1: %

2: %   SYNCHRONIZATION IN SMALLWORLD SYSTEMS

3: %

4: 

5: \documentclass[aps,prl,twocolumn,showpacs,superscriptaddress]{revtex4}

6: \usepackage[dvips]{graphicx}

7: \usepackage{amssymb}

8: %\usepackage[active]{srcltx}

9: \begin{document}

10: 

11: %\draft

12: \title{Synchronization in Small--world Systems}

13: 

14: \author{Mauricio Barahona}~\altaffiliation{Present address:

15: Dept.\ of Bioengineering, Imperial College, London SW7 2BX,

16: UK.}~\affiliation{Control and Dynamical Systems, California

17: Institute of Technology, Pasadena, CA 91125}

18: 

19: \author{Louis M.\ Pecora}

20: \affiliation{Naval Research Laboratory, Code 6340,

21:  Washington, DC 20375}

22: 

23: \date{\today}

24: 

25: 

26: \begin{abstract}

27: We quantify the dynamical implications of the small-world

28: phenomenon. We consider the generic synchronization of oscillator

29: networks of arbitrary topology, and link the linear stability of

30: the synchronous state to an algebraic condition of the Laplacian

31: of the graph. We show numerically that the addition of random

32: shortcuts produces improved network synchronizability. Further, we

33: use a perturbation analysis to place the synchronization threshold

34: in relation to the boundaries of the small-world region. Our

35: results also show that small-worlds synchronize as efficiently as

36: random graphs and hypercubes, and more so than standard

37: constructive graphs.

38: \end{abstract}

39: 

40: \pacs{}

41: 

42: \maketitle

43: 

44: 

45: Recently, Watts and Strogatz~\cite{refWS} showed that the addition

46: of a few long-range shortcuts to an otherwise locally connected

47: lattice (the "pristine world") produces a sharp reduction of the

48: average distance between arbitrary nodes. The ensuing semi-random

49: lattice was denoted a \textit{small-world} (SW) because the sudden

50: appearance of short paths occurs early on, while the system is

51: still relatively localized. This concept has wide appeal: the SW

52: property seems to be a quantifiable characteristic of many

53: real-world

54: structures~\cite{refWS,refStevelong,refDuncanbook,refbiochem},

55: both human generated (social networks, WWW, power grid), or of

56: biological origin (neural and biochemical networks).

57: 

58: A spur of ongoing research~\cite{refStevelong} has concentrated on

59: static and combinatoric

60: properties~\cite{refBA,refBW,refNMW,refNW,refMM,refKAS} of a

61: tractable SW model~\cite{refWS,refRM}. Monasson~\cite{refRM}

62: considered the SW effect on the distribution of eigenvalues of the

63: connectivity matrix (the graph Laplacian) which specifies the

64: coupling between nodes---a relevant topic for polymer

65: networks~\cite{refpolymer}. However, despite their central role in

66: real-world networks, there are fewer studies of dynamical

67: processes taking place on SW lattices. Among those, automata

68: epidemics simulations~\cite{refepidemics} and Web-browsing

69: studies~\cite{refkleinberg} have revealed the importance of

70: shortcuts. Numerical work on synchronization of Kuramoto

71: oscillators~\cite{refDuncanbook}, discrete maps~\cite{refGH} and

72: Hodgkin-Huxley neurons~\cite{refLHC} has shown improved SW

73: synchronizability, as intuitively expected. However, these

74: numerical examples are not generic, and fail to provide insight

75: into how the SW property influences the dynamics.

76: 

77: 

78: In this paper, we explicitly link the SW addition of random

79: shortcuts to the synchronization of networks of coupled dynamical

80: systems. This is an example of dynamics {\it on}

81: networks---leaving aside the distinct problem of evolution {\it

82: of} networks here. By using a generic synchronization

83: formulation~\cite{refPC,refFJM} to factor out the connectivity of

84: the network, we identify the synchronization threshold with an

85: algebraic condition of the graph Laplacian. Through numerics and

86: analysis, we quantify how the SW scheme improves the

87: synchronizability of the pristine world, mainly as a result of the

88: steep increase of the first-non-zero eigenvalue (FNZE). The

89: synchronization threshold is found to lie in the SW

90: region~\cite{refepidemics,refDuncanbook}, but does not coincide

91: with its onset---it can in fact be linked to the effective

92: randomization that ends SW. Within this framework, we show that

93: the synchronization efficiency of semi-random SW networks is

94: higher than standard deterministic graphs, and comparable to both

95: fully random and ideal constructive graphs.

96: 

97: 

98: \begin{figure}

99: \vspace*{-.4in}

100: \includegraphics[width=3.5in]{finfig1.eps}

101: \vspace*{-2.2in} \caption{Four typical master stability functions

102: (scaled for clearer visualization) for R\"ossler systems: chaotic

103: (bold) and periodic (regular lines); with $y$-coupling (dashed)

104: and $x$-coupling (solid lines). Here we consider the $x$-coupled

105: chaotic case (solid bold) with a negative well between

106: $(\alpha_1,\alpha_2)$.

107: %Note how only the chaotic

108: %$y$ coupled systems remain

109: %in stable synchronization at large coupling.

110: \label{fig:masterfunction}}

111: \end{figure}

112: 

113: Consider $n$ identical dynamical systems (placed at the nodes of a

114: graph) that are linearly and symmetrically coupled (as represented

115: by the edges of the undirected graph) with global coupling

116: strength $\sigma$. The topology of the graph can be encoded in the

117: Laplacian matrix $G$, a symmetric matrix with zero row-sum and

118: real spectrum $\{\theta_k\}, \; k=0,1,\ldots, n-1$. A general

119: linear stability criterion for the synchronized state of the

120: system~\cite{refPC,refFJM} is given by the negativity of the

121: master stability function $\lambda_{\rm{max}} (\sigma \, \theta_k)

122: < 0,\; \forall k$. This $\lambda_{\rm{max}}$ is a characteristic

123: of the particular dynamics at the nodes but, crucially, a large

124: class of oscillatory systems (chaotic, periodic and quasiperiodic)

125: have master stability functions with generic

126: features~\cite{refFJM}. In particular, for several chaotic systems

127: $\lambda_{\rm{max}}$ has a single deep well, as depicted in

128: Fig.~\ref{fig:masterfunction}. (We remark that this analysis is

129: quite general: it can be extended~\cite{refFJM} to eliminate the

130: zero row-sum constraint, and to comprise nonlinear coupling and

131: more general synchronization criteria.) Stability is thus ensured

132: by tuning the coupling $\sigma$ to try and place the entire

133: spectrum of transverse eigenvalues (times $\sigma$) in the deep,

134: stable region: $\sigma \, \theta_k \in (\alpha_1,\alpha_2)$. This

135: leads to an algebraic condition for the existence of a linearly

136: stable synchronous state: a network is synchronizable if

137: \begin{equation}

138: \theta_{\rm{max}}/\theta_1 < \alpha_2/\alpha_1 \equiv \beta,

139: \label{eq:critcond}

140: \end{equation}

141: where $\theta_1$ is the FNZE and $\theta_{\rm{max}}$ is the

142: maximum eigenvalue of the Laplacian $G$. The figure of merit

143: ($\beta$) ranges from 5 to 100 for a variety of oscillators (e.g.,

144: Lorenz, R\"ossler, double scroll).

145: 

146: Small-worlds are generated from a pristine world: a $k$-cycle of

147: $n$ nodes and range $k$, each node coupled to its $2 \, k$ nearest

148: neighbors for a total of $n k$ edges~\cite{refRM}. The Laplacian

149: of this graph $G^0$ is a banded circulant matrix with non-zero

150: elements on the main diagonal and the $2 \,k$ adjacent diagonals:

151: $G^0_{ii}=2 k$ and $G^0_{ij}=G^0_{ji}=-1$ with $(i+1)\bmod n \leq

152: j \leq (i+k) \bmod n$, and $1 \leq i \leq n$. The SW scheme dopes

153: the pristine world by adding $n s$ edges picked at random from the

154: $n (n-2k-1)/2$ remaining pairs. Each new edge between nodes $l$

155: and $ m$ adds off-diagonal $\Delta G_{lm}=\Delta G_{ml}=-1$ and

156: on-diagonal $\Delta G_{ll}=\Delta G_{mm}=1$ contributions to the

157: Laplacian, thus preserving the null row-sum and the bidirectional

158: coupling. The average number of shortcuts per node ($s$) is

159: related to other measures of randomness ($p$ and $q$) used

160: previously~\cite{refWS,refRM}: $s \equiv kp \equiv q

161: (n-2k-1)/(2n)$.

162: 

163: \begin{figure}

164: \includegraphics[width=3.3in]{finfig2.eps}

165: \caption{Synchronizability thresholds $s_{\rm{sync}}(\circ)$ for

166: graphs with $n$ nodes ($n=300, 400, 500, 1000$) and range $k \in

167: [1,70]$, numerically averaged over 1000 realizations. The solid

168: lines correspond to the analytical Eq.~(\ref{eq:analyticapprox}),

169: valid in the range $n^{1/3} < k < k_{\min}(n)$. For most

170: parameters, $s_{\rm{sync}}$ lies within the small-world region

171: between the dashed lines ($ s_L < s < s_C $) depicted here for

172: $n=1000$, but it is distinct from the SW onset $s_L$. Note how

173: synchronization is achievable without random shortcuts by

174: increasing the deterministic range up to $k_{\rm{min}}(n)$ (see

175: Fig.~{\ref{fig:semirandom}}). Inset: decay of the average distance

176: $L$, clustering $C$, and eigenratio (squares) as shortcuts are

177: added to a pristine world of range $k=20$ and $n=500$. We define

178: $s_L$ and $s_C$ as the points where $L$ and $C$ are 75\% of the

179: pristine world value; $s_{\rm{sync}}$ is the point where the

180: eigenratio $\theta_{\rm{max}}/\theta_1 = \beta \equiv 37.85$. }

181: \label{fig:threshold}

182: \end{figure}

183: 

184: The numerical results in Fig.~\ref{fig:threshold} illustrate the

185: SW effect on the synchronization of networks of different size and

186: range. For concreteness, all our numerics have been performed for

187: a network of identical $x$-coupled R\"ossler chaotic oscillators

188: with $\beta \simeq 37.85$. Similarly to other locally connected

189: networks, pristine worlds have a large eigenratio

190: $\theta_{\rm{max}}/\theta_1$ (i.e., they are difficult to

191: synchronize). However, as $s$ is increased the eigenratio falls

192: sharply until, at a value $s_{\rm{sync}}$, the

193: condition~(\ref{eq:critcond}) is reached (i.e., the addition of

194: shortcuts makes it synchronizable). The dependence of

195: $s_{\rm{sync}}$ on the network parameters $\{n,k\}$ is notably

196: complicated. First, there appears to be an optimal range $k \simeq

197: 4$ for which the SW is most efficient. Moreover, the

198: synchronization threshold $s_{\rm{sync}}$ lies in the small-world

199: region but does not seem to coincide with its onset. The SW onset

200: ($s_L$) is defined~\cite{refWS,refDuncanbook} by the decay of the

201: average graph distance~\cite{refNMW} $L(s) \simeq (n/k) \, f(n

202: s)$, where $f(x)=\arg \tanh \left (x/\sqrt{x^2+2x} \right ) /

203: \sqrt{4(x^2+2 x)}$. Fixing $L(s_L)/L(0) = 3/4$, we obtain $s_L

204: \simeq 1.061/n, \: n k \gg 1$. The end of the SW region ($s_C$)

205: corresponds to the effective graph randomization through the loss

206: of transitivity~\cite{refWS,refNSW,refDuncanbook}, as given by the

207: decay of the clustering coefficient

208: $C$~\cite{refBW,refMarkpersonal}: $C(s)/C(0) \simeq (2k-1)/(2k

209: (1+s/k)^2 -1), \: n \gg 1$. Again, fixing $C(s_C)/C(0) = 3/4$, we

210: obtain $s_C \simeq k \left (-1+ [(8k-1)/(6k-3)]^{-1/2} \right )

211: \simeq 0.155 \; k$. The synchronization threshold generally lies

212: between these two boundaries which scale differently with $n$ and

213: $k$: $s_L \simeq 1.061/n < s_{\rm{sync}} < s_{C} \simeq 0.155 \;

214: k$.

215: 

216: We can gain insight into the synchronization threshold through an

217: analytical perturbation of the eigenratio of the SW Laplacian $G=

218: G^0 + G^r$. Here, $G^0$ is the deterministic Laplacian of the

219: pristine world, and $G^r$ is the stochastic Laplacian for the

220: random shortcuts: $G^r_{ij}=G^r_{ji}=-\xi_{ij}$ (for $i+k+1 \leq j

221: \leq \min \{n,n-k+i-1\}$ with $1 \leq i \leq n-k+1$); $G^r_{ij}=0$

222: (otherwise); and $G^r_{ii}=-\sum_{j=1}^n G^r_{ij}$ (for $1 \leq i

223: \leq n$). The $\xi_{ij}$ are $n(n-2k-1)/2 \,$ i.i.d. Bernoulli

224: random variables which take the value 1 with probability $q/n

225: \equiv 2 s/(n-2k-1)$ (and the value 0 with probability $1-q/n$).

226: %This preserves the Laplacian structure.

227: The circulant $G^0$ is Fourier-diagonalizable~\cite{refRM} with

228: spectrum $\theta^0_j = (2k+1)- \sin [(2k+1) \pi j/n] /\sin [\pi

229: j/n], \,\, 1 \leq j \leq n-1$, (plus $\theta^0_0=0$ of any

230: Laplacian). The FNZE and the maximum eigenvalue of the unperturbed

231: lattice are:

232: \begin{eqnarray}

233: \theta^0_1 &\simeq& 2 \pi^2 k (k+1) (2k+1)/(3 n^2), \;\; k \ll n

234: \label{eq:fnze0}\\

235: \theta^0_{\rm{max}} &\simeq& (2k+1) +

236: \csc \left[\frac{3 \pi/2}{2k+1} \right] \label{eq:emax0_1} \\

237: &\simeq& (2k+1) [1+2/3 \pi], \;\; k \gg 1, \label{eq:emax0}

238: \end{eqnarray}

239: where (\ref{eq:emax0_1}) follows from a continuum approximation.

240: 

241: 

242: Following an ``honest'' treatment~\cite{refBoyce} with $G^r$ as

243: the perturbation, we treat the analytical expressions of the

244: doubly degenerate eigenvalues as random variables to obtain their

245: expectations. We postpone the detailed

246: calculations~\cite{refuslong} and sketch here the main results.

247: After some stochastic calculus, the expectations of the

248: eigenvalues of the SW Laplacian to second order are shown to be:

249: \begin{eqnarray}

250: {\mathcal E}\theta^{(1)}_i &\simeq& q \pm \sqrt{3 \pi q/4 n}

251: \label{eq:expectation1}\\

252: {\mathcal E}\theta^{(2)}_i &\simeq& \frac{2 q}{n}

253: {{\sum_{m=1}^n}^\prime (\theta^0_i-\theta^0_m})^{-1},

254: \label{eq:expectation2}

255: %%%%

256: %%%  in one line...

257: %{\mathcal E}\theta_i \simeq \theta_i^0+ \left( q \pm \sqrt{\frac{3

258: %\pi q}{4 n}} \right ) + {{\sum_{m=1}^n}^\prime \frac{ 2

259: %q/n}{\theta^0_i-\theta^0_m}}, \label{eq:expectation2}

260: \end{eqnarray}

261: for $q/n$ and $k/n$ small. To improve the accuracy of

262: $s_{\rm{sync}}$, we have also obtained an approximation

263: to~(\ref{eq:expectation2}) for FNZE:

264: \begin{eqnarray}

265: {\mathcal E}\theta^{(2)}_1 \simeq \frac{-2 q}{K^3} \left [ \frac{9

266: n}{\pi^2} + K^2 - \left( \frac{7}{5}+\frac{6}{\pi^2}\right) K -

267: \frac{2}{\pi} \right ] , \label{eq:approx22}

268: \end{eqnarray}

269: where $K=2 k +1$. Eqns.

270: (\ref{eq:critcond}),(\ref{eq:fnze0}),(\ref{eq:emax0}),(\ref{eq:expectation1}),

271: and (\ref{eq:approx22}) are then used to obtain an estimate of

272: $s_{\rm{sync}}$ as the solution of an algebraic equation involving

273: only $n$ and $k$:

274: \begin{equation}

275: \theta^0_{\rm{max}} + {\mathcal E}\theta^{(1)}_{\rm{max}} = \beta

276: \left( \theta^0_1 +{\mathcal E}\theta^{(1)}_1 + {\mathcal

277: E}\theta^{(2)}_1 \right). \label{eq:analyticapprox}

278: \end{equation}

279: As shown in Fig.~\ref{fig:threshold}, this approximates well our

280: numerics for $n^{1/3} < k \ll n$,  where the

281: Rayleigh-Schr\"odinger perturbation expansion is valid.

282: 

283: 

284: Using~(\ref{eq:analyticapprox}), we can obtain~\cite{refuslong} a

285: first order estimate for the \textit{maximum} of the

286: synchronization threshold $s_{\rm{sync}}^*$ in the valid range.

287: The maximum occurs at $k^* \simeq n \sqrt{(1+2/3 \pi)/2 \pi^2

288: \beta}$ with the asymptotic value $s_{\rm{sync}}^* \simeq (2+4/3

289: \pi) (1-2 k^*/n)/3 (\beta -1)$. Therefore, $s_{\rm{sync}} < s_C \;

290: [2 (1+2/3 \pi)/(2 \surd{3} -3) (\beta -1)]$ is bounded by the end

291: of the SW region but linked to it. For $k<n^{1/3}$, the eigenvalue

292: bunching (quasi-degeneracy) in the pristine lattice renders the

293: doubly degenerate perturbation invalid. We are developing another

294: approximation to quantify the behavior in this limit, but our

295: numerics~\cite{refuslong} indicate that the dependence of

296: $s_{\rm{sync}}$ with $n$ is sub-logarithmic. This confirms that

297: the synchronizability is most effectively improved for small-range

298: networks (Fig.~\ref{fig:threshold}).

299: 

300: \begin{figure}

301: \includegraphics[width=3.4in]{finfig3.eps}

302: \caption{Eigenratio decay in a $n=100$ lattice as $f$ edges are

303: added following purely deterministic, semi-random (SW), and purely

304: random schemes. Networks become synchronizable below the dashed

305: horizontal line ($\beta$). The squares (numerical) and the solid

306: line (Eq.~(\ref{eq:synchropristine})) show the decrease of the

307: eigenratio of pristine worlds ($k$-cycles) through the

308: deterministic addition of short-range connections---for $n=100$,

309: networks with $k \geq 7$ are synchronizable. The semi-random SW

310: approach (dots, shown for $k=1,2,4,6,10,14$) is more efficient in

311: producing synchronization. The dot-dashed line corresponds to

312: purely random graphs (RG, Eq.~(\ref{eq:synchrorandom})), which

313: become \textit{almost surely} disconnected at $f \simeq 2 \ln

314: n/(n+2 \ln n) = 0.0843$ (thus, with $\theta^{RG}_1=0$ and

315: unsynchronizable). The merging of the SW and RG behaviors as $f

316: \to 1$ is the dynamical analogue of the effective randomization

317: that leads to $s_C$.} \label{fig:semirandom}

318: \end{figure}

319: 

320: How efficient is the addition of random shortcuts for

321: synchronization? We have compared the semi-random SW approach with

322: purely random and purely deterministic schemes. An example of the

323: latter is the synchronization of pristine worlds through the

324: increase of the range $k$. From (\ref{eq:fnze0}) and

325: (\ref{eq:emax0}), the eigenratio of a pristine lattice is

326: \begin{equation}

327: \frac{\theta^0_{\rm{max}}}{\theta^0_{1}}

328: \simeq \frac{3 \pi+2}{2 \pi^3} \, \frac {n^2}{ k (k+1)}.

329: \label{eq:synchropristine}

330: \end{equation}

331: Therefore, $n$ nodes can be synchronized in a $k$-cycle if $k

332: > k_{\rm{min}} \simeq n \sqrt{(3 \pi +2)/2 \pi^3 \beta}$.

333: (Note in Fig.~\ref{fig:threshold} the consistency of our

334: analytical approximation (\ref{eq:analyticapprox}):

335: $s_{\rm{sync}}=0$ precisely at $k_{\rm{min}}$.) For purely random

336: graphs ${\mathcal G}_{n,f}$~\cite{refmohar},

337: \begin{equation}

338: \frac{\theta^{RG}_{\rm{max}}}{\theta^{RG}_1} \simeq

339: \frac{nf-\sqrt{2f(1-f)n \ln n}}{nf+\sqrt{2f(1-f)n \ln n}}\, ,

340: \label{eq:synchrorandom}

341: \end{equation}

342: where $f$ is the number of edges measured as a fraction of the

343: complete graph. These are compared with SW graphs in

344: Fig.~\ref{fig:semirandom}.

345: 

346: We remark on several observations regarding

347: Fig.~\ref{fig:semirandom}. First, the SW addition of shortcuts is

348: more efficient than the deterministic addition of short-range

349: layers. Second, the effective randomization of SW lattices with

350: edge addition translates into converging synchronization behaviors

351: of random and SW networks at large $f$. The $f \to 1$ region is

352: thus \textit{robustly} stable: cutting connections from the fully

353: connected graph has very little effect on synchronization

354: stability---not until over 90\% are cut (for $k$ small) does the

355: eigenratio begin to change drastically. Moreover, if we interpret

356: the number of edges needed to synchronize $n$ nodes as a simple

357: measure of ``cost'', adding many connections buys little extra

358: stability beyond the small-world regime. Finally, it can be

359: shown~\cite{refuslong} that the general trends of the eigenratio

360: in Fig.~\ref{fig:semirandom} (namely, ``hyperbolic'' dependence

361: for $f$ small, and near independence for $f$ large) can be

362: predicted with the naive perturbation result that both $\theta_1$

363: and $\theta_{\rm max}$ change linearly with the number of added

364: connections.

365: 

366: \begin{figure}

367: \includegraphics[width=3.4in]{finfig4.eps}

368: \caption{``Cost'' of synchronization measured as the number of

369: edges needed to synchronize a lattice of $n$ chaotic R\"ossler

370: systems arranged in different topologies: deterministic graphs

371: ($k$-wheels, $k$-cycles, $k$-M\"obius ladders, bipartites,

372: hypercubes), random graphs, and small-worlds ($\blacklozenge$).

373: Small-worlds scale favorably compared to deterministic structures

374: (and comparably to the ideal and largely unrealizable hypercubes).

375: Also, SW graphs with small range $k$ are as cost-efficient as

376: random graphs but demand less (algorithmic) storage memory.}

377: \label{fig:design}

378: \end{figure}

379: 

380: 

381: We have also compared the synchronization cost (in edges) for SW

382: systems and regular (constructive) lattices

383: (Fig.~\ref{fig:design}). For $x$-R\"ossler systems in a $k$-cycle,

384: this cost tends to $f=2 k_{\rm{min}}/(n-1) \simeq 0.140$. Other

385: constructive lattices~\cite{refSchwenk,refuslong} also tend to

386: constant fractions: $f=0.252$ (for $k$-wheels), $f=0.070$ (for

387: $k$-M\"obius ladders), and $f=0.053$ (for the most economical

388: bipartite graph). In all those cases, the cost of synchronization

389: is high: the necessary number of edges scales like $\sim n^2$,

390: just like the complete graph. At the other end of deterministic

391: graphs lies the quasi-optimal (though virtually unrealizable)

392: hypercube, which is always synchronizable with a number of edges

393: $f \sim \log_2 n /n$. This behavior is similar to that of random

394: graphs: from Eq.~(\ref{eq:synchrorandom}) {\it almost sure}

395: synchronization of ${\mathcal G}_{n,f}$ is asymptotically achieved

396: when $f \sim \ln n/n$. Remarkably, Fig.~\ref{fig:design} shows

397: that the cost of synchronizing small-$k$ SW networks is low, i.e.,

398: comparable to that of random graphs and

399: hypercubes~\cite{refChungmatching}.

400: 

401: 

402: These results hint at research that could deepen the connections

403: between topology and dynamics on networks. With a view to improved

404: design, the dynamic eigenratio criterion can be related to other

405: graph-theoretical properties (e.g., connectivity, diameter, and

406: convergence of Markov chains)~\cite{refChung}. Moreover, other

407: measures of cost (e.g., robustness under edge deletion) should be

408: considered as possible design constraints. Finally, recent results

409: could lead to extensions of this work to incorporate more general

410: concepts of stability~\cite{refpablo}, and broader definitions of

411: small-world lattices~\cite{refNSW}.

412: 

413: We thank Steve Strogatz for his deep and insightful involvement in

414: this work, and Mark Newman for sharing computer code and

415: unpublished results.

416: 

417: \bibliography{SW4}

418: 

419: 

420: \end{document}

421: