & 1.02 & 1.14 & 0.52 & 1.16 & 0.92 \\
482:  sel16539 & 1.42 & 1.06 & 1.25 & 0.50 & 1.08 & 0.65 \\
483:  sel16773 & 1.94 & 1.00 & 0.99 & 0.44 & 1.05 & 0.96 \\
484:  sel16786 & 1.42 & 1.00 & 1.19 & 0.56 & 1.04 & 0.77 \\
485:  sel16795 & 1.18 & 0.98 & 1.08 & 0.73 & 0.96 & 0.99 \\
486:  sel17453 & 1.38 & 1.01 & 1.02 & 0.56 & 0.98 & 0.81 \\
487: \\
488:  $H_{min}$ & 1.18 & 0.98 & 0.98 & 0.44 & 0.88 & 0.65 \\
489:  $H_{max}$ & 2.25 & 1.06 & 1.48 & 0.77 & 1.16 & 1.11 \\
490: \\
491: sel30 & 1.29 & 1.11\footnotemark[2] & 1.09 & 0.65 & 0.72\footnotemark[1] & 1.09 \\
492: sel31 & 0.96\footnotemark[1] & 1.08\footnotemark[2] & 1.17 & 1.23\footnotemark[2] & 0.94 & 0.62\footnotemark[1] \\
493: sel32	& 1.39   & 1.14\footnotemark[2]  &  1.12  & 1.02\footnotemark[2]   &  0.69\footnotemark[1]   & 0.90 \\
494: sel33	& 1.05\footnotemark[1] & 0.99 & 1.00 & 0.86\footnotemark[2] & 0.82\footnotemark[1] & 0.99 \\
495: sel34	& 2.11   & 1.29\footnotemark[2]  & 1.11  &  0.42\footnotemark[1] &	 0.78\footnotemark[1]   &  0.67 \\
496: sel35	& 1.00\footnotemark[1] &  1.00  &  0.96\footnotemark[1]  &  1.01\footnotemark[2]   &   1.05   &  	 1.08 \\
497: sel36	& 1.02\footnotemark[1] & 1.02  & 1.04  &  0.92\footnotemark[2]   &  1.00   &  0.88 \\
498: sel37	& 1.07\footnotemark[1] & 1.18\footnotemark[2]  & 1.07  &  0.55   & 0.75\footnotemark[1]  & 0.65 \\
499: sel38	& 0.99\footnotemark[1] & 1.09\footnotemark[2]  &  1.13  & 1.37\footnotemark[2] & 0.89 & 1.04 \\
500: sel39	& 0.96\footnotemark[1] & 1.02  & 1.06 &  2.93\footnotemark[2] & 0.92 & 0.90 \\
501: sel40	& 1.01\footnotemark[1] & 1.00  & 0.93\footnotemark[1] & 0.78\footnotemark[2] & 0.93 & 1.29\footnotemark[2] \\
502: sel41	& 1.07\footnotemark[1] &  1.04  & 1.02 & 1.07\footnotemark[2] & 0.84\footnotemark[1] & 0.96 \\
503: sel42	& 1.63 & 1.08\footnotemark[2] & 1.23 & 0.42\footnotemark[1] & 1.06 & 0.67 \\
504: sel43	& 2.71\footnotemark[2] & 1.11\footnotemark[2] & 1.05 & 0.56 & 0.76\footnotemark[1] & 0.89 \\
505: sel44	& 0.91\footnotemark[1] & 0.95\footnotemark[1] & 0.88\footnotemark[1] & 2.24\footnotemark[2] & 1.46\footnotemark[2] &  1.32\footnotemark[2] \\
506: sel45	& 0.98\footnotemark[1] &  1.24\footnotemark[2]  & 1.29 &  0.98\footnotemark[2] &  	 0.86\footnotemark[1] & 0.79 \\
507: sel46	& 1.03\footnotemark[1] &  1.01 &  1.03 &  1.00\footnotemark[2] & 0.84\footnotemark[1] &  1.01 \\
508: sel47	& 1.56   & 0.97\footnotemark[1] & 1.03 &  0.45 & 0.97 &  1.01 \\
509: sel48	& 0.82\footnotemark[1] & 1.18\footnotemark[2] & 1.44 & 1.48\footnotemark[2] & 0.68\footnotemark[1] & 0.73 \\
510: sel49	& 0.93\footnotemark[1]  & 1.11\footnotemark[2] & 0.96\footnotemark[1] & 1.22\footnotemark[2] & 0.70\footnotemark[1] &  1.14\footnotemark[2] \\
511: sel50	& 1.05\footnotemark[1] & 0.98 & 0.98 & 0.93\footnotemark[2] & 1.23\footnotemark[2] & 1.50\footnotemark[2] \\
512: sel51	& 1.25  &  1.01 & 0.97\footnotemark[1] &  1.05\footnotemark[2] & 1.24\footnotemark[2] & 0.91 \\
513: sel52	& 1.50  &  1.16\footnotemark[2] &  1.22 & 1.00\footnotemark[2] & 0.73\footnotemark[1] &  0.68 \\
514: sel17152 & 1.64 &   	 1.01  &  1.04  &  0.90\footnotemark[2]  &  1.01 &	 0.97 \\
515: 
516: 
517: \end{tabular}
518: \footnotetext[1]{These values are smaller than the minimum ($H_{min}$) value of $\overline{\delta
519: S}_{shuf}/ \overline{{\delta S}}$ in H for each range}
520: \footnotetext[2]{These values are larger than the maximum ($H_{max}$) value of $\overline{\delta
521: S}_{shuf}/ \overline{{\delta S}}$ in H for each range} 
522: \end{ruledtabular}
523: \end{table*}
524:  
525: A further inspection of Table \ref{tab1} leads to the following additional comment: When 
526: investigating the RR-intervals {\em alone} (which can be detected automatically more easily and 
527: precisely than the other intervals), i.e., studying $\nu_s(RR)$ and  $\nu_L(RR)$, 
528: the vast majority of SD (22 out of 24 cases) can
529:  be distinguished from H (only two SD, i.e., 
530: sel30 and sel47, obey the corresponding H-limits).
531:  Specifically, concerning $\nu_s(RR)$, 15 SD have values 
532: smaller than $H_{min}=1.18$, while only one SD (i.e., sel43) has a value exceeding
533:  $H_{max}=2.25$; as for $\nu_L(RR)$, 18 SD exceed $H_{max}=0.77$,
534:  while only 2 SD (i.e., sel34 and sel42) have values smaller than $H_{min}=0.44$.
535: 
536: 
537: In what remains, we  proceed to a tentative physical interpretation of the above results, the main 
538: feature of which focuses on the fact that most SD simultaneously
539:  have $\nu_s(RR)$-values smaller than $H_{min}(=1.18)$ and $\nu_L(RR)$-values 
540: exceeding $H_{max}(=0.77)$. The RR time-series 
541: of healthy subjects are characterized by high 
542: complexity (e.g., \cite{iva01,PNAS02}); this, if we 
543: recall that in a Markovian series we intuitively expect $\delta S_{shuf} / \delta S=1$ 
544: (and hence $\nu_s=1$ and $\nu_L=1$), is compatible with the fact that in 
545: {\em all} H {\em both} $\nu_s(RR)$ and $\nu_L(RR)$ distinctly differ from unity (see Table \ref{tab1}). 
546: We now turn to SD by considering that for individuals at 
547: high risk of sudden death the fractal physiological organization (long range correlations) 
548: breaks down and this is often accompanied by emergence of {\em uncorrelated randomness} (see 
549: \cite{PNAS02} and references therein). It is therefore naturally expected that in SD the values of 
550: $\nu_s(RR)$ and $\nu_L(RR)$ become closer to the Markovian value (i.e., unity) compared to H; thus, in SD, $\nu_s(RR)$ naturally becomes smaller than the value 1.18 (the corresponding $H_{min}$-limit) and $\nu_L(RR)$ larger than 0.77 (the corresponding $H_{max}$-limit). 
551: 
552: 
553: We now focus on the following important property of H: although {\em both} 
554: $\nu_s(RR)$ and $\nu_L(RR)$ differ from unity, as mentioned, they systematically behave 
555: {\em differently}, i.e., $\nu_s(RR) > 1$ while $\nu_L(RR) < 1$. The exact origin of the 
556: latter difference has not yet been identified with certainty, but the following comments 
557: might be relevant: First, in the frame of the frequency-domain characteristics of heart 
558: rate variability (e.g., \cite{sha03}), we may state that $\nu_s(RR)$ and $\nu_L(RR)$ are 
559: associated with the high-frequency (HF, 0.15-0.4 Hz) and low-frequency (LF, 0.015-0.15 Hz) 
560: range in the RR tachogram (``instantaneous'' heart rate,1/RR). An important difference on the 
561: effect of the sympathetic and parasympathetic modulation of the RR-intervals has been 
562: noticed (e.g., see \cite{sha03} and references therein): Sympathetic tone is believed to 
563: influence the LF component whereas {\em both} sympathetic and parasympathetic activity 
564: have an effect on the HF component (recall that our results show $\nu_s(RR) > \nu_L(RR)$). 
565: Second, at short time scales (high frequencies), it has been suggested \cite{pen95} that 
566: we have relatively {\em smooth} heartbeat oscillations associated with respiration 
567: (e.g., 15 breaths per minute corresponds to a 4 sec oscillation with a peak in the power 
568: spectrum at 0.25 Hz, see \cite{sha03}); this is lost upon randomizing the consecutive 
569: intervals $Q_k$, thus probably 
570: leading to (larger variations -compared to the original experimental data- between the durations 
571: of consecutive intervals and hence to) $\delta S_{shuf}$-values larger than $\delta S$, i.e., 
572: a $\nu_s(RR)$-value larger than unity (cf. an extension of the current analysis to a surrogate sequence for a simultaneous 
573: recording of the breath rate and the instantaneous heart rate, upon considering the points  
574: discussed in paragraph 4.6 of Ref. \cite{sch00}, could greatly 
575: contribute towards clarifying the validity of such an explanation). Such an argument, if true, 
576: cannot be applied, of course, in the longer range 50-70 beats and hence explain why the opposite 
577: behavior, i.e., $\delta S_{shuf}<\delta S$, then holds. The latter finding must be inherently 
578: connected to the nature itself of the long range correlations. The existence of the latter 
579: is pointed out from the fact that (in this range also) the RR-intervals result in $\delta S$-values 
580: ($\sim10^{-3}$) which significantly differ from the Markovian $\delta S$-value ($\sim 10^{-2}$) (cf. 
581: the existence 
582: of the long range correlations in the heart rate variability has been independently established 
583: by several applications of the detrended fluctuation analysis, e.g., see \cite{pen95}, \cite{PNAS02} and references therein).
584: 
585: 
586: A simplified interpretation of the results of Fig.\ref{fig6}, and in particular the reason 
587: why for the QT intervals the quantity $\delta S$ is larger for the SD than for the H,
588:  could be attempted if we consider that: 
589: (i) $S$ could be thought as a measure of the ``disorder'' (in the consecutive intervals) (ii) the essence 
590: of the natural time-domain analysis is built on the variation of the durations of consecutive pulses, 
591: and (iii) it has been clinically observed (e.g., see Ref.\cite{p16}) that the QT interval (which corresponds to 
592: the time in which the heart in each beat ``recovers'' -electrically speaking- from the previous 
593: excitation) exhibits frequent prolonged values before cardiac death. 
594: Thus,  when a time-window is sliding on an H-ECG, it is intuitively 
595: expected to find, more or less, the same $S$-values (when sweeping through 
596: various parts of the ECG) and hence a small $\delta S$-value is envisaged. By the same token, 
597: in an SD-ECG, we expect that, in view of the short-long-short sequences
598: of the QT-intervals, the corresponding
599:  $S$-values will be much different (compared to H), thus leading to a larger $\delta S$-value
600:  (cf. in the 
601: same frame we may also understand why the $\sigma / \mu$ values -and hence $\delta S_{shuf}$, see 
602: Eq. (\ref{eq1})- are larger in SD than those in H, as shown in Fig.\ref{fig6}). 
603: The distinction between SD and H could be also
604:  understood in the context of dynamic phase-transitions (critical phenomena), as follows:
605:  In SD, since the dynamic phase transition (cardiac arrest) is approached, the {\em fluctuations}
606:  of $S$ are expected to become larger, thus reflecting larger $\delta S$; 
607: such intense fluctuations are not expected, of course, for H.
608: 
609: 
610: \section{conclusions} \label{sec6}
611: The main point emerged in the surrogate data analysis presented in this paper, is the key-role of the 
612: quantity  $\delta S_{shuf} / \delta S$. This ratio:
613: 
614: 1. reveals the non-Markovianity  in all three types of signals analyzed here, i.e., SES activities, 
615: AN and ECG. In a Markovian case  we have $\delta S_{shuf} = \delta S$, but the reverse is 
616: not always valid; it may happen that $\delta S_{shuf} / \delta S = 1$, although $\delta S$- 
617: (and $\delta S_{shuf}$-) value drastically differ from that of the Markovian (this is the case of ECG).
618: 
619: 2. differentiates the ECG of healthy humans (H) from those suffered from sudden cardiac death (SD).
620: More precisely, in SD, the $\delta S_{shuf}/\delta S$ values of the RR (i.e. beat to beat) intervals 
621: become closer to the Markovian value (i.e., unity) compared to those in H. Furthermore, 
622: in SD, {\em both} $\delta S$ -and $\delta S_{shuf}$- values of the QT interval (   corresponding to the
623: time in which the heart ``recovers'' from its previous excitation) are larger than those in H.  
624: 
625: As for the physical meaning of  $\delta S_{shuf}$ in ECG, it was shown to be a measure of $\sigma / \mu$.
626: 
627: \appendix*
628: \section{Interrelation between $\delta S_{shuf}$ and $\sigma /\mu$ in the case of iid}
629: 
630:  If we consider a time-series $Q_k$, where $Q_k \geq 0, k=1, 2,\ldots N$, we obtain
631: the quantities $p_k=Q_k/\sum_{l=1}^N Q_l$, which satisfy the necessary conditions\cite{abr70}:
632: $p_k\geq 0, \sum_{k=1}^N p_k=1$ to be considered as point probabilities. We then define\cite{var01,var02,var03b} the moments of the natural time $\chi_k=k/N$ as $\langle \chi^q \rangle= \sum_{k=1}^N (k/N)^q p_k$ and
633: the entropy  $S \equiv \langle \chi \ln \chi \rangle -\langle \chi \rangle 
634: \ln \langle \chi \rangle$, where $\langle \chi \ln \chi \rangle=\sum_{k=1}^N  (k/N) \ln (k/N) p_k$.
635: This Appendix is solely focused on a uniform distribution in the natural time-domain.
636: 
637: We now consider the case when $Q_k$ are independent and identically distributed (iid) positive random variables. It then follows that the expectation value $\text{E}(p_k)=\text{E} [Q_k/\sum_{l=1}^N Q_l]$ of $p_k$  equals $1/N$:
638: \begin{equation}
639: \text{E}(p_k)=\frac{1}{N}.
640: \label{a1}
641: \end{equation}
642: Equation (\ref{a1}) results from the fact that, since $Q_k$ are iid, we have:  
643: $\text{E}[ \sum_{k=1}^N Q_k/\sum_{l=1}^N Q_l]=1=N \text{E}(p_k)$. 
644: For the purpose
645: of our calculations the relation between the variance of $p_k$, 
646: $\text{Var}(p_k)=\text{E} [ (p_k-1/N)^2]$, and the covariance of $p_k$ and $p_l$,
647: $\text{Cov} (p_k,p_l)= \text{E}[(p_k-1/N)(p_l-1/N)]$,  is of 
648: central importance.  Using the constraint 
649: $\sum_{k=1}^N p_k=1$,  leading to $p_k-1/N=\sum_{l\neq k}(1/N-p_l)$, and the 
650: fact that $Q_k$ are iid, we obtain 
651: $\text{E} \left [ (p_k-1/N)^2\right ]=\text{E}\left [(p_k-1/N)\sum_{l\neq k}(1/N-p_l)\right ]=-(N-1)
652: \text{E}\left[ (p_k-1/N)(p_l-1/N) \right]$. Thus, we get 
653: \begin{equation}
654: \text{Cov} (p_k,p_l)=-\frac{ \text{Var}(p_k)}{N-1}.
655: \label{a2}
656: \end{equation}
657: The $N$-dependence of $\text{Var} (p_k)$ is obtained from  
658: \begin{equation}
659: \text{Var} (p_k) = \frac{1}{N^2} \text{E} \left[ \left( \frac{N Q_k}{\sum_{l=1}^NQ_l}-1 \right)^2 \right],
660: \label{a3a}
661: \end{equation}
662: where the quantity  $\text{E} [ ( N Q_k/\sum_{l=1}^NQ_l-1)^2 ]$ is {\em asymptotically} $N$-independent. 
663: The latter arises as follows: If $\text{E} (Q_k)=\mu$ 
664: and $\text{Var} (Q_k)=\sigma^2 (< \infty)$,  as a result of the central limit 
665: theorem\cite{fel71}, we have $\text{E}(\sum_{k=1}^NQ_k/N)=\mu$ and
666: $\text{Var}(\sum_{k=1}^NQ_k/N)=\sigma^2/N$. The latter two equations, for large enough $N$ imply that 
667: $\text{E} [ ( N Q_k/\sum_{l=1}^NQ_l-1)^2 ]\approx \text{E} [ ( Q_k/\mu-1)^2 ]=\sigma^2/\mu^2$.
668: Thus, Eq.(\ref{a3a}) becomes
669: \begin{equation}
670: \text{Var} (p_k) = \frac{\sigma^2}{N^2\mu^2}. 
671: \label{a3}
672: \end{equation} 
673: 
674: We now turn to the statistical properties of $\langle \chi^q \rangle$. Using Eq.(\ref{a1}),
675: we have
676: \begin{equation}
677: \text{E} [ \langle \chi^q \rangle ]=\sum_{k=1}^N \left(\frac{k}{N}\right)^q \frac{1}{N}.
678: \label{a4}
679: \end{equation}
680: which,  since\cite{gra80} $\sum_{k=1}^N k^q=N^{q+1}/(q+1)+N^q/2+o(N^q)$, reveals that
681: $\text{E} [ \langle \chi^q \rangle ]$  is again asymptotically $N$-independent because it 
682: approaches the value $1/(q+1)$ with a ``small'' $1/(2N)$ correction. 
683: The variance 
684: $\text{Var}[ \langle \chi^q \rangle ][=(\delta \langle \chi^q \rangle)^2]$
685: \begin{equation}
686: \text{Var}[ \langle \chi^q \rangle ]=\text{E} \left \{  \left[ \sum_{k=1}^N \left(\frac{k}{N}\right)^q \left( p_k-\frac{1}{N}\right ) \right]^2 \right \},
687: \label{eqvar}
688: \end{equation} 
689: after expanding the square and using Eqs.(\ref{a2}) and (\ref{a3}), becomes:
690: \begin{eqnarray}
691: \text{Var}[ &\langle \chi^q \rangle & ]= 
692: \sum_{k=1}^N\left( \frac{k}{N} \right)^{2q}\frac{\sigma^2}{N^2\mu^2} \nonumber \\
693: &-& \frac{\sigma^2}{(N-1)N^2\mu^2} \sum_{k=1}^N\left( \frac{k}{N} \right)^{q} \sum_{l=1,l\neq k}^N \left( \frac{l}{N} \right)^{q}.
694: \label{a5} 
695: \end{eqnarray}
696: which, using Eq.(\ref{a4}), finally leads to:
697: \begin{equation}
698: \text{Var}[ \langle \chi^q \rangle ]=\frac{\sigma^2}{(N-1)\mu^2} 
699: \left\{ \text{E} [ \langle \chi^{2q} \rangle ]- \text{E}^2 [ \langle \chi^{q} \rangle ]
700: \right\}.
701: \label{fd}
702: \end{equation}
703: 
704: The proof of Eq.(\ref{fd}) can be generalized for all linear functionals of $p_k$ of 
705: the form $\langle f(\chi) \rangle=\sum_{k=1}^N f(k/N) p_k$ and yields:
706: \begin{equation}
707: \text{Var}[ \langle f(\chi) \rangle ]= \frac{\sigma^2 \left\{ \text{E} [ \langle f^2(\chi) \rangle ]- \text{E}^2 [ \langle f(\chi ) \rangle ]\right\}}{(N-1)\mu^2}.
708: \label{eq8a}
709: \end{equation}
710: In Fig.(\ref{FigX}), we compare the theoretical result of Eq.(\ref{fd}) 
711: with synthetic (Gaussian) data which have values of  $\mu$, $\sigma$ 
712: and size ($\approx 1000$) similar to those in ECG. Note that when one uses 
713: the estimator $(\delta X )^2=\sum (X-\overline{X})^2/N$, instead of the unbiased 
714: estimator  $(\delta X )^2=\sum (X-\overline{X})^2/(N-1)$, in order 
715: to find the sample variance, $N$ should replace
716: $N-1$ in Eq.(\ref{fd}).
717:  
718: \begin{figure}[b]
719: \includegraphics{Figure7}% Here is how to import EPS art
720: \caption{\label{FigX} (Color) Comparison of the theoretical estimations (solid lines) of
721:  $\delta \langle \chi \rangle$ and $\delta S$ resulting from Eqs.(\ref{fd}) and (\ref{ds}), respectively, with the 
722: values obtained (plus and cross, respectively) using a Gaussian sample having values of $\mu$, $\sigma$ 
723: and size ($\approx 1000$) similar to those in ECG. Here, as well as throughout 
724: the paper, the estimator $(\delta X )^2=\sum (X-\overline{X})^2/N$ was used for the calculation 
725: of the sample variance
726: of the synthetic data, and thus
727: $N-1$ was replaced by $N$ in Eqs.(\ref{fd}) and (\ref{ds}).}
728: \end{figure}  
729: 
730: 
731: 
732: We now proceed to the statistical properties of the entropy\cite{var01,var03b} $S = \langle \chi \ln \chi \rangle -\langle \chi \rangle \ln \langle \chi \rangle$. The expectation value 
733: \begin{equation}
734: \text{E}(S)=\text{E} \left[ \sum_{k=1}^N \frac{k}{N} \ln \left( \frac{k}{N} \right) p_k -
735: \sum_{k=1}^N \frac{k}{N} p_k \ln \left( \sum_{l=1}^N \frac{l}{N} p_l \right) \right]
736: \end{equation}
737: can be evaluated as follows: we add and subtract the term $\sum_{k=1}^N \frac{k}{N} p_k \ln \left[ \sum_{l=1}^N \left( \frac{l}{N} \right)\frac{1}{N} \right]$, and then expand 
738: the resulting term $\ln [1+ \sum_{l=1}^N \frac{l}{N} (p_l-\frac{1}{N})/\sum_{l=1}^N \frac{l}{N^2}]$ to first order in $(p_l-\frac{1}{N})$; finally, using Eq.(\ref{fd}), we obtain
739: \begin{eqnarray}
740: \text{E}(S)&=&\sum_{k=1}^N \frac{k}{N^2} \ln \left( \frac{k}{N} \right) -\sum_{k=1}^N \frac{k}{N^2} \ln \left( \sum_{l=1}^N \frac{l}{N^2}  \right) \nonumber \\
741: &-& \frac{ \sigma^2 \left[ \sum_{k=1}^N k^2/N^3-(\sum_{k=1}^N k/N^2)^2\right]}{(N-1)\mu^2 \sum_{l=1}^N l/N^2}.
742: \label{as}
743: \end{eqnarray}
744: This equation reveals that $\text{E}(S)$  depends slightly on $\sigma/\mu$;
745: upon increasing $N$ the last term
746: of Eq.(\ref{as}) decays  
747: as $1/N$ (cf. the  sums in the numerator and the denominator 
748: are of the form $E [ \langle \chi^{q} \rangle ]$, for $q$=1 and 2, and 
749: asymptotically lead to a constant $1/(q+1)$, see the relevant discussion after Eq.(\ref{a4})).
750: 
751: To simplify the calculation of the variance of the entropy $\text{Var}(S)$, 
752: we  define the two linear functionals
753: \begin{equation}
754: m[x_k]=\sum_{k=1}^N \frac{k}{N} x_k,
755: \label{a11}
756: \end{equation} 
757: \begin{equation}
758: L[x_k,\xi]=\sum_{k=1}^N \frac{k}{N} \ln \left( \frac{k}{\xi N} \right) x_k,
759: \label{a12}
760: \end{equation}
761: %\begin{equation}
762: %L_2[{x_k},{y_k}]=\sum_{k=1}^N \frac{k}{N}x_k \ln \left( \sum_{l=1}^N \frac{l}{N} y_l \right),
763: %\end{equation}
764: and the constant time-series $\mathbb{K}=\{ x_k \} : x_k=1/N,\, k=1, 2, \ldots N $. Note that for 
765: both functionals $m[x_k]$ and $L[x_k,\xi]$, in view of their linearity, we have
766: \begin{equation}
767: \text{E}\left \{m[p_k-\frac{1}{N}]\right \}=\text{E}\left \{L[p_k-\frac{1}{N},\xi]\right \}=0. 
768: \label{eqa}
769: \end{equation}
770: Using Eqs.(\ref{a11}) and (\ref{a12}) the entropy can be written, in compact form, as follows:
771: \begin{equation}
772: S=L\left[ p_k,m [p_k] \right],
773: \label{s0}
774: \end{equation}
775: and its expectation value is written as
776: \begin{equation}
777: \text{E} (S)=L[\mathbb{K},1]-m[\mathbb{K}] \ln m[\mathbb{K}]-\frac{\sigma^2
778:  \kappa_{1,u}}{(N-1)\mu^2 m[\mathbb{K}]},
779: \end{equation} 
780: where $\kappa_{1,u}=\text{E} [ \langle \chi^{2} \rangle ]- \text{E}^2 [ \langle \chi \rangle ]$.
781: 
782: The variance of the entropy $\text{Var}(S)=(\delta S)^2$
783:  can then be found by adding 
784: and subtracting the terms 
785: $m[p_k] \ln m[\mathbb{K}]$ and $m[p_k-1/N]$
786: and using the expansion 
787: $m[p_k]\ln \frac{m[p_k]}{m[\mathbb{K}]}=m[p_k]m[p_k-1/N]/m[\mathbb{K}]$;
788: this gives: 
789: \begin{widetext}
790: \begin{eqnarray}
791: \text{Var}(S)&=&\text{E} \left\{ \left( L[p_k,1]-m[p_k]\ln m[p_k] -L[\mathbb{K},1]
792: +m[\mathbb{K}]\ln m[\mathbb{K}] +\frac{\sigma^2 \kappa_{1,u}}{(N-1)\mu^2 m[\mathbb{K}]} \right)^2 \right\}, 
793: \nonumber \\
794: %\text{Var}(S)
795: &=&\text{E} \left\{ \left( L[p_k-\frac{1}{N},1]-m[p_k]\ln \frac{m[p_k]}{m[\mathbb{K}]}
796: +m[\frac{1}{N}-p_k]\ln m[\mathbb{K}] +\frac{\sigma^2 \kappa_{1,u}}{(N-1)\mu^2 m[\mathbb{K}]} \right)^2 \right\},  \nonumber \\
797: %\text{Var}(S)
798: &=&\text{E} \left\{ \left( L \left [p_k-\frac{1}{N},m[\mathbb{K}] \right]-\frac{m[p_k]m[p_k-\frac{1}{N}]}{m[\mathbb{K}]} 
799: +\frac{\sigma^2 \kappa_{1,u}}{(N-1)\mu^2 m[\mathbb{K}]} \right)^2 \right\},  \nonumber \\
800: %\text{Var}(S)
801: &=&\text{E} \left\{ \left( L \left[ p_k-\frac{1}{N},m[\mathbb{K}] \right] -m[p_k-\frac{1}{N}]-\frac{m^2[p_k-\frac{1}{N}]}{m[\mathbb{K}]} 
802: +\frac{\sigma^2 \kappa_{1,u}}{(N-1)\mu^2 m[\mathbb{K}]} \right)^2 \right\},   \nonumber \\
803: %\text{Var}(S)
804: &=&\text{E} \left\{ \left( L \left[ p_k-\frac{1}{N},m[\mathbb{K}] e \right] -\frac{m^2[p_k-\frac{1}{N}]}{m[\mathbb{K}]} 
805: +\frac{\sigma^2 \kappa_{1,u}}{(N-1)\mu^2 m[\mathbb{K}]} \right)^2 \right\}. 
806: \label{a13}
807: \end{eqnarray}
808: Expanding the square in Eq.(\ref{a13}), and using Eq.(\ref{eqa}), we find  
809: \begin{eqnarray}
810: \text{Var}(S)= \text{E} \Biggl( L^2 \left [p_k-\frac{1}{N},m[\mathbb{K}] e \right ]
811: &+& 2L \left [p_k-\frac{1}{N},m[\mathbb{K}] e \right ] 
812: \frac{m^2[p_k-\frac{1}{N}]}{m[\mathbb{K}]} \nonumber \\
813: &+&\frac{m^4[p_k-\frac{1}{N}]}{m^2[\mathbb{K}]}
814: - \frac{\sigma^4 \kappa^2_{1,u}}{(N-1)^2\mu^4 m^2[\mathbb{K}]}
815: \Biggr).\label{a14} 
816: \end{eqnarray}
817: {\em If} we assume that the distribution of $Q_k$ is {\em skewnessless}, i.e. $\text{E} [(Q_k-\mu)^3]=0$, 
818: the expectation value of the second term in Eq.(\ref{a14}) vanishes, 
819: whereas the third and the fourth terms are of order $1/N^2$ and hence negligible with 
820: respect to the first term. Thus,
821: \begin{equation}
822: \text{Var}(S)=\text{E} \left(  L^2 \left[  p_k-\frac{1}{N}, m[\mathbb{K}] e \right] \right),
823: \label{eqx}
824: \end{equation}
825: which can be explicitly  written as follows 
826: \begin{equation} 
827: \text{Var}(S)=
828: \text{E} \left\{ \left[ \sum_{k=1}^N \frac{k}{N} \ln \left( \frac{k}{m[\mathbb{K}] N e} \right) 
829:  \left( p_k-\frac{1}{N} \right) 
830: \right]^2 \right\}.
831: \label{afa}
832: \end{equation}
833: The right side of Eq.(\ref{afa}) becomes similar to Eq.(\ref{eqvar}), if we replace
834: $\chi^q$ by  $\chi \ln (\chi/m[\mathbb{K}]e)$; thus 
835: after expanding the square and using Eqs.(\ref{a2}) and (\ref{a3}), we finally obtain
836: \begin{equation}
837: \text{Var}(S)=\frac{\sigma^2}{(N-1)\mu^2} \left[ 
838:  \sum_{k=1}^N  \left( \frac{k}{N} \ln  \frac{N k}{e\sum_{k=1}^N k}  \right)^2
839: \frac{1}{N}-
840: \left( \sum_{k=1}^N \frac{k}{N^2} \ln \frac{N k}{e\sum_{k=1}^N k}  \right)^2 
841: \right] .
842: \label{ds}
843: \end{equation}
844: \end{widetext}
845: A comparison of Eqs.(\ref{eqx}) and (\ref{s0}) reveals  the following: in order to find the entropy fluctuation $\delta S$,
846: one simply has to replace in Eq. (\ref{s0}) $m[p_k]$ with $m[\mathbb{K}]e$ and then
847: directly take its variance according to Eq.(\ref{eq8a}). Equation (\ref{ds}) 
848: is just Eq.(\ref{eq1}) of the main text.
849: 
850: 
851: 
852:  
853: \bibliographystyle{apsrev}
854: \begin{thebibliography}{26}
855: \expandafter\ifx\csname natexlab\endcsname\relax\def\natexlab#1{#1}\fi
856: \expandafter\ifx\csname bibnamefont\endcsname\relax
857:   \def\bibnamefont#1{#1}\fi
858: \expandafter\ifx\csname bibfnamefont\endcsname\relax
859:   \def\bibfnamefont#1{#1}\fi
860: \expandafter\ifx\csname citenamefont\endcsname\relax
861:   \def\citenamefont#1{#1}\fi
862: \expandafter\ifx\csname url\endcsname\relax
863:   \def\url#1{\texttt{#1}}\fi
864: \expandafter\ifx\csname urlprefix\endcsname\relax\def\urlprefix{URL }\fi
865: \providecommand{\bibinfo}[2]{#2}
866: \providecommand{\eprint}[2][]{\url{#2}}
867: 
868: \bibitem[{\citenamefont{Varotsos et~al.}(2001)\citenamefont{Varotsos, Sarlis,
869:   and Skordas}}]{var01}
870: \bibinfo{author}{\bibfnamefont{P.~A.} \bibnamefont{Varotsos}},
871:   \bibinfo{author}{\bibfnamefont{N.~V.} \bibnamefont{Sarlis}},
872:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{E.~S.}
873:   \bibnamefont{Skordas}}, \bibinfo{journal}{Practica of Athens Academy}
874:   \textbf{\bibinfo{volume}{76}}, \bibinfo{pages}{294} (\bibinfo{year}{2001}).
875: 
876: \bibitem[{\citenamefont{Varotsos et~al.}(2002)\citenamefont{Varotsos, Sarlis,
877:   and Skordas}}]{var02}
878: \bibinfo{author}{\bibfnamefont{P.~A.} \bibnamefont{Varotsos}},
879:   \bibinfo{author}{\bibfnamefont{N.~V.} \bibnamefont{Sarlis}},
880:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{E.~S.}
881:   \bibnamefont{Skordas}}, \bibinfo{journal}{Phys. Rev. E}
882:   \textbf{\bibinfo{volume}{66}}, \bibinfo{pages}{011902}
883:   (\bibinfo{year}{2002}).
884: 
885: \bibitem[{\citenamefont{Varotsos
886:   et~al.}(2003{\natexlab{a}})\citenamefont{Varotsos, Sarlis, and
887:   Skordas}}]{var03b}
888: \bibinfo{author}{\bibfnamefont{P.~A.} \bibnamefont{Varotsos}},
889:   \bibinfo{author}{\bibfnamefont{N.~V.} \bibnamefont{Sarlis}},
890:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{E.~S.}
891:   \bibnamefont{Skordas}}, \bibinfo{journal}{Phys. Rev. E}
892:   \textbf{\bibinfo{volume}{68}}, \bibinfo{pages}{031106}
893:   (\bibinfo{year}{2003}{\natexlab{a}}).
894: 
895: \bibitem[{\citenamefont{Varotsos
896:   et~al.}(2003{\natexlab{b}})\citenamefont{Varotsos, Sarlis, and
897:   Skordas}}]{var03}
898: \bibinfo{author}{\bibfnamefont{P.~A.} \bibnamefont{Varotsos}},
899:   \bibinfo{author}{\bibfnamefont{N.~V.} \bibnamefont{Sarlis}},
900:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{E.~S.}
901:   \bibnamefont{Skordas}}, \bibinfo{journal}{Phys. Rev. E}
902:   \textbf{\bibinfo{volume}{67}}, \bibinfo{pages}{021109}
903:   (\bibinfo{year}{2003}{\natexlab{b}}).
904: 
905: \bibitem[{\citenamefont{Schreiber and Schmitz}(2000)}]{sch00}
906: \bibinfo{author}{\bibfnamefont{T.}~\bibnamefont{Schreiber}} \bibnamefont{and}
907:   \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Schmitz}},
908:   \bibinfo{journal}{Physica D} \textbf{\bibinfo{volume}{142}},
909:   \bibinfo{pages}{346} (\bibinfo{year}{2000}).
910: 
911: \bibitem[{\citenamefont{Chang et~al.}(1995)\citenamefont{Chang, Sauer, and
912:   Schiff}}]{cha95}
913: \bibinfo{author}{\bibfnamefont{T.}~\bibnamefont{Chang}},
914:   \bibinfo{author}{\bibfnamefont{T.}~\bibnamefont{Sauer}}, \bibnamefont{and}
915:   \bibinfo{author}{\bibfnamefont{S.~J.} \bibnamefont{Schiff}},
916:   \bibinfo{journal}{Chaos} \textbf{\bibinfo{volume}{5}}, \bibinfo{pages}{376}
917:   (\bibinfo{year}{1995}).
918: 
919: \bibitem[{\citenamefont{Siwy et~al.}(2001)\citenamefont{Siwy, Mercik, Weron,
920:   and Ausloos}}]{siw01}
921: \bibinfo{author}{\bibfnamefont{Z.}~\bibnamefont{Siwy}},
922:   \bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{Mercik}},
923:   \bibinfo{author}{\bibfnamefont{K.}~\bibnamefont{Weron}}, \bibnamefont{and}
924:   \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Ausloos}},
925:   \bibinfo{journal}{Physica A} \textbf{\bibinfo{volume}{297}},
926:   \bibinfo{pages}{79} (\bibinfo{year}{2001}).
927: 
928: \bibitem[{\citenamefont{Goldberger et~al.}(2000)\citenamefont{Goldberger,
929:   Amaral, Glass, Hausdorff, Ivanov, Mark, Mictus, Moody, Peng, and
930:   Stanley}}]{GOL00}
931: \bibinfo{author}{\bibfnamefont{A.~L.} \bibnamefont{Goldberger}},
932:   \bibinfo{author}{\bibfnamefont{L.~A.~N.} \bibnamefont{Amaral}},
933:   \bibinfo{author}{\bibfnamefont{L.}~\bibnamefont{Glass}},
934:   \bibinfo{author}{\bibfnamefont{J.~M.} \bibnamefont{Hausdorff}},
935:   \bibinfo{author}{\bibfnamefont{P.~C.} \bibnamefont{Ivanov}},
936:   \bibinfo{author}{\bibfnamefont{R.~G.} \bibnamefont{Mark}},
937:   \bibinfo{author}{\bibfnamefont{J.~E.} \bibnamefont{Mictus}},
938:   \bibinfo{author}{\bibfnamefont{G.~B.} \bibnamefont{Moody}},
939:   \bibinfo{author}{\bibfnamefont{C.-K.} \bibnamefont{Peng}}, \bibnamefont{and}
940:   \bibinfo{author}{\bibfnamefont{H.~E.} \bibnamefont{Stanley}},
941:   \bibinfo{journal}{Circulation} \textbf{\bibinfo{volume}{101}},
942:   \bibinfo{pages}{e215 (see also \url{\http://www.physionet.org})}
943:   (\bibinfo{year}{2000}).
944: 
945: \bibitem[{\citenamefont{Chialvo}(2002)}]{chia02}
946: \bibinfo{author}{\bibfnamefont{D.~R.} \bibnamefont{Chialvo}},
947:   \bibinfo{journal}{Nature} \textbf{\bibinfo{volume}{419}},
948:   \bibinfo{pages}{263} (\bibinfo{year}{2002}).
949: 
950: \bibitem[{\citenamefont{Glass}(2001)}]{p2}
951: \bibinfo{author}{\bibfnamefont{L.}~\bibnamefont{Glass}},
952:   \bibinfo{journal}{Nature} \textbf{\bibinfo{volume}{410}},
953:   \bibinfo{pages}{277} (\bibinfo{year}{2001}).
954: 
955: \bibitem[{\citenamefont{Zebrowski et~al.}(2000)\citenamefont{Zebrowski,
956:   Poplawska, Baranowski, and Buchner}}]{p11}
957: \bibinfo{author}{\bibfnamefont{J.~J.} \bibnamefont{Zebrowski}},
958:   \bibinfo{author}{\bibfnamefont{W.}~\bibnamefont{Poplawska}},
959:   \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Baranowski}},
960:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{T.}~\bibnamefont{Buchner}},
961:   \bibinfo{journal}{Chaos, Solitons and Fractals}
962:   \textbf{\bibinfo{volume}{11}}, \bibinfo{pages}{1061} (\bibinfo{year}{2000}).
963: 
964: \bibitem[{EPA({\natexlab{a}})}]{EPAPS03}
965: \eprint{See EPAPS Document No. E-PLEEE8-68-116309 for additional information.
966:   This document may be retrieved via the EPAPS homepage
967:   (\url{http://www.aip.org/pubservs/epaps.html}) or from \url{ftp.aip.org} in
968:   the directory /epaps/. See the EPAPS homepage for more information.}
969: 
970: \bibitem[{\citenamefont{Laguna et~al.}(1997)\citenamefont{Laguna, Mark,
971:   Goldberger, and Moody}}]{LAG97}
972: \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Laguna}},
973:   \bibinfo{author}{\bibfnamefont{R.~G.} \bibnamefont{Mark}},
974:   \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Goldberger}},
975:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{G.~B.} \bibnamefont{Moody}},
976:   in \emph{\bibinfo{booktitle}{Computers in Cardiology}}
977:   (\bibinfo{publisher}{IEEE Computer Society Press},
978:   \bibinfo{address}{Piscataway, NJ}, \bibinfo{year}{1997}),
979:   vol.~\bibinfo{volume}{24}, p. \bibinfo{pages}{673}.
980: 
981: \bibitem[{\citenamefont{Motulsky}(1995)}]{mot95}
982: \bibinfo{author}{\bibfnamefont{H.}~\bibnamefont{Motulsky}},
983:   \emph{\bibinfo{title}{Intuitive Biostatistics}} (\bibinfo{publisher}{Oxford
984:   University Press}, \bibinfo{address}{New York}, \bibinfo{year}{1995}).
985: 
986: \bibitem[{\citenamefont{Goldberger et~al.}(2002)\citenamefont{Goldberger,
987:   Amaral, Hausdorff, Ivanov, Peng, and Stanley}}]{PNAS02}
988: \bibinfo{author}{\bibfnamefont{A.~L.} \bibnamefont{Goldberger}},
989:   \bibinfo{author}{\bibfnamefont{L.~A.~N.} \bibnamefont{Amaral}},
990:   \bibinfo{author}{\bibfnamefont{J.~M.} \bibnamefont{Hausdorff}},
991:   \bibinfo{author}{\bibfnamefont{P.~C.} \bibnamefont{Ivanov}},
992:   \bibinfo{author}{\bibfnamefont{C.-K.} \bibnamefont{Peng}}, \bibnamefont{and}
993:   \bibinfo{author}{\bibfnamefont{H.~E.} \bibnamefont{Stanley}},
994:   \bibinfo{journal}{Proc. Natl. Acad. Sci. USA} \textbf{\bibinfo{volume}{99}},
995:   \bibinfo{pages}{2467} (\bibinfo{year}{2002}).
996: 
997: \bibitem[{\citenamefont{Jan\'{e} et~al.}(1997)\citenamefont{Jan\'{e}, Blasi,
998:   J.~Garcia, and Laguna}}]{jan97}
999: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Jan\'{e}}},
1000:   \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Blasi}},
1001:   \bibinfo{author}{\bibfnamefont{J.}~\bibnamefont{J.~Garcia}},
1002:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Laguna}},
1003:   in \emph{\bibinfo{booktitle}{Computers in Cardiology}}
1004:   (\bibinfo{publisher}{IEEE Computer Society Press},
1005:   \bibinfo{address}{Piscataway, NJ}, \bibinfo{year}{1997}),
1006:   vol.~\bibinfo{volume}{24}, p. \bibinfo{pages}{295}.
1007: 
1008: \bibitem[{\citenamefont{Laguna et~al.}(1990)\citenamefont{Laguna, Thakor,
1009:   Caminal, Jan\'{e}, and Yoon}}]{lag90}
1010: \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Laguna}},
1011:   \bibinfo{author}{\bibfnamefont{N.~V.} \bibnamefont{Thakor}},
1012:   \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Caminal}},
1013:   \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Jan\'{e}}}, \bibnamefont{and}
1014:   \bibinfo{author}{\bibfnamefont{H.~R.} \bibnamefont{Yoon}},
1015:   \bibinfo{journal}{Medical \& Biological Engineering \& Computing}
1016:   \textbf{\bibinfo{volume}{28}}, \bibinfo{pages}{67} (\bibinfo{year}{1990}).
1017: 
1018: \bibitem[{\citenamefont{Laguna et~al.}(1994)\citenamefont{Laguna, Jan\'{e}, and
1019:   Caminal}}]{lag94}
1020: \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Laguna}},
1021:   \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Jan\'{e}}}, \bibnamefont{and}
1022:   \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Caminal}},
1023:   \bibinfo{journal}{Computers and Biomedical Research}
1024:   \textbf{\bibinfo{volume}{27}}, \bibinfo{pages}{45} (\bibinfo{year}{1994}).
1025: 
1026: \bibitem[{\citenamefont{Ivanov et~al.}(2001)\citenamefont{Ivanov, Amaral,
1027:   Goldberger, Havlin, Rosenblum, Stanley, and Struzik}}]{iva01}
1028: \bibinfo{author}{\bibfnamefont{P.~C.} \bibnamefont{Ivanov}},
1029:   \bibinfo{author}{\bibfnamefont{L.~A.~N.} \bibnamefont{Amaral}},
1030:   \bibinfo{author}{\bibfnamefont{A.~L.} \bibnamefont{Goldberger}},
1031:   \bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{Havlin}},
1032:   \bibinfo{author}{\bibfnamefont{M.~G.} \bibnamefont{Rosenblum}},
1033:   \bibinfo{author}{\bibfnamefont{H.~E.} \bibnamefont{Stanley}},
1034:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{Z.~R.}
1035:   \bibnamefont{Struzik}}, \bibinfo{journal}{Chaos}
1036:   \textbf{\bibinfo{volume}{11}}, \bibinfo{pages}{641} (\bibinfo{year}{2001}).
1037: 
1038: \bibitem[{\citenamefont{McSharry et~al.}(2003)\citenamefont{McSharry, Clifford,
1039:   Tarassenko, and Smith}}]{sha03}
1040: \bibinfo{author}{\bibfnamefont{P.~E.} \bibnamefont{McSharry}},
1041:   \bibinfo{author}{\bibfnamefont{G.~D.} \bibnamefont{Clifford}},
1042:   \bibinfo{author}{\bibfnamefont{L.}~\bibnamefont{Tarassenko}},
1043:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{L.~A.} \bibnamefont{Smith}},
1044:   \bibinfo{journal}{IEEE Trans. Biomed. Eng.} \textbf{\bibinfo{volume}{550}},
1045:   \bibinfo{pages}{289} (\bibinfo{year}{2003}).
1046: 
1047: \bibitem[{\citenamefont{Peng et~al.}(1995)\citenamefont{Peng, Havlin, and
1048:   Stanley}}]{pen95}
1049: \bibinfo{author}{\bibfnamefont{C.-K.} \bibnamefont{Peng}},
1050:   \bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{Havlin}}, \bibnamefont{and}
1051:   \bibinfo{author}{\bibfnamefont{H.~E.} \bibnamefont{Stanley}},
1052:   \bibinfo{journal}{Chaos} \textbf{\bibinfo{volume}{5}}, \bibinfo{pages}{82}
1053:   (\bibinfo{year}{1995}).
1054: 
1055: \bibitem[{\citenamefont{Khan}(2002)}]{p16}
1056: \bibinfo{author}{\bibfnamefont{I.~A.} \bibnamefont{Khan}},
1057:   \bibinfo{journal}{Am. Heart J.} \textbf{\bibinfo{volume}{143}},
1058:   \bibinfo{pages}{7} (\bibinfo{year}{2002}).
1059: 
1060: \bibitem[{\citenamefont{Abramowitz and Stegun}(1970)}]{abr70}
1061: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Abramowitz}} \bibnamefont{and}
1062:   \bibinfo{author}{\bibfnamefont{I.~A.} \bibnamefont{Stegun}},
1063:   \emph{\bibinfo{title}{Handbook of Mathematical Functions}}
1064:   (\bibinfo{publisher}{Dover}, \bibinfo{address}{New York},
1065:   \bibinfo{year}{1970}).
1066: 
1067: \bibitem[{\citenamefont{Feller}(1971)}]{fel71}
1068: \bibinfo{author}{\bibfnamefont{W.}~\bibnamefont{Feller}},
1069:   \emph{\bibinfo{title}{An Introduction to Probability Theory and Its
1070:   Applications, Vol. II}} (\bibinfo{publisher}{Wiley}, \bibinfo{address}{New
1071:   York}, \bibinfo{year}{1971}).
1072: 
1073: \bibitem[{\citenamefont{Gradsteyn and Ryzhik}(1980)}]{gra80}
1074: \bibinfo{author}{\bibfnamefont{I.~S.} \bibnamefont{Gradsteyn}}
1075:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{I.~M.}
1076:   \bibnamefont{Ryzhik}}, \emph{\bibinfo{title}{Table of Integrals, Series and
1077:   Products}} (\bibinfo{publisher}{Academic Press}, \bibinfo{address}{San
1078:   Diego}, \bibinfo{year}{1980}).
1079: 
1080: \bibitem[{EPA({\natexlab{b}})}]{EPAPS}
1081: \eprint{See EPAPS Document No. E-PLEEE8-69-107405 for additional information. A direct link 
1082: to this document may be mfound in the online article's HTML reference section. The document
1083: may also be readed via via the EPAPS homepage
1084:   (\url{http://www.aip.org/pubservs/epaps.html}) or from \url{ftp.aip.org} in
1085:   the directory /epaps/. See the EPAPS homepage for more information. }
1086: 
1087: \end{thebibliography}
1088: 
1089: 
1090: 
1091: \end{document}
1092: 
1093: 
1094: 
1095: 
1096: 
1097: 
1098: 
1099: 
1100: 
1101: 
1102: 
1103: 
1104: