1: \documentclass[article]{revtex4}
2: %\setlength{\textwidth}{27pc}
3: %\setlength{\textheight}{43pc}
4: \usepackage{graphicx}
5: % Include figure files \usepackage{dcolumn}
6: %Align table columns on decimal point \usepackage{bm}
7: % bold math
8: %\documentstyle[prl,aps,epsf]{revtex}
9:
10: \begin{document}
11: \title{Phase Transition in the Aldous-Shields Model of Growing Trees}
12:
13: \author{David S. Dean $^{1}$ and Satya N. Majumdar $^{2}$}
14: \address{
15: {\small $^1$ Laboratoire de Physique Th\'eorique, UMR CNRS 5152, IRSAMC, Universit\'e
16: Paul Sabatier, 118 route de Narbonne, 31062 Toulouse Cedex 04, France}\\
17: {\small $^2$ Laboratoire de Physique Th\'eorique et Mod\`eles Statistiques, UMR CNRS
18: 8626,Universit\'e Paris Sud, B\^at 100, 91045 Orsay Cedex, France}
19: }
20:
21:
22: \date{17 October 2005}
23:
24: \begin{abstract}
25: We study analytically the late time statistics of the number of
26: particles in a growing tree model introduced by Aldous and Shields. In
27: this model, a cluster grows in continuous time on a binary Cayley
28: tree, starting from the root, by absorbing new particles at the empty
29: perimeter sites at a rate proportional to $c^{-l}$ where $c$ is a
30: positive parameter and $l$ is the distance of the perimeter site from
31: the root. For $c=1$, this model corresponds to random binary search
32: trees and for $c=2$ it corresponds to digital search trees in computer
33: science. By introducing a backward Fokker-Planck approach, we
34: calculate the mean and the variance of the number of particles at
35: large times and show that the variance undergoes a `phase transition' at
36: a critical value $c=\sqrt{2}$. While for $c>\sqrt{2}$ the variance is
37: proportional to the mean and the distribution is normal, for
38: $c<\sqrt{2}$ the variance is anomalously large and the distribution is
39: non-Gaussian due to the appearance of extreme fluctuations.
40: The model is generalized to one where growth occurs on a tree with $m$
41: branches and, in this more general case, we show that the critical
42: point occurs at $c=\sqrt{m}$.
43: \bigskip
44:
45:
46: \noindent {\bf Keywords:} Search trees, Growth models, Phase
47: transitions
48:
49: %\noindent {Pacs Numbers: 02.50.-r, 89.20.Ff, 89.75.Hc}
50:
51: \end{abstract}
52: \maketitle
53:
54: \section{Introduction}
55:
56:
57: Growing clusters are ubiquitous in nature and they exhibit fascinating
58: structures and patterns. Examples range from natural fractals, such
59: as snowflakes and soots, to artificial structures such as networks,
60: for example the Internet and social networks.
61: Various growth models have been studied
62: extensively by physicists over the last three
63: decades~\cite{Growth}. In these models growth starts from a single
64: seed site and proceeds via absorbing new particles into the cluster
65: according to certain specified rules. Different growth rules give rise
66: to different growth models, examples being the Eden model~\cite{Eden},
67: invasion percolation~\cite{IP}, diffusion limited
68: aggregation~\cite{DLA} and the growing network models~\cite{Networks}
69: which have recently received much attention. There are two reasons
70: why many of these growth models are often studied on a Cayley tree (or
71: on the Bethe lattice)~\cite{Edentree}. First, the tree structure of
72: the Bethe lattice mimics a Euclidean lattice in the limit of high
73: dimensions where the mean field theory often becomes exact. Secondly,
74: the absence of loops on the Cayley tree often allows one to obtain
75: exact analytical solutions which are very difficult to obtain on a
76: regular $d$-dimensional lattice. There is yet another compelling
77: motivation for studying these growth models on a Cayley tree and this
78: comes from computer science. `Storing and Search' of data is a very
79: important area of computer science~\cite{Knuth}. Incoming data to a
80: computer is usually stored on a Cayley tree by using various data
81: storage algorithms and the tree so grown is called a `search
82: tree'~\cite{Knuth}. Different algorithms lead to different search
83: trees and in some cases, as explained below, the rules of growth of a
84: search tree can be shown to be exactly equivalent to a `physical'
85: growth model on the tree. Thus the study of these physical growth
86: models on a Cayley tree provides important insights into data storage
87: in computer science.
88:
89: As a first example of this equivalence between a physical growth model
90: and a search tree, we show here that the Eden model on a binary Cayley
91: tree is exactly equivalent to the random binary search tree
92: (RBST). Consider the Eden model on a binary Cayley tree where the
93: growth starts from the root~\cite{Edentree}. At the first step, a
94: particle gets absorbed at the root, thus forming a cluster of size
95: $1$. This cluster has now two empty neighbors which defines the
96: perimeter of the cluster. At the next step, a new particle will get
97: absorbed at any of these two perimeter sites chosen with equal
98: probability, thus forming a cluster of size $2$. The subsequent growth
99: occurs following the same rule, namely a new particle gets absorbed at
100: any of the perimeter sites chosen with equal probability. In Fig. (1),
101: we show a cluster after $4$ steps where the black sites denote the
102: cluster and the shaded sites denote the current perimeter sites that
103: are available for subsequent growth.
104: \begin{figure}
105: \includegraphics[width=4in]{eden.eps}
106: \caption{\label{fig:eden} An Eden cluster of size $4$ on a Cayley
107: tree. The black sites form the cluster and the shaded sites form the
108: perimeter. At the next step, growth can occur at any of the $5$ shaded
109: perimeter sites with equal probability $1/5$. }
110: \end{figure}
111: Fig. (2) shows all possible Eden clusters of size $3$ and their
112: associated statistical weights.
113: \begin{figure}
114: \includegraphics[width=4in]{eden3.eps}
115: \caption{\label{fig:eden3}All possible Eden clusters of size $3$ on a
116: tree and their associated statistical weights w.}
117: \end{figure}
118:
119: On the other hand, a binary search tree in computer science is
120: constructed by the following simple
121: algorithm~\cite{Knuth,Mahmoud,Review}. Imagine that we have a data
122: string consisting of $N$ items which are labeled by the $N$ integers:
123: $\{1,2,\dots,N\}$. These could be the months of the year or the names
124: of people etc. Let us assume that this data appears in a particular
125: order, say $\{6,4,5,8,9,1,2,10,3,7\}$ for $N=10$ integers. This data
126: is first stored on a binary tree following the simple dynamical rule:
127: the first item $6$ is stored at the root of the tree (see
128: Fig. (\ref{fig:bst})). The next item in the string is $4$. We compare
129: it with $6$ at the root and since $4<6$, we store $4$ in the left
130: daughter node of the root. Had it been bigger than the root item $6$,
131: we would have stored it in the right daughter node. The next item in
132: the string is $5$. We again start from the root, see that $5<6$, so we
133: go to the left branch. There we encounter $4$ and we find $5>4$, so we
134: go the right daughter node of $4$. This process is continued till all
135: the $N=10$ items are assigned their nodes and we get a unique binary
136: search tree (BST) (see Fig. (\ref{fig:bst})) for this particular data
137: string $\{6,4,5,8,9,1,2,10,3,7\}$.
138: \begin{figure}
139: \includegraphics[width=4in]{bi1.eps}
140: \caption{\label{fig:bst} The binary search tree associated with the
141: data string $\{6,4,5,8,9,1,2,10,3,7\}$. }
142: \end{figure}
143: Usually the data arrives at a computer in random order. To study this
144: situation, one considers the simplest model called the `random binary
145: search tree' (RBST) model where one assumes that the incoming data
146: string can arrive in any of the $N!$ possible orders or sequences,
147: each with equal probability~\cite{Review}. For each of these
148: sequences, one has a binary tree. For example, in Fig. (4), we show the
149: binary trees for $N=3$ along with their associated probabilities.
150: \begin{figure}
151: \includegraphics[width=4in]{bst1.eps}
152: \caption{\label{fig:bst1} All possible random binary search trees for
153: a data of size $N=3$ and their associated statistical weights.}
154: \end{figure}
155:
156: Comparing Fig. (2) and Fig. (4), one sees immediately that the Eden
157: trees after 3 steps have exactly the same configurations and
158: statistical weights as the random binary search trees with data size
159: $N=3$. This analogy can be easily extended to all $N$.
160: The key point is that after $(n-1)$ steps there are $(n-1)$ occupied sites
161: in the Eden cluster and $n$ perimeter sites (this is easy to understand as
162: the addition of a new occupied site eliminates one old perimeter
163: site while creating two new perimeter sites).
164: The probability of subsequent growth
165: at step $n$ at any of these perimeter sites is $p_n=1/n$. Thus the
166: statistical weight of a cluster of $N$ sites formed by a specific
167: history of growth is simply $w=p_1\,p_2\,\dots p_N=1/N!$, which is the
168: same as in the RBST model. Thus the Eden model on the Cayley tree is
169: exactly equivalent to the RBST.
170:
171: Another popular search tree model is known as the `digital search
172: tree' (DST) which is constructed by the following
173: rule~\cite{Knuth,Mahmoud,Sedgewick,FS,Pittel,FR,KS,SM1}. Consider
174: again a binary Cayley tree each node of which can contain at most one
175: entry. One starts with an empty tree and the data is stored
176: sequentially. The first data item is stored in the root of the
177: tree. The next one arrives at the root and finding it occupied, moves
178: to any of the two empty daughter nodes chosen at random and occupies
179: that node. Then the next item arrives and again it starts at the node,
180: chooses any of its two daughters randomly and moves there. If the
181: chosen daughter is empty it occupies it. If the chosen daughter is
182: already occupied, it again chooses one of its two descendants at
183: random and moves there. Thus at any stage, a new entry starts at the
184: root and performs a random walk (to the left or to the right daughter
185: with equal probability) down the tree till it finds an empty node and
186: occupies the node. Thus one obtains again a growing tree where at any
187: stage growth can occur at any of the perimeter sites, but now the
188: growth probability at a perimeter site $a$ is $p_a \propto 2^{-l_a}$
189: where $l_a$ is the distance of the perimeter site from the root. The
190: DST is an important tree structure in computer science and has been
191: studied extensively. In particular, it turns out the DST is a natural
192: tree representation~\cite{AS,KS} of the data compression algorithm due
193: to Ziv and Lempel~\cite{LZ}. Recently it was shown that a diffusion
194: limited aggregation model introduced by Bradley and Strenski~\cite{BS}
195: in physics is exactly equivalent the the DST model in computer science
196: and a variety of exact results were obtained by exploiting this
197: connection~\cite{SM1}.
198:
199: The examples above illustrate a profound link between growth models
200: and the dynamics of search tree formation in computer science. Note
201: that the two search tree models discussed above, the RBST and the DST,
202: can be considered as special cases of a general growth model where
203: growth occurs ( i.e. a new particle gets absorbed) at any of the
204: available perimeter sites $a$ with a growth probability $p_a\propto
205: c^{-l_a}$ where $c$ is a constant positive parameter and $l_a$ is the
206: distance of the perimeter site $a$ from the root of the tree. The RBST
207: (equivalently the Eden model) corresponds to $c=1$ so that all
208: perimeter sites have equal probability to absorb a particle. The DST,
209: on the other hand, corresponds to $c=2$ as discussed above. It is then
210: useful and interesting to study this general growth model parametrized
211: by $c$ and ask if there are any qualitative changes in the statistical
212: properties of the growth clusters as one varies the parameter $c$
213: continuously. Indeed, Aldous and Shields studied a continuous-time
214: version of this generalized growth model~\cite{AS}. Note that in the
215: two models discussed above time is {\em discrete} and is equal to the
216: number of particles in the tree. In the version of the model studied
217: by Aldous and Shields, time is considered {\em continuous} and growth
218: occurs at any of the available perimeter sites say the site $a$ with a
219: {\em rate} proportional to $c^{-l_a}$ where $c$ is a positive
220: parameter. In this continuous-time model, the total number of
221: particles in the tree at time $t$ is thus a random variable, unlike in
222: the discrete time version. Thus, while the discrete-time model has a
223: constant particle number ensemble, the continuous-time model has a
224: constant time ensemble, much like the canonical and the grand
225: canonical ensemble in statistical physics. Asymptotically at long
226: times, both the discrete-time and the continuous-time versions of the
227: model are expected behave in a similar fashion. Henceforth, we will
228: consider in this paper only the continuous-time version \`a la
229: Aldous-Shields, since it is, from a technical point of view, easier to
230: study than its discrete-time counterpart.
231:
232: The question naturally arises whether the statistical properties of
233: the growing clusters in this model undergo any qualitative change of
234: behavior as one tunes the parameter $c$ continuously. Indeed, Aldous
235: and Shields established rigorous probabilistic bounds to show that the
236: nature of the fluctuations (variance) in the number of particles in
237: the tree at time $t$ is qualitatively different for $c<\sqrt{2}$ and
238: $c>\sqrt{2}$. While for $c>\sqrt{2}$ the central limit theorem holds
239: and the total number of particles has a limiting Gaussian
240: distribution~\cite{AS}, for $c<\sqrt{2}$ the central limit theorem
241: breaks down due to the appearance of anomalously large
242: fluctuations. Thus, there is a sharp {\em phase transition} in the
243: nature of the fluctuations at a critical value $c=\sqrt{2}$. However,
244: the mechanism responsible for this phase transition and even the
245: explicit quantitative behavior of the fluctuations above, below, or at
246: the critical point were not easy to obtain within the rigorous
247: probabilistic analysis of Aldous and Shields. The principal purpose of
248: this paper is to provide a detailed quantitative
249: understanding of this rather `peculiar' phase transition. Our method,
250: completely different from the original approach of Aldous and Shields,
251: employs a backward Fokker-Planck formalism. The advantage of this
252: method is that one can obtain exact asymptotic results
253: explicitly. Moreover, our analysis also shows that the mathematical
254: mechanism behind this phase transition is similar to the phase
255: transitions found recently in the variance of the number of nodes
256: needed to store data on a $m$-ary search tree (where $m$ is the number
257: of branches) at the critical value
258: $m=26$~\cite{CH,DM1,Review,CP,FN,FFN,TH} and also in the variance of
259: the number of splitting events in a $D$-dimensional fragmentation
260: model at the critical value $D_c= \pi/{
261: {\sin}^{-1}\left(1/\sqrt{8}\right)}=8.69363\ldots $~\cite{DM1,Review}.
262:
263: The layout of the paper is as follows. In the next Section (II), we
264: define the model precisely and summarize the main results. We study
265: here a generalized Aldous-Shields model where the growth takes place
266: on a Cayley tree with $m$ branches. In Section III, we derive the
267: evolution equations for the mean and variance of the number of
268: occupied sites as a function of time via a backward Fokker Planck
269: technique. A simple scaling analysis is then carried out to determine
270: the temporal growth exponents. In Section IV a more thorough analysis
271: of the evolution equations is provided that enables us to obtain
272: explicitly not just the growth exponents, but also exact expressions
273: for various amplitudes and prefactors that include interesting
274: log-periodic oscillations. We conclude with a summary and a
275: discussion of open questions in the last section.
276:
277:
278: \section{The Model and the Results}
279:
280: We consider a generalized Aldous-Shields model where growth occurs on
281: a Cayley tree (rooted at $O$) with $m$ branches (see
282: Fig.~\ref{fig:mary}). Aldous and Shields studied~\cite{AS} only the
283: binary case $m=2$. Initially the tree is empty and growth occurs in
284: continuous time starting from the root $O$. At any instant $t$, one
285: first identifies the available perimeter nodes. A node $a$ at time $t$
286: is a perimeter node if it is empty at $t$ but its parent node is
287: occupied at $t$ (see Fig.~\ref{fig:mary}). Subsequently, in a small
288: time interval $\Delta t$, a perimeter node $a$ either absorbs a
289: particle with probability $c^{-l_a}\, \Delta t$ or remains unoccupied
290: with probability $(1- c^{-l_a}\, \Delta t)$, where $l_a$ is the depth
291: of the perimeter node $a$, i.e. its distance from the root $O$. This
292: growth process occurs simultaneously at all the perimeter nodes. Thus,
293: the total number of particles $n(t)$ in the tree rooted at $O$ is
294: clearly not a fixed number at a given time $t$, instead it is a random
295: variable in the sense that the value of $n(t)$ differs from one
296: history of evolution to another. We are interested in computing the
297: statistics of $n(t)$ at large times $t$.
298: \begin{figure}
299: \includegraphics[width=4in]{mary.eps}
300: \caption{\label{fig:mary} The growth of the Aldous-Shields model on a
301: tree with $m$ branches and rooted at $O$. The filled circles are
302: occupied nodes and the shaded ones are the perimeter nodes where
303: growth can occur subsequently. For example the site marked $a$ is a
304: typical perimeter site at a distance $l_a=3$ from the root $O$ of the
305: tree.}
306: \end{figure}
307:
308: In this model, we have two parameters $m$ and $c$. It is useful to
309: first summarize our main results. Using a backward Fokker-Planck
310: approach we derive an exact evolution equation for the generating
311: function,
312: \begin{equation}
313: G(\mu, t) =\langle \exp\left(-\mu\, n(t)\right)\rangle =
314: \sum_{n=0}^{\infty} e^{-\mu\, n} P(n,t)
315: \label{defG}
316: \end{equation}
317: where the angle brackets denote an average over all histories of the
318: evolution process and $P(n,t)$ is the probability distribution of $n$
319: at time $t$. We show that $G(\mu,t)$ evolves via the equation
320: \begin{equation}
321: \frac{ dG(\mu,t)}{dt} = - G(\mu,t) + e^{-\mu}\, G^m (\mu, t/c),
322: \label{eqG}
323: \end{equation}
324: starting from the initial condition $G(\mu,0)=1$. By differentiating
325: $G(\mu,t)$ with respect to $\mu$ and putting $\mu=0$, one can also
326: derive the evolution equations for all the moments of $n(t)$. The
327: equation (\ref{eqG}) is nonlinear and nonlocal in time for generic
328: values of $c$ and $m$, and is thus difficult to solve exactly, except
329: for the $c=1$ case when it becomes local. However, we were able to
330: compute exactly the asymptotic large time behaviors of the mean and
331: the variance of $n(t)$ for arbitrary $m$ and $c$. Below we present
332: our results for the three different cases $c=1$, $c<1$ and $c>1$
333: separately.
334: \vspace{0.3cm}
335:
336: \noindent {\bf The case $c=1$:} In this case our model is precisely
337: the continuous-time version of the Eden model. This case $c=1$ is
338: exactly solvable since the evolution equation (\ref{eqG}) becomes
339: local in time. We solved for $G(\mu,t)$ and obtained the following
340: explicit result for the distribution $P(n,t)$ for all $m$ and $t$
341: \begin{equation}
342: P(n,t)=
343: \frac{\Gamma\left(n+\frac{1}{m-1}\right)}{\Gamma\left(\frac{1}{m-1}\right)\,
344: \Gamma(n+1)}\, e^{-t}\, \left[1-e^{-(m-1)\,t}\right]^{n}
345: \label{csol1}
346: \end{equation}
347: where $\Gamma(x)$ is the standard Gamma function. The mean number of
348: particles $M(t) = \langle n(t)\rangle$ increases exponentially in time
349: for all $m>1$,
350: \begin{equation}
351: M(t) = \frac{1}{m-1}\left[\exp\left( (m-1)t\right) -1\right].
352: \label{eqc1}
353: \end{equation}
354: For the special case $m=1$ (a line with a constant rate of
355: deposition), $M(t)=t$ and the distribution $P(n,t)= e^{-t}
356: t^{n}/{n}!$, obtained from Eq. (\ref{csol1}) by taking the limit $m\to
357: 1$, is purely Poissonian as expected.
358:
359: \vspace{0.4cm}
360:
361: \noindent {\bf The case $c<1$:} Since the growth rate at a perimeter
362: node $a$ is proportional to $c^{-l_a}$ where $l_a$ is the distance of
363: the node from the root $O$, it is clear that for $c<1$, farther a
364: perimeter node is from the root, the larger is its probability to get
365: occupied. Thus the cluster grows in a rather ramified manner where
366: long branches grow faster than the short branches. In this case we
367: expect that the mean number of sites grows at least exponentially. But
368: since physically this case is of little interest, we do not discuss it
369: further in this paper.
370:
371: \vspace{0.4cm}
372:
373: \noindent {\bf The case $c>1$:} We now come to the physically most
374: relevant case $c>1$. We show that in this case there is a sharp phase
375: transition in the asymptotic statistics of $n(t)$ across the
376: critical line $c=\sqrt{m}$ in the $(m,c)$ plane with $m>1$ and $c>1$.
377: We calculated exactly the asymptotic time dependence of the mean
378: $M(t)=\langle n(t)\rangle$ and the variance $V(t)= \langle
379: n^2(t)\rangle - M^2(t)$, for all values of $m>1$ and $c>1$. We show
380: that while the fluctuations are normal (i.e. the variance of $n(t)$ is
381: proportional to its mean) for $c>\sqrt{m}$, they are anomalously large
382: for $c<\sqrt{m}$. Even though we have two parameters $c$ and $m$, it
383: turns out that the asymptotic behaviors can be described in terms of
384: the single growth parameter
385: \begin{equation}
386: \alpha = \frac{\ln(m)}{\ln(c)}
387: \label{eqalpha}
388: \end{equation}
389: where $\alpha>0$ since $c>1$. In terms of $\alpha$, the phase
390: transition takes phase at the critical value $\alpha_c=2$. The normal
391: phase for $c>\sqrt{m}$ corresponds to $\alpha<\alpha_c=2$ and the
392: anomalous phase for $c<\sqrt{m}$ corresponds to $\alpha>\alpha_c=2$.
393: \begin{figure}
394: \includegraphics[width=4in]{aphd.eps}
395: \caption{\label{fig:aphd} The phase transition as a function of
396: $\alpha=\ln(m)/\ln(c)>0$. The critical point at $\alpha_c=2$
397: separates the phase ($\alpha<2$) with normal fluctuations from the
398: phase ($\alpha>2$) with anomalously large fluctuations.}
399: \end{figure}
400:
401: More precisely, we find that for large $t$, the mean $M(t)$ grows as a
402: power law (up to corrections periodic in $\ln(t)$),
403: \begin{equation}
404: M(t) \sim A\, t^{\alpha},
405: \label{eqm}
406: \end{equation}
407: and we provide an explicit expression for the amplitude $A$. The
408: variance, on the other hand, has different behaviors for $c<\sqrt{m}$
409: and $c>\sqrt{m}$ or equivalently for $\alpha>2$ and $\alpha<2$. We
410: show that the variance at large times $t$, again up to log-periodic
411: corrections, grows as
412: \begin{eqnarray}
413: V(t) &\simeq & B'\, t^{\alpha} \,\,\,\,\quad\quad \rm{for} \quad\quad
414: \alpha <2 \label{eqv2}\\ &\simeq & B_c\, t^2 \ln(t) \,\,\, \rm{for}
415: \quad\quad \alpha=\alpha_c=2 \label{eqvc}\\ &\simeq & B\,
416: t^{2\alpha-2} \,\,\,\,\quad \rm{for} \quad\quad \alpha >2 .
417: \label{eqv1}
418: \end{eqnarray}
419: We also provide exact expressions for the amplitudes $B'$, $B_c$ and
420: $B$. In order to use these continuous-time results for the
421: discrete-time model where the `time' is same as the number of
422: particles, it is instructive to eliminate the explicit $t$ dependence
423: in the results for the variance and instead express it as a function
424: of the mean number of particles. Eliminating $t$ between
425: Eqs. (\ref{eqm}), (\ref{eqv2}), (\ref{eqvc}) and (\ref{eqv1}), we get
426: \begin{eqnarray}
427: V(t) &\sim & C'\, M(t) \,\,\,\quad\quad \rm{for} \quad\quad \alpha <2
428: \label{eqvm2}\\ &\sim & C_c\, M(t) \ln(M(t))\,\,\, \rm{for}\quad
429: \alpha=\alpha_c=2 \label{eqvmc}\\ &\sim & C\, M(t)^{2-2/\alpha}
430: \,\,\,\,\, \rm{for} \quad\quad \alpha >2 . \label{eqvm1}
431: \end{eqnarray}
432: Explicit expressions for the amplitudes $C'$, $C_c$ and $C$ are
433: likewise provided.
434:
435: We thus see that for $\alpha <2$ the fluctuations of $n(t)$ about its
436: mean value, denoted by $\Delta n(t)=\sqrt{V(t)}$, are of order $
437: M^{1/2}$ as is the case for a normal Gaussian or Poisson
438: distribution. However for $\alpha >2$ we find that $\Delta n(t) \sim
439: M^{1 -{1/\alpha}}$ for large $M$. For $\alpha >2$, we have that $1-
440: {1/\alpha} > {1/2}$ and hence the relative fluctuations about the mean
441: become larger as we cross the threshold $\alpha = 2$ from below. The
442: phase $\alpha <2$, or equivalently $c>\sqrt{m}$ corresponds to a
443: region of slower growth where the central limit theorem holds and the
444: distribution of $n(t)$ is asymptotically normal. On the other hand,
445: $\alpha >2$ marks a phase where rapid growth tends to occur along a
446: single branch resulting in anomalously large fluctuations. Thus the
447: statistics of $n(t)$ in this phase is dominated by extreme
448: fluctuations. The nature of this phase transition is thus very
449: similar to the ones recently reported in $m$-ary search
450: trees~\cite{CH,DM1,Review,CP,FN,FFN,TH} and a related fragmentation
451: model~\cite{DM1,Review}.
452:
453: We end this section with a remark on the usage of the term `phase transition'.
454: The `phase transition' observed in this model refers to the abrupt change
455: of the variance (and also that of the full distribution) of the number of particles $n(t)$
456: in the tree as one changes the parameter $\alpha=\ln(m)/\ln(c)$ through
457: its critical value $\alpha_c=2$. This may not correspond to the traditional definition
458: of `phase transition' used in equilibrium statistical mechanics, e.g. the divergence
459: of a correlation length as one approaches a critical point as in second order
460: phase transition. The `phase transition' in the Aldous-Shields model is
461: closer to the change of behavior one observes in the diffusion
462: of a L\'evy walker. A L\'evy walker jumps, at each step, by a random length
463: $l$ drawn from a power law distribution, $p(l)\sim l^{-(1+\gamma)}$ for large $l$
464: with $\gamma>1$ (required for normalization). It is well known\cite{BG} that
465: the root mean square displacement of the particle after $n$ steps
466: $\sqrt{\langle R_n^2 \rangle } \sim n^{1/2}$ for large $n$ only when $\gamma>2$, i.e. one gets
467: normal
468: diffusion and the asymptotic position of the
469: walker is distributed normally. On the other hand, for $1<\gamma<2$ one gets anomalously large
470: diffusion, $\sqrt{\langle R_n^2\rangle} \sim n^{1/\gamma}$ for large $n$
471: and the asymptotic distribution of the position of the walker
472: is non-Gaussian. Thus
473: there is a change of behavior at the critical value $\gamma_c=2$. The change of behavior
474: in the variance of the number of particles in the Aldous-Shields model
475: at the critical value $\alpha_c=2$ is thus similar in nature to the
476: the change of behavior seen for the L\'evy diffusion at $\gamma_c=2$, rather
477: than the standard `phase transition' observed in critical phenomena.
478:
479: \section{Derivation of the evolution equations}
480:
481: In this section we will derive the evolution equation for the
482: probability distribution, and in particular the evolution equations
483: for the mean and the variance, of the total number of occupied sites
484: $n(t)$ at time $t$ in a tree rooted at the site $O$. The root $O$, by
485: definition, has level or depth $l_0 = 0$. The method of derivation is
486: based on a backward Fokker Planck formalism which involves considering
487: the future evolution of $n(t)$ conditioned on what happens in the
488: first infinitesimal time interval $(0,\Delta t)$. Since our final aim
489: is to derive a recursion relation for the evolution process, it is
490: convenient to first derive the evolution equation for the number of
491: particles $n_a(t)$ in a subtree rooted, say, at any arbitrary site
492: $a$. By definition, $n_a(t)$ includes the particle at the root $a$.
493: The number of particles in the full tree is just a special case when
494: the site $a$ is chosen to be the original root $O$ of the full tree.
495: We now count the `local' time $t$ (for this subtree) from the instant
496: the site $a$ becomes a potential growth site, so that, by definition,
497: $n_a(0)=0$. Clearly, at any given $t$, the distribution $P(n_a,t)$ of
498: $n_a(t)$ depends only on $t$ and $l_a$, the depth of the site
499: $a$. This means that one can write
500: \begin{equation}
501: \langle F(n_a(t))\rangle = \sum_{n_a=0}^{\infty} F(n_a)\, P(n_a,t)=
502: f(t; l_a)
503: \label{inv}
504: \end{equation}
505: where $F(x)$ is any arbitrary function.
506:
507: Consider now the site $a$ with its descendants $a_1,\ a_2, \cdots,
508: a_m$, where $a$ is a potential growth site at $t=0$. By definition $a$
509: is unoccupied at $t=0$ and thus $a_1,\ a_2, \cdots, a_m$ are not
510: potential growth sites at $t=0$. In the first infinitesimal time
511: interval $(0,\Delta t)$ there are two possibilities: (i) either no
512: particle fills the potential growth site $a$, thus
513: $n_a(t)=n_a(t-\Delta t)$. This happens with probability
514: $1-c^{-l_a}\Delta t$. (ii) the other possibility is that the potential
515: growth site $a$ is filled by a particle with probability $
516: c^{-l_a}\Delta t$ and as a consequence the number of particles in the
517: subtree rooted at $a$ is increased by one and the daughter nodes
518: $a_1,\ a_2, \cdots, a_m$ all become potential growth
519: sites. Mathematically we can write the above evolution in the
520: following way: in the time interval $(0,\Delta t)$
521: \begin{equation}
522: n_a(t) = n_a(t-\Delta t) (1-I) + \left[1 + \sum_{i=1}^m
523: n_{a_i}(t-\Delta t)\right] I \label{master}
524: \end{equation}
525: where $I$ is a random variable which takes the value $1$ with
526: probability $c ^{-l_a}\Delta t$ and $0$ with probability $1-c
527: ^{-l_a}\Delta t$. Taking the expectation of Eq. (\ref{master}) with
528: respect to $I$ and the subsequent growth process in the remaining time
529: $t-\Delta t$ we obtain, upon taking the limit $\Delta t \to 0$,
530: \begin{equation}
531: \frac{d}{dt}\langle n_a(t)\rangle = c^{-l_a}\left[ 1 -\langle
532: n_a(t)\rangle + \sum_{i=1}^m \langle n_{a_i}(t)\rangle \right].
533: \end{equation}
534:
535: We now use the property that the statistics of the number of particles
536: in a subtree rooted at level $a$ depends only on $l_a$ as encoded in
537: Eq. (\ref{inv}) to obtain
538: \begin{equation}
539: \frac{d}{dt}M(t;l_a) = c^{-l_a}\left[ 1 -M(t;l_a) + m M(t;l_a+1)
540: \right].\label{eqm1}
541: \end{equation}
542: where we have defined
543: \begin{equation}
544: M(t;l_a) = \langle n_a(t)\rangle
545: \end{equation}
546: and have used the fact that by definition $l_{a_i} = l_a +1$ if $a_i$
547: is a daughter of the node $a$. Note that the root $O$ of the full tree
548: has depth $l_O=0$. Thus the mean $M(t)=\langle n(t) \rangle$ of the
549: total number of particles in the full tree rooted at $O$ is simply,
550: $M(t)= M(t,0)$. Thus to obtain $M(t)$, our strategy is to find the
551: solution $M(t,l_a)$ of Eq. (\ref{eqm1}) for arbitrary $l_a$ and
552: eventually put $l_a=0$.
553:
554: The next step is to notice that as one goes down a level in the tree
555: the growth rate is reduced by a factor of $c$ which amounts to
556: rescaling time by a factor of $1/c$. In the notation of
557: Eq. (\ref{inv}) this means that one may write
558: \begin{equation}
559: f(t; l_a + 1) = f\left(\frac{t}{c}; l_a\right).\label{scal}
560: \end{equation}
561: Next we put $l_a=0$ in Eq. (\ref{eqm1}), use the definition $M(t;l_O)
562: = M(t;0) = M(t)$ and also the scaling property in Eq. (\ref{scal}) to
563: obtain,
564: \begin{equation}
565: \frac{d}{dt}M(t) = 1 -M(t) + m M\left(\frac{t}{c}\right).\label{eqM}
566: \end{equation}
567: This equation, supplemented by the boundary condition $M(0)=0$, then
568: describes the evolution of the mean number of occupied sites.
569:
570: An equation for the variance of $n(t)$ can be derived in a similar
571: fashion. The starting point is obtained by squaring the stochastic
572: evolution equation (\ref{master}) and then taking the expectation over
573: $I$ and the evolution in the remaining time $t-\Delta t$. This yields
574:
575: \begin{equation}
576: \langle n^2_a(t)\rangle = \langle n^2_a(t-\Delta t)\rangle (
577: 1-c^{-l_a}\Delta t) + \left[1 + 2\langle \sum_{i=1}^m n_{a_i}(t-\Delta
578: t)\rangle + \langle \left(\sum_{i=1}^m n_{a_i}(t-\Delta
579: t)\right)^2\rangle \right]c^{-l_a}\Delta t .
580: \label{eqvs}
581: \end{equation}
582: We now use the fact that the subtrees rooted at sites at the same
583: level are statistically independent and so
584: \begin{equation}
585: \langle n_{a_i}(t)n_{a_j}(t)\rangle = \langle n_{a_i}(t)\rangle
586: \langle n_{a_j}(t)\rangle \ {\rm{for}}\ i\neq j.
587: \end{equation}
588: Now defining the variance of the number of sites occupied in the tree
589: rooted at $0$ as
590: \begin{equation}
591: V(t) = \langle n^2(t)\rangle - \langle n(t)\rangle^2
592: \end{equation}
593: and using the scaling relation Eq. (\ref{scal}), after some elementary
594: algebra, we obtain
595: \begin{equation}
596: \frac{d}{dt}V(t) = \left(\frac{d}{dt}M(t)\right)^2 -V(t) +m
597: V\left(\frac{t}{c}\right).
598: \label{eqV}
599: \end{equation}
600: The boundary condition for this equation is clearly $V(0) = 0$.
601:
602: Another way to obtain the equations for $M$ and $V$ is by deriving
603: directly an evolution equation for the generating function $G(\mu,t)$
604: of $n(t)$ defined as in Eq. (\ref{defG}). Following exactly the same
605: backward Fokker-Planck strategy as used for the mean, it is
606: straightforward to show that $G(\mu,t)$ evolves by the nonlinear
607: nonlocal equation (\ref{eqG}). The moment equations, and equations
608: for the higher moments, Eq. (\ref{eqM}) and Eq. (\ref{eqV}) can be
609: obtained by differentiating Eq. (\ref{eqG}) with respect to $\mu$ the
610: appropriate number of times and setting $\mu =0$ at the end.
611:
612: The evolution equation (\ref{eqG}) is difficult to solve explicitly
613: for generic values of $c$ and $m$ since it is a nonlinear (for $m\ne
614: 1$) and nonlocal (for $c\ne 1$) equation. However, exact results can
615: be derived in a few cases that we consider below. The asymptotic
616: solution for the mean and variance for generic $m$ and $c$ will be
617: presented later in the next section.
618:
619: \subsection{Exact Solution for the Eden Growth $c=1$}
620: The case $c=1$ corresponds to the Eden model where growth occurs at
621: any of the available perimeter sites with equal rates. For $c=1$,
622: Eq. (\ref{eqG}) becomes local in time $t$ and can be explicitly
623: solved. We find that for all $m\ge 1$
624: \begin{equation}
625: G(\mu,t) = e^{-t}\, \left[1- e^{-\mu}\,
626: \left(1-e^{-(m-1)t}\right)\right]^{-1/(m-1)}.
627: \label{c1solgf}
628: \end{equation}
629: Expanding the r.h.s. of Eq. (\ref{c1solgf}) in powers of $e^{-\mu}$ as
630: in Eq. (\ref{defG}), one can then read off the distribution $P(n,t)$
631: explicitly as in Eq. (\ref{csol1}). The mean number of particles grows
632: exponentially for all $m>1$ as in Eq. (\ref{eqc1}). Similarly, one can
633: compute the variance $V(t)$. We find
634: \begin{equation}
635: V(t) = \frac{1}{(m-1)}\,e^{(m-1)t}\left(e^{(m-1)t}-1\right)\,.
636: \label{varc1}
637: \end{equation}
638: For a fixed $m$, if one takes the limit of large $t$ and large $n$
639: keeping the product $n\,e^{-(m-1)t}$ fixed in Eq. (\ref{csol1}), one
640: finds an asymptotic distribution
641: \begin{equation}
642: P(n,t) \simeq \frac{e^{-t}}{\Gamma\left(\frac{1}{(m-1)}\right)}\,
643: n^{-(m-2)/(m-1)}\, \exp\left[-n\, e^{-(m-1)t}\right].
644: \label{distc1}
645: \end{equation}
646: Thus, to the leading order, the distribution $P(n,t)$ decays
647: exponentially for large $n$ over a characteristic size $n^*\sim
648: e^{(m-1)t}$ that grows exponentially with time $t$. Interestingly, the
649: distribution has a sub-leading power law tail (in addition to the
650: leading exponential tail) $n^{-\phi}$ where the exponent
651: $\phi=(m-2)/(m-1)$ depends continuously on $m$.
652:
653: For the special case $m=1$, where we have just a line of sites and the
654: particles arrive at an empty available site at a constant rate $1$, we
655: get from Eq. (\ref{eqc1}), $M(t)=t$. The full distribution, from
656: Eq. (\ref{csol1}), becomes a Poisson distribution $P(n,t)= e^{-t}
657: t^{n}/{n!}$ as expected.
658:
659: \subsection{Exact Solution for the Digital Search Tree Growth $c=m$}
660:
661: The case $c=m$ corresponds to the case where particles arrive at a
662: constant rate at the root $O$ and then each carries out a random walk
663: down the tree until it finds a free site to occupy. During its
664: downward journey in the tree the particle, after arriving at any
665: occupied site, chooses one of its $m$ descendants at random. This is
666: precisely the algorithm for constructing a $m$-ary digital search
667: tree~\cite{Knuth,Sedgewick,AS}. If the rate at which the particles
668: arrive at the root $O$ is one then the total number of particles in
669: the tree at time $t$, $n(t)$, is clearly a random variable with Poisson
670: distribution
671: \begin{equation}
672: P(n=k,t) = \frac{t^k\exp(-t)}{ k!}, \label{eqPosd}
673: \end{equation}
674: where $k=0,1,2\cdots$ is a positive integer.
675: This yields
676: \begin{equation}
677: M(t) = t; \ V(t)= t, \label{poisson}
678: \end{equation}
679: which we see immediately are the solutions to Eq (\ref{eqM}) and Eq.
680: (\ref{eqV}). Furthermore we see that the generating function
681: $G(\mu,t)$ for a Poisson distribution is given by
682: \begin{equation}
683: G(t,\mu) = \exp\left(-t+ t \exp(-\mu)\right).
684: \end{equation}
685: It is easy to check that indeed this solves Eq. (\ref{eqG}) in the
686: case $m=c$.
687:
688: \subsection{ A self-consistent scaling approach for the leading asymptotic growth of the mean and the
689: variance for $c>1$ and $m>1$}
690:
691: The late time asymptotic behavior of Eq. (\ref{eqM}) and
692: Eq. (\ref{eqV}) for $c>1$ and $m>1$ may be deduced quite simply by
693: making a self-consistent ansatz for the late time behavior of $M$ and
694: $V$. First consider Eq. (\ref{eqM}). We make the ansatz
695: \begin{equation}
696: M(t) \simeq A\, t^\alpha .
697: \end{equation}
698: Substituting this into Eq. (\ref{eqM}) we may neglect the derivative
699: term on the l.h.s. and assuming that $\alpha >0$ (i.e. $c>1$),
700: matching the coefficients of $t^{\alpha}$ gives
701: \begin{equation}
702: \frac{m}{c^{\alpha}} - 1 = 0,
703: \end{equation}
704: which yields
705: \begin{equation}
706: \alpha = \frac{\ln(m)}{ \ln(c)}.
707: \end{equation}
708: For non-trivial tree structures we are always in the situation where
709: $m\geq 2$ and for the above solution to make sense we require that $c
710: >1$ to have a positive exponent $\alpha$. While this simple minded
711: scaling approach yields the correct power law growth of $M(t)\simeq
712: A\, t^{\alpha}$ for $c>1$, it does not provide us the value of the
713: amplitude $A$. To derive an exact expression for $A$, we need to solve
714: the full nonlocal equation (\ref{eqM}) at late times, and this will be
715: carried out in the next section.
716:
717: Let us make a similar power law ansatz for the late time behavior of
718: $V(t)$
719: \begin{equation}
720: V(t) \simeq B\, t^{\beta}.
721: \end{equation}
722: Substituting this into Eq. (\ref{eqV}) and neglecting the derivative
723: term we obtain
724: \begin{equation}
725: - B t^{\beta} + B m \frac{t^\beta}{c^{\beta}} + A^2 \alpha^2
726: t^{2\alpha -2}=0.
727: \label{match1}
728: \end{equation}
729: Asymptotically there are two ways to satisfy this equation. First if
730: we assume a priori that $\beta > 2\alpha -2$ then the first two terms
731: in Eq. (\ref{match1}) must cancel leading to $c^{\beta}=m$, i.e.
732: $\beta = \ln(m)/\ln(c)=\alpha$. The a posteriori condition that this
733: solution is valid is thus $\alpha > 2\alpha -2$, which means $\alpha <
734: 2$. The second possibility is that all three terms contribute and thus
735: $\beta = 2\alpha -2$. In this case we find that
736: \begin{equation}
737: B = \frac{A^2 \alpha^2}{1- \frac{m}{c^{2\alpha -2}}} = \frac{A^2
738: \alpha^2}{1- c^{2-\alpha}},
739: \label{reab}
740: \end{equation}
741: and in obtaining the last equality in
742: Eq. (\ref{reab}) we have used $m=c^{\alpha}$. However for this
743: solution to make sense we must have that $B>0$ because $V(t)$
744: is clearly positive, consequently Eq. (\ref{reab}) can only hold
745: when $\alpha>2$ (since $c>1$).
746:
747: This simple minded scaling approach thus indicates that there is a
748: phase transition in the late time behavior of the variance $V(t)$ at
749: the critical parameter value $\alpha_c=2$. For $\alpha<2$, we have
750: $V(t)\sim B'\, t^{\alpha}$ where the amplitude $B'$ can not be
751: determined by the scaling approach. On the other hand, for $\alpha>2$
752: the scaling approach indicates $V(t)\sim B\, t^{2\alpha-2}$ and
753: moreover it provides a relationship between the amplitudes $B$ and $A$
754: (of the mean) via Eq. (\ref{reab}). The critical point $\alpha_c =2$
755: thus separates the region of normal growth $\alpha<2$ (or equivalently
756: $c>\sqrt{m}$), where $V(t)\sim M(t)$, from the the region $\alpha>2$
757: (i.e. $c<\sqrt{m}$) where the variance grows anomalously faster $V(t)
758: \sim [M(t)]^{2-2/\alpha}$. In the next section, we will see that the
759: analysis of the full nonlocal equations (\ref{eqM}) and (\ref{eqV})
760: indeed corroborates theses scaling results, and in addition produces exact
761: expressions for all the amplitudes.
762:
763: Before proceeding to the full analysis of Eqs. (\ref{eqM}) and
764: (\ref{eqV}) in the next section for generic $c>1$ and $m$, it is
765: instructive to note that analytic progress is also possible for
766: Eq. (\ref{eqM}) in the case where $\alpha=\ln(m)/\ln(c)$ is a positive
767: integer. This includes, in particular, the critical point
768: $\alpha=2$. We make the following ansatz
769: \begin{equation}
770: M(t) = \sum_{k=1}^{\infty} b_n t^n,
771: \end{equation}
772: where the term $k=0$ in the above sum is omitted in order to respect
773: the initial condition $M(0) =0$. Matching powers of $t$ on
774: substituting this ansatz into Eq. (\ref{eqM}) yields
775: \begin{eqnarray}
776: b_1 &=& 1 \nonumber \\ b_{k+1} &=& b_{k} \frac{1}{k+1}
777: \left(\frac{m}{c^{k}}-1\right)\ {\rm for}\quad k > 1 .
778: \end{eqnarray}
779: We thus see that if there exists a positive integer $k^*$ such that
780: $\alpha=\ln(m)/\ln(c)=k^*$ then $b_k = 0$ for all $k>k^*=\alpha$ and
781: we have found the solution to Eq. (\ref{eqM}) in these cases. At late
782: times the leading order behavior is thus dominated by the term
783: containing $t^{\alpha}$ and we get
784: \begin{eqnarray}
785: M(t) &\sim& \frac{t^{\alpha}}{\alpha
786: !}\left(\frac{m}{c^{\alpha-1}}-1\right)
787: \left(\frac{m}{c^{\alpha-2}}-1\right) \cdots
788: \left(\frac{m}{c}-1\right) \nonumber\\ &=& \frac{t^{\alpha}}{ \alpha
789: !}\left({c}^{\alpha-1} -1\right) \left({c}^{\alpha-2} -1\right)\cdots
790: \left({c}-1\right).
791: \label{eqmint}
792: \end{eqnarray}
793: In particular, at the critical point $\alpha=2$, we get for large $t$
794: \begin{equation}
795: M(t)\simeq \frac{(c-1)}{2}\, t^2
796: \label{critint1}
797: \end{equation}
798: Thus, at this special point $\alpha=2$, we have even managed to
799: compute the amplitude $A=(c-1)/2$ of the mean $M(t)\sim A\, t^2$
800: exactly. In the case $\alpha >2$, the behavior of the variance $V(t)$
801: now follows immediately from Eq. (\ref{reab}). Finally, exactly at the
802: critical point $\alpha =2$, we may asymptotically solve
803: Eq. (\ref{eqV}) with the ansatz $V = B\, t^2 \ln(t)$ to yield
804: \begin{equation}
805: B = \frac{4}{ \ln(c)}A^2 = \frac{(c-1)^2}{ \ln(c)},
806: \label{critint2}
807: \end{equation}
808: where in the last line of Eq. (\ref{critint2}) we have used
809: $A=(c-1)/2$.
810:
811: \section{General solution of the evolution equations of the Mean and the Variance}
812:
813: The full solutions to the nonlocal and nonlinear differential
814: equations of the type in Eqs. (\ref{eqM},\ref{eqV}) are rather
815: difficult to obtain completely. Here we obtain the exact asymptotic
816: solutions following an approach similar to the one used by Flajolet
817: and Richmond ~\cite{FR} in solving a class of difference-differential
818: equations arising in the context of digital search trees.
819:
820: \subsection{ Solution for the mean $M(t)$}
821: We start by the analysis of Eq. (\ref{eqM}) assuming $c>1$ and
822: $m>1$. Taking the Laplace transform of Eq. (\ref{eqM}) we obtain
823: \begin{equation}
824: s{\tilde M}(s) = \frac{1}{ s} -{\tilde M}(s) + mc{\tilde M}(cs),
825: \end{equation}
826: where
827: \begin{equation}
828: {\tilde M}(s) = \int_0^\infty dt \ \exp(-st) M(t)
829: \end{equation}
830: and we have used the initial condition $M(0) = 0$. The above may be
831: written as
832: \begin{equation}
833: {\tilde M}(s) = \frac{1}{s (s+1)} +\frac{mc}{ (s+1)} {\tilde M}(cs).
834: \label{eqms1}
835: \end{equation}
836: Now as ${\tilde M}(s)$ should go to zero as $s\to \infty$ and we are
837: considering the case $c>1$, we solve Eq. (\ref{eqms1}) by iteration
838: finding
839: \begin{equation}
840: {\tilde M}(s) = \frac{1}{ s}\sum_{j=0}^{\infty} \frac{m^j}{
841: (1+s)(1+cs) \cdots (1+ c^j s)}.
842: \label{eqms2}
843: \end{equation}
844: Note that taking the limit $s\to 0$ is not straightforward in
845: Eq. (\ref{eqms2}). This is because if we set $s=0$ in the sum on the
846: r.h.s of Eq. (\ref{eqms2}), the sum diverges since $m>1$.
847: Following~\cite{FR} we introduce the function
848: \begin{equation}
849: Q(u) = \prod_{l=0}^{\infty} \left(1+ \frac{u}{c^{l}}\right).
850: \label{Qdef}
851: \end{equation}
852: Thus $Q(s/c)= (1+s/c)(1+s/c^2)(1+s/c^3)\ldots$. On the other hand,
853: \begin{eqnarray}
854: Q(c^js)= \prod_{j=0}^{\infty} (1+ c^{j-l}s) &=& (1+c^j
855: s)(1+c^{j-1}s)\cdots (1+cs)(1+s)(1+s/c)(1+s/c^2)\ldots \nonumber \\
856: &=& (1+s)(1+cs)(1+c^2s)\cdots (1+c^j s) Q(s/c).
857: \label{Qs1}
858: \end{eqnarray}
859: Thus, one can rewrite the product $(1+s)(1+cs)\cdots(1+c^js)= Q(c^j
860: s)/Q(s/c)$. Using this in Eq. (\ref{eqms2}) we get
861: \begin{equation}
862: {\tilde M}(s) = \frac{Q({s/c})}{s}\ H(s),
863: \label{Ms1}
864: \end{equation}
865: where
866: \begin{equation}
867: H(s) = \sum_{j =0}^{\infty}\frac{m^j}{ Q(c^js)}.
868: \label{hs1}
869: \end{equation}
870: The next step is to take the Mellin transform of $H(s)$ defined as
871: \begin{equation}
872: H^*(x) = \int_0^\infty ds\ s^{x-1}\ H(s).
873: \label{hs2}
874: \end{equation}
875: Substituting $H(s)$ from Eq. (\ref{hs1}) in the definition in
876: Eq. (\ref{hs2}) we get
877: \begin{eqnarray}
878: H^*(x)& =& \sum_{j=0}^{\infty} m^j \int_0^{\infty}
879: \frac{s^{x-1}}{Q(c^j s)} ds \nonumber \\ &=& \sum_{j=0}^{\infty}
880: (mc^{-x})^{j} \int_0^{\infty} \frac{\sigma^{x-1}}{Q(\sigma)}\,d\sigma
881: \nonumber \\ &=& \frac{h^*(x)}{ 1-mc^{-x}},
882: \label{mt1}
883: \end{eqnarray}
884: where
885: \begin{equation}
886: h^*(x) = \int_0^\infty d\sigma\ \frac{ \sigma^{x-1}}{
887: Q(\sigma)}\label{eqh}
888: \end{equation}
889: and in evaluating the sum over $j$ we have assumed ${\rm Re}(mc^{-x}
890: <1)$ or equivalently ${\rm Re}(x) > \ln(m)/\ln(c)=\alpha$. We also
891: notice that $h^*(x)$ has no poles for ${\rm Re}(x) > 0$, and that the
892: poles of $1/(1- mc^{-x})$ are at $x_k = \alpha + 2\pi i k/\ln(c)$
893: where $k=0,\pm 1,\pm 2, \ldots$ runs over all integers. All the poles
894: of $H^*(x)$ are thus to the left of the line ${\rm Re}(x) = \alpha$.
895:
896: The inversion formula for the Mellin transform is given by
897: \begin{equation}
898: H(s) = {1\over 2\pi i}\int_{-i\infty + d}^{i\infty + d} dx \ H^*(x)
899: s^{-x},
900: \end{equation}
901: where the above limits denote an integration up the imaginary axis to
902: the right of all the poles of $H^*$, therefore we chose limits with
903: $d>\alpha$. The contour may be closed in the left half plane (we
904: assume that the integrand vanishes in the region ${\rm Re}(x)\to
905: -\infty$) and we can thus evaluate the inverse Mellin transform in
906: terms of the residues of the poles to the left of ${\rm Re}(x) \leq
907: \alpha$, {\em i.e}
908: \begin{equation}
909: H(s) = \sum_{\rm{poles}} {\rm Res}\left[\frac{h^*(x) s^{-x}}{1-
910: \exp\left (\ln(m)-x\ln(c)\right)}\right]
911: \label{imt1}
912: \end{equation}
913: where $\rm{Res}$ denotes the residue at the pole in question.
914:
915: The large time behavior of $M(t)$ is determined by the small $s$
916: behavior of ${\tilde M}(s)$. Now at small $s$ the dominant behavior
917: clearly comes from the poles $x_k= \alpha + 2\pi i k/\ln(c)$ running
918: up the imaginary axis, any pole coming to the left of this line of
919: poles will be higher order in $s$. We evaluate the residues in
920: Eq. (\ref{imt1}), substitute the resulting $H(s)$ in Eq. (\ref{Ms1})
921: and then take the limit $s\to 0$ to obtain the following asymptotic
922: result
923: \begin{equation}
924: {\tilde M}(s) \sim \frac{1}{ s^{\alpha+1}\ln(c)}\left[ h^*(\alpha) +
925: \sum_{k\neq 0} h^*(\alpha + 2\pi i k/\ln(c))s^{\frac{2\pi i k}{
926: \ln(c)}}\right],\label{eqHs}
927: \end{equation}
928: where we have used the fact $Q(0)=1$. Note that from Eq. (\ref{eqh})
929: and Eq. (\ref{Qdef}), we have
930: \begin{equation}
931: h^*(\alpha)= \int_0^{\infty}
932: \frac{\sigma^{\alpha-1}\,d\sigma}{(1+\sigma)(1+\sigma/c)(1+\sigma/c^2)\ldots}
933: = \frac{\pi}{ \sin(\pi \alpha)} \prod_{k=1}^{\infty} \frac{1-
934: c^{\alpha-k}}{1-c^{-k}},
935: \label{rameq}
936: \end{equation}
937: where the last equality follows from an identity due to
938: Ramanujan~\cite{ram}. Note that this identity explicitly shows that
939: the function $h^*(x)$ has simple poles at the negative integers and
940: zero but no poles for ${\rm Re}(x) >0$ as was stated before.
941:
942: To extract the leading asymptotic behavior of $M(t)$ for large $t$,
943: let us first divide the Laplace transform ${\tilde M}(s)$ into two
944: parts, ${\tilde M}(s)= {\tilde M_p}(s) +{\tilde M_l}(s)$ where
945: ${\tilde M_p}(s)$ denote the first term on the r.h.s. of
946: Eq. (\ref{eqHs}) and ${\tilde M_l}(s)$ corresponds to the remaining
947: sum over $k\ne 0$. Subsequently the inverse Laplace transform $M(t)=
948: M_p(t) + M_l(t)$ can also be divided into two parts. The term
949: ${\tilde M_p}(s)$ has a pure algebraic form, thus its inverse $M_p(t)$
950: has a pure power law growth,
951: \begin{equation}
952: M_p(t) \sim A\ t^{\alpha},
953: \label{Mpt1}
954: \end{equation}
955: where the constant $A$ can be evaluated as follows. If $M_p(t)$ has
956: the form in Eq. (\ref{Mpt1}), its Laplace transform is ${\tilde
957: M_p}(s)= A\, \Gamma(1+\alpha)\, s^{-(1+\alpha)}$. Comparing this with
958: the first term in Eq. (\ref{eqHs}) gives
959: \begin{equation}
960: A = \frac{h^*(\alpha)}{\ln (c) \Gamma(1+\alpha)} = \frac{\pi}{ \ln(m)
961: \Gamma(\alpha) \sin(\pi \alpha)} \prod_{k=1}^{\infty}
962: \frac{1-c^{\alpha-k}}{1-c^{-k}},
963: \label{A1}
964: \end{equation}
965: where we have used $\Gamma(1+\alpha)= \alpha \Gamma(\alpha)$, the
966: definition $\alpha= \ln(m)/\ln(c)$ and the explicit form of
967: $h^*(\alpha)$ from Eq. (\ref{rameq}). Here we note that when $\alpha$
968: is an integer, it can be verified that Eq. (\ref{A1}) agrees with
969: Eq. (\ref{eqmint}) derived for discrete values of $\alpha$ in the
970: previous section.
971:
972: The second contribution to ${\tilde M}(s)$, $\tilde{M}_l(s)$ is given
973: by a Fourier series in $\ln(s)$. The inverse Laplace transform of this
974: term is difficult to obtain fully but it is easy to see that it gives
975: rise to a late time behavior of the form
976: \begin{equation}
977: M_l(t) \sim {A\, t^{\alpha}} g\left(\ln(t)\right)
978: \end{equation}
979: where $g(x)$ is a periodic function of $x$. The final asymptotic
980: result for large $t$ is thus
981: \begin{equation}
982: M(t)=M_p(t)+M_l(t) \simeq {A\, t^{\alpha}}\left[1+
983: g\left(\ln(t)\right)\right]. \label{eqmtb}
984: \end{equation}
985: This exact result thus not only confirms the dominant power-law
986: scaling predicted in section (III) up to log-periodic oscillations,
987: but also provides an explicit formula for the amplitude $A$ as in
988: Eq. (\ref{A1}). For example, let us consider the binary case
989: $m=2$. For the case, when $c=1$, the formula in Eq. (\ref{A1}) gives
990: $A=1$, thus $M(t)\simeq t$ for large $t$. On the other hand, for
991: $m=2$, when $c=\sqrt{2}$ (the critical point), one can show from
992: Eq. (\ref{A1}) that $A=(\sqrt{2}-1)/2$ and $M(t) \simeq (\sqrt{2}-1)
993: t^2 /2$ for large $t$.
994:
995: \subsection{ Solution for the variance $V(t)$}
996:
997: We now examine the asymptotic behavior of the variance $V(t)$ for
998: large $t$ using a similar formalism. The evolution equation
999: (\ref{eqV}) for the variance $V(t)$ is similar to that for the mean
1000: $M(t)$ in Eq. (\ref{eqM}) except that the source term in
1001: Eq. (\ref{eqV}) is $(dM/dt)^2$, different from the source term $1$ in
1002: Eq. (\ref{eqM}). Solution of Eq. (\ref{eqV}) thus requires an
1003: explicit knowledge of how $M(t)$ behaves with time. Taking the
1004: Laplace transform, ${\tilde V}(s)=\int_0^{\infty} V(t) e^{-st}\, dt$
1005: in Eq. (\ref{eqV}) and using $V(0)=0$ we obtain
1006: \begin{equation}
1007: s{\tilde V}(s) = S(s) -{\tilde V}(s) + mc{\tilde V}(cs),
1008: \label{eqS}
1009: \end{equation}
1010: where
1011: \begin{equation}
1012: S(s) = \int_0^\infty dt \exp(-st) \left( \frac{d M}{dt} \right)^2.
1013: \label{ss}
1014: \end{equation}
1015: Rearranging Eq. (\ref{eqS}) gives
1016: \begin{equation}
1017: {\tilde V}(s)= \frac{S(s)}{1+s} + \frac{mc}{1+s}\, {\tilde V}(cs)
1018: \label{vs1}
1019: \end{equation}
1020: which can be iterated to yield
1021: \begin{equation}
1022: {\tilde V}(s) = \sum_{j=0}^{\infty} \frac{(mc)^j}{(1+s)(1+cs)\cdots
1023: (1+c^js)}\, S(c^j s).
1024: \label{vs2}
1025: \end{equation}
1026: Using $m=c^{\alpha}$ and the function $Q(u)$ defined in
1027: Eq. (\ref{Qdef}) we can rewrite Eq. (\ref{vs2}) as
1028: \begin{equation}
1029: {\tilde V}(s) = Q(s/c) H_1(s)
1030: \label{vs3}
1031: \end{equation}
1032: where
1033: \begin{equation}
1034: H_1(s) = \sum_{j=0}^{\infty} \left(c^{1+\alpha}\right)^{j}
1035: \frac{S(c^js)}{Q(c^j s)}.
1036: \label{h1s1}
1037: \end{equation}
1038:
1039: The next step is to take the Mellin transform
1040: $H_1^*(x)=\int_0^{\infty} H_1(s) s^{x-1} ds$ of Eq. (\ref{h1s1}) which
1041: gives, after a change of variable $c^j s\to s$ in the integration
1042: \begin{equation}
1043: H_1^*(x) = \sum_{j=0}^{\infty} \left(c^{1+\alpha-x}\right)^{j}
1044: \int_0^{\infty} \frac{S(s)}{Q(s)} s^{x-1} ds .
1045: \label{h1x1}
1046: \end{equation}
1047: Let us first assume that the integral
1048: \begin{equation}
1049: h_1^*(x) = \int_0^{\infty} \frac{S(s)}{Q(s)} s^{x-1} ds
1050: \label{h1x2}
1051: \end{equation}
1052: exists (the conditions for which will be stated later). Then, for
1053: ${\rm Re}(x)>1+\alpha$, the geometric sum in Eq. (\ref{h1x1})
1054: converges (since $c>1$) and we get
1055: \begin{equation}
1056: H_1^*(x) = \frac{h_1^*(x)}{1- c^{1+\alpha-x}}.
1057: \label{h1x3}
1058: \end{equation}
1059: Inverting this Mellin transform we get
1060: \begin{equation}
1061: H_1(s) = \frac{1}{2\pi i} \int_{-i\infty +d}^{i \infty +d}
1062: \frac{h_1^*(x)}{1- c^{1+\alpha-x}} s^{-x} dx = \sum_{\rm poles} {\rm
1063: Res}\left[ \frac{h_1^*(x)}{1- c^{1+\alpha-x}} s^{-x}\right],
1064: \label{h1s2}
1065: \end{equation}
1066: where the poles are at $x_k= 1+\alpha - 2\pi ik/{\ln(c)}$ with
1067: $k=0,\pm1,\pm2,\ldots$. In Eq. (\ref{h1s2}) the integration is along
1068: the imaginary axis to the right of all the poles and then we close the
1069: contour over the left half plane. Evaluating the residues and
1070: substituting the results in Eq. (\ref{vs3}) we get
1071: \begin{equation}
1072: {\tilde V}(s) = \frac{Q(s/c)}{s^{1+\alpha}
1073: \ln(c)}\left[h_1^*(1+\alpha) + \sum_{k\ne 0} h_1^*\left(1+\alpha-2\pi
1074: ik/\ln(c)\right)\right],
1075: \label{vs4}
1076: \end{equation}
1077: where $h_1^*(x)$ is given by Eq. (\ref{h1x2}) assuming that it exists.
1078:
1079: We now need to invert the Laplace transform in Eq. (\ref{vs4}) to
1080: evaluate $V(t)$. For large $t$, as usual, the dominant contribution
1081: will come from the small $s$ behavior of ${\tilde V}(s)$. Using
1082: $Q(0)=1$ and assuming $h_1^*(1+\alpha-2\pi ik/\ln(c))$ exists for all
1083: $k=0,\pm 1, \pm 2\ldots$, it is clear from the small $s$ behavior of
1084: ${\tilde V}(s)$ in Eq. (\ref{vs4}) that for large $t$
1085: \begin{equation}
1086: V(t) \simeq B'\, t^{\alpha}[1+ G(\ln(t))],
1087: \label{vta1}
1088: \end{equation}
1089: where $G(x)$ is a periodic function in $x$ and the amplitude $B'$ can
1090: be read off as
1091: \begin{equation}
1092: B'= \frac{h_1^*(1+\alpha)}{\Gamma(1+\alpha)\ln(c)};\,\,\, {\rm
1093: where}\,\,\,\, h_1^*(1+\alpha)=\int_0^{\infty} \frac{S(s)}{Q(s)}
1094: s^{\alpha} ds.
1095: \label{b'1}
1096: \end{equation}
1097:
1098: Having obtained the results in Eqs. (\ref{vta1}) and (\ref{b'1}), we
1099: need to investigate when they are valid. These results are valid as
1100: long as the integral $h_1^*(1+\alpha)$ in Eq. (\ref{b'1}) exists. The
1101: existence of this integral depends on the small $s$ behavior of the
1102: source function $S(s)$ defined in Eq. (\ref{ss}). Using the asymptotic
1103: behavior of $M(t)$ from Eq. (\ref{eqmtb}) we find that for large $t$
1104: \begin{equation}
1105: \left(\frac{dM}{dt}\right)^2 \simeq A^2\alpha^2 t^{2\alpha-2} \left[1+
1106: g_1(\ln(t)\right],
1107: \label{dmdt2}
1108: \end{equation}
1109: where $g_1(x)$ is a periodic function in $x$. Substituting this large
1110: $t$ behavior of $(dM/dt)^2$ in Eq. (\ref{ss}), it follows that, in the
1111: case $\alpha < 1/2$, the integral converges to a nonzero constant as
1112: $s\to 0$. On the other hand, for $\alpha>1/2$, the integral diverges
1113: as $S(s)\simeq A^2\alpha^2 \Gamma(2\alpha-1) s^{-(2\alpha-1)}$ as
1114: $s\to 0$. Up to the log-periodic oscillations, the leading behavior of
1115: $S(s)$ for small $s$ can be summarized as follows
1116: \begin{eqnarray}
1117: S(s) &\simeq & C_0 s^{-(2\alpha-1)}\,\,\,\quad {\rm for}\quad\quad
1118: \alpha>1/2
1119: \label{aless} \\
1120: &\simeq & -\ln(s) \,\,\,\quad {\rm for}\quad\quad \alpha=1/2
1121: \label{aequal} \\ &\simeq & A_1 \,\,\,\quad\quad {\rm for}\quad\quad
1122: \alpha<1/2
1123: \label{amore}
1124: \end{eqnarray}
1125: where
1126: \begin{equation}
1127: C_0= A^2\alpha^2 \Gamma(2\alpha-1)
1128: \label{Ceq}
1129: \end{equation}
1130: is a positive constant for $\alpha>1/2$. Also, $A_1= \int_0^{\infty}
1131: (dM/dt)^2 dt$ is a constant that depends on the full form of $M(t)$
1132: and not just on its asymptotic behavior since for $\alpha<1/2$ the
1133: integral is convergent. Substituting this small $s$ behavior of
1134: $S(s)$ into the integral giving $h_1^*(1+\alpha)$ in Eq. (\ref{b'1})
1135: and using $Q(0)=1$, it is clear that the integral exists (no
1136: divergence from the small $s$ limit) only for $\alpha<2$. For
1137: $\alpha>2$, the integral does not exist since the integrand for small
1138: $s$ scales as $s^{1-\alpha}$. Thus the results in Eqs. (\ref{vta1})
1139: and (\ref{b'1}) hold only for $\alpha<2$.
1140:
1141: For $\alpha>2$, the above analysis breaks down and we need to employ a
1142: different method. We now go back to our starting equations (\ref{vs3})
1143: and (\ref{h1s1}). It turns out that for $\alpha>2$, we can actually
1144: extract the leading small $s$ behavior directly from these two
1145: equations. We directly substitute in Eq. (\ref{h1s1}) the leading
1146: small $s$ behavior of $S(s)\simeq C_0 s^{-(2\alpha-1)}$ from
1147: Eq. (\ref{aless}) where $C_0=A^2\alpha^2
1148: \Gamma(2\alpha-1)$. Additionally we use $Q(0)=1$. Eqs. (\ref{vs3}) and
1149: (\ref{h1s1}) then yield in the $s\to 0$ limit
1150: \begin{equation}
1151: {\tilde V}(s) \simeq \frac{C_0}{s^{2\alpha-1}} \sum_{j=0}^{\infty}
1152: \left(c^{2-\alpha}\right)^j = \frac{C_0}{s^{2\alpha-1} (1-
1153: c^{2-\alpha})}
1154: \label{vs5}
1155: \end{equation}
1156: where we have used $\alpha>2$ which ensures that the sum in
1157: Eq. (\ref{vs5}) is convergent. Inverting the Laplace transform, we
1158: then get the large $t$ behavior of $V(t)$ for $\alpha>2$
1159: \begin{equation}
1160: V(t) \simeq B\, t^{2\alpha-2};\,\,\,\, {\rm where}\,\,\,\, B=
1161: \frac{\alpha^2 A^2}{1-c^{2-\alpha}},
1162: \label{vs6}
1163: \end{equation}
1164: where the constant $A$ is given in Eq. (\ref{A1}). Note that this
1165: result in Eq. (\ref{vs6}) for $\alpha>2$ is in perfect agreement with
1166: the self-consistent scaling approach used in Section-III.
1167:
1168: At the critical point $\alpha=2$, the analysis is more
1169: delicate. However, from the scaling approach of Section-III, we
1170: already know that for large $t$, $V(t)\simeq B_c\, t^2 \ln(t)$ with
1171: $B_c= (c-1)^2/\ln (c)$ as in Eq. (\ref{critint2}). Thus, the
1172: asymptotic behavior of the variance $V(t)$ can be summarized as
1173: \begin{eqnarray}
1174: V(t) &\simeq & B'\, t^{\alpha} \,\,\,\quad \rm{for} \quad \alpha <2 \\
1175: &\simeq & B_c\, t^2 \ln(t) \,\,\,\quad \rm{for} \quad
1176: \alpha=\alpha_c=2 \\ &\simeq & B\, t^{2\alpha-2} \,\,\,\,\,\quad
1177: \rm{for} \quad \alpha >2 , \label{varat}
1178: \end{eqnarray}
1179: where the three amplitudes are given by
1180: \begin{eqnarray}
1181: B'&=& \frac{1}{\Gamma(1+\alpha)\ln(c)}\int_0^{\infty}\frac{S(s)}{Q(s)}
1182: s^{\alpha} ds \nonumber \\ B_c &=& (c-1)^2/{\ln (c)} \nonumber \\ B
1183: &=& \frac{\alpha^2 A^2}{1-c^{2-\alpha}} \label{bs}
1184: \end{eqnarray}
1185: where $A$ is given in Eq. (\ref{A1}). Note that computing the
1186: amplitude $B'$ explicitly requires an integration over the full source
1187: function $S(s)$ which is not so easy. Eliminating the time $t$
1188: between $M(t) \simeq A\, t^{\alpha}$ and $V(t)$ in Eq. (\ref{varat}),
1189: one can express the variance $V$ as a function of the mean $M$ for
1190: large $M$ as in Eqs. (\ref{eqvm2}), (\ref{eqvmc}) and (\ref{eqvm1})
1191: and one can read off the constants $C'$, $C_c$ and $C$ in terms of
1192: $B'$, $B_c$ and $B$ and the amplitude $A$ of the mean given in
1193: Eq. (\ref{A1}).
1194:
1195: Let us end this section with a remark on the mathematical mechanism
1196: responsible for the phase transition in the variance of the number of
1197: particles in this Aldous-Shields model. We note that the exact
1198: evolution equations (\ref{eqM}) and (\ref{eqV}) respectively for the
1199: mean and the variance are very similar---they are both linear and
1200: nonlocal in time, the only difference is in the source term. For the
1201: mean $M(t)$ in Eq. (\ref{eqM}), the source term is a constant $1$ (the
1202: first term on the r.h.s of Eq. (\ref{eqM})). On the other hand, for
1203: the variance, the source term $(dM/dt)^2$ in Eq. (\ref{eqV}) depends
1204: on the evolution of the mean. Thus, the mean feeds into the variance
1205: equation as an external source term leading to a competition between
1206: the growth induced by this external source term and the growth induced
1207: internally by the remaining two terms on the r.h.s of
1208: Eq. (\ref{eqV}). This competition between the external and the
1209: internal source is finally responsible for the phase transition in the
1210: asymptotic growth of $V(t)$. For $\alpha<2$, the internal source term
1211: wins out and the variance grows similarly as the mean, $V(t) \sim M(t)
1212: \sim t^{\alpha}$, leading to the normal phase. On the other hand, for
1213: $\alpha>2$, the external source term wins out leading to a faster
1214: growth $V(t)\sim t^{2\alpha-2}\sim [M(t)]^{2-2/\alpha}$ characterizing
1215: anomalously large fluctuations. We note that a similar mechanism
1216: namely a ``{\it competition between the internal source and the external
1217: driving}" was shown to be responsible for phase transitions in
1218: fluctuations in a class of fragmentation problems studied
1219: recently\cite{DM1,Review}. In these fragmentation problems, it was
1220: shown that the mean and the variance evolved via similar looking
1221: equations\cite{DM1,Review}, except that they differed in their
1222: respective source terms---the variance equation had a source term
1223: driven by the mean, in much the same way as in the Aldous-Shields
1224: model discussed here. Thus, it seems that this phase transition in
1225: fluctuations is quite generic as it occurs in a large class of
1226: problems and the mathematical mechanism responsible for it is as
1227: identified above.
1228:
1229:
1230: \section{Summary and Conclusion}
1231:
1232: In this paper we have studied analytically a growing tree model
1233: introduced by Aldous and Shields. In this model, growth occurs in
1234: continuous time. One starts at $t=0$ with an empty Cayley tree with
1235: $m$ branches rooted at $0$ and the tree grows, starting from the root
1236: site, by absorbing particles in continuous time. Each site can occupy
1237: at most one particle. At a given instant $t$, growth can occur only at
1238: the perimeter sites with a rate $c^{-l}$ where $c$ is positive
1239: parameter and $l$ is the distance of the perimeter site from the root
1240: of the tree. For $c=1$ this model is isomorphic to a continuous-time
1241: Eden model on a tree and also corresponds to the random binary search
1242: tree problem in computer science. For $c=2$ this model corresponds to
1243: the digital search tree problem in computer science.
1244:
1245: We have introduced a backward Fokker-Planck approach that enabled us
1246: to study analytically the statistics of the total number of particles
1247: $n(t)$ in the tree at large time $t$. We have shown that at large $t$,
1248: while the mean number of particles grows as a power law in time,
1249: $M(t)\simeq A\, t^{\alpha}$ with $\alpha=\ln(m)/\ln(c)$ for all $c>1$,
1250: the variance $V(t)$ of the number of particles has two different
1251: behaviors depending on the value of the parameter $\alpha$. While for
1252: $\alpha<2$ $V(t)\sim M(t)$ for large $t$, for $\alpha>2$ the variance
1253: grows anomalously quickly: $V(t)\sim [M(t)] ^{2-2/\alpha}$. We have
1254: identified the mathematical mechanism behind this phase transition at
1255: the critical value $\alpha_c=2$ and shown that it is qualitatively
1256: similar to the phase transitions recently encountered in a search tree
1257: problem and also in a related fragmentation problem. Essentially, for
1258: $\alpha<2$, the typical value of $n(t)$ grows in the same way as the
1259: average and the distribution is asymptotically normal whereas for
1260: $\alpha>2$, the typical value does not grow the same way as the average and
1261: the distribution is characterized by large fluctuations caused
1262: by the faster growth of a single branch of the tree.
1263:
1264: We obtained detailed analytical results for the first two moments of
1265: the number of particles for generic values of the parameter
1266: $\alpha$. However, we were able to calculate the full asymptotic
1267: distribution of the number of particles only for two specific values
1268: of $\alpha$, namely for $\alpha=1$ ($c=m$) and $\alpha\to \infty$
1269: ($c=1$). Fortunately these two representative values, where an exact
1270: solution is possible, fall respectively on either side of the critical
1271: point $\alpha_c=2$. Our exact solution shows that for $\alpha=1$
1272: ($<\alpha_c=2$) the distribution $P(n,t)$ is Poisson and hence is
1273: asymptotically normal for large $n$. On the other hand for $\alpha\to
1274: \infty$, the asymptotic distribution is certainly non-Gaussian,
1275: $P(n,t) \sim n^{-\phi} \exp[-n e^{-(m-1)t}]$ where the exponent
1276: $\phi=(m-2)/(m-1)$ depends on $m$. The calculation of the distribution
1277: for other values of $\alpha$ remains a challenging problem.
1278:
1279: While we have studied this growth model on a tree because of its
1280: connections to the search tree problems as mentioned in the
1281: introduction, it is of general interest to study this growth problem
1282: on a regular Euclidean lattice, e.g. on a hyper-cubic lattice in $d$
1283: dimensions. In this lattice model, the cluster will grow similarly
1284: from a seed site at the origin. At a given instant, growth can occur
1285: at any of the available surface sites with a rate $c^{-r}$ where $r$
1286: is the Euclidean distance of the surface site from the origin. One can
1287: then investigate the statistics of the total number of particles in
1288: the cluster after time $t$. It is easy to make a scaling argument for
1289: the growth of the mean number of particles $M(t)$. Assuming that the
1290: cluster is compact with a typical radius $R(t)$ at time $t$, we have
1291: $M(t)\sim [R(t)]^{d}$. Also, the mean number of surface sites
1292: $N_p(t)\sim [R(t)]^{d-1}$. By the growth rule, $dM/dt \sim N_p(t)
1293: c^{-R(t)}$ for large $t$. This predicts $R(t)\sim \ln(t)$ for large
1294: $t$ and hence the mean number of particles grows very slowly as
1295: $M(t)\sim [\ln (t)]^{d}$ for large $t$. An interesting open question
1296: for future studies is whether, in finite dimensional lattice models,
1297: the variance exhibits a phase transition, similar to that seen on the tree,
1298: for some critical value of the parameter $c$?
1299:
1300:
1301: \begin{thebibliography}{99}
1302:
1303: \bibitem{Growth} For a review of growth models, see L.M. Sander, in
1304: {\em Solids Far from Equilibrium}, edited by C. Godr\`eche (Cambridge
1305: University, Cambridge, 1992).
1306:
1307: \bibitem{Eden} M. Eden, in Proceedings of the Fourth Berkeley
1308: Symposium on Mathematical Statistics and Probability, edited by
1309: J. Neyman (University of California Press, Berkeley, 1961), Vol. IV,
1310: p. 223.
1311:
1312: \bibitem{IP} D. Wilkinson and J.F. Willemsen, Invasion percolation: a
1313: new form of percolation theory, J. Phys. A: Math. Gen. {\bf 16},
1314: 3365-3376 (1983).
1315:
1316: \bibitem{DLA} T.A. Witten and L.M. Sander, Diffusion-limited
1317: aggregation, a kinetic critical phenomenon, Phys. Rev. Lett. {\bf 47},
1318: 1400-1403 (1981).
1319:
1320: \bibitem{Networks} R. Albert and A.-L. Barabasi, Statistical mechanics
1321: of complex networks, Rev. Mod. Phys. {\bf 74}, 47-97 (2001).
1322:
1323: \bibitem{Edentree} J. Vannimenus, B. Nickel and V. Hakim, Models of
1324: cluster growth on the Cayley tree, Phys. Rev. B {\bf 30}, 391-399
1325: (1984).
1326:
1327: \bibitem{Knuth} D.E. Knuth, {\it The Art of Computer Programming,
1328: Sorting and Searching}, 2nd ed. (Addison-Wesley, Reading, MA, 1998),
1329: vol. 3.
1330:
1331: \bibitem{Mahmoud} H. Mahmoud, {\em Evolution of Random Search Trees}
1332: (Wiley, New York, 1992).
1333:
1334: \bibitem{Review} For a recent review on search trees for physicists,
1335: see S.N. Majumdar, D.S. Dean and P.L. Krapivsky, Understanding search
1336: trees via statistical physics, in {\em Proceedings of the STATPHYS 22}
1337: (Bangalore, India, 2004), Pramana-J. Phys. {\bf 64}, 1175-1189 (2005)
1338: (also available at http://xxx.arXiv.org/cond-mat/0410498); see also
1339: S.N. Majumdar and P.L. Krapivsky, Extreme value statistics and
1340: traveling fronts: application to computer science, Phys. Rev. E {\bf
1341: 65}, 036127 (2002).
1342:
1343: \bibitem{Sedgewick} R. Sedgewick, {\em Algorithms}, 2nd ed.
1344: (Addison-Wesley, Reading, MA, 1988), p 245.
1345:
1346: \bibitem{FS} P. Flajolet and R. Sedgewick, Digital search trees
1347: revisited, SIAM J. Comput. {\bf 15}, 748-767 (1986).
1348:
1349: \bibitem{Pittel} B. Pittel, Asymptotic growth of a class of random
1350: trees, Ann. Probab. {\bf 13}, 414-427 (1985); Paths in a random
1351: digital search tree: limiting distributions, Adv. Appl. Probab. {\bf
1352: 18}, 139-155 (1986).
1353:
1354: \bibitem{FR} P. Flajolet and B. Richmond, Generalized digital trees
1355: and their difference--differential equations Random
1356: Struct. Algorithms, {\bf 3}, 305-320 (1992).
1357:
1358: \bibitem{KS} For a recent review see C. Knessl and W. Szpankowski,
1359: Asymptotic behavior of the height in a digital search tree and the
1360: longest phrase of the Lempel-Ziv scheme, SIAM J. Comput. {\bf 30},
1361: 923-964 (2000).
1362:
1363: \bibitem{SM1} S.N. Majumdar, Traveling front solutions to directed
1364: diffusion limited aggregation, digital search trees, and the
1365: Lempel-Ziv data compression algorithm, Phys. Rev. E {\bf 68}, 026103
1366: (2003).
1367:
1368: \bibitem{AS} D. Aldous and P. Shields, A diffusion limit for a class
1369: of randomly-growing binary trees, Probab. Theory Relat. Fields {\bf
1370: 79}, 509-542 (1988).
1371:
1372: \bibitem{LZ} J. Ziv and A. Lempel, A universal algorithm for
1373: sequential data compression, IEEE Trans. Inf. Theory {\bf 23}, 337-343
1374: (1977).
1375:
1376: \bibitem{BS} R.M. Bradley and P.N. Strenski, Directed aggregation on
1377: the Bethe lattice: scaling, mappings, and universality, Phys. Rev. B
1378: {\bf 31}, 4319-4328 (1985).
1379:
1380: \bibitem{CH} H-H Chern and H-K Hwang, Phase changes in random m-ary
1381: search trees and generalized quicksort, Random. Struct. Algorithms
1382: {\bf 19}, 316-358 (2001).
1383:
1384: \bibitem{DM1} D.S. Dean and S.N. Majumdar, Phase transition in a
1385: random fragmentation problem with applications to computer science,
1386: J. Phys. A: Math. Gen. {\bf 35}, L501-L507 (2002).
1387:
1388: \bibitem{CP} B. Chauvin and N. Pouyanne, m-ary search trees when $m <=
1389: 27$: A strong asymptotics for the space requirements, Random
1390: Struct. Algorithms {\bf 24}, 133-154 (2004).
1391:
1392: \bibitem{FN} J.A. Fill and N. Kapur, Transfer theorems and asymptotic
1393: distributional results for m-ary search trees, Random Struct.
1394: Algorithms {\bf 26}, 359-391 (2005).
1395:
1396: \bibitem{FFN} J.A. Fill, P. Flajolet, and N. Kapur, Singularity
1397: analysis, Hadamard products and tree recurrences,
1398: J. Comput. Appl. Math. {\bf 174}(2): 271-313 (2005).
1399:
1400: \bibitem{TH} M. Ghorbel and T. Huillet, Fragment size distributions in
1401: random fragmentations with cut-off, Stat. Proba. Lett. {\bf 71},
1402: 47-60 (2005); T. Huillet, Statistical aspects of random fragmentation,
1403: J. Comput. Appl. Math. {\bf 181} (2), 364-387 (2005).
1404:
1405: \bibitem{BG} see for example, J.-P. Bouchaud and A. Georges,
1406: Anomalous diffusion in disordered media, Phys. Rep. {\bf 195}, 127-293 (1990).
1407:
1408: \bibitem{ram} G.H. Hardy, {\it Ramanujan: Twelve Lectures on Subjects
1409: Suggested by his Life and Work}, (Chelsea Publishing Company, New
1410: York, 1978).
1411:
1412: \end{thebibliography}
1413:
1414: \end{document}
1415: